Upload
nguyennhan
View
235
Download
1
Embed Size (px)
Citation preview
Page 1 (75)
Project Number: CELTIC / CP5-026
Project Title: Wireless World Initiative New Radio – WINNER+
Document Type: PU (Public)
Document Identifier: D1.8
Document Title: D1.8 Intermediate Report on CoMP (Coordinated Multi-Point) and Relaying in the
Framework of CoMP
Source Activity: WP1
Editor: Sylvie Mayrargue
Authors: Mauro Boldi, Carmen Botella, Federico Boccardi, Valeria D’Amico, Eric Hardouin, Magnus Olsson, Harri Pennanen, Peter Rost,
Valentin Savin, Tommy Svensson, Antti Tölli
Status / Version: Stable 0.1
Date Last changes: 20.01.10
File Name: D1.8.doc
Abstract:
This deliverable is an intermediate report on CoMP (Coordinated Multi-Point) and on Relaying in the Framework of CoMP. It describes the second set of innovations encompassing concepts about promising principles or ideas as well as detailed innovative techniques in the context of the work towards the WINNER+ system concept. For each concept, the associated benefits as well as the corresponding requirements on the system architecture and protocols, measurements and signalling, are considered.
Regarding CoMP algorithms, focus is put on schemes with reduced requirements in terms of backhauling considering two categories: “Coordinated Beamforming”, and “Joint Processing”. As for relaying, a relay-assisted interference channel, and a distributed LDPC coding for a Decode and Forward relay are introduced.
Keywords:
Coordinated Beamforming, Coordinated multipoint systems, Joint Processing, Relaying
Document History:
05.05.2009 1rst
Draft version
04.08.2009 2nd
Draft version
02.10.2009 Clean version
WINNER+ D1,8
Version: 1.0 Page 2 (75)
Table of Contents
1. Introduction ............................................................................................ 8
2. Investigations on CoMP issues .......................................................... 10
2.1 Assessment of transmission strategies in a RoF based CoMP architecture ......................... 10
2.2 Coordinated beamforming concepts .................................................................................. 14
2.2.1 Centralized non-codebook based coordinated beamforming ....................................... 16
2.2.2 Decentralized non-codebook based coordinated beamforming .................................... 18
2.2.3 Codebook-based coordinated beamforming ............................................................... 22
2.3 Joint processing with relaxed requirement ........................................................................ 25
2.3.1 Performance of joint processing schemes under varying CSI requirements ................. 26
2.3.2 Joint processing with reduced backhaul requirement by MAC coordination ................ 27
2.3.3 A generalized method for joint design of linear transceivers with CoMP transmission 31
3. Investigations on relaying in the framework of CoMP ..................... 35
3.1 Impact of interference on design and performance of relaying protocols ............................ 35
3.2 Distributed LDPC coding for the single relay channel ....................................................... 41
4. Conclusion ........................................................................................... 46
5. References ........................................................................................... 48
A. Appendix .............................................................................................. 52
A.1 System level performance evaluation of coordinated beamforming and joint processing .... 52
A.2 Further details and performance evaluation of decentralized coordinated beamforming ..... 56
A.3 Simulation conditions details for codebook-based coordinated beamforming .................... 60
A.4 Performance investigation of joint processing schemes considering area coverage. ............ 61
A.5 Joint processing with reduced backhaul requirement by MAC coordination....................... 67
A.6 Further details and performance evaluation of generalized CoMP transmission method ..... 69
WINNER+ D1,8
Version: 1.0 Page 3 (75)
Authors
Partner Name Phone / Fax / e-mail
CEA-LETI Sylvie Mayrargue Phone: +33 (0)4 38 78 62 42
Fax: +33 (0)4 38 78 65 86
e-mail: [email protected]
Valentin Savin Phone: +33 (0)4 38 78 07 11
Fax :: +33 (0)4 38 78 65 86
e-mail: [email protected]
Alcatel-Lucent UK Federico Boccardi Phone: +44 (0)1793776670
Fax: +44 (0)1793776725
e-mail: [email protected]
Chalmers University Tommy Svensson Phone: +46 31 772 1823
of Technology Fax: +46 31 772 1748
e-mail: [email protected]
Carmen Botella Phone: +46 31 772 1885
Fax: +46 31 772 1748
e-mail: [email protected]
Ericsson AB Magnus Olsson Phone: +46 10 71 30774
Fax: +46 10 71 31480
e-mail: [email protected]
France Telecom Eric Hardouin Phone: +33 1 45 29 44 16
Fax: +33 1 45 29 45 34
e-mail: [email protected]
Telecom Italia Lab Mauro Boldi Phone: +39 011 228 7771
Fax:+390112285224
e-mail: [email protected]
Valeria D’Amico Phone: +39 011 228 7544
Fax:+390112285224
e-mail: [email protected]
Bruno Melis Phone: +39 011 228 7121
Fax:+390112285224
e-mail: [email protected]
University of Oulu Antti Tölli Phone: +358 8 553 2986
WINNER+ D1,8
Version: 1.0 Page 4 (75)
Fax: +358 8 553 2845
e-mail: [email protected]
Harri Pennanen Phone: +358 8 553 2854
Fax: +358 8 553 2845
e-mail: [email protected]
Technische Universität Peter Rost Phone: +49 351 463 41042
Dresden/ Vodafone Chair Fax: +49 351 463 41099
e-mail: [email protected]
WINNER+ D1,8
Version: 1.0 Page 5 (75)
Executive Summary
This deliverable presents the innovative concepts on Coordinated Multi Point (CoMP) identified by the
Innovation Group and studied during year two. These concepts are candidate for potential inclusion in the
WINNER+ system concept. They consist of promising principles or ideas, and include detailed
innovative techniques if already available. State-of-the-art reviews are provided for each identified
concept, as well as first considerations about the requirements on the system, especially regarding
measurements and signalling, and architecture and protocols.
CoMP transmission and reception is studied, which refers to a system where the transmission and/or
reception at multiple, geographically separated antenna sites is dynamically coordinated in order to
improve system performance. CoMP is seen as one of the most promising means to provide high data rate
services for a large number of users, with a high spectral efficiency over the entire cell area. In year one,
WINNER+ studied different aspects of CoMP extensively. These included system architectures, different
approaches and algorithms for performing the coordination, and the requirements in terms of
measurements, signalling, backhauling constraints, etc. these put on the system. The work is reported in
[WIN+D14]. Much attention was put on joint processing/transmission schemes, in which users data are
shared between cooperating base stations (BSs). It was found that these schemes have potential to provide
significant performance gains, however at the price of high requirements on the backhaul links in terms of
latency and capacity since user data, channel state information (CSI), and precoding weights need to be
shared among the transmission points. During year two more focus has been put on another category of schemes, i.e, coordinated beamforming where each user data are available at a single BS. Only
information such as channel state or other quality indicators are shared by cooperating BSs, as well as
scheduling decisions and/or generated beams. Another topic studied during year 2 has been ways to
reduce the backhauling requirements for schemes based on joint processing. In both cases, clustering (i.e.
determining which BSs will cooperate) is key to algorithmic performance.
The introduction of Relay Nodes (RN) which are controlled by the network allows to use them as part of
a CoMP system. The RNs can be used to extend the actual coverage or to densify the actual network to
enhance the user throughput at the cell edge. Relay nodes are connected to a BS via over-the-air in-band
links (e.g. specific control channels or in-band backhaul, depending on the relay type), enabling a tight
coordination but at the price of a possible delay between the coordinated nodes. In year one, different
aspects of coordinated relaying schemes were investigated, e.g. coding schemes and schedulers taking
relaying into account. The conclusion was that it is worthwhile investigating relaying in the framework of
CoMP further. Hence, various cooperative relaying schemes have been studied in year two. The first
innovation considers a relay-assisted interference channel with two communication pairs and one relay
node for each pair. Various cooperative and non-cooperative schemes are compared. The second
innovation considers distributed Low Density Parity Codes (LDPC) coding for a Decode and Forward
(DF) type relay.
WINNER+ D1,8
Version: 1.0 Page 6 (75)
List of acronyms and abbreviations
3GPP 3rd Generation Partnership Project
AF Amplify-and-Forward
AWGN Additive White Gaussian Noise
BC Broadcast Channel
BS Base Station
CBF Codebook Based Beamforming
CDF Cumulative Distribution Function
CJP Centralized Joint Processing
CDMA Code Division Multiple Access
CoMP Coordinated MultiPoint
CSI Channel State Information
CU Central Unit
DAS Distributed Antenna System
DF Decode-and-Forward
DJP Distributed Joint Processing
DL Downlink
DPC Dirty Paper Coding
ETW Etkin-Tse-Wang
E-UTRA Evolved UMTS Terrestrial Radio Access
FDD Frequency Division Duplex
HK Han-Kobayashi
IC Interference Channel
ICI Inter-Cell Interference
IEEE Institute of Electrical and Electronics Engineers
ITU International Telecommunication Union
LDPC Low-density Parity-check
LI-PMI Least Interfering PMI
LMMSE Linear Minimum Mean Square Error
LOS Line Of Sight
LTE Long Term Evolution of 3GPP mobile system
LTE-A LTE-Advanced
MAC Medium Access Control / Multiple Access Channel
MI-PMI Most Interfering PMI
MIMO Multiple-Input Multiple-Output
MMSE Minimum Mean Square Error
MRC Maximum Ratio Combining
MuBF Multi-user Beamforming
OFDM Orthogonal Frequency Division Multiplexing
OFDMA Orthogonal Frequency Division Multiple Access
PHY Physical Layer
PJP Partial Joint Processing
PMI Precoding Matrix Index
QPSK Quadrature Phase Shift Keying
RAN Radio Access Network
RoF Radio over Fibre
RN Relay Node
RRM Radio Resource Management
RRU Remote Radio Unit
RU Remote Unit
RX Receive
SDMA Spatial Division Multiple Access
WINNER+ D1,8
Version: 1.0 Page 7 (75)
SIC Successive Interference Cancellation
SINR Signal to Interference plus Noise Ratio
SNR Signal to Noise Ratio
SOCP Second Order Cone Program
TC Turbo Code
TDD Time Division Duplex
TDMA Time Division Multiple Access
TTI Transmission Time Interval
TX Transmitter
UE User Equipment
UL Uplink
UMTS Universal Mobile Telecommunications System
UT User Terminal
WiMAX Worldwide Interoperability for Microwave Access
ZF Zero-Forcing
ZF-DPC Zero-Forcing Dirty Paper Coding
WINNER+ D1,8
Version: 1.0 Page 8 (75)
1. Introduction
Future cellular networks will need to provide high data rate services for a large number of users, which
requires a high spectral efficiency over the entire cell area. In order to achieve this, it is important that the
radio interface is robust to interference and in particular inter-cell interference (ICI) which appears when
the same radio resources is re-used in different cells in an uncoordinated way. Naturally, ICI particularly
degrades the performance of users located in the cell edge areas, which creates a performance discrepancy
between cell edge and inner cell users. One possible means to alleviate this performance discrepancy is to
employ Coordinated Multi Point (CoMP) transmission and reception, which refers to a system where the
transmission and/or reception at multiple, geographically separated antenna sites is dynamically
coordinated in order to improve system performance.
The CoMP framework encompasses all the system designs allowing tight coordination between multiple
radio access points for transmission and/or reception. Three types of coordinated entities can be considered, as stated in [WIN+D14] and depicted in HFigure 1-1:
Remote radio units (RRU);
Cells, which involve intra-BS or inter-BS coordination;
Relay nodes (RNs).
Figure 1-1: Different instances of systems able to implement CoMP.
The coordination can either be distributed, by means of direct communication between the different sites,
or by means of a central coordinating node.
At a high level, downlink coordination schemes can be divided into two categories (this classification mostly follows the one 3GPP adopted in the study item for Long Term Evolution Advanced (LTE-A)
[3GPP36814]):
Coordinated scheduling and/or beamforming
Joint processing/transmission
The first category is characterized by that data to a single user equipment (UE) is instantaneously
transmitted from one of the transmission points, and that scheduling decisions and/or generated beams are coordinated in order to control the created interference. The main advantages of these schemes compared
to schemes involving joint processing/transmission (see below) are that the requirements on the
coordination links and on the backhaul are significantly reduced, since typically
only information on scheduling decisions and/or generated beams (and information needed for
their generation) need to be coordinated, and
user data do not need to be made available at the coordinated transmission points, since there is
only one serving transmission point for one particular UE.
The second category, joint processing/transmission, is characterized by that data to a single UE is
simultaneously transmitted from multiple transmission points, e.g. to (coherently or non-coherently)
improve the received signal quality and/or cancel actively interference for other UEs. This category of schemes puts higher requirements on the coordination links and the backhaul since user data need to be
made available at the multiple coordinated transmission points. The amount of data to be exchanged over
the coordination links is also larger, e.g. channel knowledge and computed transmission weights.
WINNER+ D1,8
Version: 1.0 Page 9 (75)
In year 1, WINNER+ studied different aspects of CoMP extensively. These included system
architectures, different approaches and algorithms for performing the coordination, and the requirements
in terms of measurements, signalling, backhauling constraints, etc. these put on the system. The work is
reported in [WIN+D14]. Much attention was put on the second category of coordination schemes, i.e., joint processing/transmission. It was found that these schemes have potential to provide significant
performance gains, however to the cost of high requirements on the backhaul links in terms of latency and
capacity since user data, channel state information (CSI), and precoding weights needed to be shared
among the transmission points.
Hence, during year 2 WINNER+ put more focus on the first category of schemes, i.e., coordinated
beamforming, but also on how to reduce the backhauling requirements for schemes based on joint
processing. It’s worth noting that these requirements are being closely evaluated to assess the actual
feasibility of all the possible CoMP schemes, both from a technical and also from an economical point of
view. If low-latency exchanges are needed between BSs, then CoMP schemes will not be able to be
deployed everywhere but only in the areas where a suitable backhaul is present, or they will require
important investments from the operator to upgrade the backhaul. This could somehow limit the practical usability of the techniques.
Either in coordinated beamforming and in joint processing case, the notion of clustering is key to
coordinating beamforming and relaxed coherent joint processing. Indeed, “CoMP” means that several
base stations share some knowledge about users. However, as the number of users and BSs increase, the
signaling overhead required for the inter-base information exchange and the amount of feedback needed
from the users also increase. Therefore, cooperation should be restrained to a limited number of BSs. To
achieve this goal, the network is thus divided into clusters of cooperative cells. Cluster selection is
obviously a key to cooperation algorithms performance. Cluster formation may be static [Ven07] [BH07],
if the clusters remain fixed in time, or dynamic [PGH08]. Selection may be performed by a central entity,
i.e. in a network-centric manner, or in a per-user way, i.e. in a user-centric manner. Usually, network-centric clustering divides the network into a set of disjoint cluster of BSs, that is, one BS can belong only
to one cluster [BHA08], [PGH08], [MF07a] and [MF07b]. In contrast, in the user-centric clustering
approaches, one BS may belong to more than one cluster, depending on the parameter under
consideration [PBG+04], [PBG+08]. From the user point of view, this means that, in a given cell, each
user may have a different set of cooperating BSs. The concept of clustering, as used in WINNER+, is
closely related to those of CoMP cooperative sets and/or measurement sets as defined in 3GPP
[3GPP36814].
These different approaches are all considered in this document. Further details of these studies are given
in Sections 2.2 and 2.3. In addition, work has been also conducted on different aspects related to relaying
in the framework of CoMP, as can be seen in Section 3. It should also be mentioned that CoMP issues
such as scheduling and Radio Resource Management (RRM) are only implicitly dealt with in this document. They are considered in [WIN+D1.5] where upper layers are the main focus.
WINNER+ D1,8
Version: 1.0 Page 10 (75)
2. Investigations on CoMP issues
As already mentioned above, the focus during year 2 is on CoMP solutions that have relaxed
requirements in terms of backhauling and complexity compared to the full blown joint processing
schemes that were studied during year 1.
The first area to be considered is transmission strategies in a Radio over Fiber (RoF) based CoMP architecture. In this study, different low-complexity transmission strategies in a distributed antenna
system based CoMP scheme are evaluated. More details can be found in Section 2.1 below.
The second area in focus is coordinated beamforming. Some different concepts for this are studied. Both
centralized and decentralized as well as non-codebook based and codebook based approaches are
investigated. Further details are given below in Section 2.2.
As to coherent multi-user multi-cell precoding techniques, reduced complexity approaches are proposed
in Section 2.3.
2.1 Assessment of transmission strategies in a RoF based CoMP architecture
2.1.1 Description
An example of a possible architecture of RoF based CoMP in a multi-cell environment can be a scenario
where a cell is covered by a RoF Central Unit (CU) and a number of distributed antenna modules called Remote Units (RU). The main processing modules are typically performed at the central unit which is
connected with the distributed antennas.
In traditional cellular systems the same area is covered by only a single high-power base station. In one
possible implementation of a cellular distributed CoMP system a certain number of antenna modules is
used to cover the same area, each adopting a lower power level, as illustrated in HFigure 2-1. The actual
number of distributed antenna modules would be determined by coverage, user densities, and other environmental factors. The entire cell coverage, as shown in Figure 2-1, is schematically represented by
an ideal circle containing smaller areas corresponding to the distributed nodes RU coverage spots. One
surrounding tier of six interfering cells has been considered in a unitary frequency reuse scheme,.
As it is known CoMP can include a centrally located or distributed processing among the collaborating
nodes, but at a first and simple level CoMP introduces multiple nodes transmissions schemes, also known as Distributed Antenna Systems (DAS). Evaluations regarding the distribution criteria of a CoMP scheme
are of high relevance, representing the starting point before evolving towards more efficient processing.
In particular, in distributed DAS CoMP, there are several possible transmission strategies using multiple
distributed antenna modules. Although many methods are possible, in the suggested proposal two
transmission strategies will be considered:
the power weighted transmission scheme
the single transmit selection scheme
The power weighted transmission scheme is based on the simultaneous transmission from the central unit
and all the distributed remote units, while maintaining constant the overall total transmitted power. The
most convenient power distribution scheme among the different nodes is chosen.
In the single transmit selection scheme, a certain method to choose from the nodes the one that should
transmit has to be found; a single node only is turned on depending on the outcome of the selected
method. Many different selection methods can be found in literature, such as maximizing the Signal to
Interference plus Noise Ratio (SINR) or capacity. A possible and simple scheme that will be adopted in
the following is the criterion of minimizing the propagation loss, that minimizes the required transmit
power (and hence the interference caused to other cells).
The use of distributed antenna systems is expected to reduce ICI and to improve SINR especially for
users near cell boundaries, which normally are performance limiting users, compared to conventional
cellular systems in an interference-limited multi-cell environment. As a result, distributed antenna
systems achieve lower symbol error probability and higher capacity than conventional cellular systems. It
will follow that distributed antenna architectures could appear to be one possible effective solution for
reducing ICI in an interference-limited cellular environment.
WINNER+ D1,8
Version: 1.0 Page 11 (75)
Figure 2-1: The Distributed Antenna System (DAS CoMP) scheme.
History Continuation of the work in IR1.2,
Section 5.2.1
Duplexing mode FDD and TDD
Clustering mode (1) Static in power weighted scheme and
dynamic in transmit selection scheme
Clustering mode (2) User-centric
Codebook-based No
Data exchanges: users data Data plane transferred to all RUs in
power weighted scheme and to best RU in transmit selection scheme
Data exchanges: Channels Impulse
Responses
No
Data exchanges : others
Data exchange rate: slow or fast
None
We will derive the general downlink system model valid for a network made of C cells, with M single
antenna mobile users per cell, that can be referred both to a traditional cellular network (in which only 1
transmitter per cell is used) and to a CoMP network (in which more than 1 transmitter per cell is used). In
the model formulation we will refer to a CoMP system made of R single antenna remote units.
In general, the downlink system model of a CoMP cellular network can be expressed as:
nxPHy (2.1)
where M,C R C H is the channel matrix,
C R,M C P is the CoMP processing matrix, which depends
on the particular chosen CoMP configuration, ,1
1 2, ,..,T M
Mx x x C x is a complex vector that
contains the transmitted signal, ,1
1 2 M ,y ,.., yT My C y is a complex vector that contains the M
signals received by the M mobile users, and ,1MCn is the noise vector. The element ( )k
ijh of the
matrix H is the complex channel gain between any of the mobile users i (with 1 i M ) and any of
the remote units j (with 1 j R ), belonging to any of the considered cells k (with 1 k C ). The
element ( )k
jip of the matrix P is the square root of the transmission power applied at any of the
remote units j (with 1 j R ), transmitting towards any of the mobile users i (with 1 i M ),
belonging to any of the considered cells k (with 1 k C ). The model of equation (2.1) makes the
hypotheses of a flat fading channel as it occurs for example for each subcarrier in an Orthogonal Frequency Division Multiple Access (OFDMA) based system. The model given by the equation (2.1) can
be easily extended to the case of remote units and mobile terminals equipped with multiple antennas.
In the following we derive, as a means of example, the expression of equation (2.1) for the particular case
of 2C cells in the network topology (the respective sub-matrixes are separated by dashed lines),
WINNER+ D1,8
Version: 1.0 Page 12 (75)
3R remote units equipped with a single antenna and 2M mobile users equipped with a single
antenna. In such hypothesis, equation (2.1) becomes:
2
1
2
1
)2(
32
)2(
31
)2(
22
)2(
21
)2(
12
)2(
11
)1(
32
)1(
31
)1(
22
)1(
21
)1(
12
)1(
11
)2(
23
)2(
22
)2(
21
)1(
23
)1(
22
)1(
21
)2(
13
)2(
12
)2(
11
)1(
13
)1(
12
)1(
11
2
1
2
1
)2(
)1()2()1(
2
1
n
n
x
x
pp
pp
pp
pp
pp
pp
hhhhhh
hhhhhh
n
n
x
x
y
y
P
PHH
(2.2)
where iy is the signal received at the i-th terminal and jx is the signal transmitted by the j-th remote
unit.
The particular expression of C R,M C P for CoMP transmit selection and for CoMP power weighted
schemes, when only one user per cell is considered, is as follows:
)2(
32
)2(
22
)2(
12
)1(
31
)1(
21
)1(
11
_
)2(
)1(
_
0
0
0
0
0
0
,
00
0
00
00
0
00
p
p
p
p
p
p
p
p
pwCoMP
ij
ij
selCoMPPP . (2.3)
In the above expression (2.3):
for the COMP transmit selection scheme (COMP_sel): the only non null term ( ) , 1,2k
jip for k is selected by the criterion, for instance, of minimizing the path-loss
(PL) between any of the remote units transmitting towards the considered mobile user: ( ) min (PL(j,i)), 1,2k
jij
p for k , for each given mobile user i ;
for the COMP power weighted scheme (COMP_pw): ( )
1
, 1,2R
k
ji BS
j
p P for k
.
where BSP is the total transmit power of the conventional base station.
