EURASIP Journal on Wireless Communications and Networking

Satellite Communications

Guest Editors: Ray E. Sheriff, Anton Donner,

and Alessandro Vanelli-Coralli
Satellite Communications
EURASIP Journal on
Wireless Communications and Networking

Satellite Communications
Guest Editors: Ray E. Sheriff, Anton Donner,
and Alessandro Vanelli-Coralli
Copyright © 2007 Hindawi Publishing Corporation. All rights reserved.

This is a special issue published in volume 2007 of “EURASIP Journal on Wireless Communications and Networking.” All articles are
open access articles distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and
reproduction in any medium, provided the original work is properly cited.
Editor-in-Chief
Luc Vandendorpe, Université Catholique de Louvain, Belgium
Associate Editors
Thushara Abhayapala, Australia
Mohamed H. Ahmed, Canada
Farid Ahmed, USA
Alagan Anpalagan, Canada
Anthony Boucouvalas, Greece
Lin Cai, Canada
Biao Chen, USA
Yuh-Shyan Chen, Taiwan
Pascal Chevalier, France
Chia-Chin Chong, South Korea
Huaiyu Dai, USA
Soura Dasgupta, USA
Ibrahim Develi, Turkey
Petar M. Djurić, USA
Mischa Dohler, France
Abraham O. Fapojuwo, Canada
Michael Gastpar, USA
Alex Gershman, Germany
Wolfgang Gerstacker, Germany
David Gesbert, France
Fary Z. Ghassemlooy, UK
Christian Hartmann, Germany
Stefan Kaiser, Germany
G. K. Karagiannidis, Greece
Chi Chung Ko, Singapore
Visa Koivunen, Finland
Richard Kozick, USA
Bhaskar Krishnamachari, USA
S. Lambotharan, UK
Vincent Lau, Hong Kong
David I. Laurenson, UK
Tho Le-Ngoc, Canada
Wei Li, USA
Yonghui Li, Australia
Tongtong Li, USA
Zhiqiang Liu, USA
Stephen McLaughlin, Scotland
Sudip Misra, Canada
Marc Moonen, Belgium
Eric Moulines, France
Sayandev Mukherjee, USA
Kameswara Rao Namuduri, USA
Amiya Nayak, Canada
A. Pandharipande, The Netherlands
Athina Petropulu, USA
A. Lee Swindlehurst, USA
Sergios Theodoridis, Greece
George S. Tombras, Greece
Lang Tong, USA
Athanasios V. Vasilakos, Greece
Weidong Xiang, USA
Yang Xiao, USA
Xueshi Yang, USA
Lawrence Yeung, Hong Kong
Dongmei Zhao, Canada
Weihua Zhuang, Canada
Contents
Satellite Communications, Ray E. Sheriff, Anton Donner, and Alessandro Vanelli-Coralli
Volume 2007, Article ID 58964, 2 pages

Multi-Satellite MIMO Communications at Ku-Band and Above: Investigations on Spatial Multiplexing

for Capacity Improvement and Selection Diversity for Interference Mitigation, Konstantinos P. Liolis,
Athanasios D. Panagopoulos, and Panayotis G. Cottis
Volume 2007, Article ID 59608, 11 pages

Investigations in Satellite MIMO Channel Modeling: Accent on Polarization, Péter Horváth,

George K. Karagiannidis, Peter R. King, Stavros Stavrou, and István Frigyes
Volume 2007, Article ID 98942, 10 pages

Performance Analysis of SSC Diversity Receivers over Correlated Ricean Fading Satellite Channels,
Petros S. Bithas and P. Takis Mathiopoulos
Volume 2007, Article ID 25361, 9 pages

Advanced Fade Countermeasures for DVB-S2 Systems in Railway Scenarios, Stefano Cioni,
Cristina Párraga Niebla, Gonzalo Seco Granados, Sandro Scalise, Alessandro Vanelli-Coralli,
and Marı́a Angeles Vázquez Castro
Volume 2007, Article ID 49718, 17 pages

Capacity Versus Bit Error Rate Trade-Off in the DVB-S2 Forward Link, Matteo Berioli,
Christian Kissling, and Rémi Lapeyre
Volume 2007, Article ID 14798, 10 pages

Frequency Estimation in Iterative Interference Cancellation Applied to Multibeam Satellite Systems,

J. P. Millerioux, M. L. Boucheret, C. Bazile, and A. Ducasse
Volume 2007, Article ID 62310, 12 pages

A QoS Architecture for DVB-RCS Next Generation Satellite Networks, Thierry Gayraud and
Pascal Berthou
Volume 2007, Article ID 58484, 9 pages

Maximum Likelihood Timing and Carrier Synchronization in Burst-Mode Satellite Transmissions,

Michele Morelli and Antonio A. D’Amico
Volume 2007, Article ID 65058, 8 pages

Burst Format Design for Optimum Joint Estimation of Doppler-Shift and Doppler-Rate in Packet
Satellite Communications, Luca Giugno, Francesca Zanier, and Marco Luise
Volume 2007, Article ID 29086, 12 pages

TCP-Call Admission Control Interaction in Multiplatform Space Architectures, Georgios Theodoridis,

Cesare Roseti, Niovi Pavlidou, and Michele Luglio
Volume 2007, Article ID 23923, 8 pages

Efficient Delay Tracking Methods with Sidelobes Cancellation for BOC-Modulated Signals,
Adina Burian, Elena Simona Lohan, and Markku Kalevi Renfors
Volume 2007, Article ID 72626, 20 pages
Analysis of Filter-Bank-Based Methods for Fast Serial Acquisition of BOC-Modulated Signals,
Elena Simona Lohan
Volume 2007, Article ID 25178, 12 pages
Hindawi Publishing Corporation
EURASIP Journal on Wireless Communications and Networking
Volume 2007, Article ID 58964, 2 pages
doi:10.1155/2007/58964

Editorial
Satellite Communications

Ray E. Sheriff,1 Anton Donner,2 and Alessandro Vanelli-Coralli3

1 Mobileand Satellite Communications Research Centre, School of Engineering, Design and Technology, University of Bradford,
Richmond Road Bradford BD7 1DP, UK
2 German Aerospace Center, Institute of Communications and Navigation, Oberpfaffenhofen, 82234 Wessling, Germany
3 ARCES, University of Bologna, Via Toffano 2, 40125 Bologna, Italy

Received 28 November 2007; Accepted 9 December 2007

Copyright © 2007 Ray E. Sheriff et al. This is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

We are delighted to bring to you this special issue on satel-
lite communications, which we have prepared as part of the
spreading of excellence remit of the satellite communica-
tions network of excellence (SatNEx). The SatNEx project,
which began in 2004, is funded for five years under the Euro-
pean Union’s Sixth Framework Programme (FP6) Informa-
tion Society Technologies (IST) Thematic Area. Led by the
German Aerospace Center, SatNEx brings together a network
of 24 partners, distributed throughout Europe, with mem-
bership drawn from ten countries.
The philosophy underlying the SatNEx approach re-
volves around the selection of focused actions under Joint
Programmes of Activities, which are carried out collectively
by the partners and include research, integration, and dis-
semination activities. Training represents an important part
of the SatNEx remit and is supported through a number of
initiatives including the hosting of internship projects and an
annual summer school.
The call for papers resulted in a high number of submis-
sions, from which we have been able to select 12 excellent
papers dealing with the different aspects of satellite commu-
nications and navigation.
Multiple-input multiple-output (MIMO) techniques are
attracting a considerable amount of attention from within
the terrestrial wireless community. The first paper of this spe-
cial issue, “Multisatellite MIMO communications at Ku band
and above: investigations on spatial multiplexing for capac-
ity improvement and selection diversity for interference mit-
igation,” considers the application of such technology over a
satellite platform operating in the Ku band and above. The
paper considers how MIMO can be used to increase capac-
ity by using a satellite spatial multiplexing system and how
antenna selection can be used to mitigate interference. The
next paper “Investigations in satellite MIMO channel model-
ing: accent on polarization” looks at MIMO systems from the
polarization diversity point of view and dwells on the satellite
cooperative communication concepts.
Switch and stay combining (SSC) is a form of diversity
technique used in digital receivers to compensate for fade
events introduced by the mobile channel. The third paper
“Performance analysis of SSC diversity receivers over corre-
lated Ricean fading satellite channels” investigates the per-
formance of dual-branch SSC receivers for different fading
channel characteristics.
The next four papers deal with the emerging scenario
of mobile digital video broadcasting (DVB-S2 and RCS mo-
bile). Alternative approaches to counteracting fading chan-
nels introduced when operating in a train environment re-
ceiving satellite DVB-S2 are presented in the paper “Ad-
vanced fade countermeasures for DVB-S2 systems in railway
scenarios.” Here, as a result of simulation analysis, antenna
diversity and packet-level forward error correction mecha-
nisms are proposed and their impact is evaluated with respect
to the receiver design and system complexity. The theme of
DVB-S2 is continued with the paper “Capacity versus bit er-
ror rate trade-off in the DVB-S2 forward link,” which inves-
tigates how satellite capacity can be optimised for DVB-S2
transmissions. The DVB return channel via satellite (DVB-
RCS) is then addressed in “Frequency estimation in iterative
interference cancellation applied to multibeam satellite sys-
tems,” which considers the application of interference cancel-
lation on the reverse link of a multibeam satellite system, us-
ing DVB-RCS with convolutional coding as an example. The
paper “A QoS architecture for DVB-RCS next-generation
satellite networks” proceeds to design and emulate a quality-
of-service (QoS) architecture that demonstrates using real
multimedia applications how QoS can be supported over a
DVB-RCS network.
2 EURASIP Journal on Wireless Communications and Networking

Synchronization aspects are dealt with in “Maximum

likelihood timing and carrier synchronization in burst-mode
satellite transmissions.” The paper addresses the problem of
achieving synchronisation for a burst-mode satellite trans-
mission over an AWGN channel. The subject of burst trans-
mission continues with the paper “Burst format design for
optimum joint estimation of Doppler-shift and Doppler-
rate in packet satellite communications,” which considers
optimising the burst-format of packet-oriented transmis-
sions by proposing very-low-complexity algorithms for car-
rier Doppler-shift and Doppler-rate estimation.
A network comprising satellite and high-altitude plat-
forms is considered in the paper “TCP-call admission con-
trol interaction in multiplatform space architectures.” Cross-
layer techniques are implemented by means of TCP feeding
back into call admission control (CAC) procedures for the
purpose of prevention of congestion and improvement in
QoS.
Finally, since navigation is an extremely important part
of the satellite system family, we have included two papers.
The first paper “Efficient delay tracking methods with side-
lobes cancellation for BOC-modulated signals” deals with bi-
nary offset carrier (BOC) modulation, which is adopted in
typical navigation systems. The paper considers how to im-
prove the tracking of the main lobe of the BOC-modulated
signal by using sidelobe suppression techniques. An alterna-
tive approach based on filter bank processing is presented in
“Analysis of filter-bank-based methods for fast serial acqui-
sition of BOC-modulated signals” to conclude the special is-
sue.

ACKNOWLEDGMENTS
It has been a pleasure for us to have put together this spe-
cial issue, which we hope you will find interesting. We would
like to thank the editorial staff at Hindawi for their sup-
port and assistance during the preparation of this special is-
sue. We would like to thank the contributing authors for the
excellent quality of their submissions and our SatNEx col-
leagues for their valuable assistance in the reviewing of pa-
pers. SatNEx is partially funded by the European Commis-
sion under the Sixth Framework Programme. Further in-
formation on SatNEx can be found on the project web site:
http://www.satnex.org/.

Ray E. Sheriff
Anton Donner
Alessandro Vanelli-Coralli
Hindawi Publishing Corporation
EURASIP Journal on Wireless Communications and Networking
Volume 2007, Article ID 59608, 11 pages
doi:10.1155/2007/59608

Research Article
Multi-Satellite MIMO Communications at Ku-Band and
Above: Investigations on Spatial Multiplexing for Capacity
Improvement and Selection Diversity for
Interference Mitigation

Konstantinos P. Liolis, Athanasios D. Panagopoulos, and Panayotis G. Cottis

Wireless & Satellite Communications Group, School of Electrical and Computer Engineering, National Technical University of Athens
(NTUA), 9 Iroon Polytechniou Street, Zografou, Athens 15780, Greece
Received 28 August 2006; Revised 2 March 2007; Accepted 13 May 2007

Recommended by Alessandro Vanelli-Coralli

This paper investigates the applicability of multiple-input multiple-output (MIMO) technology to satellite communications at the
Ku-band and above. After introducing the possible diversity sources to form a MIMO matrix channel in a satellite environment,
particular emphasis is put on satellite diversity. Two specific different topics from the field of MIMO technology applications to
satellite communications at these frequencies are further analyzed: (i) capacity improvement achieved by MIMO spatial multi-
plexing systems and (ii) interference mitigation achieved by MIMO diversity systems employing receive antenna selection. In the
first case, a single-user capacity analysis of a satellite 2 × 2 MIMO spatial multiplexing system is presented and a useful analytical
closed form expression is derived for the outage capacity achieved. In the second case, a satellite 2 × 2 MIMO diversity system with
receive antenna selection is considered, adjacent satellite cochannel interference on its forward link is studied and an analytical
model predicting the interference mitigation achieved is presented. In both cases, an appropriate physical MIMO channel model is
assumed which takes into account the propagation phenomena related to the frequencies of interest, such as clear line-of-sight op-
eration, high antenna directivity, the effect of rain fading, and the slant path lengths difference. Useful numerical results obtained
through the analytical expressions derived are presented to compare the performance of multi-satellite MIMO systems to relevant
single-input single-output (SISO) ones.

Copyright © 2007 Konstantinos P. Liolis et al. This is an open access article distributed under the Creative Commons Attribution
License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly
cited.

1. INTRODUCTION MIMO both as a research topic and as a commercially viable

Multiple-input multiple-output (MIMO) technology has re-
cently emerged as one of the most significant technical
breakthroughs in modern digital communications due to its
promise of very high data rates at no cost of extra spectrum
and transmit power [1, 2]. Wireless communication can be
benefited from MIMO signaling in two different ways: spa-
tial multiplexing and diversity. In the former case, indepen-
dent data is transmitted from separate antennas, and aiming
at maximizing throughput (i.e., linear capacity growth with
the number of antennas can be achieved). In the latter case,
the same signal is transmitted along multiple (ideally) inde-
pendently fading paths aiming at improving the robustness
of the link in terms of each user BER performance. These
advantages have been largely responsible for the success of
technology in terrestrial communications [1, 2].
The appealing gains obtained by MIMO techniques in
terrestrial networks generate a further interest in investigat-
ing the possibility of applying the same principle in satel-
lite networks, as well. However, the underlying differences
between the terrestrial and the satellite channels make such
applicability a non straightforward matter and, therefore, a
rather challenging subject. In this case, one of the funda-
mental problems is the difficulty of generating a completely
independent fading profile over the space segment. In satel-
lite communications, due to the huge free space losses along
the earth-space link, line-of-sight (LOS) operation is usually
deemed a practical necessity. However, this is not the typ-
ical case in terrestrial communications where rich scatter-
ing and non-LOS environments with multipath propagation
2 EURASIP Journal on Wireless Communications and Networking
are encountered. Thus, placing multiple antennas on a sin-
gle satellite does not seem a suitable choice in order to ex-
ploit the MIMO channel capabilities. In fact, the absence of
scatterers in the vicinity of the satellite leads to an inherent
rank deficiency of the MIMO channel matrix. Therefore, at a
first glance, the applicability of MIMO technology to satellite
channels does not seem well justified.
The objective of this paper is in line with some other re-
cent research efforts [4–8, 12–16] casting further light in this
regard. These studies have been mainly concerned with the
possible diversity sources that can be exploited in satellite
communications to form a MIMO matrix channel. A cate-
gorization of these diversity sources follows.
(i) Site diversity, where multiple cooperating terminal
stations (TSs), sufficiently separated from each other, are in
communication with a single satellite. So far, it has only been
studied as an efficient rain fade mitigation technique at the
Ku (12/14 GHz), Ka (20/30 GHz), and Q/V (40/50 GHz) fre-
quency bands because of its very low achievable spatial cor-
relation due to rain [3]. However, due to the enormous slant
path lengths associated, the required separation distance be-
tween the multiple TSs to ensure ideally independent fading
profile is of the order of several km, which rather hinders its
practical interest in MIMO applications.
(ii) Satellite (or orbital) diversity, where multiple satel-
lites, sufficiently separated in orbit to provide (ideally) in-
dependently fading channels, communicate with a single TS
equipped with either multiple antennas or even a single mul-
tiple-input antenna. So far, it has been studied mostly as an
efficient rain fade mitigation technique in Ku-, Ka-, and Q/V-
band satellite communications [3] and, also, recently, as a
candidate to form satellite MIMO matrix channels at high
(i.e., Ku, Ka, and Q/V) [4, 5] as well as at low frequency
bands, such as L (1/2 GHz) and S (2/4 GHz) [6–8]. Also, it
is worthwhile noting that it is already successfully employed
in the continental US digital audio radio services (DARS),
mobile systems, Sirius and XM satellite radio, operating at
the S-band [9]. Satellite diversity provides a rather practical
solution of reasonable complexity since the multiple received
signals at the single TS can easily be combined due to the
colocation of the antennas. However, an inherent problem
of this scheme, apart from the costly utilization of multiple
satellites, is the asynchronism of the multiple transmitted sig-
nals at the TS receiver, which comes as a result of the prop-
agation delay difference due to the wide separation between
the satellites. A similar problem is dealt with and solutions
are proposed in several papers mainly concerning distributed
sensor networks, such as in [10]. To the authors’ knowledge,
for the more complicated satellite case—due to the much
larger and variable delay difference—the only relevant solu-
tion proposed so far is reported in [5].
(iii) Polarization diversity, where a single dual-orthogonal
polarized satellite communicates with a single TS equipped
with a dual-orthogonal polarized antenna. Its principle is
based on the polarization sensitivity of the reflection and
diffraction processes, which causes random signal fading at
the TS receiver. It represents a solution of rather practical
interest due to the recent developments in MIMO compact
antennas (see, e.g., [11]) which allow for compact MIMO
setups. It has already been examined as a promising solu-
tion to shape MIMO channels in S-band land mobile satellite
communications [7, 12–16]. Its main advantage over satellite
diversity is the elimination of any additional cost associated
with the utilization of multiple satellites. It also bypasses the
asynchronism problem associated with the distributed na-
ture of satellite diversity. However, it can be disadvantageous
to satellite diversity especially in satellite networks operating
at high-frequency bands (i.e., Ku, Ka, and Q/V), which are
affected by the highly correlated rainfall medium and, also,
in case of large blockages resulting in hard system failures
(i.e., on/off channel phenomena). Moreover, as concluded in
[13], polarization diversity can only increase the transmis-
sion rate of a satellite communication system by a factor of
two, whereas in multi-satellite systems, satellite diversity can
result in m-fold capacity increase, where m is the number of
satellites occupied.
This paper focuses particularly on dual-satellite MIMO
communication systems employing satellite diversity. More-
over, emphasis is put on the less congested high-frequency
bands, such as Ku and above. At these frequencies, multi-
path propagation is insignificant. However, by virtue of satel-
lite diversity, MIMO can be considered to effectively exploit
the rainfall spatial inhomogeneity instead. A physical 2 × 2
MIMO satellite channel model is assumed taking into ac-
count the relevant propagation phenomena, such as clear
LOS operation, high antenna directivity, rain fading, and
rainfall spatial inhomogeneity [3, 17]. This model is flexi-
ble and can be applied on a global scale since it has physical
inputs obtained by regression fitting analysis on the ITU-R
rainmaps [18] and is based on general assumptions about
the rain process [17]. Moreover, it incorporates the general
case of an ordered MIMO satellite channel (due to the slant
path lengths difference). To this end, the resulting propaga-
tion delay offset is assumed to be properly taken into account
at the TS receiver. A possible practical solution to this prob-
lem might be the one implemented in [5] according to which
matched filters are first applied to the received signals for the
detection of the propagation delay offset, which is then fed to
a timing aligner. Subsequently, the proposed timing aligner
eliminates the delay offset by adjusting the timing of a signal
parallel-to-serial converter. The study of more efficient solu-
tions to the asynchronism problem associated with satellite
diversity, although rather challenging, is out of the scope of
this paper and will be the subject of a future work.
In the first part of this work, emphasis is put on a satellite
2 × 2 MIMO spatial multiplexing system and on its possi-
ble capacity improvement with respect to the relevant SISO
system. The term “spatial multiplexing” refers to the trans-
mission of independent data streams from the multiple sep-
arate satellites [1, 2]. Well-known results obtained from the
MIMO literature [19, 20] are applied here for the capacity
analysis of such a 2 × 2 MIMO system. The figure of merit
used to characterize the resulting MIMO fading channel is
the outage capacity [1], for which an analytical closed form
expression is provided. Note that such analytical expressions
are extremely hard to obtain even in the well-established field
Konstantinos P. Liolis et al.
S1
d1 , AR1
Δθ
TS
(a)
Figure 1: (a) Configuration of a dual-satellite 2 × 2 MIMO channel. Individual satellites S1 and S2 transmit either independent data streams
(MIMO spatial multiplexing system, Section 3) or the same signal over the multiple (ideally) independently fading paths (MIMO diversity
system, Section 4), (b) associated elevation angles.
of MIMO theory due to the intractability of the outage ca-
pacity distribution [2].
In the second part, a satellite 2 × 2 MIMO diversity sys-
tem employing receive antenna selection is examined, and
issues specifically related to cochannel interference (CCI) are
addressed from a propagation point of view. The term “di-
versity” refers to the transmission of the same signal over the
multiple (ideally) independently fading paths [1, 2]. Receive
antenna selection is a low-cost, low-complexity approach to
benefit from many of the advantages of MIMO technology
while, at the same time, bypassing the multiple RF chains
associated with multiple antennas at the receiver, which are
costly in terms of size, power, and hardware [21]. The inter-
ference analysis presented here is quite different from con-
ventional communication-oriented approaches followed in
standard MIMO theory [1]. Attention is paid to the CCI
problems arising on the forward link of such a 2 × 2 MIMO
satellite system due to differential rain attenuation from an
adjacent satellite [22]. To deal with the statistical behaviour
of the signal-to-interference ratio (SIR) introduced by the
rainfall spatial inhomogeneity, the concept of unacceptable
interference probability1 [23, 24] is employed here. An ana-
lytical prediction model concerning the interference mitiga-
tion achieved by the proposed satellite 2 × 2 MIMO diversity
system is provided.
The rest of the paper is organized as follows. Section 2
presents the channel model adopted for MIMO satellite com-
munications at the Ku-band and above. Section 3 provides a
communication-based capacity analysis for a satellite 2 × 2
MIMO spatial multiplexing system. A propagation-oriented
1 Note that the concept of the “unacceptable interference probability
(UIP)” in this paper is exactly the same as that of the “acceptable interfer-
ence probability (AIP)” employed in [23, 24]. Their only difference con-
cerns their nomenclature.
3
S2
To S1 To S2
d2 , AR2
ϕ2
ϕ1
TS
(b)
analysis for the possible interference mitigation achieved by a
satellite 2 × 2 MIMO diversity system with receive antenna se-
lection is presented in Section 4. Useful numerical results ob-
tained for both the above satellite MIMO applications con-
sidered are provided in Section 5. Section 6 concludes the
paper.
2. MIMO SATELLITE CHANNEL MODEL
Figure 1 depicts the configuration of a dual-satellite MIMO
communication channel at the Ku-band and above. The TS
is equipped with two colocated highly directive antennas and
communicates with two satellites, S1 and S2 , subtending an
angle Δθ to the TS, large enough that the spatial correlation
due to rain along the relevant slant paths is as low as possible.
The normalized radiation pattern of each TS antenna, de-
noted by GR (·), is compatible with the ITU-R specifications
[25] and is shown in Figure 2.2 The lengths of slant paths
Si -TS are denoted by di (i = 1, 2) and the random variables
(RVs) associated with the respective rain induced attenua-
tions (in dB) are denoted by ARi (i = 1, 2). In general, the
two slant paths Si -TS have different elevation angles denoted
by φi (i = 1, 2), respectively.
Assuming that clear LOS between the TS and each satel-
lite Si exists, that each TS antenna is at boresight with the
corresponding satellite Si (i = 1, 2) and that rain attenuation
is the major fading mechanism, the path gain for each Si -TS
link is modeled as
 
gi ∝ GR 0◦ · di−2 · 10−AR i /10 (i = 1, 2). (1)
2 Note that the analyses presented hereafter are quite general and, therefore,
may incorporate other TS antenna radiation patterns, as well.
4 EURASIP Journal on Wireless Communications and Networking

0 1

Spatial correlation coefficient due to rain, ρ12

−5 0.9
TS antenna normalized gain, GR (dB)

−10
0.8
−15
0.7
−20
0.6
−25
0.5
−30

−35 0.4

−40 0.3
−100 −80 −60 −40 −20 0 20 40 60 80 100 0 20 40 60 80 100 120 140 160 180
Off-axis angle (deg) Angular separation, Δθ (deg)

Figure 2: Normalized radiation pattern of each TS antenna com-
patible with ITU-R specifications [25].
Hence, the total path loss along each Si -TS link (in dB) is
Figure 3: Spatial correlation coefficient due to rain ρ12 versus an-
gular separation Δθ for a dual-satellite MIMO channel operating in
Atlanta, GA, at the Ka-band with satellite elevation angles φ1 = 45◦
and φ2 = 40◦ .

Ai = FSLi + ARi (i = 1, 2), (2)

where FSLi = 10 log10 (4πdi is the free space loss along
each link, c the speed of light, and f the operating fre-
quency. Note that the fundamental assumptions concerning
the modeling of the rain attenuation RVs ARi (i = 1, 2) are
the same as those analytically presented in [17]. The convec-
tive raincell model employing Crane’s assumptions is used
for the description of the vertical variation of the rainfall
structure [17]. Based on this assumption, if Δθ is sufficiently
large, the spatial correlation coefficient between the RVs ARi
is relatively low and, thus, an (ideally) decorrelated MIMO
satellite channel is possible. To this end, an illustrative quan-
titative example is presented in Figure 3, which depicts the
spatial correlation coefficient due to rain ρ12 versus Δθ for a
dual-satellite MIMO channel operating in Atlanta, GA, USA
at the Ka-band with satellite elevation angles φ1 = 45◦ and
φ2 = 40◦ .
Based on the above and, also, assuming frequency nonse-
lective fading, the resulting MIMO channel matrix H is given
by
 
h h
(i = 1, 2). Finally, the assumption of independent identically
f /c)2 distributed (i.i.d) elements of H, often made in conventional
terrestrial MIMO systems, cannot be made here, since there
is a relatively high spatial correlation due to rain.
3. SATELLITE MIMO SPATIAL MULTIPLEXING SYSTEM:
CAPACITY ANALYSIS
In this Section, the two satellites Si (i = 1, 2) depicted in
Figure 1 are assumed to transmit different and independent
data streams (i.e., spatial multiplexing is investigated). The
channel H is considered perfectly known to the TS receiver
(via training and tracking), while at the transmit side, both
satellites are assumed to have no channel knowledge. In the
absence of channel state information (CSI) at the transmit
side, equal power allocation to the two satellites is a reason-
able and rather practical choice, due to the distributed na-
ture of the system. Therefore, from the standard MIMO the-
ory, the following well-known formula for the capacity (in
bps/Hz) of MIMO channels is adopted [19, 20]:

 2

H = 11 12 PT P
h21 h22 C = log2 det I2 + HHH = log2 1 + T λi ,
⎡
⎤ 2N0 i=1
2N 0
√ j2πd1 f (3)
⎢ g1 exp 0 ⎥ (4)
=⎢ c
⎥
⎣ √ j2πd2 f ⎦.
0 g2 exp where I2 is the 2 × 2 identity matrix, PT the total average
c
power available at the transmit side,3 N0 the noise spectral
The diagonal structure of H is due to the high directivity of
the TS antennas and the large value of Δθ. In MIMO ter-
3 Note that PT is the sum transmit power of all transmitting satellites Si re-
minology, channels with diagonal H matrix are known as
parallel MIMO channels. Further details about such chan- gardless of their number. This means that in both the dual-satellite MIMO
case and the single satellite SISO case, the total available transmit power
nels can be found in [26]. Moreover, as opposed to standard is constant and equal to PT . This is ensured employing the normalization
MIMO theory [1, 2], H is not normalized here (i.e., ordered factor “2” in (4), which allows for a fair comparison between the relevant
MIMO channel) due to the different slant path lengths di MIMO and SISO cases.
Konstantinos P. Liolis et al. 5

density at the TS receiver input, and λi (i = 1, 2) the positive The quantities AmRi , SaRi (i = 1, 2), encountered in (8)–(11),
eigenvalues of the matrix HHH (the superscript H stands for are the statistical parameters of the lognormal RVs ARi (i =
conjugate transposition). 1, 2) given by [17]
Taking into account the channel modeling assumptions,  
Hi    
(4) is written as S2aRi = ln 1 + exp b2 S2r − 1 (i = 1, 2),
L2Di
2 2 (12)

b Sr − S2aRi

2
 
C= log2 1 + 0.5SNRCSi 10−ARi /10 , (5) AmRi = aRbm LDi exp (i = 1, 2),
i=1 2
where SNRCSi (i = 1, 2) are the nominal SNR values under where LDi (i = 1, 2) are the projections of the effective path
clear sky conditions. Based on the path gain model given in lengths Li (i = 1, 2) [17] on the earth surface, Hi (i = 1, 2) are
(1), the SNRCSi values (in dB) are related through spatial parameters related to each path of length LDi (i = 1, 2)

which may be found in [17], and a, b are constants depend-
d2
SNRCS1 − SNRCS2 = 20 log10 . (6) ing on the operating frequency f , the polarization tilt angle,
d1 the temperature, and the rainfall characteristics over the ser-
Equation (5) provides an expression for the instantaneous viced area. Rm , Sr are the lognormal statistical parameters of
capacity of a deterministic 2 × 2 MIMO channel H. How- the rainfall rate R (in mm/hr). A reliable database of rainfall
ever, since the rainfall introduces slow fading and stochastic statistics for any geographical location on earth is provided
behaviour over the channel H, the appropriate statistic mea- by ITU-R in [18] and is used throughout the present work as
sure to characterize the resulting fading channel is the outage an input to the simulations performed in order to determine
capacity defined by [1] the values of Rm , Sr .
 
P C ≤ Cout,q = q, (7)
4. SATELLITE MIMO DIVERSITY SYSTEM
where Cout,q is the information rate guaranteed for (1− WITH RECEIVE ANTENNA SELECTION:
q)100% of the channel realizations. INTERFERENCE ANALYSIS
Consider the RV transformation
     In this section, the two satellites Si (i = 1, 2) depicted in
ln ARi − ln AmRi Figure 1 are assumed to transmit the same signal over the
ui = (i = 1, 2) (8)
SaRi (ideally) independently fading paths Si -TS (i = 1, 2) (i.e., di-
which relates the lognormal rain attenuation RVs ARi (i = versity is investigated). To alleviate the high cost and com-
1, 2) to the normalized normal RVs ui (i = 1, 2). Substituting plexity associated with multiple RF chains, the dual-antenna
(5) into (7) and after some straightforward algebra, the fol- TS receiver is equipped with only one RF chain and performs
lowing analytical closed form expression for the outage ca- antenna selection, that is, the 2 × 2 MIMO satellite system
pacity is obtained: assumed employs receive selection diversity [21]. Therefore,
  the TS receiver detects the signal related to the path with the
P C ≤ Cout,q highest SNR. Under the constraint of only one RF chain at
 +∞   the receiver, in order to know all SNRs simultaneously for
1   uB − ρn12 u1
= du1 fU1 u1 erfc    = q, optimal selection, a training signal in a preamble to the trans-
2 uA 2
2 1 − ρn12 mitted data is assumed. During this preamble, the TS receiver
(9) scans the two antennas, finds that one with the highest SNR,
and selects it for reception of the next data burst. Thus, only
where erfc(·) is the complementary error function, fU1 (u1 )
a few more training bits are required instead of additional RF
the probability density function (pdf) of the normal distri-
chains.
bution, ρn12 the logarithmic correlation coefficient between
Particular emphasis is put on possible interference mit-
the normal RVs ui (i = 1, 2) [17] and uA , uB are analytically
igation offered by the proposed satellite 2 × 2 MIMO di-
given by
versity system. In this regard, a propagation-based analy-
     
uA = ln 10 log10 0.5SNRCS2 − 10 log10 2Cout,q − 1 sis is performed which is quite different from conventional
 
communication-oriented approaches followed in standard
− ln AmR2 SaR2 , MIMO theory [1]. Specifically, the effect of rainfall on the
interference analysis is taken into account and the differential
rain attenuation related to an adjacent satellite is considered
(10) as the dominant cause of the SIR degradation [22]. Such an
   
uB = ln 10 log10 0.5SNRCS1 interference problem is further aggravated due to the spa-
  tial inhomogeneity of the rainfall medium. It constitutes a
+ 10 log10 1 + 0.5SNRCS2 10−AmR2 exp(u1 SaR2 )/10 typical interference scenario, especially over congested urban
 
− 10 log10 2Cout,q − 1 − 0.5SNRCS2 10−AmR2 exp(u1 SaR2 )/10 areas, where the increased demand for link capacity and ra-
  dio coverage imposes the coexistence of many satellite radio
− ln AmR1 SaR1 . links over the same geographical and spectral area. In the fol-
(11) lowing, an analytical prediction model is presented, which
6 EURASIP Journal on Wireless Communications and Networking

S1 S3 S2

To S3
To S2
To S1
d3 , AR3
d1 , AR1 d2 , AR2

Δψ
Δθ
ϕ3
ϕ2
ϕ1

TS TS
(a) (b)

Figure 4: (a) Configuration of the satellite 2 × 2 MIMO diversity system assumed and the interference scenario on its forward link, (b)
associated elevation angles.

quantifies the adjacent satellite CCI mitigation achieved by is true. Assuming that
the proposed 2 × 2 MIMO system with respect to the corre-    
sponding SISO one. Ωi = Ai < Md , Ai < A j (i, j) = (1, 2), (2, 1) (15)
Figure 4 depicts the configuration of the assumed inter-
ference scenario on the forward link of a satellite 2 × 2 MIMO denotes the event that “the TS is serviced by the correspond-
diversity system operating at the Ku-band and above and ing satellite Si (i = 1, 2),” it becomes clear that, due to selec-
employing receive antenna selection. The satellites S1 and tion diversity,
S2 constitute the dual-satellite transmit part of the MIMO Ω = Ω1 ∪ Ω 2 ,
system also depicted in Figure 1. Another cochannel satellite (16)
(denoted by S3 ), which may belong to either the same or to Ω1 ∩ Ω2 = ∅ .
another satellite network, is close in orbit to S1 . Thus, CCI
Therefore, the probability that the system is available (see the
problems may arise on the forward link of the 2 × 2 MIMO
appendix) can be expressed as
satellite system. S1 and S3 subtend an angle Δψ to TS. The
length of the slant path S3 -TS is denoted by d3 , while its ele-    
P(Ω) = P Ω1 + P Ω2 . (17)
vation angle is φ3 . The RV associated with the rain induced
attenuation along the interfering path S3 -TS (in dB) is de- While the satellite 2 × 2 MIMO diversity system is avail-
noted by AR3 . able (i.e., when either Ω1 or Ω2 are true), it might suffer from
Due to selection diversity at the TS receiver, the antenna CCI originating from the adjacent satellite S3 . If SIRd and rd
with the maximum SNR is selected. In mathematical terms, denote the SIR and the minimum acceptable SIR threshold
the same statement is expressed as of the MIMO diversity system, respectively (both measured
    at the output of the TS selection combiner), the probability
SNRout = max SNR1 , SNR2 ⇐⇒ Aout = min A1 , A2 , of the event that “the system is interfered while being avail-
(13) able” can be mathematically expressed based on the above
considerations as
where SNRi = SNRCSi − ARi (i = 1, 2) is the SNR at each TS
 
antenna under rain fades and Ai (i = 1, 2) the total path loss UIPd = P SIRd < rd , Ω
along each Si -TS link (i = 1, 2). SNRout corresponds to Aout    
= P SIRd1 < rd , Ω1 + P SIRd2 < rd , Ω2 (18)
which determines the output of the selection combiner at ev-
ery instant. The proposed scheme requires only the knowl- = P1 + P2 ,
edge of the wanted signals’ channels at the receiver, whereas
where UIPd is the so-called unacceptable interference proba-
knowledge of the interferer’s channel is not necessary. More-
bility (UIP) [23, 24], and the quantities SIRdi (i = 1, 2) are
over, no CSI is required at the transmit side. If Md denotes
expressed (in dB) as
the diversity system margin associated with the system avail-
ability pavail (see the appendix), the satellite MIMO diversity SIRd = SIRdi = SIRCSi − ARi + AR3 (i = 1, 2). (19)
system is considered available when the probabilistic event
  In (19), SIRCSi (i = 1, 2) is the nominal SIR value under
Ω = Aout < Md (14) clear sky conditions. In propagation terminology, ARi − AR3
Konstantinos P. Liolis et al. 7

(i = 1, 2) is known as the differential rain attenuation (DRA) the other two normal RVs u1 , u2 and can be expressed in
[22]. Based on (19), when DRA becomes sufficiently large terms of the logarithmic correlation coefficients ρni j ((i, j) =
due to the spatial inhomogeneity of the rainfall medium, se- (1, 2), (1, 3), (2, 3)) as [17, 27]
vere CCI problems may arise aggravating the SIRd distribu-
tion on the forward link of the proposed satellite 2 × 2 MIMO ρn13 − ρn12 ρn23 ρn23 − ρn12 ρn13
diversity system. To this end, UIPd is proposed as an efficient μ3/1,2 = 2 u1 + 2 u2 ,
metric to deal with the statistical behaviour of the SIRd and, 1 − ρn12 1 − ρn12
(23)
together with rd , they constitute a pair of design specifica- 2
1 − ρn12 2
− ρn13 2
− ρn23 + 2ρn12 ρn13 ρn23
2
tions concerning interference. Every user must comply with σ3/1,2 = 2 .
1 − ρn12
these specifications, given the QoS specified by the event Ω
related to the system availability (see the appendix).
The quantities SIRCSi (i = 1, 2) encountered in (19) are 5. NUMERICAL RESULTS AND DISCUSSION
given by
  The previous analyses have been applied for the prediction
SIRCSi = SIR∗i − GR θi (i = 1, 2), (20) of possible capacity improvement and interference mitiga-
where θi (i = 1, 2) are the off-axis angles formed by the in- tion achieved by the proposed satellite 2 × 2 MIMO spa-
terfering link S3 -TS and the wanted links Si -TS (i = 1, 2) in tial multiplexing and diversity systems, respectively, and for
the radiation pattern of the TS antennas. From Figure 4, it comparison to the relevant SISO cases. To this end, the base-
follows that θ1 = Δψ and θ2 = Δθ − Δψ. Also, in (20), SIR∗i line configuration scenario considers a TS located in At-
(i = 1, 2) are the relevant SIR values of the interfered links lanta, GA, and communicating with geostationary satellites
Si -TS (i = 1, 2) when θi = 1◦ , and correspond to the nominal S1 (φ1 = 45◦ ) and S2 (φ2 = 40◦ ). The angular separation as-
CCI levels. Based on the channel model assumed, their inter- sumed is Δθ =40◦ , which results in a spatial correlation coef-
relationship is defined through (6) by simply substituting the ficient of rain attenuation ρ12 = 0.6 (see Figure 3). Moreover,
SNRCSi by SIR∗i . regarding the interference scenario, an adjacent geostation-
Extending the transformation given in (8) to include also ary satellite S3 (φ3 = 45◦ ), separated from S1 by Δψ =10◦ , is
the interfering link S3 -TS (i.e., for i = 1, 2, 3) and making the considered to cause CCI problems on the forward link of the
channel modeling assumptions, the probabilities Pi (i = 1, 2) satellite 2 × 2 MIMO diversity system.
encountered in (18) after some straightforward algebra are First, the validity of the proposed analytical model in (9),
evaluated, that is, predicting the outage capacity achieved by a satellite 2 × 2
MIMO spatial multiplexing system, is numerically verified.
 uDi  +∞
  The effect of various geometrical and operational system pa-
Pi = du1 du2 fU1 U2 u1 , u2
uCi u1 rameters on the outage capacity distribution is also exam-

 (21) ined.
1 uEi − μ3/1,2
× 1− erfc √ (i = 1, 2), Figure 5 shows the dependence of the 1% outage capac-
2 2σ3/1,2
ity of the assumed 2 × 2 MIMO satellite system on the SNR.4
where fU1 U2 (u1 , u2 ) is the pdf of the two-dimensional joint The baseline configuration scenario is adopted, whereas the
normal distribution. operating frequency band assumed is Ka (i.e., f = 20 GHz).
For i = 1, 2, the rest of the parameters encountered in For the sake of comparison, the capacity of the relevant SISO
(21) are system is also plotted. Together with the analytical results
     obtained from the analytical closed form expression in (9),
ln xdi − ln AmRi Monte Carlo simulation results are also plotted for verifica-
uCi = ,
SaRi tion. The agreement observed between the analytical and the
⎧
⎪
⎪ rd > SIRCSi , simulation results is very good over the whole SNR range.
⎨0,
  As can be seen, the difference between the relevant MIMO
xdi = ⎪ SIRCSi − rd cos φi , SIRCSi +FSLi − Md<rd ≤ SIRCSi ,
⎩M − FSL  cos φ ,
⎪
rd ≤ SIRCSi + FSLi − Md ,
and SISO curves diminishes at very low SNR levels while
d i i it becomes significant as the SNR increases. As an illustra-
      tion, for SNR = 10 dB, the spectral efficiency achieved by
ln Md − FSLi cos φi − ln AmRi
uDi = , the MIMO system is 4.84 bps/Hz, whereas the one achieved
SaRi
  by the SISO system is 3.23 bps/Hz. This constitutes, approx-


exp ui SaRi AmRi imately, a 50% increase in user data rate obtained by MIMO
uEi = ln − SIRCSi + rd cos φ3
spatial multiplexing. For SNR = 20 dB, the respective per-
cos φi
 formance figures obtained are 10.95 bps/Hz and 6.41 bps/Hz
  
− ln AmR3 SaR3 . corresponding to, approximately, a 71% increase in user data
(22)
4 Note that the clear sky SNR of strong eigenmode, SNRCS1 , has been par-
AmRi , SaRi (i = 1, 2, 3) are analytically given in (12). Fur- ticularly considered. However, due to the enormous slant path lengths as-
thermore, μ3/1,2 and σ3/1,2 are the statistical parameters of sociated, the resulting difference between SNR CSi (i = 1, 2) is minimum
the conditional distribution of the normal RV u3 given see (6) and, therefore, any of the two SNRCSi can be used as x-coordinates.
8 EURASIP Journal on Wireless Communications and Networking

18 tem on the SNR, the angular separation Δθ, the operating

16 frequency f , the capacity outage probability q, and the cli-
matic conditions over the serviced area. All the results pre-
1% outage capacity (bps/Hz)

14 sented here have been obtained employing (9). The baseline

12 configuration scenario is adopted. The rest of the relevant pa-
2 × 2 MIMO rameters assumed as well as the deviations from the baseline
10 scenario are indicated on Figure 6. As can be seen, as either q
8 decreases or f increases or as the rain conditions over the ser-
SISO viced area become heavier, the rain fading becomes more se-
6 vere and, therefore, the outage capacity achieved by the 2 × 2
4 MIMO satellite system decreases. Moreover, as the angular
separation Δθ increases (from 40◦ to 60◦ ), the spatial corre-
2 lation coefficient due to rainfall medium ρ12 decreases cor-
0 respondingly (from 0.6 to 0.5, see Figure 3), and the outage
0 5 10 15 20 25 30 capacity achieved increases.
SNR (dB) In the following, the proposed analytical model in (21)
Analytical expression (9) predicting the interference mitigation achieved by a satellite
Monte Carlo simulation 2 × 2 MIMO diversity system with receive antenna selection
is numerically verified. The effect of various geometrical and
Figure 5: 1% outage capacity versus SNR for a satellite 2 × 2 MIMO operational system parameters on the forward link SIR dis-
spatial multiplexing system. Relevant SISO case is also plotted for tribution is also examined.
comparison. Verification of analytical closed form expression in (9) Figure 7 shows the dependence of the UIP of the assumed
through Monte Carlo simulation.
2 × 2 MIMO satellite system on the SIR, the system avail-
ability pavail , and the operating frequency band. Particularly,
18 two different values of system availability, pavail = 99.9%
16
and 99.99%, and two different operating frequencies, f =
12 GHz and 20 GHz, are assumed. For the sake of compar-
2 × 2 MIMO system (bps/Hz)
Outage capacity achieved by

14 ison, the UIP of the relevant SISO systems is also plotted.

12
10
8 indicated on Figure 7. It is obvious that, due to rain, an SIR
6
4 ther indicates that satellite systems operating at higher avail-
2
0 2 × 2 MIMO diversity system over the SISO one is signifi-
0 5 10 15 20
SNR (dB)
q = 1%, Δθ = 40◦ , Ka-band, Atlanta
q = 0.1%, Δθ = 40◦ , Ka-band, Atlanta
q = 1%, Δθ = 40◦ , Ku-band, Atlanta
q = 1%, Δθ = 40◦ , Ka-band, Singapore
q = 1%, Δθ = 60◦ , Ka-band, Atlanta
Figure 6: Outage capacity versus SNR for a satellite 2 × 2 MIMO
spatial multiplexing system. Effect of capacity outage probability q,
angular separation Δθ, operating frequency f , and climatic condi-
tions over the serviced area.
rate. Therefore, the capacity gain obtained by the proposed
satellite 2 × 2 MIMO spatial multiplexing system over the
SISO system turns out to be significant for no additional
transmit power or bandwidth expenditure.
Figure 6 shows the dependence of the outage capacity
achieved by a satellite 2 × 2 MIMO spatial multiplexing sys-
The baseline configuration scenario is adopted. The nomi-
nal CCI level assumed is SIR∗1 = 20 dB, whereas the rest of
the parameters encountered in the interference analysis are
degradation is observed for the same UIP level, which be-
comes more severe as either pavail or f increases. This fur-
abilities or at higher-frequency bands are more sensitive to
interference. The SIR improvement achieved by the satellite
25 30
cant, especially for high pavail and high f . As an illustration,
for UIP = 0.001%, the interference mitigation obtained is
0.67 dB at the Ka-band and for a 99.9% availability, 1.60 dB
at the Ku-band and for a 99.99% availability, and 3.52 dB at
the Ka-band and for a 99.99% availability.
Figure 8 quantifies the SIR improvement achieved by a
satellite 2 × 2 MIMO diversity system employing receive
antenna selection with respect to the relevant SISO one.
Specifically, the difference (in dB) between the respective
SIR thresholds achieved at the TS receiver input for UIP =
0.001% is plotted versus the angular separation Δθ. Two
areas with different climatic conditions are considered, At-
lanta, GA, and Athens, Greece. The operating frequency, sys-
tem availability, and nominal CCI level assumed are 20 GHz,
99.99%, and SIR∗1 = 20 dB, respectively, while the rest of
the parameters are the same as those of the baseline con-
figuration scenario. As Δθ increases, the interference miti-
gation level achieved becomes higher. Moreover, it can easily
be observed that the SIR improvement obtained in Atlanta,
Konstantinos P. Liolis et al. 9

100 100

Unacceptable interference probability (UIP)

Unacceptable interference probability (UIP)

10−1 10−1

10−2 10−2
Ka-band,
10−3 pavail = 99.9% 10−3 Δψ = 4◦ Δψ = 5◦
Ku-band,
pavail = 99.99%
10−4 10−4
Ka-band,
pavail = 99.99%
10−5 10−5

10−6 10−6
2 4 6 8 10 12 14 16 18 20 7 9 11 13 15 17 19 20
SIR (dB) SIR (dB)

SISO SISO
2 × 2 MIMO 2 × 2 MIMO

Figure 7: UIP versus SIR for a satellite 2 × 2 MIMO diversity sys- Figure 9: UIP versus SIR for a satellite 2 × 2 MIMO diversity sys-
tem employing receive antenna selection. Relevant SISO case is also tem employing receive antenna selection. Relevant SISO case is also
plotted for comparison. Effect of system availability pavail , operating plotted for comparison. Effect of angular separation ΔΨ.
frequency f , and rain climatic conditions over the serviced area.

2.5 ters assumed are the same as those in the baseline configura-
tion scenario, with the exception of a different angular sepa-
Atlanta, GA ration Δψ, that is, Δψ = 5◦ is now assumed. Operation of the
SIR improvement achieved through

2
system at the Ka-band and for a 99.99% availability is con-
sidered. To obtain the necessary QoS for UIP = 0.001%, sup-
MIMO diversity (dB)

1.5 pose that an SIR threshold of 10 dB must be overcome. In the

1 an SIR level of 11.2 dB is obtained for UIP = 0.001%, thus
0.5
0
20
Figure 8: SIR improvement achieved by a satellite 2 × 2 MIMO di-
versity system with receive antenna selection over the relevant SISO
system versus angular separation Δθ. Effect of rain climatic condi-
tions over the serviced area.
GA, is much higher than that in Athens, Greece, due to the
corresponding heavier rain conditions.
For various obvious reasons, there is a tendency to place
satellites in orbit close to each other. Due to the increased
CCI, adjacent satellite networks cannot usually operate un-
der certain SIR specifications. The proposed MIMO diversity
system may overcome this problem by adequately increasing
SIR in the presence of adjacent CCI. To demonstrate this, a
satellite 2 × 2 MIMO diversity system together with its rele-
vant SISO case are considered in Figure 9. The input parame-
SISO case, when the angular separation between the wanted
satellite S1 and the adjacent interfering one S3 is Δψ = 5◦ ,
satisfying the QoS requirement. If the interfering satellite S3
is closer in orbit to S1 , so that their angular separation is re-
duced to Δψ = 4◦ , the SIR level in the SISO case falls down
Athens, GR to 9.8 dB, thus failing to satisfy the QoS requirement. Em-
ploying the proposed 2 × 2 MIMO satellite system, the SIR
30 40 50 60 70 80 90 achieved when Δψ = 4◦ is 11.32 dB, thus remaining above
Angular separation, Δθ (deg) the QoS threshold. This is another advantage of the proposed
satellite MIMO diversity system, allowing the closer installa-
tion of satellites in orbit.
6. CONCLUSIONS
In this paper, the applicability of MIMO technology to satel-
lite communication systems operating at the Ku-band and
above is investigated. Emphasis is put on satellite diversity as
a potential candidate to form a MIMO matrix channel in the
satellite environment. The relevant propagation phenomena
at the frequencies of interest have been considered through
an appropriate physical channel model, which takes into ac-
count clear LOS operation, high antenna directivity at the TS
receiver, the effect of rain fading, and the slant path lengths
difference. Also, as it may accept physical inputs from the
ITU-R rainmaps, it is flexible and can be applied on a global
scale.
10 EURASIP Journal on Wireless Communications and Networking

Useful analytical results are presented for two different After straightforward algebra, (.2) yields
applications of MIMO technology:
(i) capacity improvement in a satellite 2 × 2 MIMO spatial pavail · 100%
multiplexing system,  +∞  
  uF2 − ρn12 u1 (.4)
(ii) interference mitigation in a satellite 2 × 2 MIMO di- = 1 − 0.5 du1 fU1 u1 erfc    .
2
versity system with receive antenna selection.
uF1 2 1 − ρn12

In the first application, significant capacity gains of the

MIMO system over the relevant SISO one are demonstrated, ACKNOWLEDGMENTS
especially for moderate and high SNR levels. The practical
case when no CSI is available at the transmitters of the two The authors are indebted to the three anonymous review-
individual satellites is considered. A useful closed form ex- ers whose constructive comments helped to significantly im-
pression for the outage capacity achieved by 2 × 2 MIMO prove the initial version of this paper. Moreover, the first au-
satellite systems is provided and successfully verified through thor would like to thank Professor Bhaskar D. Rao from Uni-
Monte Carlo simulations. Such an expression is extremely versity of California, San Diego, USA, for the fruitful discus-
hard to obtain even in the well-established field of MIMO sions they had on the first part of this work.
theory, is applicable over a large SNR range, and can incorpo-
rate the effect of various geometrical and operational system REFERENCES
parameters on the outage capacity distribution.
In the second application, the receive antenna selection [1] A. J. Paulraj, D. A. Gore, R. U. Nabar, and H. Bölcskei, “An
scheme employed in the satellite MIMO system assumed is overview of MIMO communications—a key to gigabit wire-
considered to counteract CCI problems over its forward link. less,” Proceedings of the IEEE, vol. 92, no. 2, pp. 198–218, 2004.
SIR gain of several dB is demonstrated in the numerical re- [2] D. Gesbert, M. Shafi, D.-S. Shiu, P. J. Smith, and A. Naguib,
sults. An analytical propagation model for the calculation of “From theory to practice: an overview of MIMO space-time
the interference mitigation achieved is presented, which is coded wireless systems,” IEEE Journal on Selected Areas in
flexible and can incorporate the influence of various geomet- Communications, vol. 21, no. 3, pp. 281–302, 2003.
rical and operational system parameters on the SIR distribu- [3] A. D. Panagopoulos, P.-D. M. Arapoglou, and P. G. Cottis,
“Satellite communications at Ku, Ka, and V bands: propaga-
tion.
tion impairments and mitigation techniques,” IEEE Commu-
nications Surveys and Tutorials, vol. 6, no. 3, pp. 2–14, 2004.
APPENDIX [4] K. P. Liolis, A. D. Panagopoulos, and P. G. Cottis, “Outage ca-
pacity statistics of MIMO satellite networks operating at Ka
CALCULATION OF SATELLITE 2 × 2 MIMO band and above,” in Proceedings of the 12th Ka and Broadband
DIVERSITY SYSTEM MARGIN Md Communications Conference, Naples, Italy, September 2006.
[5] F. Yamashita, K. Kobayashi, M. Ueba, and M. Umehira,
Every user in the assumed satellite 2 × 2 MIMO diversity sys- “Broadband multiple satellite MIMO system,” in Proceedings
tem employing receive antenna selection must comply with a of the 62nd IEEE Vehicular Technology Conference (VTC ’05),
certain availability percentage pavail related to a diversity sys- pp. 2632–2636, Dallas, Tex, USA, September 2005.
tem margin Md : [6] P. R. King and S. Stavrou, “Land mobile-satellite MIMO ca-
pacity predictions,” Electronics Letters, vol. 41, no. 13, pp. 749–
  751, 2005.
pavail · 100% = P(Ω) = P Aout < Md [7] T. Hult and A. Mohammed, “MIMO antenna applications
    for LEO satellite communications,” in Proceedings of the 3rd
= P min A1 , A2 < Md
  ESA International Workshop of the European COST 280 Action,
= 1 − P A1 > Md , A2 > Md Prague, Czech Republic, June 2005.
  [8] C. Martin, A. Geurtz, and B. Ottersten, “Spectrally efficient
= 1 − P AR1 > Md − FSL1 , AR2 > Md − FSL2 .
mobile satellite real-time broadcast with transmit diversity,”
(.1) in Proceedings of the 60th IEEE Vehicular Technology Confer-
ence (VTC ’04), vol. 6, pp. 4079–4083, Los Angeles, Calif, USA,
Considering the transformation given in (8), relating the log-
September 2004.
normal rain attenuation RVs ARi (i = 1, 2) to the normalized [9] C. Faller, B.-H. Juang, P. Kroon, H.-L. Lou, S. A. Ramprashad,
normal RVs ui (i = 1, 2), and the channel modeling assump- and C.-E. W. Sundberg, “Technical advances in digital audio
tions, pavail is expressed as radio broadcasting,” Proceedings of the IEEE, vol. 90, no. 8, pp.
 +∞  +∞ 1303–1333, 2002.
 
pavail · 100% = 1 − du1 du2 fU1 U2 u1 , u2 , (.2) [10] J. Mietzner and P. A. Hoeher, “Distributed space-time codes
uF1 uF2 for cooperative wireless networks in the presence of different
propagation delays and path losses,” in Proceedings of Sensor
where Array and Multichannel Signal Processing Workshop (SAM ’04),
      pp. 264–268, Barcelona, Spain, July 2004.
ln Md − FSLi cos φi − ln AmRi
uFi = (i = 1, 2). [11] B. N. Getu and J. B. Andersen, “The MIMO cube—a compact
SaRi MIMO antenna,” IEEE Transactions on Wireless Communica-
(.3) tions, vol. 4, no. 3, pp. 1136–1141, 2005.
Konstantinos P. Liolis et al. 11

[12] I. Frigyes and P. Horváth, “Polarization-time coding in satel-

lite links,” IEEE Satellite and Space Communications Newsletter,
vol. 15, no. 2, pp. 6–8, 2005.
[13] P. Horváth and I. Frigyes, “SAT02-6: application of the
3D polarization concept in satellite MIMO systems,” in
Proceedings of IEEE Global Telecommunications Conference
(GLOBECOM ’06), pp. 1–5, San Francisco, Calif, USA,
November 2006.
[14] P. R. King and S. Stavrou, “Capacity improvement for a land
mobile single satellite MIMO system,” Antennas and Wireless
Propagation Letters, vol. 5, no. 1, pp. 98–100, 2006.
[15] M. Sellathurai, P. Guinand, and J. Lodge, “Space-time cod-
ing in mobile satellite communications using dual-polarized
channels,” IEEE Transactions on Vehicular Technology, vol. 55,
no. 1, pp. 188–199, 2006.
[16] G. Taricco, E. Viterbo, and E. Biglier, “MIMO transmission for
mobile satellite communication systems: a review,” in Proceed-
ings of the 8th International Workshop on Signal Processing for
Space Communications (SPSC ’03), Catania, Italy, September
2003.
[17] A. D. Panagopoulos and J. D. Kanellopoulos, “Prediction of
triple-orbital diversity performance in Earth-space commu-
nication,” International Journal of Satellite Communications,
vol. 20, no. 3, pp. 187–200, 2002.
[18] ITU-R Recommendation P.837-4, “Characteristics of Precipi-
tation for Propagation Modeling,” Geneva, Switzerland, 2003.
[19] G. J. Foschini and M. J. Gans, “On limits of wireless commu-
nications in a fading environment when using multiple an-
tennas,” Wireless Personal Communications, vol. 6, no. 3, pp.
311–335, 1998.
[20] I. E. Telatar, “Capacity of multi-antenna Gaussian channels,”
European Transactions on Telecommunications, vol. 10, no. 6,
pp. 585–595, 1999.
[21] S. Sanayei and A. Nosratinia, “Antenna selection in MIMO sys-
tems,” IEEE Communications Magazine, vol. 42, no. 10, pp. 68–
73, 2004.
[22] J. D. Kanellopoulos, S. Ventouras, and C. N. Vazouras, “A re-
vised model for the prediction of differential rain attenua-
tion on adjacent Earth-space propagation paths,” Radio Sci-
ence, vol. 28, no. 6 part 2, pp. 1071–1086, 1993.
[23] P.-D. M. Arapoglou, A. D. Panagopoulos, J. D. Kanellopou-
los, and P. G. Cottis, “Intercell radio interference studies in
CDMA-based LMDS networks,” IEEE Transactions on Anten-
nas and Propagation, vol. 53, no. 8, pp. 2471–2479, 2005.
[24] A. D. Panagopoulos, P.-D. M. Arapoglou, J. D. Kanellopoulos,
and P. G. Cottis, “Intercell radio interference studies in broad-
band wireless access networks,” IEEE Transactions on Vehicular
Technology, vol. 56, no. 1, pp. 3–12, 2007.
[25] ITU-R Recommendation S.580-6, “Radiation Diagrams for
Use as Design Objectives for Antennas of Earth Stations Op-
erating with Geostationary Satellites,” Geneva, Switzerland,
2004.
[26] P. Horváth and I. Frigyes, “Application of the MIMO concept
in millimeter-wave broadband wireless access networks,” In-
ternational Journal of Wireless Information Networks, vol. 11,
no. 4, pp. 217–225, 2004.
[27] A. Papoulis and S. U. Pillai, Probability, Random Variables and
Stochastic Processes, McGraw-Hill, Englewood Cliffs, NJ, USA,
4th edition, 2002.
Hindawi Publishing Corporation
EURASIP Journal on Wireless Communications and Networking
Volume 2007, Article ID 98942, 10 pages
doi:10.1155/2007/98942

Research Article
Investigations in Satellite MIMO Channel Modeling:
Accent on Polarization

Péter Horváth,1 George K. Karagiannidis,2 Peter R. King,3 Stavros Stavrou,3 and István Frigyes1
1 Department of Broadband Infocommunications and Electromagnetic Theory, Budapest University of Technology and Economics,
H-1111 Budapest, Hungary
2 Division of Telecommunications, Department of Electrical and Computer Engineering, Aristotle University of Thessaloniki,

54124 Thessaloniki, Greece

3 Centre for Communication Systems Research, University of Surrey, Guildford, Surrey GU2 7XH, UK

Received 30 September 2006; Accepted 19 March 2007

Recommended by Ray E. Sheriff

Due to the much different environment in satellite and terrestrial links, possibilities in and design of MIMO systems are rather
different as well. After pointing out these differences and problems arising from them, two MIMO designs are shown rather well
adapted to satellite link characteristics. Cooperative diversity seems to be applicable; its concept is briefly presented without a de-
tailed discussion, leaving solving particular satellite problems to later work. On the other hand, a detailed discussion of polarization
time-coded diversity (PTC) is given. A physical-statistical model for dual-polarized satellite links is presented together with mea-
suring results validating the model. The concept of 3D polarization is presented as well as briefly describing compact 3D-polarized
antennas known from the literature and applicable in satellite links. A synthetic satellite-to-indoor link is constructed and its elec-
tromagnetic behavior is simulated via the FDTD (finite-difference time-domain) method. Previous result of the authors states that
in 3D-PTC situations, MIMO capacity can be about two times higher than SIMO (single-input multiple-output) capacity while a
diversity gain of nearly 2 × 3 is further verified via extensive FDTD computer simulation.

