ADAPTIVE ANTENNAS FOR CDMA WIRELESS NETWORKS A DISSERTATION SUBMITTED TO THE DEPARTMENT OF ELECTRICAL ENGINEERING AND THE COMMITTEE ON GRADUATE STUDIES OF STANFORD UNIVERSITY IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY By Ayman F. Naguib August 1996
195
Embed
Thesis Adaptive Antennas For CDMA WIRELESS Networks
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
ADAPTIVE ANTENNAS FOR CDMA WIRELESSNETWORKS
A DISSERTATION
SUBMITTED TO THE DEPARTMENT OF ELECTRICAL ENGINEERING
AND THE COMMITTEE ON GRADUATE STUDIES
OF STANFORD UNIVERSITY
IN PARTIAL FULFILLMENT OF THE REQUIREMENTS
FOR THE DEGREE OF
DOCTOR OF PHILOSOPHY
By
Ayman F. Naguib
August 1996
c Copyright by Ayman F. Naguib 1996
All Rights Reserved
ii
I certify that I have read this thesis and that in my opinion it
is fully adequate, in scope and in quality, as a dissertation for
the degree of Doctor of Philosophy.
Arogyaswami Paulraj(Principal Adviser)
I certify that I have read this thesis and that in my opinion it
is fully adequate, in scope and in quality, as a dissertation for
the degree of Doctor of Philosophy.
Thomas Kailath
I certify that I have read this thesis and that in my opinion it
is fully adequate, in scope and in quality, as a dissertation for
the degree of Doctor of Philosophy.
Donald Cox
Approved for the University Committee on Graduate Stud-
ies:
iii
Abstract
Wireless cellular communication is witnessing a rapid growth in markets, technology, and
range of services. A major current thrust for cellular communication systems is improved
economics through enhanced coverage early in the life cycle of a network and high spectrum
efficiency later in the life cycle. An attractive approach for economical, spectrally efficient,
and high quality digital cellular and personal communication services (PCS) is the use of
7.4 Power-controlled received signal vs. simulated Rayleigh fading: fd =100 Hz, K = 5, L = 4. . . . . . . . . . . . . . . . . . . . . . . . . . 147
7.5 Power control error vs. number of paths: K = 5. . . . . . . . . . . . . 148
7.6 Power control error vs. power step size: L = 2,K = 5. . . . . . . . . . 149
7.7 Power control error vs. loop delay: L = 2,K = 5. . . . . . . . . . . . 150
7.8 Power control error vs. forward link error rate: L = 2, K = 5. . . . . . 151
7.9 Power control error vs. angle spread: L = 4, and K = 5. . . . . . . . . 152
7.10 Power control error vs. angle spread: L = 4, and K = 9. . . . . . . . . 153
xv
Chapter 1
Introduction
The realization of wireless communications providing high-speed and high-quality informa-
tion exchange between two portable terminals that might be located anywhere in the world
is the new communications challenge for the next decade. The great popularity of cordless
phones, cellular phones, radio paging, and other emerging portable communication tech-
nologies demonstrates a great demand for such services. For example, from 1990 to 1994,
the number of cellular telephone users in the US has grown from 5.1 million to 23.3 million
subscribers, and by the year 2000 it is projected that the number will nearly double to 46.9
million subscribers.
What has emerged from the world wide research and development activity in this area
is the need for new technology advances to meet such an explosive growth. These include:
• New techniques to improve the quality and spectral efficiency of communication over
wireless channels.
• Better techniques for sharing the limited spectrum to accommodate different wireless
services.
• New signal processing techniques to implement various functions of the cellular sys-
tem.
1
Chapter 1. Introduction 2
Figure 1.1: If I had a cellular telephone ... (Herman c Jim Unger [1]. Reprinted with per-mission of Universal Press Syndicate. All rights reserved)
The physical limitations of the wireless channel present a fundamental technical chal-
lenge for reliable communications. The channel is susceptible to time-variant noise, inter-
ference, and multipath. Moreover, radio spectrum is now a limited resource, and even with
the recent increase in the spectrum allocation for wireless services, it will be stretched out
to its capacity limit to accommodate various current and emerging wireless services.
Also, limitation in the power and size of the communication and computing device is
another major design consideration. Most personal communications and wireless services
devices are meant to be carried in a briefcase or a pocket and, therefore, must be small and
lightweight, which translates to a low power requirement, since small batteries must be used.
Chapter 1. Introduction 3
However, many of the signal processing techniques used for efficient spectral utilization
demand significant processing power, precluding the use of low power devices. Continuing
advances in VLSI and integrated circuit technology for low power applications will provide
a partial solution to this problem. However, placing most of the signal processing burden
on fixed location sites with large power resources has and will continue to be the common
trend in wireless systems design.
Cellular communication systems involve two radio links: the reverse link from the mo-
bile to the base station, and the forward link from the base station to the mobile. In this the-
sis, we study the use of multiple antennas at the base station and the associated advanced
signal processing for CDMA wireless systems and the effect of such processing on system
performance measures such as capacity, range, and mobile transmit power. In this work we
will focus on the reverse (mobile to base) link.
1.1 Cellular Systems and Standards
In order to accommodate the demand for wireless communication services, efficient use
of limited available frequency spectrum is essential. Cellular systems exploit the power fall-
off with distance of a transmitted signal to reuse a communication channel (where a com-
munication channel can be either a frequency band, a time slot, or a unique code) at another
spatially separated location. The coverage area is divided into smaller regions or cells, each
containing a subset of the mobile users of the cellular system. In each cell, only one user is
assigned to a particular communication channel. Operation within a cell is controlled by a
base station, which is responsible for serving calls to and from users located in their respec-
tive cells. The base stations are connected to the mobile telephone switching office (MTSO)
that serves as a controller to a group of base stations and as an interface between the mobile
users and the fixed public switching telephone network (PSTN). When a mobile user crosses
Chapter 1. Introduction 4
µT1/ Wave
Base Station
Mobile
MTSO
PSTN
Co
Figure 1.2: Overview of a cellular telephone network
the boundary between two cells, its communication channel is switched, or handed off, to
the base station in the new cell. The shape of each cell is determined by the power footprint
of the transmitting base station. An illustration of a cellular telephone network is shown in
Figure 1.2.