2.1.2 State of the art
A first interesting paper by Choi-Andrews can be considered as reference ([CA07]) for DAS showing that
in addition to coverage improvements, DAS can also have potential advantages such as reduced power
and increased system capacity in a single cell environment. The paper analytically quantifies downlink
capacity of multi-cell DAS for two transmission strategies: selection diversity (where just one of the
distributed antennas is used) and blanket transmission (where all antennas in the cell broadcast data). In
particular, in the paper, the results have been drawn out analytically under some assumptions: the ergodic
capacity of cellular DAS versus the normalized distance from the home base station is derived when the CSI is known at the receiver end. In deriving the results, six RUs have been considered in the DAS
configuration. The transmit power of each distributed antenna module is a fraction of the transmit power
of the home base station. No fast fading modelling is considered when deriving the results. The
conclusions derived from this paper are that the single transmit selection scheme achieves the highest
throughput owing to the ICI reduction and macroscopic selection diversity.
WINNER+ D1,8
Version: 1.0 Page 13 (75)
A second interesting paper can be considered as reference [SWO08]. In this paper the performance of a
DAS under time-varying frequency-selective fading based on a realistic channel model is investigated.
Specifically, by shifting the hexagonal cellular layout in the conventional system and using sectorized
antennas instead of omnidirectional antennas at each BS, the performance in terms of outage probability
and outage capacity improves by a large extent without a need for additional BS towers. The results in the
paper show that with the same total transmit power and bandwidth, DAS can reduce the ICI in a multi-
cell environment and improve the outage capacity especially near the cell boundary.
2.1.3 Expected performance or benefits
In this section some results of the performance of the above mentioned CoMP systems, obtained by
means of computer simulations, will be provided. A dedicated software tool has been developed. In order
to calibrate such a tool and to gain some initial insight in the DAS scheme, preliminary simulations have
been run aligning to the hypothesis used in [CA07]. The propagation effects encountered between either
the BS or the RUs towards the UE are modelled in the simulator and include path loss, shadowing and
fast fading. The path loss for a given distance is calculated by means of the Walfish-Ikegami analytical
model. The probability distribution function of the shadowing is assumed lognormal, so that the
shadowing expressed in dB can be modelled as a Gaussian random variable. Finally the fast fading is
modeled by randomly extracting for each user and in each simulation snapshot one channel matrix. As a
first approximation the channel coefficients have been considered independent. This analysis quantifies downlink capacity of multi-cell CoMP (calculated based on the ergodic Shannon capacity) for the two
considered transmission strategies: transmit selection scheme (where just one of the distributed antennas
is used) and power weighted scheme (where all antennas in the cell broadcast data). In particular, in the
paper, the results have been drawn out analytically under some assumptions:
the CSI is known at the receiver end.
six RUs have been considered in the DAS CoMP configuration and each RU is evenly located
on a common circle with radius cRr3
2 , where cR is the cell radius.
the transmit power of each distributed antenna module is 0.1P and the transmit power of the
home base station is 0.4P in DAS CoMP whereas the transmit power of the base station in the
conventional cellular system is P .
in the case of the power weighted scheme, the signals transmitted by the different RUs are
combined non coherently at the receiver end.
Starting from the same assumptions the simulations to derive a first estimation of the overall capacity in
the scenario of RoF based DAS CoMP have been performed. In the following HFigure 2-2, the ergodic
capacity versus the normalized distance from the home base station has been evaluated by means of
simulations. The single user capacity has been depicted as a function of the normalized distance of the UE
from the home base station moving from the cell center 0r to the cell edge 1r .
From the simulation results, it is clear that the transmit selection scheme achieves the highest throughput
due to the ICI reduction and macroscopic selection diversity. Although the achieved throughput of the
power weighted scheme in the cellular DAS is slightly lower than that of conventional omni-directional
cellular system near the home base station, due to reduced transmit power, the achieved throughput of the
power weighted scheme in DAS has substantially higher throughput beyond the normalized distance of
around 2
1r , and obtaining its maximum value in correspondence to the position of the RUs at
3
2r .
The presence of the latter local maximum (see Figure 2-2), in the case of DAS system deployment, is due
to the fact that when the user is moving towards the cell edge, the presence of the RUs (positioned at
3
2r ) enables to achieve a better throughput performance near the RU positions with respect to what is
achievable with a classic omnidirectional system.
WINNER+ D1,8
Version: 1.0 Page 14 (75)
Figure 2-2: Ergodic capacity vs. the normalized distance from the home base station
2.1.4 Expected requirements on signalling and measurements
The proposed evaluation is based on theoretical distributed schemes aiming at the most convenient in
terms of capacity of the system. These schemes could require exchange of information in order to perform
the best selection, both in the case of power weighted or single transmit selection schemes; however, the
overall amount of exchanged data would be very small and suited with the capacity offered by RoF
architectures, to be considered as straightforward enablers.
2.1.5 Expected requirements on architecture and protocols
The presented schemes are quite generic, but they are studied in order to be applied to RoF architectures.
As a consequence the foreseen impacts are those of RoF schemes introduction, widely described in
[WIN+D14].
2.2 Coordinated beamforming concepts
Assuming linear transceiver processing, a CoMP system with N antennas is ideally able to accommodate
up to N streams without becoming interference limited. The inter-stream interference can be controlled or
even completely eliminated by a proper precoder selection. This is especially true in the coherent joint
processing case, where user data is conveyed from multiple BS antenna heads over a large virtual
Multiple Input Multiple Output (MIMO) channel.
The coherent multi-user multi-cell precoding techniques, however, have high requirements in terms of
signalling and measurements. In addition to the complete channel knowledge of all jointly processed links, a tight synchronisation across the transmitting nodes and centralized entities performing scheduling
and computation of joint precoding weights is required in order to avoid carrier phase drifting at different
transmit nodes. A large amount of data needs to be exchanged between the network nodes. Thus, high-
speed links, such as optical fibres or dedicated radio links, are needed.
Another form of coordinated transmission is a dynamic multi-cell scheduling and interference avoidance,
where the network nodes coordinate their transmissions (precoder design, scheduling) in order to
minimize the inter-cell interference. The carrier phase coherence between the transmit nodes is not
required, since each data stream is transmitted from a single transmission point. Thus, the non-coherent
coordinated multi-cell transmission approaches have somewhat looser requirements on the coordination
WINNER+ D1,8
Version: 1.0 Page 15 (75)
and the backhaul, but could potentially still need centralised resource management mechanisms. This
family of methods we refer to as coordinated beamforming.
Coordinated beamforming can be carried out in different ways, some examples are given in [VAL+06],
[SCW+07], and [YSL+08]. In WINNER+, both centralized and decentralized as well as non-codebook
based and codebook based approaches are investigated.
One of the earliest studies that can be considered to fall into the category of coordinated beamforming
considered the minimum power beamformer design, also known as the sum power minimization under
the minimum SINR constraint per user. This problem has been extensively studied over the past decade.
In order to solve this problem for single-antenna users, the uplink-downlink SINR duality was utilized in [RLT98] and [VM99] [SB04] for multi-cell and single-cell cases, respectively. The duality property was
utilized to develop iterative algorithms for calculating the optimal beamformers and power allocations.
Furthermore, an optimal solution for the problem of maximizing the jointly achievable SINR margin
under a total transmitted power based on the uplink-downlink duality was proposed in [SB04]. A
modified form of the single-cell multiuser beamforming algorithm in [SB04] is presented in section 2.2.1
for coordinated beamforming concept, in which the beamforming vectors and power allocation are found
to maximize the jointly-achievable SINR margin under per-transmitter power constraints
It was shown in [BO99] and [BO01] that the minimum power beamforming problem can be formulated as
a second order cone program (SOCP) [BV04] for rank one channels (a single receive antenna per user or
fixed receiver beamformers). A coordinated multi-cell scenario was considered, where all the transceivers are jointly optimized while considering the other-cell transmissions as inter-cell interference. Similarly to
[BO99] and [BO01], the power minimization problem was cast as an SOCP in [WES06]. In addition, an
efficient algorithm based on fixed point iterates was proposed. Furthermore, the worst case SINR was
maximized subject to a total power constraint based on generalized eigenvalue problem in [WES06]. The
SINR optimization problem can also be carried out via power optimization using bisection method. A
similar approach based on bisection method was proposed in [SVH06].
The solutions above need complete channel knowledge between all pairs of users and BSs, and hence,
they require centralized resource management mechanisms. It is fair to assume that each BS can measure
at least the channels of all cell edge users, independent of which BS they are identified with, for example,
during the uplink (UL) transmission phase of the Time Division Duplex (TDD) frame. In such a case,
each BS could simply form nulls (zero forcing) towards a set of users served by other BSs while optimizing the transmission for the set of served users. In a more general form of operation, each BS can
employ less restrictive interference balancing criteria (allowing some controlled interference), and take
that interference into account when designing the precoders in the adjacent BSs. This obviously requires
some extra signalling across the backhaul network but may result in improved performance in certain
scenarios.
An efficient iterative algorithm based on uplink-downlink duality was introduced in [DY08]. The multi-
cell minimum power beamformer design problem was solved via a dual uplink problem, where the
downlink beamformers are designed locally based on the reciprocal uplink channels and virtual uplink
powers of all users. This allows also for a distributed implementation, where virtual uplink powers are
exchanged between BS nodes in a coordinated manner. An alternative distributed method is presented in section 2.2.2 in which the optimal minimum power beamformers can be obtained locally at each BS
relying on inter-cell interference terms exchange between adjacent BSs.
An alternative scheme compared to those previously reported makes use of a report from the UE about
the beams that the neighbouring BSs should avoid or favour in order to limit the interference. Such an
approach is only possible when codebook-based beamforming is employed, so that the beams that a
potential interferer can use are known in advance at the UE. The report of the interfering beams requires
only a modest increase in feedback overhead since only the index of the beam(s) needs to be reported.
Nevertheless, low-latency communication links between the coordinated BSs are required in order to
exchange the information about the served UEs at a given time instant. This concept has been presented
in [WIN+D14], and is further investigated in section 2.2.3.
WINNER+ D1,8
Version: 1.0 Page 16 (75)
2.2.1 Centralized non-codebook based coordinated beamforming
The first coordinated beamforming concept in this deliverable assumes fixed coordination clusters, where
the coordination is carried out in a centralized fashion. The coordination has the following main
characteristics:
(i) Transmit powers and beamforming weights are jointly adapted for all UEs in the coordination
cluster in order to minimize caused interference,
(ii) Non-coherent, i.e. non-frequency dependent: beamforming weights are constant for all
frequencies,
(iii) Each UE receives transmissions from a single transmission point, i.e. no cooperative joint
transmission from multiple transmission points takes place.
(iii) Spatial Division Multiple Access (SDMA): one transmission point serves multiple UEs as in
multiuser MIMO.
Hence, this coordinated beamforming scheme can simply be viewed as coordinated multiuser MIMO
where the beamforming weights are jointly adjusted for all transmission points in the coordination cluster
in order to minimize the caused interference.
History N/a
Duplexing mode FDD or TDD
Clustering mode (1) Static (but can also be implemented in a
dynamic fashion)
Clustering mode (2) Network centric
Codebook-based No
Data exchanges: users data No
Data exchanges: Channels Impulse
Responses
No
Data exchanges : others
Data exchange rate: slow or fast
Transmit channel correlation, precoding
weights
Slow rate
2.2.1.1 Description
Consider a coordination cluster with N transmission points, each equipped with an antenna array
comprising TM antenna elements, and K UEs equipped with RM receive antennas. Assuming
Orthogonal Frequency Division Multiplexing (OFDM) transmission, the baseband signal received by UE i for frequency f in the downlink is given by
( ) ( ) ( ) ( )i i iy f H f x f w f (2.4)
where the R TM NM matrix )( fH i is the composite channel between all transmission points and UE i;
)( fwi represents thermal noise, including inter-cluster interference which is assumed to be white
Gaussian distributed with covariance 2( ) ( )
H
f i i iE w f w f I ; and )( fx is an 1TNM vector
representing the sum signal transmitted from all transmission points, given by
K
iiii
K
ii
fsupfxfx11
)()()( (2.5)
where ( )is f is a modulation symbol (drawn from a unit-variance symbol alphabet) which is transmitted
to UE i using the transmit power and beamforming vectors ip and iu , respectively. The beamforming
vectors are normalized to have unit power
1H
i iE u u for i =1,2,…K (2.6)
The average SINR for UE i, assuming Maximum Ratio Combining (MRC) reception, is given by
2
),(SINRi
ikki
H
kk
ii
H
ii
iuRup
uRuppU
(2.7)
WINNER+ D1,8
Version: 1.0 Page 17 (75)
where U is the matrix of transmit beamformers given by KuuuU ,, 21 , p is the vector of transmit
powers, and iR is the transmit correlation matrix for UE i, given by
Niiii
H
ii RRRdiagfHfHR ,2,1,f ,,,)()(E . (2.8)
since the transmission points are separated by large distances and therefore their antennas are mutually
uncorrelated.
The considered coordinated beamforming scheme is a modified form of the multiuser beamforming
(MuBF) algorithm in [SB04] in which the beamforming vectors and power allocation are found to
maximize the jointly-achievable SINR margin:
i
i
KipU
pUC
),(SINRminmax
],,2,1[, (2.9)
under per-transmitter power constraints
NnPpnSi
i ,2,1,max
where i is the target SINR for UE i, nS is the set of UEs connected to transmission point n and maxP is
the maximum transmission point transmit power.
The beamforming vectors and downlink power allocations are found iteratively using the uplink-
downlink duality theorem [RLT98] [VM99] [SB04] which states that the downlink broadcast channel has
a virtual dual uplink multiple access channel which has the same SINR achievable regions as the
downlink and the same beamforming vectors achieve the SINRs in both links.
2.2.1.2 Expected performance or benefits
As for most beamforming schemes it can be expected that the most appropriate scenarios for coordinated
beamforming have low angular spreads which are found in urban, suburban and rural environments with above roof-top antenna deployment. Due to low feedback requirements (see below) coordinated
beamforming is especially interesting in Frequency Division Duplex (FDD) mode, where joint processing
might be difficult to realize in a practical system due to the high feedback requirements. It can also be
expected that coordinated beamforming is robust to user mobility, hence making it an interesting
alternative both in low and high mobility scenarios.
A detailed system level performance evaluation of the herein described coordinated beamforming scheme is provided in Appendix A.1, where its performance is also compared to that of joint processing based on
zero-forcing (ZF) precoding that was evaluated in Appendix B.7 in [WIN+D14]. A summary of the
results provided in Appendix A.1 is given in Table 2-1.
Table 2-1: Summary of system level performance results
Scenario Transmission scheme Cell spectral efficiency
[bps/Hz/cell]
Cell edge user spectral
efficiency [bps/Hz]
3GPP Case 1
No CoMP 2.56 0.074
Joint processing based on ZF
3.81 0.108
Coordinated
beamforming
3.01 0.078
ITU Urban Macro
No CoMP 1.32 0.035
Joint processing based
on ZF
1.20 0.026
Coordinated
beamforming
1.97 0.053
ITU Rural Macro
No CoMP 1.52 0.050
Joint processing based
on ZF
1.05 0.019
Coordinated
beamforming
2.64 0.071
WINNER+ D1,8
Version: 1.0 Page 18 (75)
As can be expected, in the low mobility scenario 3GPP Case 1 [3GPP25814] joint processing based on
ZF performs very well; the gain over the reference non-CoMP system based on LTE release 8 is almost
50% both in cell spectral efficiency as well as cell edge user spectral efficiency. Coordinated
beamforming does also perform reasonably well here; the gain over the non-CoMP is 15-20% in cell
spectral efficiency, while the cell edge performance is only slightly better.
In scenarios with higher user mobility, e.g. International Telecommunications Union (ITU)Urban Macro
[ITURM2135] where the users are moving in vehicles at 30 km/h, it can be seen that the performance of
joint processing based on ZF degrades and is actually slightly worse than that of the reference non-CoMP
system. This is of course due to the fact that short-term channel state information needed for the ZF
precoding gets outdated, hence resulting in that the applied precoding weights are invalid. Coordinated
beamforming, on the other hand, is robust to user mobility and works well in this scenario. The gain over
the non-CoMP system is in the order of 50%, both in cell spectral efficiency and cell edge user
performance. The relative gain over the non-CoMP is higher than in 3GPP Case 1, which most probably
is explained by the fact that we now have outdoor users, and also a line-of-sight (LoS) component in the
channel model, which together makes it easier to exploit the directivity properties of the beamforming.
Finally, in the ITU Rural Macro scenario where the users are moving at 120 km/h, it can be seen that the
joint processing based on ZF breaks down even further, and that the performance now is far below that of the non-CoMP system. Again, it is demonstrated that coordinated beamforming is robust to user mobility,
and the gain over the non-CoMP system is now 70% in cell spectral efficiency and about 40% in cell edge
user spectral efficiency.
2.2.1.3 Expected requirements on signalling and measurements
The considered coordinated beamforming scheme requires transmit channel correlation estimates for
antenna arrays at all transmission points in the coordination cluster. These estimates can be obtained by
means of UE-specific uplink sounding reference signals. In case of TDD it is straightforward to apply the
estimated correlation matrices for downlink transmission, while for FDD it may be necessary to perform a
frequency translation of them. Methods for this can be found in e.g. [CHC04]. The uplink sounding
reference signals can be broadband as frequency-dependent channel information is not needed to estimate
transmit channel correlation. Frequency-dependent channel information is only needed for the serving
transmission point for channel-dependent link adaptation. The reference signals may have low time
density as transmit channel correlation does not change with fast fading.
2.2.1.4 Expected requirements on architecture and protocols
Since each UE receives data from a single transmission point, there is no need to share user data between transmission points. The information that need to be exchanged between the transmission points and the
central processor are the transmit channel correlations that are estimated at each transmission point, and
the resulting computed precoding weights. Since the transmit channel correlation does not change with
fast fading and the precoding weights are non-frequency dependent, the capacity requirements on the
backhaul connection is rather limited, however, the latency has to be sufficiently low in order to cope
with scheduling updates.
2.2.2 Decentralized non-codebook based coordinated beamforming
The second concept is a distributed solution for the coordinated multi-cell multi-antenna minimum power
beamformer design problem with single-antenna users [TPK09b] [TPK09c]. The minimum power
beamformers are obtained locally at each BS relying on limited backhaul information exchange between
adjacent BSs. Hence, this concept operates in a decentralized manner in contrast to the concept described
in Section 2.2.1 above.
The original minimum power beamformer design problem is reformulated such that the BSs are coupled
by real-valued inter-cell interference terms. The coupled interference terms are handled by taking local
copies of the terms at each BS and enforcing consistency between them. Thus, the coupling in the interference terms is transferred to coupling in the consistency constraints, which can then be decoupled
by a standard dual decomposition approach leading to a distributed algorithm. The proposed method is
able to guarantee feasible solutions even if the interference information is outdated or incomplete, at the
possible cost of increased sum power. In addition, the proposed approach allows for a number of special
cases, where the backhaul information exchange is reduced at the cost of somewhat sub-optimal
performance.
WINNER+ D1,8
Version: 1.0 Page 19 (75)
History New
Duplexing mode TDD
Clustering mode (1) Static
Clustering mode (2) Network centric
Codebook-based No
Data exchanges: users data No
Data exchanges: Channels Impulse
Responses
No
Data exchanges : others?
Data exchanges rate: slow or fast
Real-valued inter-cell interference
terms
Fast rate
2.2.2.1.1 Description
Consider a cellular system that consists of BN BSs with TN transmit antennas and K user each with
single receive antenna. A set U with size UK includes all users active at the given time instant,
while a subset Ub U includes the users bk U allocated to BS b , B,...,1 Nb . The index of serving
BS for user k is denoted as kb . The signal ky received by the user k consists of the desired signal,
intra-cell and inter-cell interference, and it can be expressed as
k
bb k
kbkbkb
ki
ibkbkbkbkbkbk naaay
k bkb
kkkkkk
U
,,,
\U
,,,,,, xhxhxh
where the vector TC,N
kb x is the transmitted signal from the b ’th BS to user k , kn ~ ),0( 0NCN
represents the additive noise sample with noise power density 0N , and T1,, C
Nkbkba
h is the channel
vector from BS b to user k with large-scale fading coefficient kba , . The elements of kb,h are
normalized to have unitary variance. The transmitted vector for user k is generated at BS b as
kkbkb d,, mx , where TC,N
kb m is the pre-coding vector and kd is the normalized complex data
symbol.
Let TC,N
kb m be an arbitrary transmit beamformer for the user k from BS b . By denoting the inter-
cell interference term from BS b to user k as kb,ζ , and relaxing the term as 2
U
,,,2,ζ
bi
ibkbkbkb a mh ,
the SINR formula can be written as
K
ki
ibkbkb
bb
kb
kbkbkb
k
kb
kkk
k
kkk
aN
a
\U
2
,,,2
,0
2
,,,
ζ mh
mh (2.10)
where the index of the serving BS for the user k is denoted as kb .
The system optimization objective is to minimize the total transmitted power subject to fixed user-
specific SINR constraints kk . This problem can be formulated as
B
U
2
2,
2,
2
U
,,,
k,
1
,...,1,
,Uζ
k s.t.
min B
NbP
bka
P
b
k
kb
bkb
i
ibkbkb
k
N
b
b
b
b
m
mh
(2.11)
where the variables are B,...1, NbPb , KkN
kb ,...1,C T, m and bk bkb ,Uζ , . Vector ζ is
defined including all the inter-cell interference terms as
WINNER+ D1,8
Version: 1.0 Page 20 (75)
T
UU,1U,UU1,1U1,BBBBB111
ζ,...,ζ,...,ζ,...,ζ
NNN NNζ , where bb k UU . The second constraint
guarantees that the interference generated from a given BS b cannot exceed the user specific thresholds
bkb k Uζ , .
Observe that BSs are coupled in the SINR constraints by the terms kb,ζ . If the interference terms were
fixed, the problem would decouple and the transmitted power could be separately minimized at each BS.
The coupled SINR constraints are addressed by introducing local auxiliary variables bζ , as well as,
additional consistency constraints that require the local variables to be equal. This results in the following
optimization problem:
b
NbP
bka
bk
P
b
b
k
kb
bb
kb
i
ibkbkb
bkb
k
N
b
b
b
b
,
,...,1,
,Uζ
U , s.t.
min
B
U
2
2,
2,
2
U
,,,
1
B
ζζ
m
mh
(2.12)
where the variables bP , KkN
kb ,...1,C T, m and b
ζ are local for each BS b , and bk is as k but
with bkb,ζ instead of kb,ζ .
In order to obtain a distributed algorithm, a dual decomposition approach is taken where the consistency
constraints are relaxed by forming the partial Lagrangian as
B
U
2
2,
2,
2
U
,,,
1
T
1
,...,1,
,Uζ
U , s.t.
min BB
NbP
bka
bk
P
b
k
kb
bb
kb
i
ibkbkb
bkb
k
bN
b
b
N
b
b
b
b
m
mh
ζζν
(2.13)
where the variables are B,...1, NbPb , KkN
kb ,...1,C T, m and b
ζ , and bν are consistency prices
for each BS b . For fixed bν , the distributed version of the problem reduces to
B
U
2
2,
2,
2
U
,,,
T
,...,1,
,Uζ
U , s.t.
min
NbP
bka
bk
P
b
k
kb
bb
kb
i
ibkbkb
bkb
k
bbb
b
b
m
mh
ζν
(2.14)
where the variables are the same as in the previous problem. The resulting convex sub-problems can be
locally and independently solved as SOCPs in each BS b with the knowledge of consistency prices bν
(see, e.g., [TPK09a] for more details). The master problem for the dual decomposition is
B
1
T maximize
N
b
bbbg ζνν (2.15)
where the variables are bbν , and where bbg ν is the dual function achieved as the minimum value of
the partial Lagrangian for given bν . This can be solved iteratively with the following updates [Boy04]
[PC06]:
btttt bbb ,1 ζζνν (2.16)
WINNER+ D1,8
Version: 1.0 Page 21 (75)
where t is the iteration index, is a positive step-size and B
B...1 1B N
NttNt 1ζζζ is the
average of all BS-specific inter-cell interference vectors, which need to be exchanged between coupled BSs. Since the original problem is convex, the algorithm is guaranteed to converge to the optimal
(centralized) solution, where bb ζζ , if the step size is sufficiently small [Boy04] [PC06], and the
channels kbkb ,, h remain constant during the iterations.