Copyright © 2007 Péter Horváth et al. This is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

1. INTRODUCTION is present in satellite-to-mobile or satellite-to-indoor links.

It is more or less a commonplace statement that in the wire-
less technology of recent years, systems applying multiple-
transmit and multiple-receive antennas (MIMO, multiple-
input multiple-output) have become one of the few meth-
ods of real innovation. Space-time processing, in particular
space-time coding (STC) techniques as applied to MIMO
systems in a multipath environment, results in significant
improvement both in transmission capacity and reliability.
It turns out that there are significant differences between ter-
restrial and satellite multipath channels; these result in signif-
icant differences in MIMO applications as well. In this paper,
we deal with some special problems raised by special charac-
teristics of satellite links.
In terrestrial applications of MIMO, the basic method
to diversify channels is with the additional dimension of
space, that is, antennas are displaced spatially from each
other, resulting in space-time processing. In addition, multi-
path channels and relevant fading characteristics—Rayleigh,
Rice, Suzuki, and so forth—are assumed. A similar situation
Among others, in [1] it is experimentally verified that the
LEO satellite-to-indoor channel has nearly exactly Rayleigh
character at any fixed indoor spot. More precise models are
available (Loo, Corrazza, etc.) well describing the multipath
behavior and not differing much from the terrestrial case.
Consequently, similar-to-terrestrial results can be foreseen in
satellite links of appropriate design. However, due to the very
huge length of the radio path, transmit and/or receive anten-
nas must be placed at significant distances from each other
in order to ensure that the various paths are really diverse.
To achieve this in principle generalization of satellite diver-
sity and site diversity would be candidates in forming MIMO
channels. (Note that in satellite diversity, there are two or
more satellites transmitting/receiving the same signal; in site
diversity there are two or more Earth stations.) These would
make original space time processing possible: both ground
and satellite terminals are in this case remote from each other
and so are their antennas. Of course the original concept of
site diversity can be excluded in the present—mostly hand-
held mobile/indoor—situations.
2 EURASIP Journal on Wireless Communications and Networking
In one class of cases, the ground terminals are located on-
board large objects, such as trains, ships, or aircrafts. Large-
antenna distances are possible then, realizing diverse routes.
Multipath, on the other hand, is nonexistent or very sparse.
Difference of LOS route lengths must be in such a case at
least λ/16 · · · λ/4. Site diversity might be applicable then, if
as a rough estimate, terminal antennas can be placed at a
distance of b = 35 m from each other. (For that figure, an
LEO satellite and 30 GHz carrier frequency were assumed;
note that b is proportional to the square root of satellite
distance × wavelength.)
Satellite diversity for space-time processing would fulfill
the requirement of uncorrelated channels and so it would be
applicable. There is a few papers dealing with this topic; for
example, [2] gives a physical-statistical model for satellite-to-
urban and satellite-to-highway channel and computes capac-
ity of a 2 × 2 MIMO system. In [3], a satellite-diversity MIMO
system and its system aspects are investigated. Further papers
on satellite MIMO are, among others, [4, 5].
There exists, however, at least one problem not present
in terrestrial systems, that is, that of synchronization. In ter-
restrial MIMO systems, both the group of transmit anten-
nas and that of receive antennas are at distances from each
other in the order of a wavelength. Consequently, the path
lengths of the diversity routes are very closely identical, and
thus signals arriving from the transmitter to the receiver are
synchronous. This makes identification and decoding of the
received signals rather easy. In the case of satellite diversity,
the satellites serving as diversity terminals are very far from
each other. Thus difference of path lengths and so delays be-
tween the satellites and the ground terminal can be very high
and highly variable. (This variability is self-evidently existing
in the case of LEO satellites but very likely also in the GEO
case.) As a consequence, the arrival time of signals from two
satellites (forming part of a single code word) can be shifted
by tens or hundreds of symbol times relative to each other.
Synchronization of the received signals is in this case rather
complicated—both acquisition and tracking. Reference [2]
or [3] or other satellite/MIMO papers known by the authors
do not deal with this problem. General aspects of it are dealt
with, for example, in [6–8], taking explicitly, however, short-
range, that is, terrestrial situations only into account.
An alternative possible solution could be cooperative
satellite diversity (CSD). In general, cooperative relaying sys-
tems have a source node (e.g., a terrestrial mobile terminal
(TMT)) multicasting a message to a number of cooperative
relays (satellites (SAT)), which in turn resend a processed ver-
sion to the intended destination node (another TMT). The
destination node combines the signal received from the re-
lays, possibly also taking into account the source’s original
signal. Recently, it has been shown that cooperative diversity
systems provide an effective way of improving spectral and
power efficiencies of the wireless networks without the ad-
ditional complexity of multiple antennas [7–11]. However, a
study on CSD systems, where the relays are satellites, to the
best of the authors’ knowledge does not exist in the literature.
A third possible method is to apply compact antennas,
in which case the synchronization problem is nonexistent.
Compact antennas with low radiator spacing and dimensions
as small as λ/20 or so are described, for example, in [12–
14]. These antennas were mainly developed for application
in handheld terminals, in which the available space is very
limited. In the case of onboard antennas, the whole antenna
need not be small, however, the radiator elements need to be
colocated, that is, their ports need to be very close to each
other. Note that polarization, and in many cases the 3D char-
acter of it, has a significant role in each of the known compact
antennas.
In this paper, the concept of cooperative satellite diversity
is briefly introduced, without, however, a detailed discussion;
this is done in Section 2. Polarization diversity and the appli-
cation of space-time coding concepts in polarization diver-
sity are dealt with in Section 3. (In analogy to the name STC,
we call that polarization time coding (PTC). Note that ac-
cording to the authors’ understanding, the term STC is used
to distinguish a transmit-and-receive-space-diversity situa-
tion from a simple receive diversity. The same understanding
is applied in this paper; so we will call our topic PTC even if
particular coding problems are not at all dealt with but coded
signals are assumed.) Section 3.1 deals with dual-polarized
MIMO channels, stating a physical-statistical model, pre-
senting measuring results and validating the model; in this
discussion conventional dual-polarized antennas are applied.
In Section 3.2, PTC antennas of 3-dimensional polarization
are dealt with, introducing the concept of 3D polarization,
presenting a few compact MIMO antennas and showing
the essential difference between terrestrial and satellite links
from the point of view of 3D PTC. In Section 4, electro-
magnetic simulation results are given; in these it is verified
that application of the FDTD method is suitable to investi-
gate MIMO channel characteristics of very complex environ-
ments; capacity as well as diversity behavior are presented;
these verify (at least for the present example) the statements
of Section 3.2 and of the authors’ references [15, 16]. Con-
clusions are drawn in Section 5.
2. A FEW WORDS ON COOPERATIVE
SATELLITE DIVERSITY
In general, cooperative relaying systems have a source node
(e.g., TMT) multicasting a message to a number of cooper-
ative relays (SAT), which in turn resend a processed version
to the intended destination node (another TMT). The des-
tination node combines the signal received from the relays,
possibly also taking into account the source’s original signal.
An example of a CSD system with two satellite relays is shown
in Figure 1.
The idea of merging cooperation with space-time coding
resulted in the so-called distributed or cooperative space-time
coding (CSTC). Compared to the conventional space-time
coding with collocated antennas, CSTC can be implemented
when transmitter and relays share their antennas to create a
virtual transmit array.
A possible cooperation scenario is applied for the con-
figuration of Figure 1, proposed in [9] as TMT1 communi-
cates with SAT1 and SAT2 in a broadcasting mode during
Péter Horváth et al. 3

3.1.1. Channel model construction

SAT1 The following dual-polarized physical-statistical LMS MI-
MO channel model is an extension to the multiple-satellite
LMS MIMO model presented in [2]. In the present paper, a
single satellite containing right-(RHCP) and left-hand circu-
lar polarization (LHCP) antennas communicates with a mo-
TMT1 bile vehicle, also containing RHCP and LHCP antennas. Note
TMT2
that taking into account the spherical symmetry of polariza-
tion states on the Poincaré sphere, actual choice of two or-
SAT2 thogonal polarizations does not have too much significance
[21].
Channel model construction is described in [2]. Addi-
Figure 1: A virtual array: 2 satellites and 2 terminals. tional insertion of polarization properties is achieved as fol-
lows. When the LOS path is unobstructed (clear), simple
path loss is applied to the copolar channels and cross-polar
channels are discarded. When the LOS path is blocked by a
building (blocked), rooftop diffraction is applied to both the
the first signaling interval and there is no transmission from co- and cross-polar channels; the cross-polar component is
SAT1 or SAT2 to TMT2 within this time interval. In the sec- scaled below the copolar component as observed from mea-
ond signaling interval, both SAT1 and SAT2 communicate sured data. When the LOS path is shadowed by vegetation
with TMT2. This scenario assumes perfect knowledge of the (tree), attenuation is applied to this path based on the dis-
channel fading coefficients at the receiver side of TMT2 and tance traversed through the tree and using a typical attenu-
synchronization as an a priori condition. However, the delays ation factor of −1.3 dB per meter [22]. Similarly, the cross-
due to distance between SAT1 and SAT2 (and the different lo- polar component is scaled below the copolar component.
cal oscillators at SAT1 and SAT2) make cooperative diversity It is assumed in this model that the LOS paths are fully
asynchronous in nature. correlated between co- and cross-polar channels, and that the
Several methods have been proposed to apply CSTC, in diffuse multipath components are fully uncorrelated between
the presence of asynchronity between relays (see [17, 18] and co- and cross-polar channels. This simplification is represen-
references therein). However, a theoretical analysis on the ef- tative of many, but not all, real practical channels; a full pre-
fect of the (high) asynchronity in cooperative satellite diver- sentation of measured satellite MIMO channel correlation is
sity systems does not exist in the literature. Such an analysis provided in [23].
is out of the scope of the present paper and is left for further The high-resolution time-series data αM,N between each
study. satellite antenna M and each mobile antenna N can be de-
fined as follows:
⎧
⎪ jkdM,N
3. POLARIZATION-TIME CODING IN SATELLITE ⎪PM,N e
⎪
⎪
⎪
⎪
COMMUNICATIONS ⎪
⎪  n
⎪
⎪
⎪
⎪ +b Ti Γi PM,N,i e jkdM,N,i clear co-polar
⎪
⎪
⎪
⎪ i=1
⎪
⎪  n
3.1. Physical-statistical model for the dual polarized ⎪
⎪
⎪
⎪ b Ti Γi PM,N,i e jkdM,N,i clear cross-polar
LMS MIMO channel ⎪
⎪
⎪
⎪ i=1
⎪
⎪
⎪DM,N PM,N e jkdM,N
⎪
In [19], a basic investigation of PTC was presented, using ⎪
⎪
⎪
⎪
⎪
⎪  n
a simple theoretical MIMO channel model. It was assumed ⎪
⎪
⎪ +b Ti Γi PM,N,i e jkdM,N,i
⎪ block co-polar
that in a multipath environment—of whatever polarization ⎪
⎪
⎪
⎪
the transmit antenna(s) is (are)—the received signal is of ⎨ i=1
jkdM,N
completely random polarization, that is, any state of polar- αM,N = ⎪Sb DM,N PM,N e
⎪
⎪ ⎪
⎪  n
ization is equally likely. With a simulation study, we did show ⎪
⎪
⎪
⎪ +b Ti Γi PM,N,i e jkdM,N,i block cross-polar
that applying normal dual-polarized antennas at both ter- ⎪
⎪
⎪
⎪ i=1
minals and transmitting Alamouti-type coded signals [20], ⎪
⎪
⎪
⎪ T P M,N e
jkdM,N
there is a 2 × 1 or 2 × 2 diversity effect if polarization of the re- ⎪
⎪
M,N
⎪
⎪
⎪
⎪  n
ceived signals is fully correlated or completely uncorrelated, ⎪
⎪
⎪
⎪
⎪
+b Ti Γi PM,N,i e jkdM,N,i tree co-polar
respectively. Incidentally, polarization characteristics are de- ⎪
⎪
⎪
⎪ i =1
scribed there via Stokes parameters and related concepts. In ⎪
⎪ St TM,N PM,N e jkdM,N
⎪
⎪
order to assess the benefits of MIMO techniques applied to ⎪
⎪
⎪
⎪  n
⎪
⎪
mobile satellite links, real channel data or accurate channel ⎪
⎪ +b Ti Γi PM,N,i e jkdM,N,i tree cross-polar
⎩
models are required. In this section, a physical-statistical 2 × 2 i=1
dual-polarized MIMO channel model is presented. (1)
4 EURASIP Journal on Wireless Communications and Networking
where PM,N is the LOS path loss between satellite antenna M
and moving mobile antenna N, k is the wavenumber, n is
the total number of valid scatterers, Ti is the tree attenuation
applied to a reflected contribution from scatterer i, Γi is the
complex reflection coefficient at scatterer i, PM,N,i is the path
loss from satellite antenna M to moving mobile antenna N
via scatterer i, dM,N,i is the distance between satellite antenna
M and moving mobile antenna N via scatterer i, DM,N is the
LOS diffraction loss, and TM,N is the LOS tree loss. The terms
Sb and St account for the attenuation of the cross-polar terms
for blocked and tree-shadowed conditions, respectively and
are derived from measured data. The term b is a clutter factor
parameter also derived from measurements in each environ-
ment.
3.1.2. Measurement campaign
Extensive measurements were carried out in Guildford, UK,
where an artificial platform situated on a hilltop (acting as
the satellite), containing directional RHCP and LHCP patch
antennas, communicated with a mobile van fitted with om-
nidirectional RHCP and LHCP antennas. Further details of
the experiment are given in [23, 24].
Two of the measured environments were modeled: (a)
tree-lined road/highway, characterized by a high likelihood
of dense tree matter at either side of the road with occasional
clearings and occasional two-storey houses beyond the veg-
etation, and (b) urban, characterized by densely placed two-
to-four-storey buildings and sporadic tree matter.
3.1.3. Model output and validation
The model was optimized by fitting its parameters to the
measured data. The model is capable of producing statisti-
cally accurate wideband channel time-series data and first-
and second-order statistics. In this paper, the first-order
statistics of the model are presented showing their validation
against measured data. Validation of second-order statistics,
not relevant to the diversity gain analysis presented below, is
a work to be published.
An example of the copolar model output high-resolution
path loss time-series data is shown in Figure 2. Similar data
were obtained between each mobile antenna and satellite, for
both polarizations.
Data were collected using three samples per wavelength
in the model and measurement campaign, ensuring a sam-
pling frequency well over twice the maximum Doppler fre-
quency.
The narrowband first-order modeled and measurement
data are compared. Cumulative distribution functions of co-
and cross-polar channels for highway and urban environ-
ments are shown in Figure 3. The 2 × 2 dual-polarized MIMO
channel matrix data were also used to estimate the diversity
gain from a 1 × 2 maximum ratio receive combining system,
a 2 × 1 polarization time block code approach [20], and a
2 × 2 polarization time block code system. An example from
the highway environment data is shown in Figure 4.
10
Received power (dB)
Urban
0
−10
−20
−30
−40
0 200 400 600 800 1000 1200 1400 1600 1800 2000
Mobile position (m)
(a)
10
Received power (dB)
Highway
0
−10
−20
−30
−40
0 200 400 600 800 1000 1200 1400 1600 1800 2000
Mobile position (m)
(b)
Figure 2: Example copolar time-series data of model.
3.1.4. A short concluding remark on this model
This model can be used to generate more statistically accu-
rate channel data, which can be used to evaluate the perfor-
mance of polarization time channel codes and algorithms,
and therefore evaluate the capacity and diversity benefits of
MIMO techniques applied to LMS systems. However, it mod-
els usual double-polarized channels/systems only, resulting
in at most 4-fold diversity gain and 2-fold increase in capac-
ity. Taking the generalized 3-dimensional (3D) character of
wave polarization state into account (and applying relevant
antennas), diversity gain can be increased. In terrestrial ap-
plications, capacity can also be increased, however, as we did
show in [15] and briefly discuss here as well, this is not the
case in satellite links. 3D polarization and its application in
PTC will be dealt with in what follows. Note that important
practical issues, like possible loss of capacity due to polar-
ization mismatch, and practical antenna configurations are
beyond the scope of the present paper.
3.2. PTC with 3D polarization satellite antennas
3.2.1. The concept of 3D polarization
Polarization state is characteristic to an electromagnetic
wave. Plane waves are TEM, that is, electric and magnetic
field vectors are in the plane perpendicular to the direction
of propagation. Thus, polarization is a 2-dimensional phe-
nomenon and 2 orthogonal polarization states exist. 2D po-
larization state of a wave, polarization properties of an an-
tenna, as well as functioning of conventional polarization di-
versity and conventional PTC can well be described by the
classical Stokes parameters. (For details see, e.g., [19, 25] for
Péter Horváth et al. 5

1 100

P(fade depth < abscissa)

P(fade depth < abscissa)

0.9 10−1

0.8 10−2
−20 −10 0 10 −45 −40 −35 −30 −25 −20
Power relative to FSL (dB) Power relative to FSL (dB)

Measured copolar Measured copolar

Measured X-polar Measured X-polar
Modeled copolar Modeled copolar
Modeled X-polar Modeled X-polar
(a) (b)

1 100
P(fade depth < abscissa)

10−1
P(fade depth < abscissa)

0.9

10−2

0.8
−20 −15 −10 −5 0 −45 −40 −35 −30 −25 −20
Power relative to FSL (dB) Power relative to FSL (dB)

Measured copolar Measured copolar

Measured X-polar Measured X-polar
Modeled copolar Modeled copolar
Modeled X-polar Modeled X-polar
(c) (d)

Figure 3: Comparison of modeled and measured cumulative distributions; upper figures: highway channel; lower figures: urban channel.

application. It is also mentioned that Stokes parameters form
a 4-vector in a Minkowskian space; their transformation, e.g.,
by scatterers or polarization filters, is a Lorentz transforma-
tion [26]; these properties, however, are not used in this dis-
cussion.)
In the case of multipath propagation (or if the direction
of propagation is unknown), wave polarization is a 3D phe-
nomenon. In that case, the number of orthogonal polariza-
tion states is 3. This can increase the number of orthogo-
nal channels to 3 if these are discriminated by polarization
6 EURASIP Journal on Wireless Communications and Networking
100
10−1
Bit error rate (BER)
10−2
10−3
10−4
10−5
10−6
0 10 20 30 40 50
Eb /N0 (dB)
No diversity PTBC (2 Tx, 1 Rx)
MRRC (1 Tx, 2 Rx) PTBC (2 Tx, 2 Rx)
Figure 4: Bit error rate curves for highway environment.
only; as far as known by the authors, reference [27] was the
first drawing the attention of the MIMO community to this
fact. Combining antenna polarization and radiation pattern
in discriminating channels, this number can be significantly
higher, as this will be briefly discussed in the following sub-
section.
(Note that Stokes parameters together with their symme-
try and invariance properties can be generalized to the 3D
case as well [28]. It is not known by the authors, however,
if these were ever applied in MIMO or communication an-
tenna problems.)
3.2.2. Compact MIMO antennas
If the degree of asynchronism arising in multisatellite-to-
ground links is too high so that synchronization or cooper-
ative diversity is not possible or is too complicated, MIMO
antennas have to be colocated onboard a single satellite. This
situation is similar although not identical to handheld termi-
nals. Like in that case, space is not an available dimension for
diversifying multiple signals: polarization and antenna pat-
tern are only available. It is different on the other hand as
available space is not as much limited as in the case of hand-
held terminals; so the antennas can be large, and aperture or
array antennas of sufficiently high gain can be applied. In re-
cent times, there is a significant progress in the field of com-
pact multielement antennas. We mention three new struc-
tures investigated in the literature.
Reference [12] deals with what is sometimes called a
vector element antenna. This contains 6 rectangular placed
Hertzian dipoles, 3 electric and 3 magnetic. Rectangular elec-
tric and rectangular magnetic dipoles as well as electrical
dipoles parallel to magnetic are fully uncorrelated, while rect-
angular placed electric to magnetic dipoles are of zero or of
very low correlation; the latter is due to different angular pat-
terns. Thus in the case of very rich scattering environment,
6-fold receive diversity gain can be achieved or in principle
even 6 × 6 diversity gain if both the transmitter and the re-
ceiver operate with vector element antennas. Increase in ca-
pacity, however, cannot be more than 4-fold, as shown by
[29].
In [13], the so-called MIMO cube is dealt with. This con-
tains 12 electric dipoles arranged at the edges of a cube.
Cube-to-cube capacity and other parameters are computed,
showing surprisingly good performance; note, however, that
even very small cubes are investigated, (cube edges as short as
0.05λ) the problem of superdirectivity is not stressed in that
paper.
In [14], behaviors of three colocated monopole and
dipole antennas are investigated, versus their mutual angles,
60 via simulation. It is shown that their performance is very
close to ideally orthogonal ones and also that the main cause
of achieving that is their different polarizations rather than
different angular patterns.
3.2.3. Compact antennas and 3D polarization in satellites
There is a significant difference between the environment
of a terrestrial multipath link and a satellite multipath link.
In Figure 5, terrestrial multipath links for indoor or mo-
bile communication are schematically shown. The system de-
picted in Figure 5(a) is of double-bounce scattering, whereas
that of Figure 5(b) is of single bounce. “Compact anten-
nas” are used in both terminals—as an example realized in
the form of triple dipoles. It is self-evident from Figure 5(a)
that waves are arriving to the receive antenna from multiple
directions—resulting in three orthogonal polarization com-
ponents. But the case is similar in situations like Figure 5(b);
this is due to the relatively short distance—characteristic in
terrestrial, in particular in indoor links.
A satellite-to-indoor/mobile link, shown in Figure 6, is
much different, as in this case terminals are (i) very far from
each other and (ii) scatterers are very far from one of these.
Due to (i), antenna must be of high gain, shown in the figure
as an aperture. And, due to (ii), TEM waves travel between
the satellite and the neighborhood of the ground terminal.
Propagation is multipath only in that—relatively short—
distance. The aperture itself can be realized either as a dish
or as an array. It could be illuminated by any 3D polarized
wave, however, only the 2D component of that would travel
towards the ground terminal.
Based on this fact, we have shown in [15] that in a satel-
lite link relative to the single-channel case, only a 2-fold in-
crease of capacity can be achieved by PTC. This is in con-
trast to the terrestrial case in which this increase is 4-fold.
In more details, while any small multielement antenna can
be applied in the ground terminal, onboard one satellite at
most conventional double-polarized antennas are applicable,
or more precisely, are reasonable. On the other hand, diver-
sity can take the full advantage of the capabilities of multi-
ple antennas if these are applied in the ground terminal. As
a consequence of these, this type of channel is asymmetric:
the downlink is a double-input multiple-output channel, the
uplink is its inverse, that is, multiple-input double-output.
Péter Horváth et al. 7

Window
t(t) r(t)

x = 2.8 m
Incident O1 O2 O3
wave

Scattering Scattering
medium medium
(a)
y = 4.5 m

Figure 7: A satellite-to-mobile/indoor link.

t(t) r(t)

In the next section, applying electromagnetic simulation

we verify the capacity and the diversity characteristics as
stated above.
Scattering
medium
4. FDTD SIMULATION OF
(b) A SATELLITE-TO-INDOOR LINK

Figure 5: Terrestrial multipath links with compact MIMO anten-
nas in scattering media; (a) double-bounce scattering; (b) single
bounce.
t(t)
r(t)
Plane wave
Aperture
Scattering
medium
Figure 6: A satellite-to-mobile/indoor link.
This has the consequence that from the coding point of view,
the system is not uniform. If as an example, space-time block
coding of the Alamouti type or orthogonal space-time block
coding (OSTBC) is chosen, RC = 1 can be applied downlink,
however in the uplink RC = 1/2 or at most RC = 3/4 can
only be achieved. (RC designates the coding rate.) It is ques-
tionable if this can be accepted from the frequency economy
point of view. If not, only two of the three or more antennas
are used in the uplink transmitter. Note that other types of
coding can give different results.
On the other hand, the number of diversity routes is
increased—say up to 2 × 3. (This is valid if terminal antenna
is a tripole; with a vector element antenna, this is 2 × 6, with
a MIMO cube even 2 × 12.)
In order to assess the performance of using three orthog-
onally polarized antennas in a satellite-to-indoor scenario,
some simulations were performed using full-wave electro-
magnetic tools. The FDTD method [30] was used to calculate
the time-dependent electromagnetic field inside a typical of-
fice room where the mobile terminal is assumed to be placed.
The office dimensions were 2.8 m × 4.5 m × 3.0 m (x, y, z),
where the floor and the ceiling are lying in and parallel to the
x-y plane, respectively, as seen in Figure 7. In the simulation,
the furniture and the walls of the room are modeled by re-
alistic material properties (brick walls, wooden and metallic
furniture, and some plastic objects). These objects of vari-
ous geometries are nearly uniformly distributed in the room.
Linear orthogonally polarized plane waves enter the room
through the window and through the external wall; one po-
larization during the first simulation run and the other one
during a subsequent run. This method allows us to split the
channel response according to the incoming polarizations.
The waveform is a modulated Gaussian pulse centered at
1.2 GHz, entering through the x-z plane at y = 0 m.
The electric field components (Ex , E y , and Ez ) are
recorded at various spots in the room. We use these field
components directly to draw conclusions about the signals
(voltages) which three antennas would produce if they would
be placed at a given observation point. Although this ap-
proach does not consider the current distribution on elec-
trically long antennas, mutual coupling, scattering by the an-
tennas, and so forth, previous FDTD studies demonstrated
that only a very low crosstalk exists between three thin-wire
half-wave dipoles which are mounted parallel to the coor-
dinate axes in an empty room [16]. Therefore, the results
can be regarded as realistic, for short orthogonally mounted
dipoles. The field components are recorded along various
x-z cross-sections of the room, at three different observa-
tion planes (O1 at y = 1.5 m, O2 at y = 2.4 m, and O3
8 EURASIP Journal on Wireless Communications and Networking

at y = 4 m), representing different propagation environ- 1

ments due to different shadowing and angle-of-incidence pa- 0.9
rameters. At each of the three planes, about 800 points were
0.8
observed, spaced 7.5 cm apart in both x and z directions. In a

Probability (CEP < abscissa)

first scenario (S1), the incident waves arrive horizontally (at 0.7
0 elevation and parallel to y-axis). In a second scenario (S2), 0.6
the elevation was chosen to be 30 degrees and the azimuth
0.5
angle 20 degrees off the y-axis. Thus, in the latter case, the
line of sight is blocked at the points of O2 and O3. For each 0.4
scenario, two simulation runs yielded 6 time functions of the 0.3
fields (Ex , E y , and Ez when using the one or the other po-
larization). From the observed fields, which were regarded as 0.2
received voltages according to the reasoning presented above, 0.1
signal portions weaker than a designated noise level, chosen 0
to be −15 dB relative to the maximum power level, were dis- 0 2 4 6 8 10 12 14 16 18 20
carded. Then the envelope of the received signals was calcu- Capacity (bits/s/Hz)
lated. Based on these data, three statistical parameters were
nT = 1; nR = 1
derived for both Scenarios 1 and 2. First, the equal-power nT = 2; nR = 2
capacity [31, Equation (4)], was calculated and its CDF was nT = 2; nR = 3
determined. In Figures 8 and 11, the capacity CDF curves
are shown for S1 and S2, respectively. As expected, at low Figure 8: CDF of the equal-power capacity (Scenario 1).
outage, levels the capacity of the dual-polarized TX, dual-
polarized RX antenna, (2, 2) and (2, 3) systems is about twice
that of the (1, 1) SISO system, and the difference between the
1
(2, 2) and the (2, 3) systems is rather small. In order to as-
sess the diversity performance, the envelope correlation [32] 0.9
was determined between the received signals (latter being the 0.8
correlation coefficient between the envelopes of the received
Probability (ρe < abscissa)

signals). Their CDFs are shown in Figures 9 and 12. As ex- 0.7
pected, in Scenario 2, lower (even negative) correlation is to 0.6
be expected. Additionally, the relative received signal power 0.5
for the (1, 1), (2, 2), and (2, 3) systems and their CDF was also
determined, which results are shown in Figures 10 and 13 0.4
for the scenarios in consideration. Note that the confidence 0.3
for very low-probability (less than 0.01 or so) portions of the
0.2
curve might be low due to the relatively low number (about
2000) of observations, but still validates the claim based on 0.1
the higher probability portion of the curves. 0
−0.2 0 0.2 0.4 0.6 0.8 1
Envelope correlation
5. CONCLUSIONS
ρH y ρV y
ρHz ρVx
The main statement of this paper is that the generalized
coded form of polarization diversity is a very good—maybe
the best—way to apply the MIMO concept in multipath Figure 9: CDF of the envelope correlation (Scenario 1).
satellite links. Two main contributions are related to the
modeling of the conventional (2D) polarization diversity
channel and to the investigation via simulation of the 3D
MIMO channel, respectively. (The relevant signal processing acteristics are investigated. The main purpose of this study
is called here PTC.) was to verify (for this example) the findings of two of these
Concerning the first of these (modeling), a physical sta- authors [15] about the capacity and diversity characteristics
tistical model is given for the urban and the highway satellite of this type of channels. Results of this simulation are as fol-
mobile channels. Besides giving a validated model, it veri- lows.
fies once again the authors’ conviction that the best type of a From the capacity point of view, (i) the difference be-
multipath channel model is of the physical-statistical type. tween the 2 × 2 and the 2 × 3 cases is negligible (as stated in
Concerning the second of these (simulation), a very ex- [15]); and (ii) with high probability capacity of the MIMO,
tensive simulation study is carried out about the 3D polar- the situation is nearly exactly 2-times as high as that of the
ization characteristics of the satellite multipath channel. A SISO case, again in accordance with [15]. (Note that with low
synthetic satellite-to-indoor link is simulated and PTC char- probability, this difference is higher.)
Péter Horváth et al. 9

100 1
0.9
0.8

Probability (ρe < abscissa)

Probability (Pr < abscissa)

10−1 0.7
0.6
0.5

10−2 0.4
0.3
0.2

10−3 0.1
0
−50 −45 −40 −35 −30 −25 −20 −15 −10 −0.2 0 0.2 0.4 0.6 0.8 1
Combined received power (dBm) Envelope correlation

nT = 1; nR = 1 ρH y ρV y
nT = 2; nR = 2 ρHz ρVx
nT = 2; nR = 3
Figure 12: CDF of the envelope correlation (Scenario 2).
Figure 10: CDF of the received power (Scenario 1).

1 100
0.9

0.8
Probability (CEP < abscissa)

Probability (Pr < abscissa)

0.7 10−1

0.6

0.5
0.4 10−2

0.3
0.2
0.1 10−3

0
0 2 4 6 8 10 12 14 16 18 −50 −45 −40 −35 −30 −25 −20 −15 −10
Capacity (bits/s/Hz) Combined received power (dBm)

nT = 1; nR = 1 nT = 1; nR = 1
nT = 2; nR = 2 nT = 2; nR = 2
nT = 2; nR = 3 nT = 2; nR = 3

Figure 11: CDF of the equal-power capacity (Scenario 2). Figure 13: CDF of the received power (Scenario 2).

To characterize the diversity performance, CDF of the re-
ceived power in the various situations is investigated; result
shows that 3-fold (i.e., 3D) polarization diversity yields sig-
nificantly higher received power than the 2-fold diversity (or
the nondiversity case).
From the simulation point of view, this study shows that
the FDTD method is very well applicable to investigate in an
exact way such extremely complex structures as the one here.
A statement of this paper (stated but not discussed in detail)
talking about satellite-diversity-MIMO, the problems briefly
dealt with in Section 3, that is, the effect of extremely large
and variable difference between the path-lengths of MIMO
branches must be taken into account.
ACKNOWLEDGMENTS
This work was done in the framework of and is supported by
the project SatNEx of the EU IST FP6 Program. Their sup-
port is gratefully acknowledged.
10
REFERENCES
[1] Z. Bodnar, Z. Herczku, J. Berces, et al., “A detailed experimen-
tal study of the LEO satellite to indoor channel characteristics,”
International Journal of Wireless Information Networks, vol. 6,
no. 2, pp. 79–91, 1999.
[2] P. R. King, B. G. Evans, and S. Stavrou, “Physical-statistical
model for the land mobile-satellite channel applied to satel-
lite/HAP MIMO,” in Proceedings of the 11th European Wireless
Conference, vol. 1, pp. 198–204, Nicosia, Cyprus, April 2005.
[3] T. Hult and A. Mohammed, “MIMO antenna applications for
LEO satellite communications,” in Proceedings of the 3rd ESA
International Workshop of the European COST 280, Prague,
Czech Republic, June 2005.
[4] F. Yamashita, K. Kobayashi, M. Ueba, and M. Umehira,
“Broadband multiple satellite MIMO system,” in Proceedings
of the 62nd IEEE Vehicular Technology Conference (VTC ’05),
vol. 4, pp. 2632–2636, Dallas, Tex, USA, September 2005.
[5] K. Liolis, A. Panagopoulos, and P. Cottis, “Outage capacity
statistics of MIMO satellite networks operating at Ka band and
above,” in Proceedings of the 12th Ka and Broadband Commu-
nications Conference, Naples, Italy, September 2006.
[6] J. Mietzner and P. A. Hoeher, “Distributed space-time codes
for cooperative wireless networks in the presence of different
propagation delays and path losses,” in Proceedings of the IEEE
Sensor Array and Multichannel Signal Processing Workshop, pp.
264–268, Barcelona, Spain, July 2004.
[7] J. N. Laneman and G. W. Wornell, “Distributed space-time-
coded protocols for exploiting cooperative diversity in wireless
networks,” IEEE Transactions on Information Theory, vol. 49,
no. 10, pp. 2415–2425, 2003.
[8] M. Janani, A. Hedayat, T. E. Hunter, and A. Nosratinia,
“Coded cooperation in wireless communications: space-time
transmission and iterative decoding,” IEEE Transactions on
Signal Processing, vol. 52, no. 2, pp. 362–371, 2004.
[9] A. Sendonaris, E. Erkip, and B. Aazhang, “User cooperation
diversity—part I: system description,” IEEE Transactions on
Communications, vol. 51, no. 11, pp. 1927–1938, 2003.
[10] A. Sendonaris, E. Erkip, and B. Aazhang, “User cooper-
ation diversity—part II: implementation aspects and per-
formance analysis,” IEEE Transactions on Communications,
vol. 51, no. 11, pp. 1939–1948, 2003.
[11] H. T. Cheng, H. Mheidat, M. Uysal, and T. M. Lok, “Dis-
tributed space-time block coding with imperfect channel esti-
mation,” in Proceedings of the IEEE International Conference on
Communications (ICC ’05), vol. 1, pp. 583–587, Seoul, South
Korea, May 2005.
[12] T. Svantesson, M. A. Jensen, and J. W. Wallace, “Analysis of
electromagnetic field polarizations in multiantenna systems,”
IEEE Transactions on Wireless Communications, vol. 3, no. 2,
pp. 641–646, 2004.
[13] B. N. Getu and J. B. Andersen, “The MIMO cube—a compact
MIMO antenna,” IEEE Transactions on Wireless Communica-
tions, vol. 4, no. 3, pp. 1136–1141, 2005.
[14] L. Dong, H. Choo, R. W. Heath Jr., and H. Ling, “Simulation of
MIMO channel capacity with antenna polarization diversity,”
IEEE Transactions on Wireless Communications, vol. 4, no. 4,
pp. 1869–1873, 2005.
[15] P. Horváth and I. Frigyes, “Application of the 3D polariza-
tion concept in satellite MIMO systems,” in Proceedings of
the 49th Annual IEEE Global Telecommunications Conference
(GLOBECOM ’06), San Francisco, Calif, USA, November-
December 2006.
EURASIP Journal on Wireless Communications and Networking
[16] P. Horváth and I. Frigyes, “Investigation of the polarization
properties of satellite channels with multiple antennas,” in
Proceedings of the 1st European Conference on Antennas and
Propagation (EuCAP ’06), Nice, France, November 2006.
[17] P. Elia and P. Kumar, “Constructions of cooperative diversity
schemes for asynchronous wireless networks,” in Proceedings
of IEEE International Symposium on Information Theory, pp.
2724–2728, Seattle, Wash, USA, July 2006.
[18] S. Wei, D. L. Goeckel, and M. C. Valenti, “Asynchronous co-
operative diversity,” IEEE Transactions on Wireless Communi-
cations, vol. 5, no. 6, pp. 1547–1557, 2006.
[19] I. Frigyes and P. Horváth, “Polarization-time coding in satellite
links,” IEEE Satellite and Space Newsletter, vol. 15, no. 2, pp. 6–
8, 2005.
[20] S. M. Alamouti, “A simple transmit diversity technique for
wireless communications,” IEEE Journal on Selected Areas in
Communications, vol. 16, no. 8, pp. 1451–1458, 1998.
[21] I. Frigyes, B. G. Molnár, Z. Herczku, and Z. Bodnár, “Antenna
gain and polarization effects in wireless links—accent on LEO
satellites,” Space Communications, vol. 19, no. 3-4, pp. 199–
208, 2004.
[22] I. H. Cavdar, H. Dincer, and K. Erdogdu, “Propagation mea-
surements at L-band for land mobile satellite link design,” in
Proceedings of the 7th Mediterranean Electrotechnical Confer-
ence (MELECON ’94), vol. 3, pp. 1162–1165, Antalya, Turkey,
April 1994.
[23] P. R. King and S. Stavrou, “Low elevation wideband land mo-
bile satellite MIMO channel characteristics,” to appear in IEEE
Transactions on Wireless Communications.
[24] P. R. King and S. Stavrou, “Capacity improvement for a land
mobile single satellite MIMO system,” IEEE Antennas and
Wireless Propagation Letters, vol. 5, no. 1, pp. 98–100, 2006.
[25] M. Born and E. Wolf, Principles of Optics, Cambridge Univer-
sity Press, Cambridge, UK, 1998.
[26] D. Han, Y. S. Kim, and M. E. Noz, “Stokes parameters as a
Minkowskian four-vector,” Physical Review E, vol. 56, no. 5,
pp. 6065–6076, 1997.
[27] M. R. Andrews, P. P. Mitra, and R. de Carvalho, “Tripling the
capacity of wireless communications using electromagnetic
polarization,” Nature, vol. 409, no. 6818, pp. 316–318, 2001.
[28] J. J. Gil, J. M. Correas, P. A. Melero, and C. Ferreira, “Gen-
eralized polarization algebra,” http://www.unizar.es/galdeano/
actas pau/PDFVIII/pp161-167.pdf.
[29] T. L. Marzetta, “Fundamental limitations on the capacity of
wireless links that use polarimetric antenna arrays,” in Pro-
ceedings of IEEE International Symposium on Information The-
ory, p. 51, Lausanne, Switzerland, June-July 2002.
[30] A. Taflove and S. C. Hagness, Computational Electrodynamics:
The Finite-Difference-Time-Domain Method, Artech House,
Norwood, Mass, USA, 2006.
[31] D. Gesbert, M. Shafi, D.-S. Shiu, P. J. Smith, and A. Naguib,
“From theory to practice: an overview of MIMO space-time
coded wireless systems,” IEEE Journal on Selected Areas in
Communications, vol. 21, no. 3, pp. 281–302, 2003.
[32] R. G. Vaughan and J. B. Andersen, “Antenna diversity in mo-
bile communications,” IEEE Transactions on Vehicular Tech-
nology, vol. 36, no. 4, pp. 149–172, 1987.
Hindawi Publishing Corporation
EURASIP Journal on Wireless Communications and Networking
Volume 2007, Article ID 25361, 9 pages
doi:10.1155/2007/25361

Research Article
Performance Analysis of SSC Diversity Receivers over
Correlated Ricean Fading Satellite Channels

Petros S. Bithas and P. Takis Mathiopoulos

Institute for Space Applications and Remote Sensing, National Observatory of Athens, Metaxa and Vas. Pavlou Street,
15236 Athens, Greece
Received 3 October 2006; Revised 23 February 2007; Accepted 6 April 2007

Recommended by Ray E. Sheriff

This paper studies the performance of switch and stay combining (SSC) diversity receivers operating over correlated Ricean fading
satellite channels. Using an infinite series representation for the bivariate Ricean probability density function (PDF), the PDF of
the SSC output signal-to-noise ratio (SNR) is derived. Capitalizing on this PDF, analytical expressions for the corresponding cu-
mulative distribution function (CDF), the moments of the output SNR, the moments generating function (MGF), and the average
channel capacity (CC) are derived. Furthermore, by considering several families of modulated signals, analytical expressions for
the average symbol error probability (ASEP) for the diversity receivers under consideration are obtained. The theoretical analy-
sis is accompanied by representative performance evaluation results, including average output SNR (ASNR), amount of fading
(AoF), outage probability (Pout ), average bit error probability (ABEP), and average CC, which have been obtained by numerical
techniques. The validity of some of these performance evaluation results has been verified by comparing them with previously
known results obtained for uncorrelated Ricean fading channels.

Copyright © 2007 P. S. Bithas and P. T. Mathiopoulos. This is an open access article distributed under the Creative Commons
Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is
properly cited.

1. INTRODUCTION
The mobile terrestrial and satellite communication channel
is particularly dynamic due to multipath fading propaga-
tion, having a strong negative impact on the average bit er-
ror probability (ABEP) of any modulation scheme [1]. Di-
versity is a powerful communication receiver technique used
to compensate for fading channel impairments. The most
important and widely used diversity reception methods em-
ployed in digital communication receivers are maximal-ratio
combining (MRC), equal-gain combining (EGC), selection
combining (SC), and switch and stay combining (SSC) [2].
For SSC diversity considered in this paper, the receiver se-
lects a particular branch until its signal-to-noise ratio (SNR)
drops below a predetermined threshold. When this happens,
the combiner switches to another branch and stays there re-
gardless of whether the SNR of that branch is above or be-
low the predetermined threshold. Hence, among the above-
mentioned diversity schemes, SSC is the least complex and
can be used in conjunction with coherent, noncoherent, and
differentially coherent modulation schemes. It is also well
known that in many real life communication scenarios the
combined signals are correlated [2, 3]. A typical example for
such signal correlation exists in relatively small-size mobile
terminals where typically the distance between the diversity
antennas is short. Due to this correlation between the signals
received at the diversity branches there is a degradation in the
achievable diversity gain.
The Ricean fading distribution is often used to model
propagation paths consisting of one strong direct line-of-
sight (LoS) signal and many randomly reflected and usually
weaker signals. Such fading environments are typically en-
countered in microcellular and mobile satellite radio links
[2]. In particular for mobile satellite communications the
Ricean distribution is used to accurately model the mo-
bile satellite channel for single- [4] and clear-state [5] chan-
nel conditions. Furthermore, in [6] it was depicted that the
Ricean K-factor characterizes the land mobile satellite chan-
nel during unshadowed periods.
The technical literature concerning diversity receivers op-
erating over correlated fading channels is quite rich, for ex-
ample, see [7–13]. In [7] expressions for the outage probabil-
ity (Pout ) and the ABEP of dual SC with correlated Rayleigh
fading were derived either in closed form or in terms of
2 EURASIP Journal on Wireless Communications and Networking
single integrals. In [8] the cumulative distribution functions
(CDF) of SC, in correlated Rayleigh, Ricean, and Nakagami-
m fading channels were derived in terms of single-fold in-
tegrals and infinite series expressions. In [9] the ABEP of
dual-branch EGC and MRC receivers operating over corre-
lated Weibull fading channels was obtained. In [10] the per-
formance of MRC in nonidentical correlated Weibull fad-
ing channels with arbitrary parameters was evaluated. In
[11] an analysis for the Shannon channel capacity (CC) of
dual-branch SC diversity receivers operating over correlated
Weibull fading was presented. In [12], infinite series expres-
sions for the capacity of dual-branch MRC, EGC, SC, and
SSC diversity receivers over Nakagami-m fading channels
have been derived.
Past work concerning the performance of SSC operat-
ing over correlated fading channels can be found in [14–
17]. One of the first attempts to investigate the performance
of SSC diversity receivers operating over independent and
correlated identical distributed Ricean fading channels was
made in [14]. However, in this reference only noncoher-
ent frequency shift keying (NCFSK) modulation was con-
sidered and its ABEP has been derived in an integral rep-
resentation form. In [15] the performance of SSC diversity
receivers was investigated for different fading channels, in-
cluding Rayleigh, Nakagami-m and Ricean, and under dif-
ferent channel conditions but dealt mainly with uncorre-
lated fading. For correlated fading in this reference only the
Nakagami-m distribution was studied. In [16] the moments
generating function (MGF) of SSC was presented in terms of
a finite integral representation for the correlated Nakagami-
m fading channel. In [17] expressions for the average output
SNR (ASNR), amount of fading (AoF) and Pout for the cor-
related log-normal fading channels have been derived.
All in all, the problem of theoretically analyzing the per-
formance of SSC over correlated Ricean fading channels has
not yet been thoroughly addressed in the open technical lit-
erature. The main difficulty in analyzing the performance of
diversity receivers in correlated Ricean fading channels is the
complicated form of the received signal bivariate probability
density function (PDF), see [14, Equation (17)], and the ab-
sence of an alternative and more convenient expression for
the multivariate distribution. An efficient solution to these
difficulties is to employ an infinite series representation for
the bivariate PDF, such as those that were proposed in [18]
or [19]. Such an approach was used in [20] to analyze the per-
formance of MRC, EGC, and SC in the presence of correlated
Ricean fading. Similarly here the most important statistical
metrics and the capacity of SSC diversity receivers operat-
ing over correlated Ricean fading channels will be studied. In
particular, we derive the PDF, CDF, MGF, moments and the
average CC of such receivers operating over correlated Ricean
fading channels. Furthermore, analytical expressions for the
average symbol error probability (ASEP) of several modula-
tion schemes will be obtained. Capitalizing on these expres-
sions, a detailed performance analysis for the Pout , ASNR,
AoF, and ASEP/ABEP will be presented.
The remainder of this paper is organized as follows. Af-
ter this introduction, in Section 2 the system model is intro-
duced. In Section 3, the SSC received signal statistics are pre-
sented, while in Section 4 the capacity is obtained. Section 5
contains the derivation of the most important performance
metrics of the SSC output SNR. In Section 6, various numer-
ical evaluation results are presented and discussed, while the
conclusions of the paper can be found in Section 7.
2. SYSTEM MODEL
By considering a dual-branch SSC diversity receiver operat-
ing over a correlated Ricean fading channel, the baseband re-
ceived signal at the th ( = 1 and 2) input branch can be
mathematically expressed as
ζ = sh + n . (1)
In the above equation, s is the transmitted complex sym-
bol, h is the Ricean fading channel complex envelope with
magnitude R = |h |, and n is the additive white Gaus-
sian noise (AWGN) having single-sided power spectral den-
sity of N0 . The usual assumption for ideal fading phase esti-
mation is made, and hence, only the distributed fading enve-
lope and the AWGN affect the received signal. Moreover, the
AWGN is assumed to be uncorrelated between the two diver-
sity branches. The instantaneous SNR per symbol at the th
input branch is γ = R2 Es /(2N0 ), where Es = E|s|2  is the
transmitted average symbol energy, where E· denoting ex-
pectation and | · | absolute value. The corresponding average
SNR per symbol at both input branches is γ = ΩEs /N0 , where
Ω = ER2 . The PDF of the SNR of the Ricean distribution
is given by [2, Equation (2.16)]
 
1+K (1 + K)
fγ (γ) = exp − K − γ
γ γ
   (2)
K(K + 1) 1/2
× I0 2 γ ,
γ
where K is the Ricean K-factor defined as the power ratio
of the specular signal to the scattered signals and I0 (·) is the
zeroth-order modified Bessel function of the first kind [21,
Equation (8.406)]. The CDF of γ is given by [14, Equation
(8)]
  
√ 2(1 + K)
Fγ (γ) = Q1 2K, γ , (3)
γ
where Q1 (·) is the first-order Marcum-Q function [2, Equa-
tion (4.33)].
The joint PDF of γ1 and γ2 , presented in [14, Equation
(17)], can be expressed in terms of infinite series by follow-
ing a similar procedure as for deriving [18, Equation (9)].
Hence, substituting I0 (·) with its infinite series representa-
tion [21, Equation (8.445)], expanding the term [γ1 + γ2 +
√
2 γ1 γ2 cos(θ)]i using the multinomial identity [22, Equa-
tion (24.1.2)], using [21, Equation (3.389/1)] and after some
P. S. Bithas and P. T. Mathiopoulos 3

mathematical manipulations the joint PDF of γ1 , γ2 can be Hence, by substituting (4) in (7) and using [21, Equation
expressed as (3.351/2-3)], these integrals can be solved and rssc (γ) can be
expressed as
∞

 
  ∞
fγ1 ,γ2 γ1 , γ2 = A exp − β1 γ1 + γ2  
i,h=0
rssc (γ) = A exp − β1 γ γβ2 −1/2
v1 +v2 +v3 =i i,h=0
v1 +v2 +v3 =i
 β2 −1 β3 −1 β −1/2 β3 −1/2 
    
× Bγ1 γ2 + Cγ−1 γ1 2 γ2 Bγ β3 , β1 γτ Cγ β3 + 1/2, β1 γτ
× √ β + β +1/2 ,
(4) γβ1 3 γβ1 3
(8)
with

 where γ(·, ·) is the lower incomplete Gamma function [21,
2v3 +2h−1 (1 + K)1+β4 ρ2h K i exp − 2K/(1 + ρ) Equation (8.350/1)].
A= √  1+2h ,
πγ1+β4 1 − ρ2 v1 !v2 !v3 !i!(1 + ρ)2i
3.2. Cumulative distribution function (CDF)


  
1 + (−1)v3 Γ h + 1 + v3 /2 Similar to [23, Equation (20)], the CDF of γssc , Fγssc (γ), is
B=     ,
Γ h + 1 + v3 /2 Γ 1 + 2h given by
   

   Fγssc (γ) = Pr γτ ≤ γ1 ≤ γ + Pr γ2 < γτ ∧ γ1 < γ (9)
− 1 + (−1)v3 2ρ(1 + K)Γ 1 + h + v3 /2
C=  
   ,
ρ2 − 1 Γ(2 + 2h)Γ h + 3 + v3 /2 which after some manipulations can be expressed in terms of
CDFs as
⎧  
(1 + K) v3 ⎪
⎨Fγ1 ,γ2 γ, γτ , γ ≤ γτ ,
β1 =   , β 2 = v1 + + h + 1,
1 − ρ2 γ 2 Fγssc (γ) = ⎪    
⎩Fγ (γ) − Fγ γτ + Fγ ,γ γ, γτ , γ > γτ .
1 2

v (10)
β3 = v2 + 3 + h + 1, β4 = i + 2h + 1,
2  γ  γτ
(5) Hence, by substituting (4) in Fγ1 ,γ2 (γ, γτ ) = 0 0 fγ1 ,γ2 (γ1 ,
γ2 )dγ1 dγ2 using [21, Equation (3.351/1)], Fγ1 ,γ2 (γ, γτ ) can be
where Γ(·) is the Gamma function [21, Equation (8.310/1)] derived as
and ρ is the correlation coefficient between γ1 and γ2 . It can ∞

  A
be proved that the above infinite series expression always Fγ1 ,γ2 γ, γτ = β +β
converges [18]. i,h=0 β1 2 3
v1 +v2 +v3 =i

   
3. RECEIVED SIGNAL STATISTICS × Bγ β2 , β1 γ γ β3 , β1 γτ

In this section, the most important statistical metrics,    

C 1 1
namely, the PDF, CDF, MGF, and moments of dual branch + γ β2 + , β1 γ γ β3 + , β1 γτ .
SSC output SNR diversity receivers operating over correlated β1 γ 2 2
Ricean fading channels will be presented. (11)
In order to verify the validity of the above derivations,
3.1. Probability density function (PDF) (10) and (11) have been numerically evaluated for the spe-
cial case of uncorrelated, that is, ρ = 0, Ricean fading chan-
Let γssc be the instantaneous SNR per symbol at the output of nels. The resulting CDF was found to be identical to the same
the SSC and γτ the predetermined switching threshold. Fol- CDF presented in [2, Equation 9.273], which was derived us-
lowing [15], the PDF of γssc , fγssc (γ), is given by ing a different mathematical approach as a closed-form ex-
⎧ pression.
⎪
⎨rssc (γ), γ ≤ γτ ,
fγssc (γ) = ⎪ (6) 3.3. Moments generating function (MGF)
⎩r (γ) + f (γ), γ > γτ .
ssc γ
Based on (6), the MGF of γssc , Mγssc (s) = Eexp(−sγssc ), [24,
Moreover, rssc (γ) is given in [23, Equation (21b)] as Equation (5.62)], can be expressed in terms of two integrals
 γτ as
  ∞
rssc (γ) = fγ1 γ2 γ, γ2 dγ2 Mγssc (s) = exp(−sγ)rssc (γ)dγ
0
0
∞ ∞ ∞
   
= fγ1 γ2 γ, γ2 dγ2 − fγ1 γ2 γ, γ2 dγ2 . + exp(−sγ) fγ (γ)dγ = I1 + I2 .
0 γτ γτ
(7) (12)
4 EURASIP Journal on Wireless Communications and Networking

Using [21, Equation (3.381/4)], I1 can be expressed in terms where BW is transmission bandwidth of the signal in Hz.
of infinite series as Hence, substituting (6) in (18), C becomes
∞
   ∞ ∞
Γ β2 −β3  
I1 = A  β2 Bβ1 γ β3 , β1 γτ C= log2 (1 + γ)rssc (γ)dγ + log2 (1 + γ) fγ (γ)dγ
β1 + s 0 γτ
i,h=0 (19)
v1 +v2 +v3 =i
 = I5 + I6 .
 
−β3 −1/2 Γ β2 +1/2 1 
  
+Cβ1  β2 +1/2 γ β3+ , β1 γτ . By representing ln(1 + γ) = G1,2 1,1
β1 +s 2 2,2 γ | 1,0 , [27, Equation
 
(13) (01.04.26.0003.01)], and exp(−γ) = G1,00,1 γ | − , [27, Equa-
0

 tion (01.03.26.0004.01)], where G(·) is Meijer’s G-function

Setting ψ = 2γ[(1 + K)/γ + s] and using [2, Equation [21, Equation (9.301)] and using [28], I5 can be solved as
(4.33)], I2 can be solved as      
∞

   A γ β3 , β1 γτ 1,3 1  1, 1, 1 − β2
I5 = B G3,2
2K(1 + K) 
 2(1 + K + γ s)γτ ln 2 β +β
β1 3 2 β1  1, 0
I 2 = Q1 , i,h=0
v1 +v2 +v3 =i
1 + K + γs γ
 
(14)  
K(1 + K) (1 + K) exp(−K) γ β3 + 1/2, β1 γτ
× exp . +C β +β2 +3/2
1 + K + sγ 1 + K + γs β1 3
   
3.4. Moments 1,3 1  1, 1, 1 − β2 .
× G3,2 
β1 1, 0
Based on (6), the moments for γssc , μγssc (n) = Eexp(γssc
n ),
(20)
[24, Equation (5.38)], can be expressed in terms of two inte-
grals as Due to the very complicated nature of I6 , it is very difficult,
∞ ∞ if not impossible, to derive a closed-form solution for this
μγssc (n) = γn rssc (γ)dγ + γn fγ (γ)dγ integral. However, I6 can be evaluated via numerical inte-
0 γτ
(15) gration using any of the well-known mathematical software
packages, such as MATHEMATICA or MATLAB.
= I3 + I4 .

Using again [21, Equation (3.381/4)], I3 can be expressed in 5. PERFORMANCE ANALYSIS

terms of infinite series as In this section a detailed performance analysis, in terms of
∞    Pout , ASEP, ASNR and AoF, for SSC diversity receivers operat-
  Γ n + β2
I3 = A Bγ β3 , β1 γτ β +β3 +n ing over correlated Ricean fading channels will be presented.
i,h=0 β1 2
v1 +v2 +v3 =i
    5.1. Outage probability (Pout )
Cγ β3 + 1/2, β1 γτ 1
+ β +β3 +n+1 Γ n + β2 + .
β1 2 2 Pout is the probability that the output SNR falls below a pre-
(16) determined threshold γth , Pout (γth ), and can be obtained by
replacing γ with γth in (10) as

Setting φ = 2γ(1 + K)/γ in I4 , using [2, Equation    
Pout γth = Fγssc γth . (21)
(4.104)], after some straight-forward mathematical manip-
ulations, yields
   5.2. Average symbol error probability (ASEP)
γn−1  2(1 + K)γ
 τ
I4 = n Q 2n+1,0 K, , (17)
2 (1 + K)n−1 γ The ASEP, P se , can be evaluated directly by averaging the con-
ditional symbol error probability, Pe (γ), over the PDF of γssc
where Qm,n (·, ·) is the Nuttal Q-function defined in [25]. [29]
∞
4. CHANNEL CAPACITY (CC) P se = Pe (γ) fγssc (γ)dγ. (22)
0
CC is a well-known performance metric providing an upper For different families of modulation schemes, Pe (γ) can
bound for maximum errorless transmission rate in a Gaus- be obtained as follows.
sian environment. The average CC, C, is defined as [26] (i) For binary phase shift keying (BPSK) and square M-
∞ ary quadrature amplitude modulation (QAM) signaling for-
Δ 
C = BW log2 (1 + γ) fγssc (γ)dγ, (18) mats and for high-input SNR, Pe (γ) = D erfc( Eγ), where
0
P. S. Bithas and P. T. Mathiopoulos 5

erfc(·) is the complementary error function [21, Equation In (25), Γ(·, ·) is the upper incomplete Gamma function [22,
(8.250/1)] and D, E are constants the values of which depend Equation (6.51)].
on the specific modulation scheme under consideration. Us- (ii) For noncoherent binary frequency shift keying
ing this expression, by substituting (6) in (22), yields (BFSK) and binary differential phase shift keying (BDPSK),
∞  ∞  Pe (γ) = D exp(−Dγ). Similar to the derivation of (12), that
P se = D erfc Eγ rssc (γ)dγ + D erfc Eγ fγ (γ)dγ is, using [21, Equation (3.381/4)] and [2, Equation (4.33)],
0 γτ P se can be expressed as
= I7 + I8 . ∞

(23) P se = AD
i,h=0
 √   v1 +v2 +v3 =i
Expressing erfc( Eγ) = π −1 G2,0 1
1,2 Bγ | 0,1/2 , [27, Equation   
  Γ β2 B  
(06.27.26.0006.01)], and exp(−γ) = − , [27, Equa- G1,0
0,1 γ|0 ×  β2 β γ β3 , β1 γτ
tion (01.03.26.0004.01)], using [28] and after some straight- β1 + E β1 3
forward mathematical manipulations I7 can be expressed as    
CΓ β2 + 1/2 1
  + β2 +1/2 β3 +1/2 γ β3 + , β1 γτ
∞
ADΓ β2 + 1/2 β1 + E β1 2
I7 = √ 
β
πβ1 3 Eβ2  
i,h=0
v1 +v2 +v3 =i 2K(1 + K)   2(1 + K + γE)γτ
+ Q1 ,
   1 + K + γE γ
BΓ β2    
×   γ β3 , β1 γτ K(1 + K) (1 + K) exp(−K)
Γ β2 + 1 × exp .
1 + K + γE 1 + K + γE
 
1 β1 (26)
× 2 F1 β2 , β2 + ; β2 + 1; −
2 E
(iii) For Gray encoded M-ary PSK and M-ary DPSK,
    Λ
Cγ β3 + 1/2, β1 γτ Γ β2 + 1 Pe (γ) = D 0 exp[−E(θ)γ]dθ, where Λ is constant. Thus, P se
+  
 1/2 3 can be expressed as
β1 E Γ β2 +
2 ∞

  P se = AD
1 3 β1 i,h=0
× 2 F1 β2 + , β2 + 1; β2 + ; − v1 +v2 +v3 =i
2 2 E
(24)    Λ  
Bγ β3 , β1 γτ Γ β2
× β
β2 dθ
with 2 F1 (·, ·; ·; ·) being Gauss Hypergeometric
 function [21, β1 3 0 β1 + E(θ)
∞
Equation (9.100)]. Moreover, I8 = 0 D erfc( Eγ) fγ (γ)dγ −  
 γτ  Cγ β3 + 1/2, β1 γτ
+
0 D erfc( Eγ) fγ (γ)dγ = I8,a − I8,b . Hence, substituting β +1/2
β1 3
again I0 (·) with its infinite series representation [21, Equa-
Λ  
tion (8.445)], I8,a can be solved with the aid of [28] and I8,b Γ β2 + 1/2)
using [27, Equation (06.27.21.0019.01)]. Thus, using these ×
β2 +1/2 dθ
0 β1 + E(θ)
solutions of I8,a and I8,b and after some mathematical ma-
nipulations, I8 can be expressed as in (25): Λ   
2K(1 + K)  2g(θ)γτ
∞  k + Q1 ,
D(1 + K) exp(−K) K(K + 1) 0 g(θ) γ
I8 = (k!)−2  
γ k=0
γ K(1 + K) (1 + K) exp(−K)
× exp dθ,
 g(θ) g(θ)
Γ(k + 1)Γ(k + 3/2) (27)
× √ k+1
πE Γ(k + 2)
  where g(θ) = 1 + K + γE(θ). The above finite integrals can be
3 1+K easily evaluated via numerical integration.
× 2 F1 k + 1, k + ; k + 2; −
2 γE
√ ∞
ρ 5.3. Average output SNR (ASNR) and
2 E/π − (1 + K)/γ Eρ
−
 k+3/2 amount of fading (AoF)
β1 1 − ρ 2 ρ=0
(2ρ + 1)ρ!
The ASNR, γssc , is a useful performance measure serving as
 

3 (1+K)γτ Γ k+1, (1+K)γτ /γ an excellent indicator for the overall system fidelity and can
× Γ k+ +ρ, −
 k+1 . be obtained from the first-order moment of γssc as
2 γ 2 β1 1 − ρ 2
(25) γssc = μγssc (1). (28)
6 EURASIP Journal on Wireless Communications and Networking

1.3
0.7

1.25
0.6
Normalized average output SNR

Amount of fading (AoF)

1.2
0.5

1.15
0.4

1.1
0.3

1.05
0.2

1
1 2 3 4 5 6 7 8 9 1 2 3 4 5 6 7 8
Ricean K-Factor Ricean K-Factor
ρ = 0.1 ρ = 0.7 ρ = 0.1 ρ = 0.7
ρ = 0.3 ρ = 0.9 ρ = 0.3 ρ = 0.9
ρ = 0.5 ρ = 0.5

Figure 1: Normalized average output SNR (ASNR) versus the Figure 2: Amount of fading (AoF) versus the Ricean K-factor for
Ricean K-factor for several values of the correlation coefficient ρ. several values of the correlation coefficient ρ.

Δ
The AoF, defined as AoF = var(γssc )/γ2ssc , is a unified mea-
sure of the severity of the fading channel [2] and gives an
insight to the performance of the entire system. It can be ex-
pressed in terms of first- and second-order moments of γssc
as
μγssc (2)
AoF = − 1.
μγssc (1)2
6. PERFORMANCE EVALUATION RESULTS
Using the previous mathematical analysis, various perfor-
mance evaluation results have been obtained by means of
numerical techniques and will be presented in this section.
Such results include performances for the ASNR, AoF, Pout ,
ABEP1 , and C and will be presented for different Ricean
channel conditions, that is, different values for K and ρ, as
well as for various modulation schemes.
In Figures 1 and 2 the normalized ASNR (γssc /γ) and AoF
are plotted as functions of the Ricean K-factor for several val-
ues of the correlation coefficient ρ. These performance eval-
uation results have been obtained by numerically evaluating
(15)–(17), (28), and (29). The results presented in Figure 1
1 For the consistency of the presentation from now on instead of the ASEP
the ABEP performance will be used. As it is well known [2] for M-ary
(M > 2) modulation schemes, assuming Gray encoding, the ABEP can
be simply obtained from the ASEP as P be ∼ = P se / log2 M, since Es =
Eb log2 M, where Eb denotes the transmitted average bit energy.
show that as K increases, that is, the severity of the fading de-
creases, and/or ρ increases, the normalized ASNR decreases,
resulting in a reduced diversity gain. We note that similar ob-
servations have been made in [3, 30]. Furthermore, the re-
sults presented in Figure 2 indicate that as K increases and/or
ρ decreases, AoF is degraded.
Next the ABEP has been obtained using (23)–(27). In
(29) Figures 3 and 4 the ABEP is plotted as a function of the av-
erage input SNR per bit, that is, γb = γ/ log2 M, for several
values of K. Figure 3 considers the performance of DBPSK,
BPSK, and M-ary PSK signaling formats and ρ = 0.5. As
expected, when K increases, the ABEP improves and BPSK
exhibits the best performance. Figure 4 presents the ABEP
of 16-QAM for different values of ρ and K. For comparison
purposes, the performance of an equivalent single receiver,
that is, without diversity, is also included. Similar to the pre-
vious cases, it is observed that the ABEP improves as K in-
creases and/or ρ decreases, while significant overall perfor-
mance improvement, as compared to the no-diversity case,
is also noted.
In Figure 5, Pout is plotted as a function of the normalized
outage threshold per bit, γth /γb , for several values of K and
ρ. These performance results have been obtained by numer-
ically evaluating (10), (11), and (21) and for ρ = 0 they are
identical to the ones obtained by using [2, Equation 9.273].
It is observed that Pout decreases, that is, the outage perfor-
mance improves, as K increases and/or ρ decreases.
Finally, the normalized average CC can be obtained as
C! = C/BW (in b/s/Hz) by employing (19) and (20). In
Figure 6, C! is plotted as a function of γb for several values
P. S. Bithas and P. T. Mathiopoulos 7

10−1
K =1
Average bit error probability (ABEP)

10−1 K =1

Outage probability (Pout )

10−2
K =8
10−2

K =4

K =8
10−3
10−3

10−4 10−4
−5 0 5 10 15 20 −10 −7.5 −5 −2.5 0 2.5 5
Average input SNR per bit (dB) γth /γb
DBPSK 8-PSK ρ=0
BPSK 16-PSK ρ = 0.4
ρ = 0.8

Figure 3: Average bit error probability (ABEP) versus average in- Figure 5: Outage probability (Pout ) versus the normalized average
put SNR per bit for DBPSK, BPSK, and M-PSK (M = 8 and 16) input SNR per bit for several values of the Ricean K-factor and the
signaling formats, for different values of the Ricean K-factor. correlation coefficient ρ.

10−1 3
No diversity
Normalized average channel capacity (b/s/Hz)
Average bit error probability (ABEP)

2.5
10−2

10−3
ρ = 0.6 1.5

10−4 1
ρ = 0.2

0.5
10−5
−5 0 5 10 15 20 −4 −2 0 2 4 6 8 10
Average input SNR per bit (dB) Average input SNR per bit (dB)

K =1 ρ = 0.1 ρ = 0.9
K =4 ρ = 0.4 No diversity
K =8 ρ = 0.7

Figure 4: Average bit error probability (ABEP) versus average input Figure 6: Normalized average channel capacity (C/BW) versus the
SNR per bit for 16-QAM signaling format for different values of the average input SNR per bit for several values of the correlation coef-
Ricean K-factor and the correlation coefficient ρ. ficient ρ.
8 EURASIP Journal on Wireless Communications and Networking
of ρ and for K = 1. These results illustrate that as ρ increases,
C! decreases, as expected [12], and the receiver without diver-
sity has always the worst performance.
7. CONCLUSIONS
In this paper, the performance of dual branch SSC diversity
receivers operating over correlated Ricean fading channels
has been studied. By deriving a convenient expression for
the bivariate Ricean PDF, analytical formulae for the most
important statistical metrics of the received signals and the
capacity of such receivers have been obtained. Capitalizing
on these formulas, useful expressions for a number of per-
formance criteria have been obtained, such as ABEP, Pout ,
ASNR, AoF, and average CC. Various performance evalua-
tion results for different fading channel conditions have been
also presented and compared with equivalent performance
results of receivers without diversity.
ACKNOWLEDGMENTS
This work has been performed within the framework of
the Satellite Network of Excellence (SatNEx-II) project (IST-
027393), a Network of Excellence (NoE) funded by European
Commission (EC) under the FP6 program.
REFERENCES
[1] T. S. Rappaport, Wireless Communications, Prentice-Hall PTR,
Upper Saddle River, NJ, USA, 2002.
[2] M. K. Simon and M.-S. Alouini, Digital Communication over
Fading Channels, John Wiley & Sons, New York, NY, USA,
2nd edition, 2005.
[3] N. C. Sagias and G. K. Karagiannidis, “Gaussian class mul-
tivariate Weibull distributions: theory and applications in
fading channels,” IEEE Transactions on Information Theory,
vol. 51, no. 10, pp. 3608–3619, 2005.
[4] G. E. Corazza and F. Vatalaro, “A statistical model for land
mobile satellite channels and its application to nongeostation-
ary orbit systems,” IEEE Transactions on Vehicular Technology,
vol. 43, no. 3, part 2, pp. 738–742, 1994.
[5] H. Wakana, “A propagation model for land mobile satellite
communications,” in Proceedings of IEEE Antennas and Propa-
gation Society International Symposium, vol. 3, pp. 1526–1529,
London, Ont, Canada, June 1991.
[6] E. Lutz, D. Cygan, M. Dippold, F. Dolainsky, and W. Papke,
“The land mobile satellite communication channel-recording,
statistics, and channel model,” IEEE Transactions on Vehicular
Technology, vol. 40, no. 2, pp. 375–386, 1991.
[7] M. K. Simon and M.-S. Alouini, “A unified performance anal-
ysis of digital communication with dual selective combin-
ing diversity over correlated Rayleigh and Nakagami-m fad-
ing channels,” IEEE Transactions on Communications, vol. 47,
no. 1, pp. 33–43, 1999.
[8] Y. Chen and C. Tellambura, “Distribution functions of selec-
tion combiner output in equally correlated Rayleigh, Rician,
and Nakagami-m fading channels,” IEEE Transactions on Com-
munications, vol. 52, no. 11, pp. 1948–1956, 2004.
[9] G. K. Karagiannidis, D. A. Zogas, N. C. Sagias, S. A. Kot-
sopoulos, and G. S. Tombras, “Equal-gain and maximal-ratio
combining over nonidentical Weibull fading channels,” IEEE
Transactions on Wireless Communications, vol. 4, no. 3, pp.
841–846, 2005.
[10] M. H. Ismail and M. M. Matalgah, “Performance of dual
maximal ratio combining diversity in nonidentical correlated
Weibull fading channels using Padé approximation,” IEEE
Transactions on Communications, vol. 54, no. 3, pp. 397–402,
2006.
[11] N. C. Sagias, “Capacity of dual-branch selection diversity re-
ceivers in correlative Weibull fading,” European Transactions
on Telecommunications, vol. 17, no. 1, pp. 37–43, 2006.
[12] S. Khatalin and J. P. Fonseka, “Capacity of correlated
Nakagami-m fading channels with diversity combining tech-
niques,” IEEE Transactions on Vehicular Technology, vol. 55,
no. 1, pp. 142–150, 2006.
[13] C.-D. Iskander and P. T. Mathiopoulos, “Performance of
dual-branch coherent equal-gain combining in correlated
Nakagami-m fading,” Electronics Letters, vol. 39, no. 15, pp.
1152–1154, 2003.
[14] A. A. Abu-Dayya and N. C. Beaulieu, “Switched diversity on
microcellular Ricean channels,” IEEE Transactions on Vehicular
Technology, vol. 43, no. 4, pp. 970–976, 1994.
[15] Y.-C. Ko, M.-S. Alouini, and M. K. Simon, “Analysis and opti-
mization of switched diversity systems,” IEEE Transactions on
Vehicular Technology, vol. 49, no. 5, pp. 1813–1831, 2000.
[16] C. Tellambura, A. Annamalai, and V. K. Bhargava, “Unified
analysis of switched diversity systems in independent and cor-
related fading channels,” IEEE Transactions on Communica-
tions, vol. 49, no. 11, pp. 1955–1965, 2001.
[17] M.-S. Alouini and M. K. Simon, “Dual diversity over corre-
lated log-normal fading channels,” IEEE Transactions on Com-
munications, vol. 50, no. 12, pp. 1946–1959, 2002.
[18] D. A. Zogas and G. K. Karagiannidis, “Infinite-series repre-
sentations associated with the bivariate Rician distribution
and their applications,” IEEE Transactions on Communications,
vol. 53, no. 11, pp. 1790–1794, 2005.
[19] M. K. Simon, Probability Distributions Involving Gaussian
Random Variables: A Handbook for Engineers and Scientists,
Kluwer Academic Publishers, Norwell, Mass, USA, 2002.
[20] P. S. Bithas, N. C. Sagias, and P. T. Mathiopoulos, “Dual diver-
sity over correlated Ricean fading channels,” Journal of Com-
munications and Networks, vol. 9, no. 1, pp. 67–74, 2007.
[21] I. S. Gradshteyn and I. M. Ryzhik, Table of Integrals, Series,
and Products, Academic Press, New York, NY, USA, 6th edi-
tion, 2000.
[22] M. Abramowitz and I. A. Stegun, Handbook of Mathemati-
cal Functions with Formulas, Graphs and Mathematical Tables,
Dover, New York, NY, USA, 9th edition, 1972.
[23] A. A. Abu-Dayya and N. C. Beaulieu, “Analysis of switched di-
versity systems on generalized-fading channels,” IEEE Transac-
tions on Communications, vol. 42, no. 11, pp. 2959–2966, 1994.
[24] A. Papoulis, Probability, Random Variables, and Stochastic Pro-
cesses, McGraw-Hill, New York, NY, USA, 2nd edition, 1984.
[25] A. Nuttal, “Some integrals involving the Q-function,” Tech.
Rep. 4297, Naval Underwater Systems Center, New London,
Conn, USA, April 1972.
[26] W. C. Y. Lee, “Estimate of channel capacity in Rayleigh fad-
ing environment,” IEEE Transactions on Vehicular Technology,
vol. 39, no. 3, pp. 187–189, 1990.
[27] “The Wolfram functions site,” http://functions.wolfram.com/.
[28] V. S. Adamchik and O. I. Marichev, “The algorithm for cal-
culating integrals of hypergeometric type functions and its
realization in REDUCE system,” in Proceedings of Interna-
tional Symposium on Symbolic and Algebraic Computation (IS-
SAC ’90), pp. 212–224, Tokyo, Japan, August 1990.
P. S. Bithas and P. T. Mathiopoulos 9

[29] N. C. Sagias, D. A. Zogas, and G. K. Karagiannidis, “Selection

diversity receivers over nonidentical Weibull fading channels,”
IEEE Transactions on Vehicular Technology, vol. 54, no. 6, pp.
2146–2151, 2005.
[30] P. S. Bithas, G. K. Karagiannidis, N. C. Sagias, P. T. Mathiopou-
los, S. A. Kotsopoulos, and G. E. Corazza, “Performance analy-
sis of a class of GSC receivers over nonidentical Weibull fading
channels,” IEEE Transactions on Vehicular Technology, vol. 54,
no. 6, pp. 1963–1970, 2005.
Hindawi Publishing Corporation
EURASIP Journal on Wireless Communications and Networking
Volume 2007, Article ID 49718, 17 pages
doi:10.1155/2007/49718

Research Article
Advanced Fade Countermeasures for DVB-S2 Systems in
Railway Scenarios

Stefano Cioni,1 Cristina Párraga Niebla,2 Gonzalo Seco Granados,3 Sandro Scalise,2
Alessandro Vanelli-Coralli,1 and Marı́a Angeles Vázquez Castro3
1 ARCES, University of Bologna, Via Toffano 2, 40125 Bologna, Italy
2 German Aerospace Center (DLR), Institute of Communications and Navigation, Postfach 1116, 82230 Wessling, Germany
3 Department of Telecommunications and Systems Engineering, Universitat Autònoma de Barcelona, Campus Universitari, s/n,

08193 Bellatera, Barcelona, Spain

Received 22 October 2006; Accepted 3 June 2007

Recommended by Ray E. Sheriff

This paper deals with the analysis of advanced fade countermeasures for supporting DVB-S2 reception by mobile terminals
mounted on high-speed trains. Recent market studies indicate this as a potential profitable market for satellite communications,
provided that integration with wireless terrestrial networks can be implemented to bridge the satellite connectivity inside railway
tunnels and large train stations. In turn, the satellite can typically offer the coverage of around 80% of the railway path with existing
space infrastructure. This piece of work, representing the first step of a wider study, is focusing on the modifications which may
be required in the DVB-S2 standard (to be employed in the forward link) in order to achieve reliable reception in a challenging
environment such as the railway one. Modifications have been devised trying to minimize the impact on the existing air interface,
standardized for fixed terminals.