In the US, cellular telephone service is provided mostly by the AMPS (Advanced Mobile
Phone Service) system [2-4] which evolved from extensive research at Bell Laboratories in
the 1970’s. In this system analog voice signals are frequency modulated onto carriers in the
800 MHz band (824-849 for the downlink and 869-894 for uplink). To prevent significant
adjacent-channel interference, each user channel is allotted a 30 kHz bandwidth. Only a
certain set of the carrier frequencies is available in a given cell, and neighboring cells must
use different sets of carriers. This method of frequency reuse results in reduced cochannel
interference, but the resulting system capacity is only one call per 210 kHz [4]. The increas-
ing demand for cellular telephone service has pushed this technology to its capacity limits
Chapter 1. Introduction 5
1 2 3 4 5 6 7 8
Frame = 4.6155 ms
Data
26F
1
F
1Train
26
Data
26
T3
Guard8.25
T3
Slot = 577 sµ
T: Tail bitsF: FlagTrain: Equalizer training sequenceGuard: Guard time interval
Figure 1.3: GSM frame and time-slot structure
and resulted in proposals for a ”second generation” of cellular systems, employing digital
modulation techniques and alternative multiple access techniques. A brief description of
several of these techniques follows.
1.1.1 European Global System for Mobile (GSM)
The Pan-European standard for digital cellular telephony, called Global System for Mobile
Communications (GSM) has two objectives: pan European roaming , which offers compat-
ibility through the European continent, and interaction with the integrated service digital
network (ISDN). The first commercial GSM system, called D2, was implemented in Ger-
many in 1992.
Chapter 1. Introduction 6
The bands allocated to GSM are 890-915 MHz for downlink and 935-960 MHz for up-
link. The band in either direction is divided into 124 frequency channels, each with car-
riers spaced 200 kHz apart. Each cell site in a GSM system has a fixed number of fre-
quency channels (two way) ranging from only one to usually not more than 15. In GSM,
an FDMA/TDMA (frequency division multiple access/time division multiple access) radio
channel structure is used. Each FDMA frequency channel supports a multiple user signal
format. Each 200 kHz FDMA channel uses an aggregate bit rate of 270.833 kbits/sec, car-
ried over the radio interface using GMSK (Gaussian-filtered minimum-shift keying) mod-
ulation with a bandwidth-time product of 0.3. The use of a transmitted bit rate as high as
270 kbits/sec requires the implementation of adaptive equalization techniques to deal with
channel multipath, and GSM specifications require that equipment be built to accommodate
RMS delay spread up to 16 µs. The GSM standard provides for the use of slow frequency-
hopping as a means of reducing other user interference, though no GSM systems currently
implements this technique.
The 270-kbits/sec data stream in each FDMA channel is divided into 8 fixed-assignment
TDMA channels or time slots termed logical channels. Each slot is 577 µs, which cor-
responds to the transmission time for about 156.25 bits, though only 148 bits are actually
transmitted in each slot. The remaining time, 8.25 bits time duration or about 30.5 µs, is
guard time in which no signal is transmitted to prevent overlapping of signal bursts arriv-
ing at a base station from different mobile terminals. Figure 1.3 shows the GSM frame and
time-slot structure. The logical channels are organized into a hierarchical frame structure
that provides each mobile terminal with a two-way traffic channel and a separate two-way
control channel. The numbering of time slots is offset between two directions on the down-
link to prevent a mobile terminal from transmitting and receiving at the same time. More
details on the GSM frame hierarchy and network issues can be found in [4-7].
Chapter 1. Introduction 7
1 2 3 4 5 6
Frame = 40 ms
G6
R6
Data16
Synch28
Data122
CCH12
DVCC12
Data122
Mobile to Base
Slot = 6.67 ms
Synch28
CCH12
Data130
DVCC12
Data130
RSVD12
Base to Mobile
G: Guard timeR: Ramp timeSynch: Synchronization
DVCC: Digital verification color codeRSVD: Reserved for future useCCH: control channel
Figure 1.4: IS-54 frame and time-slot structure
1.1.2 North American TDMA Digital Cellular (IS-54)
The North American TDMA Digital Cellular standard (IS-54) has been developed by
the digital cellular standards subcommittee as a replacement for the existing analog AMPS
system. As with GSM, the IS-54 radio channel structure is a combination of FDMA and
TDMA, with user traffic and control channels built upon the logical channels provided by
TDMA time slots. The designated frequency channels in IS-54 are the same as those in
Chapter 1. Introduction 8
AMPS, with carriers spaced 30 kHz apart. This was done so as to allow conversion of in-
dividual analog channels to digital operation. In the initial phase of TDMA implementa-
tion, the mobile phones are dual-mode devices, capable of operating on either AMPS analog
channels or TDMA digital channels, and cellular operators are required to continue support-
ing AMPS users as digital service is introduced.
The IS-54 standard specifies the radio channel modulation as π/4-shift DQPSK (dif-
ferentially encoded phase shift keying), to be implemented with square-root raised cosine
filtering and rolloff parameter 0.35. The principal advantage of this modulation scheme is
bandwidth efficiency. The channel transmission rate is 48.6 kbits/sec; and with a channel
spacing of 30 kHz, this yields channel utilization of 1.62 bits/sec/Hz, a 20% improvement
over GSM. The principal disadvantage of a linear modulation scheme is a lower efficiency
of linear transmitter power amplifiers as compared to nonlinear amplifiers that can be used
with constant envelope modulation. This is reflected in the size and weight of the hand set.
In IS-54, each 30-kHz digital channel has a transmission rate of 48.6 kbits/sec. The 48.6-
kbits/sec stream is divided into 6 TDMA channels of 8.1 kbits/sec each. The IS-54 time slot
and frame format, as shown in Figure 1.4, is much simpler than that of the GSM standard.
