This action might not be possible to undo. Are you sure you want to continue?
Satellite Communications
Guest Editors: Ray E. Sheriff, Anton Donner,
and Alessandro VanelliCoralli
Satellite Communications
EURASIP Journal on
Wireless Communications and Networking
Satellite Communications
Guest Editors: Ray E. Sheriff, Anton Donner,
and Alessandro VanelliCoralli
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.
EditorinChief
Luc Vandendorpe, Universit´ e 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
YuhShyan Chen, Taiwan
Pascal Chevalier, France
ChiaChin Chong, South Korea
Huaiyu Dai, USA
Soura Dasgupta, USA
Ibrahim Develi, Turkey
Petar M. Djuri´ c, 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 LeNgoc, 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 VanelliCoralli
Volume 2007, Article ID 58964, 2 pages
MultiSatellite MIMOCommunications at KuBand 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 MIMOChannel Modeling: Accent on Polarization, P´ eter Horv´ ath,
George K. Karagiannidis, Peter R. King, Stavros Stavrou, and Istv´ an 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 DVBS2 Systems in Railway Scenarios, Stefano Cioni,
Cristina P´ arraga Niebla, Gonzalo Seco Granados, Sandro Scalise, Alessandro VanelliCoralli,
and Mar´ıa Angeles V´ azquez Castro
Volume 2007, Article ID 49718, 17 pages
Capacity Versus Bit Error Rate TradeOﬀ in the DVBS2 Forward Link, Matteo Berioli,
Christian Kissling, and R´ emi Lapeyre
Volume 2007, Article ID 14798, 10 pages
Frequency Estimation in Iterative Interference Cancellation Applied to MultibeamSatellite Systems,
J. P. Millerioux, M. L. Boucheret, C. Bazile, and A. Ducasse
Volume 2007, Article ID 62310, 12 pages
A QoS Architecture for DVBRCS Next Generation Satellite Networks, Thierry Gayraud and
Pascal Berthou
Volume 2007, Article ID 58484, 9 pages
MaximumLikelihood Timing and Carrier Synchronization in BurstMode Satellite Transmissions,
Michele Morelli and Antonio A. D’Amico
Volume 2007, Article ID 65058, 8 pages
Burst Format Design for OptimumJoint Estimation of DopplerShift and DopplerRate in Packet
Satellite Communications, Luca Giugno, Francesca Zanier, and Marco Luise
Volume 2007, Article ID 29086, 12 pages
TCPCall Admission Control Interaction in MultiplatformSpace 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 BOCModulated Signals,
Adina Burian, Elena Simona Lohan, and Markku Kalevi Renfors
Volume 2007, Article ID 72626, 20 pages
Analysis of FilterBankBased Methods for Fast Serial Acquisition of BOCModulated 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 VanelliCoralli
3
1
Mobile and 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.
Multipleinput multipleoutput (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 MIMOsystems fromthe
polarization diversity point of viewand 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 dualbranch SSC receivers for diﬀerent fading
channel characteristics.
The next four papers deal with the emerging scenario
of mobile digital video broadcasting (DVBS2 and RCS mo
bile). Alternative approaches to counteracting fading chan
nels introduced when operating in a train environment re
ceiving satellite DVBS2 are presented in the paper “Ad
vanced fade countermeasures for DVBS2 systems in railway
scenarios.” Here, as a result of simulation analysis, antenna
diversity and packetlevel forward error correction mecha
nisms are proposed and their impact is evaluated with respect
to the receiver design and system complexity. The theme of
DVBS2 is continued with the paper “Capacity versus bit er
ror rate tradeoﬀ in the DVBS2 forward link,” which inves
tigates how satellite capacity can be optimised for DVBS2
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 DVBRCS with convolutional coding as an example. The
paper “A QoS architecture for DVBRCS nextgeneration
satellite networks” proceeds to design and emulate a quality
ofservice (QoS) architecture that demonstrates using real
multimedia applications how QoS can be supported over a
DVBRCS network.
2 EURASIP Journal on Wireless Communications and Networking
Synchronization aspects are dealt with in “Maximum
likelihood timing and carrier synchronization in burstmode
satellite transmissions.” The paper addresses the problem of
achieving synchronisation for a burstmode 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 Dopplershift and Doppler
rate in packet satellite communications,” which considers
optimising the burstformat of packetoriented transmis
sions by proposing verylowcomplexity algorithms for car
rier Dopplershift and Dopplerrate estimation.
A network comprising satellite and highaltitude plat
forms is considered in the paper “TCPcall 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 BOCmodulated 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 BOCmodulated
signal by using sidelobe suppression techniques. An alterna
tive approach based on ﬁlter bank processing is presented in
“Analysis of ﬁlterbankbased methods for fast serial acqui
sition of BOCmodulated 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 VanelliCoralli
Hindawi Publishing Corporation
EURASIP Journal on Wireless Communications and Networking
Volume 2007, Article ID 59608, 11 pages
doi:10.1155/2007/59608
Research Article
MultiSatellite MIMOCommunications at KuBand 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 VanelliCoralli
This paper investigates the applicability of multipleinput multipleoutput (MIMO) technology to satellite communications at the
Kuband 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 singleuser 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 lineofsight 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 multisatellite MIMO systems to relevant
singleinput singleoutput (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
Multipleinput multipleoutput (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
MIMO both as a research topic and as a commercially viable
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 earthspace link, lineofsight (LOS) operation is usually
deemed a practical necessity. However, this is not the typ
ical case in terrestrial communications where rich scatter
ing and nonLOS 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
tipleinput 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 Sband [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 asynchronismof 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 dualorthogonal
polarized satellite communicates with a single TS equipped
with a dualorthogonal 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 Sband 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 highfrequency 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 multisatellite systems, satellite diversity can
result in mfold capacity increase, where m is the number of
satellites occupied.
This paper focuses particularly on dualsatellite MIMO
communication systems employing satellite diversity. More
over, emphasis is put on the less congested highfrequency
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 ITUR
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
paralleltoserial 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]. Wellknown 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 wellestablished ﬁeld
Konstantinos P. Liolis et al. 3
S
1
S
2
d
1
, A
R1
d
2
, A
R2
Δθ
TS
(a)
T
o
S
1
T
o
S
2
ϕ
2
ϕ
1
TS
(b)
Figure 1: (a) Conﬁguration of a dualsatellite 2 ×2 MIMO channel. Individual satellites S
1
and S
2
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 lowcost, lowcomplexity 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 communicationoriented 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 signaltointerference ratio (SIR) introduced by the
rainfall spatial inhomogeneity, the concept of unacceptable
interference probability
1
[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 MIMOsatellite com
munications at the Kuband and above. Section 3 provides a
communicationbased capacity analysis for a satellite 2 × 2
MIMO spatial multiplexing system. A propagationoriented
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.
analysis for the possible interference mitigation achieved by a
satellite 2×2 MIMOdiversity systemwith 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. MIMOSATELLITE CHANNEL MODEL
Figure 1 depicts the conﬁguration of a dualsatellite MIMO
communication channel at the Kuband and above. The TS
is equipped with two colocated highly directive antennas and
communicates with two satellites, S
1
and S
2
, 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 G
R
(·), is compatible with the ITUR speciﬁcations
[25] and is shown in Figure 2.
2
The lengths of slant paths
S
i
TS are denoted by d
i
(i = 1, 2) and the random variables
(RVs) associated with the respective rain induced attenua
tions (in dB) are denoted by A
Ri
(i = 1, 2). In general, the
two slant paths S
i
TS have diﬀerent elevation angles denoted
by φ
i
(i = 1, 2), respectively.
Assuming that clear LOS between the TS and each satel
lite S
i
exists, that each TS antenna is at boresight with the
corresponding satellite S
i
(i = 1, 2) and that rain attenuation
is the major fading mechanism, the path gain for each S
i
TS
link is modeled as
g
i
∝G
R
_
0
◦
_
· d
−2
i
· 10
−ARi
/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
−100 −80 −60 −40 −20 0 20 40 60 80 100
Oﬀaxis angle (deg)
−40
−35
−30
−25
−20
−15
−10
−5
0
T
S
a
n
t
e
n
n
a
n
o
r
m
a
l
i
z
e
d
g
a
i
n
,
G
R
(
d
B
)
Figure 2: Normalized radiation pattern of each TS antenna com
patible with ITUR speciﬁcations [25].
Hence, the total path loss along each S
i
TS link (in dB) is
A
i
= FSL
i
+ A
Ri
(i = 1, 2), (2)
where FSL
i
= 10 log
10
(4πd
i
f /c)
2
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 A
Ri
(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 A
Ri
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 ρ
12
versus Δθ for a
dualsatellite MIMO channel operating in Atlanta, GA, USA
at the Kaband with satellite elevation angles φ
1
= 45
◦
and
φ
2
= 40
◦
.
Based on the above and, also, assuming frequency nonse
lective fading, the resulting MIMO channel matrix His given
by
H =
_
h
11
h
12
h
21
h
22
_
=
⎡
⎢
⎢
⎣
√
g
1
exp
_
j2πd
1
f
c
_
0
0
√
g
2
exp
_
j2πd
2
f
c
_
⎤
⎥
⎥
⎦
.
(3)
The diagonal structure of H is due to the high directivity of
the TS antennas and the large value of Δθ. In MIMO ter
minology, channels with diagonal H matrix are known as
parallel MIMO channels. Further details about such chan
nels can be found in [26]. Moreover, as opposed to standard
MIMO theory [1, 2], H is not normalized here (i.e., ordered
MIMO channel) due to the diﬀerent slant path lengths d
i
0 20 40 60 80 100 120 140 160 180
Angular separation, Δθ (deg)
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
S
p
a
t
i
a
l
c
o
r
r
e
l
a
t
i
o
n
c
o
e
ﬃ
c
i
e
n
t
d
u
e
t
o
r
a
i
n
,
ρ
1
2
Figure 3: Spatial correlation coeﬃcient due to rain ρ
12
versus an
gular separation Δθ for a dualsatellite MIMO channel operating in
Atlanta, GA, at the Kaband with satellite elevation angles φ
1
= 45
◦
and φ
2
= 40
◦
.
(i = 1, 2). Finally, the assumption of independent identically
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 MIMOSPATIAL MULTIPLEXINGSYSTEM:
CAPACITY ANALYSIS
In this Section, the two satellites S
i
(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 wellknown formula for the capacity (in
bps/Hz) of MIMO channels is adopted [19, 20]:
C = log
2
det
_
I
2
+
P
T
2N
0
HH
H
_
=
2
_
i=1
log
2
_
1 +
P
T
2N
0
λ
i
_
,
(4)
where I
2
is the 2 × 2 identity matrix, P
T
the total average
power available at the transmit side,
3
N
0
the noise spectral
3
Note that P
T
is the sum transmit power of all transmitting satellites S
i
re
gardless of their number. This means that in both the dualsatellite MIMO
case and the single satellite SISO case, the total available transmit power
is constant and equal to P
T
. This is ensured employing the normalization
factor “2” in (4), which allows for a fair comparison between the relevant
MIMO and SISO cases.
Konstantinos P. Liolis et al. 5
density at the TS receiver input, and λ
i
(i = 1, 2) the positive
eigenvalues of the matrix HH
H
(the superscript
H
stands for
conjugate transposition).
Taking into account the channel modeling assumptions,
(4) is written as
C =
2
_
i=1
log
2
_
1 + 0.5SNR
CSi
10
−ARi /10
_
, (5)
where SNR
CSi
(i = 1, 2) are the nominal SNR values under
clear sky conditions. Based on the path gain model given in
(1), the SNR
CSi
values (in dB) are related through
SNR
CS1
−SNR
CS2
= 20 log
10
_
d
2
d
1
_
. (6)
Equation (5) provides an expression for the instantaneous
capacity of a deterministic 2 × 2 MIMO channel H. How
ever, since the rainfall introduces slow fading and stochastic
behaviour over the channel H, the appropriate statistic mea
sure to characterize the resulting fading channel is the outage
capacity deﬁned by [1]
P
_
C ≤ C
out,q
_
= q, (7)
where C
out,q
is the information rate guaranteed for (1−
q)100% of the channel realizations.
Consider the RV transformation
u
i
=
_
ln
_
A
Ri
_
−ln
_
A
mRi
__
S
aRi
(i = 1, 2) (8)
which relates the lognormal rain attenuation RVs A
Ri
(i =
1, 2) to the normalized normal RVs u
i
(i = 1, 2). Substituting
(5) into (7) and after some straightforward algebra, the fol
lowing analytical closed form expression for the outage ca
pacity is obtained:
P
_
C ≤ C
out,q
_
=
1
2
_
+∞
uA
du
1
f
U1
_
u
1
_
erfc
_
u
B
−ρ
n12
u
1
_
2
_
1 −ρ
2
n12
_
_
= q,
(9)
where erfc(·) is the complementary error function, f
U1
(u
1
)
the probability density function (pdf) of the normal distri
bution, ρ
n12
the logarithmic correlation coeﬃcient between
the normal RVs u
i
(i = 1, 2) [17] and u
A
, u
B
are analytically
given by
u
A
=
_
ln
_
10 log
10
_
0.5SNR
CS2
_
−10 log
10
_
2
Cout,q
−1
__
−ln
_
A
mR2
_
_
_
S
aR2
,
(10)
u
B
=
_
ln
_
10 log
10
_
0.5SNR
CS1
_
+ 10 log
10
_
1 + 0.5SNR
CS2
10
−Am
R2
exp(u1Sa
R2
)/10
_
−10 log
10
_
2
Cout,q
−1−0.5SNR
CS2
10
−Am
R2
exp(u1Sa
R2
)/10
__
−ln
_
A
mR1
_
_
_
S
aR1
.
(11)
The quantities A
mRi
, S
aRi
(i = 1, 2), encountered in (8)–(11),
are the statistical parameters of the lognormal RVs A
Ri
(i =
1, 2) given by [17]
S
2
aRi
= ln
_
1 +
H
i
L
2
Di
_
exp
_
b
2
S
2
r
_
−1
_
_
(i = 1, 2),
A
mRi
= aR
b
m
L
Di
exp
_
b
2
S
2
r
−S
2
aRi
2
_
(i = 1, 2),
(12)
where L
Di
(i = 1, 2) are the projections of the eﬀective path
lengths L
i
(i = 1, 2) [17] on the earth surface, H
i
(i = 1, 2) are
spatial parameters related to each path of length L
Di
(i = 1, 2)
which may be found in [17], and a, b are constants depend
ing on the operating frequency f , the polarization tilt angle,
the temperature, and the rainfall characteristics over the ser
viced area. R
m
, S
r
are the lognormal statistical parameters of
the rainfall rate R (in mm/hr). A reliable database of rainfall
statistics for any geographical location on earth is provided
by ITUR in [18] and is used throughout the present work as
an input to the simulations performed in order to determine
the values of R
m
, S
r
.
4. SATELLITE MIMODIVERSITY SYSTEM
WITHRECEIVE ANTENNA SELECTION:
INTERFERENCE ANALYSIS
In this section, the two satellites S
i
(i = 1, 2) depicted in
Figure 1 are assumed to transmit the same signal over the
(ideally) independently fading paths S
i
TS (i = 1, 2) (i.e., di
versity is investigated). To alleviate the high cost and com
plexity associated with multiple RF chains, the dualantenna
TS receiver is equipped with only one RF chain and performs
antenna selection, that is, the 2 × 2 MIMO satellite system
assumed employs receive selection diversity [21]. Therefore,
the TS receiver detects the signal related to the path with the
highest SNR. Under the constraint of only one RF chain at
the receiver, in order to know all SNRs simultaneously for
optimal selection, a training signal in a preamble to the trans
mitted data is assumed. During this preamble, the TS receiver
scans the two antennas, ﬁnds that one with the highest SNR,
and selects it for reception of the next data burst. Thus, only
a few more training bits are required instead of additional RF
chains.
Particular emphasis is put on possible interference mit
igation oﬀered by the proposed satellite 2 × 2 MIMO di
versity system. In this regard, a propagationbased analy
sis is performed which is quite diﬀerent from conventional
communicationoriented approaches followed in standard
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
as the dominant cause of the SIR degradation [22]. Such an
interference problem is further aggravated due to the spa
tial inhomogeneity of the rainfall medium. It constitutes a
typical interference scenario, especially over congested urban
areas, where the increased demand for link capacity and ra
dio coverage imposes the coexistence of many satellite radio
links over the same geographical and spectral area. In the fol
lowing, an analytical prediction model is presented, which
6 EURASIP Journal on Wireless Communications and Networking
S
1
S
3
S
2
d
1
, A
R1
d
3
, A
R3
d
2
, A
R2
Δθ
Δψ
TS
(a)
T
o
S
1
T
o
S
3
T
o
S
2
ϕ
3
ϕ
2
ϕ
1
TS
(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
the proposed 2 × 2 MIMO system with respect to the corre
sponding SISO one.
Figure 4 depicts the conﬁguration of the assumed inter
ference scenario on the forward link of a satellite 2×2 MIMO
diversity system operating at the Kuband and above and
employing receive antenna selection. The satellites S
1
and
S
2
constitute the dualsatellite transmit part of the MIMO
system also depicted in Figure 1. Another cochannel satellite
(denoted by S
3
), which may belong to either the same or to
another satellite network, is close in orbit to S
1
. Thus, CCI
problems may arise on the forward link of the 2 × 2 MIMO
satellite system. S
1
and S
3
subtend an angle Δψ to TS. The
length of the slant path S
3
TS is denoted by d
3
, while its ele
vation angle is φ
3
. The RV associated with the rain induced
attenuation along the interfering path S
3
TS (in dB) is de
noted by A
R3
.
Due to selection diversity at the TS receiver, the antenna
with the maximum SNR is selected. In mathematical terms,
the same statement is expressed as
SNR
out
= max
_
SNR
1
, SNR
2
_
⇐⇒A
out
= min
_
A
1
, A
2
_
,
(13)
where SNR
i
= SNR
CSi
−A
Ri
(i = 1, 2) is the SNR at each TS
antenna under rain fades and A
i
(i = 1, 2) the total path loss
along each S
i
TS link (i = 1, 2). SNR
out
corresponds to A
out
which determines the output of the selection combiner at ev
ery instant. The proposed scheme requires only the knowl
edge of the wanted signals’ channels at the receiver, whereas
knowledge of the interferer’s channel is not necessary. More
over, no CSI is required at the transmit side. If M
d
denotes
the diversity system margin associated with the system avail
ability p
avail
(see the appendix), the satellite MIMO diversity
system is considered available when the probabilistic event
Ω =
_
A
out
< M
d
_
(14)
is true. Assuming that
Ω
i
=
_
A
i
< M
d
, A
i
< A
j
_ _
(i, j) = (1, 2), (2, 1)
_
(15)
denotes the event that “the TS is serviced by the correspond
ing satellite S
i
(i = 1, 2),” it becomes clear that, due to selec
tion diversity,
Ω = Ω
1
∪Ω
2
,
Ω
1
∩Ω
2
= ∅.
(16)
Therefore, the probability that the system is available (see the
appendix) can be expressed as
P(Ω) = P
_
Ω
1
_
+ P
_
Ω
2
_
. (17)
While the satellite 2 × 2 MIMO diversity system is avail
able (i.e., when either Ω
1
or Ω
2
are true), it might suﬀer from
CCI originating from the adjacent satellite S
3
. If SIR
d
and r
d
denote the SIR and the minimum acceptable SIR threshold
of the MIMO diversity system, respectively (both measured
at the output of the TS selection combiner), the probability
of the event that “the system is interfered while being avail
able” can be mathematically expressed based on the above
considerations as
UIP
d
= P
_
SIR
d
< r
d
, Ω
_
= P
_
SIR
d1
< r
d
, Ω
1
_
+ P
_
SIR
d2
< r
d
, Ω
2
_
= P
1
+ P
2
,
(18)
where UIP
d
is the socalled unacceptable interference proba
bility (UIP) [23, 24], and the quantities SIR
di
(i = 1, 2) are
expressed (in dB) as
SIR
d
= SIR
di
= SIR
CSi
−A
Ri
+ A
R3
(i = 1, 2). (19)
In (19), SIR
CSi
(i = 1, 2) is the nominal SIR value under
clear sky conditions. In propagation terminology, A
Ri
− A
R3
Konstantinos P. Liolis et al. 7
(i = 1, 2) is known as the diﬀerential rain attenuation (DRA)
[22]. Based on (19), when DRA becomes suﬃciently large
due to the spatial inhomogeneity of the rainfall medium, se
vere CCI problems may arise aggravating the SIR
d
distribu
tion on the forward link of the proposed satellite 2×2 MIMO
diversity system. To this end, UIP
d
is proposed as an eﬃcient
metric to deal with the statistical behaviour of the SIR
d
and,
together with r
d
, they constitute a pair of design speciﬁca
tions concerning interference. Every user must comply with
these speciﬁcations, given the QoS speciﬁed by the event Ω
related to the system availability (see the appendix).
The quantities SIR
CSi
(i = 1, 2) encountered in (19) are
given by
SIR
CSi
= SIR
∗
i
−G
R
_
θ
i
_
(i = 1, 2), (20)
where θ
i
(i = 1, 2) are the oﬀaxis angles formed by the in
terfering link S
3
TS and the wanted links S
i
TS (i = 1, 2) in
the radiation pattern of the TS antennas. From Figure 4, it
follows that θ
1
= Δψ and θ
2
= Δθ −Δψ. Also, in (20), SIR
∗
i
(i = 1, 2) are the relevant SIR values of the interfered links
S
i
TS (i = 1, 2) when θ
i
= 1
◦
, and correspond to the nominal
CCI levels. Based on the channel model assumed, their inter
relationship is deﬁned through (6) by simply substituting the
SNR
CSi
by SIR
∗
i
.
Extending the transformation given in (8) to include also
the interfering link S
3
TS (i.e., for i = 1, 2, 3) and making the
channel modeling assumptions, the probabilities P
i
(i = 1, 2)
encountered in (18) after some straightforward algebra are
evaluated, that is,
P
i
=
_
uDi
uCi
du
1
_
+∞
u1
du
2
f
U1U2
_
u
1
, u
2
_
×
_
1 −
1
2
erfc
_
u
Ei
−μ
3/1,2
√
2σ
3/1,2
__
(i = 1, 2),
(21)
where f
U1U2
(u
1
, u
2
) is the pdf of the twodimensional joint
normal distribution.
For i = 1, 2, the rest of the parameters encountered in
(21) are
u
Ci
=
_
ln
_
x
di
_
−ln
_
A
mRi
__
S
aRi
,
x
di
=
⎧
⎪
⎪
⎨
⎪
⎪
⎩
0, r
d
> SIR
CSi
,
_
SIR
CSi
−r
d
_
cos φ
i
, SIR
CSi
+FSL
i
−M
d
<r
d
≤SIR
CSi
,
_
M
d
−FSL
i
_
cos φ
i
, r
d
≤ SIR
CSi
+ FSL
i
−M
d
,
u
Di
=
_
ln
__
M
d
−FSL
i
_
cos φ
i
_
−ln
_
A
mRi
__
S
aRi
,
u
Ei
=
_
ln
__
exp
_
u
i
S
aRi
_
A
mRi
cos φ
i
−SIR
CSi
+ r
d
_
cos φ
3
_
−ln
_
A
mR3
_
_
_
S
aR3
.
(22)
A
mRi
, S
aRi
(i = 1, 2, 3) are analytically given in (12). Fur
thermore, μ
3/1,2
and σ
3/1,2
are the statistical parameters of
the conditional distribution of the normal RV u
3
given
the other two normal RVs u
1
, u
2
and can be expressed in
terms of the logarithmic correlation coeﬃcients ρ
ni j
((i, j) =
(1, 2), (1, 3), (2, 3)) as [17, 27]
μ
3/1,2
=
ρ
n13
−ρ
n12
ρ
n23
1 −ρ
2
n12
u
1
+
ρ
n23
−ρ
n12
ρ
n13
1 −ρ
2
n12
u
2
,
σ
2
3/1,2
=
1 −ρ
2
n12
−ρ
2
n13
−ρ
2
n23
+ 2ρ
n12
ρ
n13
ρ
n23
1 −ρ
2
n12
.
(23)
5. NUMERICAL RESULTS ANDDISCUSSION
The previous analyses have been applied for the prediction
of possible capacity improvement and interference mitiga
tion achieved by the proposed satellite 2 × 2 MIMO spa
tial multiplexing and diversity systems, respectively, and for
comparison to the relevant SISO cases. To this end, the base
line conﬁguration scenario considers a TS located in At
lanta, GA, and communicating with geostationary satellites
S
1
(φ
1
= 45
◦
) and S
2
(φ
2
= 40
◦
). The angular separation as
sumed is Δθ=40
◦
, which results in a spatial correlation coef
ﬁcient of rain attenuation ρ
12
= 0.6 (see Figure 3). Moreover,
regarding the interference scenario, an adjacent geostation
ary satellite S
3
(φ
3
= 45
◦
), separated from S
1
by Δψ=10
◦
, is
considered to cause CCI problems on the forward link of the
satellite 2 ×2 MIMO diversity system.
First, the validity of the proposed analytical model in (9),
predicting the outage capacity achieved by a satellite 2 × 2
MIMO spatial multiplexing system, is numerically veriﬁed.
The eﬀect of various geometrical and operational system pa
rameters on the outage capacity distribution is also exam
ined.
Figure 5 shows the dependence of the 1% outage capac
ity of the assumed 2 ×2 MIMO satellite system on the SNR.
4
The baseline conﬁguration scenario is adopted, whereas the
operating frequency band assumed is Ka (i.e., f = 20 GHz).
For the sake of comparison, the capacity of the relevant SISO
system is also plotted. Together with the analytical results
obtained from the analytical closed form expression in (9),
Monte Carlo simulation results are also plotted for veriﬁca
tion. The agreement observed between the analytical and the
simulation results is very good over the whole SNR range.
As can be seen, the diﬀerence between the relevant MIMO
and SISO curves diminishes at very low SNR levels while
it becomes signiﬁcant as the SNR increases. As an illustra
tion, for SNR = 10 dB, the spectral eﬃciency achieved by
the MIMO system is 4.84 bps/Hz, whereas the one achieved
by the SISO system is 3.23 bps/Hz. This constitutes, approx
imately, a 50% increase in user data rate obtained by MIMO
spatial multiplexing. For SNR = 20 dB, the respective per
formance ﬁgures obtained are 10.95 bps/Hz and 6.41 bps/Hz
corresponding to, approximately, a 71% increase in user data
4
Note that the clear sky SNR of strong eigenmode, SNR
CS1
, has been par
ticularly considered. However, due to the enormous slant path lengths as
sociated, the resulting diﬀerence between SNR
CSi
(i = 1, 2) is minimum
see (6) and, therefore, any of the two SNR
CSi
can be used as xcoordinates.
8 EURASIP Journal on Wireless Communications and Networking
0 5 10 15 20 25 30
SNR (dB)
0
2
4
6
8
10
12
14
16
18
1
%
o
u
t
a
g
e
c
a
p
a
c
i
t
y
(
b
p
s
/
H
z
)
Analytical expression (9)
Monte Carlo simulation
2 ×2 MIMO
SISO
Figure 5: 1% outage capacity versus SNR for a satellite 2×2 MIMO
spatial multiplexing system. Relevant SISO case is also plotted for
comparison. Veriﬁcation of analytical closed form expression in (9)
through Monte Carlo simulation.
0 5 10 15 20 25 30
SNR (dB)
0
2
4
6
8
10
12
14
16
18
O
u
t
a
g
e
c
a
p
a
c
i
t
y
a
c
h
i
e
v
e
d
b
y
2
×
2
M
I
M
O
s
y
s
t
e
m
(
b
p
s
/
H
z
)
q = 1%, Δθ = 40
◦
, Kaband, Atlanta
q = 0.1%, Δθ = 40
◦
, Kaband, Atlanta
q = 1%, Δθ = 40
◦
, Kuband, Atlanta
q = 1%, Δθ = 40
◦
, Kaband, Singapore
q = 1%, Δθ = 60
◦
, Kaband, 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
tem on the SNR, the angular separation Δθ, the operating
frequency f , the capacity outage probability q, and the cli
matic conditions over the serviced area. All the results pre
sented here have been obtained employing (9). The baseline
conﬁguration scenario is adopted. The rest of the relevant pa
rameters assumed as well as the deviations from the baseline
scenario are indicated on Figure 6. As can be seen, as either q
decreases or f increases or as the rain conditions over the ser
viced area become heavier, the rain fading becomes more se
vere and, therefore, the outage capacity achieved by the 2 ×2
MIMO satellite system decreases. Moreover, as the angular
separation Δθ increases (from 40
◦
to 60
◦
), the spatial corre
lation coeﬃcient due to rainfall medium ρ
12
decreases cor
respondingly (from 0.6 to 0.5, see Figure 3), and the outage
capacity achieved increases.
In the following, the proposed analytical model in (21)
predicting the interference mitigation achieved by a satellite
2 × 2 MIMO diversity system with receive antenna selection
is numerically veriﬁed. The eﬀect of various geometrical and
operational system parameters on the forward link SIR dis
tribution is also examined.
Figure 7 shows the dependence of the UIP of the assumed
2 × 2 MIMO satellite system on the SIR, the system avail
ability p
avail
, and the operating frequency band. Particularly,
two diﬀerent values of system availability, p
avail
= 99.9%
and 99.99%, and two diﬀerent operating frequencies, f =
12 GHz and 20 GHz, are assumed. For the sake of compar
ison, the UIP of the relevant SISO systems is also plotted.
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
indicated on Figure 7. It is obvious that, due to rain, an SIR
degradation is observed for the same UIP level, which be
comes more severe as either p
avail
or f increases. This fur
ther indicates that satellite systems operating at higher avail
abilities or at higherfrequency bands are more sensitive to
interference. The SIR improvement achieved by the satellite
2 × 2 MIMO diversity system over the SISO one is signiﬁ
cant, especially for high p
avail
and high f . As an illustration,
for UIP = 0.001%, the interference mitigation obtained is
0.67 dB at the Kaband and for a 99.9% availability, 1.60 dB
at the Kuband and for a 99.99% availability, and 3.52 dB at
the Kaband 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
2 4 6 8 10 12 14 16 18 20
SIR (dB)
10
−6
10
−5
10
−4
10
−3
10
−2
10
−1
10
0
U
n
a
c
c
e
p
t
a
b
l
e
i
n
t
e
r
f
e
r
e
n
c
e
p
r
o
b
a
b
i
l
i
t
y
(
U
I
P
)
SISO
2 ×2 MIMO
Kaband,
p
avail
= 99.9%
Kuband,
p
avail
= 99.99%
Kaband,
p
avail
= 99.99%
Figure 7: UIP versus SIR for a satellite 2 × 2 MIMO diversity sys
tem employing receive antenna selection. Relevant SISO case is also
plotted for comparison. Eﬀect of system availability p
avail
, operating
frequency f , and rain climatic conditions over the serviced area.
20 30 40 50 60
70
80 90
Angular separation, Δθ (deg)
0
0.5
1
1.5
2
2.5
S
I
R
i
m
p
r
o
v
e
m
e
n
t
a
c
h
i
e
v
e
d
t
h
r
o
u
g
h
M
I
M
O
d
i
v
e
r
s
i
t
y
(
d
B
)
Atlanta, GA
Athens, GR
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 SISOcase are considered in Figure 9. The input parame
7 9 11 13 15 17 19 20
SIR (dB)
10
−6
10
−5
10
−4
10
−3
10
−2
10
−1
10
0
U
n
a
c
c
e
p
t
a
b
l
e
i
n
t
e
r
f
e
r
e
n
c
e
p
r
o
b
a
b
i
l
i
t
y
(
U
I
P
)
SISO
2 ×2 MIMO
Δψ = 4
◦
Δψ = 5
◦
Figure 9: UIP versus SIR for a satellite 2 × 2 MIMO diversity sys
tem employing receive antenna selection. Relevant SISO case is also
plotted for comparison. Eﬀect of angular separation ΔΨ.
ters assumed are the same as those in the baseline conﬁgura
tion scenario, with the exception of a diﬀerent angular sepa
ration Δψ, that is, Δψ = 5
◦
is now assumed. Operation of the
system at the Kaband and for a 99.99% availability is con
sidered. To obtain the necessary QoS for UIP = 0.001%, sup
pose that an SIR threshold of 10 dB must be overcome. In the
SISO case, when the angular separation between the wanted
satellite S
1
and the adjacent interfering one S
3
is Δψ = 5
◦
,
an SIR level of 11.2 dB is obtained for UIP = 0.001%, thus
satisfying the QoS requirement. If the interfering satellite S
3
is closer in orbit to S
1
, so that their angular separation is re
duced to Δψ = 4
◦
, the SIR level in the SISO case falls down
to 9.8 dB, thus failing to satisfy the QoS requirement. Em
ploying the proposed 2 × 2 MIMO satellite system, the SIR
achieved when Δψ = 4
◦
is 11.32 dB, thus remaining above
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 Kuband 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
ITUR 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
applications of MIMO technology:
(i) capacity improvement in a satellite 2×2 MIMO spatial
multiplexing system,
(ii) interference mitigation in a satellite 2 × 2 MIMO di
versity system with receive antenna selection.
In the ﬁrst application, signiﬁcant capacity gains of the
MIMO system over the relevant SISO one are demonstrated,
especially for moderate and high SNR levels. The practical
case when no CSI is available at the transmitters of the two
individual satellites is considered. A useful closed form ex
pression for the outage capacity achieved by 2 × 2 MIMO
satellite systems is provided and successfully veriﬁed through
Monte Carlo simulations. Such an expression is extremely
hard to obtain even in the wellestablished ﬁeld of MIMO
theory, is applicable over a large SNRrange, and can incorpo
rate the eﬀect of various geometrical and operational system
parameters on the outage capacity distribution.
In the second application, the receive antenna selection
scheme employed in the satellite MIMO system assumed is
considered to counteract CCI problems over its forward link.
SIR gain of several dB is demonstrated in the numerical re
sults. An analytical propagation model for the calculation of
the interference mitigation achieved is presented, which is
ﬂexible and can incorporate the inﬂuence of various geomet
rical and operational system parameters on the SIR distribu
tion.
APPENDIX
CALCULATIONOF SATELLITE 2 ×2 MIMO
DIVERSITY SYSTEMMARGINM
d
Every user in the assumed satellite 2×2 MIMO diversity sys
tem employing receive antenna selection must comply with a
certain availability percentage p
avail
related to a diversity sys
tem margin M
d
:
p
avail
· 100%= P(Ω) = P
_
A
out
< M
d
_
= P
_
min
_
A
1
, A
2
_
< M
d
_
= 1 −P
_
A
1
> M
d
, A
2
> M
d
_
= 1 −P
_
A
R1
> M
d
−FSL
1
, A
R2
> M
d
−FSL
2
_
.
(.1)
Considering the transformation given in (8), relating the log
normal rain attenuation RVs A
Ri
(i = 1, 2) to the normalized
normal RVs u
i
(i = 1, 2), and the channel modeling assump
tions, p
avail
is expressed as
p
avail
· 100%= 1 −
_
+∞
uF1
du
1
_
+∞
uF2
du
2
f
U1U2
_
u
1
, u
2
_
, (.2)
where
u
Fi
=
_
ln
__
M
d
−FSL
i
_
cos φ
i
_
−ln
_
A
mRi
__
S
aRi
(i = 1, 2).
(.3)
After straightforward algebra, (.2) yields
p
avail
· 100%
= 1 −0.5
_
+∞
uF1
du
1
f
U1
_
u
1
_
erfc
_
u
F2
−ρ
n12
u
1
_
2
_
1 −ρ
2
n12
_
_
.
(.4)
ACKNOWLEDGMENTS
The authors are indebted to the three anonymous review
ers whose constructive comments helped to signiﬁcantly im
prove the initial version of this paper. Moreover, the ﬁrst au
thor would like to thank Professor Bhaskar D. Rao from Uni
versity of California, San Diego, USA, for the fruitful discus
sions they had on the ﬁrst part of this work.
REFERENCES
[1] A. J. Paulraj, D. A. Gore, R. U. Nabar, and H. B¨ olcskei, “An
overview of MIMO communications—a key to gigabit wire
less,” Proceedings of the IEEE, vol. 92, no. 2, pp. 198–218, 2004.
[2] D. Gesbert, M. Shaﬁ, D.S. Shiu, P. J. Smith, and A. Naguib,
“From theory to practice: an overview of MIMO spacetime
coded wireless systems,” IEEE Journal on Selected Areas in
Communications, vol. 21, no. 3, pp. 281–302, 2003.
[3] A. D. Panagopoulos, P.D. M. Arapoglou, and P. G. Cottis,
“Satellite communications at Ku, Ka, and V bands: propaga
tion impairments and mitigation techniques,” IEEE Commu
nications Surveys and Tutorials, vol. 6, no. 3, pp. 2–14, 2004.
[4] K. P. Liolis, A. D. Panagopoulos, and P. G. Cottis, “Outage ca
pacity statistics of MIMO satellite networks operating at Ka
band and above,” in Proceedings of the 12th Ka and Broadband
Communications Conference, Naples, Italy, September 2006.
[5] 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),
pp. 2632–2636, Dallas, Tex, USA, September 2005.
[6] P. R. King and S. Stavrou, “Land mobilesatellite MIMO ca
pacity predictions,” Electronics Letters, vol. 41, no. 13, pp. 749–
751, 2005.
[7] 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 Action,
Prague, Czech Republic, June 2005.
[8] C. Martin, A. Geurtz, and B. Ottersten, “Spectrally eﬃcient
mobile satellite realtime broadcast with transmit diversity,”
in Proceedings of the 60th IEEE Vehicular Technology Confer
ence (VTC ’04), vol. 6, pp. 4079–4083, Los Angeles, Calif, USA,
September 2004.
[9] C. Faller, B.H. Juang, P. Kroon, H.L. Lou, S. A. Ramprashad,
and C.E. W. Sundberg, “Technical advances in digital audio
radio broadcasting,” Proceedings of the IEEE, vol. 90, no. 8, pp.
1303–1333, 2002.
[10] J. Mietzner and P. A. Hoeher, “Distributed spacetime codes
for cooperative wireless networks in the presence of diﬀerent
propagation delays and path losses,” in Proceedings of Sensor
Array and Multichannel Signal Processing Workshop (SAM ’04),
pp. 264–268, Barcelona, Spain, July 2004.
[11] 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.
Konstantinos P. Liolis et al. 11
[12] I. Frigyes and P. Horv´ ath, “Polarizationtime coding in satel
lite links,” IEEE Satellite and Space Communications Newsletter,
vol. 15, no. 2, pp. 6–8, 2005.
[13] P. Horv´ ath and I. Frigyes, “SAT026: 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, “Spacetime cod
ing in mobile satellite communications using dualpolarized
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
tripleorbital diversity performance in Earthspace commu
nication,” International Journal of Satellite Communications,
vol. 20, no. 3, pp. 187–200, 2002.
[18] ITUR Recommendation P.8374, “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 multiantenna Gaussian channels,”
European Transactions on Telecommunications, vol. 10, no. 6,
pp. 585–595, 1999.
[21] S. Sanayei and A. Nosratinia, “Antenna selection in MIMOsys
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 Earthspace 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
CDMAbased 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] ITUR Recommendation S.5806, “Radiation Diagrams for
Use as Design Objectives for Antennas of Earth Stations Op
erating with Geostationary Satellites,” Geneva, Switzerland,
2004.
[26] P. Horv´ ath and I. Frigyes, “Application of the MIMO concept
in millimeterwave 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, McGrawHill, 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 MIMOChannel Modeling:
Accent on Polarization
P´ eter Horv´ ath,
1
George K. Karagiannidis,
2
Peter R. King,
3
Stavros Stavrou,
3
and Istv´ an Frigyes
1
1
Department of Broadband Infocommunications and Electromagnetic Theory, Budapest University of Technology and Economics,
H1111 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
timecoded diversity (PTC) is given. A physicalstatistical model for dualpolarized 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 3Dpolarized
antennas known from the literature and applicable in satellite links. A synthetic satellitetoindoor link is constructed and its elec
tromagnetic behavior is simulated via the FDTD (ﬁnitediﬀerence timedomain) method. Previous result of the authors states that
in 3DPTC situations, MIMO capacity can be about two times higher than SIMO (singleinput multipleoutput) capacity while a
diversity gain of nearly 2 ×3 is further veriﬁed via extensive FDTD computer simulation.
Copyright © 2007 P´ eter Horv´ ath 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
It is more or less a commonplace statement that in the wire
less technology of recent years, systems applying multiple
transmit and multiplereceive antennas (MIMO, multiple
input multipleoutput) have become one of the few meth
ods of real innovation. Spacetime processing, in particular
spacetime 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 spacetime processing. In addition, multi
path channels and relevant fading characteristics—Rayleigh,
Rice, Suzuki, and so forth—are assumed. A similar situation
is present in satellitetomobile or satellitetoindoor links.
Among others, in [1] it is experimentally veriﬁed that the
LEO satellitetoindoor 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, similartoterrestrial 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 spacetime 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 physicalstatistical model for satelliteto
urban and satellitetohighway channel and computes capac
ity of a 2×2 MIMOsystem. In [3], a satellitediversity 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 selfevidently 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 lowradiator 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 spacetime 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 transmitandreceivespacediversity 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 dualpolarized
MIMO channels, stating a physicalstatistical model, pre
senting measuring results and validating the model; in this
discussion conventional dualpolarized antennas are applied.
In Section 3.2, PTC antennas of 3dimensional 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 FEWWORDS ONCOOPERATIVE
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 CSDsystemwith two satellite relays is shown