Copyright © 2007 Stefano Cioni et al. This is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

1. INTRODUCTION Complementing to satellite multimedia to fixed termi-

Satellite communications developed to a tremendous global
success in the field of analog and then digital audio/TV
broadcasting by exploiting the inherent wide-area coverage
for the distribution of content. It appeared a “natural” con-
sequence to extend the satellite services for point-to-point
multimedia applications, by taking advantage of the ability of
satellite to efficiently distribute multimedia information over
very large geographical areas and of the existing/potential
large available bandwidth in the Ku/Ka band. Particularly in
Europe, due to the successful introduction of digital video
broadcasting via satellite (DVB-S) [1], a promising techni-
cal fundament has been laid for the development of satel-
lite communications into these new market opportunities
using the second generation of DVB-S [2], commonly re-
ferred to as DVB-S2, as well as return channel via satellite
(DVB-RCS) [3] standards. Thus, for satellite systems cur-
rently under development and being designed to support
mainly multimedia services, the application of the DVB-S2,
for the high-capacity gateway-to-user (forward) links and of
DVB-RCS for the user-to-gateway (return) links, is widely
accepted.
nals, people are getting more and more used to broadband
communications on the move. Mobile telephones subscrip-
tions have exceeded fixed line subscription in many coun-
tries. Higher data rates for mobile devices are provided
by new standards such as UMTS, high-speed packet access
(HSPA), prestandardized version of mobile WiMAX, and, in
case of broadcast applications, digital video broadcasting for
handhelds (DVB-H) [5].
At present, broadband access (e.g., to the Internet) and
dedicated point-to-point links (for professional services) are
primarily supplied by terrestrial networks. Broadband sat-
coms services are still a niche market, especially for mobile
users. In this context, many transport operators announce
the provision of TV services in ships, trains, buses, and air-
crafts. Furthermore, Internet access is offered to passengers.
With IP connectivity, also radio interfaces for GSM can be
implemented for such mobile platforms by using satellite
connectivity for backhauling.
Thus, DVB-S2/RCS appears an ideal candidate to be in-
vestigated for mobile usage, as it can ideally combine digital
TV broadcast reception in mobile environments (airTV, lux-
ury yachts, trains, etc.) and IP multimedia services.
2 EURASIP Journal on Wireless Communications and Networking
However, the aforementioned standards have not been
designed for mobile use. Collective terminals installed in a
mobile platform, such as train, ship, or aircraft, are exposed
to a challenging environment that will impact the system per-
formance considering the current standard in absence of any
specific provision.
Mobile terminals will have to cope in general with strin-
gent frequency regulations (especially in Ku band), Doppler
effects, frequent handovers, and impairments in the synchro-
nization acquisition and maintenance. Furthermore, the rail-
way scenario is affected by shadowing and fast fading due
to mobility, such as, for example, the deep and frequent
fades due to the presence of metallic obstacles along electri-
fied lines providing power to the locomotive1 [6] and long
blockages due to the presence of tunnels and large train sta-
tions. This suggests that hybrid networks, that is, interwork-
ing satellite and terrestrial components, are essential in order
to keep service availability.
In this context, this paper is focused on proposing and
evaluating fade countermeasures to compensate the impact
of fade sources in the railway scenario, that is, shadowing,
fast fading, and power arches, excluding tunnels which will
be address at a later stage. In particular, antenna diversity and
packet level forward error correction (FEC) are investigated.
The rest of the paper is organized as follows: Section 2
discusses the potential of opening the current DVB-S2/RCS
standards to provide mobile services efficiently. Section 3
presents the peculiarities of the trains’ scenario and discusses
the different aspects that can impact the system performance.
Section 4 describes the fade countermeasures proposed in
this paper. Section 5 introduces the simulation platforms in
which the proposed fade countermeasures are evaluated and
Section 6 presents and discusses the obtained results. Finally,
Section 7 draws the conclusions of this work.
2. THE VISION: A NEW DVB-S2/RCS STANDARD FOR
MOBILE COLLECTIVE TERMINALS
The large capacity of DVB-S2/RCS systems can efficiently ac-
commodate broadcast services (e.g., digital TV) and unicast
IP multimedia interactive services to fixed terminals. How-
ever, the increasing interest on broadband mobile services
suggests that the natural evolution of DVB-S2/RCS standard
to cover new market needs goes towards the support of mo-
bile terminals.
In particular, the required antenna performance in Ku
(10–12 GHz) and Ka (20–30 GHz) bands focuses the mar-
ket opportunities of DVB-S2/RCS onto mobile terminals in
collective transportation means. Actually, transport opera-
tors are starting to announce the provision of TV services in
ships, trains, buses, and aircrafts, and broadband IP connec-
tivity, for passengers. For the specific case of trains, broad-
band services can provided using satellite systems, cellular
connectivity or dedicated trackside installations.
1 Hereafter referred to as “power arches,” for the sake of brevity.
As summarized in Table 1, none of these alternatives
alone represents a satisfactory solution. As a matter of fact,
deployed or upcoming commercial services are based on
combinations of different access technologies. In this light,
a satellite access based on an open standard can have very
significant benefits in terms of interoperability (achieved for
DVB-S2/RCS through SatLabs Qualification Program) and
competition, thus benefiting from availability of fully com-
patible terminals from multiple vendors and reducing the
cost of terminals.
However, the aforementioned DVB standards have been
designed for fixed terminals. To cope with these new market
opportunities, DVB TM-RCS has investigated how the cur-
rent DVB-RCS standard could be applied to mobile applica-
tions. A white paper on the applicability of DVB-RCS to mo-
bile services was prepared and a technical annex was added
to the implementation guidelines document [4]. This annex
states the boundary conditions and limitations under which
the existing standard could be used in mobile environment,
considering the impact of mobility in terminal synchroniza-
tion and demodulator performance in forward and return
links. Furthermore, a survey on applicable regulations and a
brief analysis on DVB-RCS features that can be used for mo-
bility management are provided, the latter referring to inter-
beam handover only.
Thus, the DVB-RCS guideline cannot support the full
adaptability to mobile environments and hence the applica-
ble services and scenarios happen to be very limited. Fur-
thermore, additional issues related to mobility are not fully
solved, such as handling of nonline-of-sight (nLOS) channel
conditions, which will require the interworking with terres-
trial gap fillers in the railway scenario due to the presence of
tunnels. In addition, even if DVB-RCS features to be applied
for mobility management are analyzed, a determined mech-
anism or protocol should be specified in order to ensure in-
teroperability. Finally, the impact of control signals loss (due
to deep fades or handover) is not negligible. For instance, the
loss of terminal burst time plan (TBTP) tables damages the
operation of the resource management, essential in the re-
turn link for a coordinated access to the radio resources.
As a matter of fact, mobile services could be more effi-
ciently supported if the present standards could be improved
for mobile scenarios. The reopening of the standard2 would
allow for the specification of methods for improving the link
reliability in mobile environments (e.g., packet level FEC),
handover protocols, interfaces to terrestrial gap fillers (even
using terrestrial mobile technologies), improved mobility-
aware signalling and resource management, and so forth.
In this context, a number of R and D initiatives are on-
going with the aim at investigating enhancements of the
DVB-S2/RCS standards for the efficient support of mobil-
ity. Among those, the SatNEx network of excellence has set
up a dedicated working group investigating different aspects
related to mobility in DVB-S2/RCS. The first results of this
activity in the field of forward link reliability for the rail-
way scenario are presented in this paper. For the return link,
2 Envisaged at the time of writing.
Stefano Cioni et al. 3

Table 1: Pros and cons of different solutions for providing broadband services on trains.

Type of
Examples Pros Cons
technology
(i) No new trackside (i) Available tracking antennas and
infrastructure—quick to efficient satcom modems expensive
deploy, project costs may be
lower on long distance routes
DVB-S2/RCS
Satellite Proprietary systems, (ii) Dedicated bandwidth available (ii) High variable cost per MB
for example, ViaSat
(iii) Performance easy to predict (iii) Return bandwidth constrained
depending on satellite visibility by antenna size
(iv) Not affected by borders—good (iv) Satellite visibility seriously
for international trains restricted on some rail routes
(i) Equipment is small and cheap (i) Geographic coverage of UMTS
limited for years to come
(ii) Usage is cheap (50–75 C per month (ii) Coverage of railway lines often
flat rate) worse than roads
GPRS
EDGE (iii) Data rates improving year on year (iii) GPRS/EDGE not really fast enough
Cellular UMTS (iv) Competitive supply—3 or 4 network (iv) Inverse relationship between
HSUPA/HSDPA operators in most countries throughput and train speed
(EV-DO)
(v) No QoS guarantees—affected by
network congestion at peak times
(vi) Organized country by country—data
roaming charges are punitive
(i) High data rates possible (i) Existing standards not designed to
support fast-moving terminals
(ii) Can control bandwidth and QoS (ii) Proprietary equipment is more
expensive
Flash OFDM (iii) On-train equipment relatively (iii) No suitable public services yet in
Trackside IEEE 802.11 inexpensive licensed bands—will licence-holders be
IEEE 802.16 (WiMAX) allowed to provide mobile services?

(iv) No volume-related usage costs (iv) Licence-exempt bands are low power,
thus limited range

(v) Infrastructure deployment (especially

trackside) is expensive and time consuming

analogue solutions have to be devised, which are however not
in the scope of the present work.
3. THE RAILWAY SCENARIO, A CHALLENGING
ENVIRONMENT
3.1. Overview
The land mobile satellite channel (LMSC) has been widely
studied in the literature [7]. Several measurement campaigns
have been carried out and several narrow and wideband
models have been proposed for a wide range of frequencies,
including Ku [8] and Ka [9] bands. Nevertheless, for the spe-
cific case of the railway environment, only few results are
presented in [10] as a consequence of a limited trial cam-
paign using a narrowband test signal at 1.5 GHz, performed
more than 10 years ago in the north of Spain. These results
represent a very interesting reference, although no specific
channel model has been extracted from the collected data.
After an initial qualitative analysis, the railway environment
appears to differ substantially with respect to the scenarios
normally considered when modelling the LMSC. Excluding
railway tunnels and areas in the proximity of large railway
stations, one has to consider the presence of several metallic
obstacles like power arches (Figure 1, left uppermost), posts
with horizontal brackets (Figure 1, left lowermost), which
may be often grouped together (Figure 1, rightmost), and
catenaries, that is, electrical cables, visible in all the afore-
mentioned figures.
The results of direct measurements performed along the
Italian railway and aiming to characterize these peculiar ob-
stacles are reported in [6] and references herein. In summary,
4 EURASIP Journal on Wireless Communications and Networking
Figure 1: Nomenclature of railway specific obstacles.
the attenuation introduced by the catenaries (less than 2 dB)
and by posts with brackets (2-3 dB) is relatively low and can
be easily compensated by an adequate link margin. On the
other hand, the attenuation introduced by the power arches
goes, depending on the geometry, the radiation pattern of the
RX antenna, and the carrier frequency, down to values much
greater than 10 dB.
3.2. Modelling
Even if the layout and exact geometry of such obstacles can
significantly change depending on the considered railway
path, it turned out from previous works that the attenuation
introduced by these kind of obstacles can be accurately mod-
elled using knife-edge diffraction theory [11]: in presence of
an obstacle having one infinite dimension (e.g., mountains
or high buildings), the knife-edge attenuation can be com-
puted as the ratio between the received field in presence of
the obstacle and the received field in free space conditions. In
the case addressed here, as shown in Figure 2 (left), the obsta-
cle has two finite dimensions, and the received field is hence
the sum of the contributions coming from both sides of the
obstacle. Therefore, the resulting attenuation can be written
as follows:
As (h)
  ∞ 
1   
= √ G α1 (h) 
 e − j(π/2)v 2
dv
 System designers can resort to different approaches to coun-
2Gmax Kh
 
  K(h−d) − j(π/2)v2 
+ G α2 (h) 
 e dv  ,
 QoS level. A possible classification of fade countermeasure is
−∞

2a+b
K=
λ a·b
where λ is the wavelength, a is the distance between the re-
ceiving antenna and the obstacle, b is the distance between
the obstacle and the satellite, h is the height of the obstacle
above the line-of-sight (LOS), and d is the width of the ob-
stacle. Finally, the usage of a directive antenna with radiation
pattern G(α) has to be considered. This implies an additional
attenuation due to the fact that whenever the two diffracted
rays reach the receiving antenna with angles α1 and α2 as
shown in Figure 2(left), the antenna shows a gain less than
the maximum achievable (Gmax ) and depending on the vari-
able h, which is directly related to the space covered by the
train.
In absence of a channel model directly extracted from
measurements in the railway environment, it is a common
practice to model the so-called “railroad satellite channel”
by superimposing (i.e., multiplying) the statistical fades re-
produced by a Markov model (see [8, 9]) with the space-
periodic fades introduced by the electrical trellises obtainable
by means of the above equation. Values of the parameters
in Figure 2, as well as the space separation between subse-
quent electrical trellises, depend on the considered railway.
Finally, the considered receiving antennas are modelled with
high directivity in order to achieve large gain and at the same
time to reduce the received multipath components with large
angular spread. Hence, as reported in [12], the key parame-
ter becomes the antenna beamwidth which describes in the
frequency domain the Doppler power spectrum density of
the satellite fading channel. In this paper, the highly direc-
tive antennas are modelled with the reasonable value of the
beamwidth in the order of 5 degrees.
3.3. Need for fade countermeasures and gap fillers
The periodical fading events induced by power arches (PA)
result in a physical error floor that limits the performance of
the DVB-S2 system to unacceptable quality of service (QoS)
levels. In Figure 3, the baseband frame (BBFRAME) error
rate is reported in LOS conditions, for train speed equal to
300 km/h, and in the presence of power arches, when the re-
ceiver has only one receiving antenna and does not adopt
any packet level FEC technique. The error floor value is
about 0.0117, corresponding to the ratio between the du-
ration of PA induced fading events, that is, 6 msilliseconds
at 300 km/h, and the time between two fading events, that
is, 600 msilliseconds at 300 km/h. Considering the case of
27.5 Mbaud, the DVB-S2 BBFRAME duration is less than
1 msillisecond, therefore when the receiving antenna is ob-
scured by a power arch, transmitted packets are completely
lost unless fade countermeasures are adopted.
4. ADVANCED FADE COUNTERMEASURES
teract deep fading conditions and to guarantee an acceptable
(1)
between those techniques that need a return channel (from
the user to the network) to require a change in the transmis-
, sion mode or a retransmission of the lost information, and
those that do not rely on a return channel and are therefore
more suitable for unidirectional delivery, such as multicast
or broadcast applications. The latter class of techniques is of
great interest for the collective railway application considered
in this work, for which return channel-based approaches,
such as automatic repeat request (ARQ) or adaptive coding
and modulation (ACM) techniques, are not doable. In par-
ticular, antenna diversity and packet level FEC techniques are
considered in the following.
Stefano Cioni et al. 5

1E + 0

1E − 01

BBFRAME error rate

b Power arches floor
1E − 02

h-d h
1E − 03
E2 /E0 a E1 /E0 v

α1
α2 1E − 04
1 3 5 7 9 11 13 15 17 19 21
Eb /N0 (dB)
1/2 - QPSK (LOS, FAST, noPA)
2/3 - 8PSK (LOS, FAST, noPA)
(a)
3/4 - 16APSK (LOS, FAST, noPA)
5/6 - 16APSK (LOS, FAST, noPA)
d = 0.4 m, a = 2.5 m 1/2 - QPSK (LOS, FAST, PA)
5
2/3 - 8PSK (LOS, FAST, PA)
0 3/4 - 16APSK (LOS, FAST, PA)
5/6 - 16APSK (LOS, FAST, PA)
−5

−10
Figure 3: BBFRAME error rate for DVB-S2 in the presence of
power arch blockage events. LOS propagation conditions and train
Attenuation (dB)

−15 speed set to 300 km/h.

−20

−25
antennas, and assuming perfect compensation of time delays
−30 of the two replicas, the combined signal can be written as
−35
rc (t) = w1 r1 (t) + w2 r2 (t), (2)
−40

−45
−2.5 −2 −1.5 −1 −0.5 0 0.5
h (m)
0.6 m
0.4 m
0.2 m
(b)
Figure 2: Knife-edge diffraction model applied to the railway sce-
nario and possible attenuation caused by power arches at Ku band
for different antenna diameters.
4.1. Antenna diversity
The adoption of multiple receiving antennas to counteract
power arch obstructions in railway environment has been re-
cently proposed and investigated in [13, 14]. Antenna diver-
sity is used to provide different replica of the received signal
to the detector for combination or selection. If the receiving
antennas are sufficiently spaced, the received signals fade in-
dependently on each antenna thus providing multiple diver-
sity branches that can be linearly or nonlinearly combined to
improve detection reliability. There are mainly three types of
linear diversity combining approaches: selection, maximal-
ratio, and equal-gain combining. Considering two receiving
where wi and ri (t), i = 1, 2, are the combing weights and
1 1.5 2 2.5 the received signals, respectively. The received signals at each
antenna is
ri (t) = αi s0 (t) + ni (t), (3)
where s0 (t) is the transmitted signal, αi is the time variant
fading envelope over the ith antenna, and ni (t) is the thermal
noise.
The simplest combining scheme is the signal selection
Combining (SC), in which the branch-signal with the largest
amplitude or signal-to-noise ratio (SNR) is the one selected
for demodulation. In this case, wi will be 1 or 0 if the
ith power branch is the largest or the smallest, respectively.
Clearly, SC is bounded by the performance of the single re-
ceiving antenna in absence of fading, that is, there is no di-
versity gain when the two antennas experience good chan-
nel conditions at the same time. Maximum-ratio combin-
ing (MRC), although requiring a larger complexity at the
receiver, allows for the exploitation of the diversity gain. In
fact, MRC scheme provides for the maximum output SNR.
According to the optimum combination criterion, the signal
weights are directly proportional to the fading amplitude and
inversely proportional to the noise power, Ni , as follows:
αi
wi = . (4)
Ni
Another technique, often used because it does not require
channel fading strength estimation, is equal gain combining
6 EURASIP Journal on Wireless Communications and Networking
(EGC) in which the combination weights are all set to one,
thus leading to a simpler but suboptimal approach. Clearly,
SC and MRC (or EGC) represent the two extremes in diver-
sity combining strategy with respect to the complexity point
of view and the number of signals used for demodulation
process. Furthermore, the classical combining formula can
be generalized for nonconstant envelope modulations such
as 16-APSK or 32-APSK (amplitude and phase shift keying)
and integrated with the soft demodulator that computes the
channel a posteriori information to feed the low density par-
ity check (LDPC) FEC decoder. The maximum likelihood a
priori information for a single receiver antenna given by
4.2. Packet level FEC
4.2.1. The concept of packet level FEC
Reliable transmission occurs when all recipients correctly re-
ceive the transmitted data. This target can be achieved by op-
erating at different layers of the protocol stack and in dif-
ferent ways. Retransmission techniques allow that lost pack-
ets are retransmitted to the receivers, while packet level FEC
schemes create redundant packets that permit to reconstruct
the lost ones at the receiver side, with a very beneficial in-
put on the final end-to-end delay. In fact, as detailed in [15],
the additional delay introduced by packet level encoding and
 decoding is always lower than the delay deriving from any
Pr bi = 0 | rk retransmission scheme.
log

Pr bi = 1 | rk Regarding the retransmission schemes, efficient proto-

  2  (5) cols should limit the use of acknowledgement- (ACK-) based
si ∈S0 exp −rk − αk si  /N0 mechanisms because they introduce heavy feedback traffic
= log
  2 
si ∈S1 exp −rk − αk si  /N0 towards the sender, thus increasing the congestion of reverse
link that, typically, has a reduced capacity with respect to
forward link. Negative acknowledgement- (NACK-) based
can be extended for L receiving antennas, according to the approaches are hence particularly interesting. In combina-
MRC principle, as follows: tion with (or in alternative to) the traditional retransmission
schemes, packet level FEC can be added on top of physical

 layer FEC, in order to achieve the same level of reliability with
Pr bi = 0 | rk a reduced number of retransmissions. This might be partic-
log

Pr bi = 1 | rk ularly useful if resources on the return link need to be saved


L
 p
r p 2
si ∈S0 exp −
p=0 k k si  /N0
−α
= log

L  p
r p 2
si ∈S1 exp − p=0 k k si  /N0
−α
where rk is the received sample at time k, αk is the true or
the estimated channel coefficient, and S0 and S1 are the sets
of symbols which have “0” or “1” in the ith position, respec-
tively.
In the configuration proposed in this work, we adopt
MRC combining with two antennas. The antennas are placed
on the same coach so as to reduce the costs of installa-
tion and the connection length. The antenna spacing is cho-
sen as a function of the distance between two consecutive
power arches so as to guarantee that only one antenna at
a time can be obscured. Accordingly, the distance between
the two antennas is about 15 m. Considering the maximum
train speed (about 300 km/h), this translates into the fact
that power-arch blockage on a single antenna lasts for about
7 msilliseconds, and it hits the second antenna after about
180 msilliseconds. Therefore, it is reasonable to assume that
there is enough time for the combining circuit to react and
maintain constant signal connection. A drawback of this ap-
proach is that the receiving chain will be duplicated in or-
der to maintain connection and avoid frequent reacquisitions
process with the consequent loss of packet. As proposed in
[14], the solution which considers the presence of a second
receiving antenna is depicted in Figure 4. The gray blocks
represent the subsystems that need to be duplicated in the
two antenna case. Further details on the digital receiver are
described in Section 5.1.
p   (smaller number of NACKs or no NACKs are needed at all),
or when multiple lost packets are recovered with the retrans-
p  ,
mission of a lower number of redundant packets. Basically,
(6) h redundancy packets are added to each group of k informa-
tion packets, thus resulting in the transmission of n = k + h
packets. These packets are finally transferred to the physi-
cal layer, which adds independent channel coding to each of
them. This principle is described in Figure 5.
At the physical layer, the bits affected by low noise lev-
els can be corrected by the physical layer FEC, so that the
related packets are passed to the higher layer as “correct.” If
the noise level exceeds the correcting capability of the phys-
ical layer, the received bit cannot be properly decoded, but
the failure to decode can be usually detected with a very high
reliability. Since erroneous packets are not propagated to the
higher layers, we have an erasure channel. The system can use
the redundancy packets to recover these erasures. By using
maximum distance separable (MDS) codes, like the Reed-
Solomon, it is possible to reconstruct the original informa-
tion if at least k out of n packets are correctly received. There-
fore, the receiver can cope with erasures, as long as they result
in a total loss not exceeding h packets, independently from
where the erasures occurred. LDPC codes and their deriva-
tions might be also used because of their low complexity and
greater flexibility, thus permitting to encode larger files, al-
though a small inefficiency, depending on the code design
and typically around 5%–10%, will be taken into account.
If packet level FEC is implemented at IP or data link layer,
very near to the physical channel, no change in the trans-
port and network layers protocols and in the physical layer
are necessary. This solution presents the additional advantage
that it can be adapted to the propagation channel conditions
Stefano Cioni et al. 7

Frame
synch

Received
signal Matched Symbol Data
from sampling DeMUX Buffer
filter
antenna no. 1
Preamble / Noise level N01
Frequency From second
Timing pilots estimation antenna
acquisition recovery θ01 Signal
Digital α1k combiner
AGAC
θ01 Lock
Buffer detector Hard/soft
demodulator

θk1
Freq./phase
tracking
De-
interleaver

LDPC/BCH
decoder

Figure 4: Receiver block diagram with antenna diversity.

n packets

k data packets (group) h redundancy packets

1 2 ··· k k+1 ··· k+h Data link/IP layer

Channel coding

1 2 ··· k k+1 ··· k+h Physical layer

Transmission

Figure 5: Packet level FEC principle.

by choosing n, so that the interleaver size is long enough to
compensate the channel outages. However, different protec-
tion for individual transfers (e.g., specific files) is not possi-
ble (although different QoS classes may be supported), extra
memory is required, and additional delays must be properly
handled.
For the forward link, the usage of packet level FEC is
especially powerful in allowing online variable coding ap-
proaches, which can be fine tuned in a closed-loop approach.
Based upon the “history” of the link, appropriate redun-
dancy can be easily added. Packet level FEC has then impact
on different layers.
(i) The requirements on control loops can be lessened, for
example, power control and or adaptive coding and
modulation control, if a loss of up to h packets can tol-
erated.
(ii) The typical fade structure of a link can be measured
and accordingly coding with the correct profile added.
(iii) Different QoS classes with different redundancy pro-
files can be supported. Furthermore, redundancy
packets for low-priority traffic can be put in a special
queue, which is served only if free capacity is available
and, in turn, increased redundancy can be sent during
handovers, minimizing the overall probability of lost
packets.
(iv) Different IP-based access methods can be used in par-
allel, improving the link reliability if different redun-
dancy is sent via different access methods.
4.2.2. The GSE-FEC method
When moving to the concrete applicability of this scheme to
the scenario under consideration, even though the fact that
IP packets have three sizes that are the most common ones,
the fact that IP packet size can actually take any value up
to a maximum value (typically 64 Kbytes) represents a clear
8 EURASIP Journal on Wireless Communications and Networking
IP packets
GSE
FEC matrix
encapsulation
Figure 6: Steps involved in GSE-FEC.
difficulty in applying packet level FEC (PL-FEC). The funda-
mental difficulty comes from the fact that most codes take as
input a fixed amount of data, from which they compute the
redundancy bytes. As a given number of IP packets corre-
spond to a variable amount of data depending of their sizes,
codes needing a fixed amount of data cannot be directly ap-
plied. One possible solution is to use codes that can be eas-
ily adapted to different input sizes; however, this comes at
the price of a much more complex encoding and decoding
process. Another solution has been proposed in the DVB-H
standard [16]. In this case, units of constant length are built
by interleaving IP packets and, therefore, codes with fixed in-
put size can be easily applied. It is worth noting that those
units are not built by concatenating IP packets but by inter-
leaving them. However, interleaving is this case must not be
understood as it is typical in physical layer coding, where it
means that data is written in one direction in a matrix and
it is read in the orthogonal direction for transmitting. In PL-
FEC, we understand interleaving as computing the redun-
dancy in an orthogonal direction to the writing direction of
the data; however, in this case the writing and reading direc-
tions coincide. This kind of interleaving is advantageous be-
cause the redundancy is computed across a large number of
packets. Thus, a fade event may destroy one or several pack-
ets but not the majority of them, assuming that the system
is well dimensioned, so the added redundancy can effectively
help in recovering the destroyed packets.
DVB-H also provides a solution for encapsulating the
coded IP packets for transmission over DVB-T. The solution
is based on the use of multiprotocol encapsulation (MPE)
combined with MPEG. Although it would be possible to
adapt the same approach for DVB-S2, it presents a number
of drawbacks, such as lack of flexibility, low encapsulation
efficiency, delay constraints. A new encapsulation protocol
call generic stream encapsulation (GSE) has been recently de-
fined [17]. It is a very flexible protocol applicable to several
physical layer standards. It overcomes most of the limitations
of MPE-MPEG. GSE is especially suitable for transmitting IP
packets through the generic stream interface mode of DVB-
S2, and it has been proposed for the second generation of
Terrestrial digital video broadcasting (DVB-T2) as well. GSE
also efficiently supports the ACM functionalities of DVB-S2
and facilitates the provision of QoS guarantees because it re-
duces the constraints on the scheduling operation.
It can be deducted from the previous discussion that the
implementation of PL-FEC consists of two main processes:
the encoding the IP packets and, second, the encapsulation
of the result of the encoding process in order to adapt it to
the underlying transmission system. In DVB-H, the first pro-
cess consists in arranging the IP packets in a matrix (here-
after called FEC matrix) and applying a Reed-Solomon code,
BBFRAME assembly BBFRAMEs
BBFRAME
using one or several
padding
GSE units
while the second process employs MPE-MPEG. The whole
implementation is called MPE-FEC in DVB-H. Our proposal
for DVB-S2 is based on keeping the same first process as in
DVB-H, whereas it employs GSE in the second process. This
proposal for applying PL-FEC in DVB-S2 is named GSE-
FEC.
A block diagram of GSE-FEC is depicted in Figure 6. The
incoming IP packets are arranged in the so-called FEC ma-
trix, where also the packet-level redundancy is added. The
filling of the FEC matrix and the encoding are done in the
same way as in DVB-H. For the sake of completeness, this
will be briefly described below. Next, each IP packet is en-
capsulated using GSE, and this represents one of the novel
aspects of our proposal. Each IP packet may be fragmented
into several GSE units or it may also be sent unfragmented.
Subsequently, the maximum number of GSE units that can
be fitted inside a BBFRAME is concatenated and introduced
in the BBFRAME. The size of the BBFRAME depends on the
combination of coding rate and modulation scheme (MOD-
COD) adopted by the DVB-S2 modem, so the number of
GSE units that can be concatenated also depends on the
MODCOD. By making the GSE units small enough to have
the required flexibility, but large enough in order not to pe-
nalize encapsulation efficiency, this method provides an easy
mechanism to adapt the output of the packet-level FEC to the
variations of the physical layer. Moreover, note that padding
is not applied inside the GSE unit but only at BBFRAME level
if the size of the BBFRAME does not coincide with that of the
concatenation of the GSE units.
The IP packets are placed one after another along the
columns of the FEC matrix, see Figure 7. Each IP packet may
be split among two or more columns. Only the first block of
the matrix, from column 1 to 191, can be filled in with IP
packets. The second block of the matrix, from column 192 to
255, carries the redundancy information, which is computed
by a Reed-Solomon (255,191) code applied to the first block
on a row basis. Each column in the second block is encap-
sulated individually using GSE, whereas in the first block the
GSE encapsulation is performed on an IP packet basis. In the
baseline operation, padding is only applied in the first block
to account for the fact that an additional IP packet may not
be fitted without overrunning the 191 columns and all 64 re-
dundancy columns are transmitted. The code can be made
weaker (i.e., with higher rate) by puncturing some of the re-
dundancy columns, which are then not transmitted and are
considered as unreliable bytes in the decoding process. The
code can also be made more robust (i.e., with lower rate)
by padding with zeros columns in the first block and, hence,
leaving less space for IP packets. The padded columns are not
transmitted but they are used in the encoding process. In the
decoding process, they are considered as reliable.
Stefano Cioni et al. 9

Coding direction
FEC matrix

Writing direction
1 2 3 188 189 190 191 192 193 254 255

IP packet 1 (cont.)

Last IP packet (cont.)

IP packet 2 (cont.)

2nd redundancy column

1st redundancy column

Punctured column

Punctured column
Column size

IP packet 1

Padding

Padding

Padding
··· ···

IP packet 2

Padding
IP packet 3

Data submatrix Redundancy submatrix

IP packet encapsulation with GSE Percolumn GSE encapsulation

Figure 7: Arrangement of IP packets for FEC encoding.

After GSE encapsulation, the GSE packets are introduced merger/slicer that, depending on the applications, allocates
in BBFRAMEs and transmitted. On the receive side, erro- a number of input bits equal to the maximum data field ca-
neous BBFRAMEs are detected by checking the CRC. The pacity. In this way, user packets are broken in subsequent
receiver reconstructs the FEC matrix and marks any column data fields, or an integer number of packets are allocated in
that is totally or partially received by means on an erroneous it. Then, a fixed length base-band header (BBHEADER) of
BBFRAME as unreliable. Finally, if the reconstructed FEC 80 bits is inserted in front of the data field, describing its for-
matrix has no more than 64 unreliable columns, the code mat. For example, it reports to the decoder the input streams
can correctly compute all bytes in the matrix. If there are format, the mode adaptation type and the roll-off factor.
more than 64 unreliable columns, the code cannot correct The efficiency loss introduced by this header varies from
anything, and only those columns received by means of cor- 0.25% to 1% for long and short codeword lengths, respec-
rect BBFRAMEs will be correct. tively. The role of stream adaptation is to provide padding
when needed, in order to complete a constant length frame,
5. SIMULATION SCENARIOS and scrambling. Padding is applied when the user data avail-
able for transmission are not sufficient to completely fill a
In the following, the simulation platforms used to evaluate BBFRAME, or when more than one packet have to be allo-
the performance of DVB-S2 with advanced fade countermea- cated in a BBFRAME. The built frame is randomized using a
sures in the railway environment as described in Section 3 are scrambling sequence generated by the pseudorandom binary
duly detailed. sequence described by the polynomial (1 + X 14 + X 15 ). After
this scrambling, each BBFRAME is processed by the forward
5.1. Advanced physical layer simulation platform error correction (FEC) encoder which is carried out by the
concatenation of a Bose-Chaudhuri-Hocquenghem (BCH)
To cover a rather large set of spectral efficiency, four MOD- outer code and an LDPC inner code. Available code-rates
CODs have been considered: 1/2-QPSK, 2/3-8PSK, 3/4- for the inner code are 1/4, 1/3, 2/5, 1/2, 3/5, 2/3, 3/4, 4/5,
16APSK, and 5/6-16APSK. The LOS channel condition 5/6, 8/9, and 9/10. Depending on the application area, code-
(Rice factor equal to 17.4 dB) and the train speed equal to words can have length NLDPC = 64800 bits or 16200 bits. In
300 km/h have been simulated. Equally spaced power arches the following, the case of 64800 bits is considered. Regard-
with a separation of 50 m have been included in some sce- ing the modulation format, each coded BBFRAME can be
narios, with a duty cycle of 1%, corresponding to a width of mapped onto QPSK, 8PSK, 16APSK, or 32APSK constella-
0.5 m in accordance with Figure 2. The symbol rate was fixed tions. Modulated streams enter in the physical layer framing
to 27.5 Mbaud. where physical layer signalling and pilot symbols are inserted.
The considered DVB-S2 physical layer transmitter [2] is For energy dispersal, another scrambling sequence is applied
depicted in Figure 8. A continuous stream of MPEG pack- to the entire physical layer frame (PLFRAME). The system
ets passes through the mode adaptation which provides has been designed to provide a regular PLFRAME structure,
input stream interfacing. This data flow is passed to the based on slots of M = 90 modulated symbols, which allow
10
Single/multiple
input data streams
BB
Input signaling
interface no. 1
. Merge Stream
. LDPC
. slicer adapter
Input
interface no. n Mode & stream
adaptation
BBFRAME
Figure 8: DVB-S2 physical layer transmitter block diagram (taken from [2]).
reliable receiver synchronization on the FEC block struc-
ture. The first slot, PLHEADER, is devoted to physical layer
signalling, including start-of-frame (SOF) delimitation and
MODCOD definition. Receiver channel estimation is facil-
itated by the introduction of a set of P = 36 pilot sym-
bols, that are inserted every 16 slots. In addition, a pilot-
less transmission mode is also available, ensuring greater sys-
tem capacity. Finally, for shaping purposes, a squared-root
raised cosine (SRRC) filter with variable roll-off factors (0.2,
or 0.25, or 0.35) is considered. To cope with the intrinsic
nonlinearity of the on-board high power amplifier (HPA),
a purposely designed predistortion technique is considered.
In particular, a fractional predistortion technique based on
a lookup table (LUT) approach is considered which operates
right after the shaping filter [18]. The fractional predistorter,
which is a digital waveform predistorter, acts on the signal
samples for precompensating the HPA AM/AM and AM/PM
characteristics and mitigating the impact of non linear dis-
tortion. In particular, the signal is processed by means of
the LUT, which stores the inverted HPA coefficients com-
puted offline through analytic inversion of a proper HPA
model. The steps needed to obtain LUT coefficients are the
following: HPA model selection, parameter extrapolation, an-
alytical model inversion, and LUT construction. Regarding the
first step, a simple yet robust empirical model is the clas-
sic Saleh model [18]. Given the measured HPA character-
istics, the second step can be performed by minimizing the
energy of the difference between the modelled and the ex-
perimental HPA curves (MMSE criterion). These parameters
are then applied to the analytically inverted characteristics,
so as to obtain the analytical predistortion transfer function.
The last step is the quantization of the analytical curve in
order to store it into the LUT. The adopted strategy is lin-
ear in power indexing, that is, table entries are uniformly
spaced along the input signal power range, yielding denser
table entries for larger amplitudes, where nonlinear effects
reside.
The proposed digital receiver architecture is depicted in
Figure 4. In particular, several subsystems are present in or-
der to coherently demodulate and combine the received sig-
nals. The first coarse correction regards the carrier frequency,
EURASIP Journal on Wireless Communications and Networking
1/4, 1/3, 2/5, Roll-off factors:
1/2, 3/5, 2/3, PL signaling α = 0.2,
3/4, 4/5, 5/6, pilot symbols α = 0.25,
8/9, 9/10 α = 0.35
BCH QPSK
8PSK Scram BB
16APSK bler filter
bit interleaver
32APSK
Dummy
frame
FEC coding Mapping PL framing Modulation
FECFRAME PLFRAME To the RF
satellite
channel
which allows match filtering with minimal intersymbol in-
terference regrowth; then the subsequent block deals with
clock recovery for timing adjustment, performed by a digi-
tal interpolator. The demultiplexer is used to separate pilots
from data symbols in a PLFRAME. The pilot symbol stream
is used by the following four subsystems: the noise level esti-
mator, the digital automatic gain and angle control (AGAC),
the block in charge of tracking the residual frequency offset
and carrier phase, and finally the coarse frequency acquisi-
tion loop (not performed). On the other path, the data sym-
bols, softly combined with the last equation of Section 4.1,
feed the hard/soft demodulator. The demodulator provides
the hard decisions on data symbols as a feed-back for car-
rier frequency and phase tracking, and computes the soft ini-
tial a posteriori probability (APP) on the received informa-
tion bits. Finally, the APPs are deinterleaved and given to the
LDPC-BCH decoder. As far as frame synchronization and
frequency acquisition are considered, that is, dashed white
blocks in Figure 4, they are not considered in the simula-
tion chain because the receiver behaviour is assessed during
steady state.
5.2. Packet level coding simulation platform
A simulation platform to analyze the performance of GSE-
FEC has been developed. Given that this performance as-
sessment entails many layers, in particular, from the physical
to the network layers, of the protocol stack, a modular ap-
proach has been considered as the only feasible way to de-
velop the platform. The physical-layer simulator described
in the previous section interfaces with the packet-level sim-
ulator shown in Figure 9. This takes as input a stream of
IP packets and applies the GSE-FEC encoding technique as
described above, generating a sequence of BBFRAMEs. At
this point, the output of the physical-layer simulator is used
to mark the BBFRAMEs as correctly or wrongly received.
Next, the GSE-FEC decoding process is applied. The effect
of the BBFRAMEs on the GSE units and subsequently on the
columns of the reconstructed FEC matrix is calculated. Then,
the correction capability of the Reed-Solomon code is taken
into account to eliminate, if possible, the unreliable columns
Stefano Cioni et al. 11

1E + 0
IP PER
calculation
1E − 01

Mapping to
Power arches floor

PER
correct/wrong 1E − 02
IP packets
Traffic generation
1E − 03
Corrected FEC
IP packets matrix
1E − 04
−2 −1 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
GSE-FEC FEC decoding
Eb /N0 (dB)
1/2 - QPSK (LOS, FAST, noPA)
Mapping to 2/3 - 8PSK (LOS, FAST, noPA)
BBFRAMEs correct/wrong 3/4 - 16APSK (LOS, FAST, noPA)
FEC matrix 5/6 - 16APSK (LOS, FAST, noPA)
columns
1/2 - QPSK (LOS, FAST, MRC)
2/3 - 8PSK (LOS, FAST, MRC)
Mapping to 3/4 - 16APSK (LOS, FAST, MRC)
correct/wrong 5/6 - 16APSK (LOS, FAST, MRC)
Selective BBFRAME
corruption GSE units
Figure 10: MRC performance in LOS channel condition and train
speed equal to 300 km/h.

Physical-layer Time series of

simulation correct/wrong BBFRAMEs
Figure 9: Simulation platform at IP-BBFRAME level.
of the FEC matrix. Finally, the list of IP packets affected by
the unreliable columns (an IP packet is considered wrong if
any part of it falls inside an unreliable column which cannot
be corrected) is obtained and the packet error rate (PER) at
IP level is computed.
The packet-level simulator is useful to assess very quickly
the performance of different parameter configurations of
the GSE-FEC since different combinations can be simulated
without the need of repeating the time-consuming physical
layer simulations. The main parameters of GSE-FEC to be
designed are the following:
(i) size of the columns of the FEC matrix,
(ii) size of GSE units,
(iii) number of padding columns in the first part of the FEC
matrix,
(iv) number of punctured redundancy columns.
The effect of varying some of these parameters will be shown
in the numerical results section.
6. RESULTS
6.1. Antenna diversity
Numerical results have been obtained by considering the
entire transmit-receive chain described in Section 5.1. The
introduction of the second receiving antenna adopting the
MRC technique is reported in Figure 10. The most impor-
tant result is that the MRC solution completely eliminates
the error floor with respect to the single antenna case (see
for comparison Figure 3). Secondly, it will be observed that
instead of a constant 3- dB gain for all Eb /N0 values, three
different working regions can be distinguished. In particular,
BBFRAME error rates curves are characterized by two water-
fall regions separated by a short floor. This unexpected be-
haviour has a theoretical explanation that has been treated
in details in [14]. Here, we limit the discussion to a numeri-
cal example. Let us consider MODCOD = 1/2-QPSK and a
working Eb /N0 = 0 dB, when a PA blockage event occurs, the
“nonobscured” antenna has not a sufficient SNR to reliably
decode the received MPEG packets, thus generating an error
floor at that Eb /N0 . The second waterfall region starts only
for Eb /N0 values larger than 1 dB, when, as a matter of fact,
a single antenna receiver has sufficient margin to correctly
decode. This consideration can also be extended to all other
MODCOD configurations. Notably, the short floor value is
twice the floor value obtained with one receiving antenna;
this is determined by the fact that there are two blockage
events between two consecutive PA, that is, one per receiv-
ing antenna.
6.2. Packet level FEC
The objective of the following analysis is twofold: first, to
provide a guideline for an appropriate choice of the column
size of the FEC matrix, which is the key parameter in the
GSE-FEC method; second, to analyze the performance of
GSE-FEC under various configurations. In all cases, a sce-
nario with line-of-sight propagation has been used.
12 EURASIP Journal on Wireless Communications and Networking

1/2-QPSK 2/3-8PSK 5/6-16APSK

1 1 1

0.9 0.9 0.9

0.8 0.8 0.8

0.7 0.7 0.7

0.6 0.6 0.6

Probability

Probability

Probability
0.5 0.5 0.5

0.4 0.4 0.4

0.3 0.3 0.3

0.2 0.2 0.2

0.1 0.1 0.1

0 0 0
0 2 4 6 8 10 12 14 0 2 4 6 8 10 12 14 0 2 4 6 8 10 12 14
Error burst length Error burst length Error burst length
(a) (b) (c)

Figure 11: Histogram of the BBFRAME error burst length for two different MODCOD modes and target BBFRAME error rate equal to
0.02.

6.2.1. Dimensioning the FEC matrix FRAMEs and hence BBFRAMEs are affected by each power
arch.
In order to present the procedure to compute the col-
First of all, it is worth remarking that the appropriate size
umn size of the FEC matrix, we consider a numerical exam-
of the FEC matrix depends on the length of the bursts of
ple. We use for instance the least efficient MODCOD, that
erroneous BBFRAMEs. It is clear that longer bursts will re-
is, 1/2-QPSK. It can be seen in Figure 11 that the maximum
quire larger FEC matrices to avoid that the number of wrong
error burst length due to power arches is 7 BBFRAMEs. In
columns exceeds the correction capability of the code. There-
this MODCOD, each BBFRAME has a data field of length
fore, the design of the height of the FEC matrix should be
32128 bits [2], which is equal to 4016 bytes. Therefore, a
derived from an analysis of the length of the error bursts.
burst of 7 BBFRAMES corresponds to 28112 bytes. We con-
Figure 11 shows the histogram of the length of the bursts for
sider that this amount of bytes should correspond to less
some particular MODCOD modes for the scenario described
than 30 columns in the FEC matrix. The value of 30 has
above. In all modes besides the two shown in Figure 11,
been chosen arbitrarily. It is nevertheless a reasonable num-
it is observed that the distribution is bimodal. The bursts
ber since the objective is to leave a margin with respect to the
of short length (typically between 1 and 4 BBFRAMEs) are
64 columns that the code can correct (assuming no punctur-
due to random errors caused by noise, whereas the rest of
ing) so as to be able to cope with errors caused by noise as
bursts are caused by the power arches. Second, the higher
well. Therefore, the column size of the FEC matrix should
the modulation order, the longer the error bursts produced
fulfil
by power arches are. This is justified by the fact that ac-
cording to the DVB-S2 standard, BBFRAMEs are coded and
converted into FECFRAMEs, which have constant length 30Lc ≥ 28112 =⇒ Lc ≥ 938 bytes, (7)
in bits regardless of the used modulation [2]. The bits in
the FECFRAME are transformed by the modulator into where Lc is the number of rows (i.e., the length of each col-
bytes in the PLFRAMEs. Higher modulations need fewer umn) of the FEC matrix in bytes. In the previous compu-
symbols and, hence, less time to transmit an FECFRAME. tation, we have not taken into account the overhead intro-
The duration of the fade event caused by a power arch duced by GSE since it is small and we are only interested
only depends on the speed of the train, which we have in obtaining an approximate value for the column size. If
considered to be 300 km/h throughout the rest of the pa- the same calculation is repeated for the most efficient MOD-
per. Therefore, the shorter the PLFRAME, the more PL- COD, that is, 5/6-16APSK, the result is Lc ≥ 2912 bytes. The
Stefano Cioni et al. 13

1/2-QPSK 2/3-8PSK
0.12 0.12

0.1 0.1
IP packet error rate

IP packet error rate

0.08 0.08

0.06 0.06

0.04 0.04

0.02 0.02

0 0
0 1000 2000 3000 4000 5000 6000 0 1000 2000 3000 4000 5000 6000
Column size (bytes) Column size (bytes)
(a) (b)

3/4-16APSK 5/6-16APSK
0.12 0.14

0.1 0.12

0.1
IP packet error rate

IP packet error rate

0.08
0.08
0.06
0.06
0.04
0.04

0.02 0.02

0 0
0 1000 2000 3000 4000 5000 6000 0 1000 2000 3000 4000 5000 6000
Column size (bytes) Column size (bytes)
(c) (d)

Figure 12: Comparison of the IP packet error rate for different ACM modes in a channel with BBFRAME error rate equal to 12% (circles →
results without any kind of PL-FEC, squares → results with GSE-FEC).

results for the intermediate MODCODs, 2/3-8PSK and 3/4-
16APSK, are 1790 and 2618 bytes, respectively.
We conclude from this discussion that the appropriate
size of the FEC matrix strongly depends on the error burst
length caused by the power arches, which in its turn depends
on the train speed. The lower the train speed is, the longer
the bursts are and the taller the FEC matrix must be. How-
ever, the size of the FEC matrix cannot be increased arbitrar-
ily because it has an impact on the delay of GSE-FEC process
and, on top of that, because more errors due to noise appear
inside the FEC matrix. These errors may risk the correction
capability of the code, as will be seen below. Therefore, the
performance of GSE-FEC may be limited for low train speeds
since it is not possible to combat simultaneously very long er-
ror bursts due to power arches and a large amount of random
errors due to noise.
6.2.2. Performance analysis
Dependence on the size of the FEC matrix
The IP packet error rate as a function of the column size for
different MODCODs is shown in Figures 12 and 13. The con-
sidered columns sizes and the corresponding number of GSE
units used to encapsulate each RS redundancy column are
listed in Table 2. The number of GSE units per column has
been selected in such a way that the size of the units is small
enough to limit the amount of padding in the BBFRAMEs,
but large enough not to penalize encapsulation efficiency
(encapsulation efficiency is out of the scope of this work and
will be analyzed in a follow-on paper). A fixed IP packet
length equal to 576 bytes has been considered.
14 EURASIP Journal on Wireless Communications and Networking

1/2-QPSK 2/3-8PSK
0.025 0.025

0.02 0.02
IP packet error rate

IP packet error rate

0.015 0.015

0.01 0.01

0.005 0.005

0 0
0 1000 2000 3000 4000 5000 6000 0 1000 2000 3000 4000 5000 6000
Column size (bytes) Column size (bytes)
(a) (b)

3/4-16APSK 5/6-16APSK
0.02 0.03

0.025
0.015
IP packet error rate

IP packet error rate

0.02

0.01 0.015

0.01
0.005
0.005

0 0
0 1000 2000 3000 4000 5000 6000 0 1000 2000 3000 4000 5000 6000
Column size (bytes) Column size (bytes)
(c) (d)

Figure 13: Comparison of the IP packet error rate for different ACM modes in a channel with BBFRAME error rate equal to 2% (circles →
results without any kind of PL-FEC, squares → results with GSE-FEC).

Figures 12 and 13 also compare the results obtained when
GSE-FEC is used and when no packet-level FEC is applied.
The baseline GSE-FEC is employed, that is to say, no ad-
ditional padding has been used in the first 191 columns
and no puncturing of the last 64 columns has been per-
formed. The case of no packet-level FEC follows the same
architecture as for GSE-FEC, depicted in Figures 6 and 7.
The difference is that the 255 columns of the FEC ma-
trix are filled with IP packets and no redundancy is intro-
duced into it. Figure 12 was obtained when the physical-
layer simulator was tuned to provide a BBFRAME error rate
around 0.12, whereas Figure 13 was obtained for a value
of 0.02.
In the case of no packet-level FEC, the IP PER is almost
insensitive to changes in the column size and its value is very
close to the BBFRAME error rate, as expected. It is very in-
teresting to observe that the proposed scheme, GSE-FEC, ef-
fectively reduces the IP PER and, in many configurations, the
IP PER is exactly zero.3 This means that, in those cases, all
3 Note that the simulation duration was equal to 5000 BBFRAMEs, so we
can only say that the IP PER is not worse than 2 × 10−5 .
IP packets were correctly received in spite of the fact that the
BBFRAME error rate is higher than 10%.
For small column sizes, the IP PER decreases as the col-
umn size increases. This behaviour is in line with the discus-
sion at the beginning of this section: when the FEC matrix
is too small, a power arch causes errors in a portion of the
matrix that is too large to be corrected by the code. The IP
PER decreases until it reaches a minimum, which is attained
at a column length that is well approximated by the previ-
ous back-of-the-envelope calculations. If the column length
is increased further, the IP PER increases because the correc-
tion capability of the code is fixed and equal to 64 columns,
but the size of the FEC matrix becomes larger and, hence,
the number of errors due to noise increases. This behaviour
is visible in Figure 12, but not in Figure 13. The reason is
that the later figure corresponds to a scenario with very high
signal-to-noise ratio, and BBFRAME errors are almost only
caused by power arches.
Dependence on the IP packet length
The effect of different IP packet lengths is shown in Figure 14.
In this case, the column size of the FEC matrix is fixed
Stefano Cioni et al.
1/2-QPSK
0.16
0.14
0.12
IP packet error rate
0.1
0.08
0.06
0.04
0.02
0 0
0 500 1000
IP packet length (bytes)
No PL-FEC
Column size: 1024 bytes
Column size: 4096 bytes
(a)
Figure 14: Dependence of the IP packet error rate with the IP packet size for two column sizes (1024 and 4096 bytes) and two MODCOD
modes (1/2-QPSK and 5/6-16APSK).
and equal to 1024 or 4096 bytes. The general trend is that
the IP PER slightly increases as the IP size increases. There
are however some lengths, such as 576 bytes, that are espe-
cially favourable. This happens because for those lengths an
integer number of IP packets fit in an integer number of
columns of the FEC matrix. For instance, it is fulfilled that
576 × 16 = 1024 × 9, which means that 16 IP packets of
length 576 bytes fit in 9 columns of length 1024 bytes. As this
perfect fitting reduces the ratio of IP packets that are split
across two columns, the number of IP packets corrupted by
a wrong column is also reduced on average. If the length of
IP packets follows a certain distribution, as it happens with
real traffic, the IP PER can be obtained by computing an av-
erage of the values shown in Figure 14. This average would
be computed by weighting the IP PER for a given length by
the frequency of occurrence of that length.
Conclusions on GSE-FEC results
The analysis of the GSE-FEC and the corresponding numer-
ical results has shown that the column size is a key design
parameter. Long columns appropriate to obtain low IP PER
when the duration of the fade events caused by power arches
is large (e.g., when the train is moving slowly) or when very
spectrally efficient MODCODs are used; but this comes at
the price of a large encoding and decoding delay, and an in-
creased sensitivity to random BBFRAME errors caused by
noise and interference. Therefore, the column size must be
selected as the result of a tradeoff between competing goals;
15
5/6-16APSK
0.16
0.14
0.12
IP packet error rate
0.1
0.08
0.06
0.04
0.02
1500 0 500 1000 1500
IP packet length (bytes)
No PL-FEC
Column size: 1024 bytes
Column size: 4096 bytes
(b)
it is not possible to propose a single value appropriate for
all scenarios. We consider that the column size must be an
adaptive parameter, which is changed in response to vari-
ations of the propagation conditions, train speed, and so
forth. This adaptation would constitute an example of cross-
layer optimization, whereby a link layer parameter (i.e., the
column size of the FEC matrix) is adapted as function of
the physical-layer conditions. The padding and puncturing
of columns in the FEC matrix are other degrees of freedom
that can be exploited in the parameterization of GSE-FEC.
A detailed analysis of these aspects is a subject for further
research.
6.3. Comparative analysis
As it can be seen from the results presented in the last two
sections, very satisfactory results to ensure reliable reception
can be obtained with both techniques. In the case of antenna
diversity, this does not penalize the overall system efficiency,
although some additional complexity in the receiver imple-
menting the MRC scheme will be accounted for. However,
the main issue to be addressed in the practice is represented
by the installation of two antennas. Many experiments and
trials have shown that this is a very critical point, since anten-
nas suitable for installation on trains are subject to very strict
requirements in terms of pointing accuracy, size, and ro-
bustness against mechanical vibrations, wind, pressure gra-
dients when entering or exiting a tunnel, and so forth. With
current antenna technologies, a relatively high failure rate
16
Table 2: Parameters of the GSE-FEC algorithm.
FEC-matrix column
256 512 768 1024 2048 3072 4096 5120
size (bytes)
GSE units per column 1 1 1 2 2 3 4 5 (www.satnex.org), FP6 Contract IST-507052.
of mechanical components included in the antenna plat-
form has to be expected. Furthermore, train operators are
extremely keen on keeping the installation and maintenance
procedures as simple as possible. For all these reasons, addi-
tional countermeasures must be also investigated as possible
complement to the presence of two antennas (e.g., in case
one antenna suddenly breaks and no immediate replacement
is possible).
Although it has been shown that the dimensioning of
packet level FEC is a complex task, that will be carried out
following a cross-layer approach, the results presented in the
previous section confirm that also this technique, if properly
designed, can guarantee reliable reception at the expenses of
a limited increase in the system complexity and overhead.
The concrete solution presented in this paper has been es-
pecially devised taking into account the architectural con-
straints introduced by the latest encapsulation scheme (GSE)
currently being proposed for future DVB systems. Clearly,
packet level FEC results in a reduction of the overall spectral
efficiency of approximately 33% with the adopted RS code,
partially compensated by the migration to a more efficient
encapsulation scheme such as GSE.
7. CONCLUSIONS
To conclude, two countermeasures are thoroughly analyzed
in this paper: antenna diversity and a packet-level forward
error correction mechanism especially tailored to DVB-S2,
named GSE-FEC. Simulations have shown the excellent per-
formance of both approaches, while they have complemen-
tary features in terms of hardware complexity, delay, and
bandwidth efficiency. Generally speaking, the results in this
paper show that effective countermeasures to compensate the
impairments of the railroad satellite channel are possible and
can be integrated into the existing DVB-S2 standard with a
limited to moderate impact on the receiver design and on the
system complexity. In fact, to support antenna diversity, the
receiver structure will be modified as depicted in Figure 4,
whereas for packet level FEC a software implementation may
be considered.
Further topics to be addressed in order to conclude the
analysis of the forward link are the following:
(i) cross-layer optimization of all the relevant parameters
(MODCODs and GSE-FEC), taking also into account
nLOS channel conditions and the usage of ACM to
compensate for slower fades due to atmospherical ef-
fects,
(ii) inclusion of mechanizm(s) to support QoS and study
of their integration and interaction with the proposed
GSE-FEC scheme.
EURASIP Journal on Wireless Communications and Networking
ACKNOWLEDGMENT
This work was supported and partially funded by Sat-
NEx, the Satellite Communications Network of Excellence
REFERENCES
[1] EN 300 421 v1.1.2: Digital Video Broadcasting (DVB); Fram-
ing structure, channel coding and modulation for 11/12 GHz
satellite services.
[2] ETSI EN 302 307 v1.1.1: Digital Video Broadcasting (DVB):
Second generation framing structure, channel coding and
modulation system for Broadcasting, Interactive Services,
News Gathering and other broadband satellite applications.
[3] ETSI EN 301 790 v1.4.1: Digital Video Broadcasting (DVB):
Interaction channel for satellite distribution systems.
[4] ETSI TR 101 790 v1.3.1: Digital Video Broadcasting (DVB):
Interaction channel for satellite distribution systems; Guide-
lines for the use of EN 301 790.
[5] ETSI EN 302 304 v1.1.1: Digital Video Broadcasting (DVB);
Transmission System for Handheld Terminals (DVB-H).
[6] S. Scalise, R. Mura, and V. Mignone, “Air interfaces for satellite
based digital TV broadcasting in the Railway environment,”
IEEE Transactions on Broadcasting, vol. 52, no. 2, pp. 158–166,
2006.
[7] E. Lutz, M. Werner, and A. Jahn, Satellite Systems for Per-
sonal and Broadband Communications, Springer, New York,
NY, USA, 2000.
[8] S. Scalise, H. Ernst, and G. Harles, “Measurement and mod-
elling of the land mobile satellite channel at Ku-band,” to ap-
pear in IEEE Transactions on Vehicular Technology.
[9] E. Kubista, F. P. Fontan, M. A. V. Castro, S. Buonomo,
B. R. Arbesser-Rastburg, and J. P. V. Polares Baptista, “Ka-
band propagation measurements and statistics for land mobile
satellite applications,” IEEE Transactions on Vehicular Technol-
ogy, vol. 49, no. 3, pp. 973–983, 2000.
[10] A. Benarroch and L. Mercader, “Signal statistics obtained form
a LMSS experiment in Europe with the MARECS satellite,”
IEEE Transactions on Communications, vol. 42, no. 2–4, pp.
1264–1269, 1994.
[11] G. Sciascia, S. Scalise, H. Ernst, and R. Mura, “Statistical char-
acterization of the railroad satellite channel at Ku-band,” in
Proceedings of the International Workshop of Cost Actions 272
and 280, Noordwijk, The Netherlands, May 2003.
[12] S. Scalise, O. Lücke, and E. V. Torralbo, “A link availability
channel model for the railroad satellite channel,” in Proceed-
ings of 24th AIAA International Communications Satellite Sys-
tems Conference (ICSSC ’06), vol. 1, pp. 305–317, San Diego,
Calif, USA, June 2006.
[13] S. Cioni, G. E. Corazza, and A. Vanelli-Coralli, “Antenna di-
versity for DVB-S2 mobile services in Railway environments,”
to appear in Journal of Satellite Communications and Networks,
special issue on ASMS Conference.
[14] S. Cioni, M. Berdondini, G. E. Corazza, and A. Vanelli-Coralli,
“Antenna diversity for DVB-S2 mobile services in Railway en-
vironments,” in Proceedings of the 3rd Advanced Satellite Mobile
Systems (ASMS) Conference, Herrsching am Ammersee, Ger-
many, May 2006.
[15] S. Cioni, A. Vanelli-Coralli, C. Párraga Niebla, S. Scalise, G.
Seco Granados, and M.A. Vázquez Castro, “Antenna diver-
sity and GSE-based packet level FEC for DVB-S2 systems
in Railway scenarios,” in Proceedings 25th AIAA International
Stefano Cioni et al. 17

Communications Satellite Systems Conference, Seoul, South Ko-

rea, April 2007.
[16] ETSI TR 102 377 v1.2.1: Digital Video Broadcasting (DVB);
DVB-H Implementation Guidelines.
[17] DVB Blue Book A116 - Generic Stream Encapsulation
Specification. http://www.dvb.org/technology/bluebooks/a116
.tm3762r1.gbs0436r10.GSE spec.pdf.
[18] P. Salmi, M. Neri, and G. E. Corazza, “Design and perfor-
mance of predistortion techniques in Ka-band satellite net-
works,” in Proceedings of the 22nd AIAA International Commu-
nications Satellite Systems Conference and Exhibit (ICSSC ’04),
vol. 1, pp. 281–291, Monterey, Calif, USA, May 2004.
Hindawi Publishing Corporation
EURASIP Journal on Wireless Communications and Networking
Volume 2007, Article ID 14798, 10 pages
doi:10.1155/2007/14798

Research Article
Capacity Versus Bit Error Rate Trade-Off in
the DVB-S2 Forward Link

Matteo Berioli, Christian Kissling, and Rémi Lapeyre

German Aerospace Center (DLR), Institute of Communications and Navigation, Oberpfaffenhofen, 82234 Wessling, Germany

Received 5 October 2006; Accepted 12 March 2007

Recommended by Ray E. Sheriff

The paper presents an approach to optimize the use of satellite capacity in DVB-S2 forward links. By reducing the so-called safety
margins, in the adaptive coding and modulation technique, it is possible to increase the spectral efficiency at expenses of an
increased BER on the transmission. The work shows how a system can be tuned to operate at different degrees of this trade-off,
and also the performance which can be achieved in terms of BER/PER, spectral efficiency, and interarrival, duration, strength of
the error bursts. The paper also describes how a Markov chain can be used to model the ModCod transitions in a DVB-S2 system,
and it presents results for the calculation of the transition probabilities in two cases.

Copyright © 2007 Matteo Berioli et al. This is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

1. INTRODUCTION mitigation techniques like power control result in an inef-

The original DVB-S standard dates back to 1995 and was in-
tended for delivery of broadcasting services, the underlying
transport stream of DVB-S was defined to be MPEG-2. DVB-
S2 [1] is the second generation of the DVB-S standard and
comprises a variety of new features. It can be used for provi-
sion of HDTV (high definition television) but it also allows
for transportation of different multimedia streams such as,
for example, internet traffic, audio and video streaming and
file transfers with support of different input stream formats
such as IP, ATM, single/multiple MPEG streams or generic
bit streams, both for broadcast and unicast transmissions.
For the support of interactive applications a return channel
is necessary which can be provided by DVB-RCS [2]. DVB-
S2 can achieve a capacity increase of up to 30% under the
same transmission conditions compared to the older DVB-S
standard what is achieved by applying higher order modu-
lation schemes and by the use of low density parity check
codes (LDPC) and Bose-Chaudhuri-Hochquenghem (BCH)
codes.
The real novelty introduced by DVB-S2 was the pos-
sibility to use adaptive coding and modulation (ACM). In
traditional nonadaptive systems the link dimensioning has
to be made under considerations of service availability and
worst case channel assumptions due to the deep fades caused
by atmospheric effects; as a consequence the classical fade
ficient use of the system capacity since most of the time more
transponder power than necessary is used. On the other hand
in case of ACM, if a terminal is able to inform the gateway of
its particular channel conditions (by means of a proper re-
turn link) the gateway can select an appropriate waveform,
coding and modulation, to best exploit the spectrum and at
the same time to overcome the channel impairments.
An efficient exploitation of the expensive satellite capac-
ity has always been a key factor in the development of the
satellite market, and the improvements brought by DVB-S2
give promising perspectives for the future of satellite commu-
nications. Nevertheless it is important to keep improving the
exploitation of the satellite bandwidth, in order to guaran-
tee reduced costs for all satellite services (broadcast, Internet,
etc.). The aim of this work is to go one step further in this
trend and to try to optimize the throughput and the spec-
trum efficiency in DVB-S2 forward links.
Today DVB-S2 links offer to the higher-layers protocols
a terrestrial-like transmission medium, with recommended
PERs around 10−7 . This is of course an excellent result, but
not all services at higher layers require to reach such out-
standing performance. This is in particular true for Internet
and multimedia services [3].
Some audio codecs (e.g., AMR [4]) can typically accom-
modate packet losses with only a small impact in quality, and
up to 15% failures before the speech is severely degraded.
2 EURASIP Journal on Wireless Communications and Networking
Other modern media codecs (e.g., MPEG-4 [5]) have been
designed to be highly resilient to residual errors in the in-
put bit-stream, to detect and localize errors within the packet
payload, and to employ concealment techniques, like for in-
stance interframe interpolation, that hide errors from a hu-
man user. These codecs offer acceptable quality at a resid-
ual BER poorer than 10−3 , and some at poorer than 10−2
[6]. In order to support these error-tolerant codecs, the IETF
has also standardized a new multimedia transport proto-
col, UDP-Lite [7], that allows to specify the required level of
payload protection, while maintaining end-to-end delivery
checks (verification of intended destination, IP header fields
and overall length).
When these services are operating over the satellite con-
nection, it is convenient to reduce the quality of the trans-
mission in DVB-S2 forward links, by allowing higher BERs,
in order to increase the precious capacity and the through-
put. The first motivation for this is to make use of cross-layer
mechanisms by voluntarily allowing higher bit error rates
which can be compensated with error correction at higher
layers. A second motivation to allow for higher BERs is that
not all applications have the same stringent BER require-
ments. This represents a natural trade-off between errors and
capacity. The present work analyzes this trade-off, proposes
a way to tune the system parameters in order to work in
optimal conditions, and investigates the performance of the
system in this situation. The work is organized as follows:
Section 2 presents the background and the scenario of the
subject, Section 3 describes the main ideas of the paper and
the original approach to the problem, Section 4 evaluates the
performance of a system operating in the suggested condi-
tions, and Section 5 drives the conclusions of the paper.
2. BACKGROUND AND SCENARIO
2.1. Overview of DVB-S2
The second generation of DVB-S provides a new way of fade
mitigation by means of adaptation of the coding and modu-
lation (ACM) to the different channel states. This of course
implies the need for every terminal to signal its perceived
channel state back to the gateway which can then make a
frame-by-frame decision of the modulation and coding com-
bination (ModCod) to be applied based on these measure-
ments. DVB-S2 offers a broad range of modulations and cod-
ings for ACM. The supported modulation schemes comprise
QPSK, 8-PSK, 16-APSK, 32-APSK and considered coding
rates are 1/4, 3/4, 1/3, 2/5, 3/5, 4/5, 1/2, 5/6, 8/9, 9/10. The
possibility to select the modulation and coding for an indi-
vidual destination allows to make a more efficient use of the
system capacity since transmission in a higher-order modu-
lation in combination with a low coding rate (e.g., for clear
sky conditions) allows to transmit more bits per symbol than
a low-order modulation with high coding rate (e.g., for rainy
channels). In this way it is possible to use individually for ev-
ery ground terminal (or for every group of terminals in the
same spot beam) the highest possible modulation scheme
and the lowest coding rate which still allows to cope with
the channel impairments to provide low BER. A destination
with a bad channel state can thus use a very robust modula-
tion and coding pair (ModCod) while other terminals with a
very good channel state can still transmit in highly efficient
ModCods. The adaptive selection of the best suited ModCod
results in an increased net data throughput while terminals
in bad channel conditions are still able to receive their data
since they can use ModCods with lower-order modulation
and higher coding (but at the cost of lower spectral efficiency
and thus lower throughput).
As can be seen in Figure 1 the system architecture of
DVB-S2 is subdivided into six main components [1]. The
mode adaptation subsystem provides an interface to the ap-
plication specific data stream formats and also contains a
CRC-8 error detection coding scheme. It is possible to merge
different input streams together and to segment them into
the so-called data fields which are the payload part of the so-
called baseband-frames (BBFRAME) created at the output
of the consecutive stream adaptation module. Buffers store
data until they are processed by the merger/slicer and in case
not enough data is available to fill a data field or if it is re-
quired to have only an integer number of packets in a frame
(in general integer number of packets will not perfectly fit
into a frame but their payload sum will always be slightly
smaller or larger than the data field), the unused space can be
padded, this operation is accomplished by the stream adap-
tation subsystem. In order to complete the baseband frame
(BBFRAME) additional header information (BBHEADER)
is added in front of the data field and scrambling of header
and payload is applied. The final BBFRAME structure is il-
lustrated in Figure 2.
The consecutive FEC encoding block performs outer and
inner coding and bit interleaving. The coding scheme which
is used is selected based on the channel measurements re-
ceived from the terminals the data of which is contained in
the frame. The outcome of this module, called forward error
correction frame (FECFRAME), is shown in Figure 3. The
FECFRAMEs can either have a length of 16200 bits for short
frames or 64800 bits for normal frames. Since the length
of the encoded frame is fixed, this means that the length
of the payload in the underlying BBFRAME changes with
the applied coding. For applying higher-order modulation
schemes the subsequent mapping block performs a serial-to-
parallel conversion. The mapper chooses the applied mod-
ulation schemes again based on the channel measurements
for the destination(s) of the data contained in the frame. The
outcome of the mapping of the data into symbols is called
an XFECFRAME which is afterwards formed into a physical
layer frame (PLFRAME) after pilots and PL signalling have
been inserted and after final scrambling for optimization of
energy dispersal. In case no XFECFRAMES are provided by
the preceding subsystems, the PLFRAMING module inserts
the so-called DUMMY PLFRAMES to provide a continuous
TDM stream on the link. To allow every terminal indepen-
dent of its channel state to receive the PLHEADER informa-
tion (which also contains the used modulation and coding
scheme for the underlying frame) this header is always mod-
ulated with BPSK.
Matteo Berioli et al. 3

Mode adaptation
Data BB
Single signalling
input Input Null-packet Dotted subsystems are
Input stream CRC-8
stream interface synchroniser deletion encoder Buffer not relevant for
ACM (ACM, TS) single transport stream
command Merger broadcasting
slicer applications
Multiple
input Null-packet
Input Input stream CRC-8 Buffer
streams deletion
interface synchroniser encoder
(ACM, TS)

QPSK,
8PSK, PL signalling & α = 0.35, 0.25, 0.2
Rates 1/4, 1/3, 2/5 16APSK, pilot insertion
1/2, 3/5, 2/3, 3/4, 4/5, 32APSK
5/6, 8/9, 9/10
I PL
Bit scram BB filter
BB BCH LDPC Bit mapper Q BLER and
Padder scram encoder encoder inter- into quadrature
BLER (nbch , kbch ) (nldpc , kldpc ) constellations Dummy
leaver modulation
PLFRAME
Stream insertion
adaptation FEC encoding Mapping PLFRAMING Modulation

LP stream for To the RF

BBHEADER BC modes
BBFRAME FECFRAME PLFRAME satellite
data field
channel

Figure 1: DVB-S2 system architecture [1].