The 40 ms frame is composed of six 6.67-ms time slots corresponding to 324 bits each.
Although needed in only a few specific places for 48.8 kbits/sec transmission rate, a delay
spread equalizer is required in the IS-54 standard.
1.1.3 North American CDMA Digital Cellular (IS-95)
As an alternative to IS-54’s TDMA standard for digital cellular, a new digital cellular system
based on code-division multiple access (CDMA) technology was proposed [8]. In contrast
to the IS-54 standard, which uses the same set of carrier frequencies, spaced 30 kHz apart, as
are used by the analog AMPS system, the CDMA system uses spread spectrum signals with
1.2288 MHz spreading bandwidth, a frequency span equivalent to 41 AMPS channels. (The
Chapter 1. Introduction 9
forward and reverse links actually use separate carrier frequencies, spaced 45 MHz apart.)
Clearly, this design does not lend itself to selective channel replacement of analog AMPS,
unlike the TDMA IS-54 system. Instead, large blocks of channels will be replaced at one
time by the CDMA system [5]. Therefore, the IS-95 standard specifies a dual-mode design
of the mobile station so that the CDMA digital cellular units can be still used in areas with
existing analog cellular coverage.
As one of the goals of this thesis is to apply advanced array signal processing techniques
derived therein to CDMA systems, a full description of the IS-95 CDMA system will be
given later in Chapter 5.
1.2 Why Antenna Arrays?
As mentioned earlier, the goal of wireless communication systems is to provide a wide va-
riety of wireless services such as voice, data, facsimile, and electronic mail. The current
design trend for those systems is to deploy a large number of small cells, each served by a
base station using a conventional antenna system. While this might provide a solution for
the coverage problem, it has a number of significant drawbacks. A large number of small
cells will result in increased number of base station equipment, increased networking and
coordination requirements, increased handoff, and reduced trunking efficiency.
Optimum combining (or beamforming) and adaptive signal processing with antenna ar-
rays at the base station can be used to improve system coverage, base station capacity, and
link quality for moderate to large-sized cells. With antenna arrays, the base station can re-
duce or suppress the interference due to cochannel users (either from adjacent cells or same
cell), and through spatial diversity and beamforming gain, it provides improved desired sig-
nal level. This improvement that antenna arrays bring can be traded off in a number of ways.
Chapter 1. Introduction 10
For example, when a network is first deployed, capacity is not an issue (especially in ru-
ral areas). As the system matures, more and more users will request wireless services and
the need for more capacity develops. In this later stage, the gain due to spatial processing
with antenna arrays can facilitate a denser use of the available bandwidth and, therefore,
an increase in system capacity. In the early stages of system deployment, service providers
strive to minimize the initial deployment cost by covering a given service area with the least
number of base stations. The gain provided by spatial processing with antenna arrays can
be used to help meet this coverage goal. This gain can also be used to improve the signal
quality itself, which leads to low outage probability, better voice quality, and the ability to
support higher data rates for wireless data services. Other possible benefits of spatial pro-
cessing with antenna arrays include the ability to reduce the loss in performance when low
power terminals co-exist with high power terminals.
1.3 Thesis Overview
In this introductory chapter, we have attempted to lay the background for the subject mate-
rial of this thesis. The first half of Chapter 2 is an overview of radio channel modeling that
describes some problems associated with a wireless radio channel. The models described
therein are for scalar channels. To support the core subject of this thesis, which is the use of
adaptive antenna arrays at the base station, in the second half of Chapter 2 we develop a sta-
tistical vector channel model based on the physical propagation environment and its statis-
tical properties. In Chapter 3, we start by reviewing previous techniques and the conditions
under which they are applicable for optimum beamforming. We then articulate the reasons
why those techniques are not applicable for wireless CDMA systems. We then introduce
our space-time approach for estimating the channel vector and for optimum beamforming,
and construct a space-time receiver model which we call a ”Beamformer-RAKE” to exploit
Chapter 1. Introduction 11
both the temporal and spatial structure of the received multipath signal to maximize perfor-
mance. In Chapter 4, we develop a recursive algorithm for estimating the vector channel and
the optimum beamforming weights. We then study its tracking performance under realistic
channel conditions.
Chapters 5 and 6 describe an application of the above space-time processing techniques
to the current CDMA standard IS-95. Chapter 5 gives a brief description of the proposed
IS-95 system, with more emphasis on describing the reverse link modem. In Chapter 6, we
present an overall base station receiver structure with antenna arrays and describe its build-
ing blocks. We then study the performance of the proposed receiver in terms of the uncoded
bit error rate (BER) as a function of loading (number of users) under different channel con-
ditions. In our analysis, we assume perfect channel estimates and do not include tracking
errors. As we mentioned earlier, closed-loop power control is a key element of the current
implementation of the IS-95 system. In Chapter 7, we describe a method for closed-loop
control. We also present a simulation study to characterize the performance of closed-loop
power control under different operating scenarios. Chapter 8 contains concluding remarks
and a summary of the thesis. In Appendix A, we derive the fading correlation for both scalar
and vector multipath channels. In Appendix B, we derive the probability distributions used
in the analysis in Chapter 6.
Several of the main ideas in this thesis have already been published by the author in the
course of his research. The work on the vector channel model has been partially published
in [9]. The idea of space-time processing and ”Beamformer-RAKE” is presented in [10, 11].
The performance enhancement in CDMA systems due to the use of antenna arrays has been
published in [12-14]. The recursive algorithm for estimating and tracking the channel vector
is presented in [15, 16]. The proposed base station receiver architecture and its performance
analysis are presented in [17-19]. Results on the power control performance are presented
in [20].