in Figure 1.
The idea of merging cooperation with spacetime coding
resulted in the socalled distributed or cooperative spacetime
coding (CSTC). Compared to the conventional spacetime
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
P´ eter Horv´ ath et al. 3
SAT1
TMT1
TMT2
SAT2
Figure 1: A virtual array: 2 satellites and 2 terminals.
the ﬁrst signaling interval and there is no transmission from
SAT1 or SAT2 to TMT2 within this time interval. In the sec
ond signaling interval, both SAT1 and SAT2 communicate
with TMT2. This scenario assumes perfect knowledge of the
channel fading coeﬃcients at the receiver side of TMT2 and
synchronization as an a priori condition. However, the delays
due to distance between SAT1 and SAT2 (and the diﬀerent lo
cal oscillators at SAT1 and SAT2) make cooperative diversity
asynchronous in nature.
Several methods have been proposed to apply CSTC, in
the presence of asynchronity between relays (see [17, 18] and
references therein). However, a theoretical analysis on the ef
fect of the (high) asynchronity in cooperative satellite diver
sity systems does not exist in the literature. Such an analysis
is out of the scope of the present paper and is left for further
study.
3. POLARIZATIONTIME CODINGINSATELLITE
COMMUNICATIONS
3.1. Physicalstatistical model for the dual polarized
LMS MIMOchannel
In [19], a basic investigation of PTC was presented, using
a simple theoretical MIMO channel model. It was assumed
that in a multipath environment—of whatever polarization
the transmit antenna(s) is (are)—the received signal is of
completely random polarization, that is, any state of polar
ization is equally likely. With a simulation study, we did show
that applying normal dualpolarized antennas at both ter
minals and transmitting Alamoutitype coded signals [20],
there is a 2×1 or 2×2 diversity eﬀect if polarization of the re
ceived signals is fully correlated or completely uncorrelated,
respectively. Incidentally, polarization characteristics are de
scribed there via Stokes parameters and related concepts. In
order to assess the beneﬁts of MIMO techniques applied to
mobile satellite links, real channel data or accurate channel
models are required. In this section, a physicalstatistical 2×2
dualpolarized MIMO channel model is presented.
3.1.1. Channel model construction
The following dualpolarized physicalstatistical LMS MI
MO channel model is an extension to the multiplesatellite
LMS MIMO model presented in [2]. In the present paper, a
single satellite containing right(RHCP) and lefthand circu
lar polarization (LHCP) antennas communicates with a mo
bile vehicle, also containing RHCP and LHCP antennas. Note
that taking into account the spherical symmetry of polariza
tion states on the Poincar´ e sphere, actual choice of two or
thogonal polarizations does not have too much signiﬁcance
[21].
Channel model construction is described in [2]. Addi
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 crosspolar
channels are discarded. When the LOS path is blocked by a
building (blocked), rooftop diﬀraction is applied to both the
co and crosspolar channels; the crosspolar component is
scaled below the copolar component as observed from mea
sured data. When the LOS path is shadowed by vegetation
(tree), attenuation is applied to this path based on the dis
tance traversed through the tree and using a typical attenu
ation factor of −1.3 dB per meter [22]. Similarly, the cross
polar component is scaled below the copolar component.
It is assumed in this model that the LOS paths are fully
correlated between co and crosspolar channels, and that the
diﬀuse multipath components are fully uncorrelated between
co and crosspolar channels. This simpliﬁcation is represen
tative of many, but not all, real practical channels; a full pre
sentation of measured satellite MIMO channel correlation is
provided in [23].
The highresolution timeseries data α
M,N
between each
satellite antenna M and each mobile antenna N can be de
ﬁned as follows:
α
M,N
=
⎧
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎨
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎩
P
M,N
e
jkdM,N
+b
n
i=1
T
i
Γ
i
P
M,N,i
e
jkdM,N,i
clear copolar
b
n
i=1
T
i
Γ
i
P
M,N,i
e
jkdM,N,i
clear crosspolar
D
M,N
P
M,N
e
jkdM,N
+b
n
i=1
T
i
Γ
i
P
M,N,i
e
jkdM,N,i
block copolar
S
b
D
M,N
P
M,N
e
jkdM,N
+b
n
i=1
T
i
Γ
i
P
M,N,i
e
jkdM,N,i
block crosspolar
T
M,N
P
M,N
e
jkdM,N
+b
n
i=1
T
i
Γ
i
P
M,N,i
e
jkdM,N,i
tree copolar
S
t
T
M,N
P
M,N
e
jkdM,N
+b
n
i=1
T
i
Γ
i
P
M,N,i
e
jkdM,N,i
tree crosspolar
(1)
4 EURASIP Journal on Wireless Communications and Networking
where P
M,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, T
i
is the tree attenuation
applied to a reﬂected contribution from scatterer i, Γ
i
is the
complex reﬂection coeﬃcient at scatterer i, P
M,N,i
is the path
loss from satellite antenna M to moving mobile antenna N
via scatterer i, d
M,N,i
is the distance between satellite antenna
M and moving mobile antenna N via scatterer i, D
M,N
is the
LOS diﬀraction loss, and T
M,N
is the LOS tree loss. The terms
S
b
and S
t
account for the attenuation of the crosspolar terms
for blocked and treeshadowed 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)
treelined road/highway, characterized by a high likelihood
of dense tree matter at either side of the road with occasional
clearings and occasional twostorey houses beyond the veg
etation, and (b) urban, characterized by densely placed two
tofourstorey 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 timeseries data and ﬁrst
and secondorder statistics. In this paper, the ﬁrstorder
statistics of the model are presented showing their validation
against measured data. Validation of secondorder statistics,
not relevant to the diversity gain analysis presented below, is
a work to be published.
An example of the copolar model output highresolution
path loss timeseries 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 ﬁrstorder modeled and measurement
data are compared. Cumulative distribution functions of co
and crosspolar channels for highway and urban environ
ments are shown in Figure 3. The 2×2 dualpolarized 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.
−40
−30
−20
−10
0
10
R
e
c
e
i
v
e
d
p
o
w
e
r
(
d
B
)
0 200 400 600 800 1000 1200 1400 1600 1800 2000
Mobile position (m)
Urban
(a)
−40
−30
−20
−10
0
10
R
e
c
e
i
v
e
d
p
o
w
e
r
(
d
B
)
0 200 400 600 800 1000 1200 1400 1600 1800 2000
Mobile position (m)
Highway
(b)
Figure 2: Example copolar timeseries 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
MIMOtechniques applied to LMS systems. However, it mod
els usual doublepolarized channels/systems only, resulting
in at most 4fold diversity gain and 2fold increase in capac
ity. Taking the generalized 3dimensional (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 3Dpolarization satellite antennas
3.2.1. The concept of 3Dpolarization
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 2dimensional 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
P´ eter Horv´ ath et al. 5
0.8
0.9
1
P
(
f
a
d
e
d
e
p
t
h
<
a
b
s
c
i
s
s
a
)
−20 −10 0 10
Power relative to FSL (dB)
Measured copolar
Measured Xpolar
Modeled copolar
Modeled Xpolar
(a)
10
−2
10
−1
10
0
P
(
f
a
d
e
d
e
p
t
h
<
a
b
s
c
i
s
s
a
)
−45 −40 −35 −30 −25 −20
Power relative to FSL (dB)
Measured copolar
Measured Xpolar
Modeled copolar
Modeled Xpolar
(b)
0.8
0.9
1
P
(
f
a
d
e
d
e
p
t
h
<
a
b
s
c
i
s
s
a
)
−20 −15 −10 −5 0
Power relative to FSL (dB)
Measured copolar
Measured Xpolar
Modeled copolar
Modeled Xpolar
(c)
10
−2
10
−1
10
0
P
(
f
a
d
e
d
e
p
t
h
<
a
b
s
c
i
s
s
a
)
−45 −40 −35 −30 −25 −20
Power relative to FSL (dB)
Measured copolar
Measured Xpolar
Modeled copolar
Modeled Xpolar
(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 4vector 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
10
−6
10
−5
10
−4
10
−3
10
−2
10
−1
10
0
B
i
t
e
r
r
o
r
r
a
t
e
(
B
E
R
)
0 10 20 30 40 50 60
E
b
/N
0
(dB)
No diversity
MRRC (1 Tx, 2 Rx)
PTBC (2 Tx, 1 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 MIMOantennas
If the degree of asynchronism arising in multisatelliteto
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,
6fold 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 4fold, as shown by
[29].
In [13], the socalled MIMO cube is dealt with. This con
tains 12 electric dipoles arranged at the edges of a cube.
Cubetocube 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,
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 3Dpolarization 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 systemde
picted in Figure 5(a) is of doublebounce 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 selfevident 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 satellitetoindoor/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 singlechannel case, only a 2fold in
crease of capacity can be achieved by PTC. This is in con
trast to the terrestrial case in which this increase is 4fold.
In more details, while any small multielement antenna can
be applied in the ground terminal, onboard one satellite at
most conventional doublepolarized 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 doubleinput multipleoutput channel, the
uplink is its inverse, that is, multipleinput doubleoutput.
P´ eter Horv´ ath et al. 7
t(t)
Scattering
medium
Scattering
medium
r(t)
(a)
t(t)
Scattering
medium
r(t)
(b)
Figure 5: Terrestrial multipath links with compact MIMO anten
nas in scattering media; (a) doublebounce scattering; (b) single
bounce.
t(t)
Plane wave
Aperture
Scattering
medium
r(t)
Figure 6: A satellitetomobile/indoor link.
This has the consequence that from the coding point of view,
the system is not uniform. If as an example, spacetime block
coding of the Alamouti type or orthogonal spacetime block
coding (OSTBC) is chosen, R
C
= 1 can be applied downlink,
however in the uplink R
C
= 1/2 or at most R
C
= 3/4 can
only be achieved. (R
C
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.)
Incident
wave
Window
O1 O2 O3
y = 4.5 m
x
=
2
.
8
m
Figure 7: A satellitetomobile/indoor link.
In the next section, applying electromagnetic simulation
we verify the capacity and the diversity characteristics as
stated above.
4. FDTDSIMULATIONOF
A SATELLITETOINDOOR LINK
In order to assess the performance of using three orthog
onally polarized antennas in a satellitetoindoor scenario,
some simulations were performed using fullwave electro
magnetic tools. The FDTDmethod [30] was used to calculate
the timedependent 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
xy 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 xz plane at y = 0 m.
The electric ﬁeld components (E
x
, E
y
, and E
z
) 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 thinwire
halfwave 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
xz crosssections 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
ments due to diﬀerent shadowing and angleofincidence pa
rameters. At each of the three planes, about 800 points were
observed, spaced 7.5 cm apart in both x and z directions. In a
ﬁrst scenario (S1), the incident waves arrive horizontally (at
0 elevation and parallel to yaxis). In a second scenario (S2),
the elevation was chosen to be 30 degrees and the azimuth
angle 20 degrees oﬀ the yaxis. Thus, in the latter case, the
line of sight is blocked at the points of O2 and O3. For each
scenario, two simulation runs yielded 6 time functions of the
ﬁelds (E
x
, E
y
, and E
z
when using the one or the other po
larization). From the observed ﬁelds, which were regarded as
received voltages according to the reasoning presented above,
signal portions weaker than a designated noise level, chosen
to be −15 dB relative to the maximum power level, were dis
carded. Then the envelope of the received signals was calcu
lated. Based on these data, three statistical parameters were
derived for both Scenarios 1 and 2. First, the equalpower
capacity [31, Equation (4)], was calculated and its CDF was
determined. In Figures 8 and 11, the capacity CDF curves
are shown for S1 and S2, respectively. As expected, at low
outage, levels the capacity of the dualpolarized 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
(2, 2) and the (2, 3) systems is rather small. In order to as
sess the diversity performance, the envelope correlation [32]
was determined between the received signals (latter being the
correlation coeﬃcient between the envelopes of the received
signals). Their CDFs are shown in Figures 9 and 12. As ex
pected, in Scenario 2, lower (even negative) correlation is to
be expected. Additionally, the relative received signal power
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
for the scenarios in consideration. Note that the conﬁdence
for very lowprobability (less than 0.01 or so) portions of the
curve might be low due to the relatively low number (about
2000) of observations, but still validates the claim based on
the higher probability portion of the curves.
5. CONCLUSIONS
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
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
is called here PTC.)
Concerning the ﬁrst of these (modeling), a physical sta
tistical model is given for the urban and the highway satellite
mobile channels. Besides giving a validated model, it veri
ﬁes once again the authors’ conviction that the best type of a
multipath channel model is of the physicalstatistical type.
Concerning the second of these (simulation), a very ex
tensive simulation study is carried out about the 3D polar
ization characteristics of the satellite multipath channel. A
synthetic satellitetoindoor link is simulated and PTC char
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
P
r
o
b
a
b
i
l
i
t
y
(
C
E
P
<
a
b
s
c
i
s
s
a
)
0 2 4 6 8 10 12 14 16 18 20
Capacity (bits/s/Hz)
n
T
= 1; n
R
= 1
n
T
= 2; n
R
= 2
n
T
= 2; n
R
= 3
Figure 8: CDF of the equalpower capacity (Scenario 1).
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
P
r
o
b
a
b
i
l
i
t
y
(
ρ
e
<
a
b
s
c
i
s
s
a
)
−0.2 0 0.2 0.4 0.6 0.8 1
Envelope correlation
ρ
Hy
ρ
Hz
ρ
Vy
ρ
Vx
Figure 9: CDF of the envelope correlation (Scenario 1).
acteristics are investigated. The main purpose of this study
was to verify (for this example) the ﬁndings of two of these
authors [15] about the capacity and diversity characteristics
of this type of channels. Results of this simulation are as fol
lows.
From the capacity point of view, (i) the diﬀerence be
tween the 2 ×2 and the 2 ×3 cases is negligible (as stated in
[15]); and (ii) with high probability capacity of the MIMO,
the situation is nearly exactly 2times as high as that of the
SISOcase, again in accordance with [15]. (Note that with low
probability, this diﬀerence is higher.)
P´ eter Horv´ ath et al. 9
10
−3
10
−2
10
−1
10
0
P
r
o
b
a
b
i
l
i
t
y
(
P
r
<
a
b
s
c
i
s
s
a
)
−50 −45 −40 −35 −30 −25 −20 −15 −10
Combined received power (dBm)
n
T
= 1; n
R
= 1
n
T
= 2; n
R
= 2
n
T
= 2; n
R
= 3
Figure 10: CDF of the received power (Scenario 1).
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
P
r
o
b
a
b
i
l
i
t
y
(
C
E
P
<
a
b
s
c
i
s
s
a
)
0 2 4 6 8 10 12 14 16 18
Capacity (bits/s/Hz)
n
T
= 1; n
R
= 1
n
T
= 2; n
R
= 2
n
T
= 2; n
R
= 3
Figure 11: CDF of the equalpower capacity (Scenario 2).
To characterize the diversity performance, CDF of the re
ceived power in the various situations is investigated; result
shows that 3fold (i.e., 3D) polarization diversity yields sig
niﬁcantly higher received power than the 2fold 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 satellitediversityMIMO, the problems brieﬂy
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
P
r
o
b
a
b
i
l
i
t
y
(
ρ
e
<
a
b
s
c
i
s
s
a
)
−0.2 0 0.2 0.4 0.6 0.8 1
Envelope correlation
ρ
Hy
ρ
Hz
ρ
Vy
ρ
Vx
Figure 12: CDF of the envelope correlation (Scenario 2).
10
−3
10
−2
10
−1
10
0
P
r
o
b
a
b
i
l
i
t
y
(
P
r
<
a
b
s
c
i
s
s
a
)
−50 −45 −40 −35 −30 −25 −20 −15 −10
Combined received power (dBm)
n
T
= 1; n
R
= 1
n
T
= 2; n
R
= 2
n
T
= 2; n
R
= 3
Figure 13: CDF of the received power (Scenario 2).
dealt with in Section 3, that is, the eﬀect of extremely large
and variable diﬀerence between the pathlengths 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 EURASIP Journal on Wireless Communications and Networking
REFERENCES
[1] Z. Bodnar, Z. Herczku, J. Berces, et al., “A detailed experimen
tal study of the LEOsatellite 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, “Physicalstatistical
model for the land mobilesatellite 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 MIMOsatellite 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 spacetime 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 spacetime
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: spacetime
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 spacetime 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. Horv´ ath 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.
[16] P. Horv´ ath 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. Horv´ ath, “Polarizationtime 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. Moln´ ar, Z. Herczku, and Z. Bodn´ ar, “Antenna
gain and polarization eﬀects in wireless links—accent on LEO
satellites,” Space Communications, vol. 19, no. 34, pp. 199–
208, 2004.
[22] I. H. Cavdar, H. Dincer, and K. Erdogdu, “Propagation mea
surements at Lband 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 fourvector,” 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/pp161167.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, JuneJuly 2002.
[30] A. Taﬂove and S. C. Hagness, Computational Electrodynamics:
The FiniteDiﬀerenceTimeDomain 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 spacetime
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 signaltonoise 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 (P
out
), 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 maximalratio
combining (MRC), equalgain 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 signaltonoise 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 smallsize 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 lineof
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 clearstate [5] chan
nel conditions. Furthermore, in [6] it was depicted that the
Ricean Kfactor 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 (P
out
) 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 singlefold in
tegrals and inﬁnite series expressions. In [9] the ABEP of
dualbranch 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
dualbranch SC diversity receivers operating over correlated
Weibull fading was presented. In [12], inﬁnite series expres
sions for the capacity of dualbranch MRC, EGC, SC, and
SSC diversity receivers over Nakagamim 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, Nakagamim and Ricean, and under dif
ferent channel conditions but dealt mainly with uncorre
lated fading. For correlated fading in this reference only the
Nakagamim 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 P
out
for the cor
related lognormal 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 CCof 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 P
out
, 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. SYSTEMMODEL
By considering a dualbranch 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 singlesided power spectral den
sity of N
0
. 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
= R
2
E
s
/(2N
0
), where E
s
= Es
2
is the
transmitted average symbol energy, where E· denoting ex
pectation and  ·  absolute value. The corresponding average
SNRper symbol at both input branches is γ = ΩE
s
/N
0
, where
Ω = ER
2
. The PDF of the SNR of the Ricean distribution
is given by [2, Equation (2.16)]
f
γ
(γ) =
1 + K
γ
exp
_
−K −
(1 + K)
γ
γ
_
×I
0
_
2
¸
K(K + 1)
γ
γ
1/2
_
,
(2)
where K is the Ricean Kfactor deﬁned as the power ratio
of the specular signal to the scattered signals and I
0
(·) is the
zerothorder modiﬁed Bessel function of the ﬁrst kind [21,
Equation (8.406)]. The CDF of γ is given by [14, Equation
(8)]
F
γ
(γ) = Q
1
_
√
2K,
¸
2(1 + K)
γ
γ
_
, (3)
where Q
1
(·) is the ﬁrstorder MarcumQ function [2, Equa
tion (4.33)].
The joint PDF of γ
1
and γ
2
, 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 I
0
(·) with its inﬁnite series representa
tion [21, Equation (8.445)], expanding the term [γ
1
+ γ
2
+
2
√
γ
1
γ
2
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 γ
1
, γ
2
can be
expressed as
f
γ1,γ2
_
γ
1
, γ
2
_
=
∞
_
i,h=0
v1+v2+v3=i
Aexp
_
−β
1
_
γ
1
+ γ
2
__
×
_
Bγ
β2−1
1
γ
β3−1
2
+ Cγ
−1
γ
β2−1/2
1
γ
β3−1/2
2
_
(4)
with
A=
2
v3+2h−1
(1 + K)
1+β4
ρ
2h
K
i
exp
_
−2K/(1 + ρ)
_
√
πγ
1+β4
_
1 −ρ
2
_
1+2h
v
1
!v
2
!v
3
!i!(1 + ρ)
2i
,
B =
_
1 + (−1)
v3
_
Γ
_
h +
_
1 + v
3
_
/2
_
Γ
_
h + 1 + v
3
/2
_
Γ
_
1 + 2h
_ ,
C =
_
−1 + (−1)
v3
_
2ρ(1 + K)Γ
_
1 + h + v
3
/2
_
_
ρ
2
−1
_
Γ(2 + 2h)Γ
_
h +
_
3 + v
3
_
/2
_ ,
β
1
=
(1 + K)
_
1 −ρ
2
_
γ
, β
2
= v
1
+
v
3
2
+ h + 1,
β
3
= v
2
+
v
3
2
+ h + 1, β
4
= i + 2h + 1,
(5)
where Γ(·) is the Gamma function [21, Equation (8.310/1)]
and ρ is the correlation coeﬃcient between γ
1
and γ
2
. It can
be proved that the above inﬁnite series expression always
converges [18].
3. RECEIVEDSIGNAL STATISTICS
In this section, the most important statistical metrics,
namely, the PDF, CDF, MGF, and moments of dual branch
SSC output SNR diversity receivers operating over correlated
Ricean fading channels will be presented.
3.1. Probability density function (PDF)
Let γ
ssc
be the instantaneous SNR per symbol at the output of
the SSC and γ
τ
the predetermined switching threshold. Fol
lowing [15], the PDF of γ
ssc
, f
γssc
(γ), is given by
f
γssc
(γ) =
⎧
⎪
⎨
⎪
⎩
r
ssc
(γ), γ ≤ γ
τ
,
r
ssc
(γ) + f
γ
(γ), γ > γ
τ
.
(6)
Moreover, r
ssc
(γ) is given in [23, Equation (21b)] as
r
ssc
(γ) =
_
γτ
0
f
γ1γ2
_
γ, γ
2
_
dγ
2
=
_
∞
0
f
γ1γ2
_
γ, γ
2
_
dγ
2
−
_
∞
γτ
f
γ1γ2
_
γ, γ
2
_
dγ
2
.
(7)
Hence, by substituting (4) in (7) and using [21, Equation
(3.351/23)], these integrals can be solved and r
ssc
(γ) can be
expressed as
r
ssc
(γ) =
∞
_
i,h=0
v1+v2+v3=i
Aexp
_
−β
1
γ
_
γ
β2−1/2
×
_
Bγ
_
β
3
, β
1
γ
τ
_
√
γβ
β3
1
+
Cγ
_
β
3
+ 1/2, β
1
γ
τ
_
γβ
β3+1/2
1
_
,
(8)
where γ(·, ·) is the lower incomplete Gamma function [21,
Equation (8.350/1)].
3.2. Cumulative distribution function (CDF)
Similar to [23, Equation (20)], the CDF of γ
ssc
, F
γssc
(γ), is
given by
F
γssc
(γ) = Pr
_
γ
τ
≤ γ
1
≤ γ
_
+ Pr
_
γ
2
< γ
τ
∧γ
1
< γ
_
(9)
which after some manipulations can be expressed in terms of
CDFs as
F
γssc
(γ) =
⎧
⎪
⎨
⎪
⎩
F
γ1,γ2
_
γ, γ
τ
_
, γ ≤ γ
τ
,
F
γ
(γ) −F
γ
_
γ
τ
_
+ F
γ1,γ2
_
γ, γ
τ
_
, γ > γ
τ
.
(10)
Hence, by substituting (4) in F
γ1,γ2
(γ, γ
τ
) =
_
γ
0
_
γτ
0
f
γ1,γ2
(γ
1
,
γ
2
)dγ
1
dγ
2
using [21, Equation (3.351/1)], F
γ1,γ2
(γ, γ
τ
) can be
derived as
F
γ1,γ2
_
γ, γ
τ
_
=
∞
_
i,h=0
v1+v2+v3=i
A
β
β2+β3
1
×
_
Bγ
_
β
2
, β
1
γ
_
γ
_
β
3
, β
1
γ
τ
_
+
C
β
1
γ
γ
_
β
2
+
1
2
, β
1
γ
_
γ
_
β
3
+
1
2
, β
1
γ
τ
_
_
.
(11)
In order to verify the validity of the above derivations,
(10) and (11) have been numerically evaluated for the spe
cial case of uncorrelated, that is, ρ = 0, Ricean fading chan
nels. The resulting CDF was found to be identical to the same
CDF presented in [2, Equation 9.273], which was derived us
ing a diﬀerent mathematical approach as a closedform ex
pression.
3.3. Moments generating function (MGF)
Based on (6), the MGF of γ
ssc
, M
γssc
(s) = Eexp(−sγ
ssc
), [24,
Equation (5.62)], can be expressed in terms of two integrals
as
M
γssc
(s) =
_
∞
0
exp(−sγ)r
ssc
(γ)dγ
+
_
∞
γτ
exp(−sγ) f
γ
(γ)dγ = I
1
+ I
2
.
(12)
4 EURASIP Journal on Wireless Communications and Networking
Using [21, Equation (3.381/4)], I
1
can be expressed in terms
of inﬁnite series as
I
1
=
∞
_
i,h=0
v1+v2+v3=i
A
_
Γ
_
β
2
_
_
β
1
+ s
_
β2
Bβ
−β3
1
γ
_
β
3
, β
1
γ
τ
_
+Cβ
−β3−1/2
1
Γ
_
β
2
+1/2
_
_
β
1
+s
_
β2+1/2
γ
_
β
3
+
1
2
, β
1
γ
τ
_
_
.
(13)
Setting ψ =
_
2γ[(1 + K)/γ + s] and using [2, Equation
(4.33)], I
2
can be solved as
I
2
= Q
1
_¸
2K(1 + K)
1 + K + γs
,
¸
¸
_
2(1 + K + γ s)γ
τ
γ
_
×exp
_
K(1 + K)
1 + K + sγ
_
(1 + K) exp(−K)
1 + K + γs
.
(14)
3.4. Moments
Based on (6), the moments for γ
ssc
, μ
γssc
(n) = Eexp(γ
n
ssc
),
[24, Equation (5.38)], can be expressed in terms of two inte
grals as
μ
γssc
(n) =
_
∞
0
γ
n
r
ssc
(γ)dγ +
_
∞
γτ
γ
n
f
γ
(γ)dγ
= I
3
+ I
4
.
(15)
Using again [21, Equation (3.381/4)], I
3
can be expressed in
terms of inﬁnite series as
I
3
=
∞
_
i,h=0
v1+v2+v3=i
A
_
Bγ
_
β
3
, β
1
γ
τ
_Γ
_
n + β
2
_
β
β2+β3+n
1
+
Cγ
_
β
3
+ 1/2, β
1
γ
τ
_
β
β2+β3+n+1
1
Γ
_
n + β
2
+
1
2
__
.
(16)
Setting φ =
_
2γ(1 + K)/γ in I
4
, using [2, Equation
(4.104)], after some straightforward mathematical manip
ulations, yields
I
4
=
γ
n−1
2
n
(1 + K)
n−1
Q
2n+1,0
_
K,
¸
¸
_
2(1 + K)γ
τ
γ
_
, (17)
where Q
m,n
(·, ·) is the Nuttal Qfunction deﬁned in [25].
4. CHANNEL CAPACITY (CC)
CC is a wellknown performance metric providing an upper
bound for maximum errorless transmission rate in a Gaus
sian environment. The average CC, C, is deﬁned as [26]
C
Δ
= BW
_
∞
0
log
2
(1 + γ) f
γssc
(γ)dγ, (18)
where BW is transmission bandwidth of the signal in Hz.
Hence, substituting (6) in (18), C becomes
C =
_
∞
0
log
2
(1 + γ)r
ssc
(γ)dγ +
_
∞
γτ
log
2
(1 + γ) f
γ
(γ)dγ
= I
5
+ I
6
.
(19)
By representing ln(1 + γ) = G
1,2
2,2
_
γ 
1,1
1,0
_
, [27, Equation
(01.04.26.0003.01)], and exp(−γ) = G
1,0
0,1
_
γ 
0
−
_
, [27, Equa
tion (01.03.26.0004.01)], where G(·) is Meijer’s Gfunction
[21, Equation (9.301)] and using [28], I
5
can be solved as
I
5
=
∞
_
i,h=0
v1+v2+v3=i
A
ln2
_
B
γ
_
β
3
, β
1
γ
τ
_
β
β3+β2
1
G
1,3
3,2
_
1
β
1
¸
¸
¸
¸
1, 1, 1 −β
2
1, 0
_
+ C
γ
_
β
3
+ 1/2, β
1
γ
τ
_
β
β3+β2+3/2
1
×G
1,3
3,2
_
1
β
1
¸
¸
¸
¸
1, 1, 1 −β
2
1, 0
_
_
.
(20)
Due to the very complicated nature of I
6
, it is very diﬃcult,
if not impossible, to derive a closedform solution for this
integral. However, I
6
can be evaluated via numerical inte
gration using any of the wellknown mathematical software
packages, such as MATHEMATICA or MATLAB.
5. PERFORMANCE ANALYSIS
In this section a detailed performance analysis, in terms of
P
out
, ASEP, ASNRand AoF, for SSCdiversity receivers operat
ing over correlated Ricean fading channels will be presented.
5.1. Outage probability (P
out
)
P
out
is the probability that the output SNR falls below a pre
determined threshold γ
th
, P
out
(γ
th
), and can be obtained by
replacing γ with γ
th
in (10) as
P
out
_
γ
th
_
= F
γssc
_
γ
th
_
. (21)
5.2. Average symbol error probability (ASEP)
The ASEP, P
se
, can be evaluated directly by averaging the con
ditional symbol error probability, P
e
(γ), over the PDF of γ
ssc
[29]
P
se
=
_
∞
0
P
e
(γ) f
γssc
(γ)dγ. (22)
For diﬀerent families of modulation schemes, P
e
(γ) can
be obtained as follows.
(i) For binary phase shift keying (BPSK) and square M
ary quadrature amplitude modulation (QAM) signaling for
mats and for highinput SNR, P
e
(γ) = Derfc(
_
Eγ), where
P. S. Bithas and P. T. Mathiopoulos 5
erfc(·) is the complementary error function [21, Equation
(8.250/1)] and D, E are constants the values of which depend
on the speciﬁc modulation scheme under consideration. Us
ing this expression, by substituting (6) in (22), yields
P
se
=
_
∞
0
Derfc
__
Eγ
_
r
ssc
(γ)dγ +
_
∞
γτ
Derfc
__
Eγ
_
f
γ
(γ)dγ
= I
7
+ I
8
.
(23)
Expressing erfc(
_
Eγ) =
√
π
−1
G
2,0
1,2
_
Bγ 
1
0,1/2
_
, [27, Equation
(06.27.26.0006.01)], and exp(−γ) = G
1,0
0,1
_
γ 
0
−
_
, [27, Equa
tion (01.03.26.0004.01)], using [28] and after some straight
forward mathematical manipulations I
7
can be expressed as
I
7
=
∞
_
i,h=0
v1+v2+v3=i
ADΓ
_
β
2
+ 1/2
_
√
πβ
β3
1
E
β2
×
_
BΓ
_
β
2
_
Γ
_
β
2
+ 1
_γ
_
β
3
, β
1
γ
τ
_
×
2
F
1
_
β
2
, β
2
+
1
2
; β
2
+ 1; −
β
1
E
_
+
Cγ
_
β
3
+ 1/2, β
1
γ
τ
_
Γ
_
β
2
+ 1
_
_
β
1
E
_
1/2
Γ
_
β
2
+
3
2
_
×
2
F
1
_
β
2
+
1
2
, β
2
+ 1; β
2
+
3
2
; −
β
1
E
_
_
(24)
with
2
F
1
(·, ·; ·; ·) being Gauss Hypergeometric function [21,
Equation (9.100)]. Moreover, I
8
=
_
∞
0
Derfc(
_
Eγ) f
γ
(γ)dγ−
_
γτ
0
Derfc(
_
Eγ) f
γ
(γ)dγ = I
8,a
− I
8,b
. Hence, substituting
again I
0
(·) with its inﬁnite series representation [21, Equa
tion (8.445)], I
8,a
can be solved with the aid of [28] and I
8,b
using [27, Equation (06.27.21.0019.01)]. Thus, using these
solutions of I
8,a
and I
8,b
and after some mathematical ma
nipulations, I
8
can be expressed as in (25):
I
8
=
D(1 + K) exp(−K)
γ
∞
_
k=0
(k!)
−2
_
K(K + 1)
γ
_
k
×
_
Γ(k + 1)Γ(k + 3/2)
√
πE
k+1
Γ(k + 2)
×
2
F
1
_
k + 1, k +
3
2
; k + 2; −
1 + K
γE
_
−
2
√
E/π
_
β
1
_
1 −ρ
2
__
k+3/2
∞
_
ρ=0
_
−(1 + K)/γ
_
ρ
E
ρ
(2ρ + 1)ρ!
×Γ
_
k+
3
2
+ρ,
(1+K)γ
τ
γ
_
−
Γ
_
k+1, (1+K)γ
τ
/γ
_
2
_
β
1
_
1−ρ
2
__
k+1
_
.
(25)
In (25), Γ(·, ·) is the upper incomplete Gamma function [22,
Equation (6.51)].
(ii) For noncoherent binary frequency shift keying
(BFSK) and binary diﬀerential phase shift keying (BDPSK),
P
e
(γ) = Dexp(−Dγ). Similar to the derivation of (12), that
is, using [21, Equation (3.381/4)] and [2, Equation (4.33)],
P
se
can be expressed as
P
se
=
∞
_
i,h=0
v1+v2+v3=i
AD
×
_
Γ
_
β
2
_
B
_
β
1
+ E
_
β2
β
β3
1
γ
_
β
3
, β
1
γ
τ
_
+
CΓ
_
β
2
+ 1/2
_
_
β
1
+ E
_
β2+1/2
β
β3+1/2
1
γ
_
β
3
+
1
2
, β
1
γ
τ
__
+ Q
1
_¸
2K(1 + K)
1 + K + γE
,
¸
¸
_
2(1 + K + γE)γ
τ
γ
_
×exp
_
K(1 + K)
1 + K + γE
_
(1 + K) exp(−K)
1 + K + γE
.
(26)
(iii) For Gray encoded Mary PSK and Mary DPSK,
P
e
(γ) = D
_
Λ
0
exp[−E(θ)γ]dθ, where Λ is constant. Thus, P
se
can be expressed as
P
se
=
∞
_
i,h=0
v1+v2+v3=i
AD
×
_
Bγ
_
β
3
, β
1
γ
τ
_
β
β3
1
_
Λ
0
Γ
_
β
2
_
_
β
1
+ E(θ)
_
β2
dθ
+
Cγ
_
β
3
+ 1/2, β
1
γ
τ
_
β
β3+1/2
1
×
_
Λ
0
Γ
_
β
2
+ 1/2)
_
β
1
+ E(θ)
_
β2+1/2
dθ
_
+
_
Λ
0
Q
1
_¸
2K(1 + K)
g(θ)
,
¸
¸
_
2g(θ)γ
τ
γ
_
×exp
_
K(1 + K)
g(θ)
_
(1 + K) exp(−K)
g(θ)
dθ,
(27)
where g(θ) = 1+K +γE(θ). The above ﬁnite integrals can be
easily evaluated via numerical integration.
5.3. Average output SNR (ASNR) and
amount of fading (AoF)
The ASNR, γ
ssc
, is a useful performance measure serving as
an excellent indicator for the overall system ﬁdelity and can
be obtained from the ﬁrstorder moment of γ
ssc
as
γ
ssc
= μ
γssc
(1). (28)
6 EURASIP Journal on Wireless Communications and Networking
1
1.05
1.1
1.15
1.2
1.25
1.3
N
o
r
m
a
l
i
z
e
d
a
v
e
r
a
g
e
o
u
t
p
u
t
S
N
R
1 2 3 4 5 6 7 8 9
Ricean KFactor
ρ = 0.1
ρ = 0.3
ρ = 0.5
ρ = 0.7
ρ = 0.9
Figure 1: Normalized average output SNR (ASNR) versus the
Ricean Kfactor for several values of the correlation coeﬃcient ρ.
The AoF, deﬁned as AoF
Δ
= var(γ
ssc
)/γ
2
ssc
, 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 secondorder moments of γ
ssc
as
AoF =
μ
γssc
(2)
μ
γssc
(1)
2
−1. (29)
6. PERFORMANCE EVALUATIONRESULTS
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, P
out
,
ABEP
1
, 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 (γ
ssc
/γ) and AoF
are plotted as functions of the Ricean Kfactor 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 Mary
(M > 2) modulation schemes, assuming Gray encoding, the ABEP can
be simply obtained from the ASEP as P
be
∼
= P
se
/ log
2
M, since E
s
=
E
b
log
2
M, where E
b
denotes the transmitted average bit energy.
1 2 3 4 5 6 7 8
Ricean KFactor
ρ = 0.1
ρ = 0.3
ρ = 0.5
ρ = 0.7
ρ = 0.9
0.2
0.3
0.4
0.5
0.6
0.7
A
m
o
u
n
t
o
f
f
a
d
i
n
g
(
A
o
F
)
Figure 2: Amount of fading (AoF) versus the Ricean Kfactor for
several values of the correlation coeﬃcient ρ.
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
Figures 3 and 4 the ABEP is plotted as a function of the av
erage input SNR per bit, that is, γ
b
= γ/ log
2
M, for several
values of K. Figure 3 considers the performance of DBPSK,
BPSK, and Mary 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 16QAM 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 nodiversity case,
is also noted.
In Figure 5, P
out
is plotted as a function of the normalized
outage threshold per bit, γ
th
/γ
b
, 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 P
out
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 γ
b
for several values
P. S. Bithas and P. T. Mathiopoulos 7
10
−4
10
−3
10
−2
10
−1
A
v
e
r
a
g
e
b
i
t
e
r
r
o
r
p
r
o
b
a
b
i
l
i
t
y
(
A
B
E
P
)
DBPSK
BPSK
8PSK
16PSK
−5 0 5 10 15 20
Average input SNR per bit (dB)
K = 1
K = 8
Figure 3: Average bit error probability (ABEP) versus average in
put SNR per bit for DBPSK, BPSK, and MPSK (M = 8 and 16)
signaling formats, for diﬀerent values of the Ricean Kfactor.
10
−5
10
−4
10
−3
10
−2
10
−1
A
v
e
r
a
g
e
b
i
t
e
r
r
o
r
p
r
o
b
a
b
i
l
i
t
y
(
A
B
E
P
)
K = 1
K = 4
K = 8
−5 0 5 10 15 20
Average input SNR per bit (dB)
ρ = 0.2
ρ = 0.6
No diversity
Figure 4: Average bit error probability (ABEP) versus average input
SNR per bit for 16QAM signaling format for diﬀerent values of the
Ricean Kfactor and the correlation coeﬃcient ρ.
10
−4
10
−3
10
−2
10
−1
1
O
u
t
a
g
e
p
r
o
b
a
b
i
l
i
t
y
(
P
o
u
t
)
−10 −7.5 −5 −2.5 0 2.5 5
γ
th
/γ
b
ρ = 0
ρ = 0.4
ρ = 0.8
K = 4
K = 8
K = 1
Figure 5: Outage probability (P
out
) versus the normalized average
input SNR per bit for several values of the Ricean Kfactor and the
correlation coeﬃcient ρ.
0.5
1
1.5
2
2.5
3
−4 −2 0 2 4 6 8 10
Average input SNR per bit (dB)
N
o
r
m
a
l
i
z
e
d
a
v
e
r
a
g
e
c
h
a
n
n
e
l
c
a
p
a
c
i
t
y
(
b
/
s
/
H
z
)
ρ = 0.1
ρ = 0.4
ρ = 0.7
ρ = 0.9
No diversity
Figure 6: Normalized average channel capacity (C/BW) versus the
average input SNR per bit for several values of the correlation coef
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, P
out
,
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 (SatNExII) project (IST
027393), a Network of Excellence (NoE) funded by European
Commission (EC) under the FP6 program.
REFERENCES
[1] T. S. Rappaport, Wireless Communications, PrenticeHall 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 channelrecording,
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 Nakagamim 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 Nakagamimfading 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, “Equalgain and maximalratio
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 Pad´ e approximation,” IEEE
Transactions on Communications, vol. 54, no. 3, pp. 397–402,
2006.
[11] N. C. Sagias, “Capacity of dualbranch 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
Nakagamim 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
dualbranch coherent equalgain combining in correlated
Nakagamim fading,” Electronics Letters, vol. 39, no. 15, pp.
1152–1154, 2003.
[14] A. A. AbuDayya 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 lognormal fading channels,” IEEE Transactions on Com
munications, vol. 50, no. 12, pp. 1946–1959, 2002.
[18] D. A. Zogas and G. K. Karagiannidis, “Inﬁniteseries 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. AbuDayya and N. C. Beaulieu, “Analysis of switched di
versity systems on generalizedfading channels,” IEEE Transac
tions on Communications, vol. 42, no. 11, pp. 2959–2966, 1994.
[24] A. Papoulis, Probability, Random Variables, and Stochastic Pro
cesses, McGrawHill, New York, NY, USA, 2nd edition, 1984.
[25] A. Nuttal, “Some integrals involving the Qfunction,” 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 DVBS2 Systems in
Railway Scenarios
Stefano Cioni,
1
Cristina P´ arraga Niebla,
2
Gonzalo Seco Granados,
3
Sandro Scalise,
2
Alessandro VanelliCoralli,
1
and Mar´ıa Angeles V´ azquez Castro
3
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 Aut` onoma 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 DVBS2 reception by mobile terminals
mounted on highspeed 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 DVBS2 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
Satellite communications developed to a tremendous global
success in the ﬁeld of analog and then digital audio/TV
broadcasting by exploiting the inherent widearea coverage
for the distribution of content. It appeared a “natural” con
sequence to extend the satellite services for pointtopoint
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 (DVBS) [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 DVBS [2], commonly re
ferred to as DVBS2, as well as return channel via satellite
(DVBRCS) [3] standards. Thus, for satellite systems cur
rently under development and being designed to support
mainly multimedia services, the application of the DVBS2,
for the highcapacity gatewaytouser (forward) links and of
DVBRCS for the usertogateway (return) links, is widely
accepted.
Complementing to satellite multimedia to ﬁxed termi
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, highspeed packet access
(HSPA), prestandardized version of mobile WiMAX, and, in
case of broadcast applications, digital video broadcasting for
handhelds (DVBH) [5].
At present, broadband access (e.g., to the Internet) and
dedicated pointtopoint 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, DVBS2/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 systemper
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 locomotive
1
[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 DVBS2/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 systemperformance.
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 NEWDVBS2/RCS STANDARDFOR
MOBILE COLLECTIVE TERMINALS
The large capacity of DVBS2/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 DVBS2/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 DVBS2/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
DVBS2/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 TMRCS has investigated how the cur
rent DVBRCS standard could be applied to mobile applica
tions. A white paper on the applicability of DVBRCS 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 DVBRCS features that can be used for mo
bility management are provided, the latter referring to inter
beam handover only.
Thus, the DVBRCS 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 nonlineofsight (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 DVBRCS 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 standard
2
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
DVBS2/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 DVBS2/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
technology
Examples Pros Cons
Satellite
DVBS2/RCS
Proprietary systems,
for example, ViaSat
(i) No new trackside
infrastructure—quick to
deploy, project costs may be
lower on long distance routes
(i) Available tracking antennas and
eﬃcient satcom modems expensive
(ii) Dedicated bandwidth available (ii) High variable cost per MB
(iii) Performance easy to predict
depending on satellite visibility
(iii) Return bandwidth constrained
by antenna size
(iv) Not aﬀected by borders—good
for international trains
(iv) Satellite visibility seriously
restricted on some rail routes
Cellular
GPRS
EDGE
UMTS
HSUPA/HSDPA
(EVDO)
(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
ﬂat rate)
(ii) Coverage of railway lines often
worse than roads
(iii) Data rates improving year on year (iii) GPRS/EDGE not really fast enough
(iv) Competitive supply—3 or 4 network
operators in most countries
(iv) Inverse relationship between
throughput and train speed
(v) No QoS guarantees—aﬀected by
network congestion at peak times
(vi) Organized country by country—data
roaming charges are punitive
Trackside
Flash OFDM
IEEE 802.11
IEEE 802.16 (WiMAX)
(i) High data rates possible (i) Existing standards not designed to
support fastmoving terminals
(ii) Can control bandwidth and QoS (ii) Proprietary equipment is more
expensive
(iii) Ontrain equipment relatively
inexpensive
(iii) No suitable public services yet in
licensed bands—will licenceholders be
allowed to provide mobile services?
(iv) No volumerelated usage costs (iv) Licenceexempt 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 (23 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 knifeedge diﬀraction theory [11]: in presence of
an obstacle having one inﬁnite dimension (e.g., mountains
or high buildings), the knifeedge 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:
A
s
(h)
=
1
√
2G
max
_
G
_
α
1
(h)
_
¸
¸
¸
¸
_
∞
Kh
e
−j(π/2)v
2
dv
¸
¸
¸
¸
+ G
_
α
2
(h)
_
¸
¸
¸
¸
_
K(h−d)
−∞
e
−j(π/2)v
2
dv
¸
¸
¸
¸
_
,
K =
¸
2
λ
a + b
a · b
,
(1)