80 bits DFL Kbch -DFL-80 eral hundreds milliseconds (250 milliseconds). This means
that though the order of magnitude for the propagation de-
BBHEADER Data field Padding lay allows for a compensation of very slow changing channel
BBFRAME (Kbch bits) effects, like rain attenuation, it is too long to compensate fast,
high-frequent changes in the SNR as those caused by scintil-
Figure 2: Structure of a BBFRAME [1]. lation, this will be explained in the next section.

Nbch = kldpc 2.2. Channel modelling

Kbch Nbch -Kbch
BBFRAME BCHFEC
(nldpc bits)
Figure 3: Structure of a FECFRAME [1].
For the selection of a ModCod that is adapting to the in-
dividual experienced channel states of the terminals, a return
link must be provided to give feedback information about
the measured channel states to the gateway. The gateway can
then use this information to select a ModCod that suits trans-
missions in this channel state. This means the ModCod is
selected to provide a quasi-error-free transmission as long as
the critical SNR (signal-to-noise ratio) demodulation thresh-
old for this ModCod (thrdem (ModCod)) is not crossed. If the
signal drops below thrdem (ModCod) then the BER will dras-
tically increase due to the nature of the applied LDPC and
BCH coding of having very steep BER-versus-SNR curves. In
the GEO-stationary scenario investigated here, the propaga-
tion delay of the information feedback from the terminal to
the gateway takes relatively long and it is in the order of sev-
nldpc -kldpc
The selection of a ModCod scheme for transmission is very
decisive for the performance of the system in terms of net
LDPCFEC data rate, bit errors and, respectively, packet errors. If the
ModCods are selected too aggressively (meaning selection of
ModCods with a too high modulation scheme and a too low
coding) the transmission will result in a drastically higher
PER. On the other hand, selection of safe ModCods (mean-
ing a ModCod with a modulation lower than what would be
necessary and a coding higher than necessary) will result in
inefficiencies which reflects in a lower net data rate. In or-
der to evaluate the influence of different parameters for the
ModCod selection it is important to have a realistic chan-
nel model. The channels in satellite systems face mainly two
sources of signal fading, rain attenuation and scintillation.
The effect of rain attenuation is very significant for systems
operating in K-band where the signal is attenuated by ab-
sorbing effects of the water. The second effect coming along
with rain attenuation is scintillation which is basically a high
frequent distortion of the signal amplitude and phase caused
by small-scale irregularities in electron density in the iono-
sphere [8].
The scintillation in K-band can be considered to be a nor-
mal distributed random variable with a non linear spectrum
(see [9, 10]) as shown in Figure 4. The standard deviation
4 EURASIP Journal on Wireless Communications and Networking
Power spectral density of rain attenuation and scintillation
104
102
∼ f −2
PSD (dB2 /Hz)
100
10−2
10−4
10−6
10−4 fa 10−3 10−2 10−1
Frequency (Hz)
Rain attenuation
Scintillation
Figure 4: Attenuation and scintillation spectrum (typical values:
fa ≈ 10−4 Hz, fs ≈ 0.1 − 0.65 Hz).
of the scintillation process can be calculated according to (1)
corresponding to the theory of Tatarskii [9] and the model of
Matricciani [10],
σ = σ0 · A5/12 .
The value σ0 is the standard deviation of the scintillation for
a rain attenuation of A[dB]. [10] suggests a typical value of
0.039 for σ0 in the frequency range of 19.77 GHz. According
to (1) the resulting scintillation standard deviation σ is then
in the order of tenths of a dB for rain attenuations smaller
than 20 dB.
Within this work the main focus is on the scintillation
effects since these cannot be compensated by signalling of
the channel states via the return channel because of the
long propagation delay of the GEO satellite. Nevertheless the
channel simulations used in the rest of this work consider
spatial correlated rain attenuation as well since the magni-
tude of the scintillation also depends on the intensity of the
rain attenuation (see (1)). Similar to the generation of the
scintillation, also the rain attenuation is created via a normal
distributed random variable whereas its spectrum has a dif-
ferent corner frequency of fa (see also Figure 4).
Figure 5 shows a channel example for the attenuation
caused by scintillation and rain for a user located at longi-
tude 8.6◦ E and latitude 52.7◦ N, in the area around Hamburg
(Germany). It can be seen here that scintillation effects occur
with a much higher frequency than regular rain attenuation
events and how rain attenuation and scintillation are corre-
lated.
2.3. ModCod switching strategies
While the rain attenuation occurs on a larger time scale scin-
tillation effects occur very rapidly. For this reason rain fad-
Scintillation and attenuation time series
of useful user (forward downlink)
0.25 4.5
0.2 4
0.15
3.5
0.1
Scintillation (dB)
Attenuation (dB)
3
0.05
2.5
0
2
−0.05
1.5
−0.1
−0.15 1
∼ f −8/3
−0.2 0.5
−0.25 0
0 1000 2000 3000 4000 5000 6000 7000 8000
fs 100 101 Time (s)
Figure 5: Example of scintillation and rain attenuation.
ing can be mitigated by mode adaptation whereas counter-
measures for scintillation require a different compensation.
For every combination of modulation and coding a threshold
thrdem (ModCod) exists which is needed to be able to decode
the frame with a quasi-zero BER. The decision of the gateway
on which ModCod will be used is thus driven by thresholds.
For switching among ModCods these thresholds could theo-
(1) retically be used directly for the decision about which Mod-
Cod will be used, but in practice this would result in frequent
transmission errors since the high frequent variations of the
channel (due to scintillation) would cause a frequent crossing
of the threshold. On the one hand, this high frequent cross-
ing cannot be compensated by signalling to the gateway, on
the other hand such signalling would also mean a high fre-
quent change of the ModCod which is as well undesirable.
To provide more reliability the minimal needed demod-
ulation thresholds thrdem (ModCod) can be replaced by
thresholds which have a certain safety margin. This means
that a lower ModCod is selected already before the critical
threshold (the threshold below which a strong increase in
bit error rate occurs) is reached. The size of the safety mar-
gin does thus determine the robustness against fast occurring
scintillation fades. On the other hand, this size of the safety
margin also influences the system performance since trans-
mission in a higher ModCod would result in a higher net
data rate. Since fast oscillations between neighboring Mod-
Cods are also possible when safety margins are used, an ad-
ditional hysteresis margin is introduced. Figure 6 illustrates
the different thresholds and margins.
Within Figure 6 the terms thrdem (N − 1) and thrdem (N)
denote the minimum SNR values which are just enough to
provide quasi error free decoding. If, for example, ModCod
N is used and the signal strength falls below the thrdem (N)
threshold, the BER will drastically increase. These thresholds
have also been called critical in [11] for this reason. If on the
other hand the signal strength increases, for example, while
using ModCod N − 1, the next higher ModCod is not selected
as soon as the demodulation threshold of the next higher
Matteo Berioli et al.
ModCod (thrdem (N) in this case) is crossed but after a higher
threshold is exceeded (threnable (N − 1)).
In case the signal strength decreases again, the ModCod is
not switched when the enabling threshold threnable is crossed,
but just when an additional hysteresis margin is exceeded.
The threshold for switching to a smaller ModCod is denoted
as thrdown (ModCod) and the size of the hysteresis margin as
(Δthrhyst (N)). The distance between the critical demodula-
tion threshold thrdem (ModCod) and the threshold that trig-
gers a downswitching thrdown (ModCod) is called the safety
margin Δthrsafety (ModCod).
The safety margin(Δthrsafety (ModCod)) can be seen as
an additional security for high frequent oscillations which
cannot be countervailed due to the long satellite propagation
delay. If the signal strength oscillates within this area no in-
crease in BER will occur since the thrdem (ModCod) thresh-
old is not crossed. The values for the safety margin and the
hysteresis margin can be varied and they can also be different
for every ModCod. Wörz et al. [11] presented a calculation
method for all aforementioned parameters which provides a
quasi error free system performance. The calculation of the
parameters in [11] mainly depends on estimated values of
the scintillation standard deviations and a numerically de-
rived function which accounts for the fact that the standard
deviation of the scintillation is also dependent on the inten-
sity of the rain attenuation.
Within the remaining parts of this work the influence of
the size of the safety margins with respect to gain or loss in
channel net efficiency and increase/decrease of BER is inves-
tigated. The term “Wörz-Schweikert safety margins” denotes
the safety margins calculated according to the algorithm pre-
sented in [11] while “zero-safety margin” denotes the fact
that no safety margin is used.
2.4. Investigated environment
Within the examined scenarios, a set of user terminals has
been located in a geographical region close to the city of
Hamburg, Germany (longitude 9.5◦ −10.5◦ E, latitude 52.5◦ −
54◦ N) within the aforementioned channel simulator. The set
comprises 38 different terminal locations whereby the chan-
nel states are sampled with 10 Hz. The investigated duration
is 7200 seconds per simulation run. In order to get statistical
significant results the simulation duration of 7200 seconds
per simulation have been extended to 60 hours. The 60 hours
channel simulation results for the 38 terminals can be seen as
2280 hours of simulated channel states for a single terminal
what in turn means that all results are based on channel in-
formation which corresponds to roughly a quarter of a year.
3. SYSTEM MODELLING
The main idea behind this study is that by reducing the safety
margin in the ModCod switching strategy it is possible to
gain in spectral efficiency, and thus to increase the net data
throughput, at the expenses of an increased BER (and con-
sequently a higher PER). In order to investigate this and to
derive a detailed quantitative estimation of this trade-off, it
5
Signal
threnable (N − 1)
Δthrhyst (N)
thrdown (N)
Δthrsafety (N)
thrdem (N)
SNIR
No switching
here because
of hysteresis
thrdown (N − 1)
Δthrsafety (N − 1)
thrdem (N − 1)
ModCod N − 1 ModCod N
Time
Figure 6: Illustration on the different thresholds.
is important to carefully describe the assumptions on which
the analysis is based, this is what is presented in this sec-
tion. Though the obtained results have a quantitative mean-
ing only considering these assumptions, it is worth stating
that their qualitative relevance has a general importance, as it
will be later explained.
Existing systems compliant with the DVB-S2 standard
can provide the higher-layers protocols with a quasi-error-
free underlaying physical layer (PER = 10−7 ). For this pur-
pose regular 8-bytes CRC (cyclic redundancy check) fields
are used to identify errors in the BBFRAME, which were not
corrected by the coding schemes (LDPC and BCH) at re-
ception. In case an error is detected in a frame, it has to be
considered that the wrong bit(s) cannot be singularly identi-
fied in the frame, so one of the two following choices can be
made:
(1) the whole frame is discarded (this is what is normally
done);
(2) the packets in the frame are passed to the higher proto-
cols with uncorrected failures (this can be done in case
the higher protocols are able to cope with errors).
These cases are very rare when high safety margins are
adopted, and systems are normally dimensioned to avoid
them, but they become more frequent if the system works
closer to the demodulation thresholds (as we defined them
in the previous section), for the reasons already explained.
If a system is dimensioned also to operate in these con-
ditions, it is important to evaluate the statistical properties
and characteristics of these situations, that is, how often they
occur and what failures they bring in comparison to the ca-
pacity gain. In order to do that we performed three levels
of analysis. They are theoretically described in this sections,
whereas the results obtained for each of them are shown and
discussed in the next one.
3.1. Markov model
The first analysis is a comparison of the new approach with
a classical one existing in literature (the already mentioned
Wörz et al. [11]), in terms of ModCod switching statistics.
6 EURASIP Journal on Wireless Communications and Networking
The best way to show the difference between the two ap-
proaches is to model the system according to a Markov chain,
where each state represents the system operating with one
particular ModCod. The chain presents two states for each
ModCod N: the good one, NG , and the bad one, NB ; so the
overall number of states is twice the total amount of allowed
ModCods (56). In the good state the SNR measured at the re-
ceiver is above the demodulating threshold for that ModCod,
and so no failures are expected, in the bad state the system
SNR is below the demodulating threshold for that ModCod,
and so failures may occur with probabilities that are not neg-
ligible.
A similar Markov chain is an excellent model, because it
summarizes very well the properties of a ModCod switch-
ing approach. So once the transition probabilities for one
particular ModCod switching criterion have been calculated
(normally by simulations), the Markov chain can be used
as a basis for all types of analysis without the need of run-
ning again computationally heavy simulations, which might
be very long in order to gather statistically meaningful data.
In this sense the calculated Markov chain (i.e., the ModCod
transition probabilities) can be considered independent from
the simulated channel conditions, only if the simulation is
long enough to represent general channel statistics. On the
other hand, it should be mentioned that the same Markov
chain depends on some parameters which might be charac-
teristic of particular cases, for example, the link budget in
clear sky, and consequently the system availability. So even if
the resulting numbers are only meaningful bearing in mind
these assumptions, the quantitative conclusions which can be
derived have general relevance, and this will be clearer in the
next section.
3.2. Error rate versus capacity trade-off
The second level of analysis describes the details of each state
of the Markov chain. The good states present quasi-error-free
conditions according to the DVB-S2 recommendations, so
PER = 10−7 . Since the BER versus SNR characteristics for
all ModCods are very steep, the BER values increase quite
rapidly when the SNR level goes below the demodulating
threshold. In particular they change of several orders of mag-
nitude within a few tenths of dB, going from BER ≈ 10−10
when SNR is close or bigger than the demodulating thresh-
olds, up to BER ≈ 10−2 when the SNR is just 0.3 dB be-
low the threshold. Each bad state represents a set of different
BERs, the proper BER is selected at each time step according
to the received level of SNR with respect to the demodulat-
ing thresholds. The exact characteristics for the BER-versus-
SNR functions, which were used in the simulations and to
derive the Markov chain parameters, were taken from [12].
In that work, end-to-end performances of the BER versus
the SNR are presented for the DVB-S2 system, the whole
communication chain is modelled and simulated, including
coding, modulation with predistortion techniques, satellite
transponder impairments, downlink, demodulation with the
synchronization, and the final LDPC and BCH decoders. In
the Markov chain, each ModCod might have its own BER
versus SNR characteristic, but to use one single function for
all ModCods already seems an excellent approximation to the
real case, so this is how it was implemented in the simulator.
PERs are derived from these BERs under consideration
of the payload length of each BBFRAME also regarding the
applied ModCod. A BBFRAME is considered as erroneous if
at least one of the payload bits is erroneous. For the rest of
this paper the term PER denotes the BBFRAME packet error
rate. Thanks to this definition of the states of Markov chain,
this model allows to derive the properties of the communi-
cation in terms of PER and BER statistics, and by knowing
the spectral efficiency associated to each ModCod it is easy to
derive an average resulting capacity.
3.3. Error bursts analysis
The third and last level of analysis goes into the details of the
failures introduced with this novel approach. In the previous
section we explained how to derive a measure of the trade-
off between average capacity and average BER (or PER). An
average measure of the BER (or PER) does not seem a very
precise information, since these failures come in bursts. The
errors are mainly due to the ModCod switchings, and they
are mostly introduced by reduced safety margins. So we want
to investigate three main properties: (i) how often the error
bursts arrive (interarrival times statistics), (ii) how long the
bursts last (duration statistics), and (iii) how deep the fades
are (i.e., how high are BER and PER during one error burst).
These three properties can be estimated thanks to the Markov
model, and this analysis produces interesting information,
which will be presented in the next section.
4. RESULTS EVALUATION
4.1. Markov model
A software simulator was developed in order to derive the
Markov model presented in the previous section. Once the
ModCod switching criterion has been specified the software
simulates the evolution over time of the system; from these
simulations we can derive statistics about the permanence in
the different ModCods for each ModCod switching criterion,
this was done by computing transition matrices and solving
them. In the following we present two full transition matrices
for two different ModCod switching criteria.
Simulations equivalent to 3 months of SNR time series
have been carried out, one using Wörz-Schweikert safety
margins, the other one using zero-safety margin bounds with
Wörz-Schweikert hysteresis bounds.
The matrices in Figures 7 and 8 represent the transition
probabilities for those two approaches, where position (i, j)
is the probability in each time step (0.1 second) to move from
state i to state j; the first line and the first column of each
ModCod represent the bad state (iB and jB ), the second one
the good state (iG and jG ). Figures 7 and 8 show the transi-
tion matrices for zero-safety margin and the Wörz-Schweikert
safety margins. The cells marked black indicate that their
content is unequal to zero. In Figure 8, we can see that the
Matteo Berioli et al. 7

1 2 3 4 5 6 7 8 9 12 11 13 14 18 19 20 21
1, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00
1 0, 01 0, 99 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00
0, 00 0, 13 0, 74 0, 13 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00
2 0, 00 0, 00 0, 00 0, 99 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00
0, 00 0, 00 0, 00 0, 14 0, 74 0, 12 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00
3 0, 00 0, 00 0, 00 0, 00 0, 00 0, 99 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00
0, 00 0, 00 0, 00 0, 00 0, 00 0, 14 0, 73 0, 13 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00
4 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 99 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00
0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 14 0, 72 0, 14 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00
5 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 99 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00
0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 13 0, 71 0, 16 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00
6 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 1, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00
0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 13 0, 71 0, 16 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00
7 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 99 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00
0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 12 0, 71 0, 16 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00
8 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 99 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00
0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 13 0, 71 0, 17 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00
9 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 99 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00
0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 01 0, 12 0, 73 0, 15 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00
12 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 1, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00
0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 12 0, 70 0, 18 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00
11 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 99 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00
0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 01 0, 11 0, 71 0, 17 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00
13 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 1, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00
0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 11 0, 70 0, 19 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00
14 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 1, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00
0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 10 0, 66 0, 23 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00
18 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 1, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00
0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 09 0, 65 0, 26 0, 00 0, 00 0, 00 0, 00
19 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 1, 00 0, 00 0, 00 0, 00 0, 00
0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 07 0, 57 0, 36 0, 00 0, 00
20 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 1, 00 0, 00 0, 00
0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 07 0, 57 0, 37
21 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 1, 00

Figure 7: Transition matrix for zero-safety margin.

1 2 3 4 5 6 7 8 9 12 11 13 14 18 19 20 21
1, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00
1 0, 01 0, 99 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00
0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0 ,00 0 ,0 0
2 0, 00 0, 00 0, 00 0, 99 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00
0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0 ,00 0 ,0 0
3 0, 00 0, 00 0, 00 0, 00 0, 00 0, 99 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00
0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0 ,00 0 ,0 0
4 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 1, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00
0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0 ,00 0 ,0 0
5 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 99 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00
0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0 ,00 0 ,0 0
6 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 1, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00
0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0 ,00 0 ,0 0
7 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 99 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00
0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0 ,00 0 ,0 0
8 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 99 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00
0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0 ,00 0 ,0 0
9 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 99 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00
0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0 ,00 0 ,0 0
12 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 1, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00
0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0 ,00 0 ,0 0
11 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 99 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00
0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0 ,00 0 ,0 0
13 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 1, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00
0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0 ,00 0 ,0 0
14 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 1, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00
0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0 ,00 0 ,0 0
18 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 1, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00
0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0 ,00 0 ,0 0
19 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 1, 00 0, 00 0, 00 0, 00 0, 00
0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0 ,00 0 ,0 0
20 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 1, 00 0, 00 0, 00
0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0 ,00 0 ,0 0
21 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 01 0, 00 0, 99

Figure 8: Transition matrix for Wörz-Schweikert safety margin.

more “stable” states are the good states. This makes sense if ity. Transitions only occur between good states and this con-
we look at the SNR time series, because switchings between firms that this approach is designed to work only in good
ModCods are quite spaced in time compared to the time step states.
of 0.1 second. Simulations estimate an average number of 4.5 For the novel approach, the zero-safety margin one, it
ModCod switchings per hour. may be interesting to derive the probability to be in each
Moreover on the diagonal, the probabilities of remaining ModCod (bad or good state). Once the transition matrix
in a bad state are not so high; this is also correct, since when for the Zero-Safety margin is solved [13], we end up with
we are in a bad state, we know that a down-switch should Figure 9 which shows a stacked probability density graph for
occur. The only bad state which has a higher stability is the good and bad states of each ModCod. This is the result of
bad state for ModCod 1B . This results from the fact that when a simulation of an equivalent of 4.5 years of SNR evolution
the SNR goes below the last demodulation threshold, the sys- over time. What we can see is that the most used ModCods
tem cannot switch to a lower ModCod, so it remains in bad are those whose demodulation threshold is just below the
state until the SNR rises again. This is basically an outage SNIR in clear sky conditions. That makes sense because most
where the DVB-S2 receiver is not available; the simulator was of the time we are in clear sky conditions, so we use the high-
designed to give a system availability of 99.96% of the time, est ModCods. We can also notice the high value of the bad
for both approaches. state in ModCod 1B , because of system unavailability. Some
With the Wörz-Schweikert scheme, shown in Figure 8, the ModCods are never used due to overlapping with other ones,
matrix is far more sparse, and no bad states are ever ac- some ModCods achieve a better spectral efficiency requiring
tive, except for ModCod 1B because of system unavailabil- less SNR.
8 EURASIP Journal on Wireless Communications and Networking

100 100
10−1
10−1 10−2
10−3 6 × 10−3
4 × 10−3
10−2 10−4

Error rate
10−5
Probability

10−3 10−6
10−7
10−4 10−8
10−9
10−5 10−10
−3 −2.5 −2 −1.5 −1 −0.5 0 0.5
Multiplying factor on Schweikert-Wörz safety margin
10−6
2 4 6 8 10 12 14 16 18 20 22 24 26 28 Packet error rate
ModCods Bit error rate

Bad states
Good states Figure 10: Packet error rate (PER) versus Safety margin.

Figure 9: State probabilities.

3.5
Max @ − 1.5 = 3.12 b/Hz/s Max @ − 0.4 = 3.01 b/Hz/s
Average spectral efficiency (b/s/Hz)

4.2. Error rate versus capacity trade-off 3

This section presents the main results which are obtained 2.5
when reducing the safety margins, in terms of increase spec-
tral efficiency and increase errors. The starting point is the set 2
of threshold selected by Wörz-Schweikert; this set guarantees 1.5
a quasi-error-free system operation. We try to proportionally
reduce those margins and even to have negative margins, to 1
see how the system performs. The x axis in Figures 10 and 11
0.5
represents factors to be multiplied to the Wörz-Schweikert
set to get the tested thresholds. This means that for multiply- 0
ing factor 1 we have the Wörz-Schweikert set, for the factor −3 −2.5 −2 −1.5 −1 −0.5 0 0.5 1
0, we have the zero-safety margin approach, and for negative Multiplying factor on Schweikert-Wörz safety margin
values of the factors we are testing thresholds which are be- Gross efficiency
low those thresholds recommended by the DVB-S2 standard. Net efficiency (without bit aggregation)
This may seem strange, but it will appear clear how useful Net efficiency (with bit aggregation)
this is to show that there is a trade-off between errors and
increase in capacity. Figure 11: Average spectral efficiency versus safety margin.
Figure 10 shows (as expected) that the PER objective of
10−7 is achieved already before the Wörz-Schweikert bounds.
This is not surprising since the model has been designed to
do so. As expected as well, PER and BER are fast-growing up and −1.5 at the bit level. Corresponding values of PER/BER
to 1 when the safety margin becomes negative. A surprising at these maxima are 6 · 10−3 and 4 · 10−3 . The two curves
fact here is that there are possibilities to achieve the goal PER represent the two ways of operating described in Section 3:
even for margins which are 0.4 times the Wörz-Schweikert bit aggregation is when failures cause BBFRAME discard, no
safety margins. That means that those Wörz-Schweikert mar- bit aggregation means when the frame is passed to the higher
gins may not be the optimum selection. layers with failures. It should be noted that for bit error ag-
Figure 11 shows the core result of this work. A trivial gregation (see Figure 11) the PER (see Figure 10) is the rele-
thing is that the gross capacity (total amount of received bits vant result since in case of a bit error the complete BBFRAME
with failures) is still increasing when we go for lower and is discarded. Without consideration of bit error aggregation,
lower bounds, because of course we are using less and less the BER is the relevant result since erroneous bits within the
robust ModCods that provide better spectral efficiency. The BBFRAME are expected to be corrected by the higher layers.
very interesting point comes with the fact that the net ca- This means that a system which wants to have the indicated
pacity (throughput of correct bits) shows a maximum in the throughput with or without bit aggregation, is operating at
negative part of the scaling factor: −0.4 at the packet level those PER/BER.
Matteo Berioli et al. 9
SNIR at receiver and ModCod selection
100
10−1 0.8
10−2

MPEG-2 packet error rate

10−3 0.7
10−4

Events per hour

0.6
10−5 0.5
10−6
9 10−7 0.4
8 0.3
7 0.2
6 0.1
5 0
1e − 6
4 1e − 5
3 1e − 4

PE
1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 1e − 3

R
×104
Time in samples (0.1 s) 1e − 2

lev
> 0.5
0.5

e
1e − 1 0.4

l
0.2 0.3
0.1 tion (s)
Figure 12: PER with SNR and ModCod selection for zero-safety Fade dura
margin.
Figure 14: State probabilities.

Probability density function for interarrival times

100
We see that in 50% of cases, the time between two er-
ror bursts is in the range of 0–100 seconds. This distribution
comes from the fact that during a rain fade, ModCods are
10−1 switched down one by one, and as we saw on Figure 12, error
peaks often occur at every down-switch. The question is now
Probability

what is the duration/severity of these peaks?

Figure 14 shows the number of fade events per hour us-
ing zero-safety margins, sorted by their duration and PER
10−2 strength. A sequence of samples is considered as one fade
event if the associated PER is exceeding a given level. For
each PER level, Figure 14 shows the number of fades per hour
which exceed this PER level. For this graph, we have 6 dif-
10−3
ferent PER levels, and the fade events are distributed among
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000 5500 their duration. We can see that the shorter fades are the ones
Interarrival time of error bursts (event where PER > 10−6 ) (s) that occur most of the time. This comes from the fact that for
a normal process like scintillation, the probability of having a
Figure 13: Interarrival times distribution. fade is decreasing exponentially with its duration. There are
peaks for each PER level at 0.5 second, this is due to the fact
that independently from how long the fade would be, in the
worst case the system can switch to a lower ModCod within
If a system can cope with these error rates, then it may be
half a second (twice the GEO propagation delay), which is
interesting to design it with lower safety margins than those
the time needed to signal to the gateway the fading situation
in the Wörz-Schweikert strategy, in order to gain throughput.
and to receive a new transmission with a new ModCod. So
in theory fade should not exceed 500 milliseconds, but the
4.3. Error bursts analysis last bin of this bar plot shows that even if they are rare, fades
exceeding 0.5 second do exist. There are two explanations for
To deeper investigate the quality of the transmission in case that. First, if we are in the highest ModCod of a couple of very
we reduce the safety margins, we have to look at the dis- close ModCods (in terms of demodulation threshold) and we
tribution in time of the error bursts. Figure 12 shows an enter a strong rain fade, with a steep decreasing SNR, it can
example of simulated SNR time series with corresponding happen that the SNR crosses the demodulation threshold of
PER for zero safety margin. In contrast to Figure 10 which the lowest ModCod before the system has switched down.
shows the averaged error PERs and BERs, here we investigate This results in a bad state to bad state transition, and we can
the distribution of the interarrival times between two PER see some of these cases in the transition matrix (12B to 9B or
peaks (without averaging), considering a detection threshold 13B to 11B , e.g.). A second explanation is the following, non-
of PER = 10−6 . The simulation that has led to Figure 13 has negligible contributions to this behavior are the outages due
been worked out on 4.5 years of simulated SNR, and it was to system nonavailability, that is the fades that occur in the
conducted with the zero-safety margins approach. lowest ModCod.
10 EURASIP Journal on Wireless Communications and Networking

5. CONCLUSIONS [12] E. Casini, R. De Gaudenzi, and A. Ginesi, “DVB-S2 modem

algorithms design and performance over typical satellite chan-
The possibility to have a quasi-error-free transmission chan- nels,” International Journal of Satellite Communications and
nel in DVB-S2 systems is not always an optimal solution in Networking, vol. 22, no. 3, pp. 281–318, 2004.
case the higher-layer protocols do not require such high per- [13] M. F. Neuts, Matrix-Geometric Solutions in Stochastic Models:
formance. In this case the lower layers can provide a trans- An Algorithmic Approach, Dover, Mineola, NY, USA, 1981.
mission with some resilient errors, and exploit more the
spectrum to gain in throughput. The error-capacity trade-off
can be tuned, according to the requirements of each partic-
ular system, with the adjustment of the ModCod safety mar-
gins. The paper presents the gain in spectral efficiency, which
is obtained with this method, and the statistical characteris-
tics of the “artificially” introduced error bursts, in terms of
interarrival, duration and depth (PER). One additional in-
teresting side-outcome of this work is the development of
Markov chain to model the ModCod transitions and the fail-
ure occurrence in a DVB-S2 system.

ACKNOWLEDGMENTS
This work was partly supported by EC funds SatNEx under
the FP6 IST Programme, Grant number: 507052. This work
was supported by the European Satellite Network of Excel-
lence (SatNEx).

REFERENCES
[1] ETSI EN 302 307 V1.1.2, “Digital Video Broadcasting (DVB);
second generation framing structure, channel coding and
modulation systems for broadcasting, interactive services,
news gathering and other broadband satellite applications,”
June 2006.
[2] ETSI EN 301 790 V1.4.1, “Digital Video Broadcasting (DVB);
interaction channel for satellite distribution systems,” April
2005.
[3] G. Fairhurst, M. Berioli, and G. Renker, “Cross-layer control
of adaptive coding and modulation for satellite Internet multi-
media,” International Journal of Satellite Communications and
Networking, vol. 24, no. 6, pp. 471–491, 2006.
[4] ETSI TS 126 102, “AMR Speech Codec,” 2001.
[5] ISO/IEC 14496-2, “Coding of audio-visual objects (MPEG-
4)—part 2: visual,” 2004.
[6] ETSI TR 126 975, “Performance Characterisation of the Adap-
tive Multi-Rate (AMR) Speech Codec,” 2004.
[7] L.-A. Larzon, M. Degermark, S. Pink, L.-E. Jonsson, and G.
Fairhurst, “The Lightweight User Datagram Protocol (UDP-
Lite),” IETF, RFC 3828, 2004.
[8] S. Datta-Barua, P. H. Doherty, S. H. Delay, T. Dehel, and J.
A. Klobuchar, “Ionospheric scintillation effects on single and
dual frequency GPS positioning,” in Proceedings of the 16th In-
ternational Technical Meeting of the Satellite Division of the In-
stitute of Navigation (ION GPS/GNSS ’03), pp. 336–346, Port-
land, Ore, USA, September 2003.
[9] V. I. Tatarskii, Wave Propagation in a Turbulent Medium,
McGraw-Hill, New York, NY, USA, 1961.
[10] E. Matricciani, M. Mauri, and C. Riva, “Relationship between
scintillation and rain attenuation at 19.77 GHz,” Radio Science,
vol. 31, no. 2, pp. 273–280, 1996.
[11] T. Wörz, R. Schweikert, A. Jahn, and R. Rinaldo, “Physical
layer efficiency of satellite DVB using fade mitigation tech-
niques,” in Proceedings of the International Communication
Satellite Systems Conference (ICSSC ’05), Rome, Italy, Septem-
ber 2005.
Hindawi Publishing Corporation
EURASIP Journal on Wireless Communications and Networking
Volume 2007, Article ID 62310, 12 pages
doi:10.1155/2007/62310

Research Article
Frequency Estimation in Iterative Interference Cancellation
Applied to Multibeam Satellite Systems

J. P. Millerioux,1, 2, 3, 4 M. L. Boucheret,2 C. Bazile,3 and A. Ducasse5

1 TéSA, 14-16 Port Saint-Etienne, 31000 Toulouse, France
2 Institutde Recherche en Informatique de Toulouse, Ecole Nationale Supérieure d’Electrotechnique, d’Electronique,
d’Informatique, d’Hydraulique et des Télécommunications, 2 Rue Camichel, BP 7122, 31071 Toulouse, France
3 Centre National d’Etudes Spatiales, 18 Avenue E. Belin, 31401 Toulouse Cedex 4, France
4 Ecole Nationale Supérieure des Télécommunications, 46 Rue Barrault, 75634 Paris Cedex 13, France
5 Alcatel Alenia Space, 26 Avenue J.F. Champollion, BP 1187, 31037 Toulouse, France

Received 31 August 2006; Revised 26 February 2007; Accepted 13 May 2007

Recommended by Alessandro Vanelli-Coralli

This paper deals with interference cancellation techniques to mitigate cochannel interference on the reverse link of multibeam
satellite communication systems. The considered system takes as a starting point the DVB-RCS standard with the use of convolu-
tional coding. The considered algorithm consists of an iterative parallel interference cancellation scheme which includes estima-
tion of beamforming coefficients. This algorithm is first derived in the case of a symbol asynchronous channel with time-invariant
carrier phases. The aim of this article is then to study possible extensions of this algorithm to the case of frequency offsets af-
fecting user terminals. The two main approaches evaluated and discussed here are based on (1) the use of block processing for
estimation of beamforming coefficients in order to follow carrier phase variations and (2) the use of single-user frequency offset
estimations.

Copyright © 2007 J. P. Millerioux et al. This is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

1. INTRODUCTION Other approaches have been proposed in the literature

Multiuser detection appears as a promising way to mitigate
cochannel interference (CCI) on the reverse link of multi-
beam satellite systems. It can allow considering more capac-
ity efficient frequency reuse strategies than classical systems
(in which cochannel interference is assimilated to additive
noise). However, channel estimation appears to be a criti-
cal point when performed before multiuser processing. This
paper proposes a multiuser detection scheme coupled with
channel reestimations.
This study is the continuation of the work reported in
[1]. The considered system is inspired by the DVB-RCS stan-
dard [2], with the use of convolutional coding. The algorithm
is derived for a symbol-asynchronous time-invariant chan-
nel [1]. It basically consists of a parallel interference cancel-
lation (PIC) scheme which uses hard decisions provided by
single user Viterbi decoders, and includes channel reestima-
tion. The aim of this paper is to propose results on possible
adaptations of this algorithm to the more realistic case of fre-
quency offsets affecting user terminals.
with similar contexts. In [3], an iterative decoding scheme
is proposed with a very simplified channel model and with-
out considerations on channel estimation issues. In [4, 5],
MMSE and noniterative MMSE-SIC schemes are evaluated
in a realistic context and the problem of channel estima-
tion before multiuser processing is addressed based on pi-
lot symbols. In this paper, we consider a joint multiuser
detection and channel estimation approach, which can no-
tably allow reducing the required number of pilot symbols,
and consequently lead to more spectrally efficient transmis-
sions, in particular for a burst access. Notice however that
the algorithm considered here is suboptimal. Some poten-
tially optimal algorithms have been studied in [1]. However,
they have appeared much more complex than the one con-
sidered here, and have shown a gain in performance pos-
sibly very limited, and highly dependant on the antenna
implementation.
The paper is organized as follows: the system model
and assumptions are described in Section 2, Section 3 intro-
duces the algorithm on a time-invariant channel, Section 4 is
2 EURASIP Journal on Wireless Communications and Networking

As regards to the waveform, the information bits are con-

Pilot
Information QPSK symbols volutionally encoded, and the coded bits are then mapped
Encoder Πk dk [n] onto QPSK symbols which are interleaved differently on each
bits user k mapping insertion
Tk beam. A burst of N symbols dk [n] is composed of these in-
terleaved symbols in which pilot symbols are inserted. We
(a) model the signals xk (t) as
−1
N
x1 (t)  
xk (t) = ρk e jϕk (t) dk [n]s t − nT − τk , (2)
ρk e jϕk (t) nk (t) n=0

dk [n] s(t − τk ) yk (t)

xk (t) H where T, s(t), ρk , ϕk (t), τk , denote, respectively, the symbol
duration, the normalized emitter filter response (square root
raised cosine with rolloff equal to 0.35 [2]), the amplitude of
the kth signal, its (possibly time-varying) carrier phase, and
xK (t) its time delay. The whole transmitter and channel model is
summarized in Figure 1. Notice that a single frequency refer-
(b)
ence is assumed on-board the satellite.
We define the signal-to-noise ratio (SNR) for the kth sig-
Figure 1: Transmitter and channel model.
nal as

Es  2
 = ρk .
dedicated to the study of possible adaptations with frequency  (3)
N0 k σ 2
offsets, and we draw conclusions in Section 5.
Assuming an equal SNR for all users, the carrier to interfer-
2. SYSTEM MODEL AND ASSUMPTIONS ence ratio for the kth signal can be simply defined as
  −1
2.1. Model C   2
 = hk,l  . (4)

I k
The considered context is the reverse link of a fixed-satellite l =/ k
service with a regenerative geostationary satellite, a multi-
beam coverage with a regular frequency reuse pattern [6],
and an MF-TDMA access [2]. A “slot synchronous” system 2.2. Assumptions
is assumed. Multiuser detection is performed onboard the
satellite, after frequency demultiplexing. We choose here to The algorithm is derived under the following assumptions.
work on a fictitious interference configuration characterized
(i) We assume a perfect single-user frame synchronisation
by carrier to interference ratios C/I. A more detailed presen-
and timing recovery (i.e., for the kth signal on the kth
tation can be found in [1] or [7].
beam).
We consider in the following a frequency/time slot in
(ii) The matrix H is assumed time invariant on a burst du-
the MF-TDMA frame. Notations are relative to complex en-
ration, and unknown at the receiver.
velops. ·∗ , ·T , ·H , E(·), and · ∗ · denote, respectively, the
(iii) Significant interferers are only located in adjacent
conjugate, transpose, conjugate transpose, expected value,
cochannel cells: due to the regular reuse pattern, there
and convolution operators. Consider K uplink signals asso-
are at most 6 significant interferers on a beam [6].
ciated to K different cochannel cells. Under the narrowband
assumption [8], we get Let us recall that the algorithm considered in the follow-
ing is suboptimal (see Section 1 and [1]): it only performs
y(t) = Hx(t) + n(t), (1) interference cancellation for the kth signal at the output of
the kth beam.
where x(t) = [x1 (t) · · · xK (t)]T is the K × 1 vector of re-
ceived signals, y(t) = [y1 (t) · · · yK (t)]T is the K × 1 vec-
3. ALGORITHM DESCRIPTION ON A TIME
tor of signals at the beamformer outputs, H is the K × K
INVARIANT CHANNEL
beamforming matrix (i.e., the product of the matrix of steer-
ing vectors by the matrix of beamformer coefficients), and 3.1. Synchronous case
n(t) = [n1 (t) · · · nK (t)]T is the vector of additive noises.
Without loss of generality, we consider that the matrix H To simplify the presentation, we first consider a symbol-
has its diagonal coefficients equal to 1. Additive noises are synchronous time-invariant channel, that is, τk = 0 and
additive white Gaussian noises (AWGN) with the same vari- ϕk (t) = ϕk for all k. After optimal sampling, we can then
ance σ 2 , and are characterized by a spatial covariance matrix consider the “one-shot” approach with
Rn = E(n(t)n(t)H ) which depends on the antenna imple-
mentation [1]. y[n] = Gd[n] + n[n], (5)
J. P. Millerioux et al. 3

Interference cancellation
Initial phase Estimation Interference
y1 [n] Decoding
recovery of g1,. cancellation The interference cancellation block output at the mth itera-
tion (or the decoding block input at the (m + 1)th iteration)
is for the nth symbol of the kth user
dl
(m)
[n]
From beam l, for l  
interfering on beam k ∗  (m)
(m)
yk(m+1) [n] = g
k,k
(m)
yk [n] − g
k,l dl [n] . (8)
(m)
yk [n]
l =/ k
Initial phase Estimation Interference (m+1)
yk [n]
recovery
Decoding
of gk,. cancellation
yk [n] In the case of perfect channel estimation and interfering
symbol decisions, we get
dk [n]
(m) To beam l, for k  2
interfering on beam l yk(m+1) [n] = gk,k  dk [n] + gk,k
∗
nk [n], (9)
interference is entirely removed, and the carrier phase is per-
fectly compensated.
Initial phase Estimation Interference
yK [n] Decoding
recovery of gK,. cancellation
Decoding

Figure 2: Block diagram of the receiver (synchronous case). Decoding is performed by the Viterbi algorithm, by assimi-
lating the residual interference plus noise after deinterleaving
at the decoder input to AWGN.
where
Initialization
 T     
G= g1T · · · gKT = gk,l = H diag ρk exp jϕk , For the kth user, an initial carrier phase is estimated from
 T pilot symbols on the kth beam. After phase compensation,
d[n] = d1 [n] · · · dK [n] ,
the signal received on the kth beam is sent to the decoding
 T block to initialize the iterative process.
y[n] = y1 [n] · · · yK [n] with yk [n] = yk (t) ∗ s(−t)|t=nT ,
 T
n[n] = n1 [n] · · · nK [n] with nk [n] = nk (t) ∗ s(−t)|t=nT , 3.2. Asynchronous case
 
E n[k]n[l] = δ(k − l)Rn . We now consider a symbol-asynchronous time-invariant
(6) channel, that is, τk =/ τl for k =/ l, and ϕk (t) = ϕk for all k.
We introduce
A synoptic of the receiver is given in Figure 2, where inter- −1
N
leaving and deinterleaving operations are omitted for sim-  
uk (t) = dk [n]s t − nT − τk ,
plicity. All operations are performed in parallel on the dif- n=0
ferent beams, with exchange of information from one to an- (10)
−1
N
other. The main steps are described in the following. For any  
parameter c, c
(m) denotes an estimate or a decision on c at the u
(m)
k (t) = d
k(m) [n]s t − nT − τk ,
n=0
mth iteration.
and vectors u(t) = [u1 (t) · · · uK (t)]T and u

(m) (t) =
(m) (m)
Channel estimation [u
1 (t) · · · u
K (t)] .
T

We get
The channel estimation on the kth beam is processed by
a least-square estimator using currently estimated symbols y(t) = Gu(t) + n(t), (11)
(and including pilot symbols). At the mth iteration, we get
for the kth beam where G is defined in Section 3.1. We refer to u
(m)
k (t) as the
estimated kth signal at the mth iteration.
N −1 N −1 −1
  The algorithm on the asynchronous channel is then very
g
k(m) =
(m) [n]H
yk [n]d
(m) [n]d
d
(m) [n]H . similar to the one on the synchronous channel. For the kth
n=0 n=0 beam, at the mth iteration:
(7)
(i) channel estimation is processed by a least square ap-
We only use for estimation (and consequently for interfer- proach using the estimated signals at the matched fil-
ence cancellation in (8)) estimated symbols of the useful sig- ter output u
(m) (t) ∗ s(−t) and yk (t) ∗ s(−t), syn-
nal and of adjacent interfering ones (see Section 2.2. assump- chronously sampled, with 2 samples per symbol (sam-
tion (iii)), which is not specified in the equations for the sake ples of u
(m) [n] and sam-

(m) (t) ∗ s(−t) corresponds to d
of simplicity. ples of yk (t) ∗ s(−t) corresponds to yk [n] in (7));
4 EURASIP Journal on Wireless Communications and Networking

Number of
1 2 3 Cell number C/I [dB]
interferers
1, 3 3 5
4 5 6 7
2 4 4
4, 7 3 5
8 9 10 5, 6 6 2
8, 10 5 3
11 12 13 14
9 6 2
11, 14 2 6
12, 13 4 4
(b)
(a)

Figure 3: Description of the studied configuration.

(ii) interference cancellation is processed at 1 sample per reaches the Cramer-Rao bound (CRB). This bound is more
symbol, at optimal sampling instants. precisely the phase single-user modified CRB [9], given with
our notations by
More details on the implementation can be found in [1].
3.3. Simulation results
We use for the evaluation the fictitious configuration de-
scribed in Figure 3 (which is interference configuration 2 in
[1]). We consider 14 cochannel beams. The 14 users have
an equal SNR. For each cell, assumption (iii) of Section 2.2
is perfectly respected, and interference is equally distributed
among the interfering cells: for example we have for cell 1
h1,1 = 1, h1,2 = h1,4 = h1,5 = (3 · C/I |1 )−1/2 , and other coef-
ficients of the first row of H are set to zero. We consider the
following simulation parameters.
(i) Rate 1/2 nonrecursive nonsystematic convolutional
code with constraint length 7 and generators (133,
171) in octal.
(ii) Packets of 53 information bytes (ATM cell), or 430 in-
formation symbols (with closed trellis).
(iii) 32 pilot symbols, leading finally to N = 462 transmit-
ted symbols in a burst.
Users timings τk are independent and uniformly distributed
on [0, T]. Carrier phases ϕk are independent and uniformly
distributed on [0, 2π]. Additive noises are uncorrelated. New
random interleavers and training sequences are generated at
each burst.
We consider a target bit error rate (BER) equal to 2 · 10−4 ,
which is reached on AWGN channel with perfect synchroni-
sation for Eb /N0 equal to 3.2 dB. Some results for cells 5 and
6, which are symmetric, are given in Figure 4. The algorithm
exhibits a degradation with respect to single-user reference
of 0.15 dB after 3 iterations. At first iterations, the modulus
estimate of g5,9 and g6,9 (which are symmetric) is widely bi-
ased: it is underestimated due to imperfect symbol decisions.
As the algorithm converges, this bias is removed. In the same
way, the unbiased phase estimate of g5,9 and g6,9 shows an
error standard deviation decreasing with iterations, until it
 −1
   1  
hk,l 2 Es

CRB Arg gk,l = Rd2 . (12)
2N N0
Notice that these simulation results and all the following ones
correspond to at least 20 packet errors and 200 binary errors
for each user. Consider as an example the results at iteration 3
for Eb /N0 = 2.5 dB, our evaluation of confidence intervals at
95% leads to [4.8, 5.9] · 10−3 for the BER of cell 5, [1.2, 12.1] ·
10−3 for the modulus bias of coefficient g5,1 , and [4.61, 4.89]◦
for the phase error standard deviation of coefficient g5,1 .
4. EXTENSION TO THE CASE OF
FREQUENCY OFFSETS
In geostationary systems, frequency offsets between the emit-
ter and the receiver are mainly due to frequency instabilities
of local oscillators. Considering the use of the Ka-band with
low-cost user terminals, they are inevitable. In order to help
the receiver to recover these frequency offsets, synchronisa-
tion bursts, which are periodically transmitted, are defined
in the DVB-RCS standard. However, it always remains resid-
ual frequency offsets on the traffic bursts. In case of short
bursts and low SNR, frequency and phase recovery become
a challenging task, especially with a reduced number of pilot
symbols.
In the following, we study possibilities of adaptation of
the interference cancellation algorithm to the case of fre-
quency deviations affecting user terminals. We first evaluate
the algorithm sensitivity to frequency offsets in Section 4.1.
We find that it is only suited to very low frequency offsets. We
then evaluate in Section 4.2 the use of block processing for
estimation of beamforming coefficients in order to cope with
higher frequency offsets. As this approach is shown to lead
to possible significant degradations, we finally propose and
J. P. Millerioux et al. 5

BER (cells 5 and 6)

100

10−1

10−2
BER
10−3

10−4

10−5
0 0.5 1 1.5 2 2.5 3 3.5 4
Eb /N0 (dB)

No MUD PIC 3
PIC 1 Reference
PIC 2
(a)

Modulus estimate of g5,9 and g6,9 Phase estimate of g5,9 and g6,9
12

0.4
Error standard deviation (◦ )
10
Normalized bias ()

0.3
8
0.2

0.1 6

0
4
−0.1
2 2.5 3 3.5 4 2 2.5 3 3.5 4
Eb /N0 (dB) Eb /N0 (dB)

PIC 1 PIC 1 PIC 3

PIC 2 PIC 2 CRB
PIC 3
(b) (c)

Figure 4: Results with time-invariant phases.

evaluate in Section 4.3 different schemes based on a single-
user frequency estimator.
Notice the following:
(i) we possibly consider the use of pilot symbols dis-
tributed within the burst (which is not possible while
strictly following the DVB-RCS standard);
(ii) all numerical values of frequency offsets are given for
a burst of 462 symbols (430 information symbols and
32 pilot symbols).
4.1. Algorithm sensitivity to reduced frequency offsets
We evaluate in this section the algorithm sensitivity to re-
duced frequency offsets. As a worst case (which is the clas-
sical approach for single-user phase recovery) is difficult to
define in a multiuser context, we choose here to evaluate a
mean case. We model carrier phases ϕk (t) as
ϕk (t) = ϕk + Δ fk t, (13)
for all k, where the ϕk are independent and uniformly dis-
tributed on [0, 2π], and the Δ fk T follow independent zero-
mean Gaussian distributions with standard deviation σΔ f T .
No change is performed on the algorithm, which assumes
time-invariant phases, but pilot symbols are set in the mid-
dle of the bursts (to avoid too biased initial phase estimates).
Other simulation parameters are those of Section 3.3.
Some results in term of degradation with respect to
single-user reference to reach the target BER are shown in
Figure 5. Notice that the BER is independent of the sym-
bol locations in the burst due to the use of interleavers. The
algorithm appears maintainable with σΔ f T = 10−4 , but the
degradations with σΔ f T = 2 · 10−4 are very large.
6 EURASIP Journal on Wireless Communications and Networking

1.5 4.2. Approach with reduced estimation windows for

channel estimation

In order to cope with higher frequency offsets, we use in this

Degradation (dB)

1
section a classical block processing: the channel is no more
considered invariant on the whole burst, but is considered
invariant on windows of reduced length. The algorithm is
modified in this way: channel estimation (7), which includes
0.5 carrier phase estimations, is performed on reduced windows.
Interference cancellation and phase compensation (8) is then
performed on each window using the corresponding esti-
mated coefficients gk,l .
0 Channel estimation sensitivity to frequency offsets de-
0 1 1.5 1.75
creases when the length of estimation windows decreases, be-
Standard deviation of 104 ·Δ f ·T cause the constellation rotations on a window are reduced.
Single user PIC 2 cells 5 and 6 However, sensitivity to additive noise increases when the
PIC 2 cells 4 and 7 PIC 3 cells 5 and 6 length of estimation windows decreases, because noise is av-
PIC 3 cells 4 and 7 eraged on shorter windows. The optimal length of estimation
windows then results from a tradeoff between frequency off-
Figure 5: Degradation with frequency offsets. sets and noise.
We evaluate in this section the effect of reduced estima-
tion windows without frequency offsets. Pilot symbols for
initialization are uniformly distributed on the burst. Some
1.5 results in term of degradation are shown in Figure 6. The
degradation increases when the length of windows decreases.
This is partially due to the fact that CRB for estimation of gk,l
increase while the length of windows decreases, leading to a
less-efficient interference cancellation and phase compensa-
Degradation (dB)

1
0.5
0 cells 4 and 7, as the least square estimation is performed on
32 64 128 256
Length of windows for estimation (symbol)
PIC 2 cells 4 and 7 PIC 2 cells 5 and 6
PIC 3 cells 4 and 7 PIC 3 cells 5 and 6
Figure 6: Degradation with reduced estimation windows.
By comparing the degradations in single-user and mul-
tiuser cases, we can see that they are similar for σΔ f T = 10−4
and for σΔ f T = 0 (i.e., without frequency offsets). We can
conclude that the degradation in the multiuser case with
σΔ f T = 10−4 is mainly due to imperfect user phase recovery.
Beyond σΔ f T = 10−4 , it can be observed that the degradation
in the multiuser case increases more quickly than the degra-
dation in the single-user case: interference cancellation effi-
ciency is limited. The considered algorithm is consequently
limited to about σΔ f T = 10−4 for a burst length equal to 462
symbols.
tion in (8). However, the degradation is much more impor-
tant for cells 5 and 6 than for cells 4 and 7, whereas the CRB
for channel estimation are equal in both cases (as we have
|g5,2 | = |g5,6 | = |g5,9 | = |g5,8 | = |g5,4 | = |g5,1 | = |g4,1 | =
|g4,5 | = |g4,8 |). In fact, it can be seen in Figure 7 that similarly
to single-user phase estimation, our channel estimator takes
down from the CRB with short estimation windows and low
SNR. It appears much more critical for cells 5 and 6 than for
462 7 (6 + 1) coefficients in the first case, and only 4 (3 + 1) in
the second case. This effect also appears for longer channel
estimation windows, but it is less obvious to see it.
Notice that in order to optimize the length of windows
for a given σΔ f T , we would consequently have to consider dif-
ferent lengths of windows for the different cells: the optimal
length would be shorter for cells 4 and 7 than for cells 5 and
6.
The main conclusion is that the use of reduced estima-
tion windows to cope with higher frequency deviations can
lead to a significant loss (let us recall that evaluations have
been performed in this section without frequency offsets),
particularly for cells with a high number of interferers.
4.3. Approach with single-user frequency estimations
As the previous approach does not appear sufficient to cope
with higher frequency offsets without a significant degrada-
tion, we study in this section another approach. It is based on
the use of single-user frequency estimations.
J. P. Millerioux et al. 7

Coefficients g4,5 and g7,6 Coefficients g5,6 and g6,5

45 45

40 40
Phase error standard deviation (◦ )

Phase error standard deviation (◦ )

35 35

30 30

25 25

20 20

15 15

10 10

5 5
2 2.5 3 3.5 4 4.5 5 2 2.5 3 3.5 4 4.5 5
Eb /N0 (dB) Eb /N0 (dB)

PIC 2, 32 symbols BCR, 64 symbols PIC 2, 32 symbols BCR, 64 symbols

PIC 3, 32 symbols
BCR, 32 symbols
PIC 2, 64 symbols
PIC 3, 64 symbols
(a)
PIC 2, 128 symbols PIC 3, 32 symbols
PIC 3, 128 symbols BCR, 32 symbols
BCR, 128 symbols PIC 2, 64 symbols
PIC 3, 64 symbols
(b)
PIC 2, 128 symbols
PIC 3, 128 symbols
BCR, 128 symbols

Figure 7: Channel estimation errors for different coefficients and lengths of window.

Initial PA Reduced Case a

DD frequency
Case frequency estimation
reestimations Case b
estimations windows for gk
a y n n Case c

b y y n
Windows for channel estimation
c up to IT n n y
Pilot symbols
c beyond IT — y n Information symbols
(a) (b)

Figure 8: Approach with frequency estimations: (a) operations performed, (b) distributions of pilot symbols.

4.3.1. Principle accuracy is limited due to the very low signal-to-interference-

plus-noise ratio (unless using a very high number of pilot
If a frequency estimate Δ f
k for the kth signal is available, it symbols, which decreases the spectral efficiency). Another
way is to use symbol decisions for frequency estimation if
can be included in the estimated kth signal: u
(m) k (t) ∗ s(−t) it is possible to obtain sufficiently reliable symbol decisions.
consequently becomes (u
k (t) ∗ s(−t)) exp( j2πΔ f
k t) in (7).
(m)
Many different receiver architectures can be derived. Three
Since the constellation rotations on the burst for yk (t) ∗ s(−t) examples of architectures are described and evaluated in the
and (u
(m)

k (t) ∗ s(−t)) exp( j2πΔ fk t) are potentially very close following sections.
(ideally identical if Δ f
k = Δ fk ), it is then possible to keep
large estimation windows to perform estimation in (7): us- 4.3.2. Architectures with single user
ing the whole burst allows obtaining the minimum degra- frequency estimations
dation. Clearly, this approach requires “accurate” single-user
frequency estimations, which become the hard task. Two modes are considered for single-user frequency esti-
A first possibility is to use initial frequency estimations mation: the pilot aided mode (PA), based on pilot sym-
before interference cancellation. In this case, the estimation bols, and the decision directed mode (DD), based on symbol
8 EURASIP Journal on Wireless Communications and Networking

BER (cells 5 and 6) Frequency estimate (cells 5 and 6)

10−2

Error standard deviation ()

10−3 10−4
BER

10−4

10−5
2 2.5 3 3.5 4 2 2.5 3 3.5 4
Eb /N0 (dB) Eb /N0 (dB)
No MUD PIC 3 No MUD
PIC 1 Reference
PIC 2
(a) (b)

Modulus estimates of g5,9 and g6,9 Phase estimates of g5,9 and g6,9
12

0.4
Error standard deviation (◦ )

10
Normalized bias ()

0.3
8
0.2

0.1 6

0
4
−0.1
2 2.5 3 3.5 4 2 2.5 3 3.5 4
Eb /N0 (dB) Eb /N0 (dB)

PIC 1 PIC 1 PIC 3

PIC 2 PIC 2 CRB
PIC 3
(c) (d)

Figure 9: Results with frequency estimations: σΔ f T = 2 · 10−4 , case a.

decisions. For the PA mode, pilot symbols are distributed The CRB considered for frequency estimation in DD
within the burst into 3 blocks (see Figure 8(b), cases a and mode is the single-user frequency modified CRB [9], given
b). We follow the approach of [10]. First, a mean phase by
is computed on each block of pilot symbols. Then, a least
square estimation based on these mean phases is used to  −1
estimate the frequency. For the DD mode, the principle   3 Es
CRB Δ fk T = . (14)
is the same: the burst is divided into adjacent blocks, on 2π 2 N 3 N0
which mean phases are computed using symbol decisions.
For the DD mode, frequency estimations are performed For PA frequency estimation, the CRB is different from (14)
after interference cancellation, that is, Δ f
k(m) are used to with N replaced by the number of pilot symbols (because
obtain g
k(m+1) . pilot symbols are not consecutive).
J. P. Millerioux et al. 9

BER (cells 5 and 6) Frequency estimate (cells 5 and 6)

10−2

Error standard deviation ()

10−3 10−4
BER

10−4

10−5
2 2.5 3 3.5 4 2 2.5 3 3.5 4
Eb /N0 (dB) Eb /N0 (dB)
No MUD PIC 3 No MUD PIC 2
PIC 1 Reference PIC 1 CRB
PIC 2
(a) (b)

Modulus estimates of g5,9 and g6,9 Phase estimates of g5,9 and g6,9
12

0.4
Error standard deviation (◦ )

10
Normalized bias ()

0.3
8
0.2

0.1 6

0
4
−0.1
2 2.5 3 3.5 4 2 2.5 3 3.5 4
Eb /N0 (dB) Eb /N0 (dB)

PIC 1 PIC 1 PIC 3

PIC 2 PIC 2 CRB
PIC 3
(c) (d)

Figure 10: Results with frequency estimations: σΔ f T = 2 · 10−4 , case b.

The following three cases of receiver architecture are eval- Case c

uated.
No initial frequency estimation is performed:
Case a
(i) for iterations up to IT: no frequency estimation is per-
PA initial frequency estimations are performed, no frequency formed, the estimation window for the gk is 154 sym-
reestimation is performed, the estimation window for the gk bols for all cells (see Figure 8(b));
is the whole burst. (ii) for iterations beyond IT: frequencies are reestimated
in DD mode, the estimation window for the gk is the
Case b whole burst.
PA initial frequency estimations are performed, frequencies The operations performed are summarized in Figure 8(a). In
are reestimated in DD mode at each iteration, the estimation all cases, we use 32 pilot symbols. Distributions of pilot sym-
window for the gk is the whole burst. bols are shown in Figure 8(b).
10 EURASIP Journal on Wireless Communications and Networking

BER (cells 5 and 6) Frequency estimate (cells 5 and 6)

10−2

Error standard deviation ()

10−3 10−4
BER

10−4

10−5
2 2.5 3 3.5 4 2 2.5 3 3.5 4
Eb /N0 (dB) Eb /N0 (dB)
No MUD PIC 3 PIC 1
PIC 1 PIC 4 PIC 2
PIC 2 Reference CRB
(a) (b)

Modulus estimates of g5,9 and g6,9 Phase estimates of g5,9 and g6,9
12

0.4
Error standard deviation (◦ )

10
Normalized bias ()

0.3
8
0.2

0.1 6

0
4
−0.1
2 2.5 3 3.5 4 2 2.5 3 3.5 4
Eb /N0 (dB) Eb /N0 (dB)

PIC 3 PIC 3
PIC 4 PIC 4
CRB
(c) (d)

Figure 11: Results with frequency estimations: σΔ f T = 2 · 10−4 , case c.

4.3.3. Results with σΔ f T = 2 · 10−4
We first consider in this section a target σΔ f T equal to 2 · 10−4 .
Some results are given in Figures 9, 10, and 11 (with
IT = 2) for cells 5 and 6.
In case a (Figure 9), after initial frequency estima-
tion, the frequency error standard deviation is about 10−4 .
Iterative interference cancellation works, but leads to a
degradation in term of BER, as in Section 4.1. The er-
ror standard deviation on the phase of g5,9 and g6,9 is far
from the CRB, clearly because of imperfect frequency esti-
mates.
In case b (Figure 10), DD frequency reestimations allow
to get a frequency error standard deviation close to the CRB.
Hence, the phase estimate error standard deviation of g5,9
and g6,9 is much closer to the CRB than in case a. The BER
degradation is the same as that in the case without frequency
offsets in Section 3.3.
In case c (Figure 11), interference cancellation is efficient
but converges slower than in cases a and b. Four iterations
are necessary in case c to get the BER reached with three iter-
ations in case b.
With σΔ f T = 2 · 10−4 , the most efficient architecture is
consequently architecture b. However, if architecture c leads
J. P. Millerioux et al. 11

BER (cells 5 and 6) Frequency estimate (cells 5 and 6)

10−2

Error standard deviation ()

10−3 10−4
BER

10−4

10−5
2 2.5 3 3.5 4 2 2.5 3 3.5 4
Eb /N0 (dB) Eb /N0 (dB)
No MUD PIC 4 PIC 3
PIC 1 PIC 5 PIC 4
PIC 2 Reference CRB
PIC 3
(a) (b)

Modulus estimates of g5,9 and g6,9 Phase estimates of g5,9 and g6,9
12

0.4
Error standard deviation (◦ )

10
Normalized bias ()

0.3
8
0.2

0.1 6

0
4
−0.1
2 2.5 3 3.5 4 2 2.5 3 3.5 4
Eb /N0 (dB) Eb /N0 (dB)

PIC 4 PIC 4
PIC 5 PIC 5
CRB
(c) (d)

Figure 12: Results with frequency estimations: σΔ f T = 5 · 10−4 , case c.

to a slower convergence of the algorithm, a significant advan-
tage is that it appears more suited to high-frequency offsets,
as we will see in the following section.
4.3.4. Results with σΔ f T = 5 · 10−4
We now consider a target σΔ f T equal to 5 · 10−4 .
For this range of frequency deviations, it is very difficult
to obtain reliable initial frequency estimates without a huge
number of pilot symbols. On the contrary, architecture c
appears to work. After optimization, we use IT = 3 with
window lengths for gk estimation from 60 to 100 symbols
(depending on the number of interferers of the considered
cell, Section 4.2). Some results are given in Figure 12. For
Eb /N0 equal to 3.2 dB, the block processing approach allows
obtaining a BER equal to about 8 · 10−3 at iteration 3, which is
sufficient to obtain reliable frequency estimates at the follow-
ing iterations. The degradation in terms of BER at iteration 5
is then similar to the case without frequency offsets.
Finally, notice that we have considered average BER along
the paper. Actually, this average BER can hide some complete
12
×10−3 Distribution of erroneous bits
2.5
2 [1] J. P. Millerioux, M. L. Boucheret, C. Bazile, and A. Ducasse,
Probability
1.5
1
0.5
0 [2] Digital Video Broadcasting (DVB), “Interaction channel for
1 106
Number of erroneous bits per packet
Cell 5 at 4σΔ f T , Eb /N0 = 3.5 dB
424 information bits
Figure 13: Distribution of erroneous bits: σΔ f T = 5 · 10−4 , itera-
tion 5.
failures in convergence of the algorithm on some bursts, lead-
ing to a BER on these bursts much higher than the BER aver-
aged on all bursts. These failures can result from realizations
of high-frequency offsets, from cycle slip occurrences or
simply from inaccurate frequency estimates. A simple
approach to evaluate a probability of failure is to monitor the
number of erroneous bits per burst at the algorithm output.
We consider a worst case: all frequency offsets are random
(Gaussian with a standard deviation σΔ f T ) except frequency
offset for cell 5, which is deterministic and equal to 4σΔ f T =
2 · 10−3 . The estimated distribution of the number of erro-
neous bits per burst for cell 5 at iteration 5 for Eb /N0 equal to
3.5 dB is shown in Figure 13. We define a failure occurrence
when the fraction of erroneous bits in a burst exceeds one
fourth of the total bits in the burst (106 = 53 · 8/4). We de-
duce a probability of failure equal to 2 · 10−3 . In the same way,
with a frequency offset for cell 5 equal to 3σΔ f T = 1.5 · 10−3 ,
we deduce a probability of failure equal to 10−4 .
5. CONCLUSION
We have studied in this paper an iterative multiuser detection
scheme, which includes channel estimation, suited to the re-
verse link of multibeam satellite communication systems. We
have first derived the algorithm in the case of time invari-
ant carrier phases. We have then discussed possible exten-
sions to the case of frequency offsets affecting user terminals.
Our main result is that if different approaches are possible
for the first iterations, frequency offset estimations are nec-
essary for final iterations in order to limit the degradation.
Further works will consist in evaluations (and possibly al-
gorithm modifications) with a more realistic channel model
including phase noise.
ACKNOWLEDGMENT
The authors would like to thank the reviewers for their
thoughtful and incisive comments about this paper.
EURASIP Journal on Wireless Communications and Networking
REFERENCES
“Iterative interference cancellation and channel estimation in
multibeam satellite systems,” International Journal of Satellite
Communications and Networking, vol. 25, no. 3, pp. 263–283,
2007.
212
satellite distribution systems,” December 2000, ETSI EN 301
790.
[3] M. L. Moher, “Multiuser decoding for multibeam systems,”
IEEE Transactions on Vehicular Technology, vol. 49, no. 4, pp.
1226–1234, 2000.
[4] G. Caire, M. Debbah, L. Cottatellucci, et al., “Perspectives
of adopting interference mitigation techniques in the context
of broadband multimedia satellite systems,” in Proceedings of
the 23rd AIAA International Communications Satellite Systems
Conference (ICSSC ’05), pp. 25–28, Rome, Italy, September
2005.
[5] M. Debbah, G. Gallinaro, R. Müller, R. Rinaldo, and A. Ver-
nucci, “Interference mitigation for the reverse-link of inter-
active satellite networks,” in Proceedings of the 9th Interna-
tional Workshop on Signal Processing for Space Communications
(SPSC ’06), Noordwijk, The Netherlands, September 2006.
[6] E. Lutz, M. Werner, and A. Jahn, Satellite Systems for Per-
sonal and Broadband Communications, Springer, New York,
NY, USA, 2000.
[7] J. P. Millerioux, “Techniques de détection multi-utilisateurs
pour les communications multifaisceaux par satellite,” Ph.D.
dissertation, ENST, Paris, France, September 2006.
[8] L. C. Godara, “Application of antenna arrays to mobile
communications—part II: beam-forming and direction-of-
arrival considerations,” Proceedings of the IEEE, vol. 85, no. 8,
pp. 1195–1245, 1997.
[9] A. N. D’Andrea, U. Mengali, and R. Reggiannini, “The mod-
ified Cramer-Rao bound and its application to synchroniza-
tion problems,” IEEE Transactions on Communications, vol. 42,
no. 234, pp. 1391–1399, 1994.
[10] F. Adriaensen, W. Steinert, and A. Van Doninck, “MF-TDMA
burst demodulator design with pilot symbol assisted fre-
quency estimation,” in Proceedings of the 8th ESA Interna-
tional Workshop on Signal Processing for Space Communications
(SPSC ’03), Catania, Italy, September 2003.
Hindawi Publishing Corporation
EURASIP Journal on Wireless Communications and Networking
Volume 2007, Article ID 58484, 9 pages
doi:10.1155/2007/58484

Research Article
A QoS Architecture for DVB-RCS Next Generation
Satellite Networks

Thierry Gayraud1, 2 and Pascal Berthou1, 2

1 Laboratoire d’Analyse et d’Architecture des Systèmes (LAAS-CNRS), University of Toulouse, Cedex 4, 31077 Toulouse, France
2 Toulouse University of Science, Toulouse, France

Received 1 October 2006; Revised 25 January 2007; Accepted 31 May 2007

Recommended by Ray E. Sheriff

The standardization of a return channel via satellite (DVB-RCS) and satellite community efforts in term of interoperability over
the last few years leads to quite a positive outcome: geostationary satellite networks are intended to provide broadband access to
interactive multimedia services in low-infrastructure areas. However, in order to take in account real-time multimedia traffic, an
efficient resource management scheme is still necessary to maximize the scarce uplink capacities usage. To address this capacity
issue, this paper proposes a complete DVB-RCS QoS architecture that is implemented, thanks to an emulation platform, and
evaluated with real multimedia applications. This paper first gives an overview of the QoS architecture usually used in DVB-
S/RCS satellite system, especially in layers 2 and 3. It then introduces the satellite system emulation used in the experimentation
and its calibration. The main contribution of this work focuses on the signaling principle designed to allow applications to take
benefit from the QoS features of the satellite system even if they are non-QoS aware. It is then shown how signaling in such QoS
architecture allows the user to change dynamically the QoS of his application using QoS agent and QoS server applications even
if the application is not QoS-aware. It is also given quantitative results related to such a dynamic QoS change in the experiments
done on the satellite emulation system.