Chapter 1. Introduction 12
1.4 Thesis Contributions
We have made several contributions in this thesis research to the body of communication
systems engineering. These are
• The development of a statistical vector channel model based on the physical propa-
gation environment for wireless channels when multiple antennas are used at the base
station (this work is a joint contribution with Gregory Raleigh and Suhas Diggavi);
• A new space-time processing technique for estimating the vector channel and for op-
timum beamforming in CDMA systems;
• The derivation of the space-time matched filter receiver for vector multipath signals
over AWGN channels;
• The construction of a ”Beamformer-RAKE” receiver model for exploiting both spa-
tial and temporal structure in the received CDMA multipath signal to maximize sys-
tem performance;
• A recursive algorithm for estimating the vector channel and a quantification of its
tracking performance under different channel conditions;
• An antenna array-based base station receiver architecture for the current CDMA wire-
less standard based on the ”Beamformer RAKE” structure and performance analysis
of the proposed receiver in terms of the uncoded BER under different channel condi-
tions; and
• A characterization of the performance of fixed-step closed-loop power control under
different operating scenarios.
Chapter 2
Description and Modeling of Wireless
Channels
The wireless channel in mobile radio poses a great challenge as a medium for reliable high
speed communications. When a radio signal is transmitted in a wireless channel, the wave
propagates through a physical medium and interacts with physical objects and structures
therein, such as buildings, hills, streets, trees, and moving vehicles. The collection of ob-
jects in any given physical region describes the propagation environment. The propagation
of radio waves through this environment is a complicated process that involves diffraction,
refraction, and multiple reflections. In order to analyze the performance of wireless com-
munication systems, it is necessary to define statistical models that reasonably approximate
the propagation environment. Many statistical models for scalar (single antenna) channels
have been reported. In order to analyze performance of recently proposed adaptive antenna
array techniques, it is also necessary to develop statistical channel models for the vector
channel (multiple antennas) case.
In this chapter, we characterize the wireless propagation environment. We also develop
a statistical and time-variant wireless vector channel model that is based on the physical
propagation environment.
13
Chapter 2. Description and Modeling of Wireless Channels 14
2.1 Fundamentals of Radio Propagation
A typical channel model in mobile radio communication in an urban area usually involves
an elevated fixed base-station antenna (or multiple antennas), a line-of-sight (LOS) propaga-
tion path followed by many non-LOS reflected propagation paths, and an antenna mounted
on the mobile or the portable unit. In most situations, because of natural and man-made
structures located between the mobile and the base-station in the propagation environment, a
direct LOS propagation path may not exist. An illustration of such an environment is shown
in Figure 2.1.
The mobile radiation pattern illuminates all the local structures and buildings surround-
ing the mobile that are within a few hundred wavelengths from the mobile. In addition to
the local scatterers around the mobile, there are dominant reflectors such as large buildings,
hills, towers, and other structures. These dominant reflectors couple energy from the mo-
bile and/or the local scattering structure to the base station which gives rise to a propagation
path between the base-station and the mobile. In many instances there may exist more than
one propagation path, and this situation is referred to as multipath propagation. The propa-
gation path or paths change with the movement of the mobile unit and/or the movement of
its surroundings in the propagation environment.
Even the smallest and slowest movement causes time-variant multipath , resulting in a
random time-varying signal at the base-station. As an example, assume a stationary user
near a busy highway. Although the mobile user is stationary , parts of the environment are
moving at 60 mph (miles per hour). The vehicles on the highway become ”moving reflec-
tors” for the radio signal. The movement of the mobile user and/or the movement of the sur-
roundings in the propagation environment causes a Doppler spread in the received signal,
which describes the rate of variation in the received signal level. In addition, each propa-
gation path arriving at the base-station has its own time delay and angle of arrival (AOA),
which causes delay spread and angle spread in the received signal.
Chapter 2. Description and Modeling of Wireless Channels 15
Dominant Reflector
Dominant Reflector
MobileLocal Scatterers
Base Station
Figure 2.1: Illustration of the wireless propagation environment.
In general, the mobile radio propagation in such environments is characterized by three
partially separable effects: path loss, multipath fading, and shadowing. Path loss is a func-
tion of the distance between the mobile and the base-station. As discussed earlier, multiple
signal reflections arrive at the base-station each with its own phase, which causes destruc-
tive and constructive interference. The resulting variations in the signal amplitude, called
multipath fading, vary over distances proportional to the signal wavelength; thus, this type
of signal fading is referred to as fast fading. When the number of multipath components is
large, the central limit theorem can be invoked to model the fast fading by a filtered complex
Gaussian process. Multipath fast fading is described by its envelope fading (flat fading),
Doppler spread (time-selective fading), time-delay spread (frequency selective fading), and
angle spread (space selective fading).
In addition to the multipath fast fading, the LOS and reflected paths may be attenuated
by large obstructions such as large buildings and hills that are positioned between the mobile
Chapter 2. Description and Modeling of Wireless Channels 16
hr
ht
θθ
d
d
`
r
r
Figure 2.2: Direct and indirect paths on a flat-terrain environment.
and base station. This type of fading, which is called shadowing, varies with distances that
are proportional to the sizes of the buildings (or the obstructions in general), and is thus
referred to as slow fading. Most empirical studies show that the variations in signal level
due to the slow fading follow a log-normal distribution.
2.1.1 Propagation Path Loss
A measure of interest in radio propagation is the path loss, which is defined as the ratio
between the received power Pr and the transmitted power Pt
Ld = Pr
Pt(2.1.1)
Consider the propagation of a radio wave in a flat-terrain environment. The transmitted sig-
nal may reach the receiving antenna in several ways:
• Through a direct LOS path.
• Through an indirect path consisting of the radio wave reflected by the ground.
• Through an indirect path consisting of a surface wave.