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 lineofsight (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 α
1
and α
2
as
shown in Figure 2(left), the antenna shows a gain less than
the maximum achievable (G
max
) 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 socalled “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 DVBS2 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 DVBS2 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. ADVANCEDFADE COUNTERMEASURES
System designers can resort to diﬀerent approaches to coun
teract deep fading conditions and to guarantee an acceptable
QoS level. A possible classiﬁcation of fade countermeasure is
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 channelbased 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
b
h hd
E
2
/E
0
a
E
1
/E
0
v
α
1
α
2
(a)
−45
−40
−35
−30
−25
−20
−15
−10
−5
0
5
A
t
t
e
n
u
a
t
i
o
n
(
d
B
)
−2.5 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2 2.5
h (m)
0.6 m
0.4 m
0.2 m
d = 0.4 m, a = 2.5 m
(b)
Figure 2: Knifeedge 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 equalgain combining. Considering two receiving
1E −04
1E −03
1E −02
1E −01
1E + 0
B
B
F
R
A
M
E
e
r
r
o
r
r
a
t
e
1 3 5 7 9 11 13 15 17 19 21
E
b
/N
0
(dB)
1/2  QPSK (LOS, FAST, noPA)
2/3  8PSK (LOS, FAST, noPA)
3/4  16APSK (LOS, FAST, noPA)
5/6  16APSK (LOS, FAST, noPA)
1/2  QPSK (LOS, FAST, PA)
2/3  8PSK (LOS, FAST, PA)
3/4  16APSK (LOS, FAST, PA)
5/6  16APSK (LOS, FAST, PA)
Power arches ﬂoor
Figure 3: BBFRAME error rate for DVBS2 in the presence of
power arch blockage events. LOS propagation conditions and train
speed set to 300 km/h.
antennas, and assuming perfect compensation of time delays
of the two replicas, the combined signal can be written as
r
c
(t) = w
1
r
1
(t) + w
2
r
2
(t), (2)
where w
i
and r
i
(t), i = 1, 2, are the combing weights and
the received signals, respectively. The received signals at each
antenna is
r
i
(t) = α
i
s
0
(t) + n
i
(t), (3)
where s
0
(t) is the transmitted signal, α
i
is the time variant
fading envelope over the ith antenna, and n
i
(t) is the thermal
noise.
The simplest combining scheme is the signal selection
Combining (SC), in which the branchsignal with the largest
amplitude or signaltonoise ratio (SNR) is the one selected
for demodulation. In this case, w
i
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. Maximumratio 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, N
i
, as follows:
w
i
=
α
i
N
i
. (4)
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 16APSK or 32APSK (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
log
_
Pr
_
b
i
= 0  r
k
_
Pr
_
b
i
= 1  r
k
_
_
= log
_
si ∈S0
exp
_
−
¸
¸
r
k
− α
k
s
i
¸
¸
2
/N
0
_
si ∈S1
exp
_
−
¸
¸
r
k
− α
k
s
i
¸
¸
2
/N
0
_
_ (5)
can be extended for L receiving antennas, according to the
MRC principle, as follows:
log
_
Pr
_
b
i
= 0  r
k
_
Pr
_
b
i
= 1  r
k
_
_
= log
_
si ∈S0
exp
_
−
L
p=0
_¸
¸
r
p
k
− α
p
k
s
i
¸
¸
2
/N
p
0
__
si ∈S1
exp
_
−
L
p=0
_¸
¸
r
p
k
− α
p
k
s
i
¸
¸
2
/N
p
0
__
_
,
(6)
where r
k
is the received sample at time k, α
k
is the true or
the estimated channel coeﬃcient, and S
0
and S
1
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 powerarch 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.
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 endtoend 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
retransmission scheme.
Regarding the retransmission schemes, eﬃcient proto
cols should limit the use of acknowledgement (ACK) based
mechanisms because they introduce heavy feedback traﬃc
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
approaches are hence particularly interesting. In combina
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
a reduced number of retransmissions. This might be partic
ularly useful if resources on the return link need to be saved
(smaller number of NACKs or no NACKs are needed at all),
or when multiple lost packets are recovered with the retrans
mission of a lower number of redundant packets. Basically,
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 systemcan 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 FECis 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
from
antenna no. 1
Matched
ﬁlter
Symbol
sampling
DeMUX
Data
Buﬀer
Frequency
acquisition
Timing
recovery
Preamble /
pilots
Noise level
estimation
N
1
0
θ
1
0
α
1
k
θ
1
0
θ
1
k
Digital
AGAC
Buﬀer
Lock
detector
Freq./phase
tracking
Signal
combiner
Hard/soft
demodulator
De
interleaver
LDPC/BCH
decoder
From second
antenna
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 closedloop 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 lowpriority 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 IPbased 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 GSEFEC 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
FEC matrix
GSE
encapsulation
BBFRAME assembly
using one or several
GSE units
BBFRAME
padding
BBFRAMEs
Figure 6: Steps involved in GSEFEC.
diﬃculty in applying packet level FEC (PLFEC). 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 DVBH
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.
DVBH also provides a solution for encapsulating the
coded IP packets for transmission over DVBT. 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 DVBS2, it presents a number
of drawbacks, such as lack of ﬂexibility, low encapsulation
eﬃciency, delay constraints. A new encapsulation protocol
call generic streamencapsulation (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 MPEMPEG. 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 (DVBT2) as well. GSE
also eﬃciently supports the ACM functionalities of DVBS2
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 PLFEC 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 DVBH, the ﬁrst pro
cess consists in arranging the IP packets in a matrix (here
after called FEC matrix) and applying a ReedSolomon code,
while the second process employs MPEMPEG. The whole
implementation is called MPEFECin DVBH. Our proposal
for DVBS2 is based on keeping the same ﬁrst process as in
DVBH, whereas it employs GSE in the second process. This
proposal for applying PLFEC in DVBS2 is named GSE
FEC.
A block diagram of GSEFEC is depicted in Figure 6. The
incoming IP packets are arranged in the socalled FEC ma
trix, where also the packetlevel redundancy is added. The
ﬁlling of the FEC matrix and the encoding are done in the
same way as in DVBH. 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 DVBS2 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
mechanismto adapt the output of the packetlevel FECto 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 ReedSolomon (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
W
r
i
t
i
n
g
d
i
r
e
c
t
i
o
n
FEC matrix
1 2 3 188 189 190 191 192 193 254 255
· · · · · ·
Data submatrix Redundancy submatrix
IP packet encapsulation with GSE Percolumn GSE encapsulation
C
o
l
u
m
n
s
i
z
e
I
P
p
a
c
k
e
t
1
I
P
p
a
c
k
e
t
2
I
P
p
a
c
k
e
t
1
(
c
o
n
t
.
)
I
P
p
a
c
k
e
t
3
I
P
p
a
c
k
e
t
2
(
c
o
n
t
.
)
P
a
d
d
i
n
g
L
a
s
t
I
P
p
a
c
k
e
t
(
c
o
n
t
.
)
P
a
d
d
i
n
g
P
a
d
d
i
n
g
P
a
d
d
i
n
g
1
s
t
r
e
d
u
n
d
a
n
c
y
c
o
l
u
m
n
2
n
d
r
e
d
u
n
d
a
n
c
y
c
o
l
u
m
n
P
u
n
c
t
u
r
e
d
c
o
l
u
m
n
P
u
n
c
t
u
r
e
d
c
o
l
u
m
n
Figure 7: Arrangement of IP packets for FEC encoding.
After GSE encapsulation, the GSE packets are introduced
in BBFRAMEs and transmitted. On the receive side, erro
neous BBFRAMEs are detected by checking the CRC. The
receiver reconstructs the FEC matrix and marks any column
that is totally or partially received by means on an erroneous
BBFRAME as unreliable. Finally, if the reconstructed FEC
matrix has no more than 64 unreliable columns, the code
can correctly compute all bytes in the matrix. If there are
more than 64 unreliable columns, the code cannot correct
anything, and only those columns received by means of cor
rect BBFRAMEs will be correct.
5. SIMULATIONSCENARIOS
In the following, the simulation platforms used to evaluate
the performance of DVBS2 with advanced fade countermea
sures in the railway environment as described in Section 3 are
duly detailed.
5.1. Advanced physical layer simulation platform
To cover a rather large set of spectral eﬃciency, four MOD
CODs have been considered: 1/2QPSK, 2/38PSK, 3/4
16APSK, and 5/616APSK. The LOS channel condition
(Rice factor equal to 17.4 dB) and the train speed equal to
300 km/h have been simulated. Equally spaced power arches
with a separation of 50 m have been included in some sce
narios, with a duty cycle of 1%, corresponding to a width of
0.5 m in accordance with Figure 2. The symbol rate was ﬁxed
to 27.5 Mbaud.
The considered DVBS2 physical layer transmitter [2] is
depicted in Figure 8. A continuous stream of MPEG pack
ets passes through the mode adaptation which provides
input stream interfacing. This data ﬂow is passed to the
merger/slicer that, depending on the applications, allocates
a number of input bits equal to the maximum data ﬁeld ca
pacity. In this way, user packets are broken in subsequent
data ﬁelds, or an integer number of packets are allocated in
it. Then, a ﬁxed length baseband header (BBHEADER) of
80 bits is inserted in front of the data ﬁeld, describing its for
mat. For example, it reports to the decoder the input streams
format, the mode adaptation type and the rolloﬀ factor.
The eﬃciency loss introduced by this header varies from
0.25% to 1% for long and short codeword lengths, respec
tively. The role of stream adaptation is to provide padding
when needed, in order to complete a constant length frame,
and scrambling. Padding is applied when the user data avail
able for transmission are not suﬃcient to completely ﬁll a
BBFRAME, or when more than one packet have to be allo
cated in a BBFRAME. The built frame is randomized using a
scrambling sequence generated by the pseudorandom binary
sequence described by the polynomial (1 + X
14
+ X
15
). After
this scrambling, each BBFRAME is processed by the forward
error correction (FEC) encoder which is carried out by the
concatenation of a BoseChaudhuriHocquenghem (BCH)
outer code and an LDPC inner code. Available coderates
for the inner code are 1/4, 1/3, 2/5, 1/2, 3/5, 2/3, 3/4, 4/5,
5/6, 8/9, and 9/10. Depending on the application area, code
words can have length N
LDPC
= 64800 bits or 16200 bits. In
the following, the case of 64800 bits is considered. Regard
ing the modulation format, each coded BBFRAME can be
mapped onto QPSK, 8PSK, 16APSK, or 32APSK constella
tions. Modulated streams enter in the physical layer framing
where physical layer signalling and pilot symbols are inserted.
For energy dispersal, another scrambling sequence is applied
to the entire physical layer frame (PLFRAME). The system
has been designed to provide a regular PLFRAME structure,
based on slots of M = 90 modulated symbols, which allow
10 EURASIP Journal on Wireless Communications and Networking
Single/multiple
input data streams
Input
interface no. 1
BB
signaling
Merge
slicer
Stream
adapter
Input
interface no. n
.
.
.
Mode & stream
adaptation
1/4, 1/3, 2/5,
1/2, 3/5, 2/3,
3/4, 4/5, 5/6,
8/9, 9/10
BCH
LDPC
bit interleaver
FEC coding
QPSK
8PSK
16APSK
32APSK
Mapping
PL signaling
pilot symbols
Scram
bler
Dummy
frame
PL framing
Rolloﬀ factors:
α = 0.2,
α = 0.25,
α = 0.35
BB
ﬁlter
Modulation
BBFRAME FECFRAME PLFRAME
To the RF
satellite
channel
Figure 8: DVBS2 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 startofframe (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 squaredroot
raised cosine (SRRC) ﬁlter with variable rolloﬀ factors (0.2,
or 0.25, or 0.35) is considered. To cope with the intrinsic
nonlinearity of the onboard 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,
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 feedback 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
LDPCBCH 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 physicallayer simulator described
in the previous section interfaces with the packetlevel sim
ulator shown in Figure 9. This takes as input a stream of
IP packets and applies the GSEFEC encoding technique as
described above, generating a sequence of BBFRAMEs. At
this point, the output of the physicallayer simulator is used
to mark the BBFRAMEs as correctly or wrongly received.
Next, the GSEFEC decoding process is applied. The eﬀect
of the BBFRAMEs on the GSE units and subsequently on the
columns of the reconstructed FECmatrix is calculated. Then,
the correction capability of the ReedSolomon code is taken
into account to eliminate, if possible, the unreliable columns
Stefano Cioni et al. 11
Traﬃc generation
IP packets
GSEFEC
BBFRAMEs
Selective BBFRAME
corruption
Physicallayer
simulation
Time series of
correct/wrong BBFRAMEs
IP PER
calculation
Mapping to
correct/wrong
IP packets
Corrected FEC
matrix
FEC decoding
Mapping to
correct/wrong
FEC matrix
columns
Mapping to
correct/wrong
GSE units
Figure 9: Simulation platform at IPBBFRAME 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 packetlevel simulator is useful to assess very quickly
the performance of diﬀerent parameter conﬁgurations of
the GSEFEC since diﬀerent combinations can be simulated
without the need of repeating the timeconsuming physical
layer simulations. The main parameters of GSEFEC 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 transmitreceive chain described in Section 5.1. The
introduction of the second receiving antenna adopting the
1E −04
1E −03
1E −02
1E −01
1E + 0
P
E
R
−2 −1 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
E
b
/N
0
(dB)
1/2  QPSK (LOS, FAST, noPA)
2/3  8PSK (LOS, FAST, noPA)
3/4  16APSK (LOS, FAST, noPA)
5/6  16APSK (LOS, FAST, noPA)
1/2  QPSK (LOS, FAST, MRC)
2/3  8PSK (LOS, FAST, MRC)
3/4  16APSK (LOS, FAST, MRC)
5/6  16APSK (LOS, FAST, MRC)
Power arches ﬂoor
Figure 10: MRC performance in LOS channel condition and train
speed equal to 300 km/h.
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 E
b
/N
0
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 = 1/2QPSK and a
working E
b
/N
0
= 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 E
b
/N
0
. The second waterfall region starts only
for E
b
/N
0
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
GSEFEC method; second, to analyze the performance of
GSEFEC under various conﬁgurations. In all cases, a sce
nario with lineofsight propagation has been used.
12 EURASIP Journal on Wireless Communications and Networking
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
P
r
o
b
a
b
i
l
i
t
y
0 2 4 6 8 10 12 14
Error burst length
1/2QPSK
(a)
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
P
r
o
b
a
b
i
l
i
t
y
0 2 4 6 8 10 12 14
Error burst length
2/38PSK
(b)
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
P
r
o
b
a
b
i
l
i
t
y
0 2 4 6 8 10 12 14
Error burst length
5/616APSK
(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
First of all, it is worth remarking that the appropriate size
of the FEC matrix depends on the length of the bursts of
erroneous BBFRAMEs. It is clear that longer bursts will re
quire larger FEC matrices to avoid that the number of wrong
columns exceeds the correction capability of the code. There
fore, the design of the height of the FEC matrix should be
derived from an analysis of the length of the error bursts.
Figure 11 shows the histogram of the length of the bursts for
some particular MODCODmodes for the scenario described
above. In all modes besides the two shown in Figure 11,
it is observed that the distribution is bimodal. The bursts
of short length (typically between 1 and 4 BBFRAMEs) are
due to random errors caused by noise, whereas the rest of
bursts are caused by the power arches. Second, the higher
the modulation order, the longer the error bursts produced
by power arches are. This is justiﬁed by the fact that ac
cording to the DVBS2 standard, BBFRAMEs are coded and
converted into FECFRAMEs, which have constant length
in bits regardless of the used modulation [2]. The bits in
the FECFRAME are transformed by the modulator into
bytes in the PLFRAMEs. Higher modulations need fewer
symbols and, hence, less time to transmit an FECFRAME.
The duration of the fade event caused by a power arch
only depends on the speed of the train, which we have
considered to be 300 km/h throughout the rest of the pa
per. Therefore, the shorter the PLFRAME, the more PL
FRAMEs and hence BBFRAMEs are aﬀected by each power
arch.
In order to present the procedure to compute the col
umn size of the FEC matrix, we consider a numerical exam
ple. We use for instance the least eﬃcient MODCOD, that
is, 1/2QPSK. It can be seen in Figure 11 that the maximum
error burst length due to power arches is 7 BBFRAMEs. In
this MODCOD, each BBFRAME has a data ﬁeld of length
32128 bits [2], which is equal to 4016 bytes. Therefore, a
burst of 7 BBFRAMES corresponds to 28112 bytes. We con
sider that this amount of bytes should correspond to less
than 30 columns in the FEC matrix. The value of 30 has
been chosen arbitrarily. It is nevertheless a reasonable num
ber since the objective is to leave a margin with respect to the
64 columns that the code can correct (assuming no punctur
ing) so as to be able to cope with errors caused by noise as
well. Therefore, the column size of the FEC matrix should
fulﬁl
30L
c
≥ 28112 =⇒L
c
≥ 938 bytes, (7)
where L
c
is the number of rows (i.e., the length of each col
umn) of the FEC matrix in bytes. In the previous compu
tation, we have not taken into account the overhead intro
duced by GSE since it is small and we are only interested
in obtaining an approximate value for the column size. If
the same calculation is repeated for the most eﬃcient MOD
COD, that is, 5/616APSK, the result is L
c
≥ 2912 bytes. The
Stefano Cioni et al. 13
0
0.02
0.04
0.06
0.08
0.1
0.12
I
P
p
a
c
k
e
t
e
r
r
o
r
r
a
t
e
0 1000 2000 3000 4000 5000 6000
Column size (bytes)
1/2QPSK
(a)
0
0.02
0.04
0.06
0.08
0.1
0.12
I
P
p
a
c
k
e
t
e
r
r
o
r
r
a
t
e
0 1000 2000 3000 4000 5000 6000
Column size (bytes)
2/38PSK
(b)
0
0.02
0.04
0.06
0.08
0.1
0.12
I
P
p
a
c
k
e
t
e
r
r
o
r
r
a
t
e
0 1000 2000 3000 4000 5000 6000
Column size (bytes)
3/416APSK
(c)
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14
I
P
p
a
c
k
e
t
e
r
r
o
r
r
a
t
e
0 1000 2000 3000 4000 5000 6000
Column size (bytes)
5/616APSK
(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 PLFEC, squares →results with GSEFEC).
results for the intermediate MODCODs, 2/38PSK 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 GSEFEC 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 GSEFECmay be limited for lowtrain 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 followon paper). A ﬁxed IP packet
length equal to 576 bytes has been considered.
14 EURASIP Journal on Wireless Communications and Networking
0
0.005
0.01
0.015
0.02
0.025
I
P
p
a
c
k
e
t
e
r
r
o
r
r
a
t
e
0 1000 2000 3000 4000 5000 6000
Column size (bytes)
1/2QPSK
(a)
0
0.005
0.01
0.015
0.02
0.025
I
P
p
a
c
k
e
t
e
r
r
o
r
r
a
t
e
0 1000 2000 3000 4000 5000 6000
Column size (bytes)
2/38PSK
(b)
0
0.005
0.01
0.015
0.02
I
P
p
a
c
k
e
t
e
r
r
o
r
r
a
t
e
0 1000 2000 3000 4000 5000 6000
Column size (bytes)
3/416APSK
(c)
0
0.005
0.01
0.015
0.02
0.025
0.03
I
P
p
a
c
k
e
t
e
r
r
o
r
r
a
t
e
0 1000 2000 3000 4000 5000 6000
Column size (bytes)
5/616APSK
(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 PLFEC, squares →results with GSEFEC).
Figures 12 and 13 also compare the results obtained when
GSEFEC is used and when no packetlevel FEC is applied.
The baseline GSEFEC 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 packetlevel FEC follows the same
architecture as for GSEFEC, 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 packetlevel 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, GSEFEC, 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 backoftheenvelope 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
signaltonoise 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. 15
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14
0.16
I
P
p
a
c
k
e
t
e
r
r
o
r
r
a
t
e
0 500 1000 1500
IP packet length (bytes)
1/2QPSK
No PLFEC
Column size: 1024 bytes
Column size: 4096 bytes
(a)
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14
0.16
I
P
p
a
c
k
e
t
e
r
r
o
r
r
a
t
e
0 500 1000 1500
IP packet length (bytes)
5/616APSK
No PLFEC
Column size: 1024 bytes
Column size: 4096 bytes
(b)
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 (1/2QPSK and 5/616APSK).
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 GSEFEC results
The analysis of the GSEFEC 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;
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 physicallayer conditions. The padding and puncturing
of columns in the FEC matrix are other degrees of freedom
that can be exploited in the parameterization of GSEFEC.
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 EURASIP Journal on Wireless Communications and Networking
Table 2: Parameters of the GSEFEC algorithm.
FECmatrix column
size (bytes)
256 512 768 1024 2048 3072 4096 5120
GSE units per column 1 1 1 2 2 3 4 5
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 crosslayer 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 packetlevel forward
error correction mechanism especially tailored to DVBS2,
named GSEFEC. 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 showthat eﬀective countermeasures to compensate the
impairments of the railroad satellite channel are possible and
can be integrated into the existing DVBS2 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) crosslayer optimization of all the relevant parameters
(MODCODs and GSEFEC), 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
GSEFEC scheme.
ACKNOWLEDGMENT
This work was supported and partially funded by Sat
NEx, the Satellite Communications Network of Excellence
(www.satnex.org), FP6 Contract IST507052.
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 (DVBH).
[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 Kuband,” to ap
pear in IEEE Transactions on Vehicular Technology.
[9] E. Kubista, F. P. Fontan, M. A. V. Castro, S. Buonomo,
B. R. ArbesserRastburg, 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 Kuband,” in
Proceedings of the International Workshop of Cost Actions 272
and 280, Noordwijk, The Netherlands, May 2003.
[12] S. Scalise, O. L¨ ucke, 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. VanelliCoralli, “Antenna di
versity for DVBS2 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. VanelliCoralli,
“Antenna diversity for DVBS2 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. VanelliCoralli, C. P´ arraga Niebla, S. Scalise, G.
Seco Granados, and M.A. V´ azquez Castro, “Antenna diver
sity and GSEbased packet level FEC for DVBS2 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);
DVBH 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 Kaband 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 TradeOff in
the DVBS2 Forward Link
Matteo Berioli, Christian Kissling, and R´ emi 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 DVBS2 forward links. By reducing the socalled 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 tradeoﬀ,
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 DVBS2 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
The original DVBS standard dates back to 1995 and was in
tended for delivery of broadcasting services, the underlying
transport streamof DVBS was deﬁned to be MPEG2. DVB
S2 [1] is the second generation of the DVBS 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 DVBRCS [2]. DVB
S2 can achieve a capacity increase of up to 30% under the
same transmission conditions compared to the older DVBS
standard what is achieved by applying higher order modu
lation schemes and by the use of low density parity check
codes (LDPC) and BoseChaudhuriHochquenghem (BCH)
codes.
The real novelty introduced by DVBS2 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
mitigation techniques like power control result in an inef
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 DVBS2
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 DVBS2 forward links.
Today DVBS2 links oﬀer to the higherlayers protocols
a terrestriallike 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., MPEG4 [5]) have been
designed to be highly resilient to residual errors in the in
put bitstream, 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 errortolerant codecs, the IETF
has also standardized a new multimedia transport proto
col, UDPLite [7], that allows to specify the required level of
payload protection, while maintaining endtoend 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 DVBS2 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 crosslayer
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 tradeoﬀ between errors and
capacity. The present work analyzes this tradeoﬀ, 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. BACKGROUNDANDSCENARIO
2.1. Overviewof DVBS2
The second generation of DVBS 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
framebyframe decision of the modulation and coding com
bination (ModCod) to be applied based on these measure
ments. DVBS2 oﬀers a broad range of modulations and cod
ings for ACM. The supported modulation schemes comprise
QPSK, 8PSK, 16APSK, 32APSK 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 higherorder modu
lation in combination with a low coding rate (e.g., for clear
sky conditions) allows to transmit more bits per symbol than
a loworder 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 lowerorder 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
DVBS2 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
CRC8 error detection coding scheme. It is possible to merge
diﬀerent input streams together and to segment them into
the socalled data ﬁelds which are the payload part of the so
called basebandframes (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 higherorder modulation
schemes the subsequent mapping block performs a serialto
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 socalled 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
Single
input
stream
Multiple
input
streams
Input
interface
Input stream
synchroniser
Nullpacket
deletion
(ACM, TS)
CRC8
encoder
Buﬀer
BB
signalling
Merger
slicer
Buﬀer
CRC8
encoder
Nullpacket
deletion
(ACM, TS)
Input stream
synchroniser
Input
interface
Data
ACM
command
Mode adaptation
Dotted subsystems are
not relevant for
single transport stream
broadcasting
applications
Padder
BB
scram
BLER
Stream
adaptation
Rates 1/4, 1/3, 2/5
1/2, 3/5, 2/3, 3/4, 4/5,
5/6, 8/9, 9/10
BCH
encoder
(n
bch
, k
bch
)
LDPC
encoder
(n
ldpc
, k
ldpc
)
Bit
inter
leaver
FEC encoding
QPSK,
8PSK,
16APSK,
32APSK
Bit
mapper
into
constellations
Mapping
I
Q
PL signalling &
pilot insertion
PL
scram
BLER
Dummy
PLFRAME
insertion
PLFRAMING
α = 0.35, 0.25, 0.2
BB ﬁlter
and
quadrature
modulation
Modulation
BBHEADER
data ﬁeld
BBFRAME
LP stream for
BC modes
FECFRAME
PLFRAME
To the RF
satellite
channel
Figure 1: DVBS2 system architecture [1].
80 bits DFL K
bch
DFL80
BBHEADER Data ﬁeld Padding
BBFRAME (K
bch
bits)
Figure 2: Structure of a BBFRAME [1].
N
bch
= k
ldpc
K
bch
N
bch
K
bch
n
ldpc
k
ldpc
BBFRAME BCHFEC LDPCFEC
(n
ldpc
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 quasierrorfree transmission as long as
the critical SNR(signaltonoise ratio) demodulation thresh
old for this ModCod (thr
dem
(ModCod)) is not crossed. If the
signal drops below thr
dem
(ModCod) then the BER will dras
tically increase due to the nature of the applied LDPC and
BCH coding of having very steep BERversusSNR curves. In
the GEOstationary 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
eral hundreds milliseconds (250 milliseconds). This means
that though the order of magnitude for the propagation de
lay allows for a compensation of very slow changing channel
eﬀects, like rain attenuation, it is too long to compensate fast,
highfrequent changes in the SNR as those caused by scintil
lation, this will be explained in the next section.
2.2. Channel modelling
The selection of a ModCod scheme for transmission is very
decisive for the performance of the system in terms of net
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 Kband 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 smallscale irregularities in electron density in the iono
sphere [8].
The scintillation in Kband 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
10
1
10
0
10
−1
10
−2
10
−3
10
−4
f
s
f
a
Frequency (Hz)
10
−6
10
−4
10
−2
10
0
10
2
10
4
P
S
D
(
d
B
2
/
H
z
)
Power spectral density of rain attenuation and scintillation
Rain attenuation
Scintillation
∼ f
−8/3
∼ f
−2
Figure 4: Attenuation and scintillation spectrum (typical values:
f
a
≈ 10
−4
Hz, f
s
≈ 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 = σ
0
· A
5/12
. (1)
The value σ
0
is the standard deviation of the scintillation for
a rain attenuation of A[dB]. [10] suggests a typical value of
0.039 for σ
0
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 f
a
(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
8000 7000 6000 5000 4000 3000 2000 1000 0
Time (s)
−0.25
−0.2
−0.15
−0.1
−0.05
0
0.05
0.1
0.15
0.2
0.25
S
c
i
n
t
i
l
l
a
t
i
o
n
(
d
B
)
Scintillation and attenuation time series
of useful user (forward downlink)
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
A
t
t
e
n
u
a
t
i
o
n
(
d
B
)
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
thr
dem
(ModCod) exists which is needed to be able to decode
the frame with a quasizero 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
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 thr
dem
(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 thr
dem
(N − 1) and thr
dem
(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 thr
dem
(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. 5
ModCod (thr
dem
(N) in this case) is crossed but after a higher
threshold is exceeded (thr
enable
(N −1)).
In case the signal strength decreases again, the ModCod is
not switched when the enabling threshold thr
enable
is crossed,
but just when an additional hysteresis margin is exceeded.
The threshold for switching to a smaller ModCod is denoted
as thr
down
(ModCod) and the size of the hysteresis margin as
(Δthr
hyst
(N)). The distance between the critical demodula
tion threshold thr
dem
(ModCod) and the threshold that trig
gers a downswitching thr
down
(ModCod) is called the safety
margin Δthr
safety
(ModCod).
The safety margin(Δthr
safety
(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 thr
dem
(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. W¨ orz 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 “W¨ orzSchweikert safety margins” denotes
the safety margins calculated according to the algorithm pre
sented in [11] while “zerosafety 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. SYSTEMMODELLING
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 tradeoﬀ, it
ModCod N −1 ModCod N
thr
enable
(N −1)
thr
down
(N −1)
Δthr
safety
(N −1)
thr
dem
(N −1)
Δthr
hyst
(N)
Δthr
safety
(N)
No switching
here because
of hysteresis
Signal
thr
down
(N)
thr
dem
(N)
Time
S
N
I
R
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 DVBS2 standard
can provide the higherlayers protocols with a quasierror
free underlaying physical layer (PER = 10
−7
). For this pur
pose regular 8bytes 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
W¨ orz 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 systemaccording 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, N
G
, and the bad one, N
B
; 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 tradeoff
The second level of analysis describes the details of each state
of the Markov chain. The good states present quasierrorfree
conditions according to the DVBS2 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 BERversus
SNR functions, which were used in the simulations and to
derive the Markov chain parameters, were taken from [12].
In that work, endtoend performances of the BER versus
the SNR are presented for the DVBS2 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 W¨ orzSchweikert safety
margins, the other one using zerosafety margin bounds with
W¨ orzSchweikert 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 (i
B
and j
B
), the second one
the good state (i
G
and j
G
). Figures 7 and 8 show the transi
tion matrices for zerosafety margin and the W¨ orzSchweikert
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
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
12 11 13 14 18 19 20 21
Figure 7: Transition matrix for zerosafety margin.
1 2 3 4 5 6 7 8 9
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 ,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, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 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, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 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 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 ,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, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 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, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 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, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 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, 00 0, 00 0, 00 0, 00 0, 00 0, 00 0, 00 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, 00 0, 00 0, 00 0, 00 0, 00 0, 00 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, 00 0, 00 0, 00 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, 00 0, 00 0, 00 0, 00 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, 00 0, 00 0, 00 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, 00 0, 00 0, 00 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, 00 0, 00 0, 00 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, 00 0, 00 0, 00 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, 00 0,00 0 ,00
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
18 19 20 21 12 11 13 14
Figure 8: Transition matrix for W¨ orzSchweikert safety margin.
more “stable” states are the good states. This makes sense if
we look at the SNR time series, because switchings between
ModCods are quite spaced in time compared to the time step
of 0.1 second. Simulations estimate an average number of 4.5
ModCod switchings per hour.
Moreover on the diagonal, the probabilities of remaining
in a bad state are not so high; this is also correct, since when
we are in a bad state, we know that a downswitch should
occur. The only bad state which has a higher stability is the
bad state for ModCod 1
B
. This results fromthe fact that when
the SNR goes below the last demodulation threshold, the sys
tem cannot switch to a lower ModCod, so it remains in bad
state until the SNR rises again. This is basically an outage
where the DVBS2 receiver is not available; the simulator was
designed to give a system availability of 99.96% of the time,
for both approaches.
With the W¨ orzSchweikert scheme, shown in Figure 8, the
matrix is far more sparse, and no bad states are ever ac
tive, except for ModCod 1
B
because of system unavailabil
ity. Transitions only occur between good states and this con
ﬁrms that this approach is designed to work only in good
states.
For the novel approach, the zerosafety margin one, it
may be interesting to derive the probability to be in each
ModCod (bad or good state). Once the transition matrix
for the ZeroSafety margin is solved [13], we end up with
Figure 9 which shows a stacked probability density graph for
good and bad states of each ModCod. This is the result of
a simulation of an equivalent of 4.5 years of SNR evolution
over time. What we can see is that the most used ModCods
are those whose demodulation threshold is just below the
SNIR in clear sky conditions. That makes sense because most
of the time we are in clear sky conditions, so we use the high
est ModCods. We can also notice the high value of the bad
state in ModCod 1
B
, because of system unavailability. Some
ModCods are never used due to overlapping with other ones,
some ModCods achieve a better spectral eﬃciency requiring
less SNR.
8 EURASIP Journal on Wireless Communications and Networking
28 26 24 22 20 18 16 14 12 10 8 6 4 2
ModCods
10
−6
10
−5
10
−4
10
−3
10
−2
10
−1
10
0
P
r
o
b
a
b
i
l
i
t
y
Bad states
Good states
Figure 9: State probabilities.
4.2. Error rate versus capacity tradeoff
This section presents the main results which are obtained
when reducing the safety margins, in terms of increase spec
tral eﬃciency and increase errors. The starting point is the set
of threshold selected by W¨ orzSchweikert; this set guarantees
a quasierrorfree system operation. We try to proportionally
reduce those margins and even to have negative margins, to
see how the system performs. The x axis in Figures 10 and 11
represents factors to be multiplied to the W¨ orzSchweikert
set to get the tested thresholds. This means that for multiply
ing factor 1 we have the W¨ orzSchweikert set, for the factor
0, we have the zerosafety margin approach, and for negative
values of the factors we are testing thresholds which are be
low those thresholds recommended by the DVBS2 standard.
This may seem strange, but it will appear clear how useful
this is to show that there is a tradeoﬀ between errors and
increase in capacity.
Figure 10 shows (as expected) that the PER objective of
10
−7
is achieved already before the W¨ orzSchweikert bounds.
This is not surprising since the model has been designed to
do so. As expected as well, PER and BER are fastgrowing up
to 1 when the safety margin becomes negative. A surprising
fact here is that there are possibilities to achieve the goal PER
even for margins which are 0.4 times the W¨ orzSchweikert
safety margins. That means that those W¨ orzSchweikert mar
gins may not be the optimum selection.
Figure 11 shows the core result of this work. A trivial
thing is that the gross capacity (total amount of received bits
with failures) is still increasing when we go for lower and
lower bounds, because of course we are using less and less
robust ModCods that provide better spectral eﬃciency. The
very interesting point comes with the fact that the net ca
pacity (throughput of correct bits) shows a maximum in the
negative part of the scaling factor: −0.4 at the packet level
0.5 0 −0.5 −1 −1.5 −2 −2.5 −3
Multiplying factor on SchweikertW¨ orz safety margin
10
−10
10
−9
10
−8
10
−7
10
−6
10
−5
10
−4
10
−3
10
−2
10
−1
10
0
E
r
r
o
r
r
a
t
e
Packet error rate
Bit error rate
6 ×10
−3
4 ×10
−3
Figure 10: Packet error rate (PER) versus Safety margin.
1 0.5 0 −0.5 −1 −1.5 −2 −2.5 −3
Multiplying factor on SchweikertW¨ orz safety margin
0
0.5
1
1.5
2
2.5
3
3.5
A
v
e
r
a
g
e
s
p
e
c
t
r
a
l
e
ﬃ
c
i
e
n
c
y
(
b
/
s
/
H
z
)
Max @−1.5 = 3.12 b/Hz/s Max @−0.4 = 3.01 b/Hz/s
Gross eﬃciency
Net eﬃciency (without bit aggregation)
Net eﬃciency (with bit aggregation)
Figure 11: Average spectral eﬃciency versus safety margin.
and −1.5 at the bit level. Corresponding values of PER/BER
at these maxima are 6 · 10
−3
and 4 · 10
−3
. The two curves
represent the two ways of operating described in Section 3:
bit aggregation is when failures cause BBFRAME discard, no
bit aggregation means when the frame is passed to the higher
layers with failures. It should be noted that for bit error ag
gregation (see Figure 11) the PER (see Figure 10) is the rele
vant result since in case of a bit error the complete BBFRAME
is discarded. Without consideration of bit error aggregation,
the BER is the relevant result since erroneous bits within the
BBFRAME are expected to be corrected by the higher layers.
This means that a system which wants to have the indicated
throughput with or without bit aggregation, is operating at
those PER/BER.
Matteo Berioli et al. 9
2 1.9 1.8 1.7 1.6 1.5 1.4 1.3 1.2 1.1 1
×10
4
Time in samples (0.1 s)
3
4
5
6
7
8
9
S
N
I
R
a
t
r
e
c
e
i
v
e
r
a
n
d
M
o
d
C
o
d
s
e
l
e
c
t
i
o
n
10
−7
10
−6
10
−5
10
−4
10
−3
10
−2
10
−1
10
0
M
P
E
G