Copyright © 2007 T. Gayraud and P. Berthou. This is an open access article distributed under the Creative Commons Attribution
License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly
cited.

1. INTRODUCTION Most recent commercial deployments provide either In-

Geostationary satellite access networks are expected to play,
in a near future, a decisive role in next generation networks
(NGNs) as they are intended to provide broadband access
to interactive multimedia services in low-infrastructure ar-
eas. Known as a real complementary technology in geo-
graphical locations beyond reach of terrestrial means, satel-
lite networks still suffer, in comparison to terrestrial net-
works, from long delays, scarce bandwidth resources, and
equipment costs.
The standardization of the digital video broadcasting-
return channel via satellite (DVB-RCS) in March 2000 and
the publication of the guideline document in September
2001 stand for major milestones in the development of re-
liable, efficient, and low-cost satellite equipment through the
harmonization of RCS terminals (ST) based on this open
standard. Several commercial DVB-RCS based networks are
already deployed and many efforts are done in order to en-
hance interoperability.
ternet access or mesh connectivity over a transparent geo-
stationary satellite. Fixed bandwidth contracts are generally
offered to consumers, thanks to a simple resource manage-
ment scheme. It simplifies admission control, reduces cost,
and gains experience while waiting for the standardization
of finer resource management strategies and equipment. A
lot of work on IP over satellite remains particularly in the
quality-of-service (QoS) field and the next step is, obviously,
to take benefits from DVB-RCS dynamic allocation schemes
and IP QoS architectures to cope with the satellite delay and
the scarce uplink resources.
This article proposes QoS architecture compliant with
the recommendation made by the ETSI BSM (broadband
satellite systems) working group which provides a state of the
art of existing QoS mechanisms that are applicable to broad-
band multimedia satellite systems [1].
The implementation made in a satellite emulation plat-
form represents a first attempt to evaluate a complete DVB-
RCS QoS architecture and a set of new services in a system
2 EURASIP Journal on Wireless Communications and Networking
based on either a regenerative or transparent satellite that will
be the future of satellite networks.
This paper proceeds in the following way. Section 2 gives
an overview of new trends in next generation satellite sys-
tems and sums up the principle of the DVB-RCS standard.
Section 3 describes the features of our QoS architecture. The
QoS signalling principle is explained here. Then, Section 4
shows our satellite emulation platform and an evaluation of
the new services provided by the QoS architecture, demon-
strating especially the dynamic QoS change features.
2. NEW TRENDS IN SATELLITE SYSTEMS
2.1. Forward link
The first DVB norm described a transmission scheme based
on MPEG-2 (Motion Picture Expert Group) video compres-
sion and transmission schemes, using MPEG-TS (MPEG-
transport stream). This latter was adapted for satellite sys-
tems through DVB-S (DVB transmission via satellite) that
defines series of options to send MPEG-TS packets over satel-
lite links and that is currently used for digital TV. The suc-
cess of this standard has caused its adoption for Internet ser-
vices over satellite. Then, the encapsulation of IP over MPE
(multiprotocol encapsulation) or more recently ULE (ultra
lightweight protocol) is needed. This leads to a complex net-
work stack. DVB-S2 standard [2] is intended to be a suc-
cessor of DVB-S with the same applications (TV, Internet,
etc.). It offers new coding techniques that can increase per-
formance by 25% over that of DVB-S, but is still compati-
ble with encapsulation layers as MPE or ULE. An alternative
known as GS (generic stream) intends to gain direct access to
the physical layer, avoiding the MPEG2-TS packet overhead,
but this protocol remains a work in progress.
The satellite terminals could therefore only receive DVB-
S/S2 frames from the satellite, but did not have the ability to
send any traffic towards the satellite.
2.2. Return link
In 1999, the ETSI proposed a standard for a return channel
via satellite, the DVB-RCS [3, 4], which supplements the STs
with the ability to transmit traffic towards the satellite.
According to this basis, two types of satellite can be de-
fined.
(i) Transparent satellite simply forwards the signal re-
ceived with no additional processing. A gateway (GW)
is needed on the ground to convert DVB-RCS frames
into DVB-S one. Each communication goes through
the gateway with a “star” topology. The delay to cross
the satellite network is about half a second and a dou-
ble hop (at least 1 second) is needed to connect two
satellite users.
(ii) Regenerative satellite with onboard switching payload
is able to demodulate, process, and remodulate the
traffic that goes though it and therefore to multiplex
several DVB-RCS signals into a single DVB-S one. The
associated topology could be a “star” or a “meshed”
network. The end-to-end delay decreases to only one
single hop.
Furthermore, DVB-RCS requires a medium access con-
trol (MAC) protocol because satellite terminals (ST) are able
to simultaneously access the return channel capacity. The
standard method relies on a multifrequency time division
multiple access (MF-TDMA). It basically relies on the avail-
ability of several TDMA channels (corresponding to different
carrier frequencies), each subdivided into frames and further
into timeslots of fixed length (bursts) during which the STs
are able to transmit data through MPEG2-TS or ATM traffic
bursts.
The entity responsible for this timeslot allocation within
the superframe shared by competing STs is the NCC (network
control center) that centralizes the satellite resources man-
agement. Thus it periodically broadcasts a signaling frame,
the TBTP (terminal burst time plan) that contains the infor-
mation on which STs relies to know when to transmit their
bursts.
This allocation can be dynamically modified by STs re-
quests so as to prevent from wasting satellite resources that
would be otherwise statically allocated. The implementation
of such a mechanism is generally known as bandwidth on-
demand (BoD) algorithm.
2.2.1. Bandwidth on-demand mechanisms
In order to dynamically manage the bandwidth allocation,
a bandwidth on demand protocol called demand assignment
multiple access (DAMA) is defined. It relies on the STs ability
to request frequently “capacities” to the NCC which enables a
regular update of the TBTP to fit to the STs respective traffic
load. The latter provides signaling schemes as well as MAC
QoS classes and their mapping on capacity types.
Thus, the norm defines 4 capacity categories to fit the ap-
plications needs:
(i) continuous rate assignment (CRA) which is static ca-
pacity, not subject to dynamic requests;
(ii) rate-based dynamic capacity (RBDC), which is dy-
namic rate capacity (in slots/frame), upper-bounded
by MaxRBDC, granted in response to dynamic re-
quests from the STs to track their instantaneous traffic
rate;
(iii) (absolute) volume-based dynamic capacity (VBDC and
AVDBC), which is also dynamic rate capacity (in slots),
granted in response to dynamic requests from the STs
to track their traffic queue state;
(iv) free capacity allocation (FCA), which is assigned to STs
on an “as available” basis from unused capacity.
Capacity types are vital to return path QoS support at MAC
layer; therefore, they are described in detail in the following.
Any given ST can be assigned one or a mix of the four capac-
ity types. Generally, higher priority classes of service are asso-
ciated with guaranteed capacity (CRA, RBDC), while lower
priority classes are predominantly given best effort capacity
(VBDC, FCA).
T. Gayraud and P. Berthou 3

Even if the service classes are properly defined, the allo- Application
QoS Agent
cation algorithms implemented in the NCC to fulfill the ser- QoS signalling
vices requirements are not specified. IP downstream
QoS Server
from user terminal
MF-classifier IP classes
3. QoS ARCHITECTURE

This section describes the QoS architecture we propose for Traffic shaping/policing /
DVB-S/RCS satellite systems. The main contribution is built
on return link management. Thus the downlink is generally
EF AF BE
considered not to be a bottleneck and classical traffic engi-

IP Layer
neering techniques are enough to managed the network.

Scheduling
EDF EDF
3.1. Basis of QoS in satellite systems
Transmission
To reach an optimal exploitation of uplink resources, at least allowed/denied
PQ
three functions must be implemented to provide QoS guar-
EF AF + BE
antees.
IP DVB/RCS
(i) QoS admission control consists, before the application Segmentation
interface
sends its traffic, to check that the network has enough
resources. This prevents some applications from send-
ing traffic that would otherwise lead to congestion

MAC layer
Threshold
among high priority traffic.
(ii) QoS enforcement consists in checking that the admit-
ted traffic respects its contracts, that is, that it does RT DVB NRT DVB
not use more resources than requested. This is done frames frames
by policing and shaping. Framing DAMA client
(iii) QoS differentiation consists in having several classes of
DVB-RCS frames Capacity
traffic, each class provides different behavior adapted TBTP requests
to a given service. This task is complex and needs dy- To satellite To/from NCC
namic management during the connections lifetime
and must be performed at two layers: the DVB-RCS DAMA server
and IP layers.
Figure 1: QoS architecture.
Thanks to the 5 bandwidth allocation mechanisms in-
cluded in DVB-RCS standard, the traffic differentiation is
made easily in introducing several MAC queues in the ST
stack and mapping the capacity requests over the MAC The user is able to classify its own flows in any available
queues. Then, IP DiffServ-based router architecture can be service through a dedicated agent (the QoS agent) that
setup over these new MAC services. However, this cannot be communicates with the QoS server to deliver the clas-
done without cross-layer mechanisms that ensure perfect re- sification. The goal is to exploit the capabilities offered
sources use. by the IP QoS capabilities.
An overview of this QoS architecture within the ST is given
3.2. Cross-layer architecture in Figure 1.

The QoS management, in the proposed architecture, is split

into three levels detailed in the following paragraphs. 3.3. QoS at DVB-RCS layer

(i) Satellite terminal resources: is medium access control
level, where the DVB-RCS DAMA allocates the band-
width on a fixed basis for real time applications and on
demand for other flows (nonreal-time traffic).
(ii) Class of service resources: a specific IP level module im-
plements a queue management system aiming at pro-
viding a differentiated service with regards to three ser-
vice classes. These service classes are deemed to exploit
the capabilities offered by the MAC level QoS capabil-
ities.
(iii) User level resources: this level is related to the share of
previous services resource between the different users.
QoS management at the MAC layer aims at sharing with op-
timal way the global uplink resources among the STs. Thus,
the MAC layer must be able to
(i) provide strict guarantees in terms of delay and jitter;
(ii) preserve these resources through fitting their alloca-
tion to the effective ST traffic load.
Within the ST MAC layer, the traffic is split into 2 classes
of service (CoS), DVB-RT for “real-time service” and DVB-
NRT for “nonreal-time” service, which are associated to
2 different ATM permanent virtual channels (PVC). One
(DVB-RT) benefits from static resource assignment through
4 EURASIP Journal on Wireless Communications and Networking
CRA; on the contrary, (DVB-NRT) relies on a dynamic re-
source allocation scheme also called BoD algorithm which
will be further detailed in this section.
(i) Real-time queue: the CRA consists in a fixed capacity
that is set at the ST log-on and is not subject to renego-
tiation during the ST connection lifetime. Each super-
frame contains one or more slots assigned to this con-
nection. This reserved static rate is entirely dedicated
to the DVB-RT traffic, since its high delay sensitive re-
quirements hardly tolerates throughput fluctuations.
(ii) Nonreal-time queue: the request category retained for
DVB-NRT traffic class is VBDC and FCA. Delay and
jitter tolerant traffic is supplied by the MAC scheduler
to the DAMA controller that computes the adequate
dynamic volume to request to the NCC. These requests
are sent out of band, not in traffic slots assignments,
but signaled in each SYNC slots broadcasted periodi-
cally by the NCC.
This architecture uses an original DAMA protocol that aims
to reduce the allocation delays without reducing the network
use.
As soon as an application produces a data, a free slot in
the next super frame should be available to send it. However,
the allocation done with the DAMA protocol takes at least
600 milliseconds (minimum scheduling latency—MSL). To
reduce its impact on the end-to-end delay, application needs
are anticipated in monitoring the DVB-NRT queue length
that grows conjointly. If the queue grows, the requested ca-
pacity is increased by a factor α otherwise only the minimum
is requested as. This algorithm is detailed in [5]. As shown
later, by properly setting α, the latency introduced by the BoD
algorithm can be effectively reduced.
In addition to the VDBC requests, the MAC scheduler in
the NCC distributes extra capacities to the logged STs if the
network is not congested. This last capacity category (FCA)
comes out to enhance the ST performance especially on low
loading conditions, preventing the ST from waiting at least
for the MSL to be able to transmit.
3.4. QoS at IP layer
In order to achieve a complete traffic control framework, a
classifier separates IP traffic into 3 categories:
(i) real-time: such an IP flow should be guaranteed a min-
imum bandwidth, an upper bound on queuing delay,
a mean queuing delay of a few tens of milliseconds;
(ii) nonreal-time: such an IP flow should be guaranteed a
minimum bandwidth, a mean queuing delay of a few
hundreds of milliseconds;
(iii) best-effort: all IP packets not recognized as belonging
to a particular IP flow are treated without any guaran-
tee on bandwidth or delay.
The classifier then maps the packets to the 2 MAC categories.
The overall goal of the architecture is to enforce the con-
straints for the IP categories as defined above while maxi-
mizing utilization of the available time-varying capacity.
With reference to IntServ/DiffServ traffic classes, the
best-effort (BE) traffic category supports the traditional ser-
vice offered by the Internet by default without any specific
QoS measure and whose performance are strongly impacted
by network congestion states. Real-time IP data category in-
cludes both IntServ guaranteed service class and DiffServ ex-
pedited forwarding (EF) PHB (per-hop behavior) while the
nonreal-time IP category is used for IntServ controlled load
service class and DiffServ assured forwarding (AF) PHB [6].
3.4.1. QoS enforcement
The fundamental component of the architecture is the
EDF scheduler preceded by token buckets (RC-EDF, rate-
controlled earliest deadline first) which allows fixing and up-
per bound to queuing delay and a minimum bandwidth for
separate IP flows. Namely, the presence of token buckets is
a guarantee that each IP flow will receive a minimum band-
width, given sufficient demand, equal to the relevant token
rate, while the EDF scheduler will guarantee to each packet
of an IP flow, once suitably regulated by a token bucket to be
served within a deadline equal to its associated static param-
eter.
In Figure 1, the RC-EDF components are gathered under
the appellation “traffic shaping/policing.” The traffic polic-
ing and shaping are then realized, thanks to single-rate token
buckets.
3.4.2. Layer 3/layer 2 mapping
The 3 traffic categories are served by a scheduler using a sim-
ple priority queuing (PQ) discipline. This means that
(i) packets from NRT queues are served only when RT
queues get completely emptied,
(ii) packets in the BE queue are extracted only when RT
and NRT queues are empty.
3.4.3. Application mapping
The EF traffic includes a number of real-time applications
with stringent time and bandwidth requirements such as
telephony or video conferencing. IP signaling which has very
stringent delay requirements but which is characterized by
low-data rates should use this service class too.
The AF traffic should include a number of traditional
Internet applications to be served with a satisfactory level
of service and transported over TCP. They include telnet,
HTTP, SMTP, FTP. Such applications can greatly vary in
terms of bandwidth and delay requirements. This means that
applications such as telnet or HTTP should be assured small
queuing delay though with limited bandwidth.
The BE class is designed to manage all traffic which is
not recognized as belonging to a particular user entitled to
receive better QoS or to applications with no particular delay
or bandwidth requirement. SMTP or FTP should belong to
this class.
T. Gayraud and P. Berthou 5

Satellite

IP
Hub Router
ST
IPv4 , IPv6 NCC
ST: satellite terminal QoS
MAC: medium access control MAC
segregation
NCC: network control center

Figure 2: QoS signaling principle.

3.5. QoS signaling

The link between the applications and the QoS architecture

is the QoS signaling. It allows the expression of the quality of
service requested by an application and the configuration of
the corresponding QoS provider.
In the proposed architecture (Figure 2), the QoS provider
is the ST. An application who want to take advantage of a
given IP QoS service (EF, AF, BE) must configure the satel-
lite terminal in order to redirect its packet on the appropriate
queue. A classical approach consists in statically configuring
the ST to associate a port to a service (e.g., the FTP port to
the best effort service). Usually done by the network admin-
istrator, this approach does not work with a set of new appli- Figure 3: QoS agent user interface (GUI).
cations that open unfixed port as, for instance, VoIP applica-
tions.
A more generic and “user-oriented” approach has been
proposed. The ST could be customized on the user request. Simulation and emulation both provide the opportunity
Dedicated software, the QoS agent [7], allows to associate to evaluate performances, at low-cost, on more or less realis-
a running application to one of the three defined services tic systems. When simulation needs a complete modeling of
and to send this association to the ST. It dynamically mon- the systems from applications to physical network and oper-
itors application connections and sends to the ST the 5 tu- ates in virtual time, emulation is more demonstrative since
ples source IP address, destination IP address, source port, real applications can be deployed over the model describ-
destination port, type of protocol for each of them. The ing transfer characteristics, delay, and error behaviors for in-
ST maintains an association’s table. Each incoming packet is stance.
redirected within the ST to the appropriate queue according For these reasons, the choice was made to set up a satellite
to this association’s table. emulation platform to demonstrate the network and applica-
Figure 3 shows the six connections opened by Gnomeet- tion services integration on next generation satellite systems
ing, videoconference software, and their selected services. and the possibility to interoperate with terrestrial networks.
The QoS agent can be run as a daemon to apply predeter-
mined rules without user interaction. In that case, it extends
the “classical” approach with new applications. 4.2. Test bed

4. EXPERIMENTAL MEASUREMENTS The network elements that belong to a classical satellite net-
work (Satellite, NCC, STs) are emulated individually on a
4.1. Emulation principle dedicated computer. A gigabit Ethernet interconnects them
and emulates the satellite carrier emulation. Ethernet was
Evaluating performances over real data links or networks chosen for its native broadcast abilities and also for its high
is expensive, even impossible for systems in development bandwidth capacities. Each satellite channel is mapped on a
phase. single Ethernet multicast address.
6 EURASIP Journal on Wireless Communications and Networking
4.2.1. Satellite link emulation
The satellite link emulator (SLE) simulates satellite link char-
acteristics in term of delay (and distribution); bit-error rate
(error burst frequency distribution, error burst length dis-
tribution), computed according to precalculated distribution
and based on real measurements.
Each channel crosses the link emulator to simulate the
effects of the two-way satellite link in real-time. The packets
sent from an ST to the SLE are delayed and are also subject to
a sequence of bit errors at random positions before they are
forwarded to the emulated “downlink” (a multicast address
per spot).
4.2.2. Network control center
The NCC is the core of the satellite network management. It
deals with allocating radio resources to the STs according to
their subscriber profile and available satellite resource. It cer-
tainly implements a DAMA controller, but provides also an
address resolution protocol to map IP addresses over under-
lying protocols and a QoS admission mechanism.
4.2.3. Satellite terminals
The satellite terminals are based on Linux systems. They act
as an access router interconnecting a LAN to an Internet ser-
vice provider over a satellite link. Its DVB-S/DVB-RCS in-
terfaces allow the data emission and reception and it imple-
ments the corresponding network layers. The proposed QoS
architecture is mainly located in the terminals.
The accuracy of the ST implementation is close to a pro-
totype version and makes the emulation very realistic but
also critical for its configuration. The calibration of the ser-
vices given to the users has been a touchy part of this work.
4.3. Platform calibration
As we explained in the former section, the available platform
will provide us with a prototype as close as possible to a real
DVB-S/RCS system behavior. So the quantitative results per-
formed on it are rather significant. To reach this result, the
platform has to be calibrated, that means that the right pa-
rameters have to be set to the right value. If this stage is not
properly achieved, then the results offer no interest, even if
all the parts of the emulation testbed are very accurate. So, in
this section, the basic platform configuration chosen in order
to carry out our experimental measurements is detailed.
This configuration stands for the reference scenario from
which all the calibration adjustments are done in order to
ensure the significance of platform performance. Through
this nonexhaustive list of the main platform configuration
parameters, we emphasize the huge possibility to customize
the platform which still remains, even for simplified current
commercial DVB-RCS deployments, a vital and difficult task.
4.3.1. Physical layer
The satellite diffusion properties are configured through the
SE (delay, Jitter and Losses patterns) and the return link ca-
Table 1: Basic physical and MAC layer configuration.
ST information peak rate 2048 Kbps
Physical Superframe 10 Trames
layer Frame duration 50 ms
Global DVB-RCS resource 2048 Kbps
Basic FCA None
DAMA α 1
CRA 96 Kbps
SLA ST
VBDC [FCA;1760 Kbps]
pacity segmentation scheme. The satellite emulator delay is
set to 250 milliseconds, the jitter is equal to ±1 milliseconds,
and the loss pattern is typical of a nice weather. Please note
that last notion which could sound subjective corresponds,
in the satellite emulator, to real satellite measurement traces.
4.3.2. MAC layer
The main parameters are closely linked to resource sharing
assignment from the NCC that distinguished two CoS at the
MAC layer in the ST. The ST maximum transmission rate is
shared by CRA and VBDC. Therefore, the peak transmission
rate is defined at first, then the CRA amount and finally the
DAMA configuration through the α anticipation parameter
and the FCA threshold (Table 1). The MAC queue sizes have
to respect minimum thresholds so as to prevent congestion
from occurring in the MAC Layer.
4.3.3. IP layer
Considering that EF and AF services are implemented strictly
according to the single-rate token buckets and that there is no
BE service conditioning, the main parameters of traffic con-
ditioning blocks (TCB) and IP scheduling are summarized in
Table 2.
4.4. Measurements
The measurements given in this section aims at evaluating
the dynamic QoS change mechanism. First, the experimenta-
tion done on DAMA results that have already been presented
in [5] are used to prove the right calibration of the emula-
tion testbed, so that the result obtained further is realistic.
The second part of the measurements has been performed on
a multiflow scenario involving several applications. The ob-
tained figures and tables show that the QoS architecture im-
plementing the QoS server allows the user to change dynam-
ically the QoS of one application thanks to the QoS agent.
4.4.1. Impact of FCA and α
The following study is linked to a VBR traffic source: a DIVX
streaming session. The DAMA influence cannot be neglected
in these experiments when the throughput variations can be
absorbed by the DAMA algorithm. Thus the basic DAMA
performances can be enhanced by increasing the anticipation
T. Gayraud and P. Berthou 7

Table 2: IP TCBs and Scheduler configuration.

Service EF or “voice” AF or “FTP” BE

Token bucket size 172 bytes (GSM packet size) 1500 bytes (Ethernet MTU)
No conditionning
Token rate 77.9 Kbps (∼
= 96 Kbps CRA) 77.9 Kbps
FIFO size — — 500 000 bytes
EDF delay 20 ms 50 ms —
Max latency 25 ms 500 ms 5s

100 100
90 90
80 80
70 70
60 60
CDF (%)

CDF (%)
50 50
40 40
30 30
20 20
10 10
0 0
0 0.5 1 1.5 2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8
Delay (s) Delay (s)

α=0 α=0
α = 0.5 α = 0.5
α=1 α=1

Figure 4: DAMA impact (VBR flow without FCA). Figure 5: DAMA impact on a VBR flow with FCA.

factor α which enables to bring 50% of the packets delay un- 800
der 1 seconds (Figures 4 and 5).
α=1
Under 0.5, the anticipation factor does not improve the 700
Queuing delay (ms)

end-to-end delay. If we consider this factor, the reduction of

α leads to a capacity underutilization, it will be maintained 600 α = 2/3
to 0.5 which stands for an interesting compromise between
500 α = 1/2
queuing delays and uplink utilization efficiency (Figure 6).
If FCA allocation is taken into account, an additional im- α = 1/3
400
provement can be noticed. The factor α seems to have less α=0
impact on the end-to-end delay than in the first scenario. 300
However, this 40 Kbps allocation stands for more than 10% 75 77.5 80 82.5 85 87.5 90 92.5 95 97.5 100
of the DIVX average throughput and might be considered Uplink utilization efficiency (%)
as overestimated for a single ST. Therefore, the anticipation
Figure 6: Delay versus efficiency.
factor has still a relative importance on VBR traffic which is
directly linked to its average throughput and variability.

4.4.2. Global architecture under different loading
conditions
The purposes of different tests are to measure the SATIP6
QoS performances under heterogeneous traffic flows (GSM
VoIP sessions, DIVX VoD, FTP, and web browsing) which
are mapped onto the three different SATIP6 services. As seen
previously, the voice service offers strict guarantees in terms
of delay and jitter. The AF service (FTP) ensures a minimum
throughput and protects the AF traffic from losses in con-
gestion state. Finally, the BE service is the most affected by
congestion while the satellite overload implies less capacities
for overall BE traffic and therefore higher delays and losses.
In Table 3, we can notice that voice service is not affected
by the loading conditions when the delay experimented by
FTP and BE traffic increases. Inside the DVB-NRT traffic
class, FTP traffic is protected from losses at the expense of
the BE class delay and loss ratio in congestion states.
 8 EURASIP Journal on Wireless Communications and Networking

2000 Table 4: Dynamic QoS change delay.

1800 Transition 1→2 2→3 3→4
1600 Initial One flow Two flows Two flows,
Throughput (kbits/s)

1 2 3 4 situation without QoS without QoS one with QoS

1400 Final Two flows Two flows, Two flows
1200
situation without QoS one with QoS without QoS
Delay (s) 2.8 3.5 3.15
1000

800
when it is downgraded, still longer in these two cases than in
600
the simple addition of a flow.
400 These results conclude the section dedicated to experi-
0 20000 40000 60000 80000 100000 120000 mental measurements by putting the stress on the proper
Time (ms) traffic differentiation carried out by our QoS architecture and
Throughput (kbits/s) the ability to change from one class of service to the other one
thanks to the QoS agent GUI.
Figure 7: Dynamic change of class of service using the QoS agent.
5. CONCLUSION
Table 3: SATIP6 QoS performance under different loading condi-
tions. It was proved in this paper that it was possible to specify and
implement QoS architecture for DVB-S/RCS satellite system
Network Average delay (Jitter) [ms] Losses [%] in order to provide the user with QoS guarantees even if a
load [%] BE FTP Voice BE FTP Voice non-QoS-aware application is used. MAC algorithms (such
25 293 (26) 293 (26) 283 (23) 0 0 0 as DAMA) were proved to be efficient. It was also explained
50 291 (23) 291 (24) 283 (23) 0 0 0 how to proceed to evaluate such an architecture. Using mea-
75 290 (24) 289 (24) 283 (23) 0 0 0 surement tools on a well-calibrated testbed, the global QoS
100 919 (23) 948 (24) 283 (23) 0 0 0 of the satellite system may be evaluated accurately.
125 6753 (23) 1783 (24) 283 (23) 33 0 0
Using other capacity category than VBDC (RBDC for in-
stance) is improving resource utilization especially if appli-
150 6755 (28) 1783 (24) 283 (23) 37 0 0
cations throughputs are known. Unfortunately, this is not
usual today in the Internet. In that case, we propose to use
rate-based signaling protocols (SIP) in order to set up the
right capacity requests.
4.4.3. Dynamic change of QoS
Other future work may also be done related to DVB-S2
The change of multimedia stream QoS is done thanks to QoS and new admission control mechanisms.
agent.
The different following steps are easy to find in Figure 7. ACKNOWLEDGMENTS
(i) Step 1: the scenario begins as a UDP video stream The authors wish to thank all the partners of the SATIP6
starts. This stream is sent in the BE class of service. [8] consortium: Alcatel Space (France), which is the project
(ii) Step 2: 25 seconds later, another flow is sent on the coordinator, Telecom Italia Lab (Italy), France Telecom SA
same uplink; it is done so that the uplink is now con- (France), University of Rome “La Sapienza” (Italy), Sintef
gested. The throughput of the first stream is then re- (Norway), LAAS-CNRS (France), and Alliance Qualité Logi-
duced to 800 kbps. ciel (France).
(iii) Step 3: 25 seconds later, the user decides to upgrade the
class of service of this stream and set it up to “voice.”
After traffic burst, due to the addition of the EF service REFERENCES
of 1 Mbps, and the traffic buffered before the resource [1] ETSI TR 102 157, “Satellite Earth Stations and Systems (SES);
reservation, the throughput is around 1 Mbps. Broadband Satellite Multimedia; IP Internetworking over satel-
(iv) Step 4: at t = 80 s, the stream is downgraded back to BE lite; Performance, Availability and Quality of Service,” July
and the stream throughput is then around 800 kbps. 2003.
[2] ETSI Standard TR 102 376 V1.1.1, “Digital Video Broadcasting
In Figure 7, the time needed to change from one Cos to an- (DVB); User guidelines for the second generation system for
other one for a flow could also be evaluated when the link is Broadcasting, Interactive Services, News Gathering and other
congested by other data flows. broadband satellite applications (DVB-S2)”.
The delays given in Table 4 show as usually that the delay [3] ETSI EN 301 790 V1.3.1, “Digital Video Broadcasting (DVB);
is around 3 seconds. It is less when a new flow is added on a Interaction channel for Satellite Distribution Systems,” March
link. It is longer when the QoS of a flow is upgraded and less 2003.
T. Gayraud and P. Berthou 9

[4] ETSI TR 101 790 V1.2.1, “Digital Video Broadcasting (DVB);

Interaction channel for Satellite Distribution Systems, Guide-
lines for the use of EN 301 790,” January 2003.
[5] A. Pietrabissa, T. Inzerilli, O. Alphand, et al., “Validation of
a QoS architecture for DVB-RCS satellite networks via the
SATIP6 demonstration platform,” Computer Networks, vol. 49,
no. 6, pp. 797–815, 2005.
[6] J. Heinanen, F. Baker, W. Weiss, and J. Wroclawski, “RFC2597,
Assured Forwarding PHB,” June 1999.
[7] S. Combes, O. Alphand, P. Berthou, and T. Gayraud, “Satellite
and next generation networks: QoS issues,” International Jour-
nal of Space Communications, 2006.
[8] IST SATIP6 Project (Contract IST-2001-34344).
Hindawi Publishing Corporation
EURASIP Journal on Wireless Communications and Networking
Volume 2007, Article ID 65058, 8 pages
doi:10.1155/2007/65058

Research Article
Maximum Likelihood Timing and Carrier Synchronization in
Burst-Mode Satellite Transmissions

Michele Morelli and Antonio A. D’Amico

Department of Information Engineering, Via Caruso, 56100 Pisa, Italy

Received 4 August 2006; Revised 2 March 2007; Accepted 13 May 2007

Recommended by Alessandro Vanelli-Coralli

This paper investigates the joint maximum likelihood (ML) estimation of the carrier frequency offset, timing error, and carrier
phase in burst-mode satellite transmissions over an AWGN channel. The synchronization process is assisted by a training sequence
appended in front of each burst and composed of alternating binary symbols. The use of this particular pilot pattern results into
an estimation algorithm of affordable complexity that operates in a decoupled fashion. In particular, the frequency offset is mea-
sured first and independently of the other parameters. Timing and phase estimates are subsequently computed through simple
closed-form expressions. The performance of the proposed scheme is investigated by computer simulation and compared with
Cramer-Rao bounds. It turns out that the estimation accuracy is very close to the theoretical limits up to relatively low signal-to-
noise ratios. This makes the algorithm well suited for turbo-coded transmissions operating near the Shannon limit.

Copyright © 2007 M. Morelli and A. A. D’Amico. This is an open access article distributed under the Creative Commons
Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is
properly cited.

1. INTRODUCTION tion accuracy as compared to a non-data-aided (NDA) ap-

Burst transmission of digital data and voice is widely adopted
in satellite time-division multiple-access (TDMA) networks.
In these applications the propagation medium can be reason-
ably modeled as an additive white Gaussian noise (AWGN)
channel and knowledge of carrier frequency, symbol timing,
and carrier phase is necessary for coherent demodulation
of the received waveform. In the presence of nonnegligible
phase noise and/or oscillator instabilities, differential detec-
tion is often employed to overcome the inherent difficulty
posed by the phase estimation process. Even with differen-
tial detection, however, the problem of timing and frequency
offset recovery still remains.
Depending on their topology, synchronization circuits
can be divided into two main categories: feedback and feed-
forward schemes [1, 2]. The former have good tracking ca-
pabilities but exhibit comparatively long acquisitions due to
hang-up phenomena [3–5]. The latter have shorter acqui-
sitions and, accordingly, are better suited for burst-mode
transmissions. In many cases, a preamble of known sym-
bols is appended at the beginning of each burst to assist the
synchronization process. Actually, the use of a preamble al-
lows data-aided (DA) operation and provides better estima-
proach. Even so, however, synchronization may prove diffi-
cult, especially with turbo-coded modulations operating at
relatively low signal-to-noise ratios (SNRs). Clearly, very ef-
ficient synchronization algorithms are needed in these con-
ditions [6].
The common approach to solve the synchronization
problem in burst-mode transmissions is to estimate the tim-
ing error first, and then use the time-synchronized samples
for frequency and phase recovery. Two prominent feedfor-
ward schemes for NDA timing estimation are investigated
in [7, 8]. In particular, timing estimates are derived in [7]
by searching for the maximum of an approximate version
of the likelihood function while in [8] the received signal is
sampled at some multiple of the symbol rate and a square-
law nonlinearity (SLN) is employed to wipe the modula-
tion out. As shown in [2], the method in [8] is an efficient
way of maximizing the likelihood function of [7] as long
as the bandwidth of the complex envelope of the transmit-
ted signal does not exceed the signaling rate. Since the use
of a SLN exhibits poor performance in the presence of nar-
rowband signaling, alternative methods employing absolute
value or fourth order-based nonlinearities have been devised
[9]. The main advantage of the timing estimators in [7–9] is
2 EURASIP Journal on Wireless Communications and Networking
that they can operate correctly even in the presence of carrier
frequency offsets (CFOs) as large as 20% of the symbol rate.
Frequency estimation is usually performed by exploiting
the received time-synchronized samples. A large number of
schemes proposed in the past operate in either the frequency
or time domain. The Rife and Boorstyn (R&B) algorithm
[10] belongs to the former class and provides maximum like-
lihood (ML) estimates of frequency and phase errors by look-
ing for the peak of a periodogram. Interpolation techniques
may be employed to find an explicit expression of the peak
location [11]. In the time-domain approach, suitable corre-
lations of the received samples are exploited to compute the
frequency estimates. A representative selection of schemes
derived along this line of reasoning can be found in [12–15].
These methods attain the Cramer-Rao lower bound (CRB) at
intermediate/high SNRs, but exhibit different performance
in terms of estimation range and threshold, that is, the SNR
below which large estimation errors are likely to occur.
A possible drawback of conventional frequency estima-
tion schemes as those discussed in [10, 12–15] is that they all
assume ideal timing synchronization. Their performance is
thus limited by the accuracy of the timing estimator. A DA
algorithm for the joint estimation of the carrier phase, fre-
quency offset, and timing error has been proposed in [16] by
resorting to ML arguments. In order to work properly, how-
ever, the demodulated signal must incur negligible phase ro-
tations during the preamble duration. This poses a stringent
limit to the maximum tolerable CFO, which may prevent the
application of this method to many practical situations.
In the present paper we are concerned with the joint esti-
mation of all synchronization parameters for a burst-mode
satellite system operating over AWGN channels. Since one
distinct feature of packetized transmissions is that synchro-
nization must be achieved as fast as possible, in the follow-
ing we only focus our attention to a feedforward structure.
Also, we assume that a preamble of alternating binary sym-
bols is transmitted at the beginning of each burst to facili-
tate the timing estimation task [17]. Our approach is based
on ML methods and leads to a three-step procedure. In
the first step frequency recovery is accomplished through a
mono-dimensional grid search. The estimated CFO is then
exploited in the second step to obtain a closed-form expres-
sion of the timing estimate. The final step is devoted to phase
estimation and can be skipped in case of differential data
detection. Surprisingly, no complicated multidimensional
searches are needed to jointly estimate all the unknown syn-
chronization parameters. Simulations indicate that the pro-
posed estimator is well suited for turbo-coded transmissions
since its accuracy approaches the relevant CRBs even at low
SNR values. However, it should be observed that this advan-
tage is achieved at the price of a higher computational com-
plexity as compared to other existing alternatives.
The paper is organized as follows. In Section 2 we in-
troduce the signal model and formulate the synchronization
problem. Section 3 illustrates the joint ML estimation of the
unknown parameters and discusses in some detail the prac-
tical implementation of the frequency estimator. In Section 4
we derive CRBs to characterize the ultimate accuracy of fre-
quency, timing, and phase estimates. Simulation results are
presented in Section 5 while some conclusions are drawn in
Section 6.
2. SYSTEM MODEL AND PROBLEM FORMULATION
2.1. Statement of the problem
We consider the reverse link of a satellite communication
system and assume a time-division multiple-access (TDMA)
scheme where each earth station transmits bursts of data. The
structure of a burst is detailed in Figure 1. Essentially, it con-
sists of two parts: a header section followed by a payload. The
header is further divided in two portions, namely, a synchro-
nization preamble and a unique word (UW). The preamble is
made of a sequence of training symbols which are exploited
by the receiver for carrier and symbol timing recovery. The
UW is located just after the preamble and is used for burst
identification as well as to establish the start of the payload.
The first task of the receiver is the start of burst (SoB) de-
tection, that is, the recognition of the time-of-arrival (ToA)
of a generic burst. This is normally performed through a sim-
ple noncoherent energy-detection scheme which provides a
coarse estimate of the position of each burst. Once the SoB
has been identified, the preamble is exploited for carrier and
symbol timing synchronization. This is the second task of the
receiver and represents the focus of our paper. In order to ex-
plain how synchronization can be achieved, we concentrate
on a single burst and assume that the SoB detection algo-
rithm has provided a ToA estimate with an error τ, as shown
in Figure 2. The offset τ can be decomposed as follows:
τ = ηT − εT, (1)
where T is the symbol period, η is an integer (integer delay),
and ε ∈ [−0.5, 0.5) is a real-valued parameter (fractional de-
lay). During the preamble we are interested in the estimation
of the fractional delay, because the integer delay is recovered
later by searching for the location of the UW within the burst.
The estimation of the synchronization parameters (fractional
delay, carrier phase, and frequency offsets) is performed by
observing a portion of the preamble of length NT (N is a de-
sign parameter) at the right of the assumed SoB, as shown in
Figure 2. Clearly, the total duration of the preamble has to be
larger than τ + NT. Since τ is a random variable, this condi-
tion can be practically met by a proper design of the preamble
length. Since we are not concerned with the estimation of the
integer delay, in the following we set η = 0.
2.2. Signal model
We consider a linearly modulated digital signal transmitted
over an AWGN channel. The complex envelope of the re-
ceived waveform is modeled as
r(t) = e j(2π fd t+ϕ) s(t − εT) + w(t), (2)
where s(t) is the useful signal, ϕ and fd are the carrier phase
and frequency offset, respectively, and w(t) is thermal noise
with independent real and imaginary components, each with
M. Morelli and A. A. D’Amico 3

Burst#1 Burst#2 Burst#K

···

Guard Sync.
interval preamble UW
Header Payload

Figure 1: Burst structure.

Sync. ing samples taken at t = kT/2, with 0 ≤ k ≤ 2N − 1. As the

preamble
signal is not distorted in passing through the filter, we have
UW Payload
  
k
τ NT x(k) = e j(πkν+ϕ) cos − ε π + n(k) for 0 ≤ k ≤ 2N − 1,
2
(6)
Figure 2: Start of burst estimation error.
where ν = fd T is the CFO normalized to the symbol-rate 1/T
and n(k) = nR (k) + jnI (k) is the noise contribution. Due to
two-sided power spectral density N0 . Signal s(t) is expressed the previous hypotheses, {nR (k)} and {nI (k)} are indepen-
as dent and white random sequences with the same variance
 σ 2 = (Es /N0 )−1 . As the signal component in (6) depends on
s(t) = an g(t − nT), (3) ν, ε, and ϕ, we may estimate all these parameters from the
n observation of {x(k)}. This problem is addressed in the next
section by resorting to ML methods.
where {an } are modulation symbols taken from a PSK or
QAM constellation and g(t) has a root-raised-cosine Fourier
3. MAXIMUM LIKELIHOOD ESTIMATION OF
transform with some roll-off α. To facilitate the timing esti-
THE SYNCHRONIZATION PARAMETERS
mation process, during the preamble we assume a pilot pat-
tern composed by alternating BPSK symbols +1 and −1 [17]. 3.1. Maximization of the likelihood function
Accordingly, s(t) is given by
 Bearing in mind (6), the log-likelihood function for the un-
 
2Es πt known parameters is given by
s(t) = cos (4)
T T

Λ( ν, ε , ϕ)
= −2N ln 2πσ 2
with Es denoting the signal energy per symbol interval, and 2N −1    2
1   x(k) − e j(πk ν+ϕ) k 
r(t) may be rewritten in the form − 2 

cos −
ε π 
 ,
2σ k=0 2
  
2Es j(2π fd t+ϕ) π(t − εT) (7)
r(t) = e cos + w(t)
T T (5)
where ν, ε , and ϕ are trial values of ν, ε, and ϕ, respec-
for t ∈ [0, NT]. tively. The joint ML estimate of (ν, ε, ϕ) is the location where
Λ( ν, ε , ϕ)
achieves its global maximum. Skipping irrelevant
In order to produce a discrete-time signal, the received factors and additive terms independent of ( ν, ε , ϕ), it turns
waveform is fed to an anti-aliasing filter (AAF) and sampled out that Λ( ν, ε , ϕ)
may equivalently be replaced by
at some rate fc . The filter bandwidth BAAF and the sampling
rate are chosen such that the signal component is passed −1   

2N
k
undistorted (even for the maximum frequency offset) and no = e e− j ϕ
Ψ( ν, ε , ϕ) x(k)e− jπk ν cos − ε π ,
k=0
2
aliasing occurs. Assuming that the CFO is less in magnitude
than 0.5/T, from (5) it follows that we can set BAAF = 1/T (8)
and fc = 2/T. For simplicity, in the ensuing discussion the
AAF is assumed with a brick-wall transfer function, even where e{·} denotes the real part of the enclosed quantity.
though the rectangular shape is not strictly necessary and Function Ψ( ν, ε , ϕ)
can also be rewritten as
could be easily made more realistic [18].   
For normalization
 purposes, the output of the AAF is Ψ( ν, ε , ϕ) ν) cos(π ε ) + e− jπ ν Yo ( ν) sin(π ε )
= e e− j ϕ Ye (
scaled by a factor T/2Es and we call x(k) the correspond- (9)
4 EURASIP Journal on Wireless Communications and Networking

with 1.2

−1
N Es /N0 = 10 dB
Ye ( ν) = (−1)k x(2k)e− j2πk ν , 1 N = 64
k=0
−1
N
(10)
ν
− j2πk
0.8
Yo ( ν) = k
(−1) x(2k + 1)e .
k=0

P(ν)
0.6
To ease the search for the maximum of Ψ( ν, ε , ϕ),
we rewrite
(9) in the form
0.4
   
= Z(
Ψ( ν, ε , ϕ) ν, ε ) cos ψ( ν, ε ) − ϕ , (11)
0.2
where Z( ν, ε ) is a function of ν and ε defined as

Z( ν, ε ) = Ye ( ν) cos(π ε ) + e− jπ ν Yo ( ν) sin(π ε ) (12) 0

−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5
ν
while ψ( ν, ε ) = arg{Z( ν, ε )} is the argument of Z( ν, ε ).
Clearly, for fixed ν and ε , the maximum of Ψ( ν, ε , ϕ)
is
Figure 3: Typical shape of P(ν).
achieved when the cosine factor in (11) is equal to unity,
which occurs for
 

ϕ( ν, ε ) = arg Z( ν, ε ) . (13) from which it follows that the ML estimate of the frequency
offset is given by
In this case the right-hand side of (11) reduces to |Z( ν, ε )|  
and the ML estimates of ν and ε are found by maximizing the ν = arg max P( ν)
 (20)
ν

following function:
while the timing estimate is obtained from (18) in the form
 2
Γ( ν, ε ) = 2Z( ν, ε ) (14)
1  
ε = arg A(ν) . (21)
with respect to ν and ε , where the factor 2 in the right-hand 2π
side of (14) has only been inserted to avoid a factor 1/2 in the In case of coherent detection, an estimate of the carrier phase
subsequent equations. ϕ is also necessary. This is computed as indicated in (13) after
To proceed further, we substitute (12) into (14) and ob- replacing ( ν, ε ) by (ν, ε) and reads
tain
 
 2  2   ϕ = arg Ye (ν) cos(π ε) + e− jπ ν Yo (ν) sin(π ε) . (22)
Γ( ν, ε ) = Ye ( ν) + Yo ( ν) + e e− j2π ε A( ν) , (15)
In the sequel the algorithm based on (20)–(22) is called
where A( ν) is defined as the ML estimator (MLE).
 2  2  
A( ν) = Ye ( ν) − Yo ( ν) + 2 j e e jπ ν Ye ( ν)Yo∗ ( ν) . 3.2. Remarks
(16)
(1) Contrarily to what one might fear, the maximization of
Observing that |A( ν)| = |Ye2 ( ν) + e− j2π ν Yo2 ( ν)|, we may the likelihood function Λ( ν, ε , ϕ)
needs not be made on a
rewrite (15) into the equivalent form three-dimensional domain. Actually, the location (ν, ε, ϕ)
 of
 2  2 the maximum can be found through simple steps, each in-
Γ( ν, ε ) = Ye ( ν) + Yo ( ν) volving a single synchronization parameter. In particular, the
    (17) first step requires maximizing the function P( ν) defined in
+ Ye2 ( ν) + e− j2π ν Yo2 ( ν) cos θ( ν) − 2π ε
(19) in order to get the CFO estimate ν. As discussed later,
this can be done through a grid search over the interval where
with θ( ν) = arg{A( ν)}. For a given ν, the maximum of Γ( ν, ε )
ν is expected to lie. Once ν has been obtained, timing and
is achieved by setting
phase estimates are computed in closed form as indicated
1   in (21) and (22), respectively. In summary, the difficult and
ε( ν) = arg A( ν) . (18) time-consuming part in the estimation of (ν, ε, ϕ) is the one
2π
that locates the maximum of P( ν). Once this is done, the
Substituting this result into the right-hand side of (17) yields computation of ε and ϕ becomes a trivial task.
 2  2   (2) Maximizing function P( ν) may pose some difficulty
P( ν) = Ye ( ν) + Yo ( ν) + Ye2 ( ν) + e− j2π ν Yo2 ( ν) due to the presence of many local maxima. This is clearly ev-
(19) ident from Figure 3, which illustrates a typical realization of
M. Morelli and A. A. D’Amico 5

P( ν) as obtained by simulation with N = 64, Es /N0 = 10 dB, of N and Es /N0 . For this reason, it is of interest to derive
and ν = 0.1. As discussed in [10], the global maximum can the CRB for the joint estimation of the set of parameters
be sought in two steps. The first one (coarse search) calculates η = (ν, ε, ϕ) based on the model (6).
P( ν) for a set of ν values, say { νn }, covering the uncertainty We begin by computing the entries of the Fisher infor-
range of ν and determines the location νM of the maximum mation matrix F. They are defined as [20]
over this set. The second step (fine search) makes an inter-  
polation between the samples P( νn ) and computes the local ∂2 Λ(η)
[F]i, j = −E , 1 ≤ i, j ≤ 3, (25)
maximum nearest to νM . It should be noted that the shape ∂ηi ∂η j
of P( ν) is occasionally so badly distorted by noise that its
highest peak may be far from the true ν. When this happens, where Λ(η) is the log-likelihood function in (7) and η de-
the MLE makes large errors (outliers) and the system perfor- notes the th entry of η. Substituting (7) into (25), after some
mance is highly degraded. The SNR below which the outliers manipulations we obtain
start to occur is referred to as the threshold of the estimator. ⎡    ⎤
π 2 (2N − 1) 4N − 1 − 3 cos(2πε) 0 3π 2N − 1 − cos(2πε)
(3) In practice the coarse search can be efficiently per- N ⎢ ⎥
F= 2 ⎢ 0 6π 2 0 ⎥.
formed using fast Fourier transform (FFT) techniques, as it 6σ ⎣   ⎦
3π 2N − 1 − cos(2πε) 0 6
is now explained. Starting from the observed samples {x(k)},
(26)
we first compute the following zero-padded sequences of
length KN: The CRB for the estimation of η is given by [F−1 ], . Skip-
⎧
⎨(−1)k x(2k), 0 ≤ k ≤ N − 1, ping the details, it is found that
ye (k) = ⎩
−1
0, N ≤ k ≤ NK − 1, 12 Es /N0
⎧ (23) CRB(ν) =  , (27)
⎨(−1)k x(2k + 1), π 2 N 4N 2 − 4 + 3 sin2 (2πε)
0 ≤ k ≤ N − 1,
yo (k) = ⎩
−1
0, N ≤ k ≤ KN − 1, Es /N0
CRB(ε) = , (28)
π2N
where K is a design parameter called pruning factor. Next, the

−1
FFTs of { ye (k)} and { yo (k)} are evaluated at the points (2N − 1)(4N − 1 − 3 cos ε) Es /N0
CRB(ϕ) = 2   . (29)
n KN KN N 4N 2 − 4 + 3 sin2 (2πε)
νn =
, − ≤n≤ . (24)
KN 2 2
Interestingly, for large data records we can approximate (27)
This produces the quantities {Ye ( νn )} and {Yo ( νn )}, which as
are next exploited to get {P( νn )} as indicated in (19). Finally,
the largest P( νn ) is sought and this provides the coarse fre- 3(Es /N0 )−1
CRB(ν) ≈ (30)
quency estimate. π2N 3
(4) Collecting (10) and (19), it is seen that P( ν) is peri- which represents the CRB for the estimation of the frequency
odic of unit period. This means that MLE gives unambigu- of a complex sinusoid embedded in AWGN [10].
ous frequency estimates as long as ν is confined within the
interval [−1/2, 1/2). This is the frequency estimation range of
5. SIMULATION RESULTS
MLE.
(5) Compared to the R&B algorithm [10], the MLE is In this section we report on simulation results illustrating the
more complex to implement as it requires the computation performance of MLE over an AWGN channel. Unless other-
of two FFTs instead of a single FFT. In addition to carrier syn- wise specified, the synchronization parameters vary at each
chronization, however, the MLE also provides timing recov- new simulation run and are modeled as statistically indepen-
ery. Actually, from (16) and (21) we see that computing the dent random variables with a uniform distribution. In par-
timing estimate only requires knowledge of Ye (ν) and Yo (ν). ticular, ν and ε are confined within [−0.5, 0.5) while ϕ takes
Since these quantities can easily be obtained by {Ye ( νn )} and values in the interval [−π, π). A pruning factor K = 4 is
{Yo ( νn )} through interpolation, the timing estimation task is used to compute the quantities {Ye ( νn )} and {Yo ( νn )}. Also,
accomplished with a relatively low computational effort once a parabolic interpolation is chosen in the implementation of
ν is available. However, it should be observed that the com-
 the fine search. This yields a frequency estimate in the form
plexity associated to the synchronization process is negligible


compared to that of iterative data decoding [19]. So, the re- δ P νM −1 − P νM+1
ν = νM + ν ·

, (31)
quirement for an additional FFT has only a marginal impact
on the overall receiver complexity.
4. CRB ANALYSIS
By invoking the asymptotic efficiency property of the MLE,
we expect that the accuracy of the estimates (20)–(22) ap-
proaches the corresponding CRBs for relatively large values
 
2 P νM −1 − 2P( νM ) + P νM+1
where δν = 1/KN is the distance between two adjacent sam-
ples P( νn ) while νM is the output of the coarse search. The
observation length is fixed to either N = 32 or 64. For
comparison, in the ensuing discussion we also consider a
synchronization scheme in which timing recovery is first ac-
complished by resorting to the Oerder and Meyr (O&M)
6 EURASIP Journal on Wireless Communications and Networking

0.5 10−1

0.4 Es /N0 = 10 dB
N = 64
0.3
E{ν} = ν 10−2
0.2

0.1
E{ν}

MSEε
10−3
−0.1

−0.2

−0.3 10−4
−0.4
N = 32
−0.5
−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5
10−5
ν 0 5 10 15 20
Es /N0 (dB)
Figure 4: Mean normalized frequency estimates versus ν.
O&M
MLE
0.5 CRB

0.4 Es /N0 = 10 dB
N = 64
Figure 6: MSE performance of MLE and O&M estimator with N =
0.3 32.
E{
ε} = ε
0.2
10−1
0.1
ε}
E{

−0.1
10−2
−0.2

−0.3
MSEε

−0.4 10−3
−0.5
−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5
ε
10−4
Figure 5: Mean timing phase estimates versus ε.
N = 64

10−5
algorithm [8] and carrier synchronization is next achieved 0 5 10 15 20
by applying the R&B method [10] to the time-synchronized Es /N0 (dB)
samples. The O&M operates with four samples per symbol
period. O&M
MLE
Figures 4 and 5 illustrate average frequency and timing CRB
estimates, E{ν} and E{ ε}, provided by MLE as a function of
ν and ε, respectively. The observation length is N = 64 while
Figure 7: MSE performance of MLE and O&M estimator with N =
the SNR is fixed to Es /N0 = 10 dB. The ideal lines E{ν} = ν 64.
and E{ ε} = ε are indicated as references. These results show
that ν and ε are unbiased over the full range [−0.5, 0.5).
Figures 6 and 7 compare the mean square error (MSE)
of the timing estimates, E{(ε − ε)2 }, as obtained with MLE marks. We see that MLE has the best accuracy, especially at
and O&M. The observation length is N = 32 in Figure 6 low SNRs where a significant gain is observed with respect
and N = 64 in Figure 7. Marks indicate simulation results to O&M. In particular, for N = 64 the MLE is close to the
while the thin solid lines are drawn to ease the reading of the CRB down to Es /N0 values of 0 dB, while O&M approaches
graphs. The corresponding CRBs are also shown as bench- the bound only for Es /N0 > 10 dB. This feature of the MLE
M. Morelli and A. A. D’Amico 7

10−2 100
8
6
4
10−3
2

10−4 N = 32 10−1
8

MSEϕ (rad2 )
6
MSEν

4
10−5
2 N = 32

10−6 10−2
8
6
4
10−7
N = 64 N = 64
2

10−8 10−3
0 5 10 15 20 0 5 10 15 20
Es /N0 (dB) Es /N0 (dB)

R&B R&B
MLE MLE
CRB CRB

Figure 8: MSE performance of MLE and R&B estimator with N = Figure 9: Mean square error MSEϕ versus Es /N0 with N = 32 and
32 and N = 64. N = 64.

is of great importance as it makes this estimator suitable for 6. CONCLUSIONS

turbo-coded modulations operating at very low SNRs.
Figure 8 illustrates the accuracy of the frequency esti-
mates provided by MLE and R&B with either N = 32 or
64. Again, the simulation results are compared with the rel-
evant CRBs. We see that the estimation accuracy keeps close
to the CRB down to a certain value of Es /N0 that depends
on the adopted scheme and observation length. If the SNR
is decreased further, a rapid increase in the MSE is observed.
The abscissa at which the slope of the curve starts to change
indicates the estimator threshold and is a manifestation of
the occurrence of outliers. Since large errors have disabling
effects on the system performance, the frequency estimator
must operate above threshold. The results in Figure 8 re-
veal that MLE has a lower threshold than R&B, especially for
N = 64, which translates into an increased robustness against
outliers. This fact reinforces the idea that for low SNR ap-
plications the MLE is more efficient than other conventional
synchronization schemes. As expected, the threshold is a de-
creasing function of the observation length. Actually, we see
that doubling N results into a threshold decrease of approxi-
mately 3 dB with MLE, while a gain of 2 dB is observed with
R&B.
As mentioned previously, in case of coherent detection
phase recovery is required in addition to frequency and tim-
ing synchronization. The MSE of the phase estimates pro-
vided by MLE and R&B is illustrated in Figure 9 as a func-
tion of Es /N0 . These results are qualitatively similar to those
in Figure 8. In particular, it turns out that both schemes ap-
proach the CRB at intermediate/high SNR values, but MLE
exhibits a lower threshold than R&B.
We have addressed the joint ML estimation of the carrier fre-
quency, timing error, and carrier phase in burst-mode satel-
lite transmissions. Thanks to a suitably designed pilot pat-
tern composed of alternating binary symbols (which pro-
duces two spectral lines at ±1/2T), the estimation process
can be divided into three separate steps, each devoted to the
recovery of a single synchronization parameter. In particular,
timing and phase recovery is accomplished in closed form,
whereas the measurement of the frequency offset involves a
grid search which represents the time-consuming part of the
overall synchronization procedure.
Comparisons have been made with a conventional
scheme in which timing recovery is accomplished in an NDA
fashion and carrier synchronization is next achieved by ex-
ploiting the time-synchronized samples. Computer simula-
tions indicate that the proposed ML algorithm provides more
accurate timing estimates at low SNR values. In addition, it
exhibits increased robustness against the occurrence of out-
liers in the frequency estimates. It is fair to say that these ad-
vantages are achieved at the price of a certain increase of the
processing load as compared to the conventional scheme.
The question of which method is better is not easily
answered because it depends on the different weights that
may be given to the various performance indicators, includ-
ing estimation accuracy, computational complexity, and con-
straints on the pilot pattern. It is likely that the choice will
depend on the specific application. For example, the pro-
posed algorithm seems attractive for coded transmissions as
it approaches the relevant CRBs down to very low SNRs. On
8 EURASIP Journal on Wireless Communications and Networking

the other hand, at intermediate/high SNR values the conven- [17] Y. Jiang, F.-W. Sun, and J. S. Baras, “On the performance limits
tional scheme is preferable as it achieves similar performance of data-aided synchronization,” IEEE Transactions on Informa-
with reduced complexity. tion Theory, vol. 49, no. 1, pp. 191–203, 2003.
[18] H. Meyr, M. Oerder, and A. Polydoros, “On sampling rate,
analog prefiltering, and sufficient statistics for digital re-
ceivers,” IEEE Transactions on Communications, vol. 42, no. 12,
REFERENCES
pp. 3208–3214, 1994.
[19] S. Benedetto, R. Garello, G. Montorsi, et al., “MHOMS: high-
[1] U. Mengali and A. N. D’Andrea, Synchronization Techniques speed ACM modem for satellite applications,” IEEE Wireless
for Digital Receivers, Plenum Press, New York, NY, USA, 1997. Communications, vol. 12, no. 2, pp. 66–77, 2005.
[2] H. Meyr, M. Moeneclaey, and S. Fechtel, Digital Communica- [20] S. M. Kay, Fundamentals of Statistical Signal Processing: Estima-
tion Receivers: Synchronization, Channel Estimation, and Signal tion Theory, Prentice-Hall, Englewood Cliffs, NJ, USA, 1993.
Processing, John Wiley & Sons, New York, NY, USA, 1997.
[3] K. H. Mueller and M. Müller, “Timing recovery in digital syn-
chronous data receivers,” IEEE Transactions on Communica-
tions, vol. 24, no. 5, pp. 516–531, 1976.
[4] F. M. Gardner, “A BPSK/QPSK timing-error detector for sam-
pled receivers,” IEEE Transactions on Communications, vol. 34,
no. 5, pp. 423–429, 1986.
[5] A. N. D’Andrea and M. Luise, “Optimization of symbol timing
recovery for QAM data demodulators,” IEEE Transactions on
Communications, vol. 44, no. 3, pp. 399–406, 1996.
[6] A. A. D’Amico, A. N. D’Andrea, and R. Reggiannini, “Efficient
non-data-aided carrier and clock recovery for satellite DVB at
very low signal-to-noise ratios,” IEEE Journal on Selected Areas
in Communications, vol. 19, no. 12, pp. 2320–2330, 2001.
[7] M. Moeneclaey and G. de Jonghe, “Tracking perfor-
mance comparison of two feedforward ML-oriented carrier-
independent NDA symbol synchronizers,” IEEE Transactions
on Communications, vol. 40, no. 9, pp. 1423–1425, 1992.
[8] M. Oerder and H. Meyr, “Digital filter and square timing re-
covery,” IEEE Transactions on Communications, vol. 36, no. 5,
pp. 605–612, 1988.
[9] E. Panayirci and E. K. Bar-Ness, “A new approach for evaluat-
ing the performance of a symbol timing recovery system em-
ploying a general type of nonlinearity,” IEEE Transactions on
Communications, vol. 44, no. 1, pp. 29–33, 1996.
[10] D. C. Rife and R. R. Boorstyn, “Single-tone parameter estima-
tion from discrete-time observations,” IEEE Transactions on
Information Theory, vol. 20, no. 5, pp. 591–598, 1974.
[11] D.-K. Hong and S.-J. Kang, “Joint frequency offset and car-
rier phase estimation for the return channel for digital video
broadcasting,” IEEE Transactions on Broadcasting, vol. 51,
no. 4, pp. 543–550, 2005.
[12] M. P. Fitz, “Planar filtered techniques for burst mode carrier
synchronization,” in Proceedings of IEEE Global Telecommuni-
cations Conference and Exhibition (GLOBECOM ’91), vol. 1,
pp. 365–369, Phoenix, Ariz, USA, December 1991.
[13] B. C. Lovell and R. C. Williamson, “The statistical perfor-
mance of some instantaneous frequency estimators,” IEEE
Transactions on Signal Processing, vol. 40, no. 7, pp. 1708–1723,
1992.
[14] M. Luise and R. Reggiannini, “Carrier frequency recovery
in all-digital modems for burst-mode transmissions,” IEEE
Transactions on Communications, vol. 43, no. 234, pp. 1169–
1178, 1995.
[15] U. Mengali and M. Morelli, “Data-aided frequency estimation
for burst digital transmission,” IEEE Transactions on Commu-
nications, vol. 45, no. 1, pp. 23–25, 1997.
[16] Y. Fan and P. Chakravarthi, “Joint carrier phase and symbol
timing synchronization for burst satellite communications,” in
Proceedings of the 21st Century Military Communications Con-
ference (MILCOM ’00), vol. 2, pp. 1104–1108, Los Angeles,
Calif, USA, October 2000.
Hindawi Publishing Corporation
EURASIP Journal on Wireless Communications and Networking
Volume 2007, Article ID 29086, 12 pages
doi:10.1155/2007/29086

Research Article
Burst Format Design for Optimum Joint
Estimation of Doppler-Shift and Doppler-Rate
in Packet Satellite Communications

Luca Giugno,1 Francesca Zanier,2 and Marco Luise2

1 Wiser S.r.l.–Wireless Systems Engineering and Research, Via Fiume 23, 57123 Livorno, Italy
2 Dipartimento di Ingegneria dell’Informazione, University of Pisa, Via Caruso 16, 56122 Pisa, Italy

Received 1 September 2006; Accepted 10 February 2007

Recommended by Anton Donner

This paper considers the problem of optimizing the burst format of packet transmission to perform enhanced-accuracy estimation
of Doppler-shift and Doppler-rate of the carrier of the received signal, due to relative motion between the transmitter and the
receiver. Two novel burst formats that minimize the Doppler-shift and the Doppler-rate Cramér-Rao bounds (CRBs) for the joint
estimation of carrier phase/Doppler-shift and of the Doppler-rate are derived, and a data-aided (DA) estimation algorithm suitable
for each optimal burst format is presented. Performance of the newly derived estimators is evaluated by analysis and by simulation,
showing that such algorithms attain their relevant CRBs with very low complexity, so that they can be directly embedded into new-
generation digital modems for satellite communications at low SNR.

Copyright © 2007 Luca Giugno et al. This is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

1. INTRODUCTION
Packet transmission of digital data is nowadays adopted
in several wireless communications systems such as satel-
lite time-division multiple access (TDMA) and terrestrial
mobile cellular radio. In those scenarios, the received sig-
nal may suffer from significant time-varying Doppler dis-
tortion due to relative motion between the transmitter and
the receiver. This occurs, for instance, in the last-generation
mobile-satellite communication systems based on a con-
stellation of nongeostationary low-earth-orbit (LEO) satel-
lites [1] and in millimeter-wave mobile communications for
traffic control and assistance [2]. In such situations, car-
rier Doppler-shift and Doppler-rate estimation must be per-
formed at the receiver for correct demodulation of the re-
ceived signal.
A number of efficient digital signal processing (DSP) al-
gorithms have already been developed for the estimation of
the Doppler-shift affecting the received carrier [3] and a few
algorithms for Doppler-rate estimation are also available in
the open literature [4, 5]. The issue of joint Doppler-shift
and Doppler-rate estimation has been addressed as well, al-
though to a lesser extent [6, 7]. In all the papers above, the
observed signal is either an unmodulated carrier, or con-
tains pilot symbols known at the receiver. The most common
burst format is the conventional preamble-payload arrange-
ment, wherein all pilots are consecutive and they are placed
at the beginning of the data burst. Other formats are the mi-
damble as in the GSM system [8], wherein the preamble is
moved to the center of the burst, or the so-called pilot sym-
bol assisted modulation (PSAM) paradigm [9], where the
set of pilot symbols is regularly multiplexed with data sym-
bols in a given ratio (the so-called burst overhead). Data-
aided (DA) algorithms, which exploit the information con-
tained in the pilot symbols, are routinely used to attain good
performance with small burst overhead. The recent intro-
duction of efficient channel coding with iterative detection
[10] has also placed new and more stringent requirements
for receiver synchronization on satellite modems. The car-
rier synchronizer is requested to operate at a lower signal-to-
noise ratio (SNR) than it used to be with conventional coding
[11].
Therefore, it makes sense to search for the ultimate ac-
curacy that can be attained by carrier synchronizers. It turns
out that the Cramér-Rao bounds (CRBs) for joint estima-
tions are functions of the location of the reference symbols
in the burst. The issue to find the optimal burst format that
minimizes the frequency CRB has been already addressed in
2 EURASIP Journal on Wireless Communications and Networking

-2P format- -4P format-

N/2 M N/2 N/4 M/3 N/4 M/3 N/4 M/3 N/4

a Preamble Payload Postamble c P Payload P Payload P Payload P

M + N/2 2M/3 + N/2

-3P format-
N/3 M/2 N/3 M/2 N/3 -1st 2P subburst- -2nd 2P subburst-
b P Payload P Payload P d P Payload P P Payload P

L L

Figure 1: 2P burst format, 3P burst format, and 4P burst format.