Chapter 2. Description and Modeling of Wireless Channels 17
Let τ be the time delay between the direct LOS path and the reflected path. If we assume
that τ is much smaller than the inverse bandwidth of the transmitted signal B−1 and if we
neglect the effect of surface wave attenuation1, then the received power Pr is [21]
Pr = PtGtGr
(λ
2πd
)2 ∣∣1 + Re j�φ∣∣2 (2.1.2)
where Gt and Gr are the gains of the transmitting and receiving antennas, respectively (note
that we assumed that GtGr is the same for both the direct LOS path and the reflected path),
λ is the wavelength, R is the reflection coefficient of the ground, and �φ is the phase dif-
ference between the direct LOS path and the reflected path
�φ = 2π(r + r − d)λ
= 2πλ
[(ht + hr
d
)2
+ 1
]1/2
−[(
ht − hr
d
)2
+ 1
]1/2 (2.1.3)
where ht and hr are the heights of the transmitting and receiving antennas, respectively.
Equation (2.1.2) has been shown to agree very closely with the measurements in [22]. The
ground reflection coefficient R is given by
R = sin θ− Zsin θ+ Z
(2.1.4)
where
Z =
√εr − cos2 θ/εr for vertical polarization
√εr − cos2 θ for horizontal polarization
(2.1.5)
and εr is the dielectric constant of the ground, which for earth or road surfaces is approxi-
mately that of a pure dielectric (εr = 15). For large d, we have
�φ ≈ 4πhthr
λd(2.1.6)
Also, in this case the grazing angle θ ≈ 0◦ , and therefore R ≈ −1. Moreover, R tends to
−1 for frequencies above 100 MHz and incidence angles less than 10◦, irrespective of the
1This is a valid approximation for antennas located more than a few wavelengths above the ground.
Chapter 2. Description and Modeling of Wireless Channels 18
polarization [23]. In this case (2.1.2) becomes
Pr ≈ PtGtGr
(hthr
d2
)2
(2.1.7)
Thus, in the asymptotic limit of large d, the received power falls off inversely with d4. In
[22], plots of (2.1.2) as a function of distance illustrate this asymptotic behavior. Up to a
certain critical distance dc, the wave experiences constructive and destructive interference
of the two rays. At distance dc, the final maxima is reached, after which the signal power
falls off inversely with d4. If we average out the local maxima and minima, the resulting
average power loss can be approximated by dividing the power loss curve into two regions.
For d < dc, the average power fall off with distance corresponds to free space (i.e. d2). For
d > dc, the average power fall off with distance is approximated by the fourth power law in
(2.1.7).
In practice, the attenuation of radio signals is greater than that of free space and less
than that of free space with a perfectly reflecting ground. Measurements [24] show that for
frequencies around 100 MHz, the attenuation versus distance curve has a slope comparable
to that of free space with additional loss depending on the environment. For example for a
suburban environment around 100 MHz, the attenuation curve follows free space with an
additional 8.5 dB loss. In [25], it was found that at a carrier frequency of 910 MHz in an
urban environment, the attenuation loss in dB has a steeper slope than that of free space
but not as steep as with a fourth power law. Measurements at 800 MHz [26, 27] also show
two distinctly different propagation loss slopes before and after the critical distance dc. For
d < dc, the slope is slightly less than 2 while for d > dc the slope is close to 4.
The two-ray model described above is a simple model that characterizes signal propaga-
tion in isolated areas with few reflectors such as rural roads or highways. It requires infor-
mation only about antenna heights. A more complicated model for urban area transmissions
is the dielectric canyon or ten-rays model developed by Amitay [28]. This model assumes
rectilinear streets (such as in downtown Manhattan) with buildings along both sides of the
Chapter 2. Description and Modeling of Wireless Channels 19
street and transmitter and receiver well below the tops of the buildings. The building-lined
streets act as a dielectric canyon to the propagating signal.
The power fall off in both the dielectric canyon model and urban measurements [22, 29,
30] is proportional to d−2, even at relatively large distances. Moreover, the fall off exponent
is relatively insensitive to the transmitter height, as long as the transmitter is significantly
below the building skyline. This fall off with distance squared is due to the dominance of
the wall-reflected rays, which decay as d−2, over the combination of the LOS and ground-
reflected rays (two-path model above), which decays as d−4.
2.1.2 Multipath Fast Fading
The statistical model for the fast fading of the received signal level is based on a physical
propagation environment consisting of a large number of isolated scatterers with unknown
locations and reflection properties. Let the transmitted signal be
x(t) = s(t) · e j(2π f t+φo) (2.1.8)
where s(t) is the complex baseband signal with bandwidth B, f is the carrier frequency, and
φo is an arbitrary initial phase. Without loss of generality we will assume that φo is zero. If
we assume that the mobile is moving at speed v and there is no direct LOS, and if we ignore
the receiver additive white Gaussian (AWGN) noise, the corresponding received signal at
the base station is the sum of all multipath components [21, 31]
y(t) = AL∑
i=1
Ris(t − τi)ej2π[( f+ fd cosψi)t− f τi] (2.1.9)
where A includes the effects of distance loss and antenna gains (here we assumed that the
propagation distance spread�r = maxi
ri −mini
ri is much less than the propagation distance
ri for all i and, therefore, that the attenuation with distance is the same for each component.
This is true in general if all reflectors are within the vicinity of the mobile). For the ith mul-
tipath component, R2i is the fraction of the incoming power in the ith path, τi = ri/c is the
Chapter 2. Description and Modeling of Wireless Channels 20
multipath delay where ri is the propagation distance and c is the speed of light, and fd cosψi
is the Doppler shift where fd = v/λ is the maximum Doppler shift and ψi is the direction
of the ith scatterer with respect to the mobile velocity vector. The Doppler spread fm is
given by 2 fd. These parameters vary with time. As noted in [21], the process y(t) is wide-
sense stationary with respect to ensemble averages. It is not stationary with respect to time
averages, however, and thus nonergodic. But the difference between time and ensemble av-
erages decreases as the number of paths L becomes large; thus the statistical properties will
be computed on the basis of ensemble averages. Throughout the rest of this section, we will
assume that A = 1. If we assume that the multipath delay spread defined as
T = maxi
τi − mini
τi (2.1.10)
is much less than the inverse bandwidth of the signal (T � B−1), i.e. we assume that s(t)
is a narrowband signal, then s(t − τi) ≈ s(t − τo) where τo ∈ [mini
τi,maxi
τi] [31]. Then,
we can rewrite (2.1.9) as
y(t) ≈ s(t − τo) ·(
L∑i=1
Riejφi(t)
)· e j2π f t (2.1.11)
where φi(t) = 2π( fd cosψit − f τi). The phases φi(t) modulo 2π can be modeled as i.i.d.
random variables uniformly distributed over [0,2π] [32]. Analytical results based on this
assumptions agree with measurement results in [32].