The intermediate iterates tbζ in the dual decomposition do not necessarily result in a feasible
solutions, i.e., kk for some k . However, a feasible set of beamformers kb,m can be always
achieved by using the average vector tζ for each BS. In this case, one additional sub-problem per BS
has to be solved. Thus, a feasible set of kk can be guaranteed even if the update rate of tbζ
between BSs is slower than the channel coherence time, at the possible cost of increased sum power.
The dual decomposition approach allows for a number of special cases, where the backhaul information
exchange is reduced at the cost of somewhat sub-optimal performance. Some possible scenarios are listed below:
BS-specific inter-cell interference constraint, bbkb k Uζζ ,
One common constraint for all BSs (within a given joint processing area), bkkb ,ζζ , .
Zero-forcing for the inter-cell interference, bkkb ,0ζ , .
In time-correlated fading with high mobility and/or with low backhaul information exchange rate the
interference terms kb,ζ may become quickly outdated. This may result in a high peak in the transmitted
power as the interference constraints are mismatched with the actual channel realizations. The ZF solution
( bkkb ,0ζ , ) can be always calculated in each BS based on the available local information. In some
time instants, the ZF solution would result in a lower sum power than the distributed solution relying on
the backhaul information exchange. In such a case, BS b simply sets ZF−mode(b) state active and sends
a message to its neighbours. Consequently, the operation is reverted to the normal mode as soon as the
resulting transmitted power is again below the ZF mode. The distributed algorithm with ZF mode
selection is summarized in Appendix A.2.
2.2.2.2 Expected performance or benefits
A detailed performance evaluation of the proposed concept is provided in Appendix A.2. Some of the
main results from Appendix A.2 are presented in this section. Table 2-2 presents the average sum power
of 4,2,4,, TB NNK system for 0 and 10 dB SINR target per user. Different coordinated
beamforming cases and two zero-forcing approaches are compared with coherent multi-cell beamforming
case at the cell edge.
Table 2-2: Main performance results of decentralized coordinated beamforming
Transmission scheme Average sum power
[dB]: 0dB SINR target
per user
Average sum power
[dB]: 10dB SINR target
per user
Coherent multi-cell
beamforming
-1.45 9.71
Coordinated beamforming: user-
specific interference constraint
3.90 19.44
Coordinated beamforming: BS-
specific interference constraint
4.20 19.96
Coordinated beamforming:
common interference constraint
4.88 23.96
Zero-forcing for inter-cell
interference
8.92 24.81
Zero-forcing for both intra- and
inter-cell interference
15.21 25.21
As expected, the coherent multi-cell beamforming greatly outperforms the coordinated beamforming
cases at the cell edge. All the three coordinated beamforming cases with inter-cell interference constraints
have very similar performance. Thus, the loss from sub-optimal signaling is minor. The coordinated
beamforming cases require about 5-6 dB more power than the coherent case in order to meet the 0 dB
WINNER+ D1,8
Version: 1.0 Page 22 (75)
SINR target. There is a large gain from the optimal intra-cell beamformer design (zero-forcing for inter-
cell interference) as compared to the channel inversion (zero-forcing for both intra- and inter-cell).
The loss from sub-optimal signaling increases significantly for the case with one common constraint
when the SINR target is 10 dB. However, the loss is still minor in the BS-specific constraint case. Also,
the gain from the optimal intra-cell beamformer design is greatly reduced as compared to the channel
inversion. In general, the difference between the zero-forcing and coordinated beamforming cases with
inter-cell interference constraints is reduced significantly. A more detailed performance evaluation with
wide range of numerical examples is provided in Appendix A.2.
2.2.2.3 Expected requirements on signalling and measurements
The objective of this proposal is to reduce the traffic required between the cooperating BSs so that full
CSI need not be shared and that precoders are computed locally by each BS. Information about the
allowed inter-cell interference levels taken in different cells must be exchanged between adjacent BSs.
However, a master-slave structure between adjacent BSs might be required if sophisticated user-to-BS allocation algorithms are used.
2.2.2.4 Expected requirements on architecture and protocols
The studied concept does not require tight frequency synchronization or sharing of user data between
BSs. Furthermore, the required information exchange rate between BSs can be slow if the channels are
slowly varying [TPK09b]. However, it is anticipated that existing 3GPP X2 type backhaul interface might
not be able to support the requirements of this concept.
2.2.3 Codebook-based coordinated beamforming
Coordinated beamforming aims at avoiding collisions of beams originating from neighbour cells. The
third coordinated beamforming concept studied in Winner+ is based on the codebook-based precoding
which already exists in LTE Rel-8. The main idea is to make use of a report from the UE of a Precoding
Matrix Index (PMI) indicating either the most interfering (MI) beam, or the least interfering (LI) beam
received from an interfering cell. The serving cell communicates to coordinated interfering cells the time-
frequency resources that will be used for transmission to the scheduled UE, together with the MI/LI-PMI
reported by this UE. These pieces of information will then act as constraints for the coordinated cells'
schedulers, which should try as much as possible to avoid/favour the reported MI/LI-PMI on the
associated resources for their own transmission.
This concept was presented in section 2.1.3.1.2 in [WIN+D14]. This section proposes a practical solution
to implement this concept in a decentralized way (i.e. without a central control entity), and presents
preliminary results to assess its performance.
The table below summarizes the features of the concepts that are relevant for the system concept work.
History Continuation of the work in D1.4,
Section 2.1.3.1.2
Duplexing mode FDD and TDD
Clustering mode (1) Dynamic
Clustering mode (2) User-centric
Codebook-based Yes
Data exchanges: users data No
Data exchanges: Channels Impulse
Responses
No
Data exchanges : others
Data exchanges rate : slow or fast
Request to be master including an indication of the resources on which the
UE will be served, and the PMIs to
avoid/favour.
Each time a cell-edge UE is scheduled
2.2.3.1 Description
The proposed solution is described in the following, with a particular emphasis on the following points:
UE measurements and reports
WINNER+ D1,8
Version: 1.0 Page 23 (75)
Cluster formation
Implementation of scheduling restrictions
UE measurements and reports
Each UE determines
the PMI from its serving cell that maximizes the useful signal received power
the PMI(s) from its dominant (long-term) interfering cells that either
o maximize the interference power. This type of PMI will be called Most Interfering PMI
(MI-PMI) in the following;
o or minimize the interference power. This type of PMI will be called Least Interfering
PMI (LI-PMI) in the following.
If only single-layer beamforming is considered, it can be shown that the MI/LI-PMI from cell I
maximizing/minimizing the interference power after MRC reception can be determined as follows:
2)()()()((I) minargmax/argPMIMI/LI SSHIHI
ii
uHHv
Where )( I
iv is the i-th precoding vector at interfering cell I, )( I
H is the channel matrix from interfering
cell I, )(S
H is the channel matrix from the serving cell and )(S
u is the preferred precoding vector from
the serving cell.
At least one LI/MI-PMI is reported to the serving cell for each dominant interferer (the actual number of
interferers to consider being possibly configured by the network). Several MI/LI-PMIs can be reported in
order to provide more flexibility/information to the scheduler, at the price of an increase radio feedback
overhead.
In addition, the UE reports periodically to its serving cell the cell identifiers of its dominant interferers.
The dominant interferers can be determined on a long-term basis in order to limit the feedback overhead.
Cluster formation
The clusters are formed as follows: when a cell scheduler selects a UE identified as being at the cell edge
(which can be determined from UE measurements, e.g. the Reference Signal Received Quality
standardized for LTE), the cell sends a message to the cells reported by the UE as being the most
interfering, in order to request to become their master. Such a message will be referred to as a "master
request" in the following.
The master request is accompanied with the index of the beam whose use should be avoided/preferred at
the slave cell, as well as the indication of the resources for which the constraint will apply. Note that the
master requests can be sent several transmission time intervals (TTIs) in advance if the latency of the
communication links between the BSs is not low enough. This imposes that the scheduler takes decisions
several TTIs before the actual transmission. Note that such delay still allows channel-dependent
scheduling provided the UE mobility is low, which is the situation primarily envisioned for CoMP in
LTE-Advanced.
Obviously, the master/slave role of a cell can depend on the frequency resources: on some resources the
cell could be master, whereas it would be slave on some others. For the sake of simplicity, we consider
implicitly only one set of resources in the following, where a cell can be either master over all the
resources, or slave over all the resources.
In case of reception of master requests from several potential master cell, or if a candidate master receives
a master request from another cell, then a contention resolution procedure has to be applied in order to
determine the master/slave role of each cell. Such a contention procedure is out of the scope of this study,
but it should be designed in order to ensure fairness between the cells.
Note that although it is possible that one cell be the slave of several masters, increasing the number of
masters increases the constraints on the scheduler, which may severely limit the scheduler flexibility and
thus impact the system performance. Therefore, in this study we assume that one cell can be slave of only
one master.
Implementation of scheduling restrictions
Once the master/slave roles have been established, the scheduler of the slave cell applies the restriction
communicated by the master cell:
WINNER+ D1,8
Version: 1.0 Page 24 (75)
If the coordination mode is so that the most interfering beam(s) is indicated to the slave cell, the
latter tries as much as possible to serve a UE that requests to be served in another beam.
If the coordination mode is so that the least interfering beam(s) is indicated, the slave cell tries as
much as possible to serve a UE that requests to be served in this beam.
The most interfering beam avoidance constraint is the easiest to satisfy, since it puts the lowest
requirements on the scheduler: consider a codebook with 16 codewords; if one is forbidden, 15
codewords remain allowed. It is therefore much easier to find a UE requesting to be served in one among
15 beams, than one requesting to be served in one particular least interfering beam. If the number of
reported beams is increased (at the price of a larger feedback overhead), the feasibility of satisfying the
two constraints gets closer. Nevertheless, the most interfering beam approach remains more advantageous
even in the case of several reported beams, since in case it is not possible to find a UE willing to be served in one of the allowed beams, relaxing the constraints (e.g. forbidding only the N dominant beams
out of M reported beams, with N<M) will still provide some gains compared to no coordination. On the
contrary, such a constraint relaxation is not possible with the least interfering beam approach.
In the case where no UE can be found to meet the scheduler constraints (even with constraints relaxation),
two options are possible:
1. Ignoring the constraints and serving any UE, like without coordination
2. Serving no UE at all, i.e. silencing the cell on the coordinated resources.
Option 1 leads to the same instantaneous performance as in the no coordination case, whereas option 2
ensures the protection of cell-edge UEs, at the price of a potential cell throughput reduction since some
resources are not used part of the time.
Additional considerations
If Resource Quality Indication Reference Signal (RQI-RS) is used, as proposed in [3GPP09], no
confirmation/acknowledgement is needed from the slaves to the master in order to allow for proper link
adaptation.
The coordination process requires no central control entity since it relies on a master/slave principle
between cells. Moreover, even negotiation between the cooperating cells is not necessary provided an
efficient contention resolution mechanism is used, thus enabling a low latency of the coordination
mechanism. Indeed, in that case the cell receiving the master request either accept the request and
behaves as requested, or do not accept the request: the performance will then be identical to the no
coordination case.
2.2.3.2 Expected performance or benefits
This section provides preliminary results about the gains of coordinated beamforming according to the
proposed method versus uncoordinated beamforming. The considered scenario is similar to 3GPP case 1
[3GPP25814] with full buffer traffic model. A simplified class III (Snapshot/quasi-static based) system-level simulator has been used, with the following simplifications:
Only Time Division Multiple Access (TDMA) is modelled (i.e. no OFDMA, or OFDMA with
only one physical resource block);
Frequency-flat fast fading channel coefficients are randomly drawn according to a unit-variance Rayleigh law, with no spatial nor time correlation;
No HARQ retransmission;
Perfect link adaptation;
Perfect channel estimation;
The user instantaneous spectral efficiency is computed from the post-receiver SINR via the
Shannon formula.
More details on the simulation conditions are given in Appendix A.3.
The table below summarizes the results obtained for the most interfering beam approach, for 1 and 3 MI-
PMI. Note that feedback of 1 and 3 MI-PMI for 3 coordinated cells leads to multiply the PMI feedback
overhead of LTE Rel-8 by 3 and 7, respectively. In case no UE is found to meet the scheduler constraints,
the constraints are ignored.
WINNER+ D1,8
Version: 1.0 Page 25 (75)
Table 2-3: Performance results for codebook-based beamforming (CBF) (MI-PMI approach).
No coordination CBF with 1 MI-PMI CBF with 3 MI-PMI
Cell spectral efficiency (Bit/s/Hz) 4.49 4.43 (-1%) 4.62 (+3%)
Cell-edge user spectral efficiency
(Bit/s/Hz), measured at the 5%-
tile of the user throughput
cumulative distribution function
(CDF)
0.16 0.18 (+13%) 0.19 (+18%)
Coordinated beamforming with the proposed scheme provides only moderate gains in cell-edge
performance compared to non-coordinated beamforming in a realistic coordination setup (3 coordinated
cells, no central control entity): +13% for 1 MI-PMI and +19% for 3 MI-PMI. No significant change in
cell throughput is observed. Note that these results are only preliminary, and in particular use a very
simple scheduler. A more complete performance evaluation using a full LTE-Advanced simulator will be
reported in deliverable [WIN+ D4.2].
2.2.3.3 Expected requirements on signalling and measurements
The requirements on signalling and measurements are as follows:
Measurements required at the UE:
o identification of its N (configurable by the network) dominant long term interferers –
can be updated with a low frequency, depending on the UE velocity;
o MI/LI-PMI(s) from its N dominant long term interferers – has to be done with the same
periodicity as the identification of the preferred beam from the serving cell.
Feedback from the UE to its serving cell:
o N (configurable by the network) dominant long-term interferer identifiers – low-
frequency update;
o MI/LI-PMI(s) from each of the N dominant long term interferers - same periodicity as
PMI feedback.
Coordination messages from the candidate master to the BSs managing the prospective slave
cells:
o master request message including an indication of the resources on which the UE will
be served, and the PMIs to avoid/favour - each time a cell-edge UE is scheduled.
2.2.3.4 Expected requirements on architecture and protocols
Requirements on architecture include inter-BS communication links with sufficiently low latency in order
to exchange the master request messages. A contention resolution mechanism is needed for the
master/slave role attribution.
2.2.3.5 Conclusion
This section has described a practical way to achieve coordinated beamforming in a decentralized way, in
the sense that no central control entity is needed. The proposed studied approach is based on codebook-
based beamforming. Compared to Rel-8 LTE, additional feedback of PMI(s) for each considered
coordinated interfering cell is required. In addition, fairly low-latency communication links are needed
between BSs in order to convey master request messages. Preliminary results obtained from a simplified
simulator have shown moderate gains compared to non-coordinated beamforming (cell-edge performance
gains of +13% for 1 MI-PMI and +19% for 3 MI-PMI, with no significant cell-throughput gain/loss). Note that these results have been obtained with a very simple scheduler which did not include any
optimization to account for the coordination. Evaluation results of this scheme in a complete LTE-
Advanced class III simulator will be provided in deliverable [WIN+ D42].
2.3 Joint processing with relaxed requirement
One of the major drawbacks related to the implementation of joint processing as the number of users and
BSs increases is the large signaling overhead required for the inter-base information exchange and the
amount of feedback needed from the users. Therefore, one of the main challenges is the design of
efficient algorithms and principles that could reduce the complexity requirements. To achieve this goal,
one of the areas of research is leading to solutions that restrict the joint processing to a limited number of
WINNER+ D1,8
Version: 1.0 Page 26 (75)
BSs or areas in the system. In these approaches, the network is typically divided into clusters of cells, and
the joint processing schemes are implemented within the BSs included in each cluster. The cluster
formation can be static [Ven07] [BH07], if the clusters remain fixed in time or dynamic [PGH08].
2.3.1 Performance of joint processing schemes under varying CSI requirements
In this contribution, different CoMP transmission schemes are characterized and compared within a static
cluster of BSs. Notice that the cluster of BSs under consideration can be formed using the approach of
contribution in section 2.3.2. The performance of the proposed schemes is evaluated over the cluster area,
in order to analyze the impact of a non-uniform distribution of users. This study is carried out as a first
step towards designing adaptive CoMP transmission schemes that could support mobile users. Further
details of this investigation can be found in Appendix A.4
History New
Duplexing mode FDD (TDD)
Clustering mode (1) Static
Clustering mode (2) PJP: user-centric
Codebook-based No
Data exchanges: users data -CJP: Yes
-PJP: Only within transmitting BSs
-DJP: Only during the scheduling phase
Data exchanges: Channels Impulse
Responses
-CJP: Yes
-PJP: Only within transmitting BSs
-DJP: No
Data exchanges : others
Data exchanges rate slow or fast
- CJP and PJP: precoding weights
- DJP: exchange of interference level
experienced by the user
2.3.1.1 Description
We consider a static cluster of BSs. Within the cluster, three different joint processing schemes are
considered, which result in different requirements both in terms of inter-base information exchange and
amount of feedback from the users:
Centralized Joint Processing (CJP): global CSI is available at the transmitter side, and the BSs
within the cluster jointly perform the power allocation and the design of the linear precoders.
Partial Joint Processing (PJP): a particular case of the CJP scheme, it defines different stages of
coordination between BSs. Coordination degrees are obtained arranging an active set or subset
of BSs for each user in the cluster area. Hence, a user only receives its data from the subset of
BSs included in its active set [BPG+08].
From the system point of view, three benefits are provided: feedback reduction (users only feed back channels with an acceptable quality), lower inter-base information exchange (user data is
only needed in the BSs included in its active set) and efficient distribution of power (power is
saved from poor quality channels). However, this joint processing scheme introduces multi-user
interference in the system, since less CSI is available at the central unit to design the linear
precoding matrix. It should be pointed out that a similar approach has also been proposed in
[PBG+08].
Distributed Joint Processing (DJP): BSs are only aware of their local CSI. Therefore, the
precoding and power allocation are locally implemented at each BS (distributed), but the user
may receive its data from several BSs (joint processing) depending on its given channel
conditions. This approach requires a multi-base scheduling technique to assign users to BSs
under a joint processing assumption.
2.3.1.2 Expected performance or benefits
The aim of this contribution is to further characterize by means of simulations some parameters related to the centralized, partial and distributed joint processing schemes, such as the average sum-rate per cell,
the uniformity over the cluster area of the metric under consideration, the robustness of the scheme, the
total transmitted power in the system and the backhaul and signaling requirements. The motivation
behind this analysis is the need to consider the impact of the mobility of the users. This user mobility
implies that the system configuration cannot be static in time, and that the cluster of BSs may need to
WINNER+ D1,8
Version: 1.0 Page 27 (75)
decide which is the best joint transmission scheme depending on the current users requirements (e.g.,
quality of service or service delay constraints) and the system resources (e.g., available transmit power or
backhaul constraints due to the system load).
Simulations are performed over a cluster of 3 BSs, each one equipped with an array of 3 antennas, and
considering 3 single-antenna users (see Figure A-8). Simulation results, described in Appendix A.4, show
that the differences between the schemes arise in interference-limited scenarios. The centralized joint
processing scheme outperforms the remaining schemes at the cost of higher backhaul and signalling
requirements. On the other hand, the partial joint processing scheme shows a trade-off between the
backhaul and signaling requirements and the achieved average sum-rate per cell, that is, its performance
improves as the coordination degree between BSs or the threshold value increases. The backhaul and signalling requirements of this scheme are evaluated by means of the average number of BSs that are
included in the active set of a user for different degrees of coordination between BSs. Finally, the
distributed joint processing scheme improves its performance as the system becomes interference-limited.
Regarding uniformity and robustness aspects, transmission schemes implying a joint design of the linear
precoding matrix (central and partial joint processing schemes) achieve a higher uniformity of the average
sum-rate per cell over the cluster area, especially in the interference-limited scenarios. Moreover, these schemes also show a higher robustness when computing the evaluation metric, that is, they decrease the
standard deviation of the evaluation metric regardless of the position of the user over the cluster area.
2.3.1.3 Expected requirements on signalling and measurements
Centralized Joint Processing (CJP). Assuming that global CSI is available, this approach
requires a central unit to perform the linear precoding design and the power allocation. This central unit can be an additional network element associated to the cluster of BSs, or one of the
BSs of the cluster can act as a central unit. Each user needs to feedback the estimated CSI related
to all the BSs in the cluster to its primary base station, which can be defined as the one that
provides the highest channel gain. Then, the inter-base station exchange allows to gather in the
central unit the global CSI and the user data, in order to perform the joint processing.
Partial Joint Processing (PJP). In this approach, the user only receives its data from the BSs included in its active set. Therefore, the amount of user data that needs to be exchanged between
BSs and/or the central unit is reduced. In order to define the active set of BSs for a given user,
the user estimates the average gain of the received channels, one from each base station, and
defines its reference link or strongest channel, associated to a given base station. Then, the user
compares the channel gains related to the remaining BSs with the reference link, and includes
these BSs in its active set only if their channel gains are above a relative threshold, with respect
to the strongest channel. By doing so, BSs related to poor quality channels do not transmit to the
user and the cluster becomes partially coordinated. The threshold value is specified by the
cluster, and different degrees or stages of coordination can be obtained by modifying its value.
Distributed Joint Processing (DJP). Each base station only needs local CSI in order to design
the lineal precoding matrix and the power allocation. However, in a first step, global CSI is
required to perform the multibase scheduling mechanism. Depending on the system
requirements, this process can be carried out by a central unit (external or related to one base
station), or can be performed using decentralized approaches [PHG08]. Backhaul overhead is
significantly reduced (both the exchange of user data and CSI are reduced).
2.3.1.4 Expected requirements on architecture and protocols
The centralized joint processing scheme requires a central unit to perform the linear precoding design and
the power allocation. The partial joint processing scheme decreases the amount of data that needs to be
exchanged between the BSs and/or the central unit, but still requires the use of a central unit. Finally, the
distributed joint processing approach requires in a first step to perform a centralized multibase scheduling
mechanism and the exchange between BSs of the interference level experienced by the user.
2.3.2 Joint processing with reduced backhaul requirement by MAC coordination
In this paragraph an approach for maximizing the weighted sum-rate is proposed, in a downlink
transmission with multiple cells, i.e. a joint processing solution. A central unit, based on the scheduler
requirements and on the channel estimates, jointly forms the clusters, selects the users and calculates the
beamforming coefficients and the power allocations. Such an approach can be seen as an extension of
WINNER+ D1,8
Version: 1.0 Page 28 (75)
[PGH08], with the difference that users and clustering are jointly selected in order to maximize the
weighted sum-rate.
In this document a particular embodiment of the proposal using a zero-forcing beamformer is presented,
even if it is worth noting that the proposal applies to other beamformer techniques as well.