[12–14], but only for joint carrier phase/Doppler-shift es-
timation. The novelty of the paper is to extend the anal-
ysis to the joint carrier phase/Doppler-shift and Doppler-
rate estimation. It is known [12–15] that the preamble-
postamble format (2P format) described in the sequel min-
imizes the frequency CRB with no Doppler-rate, and with
constraints on the total training block length and on the
burst overhead of the signal. We demonstrate here that such
format is optimal in the presence of Doppler-rate as well,
and that the Doppler-rate CRB is minimized by estima-
tion over three equal-length blocks of reference symbols that
are equally spaced by data symbols (3P format). We also
show that other formats are very close to optimality (4P for-
mat).
In addition to computation of the burst, we also in-
troduce new high-resolution and low-complexity carrier
Doppler-shift and Doppler-rate DA estimation algorithms
for such optimal burst formats.
The paper is organized as follows. In Section 2, we
first outline the received signal model affected by Doppler
distortions. Next, in Section 3 we present and analyze a
low-complexity DA Doppler-shift estimator for the optimal
2P format. Extensions of this algorithm for joint carrier
phase/Doppler-shift and Doppler-rate estimation for the 2P
format, the 3P format, and the sub-optimum 4P format, are
introduced in Sections 4 and 5, respectively. Finally, some
conclusions are drawn in Section 6.
2. SIGNAL MODEL
In this paper, we take into consideration three different data
burst formats as depicted in Figure 1.
In all cases, the total number of pilot symbols that are
known to the receiver is equal to N, and the total length of
the “data payload” fields that contain information symbols is
equal to M. The formats differ for the specific pilots arrange-
ment in two/three/four groups of N/2, N/3, N/4 consecutive
pilot symbols equally spaced by data symbols. Hereafter we
will address them as “2P,” “3P,” “4P” formats as in Figures
1(a), 1(b), 1(c), respectively. We denote also with L = N + M
the overall burst length, and with η the burst overhead, that
is, the ratio between the total number of pilot symbols and
the total number of symbols within the burst:
N N 1
η= = = . (1)
L N +M 1 + M/N
We also assume BPSK/QPSK data modulation for the pilot
fields, and additive white Gaussian noise (AWGN) channel
with no multipath. Filtering is evenly split between transmit-
ter and receiver, and the overall channel response is Nyquist.
Timing recovery is ideal but the received signal is affected by
time-varying Doppler distortion. Filtering the received wave-
form with a matched filter and sampling at symbol rate at
the zero intersymbol interference instants yields the follow-
ing discrete-time signal:
L−1 L−1
z(k) = ck e jϕk + n(k), k=− , . . . , 0, . . . , ,
2 2
(2)
where
ϕk = θ + 2πνkT + παk2 T 2 (3)
is the instantaneous carrier excess phase, {ck } are unit-energy
(QPSK) data symbols and L (odd) is the observation (burst)
length. Also, 1/T is the symbol rate, θ is the unknown initial
carrier phase, ν is the constant unknown carrier frequency
offset (Doppler-shift), and finally α is the constant unknown
carrier frequency rate-of-change (Doppler-rate). For signal
model (2) to be valid, we assumed that the value of the
Doppler-shift ν is much smaller than the symbol rate, and
that the value of the Doppler-rate α is much smaller than
the square of the symbol rate. The noise n(k) is a complex-
valued zero-mean WGN process with independent compo-
nents, each with variance σ 2 = N0 /(2Es ), where Es /N0 repre-
sents the ratio between the received energy-per-symbol and
the one-sided channel noise power spectral density.
Estimation of ν and α from the received signal z(k) re-
quires preliminary modulation removal from the pilot fields.
Broadly speaking, it is customary to adopt BPSK or QPSK
modulation for pilot fields, so that modulation removal is
easily carried out by letting r(k) = ck∗ z(k). The result is
 
r(k) = e jϕk + w(k), k∈K = NPi , (4)
Luca Giugno et al.
where K is the symmetric set of N time indices correspond-
ing to pilot symbols, and w(k) = ck∗ n(k) is statistically equiv-
alent to n(k). We explicitly mention here that we have cho-
sen a symmetrical range K with respect to the middle of
the burst since such arrangement decouples the estimation
of some parameters, as discussed in [12] and in Appendix B.
The signal r(k) will be considered from now on as our ob-
served signal that allows to carry out the carrier synchro-
nization functions. We show in Appendix B that the burst
formats in Figure 1 are optimum so far as the estimation of
parameters ν and α is concerned. To keep complexity low, we
will not take into consideration here “mixed,” partially blind,
methods to perform carrier synchronization that use both the
known pilot symbols and all of the intermediate data sym-
bols of the burst, like envisaged in [16] for the case of channel
estimation.
3. DOPPLER-SHIFT ESTIMATOR: FEPE ALGORITHM
We momentarily neglect the effect of the Doppler-rate α in
(4), to concentrate on the issue of Doppler-shift estimation
only. Under such hypothesis, (4) can be rewritten as follows:
r(k) = e j(θ+2πνkT) + w(k), k ∈ K.
The 2P format minimizes the CRB for Doppler-shift esti-
mation for joint carrier phase/Doppler-shift estimation [12–
15]. Conventional frequency offset estimators for consecu-
tive signal samples [3] are not directly applicable to a burst
format encompassing a preamble and a postamble. In addi-
tion, straightforward solution of a maximum-likelihood es-
timation problem for ν appears infeasible. We introduce thus
a new low-complexity algorithm suitable for the estimation
of the Doppler-shift ν in (4) with the burst format as above.
The key idea of the 2P frequency estimator is really a naive
one: we start by computing two phase estimates, the one on
the preamble section, and the other on the postamble, us-
ing the standard low-complexity maximum-likelihood (ML)
algorithm [17]:
 −(M −1)/2   (N+M −1)/2
 
θ1= arg r(k) , θ2 = arg r(k) ,
k=−(N+M −1)/2 k=(M −1)/2
where arg{·} denotes the phase of the complex-valued ar-
gument. Then we associate the two phase estimates to the
two midpoints of the preamble and postamble sections, re-
spectively, whose time distance is equal to (M + N/2)T
(Figure 1(a)). After this is done, we simply derive the fre-
quency estimate as the slope of the line that connects the two
points (−(M − 1)/2 − N/4, θ1 ) and ((M − 1)/2 + N/4, θ2 ) on
the (time, phase) plane:
θ2 − θ1 2π 2π
ν=
 2π
. (7)
2π(M + N/2)T
This simple algorithm is known as frequency estimation
through phase estimation (FEPE) [15]. The operator |x|2π re-
turns the value of x modulo 2π, in order to avoid phase am-
biguities, and is trivial to implement when operating with
3
×10−3
1.5
1
0.5
MEV
0
−0.5
−1
−1.5
−1.5 −1 −0.5 0 0.5 1 1.5
νT (Hz × s) ×10−3
Ideal Es /N0 = 20 dB
Es /N0 = 0 dB Es /N0 = 100 dB
Es /N0 = 10 dB
Figure 2: MEV of FEPE estimator for different values of ES /N0 —
simulation only. Preamble + postamble DA ML phase estimation,
N = 44, M = 385.
(5)
fixed-point arithmetic on a digital hardware. It is easy to ver-
ify that such estimator is independent of the particular ini-
tial phase θ, that vanishes when computing the phase dif-
ference at the numerator of (7). It is also clear that the
operating range of the estimator is quite narrow. In order
not to have estimation ambiguities, we have to ensure that
−π ≤ |θ2 |2π − |θ1 |2π < π, and therefore the range is bounded
to
1
|ν | ≤ . (8)
2(M + N/2)T
This relatively narrow interval does not allow to use the FEPE
algorithm for initial acquisition of a large frequency offset at
 receiver start-up. Its use is therefore restricted to fine esti-
mation of a residual offset after a coarse acquisition or com-
pensation of motion-induced Doppler-shift. Figure 2 depicts
(6) the normalized mean estimated value (MEV) curves of the
FEPE algorithm (i.e., the average estimated value E{ν} as a
function of the true Doppler-shift ν) for different values of
Es /N0 as derived by simulation. In our simulations we use
the values N = 44 and M = 385 taken from the design de-
scribed in [11], so that the overhead is η = 10% (typical for
short bursts). MEV curves show that the algorithm is unbi-
ased in a broad range around the true value (here, ν = 0). It
can be shown that this is true as long as ν2NT  1, so that
the “ancillary” estimates θ2 and θ1 are substantially unbiased
as well. Such condition is implicitly assumed in (8) since in
the practice M  N/2. The curve labeled Es /N0 = 100 dB
(which is totally unrealistic) has the only purpose of showing
the bounds of the unambiguous estimation range.
It is also easy to evaluate the estimation error variance of
the FEPE estimator. It is known in fact that θ1 and θ2 in (7)
have an estimation variance σθ2 that achieves the Cramér-Rao
4 EURASIP Journal on Wireless Communications and Networking

Bound (CRB) [17]: 10−2

1 1
σθ2 = CRB(θ) = . (9)
2 · N/2 Es /N0
Therefore, considering that the two phase estimates in (7) are
independent, we get
10−3
2 · σθ2 1 1
2
σFEPE (ν) = = .
4π 2 (M + N/2)2 T 2 4π 2 T 2 N/2(M + N/2)2 Es /N0

Normalized RMSEE
(10) VCRB (ν)

The vector CRB [18] for the frequency offset estimate in the
10−4
joint carrier phase/Doppler-shift estimation with the 2P for-
mat is derived in Appendix A and reads as follows:
3 1
VCRB2P (ν) =
 .
4π 2 T 2 (N/2) 4(N/2)2 + 3M 2 +3MN − 1 Es /N0
(11) VCRB2P (ν)
10−5
Both from the expression of the bound (11) and of the
variance (10), it is seen that the estimation accuracy has an
inverse dependence on (N/2)3 , and this is nothing new with
respect to conventional estimation on a preamble only. The
important thing is that we also have inverse dependence on
M 2 , due to the 2P format that gives enhanced accuracy (with 10−6
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
small estimation complexity) with respect to the conven-
Es /N0 (dB)
tional estimator. From (1), we also have M = N(1/η − 1),
so that the term 3M 2 dominates (N/2)2 as long as η < 1/2, 2
σFEPE (ν)
which is always verified in the practice. νT = 1 × 10−3
Therefore, the ratio between the CRB (11) and the vari- νT = 1 × 10−4
ance of the FEPE estimator is very close to 1. With N = 44
2
and M = 385, we get, for instance, σFEPE /VCRB2p = 0.99. Figure 3: RMSEE of FEPE estimator for different values of
The enhanced-accuracy feature is also apparent in the com- ES /N0 and relevant bounds—solid lines: theory—marks: simula-
parison of the VCRB2p (ν) as in (11) with the conventional tion. Preamble + postamble DA ML phase estimation, N = 44,
M = 385.
VCRB(ν) [18] for frequency estimation on a single preamble
with length N, that is obtained by letting M = 0 in (11). The
reverse of the coin is of course the reduced operating range
(8) of the estimator. 13] is presented in Table 1. It is clear that the strength of the
Figure 3 shows curves of the (symbol-rate-normalized) FEPE algorithm is its very low complexity as compared to
RMSEE (root mean  square estimation error) of the FEPE
algorithm (i.e., T E{(v − v)2 }) as a function of Es /N0 for
various values of the true offset ν. In particular, marks are
2
simulation results for σFEPE , whilst the lowermost line is the
VCRB2p (ν). We do not report the curve for (10) since it
would be totally overlapped with (11).
Performance assessment of the FEPE estimator is con-
cluded in Figure 4 with the evaluation of the sensitivity of the
RMSEE to different values of an uncompensated Doppler-
rate α. Just to have an idea of practical values of αT 2 to be en-
countered in practice, we mention that the largest Doppler-
rates in LEO satellites are of the order of 200 Hz/s [1, 19] for
a carrier frequency of 2.2 GHz, and assuming a symbol rate
of 2 Mbaud, we end up with the value αT 2 = 5.10−11 . From
simulation results, we highlight that the performance of this
algorithm is affected by α, but only in the case of a normal-
ized Doppler-rate αT 2 ≥ 10−7 , that is larger than those that
are found in the practice.
Finally, the complexity of the FEPE estimator with re-
spect to conventional methods of frequency estimation [3,
conventional algorithms.
4. DOPPLER-RATE ESTIMATORS IN 2P FRAME:
FREPE AND FREFE ALGORITHMS
We take now back into consideration the presence of a non-
negligible Doppler-rate in the received signal, modeled as in
(3)-(4). We focus again on the 2P format (Figure 1(a)), since
it is the optimal format for Doppler-shift estimation in joint
carrier phase/Doppler-shift and Doppler-rate estimation too,
as demonstrated in Appendix B. A new simple estimator for
α in the 2P format is found by a straightforward general-
ization of the FEPE approach. Assume that we further split
both the preamble and the postamble into two subsections of
equal length, and we compute four (independent) ML phase
estimates on the two subsections. We know in advance that
the time evolution of the phase is described by a parabola.
The four phase estimates can thus be used to fit a second-
order phase polynomial according to the Minimum Mean
Squared Error (MMSE) criterion; taking the origin in the
Luca Giugno et al. 5

10−3 Table 1: The FEPE computational complexity comparison.

7 (Nalg = estimation design parameter.)
6
5 Computational complexity of major
4
Doppler-shift estimation algorithms
3
Number of real products
2 Algorithm Reference LUT access
and additions
 
L&R [3] 4N Nalg + 1 − 2 1
10−4  
M&M [3] Nalg 8N − 4Nalg − 3 − 2 Nalg
Normalized RMSEE

7 S-BLUE [13] 4N 2 + 4.5N − 3 1.5N − 2

6
5 P-BLUE-2 [13] 4N − 1 1
4
3 FEPE — 2N + 3 2

ε(a, b, c) with respect to a, b, and c, we obtain

10−5
7  2
∂ε(a, b, c) 
4
6 M − 1 3N
5 = e i · ni + + = 0,
4 ∂a i=1
2 8
3  
∂ε(a, b, c) 
4
M − 1 3N
2 = e i · ni + + = 0, (14)
∂b i=1
2 8
∂ε(a, b, c) 
4
10−6
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 = ei = 0,
∂c i=1
Es /N0 (dB)

VCRB2P (ν) αT 2 = 1 × 10−7 and solving for a we get the following so-called frequency rate
αT 2 = 0 αT 2 = 2 × 10−7
αT 2 = 2 × 10−8
estimation through phase estimation (FREPE) algorithm [15]:
αT 2 = 2.5 × 10−7
αT 2 = 5 × 10−8
   
a θ4 − θ3 − θ2 − θ1
αFREPE = 2 = (15)
Figure 4: Sensitivity of FEPE estimator to different values of the T πN/2(N/2 + M)T 2
Doppler-rate αT 2 . Preamble + postamble DA ML phase estimation,
N = 44, M = 385, vT = 1.0 × 10−3 . (all differences to be intended modulo-2π). This extremely
simple approach can be viewed as a generalization of the
FEPE introduced in the previous section. In particular, by us-
first section of the preamble, we obtain the phase model ing (7), the terms

  
M − 1 3N 2 θi − θi−1
ϕP (n) = aπ n + + , i = 2, 4, (16)
2 8 2π(N/4)T
 (12)
M − 1 3N
+ 2πb n + + + c, represent two Doppler-shift estimations, the first on the
2 8
preamble and the second on the postamble, respectively,
where the regression coefficients a and b directly repre- which are spaced M + N/2 symbols apart. The Doppler-rate
sent estimates for the (normalized) carrier Doppler-rate and estimate is thus simply the difference between the two fre-
Doppler-shift, respectively, and c is an estimate for the initial quency estimates, divided by their time distance (M +N/2)T.
phase (that we are not interested into). The coefficients are The considerations above allow us to also introduce
found after observing that the MSE is written as the frequency rate estimation through frequency estimation
(FREFE) algorithm [15]

4

  2 
4
ε(a, b, c) = ϕP ni − θi = ei2 , (13) ν2 − ν1

i=1 i=1
αFREFE = , (17)
(M + N/2)T

where θi , i = 1, . . . , 4, are the above-mentioned ML phase
estimates on N/4 pilots each, and n1 = −[(M − 1)/2 + 3N/8],
n2 = −[(M − 1)/2 + N/8], n3 = [(M − 1)/2 + N/8], and
n4 = [(M − 1)/2 + 3N/8] are the four time instants that we
conventionally associate to the four estimates (the midpoints
of the four subsections). Equating to zero the derivatives of
wherein the two frequency estimates ν1 and ν2 can be ob-
tained by any conventional algorithm [3] operating sepa-
rately on the preamble and on the postamble, respectively.
We can choose for instance the L&R algorithm [20] or the
R&B algorithm [21]. Assuming that the selected algorithm
operates close enough to the CRB (as is shown in [3]), the
6 EURASIP Journal on Wireless Communications and Networking

variance of (17) is ×10−4

1.5
2 2σν2
σFREFE (α) = 1
(M + N/2)2 T 2
3 1
=   , 0.5
π 2 T 4 N/2 (N/2)2 − 1 (M + N/2)2 Es /N0

MEV
(18) 0

where we have used σν2 = 3 · (Es /N0 )−1 /[2π 2 T 2 N/2((N/2)2 − −0.5
1)] [17]. This can be compared to the variance of the FREPE
algorithm that is easily found to be −1

4 · σθ2 −1.5
2 −1.5 −1 −0.5 0 0.5 1 1.5
σFREPE (α) = ×10−4
π 2 (N/2)2 (M + N/2)2 T 4 αT 2 (Hz/s × s2 )
(19)
4 1 Ideal Es /N0 = 20 dB
= 2 4 ,
π T (N/2)3 (M + N/2)2 Es /N0 Es /N0 = 0 dB Es /N0 = 100 dB
Es /N0 = 10 dB
where now σθ2 = (Es /N0 )−1 /(N/2). The relevant vector CRB
for Doppler-rate estimate is (see Appendix B): Figure 5: MEV of FREPE estimator for different values of ES /N0 —
VCRB2P (α) simulation only. Preamble + postamble DA ML phase estimation,
N = 44, M = 385, vT = 1.0 × 10−3 .
45 1
=    .
π2T 4 (N/2)3− N/2 16(N/2) +15M +30MN/2 − 4 Es /N0
2 2

(20) ×10−4
1.5
All expressions inversely depend on (N/2)5 as in conven-
tional preamble-only estimation of the Doppler-rate [6], but 1
they also bear again inverse dependence on M 2 that gives en-
0.5
hanced accuracy. For sufficiently large values of N and M,
MEV

M  N, we have 0
2
σFREFE (α) ∼ 3 VCRBPP (α) ∼ −0.5
2 = , 2 = 1. (21)
σFREPE (α) 4 σFREFE (α)
−1
Figure 5 shows the MEV curves (i.e., E{ α}) of the FREPE al-
gorithm for different values of Es /N0 , in the case of N = 44, −1.5
M = 385, and Doppler-shift vT = 10−3 . The estimator is −1.5 −1 −0.5 0 0.5 1 1.5
αT 2 (Hz/s × s2 ) ×10−4
unbiased with an operating range equal to
1 Ideal νT = 5 × 10−3
αFREPE ≤ . (22) νT = 0 νT = 1 × 10−2
N/2(M + N/2)T 2
νT = 1 × 10−3
The sensitivity of FREPE to different uncompensated val-
ues of vT is illustrated in Figure 6 in terms of MEV. Figure 6: MEV of FREPE estimator for different values of the
The same simulations have been run also for the FREFE Doppler-shift vT—simulation only. Preamble + postamble DA ML
algorithm. In particular, Figure 7 illustrates the MEV curves phase estimation, N = 44, M = 385, Es /N0 = 10 dB.
for different values of Es /N0 and with vT = 10−3 . By using
the L&R algorithm to estimate ν1 and ν2 , the operating range
of FREFE is roughly twice that of FREPE:
5. OPTIMUM DOPPLER-RATE ESTIMATION
1
αFREFE ≤ . (23)
(N/4 + 1)(M + N/2)T 2 5.1. Odd number of pilot fields: FRE-3PE algorithm

In particular, the term [(N/2 + 1)T]−1 represents the fre-
quency pull-in range of L&R on N/2 pilots [20].
Figure 8 demonstrates that FREPE is also less sensitive
than FREFE to an uncompensated Doppler-shift. Finally,
Figure 9 shows the curve of the Doppler-rate RMSEE of
FREPE and FREFE as a function of Es /N0 , for νT = 10−3 and
αT 2 = 10−6 . The FREPE estimator loses only 10 log10 (4/3) =
1.25 dB in terms of Es /N0 with respect to the performance of
the more complex FREFE when N  1.
We turn now to the issue of optimum burst configuration
for the estimation of the Doppler-rate. We demonstrate in
Appendix B that the 3P format (Figure 1(b)) minimizes the
CRB for Doppler-rate estimation, with the usual constraints
on the total training block length and on the burst over-
head (1). In the following, we develop a new low-complexity
algorithm suitable for Doppler-rate estimation with the 3P
format. We know in advance that the time evolution of the
phase is described by a parabola. As was done for the FREPE
Luca Giugno et al. 7

×10−4 10−3
2.5
2
1.5
1
10−4
0.5
MEV

0
−0.5
−1 VCRBP (α)
−1.5

Normalized RMSEE
10−5
−2
−2.5
−2.5 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2 2.5
αT 2 (Hz/s × s2 ) ×10−4

10−6
Ideal Es /N0 = 20 dB VCRB2P (α)
Es /N0 = 0 dB Es /N0 = 100 dB
Es /N0 = 10 dB

Figure 7: MEV of FREFE estimator for different values of ES /N0 — 10−7 VCRB4P (α)
simulation only. Preamble + postamble Luise and Reggiannini, N =
VCRB3P (α)
44, M = 385, vT = 1.0 × 10−3 .

×10−4
2.5 10−8
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
2
Es /N0 (dB)
1.5
1 FREPE, αT 2 = 1 × 10−6 FRE-2FEPE, αT 2 = 1 × 10−6
0.5 FREFE, αT 2 = 1 × 10−6 FRE-3PE, αT 2 = 1 × 10−6
MEV

0
−0.5 Figure 9: RMSEE of FREPE, FREFE, FRE-3PE, and FRE-2FREPE
−1 estimators for different values of ES /N0 and relevant bounds,—solid
−1.5
lines: theory—marks: simulation. Doppler-rate algorithms: FREFE
versus FREPE versus FRE-3PE versus FRE-2FEPE, N = 44(45),
−2
M = 385(384), vT = 1.0 × 10−3 .
−2.5
−2.5 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2 2.5
αT 2 (Hz/s × s2 ) ×10−4

where θi are the above-mentioned ML phase estimates on

Ideal νT = 5 × 10−3
νT = 0 νT = 1 × 10−2
N/3 pilots each, and where n1 = −(M/2 + N/3), n2 = 0, and
νT = 1 × 10−3 n3 = (M/2 + N/3) are the three time instants that we con-
ventionally associate to the three estimates (the midpoints of
Figure 8: MEV of FREFE estimator for different values of the the three subsections). Solving for a, we get the following so-
Doppler-shift vT—simulation only. FREFE estimator preamble + called (FRE-3PE) (frequency rate estimation through 3 phase
postamble Luise and Reggiannini, N = 44, M = 385, Es /N0 = estimations) algorithm:
10 dB.

   
a 18 θ3 − θ2 − θ2 − θ1
αFRE-3PE = = (26)
T 2 π(2N + 3M − 2)2 T 2
algorithm in the 2P configuration, a simple estimator of α
in the 3P format is found by computing three (independent) (all differences to be intended modulo-2π). The estimator is
ML phase estimates on the three blocks of pilots, and then unbiased with an operating range equal to:
fitting a second-order phase polynomial. Taking the origin in
18
the first block of pilots, we obtain this time the phase model αFRE-3PE ≤ . (27)
(2N + 3M − 2)2 T 2
2 
N M N M
ϕP (n) = aπ n + + + 2πb n + + + c. (24) In our simulations (N = 45 and M = 384), |αFRE-3PE · T 2 | ≤
3 2 3 2
10−5 . This range is narrower than FREPE’s and FREFE’s in
The coefficients are found solving the following set of equa- the 2P format, but it still widely includes practical Doppler-
tions: rate values mentioned in Section 3. Figure 10 shows the MEV
  curves of the FRE-3PE algorithm for different values of
ϕP ni = θi , i = 1, . . . , 3, (25) Es /N0 , in the case of N = 45, M = 384, and Doppler-shift
8 EURASIP Journal on Wireless Communications and Networking

×10−5 ×10−5
1.5 1.5

1 1

0.5 0.5
MEV

MEV
0 0

−0.5 −0.5

−1 −1

−1.5 −1.5
−1.5 −1 −0.5 0 0.5 1 1.5 −1.5 −1 −0.5 0 0.5 1 1.5
αT 2 (Hz/s × s2 ) ×10−5 ×10−5
αT 2 (Hz/s × s2 )

Ideal Es /N0 = 20 dB Ideal νT = 5 × 10−4

Es /N0 = 0 dB Es /N0 = 100 dB νT = 0 νT = 1 × 10−3
Es /N0 = 10 dB νT = 1 × 10−4

Figure 10: MEV of FRE-3PE estimator for different values of Figure 11: MEV of FRE-3PE estimator for different values of the
ES /N0 —simulation only. 3 blocks of pilots DA ML phase estima- Doppler-shift vT—simulation only. 3 blocks of pilots, N = 45, M =
tion, N = 45, M = 384, vT = 1.0 × 10−3 . 384, ES /N0 = 10 dB.

5.2. Even number of pilot fields: FRE-2FEPE algorithm

vT = 10−4 , while Figure 11 shows the sensitivity of the MEV
to different uncompensated values of the Doppler-shift vT.
The theoretical error variance of the FRE-3PE estimator
can be easily evaluated, similarly to what was done for the
2
calculation of σFREFE (α) in Section 4:
2
182 · 6 · σθ2
σFRE-3PE (α) =
π 2 (2N + 3M − 2)4 T 4
(28)
182 · 6 1
= 2 4 , and postamble is now of length N/4, and we can derive two
π T (2N/3)(2N + 3M − 2)4 Es /N0
where now σθ2 = (Es /N0 )−1 /(2N/3). Comparing this expres-
sion with the VCRB3P (α) in (B.11) and with the variances of
the FREFE and FREPE algorithms, we note that all expres-
sions inversely depend on N 5 as in conventional preamble-
only estimation of the Doppler-rate [6]. On the other hand,
2
σFRE-3PE (α) and VCRB3P (α) inversely depend on M 4 , out-
performing the accuracy of both the traditional preamble-
only format and the 2P format (that depends on M −2 ). The
enhanced accuracy is highlighted by Figure 9, where we re-
port the simulated RMSEE (marks) of FRE-3PE, FREPE, and
FREFE versus Es /N0 . To perform a fair comparison, we also
reported the VCRBP (β), obtained in the case of estimation of
Doppler-rate in the preamble-only configuration. The FRE-
3PE algorithm attains its own CRB, and exhibits a gain of
19 dB in terms of Es /N0 with respect to the 2P format.
As a final remark, we only mention that a simple estima-
tor of Doppler-shift in the 3P format is found by applying the
FEPE algorithm to the two extreme pilot fields of the burst.
Its variance reaches the VCRB3P (ν) calculated setting x = 1
in (B.7) and (B.9), that is 1.5 dB apart from the VCRB2P (ν)
of the optimal 2P format.
When the number of pilot fields is even, the optimum burst
format turns out to be the 4P as shown in Appendix B.
We notice that the ratio of the two bounds for 3P and
4P amounts to VCRB4p (α)/VCRB3p (α) = ∼ 9720/108 · 640/
∼
51840 = 1.09 M  N, so that 4P is only slightly optimal.
A simple estimator of α in the 4P format is found by a
straightforward generalization of the FEPE and FREFE ap-
proaches. Assume that we split the burst into two 2P sub-
bursts of length (M/3 + N/2), (Figure 1(d)). Each preamble
FEPE estimates of frequency on each subburst:
θ2 − θ1 2π 2π θ4 − θ3 2π 2π
ν1 =
 2π
, ν2 =
 2π
,
2π(M/3 + N/4)T 2π(M/3 + N/4)T
(29)
where θi , i = 1, . . . , 4, are the ML phase estimates computed
on the four pilot fields of N/4 pilots each. The two Doppler-
shift estimates ν1 and ν2 are associated with the two mid-
point instants of the two 2P subbursts, whose time distance
is equal to (2M/3 + N/2)T (Figure 1(c)). Again, we estimate
the Doppler-rate as the slope of the line that connects the two
points (−(M/3 − 1/2) − N/4, ν1 ) and ((M/3 − 1/2) + N/4, ν2 )
in the (time, frequency) plane:
ν2 − ν1

αFRE-2FEPE = . (30)
(2M/3 + N/2)T
We call this algorithm FRE-2FEPE (frequency rate estimation
through two FEPE estimations) .
It is clear that the operating range of the estimator with
respect to ν comes from the application of (8) to the new
configuration and turns out to be |ν| ≤ [2(M/3 + N/4)T]−1 .
The MEV curves of FRE-2FEPE are not reported here since
Luca Giugno et al. 9

they basically mimic those in Figures 10 and 11 for the [12]. After modulation removal, the generic sample within
FRE-3PE algorithm. The estimation error variance of (30) the preamble and the postamble is given by (5).
is found to be The Fisher information matrix (FIM) [18] can be written
as
σθ2  
2
σFRE-2FEPE (α) = Fθθ Fθν
(2M/3 + N/2)2 (M/3 + N/4)2 π 2 T 2 F=
 −1 Fνθ Fνν
2 · Es /N0 ⎡    ⎤
= . ∂2 ln p(r |  
ν, θ) ∂2 ln p(r |  
ν, θ)

π 2 T 4 N(2M/3 + N/2)2 (M/3 + N/4)2 ⎢−Er

−Er

⎥
⎥
⎢ ∂ 
θ 2 ∂ 
θ∂ ν

(31) =⎢
⎢ 



 

⎥
⎥,
⎣ ∂ ln p(r | 
2 
ν, θ) ∂ ln p(r | 
2 
ν, θ) ⎦
−Er −Er
Figure 9 shows also the curves of the RMSEE of FRE-2FEPE ∂ν∂θ ∂ν2
and its respective CRB. The FRE-2FEPE algorithm reaches its (A.1)
own VCRB4p (α) and thus, as demonstrated in Appendix B, it
gains 10 log10 (7.19) = 18.5 dB in terms of Es /N0 with respect  is the probability density function of r =
where p(r | ν, θ)
to the performance of the previous algorithms with the 2P {r(k)}, k ∈ K, conditioned on (  and r(k) is a random
ν, θ),
format. Also, the FRE-2FEPE loses only 0.4 dB with respect Gaussian variable with variance equal to σ 2 = N0 /(2Es ) and
to the FRE-3PE algorithm and can thus be a valid alternative mean value equal to
to the 3P format. 
As a final remark, we briefly address the issue of Doppler- s(k) = e j(θ+2π νkT) .
 (A.2)
shift estimation in the 4P format. The best method is found
 as
Therefore, we write p(r | ν, θ)
by applying the FEPE algorithm to the two extreme pilot
fields of the burst. Its variance is close to the VCRB4P (ν) cal-   
 =
p(r | ν, θ) p rk | ν, θ
culated setting x = 1 in (B.8) and (B.9), that is 2.4 dB worse k∈K
than the VCRB2P (ν) of the optimal 2P format.  
1 1  r(k) − 
2
=  N exp − s(k) .
6. CONCLUSIONS 2πσ 2 2σ 2 k∈K
(A.3)
In this paper, we presented and analyzed some very-
low-complexity algorithms for carrier Doppler-shift and Taking the logarithm of (A.3), we obtain
Doppler-rate estimation in burst digital transmission. To 
ln p(r | ν, θ)
achieve enhanced accuracy, the burst configurations that

1 1 

r(k) 2 + 
2
minimize the CRB for the estimation of Doppler-shift and = N ln − s(k)
Doppler-rate are derived. Our analysis showed that the 2P 2πσ 2 2σ 2 k∈K
format is optimum for Doppler-shift estimation and that the  
s∗ (k)
− 2Re r(k)
3P format is optimum for Doppler-rate estimation. These
1   
two configurations can be practically thought as repetition of =C + Re r(k)s∗ (k) ,
two/three consecutive conventional (preamble-only) bursts. σ 2 k∈K
Despite preventing from real-time processing of the data pay- (A.4)
load section, the 2P and 3P formats greatly outperform the
where C is a constant term that includes all the quantities
estimation based on conventional preamble-only pilot dis-  After differentiating twice (A.4) with
tribution. Performance assessment has shown that all of the independent of ν and θ.

respect to ν and θ, calculating the expectation of the various
proposed algorithms are unbiased in practical operating con-
ditions, and that their accuracy in terms of estimation vari- terms with respect to r, we get
ance gets remarkably close to their respective CRBs down to  
a b
very low Es /N0 values. F= , (A.5)
c d
APPENDICES where

1 
 ∗ 
A. VCRB FOR JOINT CARRIER PHASE/DOPPLER-SHIFT a = (1)Er Re r(k)s (k) ,
ESTIMATION WITH 2P FORMAT σ2 k∈K

1 
 ∗ 
In this appendix, we calculate the VCRB for the error vari- b = (2πTk)Er Re r(k)s (k) ,
σ2 k∈K
ance of any unbiased estimator of Doppler-shift in the case of  (A.6)
joint estimation of phase/Doppler-shift using the preamble- 1 
 ∗ 
c = (2πTk)Er Re r(k)s (k) ,
postamble (2P) format. We explicitly mention that we have σ2 k∈K
chosen a set K of pilot locations that is symmetrical with 
1  
 ∗ 
respect to the middle of the burst, since a symmetrical K de- d = 4π 2 T 2 k2 Er Re r(k)s (k) .
σ2
couples phase from Doppler-shift estimation, as discussed in k∈K
10 EURASIP Journal on Wireless Communications and Networking

By noticing that -Symmetric format-

  NP MD NP MD NP MD NP MD NP MD NP
Er Re r(k)s∗ (k) = 1, (A.7)
P P P P P P
we obtain 0
⎡   ⎤ (a)
⎢
(1) 2πT k ⎥
1 ⎢ k∈K k∈K ⎥ NP MD NP MD NP MD NP MD NP MD NP MD NP
F= ⎢ ⎥
⎢   ⎥, (A.8)
σ2 ⎣2πT k 4π 2 T 2 k2 ⎦ P P P P P P P
k∈K k∈K
0
L
where, considering the symmetry of the range K,
(b)
 
(1) = N, k = 0, (A.9)
k∈K k∈K Figure 12: Generic symmetric burst format.
  2 
N/2 N N
k2 = 8 −6 +1
k∈K
3 2 2
  (A.10) pilot symbols, and (2xeven + 1) is the (odd) number of sub-
2 N groups of data symbols; in Figure 12(b), (2xodd + 1) is the
+ 3M + 3M 3 −1 .
2 (odd) number of subgroups of pilot symbols, and 2xodd is
the (even) number of subgroups of data symbols. In the se-
After calculation of F−1 , the VCRB for ν in case of joint quel we find the values of x that minimize the VCRBs of ν
phase/Doppler-shift estimation is found to be and α, for fixed values of L, N, and M.
In the case of joint phase/Doppler-shift/Doppler-rate es-
−1 1 1
Fνν = VCRB2P (ν) =  timation, the fisher information matrix (FIM) of the generic
2π 2 T 2 k∈K k2 Es /N0 bursts of Figure 12 can be written as
 −1 (A.11)
3 · Es /N0 ⎡ ⎤
=
. F F F
4π 2 T 2 (N/2) 4(N/2)2 + 3M 2 + 3MN − 1 ⎢ θθ θν θα ⎥
F=⎢ ⎣ Fνθ Fνν Fνα ⎦
⎥

B. OPTIMAL SYMMETRIC BURST CONFIGURATION Fαθ Fαν Fαα

⎡ ! ! !⎤
FOR JOINT CARRIER-PHASE/DOPPLER-SHIFT a a a
− −
AND DOPPLER-RATE ESTIMATION: ⎢ E r  E r  −Er  ⎥
⎢ ∂θ 2 ∂θ∂ν ∂θ∂α ⎥
2P, 3P, 4P FORMATS ⎢ ! ! ! ⎥
⎢ a a a ⎥
⎢ ⎥
= ⎢ −Er

−Er ∂

2
−Er
∂ ⎥,
This appendix addresses the optimal signal design for ⎢
⎢ ∂ ν
∂ 
θ ν ν∂α ⎥
⎥
Doppler-shift ν and Doppler-rate α estimation in the case of ⎢ ! ! ! ⎥
⎣ a a ⎦
−Er −Er a −Er
joint phase/Doppler-shift and Doppler-rate estimation when ∂α∂ν 2
∂α
the received signal is expressed by (2)–(4). The optimal train- ∂α∂θ
(B.1)
ing signal structure is developed by minimizing the vector
Cramér-Rao bounds (VCRBs) [17, 18] for ν and α, with  p(r | α  is the probability
where a = ∂2 ln p(r | α, ν, θ), ν, θ)
, 
constraints on the total training block length and on the
density function of r = {r(k)}, with k ∈ K, conditioned on
burst overhead (1) of the signal (4). In fact, the Cramér-Rao
 Now r(k) is a random Gaussian variable with vari-
(α, ν, θ).
bounds (CRBs) for joint estimations are functions of the lo-
cation of the reference symbols in the burst. ance equal to σ 2 = N0 /(2Es ) and mean equal to
The issue of finding the optimal burst format that mini- 
s(k) = e j(θ+2π νkT+απk
2T2)
mizes the frequency CRB has been already addressed in [12–  (B.2)
14], but only for joint phase/Doppler-shift estimation. We
restrict our analysis to a symmetric burst format. In the se- so that
quel, we demonstrate that this symmetry also decouples the   
 =
p(r | α, ν, θ) p rk | α, ν, θ
estimation of Doppler-shift and Doppler-rate. Our attention
k∈K
is focused on a generic burst format as in Figure 12, either  
1 1  r(k) − 
2
with an even (Figure 12(a)) or an odd (Figure 12(b)) num- =  N exp − 2
s(k) .
ber of blocks of pilots. Just to rehearse notation, we mention 2πσ 2 2σ k∈K
that the length of the burst is L symbols, N is the total num- (B.3)
ber of pilot symbols, NP is the number of reference symbols
in each subgroup, M is the total number of data symbols, As detailed in Appendix A, after taking the logarithm of
and MD is the number of data symbols in each subgroup. (B.3), and after differentiating with respect to the unknown
In Figure 12(a), 2xeven is the (even) number of subgroups of parameters, and calculating the expectation of the terms with
Luca Giugno et al. 11

respect to r, we have Note that, thanks to the symmetry of the burst, the el-
⎡    ⎤ ements Fθν , Fνθ , Fαν , Fνα are all zero, which means that
(1) 2πT k πT 2 k2 the joint phase/Doppler-shift and Doppler-shift/Doppler-
⎢ k∈K ⎥
⎢ k∈K k∈K ⎥ rate estimations are decoupled.
⎢ ⎥
⎢    ⎥ Calculating F−1 , we obtain the VCRBs for the estimation
1 ⎢
⎢ 2πT 2
k 4π T 2 2
k 2π T2 3 3⎥
k ⎥ of ν as follows:
F= 2⎢ ⎥ , (B.4)
σ ⎢⎢ k∈K k∈K k∈K ⎥
⎥
⎢ ⎥
⎢    ⎥
⎣πT 2 k2 2π 2 T 3 k3 π 2 T 4 k4 ⎦ −1 1 1
Fνν = VCRB(ν) =  , (B.9)
k∈K k∈K k∈K 2π 2 T 2 k∈K k2 Es /N0
where, thanks to the symmetry of range K,
  as the one found in (A.11) without any Doppler-rate. The
k = 0, k3 = 0. (B.5) optimal burst configuration that minimizes the VCRB for ν
k∈K k∈K is thus the 2P format found in [14] also in the presence of
We finally get the expression of the FIM matrix as Doppler-rate effects.
The VCRB for α is
⎡  ⎤
N 0 πT 2 k2
⎢ ⎥
⎢  k∈K ⎥ 2N 1
1 ⎢ 4π 2 T 2 2 ⎥ −1
Fαα = VCRB(α) = −
  2   .
F= 2⎢ 0 k 0 ⎥. (B.6)
σ ⎢ ⎥ π2T 4 k∈K k
2 − N k∈K k 4 E s /N0
⎢  k∈K  ⎥
⎣ ⎦ (B.10)
πT 2 k2 0 π2T 4 4
k
k∈K k∈K

With an even number of pilot fields (Figure 12(a)), we have −1

If we compute Fαα as a function of x through (B.7) and
    (B.8), for both configurations of Figure 12, we find that the
 xeven even 
−1 N/2x M/ 2xeven − 1 − 1 minimum for Fαα −1
is obtained with xodd = 1 in (B.8). This
k2 = 2
k∈K n=0 l=1
2 was found by exhaustive numerical evaluation with practical
2   values for M and N. We can conclude that the VCRB of the
N M error variance of any unbiased estimator of α is always mini-
+l+ + n ,
2xeven 2xeven − 1 mized for a configuration with three blocks of pilot symbols
   
even 
 −1 N/2x
xeven equally spaced by two blocks of data symbols (3P format).
4 M/ 2xeven − 1 − 1 Setting xodd = 1 in (B.8) and (B.10), the minimum VCRB
k =2
k∈K n=0 l=1
2 of the error variance of any unbiased estimator of α for the
4   optimal 3P format is thus
N M
+l+ + n
2xeven 2xeven − 1
(B.7)
VCRBmin (α)

−1
while, with an odd number of pilot fields (Figure 12(b)), we = Fαα xodd =1 = VCRB3P (α)
get  −1  2 4
 
= 9720 · Es /N0 / π T N 108 4 − 5N 2 + N 4
odd +1)−1
 
 N/(2x
+ 32MN 15N 2 − 45
k 2= 2 k2  
k∈K k=1 + 24M 2 35N 2 − 45

odd +1) 
    3 4
+ 720NM + 270M .
−1 N/(2x
xodd
N/ 2xodd + 1 − 1
+2 +l (B.11)
n=0 l=1
2
 2
N M M
+ + n+ , In order to evaluate the gain in using the 3P for-
2xodd 2xodd +1 2xodd
odd +1)−1
mat, we have compared the VCRB3P (α) to the bounds
 N/(2x
k 4= 2 k4 for α in other configurations. Figure 13 shows the ra-
k∈K k=1 tios VCRB2P (α)/VCRB3P (α), VCRB4P (α)/VCRB3P (α), and
    VCRB2P (α)/VCRB4P (α) as functions of the total number N
odd +1) 
−1 N/(2x
xodd
N/ 2xodd + 1 − 1 of pilots and with η = 10%. It is clear that for practical
+2 +l
n=0
2 values of N = 40 ÷ 70, the 3P format exhibits a gain of
l=1
  10 log(78.6) = 19 dB in terms of Es /N0 with respect to the
N M M 4 2P format and of 10 log(1.1) = 0.4 dB with respect to the 4P
+ + n+ .
2xodd 2xodd +1 2xodd format. The accuracy of the 3P format and of the 4P format
(B.8) can be thus considered almost equivalent.
12 EURASIP Journal on Wireless Communications and Networking

200 [5] M. Morelli, “Doppler-rate estimation for burst digital trans-

180 mission,” IEEE Transactions on Communications, vol. 50, no. 5,
160 pp. 707–710, 2002.
140 [6] P. M. Baggenstoss and S. M. Kay, “On estimating the angle pa-
120 rameters of an exponential signal at high SNR,” IEEE Transac-
VCRB (α) ratios

100 78.6 tions on Signal Processing, vol. 39, no. 5, pp. 1203–1205, 1991.
71.9 [7] T. J. Abatzoglou, “Fast maximum likelihood joint estima-
80
tion of frequency and frequency rate,” IEEE Transactions on
60
Aerospace and Electronic Systems, vol. 22, no. 6, pp. 708–715,
40 1986.
20 [8] T. S. Rappaport, Wireless Communications: Principles & Prac-
1.09
0 tice, Prentice-Hall, Englewood Cliffs, NJ, USA, 1999.
−20 [9] J. A. Gansman, J. V. Krogmeier, and M. P. Fitz, “Single fre-
−40 quency estimation with non-uniform sampling,” in Proceed-
0 10 20 30 40 50 60 70 80 90 100 ings of the 30th Asilomar Conference on Signals, Systems and
N Computers, vol. 1, pp. 399–403, Pacific Grove, Calif, USA,
November 1996.
VCRB2P (α)/VCRB3P (α) [10] V. Lottici and M. Luise, “Embedding carrier phase recovery
VCRB2P (α)/VCRB4P (α)
into iterative decoding of turbo-coded linear modulations,”
VCRB4P (α)/VCRB3P (α)
IEEE Transactions on Communications, vol. 52, no. 4, pp. 661–
669, 2004.
Figure 13: VCRB2P (α)/VCRB3P (α), VCRB2P (α)/VCRB4P (α) and [11] S. Benedetto, R. Garello, G. Montorsi, et al., “MHOMS: high-
VCRB4P (α)/VCRB3P (α) ratios as function of the total number of pi- speed ACM modem for satellite applications,” IEEE Wireless
lots N. Communications, vol. 12, no. 2, pp. 66–77, 2005.
[12] F. Rice, “Carrier-phase and frequency-estimation bounds for
transmissions with embedded reference symbols,” IEEE Trans-
The various VCRBs can be easily calculated from (B.10) actions on Communications, vol. 54, no. 2, pp. 221–225, 2006.
using the appropriate x and (B.7) and (B.8). We report here [13] H. Minn and S. Xing, “An optimal training signal structure for
the final expressions frequency-offset estimation,” IEEE Transactions on Communi-
cations, vol. 53, no. 2, pp. 343–355, 2005.
VCRB2P (α) [14] A. Adriaensen, A. Van Doninck, and W. Steinert, “MF-TDMA

−1 burst demodulatir design with pilot-symbol assisted frequency
= Fαα xeven =1
 −1 estimation,” in Proceedings of the 8th International Workshop
360 · Es /N0 on Signal Processing for Space Communications (SPSC ’03),
= 2 4 3  ,
Catania, Italy, September 2003.
π T N − 4N 4N 2 + 15M 2 + 15MN − 4
[15] L. Giugno and M. Luise, “Carrier frequency and frequency
rate-of-change estimators with preamble-postamble pilot
(B.12) symbol distribution,” in Proceedings of IEEE International Con-
ference on Communications (ICC ’05), vol. 4, pp. 2478–2482,
VCRB4P (α)

−1
Seoul, Korea, May 2005.
= Fαα xodd =2 [16] A. Zhuang and M. Renfors, “Combined pilot aided and de-
 −1  2 4  cision directed channel estimation for the RAKE receiver,” in
=25920 · Es /N0 / π T N 288N 4 + 1305N 3 M Proceedings of the 52nd IEEE Vehicular Technology Conference
 
+ 240NM 8M 2 − 15 (VTC ’00), vol. 2, pp. 710–713, Boston, Mass, USA, September
  2000.
+ 30N 2 77M 2 − 48
  [17] U. Mengali and A. N. D’Andrea, Synchronization Techniques
+ 32 20M 4 − 75M 2 + 36 . for Digital Receivers, Plenum Press, New York, NY, USA, 1997.
(B.13) [18] F. Gini, R. Reggiannini, and U. Mengali, “The modified
Cramér-Rao bound in vector parameter estimation,” IEEE
Transactions on Communications, vol. 46, no. 1, pp. 52–60,
REFERENCES
1998.
[1] I. Ali, N. Al-Dhahir, and J. E. Hershey, “Doppler characteriza- [19] G. R. J. Povey and J. Talvitie, “Doppler compensation and code
tion for LEO satellites,” IEEE Transactions on Communications, acquisition techniques for LEO satellite mobile radio commu-
vol. 46, no. 3, pp. 309–313, 1998. nications,” in Proceedings of the 5th International Conference on
[2] “Special Issue on communications in the intelligence trans- Satellite Systems for Mobile Communications and Navigation,
portation system,” IEEE Communications Magazine, vol. 34, pp. 16–19, London, UK, May 1996.
1996. [20] M. Luise and P. Reggiannini, “Carrier frequency recovery
[3] M. Morelli and U. Mengali, “Feedforward frequency estima- in all-digital modems for burst-mode transmissions,” IEEE
tion for PSK: a tutorial review,” European Transactions on Transactions on Communications, vol. 43, no. 2–4, pp. 1169–
Telecommunications, vol. 9, no. 2, pp. 103–116, 1998. 1178, 1995.
[4] F. Giannetti, M. Luise, and R. Reggiannini, “Simple carrier fre- [21] D. C. Rife and R. R. Boorstyn, “Single-tone parameter estima-
quency rate-of-change estimators,” IEEE Transactions on Com- tion from discrete-time observations,” IEEE Transactions on
munications, vol. 47, no. 9, pp. 1310–1314, 1999. Information Theory, vol. 20, no. 5, pp. 591–598, 1974.
Hindawi Publishing Corporation
EURASIP Journal on Wireless Communications and Networking
Volume 2007, Article ID 23923, 8 pages
doi:10.1155/2007/23923

Research Article
TCP-Call Admission Control Interaction in
Multiplatform Space Architectures

Georgios Theodoridis,1 Cesare Roseti,2 Niovi Pavlidou,1 and Michele Luglio2

1 Department of Electrical & Computing Engineering, Aristotle University of Thessaloniki, 54124 Thessaloniki, Greece
2 Dipartimento di Ingegneria Elettonica, Università degli Studi di Roma Tor Vergata, Via del Politecnico 1, 00133 Rome, Italy

Received 28 September 2006; Revised 3 March 2007; Accepted 18 May 2007

Recommended by Anton Donner

The implementation of efficient call admission control (CAC) algorithms is useful to prevent congestion and guarantee target qual-
ity of service (QoS). When TCP protocol is adopted, some inefficiencies can arise due to the peculiar evolution of the congestion
window. The development of cross-layer techniques can greatly help to improve efficiency and flexibility for wireless networks.
In this frame, the present paper addresses the introduction of TCP feedback into the CAC procedures in different nonterrestrial
wireless architectures. CAC performance improvement is shown for different space-based architectures, including both satellites
and high altitude platform (HAP) systems.

Copyright © 2007 Georgios Theodoridis et al. This is an open access article distributed under the Creative Commons Attribution
License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly
cited.

1. INTRODUCTION the very popular Internet-based applications. It presents sev-

The development of network architectures including the
space segment (GEO satellites) aims to provide telecommu-
nication services in wide geographical areas. The space seg-
ment can complement or even replace the terrestrial infras-
tructure wherever the latter either fails or is not cost effective.
As a matter of fact, along with the evolution of new tech-
nological solutions, such as high altitude platforms (HAPs)
[1], next generation networks are envisioned as the integra-
tion of several different subsystems, unconditionally interop-
erating among one another [2]. Various architectures includ-
ing combinations of GEO and HAPs are under continuous
study so as to take advantage of each segment’s most favor-
able features in terms of coverage area, easy and quick de-
ployment, robustness to failure and to disaster occurrence,
and so forth.
On the other hand, some of the protocols and techniques
supporting the communication through this heterogeneous
wireless environment could be inadequate, since they are
specifically designed for wired networks. Nevertheless, these
protocols and techniques are worth being utilized due to
their many desirable characteristics and to the fact that the
wireless path is usually only a segment of the whole route be-
tween the sender and the receiver.
One of these protocols is TCP [3, 4], which is the pre-
dominant protocol at the transport layer when dealing with
eral impairments when it is implemented in wireless envi-
ronments [5, 6]. In brief, TCP, originally designed to work
well over wired congested network, considers all losses as
an explicit indication of network congestion [7]. Therefore,
TCP control rate leads to unnecessary rate reductions and
then to severe performance degradation without taking into
account error-prone wireless links. Communication involv-
ing long-delay segments (i.e., geostationary satellites), em-
phasizes such an impairment slowing down the reversion to
the previous transmission rate.
In addition, the presence of asymmetric links may slow
down the acknowledgement flow causing problems in the
forward channel as well, since TCP misinterprets the overall
RTT increase as a congestion notification in the data direc-
tion.
In parallel, CAC has evolved into one of the most signifi-
cant bandwidth management tools in the case of both wired
and wireless networks. However, the efficiency of the CAC
process is highly dependent on the accuracy of the available
info concerning the transmission rate of the serviced connec-
tions, not only at the time instant that the CAC algorithm is
executed but also for the whole duration of these connec-
tions. In particular, the CAC algorithm must be able to make
a safe prediction regarding the availability of resources in the
long term in order to decide if a new connection can be ad-
mitted into the system [8, 9].
2 EURASIP Journal on Wireless Communications and Networking
In this frame, the present paper investigates the possibil-
ity of introducing an interaction between TCP and CAC in
several nonterrestrial wireless architectures, in order to im-
prove CAC efficiency by taking TCP dynamics into account.
More specifically, the exploitation of TCP feedback as input
for the CAC algorithm at regular time intervals has proved
to be of primary importance for maximizing the utilization
of the network resources [10]. However, as the TCP perfor-
mance is rather dependent on the characteristics of the com-
munication path, the implementation of the CAC-TCP inter-
action on different system architectures will introduce mean-
ingful improvements in all the architectures showing such
characteristics, demonstrating a more general importance of
the concept. Additionally, so far, limited work can be found
in the literature on this topic.
The paper is organized as follows: Section 2 provides a
brief analysis of the TCP driven CAC concept, while Section 3
includes an overview of possible architectures. Section 4
presents a description of the reference architectures and the
simulation scenario, along with results comparing the effi-
ciency of the proposed algorithm in the several space archi-
tectures. Finally, Section 5 summarizes the conclusions.
2. THE CONCEPT OF A TCP-BASED CAC
TCP is a transport layer protocol based on sending data pack-
ets upon reception of acknowledgement of previously sent
packets, thus guaranteeing high reliability. When the net-
work is characterized by significant round trip time (RTT),
as in the case a satellite path is included, this process can sig-
nificantly slow down data transfer.
TCP can exploit congestion control either through an
ACK counting mechanism (the actions on the sliding win-
dow are just based on the number of received ACKs) [3] or
through the byte-counting scheme (the actions on the sliding
window are based on the actual number of bytes acknowl-
edged) [11].
When the communication path is not error-free (usual
in wireless networks) TCP misinterprets the data loss due to
the harsh wireless reception conditions as congestion occur-
rence. As a consequence, for every packet loss, TCP reduces
the actual transmission rate, limiting the bandwidth utiliza-
tion of the connection far below its nominal value.
This inefficiency is meaningful in wireless networks since
the radio resource is usually scarce and expensive. Particu-
larly in GEO satellite, the large footprints limit the imple-
mentation of frequency reuse, thus reducing system capac-
ity. Therefore, achieving maximum utilization of the avail-
able bandwidth must be the primary goal of every network
configuration.
On the other hand, CAC is implemented by the network
manager as a preventive congestion control scheme. CAC al-
gorithms decide upon the admittance/rejection of new con-
nections based on the network conditions (traffic load, link
capacity, buffer size, etc.) as well as the traffic characteristics
and the QoS objectives of both the candidate and the already
active users. In this framework, the aim of CAC is twofold: (i)
to guarantee that the QoS requirements of all the admitted
users are met, and (ii) maximize revenue from the network’s
perspective, that is, optimize the utilization of the available
resources [8]. However, achieving these objectives is rather
difficult, since CAC is inherently an “in advance” procedure
and no traffic model can offer a priori a completely accu-
rate prediction, in particular considering the heterogeneity
of multimedia telecommunication traffic sources. Therefore,
real-time measurements of each connection’s load and con-
ditions are considered essential for the CAC’s effectiveness
[9].
CAC functionality is based on the concept that the used
bandwidth plus the bandwidth of the upcoming user should
be lower or equal to the total capacity. As a matter of fact the
following condition must be always respected:

N
Bi + B f ≤ c. (1)
i=1
Since always B j TCP datarate ≤ B j nominal datarate, the
exploitation of the TCP feedback leads to a decrease in the
system overall blocking probability. Moreover, the band-
width assigned to each connection is equal to the real data
rate of the connection monitored via the TCP performance.
Therefore, having maximized the average number of the
users simultaneously active in the network and having min-
imized the over-assignment of resources, the throughput of
the network, defined as the percentage of the aggregate ca-
pacity that is actually occupied by the set of active connec-
tions, is radically improved.
In this frame, the possibility to get feedback informa-
tion about TCP congestion window actual evolution would
be of primary importance in order to efficiently drive CAC
scheme. In fact, since the CAC algorithm, by taking into ac-
count the actual amount of capacity necessary to exploit all
the TCP connections, could prevent the over provision of
bandwidth to the aforementioned connections, a better uti-
lization of the network resources would be achieved. In this
way, the admission/rejection of the new user would be based
on the actual occupancy of the channel by the active users at
the time instant of a new user arrival, computed according to
the TCP congestion window state of the connections instead
of their nominal data rate. The above scheme is depicted in
the flow chart of Figure 1.
3. SUITABLE ARCHITECTURES FOR
CAC-TCP INTERACTION
The potential improvement introduced by the implementa-
tion of the integrated CAC-TCP scheme is addressed in var-
ious nonterrestrial wireless architectures, where either the
high propagation delay and/or the occurrence of transmis-
sion errors negatively impact TCP performance by leading to
an unjustified decrease of the transmission data rate.
In particular, four different architectures are introduced
and described, focusing on the potential drawbacks concern-
ing optimal TCP working.
Georgios Theodoridis et al.
Arrival of a new user A
belonging to the x QoS-class
and the y mobility group
The set of the N active users B = {B j, }, j = 1, · · · , N
TCP the transport protocol
of the connection B j
Yes
B j real datarate = B j TCP datarate;

N
BW occupied = B j datarate;
j =1
A nominal datarate + BW occupied ≤ capacity
No
A is blocked
Figure 1: Cross-layer CAC-TCP flow chart.
Core network
Figure 2: Stand-alone GEO satellite.
3.1. Stand-alone GEO satellite
A system architecture based on a stand alone GEO satellite
(Figure 2) implies a rather challenging environment for TCP
performance. Such an architecture presents a long propaga-
tion path (in average about 80 000 km end to end) along with
transmission errors quantified in terms of BER (depending on
propagation channel conditions) and link unavailability, par-
3
No
B j real datarate = B j nominal datarate;
Yes
A is admitted
ticularly meaningful in case of use of high frequencies and/or
terminal mobility.
The large latency-bandwidth product could cause two
harmful effects:
(i) the pipe size, indicating the amount on unacknowl-
edged data that can be “in-flight” in a given instant,
could exceed the buffer limits in the existing im-
plementations resulting in a suboptimal maximum
throughput;
(ii) the high latency entails a considerable time interval to
open the TCP sliding window, when a new connection
starts (slow-start algorithm). Similarly, in the case of
losses, the reaction of TCP is very slow, increasing the
time needed to return to high transmission rates.
3.2. Stand-alone HAP
HAPS are characterized by the utilization of a platform lo-
cated in the stratosphere (about 20 km from ground), al-
lowing very fast deployment, low cost, less critical commu-
nication parameters, flexible architecture but limited cov-
erage (Figure 3). The proximity of the HAP to the ground
minimizes the propagation delay, being distances compa-
rable to the ones in terrestrial wireless systems [1, 12].
4 EURASIP Journal on Wireless Communications and Networking
Core network
Figure 3: Stand-alone HAP.
Nevertheless, since HAP systems work also in millimeter-
wave bands (47/48 GHz), in that case rain attenuation and
scattering constitute a severe constraint in achieving good
TCP performance. Some studies indicate a two-state (good-
bad) Markov model as a suitable error model [13]. There-
fore, the packet-error rate (PER) experienced by the TCP can
be approximated by the probability of the bad channel con-
ditions.
Depending on the PER value, TCP congestion window
continuously stops its growth resulting in “fast retransmit
and fast recovery” (FR-FR) or even timeout expirations,
when due to the loss of a large burst of segments, sender does
not receive any feedback (i.e., duplicate ACKs). In the latter
case, TCP remains in an idle state for several seconds and re-
sets its window to one segment.
3.3. Integrated GEO satellite—HAP
In order to allow HAPS users to communicate with remote
locations, a link between the HAPS and the satellite can be
envisaged, as depicted in Figure 4.
Being the GEO-HAP segment outside the atmosphere
and in line-of-sight (LoS) conditions, errors are due to free
space losses and thermal noise and quantified in terms of
BER. On the contrary, the PER of the overall link is predom-
inately defined by the HAP-ground segment, where signifi-
cant transmission errors can occur depending on the utilized
frequency and on eventual ground terminal mobility. Thus,
from the PER point of view an integrated GEO-HAP archi-
tectures is equivalent to the stand-alone HAP case. Moreover,
the use of GEO satellite as an intermediate node introduces
long RTT, adding the drawbacks in the TCP dynamics de-
tected in the stand-alone GEO satellite scenario.
3.4. HAP constellation
Finally, we consider the architecture of Figure 5, where the
coverage area is served by a constellation of HAPs; inter-HAP
links are also set up [14]. If the data is forwarded to the des-
tination HAP via one (or more) of its neighboring HAPs, al-
though the propagation delay is kept low, the end-to-end re-
ception conditions could possibly become harsher, due to the
imperfections of the inter-HAP links. The aforementioned
imperfections could be mostly due to the stabilization prob-
lems of the platforms, which would result in corresponding
pointing difficulties regarding the optical links that are envi-
sioned for such an inter-HAP communication.
Core network
Figure 4: Integrated GEO satellite—HAP.
Then, in this scenario, TCP performance suffers mainly
from the problems arisen in the stand-alone HAP architec-
ture.
4. EVALUATION OF THE CAC-TCP INTERACTION
IN SPACE ARCHITECTURES
4.1. Reference architectures
Summarizing, TCP performance over radio links, including
one or more space systems, relies primarily on two factors:
(1) the delay imposed by the space segment (RTT),
(2) the reception error probability of the wireless space-
user channel (PER).
The adopted TCP scheme, based on ACK counting, leads to
same efficiency as achievable when using the byte-counting
algorithm [11], because all the correctly delivered TCP pack-
ets are considered immediately acknowledged by the corre-
sponding ACK (ACK are not delayed).
Thus, in order to evaluate the efficiency as well as the
necessity of a TCP driven CAC scheme, only three different
network architectures, based on the boundary conditions in
terms of RTT and PER (or both), are selected to be simulated
in the present paper. They are stand-alone GEO (Figure 2),
stand-alone HAP (Figure 3), and integrated GEO satellite-
HAP (Figure 4). In the following, the most meaningful im-
plemented features concerning the selected architectures, are
described. In all the three architectures losses affect both
ACK and TCP packet flows (ACK losses have a slight impact
on the overall performance due to the cumulative nature of
ACKs [4]).
Stand-alone GEO architecture
Data originating from the core network are forwarded via a
gateway toward the GEO satellite, which transparently redi-
rects the stream (bent-pipe satellite) to the end users. Users
are considered to be fixed and equipped with VSATs appro-
priately mounted so as to guarantee line-of-sight (LoS) con-
dition for the satellite-user link. Therefore, since the signal-
to-noise ratio (SNR) is not only maximized but also relatively
invariant due to the absence of mobility, low PER value can
Georgios Theodoridis et al.
Figure 5: HAP constellation.
be assumed. Moreover, the gateway-satellite link is typically
dimensioned to be error free.
Stand-alone HAPS
In comparison with the previous architecture, the GEO satel-
lite has been replaced with a HAP, while the data flow main-
tains the same characteristics. The proximity of the HAP to
the earth greatly decreases link latency and facilitates the con-
nectivity of mobile users. In more detail, in the GEO satellite
scenario, a mobile terminal should be equipped with high
power transmission amplifier as well as sizeable antennas, so
as to compensate the high free-space attenuation imposed
by the long propagation path. These features lead to bulky
and power consuming (limited autonomy) terminals, com-
pletely inappropriate for mobile use. On the contrary, pro-
viding access via a HAP located at an altitude of 20 km al-
lows the use of small, cost-efficient, and user-friendly de-
vices. Consequently, the stand-alone HAPS scenario consid-
ers mobile users, which are further divided into three cate-
gories based on their mobility characteristics: highway-users,
suburban-users, urban-users. In particular, highway-users
move in open areas with maximum LoS probability, while, as
the city centre (suburban and urban users) is approached, the
higher building in combination with the narrow streets hin-
der the LOS path and the received signal is the result of suc-
cessive reflections (multipath). Moreover, according to the
channel model, even in the case of a highway user, the av-
erage PER is much higher than in case of a fixed user that is
served by a GEO system.
Integrated GEO satellite—HAP
The rationale behind the integration of the two systems is
that one satellite can provide connectivity to multiple HAPs
both among each other and toward the core network, with-
out the deployment of extra infrastructures. In this case, as
described in Section 3.4, the PER of the end-to-end link is de-
termined by the PER of the user-HAP segment (equal to the
of stand-alone HAP system), while the long RTT is imposed
by the GEO-satellite segment (equal to the case of stand-
alone GEO system).
As it becomes evident, the scenario involving a stand-
alone GEO system with fixed users presents the highest RTT
and the lowest PER, while the scenario involving a stand-
alone HAP system presents the highest PER and the lowest
5
Core network
RTT. Finally, the integrated GEO-HAP scenario combines
the characteristics of both of them, that is maximum RTT
and maximum PER. Moreover, beyond the fact that these
network scenarios present a wide range of RTT and PER val-
ues, they are also the most significant in terms of services
and applications. Therefore, the analysis of these case stud-
ies can provide solid conclusions regarding the ability of the
proposed TCP-CAC interaction to improve the network per-
formance, in terms of both blocking probability and average
throughput, in a great variety of channel conditions.
4.2. Simulation scenario and parameters
All the users are classified into three QoS classes accord-
ing to the nominal rate of their connections: 128, 256, and
512 kbps. The implementation of a weighted priority CAC
scheme, as the one proposed in [15], guarantees the provi-
sion of equitable service of multiple parallel flows with dif-
ferent bandwidth requirements. According to this admission
control algorithm, the aggregate capacity of the system is di-
vided into a number of segments equal to the number of QoS
classes. The width of each segment (i.e., the capacity percent-
age assigned to each QoS class) is determined by manipulat-
ing the desired blocking probability ratio between the QoS
classes. Thus, a new flow belonging to the QoSi class is ad-
mitted to the network on the basis of the bandwidth commit-
ted to the particular QoSi class. Instead, in the case of a CAC
scheme without any prioritization based on QoS class, the
users of the higher QoS classes would be practically excluded
from the network, as it would be difficult to satisfy their ex-
cessive bandwidth needs and they would be usually blocked
in favor of users with lower data rate requirements. There-
fore, a weighted priority CAC scheme as defined in [15] has
been taken as reference in our analysis presented hereafter.
Furthermore, the TCP driven CAC scheme has been derived
from exactly the same notion, with the only difference that,
as it has been described in Section 2, the TCP-CAC algorithm
takes into account the TCP feedback of the flows instead of
their nominal data rate in the process of computing the uti-
lization of the channel and the availability of resources.
Both the TCP driven and the reference CAC scheme, have
been simulated through an offline combination of two sim-
ulation tools that run sequentially. In particular, the network
simulator ns-2 [16] is used to configure the communica-
tion scenario (nodes, link parameters, and communication
protocols) and to obtain TCP statistics. Additionally, a C++
6 EURASIP Journal on Wireless Communications and Networking
simulation tool gets as input the TCP statistics and provides
the following functionalities:
(i) it runs alternatively either the reference or the TCP
driven CAC scheme;
(ii) it calculates the instantaneous and the average
throughput of the network;
(iii) it computes the connection blocking probability for
each QoS class as well as the connection blocking prob-
ability of the network.
To reproduce a trustworthy network traffic, we have consid-
ered packet error distribution (derived at TCP level) com-
pliant to the HAP communication characteristics [13, 17],
while satellite-HAP or satellite-user terminal link have been
considered as almost error free. The latter assumption is
rather realistic since satellite gateway EIRP can be set in or-
der to counterbalance the atmosphere attenuation. Then, de-
pending on the terminal mobility, the following PER distri-
butions have been considered.
(i) Fixed and portable terminals have been assumed al-
ways in line-of-sight (LoS) with the HAP/satellite.
Thus, uniform packet loss distributions (TCP level)
are considered with relatively low mean values (10−4
for fixed terminals and 10−3 for portable terminals).
(ii) In case of mobile terminals, a two-state channel model
[13] is considered to feature the alternating LoS and
shadowing conditions. Durations of “bad” and “good”
states depend on the motion environment according to
the values reported in [17].
Furthermore, both arrival and termination of TCP connec-
tions are managed by the C++ event driven simulator as Pois-
son processes [15]. Thus, the time between two successive
arrivals of users (τ) and the duration of each admitted con-
nection (d) follow exponential distribution with mean value
1/λ and 1/μ, respectively:
1
pdf(τ) = λ · e−λ·τ , E[τ] = ,
λ CAC scheme in comparison to the basic-CAC scheme. It
(2)
−μ·d 1 allows the reader to compare the impact of the proposed
pdf(d) = μ · e , E[d] = .
μ
The parameters E[d] and E[τ] along with the aggregate num-
ber of users in the network (S) determine the traffic load of
the network (L).
4.3. Results
The impact of the TCP-CAC interaction on all the three
selected network configurations (GEO, HAPS, and GEO-
HAPS) has been evaluated in terms of blocking probability
and average throughput. Moreover, the blocking probability
and the average throughput are calculated for both the basic
and the TCP-based call admission control scheme.
Figure 6 presents the system blocking probability for dif-
ferent traffic loads. Regardless the network architecture, the
basic-CAC algorithm leads in every case to the same block-
ing probability for the whole variety of traffic loads, which is
70
60
Blocking probability (%)
50
40
30
20
10
0
7000 8000 9000 10000 11000 12000 13000
Average traffic load
GEO, TCP-CAC GEO-HAPS, basic-CAC
GEO, basic-CAC HAPS, TCP-CAC
GEO-HAPS, TCP-CAC HAPS, basic-CAC
Figure 6: Blocking probability versus average traffic load.
expected, since only the nominal data rate of both the can-
didate and the already admitted users is taken into account
during the acceptance/rejection procedure. The fluctuations
in the TCP rate caused by the latency and the errors imposed
by the different channels do not affect the admission proce-
dure and therefore the curves regarding the basic-CAC al-
gorithm for all the three scenarios completely coincide with
each other. On the contrary, TCP-CAC algorithms present
much lower blocking probability. Due to the TCP feedback,
the system is able to calculate the actual occupancy of the
available channels which is much lower than the one declared
by the users initially during their admittance. Therefore, the
unused bandwidth is reassigned to new users that would oth-
erwise be blocked.
Figure 7 presents the improvement (decrease) intro-
duced to the system blocking probability by the TCP driven
scheme on architectures with different propagation charac-
teristics. As it becomes apparent,
BP decreaseGEO < BP decreaseHAP < BP decreaseGEO-HAPS .
(3)
This can be easily explained by the fact that in the case of
the integrated GEO-HAPS system, the harsh reception en-
vironment (long latency and high reception error probabil-
ity) leads to a severe degradation of the TCP performance
and thus to an intense decrease in the actual data rate of the
TCP connections. In fact, letting x be the aggregate amount
of nominal traffic load applying for network resources and y
the amount of traffic actually forwarded through the network
channels, simulations have demonstrated that
x > yGEO > yGEO-HAPS ,
(4)
x > yHAP > yGEO-HAP .
Georgios Theodoridis et al.
100
Decrease in blocking probability (%)
90
80
70
60
50
40
30
7000 8000 9000 10000 11000
Average traffic load
GEO
GEO-HAPS
HAPS
Figure 7: Blocking probability decrease versus average traffic load.
This means that users of the GEO-HAPS network leave a
great percentage of the system resources unutilized and thus,
in comparison with other architectures, the number of users
that can be simultaneously served by a channel of given ca-
pacity is much higher (lower blocking probability).
Moreover, results shown in Figure 7 lead us to the con-
clusion that the exploitation of TCP-feedback is much more
crucial in a stand-alone HAP system (high PER, low RTT)
than in a stand-alone GEO (high RTT, low PER) configura-
tion. Then, an error prone communication path, even with
low RTT, can cause abrupt decrease in the connection trans-
mission rate.
Blocking probability and average throughput are the two
main metrics of the network performance, each dealing with
the issue of the system efficiency from a different perspec-
tive. Blocking probability must be minimized to maximize
the QoS (minimum delay) guaranteed to the users, while av-
erage throughput must be maximized to maximize revenues
for the network administrator. Figure 8 shows that there is
always a tradeoff between these two factors: increased aver-
age throughput leads to increased blocking probability, while
limitation of the blocking probability results in a low band-
width utilization. In addition, from Figure 8 it is evident that,
regardless of the network scenario, the implementation of
the integrated TCP-CAC scheme results in the same average
throughput for any given value of blocking probability. This
is due to the fact that the admission control algorithm bases
the acceptance/rejection decision upon the knowledge of the
real traffic load forwarded at that given time through the net-
work.
Therefore, a new connection is blocked only if there is
no further available bandwidth. Thus, since the availability
of resources occurs on the basis of the new connection nom-
inal rate, for a given throughput, the blocking probability is
the same for all the possible scenarios (GEO, GEO-HAPS,
HAPS). On the contrary, the basic-CAC algorithm assumes
7
100
90
Average throughput (%)
80
70
60
50
40
30
12000 13000 0.1 1 10
Blocking probability (%)
GEO, TCP-CAC GEO-HAPS, basic-CAC
GEO, basic-CAC HAPS, TCP-CAC
GEO-HAPS, TCP-CAC HAPS, basic-CAC
Figure 8: Average throughput versus blocking probability.
that the occupancy of the network capacity is equal to the
aggregate of the nominal rates of all the active users. Conse-
quently, the requests for new connections are rejected while
there is still spare bandwidth. The average throughput for a
given blocking probability relies now upon the amount of
TCP data rate degradation. Therefore, the stand-alone GEO
case presents the higher average throughput and the inte-
grated GEO-HAPS architecture the minimum one, as they
present, respectively, the minimum and the maximum de-
crease in the TCP data rate.
Finally, according to Figure 9 the lower the network
blocking probability is, the higher the gain from the utiliza-
tion of the TCP feedback is. Moreover, the gain for the sce-
narios with the worst reception conditions is higher, since the
basic-CAC algorithm severely limits the system throughput.
5. CONCLUSIONS AND FUTURE PERSPECTIVES
New and innovative wireless telecommunication architec-
tures (including HAPs and satellite segments) are identified
to provide broadband services in a cost-efficient and ubiqui-
tous manner, ensuring seamless interoperation with the ex-
isting infrastructure. To ensure network efficiency for such
architectures it is worth optimizing the performance of pro-
tocols originally designed for terrestrial networks and for
classical architectures. Cross-layer techniques are becoming
fundamental to cope with the dynamic variations character-
izing wireless environments. The present paper focuses on
optimal utilization of the precious wireless resources when
flows running TCP share the channel. Referring to 5 differ-
ent architectures based on HAP/satellite links, we have an-
alyzed the potential drawbacks leading to suboptimal end-
to-end performance. A TCP driven CAC scheme has been
proposed in order to guarantee QoS for multimedia ser-
vices with different bandwidth requirements, guarantee an
8 EURASIP Journal on Wireless Communications and Networking

110 [9] K. Shiomoto, N. Yamanaka, and T. Takahashi, “Overview of

measurement-based connection admission control methods
100
Increase in average throughput (%)

in ATM networks,” IEEE Communications Surveys and Tuto-

90 rials, vol. 2, no. 1, pp. 2–13, 1999.
[10] C. Roseti, G. Theodoridis, M. Luglio, and N. Pavlidou, “TCP
80
driven CAC scheme for HAPS and satellite integrated sce-
70 nario,” in International Workshop on High Altitude Platform
Systems (WHAPS ’05), Athens, Greece, September 2005.
60
[11] M. Allman, “TCP Congestion Control with Appropriate Byte
50 Counting (ABC),” RFC 3465, February 2003.
[12] S. Karapantazis and N. Pavlidou, “Broadband communica-
40 tions via high-altitude platforms: a survey,” IEEE Communi-
30 cations Surveys & Tutorials, vol. 7, no. 1, pp. 2–31, 2005.
[13] J. L. Cuevas-Ruı́z and J. A. Delgado-Penı́n, “Channel model
0 10 20 30 40 based on semi-Markovian processes: an approach for HAPS
systems,” in Proceedings of the 14th International Conference
Blocking probability (%)
on Electronics, Communications and Computers (CONIELE-
GEO COMP ’04), pp. 52–56, Veracruz, Mexico, February 2004.
GEO-HAPS [14] R. Miura and M. Oodo, “Wireless communications system us-
HAPS
ing stratospheric platforms: R & D program on telecom and
broadcasting system using high altitude platform stations,”
Figure 9: Average throughput increase versus blocking probability. Journal of the Communications Research Laboratory, vol. 48,
no. 4, pp. 33–48, 2001.
[15] B. M. Epstein and M. Schwartz, “Predictive QoS-based admis-
optimal resource utilization, and reduce the system blocking sion control for multiclass traffic in cellular wireless networks,”
probability, without altering the TCP standard mechanisms. IEEE Journal on Selected Areas in Communications, vol. 18,
Through simulation, we demonstrated a considerable no. 3, pp. 523–534, 2000.
[16] K. Fall and K. Varadhan, The ns manual, VINT Project, Uni-
improvement on the performance with respect to a reference
versity of California, Berkeley, Calif, USA, 2001, http://www
CAC algorithm that takes into account only QoS require- .isi.edu/nsnam/ns/ns-documentation.html.
ments and physical parameters. [17] Recommendation ITU-R P.681-6, “ITU-R P.681-6 Propaga-
tion data required for the design of Earth-space land mobile
ACKNOWLEDGMENT telecommunication systems,” January 2003.