The equivalent lowpass received signal is
y(t) ≈ s(t − τo) ·(
L∑i=1
Riejφi(t)
)(2.1.12)
In addition to the time delay, the received signal in (2.1.12) differs from the original trans-
mitted signal by the complex scale factor in the parentheses. Let
β(t) =L∑
i=1
Riejφi(t) = α(t)e jφ(t) (2.1.13)
Chapter 2. Description and Modeling of Wireless Channels 21
Since y(t) is the response of an equivalent lowpass channel to the lowpass signal s(t), it fol-
lows that the equivalent lowpass channel is described by the time-variant impulse response
[33]
h(t; τ) = δ(τ − τo)α(t)ejφ(t) (2.1.14)
If we assume that the Ri are i.i.d. and independent of the φi, then the first and second mo-
ments of β(t) are
E{β(t)} = 0 (2.1.15)
E{β(t)β∗(t + ν)} =∑
i
E{R2i }e j2π fd cosψiν (2.1.16)
If we also assume that, in addition to being i.i.d., the Ri have bounded variance, then β(t)
will approach a complex Gaussian random variable as the number of scatterers L becomes
large [34]. In this case α(t) has a Rayleigh distribution [33]
f (α) = 2ασ2
exp
(−α2
σ2
), α ≥ 0 (2.1.17)
where σ2 = E{α2}. Therefore the variation of the received signal envelope is Rayleigh,
which has also been confirmed by measurements [35, 36]. Now, if the direct LOS is not
obstructed, then α(t) will have a Rician distribution [33]
f (α) = 2ασ2
exp
(−α2 +µ2
σ2
)Io
(2αµσ2
), α ≥ 0 (2.1.18)
where µ2 is the average power in the direct LOS and In(.) is the modified Bessel function
of the n-th order [37].
Note that the real and imaginary parts of β(t) are independent,
E{Re{β(t)}Im{β(t)}} =L∑
i=1
∫ 2π
0
12π
cos(φi(t)) sin(φi(t)) dφi(t)
= 0 (2.1.19)
Note: A more general distribution for the fast fading amplitude is given by the Nakagami
distribution [32]
f (α) = 2mmα2m−1
�(m)σ2· exp
(−mα2
σ2
)(2.1.20)
Chapter 2. Description and Modeling of Wireless Channels 22
where σ2 = E{α2} and m = σ4/E{(α2 − σ2)2} and �(.) is the gamma function. This distri-
bution is general in the sense that none of the above assumptions has to hold. When the scat-
tering process generates merely diffuse wave field, then m ≈ 1 [32] and the Nakagami dis-
tribution in (2.1.20) is identical to the Rayleigh distribution in (2.1.17). When a direct com-
ponent is present, the Nakagami distribution approximates the Rice distribution in (2.1.18)
with m > 1.
It remains now to determine the time-frequency correlation behavior of β(ω, t), where
β(ω, t) is given in (2.1.13). In order to do that, let fT(τ) be the probability density function
of the time delay τi, where fT(τ) is nonzero for 0 ≤ τ < ∞ and zero otherwise. In Ap-
pendix A, it is shown that the time-frequency correlation of β(ω, t) and β(ω+�ω, t + ν)
is given by
ρβ(�ω, ν) = E{β(ω, t)β∗(ω+�ω, t + ν)} (2.1.21)
= Jo(ωdν) · FT( j�ω) (2.1.22)
where Jn(.) is the Bessel funcion of the first kind of order n [37] and FT( j�ω) is the Fourier
transform of fT(τ). Interpretation of some measured data [21] indicates that the path time
delay τi can be modeled as an exponential random variable. That is
fT(τ) = 1
Texp
(− τ
T
), τ ≥ 0 (2.1.23)
where T is mean time delay. In this case, we have
FT( j�ω) = 1 − j�ωT
1 + (�ωT )2(2.1.24)
and
ρβ(�ω, ν) = Jo(ωdν) · 1 − j�ωT
1 + (�ωT )2(2.1.25)
To a good approximation [21, 23, 38], the envelope correlation ρα(�ω, ν) is equal to
the squared magnitude of the complex signal correlation, i.e..
ρα(�ω, ν) = Cov{α(ω, t)α∗(ω+�ω, t + ν)}√Var{α(ω, t)α∗}Var{(ω+�ω, t + ν)} (2.1.26)
Chapter 2. Description and Modeling of Wireless Channels 23
Time (Symbol Period Ts)
0 200 400 600 800 1000
Fadi
ng M
agni
tude
(dB
)
-20
-15
-10
-5
0
5
10
Figure 2.3: Magnitude of a Rayleigh fading channel: fd = 80 Hz and Ts = 104.2µs.
ρα(�ω, ν) � J2o (ωdν) · |FT( j�ω)|2 (2.1.27)
Figure 2.3 shows the magnitude for a simulated multipath Rayleigh fading channel. The
mobile is assumed to be moving such that fd = 80 Hz. The transmitted signal is a balanced
QPSK signal with symbol rate of 9600 symbols/sec. Figure 2.4 shows the amplitude time-
correlation ρβ(0, ν)= Jo(2π fdν) as a function of fdν (which is the number of wavelengths
traveled by the mobile over ν sec).