The proposed technique allows a significant reduction of signaling in the backhaul due to data sharing
between cooperating base stations, while achieving a high fraction of the full coordination performance.
Duplexing mode TDD/FDD
Clustering mode (1) Dynamic
Clustering mode (2) Network centric
Codebook-based Works with or w\o codebook
Data exchanges: users data Yes
Data exchanges: Channels Impulse
Responses
Yes
Data exchanges : others
Data exchanges rate : slow or fast
Scheduling coefficients, precoding
weights
Fast rate
2.3.2.1 Description
Downlink transmission is considered. Let N be the number of single-antenna base stations in the system,
K the number of users in the network, 1, , K the quality of service weights associated to the different
users. The channel from the n-th base station to the k-th user is modelled as
knkn kn knh d
(2.17)
where kn is the fast fading coefficient, is the path-loss constant, knd is the distance in km between n-
th base-station and the k-th user, is the path-loss exponent and kn models the shadowing term.
Moreover, letnkw indicate the precoding weight used for the k-th user at the n-th base-station.
The N base stations are grouped in subsets of N
LB
clusters, where B represents the maximum
dimension of a given cluster. Coordination is allowed between base stations belonging to the same
cluster, whereas base stations belonging to different clusters are not coordinated. The clusters are disjoint,
i.e. a given base station cannot belong to more than one cluster. The base station clusters are created in a dynamic way, in other words at each time interval1 t the sets of coordinated base stations are generated in
order to maximize a given objective function.
We define , 1, ,lC t l L as the set of base station indexes belonging to the lth cluster at the time
interval t and , 1, ,lU t l L as the set of user indexes scheduled for transmission in a given cluster at
the time interval t. We define 1 , , LC t C t as a base station clustering at the tth time interval.
Assuming that the kth user at the tth time interval is scheduled for transmission in the l th cluster, the
signal received by the kth user can be written as
,ll l
l l
k kn nk k kn nj j
j U t j kn C t n C t
kn nj j k
j U tl l n C t
y t h t w t d t h t w t d t
h t w t d t n t
(2.18)
where the first term is the useful signal, the second term is the interference due to the signals sent to the users in the same cluster than user k, the third term is the interference due to the signals coming from
other clusters, and the last term is the additive white Gaussian noise.
1 In this paragraph the notion of time slot is somehow generic, and it does not refer to any LTE-related definition.
The reason for this choice is that we want to keep this framework as much general as possible, leaving eventually
to the system designer the choice of selecting the best time-frame for a practical implementation of this algorithm.
WINNER+ D1,8
Version: 1.0 Page 29 (75)
Let’s for the moment drop the dependence on the time interval t, and focus on the lth base station cluster.
Let be l a mapping between every element in
lC and a corresponding element in the set 1, , lC , and
let l a mapping between every element in
lU and a corresponding element in the set 1, , lU . The
vector of signals received by the lU users scheduled for transmission in the lth cluster can be written as
'
l l l l l y H Wd n (2.19)
where l l
l knk nh
H ,
l ll nkn k
w
W , l
l kkd
d and '
ln is a term representing the white
Gaussian noise and the interference coming from the other clusters.
TheError! Reference source not found. technique proposed in this paper does not rely on a particular
method for the choice of the precoding matrixlW , and can be applied with both full and partial channel
state information. Nevertheless, as an example of a particular embodiment, in the following lW will be
obtained by using a ZF approach, i.e. such that l l
l l U UH W I , where
l lU UI is the
l lU U identity
matrix. Assuming l lU C the ZF matrix can be chosen to the pseudo-inverse of
lH
1
H H
l l l l
W H H H . (2.20)
Then the received signal at the kth user becomes ' ,k k k ly d n k U , and the rate achievable by the k-th
user belonging to the l-th cluster is 2 2
log 1 k
nk
dP
where 2
kd kP E d
is the power allocated to that
user. The problem of finding the power allocation that maximizes the weighted sum rate under ZF
beamforming subject to per-base station power constraints can be formalized as:
1 1 2 2,
2
, , , , , max log 1
0,
,
k
d lkl nk
k
k
l
d
l L L kP k U
k U
d l
nk d n l
k U
PR U U C C
P k U
w P P n C
(2.21)
where nP is the power constraint at the nth base station of the lth cluster. This is a convex problem which
in general can be solved by using an interior point method. In order to lower the computational
complexity, in this paper a simplified solution obtained by normalizing the water filling power allocation
is considered, obtained under a sum-power constraint 1
lC
n
n
P
between the lC base stations belonging to
the lth cluster, in order to satisfy the per-base power constraints.
A star network topology is assumed, that can represent the case of multiple base stations connected with a
central unit or the case of multiple base stations connected to the network, where one of them acts as
central unit. Scheduling, base station clustering, calculation of the beamforming coefficients and power
allocation are realized in the central unit.
The proposed technique can be summarized as follows:
Phase I. Each base station sends the channel estimates to the central unit.
Phase II. Based on the channel state information and on the scheduling requirements, the central unit
jointly creates the clusters of collaborating base stations, schedules the users in these clusters and calculates the beamforming coefficients and the power allocation.
Phase III. The central unit sends to the base stations beamforming coefficients, power allocation, indexes
of the coordinated cells and indexes of the selected users. At this point, the base stations belonging to the
same cluster need to share the data of the selected users between them.
With respect to full network coordination, the proposed technique allows the reduction of signaling due to
data sharing, while requiring the same amount of signaling due to channel estimates exchanges. Anyway,
WINNER+ D1,8
Version: 1.0 Page 30 (75)
under the assumption of low to average Doppler conditions, the backhaul bandwidth required for sharing
the data between cooperating base stations is much higher than the one required for updating the channel
estimates at the central unit.
Assuming ZF precoding, the problem of jointly clustering, user selection and power allocation in a given
time interval0F0F
2 can be formalized as
1 1
1 1, , , , ,
1
max , , , , ,
, , , 1, ,
, , , 1, ,
L L
L
l L LU U C C
l
k j
k j
R U U C C
U U k j k j L
C C k j k j L
(2.22)
The objective function 1 1
1
, , , , ,L
l L L
l
R U U C C
is a function of both the base station clusters and of
the users scheduled in each cluster. Under the ZF precoding assumption the optimum
1 1, , , , ,L LU U C C must at the same time minimize the inter-cluster interference and select a quasi-
orthogonal set of users to be scheduled in each cluster.
The two constraints are related respectively to the assumption of non-overlapping clusters and to the
assumption that each user cannot be served at the same time by base stations belonging to the same cluster.
The optimal solution requires a brute force search over the sets of users and possible base station clusters.
In the following a sub-optimal approach based on the idea of restricting the search space is proposed. This approach consists in two different stages: an off-line stage, where the different clusterings are generated,
and an on-line stage, where the best clustering is chosen as a function of the scheduler requests and on the
channel state information available at the central node.
Off-line phase. The candidate clusterings are chosen off-line taking into account path loss and shadowing
(or more in general average user distribution and average channel estimates). For example for a system
with 7 cells and B=4, the following three candidate base station clusterings could be chosen based on
average measurements: {1234}, {567} - {2467}, {135}- {2356},{147}.
On-line phase. At each time interval t the central unit estimates the weighted sum-rate achievable for
each cluster. This sum-rate estimation involves the choice of a candidate set of users to be scheduled with
a brute force user selection or with a greedy user selection technique and the calculation of the power
allocation that maximizes the weighted sum-rate. Finally, the clustering that maximizes the weighted
sum-rate and the associated set of users, beamforming coefficients and power allocation are used for transmission in the tth time interval.
2.3.2.2 State of the art
Inter-cell interference in theory can be completely eliminated by full network coordination [HFV06].
Unfortunately, full network coordination is not easy to be implemented in real systems. The main problem to face is the amount of backhaul overhead required for signaling and data transmissions.
Different approaches have been considered in order to limit the coordination to only a subset of the cells
in the system. In [MF07a] and [MF07b] an approach has been considered for uplink and downlink
transmissions such that the users are divided in groups using orthogonal resources. Joint detection can be
used only between users belonging to the same group. Weak users (i.e. users at the edge of the cells) are
grouped together and the base-station coordination is realized starting from the weak users until the constraint on the backhaul is achieved. The grouping is realized considering only average channel state
information, without exploiting instantaneous dynamics of the channel. In [Ven07] a BS selection
algorithm is presented that refers to the uplink problem. The goal is to minimize the power in order to
achieve an equal-rate requirement. Power allocation, receive (linear) beamforming and cluster assignment
are jointly realized using an extension of the algorithm proposed in [RTL98]. The main limitation of this
work is the lack of diversity with respect to changing channel conditions. In [PGH08] a dynamic
clustering technique is considered for uplink transmissions in order to maximize the weighted sum-rate.
At each time interval one user per cell is selected using the round robin scheduling. At that time interval
and for those specific selected users the algorithm chooses the best bases in order to serve those users
using joint combining.
2 As already mentioned before, the dependence on the time-interval t is implicit.
WINNER+ D1,8
Version: 1.0 Page 31 (75)
2.3.2.3 Expected performance or benefits
We refer to Appendix A.5 for a detailed description of the simulation results. We anticipate that the
performance will be measured in terms of CDF of the average per-cell rate, and per-user rate. We will
show that the proposal gives an advantage on both average per cell-rate and on cell-edge rate.
2.3.2.4 Expected requirements on signalling and measurements
The proposed technique allows a significant reduction of signaling in the backhaul due to data sharing
between cooperating base stations. As explained before, this reduction is obtained by dynamically
selecting the coordination region.
It can be applied with different type of beamformers, so the level of uplink feedback is different for the
different beamforming techniques.
2.3.2.5 Expected requirements on architecture and protocols
The proposal needs a logical centralized entity to perform the joint clustering, scheduling and
beamforming. On the other hand this centralized entity can be implemented in a distributed way.
2.3.3 A generalized method for joint design of linear transceivers with CoMP
transmission
A general method for joint design of the linear transceivers with coordinated multi-cell processing subject
to per-BS power constraints is proposed for multiple antenna receivers. Two specific system optimization
objectives are considered. In the first, the minimum weighted SINR per data stream is maximized, which results in SINR balancing at the optimal solution. In the second, the weighted sum rate is maximized.
Both optimization problems are non-convex for multiple antenna receivers and only locally optimal, but
still efficient, solutions can be found.
This work is a continuation to the work in [WIN+D14] and [TPK09a], where a generalized method for
joint design of the linear transceivers was proposed for SINR balancing case with single antenna
receivers. This is now extended to consider multiple antenna receivers as well as another optimization
objective, i.e., maximizing the sum rate.
The proposed generalized method for solving different optimization problems with coordinated BS
processing can accommodate the following special cases:
Coherent multi-cell beamforming with per BS power constraints, which requires a full phase synchronism between all cooperating BSs [TCJ08a].
Coordinated single-cell beamforming case, where all the transceivers are jointly optimized while
considering the other-cell transmissions as inter-cell interference (similar to solution in [BO01]).
Any combination of above two, where the number of jointly transmitting BSs may vary between
users and/or streams.
In the coherent multi-cell beamforming case, each data stream may be transmitted from multiple BSs over
a virtual MIMO channel. This requires a full phase synchronism between all BSs. In addition, a large
amount of data needs to be exchanged on the coordination link, i.e., complete channel knowledge of all
jointly processed BS-user links need to be conveyed to the central controller and user data as well as
precoding weights need to be made available at the cooperating BSs.
In the coordinated single-cell beamforming case, each data stream is transmitted from a single BS. Hence,
a full carrier phase synchronism between transmitting BSs is not required. Furthermore, a lower amount
of data needs to be exchanged on the coordination link. In the coordinated single-cell beamforming case,
a user is typically allocated to a cell with the smallest path loss. Near the cell edge, the optimal user/beam
allocation strategy may also depend on the time varying properties of the channel. Hence, large gains
from fast user/beam allocation algorithms are potentially available for the cell edge users. The optimal BS
assignment per user/beam requires an exhaustive search over all possible combinations of user/beam
allocations. This is clearly computationally prohibitive for a large number of users and BSs. Therefore,
sub-optimal heuristic allocation algorithms are proposed. A detailed description of the proposed beam
allocation algorithms are provided in Appendix A.6.
The presented generalized method requires full channel knowledge between all BS-user links. Thus, the
solution represents an absolute upper bound for the less ideal solutions with incomplete channel
knowledge.
WINNER+ D1,8
Version: 1.0 Page 32 (75)
Since the proposed generalized CoMP transmission algorithm is able to perform any scheme between
joint processing and coordinated beamforming it enables an easy design of adaption algorithms that can
switch between joint processing and coordinated beamforming. Using this kind of adaptation algorithms
the high requirements on pure joint processing can be relaxed. Since the generalized method is designed
for a fixed cluster it is independent from clustering approaches. Hence, it can be efficiently utilized in
different clustering methods, e.g., path loss based clustering.
History Continuation of the work in D1.4,
Section 3.3.2.2
Duplexing mode TDD
Clustering mode (1) Static
Clustering mode (2) Network centric
Codebook-based No
Data exchanges: users data Coherent multi-cell beamforming: Yes
Coordinated beamforming: No
Data exchanges: Channels Impulse
Responses
Yes
Data exchanges : others
Data exchanges rate : slow or fast
Precoding weights
Fast rate
2.3.3.1 Description
Consider a cellular system that consists of BN BSs and K users. Each BS b and each user k are
equipped with TN and k
N R antennas, respectively. A set U with size UK includes all users active
at the given time instant, while a subset UU b includes the users allocated to BS bkb U, . Each user
k can be served by kM BS’s which define the joint processing set kB for the user k , and
B1,...,BB Nk . The signal vector kN
kRCy of user k consists of the desired signal, intra-cell and
inter-cell interference. It is expressed as follows
k
b i
ibkbkb
b ki
ibkbkb
b
kbkbkbk
k bkk
aaa nxHxHxHy B\B U
,,,
B
,,,
B
,,,
where the vector TC,N
kb x is the transmitted signal from the b ’th BS to user k , kn ~ ),0(R0
kNNCN I
represents the additive noise sample vector with noise power density 0N , and TRC,,
NN
kbkbka
H
TRCNN
k
is the channel matrix from BS b to user k with large-scale fading coefficient kba , . The
elements of kb,H are normalized to have unitary variance.
The transmitted vector for user k is generated at BS b as kkbkb dMx ,, , where kmNkb
TC,M is the
pre-coding matrix, T,,1 ,..., kmkk kddd is the vector of normalized complex data symbols, and
k
NMNm kk RT ,min denotes the number of active data streams.
The focus is on linear transmission schemes, where the BN transmitters send S independent streams,
URTB ,min
k kNNNS per transmit dimension. Per data stream processing is considered, where for
each data stream Sss ,...,1, the scheduler unit associates an intended user sk , with the channel matrices
s
NN
kb bsk
sB,C
RT
,
H . Note that more than one stream can be assigned to one user, i.e. the cardinality
of the set of scheduled users, Ssks ,...,1|U is less than or equal to S .
Let TC,N
sb m and skN
sR
Cw be arbitrary transmit and receive beamformers for the stream s. The
SINR of the data stream s can be expressed as
WINNER+ D1,8
Version: 1.0 Page 33 (75)
S
sii b
ibkbskbs
b
sbkbsskb
s
i
ss
s
s
aN
a
,1
2
B
,,H
,
2
20
2
B
,,H
,
mHww
mHw
(2.23)
The general system optimization objective is to maximize a function Sf ,...,1 that depends on the
individual SINR values for each stream s . This can be formulated as
Ss
NbP
s
aN
a
f
s
b
s
sb
sS
si b
ibkbskbs
b
sbkbskb
S
b
i
ss
s
ss
,...,1,1
,...,1
s.t.
,...,max
2
B,
S
2
2,
,2
1,1
,,H
,
2
20
2
,,H
,
1
w
m
mHww
mHw
(2.24)
where the variables are TC,N
sb m and skN
sR
Cw , Ss ,...,1 and bS includes all streams allocated to
BS b , i.e., bsb ks U|S . sb,m can be further split into sbsbsb p ,,, vm , where TC,N
sb v and
sbp , are the normalized transmit beamformer and the allocated power for the stream s . The weights
Sss ,...,1,0 , are used to prioritise the data streams of different users differently.
Two different system optimization objectives are considered.
Minimum weighted SINR maximization, ssSsSf 1,...,11 min,...,max
[TPK09a].
Weighted sum rate maximization,
S
s
ss
S
ssS
sf
1
21
1 1log1log,...,max
[TCJ08a].
Solutions for these optimization problems are presented in detail in Appendix A.6.
Notice that another approach for the problem of maximizing the minimum SINR per user/stream, based
on the uplink-downlink duality theorem, was considered for coordinated beamforming case in section
2.2.1. In contrast to that, the generalized approach in this section, which can accommodate any scenario
between joint processing and coordinated beamforming, utilize the conic optimization [WES06] on the
beamformer design. The non-convex problem is divided into convex sub-problems that can be optimally
solved by using standard convex optimization tools. However, the global optimality cannot be guaranteed due to the non-convexity of the original problem.
2.3.3.2 State of the art
A generalised method for joint design of linear transceivers for SINR balancing case subject to per-BS
power constraints with single antenna receivers was presented and analysed in [WIN+D14] and
[TPK09a]. The method can accommodate any scheme between coherent multi-cell beamforming and
coordinated single-cell beamforming. It is assumed that full CSI between all terminals and BSs is
communicated from the BSs to the central controller or master BS so that the precoders are computed in a centralized manner. Furthermore, the precoders need then to be communicated back to each BS involved.
2.3.3.3 Expected performance or benefits
Performance of the proposed concept is evaluated in detail in Appendix A.6. Some of the results from
Appendix A.6 are presented herein. Table 2-4 presents the ergodic sum rate of users at the cell edge for
1,2,2,2,,, RTB k
NNNK and 2,2,2,2 systems with sum rate maximization algorithm at 20 dB
single link signal to noise ratio (SNR). The following transmission schemes are compared:
1. Coherent multi-cell beamforming. 2. Coordinated beamforming with optimal beam allocation (exhaustive search)
WINNER+ D1,8
Version: 1.0 Page 34 (75)
3. Non-coordinated beamforming with optimal beam allocation (exhaustive search). The inter-cell
interference is neglected in the precoder design.
Table 2-4: Main performance results of generalized CoMP transmission method
Transmission scheme Ergodic sum rate
[bits/s/Hz]: 1R k
N
Ergodic sum rate
[bits/s/Hz]: 2R k
N
Coherent multi-cell beamforming
15.95 23.38
Coordinated beamforming
(optimal beam allocation)
10.36 17.15
Non-coordinated beamforming
(optimal beam allocation)
5.47 8.17
It can be seen that coherent multi-cell beamforming greatly outperforms all the non-coherent cases. The
difference between coherent beamforming and coordinated beamforming in terms of ergodic sum rate is
about 6 bits/s/Hz at the cell edge for both configurations of the receiver antennas, 1R k
N and 2R k
N .
This is due to the fact that the coherent multi-cell beamforming can fully eliminate the inter-cell
interference. Obviously, there is a trade-off between performance and complexity since using coordinated
beamforming the amount of data to be exchanged on the coordination link is reduced. Furthermore, a full
carrier phase synchronism between BSs is not required. It is also shown in Table 2-4 that the performance
of the coordinated beamforming is significantly better, i.e., almost 5 and 9 bits/s/Hz, than that of non-
coordinated beamforming case for 1R k
N and 2R k
N , respectively. A more detailed performance
evaluation with wide range of numerical examples is provided in Appendix A.6.
2.3.3.4 Expected requirements on signalling and measurements
The coherent multi-cell beamforming sets high requirements for signalling and measurements. A large
amount of data needs to be exchanged on the coordination links, i.e., complete CSI of all jointly
processed BS-user links to the central entity, and user data as well as precoding weights from the central controller to the cooperating BSs. Furthermore, tight phase synchronization across the cooperating BSs
and centralized controllers is required. Therefore, high speed links, i.e. optical fibres or dedicated radio
links, are required between cooperating BSs and the central controller.
Coordinated beamforming and beam allocation has looser requirements on coordination and backhaul,
i.e., tight frequency synchronization and sharing of user data between each BS are not required. However,
full channel knowledge of all jointly processed links is still needed for the ideal interference avoidance.
Hence, a centralised resource management mechanism, i.e., central controller or master BS, is still
needed.
2.3.3.5 Expected requirements on architecture and protocols
The proposed method requires a centralized resource management mechanism. Furthermore, the required
information exchange rate between the cooperating BSs and the central controller is high.
WINNER+ D1,8
Version: 1.0 Page 35 (75)
3. Investigations on relaying in the framework of CoMP
The introduction of Relay Nodes which are controlled by the network allows to use them as part of a
CoMP system. The RNs can be used to extend the actual coverage or to densify the actual network to
enhance the user throughput at the cell edge. Relay nodes are connected to a base station via over-the-air
in-band links (e.g. specific control channels or in-band backhaul, depending on the relay type), enabling a
tight coordination but at the price of a possible delay between the coordinated nodes. Hence, coordinated relaying schemes are a particular way to implement CoMP, which has to account for the specificities of
relays regarding the coordinated processing: the delay inherent to the 2-hops transmission between a base
station and a user terminal (UT), and/or the potential errors affecting the first hop.
In year 1, different aspects of coordinated relaying schemes were investigated, e.g. coding schemes and
schedulers taking relaying into account. The conclusion was that it is worthwhile to further investigate
relaying in the framework of CoMP. Hence, in this section various cooperative relaying schemes considered for year 2 are proposed. The first innovation considers a cooperative scheme, which combines
cooperative base station transmission and relaying. The second innovation considers distributed LDPC
coding for a Decode and Forward type relay..
History Continuation of the work in D1.4,
Section 3.3.2.2
Duplexing mode TDD
UL/DL UL or DL
3.1 Impact of interference on design and performance of relaying protocols
Although relaying is a promising concept it still needs to be clarified to which extent the increased
interference level reduces the actual benefits of relaying. In this section, we extend the interference channel by additional relay nodes in order to investigate the influence of interference on the design and
performance of relaying protocols.
Figure 3-1: Relay-assisted interference channel with two communication pairs and one relay node
supporting each pair.
This section presents system level results for a protocol, which combines the advantages of coordinated
base station transmission and relaying. In order to discuss this protocol, we introduce the relay-
interference channel as introduced in [RFL09] with two communication pairs (with source nodes s1 and
s2, and destination nodes d1 and d2), each supported by one relay node (r1 (resp. r2) supports the
communication between s1 (resp. s2) and d1 (resp. d2)). Each relay node is underlying the half-duplex
constraint, which implies that it is either transmitting or receiving on a particular time-frequency resource. It further considers backhaul links C{1,2} and C{2,1} between source nodes, which are used to coordinate
their transmissions. Figure 3-1 illustrates the channel. In order to explain the employed schemes, we
introduce the interference and broadcast channel in the following.
WINNER+ D1,8
Version: 1.0 Page 36 (75)
3.1.1 The Interference Channel
Figure 3-2 The standard Gaussian interference channel (f1 and f2 denote the individual encoders
and g1 and g2 denote the individual decoders). Message W1 and W2 are independently encoded and
each terminal transmits the message X1 and X2, respectively, At both destination the messages W1
and W2 are decoded using the channel outputs Y1 and Y2, respectively.