This paper has been supported by the European IST-FP6

project: “SatNEx II—Satellite Communications Network of
Excellence II.”

REFERENCES
[1] T. C. Tozer and D. Grace, “High-altitude platforms for wireless
communications,” Electronics and Communication Engineering
Journal, vol. 13, no. 3, pp. 127–137, 2001.
[2] S. Uskela, “Key concepts for evolution toward beyond 3G net-
works,” IEEE Wireless Communications, vol. 10, no. 1, pp. 43–
48, 2003.
[3] J. Postel, “Transmission Control Protocol,” IETF RFC 793,
September 1981.
[4] W. Stevens, TCP/IP Illustrated. Volume 1: The Protocols,
Addison-Wesley, Reading, Mass, USA, 1994.
[5] C. Partridge and T. J. Shepard, “TCP/IP performance over
satellite links,” IEEE Network, vol. 11, no. 5, pp. 44–49, 1997.
[6] P. Loreti, M. Luglio, R. Kapoor, et al., “Satellite systems per-
formance with TCP-IP applications,” in Proceedings of IEEE
Military Communications Conference on Communications for
Network-Centric Operations: Creating the Information Force
(MILCOM ’01), vol. 2, pp. 811–815, McLean, Va, USA, Oc-
tober 2001.
[7] W. Stevens, “TCP Slow Start, Congestion Avoidance, Fast re-
transmit, and Fast recovery Algorithms,” Internet RFC 2001,
1997.
[8] H. G. Perros and K. M. Elsayed, “Call admission control
schemes: a review,” IEEE Communications Magazine, vol. 34,
no. 11, pp. 82–91, 1996.
Hindawi Publishing Corporation
EURASIP Journal on Wireless Communications and Networking
Volume 2007, Article ID 72626, 20 pages
doi:10.1155/2007/72626

Research Article
Efficient Delay Tracking Methods with Sidelobes
Cancellation for BOC-Modulated Signals

Adina Burian, Elena Simona Lohan, and Markku Kalevi Renfors

Institute of Communications Engineering, Tampere University of Technology, P.O. Box 553, 33101 Tampere, Finland

Received 26 September 2006; Accepted 2 July 2007

Recommended by Anton Donner

In positioning applications, where the line of sight (LOS) is needed with high accuracy, the accurate delay estimation is an im-
portant task. The new satellite-based positioning systems, such as Galileo and modernized GPS, will use a new modulation type,
that is, the binary offset carrier (BOC) modulation. This type of modulation creates multiple peaks (ambiguities) in the envelope
of the correlation function, and thus triggers new challenges in the delay-frequency acquisition and tracking stages. Moreover, the
properties of BOC-modulated signals are yet not well studied in the context of fading multipath channels. In this paper, sidelobe
cancellation techniques are applied with various tracking structures in order to remove or diminish the side peaks, while keep-
ing a sharp and narrow main lobe, thus allowing a better tracking. Five sidelobe cancellation methods (SCM) are proposed and
studied: SCM with interference cancellation (IC), SCM with narrow correlator, SCM with high-resolution correlator (HRC), SCM
with differential correlation (DC), and SCM with threshold. Compared to other delay tracking methods, the proposed SCM ap-
proaches have the advantage that they can be applied to any sine or cosine BOC-modulated signal. We analyze the performances of
various tracking techniques in the presence of fading multipath channels and we compare them with other methods existing in the
literature. The SCM approaches bring improvement also in scenarios with closely-spaced paths, which are the most problematic
from the accurate positioning point of view.

Copyright © 2007 Adina Burian et al. This is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

1. INTRODUCTION to the terminal speed. Also, the satellite movement induces

Applications of new generations of Global Navigation Satel-
lite Systems (GNSS) are developing rapidly and attract a
great interest. The modernized GPS proposals have been re-
cently defined [1, 2] and the first version of Galileo (the
new European Satellite System) standards has been released
in May 2006 [3]. Both GPS and Galileo signals use direct
sequence-code division multiple access (DS-CDMA) tech-
nology, where code and frequency synchronizations are im-
portant stages at the receiver. The GNSS receivers estimate
jointly the code phase and the Doppler spreads through a
two-dimensional searching process in time-frequency plane.
This delay-Doppler estimation process is done in two phases,
first a coarse estimation stage (acquisition), followed by the
fine estimation stage (tracking). The mobile wireless chan-
nels suffer adverse effects during transmission, such as pres-
ence of multipath propagation, high level of noise, or ob-
struction of LOS by one or several closely spaced non-LOS
components (especially in indoor environments). The fading
of channel paths induces a certain Doppler spread, related
a Doppler shift, which deteriorates the performance, if not
correctly estimated and removed [4].
Since both the GPS and Galileo systems will send several
signals on the same carriers, a new modulation type has been
selected. This binary offset carrier (BOC) modulation has
been proposed in [5], in order to get a more efficient shar-
ing of the L-band spectrum by multiple civilian and military
users. The spectral efficiency is obtained by moving the signal
energy away from the band center, thus achieving a higher
degree of spectral separation between the BOC-modulated
signals and other signals which use the shift-keying mod-
ulation, such as the GPS C/A code. The BOC performance
has been studied for the GPS military M-signal [6] and later
has been also selected for the use with the new Galileo sig-
nals [3] and modernized GPS signals. The BOC modulation
is a square-wave modulation scheme, which uses the typi-
cal non-return-to-zero (NRZ) format [7]. While this type of
modulation provides better resistance to multipath and nar-
rowband interference [6], it triggers new challenges in the de-
lay estimation process, since deep fades (ambiguities) appear
2 EURASIP Journal on Wireless Communications and Networking
into the range of the ±1 chips around the maximum peak
of the correlation envelope. Since the receiver can lock on
a sidelobe peak, the tracking process has to cope with these
false lock points. In conclusion, the acquisition and track-
ing processes should counteract all these effects, and different
methods have been proposed in literature, in order to allevi-
ate multipath propagation and/or side-peaks ambiguities.
In order to minimize the influence of multipath errors,
which are the dominating error sources for many GNSS ap-
plications, several receiver-internal correlation approaches
have been proposed. During the 1990’s, a variety of receiver
architectures were introduced in order to mitigate the multi-
path for GPS C/A code or GLONASS. The traditional GPS re-
ceiver employs a delay-lock loop (DLL) with a spacing Δ be-
tween the early and late correlators of one chip. However, due
to presence of multipath, this wide DLL, which should track
the incoming signal within the receiver, is not able to align
perfectly the local code with the incoming signal, since the
presence of multipath (within a delay of 1.5 chips) creates a
bias of the zero-crossing point of the S-curve function. A first
approach to reduce the influences of code multipath is the
narrow correlator or narrow early minus-late (NEML) track-
ing loop introduced for GPS receivers by NovAtel [8]. Instead
of using a standard (wide) correlator, the chip spacing of a
narrow correlator is less than one chip (typically Δ = 0.1
chips). The lower bound on the correlator spacing depends
on the available bandwidth. Correlator spacings of Δ = 0.1
and Δ = 0.05 chips are commercially available for GPS.
Another family of tracking loops proposed for GPS are
the so-called double-delta (ΔΔ) correlators, which are the
general name for special code discriminators which are
formed by two correlator pairs instead of one [9]. Some
well-known implementations of ΔΔ concept are the high-
resolution correlator (HRC) [10], the Ashtech’s Strobe Cor-
relator [11], or the NovAtel’s Pulse Aperture Correlator [12].
Another similar tracking method with ΔΔ structure is the
Early1/Early2 tracking [13], where two correlators are lo-
cated on the early slope of the correlation function (with
an arbitrary spacing); their amplitudes are compared with
the amplitudes of an ideal reference correlation function and
based on the measured amplitudes and reference amplitudes,
a delay correction factor is calculated. The Early1/Early2
tracker shows the worst multipath performance for short-
and medium-delay multipath compared to the HRC or the
Strobe Correlator [9].
The early late slope technique [9], also called Multipath
Elimination Technology, is based on determining the slope
at both sides of autocorrelation function’s central peak. Once
both slopes are known, they can be used to perform a pseu-
dorange correction. Simulation results showed that in multi-
path environments, the early late slope technique is outper-
formed by HRC and Strobe correlators [9]. Also, it should
be mentioned that in cases of Narrow Correlator, ΔΔ, early-
late slope, or Early1/Early2 methods the BOC(n, n) modu-
lated signal outperforms the BPSK modulated signals, for
multipath delays greater than approximately 0.5 chips (long-
delay multipath) [9]. A scheme based on the slope differen-
tial of the correlation function has been proposed in [14].
This scheme employs only the prompt correlator and in pres-
ence of multipath, it has an unbiased tracking error, unlike
the narrow or strobe correlators schemes, which have a bi-
ased tracking error due to the nonsymmetric property of the
correlation output. However, the performance measure was
solely based on the multipath error envelope curves, thus its
potential in more realistic multipath environments is still an
open issue. One algorithm proposed to diminish the effect
of multipath for GPS application is the multipath estimating
delay locked loop (MEDLL) [15]. This method is different in
that it is not based on a discriminator function, but instead
forms estimates of delay and phase of direct LOS signal com-
ponent and of the indirect multipath components. It uses
a reference correlation function in order to determine the
best combinations of LOS and NLOS components (i.e., am-
plitudes, delays, phases, and number of multipaths) which
would have produced the measured correlation function.
As mentioned above, in the case of BOC-modulated sig-
nals, besides the multipath propagation problem, the side-
lobes peaks ambiguities should be also taken into account. In
order to counteract this issue, different approaches have been
introduced. One method considered in [16] is the partial
Sideband discriminator, which uses weighted combinations
of the upper and lower sidebands of received signal, to obtain
modified upper and lower signals. A “bump-jumping” algo-
rithm is presented in [17]. The “bump-jumping” discrimi-
nator tracks the ambiguous offset that arises due to multi-
peaked Autocorrelation Function (ACF), making amplitude
comparisons of the prompt peak with those of neighbor-
ing peaks, but it does not resolve continuously the ambigu-
ity issue. An alternative method of preventing incorrect code
tracking is proposed in [18]. This technique relies on sum-
mation of two different discriminator S-curves (named here
restoring forces), derived from coherent, respectively non-
coherent combining of the sidebands. One drawback is that
there is a noise penalty which increases as carrier-to-noise
ratio (CNR) decreases, but it does not seem excessive [18]. A
new approach which design a new replica code and produces
a continuously unambiguous BOC correlation is described
in [19].
The methods proposed in [16–19] tend to destroy the
sharp peak of the ACF, while removing its ambiguities. How-
ever, for accurate delay tracking, preserving a sharp peak of
the ACF is a prerequisite. An innovative unambiguous track-
ing technique, that keeps the sharp correlation of the main
peak, is proposed in [20]. This approach uses two correlation
channels, completely removing the side peaks from the corre-
lation function. However, this method is verified for the par-
ticular case of SinBOC(n, n) modulated signals, and its ex-
tension to other sine or cosine BOC signals is not straightfor-
ward. A similar method, with a better multipath resistance, is
introduced in [21].
Another approach which produces a decrease of sidelobes
from ACF is the differential correlation method, where the
correlation is performed between two consecutive outputs of
coherent integration [22].
In this paper, we analyze in details and develop further a
novel class of tracking algorithms, introduced by authors in
Adina Burian et al. 3

BOC-modulated
[23]. These techniques are named the sidelobes cancellation
1
methods (SCM), because they are all based on the idea of

code
suppressing the undesired lobes of the BOC correlation en- 0
velope and they cope better with the false lock points (ambi-
−1
guities) which appear due to BOC modulation, while keeping 0 1 2 3 4 5
the sharp shape of the main peak. It can be applied in both Chips
acquisition and tracking stages, but due to narrow width of
PRN sequence (NBOC1 = 1)
the main peak, only the tracking stage is considered here.
In contrast with the approach from [20] (valid only for sine

BOC-modulated
BOC(n, n) cases), our methods have the advantage that they 1

code
can be generalized to any sine and cosine BOC(m, n) modu- 0
lation and that they have reduced complexity, since they are
based on an ideal reference correlation function, stored at re- −1
0 1 2 3 4 5
ceiver side. In order to deal with both sidelobes ambiguities Chips
and multipath problems, we used the sidelobes cancellation
idea in conjunction with different discriminators, based on NBOC1 = 2
the unambiguous shape of ACF (i.e., the narrow correlator,

BOC-modulated
the high resolution correlator), or after applying the differ- 1

code
ential correlation method. We also introduced here an SCM 0
method with multipath interference cancellation (SCM IC),
where the SCM is used in combination with a MEDLL unit, −1
0 1 2 3 4 5
and also an SCM algorithm based on threshold comparison.
Chips
This paper is organized as follows: Section 2 describes the
signal model in the presence of BOC modulation. Section 3 NBOC1 = 3
presents several representative delay tracking algorithms,
employed for comparison with the SCM methods. Section 4 Figure 1: Examples of time-domain waveforms for sine BOC-
introduces the SCM ideas and presents the SCM usage in modulated signals.
conjunction with other delay tracking algorithms or based
solely on threshold comparison. The performance evalua-
tion of the new methods with the existing delay estimators,
in terms of root mean square error (RMSE) and mean time SinBOC(15, 10) (odd BOC-modulation order NBOC1 = 3)
to lose lock (MTLL), is done in Section 5. The conclusions together with the original PRN sequence (NBOC1 = 1) are
are drawn in Section 6. shown in Figure 1. In order to consider the cosine BOC-
modulation case, a second BOC-modulation order NBOC2 =
2 has been defined in [25], in a way that the case of sine BOC-
2. SIGNAL MODEL IN PRESENCE OF modulation corresponds to NBOC2 = 1 and the case of cosine
BOC MODULATION BOC modulation corresponds to NBOC2 = 2 (see the expres-
sions of (1) to (4)). After spreading and BOC modulation,
At the transmitter, the data sequence is first spread and the the data sequence is oversampled with an oversampled factor
pseudorandom (PRN) sequence is further BOC-modulated. of Ns , and this oversampling determines the desired accuracy
The BOC modulation is a square subcarrier modulation, in the delay estimation process. Thus, the oversampling fac-
where the PRN signal is multiplied by a rectangular sub- tor Ns represents the number of samples per BOC interval,
carrier which has a frequency multiple of code frequency. A and one chip will consists of NBOC1 NBOC2 Ns samples (i.e, the
BOC-modulated signal (sine or cosine) creates a split spec- chip period is Tc = Ns NBOC1 NBOC2 Ts , where Ts is the sam-
trum with the two main lobes shifted symmetrically from the pling rate).
carrier frequency by a value of the subcarrier frequency fsc The BOC-modulated signal sn,BOC (t) can be written, in
[5]. its most general form, as a convolution between a PRN se-
The usual notation for BOC modulation is BOC( fsc , fc ), quence sPRN (t) and a BOC waveform sBOC (t) [25]:
where fc is the chip frequency. For Galileo signals, the
BOC(m, n) notation is also used [5], where the sine and co-
sine BOC modulations are defined via two parameters m and sn,BOC (t)
n, satisfying the relationships m = fsc / fref and n = fc / fref , +∞
 
SF
 
where fref = 1.023 MHz is the reference frequency [5, 24]. = bn (−1)nNBOC1 ck,n sBOC t − nT − kTc
From the point of view of equivalent baseband signal, BOC n=−∞ k=1
modulation can be defined via a single parameter, denoted +∞ 
 SF
 
by the BOC-modulation order NBOC1 = 2m/n = 2 fsc / fc . The = sBOC (t) ⊗ bn ck,n (−1)nNBOC1 δ t − nT − kTc
factor NBOC1 is an integer number [25]. n=−∞ k=1

Examples of sine BOC-modulated waveforms for Sin- = sBOC (t) ⊗ sPRN (t),
BOC(1, 1) (even BOC-modulation order NBOC1 = 2) and (1)
4 EURASIP Journal on Wireless Communications and Networking

where bn is the nth complex data symbol, T is the symbol Examples of PSD for different BOC-modulated signals
0
period (or code epoch length) (T = SF Tc ), ck,n is the kth
chip corresponding to the nth symbol, Tc = 1/ fc is the chip
−20
period, SF is the spreading factor (i.e., for GPS C/A signal
and Galileo OS signal, SF = 1023), δ(t) is the Dirac pulse,
−40
⊗ is the convolution operator and sPRN (t) is the pseudo-

PSD (dB/Hz)
random (PRN) code sequence (including data modulation)
−60
of satellite of interest, and sBOC (·) is the BOC-modulated
signal (sine or cosine) whose expression is given in (2) to
−80
(4). We remark that the term (−1)nNBOC1 is included to take
into account also odd BOC-modulation orders, similar with
[26]. The interference of other satellites is modeled as addi- −100

tive white Gaussian noise, and, for clarity of notations, the

continuous-time model is employed here. However, the ex- −120
tension to the discrete-time model is straightforward and all −2 −1 0 1 2
presented results are based on discrete-time implementation. Frequency (MHz)
The SinBOC-CosBOC-modulated waveforms sBOC (t) are
defined as in [5, 25]: BPSK SinBOC (15, 10)
⎧ SinBOC (1, 1) CosBOC (10, 5)

⎪
⎪ NBOC1 πt
⎪
⎨sign sin for SinBOC, Figure 2: Examples of baseband PSD for BOC-modulated signals.
ssin / CosBOC (t) = ⎪ Tc

⎪
⎪
NBOC1 πt
⎩sign cos for CosBOC,
Tc
(2) SinBOC(15, 10)), the interference around the DC frequency
is not completely suppressed.
respectively, that is, for SinBOC-modulation [25], The baseband model of the received signal r(t) via a fad-
ing channel can be written as [25]
NBOC1 −1


Tc  n=
+∞ 
L
sSinBOC (t) = (−1)i pTB1 t − i , (3) r(t) = Eb e+ j2π fD t bn αn,l (t)
i=0
NBOC
n=−∞ l=1 (6)
 
and for CosBOC-modulation [25], × sn,sin / CosBOC t − τl + η(t),

NBOC1 −1 NBOC2 −1
 
where Eb is the bit or symbol energy of signal (one symbol is
sCosBOC (t) = (−1) i+k equivalent with a code epoch and typically has a duration of
i=0 k=0 (4) T = 1 ms), fD is the Doppler shift introduced by channel, L

is the number of channel paths, αn,l is the time-varying com-
Tc Tc
× pTB t − i −k . plex fading coefficient of the lth path during the nth code
NBOC1 NBOC1 NBOC2
epoch, τl is the corresponding path delay (assuming to be
In (3) and (4), pTB1 (·) is a rectangular pulse of sup- constant or slowly varying during the observation interval)
port Tc /NBOC1 and pTB (·) is a rectangular pulse of support and η(·) is the additive noise component which incorporates
Tc /NBOC1 NBOC2 . For example, the additive white noise from the channel and the interfer-
⎧ ence due to other satellites.
⎪
⎨1 Tc At the receiver, the code-Doppler acquisition and track-
if 0 ≤ t < ,
pTB (t) = ⎪ NBOC1 NBOC2 (5) ing of the received signal (i.e., estimating the Doppler shift fD
⎩0 otherwise. and the channel delay τl ) are based on the correlation with a
reference signal sref (t − τ, fD , n1 ), including the PRN code and
We remark that the bandlimiting case can also be taken into the BOC modulation (here, n1 is the considered symbol in-
account, by setting pTB (·) to be equal to the pulse shaping dex):
filter.  
Some examples of the normalized power spectral den- sref t − τ, fD , n1
sity (PSD), computed as in [25], for several sine and cosine 
SF NBOC1 −1 NBOC2 −1
 

BOC-modulated signals, are shown in Figure 2. It can be ob- = e− j2π fD t ck,n1 (−1)i+ j pTB
served that for even-modulation orders such as SinBOC(1, 1) k=−1 i=0 j =0
or CosBOC(10, 5) (currently selected or proposed by Galileo

Tc Tc
Signal Task Force), the spectrum is symmetrically split into t − n1 T − kTc − i −j − τ .
NBOC1 NBOC1 NBOC2
two parts, thus moving the signal energy away from DC fre- (7)
quency and thus allowing for less interference with the exist-
ing GPS bands (i.e., the BPSK case). Also, it should be men- Some examples of the absolute value of the ideal ACF for
tioned that in case of an odd BOC-modulation order (i.e., several BOC-modulated PRN sequences, together with the
Adina Burian et al. 5

BPSK case, are illustrated in Figure 3. As it can be observed, Ideal ACF for BOC-modulated signals
1
for any BOC-modulated signal, there are ambiguities within
the ±1 chips interval around the maximum peak. 0.9
After correlation, the signal is coherently averaged over 0.8
Nc ms, with the maximum coherence integration length dic-
0.7
tated by the coherence time of the channel, by possible resid-

Normalized ACFs
ual Doppler shift errors and by the stability of oscillators. If 0.6
the coherent integration time is higher than the coherence 0.5
time of the channel, the spectrum of the received signal will
be severely distorted. The Doppler shift due to satellite move- 0.4
ment is estimated and removed before performing the coher- 0.3
ent integration. For further noise reduction, the signal can be
0.2
noncoherently averaged over Nnc blocks; however there are
some squaring losses in the signal power due to noncoher- 0.1
ent averaging. The delay estimation is performed on a code- 0
Doppler search space, whose values are averaged correlation −1.5 −1 −0.5 0 0.5 1 1.5
functions with different time and frequency lags, with max- Chips
ima occurring at f = fD and τ = τl . BPSK SinBOC (15, 10)
SinBOC (1, 1) CosBOC (10, 5)
3. EXISTING DELAY ESTIMATION ALGORITHMS IN
Figure 3: Examples of absolute value of the ACF for BOC-
MULTIPATH CHANNELS modulated signals.
The presence of multipath is an important source of error
for GPS and Galileo applications. As mentioned before, tra-
ditionally, the multipath delay estimation block is imple- branches are combined noncoherently, and the S-curve is ob-
mented via a feedback loop. These tracking loop methods are tained as in (9),
based on the assumption that a coarse delay estimate is avail-  2 2
able at receiver, as result of the acquisition stage. The tracking
loop is refining this estimate by keeping the track of the pre-
vious estimate.
3.1. Narrow early minus late (NEML) correlator
One of the first approaches to reduce the influences of code
multipath is the narrow early minus late correlation method,
first proposed in 1992 for GPS receivers [8]. Instead of us-
ing a standard correlator with an early late spacing Δ of 1
chip, a smaller spacing (typically Δ = 0.1 chips) is used.
Two correlations are performed between the incoming sig-
nal r(t) and a late (resp., early) version of the reference code
SNEML (τ) = RLate (τ) − |REarly (τ) . (9)
The error signal given by the S-curve is fed back into
a loop filter and then into a numeric controlled oscilla-
tor (NCO) which advances or delays the timing of the ref-
erence signal generator. Figure 4 illustrates the S-curve in
single path channel, for BPSK, SinBOC(1, 1), respectively,
SinBOC(10, 5) modulated signals. The zerocrossing shows
the presence of channel path, that is, the zero delay er-
ror corresponds to zero feedback error. However, for BOC-
modulated signals, due to sidelobes ambiguities, the early late
spacing should be less than the width of the main lobe of
the ACF envelope, in order to avoid the false locks. Typically,
for BOC(m, n) modulation, this translates to approximately
srefEarly,Late (t − τ ± Δ/2), where srefEarly,Late (·) is the advanced or Δ ≤ n/4m.
delayed BOC-modulated PRN code and τ is the tentative
delay estimate. The early (resp., late) branch correlations 3.2. High-resolution correlator (HRC)
Rearly,Late (·) can be written as

The high-resolution correlator (HRC), introduced in [10],

Δ can be obtained using multiple correlator outputs from con-
REarly,Late (τ) = r(t)srefEarly,Late
Nc
These two correlators spaced at Δ (e.g., Δ = 0.1 chips) are
used in the receiver in order to form the discriminator func-
tion. If channel and data estimates are available, the NEML
loops are coherent. Typically, due to low CNR and residual
Doppler errors from GPS and Galileo systems, noncoherent
NEML loops are employed, when squaring or absolute value
are used in order to compensate for data modulation and
channel variations. The performance of NEML is best illus-
trated by the S-curve, which presents the expected value of
error as a function of code phase error. For NEML, the two
t − τ ± dt. (8)
2 ventional receiver hardware. There are a variety of combi-
nations of multiple correlators which can be used to imple-
ment the HRC concept, which yield similar performance.
The HRC provides significant code multipath mitigation for
medium and long delay multipath, compared to the con-
ventional NEML detector, with minor or negligible degrada-
tion in noise performance. It also provides substantial carrier
phase multipath mitigation, at the cost of significantly de-
graded noise performance, but, it does not provide rejection
of short delay multipath [10]. The block diagram of a non-
coherent HRC is shown in Figure 5. In contrast to the NEML
structure, two new branches are introduced, namely, a very
6 EURASIP Journal on Wireless Communications and Networking

Ideal S-curve (no multipath) for Ideal S-curve (no multipath) for two BOC-modulated signals
BOC-modulated and BPSK signals 1
1 0.8
0.8
0.6
0.6
0.4

Normalized S-curve
0.4
Normalized S-curve

0.2
0.2
0
0
−0.2
−0.2
−0.4
−0.4
−0.6
−0.6
−0.8
−0.8
−1
−1 −1.5 −1 −0.5 0 0.5 1 1.5
−1.5 −1 −0.5 0 0.5 1 1.5 Delay error (chips)
Delay error (chips)
SinBOC (1, 1)
SinBOC (1, 1) SinBOC (10, 5)
SinBOC (10, 5)
BPSK Figure 6: Ideal S-curves for noncoherent HRC with a = −1/2, for
two BOC-modulated signals and Δ = 0.1 chips.
Figure 4: Ideal S-curves for BOC-modulated and BPSK signals
(NEML, Δ = 0.1 chips).

r(t) 3.3. Multipath estimating delay locked loop (MEDLL)

I & D on
Nc msec | |2
+ +
Late code A different approach, proposed to remove the multipath ef-
I & D on fects for GPS C/A delay tracking is the multipath estima-
Nc msec | |2 − +
tion delay locked l;oop [15]. The MEDLL method estimates
Early code I & D on jointly the delays, phases, and amplitudes of all multipaths,
Nc msec | |2 +
canceling the multipath interference. Since it is not based on
Very late code
I & D on − an S-curve, it can work in both feedback and feedforward
| |2
Nc msec configurations. To the authors’ knowledge, the performance
Constant factor a of MEDLL algorithm for BOC modulated signals is still not
Very early code
NCO Loop filter well understood, therefore, would be interesting to study a
similar approach. The steps of the MEDLL algorithm (as im-
Figure 5: Block diagram for HRC tracking loop. plemented by us) are summarized bellow.
(i) Calculate the correlation function Rn (t) for the nth
transmitted code epoch. Find out the maximum peak
of the correlation function and the corresponding de-
early and, respectively, a very late branch. The S-curve for a lay τ1 , amplitude a1,n , and phase θ1,n .
noncoherent five-correlator HRC can be written as in [10]: (ii) Subtract the contribution of the calculated peak, in or-
der to have a new approximation of the correlation
 2  2 
SHRC (τ) = RLate (τ) − REarly (τ) function R(1) n (τ) = Rn (τ) − a 1,n Rref (t − τ1,n )e j θ1,n . Here
 2  2  (10) Rref (·) is the reference correlation function, in the ab-
+ a RVeryLate (τ) − RVeryEarly (τ) , sence of multipaths (which can be, for example, stored
at the receiver). Find out the new peak of the residual
where RVeryLate (·) and RVeryEarly (·) are the very late and very function R(1) n (·) and its corresponding delay τ 2,n , am-
early correlations, with the spacing between them of 2Δ plitude a2,n , and phase θ2,n . Subtract the contribution
chips, and a is a weighting factor which is typically −1/2 [10]. of the new peak of residual function from R(1) n (t) and
Examples of S-curves for HRC in the presence of a sin- find a new estimate of the first peak. For more than
gle path static channel, are shown in Figure 6, for two BOC- two peaks, the procedure is continued until all desired
modulated signals. The early late spacing is Δ = 0.1 chips peaks are estimated.
(i.e., narrow correlator), thus the main lobes around zero (iii) The previous step is repeated until a certain criterion
crossing are narrower, and it is more likely that the separa- of convergence is met, that is, when residual function
tion between multiple paths will be done more easily. is below a threshold (e.g., set to 0.5 here) or until
Adina Burian et al. 7

Ideal ACFs (no multipath) for SinBOC (1, 1)-modulated signal and thus this method has a potential in reducing the side
1
peaks ambiguities.
0.9
0.8 3.5. Nonambiguous BOC(n, n) signal tracking
0.7
(Julien&al. method)
Normalized ACF

0.6 A recent tracking approach, which removes the sidelobes

ambiguities of SinBOC(n, n) signals and offers an improved
0.5
resistance to long-delay multipath, has been introduced in
0.4 [20]. This method, referred here as Julien&al. method, af-
0.3 ter the name of the first author in [20], has emerged while
observing the ACF of a SinBOC(1, 1) signal with sine phas-
0.2
ing, and the cross correlation of SinBOC(1, 1) signal with its
0.1 spreading sequence. The ideal correlation function Rideal
BOC (·)
0 for SinBOC(1, 1)-modulated signals in the absence of multi-
−1.5 −1 −0.5 0 0.5 1 1.5 paths, can be written as [25]
Delay error (chips)

1 Tc 1 Tc
Non-coherent integration Rideal
BOC (τ) = ΛTc /2 (τ) − ΛTc /2 τ − − ΛTc /2 τ + ,
2 2 2 2
Differential correlation (12)
Figure 7: Envelope correlation function of traditional noncoher-
ent integration and differential correlation for a SinBOC(1, 1)- where ΛTc /2 (τ − α) is the value in τ of a triangular function1
modulated signal. centered in α, with a width of 1-chip, Tc is the chip period,
and τ is the code delay in chips.
The cross correlation of a SinBOC(1, 1) signal with the
the moment when introducing a new delay does not spreading pseudorandom code, for an ideal case (no multi-
improve the performance in the sense of root mean paths and ideal PRN code), can be expressed as [20]
square error between the original correlation function

1 T T
and the estimated correlation function. Rideal
BOC,PRN (τ) = ΛTc /2 τ + c + ΛTc /2 τ − c .
2 2 2
(13)
3.4. Differential correlation (DC)
Two types of DLL discriminators have been considered
Originally proposed for CDMA-based wireless communi- in [20], namely, the early-minus- late- power (EMLP) dis-
cation systems, the differential correlation method has also criminator and the dot-product (DP) discriminator. These
been investigated in context of GPS navigation system [22]. It examples of possible discriminators result from the use of
has been observed that with low and medium coherent times the combination of BOC-autocorrelation function and of
of the fading channel and in absence of any frequency error, the BOC/PRN-correlation function [20]. Based on (12) and
this approach provides better resistance to noise than the tra- (13), the ideal EMLP discriminator is constructed, as in (14),
ditional noncoherent integration methods. In DC method, where τ is the code tracking error [20]:
the correlation is performed between two consecutive out- 


2 Δ 2 Δ
puts of coherent integration. These correlation variables are Sideal ideal
EMLP (τ) = RBOC τ + − Rideal
BOC τ −
then integrated, in order to obtain a differential variable. The 2 2



differential detection variable z is given as ideal2 Δ ideal2 Δ
− RBOC,PRN τ + − RBOC,PRN τ − .
2 2

M −1
1  ∗  (14)
zDC =  y yk+1 2 , (11)
k
M−1 k=1 The alternative DP discriminator variant [20] does not
where yk , k = 1, . . . , M are the outputs of the coherent in- have a linear variation as a function of code tracking error:
tegration and M is the differential integration length. For a
Sideal
DP (τ)
fair comparison between the differential noncoherent and 


traditional noncoherent methods, here it is assumed that 2 Δ 2 Δ 2
= Rideal
BOC τ + − Rideal
BOC τ − Rideal
BOC (τ)
M = Nnc , where Nnc is the noncoherent integration length. 
2

2


Since the differential coherent correlation method was no- 2 Δ 2 Δ 2
− Rideal
BOC,PRN τ + − Rideal
BOC,PRN τ − Rideal
BOC (τ).
ticed to be more sensitive to residual Doppler errors, only 2 2
the differential noncoherent correlation is considered here. (15)
The analysis done in [22] is limited to BPSK modulation.
From Figure 7, it can be noticed that applying the DC to a
BOC-modulated signal, instead of the conventional nonco- 1 Our notation is equivalent with the notation triα (x/ y) used in [20], via
herent integration, the sidelobes envelope can be decreased, triα (τ/ y) = ΛTc /2 (τ − αTc / y).
8 EURASIP Journal on Wireless Communications and Networking
SinBOC (1, 1) modulation, ACFs of
BOC-modulated and subtracted signals
1
Continue line:
BOC-modulated signal
Dashed line:
0.5 subtracted signal
0 BOC-modulated signals. In contrast with the Julien&al.
−1.5 −1 −0.5 0 0.5 1 1.5
Delay (chips)
SinBOC (1, 1) modulation, ACF of unambiguous signal
1
Unambiguous signal
0.5
0 ceived BOC-modulated signal with the reference PRN code.
−1.5 −1 −0.5 0 0.5 1 1.5
Delay (chips)
Figure 8: SinBOC(1, 1)-modulated signal: examples of the ambigu-
ous correlation function and subtracted pulse (upper plot) and
the obtained unambiguous correlation function (lower plot), for a
single-path channel.
Since the resulting discriminators remove the effect of
SinBOC(1, 1) modulation, there are no longer false lock
points, and the narrow structure of the main correlation lobe
is preserved [20]. Indeed, the side peaks of SinBOC(1, 1)
correlation function RidealBOC (τ) have the same magnitude
and same location as the two peaks of SinBOC(1, 1)/PRN-
correlation function Rideal
BOC,PRN (τ). By subtracting the squares
of the two functions, a new synthesized correlation function
is derived and the two side peaks of SinBOC(1, 1) correlation
function are canceled almost totally, while still keeping the
sharpness of the main lobe (Figure 8). Two small negative
sidelobes appear next to the main peak (about ±0.35 chips
around the global maximum), but since they point down-
wards, they do not bring any threat [20]. The correlation val-
ues spaced at more than 0.5 chips apart from the global peak
are very close to zero, which means a potentially strong resis-
tance to long-delay multipath.
In practice, the discriminators SEMLP (τ) or SDP (τ), as
given in [20], are formed via continuous computation, at re-
ceiver side, of correlation functions RBOC (·) and RBOC,PRN (·)
values, not on the ideal ones. In practice, RBOC (·) is the
correlation between the incoming signal (in the presence of
multipaths) and the reference BOC-modulated code, and
RBOC,PRN (·) is the correlation between the incoming signal
and the pseudorandom code (without BOC modulation).
This method has been applied only to SinBOC(n, n) signals.
Moreover, instead of making use of the ideal reference func-
tion Rideal
BOC,PRN (·) (which can be computed only once and
stored at the receiver side), the correlation RBOC,PRN (·) needs
to be computed for each code epoch in [20]. Of course, in or-
der to make use of the Rideal
BOC,PRN (·) shape, we also need some
information about channel multipath profile. This will be ex-
plained in the next section.
4. SIDELOBES CANCELLATION METHOD (SCM)
In this section, we introduce unambiguous tracking ap-
proaches based on sidelobe cancellation; all these approaches
are grouped under the generic name of sidelobes cancel-
lation methods). The SCM technique removes or dimin-
ishes the threats brought by the sidelobes peaks of the
method, which is restricted to the SinBOC(n, n) case, we
will show here how to use SCM with any sine or cosine
BOC-modulated signal. The SCM approach uses an ideal
reference correlation function at receiver, which resembles
the shapes of sidelobes, induced by BOC modulation. In
order to remove the sidelobes ambiguities, this ideal refer-
ence function is subtracted from the correlation of the re-
In the Julien&al. method, the subtraction function, which
approximates the sidelobes, is provided by cross-correlating
the spreading PRN code and the received signal. Here, this
subtraction function is derived theoretically, and computed
only once per BOC signal. Then, it is stored at the receiver
side in order to reduce the number of correlation operations.
Therefore, our methods provide a less time-consuming and
simpler approach, since the reference ideal correlation func-
tion is generated only once and can be stored at receiver.
4.1. Ideal reference functions for SCM method
In this subsection, we explain how the subtraction pulses
are computed and then applied to cancel the undesired side-
lobes.
Following derivations similar with those from [25] and
intuitive deductions, we have derived the following ideal ref-
erence function to be subtracted from the received signal af-
ter the code correlation:
NBOC1 −1 NBOC1 −1 NBOC2 −1 NBOC2 −1
   
Rideal
sub (τ) =
i=0 j =0 k=0 l=0
TB
(−1)i× j+k+l ΛTB τ + (i − j)TB + (k − l) ,
NBOC2
(16)
where TB = Tc /NBOC1 NBOC2 is the BOC interval, ΛTB (·)
is the triangular function centered at 0 and with a width
of 2TB -chips, NBOC1 is the sine BOC-modulation order
(e.g., NBOC1 = 2 for SinBOC(1, 1), or NBOC1 = 4
for SinBOC(10, 5)) [25], and NBOC2 is the second BOC-
modulation factor which covers sine and cosine cases, as ex-
plained in [25] (i.e., if sine BOC modulation is employed,
NBOC2 = 1 and, if cosine BOC modulation is employed,
NBOC2 = 2).
As an example, the simplest case of SinBOC(1, 1)-
modulation (i.e., the main choice for Open Services in
Galileo), (16) becomes
    
Rideal
sub,SinBOC(1,1) (τ) = ΛTB τ − TB + ΛTB τ + TB , (17)
Adina Burian et al. 9

which is similar with Julien& al. expression of (13) with the CosBOC (10, 5) modulation, ACFs of
BOC-modulated and subtracted signals
exception of a 1/2 factor (here, TB = Tc /2). 1
The Sin- and CosBOC(m, n)-based ideal autocorrelation Continue line:
BOC-modulated signal
function can be written as [25]
Dashed line:
0.5 subtracted signal
NBOC1 −1 NBOC1 −1 NBOC2 −1 NBOC2 −1
   
Rideal
BOC (τ) =
i=0 j =0 k=0 l=0 0

−1.5 −1 −0.5 0 0.5 1 1.5
T
(−1)i+ j+k+l ΛTB τ + (i − j)TB + (k − l) B . Delay (chips)
NBOC2
(18) CosBOC (10, 5) modulation, ACF of unambiguous signal
1
Again, for SinBOC(1, 1) case, the expression of (18) reduces Unambiguous signal
to 0.5

Rideal
SinBOC(1,1) (τ) 0
    
= 2ΛTB (τ) − ΛTB τ − TBOC − ΛTB τ + TBOC ,
−1.5 −1 −0.5 0 0.5 1 1.5
(19)
Delay (chips)
which is, again, similar to Julien& al. expression of (12) with
Figure 9: CosBOC(10, 5)-modulated signal: examples of the am-
the exception of a 1/2 factor (for SinBOC(1, 1), TBOC = Tc /2, biguous correlation function and subtracted pulse (upper plot)
NBOC1 = 2 and NBOC2 = 1). and obtained unambiguous correlation function (lower plot), in a
We remark that the difference between (16) and (18) single-path channel.
stays in the power of −1 factor, that is, (16) stands for an ap-
proximation of the sidelobe effects (no main lobe included),
while (18) is the overall ACF (including both the main lobe
and the side lobes). The next step consists in canceling the ef- normalization of reference function (i.e., to find the weight
fect of sidelobes (16) from the overall correlation (18), after factors w), the peaks magnitudes of RBOC (·) function are first
normalizing them properly. found out and sorted in increased order. Then the weighting
Thus, in order to obtain an unambiguous ACF shape, the factor w is computed as the ratio between the last-but-one
squared function (Rideal
sin (·)) , (Rcos (·)) , respectively, has to
2 ideal 2 peak and the highest peak. We remark that the above algo-
be subtracted from the ambiguous squared correlation func- rithm does not require the computation of the BOC/PRN
tion as shown in correlation anymore, it only requires the computation of
 2  2
RBOC (τ) = Rn (τ) correlation. The pulses to be subtracted are
Rideal ideal
unamb (τ) = RBOC (τ) − w Rideal
sin / cos (τ) , (20) always based on the ideal functions Ridealsin / cos (τ), and therefore,
they can be computed only once (via (16)) and stored at the
where w < 1 is a weight factor used to normalize the reference receiver (in order to decrease the complexity of the tracking
function (to achieve a magnitude of 1). unit).
For example, for SinBOC(1, 1) and w = 1, we get from By comparison with Julien&al. method, here the num-
(17), (19), and (20), after straightforward computations, that ber of correlations at the receiver is reduced by half (i.e.,
   RBOC,PRN (·) computation is not needed anymore). Thus the
Rideal 2
unamb (τ) = 4 ΛTB (τ) − ΛTB (τ)ΛTB τ − TBOC SCM technique offers less computational burden (only one
  (21)
− ΛTB (τ)ΛTB τ + TBOC , correlation channel in contrast to Julien&al. method, which
uses two correlation channels).
and if we plot Rideal
unamb (τ) (e.g., see the lower plot of Figure 8),
Figures 8 and 9 show the shapes of the ideal ambigu-
we get a main narrow correlation peak, without sidelobes. ous correlation functions and of the subtracted pulses, to-
All the derivations so far were based on ideal assumptions gether with the correlation functions, obtained after subtrac-
(ideal correlation codes, single path static channels, etc.). tion (SCM method). Figure 8 exemplifies a SinBOC(1, 1)-
However, in practice, we have to cope with the real signals, modulated signal, while Figure 9 illustrates the shapes for a
so the ideal autocorrelation function RidealBOC (τ) should be re-
CosBOC(10, 5)-modulation case. As it can be observed, for
placed with the computed correlation RBOC (τ) between the both SinBOC and CosBOC modulations, the subtractions
received signal and the reference BOC-modulated pseudo- removes the sidelobes closest to the main peak, which are
random code. Thus, (20) becomes the main threats in the tracking process. Also, it should be
mentioned that the Figure 8, for a SinBOC(1, 1) modulated
 2  2
Runamb (τ) = RBOC (τ) − w Rideal
sin / cos (τ) . (22) signal, is also illustrative for the Julien&al. method, since the
shapes of correlation functions are similar with those pre-
Here comes into equation the weighting factor, since vari- sented in [20].
ous channel effects (such as noise and multipath) can mod- Equation (20) is valid for single path channels. How-
ify the levels of RBOC (τ) function. In order to perform the ever, in multipath presence, delay errors due to multipaths
10 EURASIP Journal on Wireless Communications and Networking

are likely to appear. When (22) is applied in this situation, Exemplification of SCM IC method (steps 1 to 4)
one important issue is to align the subtraction pulse to the 1
LOS peak (otherwise, the subtraction of (22) will not can-
cel the correct sidelobes). This can be done only if some ini- 0.8
tial estimate of LOS delay is obtained. For this purpose, we
employ and compare several feedback loops or feedforward 0.6
algorithms, as it will be explained next.
0.4

4.2. SCM with interference cancellation (IC) 0.2

Combining the multipath eliminating DLL concept with the 0

SCM method, we obtain an improved SCM technique with
multipath interference cancellation (SCM with IC). In this 0 10 20 30 40 50 60 70 80
method, the initial estimate of LOS delay is obtained via Samples
MEDLL algorithm. The sidelobe cancellation is applied in-
Original ACF Subtracted ideal function
side the iterative steps of MEDLL, as explained below.
Estimated CIR Unambiguous ACF
(1) Calculate the correlation function Rn (τ) between the
received signal and the reference BOC-modulated Exemplification of SCM IC method (steps 5 to 6)
code (e.g., see the continuous line, Figure 10, up- 1
per plot). Find the global maximum peak (the peak
1) of this correlation function, maxτ |Rn (τ)|, and its 0.8
corresponding delay, τ1,n , amplitude a1,n and phase
θ1,n (e.g., the peak situated at the 50th-sample delay,
0.6
Figure 10, upper plot).
(2) Compute the ideal reference function centered at τ1,n : 0.4
Rideal
sub (τ − τ
1,n ) via (16) (see the dashed line, Figure 10,
upper plot). 0.2
(3) Build an initial estimate of the channel impulse re-
sponse (CIR) based on τ1,n , a1,n , and θ1,n (e.g., the es- 0
timated CIR of peak 1, Figure 10, upper plot).
−0.2
(4) In order to remove the sidelobes ambiguities, the 0 10 20 30 40 50 60 70 80
function Rideal
sub (τ − τ
1,n ) is then subtracted from the Samples
multipath correlation function Rn (τ) and an unam-
Unambiguous ACF
biguous shape is obtained, using (22), or, equiva- Residual function
lently Rn,unamb (τ) = (Rn (τ))2 − (Rideal
sub (τ − τ
1,n ))2 . In Estimated CIR, 2nd peak
Figure 10, the unambiguous ACF Rn,unamb (·) is plot-
ted with dashed-dotted line, in both upper and lower Figure 10: Exemplification of SCM IC method, 2-paths fading
plots. channel with true channel delay at 44 and 50 samples, average path
powers [−2, 0] dB, SinBOC(1, 1)-modulated signal.
(5) Cancel out the contribution of the strongest path
and obtain the residual function R(1) n,unamb (τ) =
ideal 1,n
Rn,unamb (τ) − a1,n Runamb (τ)(τ − τ1,n )e , where
j θ

Rideal
unmab (τ) is the unambiguous reference function
given by (20). The shape of residual function is and the maximum global peak is re-estimated from
exemplified in Figure 10, lower plot (drawn with R(2)
n,unamb (τ) = (Rn,unamb (τ)) − (a
2 1,n Rideal
unamb (τ)(τ −
1,n ideal 2,n 2
continuous line).
(6) The new maximum peak of the residual function
R(1)
n,unamb is found out (e.g., at 44th-sample delay,
Figure 10, lower plot), with its corresponding de-
τ1,n )e j θ + a2,n Runamb (τ)(τ − τ2,n )e ) .
j θ
(7) The steps (3) to (6) are repeated until all desired peaks
are estimated and until the residual function is below
a threshold value. In the example of Figure 10, after 6
steps both path delays are estimated correctly.
lay τ2,n , amplitude a2,n and phase θ2,n . The con-
tributions of both peaks 1 and 2 are subtracted These steps of SCM IC method are illustrated in
from unambiguous correlation function Rn,unamb (τ) Figure 10, for 2-path fading channel.
Adina Burian et al. 11

Ideal S-curve (no multipath), SCM NEML method SinBOC (1, 1), Δ = 0.1 chips
1

envelope (meters)
Multipath error
10
0.8
0.6 0

0.4 −10
Normalized S-curve

0.2 0 0.2 0.4 0.6 0.8 1

0 Multipath spacing (chips)

−0.2 NEML correlator

−0.4
SCM NEML method

−0.6
SinBOC (10, 5), Δ = 0.1 chips

envelope (meters)
Multipath error
−0.8 10

−1 0
−1.5 −1 −0.5 0 0.5 1 1.5
Delay error (chips) −10

SinBOC (1, 1) 0 0.2 0.4 0.6 0.8 1

SinBOC (10, 5) Multipath spacing (chips)
Figure 11: SCM NEML method: ideal S-curves (no multipath), for NEML correlator
two BOC-modulation cases and Δ = 0.1 chips. SCM NEML method

Figure 12: Multipath error envelopes (in meters): NEML correlator

4.3. SCM using narrow early minus lat discriminator versus SCM NEML method, for two BOC-modulation cases and
Δ = 0.1 chips.
(SCM NEML)

After obtaining an unambiguous correlation function

Rn,unamb (τ) (as it was shown in the previous section, steps
(1) to (4)), a NEML S-curve is constructed, by forming the
early, respectively, late branches, spaced at Δ = 0.1 chips. The
S-curve is obtained in the same way as in Section 3.1, by sub-
tracting the late and early branches of unambiguous correla-
tion function,
 2  2
SCM NEML method brings a decrease in the errors of mul-
tipath envelopes, for both SinBOC(1, 1) and SinBOC(10, 5)
signals. We remark that the variations of the lower delay er-
ror envelope in the lower plot of Figure 12 are due to, on one
hand, the errors in the zero-crossing estimation algorithm,
and, on the other hand, to the fact that worse MEE is not
necessarily guaranteed when the paths are out of phase for
SSCMNEML (τ) = RLate ) − Rn,unamb (τ) .
Early
n,unamb (τ (23) the noncoherent NEML.
Examples of S-curves obtained with this method, in
4.4. SCM using high-resolution correlator
presence of a single path static channel, are presented in
discriminator (SCM HRC)
Figure 11, for two BOC-modulated signals, SinBOC(1, 1)
and SinBOC(10, 5), and a spacing of Δ = 0.1 chips. Com- In a similar manner as in previous section, the SCM method
paring with Figure 4, which presents the NEML S-curves for can be also used in conjunction with an HRC discrimina-
ambiguous signals, in Figure 11, the possibility to detect an tor, after removing the side peaks threats and obtaining an
incorrect zero crossing, due to sidelobes peaks, is decreased. unambiguous correlation function Rn,unamb (τ). Based on this
A typical measure of performance for the ability of a de- unambiguous function, an HRC S-curve is constructed, in an
lay tracking loop to deal with multipath error is the so-called analogous way as in Section 3.2:
multipath error envelope (MEE) [9, 10]. The MEE is usu-
ally computed for one direct and one reflected channel paths,  2  2
SSCMHRC (τ) = RLate ) − Rn,unamb (τ)
Early
with a certain variable spacing. The multipath errors are cal- n,unamb (τ

 VeryLate 2  VeryEarly 2 
culated for the worst-case scenario, when the two paths are + a Rn,unamb (τ) − Rn,unamb (τ) ,
added inphase (upper MEE) and have equal strength, and (24)
also, when the two paths are out of phase (lower MEE). Com-
parisons of MEEs plots, for both NEML and SCM NEML Early
where Rn,unamb (·) and RLate
n,unamb (·) are the advanced and de-
methods, are shown in Figure 12, for two BOC-modulated
layed unambiguous correlations, with a spacing between
signals. A static channel with two paths of equal amplitudes VeryEarly
and variable spacing was considered. The only interference them of Δ = 0.1 chips. The Rn,unamb (·), respectively,
VeryLate
considered here is the multipath interference, and the addi- Rn,unamb (·) are the very early and the very late unambiguous
tive white noise effect is not taken into account. As it can be correlation branches, spaced at 2Δ chips and the weighting
seen in Figure 12, comparing with the NEML correlator, the factor a = −1/2.
12 EURASIP Journal on Wireless Communications and Networking

Ideal S-curve (no multipath), SCM HRC method Ideal ACF (no multipath) for SinBOC (10, 5) modulated signal
1
0.8
0.8
0.6
0.4
Normalized S-curve

0.6

Normalized ACF
0.2
0 0.4

−0.2
0.2
−0.4

−0.6 0

−0.8
−0.2
−1
−1.5 −1 −0.5 0 0.5 1 1.5 −1.5 −1 −0.5 0 0.5 1 1.5
Delay error (chips) Delay error (chips)

SinBOC (1, 1) Ambiguous correlation SCM method

SinBOC (10, 5) Differential correlation SCM DC method

Figure 13: SCM HRC method: ideal S-curves (no multipath), for Figure 15: Envelopes of correlation functions obtained with am-
two BOCmodulation cases, with a = −1/2 and Δ = 0.1 chips. biguous correlation, DC method, SCM approach, and SCM DC
method, for a SinBOC(10, 5)-modulated signal.

SinBOC (1, 1), Δ = 0.1 chips SinBOC(10, 5) cases. As it can be noticed, there is a slight im-
provement brought by the SCM HRC method over the HRC
10
envelope (meters)
Multipath error

correlator.
5
0
4.5. SCM using differential correlation (DC) in
−5
conjunction with feedback and feedforward
−10
tracking algorithms
0 0.2 0.4 0.6 0.8 1
Multipath spacing (chips) It has been observed that the DC method has potential to de-
crease the sidelobes amplitudes, thus lowering the possibility
HRC method
SCM HRC method
to detect a wrong side peak. To enhance the performance of
the DC method, we use it in conjunction with different track-
SinBOC (10, 5), Δ = 0.1 chips ing algorithms, such as NEML or HRC methods, or with IC
method. These algorithms are applied in similar ways as ex-
envelope (meters)

10
Multipath error

5 plained in Sections 3.1, 3.2, and 3.3, on the correlation func-

0 tions obtained after performing the noncoherent DC tech-
−5 nique (Section 3.4).
−10 Also, the performance may be enhanced further, by us-
0 0.2 0.4 0.6 0.8 1 ing the SCM approach after applying the DC method. This is
Multipath spacing (chips) done in the same way as explained in previous Sections (4.2,
4.3, and 4.4), but after using first the DC method on the am-
HRC method biguous correlation function between the multipath received
SCM HRC method signal and the reference BOC-modulated code. Indeed, as il-
lustrated in Figure 15, in case of a SinBOC(10, 5) modulated
Figure 14: Multipath error envelopes (in meters): HRC method
versus SCM HRC method, for two BOC-modulation cases and signal, the combination of DC and SCM algorithms can de-
Δ = 0.1 chips. crease even further the sidelobes amplitudes, thus eliminat-
ing more ambiguities.

4.6. SCM with threshold comparison (SCM thr)

The ideal S-curves obtained with the SCM HRC method, Another approach is to test the performance of SCM tech-
for two BOC-modulation orders, are presented in Figure 13. nique using a thresholding algorithm. Starting from the un-
The MEEs performances, for both the HRC and SCM HRC ambiguous correlation function Rn,unamb (τ), an estimate of
methods, are illustrated in Figure 14, for SinBOC(1, 1) and noise variance σn2 is obtained, as the mean of the squares of
Adina Burian et al. 13

the out-of-peak values, similar to [4]. Using this estimated SinBOC (1, 1), AWGN single-path channel
100
noise variance, a linear threshold γ is computed, based on the
second peak γ2 of the ideal unambiguous correlation func-
tion Rideal 10−1
unamb (τ) (i.e., for SinBOC(1, 1) γ2 = 0.5, as seen in
Figure 3), together with the estimate of the noise variance σn2 :
 10−2

RMSE (chips)
γ = γ2 + σn2 . (25)
10−3
Then the LOS delay is estimated, based on the unambigu-
ous correlation function Rn,unamb (τ), using this threshold. If
the peak of the estimated first path is too low (i.e., ten times 10−4
lower than the global peak), then this path is discarded and
the next estimate is considered. 10−5

5. SIMULATION RESULTS 10−6

20 25 30 35 40
5.1. Additive white noise Gaussian (AWGN) channel CNR (dB-Hz)

NEML DC NEML
We first test the performance of the proposed algorithms in
Julien & al. EMLP DC SCM NEML
the ideal AWGN channel (single path), in order to check SCM NEML
whether SCM algorithm introduces a deterioration with re-
SinBOC (1, 1), AWGN single-path channel
spect to the standard narrow and high-resolution correla-
tors (it is known that NEML is able to attain the Cramer-
Rao bound in AWGN channels [8]). We will show that no
deterioration is incurred when SCM is applied. The perfor-
mance criteria are root mean square error (RMSE) and mean 102.4
time to lose lock (MTLL). The simulations were carried out
in Matlab. The MTLL is computed as the average value for
MTLL (s)

which the estimated delay tracking error of the first path

is below 1 chip. The tracking process is started, after the
coarse acquisition of the signal, assuming that we are in the
“lock” condition, that is, the delay error is strictly less than
one chip. For all presented simulations (both in this section
and in Section 5.2), the coherent integration length is set to
Nc = 20 milliseconds and the noncoherent integration is per-
formed over Nnc = 3 blocks (i.e., the total coherent and non-
coherent integration length is 60 milliseconds), and the over-
sampling factor is set to Ns = 11. We generated 5000 random
points in order to compute the RMSE and MTLL statistics.
That is, the maximum observable MTLL based on these sim-
ulations is 5000Nc Nnc = 300 s (i.e., an MTTL value of 300
seconds reflects the fact that we never lost the lock during
that particular simulation).
The AWGN results are shown for SinBOC(1, 1) case in
Figures 16 and 17, for the comparison with NEML and HRC,
respectively. As seen in these figures, SCM algorithm does not
deteriorate the performance in AWGN case, compared with
narrow and high-resolution correlators. The sidelobe cancel-
lations applied on the top of NEML and HRC give the same
results as those of the original NEML and HRC algorithms,
respectively, if the channel is single path AWGN channel (e.g.,
the differences in performance between SCM + NEML and
NEML are only at the third decimal, with NEML slightly bet-
ter).
5.2. Fading channels
In what follows, the performance of the discussed delay es-
timation algorithms is compared in multipath fading chan-
102.3
102.2
20 25 30 35 40
CNR (dB-Hz)
NEML DC NEML
Julien & al. EMLP DC SCM NEML
SCM NEML
Figure 16: Comparison of feedback delays estimation algorithms
employing the NEML discriminator and of the Julien&al. method,
as a function of CNR; upper plots: RMSE, lower plots: MTLL.
NEML and SCM NEML curves are overlapping. DC NEML and DC
SCM NEML curves are also overlapping (differences at the 3rd dec-
imal).
nels. The same performance criteria as in the previous sec-
tion are used, namely, RMSE and MTLL. Two representative
BOC-modulated signals have been selected for the simula-
tions included in this paper. The first one is the SinBOC(1, 1)
modulation, the common baseline for Galileo open service
(OS) structure, agreed by US and European negotiation.
The second one is the CosBOC(10, 5) modulation, which
has been proposed for the Galileo Public Regulated Service
(PRS) and for the current GPS M-code. In order to have fair
14 EURASIP Journal on Wireless Communications and Networking

SinBOC (1, 1), AWGN single-path channel to v = 3 km/h, while for outdoor profiles, the mobile speeds
100
of 25, 45, or 75 km/h have been selected). Two main chan-
nel profiles have been considered: either with fixed Rayleigh
10−1 distribution of all paths and with average path power of −1,
−2, 0 and −3 dB, or a 2-paths decaying power delay profile
(PDP) channel, with Rician distributions for the first path
10−2
RMSE (chips)

and Rayleigh distribution for the next path. Similar with the
AWGN case in Section5.1, during simulations, the first path
delay of the channel is assumed to be linearly increasing, with
10−3
a slope of 0.05 chips per block of Nc Nnc millisecond, thus the
tracking algorithms should capture this linear delay increase.
10−4 The successive channel path delays have a random spacing
with respect to the precedent delay, uniformly distributed be-
tween 1/(Ns NBOC1 NBOC2 ) and xmax , where xmax (in chips) is
10−5 the maximum separation between successive paths (i.e., for
20 25 30 35 40
closed-spaced paths scenario, xmax = 0.1 chips). In order to
CNR (dB-Hz)
have independent and reliable results for each method, the
HRC DC HRC search interval is different for each algorithm. which means
Julien & al. DP SCM DC HRC that once the lock is lost for one method, this will not affect
SCM HRC the other algorithms. The search window has few chips (typ-
ically between 4 and 12 chips), depending on the number
SinBOC (1, 1), AWGN single-path channel of paths, the distance between them and on the used BOC-
modulation orders. The search window is sliding around the
350 previous delay estimate and if we have erroneous estimates,
the lock is lost at some point. For the feedback algorithms
(i.e., NEML, HRC, or Julien&al. methods), the search for
zero crossing is conditioned by the previous delay estimates.
Similar with AWGN case, he coherent integration length is
200 set to Nc = 20 milliseconds, the noncoherent integration is
MTLLS (s)

performed over Nnc = 3 blocks, and the oversampling factor

100
20 22 24 26 28 30 32
CNR (dB-Hz)
HRC
Julien & al. DP
SCM HRC
Figure 17: Comparison of feedback delays estimation algorithms
employing the HRC discriminator and of the Julien&al. method,
as a function of CNR, upper plots: RMSE, lower plots: MTLL. HRC
and SCM HRC curves are overlapping; DC HRC and DC SCM HRC
curves are also overlapping (differences at the 4th decimal).
comparison, the performance of introduced feedback tech-
niques is evaluated separately from that of the feedforward
methods. The same modulation types as in Section 5.1 are
used here, namely, SinBOC(1, 1) and CosBOC(10, 5) mod-
ulations. However, the introduced SCM method can be ex-
tended to any sine or cosine BOC-modulation case.
The studied techniques have been investigated under the
assumption of indoor or outdoor Rayleigh or Rician multi-
path profiles (i.e., for indoor channel, the speed mobile is set
is set to Ns = 11.
The SCM approach is exemplified in Figure 18, for a
Rayleigh 2-paths fading channel, with equal PDP. The up-
per plot exemplifies a SinBOC(1, 1) modulation case, with
xmax = 1 chip, while the lower plot shows the original ACF,
together with subtracted pulse and unambiguous shape, for
34 36 38 a SinBOC(10, 5) case and xmax = 0.5 chips. In both cases
the threat of the sidelobes is eliminated using the SCM tech-
nique. For instance, in the SinBOC(1, 1) case, the correct de-
DC HRC
lay of first path, situated at the 70th sample (in one chip, there
SCM DC HRC
are Ns NBOC1 NBOC2 samples) is more likely to be detected, af-
ter the main sidelobe (situated at the 81th sample) is removed
by subtraction.
Figure 19 presents the RMSE and MTLL, for the feedback
algorithms which use the NEML discriminator, with an early
late spacing of Δ = 0.1 chips. The signal is SinBOC(1, 1)
modulated. Here, the Julien&al. method employs an EMLP
discriminator, as presented in Section 3.5. The channel is 4-
path outdoor Rayleigh channel, v = 75 km/h, with the most
challenging situation of closely-spaced paths (i.e., xmax = 0.1
chips). From both plots, it can be seen that both SCM-
enhanced methods (the SCM NEML and SCM DC NEML)
are performing much better than the other algorithms. Also,
the Julien&al. EMLP technique brings an improvement in the
results, comparing with both NEML and DC NEML meth-
ods, but still not approaching the performance of the SCM
algorithms.
Adina Burian et al. 15

SinBOC (1, 1), Rayleigh fading channel SinBOC (1, 1), Rayleigh channel,
with 2 paths, xmax = 1 chip speed mobile = 75 km/h, xmax = 0.1 chips

1st path true delay =

0.8 70 samples 10−0.3

0.6

RMSE (chips)
10−0.4
ACFs

0.4

0.2
10−0.5
0

−0.2 10−0.6
40 60 80 100 120 20 25 30 35 40
Delay error (samples) CNR (dB-Hz)

Ambiguous ACF NEML DC NEML

Subtracted pulse Julien & al. EMLP SCM DC NEML
Unambiguous ACF SCM NEML

SinBOC (10, 5), Rayleigh fading channel

SinBOC (1, 1), Rayleigh channel, 4 paths, xmax = 0.1 chips
with 2 paths, xmax = 0.5 chips
1
True delay =
239 samples
0.8
102
0.6
MTLL (s)
ACFs

0.4

0.2

−0.2

20 25 30 35 40
180 200 220 240 260 280 300 320
CNR (dB-Hz)
Delay error (samples)
NEML DC NEML
Ambiguous ACF Julien & al. EMLP SCM DC NEML
Subtracted pulse SCM NEML
Unambiguous ACF
Figure 19: Comparison of feedback delays estimation algorithms
Figure 18: Exemplification of SCM method for a 2-paths Rayleigh employing the NEML discriminator and of the Julien&al. method,
fading channel. Upper plot: SinBOC(1, 1)-modulated signal and as a function of CNR; SinBOC(1, 1) modulation, Rayleigh channel
xmax = 1 chip. Lower plot: SinBOC(10, 5)-modulated signal and with an average pathspower delay profile of −1, −2, 0, and −3 dB,
xmax = 0.5 chips. v = 75 km/h, closely spaced paths with xmax = 0.1 chips.