2
p
a
c
k
e
t
e
r
r
o
r
r
a
t
e
Figure 12: PER with SNR and ModCod selection for zerosafety
margin.
5500 5000 4500 4000 3500 3000 2500 2000 1500 1000 500 0
Interarrival time of error bursts (event where PER > 10
−6
) (s)
10
−3
10
−2
10
−1
10
0
P
r
o
b
a
b
i
l
i
t
y
Probability density function for interarrival times
Figure 13: Interarrival times distribution.
If a system can cope with these error rates, then it may be
interesting to design it with lower safety margins than those
in the W¨ orzSchweikert strategy, in order to gain throughput.
4.3. Error bursts analysis
To deeper investigate the quality of the transmission in case
we reduce the safety margins, we have to look at the dis
tribution in time of the error bursts. Figure 12 shows an
example of simulated SNR time series with corresponding
PER for zero safety margin. In contrast to Figure 10 which
shows the averaged error PERs and BERs, here we investigate
the distribution of the interarrival times between two PER
peaks (without averaging), considering a detection threshold
of PER = 10
−6
. The simulation that has led to Figure 13 has
been worked out on 4.5 years of simulated SNR, and it was
conducted with the zerosafety margins approach.
> 0.5
0.5
0.4
0.3
0.2
0.1
Fade duration (s)
1e −1
1e −2
1e −3
1e −4
1e −5
1e −6
P
E
R
l
e
v
e
l
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
E
v
e
n
t
s
p
e
r
h
o
u
r
Figure 14: State probabilities.
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
switched down one by one, and as we saw on Figure 12, error
peaks often occur at every downswitch. The question is now
what is the duration/severity of these peaks?
Figure 14 shows the number of fade events per hour us
ing zerosafety margins, sorted by their duration and PER
strength. A sequence of samples is considered as one fade
event if the associated PER is exceeding a given level. For
each PERlevel, Figure 14 shows the number of fades per hour
which exceed this PER level. For this graph, we have 6 dif
ferent PER levels, and the fade events are distributed among
their duration. We can see that the shorter fades are the ones
that occur most of the time. This comes from the fact that for
a normal process like scintillation, the probability of having a
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
half a second (twice the GEO propagation delay), which is
the time needed to signal to the gateway the fading situation
and to receive a new transmission with a new ModCod. So
in theory fade should not exceed 500 milliseconds, but the
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
that. First, if we are in the highest ModCod of a couple of very
close ModCods (in terms of demodulation threshold) and we
enter a strong rain fade, with a steep decreasing SNR, it can
happen that the SNR crosses the demodulation threshold of
the lowest ModCod before the system has switched down.
This results in a bad state to bad state transition, and we can
see some of these cases in the transition matrix (12
B
to 9
B
or
13
B
to 11
B
, e.g.). A second explanation is the following, non
negligible contributions to this behavior are the outages due
to system nonavailability, that is the fades that occur in the
lowest ModCod.
10 EURASIP Journal on Wireless Communications and Networking
5. CONCLUSIONS
The possibility to have a quasierrorfree transmission chan
nel in DVBS2 systems is not always an optimal solution in
case the higherlayer protocols do not require such high per
formance. In this case the lower layers can provide a trans
mission with some resilient errors, and exploit more the
spectrum to gain in throughput. The errorcapacity tradeoﬀ
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 sideoutcome of this work is the development of
Markov chain to model the ModCod transitions and the fail
ure occurrence in a DVBS2 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, “Crosslayer 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 144962, “Coding of audiovisual objects (MPEG
4)—part 2: visual,” 2004.
[6] ETSI TR 126 975, “Performance Characterisation of the Adap
tive MultiRate (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. DattaBarua, 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,
McGrawHill, 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. W¨ orz, 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.
[12] E. Casini, R. De Gaudenzi, and A. Ginesi, “DVBS2 modem
algorithms design and performance over typical satellite chan
nels,” International Journal of Satellite Communications and
Networking, vol. 22, no. 3, pp. 281–318, 2004.
[13] M. F. Neuts, MatrixGeometric Solutions in Stochastic Models:
An Algorithmic Approach, Dover, Mineola, NY, USA, 1981.
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 MultibeamSatellite Systems
J. P. Millerioux,
1, 2, 3, 4
M. L. Boucheret,
2
C. Bazile,
3
and A. Ducasse
5
1
T´eSA, 1416 Port SaintEtienne, 31000 Toulouse, France
2
Institut de Recherche en Informatique de Toulouse, Ecole Nationale Sup´erieure d’Electrotechnique, d’Electronique,
d’Informatique, d’Hydraulique et des T´el´ecommunications, 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 Sup´erieure des T´el´ecommunications, 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 VanelliCoralli
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 DVBRCS 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 timeinvariant
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 singleuser 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
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 DVBRCS stan
dard [2], with the use of convolutional coding. The algorithm
is derived for a symbolasynchronous timeinvariant 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.
Other approaches have been proposed in the literature
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 MMSESIC 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 timeinvariant channel, Section 4 is
2 EURASIP Journal on Wireless Communications and Networking
Information
bits user k
Encoder
QPSK
mapping
Π
k
Pilot
symbols
insertion
T
k
d
k
[n]
(a)
d
k
[n] s(t −τ
k
)
ρ
k
e
jϕ
k
(t)
x
1
(t)
x
k
(t)
x
K
(t)
H
n
k
(t)
y
k
(t)
(b)
Figure 1: Transmitter and channel model.
dedicated to the study of possible adaptations with frequency
oﬀsets, and we draw conclusions in Section 5.
2. SYSTEMMODEL ANDASSUMPTIONS
2.1. Model
The considered context is the reverse link of a ﬁxedsatellite
service with a regenerative geostationary satellite, a multi
beam coverage with a regular frequency reuse pattern [6],
and an MFTDMA access [2]. A “slot synchronous” system
is assumed. Multiuser detection is performed onboard the
satellite, after frequency demultiplexing. We choose here to
work on a ﬁctitious interference conﬁguration characterized
by carrier to interference ratios C/I. A more detailed presen
tation can be found in [1] or [7].
We consider in the following a frequency/time slot in
the MFTDMA frame. Notations are relative to complex en
velops. ·
∗
, ·
T
, ·
H
, E(·), and · ∗ · denote, respectively, the
conjugate, transpose, conjugate transpose, expected value,
and convolution operators. Consider K uplink signals asso
ciated to K diﬀerent cochannel cells. Under the narrowband
assumption [8], we get
y(t) = Hx(t) + n(t), (1)
where x(t) = [x
1
(t) · · · x
K
(t)]
T
is the K × 1 vector of re
ceived signals, y(t) = [y
1
(t) · · · y
K
(t)]
T
is the K × 1 vec
tor of signals at the beamformer outputs, H is the K × K
beamforming matrix (i.e., the product of the matrix of steer
ing vectors by the matrix of beamformer coeﬃcients), and
n(t) = [n
1
(t) · · · n
K
(t)]
T
is the vector of additive noises.
Without loss of generality, we consider that the matrix H
has its diagonal coeﬃcients equal to 1. Additive noises are
additive white Gaussian noises (AWGN) with the same vari
ance σ
2
, and are characterized by a spatial covariance matrix
R
n
= E(n(t)n(t)
H
) which depends on the antenna imple
mentation [1].
As regards to the waveform, the information bits are con
volutionally encoded, and the coded bits are then mapped
onto QPSKsymbols which are interleaved diﬀerently on each
beam. A burst of N symbols d
k
[n] is composed of these in
terleaved symbols in which pilot symbols are inserted. We
model the signals x
k
(t) as
x
k
(t) = ρ
k
e
jϕk(t)
N−1
_
n=0
d
k
[n]s
_
t −nT −τ
k
_
, (2)
where T, s(t), ρ
k
, ϕ
k
(t), τ
k
, 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 timevarying) carrier phase, and
its time delay. The whole transmitter and channel model is
summarized in Figure 1. Notice that a single frequency refer
ence is assumed onboard the satellite.
We deﬁne the signaltonoise ratio (SNR) for the kth sig
nal as
E
s
N
0
¸
¸
¸
¸
k
=
ρ
2
k
σ
2
. (3)
Assuming an equal SNR for all users, the carrier to interfer
ence ratio for the kth signal can be simply deﬁned as
C
I
¸
¸
¸
¸
k
=
_
_
l / =k
¸
¸
h
k,l
¸
¸
2
_
−1
. (4)
2.2. Assumptions
The algorithm is derived under the following assumptions.
(i) We assume a perfect singleuser frame synchronisation
and timing recovery (i.e., for the kth signal on the kth
beam).
(ii) The matrix His assumed time invariant on a burst du
ration, and unknown at the receiver.
(iii) Signiﬁcant interferers are only located in adjacent
cochannel cells: due to the regular reuse pattern, there
are at most 6 signiﬁcant interferers on a beam [6].
Let us recall that the algorithm considered in the follow
ing is suboptimal (see Section 1 and [1]): it only performs
interference cancellation for the kth signal at the output of
the kth beam.
3. ALGORITHMDESCRIPTIONONA TIME
INVARIANT CHANNEL
3.1. Synchronous case
To simplify the presentation, we ﬁrst consider a symbol
synchronous timeinvariant channel, that is, τ
k
= 0 and
ϕ
k
(t) = ϕ
k
for all k. After optimal sampling, we can then
consider the “oneshot” approach with
y[n] = Gd[n] + n[n], (5)
J. P. Millerioux et al. 3
yK[n]
Initial phase
recovery
Decoding
Estimation
of gK,.
Interference
cancellation
d
(m)
k
[n]
To beam l, for k
interfering on beam l
yk[n]
Initial phase
recovery
y
(m)
k
[n]
Decoding
Estimation
of gk,.
Interference
cancellation
y
(m+1)
k
[n]
d
(m)
l
[n]
From beam l, for l
interfering on beam k
y1[n]
Initial phase
recovery
Decoding
Estimation
of g1,.
Interference
cancellation
Figure 2: Block diagram of the receiver (synchronous case).
where
G =
_
g
T
1
· · · g
T
K
_
T
=
_
g
k,l
_
= Hdiag
_
ρ
k
exp
_
jϕ
k
__
,
d[n] =
_
d
1
[n] · · · d
K
[n]
_
T
,
y[n]=
_
y
1
[n] · · · y
K
[n]
_
T
with y
k
[n]=y
k
(t) ∗s(−t)
t=nT
,
n[n]=
_
n
1
[n] · · · n
K
[n]
_
T
with n
k
[n]=n
k
(t) ∗s(−t)
t=nT
,
E
_
n[k]n[l]
_
= δ(k −l)R
n
.
(6)
A synoptic of the receiver is given in Figure 2, where inter
leaving and deinterleaving operations are omitted for sim
plicity. All operations are performed in parallel on the dif
ferent beams, with exchange of information from one to an
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
mth iteration.
Channel estimation
The channel estimation on the kth beam is processed by
a leastsquare estimator using currently estimated symbols
(and including pilot symbols). At the mth iteration, we get
for the kth beam
g
(m)
k
=
_
N−1
_
n=0
y
k
[n]
d
(m)
[n]
H
__
N−1
_
n=0
d
(m)
[n]
d
(m)
[n]
H
_
−1
.
(7)
We only use for estimation (and consequently for interfer
ence cancellation in (8)) estimated symbols of the useful sig
nal and of adjacent interfering ones (see Section 2.2. assump
tion (iii)), which is not speciﬁed in the equations for the sake
of simplicity.
Interference 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
y
(m+1)
k
[n] = g
(m)
∗
k,k
_
y
k
[n] −
_
l / =k
g
(m)
k,l
d
(m)
l
[n]
_
. (8)
In the case of perfect channel estimation and interfering
symbol decisions, we get
y
(m+1)
k
[n] =
¸
¸
g
k,k
¸
¸
2
d
k
[n] + g
∗
k,k
n
k
[n], (9)
interference is entirely removed, and the carrier phase is per
fectly compensated.
Decoding
Decoding is performed by the Viterbi algorithm, by assimi
lating the residual interference plus noise after deinterleaving
at the decoder input to AWGN.
Initialization
For the kth user, an initial carrier phase is estimated from
pilot symbols on the kth beam. After phase compensation,
the signal received on the kth beam is sent to the decoding
block to initialize the iterative process.
3.2. Asynchronous case
We now consider a symbolasynchronous timeinvariant
channel, that is, τ
k
/ = τ
l
for k / = l, and ϕ
k
(t) = ϕ
k
for all k.
We introduce
u
k
(t) =
N−1
_
n=0
d
k
[n]s
_
t −nT −τ
k
_
,
u
(m)
k
(t) =
N−1
_
n=0
d
(m)
k
[n]s
_
t −nT −τ
k
_
,
(10)
and vectors u(t) = [u
1
(t) · · · u
K
(t)]
T
and u
(m)
(t) =
[ u
(m)
1
(t) · · · u
(m)
K
(t)]
T
.
We get
y(t) = Gu(t) + n(t), (11)
where G is deﬁned in Section 3.1. We refer to u
(m)
k
(t) as the
estimated kth signal at the mth iteration.
The algorithm on the asynchronous channel is then very
similar to the one on the synchronous channel. For the kth
beam, at the mth iteration:
(i) channel estimation is processed by a least square ap
proach using the estimated signals at the matched ﬁl
ter output u
(m)
(t) ∗ s(−t) and y
k
(t) ∗ s(−t), syn
chronously sampled, with 2 samples per symbol (sam
ples of u
(m)
(t)∗s(−t) corresponds to
d
(m)
[n] and sam
ples of y
k
(t) ∗s(−t) corresponds to y
k
[n] in (7));
4 EURASIP Journal on Wireless Communications and Networking
11 12 13 14
8 9 10
4 5 6 7
1 2 3
(a)
Cell number
Number of
interferers
C/I [dB]
1, 3 3 5
2 4 4
4, 7 3 5
5, 6 6 2
8, 10 5 3
9 6 2
11, 14 2 6
12, 13 4 4
(b)
Figure 3: Description of the studied conﬁguration.
(ii) interference cancellation is processed at 1 sample per
symbol, at optimal sampling instants.
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
h
1,1
= 1, h
1,2
= h
1,4
= h
1,5
= (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 τ
k
are independent and uniformly distributed
on [0, T]. Carrier phases ϕ
k
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 E
b
/N
0
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 singleuser reference
of 0.15 dB after 3 iterations. At ﬁrst iterations, the modulus
estimate of g
5,9
and g
6,9
(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 g
5,9
and g
6,9
shows an
error standard deviation decreasing with iterations, until it
reaches the CramerRao bound (CRB). This bound is more
precisely the phase singleuser modiﬁed CRB [9], given with
our notations by
CRB
_
Arg
_
g
k,l
__
=
1
2N
¸
¸
h
k,l
¸
¸
2
_
E
s
N
0
_
−1
_
Rd
2
_
. (12)
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 E
b
/N
0
= 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 g
5,1
, and [4.61, 4.89]
◦
for the phase error standard deviation of coeﬃcient g
5,1
.
4. EXTENSIONTOTHE 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 Kaband with
lowcost 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 DVBRCS 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
0 0.5 1 1.5 2 2.5 3 3.5 4
E
b
/N0 (dB)
10
−5
10
−4
10
−3
10
−2
10
−1
10
0
B
E
R
BER (cells 5 and 6)
No MUD
PIC 1
PIC 2
PIC 3
Reference
(a)
2 2.5 3 3.5 4
E
b
/N0 (dB)
−0.1
0
0.1
0.2
0.3
0.4
N
o
r
m
a
l
i
z
e
d
b
i
a
s
(
)
Modulus estimate of g5,9 and g6,9
PIC 1
PIC 2
PIC 3
(b)
2 2.5 3 3.5 4
E
b
/N0 (dB)
4
6
8
10
12
E
r
r
o
r
s
t
a
n
d
a
r
d
d
e
v
i
a
t
i
o
n
(
◦
)
Phase estimate of g5,9 and g6,9
PIC 1
PIC 2
PIC 3
CRB
(c)
Figure 4: Results with timeinvariant 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 DVBRCS 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. Algorithmsensitivity 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 singleuser phase recovery) is diﬃcult to
deﬁne in a multiuser context, we choose here to evaluate a
mean case. We model carrier phases ϕ
k
(t) as
ϕ
k
(t) = ϕ
k
+ Δf
k
t, (13)
for all k, where the ϕ
k
are independent and uniformly dis
tributed on [0, 2π], and the Δf
k
T follow independent zero
mean Gaussian distributions with standard deviation σ
Δf T
.
No change is performed on the algorithm, which assumes
timeinvariant 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
singleuser 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 σ
Δf T
= 10
−4
, but the
degradations with σ
Δf T
= 2 · 10
−4
are very large.
6 EURASIP Journal on Wireless Communications and Networking
0 1 1.5 1.75
Standard deviation of 10
4
·Δf ·T
0
0.5
1
1.5
D
e
g
r
a
d
a
t
i
o
n
(
d
B
)
Single user
PIC 2 cells 4 and 7
PIC 3 cells 4 and 7
PIC 2 cells 5 and 6
PIC 3 cells 5 and 6
Figure 5: Degradation with frequency oﬀsets.
32 64 128 256 462
Length of windows for estimation (symbol)
0
0.5
1
1.5
D
e
g
r
a
d
a
t
i
o
n
(
d
B
)
PIC 2 cells 4 and 7
PIC 3 cells 4 and 7
PIC 2 cells 5 and 6
PIC 3 cells 5 and 6
Figure 6: Degradation with reduced estimation windows.
By comparing the degradations in singleuser and mul
tiuser cases, we can see that they are similar for σ
Δf T
= 10
−4
and for σ
Δf T
= 0 (i.e., without frequency oﬀsets). We can
conclude that the degradation in the multiuser case with
σ
Δf T
= 10
−4
is mainly due to imperfect user phase recovery.
Beyond σ
Δf T
= 10
−4
, it can be observed that the degradation
in the multiuser case increases more quickly than the degra
dation in the singleuser case: interference cancellation eﬃ
ciency is limited. The considered algorithm is consequently
limited to about σ
Δf T
= 10
−4
for a burst length equal to 462
symbols.
4.2. Approach with reduced estimation windows for
channel estimation
In order to cope with higher frequency oﬀsets, we use in this
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
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 g
k,l
.
Channel estimation sensitivity to frequency oﬀsets de
creases when the length of estimation windows decreases, be
cause the constellation rotations on a window are reduced.
However, sensitivity to additive noise increases when the
length of estimation windows decreases, because noise is av
eraged on shorter windows. The optimal length of estimation
windows then results from a tradeoﬀ between frequency oﬀ
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
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 g
k,l
increase while the length of windows decreases, leading to a
lesseﬃcient interference cancellation and phase compensa
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
g
5,2
 = g
5,6
 = g
5,9
 = g
5,8
 = g
5,4
 = g
5,1
 = g
4,1
 =
g
4,5
 = g
4,8
). In fact, it can be seen in Figure 7 that similarly
to singleuser 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
cells 4 and 7, as the least square estimation is performed on
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 σ
Δf 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 singleuser 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 singleuser frequency estimations.
J. P. Millerioux et al. 7
2 2.5 3 3.5 4 4.5 5
E
b
/N
0
(dB)
5
10
15
20
25
30
35
40
45
P
h
a
s
e
e
r
r
o
r
s
t
a
n
d
a
r
d
d
e
v
i
a
t
i
o
n
(
◦
)
Coeﬃcients g
4,5
and g
7,6
PIC 2, 32 symbols
PIC 3, 32 symbols
BCR, 32 symbols
PIC 2, 64 symbols
PIC 3, 64 symbols
BCR, 64 symbols
PIC 2, 128 symbols
PIC 3, 128 symbols
BCR, 128 symbols
(a)
2 2.5 3 3.5 4 4.5 5
E
b
/N
0
(dB)
5
10
15
20
25
30
35
40
45
P
h
a
s
e
e
r
r
o
r
s
t
a
n
d
a
r
d
d
e
v
i
a
t
i
o
n
(
◦
)
Coeﬃcients g
5,6
and g
6,5
PIC 2, 32 symbols
PIC 3, 32 symbols
BCR, 32 symbols
PIC 2, 64 symbols
PIC 3, 64 symbols
BCR, 64 symbols
PIC 2, 128 symbols
PIC 3, 128 symbols
BCR, 128 symbols
(b)
Figure 7: Channel estimation errors for diﬀerent coeﬃcients and lengths of window.
Case
Initial PA
frequency
estimations
DD frequency
reestimations
Reduced
estimation
windows for g
k
a y n n
b y y n
c up to IT n n y
c beyond IT — y n
(a)
Windows for channel estimation
Case a
Case b
Case c
Pilot symbols
Information symbols
(b)
Figure 8: Approach with frequency estimations: (a) operations performed, (b) distributions of pilot symbols.
4.3.1. Principle
If a frequency estimate Δ
f
k
for the kth signal is available, it
can be included in the estimated kth signal: u
(m)
k
(t) ∗ s(−t)
consequently becomes ( u
(m)
k
(t) ∗s(−t)) exp( j2πΔ
f
k
t) in (7).
Since the constellation rotations on the burst for y
k
(t)∗s(−t)
and ( u
(m)
k
(t) ∗s(−t)) exp( j2πΔ
f
k
t) are potentially very close
(ideally identical if Δ
f
k
= Δf
k
), it is then possible to keep
large estimation windows to perform estimation in (7): us
ing the whole burst allows obtaining the minimum degra
dation. Clearly, this approach requires “accurate” singleuser
frequency estimations, which become the hard task.
A ﬁrst possibility is to use initial frequency estimations
before interference cancellation. In this case, the estimation
accuracy is limited due to the very lowsignaltointerference
plusnoise ratio (unless using a very high number of pilot
symbols, which decreases the spectral eﬃciency). Another
way is to use symbol decisions for frequency estimation if
it is possible to obtain suﬃciently reliable symbol decisions.
Many diﬀerent receiver architectures can be derived. Three
examples of architectures are described and evaluated in the
following sections.
4.3.2. Architectures with single user
frequency estimations
Two modes are considered for singleuser frequency esti
mation: the pilot aided mode (PA), based on pilot sym
bols, and the decision directed mode (DD), based on symbol
8 EURASIP Journal on Wireless Communications and Networking
2 2.5 3 3.5 4
E
b
/N0 (dB)
10
−5
10
−4
10
−3
10
−2
B
E
R
BER (cells 5 and 6)
No MUD
PIC 1
PIC 2
PIC 3
Reference
(a)
2 2.5 3 3.5 4
E
b
/N
0
(dB)
10
−4
E
r
r
o
r
s
t
a
n
d
a
r
d
d
e
v
i
a
t
i
o
n
(
)
Frequency estimate (cells 5 and 6)
No MUD
(b)
2 2.5 3 3.5 4
E
b
/N
0
(dB)
−0.1
0
0.1
0.2
0.3
0.4
N
o
r
m
a
l
i
z
e
d
b
i
a
s
(
)
Modulus estimates of g
5,9
and g
6,9
PIC 1
PIC 2
PIC 3
(c)
2 2.5 3 3.5 4
E
b
/N
0
(dB)
4
6
8
10
12
E
r
r
o
r
s
t
a
n
d
a
r
d
d
e
v
i
a
t
i
o
n
(
◦
)
Phase estimates of g
5,9
and g
6,9
PIC 1
PIC 2
PIC 3
CRB
(d)
Figure 9: Results with frequency estimations: σ
Δf T
= 2 · 10
−4
, case a.
decisions. For the PA mode, pilot symbols are distributed
within the burst into 3 blocks (see Figure 8(b), cases a and
b). We follow the approach of [10]. First, a mean phase
is computed on each block of pilot symbols. Then, a least
square estimation based on these mean phases is used to
estimate the frequency. For the DD mode, the principle
is the same: the burst is divided into adjacent blocks, on
which mean phases are computed using symbol decisions.
For the DD mode, frequency estimations are performed
after interference cancellation, that is, Δ
f
(m)
k
are used to
obtain g
(m+1)
k
.
The CRB considered for frequency estimation in DD
mode is the singleuser frequency modiﬁed CRB [9], given
by
CRB
_
Δf
k
T
_
=
3
2π
2
N
3
_
E
s
N
0
_
−1
. (14)
For PA frequency estimation, the CRB is diﬀerent from (14)
with N replaced by the number of pilot symbols (because
pilot symbols are not consecutive).
J. P. Millerioux et al. 9
2 2.5 3 3.5 4
E
b
/N0 (dB)
10
−5
10
−4
10
−3
10
−2
B
E
R
BER (cells 5 and 6)
No MUD
PIC 1
PIC 2
PIC 3
Reference
(a)
2 2.5 3 3.5 4
E
b
/N
0
(dB)
10
−4
E
r
r
o
r
s
t
a
n
d
a
r
d
d
e
v
i
a
t
i
o
n
(
)
Frequency estimate (cells 5 and 6)
No MUD
PIC 1
PIC 2
CRB
(b)
2 2.5 3 3.5 4
E
b
/N
0
(dB)
−0.1
0
0.1
0.2
0.3
0.4
N
o
r
m
a
l
i
z
e
d
b
i
a
s
(
)
Modulus estimates of g
5,9
and g
6,9
PIC 1
PIC 2
PIC 3
(c)
2 2.5 3 3.5 4
E
b
/N
0
(dB)
4
6
8
10
12
E
r
r
o
r
s
t
a
n
d
a
r
d
d
e
v
i
a
t
i
o
n
(
◦
)
Phase estimates of g
5,9
and g
6,9
PIC 1
PIC 2
PIC 3
CRB
(d)
Figure 10: Results with frequency estimations: σ
Δf T
= 2 · 10
−4
, case b.
The following three cases of receiver architecture are eval
uated.
Case a
PA initial frequency estimations are performed, no frequency
reestimation is performed, the estimation window for the g
k
is the whole burst.
Case b
PA initial frequency estimations are performed, frequencies
are reestimated in DD mode at each iteration, the estimation
window for the g
k
is the whole burst.
Case c
No initial frequency estimation is performed:
(i) for iterations up to IT: no frequency estimation is per
formed, the estimation window for the g
k
is 154 sym
bols for all cells (see Figure 8(b));
(ii) for iterations beyond IT: frequencies are reestimated
in DD mode, the estimation window for the g
k
is the
whole burst.
The operations performed are summarized in Figure 8(a). In
all cases, we use 32 pilot symbols. Distributions of pilot sym
bols are shown in Figure 8(b).
10 EURASIP Journal on Wireless Communications and Networking
2 2.5 3 3.5 4
E
b
/N0 (dB)
10
−5
10
−4
10
−3
10
−2
B
E
R
BER (cells 5 and 6)
No MUD
PIC 1
PIC 2
PIC 3
PIC 4
Reference
(a)
2 2.5 3 3.5 4
E
b
/N
0
(dB)
10
−4
E
r
r
o
r
s
t
a
n
d
a
r
d
d
e
v
i
a
t
i
o
n
(
)
Frequency estimate (cells 5 and 6)
PIC 1
PIC 2
CRB
(b)
2 2.5 3 3.5 4
E
b
/N
0
(dB)
−0.1
0
0.1
0.2
0.3
0.4
N
o
r
m
a
l
i
z
e
d
b
i
a
s
(
)
Modulus estimates of g
5,9
and g
6,9
PIC 3
PIC 4
(c)
2 2.5 3 3.5 4
E
b
/N
0
(dB)
4
6
8
10
12
E
r
r
o
r
s
t
a
n
d
a
r
d
d
e
v
i
a
t
i
o
n
(
◦
)
Phase estimates of g
5,9
and g
6,9
PIC 3
PIC 4
CRB
(d)
Figure 11: Results with frequency estimations: σ
Δf T
= 2 · 10
−4
, case c.
4.3.3. Results with σ
Δf T
= 2 · 10
−4
We ﬁrst consider in this section a target σ
Δf 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 g
5,9
and g
6,9
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 g
5,9
and g
6,9
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 σ
Δf T
= 2 · 10
−4
, the most eﬃcient architecture is
consequently architecture b. However, if architecture c leads
J. P. Millerioux et al. 11
2 2.5 3 3.5 4
E
b
/N0 (dB)
10
−5
10
−4
10
−3
10
−2
B
E
R
BER (cells 5 and 6)
No MUD
PIC 1
PIC 2
PIC 3
PIC 4
PIC 5
Reference
(a)
2 2.5 3 3.5 4
E
b
/N
0
(dB)
10
−4
E
r
r
o
r
s
t
a
n
d
a
r
d
d
e
v
i
a
t
i
o
n
(
)
Frequency estimate (cells 5 and 6)
PIC 3
PIC 4
CRB
(b)
2 2.5 3 3.5 4
E
b
/N
0
(dB)
−0.1
0
0.1
0.2
0.3
0.4
N
o
r
m
a
l
i
z
e
d
b
i
a
s
(
)
Modulus estimates of g
5,9
and g
6,9
PIC 4
PIC 5
(c)
2 2.5 3 3.5 4
E
b
/N
0
(dB)
4
6
8
10
12
E
r
r
o
r
s
t
a
n
d
a
r
d
d
e
v
i
a
t
i
o
n
(
◦
)
Phase estimates of g
5,9
and g
6,9
PIC 4
PIC 5
CRB
(d)
Figure 12: Results with frequency estimations: σ
Δf T
= 5 · 10
−4
, case c.
to a slower convergence of the algorithm, a signiﬁcant advan
tage is that it appears more suited to highfrequency oﬀsets,
as we will see in the following section.
4.3.4. Results with σ
Δf T
= 5 · 10
−4
We now consider a target σ
Δf 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 g
k
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
E
b
/N
0
equal to 3.2 dB, the block processing approach allows
obtaining a BERequal 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 BERalong
the paper. Actually, this average BER can hide some complete
12 EURASIP Journal on Wireless Communications and Networking
1 106 212
Number of erroneous bits per packet
0
0.5
1
1.5
2
2.5
×10
−3
P
r
o
b
a
b
i
l
i
t
y
Distribution of erroneous bits
Cell 5 at 4σ
Δf T
, E
b
/N0 = 3.5 dB
424 information bits
Figure 13: Distribution of erroneous bits: σ
Δf T
= 5 · 10
−4
, itera
tion 5.
failures in convergence of the algorithmon 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 highfrequency 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 σ
Δf T
) except frequency
oﬀset for cell 5, which is deterministic and equal to 4σ
Δf T
=
2 · 10
−3
. The estimated distribution of the number of erro
neous bits per burst for cell 5 at iteration 5 for E
b
/N
0
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 3σ
Δf 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.
REFERENCES
[1] J. P. Millerioux, M. L. Boucheret, C. Bazile, and A. Ducasse,
“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.
[2] Digital Video Broadcasting (DVB), “Interaction channel for
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. M¨ uller, R. Rinaldo, and A. Ver
nucci, “Interference mitigation for the reverselink 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 d´ etection multiutilisateurs
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: beamforming and directionof
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 CramerRao 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, “MFTDMA
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 DVBRCS Next Generation
Satellite Networks
Thierry Gayraud
1, 2
and Pascal Berthou
1, 2
1
Laboratoire d’Analyse et d’Architecture des Syst`emes (LAASCNRS), 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 (DVBRCS) 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 lowinfrastructure areas. However, in order to take in account realtime 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 DVBRCS 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 nonQoS 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 QoSaware. 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
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 lowinfrastructure 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 (DVBRCS) 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 lowcost satellite equipment through the
harmonization of RCS terminals (ST) based on this open
standard. Several commercial DVBRCS based networks are
already deployed and many eﬀorts are done in order to en
hance interoperability.
Most recent commercial deployments provide either In
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
qualityofservice (QoS) ﬁeld and the next step is, obviously,
to take beneﬁts from DVBRCS 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 DVBRCS 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. NEWTRENDS INSATELLITE SYSTEMS
2.1. Forward link
The ﬁrst DVB norm described a transmission scheme based
on MPEG2 (Motion Picture Expert Group) video compres
sion and transmission schemes, using MPEGTS (MPEG
transport stream). This latter was adapted for satellite sys
tems through DVBS (DVB transmission via satellite) that
deﬁnes series of options to send MPEGTS 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. DVBS2 standard [2] is intended to be a suc
cessor of DVBS with the same applications (TV, Internet,
etc.). It oﬀers new coding techniques that can increase per
formance by 25% over that of DVBS, 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 MPEG2TS 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 DVBRCS [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 DVBRCS frames
into DVBS 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 DVBRCS signals into a single DVBS one. The
associated topology could be a “star” or a “meshed”
network. The endtoend delay decreases to only one
single hop.
Furthermore, DVBRCS 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 (MFTDMA). It basically relies on the avail
ability of several TDMAchannels (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 MPEG2TS 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 ondemand 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 NCCwhich 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) ratebased dynamic capacity (RBDC), which is dy
namic rate capacity (in slots/frame), upperbounded
by MaxRBDC, granted in response to dynamic re
quests from the STs to track their instantaneous traﬃc
rate;
(iii) (absolute) volumebased 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
cation algorithms implemented in the NCC to fulﬁll the ser
vices requirements are not speciﬁed.
3. QoS ARCHITECTURE
This section describes the QoS architecture we propose for
DVBS/RCS satellite systems. The main contribution is built
on return link management. Thus the downlink is generally
considered not to be a bottleneck and classical traﬃc engi
neering techniques are enough to managed the network.
3.1. Basis of QoS in satellite systems
To reach an optimal exploitation of uplink resources, at least
three functions must be implemented to provide QoS guar
antees.
(i) QoS admission control consists, before the application
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
among high priority traﬃc.
(ii) QoS enforcement consists in checking that the admit
ted traﬃc respects its contracts, that is, that it does
not use more resources than requested. This is done
by policing and shaping.
(iii) QoS diﬀerentiation consists in having several classes of
traﬃc, each class provides diﬀerent behavior adapted
to a given service. This task is complex and needs dy
namic management during the connections lifetime
and must be performed at two layers: the DVBRCS
and IP layers.
Thanks to the 5 bandwidth allocation mechanisms in
cluded in DVBRCS standard, the traﬃc diﬀerentiation is
made easily in introducing several MAC queues in the ST
stack and mapping the capacity requests over the MAC
queues. Then, IP DiﬀServbased router architecture can be
setup over these new MAC services. However, this cannot be
done without crosslayer mechanisms that ensure perfect re
sources use.
3.2. Crosslayer architecture
The QoS management, in the proposed architecture, is split
into three levels detailed in the following paragraphs.
(i) Satellite terminal resources: is medium access control
level, where the DVBRCS DAMA allocates the band
width on a ﬁxed basis for real time applications and on
demand for other ﬂows (nonrealtime 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 Agent
Application
QoS signalling
QoS Server
MFclassiﬁer IP classes
IP downstream
from user terminal
Traﬃc shaping/policing
S
c
h
e
d
u
l
i
n
g
EF AF BE
EDF EDF
PQ
/
EF AF + BE
Segmentation
I
P
L
a
y
e
r
Transmission
allowed/denied
IP DVB/RCS
interface
M
A
C
l
a
y
e
r
Threshold
RT DVB
frames
NRT DVB
frames
Framing
DVBRCS frames
To satellite
DAMA client
DAMA server
TBTP
Capacity
requests
To/from NCC
Figure 1: QoS architecture.
The user is able to classify its own ﬂows in any available
service through a dedicated agent (the QoS agent) that
communicates with the QoS server to deliver the clas
siﬁcation. The goal is to exploit the capabilities oﬀered
by the IP QoS capabilities.
An overview of this QoS architecture within the ST is given
in Figure 1.
3.3. QoS at DVBRCS layer
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), DVBRT for “realtime service” and DVB
NRT for “nonrealtime” service, which are associated to
2 diﬀerent ATM permanent virtual channels (PVC). One
(DVBRT) beneﬁts from static resource assignment through
4 EURASIP Journal on Wireless Communications and Networking
CRA; on the contrary, (DVBNRT) relies on a dynamic re
source allocation scheme also called BoD algorithm which
will be further detailed in this section.
(i) Realtime queue: the CRA consists in a ﬁxed capacity