The general interference channel consists of N source nodes and M destination nodes. However, for the
sake of readability we focus in the following on the case N = M = 2, where two transmitters and two
receivers are considered. Consider Figure 3-2, which shows the standard form of the Gaussian
interference channel (IC) (note that any interference channel with N = M = 2 can be modeled by the
standard form).
So far the best known inner bound of the (general) interference channel has been derived by Han and Kobayashi in [HK81], where the first transmitter divides its message in two parts W(1,1) and W(1,D)
(similarly the second transmitter uses W(2,2) and W(2,D)). While the first part W(1,1) is only decoded by the
first receiver, W(1,D) is decoded by both receivers and called common message. The purpose of a common
message is to reduce the interference for the private message. Another basic element is the joint decoding
of both common messages and the own private message. Even though the rate region of Han-Kobayashi
(HK) is the best inner bound known so far, it is computationally complex to determine. The complete HK
region is the convex hull over all possible power assignments, which is difficult to determine in real-time
for each fading situation. Since complexity is a driving factor we are interested in less complex methods,
which still provide sufficient performance figures.
One such method has been presented by Etkin et al. in [ETW06], [ETW08], which has been proven to be
within 1 bit per channel use (bpcu) of channel capacity while being much simpler than the full HK coding approach. Assume the channel at both terminals is given by
Their basic idea is to introduce a decoding order and align the interference caused by each transmitter
with the noise power at the non-intended receiver, i. e., both terminals choose the private message power
such that
holds. Compared to the full HK approach, each terminal at first decodes both common messages jointly
and then each one decodes its own private message.
3.1.2 The Broadcast Channel
We further detail the discussion with the Gaussian broadcast channel (BC). The BC models the case
where one source communicates with multiple destinations. Each decoder at a particular destination
works independently of all other decoders such that no joint processing is possible. Costa introduced in
[Cos85] the idea of Dirty Paper Coding (DPC), which is the capacity achieving approach and is able to
provide the same capacity in case of interference known at the transmitter as if there were no interference.
WINNER+ D1,8
Version: 1.0 Page 37 (75)
Figure 3-3 The multiple antenna Gaussian broadcast channel with only private messages and
perfect channel knowledge at the encoder
Figure 3-3 shows the two-antenna Gaussian BC. First, the two private messages W1 and W2 are jointly
encoded giving their corresponding transmit vectors X1 and X2, respectively, as well as the channel input
vector X = X1+X2. Obviously both antenna transmissions are interfering with each other. But knowing the channel matrix at the encoder, we can subsequently apply DPC on each user data stream in order to
alleviate the interference of all previously encoded user streams. However, this approach requires
significant computational efforts and channel knowledge at the transmitter. The Gaussian IC, introduced
in the previous section, considers two transmitters, which do not have common CSI and data. Consider
now the Gaussian BC with per-antenna power constraint, which models a Gaussian IC where transmitters
are connected and able to jointly encode their channel input. Compared to the Gaussian BC, where a
duality with the Gaussian multiple access channel (MAC) can be exploited, the Gaussian BC with per-
antenna power constraint is more difficult to handle. In this section, we use the suboptimal approach
Zero-Forcing DPC (ZF-DPC) [CS03, KF+06] which uses the LQ-decomposition of the channel matrix
and has deterministic, polynomial complexity. More specifically, the channel matrix is expressed as
H=LQ, where L is lower triangular matrix and Q is a unitary matrix. By precoding with the Hermitian of Q, the channel is transformed is lower triangular form. Similar to SIC using VBLAST after QR post-
processing, DPC is used to pre-cancel the interference at the transmitter. This results in a set of
interference-free single-user links. Furthermore, the ZF-DPC approach satisfies per-antenna power
constraints.
3.1.3 The Relay-Assisted Interference Channel
One way to examine the relay-interference channel as illustrated in Figure 3-1 is to treat it as a cascade of
two BCs or ICs subsystems, which interfere with each other, i.e., we still consider the signal from the
source nodes received by both destinations. The first channel is from sources to relays (S → R) and might
exploit backhaul links between both sources in order to coordinate their transmissions. Depending on the
respective coding schemes in this first channel, the used encoding scheme for the second channel from
relays to destinations (R → D) must be selected. In case one of both channels is considered to be a BC,
the transmitting nodes exploit common information to coordinate their transmissions. When treating one of both stages as interference channel, the transmitting nodes do not exploit common information but
might employ interference mitigation techniques such as Han-Kobayashi coding [HK81] or time-division
duplex. We showed in [RFL09] that the best choice is to use DPC on the link between base stations and
relay nodes, and to use HK coding on the link between relay nodes and user terminals. Furthermore, it
turned out to be the best choice to use an adaptive approach, where users close to the base station are
served directly by base stations, and all other users are served using relay nodes. In the following, we
apply this idea to the WINNER system model defined by deliverables [WIN2D6137] and [WIN2D6112]
(and channel models defined in [WIN2D111] and [WIN2D112]) and evaluate the protocol’s performance
in the wide-area and Manhattan-area scenario.
3.1.4 Considered Protocols and Constraints
This section’s analysis considers four protocols, which are detailed in the following.
Conventional, direct transmission
The reference protocol of our analysis is a conventional direct transmission protocol, which divides the
overall usable bandwidth into an edge band and center band. Edge band resources are used to serve users
at the cell edge with high inter-cell interference. Furthermore, this protocol reserves a center band, which
is used by all cells to serve users with low inter-cell interference. The assignment of users to edge and
WINNER+ D1,8
Version: 1.0 Page 38 (75)
center band uses the expected long-term SINR, which is assumed to be perfectly known. Furthermore,
each BS is equipped with four antennas and therefore able to serve up to four users on the same time-
frequency-resource using ZF-DPC.
Multi-Cell MIMO Transmission
In order to improve the spectrum reuse, we additionally consider a multi-cell MIMO transmission
protocol where, in our case, three base stations are cooperatively transmitting (in this case the setup in
Figure 3-1 has to be appropriately modified). In order to coordinate their transmission all base stations are
connected by an unlimited backhaul, which provides the means to perfectly exchange CSI and the
individual messages at the base stations. In this particular scenario, only multi-cell MIMO is used and no
relays are used. This protocol is one way to implement the CoMP cooperation described in further detail
in the previous section.
As discussed in the previous part, we use ZF-DPC in order to guarantee a deterministic complexity. The
assignment of user terminals to the individual base stations is done based on the path loss between base
station and user terminal. Throughout the following analysis we assume that each base station has perfect
CSI of the compound channel matrix. Since each BS is equipped with four antennas, three BS can serve
up to 12 UEs on the same resource using ZF-DPC.
Relay-Only
We consider in our analysis a protocol where all users are served using relay nodes and Etkin-Tse-Wang
(ETW) coding [ETW06], [ETW08] based on path loss information. In order to ensure efficient
interference mitigation each user is assigned to the relay node with the smallest path loss towards this
user. Then the relay node with the second-smallest path loss towards this user is chosen and selected for
the ETW coding. Hence, we always group two users and two relay nodes such that they cooperate using
ETW coding. The expected bottleneck in such a system is the in-band feederlink between base station and
relay node. In order to improve the performance of the in-band feederlink, we consider that at most three
base stations form one cluster and cooperatively transmit to their assigned relay nodes. Compared to the
previously described multi-cell MIMO protocol the signaling overhead is reduced (only the power levels of the individual messages must be exchanged), because relay nodes do not move and have line-of-sight
(LOS) towards the base station, which increases the coherence time/frequency. Furthermore, the user
grouping is significantly simplified as only a small number of fixed relay nodes is considered.
Mixed Protocol using Relays and Multi-Cell MIMO Transmission
In this setup users are served either using the multi-cell MIMO protocol or using relay nodes and ETW coding. Based on the path loss to the individual nodes (a penalty for the lower transmission power of
relay nodes is included), each user is assigned to either a base station or a relay node. In this mixed
protocol either base stations or relay nodes transmit within a certain part of the spectrum, but not both.
This part does not consider the case where two relays have common message information and exchanged
CSI, i.e., we do not use the backhaul link between both relay nodes. In case we treat S → R as IC and
employ HK coding, the commonly decoded information can be used for a distributed DPC and in case we
treat S → R as BC, the S → R stage degrades to a multicast channel because both relays must receive the
same amount of information (CSI and data messages). Nevertheless, the necessity to exchange the
transmitted messages and CSI as well as to synchronize their transmissions will require a high-capacity
feeder link between sources and relay nodes, which is usually the bottleneck in a wireless communication
system with relays.
3.1.5 Numerical Results
This section compares the performance of the previously introduced protocols using
the 5%-ile throughput θ5% defined by the throughput achieved by at least 95% of all users, and
the median throughput θ defined by the throughput achieved by at least 50% of all users
In this section, we analyze the urban wide-area and Manhattan-area scenario. In the former we consider
one central site, which is surrounded by two tiers of overall 18 interfering sites. In the case of Manhattan-
area, we use overall 44 cells but only considering the results for the three inner cells. Furthermore, the
number of users per cell may vary as we keep the number of overall users constant and ensure that in the
case of wide-area on average 90 Users per km2 are randomly placed according to a uniform distribution
and 250 Users per km2 in the case of Manhattan-area. (in both scenarios this equals to about 25 users per cell on average).
WINNER+ D1,8
Version: 1.0 Page 39 (75)
The overall area is partitioned into rectangles of equal size. In case of the wide-area scenario they are of
size 30m × 30m, and in case of the Manhattan-area scenario they are of size 10m×10m. For each user the
corresponding spatial block (x, y) is determined and the average throughput θ(x, y) = Et {θ(x, y, t)}
over all users, which have been placed in this block, is used to determine the number of resources. Using
a fair scheduler the number of resources assigned to a user is proportional to 1/θ (x, y). The fact that we
use the average throughput reflects that we try to achieve a good tradeoff of high system throughput and
fairness. In order to achieve an improved fairness within the system, it would be more suitable to use the
5% quantile. In this way, a user with a high throughput variance gets even more resources in order to
improve the 5% quantile of throughput at the expense of other users’ performance.
Our analysis does not consider any user mobility. Furthermore, the Manhattan-area scenario considers
users placed both on streets and within buildings. Such a uniform distribution might not be realistic, but
allows for insights in the performance improvements through relaying.
Figure 3-4 shows the CDF of the user throughput, i.e., the probability that an arbitrary user achieves a
certain throughput. The conventional protocol with direct transmission achieves only a 5%-ile throughput
of about 0.1Mbit/s and a median throughput of about 1Mbit/s. By contrast, multi-cell MIMO transmission is able to significantly improve the performance: θ5% = 0.6Mbit/s and θ = 3.5Mbit/s, which not only
implies an improved system throughput but also that users with low data rates benefit from these
improvements. The deployment of two additional relay nodes and application of the relaying-only
protocol improves the median throughput θ = 3Mbit/s) but reduces the performance of the worst users θ5%
= 45 kBit/s), which results in a step of the CDF. Relays only improve the channel conditions for those
users at the cell edge and users close to the relay node. If only relay nodes are used and no direct BS-UT
transmission, the worst users are placed closely to the base station. Contrary to the relaying-only protocol,
the mixed protocol is able to improve both median throughput θ = 4.8Mbit/s) and worst-user performance
θ5% = 0.8Mbit/s). This time users close to a base station are served using multi-cell MIMO transmission
and users at the cell edge are assigned to relays. Surprisingly, if we deploy eight relay nodes and use the
relaying-only protocol, we do not perform much better than with two relay nodes and using the mixed approach where θ = 8Mbit/s and θ5% = 1.1Mbit/s). There are two reasons for this: in-band feederlink and
the requirement that you can only transmit as much as in the relay buffer (a very high-data rate feederlink
might be able to support higher data rates than the buffer).
.
Figure 3-4 Throughput for the wide-area scenario using in-band relays
WINNER+ D1,8
Version: 1.0 Page 40 (75)
Figure 3-5 Throughput for the Manhattan-scenario and in-band feederlinks
Figure 3-5 shows the results for the Manhattan-area scenario. We can again see that multi-cell MIMO
improves the performance compared to a conventional system. The 5%-tile throughput of multi-cell
MIMO is θ5% = 5 kBit/s and the median throughput is θ = 300 kBit/s. Compared to the previous scenario
the difference between worst and median throughput performance is much higher, which implies that a
major part of the users has very bad channel conditions towards the base stations, primarily those located
indoor. The same effect can be seen if only two relay nodes are deployed outdoor (θ5% = 3 kBit/s and θ =
500 kBit/s for the mixed protocol). Only four relay nodes (two outdoor and two indoor relays) are able to
improve performance figures for both worst user and median user performance (θ5% = 0.3Mbit/s and θ =
1.8Mbit/s). We can further see that in case no direct transmission is used the performance is the same as
for the mixed protocol, which implies that only a few users benefit from multi-cell MIMO transmission.
For practical deployments this implies that it suffices to only use relay nodes in micro-cellular
deployments, which reduces complexity and power consumption.
Figure 3-6 Throughput over area. Arrows indicate the main lobe direction and triangles indicate
the relay nodes.
Among others, relaying helps to provide a more uniform power distribution, which is reflected by a more
uniform throughput distribution as shown in Figure 3-6. Of course, the throughput around the base station
WINNER+ D1,8
Version: 1.0 Page 41 (75)
and relay nodes is higher. While the throughput around the base station quickly declines due to the
decreasing LOS probability, the slope around each relay node is much lower. Furthermore, the peak
throughput around the base station is higher than around the relay nodes due to the higher transmission
power. The lowest throughput is achieved at the cell edge and between relay nodes in one cell. In both
areas the reasons are the high path loss and the high inter/intra-cell interference.
The mixed protocol supports multi-cell MIMO transmission as well as relaying, and the node assignment
is done based on the path loss. From Figure 3-6 we can see that the probability, that a user terminal is
assigned to a relay node, increases towards the relay nodes, and the area where users are assigned to a
relay node (with high probability) is much larger than the area where user terminals are assigned to a base
station. Cell edge users are primarily served by relay nodes, which implies that the improved path loss conditions outweigh the lower transmission power of relay nodes.
Figure 3-7 Throughput-CDF and comparison of correct (fast-fading based) ETW coding (dashed
lines) and path loss-based ETW
Figure 3-7 compares the performance of
ETW coding based on perfect fast-fading-CSI at the relay node,
a binary ETW coding with possible two power assignments (either only private or common
messages), and
TDMA between relay nodes and user terminals. We can conclude from this figure that there are only minor performance improvements if fast-fading CSI
is used. The usage of path loss information significantly reduces the necessary signaling overhead and
makes ETW coding feasible. However, the results also reveal that a binary ETW coding, where either
only private or common messages are used, is able to achieve almost the same performance. This
suggests that it is sufficient to quantize the power assignment level with 1 bit. TDMA between relay
nodes and user terminals does not achieve the performance of ETW coding. Compared to TDMA, ETW
coding needs no central scheduler, which coordinates the assignment of resources to each relay node
based on the global path loss information. The power assignment for ETW coding can be done at each
relay node based on the path loss between the relay and the assigned user terminals.
3.2 Distributed LDPC coding for the single relay channel
After the authors in [SEA03a], [SEA03b] introduced the concept of cooperative diversity, many papers
proposed cooperation protocols for relay channels, that can be classified into two major categories,
namely amplify-and-forward (AF) and decode-and-forward (DF).
In the AF protocol the relays scale the signal received from the source and forward it to the destination
without other treatment. Thus, such a cooperation scheme employs a unique error correcting code, coding being only performed by the source, and decoding only by the destination.
In DF protocols the relay decodes the signal and, assuming successful decoding, it recovers the codeword
(information and parity bits) sent by the source. Two main scenarios are then possible:
WINNER+ D1,8
Version: 1.0 Page 42 (75)
[DF-1] The recovered codeword, or part of it (for instance, only the information bits) is forwarded to the
destination. Thus, the forwarded information is (part of) the information originally transmitted by the source, and the destination will benefit from the channel diversity, which improves the
signal to noise ratio on (part of) the received transmission.
[DF-2] The relay computes a new sequence of parity bits (different from the sequence of parity bits in
the recovered codeword), which is forwarded to the destination. In this case, the forwarded
information is not contained in the one originally transmitted by the source. However, the destination will benefit from the coding diversity, consisting on the knowledge of the extra
information received from the relay. We refer to this scenario as distributed coding or coded
cooperation.
The above scenarios are illustrated here below (I = information bits, P = parity bits):
Figure 3-8. Examples of cooperation scenarios under the DF protocol
The most known example of distributed coding is the one of a distributed Turbo-code (TC): the source
and the relay encode the signal using a convolutional code (the relay also has to interleave the received
signal), allowing the destination to decode the received signal using a parallel-concatenated TC [ZV03].
History New
Duplexing mode FDD or TDD
UL/DL UL/DL
3.2.1 State of the Art
LDPC codes for the single relay channel have been mostly investigated under the first DF scenario
described above. The second DF scenario (distributed coding) is, from the coding theory point of view,
closely related to incremental redundancy. Some approaches have been already proposed in the literature,
which are mainly based either on serial or parallel code concatenation ([KAA04], [CBS+07], [HD07],
[RY07]).
Serial concatenation
Consider two LDPC codes C1 and C2 with parameters (K1, N1) and (K2, N2), where Ki denotes the code
dimension and Ni denotes the code length. Assume that N1 = K2, and that the two codes are systematic
(the last condition is only assumed for simplicity). Then, these codes can be used within a distributed
coding scenario as follows:
The source encodes the sequence of information bits I, of length K1, using the code C1. Since C1
is systematic, the encoding operation generates a sequence of parity bits P1, of length N1-K1, and
the codeword (I, P1) is broadcasted to the relay and the destination.
The relay decodes the received signal and recovers the codeword (I, P1), which is then encoded
using the code C2. Since C2 is systematic, the encoding operation generates a sequence of parity
bits P2, of length N2-K2, which is forwarded to the destination.
R
S D (I, P)
(I, P) I
R
S D (I, P1)
(I, P1) P2
(a) scenario DF-1 (b) scenario DF-2
WINNER+ D1,8
Version: 1.0 Page 43 (75)
Thus, the destination receives 21
~,
~,
~PPI , representing noisy versions of the transmitted sequences (I, P1,
P2). If H1 and H2 denote the parity check matrices of the codes C1 and C2, the noisy received signal
21
~,
~,
~PPI can be decoded at the destination, by using the matrix H obtained by the concatenation of H1
and H2, as shown at Figure 3-8 (a).
Parallel concatenation
Consider two LDPC codes C1 and C2 with parameters (K1, N1) and (K2, N2), where Ki denotes the code
dimension and Ni denotes the code length. Assume that K1 = K2 = K and that the two codes are systematic
(the last condition is only assumed for simplicity). Then, these codes can be used within a distributed
coding scenario as follows:
The source encodes the sequence of information bits I, of length K, using the code C1. Since C1
is systematic, the encoding operation generates a sequence of parity bits P1, of length N1-K, and
the codeword (I, P1) is broadcasted to the relay and the destination.
The relay decodes the received signal and recovers the codeword (I, P1). It re-encodes the
information sequence I using the code C2. Since C2 is systematic, the encoding operation
generates a sequence of parity bits P2, of length N2-K, which is forwarded to the destination.
Thus, the destination receives 21
~,
~,
~PPI , representing noisy versions of the transmitted sequences (I, P1,
P2). Let H1 and H2 be the parity check matrices of the codes C1 and C2. Then the received signal
21
~,
~,
~PPI can be decoded at the destination, by using the matrix H obtained by the concatenation of H1
and H2, as shown at Figure 3-8-(b). Alternatively, the destination can attempt to decode H1 and H2
separately, exchanging periodically extrinsic information concerning information bits between the two
decoders.
Figure 3-9. Code concatenation
Obviously, many other scenarios may be considered for the above constructions: for instance, the relay can forward (I, P2) instead of P2 only, which improves the signal to noise ratio on information bits, and
also provides the extra information P2 to the destination.
It should also be noted that we consider the problem of designing LDPC codes that can be advantageously
used for coded cooperation, but our goal is not to discuss cooperation protocols (Orthogonal or Non-
Orthogonal DF, etc.) or multiplexing schemes of bit streams from source and relay (time or frequency
division multiplexing ).
From the code design point of view, the serial or parallel concatenation of two LDPC codes presents several weaknesses. It is well known that the correction capacity of LDPC codes is strongly related to the
code irregularity, which describes the distribution of the non-zero entries in the parity-check matrix.
Moreover, the density of the non-zero entries in the parity check matrix plays also a very important role
by itself, and codes with lower rates must be sparser than those with higher rates. However, this basic property is violated by the code concatenation construction, since matrix H (with lower rate) contains
both matrices H1 and H2 (with higher rates). Another drawback of the concatenation construction is that
the matrix H inherits all the cycles of H1 and H2, which can be normally avoided for codes with lower
WINNER+ D1,8
Version: 1.0 Page 44 (75)
rates. As a consequence, the performance of concatenated codes is degraded with respect to “properly
designed” (unconcatenated) LDPC codes with the same rate.
Stated in a slightly different manner, code concatenation does not allow the joint design of matrices H1
and H, where H is the (serial or parallel) concatenation of H1 by some other matrix H2, such that the
corresponding codes perform close to their channel capacities. On the other hand, the success of the DF cooperation scheme depends on the performance of both H1 and H, since that successful decoding is
required both at the relay and the destination.
3.2.2 Cooperative coding using extended LDPC codes
The purpose is to address the problem of designing capacity approaching LDPC codes for coded
cooperation. We say that the code C, of length N, is an extension of the code C1, of length N1 N, if any
codeword 1
,,1 Nxx of C1 can be extended to a codeword NN xxx ,,,,11 of C. Obviously,
(serial or parallel) code concatenation is a particular case of the above definition. Extended codes can be
used within a distributed coding scenario, for instance as follows:
The source encodes the sequence of information bits using the code C1. The encoded codeword
1
,,11 NxxX is then broadcasted to the relay and the destination.
The relay decodes the received signal and recovers the codeword 1
,,11 NxxX , which is
then extended to a codeword NN xxxX ,,,,11 of C. Then, the extra information
contained in X , that is NN xxXX ,,\ 11 1 , is forwarded to the destination.
A construction method of extended codes has been elaborated, which avoids code concatenation, and allows more degrees of freedom to optimize the joint design of H1 and H. The proposed method also
enables a higher decorrelation between the irregularities (and densities) of matrices H1 and H, making
possible the design of capacity approaching codes by density evolution methods.
3.2.3 Expected benefits
To illustrate the expected benefits, Figure 3-10 presents the asymptotic performance derived by density
evolution methods using Gaussian approximation. The additive white Gaussian noise (AWGN) relay
channel is considered, with fixed source-to-destination SNR (either -1dB – dotted curves, or -0.5dB –
solid curves); the relay successfully decodes the received signal. The SNR of the relay-to-destination link
is shown on the horizontal x-axis. Black, red and green curves correspond to the distributed coding scenario (scenario DF-2, Figure 3-2-(b)) as follows: black curves matrices (H1, H) are obtained by the
proposed design, red curves matrices (H1, H) are obtained by parallel concatenation, and green curves
matrices (H1, H) are obtained by serial concatenation. Finally, the blue curves represent the performance
of the scenario DF-1 (Figure 3-2-(a)). It should be noted that the performance shown does not correspond
to optimized codes; regular codes in all the above scenarios are used.