Figures 20 and 21 illustrate the performances of the

introduced methods using an HRC discriminator. The
Julien&al. method employs a DP discriminator, as explained
in Section 3.5. This selection is done because it has been ob-
served by simulations that the Julien&al. method employing
a DP discriminator exceeds the performance of the EMLP
discriminator; this behavior is expected since the DP ap-
proach does not vary linearly with the code tracking error
[20] as the EMLP discriminator. In Figure 20, the signal is
SinBOC(1, 1)-modulated, for a 2-path channel with Rician
distribution for the first path, a mobile speed of 25 km/h and
a large separation between successive paths xmax = 1 chip.
Figure 21 presents the case of a CosBOC(10, 5)-modulated
16 EURASIP Journal on Wireless Communications and Networking

SinBOC (1, 1), 2-paths Rician channel, CosBOC (10, 5), Rayleigh channel,
xmax = 1 chip, mobile speed = 25 km/h xmax = 0.1 chips, mobile speed = 45 km/h

10−0.3
10−0.3

10−0.4
RMSE (chips)

RMSE (chips)
10−0.4 10−0.5

10−0.6
10−0.5
10−0.7

20 25 30 35 40 20 25 30 35 40
CNR (dB-Hz) CNR (dB-Hz)

HRC DC HRC HRC DC HRC

Julien & al. DP SCM DC HRC Julien & al. DP SCM DC HRC
SCM HRC SCM HRC

SinBOC (1, 1), 2-paths Rician channel, CosBOC (10, 5), 4-paths Rayleigh channel,
xmax = 1 chip, mobile speed = 25 km/h xmax = 0.1 chips
MTLL (s)
MTLL (s)

101 101

20 25 30 35 40
20 25 30 35 40
CNR (dB-Hz)
CNR (dB-Hz)
HRC DC HRC
HRC DC HRC
Julien & al. DP SCM DC HRC
Julien & al. DP SCM DC HRC
SCM HRC
SCM HRC
Figure 21: Comparison of feedback delays estimation algorithms
Figure 20: Comparison of feedback delays estimation algorithms employing the HRC discriminator and of the Julien&al. method, as
employing the HRC discriminator and of the Julien&al. method, as a function of CNR; CosBOC(10, 5) modulation, 4-paths Rayleigh
a function of CNR; SinBOC(1, 1) modulation, 2-paths Rician chan- channel, with paths PDP of −1, −2, 0, and −3 dB, v = 45 km/h,
nel with decaying PDP of 0 and −2 dB, v = 25 km/h, maximum closely spaced paths xmax = 0.1 chips.
separation between paths xmax = 1 chip.

signal, for a 4-paths Rayleigh channel, with closely spaced
paths xmax = 0.1 chips and v = 45 km/h.
From all plots of Figures 20 and 21, it can be ob-
served that, in both RMSE and MTLL terms, there is a
small improvement brought by the DC HRC and SCM DC
HRC methods, which have similar performance. For the
SinBOC(1, 1) case, the performance of the Julien& al. DP
method exceeds those of HRC and SCM HRC algorithms,
which both give similar results. On the other hand, for the
CosBOC(10, 5) modulation, the Julien& al. DP method ap-
proaches the results provided by the HRC and SCM HRC
algorithms, which still offer a deterioration in performance
of about 1 dB, comparing to DC HRC and SCM DC HRC
methods.
Adina Burian et al. 17

SinBOC (1, 1), Rayleigh channel, SinBOC (1, 1), 2-paths Rician channel,
speed mobile = 3 km/h, xmax = 0.1 chips xmax = 0.5 chips, mobile speed = 45 km/h
100

10−0.3

RMSE (chips)
RMSE (chips)

10−0.5

10−0.7

10−1

10−0.9
20 25 30 35 40 20 25 30 35 40
CNR (dB-Hz) CNR (dB-Hz)

MEDLL DC IC MEDLL DC IC
SCM IC SCM DC IC SCM IC SCM DC IC
SCM thr. SCM thr.
SinBOC (1, 1), Rayleigh channel, SinBOC (1, 1), 2-paths Rician channel,
4 paths, xmax = 0.1 chips xmax = 0.5 chips, mobile speed = 45 km/h
102

101
MTLL (s)

MTLL (s)

101

100 100
20 25 30 35 40 20 25 30 35 40
CNR (dB-Hz) CNR (dB-Hz)

MEDLL DC IC MEDLL DC IC
SCM IC SCM DC IC SCM IC SCM DC IC
SCM thr. SCM thr.

Figure 22: Comparison of feedforward delays estimation algo-
rithms employing the MEDLL and IC methods and of the SCM
with threshold approach, as a function of CNR; SinBOC(1, 1) mod-
ulation, 4-paths indoor Rayleigh channel, with PDP of −1, −2, 0,
and −3 dB, v = 3 km/h, closely spaced paths with xmax = 0.1 chips.
Figure 23: Comparison of feedforward delays estimation algo-
rithms employing the MEDLL and IC methods and of the SCM with
threshold approach, as a function of CNR; SinBOC(1, 1) modula-
tion, 2-paths decaying PDP Rician channel, v = 45 km/h, xmax = 0.5
chips.

The comparisons between the introduced feedforward

delay estimation algorithms (the MEDLL method, the IC en-
hanced techniques and the SCM with threshold comparison
approach) are presented in Figures 22 to 25. In Figure 22,
the signal is SinBOC(1, 1)-modulated, with a indoor closely
spaced paths Rayleigh channel (xmax = 0.1 chips, v =
3 km/h). In Figure 23, the signal is also SinBOC(1, 1) modu-
lated, the channel is 2-paths with Rician distribution on first
path, v = 45 km/h and xmax = 0.5 chips.
In all plots the performance of MEDLL algorithm is ex-
ceeded by the other methods, since they eliminate or de-
crease the threats of the sidelobes. In terms of RMSE, for a
Rayleigh profile with closely-spaced paths (Figure 22, upper
plot), the performances of the SCM IC and DC IC algorithms
are exceeded by those of SCM DC IC and SCM thresholding
methods, for a CNR range from 20 to 30 dB-Hz. In case of
Figure 23, for a higher spacing between successive paths up
18 EURASIP Journal on Wireless Communications and Networking

CosBOC (10, 5), Rayleigh channel, CosBOC (10, 5), Rician channel, 2 paths, xmax = 0.5 chips
speed mobile = 3 km/h, xmax = 0.1 chips 102
10−0.3

10−0.4
101

MTLL (s)
10−0.5
RMSE (chips)

10−0.6
100

10−0.7

10−0.8 10−1
20 25 30 35 40
20 25 30 35 40 CNR (dB-Hz)
CNR (dB-Hz)
MEDLL DC IC
MEDLL DC IC SCM IC SCM DC IC
SCM IC SCM DC IC SCM thr.
SCM thr.
Figure 25: Comparison of feedforward delays estimation algo-
rithms employing the MEDLL and IC methods and of the SCM with
CosBOC (10, 5), 4-paths Rayleigh channel, threshold approach, as a function of CNR; CosBOC(10, 5) modula-
xmax = 0.1 chips tion, 2-paths decaying PDP Rician channel, v = 45 km/h, xmax = 0.5
102 chips.

(i.e., the highest MTLL) is provided by the SCM DC IC

and SCM thresholding algorithms, with an improvement of
101 about 4-5 dB-Hz comparing to SCM IC and DC IC methods,
MTLL (s)

which give similar results.

100
20 25 30
CNR (dB-Hz)
MEDLL DC IC
SCM IC SCM DC IC
SCM thr.
Figure 24: Comparison of feedforward delays estimation algo-
rithms employing the MEDLL and IC methods and of the SCM with
threshold approach, as a function of CNR; CosBOC(10, 5) modula-
tion, 4-paths indoor Rayleigh channel, v = 3 km/h, closely-spaced
paths xmax = 0.1 chips.
to 0.5 chips and a higher mobile speed, the SCM with thresh-
old comparison gives the best results, while the SCM IC and
SCM DC IC methods have similar performance, which is
still better then that of DC IC, for a range of about 20 to
33 dB-Hz.
In terms of MTLL, from both Figure 22 and Figure 23,
lower plots, can be concluded that the best performance
Figures 24 and 25 illustrate the obtained simulation re-
sults, for a CosBOC(10, 5)-modulated signals, for a 4-closely-
spaced paths indoor Rayleigh profile, respectively for a 2-
paths channel, with v = 45 km/h and a separation between
paths xmax of up to 0.5 chips. In terms of RMSE (Figure 24,
upper plot), the SCM DC IC method gives the best results,
35 40 followed by the SCM with threshold comparison and SCM
IC methods, for a CNR range of up to 33 dB-Hz. The good
performance of SCM DC IC method is expected, since for a
higher BOC-modulation order, it eliminates more sidelobes
than the other SCM methods (as illustrated in Figure 15).
The MEDLL technique is still outperformed by all the other
methods.
In terms of MTLL (Figure 24, lower and plot and
Figure 25), for both channel profile cases, the SCM with
threshold comparison and SCM DC IC approaches have
the best performance, while the SCM IC technique brings
an improvement over the DC IC case (in contrast with the
SinBOC(1, 1) situation, i.e., Figure 22). This is explicable,
since the SCM approach removes completely the sidelobes
situated near the main peak, while the DC method just de-
creases their amplitudes (Figure 15).
Figure 26 presents the effect of maximum separation be-
tween successive paths xmax , in case of feedback delay esti-
mation algorithms which use NEML discriminator, together
with the Julien&al. EMLP method. The channel has a 4-paths
indoor Rayleigh profile with the mobile speed of 4 km/h and
Adina Burian et al.
SinBOC (1, 1), 4-paths Rayleigh channel,
CNR = 35 dBHz, v = 4 km/h
10−0.3
10−0.4
RMSE (chips)
10−0.5
10−0.6
10−0.7
0 0.5 1 1.5
xmax (chips)
NEML DC NEML
Julien & al. EMLP SCM DC NEML
SCM NEML
Figure 26: Comparison of feedback delays estimation algorithms
employing the NEML discriminator and of the Julien&al. EMLP
method, as a function of separation between successive channel
paths xmax, in terms of RMSE; SinBOC(1, 1) modulation, 4-paths
Rayleigh channel, mobile speed 4 km/h, CNR = 35 dB-Hz.
the CNR is set to 35 dB-Hz. In this case, both SCM algo-
rithms provide a decreasing in error as xmax is increasing,
while the other methods have an almost linear behavior, for
xmax greater than half of chip. Also, it can be observed that the
same gap between the studied methods, at xmax = 0.1 chips,
is presented in Figure 19, upper plot.
6. CONCLUSIONS
A new tracking technique (the sidelobes cancellation
method) has been introduced, which removes or dimin-
ishes the sidelobes ambiguities of the BOC-modulated sig-
nals, while keeping the narrow width of the main lobe, which
is benefic for the tracking process. In contrast with other
methods, this algorithm has the advantage that can be ap-
plied to any sine or cosine, odd or even BOC-modulation
case. It also provides a lower complexity solution, since it
uses ideal reference correlation functions, which are gener-
ated only once and can be stored at receiver side. The per-
formance of the SCM algorithm can be enhanced if other
tracking-loop methods are used after removing the sidelobes
and the multipath problem can be alleviated, since the un-
desired effect of short delay multipath can be reduced. It
has been shown through extensive simulation results, that in
case of multipath fading channels, with both closely spaced
or long delayed paths, the introduced SCM algorithms bring
an improvement in performance compared to other consid-
ered delay tracking methods. The highest performance im-
provement comes when combining SCM technique with the
narrow EML correlator. The combination between HRC and
19
SCM does not bring substantial improvement, since HRC
has already rather good performance in multipath channels.
Also, the higher BOC-modulation order, the more advanta-
geous is to apply SCM technique in order to cope better with
the false lock points.
ACKNOWLEDGMENTS
This work was carried out in the project “Advanced Tech-
niques for Personal Navigation (ATENA)” funded by the
Finnish Funding Agency for Technology and Innovation
(Tekes). This work has also been supported by the Academy
of Finland. The authors would like to thank the anonymous
reviewers for their valuable comments to improve this paper.
2 REFERENCES
[1] J. Betz and D. Goldstein, “Candidate designs for an ad-
ditional civil signal in GPS spectral bands,” Tech. Rep.,
MITRE, Bedford, Mass, USA, 2002. http://www.mitre.org/
work/tech papers/tech papers 02/betz candidate/.
[2] B. Barker, J. Betz, J. Clark, et al., “Overview of the GPS M
code signal,” in CDROM Proceedings of the ION National Meet-
ing; Navigating into the New Millennium, Anaheim, Calif, USA,
January 2000.
[3] GJU, “Galileo Open Service—Signal in Space Interface Con-
trol Document (OS SIS ICD),” Galileo Joint Undertalikng
(GJU), http://www.galileoju.com/, May 2006.
[4] E. S. Lohan, A. Lakhzouri, and M. Renfors, “Feedforward delay
estimators in adverse multipath propagation for Galileo and
modernized GPS signals,” EURASIP Journal on Applied Signal
Processing, vol. 2006, Article ID 50971, 19 pages, 2006.
[5] J. Betz, “The offset carrier modulation for GPS moderniza-
tion,” in Proceedings of the National Technical Meeting of the In-
stitute of Navigation (ION-NTM ’99), pp. 639–648, San Diego,
Calif, USA, January 1999.
[6] J. Betz, “Design and performance of code tracking for the GPS
M code signal,” Tech. Rep., MITRE, Mclean, Va, USA, Septem-
ber 2000, http://www.mitre.org/work/tech papers/tech pa-
pers 00/betz codetracking/.
[7] J. Holmes, S. Raghavan, and S. Lazar, “Acquisition and track-
ing performance of NRZ and square wave modulated symbols
for use in GPS,” in Proceedings of the 54th Annual Meeting of the
Institue of Navigation, pp. 611–625, Denver, Colo, USA, June
1998.
[8] A. V. Dierendonck, P. Fenton, and T. Ford, “Theory and per-
formance of narrow correlator spacing in a GPS receiver,” Jour-
nal of the Institute of Navigation, vol. 39, no. 3, pp. 265–283,
1992.
[9] M. Irsigler and B. Eissfeller, “Comparison of multipath mit-
igation techniques with consideration of future signal struc-
tures,” in Proceedings of the International Technical Meeting
of the Institute of Navigation (ION-GPS/GNSS ’03), pp. 2584–
2592, Portland, Ore, USA, September 2003.
[10] A. McGraw and M. Braasch, “GNSS multipath mitigation
using high resolution correlator concepts,” in Proceedings of
the National Technical Meeting of the Institute of Navigation
(ION-NTM ’99), pp. 333–342, San Diego, Calif, USA, January
1999.
[11] L. Garin and J.-M. Rousseau, “Enhanced strobe correlator
multipath rejection for code and carrier,” in Proceedings of the
10th International Technical Meeting of the Satellite Division of
20 EURASIP Journal on Wireless Communications and Networking

the Institute of Navigation (ION-GPS ’97), vol. 1, pp. 559–568, R. Lucas, “Status of Galileo frequency and signal design,”
Kansas City, Mo, USA, September 1997. in CDROM Proceedings of the International Technical Meet-
[12] J. Jones, P. Fenton, and B. Smith, “Theory and perfor- ing of the Satellite Division of the Institute of Navigation
mance of the pulse aperture correlator,” Tech. Rep., NovA- (ION-GPS ’02), Portland, Ore, USA, September 2002.
tel, Calgary, Alberta, Canada, September 2004, http://www [25] E. S. Lohan, A. Lakhzouri, and M. Renfors, “Binary-offset-
.novatel.com/Documents/Papers/PAC.pdf. carrier modulation techniques with applications in satel-
[13] A. V. Dierendonck and M. Braasch, “Evaluation of GNSS re- lite navigation systems,” Wireless Communications and Mobile
ceiver correlation processing techniques for multipath and Computing, vol. 7, no. 6, pp. 767–779, 2006.
noise mitigation,” in Proceedings of the National Technical [26] E. Rebeyrol, C. Macabiau, L. Lestarquit, et al., “BOC power
Meeting of the Institute of Navigation (ION-NTM ’97), pp. spectrum densities,” in CDROM Proceedings of the National
207–215, Santa Monica, Calif, USA, January 1997. Technical Meeting of Institute of Navigation (ION-NTM ’05),
[14] C. Lee, S. Yoo, S. Yoon, and S. Y. Kim, “A novel multipath San Diego, Calif, USA, January 2005.
mitigation scheme based on slope differential of correlator
output for Galileo systems,” in Proceedings of the 8th Inter-
national Conference on Advanced Communication Technology
(ICACT ’06), vol. 2, pp. 1360–1363, Phoenix Park, Korea,
February 2006.
[15] R. van Nee, J. Siereveld, P. Fenton, and B. Townsend, “The
multipath estimating delay lock loop: approaching theoreti-
cal accuracy limits,” in Proceedings of IEEE Position Location
and Navigation Symposium, pp. 246–251, Las Vegas, Nev, USA,
April 1994.
[16] P. A. Bello and R. L. Fante, “Code tracking performance for
novel unambiguous M-code time discriminators,” in Proceed-
ings of the National Technical Meeting of the Institute of Nav-
igation (ION-NTM ’05), pp. 293–298, San Diego, Calif, USA,
January 2005.
[17] P. Fine and W. Wilson, “Tracking algorithms for GPS offset
carrier signals,” in Proceedings of the National Technical Meet-
ing of the Institute of Navigation (ION-NTM ’99), San Diego,
Calif, USA, January 1999.
[18] V. Lin, P. Dafesh, A. Wu, and C. Cahn, “Study of the impact
of false lock points in subcarrier modulated ranging signals
and recommended mitigation approaches,” in Proceedings of
the 59th ION Annual Meeting & CIGTF Guidance Test Sympo-
sium, pp. 156–165, Albuquerque, NM, USA, June 2003.
[19] P. Ward, “A design technique to remove the correlation ambi-
guity in binary offset carrier (BOC) spread spectrum signals,”
in Proceedings of the National Technical Meeting of the Institute
of Navigation (ION-NTM ’04), pp. 886–896, San Diego, Calif,
USA, January 2004.
[20] O. Julien, C. Macabiau, M. Cannon, and G. Lachapelle, “BOC
signal acquisition and tracking method and apparatus,” US
Patent Application Publication 2005/0270997 A1, December
2005.
[21] V. Heiries, J.-A. Avila-Rodriguez, M. Irsigler, G. Hein, E. Re-
beyrol, and D. Roviras, “Acquisition performance analysis of
composite signals for the L1 OS optimized signal,” in Proceed-
ings of the 18th International Technical Meeting of the Satel-
lite Division of the Institue of Navigation (ION-GNSS ’05), pp.
877–889, Long Beach, Calif, USA, September 2005.
[22] A. Schmid and A. Neubauer, “Differential correlation
for Galileo/GPS receivers,” in Proceedings IEEE Interna-
tional Conference on Acoustics, Speech, and Signal Processing
(ICASSP ’05), vol. 3, pp. 953 –956, Philadelphia, Pa, USA,
March 2005.
[23] A. Burian, E. S. Lohan, and M. Renfors, “Sidelobe cancellation
method for unambiguous tracking of binary-offset-carrier
modulated signals,” in CDROM Proceedings of the 3rd ESA
Workshop on Satellite Navigation User Equipment Technolo-
gies (NAVITEC ’06), Noordwijk, The Netherlands, December
2006.
[24] G. Hein, J. Godet, J.-L. Issler, J. C. Martin, T. Pratt, and
Hindawi Publishing Corporation
EURASIP Journal on Wireless Communications and Networking
Volume 2007, Article ID 25178, 11 pages
doi:10.1155/2007/25178

Research Article
Analysis of Filter-Bank-Based Methods for Fast Serial
Acquisition of BOC-Modulated Signals

Elena Simona Lohan

Institute of Communications Engineering, Tampere University of Technology, P.O. Box 553, 33101 Tampere, Finland

Received 29 September 2006; Accepted 27 July 2007

Recommended by Anton Donner

Binary-offset-carrier (BOC) signals, selected for Galileo and modernized GPS systems, pose significant challenges for the code ac-
quisition, due to the ambiguities (deep fades) which are present in the envelope of the correlation function (CF). This is different
from the BPSK-modulated CDMA signals, where the main correlation lobe spans over 2-chip interval, without any ambiguities or
deep fades. To deal with the ambiguities due to BOC modulation, one solution is to use lower steps of scanning the code phases
(i.e., lower than the traditional step of 0.5 chips used for BPSK-modulated CDMA signals). Lowering the time-bin steps entails
an increase in the number of timing hypotheses, and, thus, in the acquisition times. An alternative solution is to transform the
ambiguous CF into an “unambiguous” CF, via adequate filtering of the signal. A generalized class of frequency-based unambigu-
ous acquisition methods is proposed here, namely the filter-bank-based (FBB) approaches. The detailed theoretical analysis of
FBB methods is given for serial-search single-dwell acquisition in single path static channels and a comparison is made with other
ambiguous and unambiguous BOC acquisition methods existing in the literature.

Copyright © 2007 Elena Simona Lohan. This is an open access article distributed under the Creative Commons Attribution
License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly
cited.

1. INTRODUCTION
The modulation selected for modernized GPS and Galileo
signals is BOC modulation, often denoted as BOC(m, n),
with m = fsc / fref , n = fc / fref . Here, fc is the chip rate, fsc
is the subcarrier rate, and fref = 1.023 MHz is the reference
chip frequency (that of the C/A GPS signal) [1]. Alterna-
tively, a BOC-modulated signal can also be defined via its
BOC modulation order NBOC  2 fsc / fc [2–4]. Both sine and
cosine BOC variants are possible (for a detailed description
of sine and cosine BOC properties, see [3, 4]). The acqui-
sition of BOC-modulated signals is challenged by the pres-
ence of several ambiguities in CF envelope (here, CF refers to
the correlation between the received signal and the reference
BOC-modulated code). That is, if the so-called ambiguous-
BOC (aBOC) approach [5–7] is used (meaning that there
is no bandlimiting filtering at the receiver or that this filter
has a bandwidth sufficiently high to capture most energy of
the incoming signal), the resultant CF envelope will exhibit
some deep fades within ±1 chip interval around the correct
peak [5, 8], as it will be illustrated in Section 4. We remark
that sometimes the term “ambiguities” refers to the multi-
ple peaks within ±1 chip interval around the correct peak;
however, they are also related to the deep fades within this
interval. The terminology used here refers to the deep fades
of CF envelope.
The number of fades or ambiguities within 2-chip inter-
val depends on the NBOC order (e.g., for SinBOC, we have
2NBOC − 2 ambiguities around the maximum peak, while for
CosBOC, we have 2NBOC ambiguities [4]). The distance be-
tween successive ambiguities in the CF envelope sets an up-
per bound on the step of searching the time-bin hypotheses
(Δt)bin , in the sense that if the time-bin step becomes too
high, the main lobe of the CF envelope might be lost during
the acquisition. Typically, a step of one-half the distance be-
tween the correlation peak and its first zero value, or, equiva-
lently, one quarter of the main lobe width is generally consid-
ered [9]. For example, acquisition time-bin steps of 0.5 chips
are used for BPSK modulation (such as for C/A code of GPS),
where the width of the main lobe is 2 chips, and steps of 0.1–
0.2 chips are used for SinBOC(1,1) modulation, where the
width of the main lobe is about 0.7 chips (such as for Galileo
Open Service) [5, 10, 11].
In order to be able to increase the time-bin step (and,
thus, the speed of the acquisition process), several Filter-
Bank-Based (FBB) methods are proposed here, which exploit
2 EURASIP Journal on Wireless Communications and Networking
Time uncertainty Δtmax
. .
Frequency uncertainty Δ fmax
.. .. ···
···
One time-frequency bin
(Δ f )bin
···
Time-bin step (Δt)bin
Figure 1: Illustration of the time/frequency search space.
the property that by reducing the signal bandwidth before
correlation, we are able to increase the width of the CF
main lobe. A thorough theoretical model is given for the
characterization of the decision variable in single-path static
channels and the theoretical model is validated via sim-
ulations. The proposed FBB methods are compared with
two other existing methods in the literature: the classical
ambiguous-BOC processing (above-mentioned) and a more
recent, unambiguous-BOC technique, introduced by Fish-
man and Betz [9] (denoted here via B&F method, but also
known as sideband correlation method or BPSK-like tech-
nique) and further analyzed and developed in [2, 6, 7, 10, 11].
It is mentioned that FBB methods have also been studied by
the author in the context of hybrid-search acquisition [12].
However, the theoretical analysis of FBB methods is newly
introduced here.
2. ACQUISITION PROBLEM AND AMBIGUOUS
(ABOC) ACQUISITION
In Global Navigation Satellite Systems (GNSS) based on code
division multiple access (CDMA), such as Galileo and GPS
systems, the signal acquisition is a search process [13] which
requires replication of both the code and the carrier of the
space vehicle (SV) to acquire the SV signal. The range di-
mension is associated with the replica code and the Doppler
dimension is associated with the replica carrier. Therefore,
the signal match is two dimensional. The combination of
one code range search increment (code bin) and one velocity
search increment (Doppler bin) is a cell.
The time-frequency search space is illustrated in Figure 1.
The uncertainty region represents the total number of cells
(or bins) to be searched [13–15]. The cells are tested by cor-
relating the received and locally generated codes over a dwell
or integration time τ d . The whole uncertainty region in time
Δtmax is equal to the code epoch length. The length of the fre-
quency uncertainty region Δ fmax may vary according to the
initial information: if assisted-GPS data are available, Δ fmax
can be as small as couple of Hertzs or couple of tens of Hertzs.
If no Doppler-frequency information exists (i.e., no assis-
tance or autonomous GPS), the frequency range Δ fmax can
be as large as few tens of kHz [13].
The time-frequency bin defines the final time-frequency
error after the acquisition process and it is characterized by
one correlator output: the length of a bin in time direction
(or the time-bin step) is denoted by (Δt)bin (expressed in
chips) and the length of a bin in frequency direction is de-
noted by (Δ f )bin . For example, for GPS case, a typical value
for the (Δt)bin is 0.5 chips [13]. The search procedure can
be serial (if each bin is searched serially in the uncertainty
space), hybrid (if several bins are searched together), or fully
parallel (if one decision variable is formed for the whole un-
certainty space) [13]. This paper focuses on the serial search
approach.
One of the main features of Galileo system is the intro-
duction of longer codes than those used for most GPS sig-
nals. Also, the presence of BOC modulation creates some ad-
ditional peaks in the envelope of the correlation function, as
well as additional deep fades within ±1 chip from the main
peak. For this reason, a time-bin step of 0.5 chips is typically
not sufficient and smaller steps need to be used [5, 10, 11].
On the other hand, decreasing the time-bin step will increase
the mean acquisition time and the complexity of the receiver
[9].
In the serial search code acquisition process, one decision
variable is formed per each time-frequency bin (based on the
correlation between the received signal and a reference code),
and this decision variable is compared with a threshold in
order to decide whether the signal is present or absent. The
ambiguous-BOC (aBOC) processing means that, when form-
ing the decision variable, the received signal is directly corre-
lated with the reference BOC-modulated PRN sequence (all
the spectrum is used for both the received signal and refer-
ence code).
3. BENCHMARK UNAMBIGUOUS ACQUISITION:
B&F METHOD
The presence of BOC modulation in Galileo systems poses
supplementary constraints on code search strategies, due to
the ambiguities of the CF envelope. Therefore, better strate-
gies should be used to insure reasonable performance (acqui-
sition time and detection probabilities) as those obtained for
short codes. One of the proposed strategies to deal with the
ambiguities of BOC-modulated signals is the unambiguous
acquisition (known under several names, such as sideband
correlation method or BPSK-like technique).
The original unambiguous acquisition technique, pro-
posed by Fishman and Betz in [9, 16], and later modified
in [6, 10], uses a frequency approach, shown in Figure 2. In
what follows, we denote this technique via B&F technique,
from the initials of the main authors. The block diagrams of
the B&F method (single-sideband processing) is illustrated
in Figure 2, for upper sideband- (USB-) processing [9, 16].
The same is valid for the lower sideband- (LSB-) processing.
The main lobe of one of the sidebands of the received sig-
nal (upper or lower) is selected via filtering and correlated
with a reference code, with tentative delay τ and reference
Doppler frequency  fD . The reference code is obtained in a
Elena Simona Lohan
SinBOC(1, 1) spectrum
1
Normalized PSD
0.8
0.6
0.4
0.2
0
−4 −3 −2 −1 0 1
Frequecy (MHz)
Received BOC-modulated Upper sideband
signal filter
Reference BOC-modulated Upper sideband
PRN code filter
Figure 2: Block diagram of B&F method, single-sideband processing (here, upper sideband).
similar manner with the received signal, hence the autocor-
relation function is no longer the CF of a BOC-modulated
signal, but it will resemble the CF of a BPSK-modulated sig-
nal. However, the exact shape of the resulting CF is not iden-
tical with the CF of a BPSK-modulated signal, since some in-
formation is lost when filtering out the sidelobes adjacent to
the main lobe (this is exemplified in Section 4). This filtering
is needed in order to reduce the noise power. When the B&F
dual-sideband method is used, we add together the USB and
LSB outputs and form the dual-sideband statistic.
4. FILTER-BANK-BASED (FBB) METHODS
The underlying principle of the proposed FBB methods is
illustrated in Figure 3 and the block diagram is shown in
Figure 4. The number of filters in the filter bank is denoted
by Nfb and it is related to the number of frequency pieces per
sideband Npieces via: Nfb = 2Npieces if dual sideband (SB) is
used, or Nfb = Npieces if single SB is used. In Figure 3, the
upper plot shows the spectrum of a SinBOC(1,1)-modulated
signal, together with several filters (here Nfb = 4) which cover
the useful part of the signal spectrum (the useful part is con-
sidered here to be everything between the main spectral lobes
of the signal, including these main lobes). Alternatively, we
may select only the upper (or lower) SB of the signal (i.e.,
single-SB processing).
The filters may have equal or unequal frequency widths.
Two methods may be employed and they have been denoted
here via equal-power FBB (FBBep ), where each filter lets the
same signal’s spectral energy to be passed, thus they have un-
equal frequency widths (see upper plot of Figure 3), or equal-
frequency-width FBB (FBBefw ), where all the filters in the fil-
ter bank have the same bandwidth (but the signal power is
different from one band to another). An observation ought
to be made here with respect to these denominations: indeed,
before the correlation takes place and after filtering the in-
coming signal (via the filter bank), the noise power density
is exactly in reverse situation compared to the signal power,
3
2 3 4
Coherent and non
coherent integration
∗ Towards
Σ detection
stage
Upper sideband processing
Lower sideband processing
since the noise power depends on the filter bandwidth (i.e.,
the noise power is constant from one band to another for
the FBBefw case, and it is variable for the FBBep case). How-
ever, the incoming (filtered) signal is correlated with the ref-
erence BOC-modulated code. Thus, the noise, which may
be assumed white before the correlation, becomes coloured
noise after the correlation with BOC signal, and its spectrum
(after the correlation) takes the shape of the BOC power
spectral density. Therefore, after the correlation stage at the
receiver (e.g., immediately before the coherent integration
block), both signal power density and noise power density
are shaped by the BOC spectrum. Thus, the denominations
used here (FBBep and FBBefw ) are suited for both signal and
noise parts, as long as the focus is on the processing after the
correlation stage (as it is the case in the acquisition).
As seen in Figure 4, the same filter bank is applied to
both the signal and the reference BOC-modulated pseudo-
random code. Then, filtered pieces of the signal are corre-
lated with filtered pieces of the code (as shown in Figure 4)
and an example of the resultant CF is plotted in the lower
part of Figure 3. For reference purpose, also aBOC and B&F
cases are shown. It is noticed that, when Npieces = 1, the pro-
posed FBB methods (both FBBep and FBBefw ) become identi-
cal with B&F method, and the higher the Npieces is, the wider
the main lobe of the CF envelope becomes, at the expense of
a higher decrease in the signal power.
The block diagram in Figure 4 applies not only to FBB
methods, but also to other GPS/Galileo acquisition meth-
ods, such as single/dual SB, and ambiguous-/unambiguous-
BOC acquisition methods (i.e., aBOC corresponds to the
case when no filtering stage is applied to the received and
reference signals, while B&F corresponds to the case when
Npieces = 1). The complex outputs yi (·), i = 1, . . . , Nfb of the
coherent integration block of Figure 4 can be written as

  1 nT+Tcoh

yi τ, fD , n = ri (t)ci (t − τ)e j2π fD t dt, (1)
Tcoh nT
4 EURASIP Journal on Wireless Communications and Networking

0.7
Dual sideband processing, equal-power pieces nel is available (such as it is the case of Galileo L1 band), thus
the received signal r(t) (before filtering) has the form
0.6 
r(t) = Eb c(t − τ)e− j2π fD t + ηwb (t), (2)
0.5
where τ and fD are the delay and Doppler shift introduced
Spectrum

0.4 by the channel, ηwb (t) is the additive white Gaussian noise at
wideband level, and Eb is the bit energy.
0.3 The coherent integration outputs yi (·) are Gaussian pro-
cesses (since a filtered Gaussian processes is still a Gaussian
0.2
processes). Their mean is either 0 (if we are in an incorrect
time-frequency bin) or it is proportional to a time-Doppler
0.1
deterioration factor Eb F (Δτ, Δ fD ) [11], with a proportion-
0
ality constant dependent on the number of filters and of the
−3 −2 −1 0 1 2 3 acquisition algorithm, as it will be shown in Section 5. Here,
Frequency (MHz) F (·) is the amplitude deterioration in the correct bin due
Filter 3
to a residual time error Δτ and a residual Doppler error Δ fD
BOC PSD
Filter 1 Filter 4 [11]
Filter 2

 

 
  sin πΔ fD Tcoh

(a) 

F Δτ, Δ fD =
R Δτ
. (3)

Squared CF envelope, Npieces = 2, NBOC = 2 πΔ fD Tcoh

3.5

3 As mentioned above, Δτ = τ − τ, Δ fD = fD − fD , and R(Δτ)

is the CF value at delay error Δτ (CF is dependent on the used
2.5 algorithm, as shown in the lower plot of Figure 3). Moreover,
if we normalize the yi (·) variables with respect to their max-
2 imum power, the variance of yi (·) variables (in both the cor-
|CF|2

rect and incorrect bins) is proportional to the postintegration

1.5 noise variance

1 σ 2  10−(CNR+10log10 Tcoh )/10 , (4)

0.5
where CNR = Eb BW /N0 is the Carrier-to-Noise Ratio, ex-
pressed in dB-Hz [5, 7, 11], BW is the signal bandwidth after
0 despreading (e.g., BW = 1 kHz for GPS and Galileo signals),
−3 −2 −1 0 1 2 3
Delay error (chips) and N0 is the double-sided noise spectral power density in
the narrowband domain (after despreading or correlation on
BOC FBBep , dual SB 1 millisecond in GPS/Galileo). The proportionality constants
B&F, dual SB FBBefw , dual SB are presented in Section 5. The decision statistic Z of Figure 4
(b) is the output of noncoherent combining of Nnc Nfb complex
Gaussian variables, where Nnc is the noncoherent integration
Figure 3: Illustration of the FBB acquisition methods, SinBOC(1,1) time (expressed in blocks of Nc ms):
case. Upper plot: division into frequency pieces, via Nfb = 4 filters
(FBBep method). Lower plot: squared CF shapes for 2 FBB meth-
1 1 nc fb

N N
ods, compared with ambiguous BOC (aBOC) and unambiguous  

Z=
yi τ, fD , n
2 . (5)
Betz&Fishman (B&F) methods. Nnc Nfb n=1 i=1

We remark that the coloured noise impact on Z statistic is

where n is the symbol (or code epoch) index, T is the symbol
interval, ri (t) is the filtered signal via the ith filter, ci (t) is the
filtered reference code (note that the code c(t) before the filter
bank is the BOC-modulated spread spectrum sequence), τ
and fD are the receiver candidates for the delay and Doppler
shift, respectively, and Tcoh is the coherent integration length
(if the code epoch length is 1 millisecond, then the number of
coherent code epochs Nc may be used instead: Tcoh = Nc ms).
Without loss of generality, we may assume that a pilot chan-
similar with the impact of a white noise; the only difference
will be in the moments of Z, as discussed in Section 5.1 (since
a filtered Gaussian variable is still a Gaussian variable, but
with different mean and variance, according to the used fil-
ter). Thus, if those Gaussian variables have equal variances,
Z is a chi-square distributed variable [17, 18], whose num-
ber of degrees of freedom depends on the method and the
number of filters used. Next section presents the parameters
of the distribution of Z for each of the analyzed methods.
Elena Simona Lohan 5

Optional stage y1
Coherent Nnc
r1 (t) ||2 
FB integr. . . . Nfb Z
Rx sign. .. . . . 
. . .
. rNfb (t) ∗ Coherent Nnc

r(t) Nfb filters
c1 (t)
Ref code FB .
.
c(t) . cNfb (t)
Nfb filters
Figure 4: Block diagram of a generic acquisition block.
5. THEORETICAL MODEL FOR FBB
ACQUISITION METHODS
5.1. Test statistic distribution
As explained above, the test statistic Z for aBOC, B&F, and
proposed FBBep approaches1 is either a central or a noncen-
tral χ 2 -distributed variable with Ndeg degrees of freedom, ac-
cording to whether we have an incorrect (bin  H0 ) or a
correct (bin  H1 ) time-frequency bin, respectively. Its non-
centrality parameter λZ and its variance σ 2Z are thus given by
 
for single SB, therefore, the noncentrality parameter (which
λZ = ξ λbin
F Δτ, Δ fD
, is a measure √ of the amplitude, not of the signal power) in-
σ2
σ 2Z = ξ σ 2bin , a double number of noncoherent variables, thus the num-
Nnc
where F (·) is given in (3), σ 2 is given in (4), and ξ σ 2bin and
ξ λbin are two algorithm-dependent factors shown in Table 1
(they also depend on whether we are in a correct bin or in an
incorrect bin). We remark that the noncentrality parameter
used here is the square-root of the noncentrality parameter
defined in [17], such that it corresponds to amplitude lev-
els (and not to power levels). The relationship between the
distribution functions and their noncentrality parameter and
variance will be given in (8).
All the parameters in Table 1 have been derived by in-
tuitive reasoning (explained below), followed by a thorough
verification of the theoretical formulas via simulations. For
clarity reasons, we assumed that the bit energy is normalized
to Eb = 1 and all the signal and noise statistics are present
with respect to this normalization.
Clearly, for aBOC algorithm, ξ σ 2bin = 1 and the noncen-
trality factor ξ λbin is either 1 (in a correct bin) or 0 (in an in-
correct bin) [5, 7, 19]. Also, Ndeg = 2Nnc for aBOC, because
we add together the absolute-squared valued of Nnc complex
variables (or the squares 2Nnc real variables, coming from
real and imaginary parts of the correlator outputs). For B&F,
the noncentrality deterioration factor and the variance dete-
rioration factor depend on the normalized power per main
lobe (positive or negative) Pml of the BOC power spectral
1 The case of FBBefw is discussed separately, later in this section.
integr. yNfb ||2
∗
density (PSD) function. Pml can be easy computed analyti-
cally, using, for example, the formulas for PSD given in [3, 4]
and some illustrative examples are shown in Figure 5; the
normalization is done with respect to the total signal power,
thus Pml < 0.5.; Pml factor is normalized with respect to the
total signal power, thus Pml < 0.5 (e.g., Pml = 0.428 for Sin-
BOC(1,1)). The decrease in the signal and noise power after
the correlation in B&F method (and thus, the decrease in ξ λH1
and ξ σ 2bin parameters) is due to the fact that both the signal
and the reference code are filtered and the filter bandwidth is
adjusted to the width of the PSD main lobe. Also, in dual-
SB approaches, the signal power is twice the signal power
(6) creases by 2. Furthermore, in dual-SB approaches, we add
ber of degrees of freedom is doubled compared to single-SB
approaches.
The derivation of χ 2 parameters for FBBep is also straight-
forward by keeping in mind that the variance of the vari-
ables yi is constant for each frequency piece (the filters were
designed in such a way to let equal power to be passed
through them). Thus, the noise power decrease factor is
ξ σ 2bin = Pml /Npieces , bin = H0 , H1 , and the signal power de-

creases to Npieces (P 2ml /Npieces
2
), thus xλbin = Pml / Npieces for
√ 
single SB (and xλbin = 2Pml / Npieces for dual SB).
For FBBefw , the reasoning is not so straightforward (be-
cause the sum of squares of Gaussian variables of different
variances is no longer χ 2 distributed) and the bounds given
in Table 1 were obtained via simulations. It was noticed (via
simulations) that the noise variance in the correct and in-
correct bins is no longer the same. It was also noticed that
the distribution of FBBefw test statistic is bounded by two χ 2
distributions. Moreover, Pmaxpp is the maximum power per
piece (in the positive or in the negative frequency band). For
example, if Npieces = 2 and FBBefw approach is used for Sin-
BOC(1,1) case, the powers per piece of the positive-sideband
lobe are 0.10 and 0.34, respectively (hence, Pmaxpp = 0.34).
Again, these powers can be derived straightforwardly, via the
formulas shown in [1, 3, 4, 20].
Figure 6 compares the simulation-based complementary
CDF (i.e., 1-CDF) with theoretical complementary CDFs
for FBBep case (similar plots were obtained for aBOC,
B&F, and FBBefw but they are not included here due to
6 EURASIP Journal on Wireless Communications and Networking

Table 1: χ 2 parameters for the distribution of the decision variable Z, various acquisition methods.

Correct bin (hypothesis H1 ) Incorrect bin (hypothesis H0 )

ξ λH1 ξ σ 2H Ndeg ξ λH0 ξ σ 2H Ndeg
1 0

aBOC 1 1 2Nnc 0 1 2Nnc

Single-sideband
Pml Pml 2Nnc 0 Pml 2Nnc
B& F
Dual-sideband √
2Pml Pml 4Nnc 0 Pml 4Nnc
B&F
Single-sideband
FBBep and lower Pml Pml Pml
 2Nnc Npieces 0 2Nnc Npieces
bound of single- Npieces Npieces Npieces
sideband FBBefw
Dual-sideband
FBBep and lower √ P Pml Pml
2  ml 4Nnc Npieces 0 4Nnc Npieces
bound of dual- Npieces Npieces Npieces
sideband FBBefw
Upper bound of Pml Pmaxpp Pml
single-sideband  2Nnc Npieces 0 2Nnc Npieces
Npieces Npieces Npieces
FBBefw
Upper bound of √ P Pmaxpp Pml
dual-sideband 2  ml 4Nnc Npieces 0 4Nnc Npieces
Npieces Npieces Npieces
FBBefw

lack of space). For the simulations shown in Figure 6, 0.43

Power per main lobe of BOC-modulated signal
SinBOC(1,1) signal was used, with coherent integration
Power per main (positive or negative) lobe Pml

length Nc = 20 milliseconds, noncoherent integration length 0.42

Nnc = 2, CNR = 24 dB-Hz, number of samples per BOC
interval Ns = 4, and single-SB filter bank with 4 fil- 0.41
ters (i.e., Nfb = Npieces = 4). It was also noticed that
the bounds for FBBefw approach are rather loose. How- 0.4
ever, simulation results showed that the average behavior
of FBBefw , while keeping between the bounds, is also very 0.39
similar with the average behavior of FBBep [12], therefore,
from now on, it is possible to rely on FBBep curves alone 0.38
in order to illustrate the average performance of proposed
FBB methods. We remark that the plots of complementary 0.37
CDF were chosen instead of CDF, in order to show bet-
ter the tail matching of the theoretical and simulation-based 0.36
2 3 4 5 6 7 8 9 10 11 12
distributions.
BOC modulation order NBOC

5.2. Detection probability and Sine BOC

mean acquisition times Cosine BOC

Figure 5: Normalized power per main lobe Pml for BOC-modulated

In serial search acquisition, the detection probability per signals for various NBOC orders.
bin Pdbin (Δτ) is the probability that the decision variable Z
is higher than the decision threshold γ, provided that we
are in a correct bin (hypothesis H1 ). Similarly, the false where Fnc (·) is the CDF of a noncentral χ 2 variable and
alarm probability Pfa is the probability that the decision vari- Fc (·) is the CDF of a central χ 2 variable, given by [17]:
able is higher than γ, provided that we are in an incor-
Ndeg /2−1
 k
rect bin (hypothesis H0 ). These probabilities can be easily 2 z 1
computed based on the cumulative distribution functions Fc (z) = 1 − e−z/σ Z
k=0
σ 2Z k!
(CDFs) of Z in the correct Fnc (·) and incorrect bins Fc (·)
[11]: in incorrect bins H0
(8)
√ √ 
    λZ 2 2z
Pdbin Δτ, Δ fD = 1 − Fnc (γ, λZ ), Fnc z, λZ = 1 − QNdeg /2 ,
(7) σZ σZ
Pfa = 1 − Fc (γ), in correct bins H1
Elena Simona Lohan 7

Matching to χ 2 complementary CDF for SSB, FBB sition process if a false alarm state is reached), and the total
1
number of bins in the search space [21]:
0.9
   
0.8 2 + 2 − Pd (q − 1) 1 + Kpenalty Pfa
T acq = τd, (11)
2Pd
0.7
0.6 where τ d = Nnc Tcoh is the dwell time, q is the total num-
1-CDF

0.5
ber of bins in the search space, and Pd and Pfa are given by
(7) to (10). An example of the theoretical average detection
0.4 probability Pd compared with the simulation results is shown
0.3 in Figure 7, where a very good match is observed. The small
mismatch at high (Δt)bin for the dual B&F method can be ex-
0.2
plained by the number of points used in the statistics: about
0.1 5000 random points have been used to build such statistics,
0 which seemed enough for most of (Δt)bin ranges. However, at
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 very low detection probabilities, this number is still too small
Test statistic levels for a perfect match. However, the gap is not significant, and
Sim, non-central Sim, central
low Pd regions are not the most interesting from the analysis
Th, non-central Th, central point of view.
An example of performance (in terms of average and
Figure 6: Matching with χ 2 distributions, (complementary CDF: worst detection probabilities) of the proposed FBB methods
1-CDF), theory (th) versus simulations (sim), FBBep , Nfb = is given in Figure 8. The gap between proposed FBB methods
Npieces = 4. and aBOC method is even higher from the point of view of
the worst Pd . Here, SinBOC(1,1)-modulated signal was used,
and Nc = 20 ms, Nnc = 2. The other parameters are specified
with σ 2Z , Ndeg , and λZ given in (6) and in Table 1, and in the figures captions. The small edge in aBOC average per-
QNdeg /2 (·) being the generalized Marcum-Q function [17]. formance at around 0.7 chips is explained by the fact that a
Due to the fact that the time-bin step may be smaller than time-bin step equal to the width of the main lobe of CF en-
the 2-chip interval of the CF main lobe, we might have velope (i.e., about 0.7 chips) would give worse performance
several correct bins. The number of correct bins is: Nt = than a slightly higher or smaller steps, due to ambiguities in
 2Tc /(Δt)bin , where Tc is the chip interval. Thus, the global the CF envelope. Also, the relatively constant slope in the re-
detection probability Pd is the sum of probabilities of detect- gion of 0.7–1 chips can be explained by the combination of
ing the signal in the ith bin, provided that all the previous high time-bin steps and the presence of the deep fades in the
tested hypotheses for the prior correct bins gave a misdetec- CF: since the spacing between those deep fades is around 0.7
tion [11]: chips for SinBOC(1,1), then a time-bin step of 0.7 chips is the
worst possible choice in the interval up to 1 chip. However,
t −1
N there is no significant difference in average Pd for time-bin
   
Pd Δτ 0 = Pdbin Δτ 0 + k(Δt)bin , Δ fD steps between 0.7 and 1 chip, since two counter-effects are
k=0 superposed (and they seem to cancel each other in the region
k−1 (9)
    of 0.7 till 1 chip from the point of view of average Pd ): on
1 − Pdbin Δτ 0 + i(Δt)bin , Δ fD . one hand, increasing the time-bin step is deteriorating the
i=0
performance; on the other hand, avoiding (as much as possi-
ble) the deep fades of CF is beneficial. This fact is even more
In (9), Δτ 0 is the delay error associated with the first sam- visible from the lower plot of Figure 8, where worst-case Pd
pling point in the two-chip interval, where we have Nt cor- are shown. Clearly, having a time-bin step of about 0.7 chips
rect bins. Equation (9) is valid only for fixed sampling points. would mean that, in the worst case, we are always in a deep
However, due to the random nature of the channels, the sam- fade and lose completely the peak of the main lobe. This ex-
pling point (with respect to the channel delay) is randomly plains the minimum Pd achieved at such a step. Also, for steps
fluctuating, hence, the global Pd is computed as the expecta- higher than 1.5 chips, there is always a sampling sequence
tion E(·) over all possible initial delay errors (under uniform that will miss completely the main lobe of the envelope of CF
distribution, we simply take the temporal mean): (thus, the worst Pd will be zero).
   It is noticed that FBB methods can work with time-bin
Pd = EΔτ 0 Pd Δτ 0 , (10) steps higher than 1 chip, due to the increase in the main lobe
of the CF envelope. Moreover, the proposed FBB methods
and the worst detection probability is obtained for the worst (both single and dual SB) outperform the B&F and aBOC
sequence of sampling points: Pd,worst = minΔτ 0 (P d (Δτ 0 )). method if the step (Δt)bin is sufficiently high. Indeed, the
The mean acquisition time T acq for the serial search is higher the time-bin step, the higher is the improvement of
computed according to the global Pd , the false alarm Pfa , the FBB methods over aBOC and B&F methods. We remark that
penalty time Kpenalty (i.e., the time lost to restart the acqui- even at (Δt)bin = 1 chip, we have a significantly high Pd ,
8 EURASIP Journal on Wireless Communications and Networking

Pd at Pfa = 0.001, dual B&F, CNR = 27 dB-Hz Average Pd , Npieces = 2, CNR = 30 dB-Hz
100 1

0.9

0.8
10−1
0.7
Pd

Pd
0.6

0.5
10−2
0.4

0.3

10−3 0.2
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 0 0.5 1 1.5 2 2.5 3
(Δt)bin (chips) Time-bin step (Δt)bin

Sim, average Sim, worst aBOC Single FBB

Th, average Th, worst Single B&F Dual FBB
Dual B&F
(a)
(a)
Pd at Pfa = 0.001, dual FBBep , CNR = 27 dB-Hz, Npieces = 2
100 Pd,worst , Npieces = 2, CNR = 30 dB-Hz
1
0.9
0.8
0.7
0.6
10−1
Pd

Pd

0.5
0.4
0.3
0.2
0.1
10−2
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 0
(Δt)bin (chips) 0 0.5 1 1.5 2 2.5 3
Time-bin step (Δt)bin
Sim, average Sim, worst
Th, average Th, worst aBOC Single FBB
Single B&F Dual FBB
(b)
Dual B&F
Figure 7: Comparison between theory and simulations for Sin- (b)
BOC(1,1). Left: dual-sideband B&F method. Right: Dual-sideband
FBBep method, Npieces = 2. Nc = 10 milliseconds, Nnc = 5, CNR = Figure 8: Average (upper) and worst (lower) detection probabili-
27 dB-Hz, Ns = 5. ties versus (Δt)bin ambiguous and unambiguous BOC acquisition
methods (FBBep was used here).

due to the widening of the CF main lobe. The constant Pd
at higher time-bin steps is explained by the fact that, if the
step increases with respect to the correlation function width,
only noise is captured in the acquisition block. Thus, increas-
ing the step above a certain threshold would not change the
serial detection probability, since the decision variable will
only contains noise samples.
On the other hand, by increasing the time-bin step in
the acquisition process, we may decrease substantially the
mean acquisition time, because the number of bins in the
search space (see (11) is directly proportional to (Δt)bin . For
example, if the code epoch length is 1023 chips and only
one frequency bin is searched (assisted acquisition), q =
1023/(Δt)bin . Moreover, the computational load required
for implementing a correlator acquisition receiver per unit of
time uncertainty is inversely proportional to (Δt)2bin [9], thus,
when (Δt)bin increases, the computational load decreases.
An example regarding the needed time-bin step in or-
der to achieve a certain detection probability, at fixed CNR
and false alarm probability, is shown in what follows. We
Elena Simona Lohan 9

Step needed to achieve a target Pd = 0.9, (average case) Step needed to achieve a target Pd = 0.9, (average case)
2 1.5
(Δt)bin (chips)

(Δt)bin (chips)
1.5
1
1
0.5
0.5

0 0
25 26 27 28 29 30 31 25 26 27 28 29 30 31
CNR (dB-Hz) CNR (dB-Hz)

Dual SB, FBBep Single SB, FBBep

Dual SB, B&F Single SB, B&F
(a) (b)
Achieved MAT [s] at considered step Achieved MAT [s] at considered step
103 104

103
MAT

MAT
102
102

101 101
25 26 27 28 29 30 31 25 26 27 28 29 30 31
CNR (dB-Hz) CNR (dB-Hz)

Dual SB, FBB Single SB, FBBep

Dual SB, B&F Single SB, B&F
(c) (d)

Figure 9: Step needed to achieve a target average Pd = 0.9, at false alarm Pfa = 10−3 and corresponding mean acquisition time, SinBOC(1,1)
signal. Code length 4092 chips, penalty factor Kpenalty = 1, single frequency-bin. Npieces = 2 for FBBep . Left: dual sideband. Right: single
sideband.

assume a SinBOC(1,1)-modulated signal, a CNR = 30 dB-
Hz, and a target average detection probability of Pd = 0.9 at
Pfa = 10−3 . For these values, we need a step of (Δt)bin = 1.2
chips for the dual-sideband B&F method (which will cor-
respond to a mean acquisition time T acq = 86.24 s for sin-
gle frequency serial search and 4092-chip length code) and a
step of (Δt)bin = 1.7 chips for dual-sideband FBBep method
with Npieces = 2 (i.e., T acq = 58.14 s). Thus, the step can be
about 50% higher for dual-sideband FBB case than for dual-
sideband B&F case, and we may gain about 48% in the MAT
(i.e., MAT is 48% less in dual-SB FBB case than in dual-SB
B&F case). For single-sideband approaches, the differences
between FBB and B&F methods are smaller. An illustrative
plots is shown in Figure 9, where the needed steps and the
achievable mean acquisition times are given with respect to
CNR. We notice that FBB methods outperform B&F meth-
ods at high CNRs. Below a certain CNR limit (which, of
course, depends on the (Nc , Nnc ) pair), B&F method may
be better than FBB method.
The optimal number of pieces or filters to be used in the
filter bank depends on the CNR, on the method (single or
dual SB), and on the BOC modulation orders. From simu-
lation results (not included here due to lack of space), best
values between 2 and 6 have been observed. This is due to
the fact that a too high Npieces parameter would deteriorate
the signal power too much.
We remark that the choice of the penalty factor has not
been documented well in the literature. The penalty time se-
lection is in general related to the quality of the following
code tracking circuit. There is a wide range of values that
Kpenalty may take and no general rule about the choice of
Kpenalty has been given so far, to the author’s knowledge. For
example, in [22] a penalty factor Kpenalty = 1 was consid-
ered; in [23] simulations were carried out for Kpenalty = 2, in
[24] a penalty factor of Kpenalty = 103 was used, while in [25]
we have Kpenalty = 106 . Penalty factors with respect to dwell
times were also used in the literature, for example: Kpenalty =
105 /(Nc Nnc ) [26, 27], or Kpenalty = 107 /(Nc Nnc ) [27] (in our
simulations, Nc Nnc = 40 ms). Therefore, Kpenalty may spread
over an interval of [1, 106 ], therefore, in our simulations we
considered the 2 extreme cases: Kpenalty = 1 (Figure 9) and
Kpenalty = 106 (Figure 10). Figure 10 uses exactly the same
parameters as Figure 9, with the exception of the penalty
factor, which is now Kpenalty = 106 . For Kpenalty = 106 of
Figure 10, MAT for the dual-sideband B&F method becomes
T acq = 8.62 ∗ 104 , which is still higher than MAT for the
dual-sideband FBBep (T acq = 5.8 ∗ 104 s). Similar improve-
ments in MAT times via FBB processing (as for Kpenalty = 1)
are observed if we increase the penalty time.
The plots with respect to the receiver operating charac-
teristics (ROC) are shown in Figure 11 for a CNR of 30 dB-
Hz. ROC curves are obtained by plotting the misdetection
probability 1 − Pd versus false alarm probability Pfa [28]. The
lower the area below the ROC curves is, the better the per-
formance of the algorithm is. As seen in Figure 11, the dual
sideband unambiguous methods have the best performance.
10 EURASIP Journal on Wireless Communications and Networking

Achieved MAT [s] at considered step Achieved MAT [s] at considered step
106 107

106

MAT
MAT

105

105

104 104
25 26 27 28 29 30 31 25 26 27 28 29 30 31
CNR (dB-Hz) CNR (dB-Hz)

Dual SB, FBBep Single SB, FBBep

Dual SB, B&F Single SB, B&F
(a) (b)

Figure 10: Mean acquisition time corresponding to the step needed to achieve a target average Pd = 0.9, at false alarm Pfa = 10−3 , Sin-
BOC(1,1) signal. Code length 4092 chips, penalty factor Kpenalty = 106 , single frequency-bin. Npieces = 2 for FBBep . Left: dual sideband. Right:
single sideband.

ROC, (Δt)bin = 0.5 chips, CNR = 30 dB-Hz ROC, (Δt)bin = 1.5 chips, CNR = 30 dB-Hz
1 1
0.9 0.9

0.8 0.8
Mis-detection probability 1-Pd

Mis-detection probability 1-Pd

0.7 0.7
0.6 0.6
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0 0
10−10 10−8 10−6 10−4 10−2 10−10 10−8 10−6 10−4 10−2
False alarm probability Pfa False alarm probability Pfa

aBOC Single FBB aBOC Single FBB

Single BF Dual FBB Single BF Dual FBB
Dual BF Dual BF
(a) (b)

Figure 11: Receiver operating characteristic for CNR = 30 dB-Hz, SinBOC(1,1) signal, Nc = 20, Nnc = 2. Left: (Δt)bin = 0.5 chips; right
(Δt)bin = 1.5 chips.

At low time-bin steps (e.g., (Δt)bin = 0.5 chips), the FBB and pends on the CNR, on the integration times, and on the time-
B&F methods behave similarly, as it has been seen before also bin step and it is typically quite low (below 10−5 ).
in Figure 8. The main advantage of FBB methods is observed
for time-bin steps higher than one chip, as shown in the left 6. CONCLUSIONS
plot of Figure 11. For both time-bin steps considered here,
the single sideband unambiguous methods have a threshold This paper introduces a new class of code acquisition meth-
false alarm, below which their performance becomes worse ods for BOC-modulated CDMA signals, based on filter bank
than that of ambiguous BOC approach. This threshold de- processing. The detailed theoretical characterization of this
Elena Simona Lohan
new method has been given and theoretical curves were val-
idated via simulations. The performance comparison with
other methods (i.e., ambiguous BOC and Betz&Fishman
sideband correlator) showed that FBB techniques can be suc-
cessfully employed if the target is to increase the time-bin
step of the acquisition process and to minimize the mean ac-
quisition times and the computational load of the correlator.
ACKNOWLEDGMENTS
This work was carried out in the project “Advanced Tech-
niques for Personal Navigation (ATENA)” funded by the
Finnish Funding Agency for Technology and Innovation
(Tekes). This work has also been supported by the Academy
of Finland.
REFERENCES
[1] J. W. Betz, “The offset carrier modulation for GPS moderniza-
tion,” in Proceedings of the International Technical Meeting of
the Institute of Navigation (ION-NTM ’99), pp. 639–648, San
Diego, Calif, USA, January 1999.
[2] A. Burian, E. S. Lohan, and M. Renfors, “BPSK-like methods
for hybrid-search acquisition of Galileo signals,” in Proceedings
of the IEEE International Conference on Communications (ICC
’06), vol. 11, pp. 5211–5216, Istanbul, Turkey, June 2006.
[3] E. S. Lohan, A. Lakhzouri, and M. Renfors, “Binary-offset-
carrier modulation techniques with applications in satel-
lite navigation systems,” Wireless Communications and Mobile
Computing, vol. 7, no. 6, pp. 767–779, 2006.
[4] E. S. Lohan, A. Lakhzouri, and M. Renfors, “Spectral shap-
ing of Galileo signals in the presence of frequency offsets and
multipath channels,” in Proceedings of 14th IST Mobile & Wire-
less Communications Summit, Dresden, Germany, June 2005,
CDROM.
[5] S. Fischer, A. Guérin, and S. Berberich, “Acquisition concepts
for Galileo BOC(2,2) signals in consideration of hardware lim-
itations,” in Proceedings of the 59th IEEE Vehicular Technology
Conference (VTC ’04), vol. 5, pp. 2852–2856, Milan, Italy, May
2004.
[6] N. Martin, V. Leblond, G. Guillotel, and V. Heiries, “BOC(x,y)
signal acquisition techniques and performances,” in Proceed-
ings of the 16th International Technical Meeting of the Satellite
Division of the Institute of Navigation (ION GPS/GNSS ’03),
pp. 188–198, Portland, Ore, USA, September 2003.
[7] B. Bandemer, H. Denks, A. Hornbostel, A. Konovaltsev, and
P. R. Coutinho, “Performance of acquisition methods for
Galileo SW receivers,” in Proceedings of the European Navi-
gation Conference (ENC-GNSS ’05), Munich, Germany, July
2005, CDROM.
[8] B. C. Barker, J. W. Betz, J. E. Clark, et al., “Overview of the
GPS M code signal,” in Proceedings of the International Tech-
nical Meeting of the Institute of Navigation (ION-NTM ’00),
Anaheim, Calif, USA, January 2000, CDROM.
[9] P. Fishman and J. W. Betz, “Predicting performance of direct
acquisition for the M-code signal,” in Proceedings of the Inter-
national Technical Meeting of the Institute of Navigation (ION-
NTM ’00), pp. 574–582, Anaheim, Calif, USA, January 2000.
[10] V. Heiries, D. Roviras, L. Ries, and V. Calmettes, “Analysis of
non ambiguous BOC signal acquisition performance,” in Pro-
ceedings of the 18th International Technical Meeting of the Satel-
lite Division of the Institute of Navigation (ION-GNSS ’05),
Long Beach, Calif, USA, September 2005, CDROM.
11
[11] E. S. Lohan, “Statistical analysis of BPSK-like techniques for
the acquisition of Galileo signals,” in Proceedings of the 23rd
AIAA International Communication Systems Conference (IC-
SSC ’05), Rome, Italy, September 2005, CDROM.
[12] E. S. Lohan, “Filter-bank based technique for fast acquisition
of Galileo and GPS signals,” in Proceedings of the 17th IEEE
International Symposium on Personal, Indoor and Mobile Ra-
dio Communications (PIMRC ’06), pp. 1–5, Helsinki, Finland,
September 2006.
[13] E. D. Kaplan, Understanding GPS: Principles and Applications,
Artech House, London, UK, 1996.
[14] P. W. Ward, “GPS receiver search techniques,” in Proceedings
of the IEEE Position Location and Navigation Symposium, pp.
604–611, Atlanta, Ga, USA, April 1996.
[15] M. Katz, Code acquisition in advanced CDMA networks, Ph.D.
thesis, University of Oulu, Oulu, Finland, 2002.
[16] J. Betz and P. Capozza, “System for direct acquisition of re-
ceived signals,” US patent no. 2004/0071200 A1, April 2004.
[17] J. Proakis, Digital Communications, McGraw-Hill, New York,
NY, USA, 4th edition, 2001.
[18] R. R. Rick and L. B. Milstein, “Optimal decision strate-
gies for acquisition of spread-spectrum signals in frequency-
selective fading channels,” IEEE Transactions on Communica-
tions, vol. 46, no. 5, pp. 686–694, 1998.
[19] F. Bastide, O. Julien, C. Macabiau, and B. Roturier, “Analysis
of L5/E5 acquisition, tracking and data demodulation thresh-
olds,” in Proceedings of the International Technical Meeting of
the Satellite Division of the Institute of Navigation (ION-GPS
’02), pp. 2196–2207, Portland, Ore, USA, September 2002.
[20] S. H. Raghavan and J. K. Holmes, “Modeling and simulation of
mixed modulation formats for improved CDMA bandwidth
efficiency,” in Proceedings of the 60th IEEE Vehicular Technol-
ogy Conference (VTC ’04), vol. 6, pp. 4290–4295, Los Angeles,
Calif, USA, September 2004.
[21] J. Holmes and C. Chen, “Acquisition time performance of PN
spread-spectrum systems,” IEEE Transactions on Communica-
tions, vol. 25, no. 8, pp. 778–784, 1977.
[22] G. J. R. Povey, “Spread spectrum PN code acquisition using
hybrid correlator architectures,” Wireless Personal Communi-
cations, vol. 8, no. 2, pp. 151–164, 1998.
[23] W. Zhuang, “Noncoherent hybrid parallel PN code acquisition
for CDMA mobile communications,” IEEE Transactions on Ve-
hicular Technology, vol. 45, no. 4, pp. 643–656, 1996.
[24] E. A. Homier and R. A. Scholtz, “Hybrid fixed-dwell-time
search techniques for rapid acquisition of ultra-wideband sig-
nals,” in Proceedings of the International Workshop on Ultra-
Wideband Systems, Oulu, Finland, June 2003.
[25] B.-J. Kang and I.-K. Lee, “A performance comparison of code
acquisition techniques in DS-CDMA system,” Wireless Per-
sonal Communications, vol. 25, no. 2, pp. 163–176, 2003.
[26] O.-S. Shin and K. B. Lee, “Utilization of multipaths for spread-
spectrum code acquisition in frequency-selective Rayleigh fad-
ing channels,” IEEE Transactions on Communications, vol. 49,
no. 4, pp. 734–743, 2001.
[27] D. DiCarlo and C. Weber, “Multiple dwell serial search: perfor-
mance and application to direct sequence code acquisition,”
IEEE Transactions on Communications, vol. 31, no. 5, pp. 650–
659, 1983.
[28] S. Soliman, S. Glazko, and P. Agashe, “GPS receiver sensitiv-
ity enhancement in wireless applications,” in Proceedings of
the IEEE MTT-S International Tipical Symposium on Technolo-
gies for Wireless Applications, pp. 181–186, Vancouver, Canada,
February 1999.