that is set at the ST logon 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 DVBRT traﬃc, since its high delay sensitive re
quirements hardly tolerates throughput ﬂuctuations.
(ii) Nonrealtime queue: the request category retained for
DVBNRT 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 endtoend delay, application needs
are anticipated in monitoring the DVBNRT 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) realtime: 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) nonrealtime: such an IP ﬂow should be guaranteed a
minimum bandwidth, a mean queuing delay of a few
hundreds of milliseconds;
(iii) besteﬀort: 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 timevarying capacity.
With reference to IntServ/DiﬀServ traﬃc classes, the
besteﬀort (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. Realtime IP data category in
cludes both IntServ guaranteed service class and DiﬀServ ex
pedited forwarding (EF) PHB (perhop behavior) while the
nonrealtime 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 (RCEDF, 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 RCEDF components are gathered under
the appellation “traﬃc shaping/policing.” The traﬃc polic
ing and shaping are then realized, thanks to singlerate 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 realtime 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
lowdata 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
QoS
segregation
Satellite
ST
MAC
Hub
NCC
Router
IP
IPv4 , IPv6
ST: satellite terminal
MAC: medium access control
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
cations that open unﬁxed port as, for instance, VoIP applica
tions.
A more generic and “useroriented” approach has been
proposed. The ST could be customized on the user request.
Dedicated software, the QoS agent [7], allows to associate
a running application to one of the three deﬁned services
and to send this association to the ST. It dynamically mon
itors application connections and sends to the ST the 5 tu
ples source IP address, destination IP address, source port,
destination port, type of protocol for each of them. The
ST maintains an association’s table. Each incoming packet is
redirected within the ST to the appropriate queue according
to this association’s table.
Figure 3 shows the six connections opened by Gnomeet
ing, videoconference software, and their selected services.
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. EXPERIMENTAL MEASUREMENTS
4.1. Emulation principle
Evaluating performances over real data links or networks
is expensive, even impossible for systems in development
phase.
Figure 3: QoS agent user interface (GUI).
Simulation and emulation both provide the opportunity
to evaluate performances, at lowcost, on more or less realis
tic systems. When simulation needs a complete modeling of
the systems from applications to physical network and oper
ates in virtual time, emulation is more demonstrative since
real applications can be deployed over the model describ
ing transfer characteristics, delay, and error behaviors for in
stance.
For these reasons, the choice was made to set up a satellite
emulation platformto demonstrate the network and applica
tion services integration on next generation satellite systems
and the possibility to interoperate with terrestrial networks.
4.2. Test bed
The network elements that belong to a classical satellite net
work (Satellite, NCC, STs) are emulated individually on a
dedicated computer. A gigabit Ethernet interconnects them
and emulates the satellite carrier emulation. Ethernet was
chosen for its native broadcast abilities and also for its high
bandwidth capacities. Each satellite channel is mapped on a
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); biterror 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 twoway satellite link in realtime. 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 DVBS/DVBRCS 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. Platformcalibration
As we explained in the former section, the available platform
will provide us with a prototype as close as possible to a real
DVBS/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 DVBRCS 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.
Physical
layer
ST information peak rate 2048 Kbps
Superframe 10 Trames
Frame duration 50 ms
Global DVBRCS resource 2048 Kbps
Basic
DAMA
FCA None
α 1
SLA ST
CRA 96 Kbps
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 singlerate 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 5 s
2 1.5 1 0.5 0
Delay (s)
0
10
20
30
40
50
60
70
80
90
100
C
D
F
(
%
)
α = 0
α = 0.5
α = 1
Figure 4: DAMA impact (VBR ﬂow without FCA).
factor α which enables to bring 50% of the packets delay un
der 1 seconds (Figures 4 and 5).
Under 0.5, the anticipation factor does not improve the
endtoend delay. If we consider this factor, the reduction of
α leads to a capacity underutilization, it will be maintained
to 0.5 which stands for an interesting compromise between
queuing delays and uplink utilization eﬃciency (Figure 6).
If FCA allocation is taken into account, an additional im
provement can be noticed. The factor α seems to have less
impact on the endtoend delay than in the ﬁrst scenario.
However, this 40 Kbps allocation stands for more than 10%
of the DIVX average throughput and might be considered
as overestimated for a single ST. Therefore, the anticipation
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
1.8 1.6 1.4 1.2 1 0.8 0.6 0.4 0.2 0
Delay (s)
0
10
20
30
40
50
60
70
80
90
100
C
D
F
(
%
)
α = 0
α = 0.5
α = 1
Figure 5: DAMA impact on a VBR ﬂow with FCA.
100 97.5 95 92.5 90 87.5 85 82.5 80 77.5 75
Uplink utilization eﬃciency (%)
300
400
500
600
700
800
Q
u
e
u
i
n
g
d
e
l
a
y
(
m
s
)
α = 1
α = 2/3
α = 1/2
α = 1/3
α = 0
Figure 6: Delay versus eﬃciency.
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 DVBNRT 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
120000
100000 80000 60000 40000 20000 0
Time (ms)
400
600
800
1000
1200
1400
1600
1800
2000
T
h
r
o
u
g
h
p
u
t
(
k
b
i
t
s
/
s
)
1 2 3 4
Throughput (kbits/s)
Figure 7: Dynamic change of class of service using the QoS agent.
Table 3: SATIP6 QoS performance under diﬀerent loading condi
tions.
Network
load [%]
Average delay (Jitter) [ms] Losses [%]
BE FTP Voice BE FTP Voice
25 293 (26) 293 (26) 283 (23) 0 0 0
50 291 (23) 291 (24) 283 (23) 0 0 0
75 290 (24) 289 (24) 283 (23) 0 0 0
100 919 (23) 948 (24) 283 (23) 0 0 0
125 6753 (23) 1783 (24) 283 (23) 33 0 0
150 6755 (28) 1783 (24) 283 (23) 37 0 0
4.4.3. Dynamic change of QoS
The change of multimedia stream QoS is done thanks to QoS
agent.
The diﬀerent following steps are easy to ﬁnd in Figure 7.
(i) Step 1: the scenario begins as a UDP video stream
starts. This stream is sent in the BE class of service.
(ii) Step 2: 25 seconds later, another ﬂow is sent on the
same uplink; it is done so that the uplink is now con
gested. The throughput of the ﬁrst stream is then re
duced to 800 kbps.
(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
of 1 Mbps, and the traﬃc buﬀered before the resource
reservation, the throughput is around 1 Mbps.
(iv) Step 4: at t = 80 s, the streamis downgraded back to BE
and the stream throughput is then around 800 kbps.
In Figure 7, the time needed to change from one Cos to an
other one for a ﬂow could also be evaluated when the link is
congested by other data ﬂows.
The delays given in Table 4 show as usually that the delay
is around 3 seconds. It is less when a new ﬂow is added on a
link. It is longer when the QoS of a ﬂow is upgraded and less
Table 4: Dynamic QoS change delay.
Transition 1 →2 2 →3 3 →4
Initial
situation
One ﬂow
without QoS
Two ﬂows
without QoS
Two ﬂows,
one with QoS
Final
situation
Two ﬂows
without QoS
Two ﬂows,
one with QoS
Two ﬂows
without QoS
Delay (s) 2.8 3.5 3.15
when it is downgraded, still longer in these two cases than in
the simple addition of a ﬂow.
These results conclude the section dedicated to experi
mental measurements by putting the stress on the proper
traﬃc diﬀerentiation carried out by our QoS architecture and
the ability to change fromone class of service to the other one
thanks to the QoS agent GUI.
5. CONCLUSION
It was proved in this paper that it was possible to specify and
implement QoS architecture for DVBS/RCS satellite system
in order to provide the user with QoS guarantees even if a
nonQoSaware application is used. MAC algorithms (such
as DAMA) were proved to be eﬃcient. It was also explained
how to proceed to evaluate such an architecture. Using mea
surement tools on a wellcalibrated testbed, the global QoS
of the satellite system may be evaluated accurately.
Using other capacity category than VBDC (RBDC for in
stance) is improving resource utilization especially if appli
cations throughputs are known. Unfortunately, this is not
usual today in the Internet. In that case, we propose to use
ratebased signaling protocols (SIP) in order to set up the
right capacity requests.
Other future work may also be done related to DVBS2
and new admission control mechanisms.
ACKNOWLEDGMENTS
The authors wish to thank all the partners of the SATIP6
[8] consortium: Alcatel Space (France), which is the project
coordinator, Telecom Italia Lab (Italy), France Telecom SA
(France), University of Rome “La Sapienza” (Italy), Sintef
(Norway), LAASCNRS (France), and Alliance Qualit´ e Logi
ciel (France).
REFERENCES
[1] ETSI TR 102 157, “Satellite Earth Stations and Systems (SES);
Broadband Satellite Multimedia; IP Internetworking over satel
lite; Performance, Availability and Quality of Service,” July
2003.
[2] ETSI Standard TR 102 376 V1.1.1, “Digital Video Broadcasting
(DVB); User guidelines for the second generation system for
Broadcasting, Interactive Services, News Gathering and other
broadband satellite applications (DVBS2)”.
[3] ETSI EN 301 790 V1.3.1, “Digital Video Broadcasting (DVB);
Interaction channel for Satellite Distribution Systems,” March
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 DVBRCS 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 IST200134344).
Hindawi Publishing Corporation
EURASIP Journal on Wireless Communications and Networking
Volume 2007, Article ID 65058, 8 pages
doi:10.1155/2007/65058
Research Article
MaximumLikelihood Timing and Carrier Synchronization in
BurstMode 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 VanelliCoralli
This paper investigates the joint maximum likelihood (ML) estimation of the carrier frequency oﬀset, timing error, and carrier
phase in burstmode 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
closedform expressions. The performance of the proposed scheme is investigated by computer simulation and compared with
CramerRao bounds. It turns out that the estimation accuracy is very close to the theoretical limits up to relatively low signalto
noise ratios. This makes the algorithm well suited for turbocoded 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
Burst transmission of digital data and voice is widely adopted
in satellite timedivision multipleaccess (TDMA) networks.
In these applications the propagation mediumcan 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
hangup phenomena [3–5]. The latter have shorter acqui
sitions and, accordingly, are better suited for burstmode
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 dataaided (DA) operation and provides better estima
tion accuracy as compared to a nondataaided (NDA) ap
proach. Even so, however, synchronization may prove diﬃ
cult, especially with turbocoded modulations operating at
relatively low signaltonoise ratios (SNRs). Clearly, very ef
ﬁcient synchronization algorithms are needed in these con
ditions [6].
The common approach to solve the synchronization
problem in burstmode transmissions is to estimate the tim
ing error ﬁrst, and then use the timesynchronized 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 orderbased 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 timesynchronized 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 timedomain 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 CramerRao 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 burstmode
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 threestep procedure. In
the ﬁrst step frequency recovery is accomplished through a
monodimensional grid search. The estimated CFO is then
exploited in the second step to obtain a closedform 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 turbocoded 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. SYSTEMMODEL ANDPROBLEMFORMULATION
2.1. Statement of the problem
We consider the reverse link of a satellite communication
system and assume a timedivision multipleaccess (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 timeofarrival (ToA)
of a generic burst. This is normally performed through a sim
ple noncoherent energydetection 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 realvalued 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 UWwithin 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 fdt+ϕ)
s(t −εT) + w(t), (2)
where s(t) is the useful signal, ϕ and f
d
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
interval
Sync.
preamble
UW
Header Payload
Figure 1: Burst structure.
Sync.
preamble
UW Payload
τ
NT
Figure 2: Start of burst estimation error.
twosided power spectral density N
0
. Signal s(t) is expressed
as
s(t) =
_
n
a
n
g(t −nT), (3)
where {a
n
} are modulation symbols taken from a PSK or
QAM constellation and g(t) has a rootraisedcosine Fourier
transform with some rolloﬀ α. To facilitate the timing esti
mation process, during the preamble we assume a pilot pat
tern composed by alternating BPSK symbols +1 and −1 [17].
Accordingly, s(t) is given by
s(t) =
¸
2E
s
T
cos
_
πt
T
_
(4)
with E
s
denoting the signal energy per symbol interval, and
r(t) may be rewritten in the form
r(t) =
¸
2E
s
T
e
j(2π fdt+ϕ)
cos
_
π(t −εT)
T
_
+ w(t)
for t ∈ [0, NT].
(5)
In order to produce a discretetime signal, the received
waveform is fed to an antialiasing ﬁlter (AAF) and sampled
at some rate f
c
. The ﬁlter bandwidth B
AAF
and the sampling
rate are chosen such that the signal component is passed
undistorted (even for the maximum frequency oﬀset) and no
aliasing occurs. Assuming that the CFO is less in magnitude
than 0.5/T, from (5) it follows that we can set B
AAF
= 1/T
and f
c
= 2/T. For simplicity, in the ensuing discussion the
AAF is assumed with a brickwall transfer function, even
though the rectangular shape is not strictly necessary and
could be easily made more realistic [18].
For normalization purposes, the output of the AAF is
scaled by a factor
_
T/2E
s
and we call x(k) the correspond
ing samples taken at t = kT/2, with 0 ≤ k ≤ 2N − 1. As the
signal is not distorted in passing through the ﬁlter, we have
x(k)=e
j(πkν+ϕ)
cos
__
k
2
−ε
_
π
_
+ n(k) for 0 ≤ k ≤ 2N −1,
(6)
where ν = f
d
T is the CFOnormalized to the symbolrate 1/T
and n(k) = n
R
(k) + jn
I
(k) is the noise contribution. Due to
the previous hypotheses, {n
R
(k)} and {n
I
(k)} are indepen
dent and white random sequences with the same variance
σ
2
= (E
s
/N
0
)
−1
. As the signal component in (6) depends on
ν, ε, and ϕ, we may estimate all these parameters from the
observation of {x(k)}. This problem is addressed in the next
section by resorting to ML methods.
3. MAXIMUMLIKELIHOODESTIMATIONOF
THE SYNCHRONIZATIONPARAMETERS
3.1. Maximization of the likelihood function
Bearing in mind (6), the loglikelihood function for the un
known parameters is given by
Λ(`ν, ` ε, ` ϕ) = −2N ln
_
2πσ
2
_
−
1
2σ
2
2N−1
_
k=0
¸
¸
¸
¸
x(k) −e
j(πk`ν+` ϕ)
cos
__
k
2
− ` ε
_
π
_¸
¸
¸
¸
2
,
(7)
where `ν, ` ε, and ` ϕ are trial values of ν, ε, and ϕ, respec
tively. The joint ML estimate of (ν, ε, ϕ) is the location where
Λ(`ν, ` ε, ` ϕ) achieves its global maximum. Skipping irrelevant
factors and additive terms independent of (`ν, ` ε, ` ϕ), it turns
out that Λ(`ν, ` ε, ` ϕ) may equivalently be replaced by
Ψ(`ν, ` ε, ` ϕ) = e
_
e
−j ` ϕ
2N−1
_
k=0
x(k)e
−jπk`ν
cos
__
k
2
− ` ε
_
π
_
_
,
(8)
where e{·} denotes the real part of the enclosed quantity.
Function Ψ(`ν, ` ε, ` ϕ) can also be rewritten as
Ψ(`ν, ` ε, ` ϕ) = e
_
e
−j ` ϕ
_
Y
e
(`ν) cos(π` ε) + e
−jπ`ν
Y
o
(`ν) sin(π` ε)
__
(9)
4 EURASIP Journal on Wireless Communications and Networking
with
Y
e
(`ν) =
N−1
_
k=0
(−1)
k
x(2k)e
−j2πk`ν
,
Y
o
(`ν) =
N−1
_
k=0
(−1)
k
x(2k + 1)e
−j2πk`ν
.
(10)
To ease the search for the maximum of Ψ(`ν, ` ε, ` ϕ), we rewrite
(9) in the form
Ψ(`ν, ` ε, ` ϕ) =
¸
¸
Z(`ν, ` ε)
¸
¸
cos
_
ψ(`ν, ` ε) − ` ϕ
_
, (11)
where Z(`ν, ` ε) is a function of `ν and ` ε deﬁned as
Z(`ν, ` ε) = Y
e
(`ν) cos(π` ε) + e
−jπ`ν
Y
o
(`ν) sin(π` ε) (12)
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
achieved when the cosine factor in (11) is equal to unity,
which occurs for
ϕ(`ν, ` ε) = arg
_
Z(`ν, ` ε)
_
. (13)
In this case the righthand side of (11) reduces to Z(`ν, ` ε)
and the ML estimates of ν and ε are found by maximizing the
following function:
Γ(`ν, ` ε) = 2
¸
¸
Z(`ν, ` ε)
¸
¸
2
(14)
with respect to `ν and ` ε, where the factor 2 in the righthand
side of (14) has only been inserted to avoid a factor 1/2 in the
subsequent equations.
To proceed further, we substitute (12) into (14) and ob
tain
Γ(`ν, ` ε) =
¸
¸
Y
e
(`ν)
¸
¸
2
+
¸
¸
Y
o
(`ν)
¸
¸
2
+ e
_
e
−j2π` ε
A(`ν)
_
, (15)
where A(`ν) is deﬁned as
A(`ν) =
¸
¸
Y
e
(`ν)
¸
¸
2
−
¸
¸
Y
o
(`ν)
¸
¸
2
+ 2je
_
e
jπ`ν
Y
e
(`ν)Y
∗
o
(`ν)
_
.
(16)
Observing that A(`ν) = Y
2
e
(`ν) + e
−j2π`ν
Y
2
o
(`ν), we may
rewrite (15) into the equivalent form
Γ(`ν, ` ε) =
¸
¸
Y
e
(`ν)
¸
¸
2
+
¸
¸
Y
o
(`ν)
¸
¸
2
+
¸
¸
Y
2
e
(`ν) + e
−j2π`ν
Y
2
o
(`ν)
¸
¸
cos
_
θ(`ν) −2π` ε
_
(17)
with θ(`ν) = arg{A(`ν)}. For a given `ν, the maximumof Γ(`ν, ` ε)
is achieved by setting
ε(`ν) =
1
2π
arg
_
A(`ν)
_
. (18)
Substituting this result into the righthand side of (17) yields
P(`ν) =
¸
¸
Y
e
(`ν)
¸
¸
2
+
¸
¸
Y
o
(`ν)
¸
¸
2
+
¸
¸
Y
2
e
(`ν) + e
−j2π`ν
Y
2
o
(`ν)
¸
¸
(19)
0.5 0.4 0.3 0.2 0.1 0 −0.1 −0.2 −0.3 −0.4 −0.5
ν
0
0.2
0.4
0.6
0.8
1
1.2
P
(
ν
)
E
s
/N
0
= 10 dB
N = 64
Figure 3: Typical shape of P(ν).
from which it follows that the ML estimate of the frequency
oﬀset is given by
ν = arg max
`ν
_
P(`ν)
_
(20)
while the timing estimate is obtained from (18) in the form
ε =
1
2π
arg
_
A(ν)
_
. (21)
In case of coherent detection, an estimate of the carrier phase
ϕ is also necessary. This is computed as indicated in (13) after
replacing (`ν, ` ε) by (ν, ε) and reads
ϕ = arg
_
Y
e
(ν) cos(π ε) + e
−jπν
Y
o
(ν) sin(π ε)
_
. (22)
In the sequel the algorithm based on (20)–(22) is called
the ML estimator (MLE).
3.2. Remarks
(1) Contrarily to what one might fear, the maximization of
the likelihood function Λ(`ν, ` ε, ` ϕ) needs not be made on a
threedimensional domain. Actually, the location (ν, ε, ϕ) of
the maximum can be found through simple steps, each in
volving a single synchronization parameter. In particular, the
ﬁrst step requires maximizing the function P(`ν) deﬁned in
(19) in order to get the CFO estimate ν. As discussed later,
this can be done through a grid search over the interval where
ν is expected to lie. Once ν has been obtained, timing and
phase estimates are computed in closed form as indicated
in (21) and (22), respectively. In summary, the diﬃcult and
timeconsuming part in the estimation of (ν, ε, ϕ) is the one
that locates the maximum of P(`ν). Once this is done, the
computation of ε and ϕ becomes a trivial task.
(2) Maximizing function P(`ν) may pose some diﬃculty
due to the presence of many local maxima. This is clearly ev
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, E
s
/N
0
= 10 dB,
and ν = 0.1. As discussed in [10], the global maximum can
be sought in two steps. The ﬁrst one (coarse search) calculates
P(`ν) for a set of `ν values, say {`ν
n
}, covering the uncertainty
range of ν and determines the location `ν
M
of the maximum
over this set. The second step (ﬁne search) makes an inter
polation between the samples P(`ν
n
) and computes the local
maximum nearest to `ν
M
. It should be noted that the shape
of P(`ν) is occasionally so badly distorted by noise that its
highest peak may be far from the true ν. When this happens,
the MLE makes large errors (outliers) and the system perfor
mance is highly degraded. The SNR below which the outliers
start to occur is referred to as the threshold of the estimator.
(3) In practice the coarse search can be eﬃciently per
formed using fast Fourier transform (FFT) techniques, as it
is now explained. Starting from the observed samples {x(k)},
we ﬁrst compute the following zeropadded sequences of
length KN:
y
e
(k) =
⎧
⎨
⎩
(−1)
k
x(2k), 0 ≤ k ≤ N −1,
0, N ≤ k ≤ NK −1,
y
o
(k) =
⎧
⎨
⎩
(−1)
k
x(2k + 1), 0 ≤ k ≤ N −1,
0, N ≤ k ≤ KN −1,
(23)
where K is a design parameter called pruning factor. Next, the
FFTs of {y
e
(k)} and {y
o
(k)} are evaluated at the points
`ν
n
=
n
KN
, −
KN
2
≤ n ≤
KN
2
. (24)
This produces the quantities {Y
e
(`ν
n
)} and {Y
o
(`ν
n
)}, which
are next exploited to get {P(`ν
n
)} as indicated in (19). Finally,
the largest P(`ν
n
) is sought and this provides the coarse fre
quency estimate.
(4) Collecting (10) and (19), it is seen that P(`ν) is peri
odic of unit period. This means that MLE gives unambigu
ous frequency estimates as long as ν is conﬁned within the
interval [−1/2, 1/2). This is the frequency estimation range of
MLE.
(5) Compared to the R&B algorithm [10], the MLE is
more complex to implement as it requires the computation
of two FFTs instead of a single FFT. In addition to carrier syn
chronization, however, the MLE also provides timing recov
ery. Actually, from (16) and (21) we see that computing the
timing estimate only requires knowledge of Y
e
(ν) and Y
o
(ν).
Since these quantities can easily be obtained by {Y
e
(`ν
n
)} and
{Y
o
(`ν
n
)} through interpolation, the timing estimation task is
accomplished with a relatively low computational eﬀort once
ν is available. However, it should be observed that the com
plexity associated to the synchronization process is negligible
compared to that of iterative data decoding [19]. So, the re
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
of N and E
s
/N
0
. For this reason, it is of interest to derive
the CRB for the joint estimation of the set of parameters
η = (ν, ε, ϕ) based on the model (6).
We begin by computing the entries of the Fisher infor
mation matrix F. They are deﬁned as [20]
[F]
i, j
= −E
_
∂
2
Λ(η)
∂η
i
∂η
j
_
, 1 ≤ i, j ≤ 3, (25)
where Λ(η) is the loglikelihood function in (7) and η
de
notes the th entry of η. Substituting (7) into (25), after some
manipulations we obtain
F =
N
6σ
2
⎡
⎢
⎢
⎣
π
2
(2N−1)
_
4N−1−3 cos(2πε)
_
0 3π
_
2N−1−cos(2πε)
_
0 6π
2
0
3π
_
2N−1−cos(2πε)
_
0 6
⎤
⎥
⎥
⎦
.
(26)
The CRB for the estimation of η
is given by [F
−1
]
,
. Skip
ping the details, it is found that
CRB(ν) =
12
_
E
s
/N
0
_
−1
π
2
N
_
4N
2
−4 + 3 sin
2
(2πε)
_, (27)
CRB(ε) =
_
E
s
/N
0
_
−1
π
2
N
, (28)
CRB(ϕ) = 2
(2N −1)(4N −1 −3 cos ε)
_
E
s
/N
0
_
−1
N
_
4N
2
−4 + 3 sin
2
(2πε)
_ . (29)
Interestingly, for large data records we can approximate (27)
as
CRB(ν) ≈
3(E
s
/N
0
)
−1
π
2
N
3
(30)
which represents the CRB for the estimation of the frequency
of a complex sinusoid embedded in AWGN [10].
5. SIMULATIONRESULTS
In this section we report on simulation results illustrating the
performance of MLE over an AWGN channel. Unless other
wise speciﬁed, the synchronization parameters vary at each
new simulation run and are modeled as statistically indepen
dent random variables with a uniform distribution. In par
ticular, ν and ε are conﬁned within [−0.5, 0.5) while ϕ takes
values in the interval [−π, π). A pruning factor K = 4 is
used to compute the quantities {Y
e
(`ν
n
)} and {Y
o
(`ν
n
)}. Also,
a parabolic interpolation is chosen in the implementation of
the ﬁne search. This yields a frequency estimate in the form
ν = `ν
M
+
δ
ν
2
·
P
_
`ν
M−1
_
−P
_
`ν
M+1
_
P
_
`ν
M−1
_
−2P(`ν
M
) + P
_
`ν
M+1
_, (31)
where δ
ν
= 1/KN is the distance between two adjacent sam
ples P(`ν
n
) while `ν
M
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 0.4 0.3 0.2 0.1 0 −0.1 −0.2 −0.3 −0.4 −0.5
ν
−0.5
−0.4
−0.3
−0.2
−0.1
0
0.1
0.2
0.3
0.4
0.5
E
{
ν
}
E{ν} = ν
E
s
/N
0
= 10 dB
N = 64
Figure 4: Mean normalized frequency estimates versus ν.
0.5 0.4 0.3 0.2 0.1 0 −0.1 −0.2 −0.3 −0.4 −0.5
ε
−0.5
−0.4
−0.3
−0.2
−0.1
0
0.1
0.2
0.3
0.4
0.5
E
{
ε
}
E{ ε} = ε
E
s
/N
0
= 10 dB
N = 64
Figure 5: Mean timing phase estimates versus ε.
algorithm [8] and carrier synchronization is next achieved
by applying the R&B method [10] to the timesynchronized
samples. The O&M operates with four samples per symbol
period.
Figures 4 and 5 illustrate average frequency and timing
estimates, E{ν} and E{ ε}, provided by MLE as a function of
ν and ε, respectively. The observation length is N = 64 while
the SNR is ﬁxed to E
s
/N
0
= 10 dB. The ideal lines E{ν} = ν
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
and O&M. The observation length is N = 32 in Figure 6
and N = 64 in Figure 7. Marks indicate simulation results
while the thin solid lines are drawn to ease the reading of the
graphs. The corresponding CRBs are also shown as bench
20 15 10 5 0
E
s
/N
0
(dB)
10
−5
10
−4
10
−3
10
−2
10
−1
M
S
E
ε
O&M
MLE
CRB
N = 32
Figure 6: MSE performance of MLE and O&M estimator with N =
32.
20 15 10 5 0
E
s
/N
0
(dB)
10
−5
10
−4
10
−3
10
−2
10
−1
M
S
E
ε
O&M
MLE
CRB
N = 64
Figure 7: MSE performance of MLE and O&M estimator with N =
64.
marks. We see that MLE has the best accuracy, especially at
low SNRs where a signiﬁcant gain is observed with respect
to O&M. In particular, for N = 64 the MLE is close to the
CRB down to E
s
/N
0
values of 0 dB, while O&M approaches
the bound only for E
s
/N
0
> 10 dB. This feature of the MLE
M. Morelli and A. A. D’Amico 7
20 15 10 5 0
E
s
/N
0
(dB)
10
−8
10
−7
10
−6
10
−5
10
−4
10
−3
10
−2
M
S
E
ν
R&B
MLE
CRB
N = 64
N = 32
Figure 8: MSE performance of MLE and R&B estimator with N =
32 and N = 64.
is of great importance as it makes this estimator suitable for
turbocoded 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 E
s
/N
0
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 E
s
/N
0
. 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.
20 15 10 5 0
E
s
/N
0
(dB)
10
−3
10
−2
10
−1
10
0
2
4
6
8
2
4
6
8
2
4
6
8
M
S
E
ϕ
(
r
a
d
2
)
R&B
MLE
CRB
N = 64
N = 32
Figure 9: Mean square error MSE
ϕ
versus E
s
/N
0
with N = 32 and
N = 64.
6. CONCLUSIONS
We have addressed the joint ML estimation of the carrier fre
quency, timing error, and carrier phase in burstmode 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 timeconsuming 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 timesynchronized samples. Computer simula
tions indicate that the proposed ML algorithmprovides 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
tional scheme is preferable as it achieves similar performance
with reduced complexity.
REFERENCES
[1] U. Mengali and A. N. D’Andrea, Synchronization Techniques
for Digital Receivers, Plenum Press, New York, NY, USA, 1997.
[2] H. Meyr, M. Moeneclaey, and S. Fechtel, Digital Communica
tion Receivers: Synchronization, Channel Estimation, and Signal
Processing, John Wiley & Sons, New York, NY, USA, 1997.
[3] K. H. Mueller and M. M¨ uller, “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 timingerror 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
nondataaided carrier and clock recovery for satellite DVB at
very low signaltonoise 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 MLoriented 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. BarNess, “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, “Singletone parameter estima
tion from discretetime 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 alldigital modems for burstmode transmissions,” IEEE
Transactions on Communications, vol. 43, no. 234, pp. 1169–
1178, 1995.
[15] U. Mengali and M. Morelli, “Dataaided 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.
[17] Y. Jiang, F.W. Sun, and J. S. Baras, “On the performance limits
of dataaided synchronization,” IEEE Transactions on Informa
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,
pp. 3208–3214, 1994.
[19] S. Benedetto, R. Garello, G. Montorsi, et al., “MHOMS: high
speed ACM modem for satellite applications,” IEEE Wireless
Communications, vol. 12, no. 2, pp. 66–77, 2005.
[20] S. M. Kay, Fundamentals of Statistical Signal Processing: Estima
tion Theory, PrenticeHall, Englewood Cliﬀs, NJ, USA, 1993.
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 OptimumJoint
Estimation of DopplerShift and DopplerRate
in Packet Satellite Communications
Luca Giugno,
1
Francesca Zanier,
2
and Marco Luise
2
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 enhancedaccuracy estimation
of Dopplershift and Dopplerrate of the carrier of the received signal, due to relative motion between the transmitter and the
receiver. Two novel burst formats that minimize the Dopplershift and the Dopplerrate Cram´ erRao bounds (CRBs) for the joint
estimation of carrier phase/Dopplershift and of the Dopplerrate are derived, and a dataaided (DA) estimation algorithmsuitable
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 timedivision multiple access (TDMA) and terrestrial
mobile cellular radio. In those scenarios, the received sig
nal may suﬀer from signiﬁcant timevarying Doppler dis
tortion due to relative motion between the transmitter and
the receiver. This occurs, for instance, in the lastgeneration
mobilesatellite communication systems based on a con
stellation of nongeostationary lowearthorbit (LEO) satel
lites [1] and in millimeterwave mobile communications for
traﬃc control and assistance [2]. In such situations, car
rier Dopplershift and Dopplerrate 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 Dopplershift aﬀecting the received carrier [3] and a few
algorithms for Dopplerrate estimation are also available in
the open literature [4, 5]. The issue of joint Dopplershift
and Dopplerrate 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 preamblepayload 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 socalled 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 socalled 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 signalto
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 Cram´ erRao 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
a
b
N/2 N/2
N/3 N/3 N/3
M
2P format
3P format
Preamble Payload Postamble
M + N/2
M/2 M/2
Payload Payload
L
P P P
c
d
N/4 M/3 N/4 N/4 M/3 N/4 M/3
1st 2P subburst 2nd 2P subburst
4P format
P Payload Payload Payload P P P
2M/3 + N/2
Payload Payload
L
P P P P
Figure 1: 2P burst format, 3P burst format, and 4P burst format.
[12–14], but only for joint carrier phase/Dopplershift es
timation. The novelty of the paper is to extend the anal
ysis to the joint carrier phase/Dopplershift 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 Dopplerrate, 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 Dopplerrate as well,
and that the Dopplerrate CRB is minimized by estima
tion over three equallength 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 highresolution and lowcomplexity carrier
Dopplershift and Dopplerrate 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
lowcomplexity DA Dopplershift estimator for the optimal
2P format. Extensions of this algorithm for joint carrier
phase/Dopplershift and Dopplerrate estimation for the 2P
format, the 3P format, and the suboptimum 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:
η =
N
L
=
N
N + M
=
1
1 + M/N
. (1)
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
timevarying 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 discretetime signal:
z(k) = c
k
e
jϕk
+ n(k), k = −
L −1
2
, . . . , 0, . . . ,
L −1
2
,
(2)
where
ϕ
k
= θ + 2πνkT + παk
2
T
2
(3)
is the instantaneous carrier excess phase, {c
k
} are unitenergy
(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 (Dopplershift), and ﬁnally α is the constant unknown
carrier frequency rateofchange (Dopplerrate). For signal
model (2) to be valid, we assumed that the value of the
Dopplershift ν is much smaller than the symbol rate, and
that the value of the Dopplerrate α is much smaller than
the square of the symbol rate. The noise n(k) is a complex
valued zeromean WGN process with independent compo
nents, each with variance σ
2
= N
0
/(2E
s
), where E
s
/N
0
repre
sents the ratio between the received energypersymbol and
the onesided 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) = c
∗
k
z(k). The result is
r(k) = e
jϕk
+ w(k), k ∈ K =
_
_
N
Pi
_
, (4)
Luca Giugno et al. 3
where K is the symmetric set of N time indices correspond
ing to pilot symbols, and w(k) = c
∗
k
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 performcarrier 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. DOPPLERSHIFT ESTIMATOR: FEPE ALGORITHM
We momentarily neglect the eﬀect of the Dopplerrate α in
(4), to concentrate on the issue of Dopplershift estimation
only. Under such hypothesis, (4) can be rewritten as follows:
r(k) = e
j(θ+2πνkT)
+ w(k), k ∈ K. (5)
The 2P format minimizes the CRB for Dopplershift esti
mation for joint carrier phase/Dopplershift 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 maximumlikelihood es
timation problem for ν appears infeasible. We introduce thus
a new lowcomplexity algorithm suitable for the estimation
of the Dopplershift 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 lowcomplexity maximumlikelihood (ML)
algorithm [17]:
θ
1
=arg
_
−(M−1)/2
_
k=−(N+M−1)/2
r(k)
_
,
θ
2
=arg
_
(N+M−1)/2
_
k=(M−1)/2
r(k)
_
,
(6)
where arg{·} denotes the phase of the complexvalued 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,
θ
1
) and ((M−1)/2 + N/4,
θ
2
) on
the (time, phase) plane:
ν =
¸
¸
¸
¸
θ
2
¸
¸
2π
−
¸
¸
θ
1
¸
¸
2π
¸
¸
2π
2π(M + N/2)T
. (7)
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
−1.5
−1
−0.5
0
0.5
1
1.5
×10
−3
M
E
V
−1.5 −1 −0.5 0 0.5 1 1.5
×10
−3
νT (Hz ×s)
Ideal
E
s
/N
0
= 0 dB
E
s
/N
0
= 10 dB
E
s
/N
0
= 20 dB
E
s
/N
0
= 100 dB
Figure 2: MEV of FEPE estimator for diﬀerent values of E
S
/N
0
—
simulation only. Preamble + postamble DA ML phase estimation,
N = 44, M = 385.
ﬁxedpoint 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 ≤ 
θ
2