WINNER+ D1,8
Version: 1.0 Page 45 (75)
Figure 3-10 Comparison of the asymptotic performance of several DF schemes
WINNER+ D1,8
Version: 1.0 Page 46 (75)
4. Conclusion
Considering that future cellular wireless systems will need to provide high data rate services for a large
number of users without the use of an increased signal bandwidth, a possible strategy to reduce the
performance limiting inter-cell interference is identified in Coordinated MultiPoint systems (CoMP).
CoMP refers to a system where the transmission and/or reception at multiple, geographically separated antenna sites is dynamically coordinated in order to improve system performance. CoMP is already being evaluated as one of the most promising techniques for increasing capacity of new radio access systems in
several standard organisations. In the framework of 3GPP, CoMP is identified as a key concept in the
LTE-Advanced study item [3GPP36.814]. CoMP approaches are also considered in the WiMAX
standardisation framework, see e.g. [MHM+08] for a classification of the related contributions.
In year 1, WINNER+ studied different aspects of CoMP extensively. These included system
architectures, different approaches and algorithms for performing the coordination, and the requirements
in terms of measurements, signalling, backhauling constraints, etc. these put on the system. The work is
reported in [WIN+D14]. Much attention was put on so-called joint processing/transmission schemes, in
which the transmission to a particular UE originates from multiple transmission points. It was found that
these schemes have potential to provide significant performance gains, however at the price of high
requirements on the backhaul links in terms of latency and capacity since user data, CSI, and precoding
weights need to be shared among the transmission points. Furthermore, different aspects of coordinated
relaying schemes were investigated, e.g. coding schemes and schedulers taking relaying into account. The
conclusion was that it is worthwhile investigating relaying in the framework of CoMP further.
Considering these conclusions from year 1, the focus in year 2 was set on CoMP solutions that have
relaxed requirements in terms of backhauling and complexity compared to the full blown joint processing
schemes that were studied during year 1. In line with 3GPP, these schemes can be split into two
categories “Coordinated Beamforming”, and “Joint Processing”. Both categories have been considered in
this second year. For coordinated beamforming schemes, the transmission to a particular UE takes place
from a single transmission point, but scheduling and beamforming weights are coordinated in order to
reduce interference. In the joint processing case, which is more demanding in terms of backhauling since
users data are present at each cooperating BS, focus was put on relaxed requirements schemes. In both cases, clustering (i.e. determining which BSs will cooperate) showed to be key to algorithmic
performance.
Simulations scenarios were selected in order to allow comparisons between various proposed schemes. Thus respective benefits of Joint Processing and Coordinated Beamforming with reference to a baseline
non CoMP scenario can be evaluated. Attention was given not only to average throughput gains, but
specifically to gains for cell edge users, whose quality improvement is one of the main motivations of
CoMP techniques. To complete this comparative study, future research should include the impact of
imperfect channel estimation on algorithmic performance.
The work reported in this deliverable about CoMP tries to pave the way to a real feasibility of CoMP
schemes, both introducing and extensively studying coordinated beamforming schemes and aiming to
ease the effective implementation of joint processing techniques. It’s worth noting and well known that if
low-latency exchanges are needed between cooperating BSs, then the corresponding CoMP schemes will
not be able to be deployed everywhere but only in the areas where a suitable backhaul is present, or they
will require important investments from the operator to upgrade the backhaul. This could somehow limit the practical usability of the techniques. For each technique referred in this document, requirements
concerning amount of data exchange and data exchange rate were highlighted qualitatively (see
introductory table for each section). Upper layers algorithm and proper scheduling management schemes
are important features as well in this framework; they are extensively analyzed in [WIN+D15].
When it comes to relaying in the framework of CoMP, various cooperative relaying schemes were
considered. The first innovation considers a relay-assisted interference channel with two communication
pairs and one relay node for each pair. Various cooperative and non-cooperative schemes are compared.
The second innovation considers distributed LDPC coding for a Decode and Forward type relay.
WINNER+ D1,8
Version: 1.0 Page 47 (75)
In both areas studied in this deliverable much work is still ongoing in close cooperation with the
standardization consortia and considering the main outcomes that will come from them. Evaluations of
the proposed schemes and further investigations are foreseen, relaxing ideal assumptions and aiming to
the definition of as much widely accepted as possible coordinated systems.
WINNER+ D1,8
Version: 1.0 Page 48 (75)
5. References
[3GPP07] 3GPP TSG-RAN1#48, Orange et al, “R1-070674, LTE physical layer framework for
performance verification”, February 2007, USA.
[3GPP09] 3GPP TSG-RAN1#58, Qualcomm Europe, “R1-093141 Signaling for spatial
coordination in DL CoMP”, August 2009.
[3GPP25814] 3GPP TSG RAN, “3GPP TR.25814, Physical Layer Aspects for Evolved UTRA (Release7)”, v7.1.0 (2006-09).
[3GPP36814] 3GPP TR 36.814, "Further Advancements for E-UTRA Physical Layer Aspects",
V1.0.0, Mar. 2009.
[AGS05] K. Azarian, H. El Gamal, and P. Schniter , “On the achievable diversity-multiplexing tradeoff in half-duplex cooperative channels”, IEEE Transactions on Information
Theory, December 2005.
[Ber03] D.P. Bertsekas, “Nonlinear programming”, 2nd edition, Belmont, Mass: Athena
Scientific, 2003.
[BH07] F. Boccardi and H. Huang, “Limited downlink network coordination in cellular networks”, in Proc. of PIMRC, September 2007
[BHA08] F. Boccardi, H. Huang and A. Alexiou, “Network MIMO with reduced backhaul requirements by MAC coordination”, in Proc. of Asilomar, 2008.
[BKV+07] S. Boyd, S. J. Kim, L. Vandenberghe, and A. Hassibi, “A tutorial on geometric programming,” Optimization and Engineering, vol. 8, no. 1, pp. 67–127, 2007.
[BO99] M. Bengtsson and B. Ottersten, “Optimal downlink beamforming using semidefinite optimization,” in Proc. Annual Allerton Conf. Commun., Contr., Computing,
Monticello, IL, Sept. 22–24 1999, pp. 987–996.
[BO01] M. Bengtsson and B. Ottersten, “Optimal and suboptimal transmit beamforming,” in
Handbook of Antennas in Wireless Communications, L. C. Godara, Ed. Boca Raton,
FL: CRC Press, 2001.
[BPG+08] C. Botella, G. Piñero, A. González and M. de Diego, “Coordination in a multi-cell multi-antenna multi-user W-CDMA system: a beamforming approach”, IEEE
Transactions on Wireless Communications, November 2008.
[Boy04] S. Boyd, “Notes on decomposition methods. Course reader for convex optimization II,
Stanford,” 2008, available online: http://www.stanford.edu/class/ee364b/.
[BAS+05] K. Brueninghaus, D.Astely, T.Salzer, S. Visuri, A;Alexiou, S.Karger, G-A Serail. “Link
Performance Models for System Level Simulations of Broadband Radio Access Systems”, in Proc. IEEE Int. Symp. Pers., Indoor, Mobile Radio Commun. Berlin,
Germany, Sep. 2005.
[BV04] S. Boyd and L. Vandenberghe, “Convex Optimization”. Cambridge, UK: Cambridge
University Press, 2004.
[CA07] W. Choi, J. Andrews, “Downlink Performance and Capacity of Distributed Antenna Systems in a Multicell Environment”, IEEE Transactions on Wireless Communications,
2007.
[CBS+07] Chakrabarti, A. de Baynast, A. Sabharwal, B. Aazhang, “Low Density Parity Check Codes for the Relay Channel”, IEEE Journal on Selected Areas in Communications,
February 2007
[CS03] G. Caire and S. Shamai (Shitz), “On the achievable throughput of a multi-antenna Gaussian broadcast channel”, IEEE Transactions on Information Theory, vol. 49, no. 7,
pp. 1691-1706, July 2003
[DPS+08] E. Dahlman, S. Parkvall, J. Sköld and P. Beming, “3G Evolution – HSPA and LTE for
Mobile Broadband”, 2nd ed., Academic Press, 2008.
[DY08] H. Dahrouj and W. Yu, “Coordinated beamforming for the multi-cell multi-antenna wireless system,” in Proc. Conf. Inform. Sciences Syst. (CISS), Princeton, NJ, USA,
Mar. 2008, pp. 429–434.
[DZW07] Y. Ding, J.-K. Zhang, and K. M. Wong, “The amplify-and-forward half-duplex cooperative system: pairwise error probability and precoder design”, IEEE Transactions
on Signal Processing, February 2007
WINNER+ D1,8
Version: 1.0 Page 49 (75)
[ETW06] R. Etkin, D.N.C. Tse, and H. Wang. Gaussian interference channel capacity to within one bit: the symmetric case. In IEEE Information Theory Workshop, pages 601–605,
Chengdu, China, October 2006.
[ETW08] R. Etkin, D.N.C. Tse, and H. Wang. Gaussian interference channel capacity to within one bit. IEEE Transactions on Information Theory, 54(12):5534–5562, December 2008
[HD07] J. Hun, T. Duman, “Low Density Parity Check Codes over Wireless Relay Channels”, IEEE Transactions on Wireless Communications, September 2007
[HFV06] M. K. Karakayali, G. J. Foschini and R. A. Valenzuela, “Network coordination for Spectrally Efficient Communications in Cellular Systems”, IEEE Wireless
Communications Magazine, Aug. 2006.
[HK81] T. Han and K. Kobayashi, “A new achievable rate region for the interference channel”, IEEE Transactions on Information Theory, vol. IT-27, no. 1, pp. 49-60, January 1981
[ITURM2135] ITU-R, “Report ITU-R M.2135: Guidelines for evaluation of radio interface
technologies for IMT-Advanced”, Nov. 2008.
[KAA04] M. A. Khojastepour, N. Ahmed, B. Aazhang, “Code Design for the Relay Channel and Factor Graph Decoding”, Asilomar Conference, 2004
[KF+06] K. Karakayali, G. Foschini, R. Valenzuela, and R. Yates, “On the maximum common rate achievable in a coordinated network,” in IEEE International Conference on
Communications, vol. 9, Istanbul, Turkey, June 2006, pp. 4333-4338.
|KGB07] G.M. Kraidy, N. Gresset, and J. Boutros, “Coding for the non-orthogonal amplify-and-forward cooperative channel,” IEEE, ITW 2007
[MHM+08] “Classification on Interference Management Proposals in TGm”, Contribution to IEEE
802.16m, IEEE C802.16m-08/142r6 (Mediatek, Huawei, Motorola, LGE, Alcatel-
Lucent, Nextwave, Mitsubishi Electric, Panasonic), Mar. 2008.
[MF07a] P. Marsch and G. Fettweis, “A Framework for Optimizing the Uplink Performance of Distributed Antenna Systems under a Constrained Backhaul”, in Proc. Of ICC Apr.
2007
[MF07b] P. Marsch and G. Fettweis, “A Framework for Optimizing the Downlink Performance of Distributed Antenna Systems under a Constrained Backhaul”, in Proc. of EW, Apr.
2007.
[NBK04] R. Nabar, H. Bolcskei, F. Kneubuhler, “Fading Relay Channels: Performance Limits and Space-Time Signal Design”, IEEE journal on selected areas in communications,
August 2004
[PBG+04] G. Piñero, C. Botella, A. González, M. de Diego and N. Cardona, “Downlink power
control and beamforming for a cooperative wireless system”, in Proc. of PIMRC, 2004.
[PBG+08] A. Papadogiannis, H.J. Bang, D. Gesbert and E. Hardouin, “Downlink overhead reduction for multi-cell cooperative processing enabled wireless networks”, in Proc. of
PIMRC, 2008.
[PC06] D. P. Palomar and M. Chiang, “A tutorial on decomposition methods for network utility maximization” IEEE J. Select. Areas Commun., vol. 24, no. 8, pp. 1439?1451, Aug.
2006.
[PGH08] A. Papadogiannis, D. Gesbert and E. Hardouin “A dynamic clustering approach in wireless networks with multi-cell cooperative processing”, in Proc. of ICC 08.
[PHG08] A. Papadogiannis, E. Hardouin and D. Gesbert, “A framework for decentralizing multi-cell cooperative processing on the downlink”, in Proc. of Globecom 2008.
[RFL09] P. Rost, G. Fettweis, and J.N. Laneman, “Opportunities, Constraints, and Benefits of Relaying in the Presence of Interference”, 2009 IEEE International Conference on
Communications, Dresden, Germany, June 2009
[RLT98] F. Rashid-Farrokhi, K. Liu, and L. Tassiulas, “Transmit beamforming and power control for cellular wireless systems,” IEEE J. Select. Areas Commun., vol. 16, no. 8,
pp. 1437–1450, Oct. 1998.
[RTL98] F. Rashid-Farrokhi, L. Tassiulas and K. J. R. Liu, “Joint optimal power control and beamforming in wireless networks using antenna arrays”, IEEE Trans. Commun., vol.
46, pp. 1313-1324, Oct. 1998.
[RY07] P. Razaghi, W. Yu, “Bilayer Low-Density Parity-Check Codes for Decode-and-Forward in Relay Channels”, IEEE Transactions on Information Theory, October 2007
WINNER+ D1,8
Version: 1.0 Page 50 (75)
[SB04] M. Schubert and H. Boche, “Solution of the multiuser beamforming problem with individual SINR constraints”, IEEE Trans. Vehicular Technology, vol.53, no.1, pp.18-
28, Jan. 2004.
[SCW+07] Y. Song, L. Cai, K. Wu and H. Yang, "Collaborative MIMO Based on Multiple Base
Station Coordination", Contribution to IEEE 802.16m, IEEE C802.16m-07/162 (Alcatel-Lucent), Aug. 2007.
[SEA03a] A. Sendonaris, E. Erkip, and B. Aazhang, “User cooperation diversity. Part I. System description”, IEEE Transactions on Communications, November 2003.
[SEA03b] A. Sendonaris, E. Erkip, and B. Aazhang, “User cooperation diversity. Part II. Implementation aspects and performance analysis”, IEEE Transactions on
Communications, November 2003.
[SGH08] H. Skjevling, D. Gesbert, A. Hjorungnes, “Low-complexity distributed multibase transmission and scheduling”, EURASIP Journal on Advances in Signal Processing,
2008.
[SVH06] M. Stojnic, H. Vikalo, and B. Hassibi, “Rate maximization in multiantenna broadcast channels with linear preprocessing”, IEEE Transactions on Wireless Communications
vol. 5, No. 9, pp. 2338–2342, November 2004.
[SWO08] N.Seifi, A.Wolfgang, T.Ottonson “Downlink performance and capacity of distributed antenna systems based on realistic channel model” ITG Workshop on Smart Antennas,
2008.
[TCJ08a] A. Tölli, M. Codreanu & M. Juntti, "Cooperative MIMO-OFDM Cellular System with Soft Handover between Distributed Base Station Antennas", IEEE Transactions on
Wireless Communications, Vol. 7, No. 4, pp. 1428-1440, April 2008.
[TCJ08b] A. Tölli, M. Codreanu & M. Juntti, "Linear Multiuser MIMO Transceiver Design with Quality of Service and Per-Antenna Power Constraints", IEEE Transactions on Signal
Processing, Vol. 56, No. 7, Jul. 2008.
[TPK09a] A. Tölli, H. Pennanen, and P. Komulainen, “SINR balancing with coordinated multi-cell transmission,” in Proc. IEEE Wireless Commun. and Networking Conf., Budapest,
Hungary, April 2009.
[TPK09b] A. Tölli, H. Pennanen, and P. Komulainen, “Distributed implementation of coordinated multi-cell beamforming,” in Proc. IEEE Int. Symp. Pers., Indoor, Mobile Radio
Commun., Tokyo, Japan, Sept. 2009, pp. 1–5.
[TPK09c] A. Tölli, H. Pennanen, and P. Komulainen, “Distributed coordinated multi-cell transmission based on dual decomposition,” in Proc. IEEE Globecom, Honolulu,
Hawaii, USA, Nov. 2009.
[Ven07] S. Venkatesan, “Coordinating base stations for grater uplink spectral efficiency in a cellular systems”, in Proc. of PIMRC, Sept. 2007
[VAL+06] M. Vemula, D. Avidor, J. Ling, and C. Papadias, “Inter-cell co-ordination, opportunistic beamforming and scheduling,” in IEEE International Conference on Communications,
2006. vol. 12, June 2006, pp. 5319–5324.
[VM99] E. Visotsky and U. Madhow, “Optimum beamforming using transmit antenna arrays,” in Proc. IEEE Veh. Technol. Conf., vol. 1, Houston, TX, May 16–20 1999, pp. 851–
856.
[WES06] A. Wiesel, Y. C. Eldar, and S. Shamai, “Linear precoding via conic optimization for fixed MIMO receivers,” IEEE Trans. Signal Processing, vol. 54, no. 1, pp. 161–176,
Jan. 2006.
[WIN2D6137] IST-4-027756 WINNER II D6.13.7 “Test Scenarios and Calibration Cases Issue 2”
Dec. 2006
[WIN2D6112] IST-4-027756 WINNER II D6.11.2 “Key Scenarios and Implications for WINNER II”
Sept. 2006
[WIN2D111] IST-4-027756 WINNER II D1.1.1 “WINNER II interim channel models“ Nov 2006
[WIN2D112] IST-4-027756 WINNER II D1.1.2 “WINNER II Channel Models“ Sept 2007
[WIN+D14] CELTIC / CP5-026 WINNER+, “D1.4 Initial Report on Advanced Multiple Antenna
Systems”, Jan. 2009.
[WIN+D15] CELTIC / CP5-026 WINNER+, “D1.5 Intermediate Report on system aspects of
advanced RRM”, October 2009
WINNER+ D1,8
Version: 1.0 Page 51 (75)
[WIN+D42] CELTIC / CP5-026 WINNER+, “D4.2 Final conclusions on end-to-end performance
and sensitivity analysis” to be delivered in April 2009.
[YE06] M. Yuksel and E. Erkip, “Diversity-Multiplexing Tradeoff in Cooperative Wireless Systems”, Proceedings of the 2006 Conference on Information Sciences and Systems,
Princeton University, New Jersey, March 2006.
[YSL+08] C. I. Yeh, Y. S. Song, S. J. Lee, B.-J. Kwak, J. Kim, and D. S. Kwon, "Frame Structure
to Support Inter-cell Interference Mitigation for Downlink Traffic Channel using Co-
MIMO and FFR", Contribution to IEEE 802.16m, IEEE C802.16m-08/017r1 (ETRI),
Jan. 2008.
[ZD04] H. Zhang and H. Dai, “Cochannel interference mitigation and cooperative processing in downlink multicell multiuser MIMO networks”, EURASIP Journal on Wireless
Communications and Networking, 2004.
[ZV03] B. Zhao and M.C. Valenti , “Cooperative diversity using distributed turbo codes”, Virginia Tech Symposium on Wireless Personal Communications, Blacksburg, VA,
June 2003
WINNER+ D1,8
Version: 1.0 Page 52 (75)
A. Appendix
A.1 System level performance evaluation of coordinated beamforming and
joint processing
In Section 2.2.1 a centralized non-codebook based scheme was described as an example of a coordinated
beamforming concept. In this appendix we will provide a detailed system level performance evaluation of
this scheme, and also compare it to the performance of a joint processing scheme based on zero-forcing
(ZF) precoding that was presented and initially evaluated in [WIN+D14]. The goal is to give some hints
on the achievable performance with different CoMP approaches, and some indications on in which
scenarios the approaches are most suitable. One interesting question is whether the increased system
complexity in terms of feedback overhead and backhaul requirements posed by joint processing pays off
in terms of system performance.
A.1.1 Simulation setup
The simulations were performed with a radio network simulator with a regular cell plan and hexagonal
cell layout. Three different macro scenarios are considered; the first two of them are the ITU scenarios
Urban Macro and Rural Macro as defined by ITU-R in [ITURM2135]. Then, results are also given for the
so-called 3GPP Case 1, which is an urban macro scenario defined in [3GPP25814]. For full details of the scenarios, the interested reader is referred to the documents cited above, whereas a condensed summary
of the used models and assumptions is given in the following.
In all scenarios, a regular cell plan and hexagonal cell layout is assumed. The studied deployment
comprises 19 sites, each with three sectors (cells) per site, which means in total 57 cells are simulated. A
wrap-around technique is used to avoid border effects. The differences between the scenarios are mainly
related to the site-to-site distances, antenna heights, user mobility and locations, etc., and are summarized
in Table A-1. Simulation assumptions common for all scenarios are provided in Table A-2.
Table A-1: Scenario specific parameters
ITU Urban Macro ITU Rural Macro 3GPP Case 1
Carrier frequency 2 GHz 800 MHz 2 GHz
Site-to-site distance 500 m 1732 m 500 m
BS antenna height 25 m, above rooftop 35 m, above rooftop 32 m, above rooftop
UE speed 30 km/h 120 km/h 3 km/h
User distribution Randomly and
uniformly distributed
over area, 100%
outdoor in vehicles
Randomly and
uniformly distributed
over area, 100%
outdoor in high-speed
vehicles
Randomly and
uniformly distributed
over area, 100% indoor
Outdoor-to-indoor
penetration loss
n/a n/a 20 dB
Outdoor-to-in-car
penetration loss
9 dB 9 dB n/a
WINNER+ D1,8
Version: 1.0 Page 53 (75)
Table A-2: Simulation assumptions
Transmission
bandwidth
10 MHz
Subcarrier
spacing
15 kHz
Number of
subcarriers
320
Frequency reuse 1
Number of cells 57
Number of users
per cell
10 (on average)
Wrap-around Yes
Interference modelling
All links modeled
Transmission
schemes
Reference non-CoMP system:
Codebook-based precoded adaptive rank MIMO [DPS+08],
4x2 co-polarized antennas, two groups of BS antennas, 10 λ spacing between
groups, λ/2 spacing within groups, MMSE-SIC UE receiver.
Joint processing CoMP:
Joint transmission over 9 sectors/cells (3 sites), zero-forcing precoding, dual stream,
4x2 co-polarized antennas, 10 λ spacing at BS, MMSE-SIC UE receiver.
Coordinated beamforming CoMP:
Coordination over 9 sectors/cells (3 sites), single stream per user, SDMA with up to
4 users per cell, 4x2 co-polarized antennas, λ/2 spacing at BS, MMSE-SIC UE
receiver.
Coding Practical turbo codes
Modulation QPSK, 16QAM, 64QAM
Link adaptation Non-ideal, based on delayed feedback
Data traffic model Full buffer
Three different system configurations are simulated, all based on 4x2 antenna configurations. As
reference, a non-CoMP system based on LTE release 8 is considered. In this case, each sector is equipped
with a clustered antenna array comprising four elements divided into two groups separated 10
wavelengths and with half a wavelength separation within the groups. The transmission scheme is
codebook-based precoded adaptive rank MIMO according to LTE release 8 [DPS+08]. The second
system configuration employs joint processing CoMP based on zero-forcing (ZF) precoding. Each sector
is equipped with a four-element uniform linear antenna array with 10 wavelength element separation.
Finally, the third system configuration is also a CoMP system, but now based on coordinated
beamforming. In this configuration, the sectors are equipped with four-element uniform linear antenna
arrays with half a wavelength element separation. In all system configurations, the UEs have two antenna elements separated half a wavelength and employs an ideal Minimum Mean Square Error (MMSE)
receiver with additionally Successive Interference Cancellation (SIC) functionality.