As the signal bandwidth B increases so that T ≈ B−1, the approximation s(t − τi) ≈s(t − τo) for τo ∈ [min
iτi,max
iτi] is no longer valid. Then, the received signal is a sum
of copies of the original signal, each is delayed in time by τi and phase shifted by φi. For
wideband signals, the channel response can be approximated using Turin’s model [36] if the
incoming paths form subpath clusters. In this model, paths that are approximately the same
Chapter 2. Description and Modeling of Wireless Channels 24
Delay ( in wavelength traveled)
0 1 2 3 4 5
Am
plitu
de C
orre
lati
on ρ
(ν)
-0.4
-0.2
0.2
0.4
0.6
0.8
0.0
1.0
Figure 2.4: Fast fading amplitude correlation vs. delay (in wavelengths traveled).
length ( |τi − τ j| < B−1 ) are not resolvable at the receiver. Thus, they are combined into a
single path. If we assume a finite number of resolvable paths, then the received signal can
be written as
y(t) =L∑
l=1
x(t − τl )αl(t)ejφl (t) (2.1.28)
where L in this case represents the number of resolvable paths or subpath clusters, and αl,
φl, and τl are the amplitude, phase, and delay of each resolvable path. The complex gains
βl(t) = αl(t)e jφl (t) are independent complex Gaussian processes. Again, the equivalent
lowpass channel is described by the time-variant impulse response [33]
h(t; τ) =L∑
l=1
δ(τ − τl)αl(t)ejφl (t) (2.1.29)
Chapter 2. Description and Modeling of Wireless Channels 25
As the mobile moves in the environment, the number and position of scatterers contributing
to the received signal will change and, therefore, the time delay τl, the amplitude fade αl,
the phase φl, and the number of resolvable paths L will also change with time and, there-
fore, in addition to specifying their first order statistics, one also needs to specify their time
correlation properties. The time correlation properties for the complex path gain βl(t) are
also described by the correlation function in (2.1.22). As suggested by measurements [39],
the time delays τl are characterized by a Poisson or modified Poisson process. More details
on characterizing the above parameters for mobile radio and indoor wireless channels can
be found in [39-42].
2.1.3 Log-Normal Slow Fading
The signal fading described in the previous section results from out-of-phase combining of
different multipath components. Since these phases change by π degrees every half wave-
length, the signal amplitude changes rapidly over very short distances as the mobile moves
in the propagation environment (approximately every foot for 900 MHz signals). If we aver-
age out these local variations due to multipath over large distances, the local mean will also
change with distance due to two effects: the propagation path loss with distance described
earlier, and the changing size and geometry of the surrounding buildings and obstacles that
attenuate radio waves as it propagates through the environment. Measurements are usually
used to predict the power loss with distance due to shadowing [21, 43, 44]. Although these
measurement data depend on the environment in which the measurement was taken, most
empirical studies show that the mean signal level is approximately log-normal, i.e. the dB
value of the mean signal power is Gaussian. The following argument suggests why the log-
normal distribution might be expected [23].
Suppose that the ith obstruction has an attenuation constant ai and thickness �ri. If si−1
is the amplitude of the wave entering this obstruction and si is the wave amplitude after the
Chapter 2. Description and Modeling of Wireless Channels 26
obstruction, then
si = si−1 exp (−ai�ri) (2.1.30)
Then, it follows that the signal leaving the nth obstruction is given by
sn = so exp
(−
n∑i=1
ai�ri
)(2.1.31)
If we assume that ai and �ri vary randomly and independently from obstruction to obstruc-
tion, then as the number of obstructions gets large enough (n →∞) we can use the central
limit theorem to show that
x�= −
n∑i=1
ai�ri (2.1.32)
is approximately Gaussian. Therefore, S = 10 log10 s will have a Gaussian distribution with
mean µs and standard deviation σs. However, we should note that in a typical urban prop-
agation scenario, only few obstructions affect the signal propagation. The value of σs de-
pends on the environment and varies from 4 to 12 dB [25, 45-47].
The correlation behavior of the slow fading process is not known in general. However,
measurement data in [47] suggest that S(t) can be modeled as a first order Markov process
with autocorrelation
ρS(τ) = σ2s exp
(−v|τ|
Xc
)(2.1.33)
where v is the mobile velocity and Xc is the decorrelation distance which is a function of the
surrounding obstruction sizes and layout. Values of Xc under different measurement condi-
tions are reported in [47]. However, in a dense propagation environment such as New York
city where building sizes and heights vary considerably, the shadowing process is highly
non-stationary and an independent increments process model is more suitable [48].
Chapter 2. Description and Modeling of Wireless Channels 27
2.2 Vector Multipath Channels
To support the development of real time adaptive antenna techniques for wireless applica-
tions, which is the focus of this thesis, multiple antenna channel models (vector channels)
are needed to analyze any such proposed approach and predict its performance. Some vec-
tor channel models have been reported [49, 50]. However, these models are incomplete in
the sense that they do not simultaneously account for channel time variation and fading cor-
relation between antenna elements. This problem has been addressed in [9]. In this section
we will develop a statistical model for multipath wireless vector channels. This new model
will be shown to extend the existing scalar channel models in [21, 31, 36, 39, 41, 51-53]
to the vector channel case. We begin by defining the array response vector and the narrow-
band data model for array signal processing. Then, we develop the statistical channel model
based on the physical description of the propagation environment. In the remainder of this
chapter and throughout the thesis we will use xT and x∗ to denote the transpose and Hermi-
tian transpose of x, respectively.
2.2.1 Array Response Vector
Figure 2.5 shows a wireless communication system employing an antenna array where a
base station with K antennas receives signals from a mobile user. In order to introduce the
array response vector concept, we consider a simple scenario where we have a single user.