2π
−
θ
1

2π
< π, and therefore the range is bounded
to
ν ≤
1
2(M + N/2)T
. (8)
This relatively narrow interval does not allowto use the FEPE
algorithm for initial acquisition of a large frequency oﬀset at
receiver startup. Its use is therefore restricted to ﬁne esti
mation of a residual oﬀset after a coarse acquisition or com
pensation of motioninduced Dopplershift. Figure 2 depicts
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 Dopplershift ν) for diﬀerent values of
E
s
/N
0
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
θ
2
and
θ
1
are substantially unbiased
as well. Such condition is implicitly assumed in (8) since in
the practice M N/2. The curve labeled E
s
/N
0
= 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
θ
1
and
θ
2
in (7)
have an estimation variance σ
2
θ
that achieves the Cram´ erRao
4 EURASIP Journal on Wireless Communications and Networking
Bound (CRB) [17]:
σ
2
θ
= CRB(θ) =
1
2 · N/2
1
E
s
/N
0
. (9)
Therefore, considering that the two phase estimates in (7) are
independent, we get
σ
2
FEPE
(ν)=
2 · σ
2
θ
4π
2
(M + N/2)
2
T
2
=
1
4π
2
T
2
N/2(M + N/2)
2
1
E
s
/N
0
.
(10)
The vector CRB [18] for the frequency oﬀset estimate in the
joint carrier phase/Dopplershift estimation with the 2P for
mat is derived in Appendix A and reads as follows:
VCRB
2P
(ν)=
3
4π
2
T
2
(N/2)
_
4(N/2)
2
+ 3M
2
+3MN−1
_
1
E
s
/N
0
.
(11)
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
small estimation complexity) with respect to the conven
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,
which is always veriﬁed in the practice.
Therefore, the ratio between the CRB (11) and the vari
ance of the FEPE estimator is very close to 1. With N = 44
and M = 385, we get, for instance, σ
2
FEPE
/VCRB
2p
= 0.99.
The enhancedaccuracy feature is also apparent in the com
parison of the VCRB
2p
(ν) as in (11) with the conventional
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.
Figure 3 shows curves of the (symbolratenormalized)
RMSEE (root mean square estimation error) of the FEPE
algorithm (i.e., T
_
E{( v −v)
2
}) as a function of E
s
/N
0
for
various values of the true oﬀset ν. In particular, marks are
simulation results for σ
2
FEPE
, whilst the lowermost line is the
VCRB
2p
(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 Dopplerrate 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,
10
−6
10
−5
10
−4
10
−3
10
−2
N
o
r
m
a
l
i
z
e
d
R
M
S
E
E
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
E
s
/N
0
(dB)
σ
2
FEPE
(ν)
νT = 1 ×10
−3
νT = 1 ×10
−4
VCRB(ν)
VCRB
2P
(ν)
Figure 3: RMSEE of FEPE estimator for diﬀerent values of
E
S
/N
0
and relevant bounds—solid lines: theory—marks: simula
tion. Preamble + postamble DA ML phase estimation, N = 44,
M = 385.
13] is presented in Table 1. It is clear that the strength of the
FEPE algorithm is its very low complexity as compared to
conventional algorithms.
4. DOPPLERRATE ESTIMATORS IN2P FRAME:
FREPE ANDFREFE ALGORITHMS
We take now back into consideration the presence of a non
negligible Dopplerrate 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 Dopplershift estimation in joint
carrier phase/Dopplershift and Dopplerrate 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
−6
2
3
4
5
6
7
10
−5
2
3
4
5
6
7
10
−4
2
3
4
5
6
7
10
−3
N
o
r
m
a
l
i
z
e
d
R
M
S
E
E
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
E
s
/N
0
(dB)
VCRB
2P
(ν)
αT
2
= 0
αT
2
= 2 ×10
−8
αT
2
= 5 ×10
−8
αT
2
= 1 ×10
−7
αT
2
= 2 ×10
−7
αT
2
= 2.5 ×10
−7
Figure 4: Sensitivity of FEPE estimator to diﬀerent values of the
Dopplerrate αT
2
. Preamble + postamble DA ML phase estimation,
N = 44, M = 385, vT = 1.0 ×10
−3
.
ﬁrst section of the preamble, we obtain the phase model
ϕ
P
(n) = aπ
_
n +
M−1
2
+
3N
8
_
2
+ 2πb
_
n +
M−1
2
+
3N
8
_
+ c,
(12)
where the regression coeﬃcients a and b directly repre
sent estimates for the (normalized) carrier Dopplerrate and
Dopplershift, respectively, and c is an estimate for the initial
phase (that we are not interested into). The coeﬃcients are
found after observing that the MSE is written as
ε(a, b, c) =
4
_
i=1
_
ϕ
P
_
n
i
_
−
θ
i
_
2
=
4
_
i=1
e
2
i
, (13)
where
θ
i
, i = 1, . . . , 4, are the abovementioned ML phase
estimates on N/4 pilots each, and n
1
= −[(M−1)/2+3N/8],
n
2
= −[(M − 1)/2 + N/8], n
3
= [(M − 1)/2 + N/8], and
n
4
= [(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
Table 1: The FEPE computational complexity comparison.
(N
alg
= estimation design parameter.)
Computational complexity of major
Dopplershift estimation algorithms
Algorithm Reference
Number of real products
and additions
LUT access
L&R [3] 4N
_
N
alg
+ 1
_
−2 1
M&M [3] N
alg
_
8N −4N
alg
−3
_
−2 N
alg
SBLUE [13] 4N
2
+ 4.5N −3 1.5N −2
PBLUE2 [13] 4N −1 1
FEPE — 2N + 3 2
ε(a, b, c) with respect to a, b, and c, we obtain
∂ε(a, b, c)
∂a
=
4
_
i=1
e
i
·
_
n
i
+
_
M−1
2
+
3N
8
__
2
= 0,
∂ε(a, b, c)
∂b
=
4
_
i=1
e
i
·
_
n
i
+
_
M−1
2
+
3N
8
__
= 0,
∂ε(a, b, c)
∂c
=
4
_
i=1
e
i
= 0,
(14)
and solving for a we get the following socalled frequency rate
estimation through phase estimation (FREPE) algorithm [15]:
α
FREPE
=
a
T
2
=
_
θ
4
−
θ
3
_
−
_
θ
2
−
θ
1
_
πN/2(N/2 + M)T
2
(15)
(all diﬀerences to be intended modulo2π). This extremely
simple approach can be viewed as a generalization of the
FEPE introduced in the previous section. In particular, by us
ing (7), the terms
_
θ
i
−
θ
i−1
_
2π(N/4)T
, i = 2, 4, (16)
represent two Dopplershift estimations, the ﬁrst on the
preamble and the second on the postamble, respectively,
which are spaced M + N/2 symbols apart. The Dopplerrate
estimate is thus simply the diﬀerence between the two fre
quency estimates, divided by their time distance (M+N/2)T.
The considerations above allow us to also introduce
the frequency rate estimation through frequency estimation
(FREFE) algorithm [15]
α
FREFE
=
ν
2
−ν
1
(M + N/2)T
, (17)
wherein the two frequency estimates ν
1
and ν
2
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
σ
2
FREFE
( α) =
2σ
2
ν
(M + N/2)
2
T
2
=
3
π
2
T
4
N/2
_
(N/2)
2
−1
_
(M + N/2)
2
1
E
s
/N
0
,
(18)
where we have used σ
2
ν
= 3 · (E
s
/N
0
)
−1
/[2π
2
T
2
N/2((N/2)
2
−
1)] [17]. This can be compared to the variance of the FREPE
algorithm that is easily found to be
σ
2
FREPE
( α) =
4 · σ
2
θ
π
2
(N/2)
2
(M + N/2)
2
T
4
=
4
π
2
T
4
(N/2)
3
(M + N/2)
2
1
E
s
/N
0
,
(19)
where now σ
2
θ
= (E
s
/N
0
)
−1
/(N/2). The relevant vector CRB
for Dopplerrate estimate is (see Appendix B):
VCRB
2P
(α)
=
45
π
2
T
4
_
(N/2)
3
−N/2
__
16(N/2)
2
+15M
2
+30MN/2−4
_
1
E
s
/N
0
.
(20)
All expressions inversely depend on (N/2)
5
as in conven
tional preambleonly estimation of the Dopplerrate [6], but
they also bear again inverse dependence on M
2
that gives en
hanced accuracy. For suﬃciently large values of N and M,
M N, we have
σ
2
FREFE
( α)
σ
2
FREPE
( α)
∼
=
3
4
,
VCRB
PP
(α)
σ
2
FREFE
( α)
∼
= 1. (21)
Figure 5 shows the MEV curves (i.e., E{ α}) of the FREPE al
gorithm for diﬀerent values of E
s
/N
0
, in the case of N = 44,
M = 385, and Dopplershift vT = 10
−3
. The estimator is
unbiased with an operating range equal to
¸
¸
α
FREPE
¸
¸
≤
1
N/2(M + N/2)T
2
. (22)
The sensitivity of FREPE to diﬀerent uncompensated val
ues of vT is illustrated in Figure 6 in terms of MEV.
The same simulations have been run also for the FREFE
algorithm. In particular, Figure 7 illustrates the MEV curves
for diﬀerent values of E
s
/N
0
and with vT = 10
−3
. By using
the L&R algorithm to estimate ν
1
and ν
2
, the operating range
of FREFE is roughly twice that of FREPE:
¸
¸
α
FREFE
¸
¸
≤
1
(N/4 + 1)(M + N/2)T
2
. (23)
In particular, the term [(N/2 + 1)T]
−1
represents the fre
quency pullin range of L&R on N/2 pilots [20].
Figure 8 demonstrates that FREPE is also less sensitive
than FREFE to an uncompensated Dopplershift. Finally,
Figure 9 shows the curve of the Dopplerrate RMSEE of
FREPE and FREFE as a function of E
s
/N
0
, for νT = 10
−3
and
αT
2
= 10
−6
. The FREPE estimator loses only 10 log
10
(4/3) =
1.25 dB in terms of E
s
/N
0
with respect to the performance of
the more complex FREFE when N 1.
−1.5
−1
−0.5
0
0.5
1
1.5
×10
−4
M
E
V
−1.5 −1 −0.5 0 0.5 1 1.5
×10
−4
αT
2
(Hz/s ×s
2
)
Ideal
E
s
/N
0
= 0 dB
E
s
/N
0
= 10 dB
E
s
/N
0
= 20 dB
E
s
/N
0
= 100 dB
Figure 5: MEV of FREPE estimator for diﬀerent values of E
S
/N
0
—
simulation only. Preamble + postamble DA ML phase estimation,
N = 44, M = 385, vT = 1.0 ×10
−3
.
−1.5
−1
−0.5
0
0.5
1
1.5
×10
−4
M
E
V
−1.5 −1 −0.5 0 0.5 1 1.5
×10
−4
αT
2
(Hz/s ×s
2
)
Ideal
νT = 0
νT = 1 ×10
−3
νT = 5 ×10
−3
νT = 1 ×10
−2
Figure 6: MEV of FREPE estimator for diﬀerent values of the
Dopplershift vT—simulation only. Preamble + postamble DA ML
phase estimation, N = 44, M = 385, E
s
/N
0
= 10 dB.
5. OPTIMUMDOPPLERRATE ESTIMATION
5.1. Odd number of pilot ﬁelds: FRE3PE algorithm
We turn now to the issue of optimum burst conﬁguration
for the estimation of the Dopplerrate. We demonstrate in
Appendix B that the 3P format (Figure 1(b)) minimizes the
CRB for Dopplerrate 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 lowcomplexity
algorithm suitable for Dopplerrate 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
−2.5
−2
−1.5
−1
−0.5
0
0.5
1
1.5
2
2.5
×10
−4
M
E
V
−2.5 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2 2.5
×10
−4
αT
2
(Hz/s ×s
2
)
Ideal
E
s
/N
0
= 0 dB
E
s
/N
0
= 10 dB
E
s
/N
0
= 20 dB
E
s
/N
0
= 100 dB
Figure 7: MEV of FREFE estimator for diﬀerent values of E
S
/N
0
—
simulation only. Preamble + postamble Luise and Reggiannini, N =
44, M = 385, vT = 1.0 ×10
−3
.
−2.5
−2
−1.5
−1
−0.5
0
0.5
1
1.5
2
2.5
×10
−4
M
E
V
−2.5 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2 2.5
×10
−4
αT
2
(Hz/s ×s
2
)
Ideal
νT = 0
νT = 1 ×10
−3
νT = 5 ×10
−3
νT = 1 ×10
−2
Figure 8: MEV of FREFE estimator for diﬀerent values of the
Dopplershift vT—simulation only. FREFE estimator preamble +
postamble Luise and Reggiannini, N = 44, M = 385, E
s
/N
0
=
10 dB.
algorithm in the 2P conﬁguration, a simple estimator of α
in the 3P format is found by computing three (independent)
ML phase estimates on the three blocks of pilots, and then
ﬁtting a secondorder phase polynomial. Taking the origin in
the ﬁrst block of pilots, we obtain this time the phase model
ϕ
P
(n) = aπ
_
n+
N
3
+
M
2
_
2
+2πb
_
n+
N
3
+
M
2
_
+c. (24)
The coeﬃcients are found solving the following set of equa
tions:
ϕ
P
_
n
i
_
=
θ
i
, i = 1, . . . , 3, (25)
10
−8
10
−7
10
−6
10
−5
10
−4
10
−3
N
o
r
m
a
l
i
z
e
d
R
M
S
E
E
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
E
s
/N
0
(dB)
FREPE, αT
2
= 1 ×10
−6
FREFE, αT
2
= 1 ×10
−6
FRE2FEPE, αT
2
= 1 ×10
−6
FRE3PE, αT
2
= 1 ×10
−6
VCRB
P
(α)
VCRB
2P
(α)
VCRB
4P
(α)
VCRB
3P
(α)
Figure 9: RMSEE of FREPE, FREFE, FRE3PE, and FRE2FREPE
estimators for diﬀerent values of E
S
/N
0
and relevant bounds,—solid
lines: theory—marks: simulation. Dopplerrate algorithms: FREFE
versus FREPE versus FRE3PE versus FRE2FEPE, N = 44(45),
M = 385(384), vT = 1.0 ×10
−3
.
where
θ
i
are the abovementioned ML phase estimates on
N/3 pilots each, and where n
1
= −(M/2 + N/3), n
2
= 0, and
n
3
= (M/2 + N/3) are the three time instants that we con
ventionally associate to the three estimates (the midpoints of
the three subsections). Solving for a, we get the following so
called (FRE3PE) (frequency rate estimation through 3 phase
estimations) algorithm:
α
FRE3PE
=
a
T
2
=
18
__
θ
3
−
θ
2
_
−
_
θ
2
−
θ
1
__
π(2N + 3M−2)
2
T
2
(26)
(all diﬀerences to be intended modulo2π). The estimator is
unbiased with an operating range equal to:
¸
¸
α
FRE3PE
¸
¸
≤
18
(2N + 3M−2)
2
T
2
. (27)
In our simulations (N = 45 and M = 384), α
FRE3PE
· T
2
 ≤
10
−5
. This range is narrower than FREPE’s and FREFE’s in
the 2P format, but it still widely includes practical Doppler
rate values mentioned in Section 3. Figure 10 shows the MEV
curves of the FRE3PE algorithm for diﬀerent values of
E
s
/N
0
, in the case of N = 45, M = 384, and Dopplershift
8 EURASIP Journal on Wireless Communications and Networking
−1.5
−1
−0.5
0
0.5
1
1.5
×10
−5
M
E
V
−1.5 −1 −0.5 0 0.5 1 1.5
×10
−5
αT
2
(Hz/s ×s
2
)
Ideal
E
s
/N
0
= 0 dB
E
s
/N
0
= 10 dB
E
s
/N
0
= 20 dB
E
s
/N
0
= 100 dB
Figure 10: MEV of FRE3PE estimator for diﬀerent values of
E
S
/N
0
—simulation only. 3 blocks of pilots DA ML phase estima
tion, N = 45, M = 384, vT = 1.0 ×10
−3
.
vT = 10
−4
, while Figure 11 shows the sensitivity of the MEV
to diﬀerent uncompensated values of the Dopplershift vT.
The theoretical error variance of the FRE3PE estimator
can be easily evaluated, similarly to what was done for the
calculation of σ
2
FREFE
( α) in Section 4:
σ
2
FRE3PE
( α) =
18
2
· 6 · σ
2
θ
π
2
(2N + 3M−2)
4
T
4
=
18
2
· 6
π
2
T
4
(2N/3)(2N + 3M−2)
4
1
E
s
/N
0
,
(28)
where now σ
2
θ
= (E
s
/N
0
)
−1
/(2N/3). Comparing this expres
sion with the VCRB
3P
(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 Dopplerrate [6]. On the other hand,
σ
2
FRE3PE
( α) and VCRB
3P
(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 FRE3PE, FREPE, and
FREFE versus E
s
/N
0
. To perform a fair comparison, we also
reported the VCRB
P
(β), obtained in the case of estimation of
Dopplerrate in the preambleonly conﬁguration. The FRE
3PE algorithm attains its own CRB, and exhibits a gain of
19 dB in terms of E
s
/N
0
with respect to the 2P format.
As a ﬁnal remark, we only mention that a simple estima
tor of Dopplershift in the 3P format is found by applying the
FEPE algorithm to the two extreme pilot ﬁelds of the burst.
Its variance reaches the VCRB
3P
(ν) calculated setting x = 1
in (B.7) and (B.9), that is 1.5 dB apart from the VCRB
2P
(ν)
of the optimal 2P format.
−1.5
−1
−0.5
0
0.5
1
1.5
×10
−5
M
E
V
−1.5 −1 −0.5 0 0.5 1 1.5
×10
−5
αT
2
(Hz/s ×s
2
)
Ideal
νT = 0
νT = 1 ×10
−4
νT = 5 ×10
−4
νT = 1 ×10
−3
Figure 11: MEV of FRE3PE estimator for diﬀerent values of the
Dopplershift vT—simulation only. 3 blocks of pilots, N = 45, M =
384, E
S
/N
0
= 10 dB.
5.2. Even number of pilot ﬁelds: FRE2FEPE algorithm
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 VCRB
4p
(α)/VCRB
3p
(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
and postamble is now of length N/4, and we can derive two
FEPE estimates of frequency on each subburst:
ν
1
=
¸
¸
¸
¸
θ
2
¸
¸
2π
−
¸
¸
θ
1
¸
¸
2π
¸
¸
2π
2π(M/3 + N/4)T
, ν
2
=
¸
¸
¸
¸
θ
4
¸
¸
2π
−
¸
¸
θ
3
¸
¸
2π
¸
¸
2π
2π(M/3 + N/4)T
,
(29)
where
θ
i
, i = 1, . . . , 4, are the ML phase estimates computed
on the four pilot ﬁelds of N/4 pilots each. The two Doppler
shift estimates ν
1
and ν
2
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 Dopplerrate as the slope of the line that connects the two
points (−(M/3 −1/2) −N/4, ν
1
) and ((M/3 −1/2) +N/4, ν
2
)
in the (time, frequency) plane:
α
FRE2FEPE
=
ν
2
−ν
1
(2M/3 + N/2)T
. (30)
We call this algorithm FRE2FEPE (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 FRE2FEPE are not reported here since
Luca Giugno et al. 9
they basically mimic those in Figures 10 and 11 for the
FRE3PE algorithm. The estimation error variance of (30)
is found to be
σ
2
FRE2FEPE
( α) =
σ
2
θ
(2M/3 + N/2)
2
(M/3 + N/4)
2
π
2
T
2
=
2 ·
_
E
s
/N
0
_
−1
π
2
T
4
N(2M/3 + N/2)
2
(M/3 + N/4)
2
.
(31)
Figure 9 shows also the curves of the RMSEE of FRE2FEPE
and its respective CRB. The FRE2FEPE algorithm reaches its
own VCRB
4p
(α) and thus, as demonstrated in Appendix B, it
gains 10 log
10
(7.19) = 18.5 dB in terms of E
s
/N
0
with respect
to the performance of the previous algorithms with the 2P
format. Also, the FRE2FEPE loses only 0.4 dB with respect
to the FRE3PE algorithm and can thus be a valid alternative
to the 3P format.
As a ﬁnal remark, we brieﬂy address the issue of Doppler
shift estimation in the 4P format. The best method is found
by applying the FEPE algorithm to the two extreme pilot
ﬁelds of the burst. Its variance is close to the VCRB
4P
(ν) cal
culated setting x = 1 in (B.8) and (B.9), that is 2.4 dB worse
than the VCRB
2P
(ν) of the optimal 2P format.
6. CONCLUSIONS
In this paper, we presented and analyzed some very
lowcomplexity algorithms for carrier Dopplershift and
Dopplerrate estimation in burst digital transmission. To
achieve enhanced accuracy, the burst conﬁgurations that
minimize the CRB for the estimation of Dopplershift and
Dopplerrate are derived. Our analysis showed that the 2P
format is optimum for Dopplershift estimation and that the
3P format is optimum for Dopplerrate estimation. These
two conﬁgurations can be practically thought as repetition of
two/three consecutive conventional (preambleonly) bursts.
Despite preventing fromrealtime processing of the data pay
load section, the 2P and 3P formats greatly outperform the
estimation based on conventional preambleonly pilot dis
tribution. Performance assessment has shown that all of the
proposed algorithms are unbiased in practical operating con
ditions, and that their accuracy in terms of estimation vari
ance gets remarkably close to their respective CRBs down to
very low E
s
/N
0
values.
APPENDICES
A. VCRB FOR JOINT CARRIER PHASE/DOPPLERSHIFT
ESTIMATIONWITH2P FORMAT
In this appendix, we calculate the VCRB for the error vari
ance of any unbiased estimator of Dopplershift in the case of
joint estimation of phase/Dopplershift using the preamble
postamble (2P) format. We explicitly mention that we have
chosen a set K of pilot locations that is symmetrical with
respect to the middle of the burst, since a symmetrical K de
couples phase from Dopplershift estimation, as discussed in
[12]. After modulation removal, the generic sample within
the preamble and the postamble is given by (5).
The Fisher information matrix (FIM) [18] can be written
as
F =
_
F
θθ
F
θν
F
νθ
F
νν
_
=
⎡
⎢
⎢
⎢
⎢
⎣
−E
r
_¸
¸
¸
¸
¸
∂
2
ln p(r  `ν,
`
θ)
∂
`
θ
2
¸
¸
¸
¸
¸
_
−E
r
_¸
¸
¸
¸
¸
∂
2
ln p(r  `ν,
`
θ)
∂
`
θ∂`ν
¸
¸
¸
¸
¸
_
−E
r
_¸
¸
¸
¸
¸
∂
2
ln p(r  `ν,
`
θ)
∂`ν∂
`
θ
¸
¸
¸
¸
¸
_
−E
r
_¸
¸
¸
¸
¸
∂
2
ln p(r  `ν,
`
θ)
∂`ν
2
¸
¸
¸
¸
¸
_
⎤
⎥
⎥
⎥
⎥
⎦
,
(A.1)
where p(r  `ν,
`
θ) is the probability density function of r =
{r(k)}, k ∈ K, conditioned on (`ν,
`
θ), and r(k) is a random
Gaussian variable with variance equal to σ
2
= N
0
/(2E
s
) and
mean value equal to
`s(k) = e
j(
`
θ+2π`νkT)
. (A.2)
Therefore, we write p(r  `ν,
`
θ) as
p(r  `ν,
`
θ) =
k∈K
p
_
r
k
 `ν,
`
θ
_
=
1
_
2πσ
2
_
N
exp
_
−
1
2σ
2
_
k∈K
¸
¸
r(k) −`s(k)
¸
¸
2
_
.
(A.3)
Taking the logarithm of (A.3), we obtain
ln p(r  `ν,
`
θ)
= N ln
_
1
2πσ
2
_
−
1
2σ
2
_
k∈K
_¸
¸
r(k)
¸
¸
2
+
¸
¸
`s(k)
¸
¸
2
−2Re
_
r(k)`s
∗
(k)
__
=C +
1
σ
2
_
k∈K
Re
_
r(k)`s
∗
(k)
_
,
(A.4)
where C is a constant term that includes all the quantities
independent of `ν and
`
θ. After diﬀerentiating twice (A.4) with
respect to `ν and
`
θ, calculating the expectation of the various
terms with respect to r, we get
F =
_
a
b
c
d
_
, (A.5)
where
a
=
_
1
σ
2
_
_
k∈K
_
(1)E
r
_
Re
_
r(k)` s
∗
(k)
___
,
b
=
_
1
σ
2
_
_
k∈K
_
(2πTk)E
r
_
Re
_
r(k)` s
∗
(k)
___
,
c
=
_
1
σ
2
_
_
k∈K
_
(2πTk)E
r
_
Re
_
r(k)` s
∗
(k)
___
,
d
=
_
1
σ
2
_
_
k∈K
__
4π
2
T
2
k
2
_
E
r
_
Re
_
r(k)` s
∗
(k)
___
.
(A.6)
10 EURASIP Journal on Wireless Communications and Networking
By noticing that
E
r
_
Re
_
r(k)`s
∗
(k)
__
= 1, (A.7)
we obtain
F =
1
σ
2
⎡
⎢
⎢
⎢
⎢
⎣
_
k∈K
(1) 2πT
_
k∈K
k
2πT
_
k∈K
k 4π
2
T
2
_
k∈K
k
2
⎤
⎥
⎥
⎥
⎥
⎦
, (A.8)
where, considering the symmetry of the range K,
_
k∈K
(1) = N,
_
k∈K
k = 0, (A.9)
_
k∈K
k
2
=
N/2
3
_
8
_
N
2
_
2
−6
_
N
2
_
+ 1
+ 3M
2
+ 3M
_
3
_
N
2
_
−1
__
.
(A.10)
After calculation of F
−1
, the VCRB for ν in case of joint
phase/Dopplershift estimation is found to be
F
−1
νν
= VCRB
2P
(ν) =
1
2π
2
T
2
k∈K
k
2
1
E
s
/N
0
=
3 ·
_
E
s
/N
0
_
−1
4π
2
T
2
(N/2)
_
4(N/2)
2
+ 3M
2
+ 3MN −1
_.
(A.11)
B. OPTIMAL SYMMETRIC BURST CONFIGURATION
FOR JOINT CARRIERPHASE/DOPPLERSHIFT
ANDDOPPLERRATE ESTIMATION:
2P, 3P, 4P FORMATS
This appendix addresses the optimal signal design for
Dopplershift ν and Dopplerrate α estimation in the case of
joint phase/Dopplershift and Dopplerrate estimation when
the received signal is expressed by (2)–(4). The optimal train
ing signal structure is developed by minimizing the vector
Cram´ erRao bounds (VCRBs) [17, 18] for ν and α, with
constraints on the total training block length and on the
burst overhead (1) of the signal (4). In fact, the Cram´ erRao
bounds (CRBs) for joint estimations are functions of the lo
cation of the reference symbols in the burst.
The issue of ﬁnding the optimal burst format that mini
mizes the frequency CRB has been already addressed in [12–
14], but only for joint phase/Dopplershift estimation. We
restrict our analysis to a symmetric burst format. In the se
quel, we demonstrate that this symmetry also decouples the
estimation of Dopplershift and Dopplerrate. Our attention
is focused on a generic burst format as in Figure 12, either
with an even (Figure 12(a)) or an odd (Figure 12(b)) num
ber of blocks of pilots. Just to rehearse notation, we mention
that the length of the burst is L symbols, N is the total num
ber of pilot symbols, N
P
is the number of reference symbols
in each subgroup, M is the total number of data symbols,
and M
D
is the number of data symbols in each subgroup.
In Figure 12(a), 2x
even
is the (even) number of subgroups of
N
P
M
D
N
P
M
D
N
P
M
D
N
P
M
D
N
P
M
D
N
P
Symmetric format
0
P P P P P P
(a)
N
P
M
D
N
P
M
D
N
P
M
D
N
P
M
D
N
P
M
D
N
P
M
D
N
P
0
L
P P P P P P P
(b)
Figure 12: Generic symmetric burst format.
pilot symbols, and (2x
even
+ 1) is the (odd) number of sub
groups of data symbols; in Figure 12(b), (2x
odd
+ 1) is the
(odd) number of subgroups of pilot symbols, and 2x
odd
is
the (even) number of subgroups of data symbols. In the se
quel we ﬁnd the values of x that minimize the VCRBs of ν
and α, for ﬁxed values of L, N, and M.
In the case of joint phase/Dopplershift/Dopplerrate es
timation, the ﬁsher information matrix (FIM) of the generic
bursts of Figure 12 can be written as
F =
⎡
⎢
⎢
⎣
F
θθ
F
θν
F
θα
F
νθ
F
νν
F
να
F
αθ
F
αν
F
αα
⎤
⎥
⎥
⎦
=
⎡
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎣
−E
r
_¸
¸
¸
¸
a
∂
`
θ
2
¸
¸
¸
¸
_
−E
r
_¸
¸
¸
¸
a
∂
`
θ∂`ν
¸
¸
¸
¸
_
−E
r
_¸
¸
¸
¸
a
∂
`
θ∂` α
¸
¸
¸
¸
_
−E
r
_¸
¸
¸
¸
a
∂`ν∂
`
θ
¸
¸
¸
¸
_
−E
r
_¸
¸
¸
¸
a
∂`ν
2
¸
¸
¸
¸
_
−E
r
_¸
¸
¸
¸
a
∂`ν∂` α
¸
¸
¸
¸
_
−E
r
_¸
¸
¸
¸
a
∂` α∂
`
θ
¸
¸
¸
¸
_
−E
r
_¸
¸
¸
¸
a
∂` α∂`ν
¸
¸
¸
¸
_
−E
r
_¸
¸
¸
¸
a
∂` α
2
¸
¸
¸
¸
_
⎤
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎦
,
(B.1)
where a
= ∂
2
ln p(r  ` α, `ν,
`
θ), p(r  ` α, `ν,
`
θ) is the probability
density function of r = {r(k)}, with k ∈ K, conditioned on
(` α, `ν,
`
θ). Now r(k) is a random Gaussian variable with vari
ance equal to σ
2
= N
0
/(2E
s
) and mean equal to
`s(k) = e
j(
`
θ+2π`νkT+` απk
2
T
2
)
(B.2)
so that
p(r  ` α, `ν,
`
θ) =
k∈K
p
_
r
k
 ` α, `ν,
`
θ
_
=
1
_
2πσ
2
_
N
exp
_
−
1
2σ
2
_
k∈K
¸
¸
r(k) −`s(k)
¸
¸
2
_
.
(B.3)
As detailed in Appendix A, after taking the logarithm of
(B.3), and after diﬀerentiating with respect to the unknown
parameters, and calculating the expectation of the terms with
Luca Giugno et al. 11
respect to r, we have
F =
1
σ
2
⎡
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎣
_
k∈K
(1) 2πT
_
k∈K
k πT
2
_
k∈K
k
2
2πT
_
k∈K
k 4π
2
T
2
_
k∈K
k
2
2π
2
T
3
_
k∈K
k
3
πT
2
_
k∈K
k
2
2π
2
T
3
_
k∈K
k
3
π
2
T
4
_
k∈K
k
4
⎤
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎦
, (B.4)
where, thanks to the symmetry of range K,
_
k∈K
k = 0,
_
k∈K
k
3
= 0. (B.5)
We ﬁnally get the expression of the FIM matrix as
F =
1
σ
2
⎡
⎢
⎢
⎢
⎢
⎢
⎢
⎣
N 0 πT
2
_
k∈K
k
2
0 4π
2
T
2
_
k∈K
k
2
0
πT
2
_
k∈K
k
2
0 π
2
T
4
_
k∈K
k
4
⎤
⎥
⎥
⎥
⎥
⎥
⎥
⎦
. (B.6)
With an even number of pilot ﬁelds (Figure 12(a)), we have
_
k∈K
k
2
= 2
xeven−1
_
n=0
N/2xeven
_
l=1
_
_
M/
_
2x
even
−1
_
−1
_
2
+ l +
_
N
2x
even
+
M
2x
even
−1
_
n
_
2
,
_
k∈K
k
4
= 2
xeven−1
_
n=0
N/2xeven
_
l=1
_
_
M/
_
2x
even
−1
_
−1
_
2
+ l +
_
N
2x
even
+
M
2x
even
−1
_
n
_
4
(B.7)
while, with an odd number of pilot ﬁelds (Figure 12(b)), we
get
_
k∈K
k
2
=2
N/(2xodd+1)−1
_
k=1
k
2
+2
xodd−1
_
n=0
N/(2xodd+1)
_
l=1
_
_
N/
_
2x
odd
+ 1
_
−1
_
2
+l
+
_
N
2x
odd
+
M
2x
odd
+1
_
n+
M
2x
odd
_
2
,
_
k∈K
k
4
=2
N/(2xodd+1)−1
_
k=1
k
4
+2
xodd−1
_
n=0
N/(2xodd+1)
_
l=1
_
_
N/
_
2x
odd
+ 1
_
−1
_
2
+l
+
_
N
2x
odd
+
M
2x
odd
+1
_
n+
M
2x
odd
_
4
.
(B.8)
Note that, thanks to the symmetry of the burst, the el
ements F
θν
, F
νθ
, F
αν
, F
να
are all zero, which means that
the joint phase/Dopplershift and Dopplershift/Doppler
rate estimations are decoupled.
Calculating F
−1
, we obtain the VCRBs for the estimation
of ν as follows:
F
−1
νν
= VCRB(ν) =
1
2π
2
T
2
k∈K
k
2
1
E
s
/N
0
, (B.9)
as the one found in (A.11) without any Dopplerrate. The
optimal burst conﬁguration that minimizes the VCRB for ν
is thus the 2P format found in [14] also in the presence of
Dopplerrate eﬀects.
The VCRB for α is
F
−1
αα
= VCRB(α) = −
2N
π
2
T
4
__
k∈K
k
2
_
2
−N
k∈K
k
4
_
1
E
s
/N
0
.
(B.10)
If we compute F
−1
αα
as a function of x through (B.7) and
(B.8), for both conﬁgurations of Figure 12, we ﬁnd that the
minimum for F
−1
αα
is obtained with x
odd
= 1 in (B.8). This
was found by exhaustive numerical evaluation with practical
values for M and N. We can conclude that the VCRB of the
error variance of any unbiased estimator of α is always mini
mized for a conﬁguration with three blocks of pilot symbols
equally spaced by two blocks of data symbols (3P format).
Setting x
odd
= 1 in (B.8) and (B.10), the minimum VCRB
of the error variance of any unbiased estimator of α for the
optimal 3P format is thus
VCRB
min
(α)
= F
−1
αα
¸
¸
xodd=1
= VCRB
3P
(α)
= 9720 ·
_
E
s
/N
0
_
−1
/
_
π
2
T
4
N
_
108
_
4 −5N
2
+ N
4
_
+ 32MN
_
15N
2
−45
_
+ 24M
2
_
35N
2
−45
_
+ 720NM
3
+ 270M
4
__
.
(B.11)
In order to evaluate the gain in using the 3P for
mat, we have compared the VCRB
3P
(α) to the bounds
for α in other conﬁgurations. Figure 13 shows the ra
tios VCRB
2P
(α)/VCRB
3P
(α), VCRB
4P
(α)/VCRB
3P
(α), and
VCRB
2P
(α)/VCRB
4P
(α) as functions of the total number N
of pilots and with η = 10%. It is clear that for practical
values of N = 40 ÷ 70, the 3P format exhibits a gain of
10 log(78.6) = 19 dB in terms of E
s
/N
0
with respect to the
2P format and of 10 log(1.1) = 0.4 dB with respect to the 4P
format. The accuracy of the 3P format and of the 4P format
can be thus considered almost equivalent.
12 EURASIP Journal on Wireless Communications and Networking
−40
−20
0
20
40
60
80
100
120
140
160
180
200
V
C
R
B
(
α
)
r
a
t
i
o
s
0 10 20 30 40 50 60 70 80 90 100
N
VCRB
2P
(α)/VCRB
3P
(α)
VCRB
2P
(α)/VCRB
4P
(α)
VCRB
4P
(α)/VCRB
3P
(α)
78.6
71.9
1.09
Figure 13: VCRB
2P
(α)/VCRB
3P
(α), VCRB
2P
(α)/VCRB
4P
(α) and
VCRB
4P
(α)/VCRB
3P
(α) ratios as function of the total number of pi
lots N.
The various VCRBs can be easily calculated from (B.10)
using the appropriate x and (B.7) and (B.8). We report here
the ﬁnal expressions
VCRB
2P
(α)
= F
−1
αα
¸
¸
xeven=1
=
360 ·
_
E
s
/N
0
_
−1
π
2
T
4
_
N
3
−4N
__
4N
2
+ 15M
2
+ 15MN −4
_,
(B.12)
VCRB
4P
(α)
= F
−1
αα
¸
¸
xodd=2
=25920 ·
_
E
s
/N
0
_
−1
/
_
π
2
T
4
N
_
288N
4
+ 1305N
3
M
+ 240NM
_
8M
2
−15
_
+ 30N
2
_
77M
2
−48
_
+ 32
_
20M
4
−75M
2
+ 36
___
.
(B.13)
REFERENCES
[1] I. Ali, N. AlDhahir, and J. E. Hershey, “Doppler characteriza
tion for LEO satellites,” IEEE Transactions on Communications,
vol. 46, no. 3, pp. 309–313, 1998.
[2] “Special Issue on communications in the intelligence trans
portation system,” IEEE Communications Magazine, vol. 34,
1996.
[3] M. Morelli and U. Mengali, “Feedforward frequency estima
tion for PSK: a tutorial review,” European Transactions on
Telecommunications, vol. 9, no. 2, pp. 103–116, 1998.
[4] F. Giannetti, M. Luise, and R. Reggiannini, “Simple carrier fre
quency rateofchange estimators,” IEEE Transactions on Com
munications, vol. 47, no. 9, pp. 1310–1314, 1999.
[5] M. Morelli, “Dopplerrate estimation for burst digital trans
mission,” IEEE Transactions on Communications, vol. 50, no. 5,
pp. 707–710, 2002.
[6] P. M. Baggenstoss and S. M. Kay, “On estimating the angle pa
rameters of an exponential signal at high SNR,” IEEE Transac
tions on Signal Processing, vol. 39, no. 5, pp. 1203–1205, 1991.
[7] T. J. Abatzoglou, “Fast maximum likelihood joint estima
tion of frequency and frequency rate,” IEEE Transactions on
Aerospace and Electronic Systems, vol. 22, no. 6, pp. 708–715,
1986.
[8] T. S. Rappaport, Wireless Communications: Principles & Prac
tice, PrenticeHall, Englewood Cliﬀs, NJ, USA, 1999.
[9] J. A. Gansman, J. V. Krogmeier, and M. P. Fitz, “Single fre
quency estimation with nonuniform sampling,” in Proceed
ings of the 30th Asilomar Conference on Signals, Systems and
Computers, vol. 1, pp. 399–403, Paciﬁc Grove, Calif, USA,
November 1996.
[10] V. Lottici and M. Luise, “Embedding carrier phase recovery
into iterative decoding of turbocoded linear modulations,”
IEEE Transactions on Communications, vol. 52, no. 4, pp. 661–
669, 2004.
[11] S. Benedetto, R. Garello, G. Montorsi, et al., “MHOMS: high
speed ACM modem for satellite applications,” IEEE Wireless
Communications, vol. 12, no. 2, pp. 66–77, 2005.
[12] F. Rice, “Carrierphase and frequencyestimation bounds for
transmissions with embedded reference symbols,” IEEE Trans
actions on Communications, vol. 54, no. 2, pp. 221–225, 2006.
[13] H. Minn and S. Xing, “An optimal training signal structure for
frequencyoﬀset estimation,” IEEE Transactions on Communi
cations, vol. 53, no. 2, pp. 343–355, 2005.
[14] A. Adriaensen, A. Van Doninck, and W. Steinert, “MFTDMA
burst demodulatir design with pilotsymbol assisted frequency
estimation,” in Proceedings of the 8th International Workshop
on Signal Processing for Space Communications (SPSC ’03),
Catania, Italy, September 2003.
[15] L. Giugno and M. Luise, “Carrier frequency and frequency
rateofchange estimators with preamblepostamble pilot
symbol distribution,” in Proceedings of IEEE International Con
ference on Communications (ICC ’05), vol. 4, pp. 2478–2482,
Seoul, Korea, May 2005.
[16] A. Zhuang and M. Renfors, “Combined pilot aided and de
cision directed channel estimation for the RAKE receiver,” in
Proceedings of the 52nd IEEE Vehicular Technology Conference
(VTC ’00), vol. 2, pp. 710–713, Boston, Mass, USA, September
2000.
[17] U. Mengali and A. N. D’Andrea, Synchronization Techniques
for Digital Receivers, Plenum Press, New York, NY, USA, 1997.
[18] F. Gini, R. Reggiannini, and U. Mengali, “The modiﬁed
Cram´ erRao bound in vector parameter estimation,” IEEE
Transactions on Communications, vol. 46, no. 1, pp. 52–60,
1998.
[19] G. R. J. Povey and J. Talvitie, “Doppler compensation and code
acquisition techniques for LEO satellite mobile radio commu
nications,” in Proceedings of the 5th International Conference on
Satellite Systems for Mobile Communications and Navigation,
pp. 16–19, London, UK, May 1996.
[20] M. Luise and P. Reggiannini, “Carrier frequency recovery
in alldigital modems for burstmode transmissions,” IEEE
Transactions on Communications, vol. 43, no. 2–4, pp. 1169–
1178, 1995.
[21] D. C. Rife and R. R. Boorstyn, “Singletone parameter estima
tion from discretetime observations,” IEEE Transactions on
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
TCPCall Admission Control Interaction in
MultiplatformSpace Architectures
Georgios Theodoridis,
1
Cesare Roseti,
2
Niovi Pavlidou,
1
and Michele Luglio
2
1
Department of Electrical & Computing Engineering, Aristotle University of Thessaloniki, 54124 Thessaloniki, Greece
2
Dipartimento di Ingegneria Elettonica, Universit` a 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 crosslayer 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 spacebased 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 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
the very popular Internetbased applications. It presents sev
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 errorprone wireless links. Communication involv
ing longdelay 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 CACTCP 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 CACconcept, 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 TCPBASEDCAC
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 bytecounting scheme (the actions on the sliding
window are based on the actual number of bytes acknowl
edged) [11].
When the communication path is not errorfree (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 aimof CACis 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,
realtime 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
i=1
B
i
+ B
f
≤ c. (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 overassignment 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
CACTCP INTERACTION
The potential improvement introduced by the implementa
tion of the integrated CACTCP 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. 3
Arrival of a new user A
belonging to the x QoSclass
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
No
Yes
B
j
real datarate = B
j
TCP datarate;
BW occupied =
N
j=1
B
j
datarate;
A nominal datarate + BW occupied ≤ capacity
No
Yes
A is blocked
A is admitted
B
j
real datarate = B
j
nominal datarate;
Figure 1: Crosslayer CACTCP ﬂow chart.
Core network
Figure 2: Standalone GEO satellite.
3.1. Standalone GEOsatellite
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
ticularly meaningful in case of use of high frequencies and/or
terminal mobility.
The large latencybandwidth product could cause two
harmful eﬀects:
(i) the pipe size, indicating the amount on unacknowl
edged data that can be “inﬂight” 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 (slowstart 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. Standalone 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: Standalone 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 twostate (good
bad) Markov model as a suitable error model [13]. There
fore, the packeterror 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” (FRFR) 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 GEOsatellite—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 GEOHAP segment outside the atmosphere
and in lineofsight (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 HAPground 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 GEOHAP archi
tectures is equivalent to the standalone 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 standalone 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; interHAP
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 endtoend re
ception conditions could possibly become harsher, due to the
imperfections of the interHAP 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 interHAP communication.
Core network
Figure 4: Integrated GEO satellite—HAP.
Then, in this scenario, TCP performance suﬀers mainly
from the problems arisen in the standalone HAP architec
ture.
4. EVALUATIONOF THE CACTCP INTERACTION
INSPACE 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 bytecounting
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 standalone GEO (Figure 2),
standalone 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]).
Standalone GEOarchitecture
Data originating from the core network are forwarded via a
gateway toward the GEO satellite, which transparently redi
rects the stream (bentpipe satellite) to the end users. Users
are considered to be ﬁxed and equipped with VSATs appro
priately mounted so as to guarantee lineofsight (LoS) con
dition for the satelliteuser link. Therefore, since the signal
tonoise ratio (SNR) is not only maximized but also relatively
invariant due to the absence of mobility, low PER value can
Georgios Theodoridis et al. 5
Core network
Figure 5: HAP constellation.
be assumed. Moreover, the gatewaysatellite link is typically
dimensioned to be error free.
Standalone 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 freespace 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, costeﬃcient, and userfriendly de
vices. Consequently, the standalone HAPS scenario consid
ers mobile users, which are further divided into three cate
gories based on their mobility characteristics: highwayusers,
suburbanusers, urbanusers. In particular, highwayusers
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 GEOsatellite—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 PERof the endtoend link is de
termined by the PER of the userHAP segment (equal to the
of standalone HAP system), while the long RTT is imposed
by the GEOsatellite 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
RTT. Finally, the integrated GEOHAP 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 TCPCAC 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 QoS
i
class is ad
mitted to the network on the basis of the bandwidth commit
ted to the particular QoS
i
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 TCPCACalgorithm
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 CACscheme, have
been simulated through an oﬄine combination of two sim
ulation tools that run sequentially. In particular, the network
simulator ns2 [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 satelliteHAP or satelliteuser 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 lineofsight (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 twostate 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:
pdf(τ) = λ · e
−λ·τ
, E[τ] =
1
λ
,
pdf(d) = μ · e
−μ·d
, E[d] =
1
μ
.
(2)
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 TCPCAC 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 TCPbased call admission control scheme.
Figure 6 presents the system blocking probability for dif
ferent traﬃc loads. Regardless the network architecture, the
basicCAC algorithm leads in every case to the same block
ing probability for the whole variety of traﬃc loads, which is
0
10
20
30
40
50
60
70
B
l
o
c
k
i
n
g
p
r
o
b
a
b
i
l
i
t
y
(
%
)
7000 8000 9000 10000 11000 12000 13000
Average traﬃc load
GEO, TCPCAC
GEO, basicCAC
GEOHAPS, TCPCAC
GEOHAPS, basicCAC
HAPS, TCPCAC
HAPS, basicCAC
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 basicCAC al
gorithm for all the three scenarios completely coincide with
each other. On the contrary, TCPCAC 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
CAC scheme in comparison to the basicCAC scheme. It
allows the reader to compare the impact of the proposed
scheme on architectures with diﬀerent propagation charac
teristics. As it becomes apparent,
BP decrease
GEO
< BP decrease
HAP
< BP decrease
GEOHAPS
.
(3)
This can be easily explained by the fact that in the case of
the integrated GEOHAPS 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 > y
GEO
> y
GEOHAPS
,
x > y
HAP
> y
GEOHAP
.
(4)
Georgios Theodoridis et al. 7
30
40
50
60
70
80
90
100
D
e
c
r
e
a
s
e
i
n
b
l
o
c
k
i
n
g
p
r
o
b
a
b
i
l
i
t
y
(
%
)
7000 8000 9000 10000 11000 12000 13000
Average traﬃc load
GEO
GEOHAPS
HAPS
Figure 7: Blocking probability decrease versus average traﬃc load.
This means that users of the GEOHAPS 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 TCPfeedback is much more
crucial in a standalone HAP system (high PER, low RTT)
than in a standalone 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, fromFigure 8 it is evident that,
regardless of the network scenario, the implementation of
the integrated TCPCAC 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, GEOHAPS,
HAPS). On the contrary, the basicCAC algorithm assumes
30
40
50
60
70
80
90
100
A
v
e
r
a
g
e
t
h
r
o
u
g
h
p
u
t
(
%
)
0.1 1 10
Blocking probability (%)
GEO, TCPCAC
GEO, basicCAC
GEOHAPS, TCPCAC
GEOHAPS, basicCAC
HAPS, TCPCAC
HAPS, basicCAC
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 standalone GEO
case presents the higher average throughput and the inte
grated GEOHAPS 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
basicCAC algorithm severely limits the system throughput.
5. CONCLUSIONS ANDFUTURE PERSPECTIVES
New and innovative wireless telecommunication architec
tures (including HAPs and satellite segments) are identiﬁed
to provide broadband services in a costeﬃcient 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. Crosslayer 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
toend 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
30
40
50
60
70
80
90
100
110
I
n
c
r
e
a
s
e
i
n
a
v
e
r
a
g
e
t
h
r
o
u
g
h
p
u
t
(
%
)
0 10 20 30 40
Blocking probability (%)
GEO
GEOHAPS
HAPS
Figure 9: Average throughput increase versus blocking probability.
optimal resource utilization, and reduce the system blocking
probability, without altering the TCP standard mechanisms.
Through simulation, we demonstrated a considerable
improvement on the performance with respect to a reference
CAC algorithm that takes into account only QoS require
ments and physical parameters.
ACKNOWLEDGMENT
This paper has been supported by the European ISTFP6
project: “SatNEx II—Satellite Communications Network of
Excellence II.”
REFERENCES
[1] T. C. Tozer and D. Grace, “Highaltitude 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,
AddisonWesley, 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 TCPIP applications,” in Proceedings of IEEE
Military Communications Conference on Communications for
NetworkCentric 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.
[9] K. Shiomoto, N. Yamanaka, and T. Takahashi, “Overview of
measurementbased connection admission control methods
in ATM networks,” IEEE Communications Surveys and Tuto
rials, vol. 2, no. 1, pp. 2–13, 1999.
[10] C. Roseti, G. Theodoridis, M. Luglio, and N. Pavlidou, “TCP
driven CAC scheme for HAPS and satellite integrated sce
nario,” in International Workshop on High Altitude Platform
Systems (WHAPS ’05), Athens, Greece, September 2005.
[11] M. Allman, “TCP Congestion Control with Appropriate Byte
Counting (ABC),” RFC 3465, February 2003.
[12] S. Karapantazis and N. Pavlidou, “Broadband communica
tions via highaltitude platforms: a survey,” IEEE Communi
cations Surveys & Tutorials, vol. 7, no. 1, pp. 2–31, 2005.
[13] J. L. CuevasRu´ız and J. A. DelgadoPen´ın, “Channel model
based on semiMarkovian processes: an approach for HAPS
systems,” in Proceedings of the 14th International Conference
on Electronics, Communications and Computers (CONIELE
COMP ’04), pp. 52–56, Veracruz, Mexico, February 2004.
[14] R. Miura and M. Oodo, “Wireless communications system us
ing stratospheric platforms: R & D program on telecom and
broadcasting system using high altitude platform stations,”
Journal of the Communications Research Laboratory, vol. 48,
no. 4, pp. 33–48, 2001.
[15] B. M. Epstein and M. Schwartz, “Predictive QoSbased admis
sion control for multiclass traﬃc in cellular wireless networks,”
IEEE Journal on Selected Areas in Communications, vol. 18,
no. 3, pp. 523–534, 2000.
[16] K. Fall and K. Varadhan, The ns manual, VINT Project, Uni
versity of California, Berkeley, Calif, USA, 2001, http://www
.isi.edu/nsnam/ns/nsdocumentation.html.
[17] Recommendation ITUR P.6816, “ITUR P.6816 Propaga
tion data required for the design of Earthspace land mobile
telecommunication systems,” January 2003.
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 BOCModulated 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 satellitebased 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 delayfrequency acquisition and tracking stages. Moreover, the
properties of BOCmodulated 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 highresolution 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 BOCmodulated 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 closelyspaced 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
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
sequencecode division multiple access (DSCDMA) 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
twodimensional searching process in timefrequency plane.
This delayDoppler 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 nonLOS
components (especially in indoor environments). The fading
of channel paths induces a certain Doppler spread, related
to the terminal speed. Also, the satellite movement induces
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 Lband 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 BOCmodulated
signals and other signals which use the shiftkeying mod
ulation, such as the GPS C/A code. The BOC performance
has been studied for the GPS military Msignal [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 squarewave modulation scheme, which uses the typi
cal nonreturntozero (NRZ) format [7]. While this type of
modulation provides better resistance to multipath and nar
rowband interference [6], it triggers newchallenges 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 sidepeaks ambiguities.
In order to minimize the inﬂuence of multipath errors,
which are the dominating error sources for many GNSS ap
plications, several receiverinternal 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/Acode or GLONASS. The traditional GPS re
ceiver employs a delaylock 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 zerocrossing point of the Scurve function. A ﬁrst
approach to reduce the inﬂuences of code multipath is the
narrow correlator or narrow early minuslate (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 socalled doubledelta (ΔΔ) correlators, which are the
general name for special code discriminators which are
formed by two correlator pairs instead of one [9]. Some
wellknown 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 mediumdelay 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 BOCmodulated 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 “bumpjumping” algo
rithm is presented in [17]. The “bumpjumping” 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 Scurves (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 carriertonoise
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 fromthe 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
[23]. These techniques are named the sidelobes cancellation
methods (SCM), because they are all based on the idea of
suppressing the undesired lobes of the BOC correlation en
velope and they cope better with the false lock points (ambi
guities) which appear due to BOCmodulation, while keeping
the sharp shape of the main peak. It can be applied in both
acquisition and tracking stages, but due to narrow width of
the main peak, only the tracking stage is considered here.
In contrast with the approach from [20] (valid only for sine
BOC(n, n) cases), our methods have the advantage that they
can be generalized to any sine and cosine BOC(m, n) modu
lation and that they have reduced complexity, since they are
based on an ideal reference correlation function, stored at re
ceiver side. In order to deal with both sidelobes ambiguities
and multipath problems, we used the sidelobes cancellation
idea in conjunction with diﬀerent discriminators, based on
the unambiguous shape of ACF (i.e., the narrow correlator,
the high resolution correlator), or after applying the diﬀer
ential correlation method. We also introduced here an SCM
method with multipath interference cancellation (SCM IC),
where the SCM is used in combination with a MEDLL unit,
and also an SCM algorithm based on threshold comparison.
This paper is organized as follows: Section 2 describes the
signal model in the presence of BOC modulation. Section 3
presents several representative delay tracking algorithms,
employed for comparison with the SCM methods. Section 4
introduces the SCM ideas and presents the SCM usage in
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
to lose lock (MTLL), is done in Section 5. The conclusions
are drawn in Section 6.
2. SIGNAL MODEL INPRESENCE OF
BOC MODULATION
At the transmitter, the data sequence is ﬁrst spread and the
pseudorandom (PRN) sequence is further BOCmodulated.
The BOC modulation is a square subcarrier modulation,
where the PRN signal is multiplied by a rectangular sub
carrier which has a frequency multiple of code frequency. A
BOCmodulated signal (sine or cosine) creates a split spec
trum with the two main lobes shifted symmetrically from the
carrier frequency by a value of the subcarrier frequency f
sc
[5].
The usual notation for BOC modulation is BOC( f
sc
, f
c
),
where f
c
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
n, satisfying the relationships m = f
sc
/ f
ref
and n = f
c
/ f
ref
,
where f
ref
= 1.023 MHz is the reference frequency [5, 24].
From the point of view of equivalent baseband signal, BOC
modulation can be deﬁned via a single parameter, denoted
by the BOCmodulation order N
BOC1
= 2m/n = 2 f
sc
/ f
c
. The
factor N
BOC1
is an integer number [25].
Examples of sine BOCmodulated waveforms for Sin
BOC(1, 1) (even BOCmodulation order N
BOC1
= 2) and
1
0
−1
0 1 2 3 4 5
PRN sequence (N
BOC1
= 1)
B
O
C