The layout for the reference non-CoMP system is depicted in Figure A-1a, where each cell (a numbered
hexagon) acts independently. To reflect the fact that each cell acts independently, each cell is depicted in
a different color. The locations of the sites are represented by red circles. In Figure A-1b the layout for the
CoMP system is illustrated. For simplicity, static, pre-defined CoMP clusters are considered. Each CoMP
cluster in the system is formed by grouping 9 neighboring cells (or sectors) served by 3 sites, resulting in
coordination clusters of 9 cells (i.e. the transmissions within these 9 cells are coordinated). The cells
belonging to the same coordination cluster are depicted in the same color.
WINNER+ D1,8
Version: 1.0 Page 54 (75)
a) Reference layout (each cell acts independently) b) CoMP layout, coordination clusters of 9 cells
Figure A-1: Cell layouts used in the simulations. Note that the site-to-site distance is the same in
both cases (i.e. the cell size is the same), and that wrap-around (not shown here) is used in the
simulations in order to avoid border effects.
All users are assumed to have full buffers, and the average number of users per cell is 10. The users are
uniformly distributed across the simulated area, and each user moves at a constant scenario-specific
speed, cf. Table A-1.
There are three available modulation schemes (QPSK, 16QAM, and 64QAM) and six different channel
code rates (1/10, 1/3, 1/2, 2/3, 3/4, and 8/9). The transport format is selected to maximize the expected
throughput, but the link adaptation is non-ideal in the sense that it is based on delayed feedback. Packet
decoding error probability is modeled according to a mutual information based link-to-system interface
[BAS+05].
The simulations assume perfect channel and interference estimation at the UE. The OFDM transmission
is further modeled as perfectly orthogonal and any potential inter-symbol or inter-carrier interference
caused by channel time dispersion exceeding the cyclic prefix is neglected. Overhead such as reference
signals, e.g. for channel and interference estimation, or protocol headers are neither accounted for. It is
further assumed that all necessary channel knowledge required by the CoMP schemes is available at the
BS. However, feedback and reporting delays are modeled.
The used performance measures are the cell spectral efficiency measured in bps/Hz/cell, and the cell edge
user spectral efficiency (the 5th percentile of the normalized user throughput) measured in bps/Hz. The
first one is focused on the system performance, while the second one may be described as a user centric
performance measure and/or the fairness in the system when put in relation to the cell spectral efficiency.
A.1.2 Simulation Results
Table A-3 shows the results achieved for the 3GPP Case 1 scenario. As can be expected in this low-
mobility scenario, it can be seen that the joint processing based on ZF precoding works very well. It
achieves a cell spectral efficiency of 3.81 bps/Hz/cell, which can be compared to that of the non-CoMP
system which is 2.56 bps/Hz/cell. Also the cell edge user spectral efficiency is improved from 0.074 to
0.108 bps/Hz. Both these improvements are in the order of 50% in relative terms. Coordinated
beamforming does also perform reasonable well here; it achieves a cell spectral efficiency of 3.01
bps/Hz/cell, which is 15-20% better than the non-CoMP system. However, the cell edge performance is
only slightly better, 0.078 bps/Hz compared to 0.074 bps/Hz,
The results for ITU Urban Macro are shown in Table A-4. In this case the UE speed is 30 km/h, and it can be seen that this hits on the performance of the ZF based joint processing transmission scheme. As can be
seen, both the cell spectral efficiency and the cell edge user spectral efficiency of joint processing based
on ZF is actually slightly lower than that of the non-CoMP system based on LTE release 8. This is of
course due to the fact that short-term channel state information needed for the ZF precoding gets
outdated, hence resulting in that the applied precoding weights are invalid. Coordinated beamforming, on
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
1
2 3
4
5 6
7
8 9
10
11 12
13
14 15
16
17 18
19
20 21
22
23 24
25
26 27
WINNER+ D1,8
Version: 1.0 Page 55 (75)
the other hand, is robust to user mobility and works well in this scenario. The simulated cell spectral
efficiency is 1.97 bps/Hz/cell and the cell edge user spectral efficiency 0.053 bps/Hz, compared to 1.32
bps/Hz/cell and 0.035 bps/Hz for the non-CoMP system. This translates into a relative gain over the non-
CoMP system in the order of 50%, both in cell spectral efficiency and cell edge user performance. It can
be noted that this relative gain over the non-CoMP system is higher than in 3GPP Case 1, which most
probably is explained by the fact that we now have outdoor users, and also a line-of-sight (LoS)
component in the channel model, which together makes it easier to exploit the directivity properties of the
beamforming.
Finally, Table A-5 shows results for the ITU Rural Macro scenario where the users are moving at 120
km/h. It can be seen that the joint processing based on ZF breaks down even further, and that the performance now is far below that of the non-CoMP system. Again, it is demonstrated that coordinated
beamforming is robust to user mobility, and the gain over the non-CoMP system is significant. It achieves
a cell spectral efficiency of 2.64 bps/Hz/cell compared to 1.52 bps/Hz/cell for the non-CoMP system,
which is a relative gain in the order of 70%. In cell edge user spectral efficiency the gain is approximately
40%, an improvement from 0.050 bps/Hz to 0.071 bps/Hz.
Table A-3: System level performance results for 3GPP Case 1
No CoMP Joint processing based
on ZF precoding
Coordinated
beamforming
Cell spectral efficiency
[bps/Hz/cell]
2.56 3.81 3.01
Cell edge user spectral
efficiency [bps/Hz]
0.074 0.108 0.078
Table A-4: System level performance results for ITU Urban Macro
No CoMP Joint processing based
on ZF precoding
Coordinated
beamforming
Cell spectral efficiency
[bps/Hz/cell]
1.32 1.20 1.97
Cell edge user spectral
efficiency [bps/Hz]
0.035 0.026 0.053
Table A-5: System level performance results for ITU Rural Macro
No CoMP Joint processing based
on ZF precoding
Coordinated
beamforming
Cell spectral efficiency
[bps/Hz/cell]
1.52 1.05 2.64
Cell edge user spectral
efficiency [bps/Hz]
0.050 0.019 0.071
A.1.3 Conclusions
In this appendix we have carried out dynamic system level simulations of two different CoMP
approaches. One of the approaches is joint processing based on zero-forcing precoding that was
introduced in [WIN+D14], while the second approach is coordinated beamforming as described in
Section 2.2.1. The results show that joint processing achieves best performance in low mobility scenarios,
but that coordinated beamforming also achieves gains over the reference non-CoMP system configuration
in these scenarios. With higher user mobility, already at 30 km/h, the performance of joint processing
WINNER+ D1,8
Version: 1.0 Page 56 (75)
degrades and is similar to that of the reference non-CoMP configuration, while coordinated beamforming
is robust to user mobility and provides significant gains.
Hence, the main conclusion is that coordinated beamforming type of shemes has the potential to provide
good performance in low mobility scenarios, however, not as good as joint processing type of schemes.
On the other hand, coordinated beamforming type of schemes is much more robust to user mobility. This,
in combination with the fact that coordinated beamforming type of schemes require lower feedback
overhead and put less demanding requirements on the backhaul connections in terms of latency and
capacity than joint processing type of schemes, suggests that coordinated beamforming schemes are
promising CoMP schemes to be realized in practical networks.
Finally, it should be emphasized that the results here are based on relatively ideal assumptions and that
the performance might change as more impairments are added and when practical implementation
limitations are taken into account. Nevertheless, the conclusions should remain the same since the
impairments and implementation limitations are expected to hit harder on the joint processing type of
schemes than on the coordinated beamforming type of schemes. It should also be pointed out that the
considered schemes in this study should be seen as exemplary configurations and not necessarily suitable
for implementation in a real system. However, in order to provide overall guidelines wrt technology
potential they fulfill their purpose.
A.2 Further details and performance evaluation of decentralized coordinated
beamforming
A comprehensive description and some main performance results of decentralized CoMP transmission
were given in Section 2.2.2. In this appendix, a summary and a detailed performance evaluation of
decentralized coordinated beamforming algorithm are presented. At the end, conclusion is drawn.
A.2.1 Decentralized coordinated beamforming algorithm
The decentralized coordinated beamforming algorithm with ZF mode selection is summarized in the
following.
1) Initialize 0t and set bb 0ν to some values, e.g., 0ν 0b .
2) Solve locally the distributed problem and transmit the resulting bζ to adjacent BSs.
3) Calculate the average inter-cell interference vector tζ and update local consistency prices
1tbν .
4) Use tζ in the distributed problem to get a feasible set of beamformers kkb ,m and the
minimum power bP .
5) Calculate ZFbP and ZF
,kbm using kkb 0ζ , .
6) IF bb PP ZF THEN
IF ZF−mode(b) is not active THEN
Set ZF−mode(b) active and send a message to the adjacent BSs
ELSEIF ZF−mode(b) is active THEN
Set ZF−mode(b) inactive and send a message to the adjacent BSs
7) IF any ZF−mode(b) b is active THEN
Use beamformers kkb ZF,m
ELSE
Use beamformers kkb ,m
8) Set 1 tt and go to step 1.
A.2.2 Numerical results
A simplified multi-cell transmission scenario with frequency flat fading with Jakes Doppler spectrum and
uncorrelated channels between antennas is considered, where 4K single antenna users are served
simultaneously by 2 BSs. We denote the maximum Doppler shift as df and the signaling period for the
exchange of bζ between BSs and ST . The number of antennas at each BS is 4T N . The simulation
scenario is depicted in Figure A-2. For simplicity, we assume that the users are divided into BN groups
where the users have identical large scale fading coefficients kba , , i.e., aaaaa 4,23,22,11,1 and
WINNER+ D1,8
Version: 1.0 Page 57 (75)
the path gain to noise ratio is normalised to 102 Na . Furthermore, the distance between different user
groups is kept at identical values. The distance is defined by a parameter that fixes the ratio of path
losses between the different user groups. For example, 21,2
23,2
23,1
21,1 aaaa . When the parameter
is fixed at 0 dB, all K users are located exactly on the cell border. On the other hand, cells are
completely isolated when . In the coordinated single-cell beamforming case, each stream is
transmitted from a single BS. A user k is always allocated to a cell with the smallest path loss,
kbb
a ,B
maxarg
.
Figure A-2: Simulation scenario
First, the sub-optimality of the cases with reduced signalling in a static scenario ( 0df ) with random
channel realizations are studied. The following coordinated beamforming cases are compared:
1. Optimal user-specific interference constraint kb,ζ
2. BS-specific interference constraint, bbkb k Uζζ ,
3. Common interference constraint, bkkb ,ζζ ,
4. Zero-forcing for the inter-cell interference, bkkb ,0ζ , (optimal intra-cell beamformer design)
5. Zero-forcing for both intra- and inter-cell interference (channel inversion)
The ZF approaches are possible in the considered scenario since KN T .
Figure A-3 and Figure A-4 illustrate the average sum power of 4,2,4,, TB NNK system as a
function of , to meet 0 dB and 10 dB SINR constraints, respectively. Each user has an equal SINR
constraint. The coherent multi-cell beamforming greatly outperforms all other simulation cases when the
distance between different user groups is finite. At the cell edge, the coordinated beamforming cases
require 5-6 dB more power than the coherent case in order to meet the 0 dB SINR target, as seen in
Figure A-3. All three coordinated beamforming cases have very similar performance. Thus, the loss from sub-optimal signaling is minor. The distance does not have any impact on the zero-forcing cases since
no interference is allowed towards the other cell users. However, there is a large gain from the optimal
intra-cell beamformer design ( bkkb ,0, ζ ) as compared to the channel inversion.
The loss from sub-optimal signaling increases significantly for the case with one common constraint
when the SINR target is 10 dB. However, the loss is still minor in the BS-specific case. Also, the gain
from the optimal intra-cell beamformer design (with ZF for the inter-cell interference) is greatly reduced
as compared to the channel inversion. In general, the difference between the zero-forcing and coordinated
beamforming cases is reduced significantly.
WINNER+ D1,8
Version: 1.0 Page 58 (75)
Figure A-3: Average sum power of 4,2,4,, TB NNK system for 0 dB SINR target.
Figure A-4: Average sum power of 4,2,4,, TB NNK system for 10 dB SINR target.
Figure A-5 and Figure A-6 demonstrate the cell edge ( 0 dB) performance of the distributed algorithm
in a time-correlated fading scenario as a function of dfTS to meet 0 dB and 10 dB SINR constraints,
respectively. The results for the distributed algorithm include cases with and without the ZF transmission
possibility. Figure A-7 compares the time evolution of the distributed algorithm with the centralized and
zero-forcing cases. The results demonstrate that the distributed algorithm performs nearly as well as the
centralized solution even at relatively high velocities, especially with low SINR targets. Note that 1
S 10dfT with 2S T ms and with 2 GHz center frequency corresponds to the velocity of 27 km/h.
For the 10 dB SINR target (and higher), high velocities or low signalling rates cause occasional high
power peaks. This is reflected in the increased average power. This can be at least partially alleviated by enabling the ZF mode possibility as illustrated in Figure A-6.
WINNER+ D1,8
Version: 1.0 Page 59 (75)
Figure A-5: Average sum power of 4,2,4,, TB NNK system for 0 dB SINR target.
Figure A-6: Average sum power of 4,2,4,, TB NNK system for 10 dB SINR target.
WINNER+ D1,8
Version: 1.0 Page 60 (75)
Figure A-7: Time evolution of the distributed algorithm with 0 dB SINR target, 05.0S dfT
A.2.3 Conclusion
A decentralized solution for the coordinated multi-cell multi-antenna minimum power beamformer design
problem was proposed. The beamformers can be obtained locally at each BS relying on limited backhaul
information exchange on the allowed inter-cell interference levels between adjacent BSs. The method is
able to guarantee feasible solutions even if the interference information is outdated or incomplete. The
proposed approach allows for a number of special cases, where the signaling is reduced at the cost of
somewhat sub-optimal performance. The numerical examples demonstrated that a near-optimal performance could be achieved even with significantly reduced backhaul information exchange and with
relatively high velocities and/or low signaling rates. Occasional high peaks in the transmitted power due
to outdated interference terms can be alleviated by switching to interference nulling mode when
necessary.
A.3 Simulation conditions details for codebook-based coordinated beamforming
This appendix provides more details on the simulation conditions for the results presented in section 2.2.3
The table below summarizes the simulation conditions.
Table A-6: Simulation conditions for codebook-based coordinated beamforming.
Parameter description Value
Cellular Layout Hexagonal grid, 19 sites, 3 sectors per site
Inter-site distance (ISD) 500 m
Traffic model Full buffer
Number of UEs per cell 10 in average
Number of snapshots & TTIs per simulation 100 snapshots, 700 TTIs per snapshot
Node B
Transmission power 46 dBm
Ante
nnas
Number of TX antennas 4
Antenna gain plus cable loss 14 dBi
Antenna pattern
25,
7012min
2 dB ( is angle in degrees)
70-degree sectored beam.
Propagation
WINNER+ D1,8
Version: 1.0 Page 61 (75)
Pro
pag
atio
n
Path loss R10log6.371.128 dB ( R is distance in km)
Slow
fading
Standard deviation 8 dB
Correlation between
sites 0
Penetration loss 20 dB
Inter-cell interference Fully modelled, using the central cell method
UE
Antenna pattern 0 dBi
Number of RX-antennas 2 with MMSE spatial combining
Channel estimation Ideal
Noise figure 9 dB
Scheduler
Policy Round Robin
MIMO schemes
MIM
O
MIMO schemes Codebook-based beamforming (Single-User
MIMO) using LTE Rel-8 codebook
Cells coordination
Cluster size 3 cells: the master and the 2 dominant long-
term interferers of the UE served by the
master
Number of reported interfering beams 1, 3
Coordinated beamforming variant Most interfering beams
The coordination procedure described in section 2.2.3.is modelled as follows: at each TTI, scheduling
decisions are first performed as in the case of no coordination. The cells serving cell-edge UEs are then
identified as candidates for being a master. For each candidate master cell, the cluster is formed with the 2
dominant long-term interferers of the served UE, provided those interfering cells are not member of a
cluster already (as slave or master). Note that for the cluster formation, the candidate master cells are
picked in random order in order to ensure that the master role is fairly distributed among the cells.
A.4 Performance investigation of joint processing schemes considering area
coverage.
In this appendix, further evaluations of the joint processing schemes introduced in Section 2.3.1 are provided.
A.4.1 Centralized joint processing
Consider a cluster of K BSs, each one equipped with tN antennas, where M single-antenna users are
using a particular orthogonal dimension. When joint processing between BSs is allowed, the data to each
user is simultaneously transmitted from multiple BSs. In this case, a total of tK N antennas transmit
coordinately in the system, where tK N M . In this contribution, joint processing between BSs is
implemented by means of a joint linear precoding design and power allocation. Then, the received signals
at the M users can be expressed by means of a vector y of size 1M , as:
y = HW Px + n (A.1)
In the above expression, the matrix H of size tM K N includes the channel vectors of the system:
1
TT T
M H = h … h ,
where NtK
m C 1h stands for the channel between the m th user and the K BSs. Notice that nmC
denotes the set of m n complex matrices.
In the transmitter side, the joint processing is reflected in the choice of the W and P matrices. The
precoding matrix W of size tK N M contains the precoders designed for each of the users:
WINNER+ D1,8
Version: 1.0 Page 62 (75)
1 M W = w … w ,
where 1w NtK
m C is the precoder for the mth user. In this case, the precoder matrix W is obtained
using a ZF approach, that is, M MHW = I , where M M
I is an identity matrix of size M M .
Since tK N M , the ZF matrix is the pseudo-inverse of the channel matrix:
1
H H
W H HH (A.2)
In the above expression, H
and 1
stand for the conjugate transpose and matrix inversion
operations, respectively.
The power allocation matrix P is a diagonal matrix of size M M . In this case, the maximum
available transmit power at each base station is restricted to a maxP value. For simplicity, equal power
allocation is performed by means of the expression for matrix P given in [ZD04]:
max
21, , ( )min
M Mk K k
F
P
P I
W
(A.3)
where ( )k
W are the rows of matrix W related to the kth base station, and F
stands for the Frobenius
norm of a matrix. It should be pointed out that this power allocation is suboptimal, since it typically
results in only one base station meeting the maximum transmitted power requirement with equality, and
hence, the remaining 1K BSs transmit below the maxP value.
Finally vector x of size 1M includes the precoded information symbols and n is the receiver noise
with variance 2
, which is spatially and temporally white and is also uncorrelated with the signals.
Under these assumptions, and assuming coherent multibase reception, the evaluation metric under consideration is the average sum-rate per cell:
2
1
1log 1 SINR
M
H m
m
C EK
(A.4)
where the SINR for the mth user is given by the following expression:
2
2 2
1
SINRm m m
m M
m i i
ii m
p
p
h w
h w
(A.5)
In the above expression, mp is the power allocated to the mth user, that is, 2( , )m mmp P .
A.4.2 Partial joint processing
The partial joint processing scheme introduces a certain level of multi-user interference in the system due
to the limited CSI available in the central unit. This multi-user interference contribution can be defined by
analyzing the expression of the signal received by one user. Assume that mBS is the active set of BSs that
give service to the mth user, whereas kM is the set of users that are served by the kth base station. Note
that the cardinality of any mBS is such that 1 mBS K , whereas the cardinality of any kM ranges
between 0 kM M . Hence, for the mth user, the received signal can be expressed as a sum of the
signal of interest, multi-user interference and noise:
WINNER+ D1,8
Version: 1.0 Page 63 (75)
( ) ( ) ( )
( ,:) (: , )
( ) ( ) ( ) ( ) ( ) ( )
( ,:) (: , ) ( ,:) (: , )
1 1
,
m
jk
m m
k k k
m m m m m
k BS
MMk k k j j j
m i i i m i i i
k BS i ij BSi m
m
y p x
p x p x
n
H W
H W H W (A.6)
where mBS is the complement set of mBS .
( )
( ,:)
k
mH stands for the mth row of the tM N matrix ( )k
H ,
which includes the columns of matrix H related to the kth base station (in the case of ( )
(: , )
k
mW , the same
applies to the rows of matrix W ). Similarly, ( )k
mp is the power allocated to the mth user from the kth base
station. In the above expression, it is assumed that the channel coefficients ( )
( ,:)
j
mH included in the multi-
user interference term cannot be estimated by the mth user, since those BSs are not included in mBS .
Therefore, this term represents the multi-user interference contribution that remains in the system when
the partial joint processing scheme is implemented by the cluster of BSs.
A.4.3 Distributed joint processing
The distributed joint processing scheme assumes a local per-base station design of the linear precoding
matrix and power allocation, since only local CSI is available at each base station. Hence, the cardinality
of the set of spatially separated users that can be served by each base station in the cluster is reduced to
tN . Notice that in a multicarrier scenario, tN users can be served per subcarrier.
Joint processing between BSs is still allowed. In the system model, there are K BSs and M users, with
tK N M . Since each base station can give service only to tN users in the spatial domain, the problem
of assigning users to BSs under a joint processing assumption arises. This multibase scheduling problem
has been previously considered in [SGH08], where low-complexity algorithms have been proposed in
order to optimize a given objective function.
In this contribution, the multibase scheduling problem is solved as follows: kM includes the set of tN
users that present the highest channel gain with respect to the kth base station. This approach is similar to
the active set procedure proposed for the partial joint processing scheme. However, the decision process
of determining which base station transmits to each user is now moved towards the base station, and solved in a distributed manner. As shown in [SGH08], this multibase scheduling solution results in
different degrees or stages of coordination in the cluster depending on the distribution of the users over
the cluster area and the system parameters, i.e., each of the M users can be served by a number of BSs
that ranges from zero to K . Hence, the distributed joint processing scheme implies that a certain number
of users in the cluster may remain without service and then, some sort of fairness mechanism would be
additionally required.
The signal received by one user can still be modeled with the expression proposed for the partial joint
processing scheme, where the linear precoding matrix ( )k
W is the pseudo-inverse of the channel matrix
( )kH and the power allocation is performed equally dividing the maximum available transmit power
maxP into tN users.
A.4.4 Numerical results
We consider a cluster of 3K BSs, each one equipped with an array of 3tN antennas, and 3M
single-antenna users. The objective of the simulations is to characterize and compare the performance of
the centralized joint processing (CJP), partial joint processing (PJP) and distributed joint processing (DJP)
schemes under different evaluation metrics and a non-uniform distribution of users in the cluster area, see
Figure A-8. The cluster radius and height are 500R and 433h meters, respectively.
The channel vector between the mth user and the kth base station is modeled as:
WINNER+ D1,8
Version: 1.0 Page 64 (75)
'
mk mk s p h h ,
where the shadow fading is a random variable described by a log-normal distribution, 0,8s dBN , the
pathloss follows the 3GPP LTE model, 10 ( )148.1 37.6 logp mkdB r , and
'
mkh includes the small-
scale fading coefficients, which are i.i.d complex Gaussian values according to 0,1CN .
In the simulations, a grid of possible user locations is defined over the cluster area. Then, the users are
uniformly placed over a small area around each location, ( ,x x y y , with 16x R and
16y h ), and different metrics are evaluated. The results are averaged over 500 independent channel
realizations for each position of the grid. Different system SNR values are also simulated (reference value
for one user located at the cell-edge), in order to define noise and interference-limited scenarios.
Figure A-9 and Figure A-10 show the average sum-rate per cell for different transmission schemes in a
noise and interference-limited scenario, respectively. Results are plotted versus the normalized distance from one base station, [Distance/R] (the symmetry of the cluster area guarantees that no information is
lost when changing the 3-D plot into a 2-D plot). “PJP-10dB”, “PJP-20dB” and “PJP-40dB” plots stand
for the results of the PJP scheme when threshold values of 10, 20 or 40 dB are simulated. For comparison
purposes, results for the conventional single-base station transmission scheme, “1BS”, are also included.
Results labeled with “2BSs” are obtained when 2 BSs from the cluster transmit to each user. It should be
noticed that this is a particular case of the PJP scheme.
Comparing both figures, it can be seen that the differences between the simulated transmission schemes
arise in interference-limited scenarios (Figure A-10). The CJP scheme clearly outperforms the remaining
schemes (the “PJP-40dB” achieves the same performance since almost 3 BSs are jointly transmitting to
each user when the threshold value is fixed to 40 dB) at the cost of higher backhaul and signalling
requirements. On the other hand, the PJP scheme shows a trade-off between the backhaul and signaling requirements and the achieved average sum-rate per cell, that is, its performance improves as the
coordination degree between BSs or the threshold value increases. If we compare the performance of the
PJP curves with the “2BSs” scheme results, we can conclude that allowing a different number of BSs to
transmit to each user, depending on the user channel conditions, can also improve the average sum-rate
per cell, since the flexibility of the system is increased. Finally, it should be noticed that the DJP scheme
is not a solution for noise-limited environments. However, its performance is close to the “2BSs” and
“PJP-10dB” cases for interference-limited scenarios.
From a practical point of view, the major drawback in the spatial domain of the DJP scheme is the
limitation in the maximum number of served users per base station, which may require an additional
fairness mechanism depending on the requirements of the users. However, this can be easily done by optimizing the multibase scheduling technique, or by exploiting the subcarrier allocation.
Another interesting conclusion is the fact that each transmission scheme shows a different behavior when
taking into account aspects of fairness or uniformity of the evaluation metric over the cluster area.
Transmission schemes implying a joint design of the linear precoding matrix (CJP and PJP with
threshold 20 dB), achieve a higher uniformity of the average sum-rate per cell over the cluster area,
especially in the interference-limited scenarios as shown in Figure A-10.
On the other hand, these schemes also show different levels of robustness when computing the evaluation
metric. In this case, the increased robustness of the CJP scheme and the PJP scheme with
threshold 20 dB is revealed as a lower standard deviation of the evaluation metric regardless of the
position of the user over the cluster area, e.g., for a system SNR of 15 dB, the DJP scheme shows a 50 %
worse standard deviation over the average sum-rate per cell and less uniformity over the cluster area than
the CJP scheme.
Figure A-11 and Figure A-12 show the average number of BSs that are included in the active set of a user
in the PJP scheme for a threshold value of 10 and 20 dB, respectively (when the threshold value is fixed to 40 dB, we obtain a flat surface where the 3 BSs are always included in the active set of the user). This
parameter can be seen as a rough estimation of the backhaul and signalling requirements of the PJP
scheme. The figures confirm that the backhaul and signaling requirements of the PJP scheme depend both
WINNER+ D1,8
Version: 1.0 Page 65 (75)
on the threshold value and the user position over the cluster area, whereas they remain fixed for the CJP
and DJP transmission schemes. This result is of great importance since the threshold value may
dynamically change in time in order to fulfill the given system or user requirements. As a last remark, it
should be pointed out that also the transmitted power per base station of the PJP scheme depends on the
threshold value, e.g., for a threshold value of 10 dB, in average, a 14.14 % of the total transmitted power
in the system is saved when compared to the CJP and DJP schemes. This value decreases to 5.74 % when
the threshold value is set to 20 dB.
Figure A-8: The cluster area under consideration is the shadowed area close to the cell-edge of each
cell.
Figure A-9: Average sum-rate per cell versus the normalized distance from the base station
[Distance/R] for a system SNR of 0 dB.
WINNER+ D1,8
Version: 1.0 Page 66 (75)
Figure A-10: Average sum-rate per cell versus the normalized distance from the base station
[Distance/R] for a system SNR of 15 dB.
Figure A-11: Average number of BSs in the partial joint processing scheme transmitting to each
user versus user position in the cluster area. Threshold value = 10 dB.
WINNER+ D1,8
Version: 1.0 Page 67 (75)
Figure A-12: Average number of BSs in the partial joint processing scheme transmitting to each
user versus user position in the cluster area. Threshold value = 20 dB.
A.5 Joint processing with reduced backhaul requirement by MAC coordination
A system simulator has been developed with 19 single antenna base stations and wraparound. Each single-antenna user is dropped with uniform probability inside each cell. Fairness is guaranteed by a
proportional fairness scheduler. In other words, the quality of service weights 1, , K are the
reciprocal of the users’ average windowed rates. The reference SNR is defined as the SNR at the cell
vertex. The channel has been modelled considering Rayleigh and path loss effect. The values of the main
parameters used for the simulation are summarized in Table A-7
Table A-7. Main parameters used in the simulations.
N (number of base stations) 19
center to edge distance 1km
number of users per cell 30
number of base station clusterings 30
reference SNR 15dB
path loss exponent 4.5
prop. fair forgetting factor 0.01
-4000 -3000 -2000 -1000 0 1000 2000 3000 4000
-4000
-3000
-2000
-1000
0
1000
2000
3000
4000
1
2
3
4
5
6
7
8
9
10
11
1213
14
15
16
17
18
19
distance [m]
dis
tance [
m]
x
xx
x
x x
x
x
x
x
x
x
x
x
x
x
x
x
x
Figure A.13 Example of a clustering and of the corresponding user scheduling.
WINNER+ D1,8
Version: 1.0 Page 68 (75)
In Figure A.13 an example of a clustering and of the corresponding user allocation is shown for a given
frame and for the case B=4. Cells with the same colour belong to the same cluster, and the positions of
the users scheduled for transmission are denoted by a bold x.
In Figure A.14 and Figure A.15 the performance of the proposed algorithm is shown respectively in terms
of average rate per cell and average rate per user (sorted from the worst to the best user) for the case B=10, which corresponds to a 50% reduction in the number of base stations sharing the data of the users
scheduled for transmission in a given frame.
Four different techniques are compared: non-cooperating base stations (dotted line), static coordination
(dash-dotted line), i.e. 2N
LB
clusters of cooperating base stations are kept fixed during all the
simulation. In each cluster the users are selected for transmission using a proportional fair scheduler, dynamic coordination (dashed line), full coordination, i.e. all the N base stations cooperate together and
up to N users are scheduled for transmission in each frame with a proportional fair approach.
0 2 4 6 8 10 120
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
x [bit/s/Hz]
P[a
vera
ge r
ate
per
cell<
x]
no coord.
static coord.
dynamic coord.
full coord.
Figure A.14 CDF of the average rate per cell.
The proposed dynamic coordination technique performs near to the full coordination case, while almost
halving the number of cells that cooperate in a given frame. As a matter of fact, in the full coordination
case the number of cooperating base stations is 19, while in the dynamic coordination case the two
clusters of cooperating base stations are composed by 10 and 9 base stations. Such a reduction in the
number of base stations that cooperate at the same time allows a reduction of about 50% in terms of the
number of users whose data needs to be shared in order to allow joint transmission. Furthermore the
static coordination technique offers only a slightly improved edge-of-cell performance with respect to the
non-coordinated case (left part of Figure A.15), at the price of a reduced average rate per cell (Figure A.14). The tail of the static coordination curve in Figure A.15 is due to the attempt to serve the users that
are at the edge of two base stations belonging to two different clusters.
WINNER+ D1,8
Version: 1.0 Page 69 (75)
0 100 200 300 400 500 6000
0.1
0.2
0.3
0.4
0.5
sort(user index)
avera
ge p
er-
user
rate
[bit/s
/Hz]
no coord.
static coord.
dynamic coord.
full coord.
Figure A.15. Average rate per user vs user index (sorted from the worst to the best user).
A.6 Further details and performance evaluation of generalized CoMP
transmission method
A detailed description and some performance results of generalized CoMP transmission were given in
Section 2.3.3. This appendix presents supplementary details for beamformer design with coordinated BS
processing and introduces some beam allocation algorithms as well as presents a detailed performance
evaluation of the proposed concept. At the end, conclusion is drawn.
A.6.1 Beamformer design with coordinated BS processing
Details for solving two different optimization problems, i.e., weighted SINR balancing and weighted sum
rate maximization are provided. First, weighted SINR balancing is considered where the objective is to
keep the SINR per data stream s in fixed ratios in order to guarantee fairness between streams/users, i.e.
0 ss , and 0 has to be maximized subject to per BS power constraints. This can be formulated as
maximization of the minimum weighted SINR per stream:
,...,1, s.t.
minmax
2
2,
2
1,1
,.H
,
2
20
2
,.H
,1
,...,1
b
i
ss
s
ss
Ss
Bbsb
S
si b
ibkbskbs
b
sbkbskbs
Ss
NbP
aN
a
m
mHww
mHw
(A.7)
where the variables are TC,N
sb m and skN
sR
Cw , Ss ,...,1 . The SINR balancing problem is not
jointly convex in variables sb,m and sw . However, for a fixed Sssb ,...,1,, m it has a unique solution
given by the linear minimum mean square error (LMMSE) receiver which provides the maximum SINR
for stream s [TPK09a]. Furthermore, for a single antenna receiver or a fixed sw , optimization problem
is quasi-convex in sb,m [TCJ08b]. Thus, it can be solved by using the bisection method [Boy04]
presented in Algorithm 2 in [TCJ08b]. The constraints of the feasibility problem presented in Algorithm 2
in [TCJ08b] can be expressed as a generalized inequality with respect to the second-order cone [Boy04].
Finally, a solution for the SINR balancing problem can be found by using a coordinate ascent method
[Ber03], i.e., at each iteration, the objective is maximized with respect to one set of variables sw (or
sb,m ) while considering the other set fixed. In other words, alternating between calculation of sw using
LMMSE receiver and Algorithm 2 in [TCJ08b] until convergence for fixed sb,m and sw , respectively.
More details can be found in [TCJ08b, Algorithm 1]. The block coordinate ascent method converges to
WINNER+ D1,8
Version: 1.0 Page 70 (75)
the global optimum if the problems solved at each step have unique solutions [Ber03]. The optimal
objective value 0 for fixed sw is indeed unique, but the resulting Sssb ,...,1,, m is not guaranteed to
have a unique solution in general, due to the quasi-convexity of the original problem [TCJ08b].
Therefore, global optimality of the above method can only be guaranteed for the fixed and single-antenna
receiver cases.
Now, weighted sum rate maximization problem is considered for coordinated multi-cell transmission subject to per BS power constraints. The problem can be formulated as follows
Ssp
NbPp
Ss
paN
pa
sbss
b
s
sbsb
s
S
si b
ibibkbskbs
b
sbsbkbskb
s
S
s
s
b
i
ss
s
ss
s
,...,1,0,1,1
,...,1
,...1 , s.t.
1max
,22
B,
S
2
2,,
,1,1
2
,,,H
,
2
20
2
,,,H
,
1
vw
v
vHww
vHw
(A.8)
where the variables are TC,N
sb v , skN
sR
Cw , R,sbp and Rs , Ss ,...,1 and
T,B
T,1 ,..., sss
svvv is the normalized transmit vector over the joint processing set sB for the stream s .
The optimization problem is not convex, and, hence, the problem of finding the global optimum is intrinsically non-tractable. However, the problem can be maximized with respect to different subsets of
variables by considering the others fixed. A detailed analysis of the problem solving algorithms are given
in [TCJ08a] for coherent multi-cell MIMO transceivers. An efficient iterative solution was proposed
where each sub-step was solved as SOCPs or geometric programs [BKV07]. Even though each sub-
problem was optimally solved, only local optimal solutions could be found due to the non-convexity of
the original problem. Now, the same method can be extended to the general coordinated multi-cell
transmission by modifying the constraints of all sub-problems presented in [TCJ08a, Chapter III]. For the
sake of clarity and space, the exact details are omitted in this report.
A.6.2 Beam allocation algorithms
In the coordinated single-cell beamforming case ( ss 1B ), each stream is transmitted from a single
BS. In such a case, a user sk is typically allocated to a cell b with the smallest path loss. Near the cell
edge, however, the optimal beam allocation strategy may also depend on the time varying properties of
the channel kb,H . Thus, large gains from fast BS assignment (cell selection) algorithms are potentially
available for the cell edge users. As the optimal BS assignment per beam requires an exhaustive search
over all possible combinations of beam allocations and re-computation of the algorithm for each
allocation, sub-optimal allocation algorithms are needed in practice.
The allocation problems have often been addressed for systems with single-antenna users. When the users
have multiple receive antennas, the signal space of each user has multiple dimensions allowing for multiple beams to be allocated per user. This further complicates the allocation problem. Since the
transmit beamforming vectors , and thus, the corresponding receive beamforming vectors at each user are
affected by the set of selected users, it is impossible to know the actual receiver structure at the
transmitter before the final beam allocation. An obvious candidate for an intelligent initial guess of the
receiver matrix, and the one used in the proposed algorithms, is the optimum single user receiver, i.e., the
dominant left singular vectors of H,,,, kbkbkbkb VΛUH . Hence, each user channel kb,H is decomposed
into TR ,min NNk
virtual single-antenna users (eigenbeams) with corresponding channel gains
TR,,,,,,,,,,, ,min,...,1, NNlaak
Hlkblkbkbkb
Hlkbkblkb vHuh
where lkb ,,u , lkb ,, and lkb ,,v correspond to l th column (or diagonal term) of kb,U , kb,Λ and kb,V ,
respectively.
WINNER+ D1,8
Version: 1.0 Page 71 (75)
The aim of any beam-to-BS allocation algorithm is to select such BSs that the resulting beamformers
mutually interfere as little as possible while providing large beamforming gains towards the intended
users. A set of heuristic allocation algorithms for the coordinated beamforming case are considered in the
following:
1. Greedy selection: The algorithm consecutively selects at most R Tmin ,kk
N N channels from
the total set of R Tmin ,kk
N N channels. First, the strongest channel among all channels
lkblkb
,,,,
maxarg h is selected. Subsequently, on each step of the selection process, the channel with
the largest component orthogonal to the previously selected set of beams is chosen.
2. Maximum eigenvalue selection: The eigenvalues of the virtual channel vectors lkb ,,h are simply
sorted and at most TN streams are allocated per cell. Spatial compatibility with other channels is
not considered.
3. Eigenbeam selection using max rate criterion: A simplified exhaustive search is carried out over
all possible combinations of beam-to-BS allocations, where the beamformer sb,m for each
stream s is matched to the virtual channel vector ss lkb ,,h i.e., blkbsb P
ssST,,, vm . For
each allocation, the receivers sw and the corresponding SINR values Sss ,...,1, are
calculated. Finally, the selection of the optimal allocation is based on the maximum rate
criterion, i.e.,
S
s
slkb 1
2,,
)1(logargmax .
4. Eigenbeam selection using maxmin SINR criterion: Same as above except that the selection is
based on maxmin SINR criterion, i.e., sSslkb
,...,1,,
minargmax
.
It is worth noting that the usage of the greedy approach is rather limited since it can only be used when
TNS . Thus, it cannot be applied to the interference limited scenarios, i.e., TNS .
Figure A-16: Simulation scenario for 4 users and 2 BSs
A.6.3 Numerical results
A simplified multi-cell transmission scenario with frequency flat fading is considered where 42K
users with 21R N receive antennas are served simultaneously by 2 BSs. For simplicity, the BSs were
assumed to have an equal maximum power limit TP , i.e. bPPb T . The SNR for each user k was
based on the smallest pathloss among BN BSs and defined as 02,
BT maxSNR NaP kb
bk
. In the
simulations, the elements of the channel matrices kb,H were modelled as i.i.d. Gaussian random
variables. The simulation scenario is depicted in Figure A-16. For simplicity, we assume that the users are
divided into BN groups where the users have identical large scale fading coefficients kba , , i.e.,
aaaaa 4,23,22,11,1 and the path gain to noise ratio is normalised to 102 Na . Furthermore,
the distance between different user groups, as well as, SNR per user were kept at identical values. The
distance is defined by a parameter which fixes the ratio of path losses between the different user
groups. When the parameter is fixed at 0 dB, all K users are located exactly on the cell border. On the
other hand, cells are completely isolated when .
WINNER+ D1,8
Version: 1.0 Page 72 (75)
We study the achievable sum rate of the SINR balancing and the sum rate maximization algorithms both
with per BS power constraints and with equal weighting of streams ss 1 . The following cases were
compared by simulations:
1. Coherent multi-cell MIMO transmission ( ss BB ) with per BS power constraints.
2. Coordinated single-cell transmission ( ss 1B )
o Exhaustive search over all possible combinations of beam allocations.
o Fixed allocation, i.e., user sk is always allocated to a cell b with the smallest path loss.
o Maximum eigenvalue selection.
o Eigenbeam selection both with max rate and maxmin SINR allocation criteria.
o Greedy selection (can be applied when TNS ).
3. Non-coordinated single-cell transmission where the inter-cell interference is neglected in the
precoder design.
4. Single-cell transmission with TDMA, i.e., without inter-cell interference.
The inter-cell interference is omitted in the precoder design both in the non-coordinated transmission case
and in the TDMA case, and thus the resulting beamformers are identical for both cases. In the TDMA
case, the transmission is time multiplexed between the BSs, and hence the reception is interference free.
In Figure A-17, the sum rate maximization (solid) and the SINR balancing (dashed) algorithms are
compared in 1,2,2,4,,, RTB k
NNNK scenario in terms of the ergodic sum rate as a function of the
distance between different user groups, and for 20 dB single link SNR. As expected, the sum rate
maximization largely outperforms the SINR balancing in all simulated cases. The most notable increase
in sum rate can be observed in the coordinated single-cell beamforming case. This is due to the fact that
the sum rate maximization algorithm ends up assigning zero powers to some users and hence inherently
avoids the spatial overload.
It was already shown in [WIN+D14] and [TPK09a] that in the case of SINR balancing algorithm the
coherent multi-cell beamforming greatly outperforms all the non-coherent cases, e.g., coordinated/non-
coordinated single-cell beamforming. This is also the case for the sum rate maximization algorithm. This
is due to the fact that the coherent multi-cell beamforming can fully eliminate the inter-cell interference
unlike the single-cell beamforming methods. However, the sum rates of coherent and non-coherent schemes become asymptotically equivalent as the distance approaches the infinity, i.e., there is no gain
from the coherent coordinated multi-cell processing.
Figure A-17: Ergodic sum of user rates of 1,2,2,4,,, RTB k
NNNK system with SINR balancing
and sum rate maximization algorithms at 20 dB single link SNR
WINNER+ D1,8
Version: 1.0 Page 73 (75)
Figure A-18: Ergodic sum of user rates of 1,2,2,2,,, RTB k
NNNK and 2,2,2,2 systems with
sum rate maximization algorithm at 20 dB single link SNR
Figure A-18 shows the ergodic sum rates of systems 1,2,2,2,,, RTB k
NNNK (dashed) and 2,2,2,2
(solid) for the sum rate maximization algorithm at 20 dB single link SNR. The system with 1R k
N is
not interference limited since TNS whereas the system with 2R k
N can be spatially overloaded
( TNS ) at the cell edge. However, as mentioned previously, the sum rate maximization algorithm
inherently avoids the spatial overload by assigning zero (or near to zero) powers to some streams. Hence,
using two receive antennas instead of one large gains are available to all simulation cases as can be seen
from Figure A-18.
In Figure A-19, coordinated single-cell beamforming with different beam-to-BS allocation algorithms
with the aim of sum rate maximization are compared in 2,2,2,2,,, RTB k
NNNK scenario at 20 dB
single link SNR. Near the cell edge, the optimal beam allocation strategy depends heavily on the
properties of channel realizations between BSs and users. Hence, large gains from different beam
allocation algorithms are available for the cell edge users. It can be seen that the channel dependent beam
allocation is beneficial within about 5 dB region around the cell edge. Otherwise, a beam should be
allocated to a cell with the smallest path loss.
Sub-optimal (heuristic) beam allocation algorithms perform relatively well at the cell edge compared to
the optimal case, i.e., exhaustive search. In particular, a simple maximum eigenvalue based selection
algorithm performs nearly the same as the optimal case.
Note that coordinated single-cell beamforming with any beam allocation algorithms as well as coherent
multi-cell beamforming cases asymptotically achieve the sum capacity in 2,2,2,2,,, RTB k
NNNK
system. This is due to the fact that the two-cell scenario reduces to a two totally isolated single-cell single
user cases when the approaches infinity, and thus, the optimal power allocation and beamforming are
easy to achieve.
In Figure A-20, optimal beam allocation algorithm with coordinated/non-coordinated single-cell
beamforming cases are compared to the same beamforming cases but where a single UE is served from a
single BS only. The simulation scenario is the same as previously. It can be seen that for
2,2,2,2,,, RTB k
NNNK system optimal beam allocation algorithm provides approximately a 2 dB
WINNER+ D1,8
Version: 1.0 Page 74 (75)
and a 1 dB gain with coordinated and non-coordinated beamforming, respectively. Thus, it is beneficial
for the system if a single user with multiple receiver antennas can be served from multiple nearby BSs.
Figure A-19: Ergodic sum of user rates of 2,2,2,2,,, RTB k
NNNK system with different beam-
to-BS allocation algorithms at 20 dB single link SNR
Figure A-20: Ergodic sum of user rates of 2,2,2,2,,, RTB k
NNNK system with different beam-
to-BS allocation algorithms at 20 dB single link SNR
A.6.4 Conclusion
A generalized method for joint design of the linear transceivers with coordinated multi-cell processing
subject to per-BS power constraints was proposed for two different optimization objectives, i.e., the
weighted SINR balancing and the weighted sum rate maximization. The generalized method can
accommodate a variety of scenarios from coherent multi-cell beamforming across a large virtual MIMO
channel to a single-cell beamforming with inter-cell interference coordination and beam allocation. The
performance of SINR balancing and sum rate maximization algorithms with different multi-cell
transmission methods (coherent/non-coherent and coordinated/non-coordinated) with optimal and sub-
optimal beam allocation algorithms was numerically evaluated. The numerical results showed that the
WINNER+ D1,8
Version: 1.0 Page 75 (75)
sum rate maximization algorithm largely outperforms the SINR balancing in terms of ergodic sum rate in
all simulated cases. This is obviously due to the fact that the sum rate maximization algorithm ends up
assigning zero powers to some beams and hence inherently avoids the spatial overload. The numerical
examples demonstrated that the coherent multi-cell beamforming greatly outperforms the non-coherent
cases especially at the cell edge. However, it was shown that the coordinated single-cell beamforming
with interference avoidance and dynamic beam allocation has a good performance-complexity trade-off.
In particular, the coordinated single-cell beamforming with sub-optimal, but simple, beam allocation
algorithm, i.e., maximum eigenvalue selection, performed relatively well with reduced complexity
compared to the optimal beam allocation scheme. Furthermore, significant performance gains are
available when using multiple receive antennas.