m
o
d
u
l
a
t
e
d
c
o
d
e
Chips
1
0
−1
0 1 2 3 4 5
N
BOC1
= 2
B
O
C

m
o
d
u
l
a
t
e
d
c
o
d
e
Chips
1
0
−1
0 1 2 3 4 5
N
BOC1
= 3
B
O
C

m
o
d
u
l
a
t
e
d
c
o
d
e
Chips
Figure 1: Examples of timedomain waveforms for sine BOC
modulated signals.
SinBOC(15, 10) (odd BOCmodulation order N
BOC1
= 3)
together with the original PRN sequence (N
BOC1
= 1) are
shown in Figure 1. In order to consider the cosine BOC
modulation case, a second BOCmodulation order N
BOC2
=
2 has been deﬁned in [25], in a way that the case of sine BOC
modulation corresponds to N
BOC2
= 1 and the case of cosine
BOC modulation corresponds to N
BOC2
= 2 (see the expres
sions of (1) to (4)). After spreading and BOC modulation,
the data sequence is oversampled with an oversampled factor
of N
s
, and this oversampling determines the desired accuracy
in the delay estimation process. Thus, the oversampling fac
tor N
s
represents the number of samples per BOC interval,
and one chip will consists of N
BOC1
N
BOC2
N
s
samples (i.e, the
chip period is T
c
= N
s
N
BOC1
N
BOC2
T
s
, where T
s
is the sam
pling rate).
The BOCmodulated signal s
n,BOC
(t) can be written, in
its most general form, as a convolution between a PRN se
quence s
PRN
(t) and a BOC waveform s
BOC
(t) [25]:
s
n,BOC
(t)
=
+∞
_
n=−∞
b
n
SF
_
k=1
(−1)
nNBOC
1
c
k,n
s
BOC
_
t −nT −kT
c
_
= s
BOC
(t) ⊗
+∞
_
n=−∞
SF
_
k=1
b
n
c
k,n
(−1)
nNBOC
1
δ
_
t −nT −kT
c
_
= s
BOC
(t) ⊗s
PRN
(t),
(1)
4 EURASIP Journal on Wireless Communications and Networking
where b
n
is the nth complex data symbol, T is the symbol
period (or code epoch length) (T = S
F
T
c
), c
k,n
is the kth
chip corresponding to the nth symbol, T
c
= 1/ f
c
is the chip
period, S
F
is the spreading factor (i.e., for GPS C/A signal
and Galileo OS signal, S
F
= 1023), δ(t) is the Dirac pulse,
⊗ is the convolution operator and s
PRN
(t) is the pseudo
random (PRN) code sequence (including data modulation)
of satellite of interest, and s
BOC
(·) is the BOCmodulated
signal (sine or cosine) whose expression is given in (2) to
(4). We remark that the term (−1)
nNBOC
1
is included to take
into account also odd BOCmodulation orders, similar with
[26]. The interference of other satellites is modeled as addi
tive white Gaussian noise, and, for clarity of notations, the
continuoustime model is employed here. However, the ex
tension to the discretetime model is straightforward and all
presented results are based on discretetime implementation.
The SinBOCCosBOCmodulated waveforms s
BOC
(t) are
deﬁned as in [5, 25]:
s
sin/ CosBOC
(t) =
⎧
⎪
⎪
⎪
⎨
⎪
⎪
⎪
⎩
sign
_
sin
_
N
BOC1
πt
T
c
__
for SinBOC,
sign
_
cos
_
N
BOC1
πt
T
c
__
for CosBOC,
(2)
respectively, that is, for SinBOCmodulation [25],
s
SinBOC
(t) =
NBOC
1
−1
_
i=0
(−1)
i
p
TB
1
_
t −i
T
c
N
BOC
_
, (3)
and for CosBOCmodulation [25],
s
CosBOC
(t) =
NBOC
1
−1
_
i=0
NBOC
2
−1
_
k=0
(−1)
i+k
× p
TB
_
t −i
T
c
N
BOC1
−k
T
c
N
BOC1
N
BOC2
_
.
(4)
In (3) and (4), p
TB
1
(·) is a rectangular pulse of sup
port T
c
/N
BOC1
and p
TB
(·) is a rectangular pulse of support
T
c
/N
BOC1
N
BOC2
. For example,
p
TB
(t) =
⎧
⎪
⎨
⎪
⎩
1 if 0 ≤ t <
T
c
N
BOC1
N
BOC2
,
0 otherwise.
(5)
We remark that the bandlimiting case can also be taken into
account, by setting p
TB
(·) to be equal to the pulse shaping
ﬁlter.
Some examples of the normalized power spectral den
sity (PSD), computed as in [25], for several sine and cosine
BOCmodulated signals, are shown in Figure 2. It can be ob
served that for evenmodulation orders such as SinBOC(1, 1)
or CosBOC(10, 5) (currently selected or proposed by Galileo
Signal Task Force), the spectrum is symmetrically split into
two parts, thus moving the signal energy away from DC fre
quency and thus allowing for less interference with the exist
ing GPS bands (i.e., the BPSK case). Also, it should be men
tioned that in case of an odd BOCmodulation order (i.e.,
−2 −1 0 1 2
−120
−100
−80
−60
−40
−20
0
Frequency (MHz)
BPSK
SinBOC (1, 1)
SinBOC (15, 10)
CosBOC (10, 5)
Examples of PSD for diﬀerent BOCmodulated signals
P
S
D
(
d
B
/
H
z
)
Figure 2: Examples of baseband PSD for BOCmodulated signals.
SinBOC(15, 10)), the interference around the DC frequency
is not completely suppressed.
The baseband model of the received signal r(t) via a fad
ing channel can be written as [25]
r(t) =
_
E
b
e
+j2π fDt
n=+∞
_
n=−∞
b
n
L
_
l=1
α
n,l
(t)
×s
n,sin/ CosBOC
_
t −τ
l
_
+ η(t),
(6)
where E
b
is the bit or symbol energy of signal (one symbol is
equivalent with a code epoch and typically has a duration of
T = 1 ms), f
D
is the Doppler shift introduced by channel, L
is the number of channel paths, α
n,l
is the timevarying com
plex fading coeﬃcient of the lth path during the nth code
epoch, τ
l
is the corresponding path delay (assuming to be
constant or slowly varying during the observation interval)
and η(·) is the additive noise component which incorporates
the additive white noise from the channel and the interfer
ence due to other satellites.
At the receiver, the codeDoppler acquisition and track
ing of the received signal (i.e., estimating the Doppler shift f
D
and the channel delay τ
l
) are based on the correlation with a
reference signal s
ref
(t− τ,
f
D
, n
1
), including the PRNcode and
the BOC modulation (here, n
1
is the considered symbol in
dex):
s
ref
_
t − τ,
f
D
, n
1
_
= e
−j2π
fDt
SF
_
k=−1
c
k,n1
NBOC
1
−1
_
i=0
NBOC
2
−1
_
j=0
(−1)
i+j
p
TB
_
t −n
1
T −kT
c
−i
T
c
N
BOC1
− j
T
c
N
BOC1
N
BOC2
− τ
_
.
(7)
Some examples of the absolute value of the ideal ACF for
several BOCmodulated PRN sequences, together with the
Adina Burian et al. 5
BPSK case, are illustrated in Figure 3. As it can be observed,
for any BOCmodulated signal, there are ambiguities within
the ±1 chips interval around the maximum peak.
After correlation, the signal is coherently averaged over
N
c
ms, with the maximum coherence integration length dic
tated by the coherence time of the channel, by possible resid
ual Doppler shift errors and by the stability of oscillators. If
the coherent integration time is higher than the coherence
time of the channel, the spectrum of the received signal will
be severely distorted. The Doppler shift due to satellite move
ment is estimated and removed before performing the coher
ent integration. For further noise reduction, the signal can be
noncoherently averaged over N
nc
blocks; however there are
some squaring losses in the signal power due to noncoher
ent averaging. The delay estimation is performed on a code
Doppler search space, whose values are averaged correlation
functions with diﬀerent time and frequency lags, with max
ima occurring at f = f
D
and τ = τ
l
.
3. EXISTINGDELAY ESTIMATIONALGORITHMS IN
MULTIPATHCHANNELS
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
mented via a feedback loop. These tracking loop methods are
based on the assumption that a coarse delay estimate is avail
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. Narrowearly 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
s
refEarly,Late
(t − τ ± Δ/2), where s
refEarly,Late
(·) is the advanced or
delayed BOCmodulated PRN code and τ is the tentative
delay estimate. The early (resp., late) branch correlations
R
early,Late
(·) can be written as
R
Early,Late
( τ) =
_
Nc
r(t)s
refEarly,Late
_
t − τ ±
Δ
2
_
dt. (8)
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 Scurve, which presents the expected value of
error as a function of code phase error. For NEML, the two
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
−1.5 −1 −0.5 0 0.5 1 1.5
N
o
r
m
a
l
i
z
e
d
A
C
F
s
Chips
Ideal ACF for BOCmodulated signals
BPSK
SinBOC (1, 1)
SinBOC (15, 10)
CosBOC (10, 5)
Figure 3: Examples of absolute value of the ACF for BOC
modulated signals.
branches are combined noncoherently, and the Scurve is ob
tained as in (9),
S
NEML
( τ) =
¸
¸
R
Late
( τ)
¸
¸
2
−R
Early
( τ)
¸
¸
2
. (9)
The error signal given by the Scurve 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 Scurve 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
Δ ≤ n/4m.
3.2. Highresolution correlator (HRC)
The highresolution correlator (HRC), introduced in [10],
can be obtained using multiple correlator outputs from con
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
1
0.8
0.6
0.4
0.2
0
−0.2
−0.4
−0.6
−0.8
−1
−1.5 −1 −0.5 0 0.5 1 1.5
N
o
r
m
a
l
i
z
e
d
S

c
u
r
v
e
Delay error (chips)
Ideal Scurve (no multipath) for
BOCmodulated and BPSK signals
BPSK
SinBOC (1, 1)
SinBOC (10, 5)
Figure 4: Ideal Scurves for BOCmodulated and BPSK signals
(NEML, Δ = 0.1 chips).
I & D on
N
c
msec
I & D on
N
c
msec
I & D on
N
c
msec
I & D on
N
c
msec
Late code
Early code
Very early code
Very late code
Constant factor a
NCO Loop ﬁlter
r(t)
+
−
+
+
+
−
 
2
 
2
 
2
 
2
Figure 5: Block diagram for HRC tracking loop.
early and, respectively, a very late branch. The Scurve for a
noncoherent ﬁvecorrelator HRC can be written as in [10]:
S
HRC
( τ) =
¸
¸
R
Late
( τ)
¸
¸
2
−
¸
¸
R
Early
( τ)
¸
¸
2
+ a
_¸
¸
R
VeryLate
( τ)
¸
¸
2
−
¸
¸
R
VeryEarly
( τ)
¸
¸
2
_
,
(10)
where R
VeryLate
(·) and R
VeryEarly
(·) are the very late and very
early correlations, with the spacing between them of 2Δ
chips, and a is a weighting factor which is typically −1/2 [10].
Examples of Scurves for HRC in the presence of a sin
gle path static channel, are shown in Figure 6, for two BOC
modulated signals. The early late spacing is Δ = 0.1 chips
(i.e., narrow correlator), thus the main lobes around zero
crossing are narrower, and it is more likely that the separa
tion between multiple paths will be done more easily.
1
0.8
0.6
0.4
0.2
0
−0.2
−0.4
−0.6
−0.8
−1
−1.5 −1 −0.5 0 0.5 1 1.5
N
o
r
m
a
l
i
z
e
d
S

c
u
r
v
e
Delay error (chips)
Ideal Scurve (no multipath) for two BOCmodulated signals
SinBOC (1, 1)
SinBOC (10, 5)
Figure 6: Ideal Scurves for noncoherent HRC with a = −1/2, for
two BOCmodulated signals and Δ = 0.1 chips.
3.3. Multipath estimating delay locked loop (MEDLL)
A diﬀerent approach, proposed to remove the multipath ef
fects for GPS C/A delay tracking is the multipath estima
tion delay locked l;oop [15]. The MEDLL method estimates
jointly the delays, phases, and amplitudes of all multipaths,
canceling the multipath interference. Since it is not based on
an Scurve, it can work in both feedback and feedforward
conﬁgurations. To the authors’ knowledge, the performance
of MEDLL algorithm for BOC modulated signals is still not
well understood, therefore, would be interesting to study a
similar approach. The steps of the MEDLL algorithm (as im
plemented by us) are summarized bellow.
(i) Calculate the correlation function R
n
(t) for the nth
transmitted code epoch. Find out the maximum peak
of the correlation function and the corresponding de
lay τ
1
, amplitude a
1,n
, and phase
θ
1,n
.
(ii) Subtract the contribution of the calculated peak, in or
der to have a new approximation of the correlation
function R
(1)
n (τ) = R
n
(τ) − a
1,n
R
ref
(t − τ
1,n
)e
j
θ1,n
. Here
R
ref
(·) is the reference correlation function, in the ab
sence of multipaths (which can be, for example, stored
at the receiver). Find out the new peak of the residual
function R
(1)
n (·) and its corresponding delay τ
2,n
, am
plitude a
2,n
, and phase
θ
2,n
. Subtract the contribution
of the new peak of residual function from R
(1)
n (t) and
ﬁnd a new estimate of the ﬁrst peak. For more than
two peaks, the procedure is continued until all desired
peaks are estimated.
(iii) The previous step is repeated until a certain criterion
of convergence is met, that is, when residual function
is below a threshold (e.g., set to 0.5 here) or until
Adina Burian et al. 7
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
−1.5 −1 −0.5 0 0.5 1 1.5
N
o
r
m
a
l
i
z
e
d
A
C
F
Delay error (chips)
Ideal ACFs (no multipath) for SinBOC (1, 1)modulated signal
Noncoherent integration
Diﬀerential correlation
Figure 7: Envelope correlation function of traditional noncoher
ent integration and diﬀerential correlation for a SinBOC(1, 1)
modulated signal.
the moment when introducing a new delay does not
improve the performance in the sense of root mean
square error between the original correlation function
and the estimated correlation function.
3.4. Differential correlation (DC)
Originally proposed for CDMAbased wireless communi
cation systems, the diﬀerential correlation method has also
been investigated in context of GPS navigation system[22]. It
has been observed that with low and medium coherent times
of the fading channel and in absence of any frequency error,
this approach provides better resistance to noise than the tra
ditional noncoherent integration methods. In DC method,
the correlation is performed between two consecutive out
puts of coherent integration. These correlation variables are
then integrated, in order to obtain a diﬀerential variable. The
diﬀerential detection variable z is given as
z
DC
=
1
M−1
M−1
_
k=1
¸
¸
y
∗
k
y
k+1
¸
¸
2
, (11)
where y
k
, k = 1, . . . , M are the outputs of the coherent in
tegration and M is the diﬀerential integration length. For a
fair comparison between the diﬀerential noncoherent and
traditional noncoherent methods, here it is assumed that
M = N
nc
, where N
nc
is the noncoherent integration length.
Since the diﬀerential coherent correlation method was no
ticed to be more sensitive to residual Doppler errors, only
the diﬀerential noncoherent correlation is considered here.
The analysis done in [22] is limited to BPSK modulation.
From Figure 7, it can be noticed that applying the DC to a
BOCmodulated signal, instead of the conventional nonco
herent integration, the sidelobes envelope can be decreased,
and thus this method has a potential in reducing the side
peaks ambiguities.
3.5. Nonambiguous BOC(n, n) signal tracking
(Julien&al. method)
A recent tracking approach, which removes the sidelobes
ambiguities of SinBOC(n, n) signals and oﬀers an improved
resistance to longdelay multipath, has been introduced in
[20]. This method, referred here as Julien&al. method, af
ter the name of the ﬁrst author in [20], has emerged while
observing the ACF of a SinBOC(1, 1) signal with sine phas
ing, and the cross correlation of SinBOC(1, 1) signal with its
spreading sequence. The ideal correlation function R
ideal
BOC
(·)
for SinBOC(1, 1)modulated signals in the absence of multi
paths, can be written as [25]
R
ideal
BOC
(τ) = Λ
Tc /2
(τ) −
1
2
Λ
Tc /2
_
τ −
T
c
2
_
−
1
2
Λ
Tc /2
_
τ +
T
c
2
_
,
(12)
where Λ
Tc /2
(τ −α) is the value in τ of a triangular function
1
centered in α, with a width of 1chip, T
c
is the chip period,
and τ is the code delay in chips.
The cross correlation of a SinBOC(1, 1) signal with the
spreading pseudorandom code, for an ideal case (no multi
paths and ideal PRN code), can be expressed as [20]
R
ideal
BOC,PRN
(τ) =
1
2
_
Λ
Tc /2
_
τ +
T
c
2
_
+ Λ
Tc /2
_
τ −
T
c
2
__
.
(13)
Two types of DLL discriminators have been considered
in [20], namely, the earlyminus late power (EMLP) dis
criminator and the dotproduct (DP) discriminator. These
examples of possible discriminators result from the use of
the combination of BOCautocorrelation function and of
the BOC/PRNcorrelation function [20]. Based on (12) and
(13), the ideal EMLP discriminator is constructed, as in (14),
where τ is the code tracking error [20]:
S
ideal
EMLP
(τ) =
_
R
ideal
2
BOC
_
τ +
Δ
2
_
−R
ideal
2
BOC
_
τ −
Δ
2
__
−
_
R
ideal
2
BOC,PRN
_
τ +
Δ
2
_
−R
ideal
2
BOC,PRN
_
τ −
Δ
2
__
.
(14)
The alternative DP discriminator variant [20] does not
have a linear variation as a function of code tracking error:
S
ideal
DP
(τ)
=
_
R
ideal
2
BOC
_
τ +
Δ
2
_
−R
ideal
2
BOC
_
τ −
Δ
2
__
R
ideal
2
BOC
(τ)
−
_
R
ideal
2
BOC,PRN
_
τ +
Δ
2
_
−R
ideal
2
BOC,PRN
_
τ −
Δ
2
__
R
ideal
2
BOC
(τ).
(15)
1
Our notation is equivalent with the notation tri
α
(x/ y) used in [20], via
tri
α
(τ/ y) = Λ
Tc /2
(τ −αT
c
/ y).
8 EURASIP Journal on Wireless Communications and Networking
1
0.5
0
−1.5 −1 −0.5 0 0.5 1 1.5
SinBOC (1, 1) modulation, ACFs of
BOCmodulated and subtracted signals
Continue line:
BOCmodulated signal
Dashed line:
subtracted signal
Delay (chips)
1
0.5
0
−1.5 −1 −0.5 0 0.5 1 1.5
SinBOC (1, 1) modulation, ACF of unambiguous signal
Unambiguous signal
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
singlepath 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 R
ideal
BOC
(τ) have the same magnitude
and same location as the two peaks of SinBOC(1, 1)/PRN
correlation function R
ideal
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 longdelay multipath.
In practice, the discriminators S
EMLP
(τ) or S
DP
(τ), as
given in [20], are formed via continuous computation, at re
ceiver side, of correlation functions R
BOC
(·) and R
BOC,PRN
(·)
values, not on the ideal ones. In practice, R
BOC
(·) is the
correlation between the incoming signal (in the presence of
multipaths) and the reference BOCmodulated code, and
R
BOC,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 R
ideal
BOC,PRN
(·) (which can be computed only once and
stored at the receiver side), the correlation R
BOC,PRN
(·) needs
to be computed for each code epoch in [20]. Of course, in or
der to make use of the R
ideal
BOC,PRN
(·) shape, we also need some
information about channel multipath proﬁle. This will be ex
plained in the next section.
4. SIDELOBES CANCELLATIONMETHOD(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
BOCmodulated signals. In contrast with the Julien&al.
method, which is restricted to the SinBOC(n, n) case, we
will show here how to use SCM with any sine or cosine
BOCmodulated 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
ceived BOCmodulated signal with the reference PRN code.
In the Julien&al. method, the subtraction function, which
approximates the sidelobes, is provided by crosscorrelating
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 timeconsuming 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 SCMmethod
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:
R
ideal
sub
(τ) =
NBOC
1
−1
_
i=0
NBOC
1
−1
_
j=0
NBOC
2
−1
_
k=0
NBOC
2
−1
_
l=0
(−1)
i×j+k+l
Λ
TB
_
τ + (i − j)T
B
+ (k −l)
T
B
N
BOC2
_
,
(16)
where T
B
= T
c
/N
BOC1
N
BOC2
is the BOC interval, Λ
TB
(·)
is the triangular function centered at 0 and with a width
of 2T
B
chips, N
BOC1
is the sine BOCmodulation order
(e.g., N
BOC1
= 2 for SinBOC(1, 1), or N
BOC1
= 4
for SinBOC(10, 5)) [25], and N
BOC2
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,
N
BOC2
= 1 and, if cosine BOC modulation is employed,
N
BOC2
= 2).
As an example, the simplest case of SinBOC(1, 1)
modulation (i.e., the main choice for Open Services in
Galileo), (16) becomes
R
ideal
sub,SinBOC(1,1)
(τ) =
_
Λ
TB
_
τ −T
B
_
+ Λ
TB
_
τ + T
B
__
, (17)
Adina Burian et al. 9
which is similar with Julien& al. expression of (13) with the
exception of a 1/2 factor (here, T
B
= T
c
/2).
The Sin and CosBOC(m, n)based ideal autocorrelation
function can be written as [25]
R
ideal
BOC
(τ) =
NBOC
1
−1
_
i=0
NBOC
1
−1
_
j=0
NBOC
2
−1
_
k=0
NBOC
2
−1
_
l=0
(−1)
i+j+k+l
Λ
TB
_
τ + (i − j)T
B
+ (k −l)
T
B
N
BOC2
_
.
(18)
Again, for SinBOC(1, 1) case, the expression of (18) reduces
to
R
ideal
SinBOC(1,1)
(τ)
=
_
2Λ
TB
(τ) −Λ
TB
_
τ −T
BOC
_
−Λ
TB
_
τ + T
BOC
__
,
(19)
which is, again, similar to Julien& al. expression of (12) with
the exception of a 1/2 factor (for SinBOC(1, 1), T
BOC
= T
c
/2,
N
BOC1
= 2 and N
BOC2
= 1).
We remark that the diﬀerence between (16) and (18)
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
fect of sidelobes (16) from the overall correlation (18), after
normalizing them properly.
Thus, in order to obtain an unambiguous ACF shape, the
squared function (R
ideal
sin
(·))
2
, (R
ideal
cos
(·))
2
, respectively, has to
be subtracted from the ambiguous squared correlation func
tion as shown in
R
ideal
unamb
(τ) =
_
R
ideal
BOC
(τ)
_
2
−w
_
R
ideal
sin/ cos
(τ)
_
2
, (20)
where w < 1 is a weight factor used to normalize the reference
function (to achieve a magnitude of 1).
For example, for SinBOC(1, 1) and w = 1, we get from
(17), (19), and (20), after straightforward computations, that
R
ideal
unamb
(τ) = 4
_
Λ
2
TB
(τ) −Λ
TB
(τ)Λ
TB
_
τ −T
BOC
_
−Λ
TB
(τ)Λ
TB
_
τ + T
BOC
__
,
(21)
and if we plot R
ideal
unamb
(τ) (e.g., see the lower plot of Figure 8),
we get a main narrow correlation peak, without sidelobes.
All the derivations so far were based on ideal assumptions
(ideal correlation codes, single path static channels, etc.).
However, in practice, we have to cope with the real signals,
so the ideal autocorrelation function R
ideal
BOC
(τ) should be re
placed with the computed correlation R
BOC
(τ) between the
received signal and the reference BOCmodulated pseudo
random code. Thus, (20) becomes
R
unamb
(τ) =
_
R
BOC
(τ)
_
2
−w
_
R
ideal
sin/ cos
(τ)
_
2
. (22)
Here comes into equation the weighting factor, since vari
ous channel eﬀects (such as noise and multipath) can mod
ify the levels of R
BOC
(τ) function. In order to perform the
1
0.5
0
−1.5 −1 −0.5 0 0.5 1 1.5
CosBOC (10, 5) modulation, ACFs of
BOCmodulated and subtracted signals
Continue line:
BOCmodulated signal
Dashed line:
subtracted signal
Delay (chips)
1
0.5
0
−1.5 −1 −0.5 0 0.5 1 1.5
CosBOC (10, 5) modulation, ACF of unambiguous signal
Unambiguous signal
Delay (chips)
Figure 9: CosBOC(10, 5)modulated signal: examples of the am
biguous correlation function and subtracted pulse (upper plot)
and obtained unambiguous correlation function (lower plot), in a
singlepath channel.
normalization of reference function (i.e., to ﬁnd the weight
factors w), the peaks magnitudes of R
BOC
(·) function are ﬁrst
found out and sorted in increased order. Then the weighting
factor w is computed as the ratio between the lastbutone
peak and the highest peak. We remark that the above algo
rithm does not require the computation of the BOC/PRN
correlation anymore, it only requires the computation of
R
BOC
(τ) = R
n
(τ) correlation. The pulses to be subtracted are
always based on the ideal functions R
ideal
sin/ cos
(τ), and therefore,
they can be computed only once (via (16)) and stored at the
receiver (in order to decrease the complexity of the tracking
unit).
By comparison with Julien&al. method, here the num
ber of correlations at the receiver is reduced by half (i.e.,
R
BOC,PRN
(·) computation is not needed anymore). Thus the
SCM technique oﬀers less computational burden (only one
correlation channel in contrast to Julien&al. method, which
uses two correlation channels).
Figures 8 and 9 show the shapes of the ideal ambigu
ous correlation functions and of the subtracted pulses, to
gether with the correlation functions, obtained after subtrac
tion (SCM method). Figure 8 exempliﬁes a SinBOC(1, 1)
modulated signal, while Figure 9 illustrates the shapes for a
CosBOC(10, 5)modulation case. As it can be observed, for
both SinBOC and CosBOC modulations, the subtractions
removes the sidelobes closest to the main peak, which are
the main threats in the tracking process. Also, it should be
mentioned that the Figure 8, for a SinBOC(1, 1) modulated
signal, is also illustrative for the Julien&al. method, since the
shapes of correlation functions are similar with those pre
sented in [20].
Equation (20) is valid for single path channels. How
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,
one important issue is to align the subtraction pulse to the
LOS peak (otherwise, the subtraction of (22) will not can
cel the correct sidelobes). This can be done only if some ini
tial estimate of LOS delay is obtained. For this purpose, we
employ and compare several feedback loops or feedforward
algorithms, as it will be explained next.
4.2. SCMwith interference cancellation (IC)
Combining the multipath eliminating DLL concept with the
SCM method, we obtain an improved SCM technique with
multipath interference cancellation (SCM with IC). In this
method, the initial estimate of LOS delay is obtained via
MEDLL algorithm. The sidelobe cancellation is applied in
side the iterative steps of MEDLL, as explained below.
(1) Calculate the correlation function R
n
(τ) between the
received signal and the reference BOCmodulated
code (e.g., see the continuous line, Figure 10, up
per plot). Find the global maximum peak (the peak
1) of this correlation function, max
τ
R
n
(τ), and its
corresponding delay, τ
1,n
, amplitude a
1,n
and phase
θ
1,n
(e.g., the peak situated at the 50thsample delay,
Figure 10, upper plot).
(2) Compute the ideal reference function centered at τ
1,n
:
R
ideal
sub
(τ − τ
1,n
) via (16) (see the dashed line, Figure 10,
upper plot).
(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
timated CIR of peak 1, Figure 10, upper plot).
(4) In order to remove the sidelobes ambiguities, the
function R
ideal
sub
(τ − τ
1,n
) is then subtracted from the
multipath correlation function R
n
(τ) and an unam
biguous shape is obtained, using (22), or, equiva
lently R
n,unamb
(τ) = (R
n
(τ))
2
− (R
ideal
sub
(τ − τ
1,n
))
2
. In
Figure 10, the unambiguous ACF R
n,unamb
(·) is plot
ted with dasheddotted line, in both upper and lower
plots.
(5) Cancel out the contribution of the strongest path
and obtain the residual function R
(1)
n,unamb
(τ) =
R
n,unamb
(τ) − a
1,n
R
ideal
unamb
(τ)(τ − τ
1,n
)e
j
θ1,n
, where
R
ideal
unmab
(τ) is the unambiguous reference function
given by (20). The shape of residual function is
exempliﬁed in Figure 10, lower plot (drawn with
continuous line).
(6) The new maximum peak of the residual function
R
(1)
n,unamb
is found out (e.g., at 44thsample delay,
Figure 10, lower plot), with its corresponding de
lay τ
2,n
, amplitude a
2,n
and phase
θ
2,n
. The con
tributions of both peaks 1 and 2 are subtracted
from unambiguous correlation function R
n,unamb
(τ)
1
0.8
0.6
0.4
0.2
0
0 10 20 30 40 50 60 70 80
Samples
Exempliﬁcation of SCM IC method (steps 1 to 4)
Original ACF
Estimated CIR
Subtracted ideal function
Unambiguous ACF
1
0.8
0.6
0.4
0.2
−0.2
0
0 10 20 30 40 50 60 70 80
Samples
Exempliﬁcation of SCM IC method (steps 5 to 6)
Unambiguous ACF
Residual function
Estimated CIR, 2nd peak
Figure 10: Exempliﬁcation of SCM IC method, 2paths fading
channel with true channel delay at 44 and 50 samples, average path
powers [−2, 0] dB, SinBOC(1, 1)modulated signal.
and the maximum global peak is reestimated from
R
(2)
n,unamb
(τ) = (R
n,unamb
(τ))
2
− ( a
1,n
R
ideal
unamb
(τ)(τ −
τ
1,n
)e
j
θ1,n
+ a
2,n
R
ideal
unamb
(τ)(τ − τ
2,n
)e
j
θ2,n
)
2
.
(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.
These steps of SCM IC method are illustrated in
Figure 10, for 2path fading channel.
Adina Burian et al. 11
1
0.8
0.6
0.4
0.2
0
−0.2
−0.4
−0.6
−0.8
−1
−1.5 −1 −0.5 0 0.5 1 1.5
N
o
r
m
a
l
i
z
e
d
S

c
u
r
v
e
Delay error (chips)
Ideal Scurve (no multipath), SCM NEML method
SinBOC (1, 1)
SinBOC (10, 5)
Figure 11: SCM NEML method: ideal Scurves (no multipath), for
two BOCmodulation cases and Δ = 0.1 chips.
4.3. SCMusing narrowearly minus lat discriminator
(SCMNEML)
After obtaining an unambiguous correlation function
R
n,unamb
(τ) (as it was shown in the previous section, steps
(1) to (4)), a NEML Scurve is constructed, by forming the
early, respectively, late branches, spaced at Δ = 0.1 chips. The
Scurve is obtained in the same way as in Section 3.1, by sub
tracting the late and early branches of unambiguous correla
tion function,
S
SCMNEML
( τ) =
¸
¸
R
Late
n,unamb
( τ)
¸
¸
2
−
¸
¸
R
Early
n,unamb
( τ)
¸
¸
2
. (23)
Examples of Scurves obtained with this method, in
presence of a single path static channel, are presented in
Figure 11, for two BOCmodulated signals, SinBOC(1, 1)
and SinBOC(10, 5), and a spacing of Δ = 0.1 chips. Com
paring with Figure 4, which presents the NEML Scurves for
ambiguous signals, in Figure 11, the possibility to detect an
incorrect zero crossing, due to sidelobes peaks, is decreased.
A typical measure of performance for the ability of a de
lay tracking loop to deal with multipath error is the socalled
multipath error envelope (MEE) [9, 10]. The MEE is usu
ally computed for one direct and one reﬂected channel paths,
with a certain variable spacing. The multipath errors are cal
culated for the worstcase scenario, when the two paths are
added inphase (upper MEE) and have equal strength, and
also, when the two paths are out of phase (lower MEE). Com
parisons of MEEs plots, for both NEML and SCM NEML
methods, are shown in Figure 12, for two BOCmodulated
signals. A static channel with two paths of equal amplitudes
and variable spacing was considered. The only interference
considered here is the multipath interference, and the addi
tive white noise eﬀect is not taken into account. As it can be
seen in Figure 12, comparing with the NEML correlator, the
10
0
−10
0 0.2 0.4 0.6 0.8 1
M
u
l
t
i
p
a
t
h
e
r
r
o
r
e
n
v
e
l
o
p
e
(
m
e
t
e
r
s
)
SinBOC (1, 1), Δ = 0.1 chips
Multipath spacing (chips)
NEML correlator
SCM NEML method
10
0
−10
0 0.2 0.4 0.6 0.8 1
M
u
l
t
i
p
a
t
h
e
r
r
o
r
e
n
v
e
l
o
p
e
(
m
e
t
e
r
s
)
SinBOC (10, 5), Δ = 0.1 chips
Multipath spacing (chips)
NEML correlator
SCM NEML method
Figure 12: Multipath error envelopes (in meters): NEML correlator
versus SCM NEML method, for two BOCmodulation cases and
Δ = 0.1 chips.
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 zerocrossing 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
the noncoherent NEML.
4.4. SCMusing highresolution correlator
discriminator (SCMHRC)
In a similar manner as in previous section, the SCM method
can be also used in conjunction with an HRC discrimina
tor, after removing the side peaks threats and obtaining an
unambiguous correlation function R
n,unamb
(τ). Based on this
unambiguous function, an HRCScurve is constructed, in an
analogous way as in Section 3.2:
S
SCMHRC
( τ) =
¸
¸
R
Late
n,unamb
( τ)
¸
¸
2
−
¸
¸
R
Early
n,unamb
( τ)
¸
¸
2
+ a
_
¸
¸
R
VeryLate
n,unamb
( τ)
¸
¸
2
−
¸
¸
R
VeryEarly
n,unamb
( τ)
¸
¸
2
_
,
(24)
where R
Early
n,unamb
(·) and R
Late
n,unamb
(·) are the advanced and de
layed unambiguous correlations, with a spacing between
them of Δ = 0.1 chips. The R
VeryEarly
n,unamb
(·), respectively,
R
VeryLate
n,unamb
(·) are the very early and the very late unambiguous
correlation branches, spaced at 2Δ chips and the weighting
factor a = −1/2.
12 EURASIP Journal on Wireless Communications and Networking
1
0.8
0.6
0.4
0.2
0
−0.2
−0.4
−0.6
−0.8
−1
−1.5 −1 −0.5 0 0.5 1 1.5
N
o
r
m
a
l
i
z
e
d
S

c
u
r
v
e
Delay error (chips)
Ideal Scurve (no multipath), SCM HRC method
SinBOC (1, 1)
SinBOC (10, 5)
Figure 13: SCM HRC method: ideal Scurves (no multipath), for
two BOCmodulation cases, with a = −1/2 and Δ = 0.1 chips.
10
5
0
−5
−10
0 0.2 0.4 0.6 0.8 1
M
u
l
t
i
p
a
t
h
e
r
r
o
r
e
n
v
e
l
o
p
e
(
m
e
t
e
r
s
)
SinBOC (1, 1), Δ = 0.1 chips
Multipath spacing (chips)
HRC method
SCM HRC method
10
5
0
−5
−10
0 0.2 0.4 0.6 0.8 1
M
u
l
t
i
p
a
t
h
e
r
r
o
r
e
n
v
e
l
o
p
e
(
m
e
t
e
r
s
)
SinBOC (10, 5), Δ = 0.1 chips
Multipath spacing (chips)
HRC method
SCM HRC method
Figure 14: Multipath error envelopes (in meters): HRC method
versus SCM HRC method, for two BOCmodulation cases and
Δ = 0.1 chips.
The ideal Scurves obtained with the SCM HRC method,
for two BOCmodulation orders, are presented in Figure 13.
The MEEs performances, for both the HRC and SCM HRC
methods, are illustrated in Figure 14, for SinBOC(1, 1) and
0.8
0.6
0.4
0.2
0
−0.2
−1.5 −1 −0.5 0 0.5 1 1.5
N
o
r
m
a
l
i
z
e
d
A
C
F
Delay error (chips)
Ideal ACF (no multipath) for SinBOC (10, 5) modulated signal
Ambiguous correlation
Diﬀerential correlation
SCM method
SCM DC method
Figure 15: Envelopes of correlation functions obtained with am
biguous correlation, DC method, SCM approach, and SCM DC
method, for a SinBOC(10, 5)modulated signal.
SinBOC(10, 5) cases. As it can be noticed, there is a slight im
provement brought by the SCM HRC method over the HRC
correlator.
4.5. SCMusing differential correlation (DC) in
conjunction with feedback and feedforward
tracking algorithms
It has been observed that the DC method has potential to de
crease the sidelobes amplitudes, thus lowering the possibility
to detect a wrong side peak. To enhance the performance of
the DCmethod, we use it in conjunction with diﬀerent track
ing algorithms, such as NEML or HRC methods, or with IC
method. These algorithms are applied in similar ways as ex
plained in Sections 3.1, 3.2, and 3.3, on the correlation func
tions obtained after performing the noncoherent DC tech
nique (Section 3.4).
Also, the performance may be enhanced further, by us
ing the SCM approach after applying the DC method. This is
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
biguous correlation function between the multipath received
signal and the reference BOCmodulated code. Indeed, as il
lustrated in Figure 15, in case of a SinBOC(10, 5) modulated
signal, the combination of DC and SCM algorithms can de
crease even further the sidelobes amplitudes, thus eliminat
ing more ambiguities.
4.6. SCMwith threshold comparison (SCMthr)
Another approach is to test the performance of SCM tech
nique using a thresholding algorithm. Starting from the un
ambiguous correlation function R
n,unamb
(τ), an estimate of
noise variance σ
2
n
is obtained, as the mean of the squares of
Adina Burian et al. 13
the outofpeak values, similar to [4]. Using this estimated
noise variance, a linear threshold γ is computed, based on the
second peak γ
2
of the ideal unambiguous correlation func
tion R
ideal
unamb
(τ) (i.e., for SinBOC(1, 1) γ
2
= 0.5, as seen in
Figure 3), together with the estimate of the noise variance σ
2
n
:
γ = γ
2
+
_
σ
2
n
. (25)
Then the LOS delay is estimated, based on the unambigu
ous correlation function R
n,unamb
(τ), using this threshold. If
the peak of the estimated ﬁrst path is too low (i.e., ten times
lower than the global peak), then this path is discarded and
the next estimate is considered.
5. SIMULATIONRESULTS
5.1. Additive white noise Gaussian (AWGN) channel
We ﬁrst test the performance of the proposed algorithms in
the ideal AWGN channel (single path), in order to check
whether SCM algorithm introduces a deterioration with re
spect to the standard narrow and highresolution 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
time to lose lock (MTLL). The simulations were carried out
in Matlab. The MTLL is computed as the average value for
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
N
c
= 20 milliseconds and the noncoherent integration is per
formed over N
nc
= 3 blocks (i.e., the total coherent and non
coherent integration length is 60 milliseconds), and the over
sampling factor is set to N
s
= 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 5000N
c
N
nc
= 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, SCMalgorithmdoes not
deteriorate the performance in AWGN case, compared with
narrow and highresolution 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 AWGNchannel (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
R
M
S
E
(
c
h
i
p
s
)
SinBOC (1, 1), AWGN singlepath channel
20 25 30 35 40
10
−6
10
−5
10
−4
10
−3
10
−2
10
−1
10
0
CNR (dBHz)
NEML
Julien & al. EMLP
SCM NEML
DC NEML
DC SCM NEML
M
T
L
L
(
s
)
SinBOC (1, 1), AWGN singlepath channel
20 25 30 35 40
10
2.4
10
2.3
10
2.2
CNR (dBHz)
NEML
Julien & al. EMLP
SCM NEML
DC NEML
DC 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
BOCmodulated 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 Mcode. In order to have fair
14 EURASIP Journal on Wireless Communications and Networking
R
M
S
E
(
c
h
i
p
s
)
SinBOC (1, 1), AWGN singlepath channel
20 25 30 35 40
10
−5
10
−4
10
−3
10
−2
10
−1
10
0
CNR (dBHz)
HRC
Julien & al. DP
SCM HRC
DC HRC
SCM DC HRC
M
T
L
L
S
(
s
)
SinBOC (1, 1), AWGN singlepath channel
20 22 24 26 28 30 32 34 36 38
100
200
350
CNR (dBHz)
HRC
Julien & al. DP
SCM HRC
DC HRC
SCM DC 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 SCMHRCcurves are overlapping; DCHRCand DCSCMHRC
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 BOCmodulation 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
to v = 3 km/h, while for outdoor proﬁles, the mobile speeds
of 25, 45, or 75 km/h have been selected). Two main chan
nel proﬁles have been considered: either with ﬁxed Rayleigh
distribution of all paths and with average path power of −1,
−2, 0 and −3 dB, or a 2paths decaying power delay proﬁle
(PDP) channel, with Rician distributions for the ﬁrst path
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
a slope of 0.05 chips per block of N
c
N
nc
millisecond, thus the
tracking algorithms should capture this linear delay increase.
The successive channel path delays have a random spacing
with respect to the precedent delay, uniformly distributed be
tween 1/(N
s
N
BOC1
N
BOC2
) and x
max
, where x
max
(in chips) is
the maximum separation between successive paths (i.e., for
closedspaced paths scenario, x
max
= 0.1 chips). In order to
have independent and reliable results for each method, the
search interval is diﬀerent for each algorithm. which means
that once the lock is lost for one method, this will not aﬀect
the other algorithms. The search window has few chips (typ
ically between 4 and 12 chips), depending on the number
of paths, the distance between them and on the used BOC
modulation orders. The search window is sliding around the
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
set to N
c
= 20 milliseconds, the noncoherent integration is
performed over N
nc
= 3 blocks, and the oversampling factor
is set to N
s
= 11.
The SCM approach is exempliﬁed in Figure 18, for a
Rayleigh 2paths fading channel, with equal PDP. The up
per plot exempliﬁes a SinBOC(1, 1) modulation case, with
x
max
= 1 chip, while the lower plot shows the original ACF,
together with subtracted pulse and unambiguous shape, for
a SinBOC(10, 5) case and x
max
= 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
lay of ﬁrst path, situated at the 70th sample (in one chip, there
are N
s
N
BOC1
N
BOC2
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 closelyspaced paths (i.e., x
max
= 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
0.8
0.6
0.4
0.2
0
−0.2
40 60 80 100 120
A
C
F
s
Delay error (samples)
SinBOC (1, 1), Rayleigh fading channel
with 2 paths, x
max
= 1 chip
Ambiguous ACF
Subtracted pulse
Unambiguous ACF
1st path true delay =
70 samples
0.8
0.6
0.4
0.2
0
−0.2
180 200 220 240 260 280 300 320
A
C
F
s
Delay error (samples)
SinBOC (10, 5), Rayleigh fading channel
with 2 paths, x
max
= 0.5 chips
Ambiguous ACF
Subtracted pulse
Unambiguous ACF
True delay =
239 samples
1
Figure 18: Exempliﬁcation of SCM method for a 2paths Rayleigh
fading channel. Upper plot: SinBOC(1, 1)modulated signal and
x
max
= 1 chip. Lower plot: SinBOC(10, 5)modulated signal and
x
max
= 0.5 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
R
M
S
E
(
c
h
i
p
s
)
SinBOC (1, 1), Rayleigh channel,
speed mobile = 75 km/h, x
max
= 0.1 chips
20 25 30 35 40
10
−0.6
10
−0.5
10
−0.4
10
−0.3
CNR (dBHz)
NEML
Julien & al. EMLP
SCM NEML
DC NEML
SCM DC NEML
M
T
L
L
(
s
)
SinBOC (1, 1), Rayleigh channel, 4 paths, x
max
= 0.1 chips
20 25 30 35 40
10
2
CNR (dBHz)
NEML
Julien & al. EMLP
SCM NEML
DC NEML
SCM DC NEML
Figure 19: Comparison of feedback delays estimation algorithms
employing the NEML discriminator and of the Julien&al. method,
as a function of CNR; SinBOC(1, 1) modulation, Rayleigh channel
with an average pathspower delay proﬁle of −1, −2, 0, and −3 dB,
v = 75 km/h, closely spaced paths with x
max
= 0.1 chips.
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 2path channel with Rician
distribution for the ﬁrst path, a mobile speed of 25 km/h and
a large separation between successive paths x
max
= 1 chip.
Figure 21 presents the case of a CosBOC(10, 5)modulated
16 EURASIP Journal on Wireless Communications and Networking
R
M
S
E
(
c
h
i
p
s
)
SinBOC (1, 1), 2paths Rician channel,
x
max
= 1 chip, mobile speed = 25 km/h
20 25 30 35 40
10
−0.5
10
−0.4
10
−0.3
CNR (dBHz)
HRC
Julien & al. DP
SCM HRC
DC HRC
SCM DC HRC
M
T
L
L
(
s
)
SinBOC (1, 1), 2paths Rician channel,
x
max
= 1 chip, mobile speed = 25 km/h
20 25 30 35 40
10
1
CNR (dBHz)
HRC
Julien & al. DP
SCM HRC
DC HRC
SCM DC HRC
Figure 20: Comparison of feedback delays estimation algorithms
employing the HRC discriminator and of the Julien&al. method, as
a function of CNR; SinBOC(1, 1) modulation, 2paths Rician chan
nel with decaying PDP of 0 and −2 dB, v = 25 km/h, maximum
separation between paths x
max
= 1 chip.
signal, for a 4paths Rayleigh channel, with closely spaced
paths x
max
= 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
R
M
S
E
(
c
h
i
p
s
)
CosBOC (10, 5), Rayleigh channel,
x
max
= 0.1 chips, mobile speed = 45 km/h
20 25 30 35 40
10
−0.6
10
−0.7
10
−0.5
10
−0.4
10
−0.3
CNR (dBHz)
HRC
Julien & al. DP
SCM HRC
DC HRC
SCM DC HRC
M
T
L
L
(
s
)
CosBOC (10, 5), 4paths Rayleigh channel,
x
max
= 0.1 chips
20 25 30 35 40
10
1
CNR (dBHz)
HRC
Julien & al. DP
SCM HRC
DC HRC
SCM DC HRC
Figure 21: Comparison of feedback delays estimation algorithms
employing the HRC discriminator and of the Julien&al. method, as
a function of CNR; CosBOC(10, 5) modulation, 4paths Rayleigh
channel, with paths PDP of −1, −2, 0, and −3 dB, v = 45 km/h,
closely spaced paths x
max
= 0.1 chips.
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
R
M
S
E
(
c
h
i
p
s
)
SinBOC (1, 1), Rayleigh channel,
speed mobile = 3 km/h, x
max
= 0.1 chips
20 25 30 35 40
10
−0.7
10
−0.9
10
−0.5
10
−0.3
CNR (dBHz)
MEDLL
SCM IC
SCM thr.
DC IC
SCM DC IC
M
T
L
L
(
s
)
SinBOC (1, 1), Rayleigh channel,
4 paths, x
max
= 0.1 chips
20 25 30 35 40
10
0
10
1
CNR (dBHz)
MEDLL
SCM IC
SCM thr.
DC IC
SCM DC IC
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, 4paths indoor Rayleigh channel, with PDP of −1, −2, 0,
and −3 dB, v = 3 km/h, closely spaced paths with x
max
= 0.1 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 (x
max
= 0.1 chips, v =
3 km/h). In Figure 23, the signal is also SinBOC(1, 1) modu
lated, the channel is 2paths with Rician distribution on ﬁrst
path, v = 45 km/h and x
max
= 0.5 chips.
R
M
S
E
(
c
h
i
p
s
)
SinBOC (1, 1), 2paths Rician channel,
x
max
= 0.5 chips, mobile speed = 45 km/h
20 25 30 35 40
10
−1
10
0
CNR (dBHz)
MEDLL
SCM IC
SCM thr.
DC IC
SCM DC IC
M
T
L
L
(
s
)
SinBOC (1, 1), 2paths Rician channel,
x
max
= 0.5 chips, mobile speed = 45 km/h
20 25 30 35 40
10
1
10
2
10
0
CNR (dBHz)
MEDLL
SCM IC
SCM thr.
DC IC
SCM DC IC
Figure 23: Comparison of feedforward delays estimation algo
rithms employing the MEDLL and ICmethods and of the SCMwith
threshold approach, as a function of CNR; SinBOC(1, 1) modula
tion, 2paths decaying PDP Rician channel, v = 45 km/h, x
max
= 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 closelyspaced paths (Figure 22, upper
plot), the performances of the SCMICand DCICalgorithms
are exceeded by those of SCM DC IC and SCM thresholding
methods, for a CNR range from 20 to 30 dBHz. In case of
Figure 23, for a higher spacing between successive paths up
18 EURASIP Journal on Wireless Communications and Networking
R
M
S
E
(
c
h
i
p
s
)
CosBOC (10, 5), Rayleigh channel,
speed mobile = 3 km/h, x
max
= 0.1 chips
20 25 30 35 40
10
−0.6
10
−0.8
10
−0.7
10
−0.5
10
−0.4
10
−0.3
CNR (dBHz)
MEDLL
SCM IC
SCM thr.
DC IC
SCM DC IC
M
T
L
L
(
s
)
CosBOC (10, 5), 4paths Rayleigh channel,
x
max
= 0.1 chips
20 25 30 35 40
10
1
10
2
10
0
CNR (dBHz)
MEDLL
SCM IC
SCM thr.
DC IC
SCM DC IC
Figure 24: Comparison of feedforward delays estimation algo
rithms employing the MEDLL and ICmethods and of the SCMwith
threshold approach, as a function of CNR; CosBOC(10, 5) modula
tion, 4paths indoor Rayleigh channel, v = 3 km/h, closelyspaced
paths x
max
= 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 dBHz.
In terms of MTLL, from both Figure 22 and Figure 23,
lower plots, can be concluded that the best performance
M
T
L
L
(
s
)
CosBOC (10, 5), Rician channel, 2 paths, x
max
= 0.5 chips
20 25 30 35 40
10
1
10
2
10
0
10
−1
CNR (dBHz)
MEDLL
SCM IC
SCM thr.
DC IC
SCM DC IC
Figure 25: Comparison of feedforward delays estimation algo
rithms employing the MEDLL and ICmethods and of the SCMwith
threshold approach, as a function of CNR; CosBOC(10, 5) modula
tion, 2paths decaying PDP Rician channel, v = 45 km/h, x
max
= 0.5
chips.
(i.e., the highest MTLL) is provided by the SCM DC IC
and SCM thresholding algorithms, with an improvement of
about 45 dBHz comparing to SCM IC and DC IC methods,
which give similar results.
Figures 24 and 25 illustrate the obtained simulation re
sults, for a CosBOC(10, 5)modulated signals, for a 4closely
spaced paths indoor Rayleigh proﬁle, respectively for a 2
paths channel, with v = 45 km/h and a separation between
paths x
max
of up to 0.5 chips. In terms of RMSE (Figure 24,
upper plot), the SCM DC IC method gives the best results,
followed by the SCM with threshold comparison and SCM
IC methods, for a CNR range of up to 33 dBHz. The good
performance of SCM DC IC method is expected, since for a
higher BOCmodulation 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 x
max
, in case of feedback delay esti
mation algorithms which use NEML discriminator, together
with the Julien&al. EMLP method. The channel has a 4paths
indoor Rayleigh proﬁle with the mobile speed of 4 km/h and
Adina Burian et al. 19
R
M
S
E
(
c
h
i
p
s
)
SinBOC (1, 1), 4paths Rayleigh channel,
CNR = 35 dBHz, v = 4 km/h
0 0.5 1 1.5 2
10
−0.6
10
−0.7
10
−0.5
10
−0.4
10
−0.3
x
max
(chips)
NEML
Julien & al. EMLP
SCM NEML
DC NEML
SCM DC 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, 4paths
Rayleigh channel, mobile speed 4 km/h, CNR = 35 dBHz.
the CNR is set to 35 dBHz. In this case, both SCM algo
rithms provide a decreasing in error as x
max
is increasing,
while the other methods have an almost linear behavior, for
x
max
greater than half of chip. Also, it can be observed that the
same gap between the studied methods, at x
max
= 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 BOCmodulated 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 BOCmodulation
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
trackingloop 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
SCM does not bring substantial improvement, since HRC
has already rather good performance in multipath channels.
Also, the higher BOCmodulation 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.
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 CDROMProceedings of the IONNational Meet
ing; Navigating into the NewMillennium, 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 (IONNTM ’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 narrowcorrelator 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 (IONGPS/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
(IONNTM ’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 (IONGPS ’97), vol. 1, pp. 559–568,
Kansas City, Mo, USA, September 1997.
[12] J. Jones, P. Fenton, and B. Smith, “Theory and perfor
mance of the pulse aperture correlator,” Tech. Rep., NovA
tel, Calgary, Alberta, Canada, September 2004, http://www
.novatel.com/Documents/Papers/PAC.pdf.
[13] A. V. Dierendonck and M. Braasch, “Evaluation of GNSS re
ceiver correlation processing techniques for multipath and
noise mitigation,” in Proceedings of the National Technical
Meeting of the Institute of Navigation (IONNTM ’97), pp.
207–215, Santa Monica, Calif, USA, January 1997.
[14] C. Lee, S. Yoo, S. Yoon, and S. Y. Kim, “A novel multipath
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 Mcode time discriminators,” in Proceed
ings of the National Technical Meeting of the Institute of Nav
igation (IONNTM ’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 (IONNTM ’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 (IONNTM ’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. AvilaRodriguez, 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 (IONGNSS ’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 binaryoﬀsetcarrier
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
R. Lucas, “Status of Galileo frequency and signal design,”
in CDROM Proceedings of the International Technical Meet
ing of the Satellite Division of the Institute of Navigation
(IONGPS ’02), Portland, Ore, USA, September 2002.
[25] E. S. Lohan, A. Lakhzouri, and M. Renfors, “Binaryoﬀset
carrier modulation techniques with applications in satel
lite navigation systems,” Wireless Communications and Mobile
Computing, vol. 7, no. 6, pp. 767–779, 2006.
[26] E. Rebeyrol, C. Macabiau, L. Lestarquit, et al., “BOC power
spectrum densities,” in CDROM Proceedings of the National
Technical Meeting of Institute of Navigation (IONNTM ’05),
San Diego, Calif, USA, January 2005.
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 FilterBankBased Methods for Fast Serial
Acquisition of BOCModulated 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
Binaryoﬀsetcarrier (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 BPSKmodulated CDMA signals, where the main correlation lobe spans over 2chip 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 BPSKmodulated CDMA signals). Lowering the timebin 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 frequencybased unambigu
ous acquisition methods is proposed here, namely the ﬁlterbankbased (FBB) approaches. The detailed theoretical analysis of
FBB methods is given for serialsearch singledwell 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 = f
sc
/ f
ref
, n = f
c
/ f
ref
. Here, f
c
is the chip rate, f
sc
is the subcarrier rate, and f
ref
= 1.023 MHz is the reference
chip frequency (that of the C/A GPS signal) [1]. Alterna
tively, a BOCmodulated signal can also be deﬁned via its
BOC modulation order N
BOC
2 f
sc
/ f
c
[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 BOCmodulated 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
BOCmodulated code). That is, if the socalled 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 2chip inter
val depends on the N
BOC
order (e.g., for SinBOC, we have
2N
BOC
−2 ambiguities around the maximum peak, while for
CosBOC, we have 2N
BOC
ambiguities [4]). The distance be
tween successive ambiguities in the CF envelope sets an up
per bound on the step of searching the timebin hypotheses
(Δt)
bin
, in the sense that if the timebin step becomes too
high, the main lobe of the CF envelope might be lost during
the acquisition. Typically, a step of onehalf 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 timebin steps of 0.5 chips
are used for BPSKmodulation (such as for C/Acode 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 timebin step (and,
thus, the speed of the acquisition process), several Filter
BankBased (FBB) methods are proposed here, which exploit
2 EURASIP Journal on Wireless Communications and Networking
Time uncertainty
Δt
max
.
.
.
.
.
. · · ·
· · ·
· · ·
(
Δ
f
)
b
i
n
Timebin step (Δt)
bin
F
r
e
q
u
e
n
c
y
u
n
c
e
r
t
a
i
n
t
y
Δ
f
m
a
x
One timefrequency 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 singlepath 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
ambiguousBOC processing (abovementioned) and a more
recent, unambiguousBOC technique, introduced by Fish
man and Betz [9] (denoted here via B&F method, but also
known as sideband correlation method or BPSKlike 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 hybridsearch acquisition [12].
However, the theoretical analysis of FBB methods is newly
introduced here.
2. ACQUISITIONPROBLEMANDAMBIGUOUS
(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 timefrequency 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
Δt
max
is equal to the code epoch length. The length of the fre
quency uncertainty region Δf
max
may vary according to the
initial information: if assistedGPS data are available, Δf
max
can be as small as couple of Hertzs or couple of tens of Hertzs.
If no Dopplerfrequency information exists (i.e., no assis
tance or autonomous GPS), the frequency range Δf
max
can
be as large as few tens of kHz [13].
The timefrequency bin deﬁnes the ﬁnal timefrequency
error after the acquisition process and it is characterized by
one correlator output: the length of a bin in time direction
(or the timebin step) is denoted by (Δt)
bin
(expressed in
chips) and the length of a bin in frequency direction is de
noted by (Δf )
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 timebin 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 timebin 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 timefrequency 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
ambiguousBOC (aBOC) processing means that, when form
ing the decision variable, the received signal is directly corre
lated with the reference BOCmodulated 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 BOCmodulated signals is the unambiguous
acquisition (known under several names, such as sideband
correlation method or BPSKlike 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 (singlesideband 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
f
D
. The reference code is obtained in a
Elena Simona Lohan 3
Upper sideband processing
Lower sideband processing
Upper sideband
ﬁlter
Upper sideband
ﬁlter
Received BOCmodulated
signal
Reference BOCmodulated
PRN code
Coherent and non
coherent integration
Σ
Towards
detection
stage
∗
0
0.2
0.4
0.6
0.8
1
N
o
r
m
a
l
i
z
e
d
P
S
D
−4 −3 −2 −1 0 1 2 3 4
Frequecy (MHz)
SinBOC(1, 1) spectrum
Figure 2: Block diagram of B&F method, singlesideband processing (here, upper sideband).
similar manner with the received signal, hence the autocor
relation function is no longer the CF of a BOCmodulated
signal, but it will resemble the CF of a BPSKmodulated sig
nal. However, the exact shape of the resulting CF is not iden
tical with the CF of a BPSKmodulated 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
dualsideband method is used, we add together the USB and
LSB outputs and form the dualsideband statistic.
4. FILTERBANKBASED(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 N
fb
and it is related to the number of frequency pieces per
sideband N
pieces
via: N
fb
= 2N
pieces
if dual sideband (SB) is
used, or N
fb
= N
pieces
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 N
fb
= 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.,
singleSB processing).
The ﬁlters may have equal or unequal frequency widths.
Two methods may be employed and they have been denoted
here via equalpower FBB (FBB
ep
), 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
frequencywidth FBB (FBB
efw
), 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,
since the noise power depends on the ﬁlter bandwidth (i.e.,
the noise power is constant from one band to another for
the FBB
efw
case, and it is variable for the FBB
ep
case). How
ever, the incoming (ﬁltered) signal is correlated with the ref
erence BOCmodulated 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 (FBB
ep
and FBB
efw
) 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 BOCmodulated 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 N
pieces
= 1, the pro
posed FBB methods (both FBB
ep
and FBB
efw
) become identi
cal with B&F method, and the higher the N
pieces
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
N
pieces
= 1). The complex outputs y
i
(·), i = 1, . . . , N
fb
of the
coherent integration block of Figure 4 can be written as
y
i
_
τ,
f
D
, n
_
=
1
T
coh
_
nT+Tcoh
nT
` r
i
(t)` c
i
(t − τ)e
j2π
fDt
dt, (1)
4 EURASIP Journal on Wireless Communications and Networking
−3 −2 −1 0 1 2 3
Frequency (MHz)
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
S
p
e
c
t
r
u
m
Dual sideband processing, equalpower pieces
BOC PSD
Filter 1
Filter 2
Filter 3
Filter 4
(a)
−3 −2 −1 0 1 2 3
Delay error (chips)
0
0.5
1
1.5
2
2.5
3
3.5

C
F

2
Squared CF envelope, N
pieces
= 2, N
BOC
= 2
BOC
B&F, dual SB
FBB
ep
, dual SB
FBB
efw
, dual SB
(b)
Figure 3: Illustration of the FBB acquisition methods, SinBOC(1,1)
case. Upper plot: division into frequency pieces, via N
fb
= 4 ﬁlters
(FBB
ep
method). Lower plot: squared CF shapes for 2 FBB meth
ods, compared with ambiguous BOC (aBOC) and unambiguous
Betz&Fishman (B&F) methods.
where n is the symbol (or code epoch) index, T is the symbol
interval, ` r
i
(t) is the ﬁltered signal via the ith ﬁlter, ` c
i
(t) is the
ﬁltered reference code (note that the code c(t) before the ﬁlter
bank is the BOCmodulated spread spectrum sequence), τ
and
f
D
are the receiver candidates for the delay and Doppler
shift, respectively, and T
coh
is the coherent integration length
(if the code epoch length is 1 millisecond, then the number of
coherent code epochs N
c
may be used instead: T
coh
= N
c
ms).
Without loss of generality, we may assume that a pilot chan
nel is available (such as it is the case of Galileo L1 band), thus
the received signal r(t) (before ﬁltering) has the form
r(t) =
_
E
b
c(t − τ)e
−j2π fDt
+ η
wb
(t), (2)
where τ and f
D
are the delay and Doppler shift introduced
by the channel, η
wb
(t) is the additive white Gaussian noise at
wideband level, and E
b
is the bit energy.
The coherent integration outputs y
i
(·) are Gaussian pro
cesses (since a ﬁltered Gaussian processes is still a Gaussian
processes). Their mean is either 0 (if we are in an incorrect
timefrequency bin) or it is proportional to a timeDoppler
deterioration factor
_
E
b
F (Δ τ, Δ
f
D
) [11], with a proportion
ality constant dependent on the number of ﬁlters and of the
acquisition algorithm, as it will be shown in Section 5. Here,
F (·) is the amplitude deterioration in the correct bin due
to a residual time error Δ τ and a residual Doppler error Δ
f
D
[11]
F
_
Δ τ, Δ
f
D
_
=
¸
¸
¸
¸
R
_
Δ τ
_sin
_
πΔ
f
D
T
coh
_
πΔ
f
D
T
coh
¸
¸
¸
¸
. (3)
As mentioned above, Δ τ = τ − τ, Δ
f
D
= f
D
−
f
D
, and R(Δ τ)
is the CF value at delay error Δ τ (CF is dependent on the used
algorithm, as shown in the lower plot of Figure 3). Moreover,
if we normalize the y
i
(·) variables with respect to their max
imum power, the variance of y
i
(·) variables (in both the cor
rect and incorrect bins) is proportional to the postintegration
noise variance
σ
2
10
−(CNR+10log
10
Tcoh)/10
, (4)
where CNR = E
b
B
W
/N
0
is the CarriertoNoise Ratio, ex
pressed in dBHz [5, 7, 11], B
W
is the signal bandwidth after
despreading (e.g., B
W
= 1 kHz for GPS and Galileo signals),
and N
0
is the doublesided noise spectral power density in
the narrowband domain (after despreading or correlation on
1 millisecond in GPS/Galileo). The proportionality constants
are presented in Section 5. The decision statistic Z of Figure 4
is the output of noncoherent combining of N
nc
N
fb
complex
Gaussian variables, where N
nc
is the noncoherent integration
time (expressed in blocks of N
c
ms):
Z =
1
N
nc
1
N
fb
Nnc
_
n=1
Nfb
_
i=1
¸
¸
y
i
_
τ,
f
D
, n
_¸
¸
2
. (5)
We remark that the coloured noise impact on Z statistic is
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 chisquare 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
c(t)
Ref code
r(t)
Rx sign.
N
fb
ﬁlters
FB
N
fb
ﬁlters
FB
Optional stage
.
.
.
` c
N
fb
(t)
` c
1
(t)
.
.
. ` r
N
fb
(t)
` r
1
(t)
∗
∗
y
N
fb
y
1
Coherent
integr.
Coherent
integr. .
.
.

2

2
.
.
.
N
nc
N
nc
.
.
.
N
fb
Z
Figure 4: Block diagram of a generic acquisition block.
5. THEORETICAL MODEL FOR FBB
ACQUISITIONMETHODS
5.1. Test statistic distribution
As explained above, the test statistic Z for aBOC, B&F, and
proposed FBB
ep
approaches
1
is either a central or a noncen
tral χ
2
distributed variable with N
deg
degrees of freedom, ac
cording to whether we have an incorrect (bin H
0
) or a
correct (bin H
1
) timefrequency bin, respectively. Its non
centrality parameter λ
Z
and its variance σ
2
Z
are thus given by
λ
Z
= ξ
λbin
¸
¸
F
_
Δ τ, Δ
f
D
_¸
¸
,
σ
2
Z
= ξ
σ
2
bin
σ
2
N
nc
,
(6)
where F (·) is given in (3), σ
2
is given in (4), and ξ
σ
2
bin
and
ξ
λbin
are two algorithmdependent 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 squareroot 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 E
b
= 1 and all the signal and noise statistics are present
with respect to this normalization.
Clearly, for aBOC algorithm, ξ
σ
2
bin
= 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, N
deg
= 2N
nc
for aBOC, because
we add together the absolutesquared valued of N
nc
complex
variables (or the squares 2N
nc
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) P
ml
of the BOC power spectral
1
The case of FBB
efw
is discussed separately, later in this section.
density (PSD) function. P
ml
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 P
ml
< 0.5.; P
ml
factor is normalized with respect to the
total signal power, thus P
ml
< 0.5 (e.g., P
ml
= 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
λH
1
and ξ
σ
2
bin
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
for single SB, therefore, the noncentrality parameter (which
is a measure of the amplitude, not of the signal power) in
creases by
√
2. Furthermore, in dualSB approaches, we add
a double number of noncoherent variables, thus the num
ber of degrees of freedom is doubled compared to singleSB
approaches.
The derivation of χ
2
parameters for FBB
ep
is also straight
forward by keeping in mind that the variance of the vari
ables y
i
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
2
bin
= P
ml
/N
pieces
, bin = H
0
, H
1
, and the signal power de
creases to N
pieces
(P
2
ml
/N
2
pieces
), thus x
λbin
= P
ml
/
_
N
pieces
for
single SB (and x
λbin
=
√
2P
ml
/
_
N
pieces
for dual SB).
For FBB
efw
, 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 FBB
efw
test statistic is bounded by two χ
2
distributions. Moreover, P
maxpp
is the maximum power per
piece (in the positive or in the negative frequency band). For
example, if N
pieces
= 2 and FBB
efw
approach is used for Sin
BOC(1,1) case, the powers per piece of the positivesideband
lobe are 0.10 and 0.34, respectively (hence, P
maxpp
= 0.34).
Again, these powers can be derived straightforwardly, via the
formulas shown in [1, 3, 4, 20].
Figure 6 compares the simulationbased complementary
CDF (i.e., 1CDF) with theoretical complementary CDFs
for FBB
ep
case (similar plots were obtained for aBOC,
B&F, and FBB
efw
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 H
1
) Incorrect bin (hypothesis H
0
)
ξ
λ
H
1
ξ
σ
2
H
1
N
deg
ξ
λ
H
0
ξ
σ
2
H
0
N
deg
aBOC 1 1 2N
nc
0 1 2N
nc
Singlesideband
B& F
P
ml
P
ml
2N
nc
0 P
ml
2N
nc
Dualsideband
B&F
√
2P
ml
P
ml
4N
nc
0 P
ml
4N
nc
Singlesideband
FBB
ep
and lower
bound of single
sideband FBB
efw
P
ml
_
N
pieces
P
ml
N
pieces
2N
nc
N
pieces
0
P
ml
N
pieces
2N
nc
N
pieces
Dualsideband
FBB
ep
and lower
bound of dual
sideband FBB
efw
√
2
P
ml
_
N
pieces
P
ml
N
pieces
4N
nc
N
pieces
0
P
ml
N
pieces
4N
nc
N
pieces
Upper bound of
singlesideband
FBB
efw
P
ml
_
N
pieces
P
maxpp
N
pieces
2N
nc
N
pieces
0
P
ml
N
pieces
2N
nc
N
pieces
Upper bound of
dualsideband
FBB
efw
√
2
P
ml
_
N
pieces
P
maxpp
N
pieces
4N
nc
N
pieces
0
P
ml
N
pieces
4N
nc
N
pieces
lack of space). For the simulations shown in Figure 6,
SinBOC(1,1) signal was used, with coherent integration
length N
c
= 20 milliseconds, noncoherent integration length
N
nc
= 2, CNR = 24 dBHz, number of samples per BOC
interval N
s
= 4, and singleSB ﬁlter bank with 4 ﬁl
ters (i.e., N
fb
= N
pieces
= 4). It was also noticed that
the bounds for FBB
efw
approach are rather loose. How
ever, simulation results showed that the average behavior
of FBB
efw
, while keeping between the bounds, is also very
similar with the average behavior of FBB
ep
[12], therefore,
from now on, it is possible to rely on FBB
ep
curves alone
in order to illustrate the average performance of proposed
FBB methods. We remark that the plots of complementary
CDF were chosen instead of CDF, in order to show bet
ter the tail matching of the theoretical and simulationbased
distributions.
5.2. Detection probability and
mean acquisition times
In serial search acquisition, the detection probability per
bin P
dbin
(Δ τ) is the probability that the decision variable Z
is higher than the decision threshold γ, provided that we
are in a correct bin (hypothesis H
1
). Similarly, the false
alarm probability P
fa
is the probability that the decision vari
able is higher than γ, provided that we are in an incor
rect bin (hypothesis H
0
). These probabilities can be easily
computed based on the cumulative distribution functions
(CDFs) of Z in the correct F
nc
(·) and incorrect bins F
c
(·)
[11]:
P
dbin
_
Δ τ, Δ
f
D
_
= 1 − F
nc
(γ, λ
Z
),
P
fa
= 1 − F
c
(γ),
(7)
2 3 4 5 6 7 8 9 10 11 12
BOC modulation order N
BOC
0.36
0.37
0.38
0.39
0.4
0.41
0.42
0.43
P
o
w
e
r
p
e
r
m
a
i
n
(
p
o
s
i
t
i
v
e
o
r
n
e
g
a
t
i
v
e
)
l
o
b
e
P
m
l
Power per main lobe of BOCmodulated signal
Sine BOC
Cosine BOC
Figure 5: Normalized power per main lobe P
ml
for BOCmodulated
signals for various N
BOC
orders.
where F
nc
(·) is the CDF of a noncentral χ
2
variable and
F
c
(·) is the CDF of a central χ
2
variable, given by [17]:
F
c
(z) = 1 −
Ndeg/2−1
_
k=0
e
−z/σ
2
Z
_
z
σ
2
Z
_
k
1
k!
in incorrect bins H
0
F
nc
_
z, λ
Z
_
= 1 − Q
Ndeg/2
_
λ
Z
√
2
σ
Z
,
√
2z
σ
Z
_
in correct bins H
1
(8)
Elena Simona Lohan 7
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
Test statistic levels
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
1

C
D
F
Matching to χ
2
complementary CDF for SSB, FBB
Sim, noncentral
Th, noncentral
Sim, central
Th, central
Figure 6: Matching with χ
2
distributions, (complementary CDF:
1CDF), theory (th) versus simulations (sim), FBB
ep
, N
fb
=
N
pieces
= 4.
with σ
2
Z
, N
deg
, and λ
Z
given in (6) and in Table 1, and
Q
Ndeg/2
(·) being the generalized MarcumQ function [17].
Due to the fact that the timebin step may be smaller than
the 2chip interval of the CF main lobe, we might have
several correct bins. The number of correct bins is: N
t
=
2T
c
/(Δt)
bin
, where T
c
is the chip interval. Thus, the global
detection probability P
d
is the sum of probabilities of detect
ing the signal in the ith bin, provided that all the previous
tested hypotheses for the prior correct bins gave a misdetec
tion [11]:
P
d
_
Δτ
0
_
=
Nt −1
_
k=0
P
dbin
_
Δτ
0
+ k(Δt)
bin
, Δ
f
D
_
k−1
i=0
_
1 − P
dbin
_
Δτ
0
+ i(Δt)
bin
, Δ
f
D
__
.
(9)
In (9), Δτ
0
is the delay error associated with the ﬁrst sam
pling point in the twochip interval, where we have N
t
cor
rect bins. Equation (9) is valid only for ﬁxed sampling points.
However, due to the random nature of the channels, the sam
pling point (with respect to the channel delay) is randomly
ﬂuctuating, hence, the global P
d
is computed as the expecta
tion E(·) over all possible initial delay errors (under uniform
distribution, we simply take the temporal mean):
P
d
= E
Δτ0
_
P
d
_
Δτ
0
__
, (10)
and the worst detection probability is obtained for the worst
sequence of sampling points: P
d,worst
= min
Δτ0
(P
d
(Δτ
0
)).
The mean acquisition time T
acq
for the serial search is
computed according to the global P
d
, the false alarm P
fa
, the
penalty time K
penalty
(i.e., the time lost to restart the acqui
sition process if a false alarm state is reached), and the total
number of bins in the search space [21]:
T
acq
=
2 +
_
2 − P
d
_
(q − 1)
_
1 + K
penalty
P
fa
_
2P
d
τ
d
, (11)
where τ
d
= N
nc
T
coh
is the dwell time, q is the total num
ber of bins in the search space, and P
d
and P
fa
are given by
(7) to (10). An example of the theoretical average detection
probability P
d
compared with the simulation results is shown
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
plained by the number of points used in the statistics: about
5000 random points have been used to build such statistics,
which seemed enough for most of (Δt)
bin
ranges. However, at
very low detection probabilities, this number is still too small
for a perfect match. However, the gap is not signiﬁcant, and
low P
d
regions are not the most interesting from the analysis
point of view.
An example of performance (in terms of average and
worst detection probabilities) of the proposed FBB methods
is given in Figure 8. The gap between proposed FBB methods
and aBOC method is even higher from the point of view of
the worst P
d
. Here, SinBOC(1,1)modulated signal was used,
and N
c
= 20 ms, N
nc
= 2. The other parameters are speciﬁed
in the ﬁgures captions. The small edge in aBOC average per
formance at around 0.7 chips is explained by the fact that a
timebin step equal to the width of the main lobe of CF en
velope (i.e., about 0.7 chips) would give worse performance
than a slightly higher or smaller steps, due to ambiguities in
the CF envelope. Also, the relatively constant slope in the re
gion of 0.7–1 chips can be explained by the combination of
high timebin steps and the presence of the deep fades in the
CF: since the spacing between those deep fades is around 0.7
chips for SinBOC(1,1), then a timebin step of 0.7 chips is the
worst possible choice in the interval up to 1 chip. However,
there is no signiﬁcant diﬀerence in average P
d
for timebin
steps between 0.7 and 1 chip, since two countereﬀects are
superposed (and they seem to cancel each other in the region
of 0.7 till 1 chip from the point of view of average P
d
): on
one hand, increasing the timebin step is deteriorating the
performance; on the other hand, avoiding (as much as possi
ble) the deep fades of CF is beneﬁcial. This fact is even more
visible from the lower plot of Figure 8, where worstcase P
d
are shown. Clearly, having a timebin step of about 0.7 chips
would mean that, in the worst case, we are always in a deep
fade and lose completely the peak of the main lobe. This ex
plains the minimumP
d
achieved at such a step. Also, for steps
higher than 1.5 chips, there is always a sampling sequence
that will miss completely the main lobe of the envelope of CF
(thus, the worst P
d
will be zero).
It is noticed that FBB methods can work with timebin
steps higher than 1 chip, due to the increase in the main lobe
of the CF envelope. Moreover, the proposed FBB methods
(both single and dual SB) outperform the B&F and aBOC
method if the step (Δt)
bin
is suﬃciently high. Indeed, the
higher the timebin step, the higher is the improvement of
FBB methods over aBOC and B&F methods. We remark that
even at (Δt)
bin
= 1 chip, we have a signiﬁcantly high P
d
,
8 EURASIP Journal on Wireless Communications and Networking
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
(Δt)
bin
(chips)
10
−3
10
−2
10
−1
10
0
P
d
P
d
at P
fa
= 0.001, dual B&F, CNR = 27 dBHz
Sim, average
Th, average
Sim, worst
Th, worst
(a)
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
(Δt)
bin
(chips)
10
−2
10
−1
10
0
P
d
P
d
at P
fa
= 0.001, dual FBB
ep
, CNR = 27 dBHz, N
pieces
= 2
Sim, average
Th, average
Sim, worst
Th, worst
(b)
Figure 7: Comparison between theory and simulations for Sin
BOC(1,1). Left: dualsideband B&F method. Right: Dualsideband
FBB
ep
method, N
pieces
= 2. N
c
= 10 milliseconds, N
nc
= 5, CNR =
27 dBHz, N
s
= 5.
due to the widening of the CF main lobe. The constant P
d
at higher timebin 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 timebin step in
the acquisition process, we may decrease substantially the
mean acquisition time, because the number of bins in the
0 0.5 1 1.5 2 2.5 3
Timebin step (Δt)
bin
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
P
d
Average P
d
, N
pieces
= 2, CNR = 30 dBHz
aBOC
Single B&F
Dual B&F
Single FBB
Dual FBB
(a)
0 0.5 1 1.5 2 2.5 3
Timebin step (Δt)
bin
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
P
d
P
d,worst
, N
pieces
= 2, CNR = 30 dBHz
aBOC
Single B&F
Dual B&F
Single FBB
Dual FBB
(b)
Figure 8: Average (upper) and worst (lower) detection probabili
ties versus (Δt)
bin
ambiguous and unambiguous BOC acquisition
methods (FBB
ep
was used here).
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)
2
bin
[9], thus,
when (Δt)
bin
increases, the computational load decreases.
An example regarding the needed timebin 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
25 26 27 28 29 30 31
CNR (dBHz)
0
0.5
1
1.5
2
(
Δ
t
)
b
i
n
(
c
h
i
p
s
)
Step needed to achieve a target P
d
= 0.9, (average case)
Dual SB, FBB
ep
Dual SB, B&F
(a)
25 26 27 28 29 30 31
CNR (dBHz)
0
0.5
1
1.5
(
Δ
t
)
b
i
n
(
c
h
i
p
s
)
Step needed to achieve a target P
d
= 0.9, (average case)
Single SB, FBB
ep
Single SB, B&F
(b)
25 26 27 28 29 30 31
CNR (dBHz)
10
1
10
2
10
3
M
A
T
Achieved MAT [s] at considered step
Dual SB, FBB
Dual SB, B&F
(c)
25 26 27 28 29 30 31
CNR (dBHz)
10
1
10
2
10
3
10
4
M
A
T
Achieved MAT [s] at considered step
Single SB, FBB
ep
Single SB, B&F
(d)
Figure 9: Step needed to achieve a target average P
d
= 0.9, at false alarm P
fa
= 10
−3
and corresponding mean acquisition time, SinBOC(1,1)
signal. Code length 4092 chips, penalty factor K
penalty
= 1, single frequencybin. N
pieces
= 2 for FBB
ep
. 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 P
d
= 0.9 at
P
fa
= 10
−3
. For these values, we need a step of (Δt)
bin
= 1.2
chips for the dualsideband B&F method (which will cor
respond to a mean acquisition time T
acq
= 86.24 s for sin
gle frequency serial search and 4092chip length code) and a
step of (Δt)
bin
= 1.7 chips for dualsideband FBB
ep
method
with N
pieces
= 2 (i.e., T
acq
= 58.14 s). Thus, the step can be
about 50% higher for dualsideband 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 dualSB FBB case than in dualSB
B&F case). For singlesideband 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 (N
c
, N
nc
) 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 N
pieces
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
K
penalty
may take and no general rule about the choice of
K
penalty
has been given so far, to the author’s knowledge. For
example, in [22] a penalty factor K
penalty
= 1 was consid
ered; in [23] simulations were carried out for K
penalty
= 2, in
[24] a penalty factor of K
penalty
= 10
3
was used, while in [25]
we have K
penalty
= 10
6
. Penalty factors with respect to dwell
times were also used in the literature, for example: K
penalty
=
10
5
/(N
c
N
nc
) [26, 27], or K
penalty
= 10
7
/(N
c
N
nc
) [27] (in our
simulations, N
c
N
nc
= 40 ms). Therefore, K
penalty
may spread
over an interval of [1, 10
6
], therefore, in our simulations we
considered the 2 extreme cases: K
penalty
= 1 (Figure 9) and
K
penalty
= 10
6
(Figure 10). Figure 10 uses exactly the same
parameters as Figure 9, with the exception of the penalty
factor, which is now K
penalty
= 10
6
. For K
penalty
= 10
6
of
Figure 10, MAT for the dualsideband B&F method becomes
T
acq
= 8.62 ∗ 10
4
, which is still higher than MAT for the
dualsideband FBB
ep
(T
acq
= 5.8 ∗ 10
4
s). Similar improve
ments in MAT times via FBB processing (as for K
penalty
= 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−P
d
versus false alarm probability P
fa
[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
25 26 27 28 29 30 31
CNR (dBHz)
10
4
10
5
10
6
M
A
T
Achieved MAT [s] at considered step
Dual SB, FBB
ep
Dual SB, B&F
(a)
25 26 27 28 29 30 31
CNR (dBHz)
10
4
10
5
10
6
10
7
M
A
T
Achieved MAT [s] at considered step
Single SB, FBB
ep
Single SB, B&F
(b)
Figure 10: Mean acquisition time corresponding to the step needed to achieve a target average P
d
= 0.9, at false alarm P
fa
= 10
−3
, Sin
BOC(1,1) signal. Code length 4092 chips, penalty factor K
penalty
= 10
6
, single frequencybin. N
pieces
= 2 for FBB
ep
. Left: dual sideband. Right:
single sideband.
10
−10
10
−8
10
−6
10
−4
10
−2
False alarm probability P
fa
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
M
i
s

d
e
t
e
c
t
i
o
n
p
r
o
b
a
b
i
l
i
t
y
1

P
d
ROC, (Δt)
bin
= 0.5 chips, CNR = 30 dBHz
aBOC
Single BF
Dual BF
Single FBB
Dual FBB
(a)
10
−10
10
−8
10
−6
10
−4
10
−2
False alarm probability P
fa
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
M
i
s

d
e
t
e
c
t
i
o
n
p
r
o
b
a
b
i
l
i
t
y
1

P
d
ROC, (Δt)
bin
= 1.5 chips, CNR = 30 dBHz
aBOC
Single BF
Dual BF
Single FBB
Dual FBB
(b)
Figure 11: Receiver operating characteristic for CNR = 30 dBHz, SinBOC(1,1) signal, N
c
= 20, N
nc
= 2. Left: (Δt)
bin
= 0.5 chips; right
(Δt)
bin
= 1.5 chips.
At low timebin steps (e.g., (Δt)
bin
= 0.5 chips), the FBB and
B&F methods behave similarly, as it has been seen before also
in Figure 8. The main advantage of FBB methods is observed
for timebin steps higher than one chip, as shown in the left
plot of Figure 11. For both timebin steps considered here,
the single sideband unambiguous methods have a threshold
false alarm, below which their performance becomes worse
than that of ambiguous BOC approach. This threshold de
pends on the CNR, on the integration times, and on the time
bin step and it is typically quite low (below 10
−5
).
6. CONCLUSIONS
This paper introduces a new class of code acquisition meth
ods for BOCmodulated CDMA signals, based on ﬁlter bank
processing. The detailed theoretical characterization of this
Elena Simona Lohan 11
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 timebin
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 (IONNTM ’99), pp. 639–648, San
Diego, Calif, USA, January 1999.
[2] A. Burian, E. S. Lohan, and M. Renfors, “BPSKlike methods
for hybridsearch 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, “Binaryoﬀset
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. Gu´ erin, 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 (ENCGNSS ’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 (IONNTM ’00),
Anaheim, Calif, USA, January 2000, CDROM.
[9] P. Fishman and J. W. Betz, “Predicting performance of direct
acquisition for the Mcode 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 (IONGNSS ’05),
Long Beach, Calif, USA, September 2005, CDROM.
[11] E. S. Lohan, “Statistical analysis of BPSKlike 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, “Filterbank 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, McGrawHill, New York,
NY, USA, 4th edition, 2001.
[18] R. R. Rick and L. B. Milstein, “Optimal decision strate
gies for acquisition of spreadspectrum 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 (IONGPS
’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
spreadspectrum 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 PNcode acquisition
for CDMAmobile communications,” IEEE Transactions on Ve
hicular Technology, vol. 45, no. 4, pp. 643–656, 1996.
[24] E. A. Homier and R. A. Scholtz, “Hybrid ﬁxeddwelltime
search techniques for rapid acquisition of ultrawideband 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 DSCDMA 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
spectrumcode acquisition in frequencyselective 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 MTTS International Tipical Symposium on Technolo
gies for Wireless Applications, pp. 181–186, Vancouver, Canada,
February 1999.
This action might not be possible to undo. Are you sure you want to continue?
We've moved you to where you read on your other device.
Get the full title to continue reading from where you left off, or restart the preview.