We also assume that this user can be represented by a point source and that waves arriving
at the array can be considered planar (i.e. no multipath). Once we establish a model for the
received signal for this case, the general model for multipath and multiple users case can
be simply obtained by the ”superposition” principle. In the following discussion,
we will use the so-called narrowband data model for array signal processing to obtain the
model for the received signal vector at the array. This model inherently assumes that as the
signal wavefront propagates across the array, the envelope of the signal remains essentially
Chapter 2. Description and Modeling of Wireless Channels 28
Mobile User
Antenna 1
Antenna 2
Antenna K
Base-Station
Receiver Processor
User Signal
Figure 2.5: A Wireless communication system employing antenna arrays.
unchanged. The term narrowband is used here since the assumption of a slowly varying
signal envelope is most often satisfied when either the signals or the antennas have a band-
width that is small relative to the carrier frequency f . However, this assumption can also
be satisfied by wideband signals, as is the case in CDMA, provided that the frequency re-
sponse of each antenna is approximately flat over the signal bandwidth, and provided that
the propagation time across the array is small compared to the inverse bandwidth of the sig-
nal B−1. Under these assumptions, we can write the complex base-band representation of a
real narrowband received signal at the kth antenna as [54]
xk(t) = Hk(ω)e− jωτk s(t)+ nk(t) (2.2.1)
where s(t) is the complex baseband transmitted signal, Hk(ω) is the frequency response of
the kth antenna, τk is the propagation time delay, nk(t) is the additive white Gaussian noise
(AWGN), and ω = 2π f is the carrier frequency. We can write (2.2.1) as
x(t) = v(�)s(t)+ n(t) (2.2.2)
Chapter 2. Description and Modeling of Wireless Channels 29
v(θ)
Array Manifold
Figure 2.6: A one-dimensional array manifold.
where x(t) is the array output vector and n(t) is the additive white Gaussian noise vector.
The vector v(�) ∈ C K×1 is the array response vector given by
v(�) = [H1(ω)e
− jωτ1 H2(ω)e− jωτ2 · · · HK(ω)e
− jωτK]T
(2.2.3)
and is a function of the parameter vector � which might include for example the location of
the mobile in some coordinate system, the signal carrier frequency, polarization angles, etc.
If there are p < K elements (different parameters) in v(�), then� will trace a p-dimensional
surface in C K as � is varied over the parameter space. This surface is referred to as the array
manifold and is denoted mathematically as
A = {v(�) : � ∈ Θ} (2.2.4)
where Θ denotes the set of all possible parameter vectors.
Throughout this thesis, we will assume that we have identical antennas in the array. In
this case Hk(ω) will be the same for all antennas. Also, all antennas will have the same
Chapter 2. Description and Modeling of Wireless Channels 30
response in any given direction. In this case, the array response vector is parameterized by
the angular carrier frequency ω and the time delays τ1, τ2, · · · , τk, which can be shown to be
a function of the mobile position with respect to each antenna. To simplify the discussion
we also assume that the mobile is in the far-field and that the mobile and the antenna array
are in the same plane2. Under the above assumptions, the parameter vector contains only
the azimuth θ and the array manifold A is a one-dimensional rope winding through C K as
illustrated in Figure 2.6. In this case, by redefining the signal s(t) (by denoting H(ω)s(t)
as s(t) or by including H(ω) within the channel response), then
v(θ) = [e− jωτa
1 e− jωτa2 · · ·e− jωτa
K]T
(2.2.5)
where the time delays τa1, τ
a2, · · · , τa
K are now with respect to a given reference point and are
functions of the array geometry and θ. For example, for a uniform linear array (ULA) of
identical sensors as illustrated in Figure 2.7, taking the 1st antenna as a reference point, the
time delay τak is given by
τak = (k − 1)
d sinθc
, k = 1 · · · K (2.2.6)
where θ is the angle between the arriving signal and the normal to the array, and d is the
spacing between sensors. In this case the array response vector v(θ) is given by
v(θ) =
1
e− j2πd sin θ/λ
e− j4πd sin θ/λ
...
e− j2(K−1)πd sin θ/λ
. (2.2.7)
2These assumptions may be relaxed at the expense of complicating the array response vector model. Inthe case of near-field and 3-D arrays, three parameters are required to define the parameter vector � which areazimuth, elevation and range.
Chapter 2. Description and Modeling of Wireless Channels 31
Mobile
θ
12345d
Figure 2.7: The uniform linear array scenario.
2.2.2 Vector Channel Modeling
In the vector channel model developed below, the mobile to base station link is emphasized.
As stated earlier, for simplicity, only azimuth angles are considered in the propagation ge-
ometry, but the results can be generalized to three dimensions. It is assumed that the mobile
antenna radiates uniformly in azimuth. The propagation environment under consideration
is densely populated with both natural and man-made structures. An illustration of the vec-
tor channel propagation environment is shown in Figure 2.8. The mobile radiation pattern
illuminates all local scattering structures, or local reflectors, surrounding the mobile that
are within few hundred wavelengths from the mobile. Radiation from these local reflectors
and/or the mobile reaches the base station either directly or by reflection from large reflect-
ing objects in the environment such as large buildings and hills. These objects are termed
dominant reflectors. Scalar channel models do not explicitly include the effects of these
Chapter 2. Description and Modeling of Wireless Channels 32
Local Scattering Structure
Mobile
Dominant Reflector #1
Dominant Reflector #2
Dominant Reflector #3
Base Station Antenna Array
v
∆∆
Figure 2.8: An illustration of the vector channel.
dominant reflectors. For the vector channel, the angle of arrival of each reflected wave with
respect to the base station coordinate system is determined by the physical position of the
dominant reflector with respect to the base. An illustration of the vector multipath channel
is shown in Figure 2.8. Let s(t) be the complex baseband transmitted signal. The complex
baseband received signal vector at the base station antenna array can be written as
Chapter 2. Description and Modeling of Wireless Channels 35
mnθ
dmn
r
rm
rnmnθr
dmn /2 dmn /2
Reference Point
Sensor m Sensor n
Figure 2.9: Model geometry.
where θm,n is the mean angle of arrival measured with respect to the normal to the line joining
the two sensors as shown in Figure 2.9 and zm,n and δm,n are defined in (A.2.13).For ω1 =ω2 = ω and ν = 0, we have Ra = Rs = Ra, δmn = 0, and zm,n = ωdm,n/c where dm,n is the
distance between the mth and nth sensors.
Figures 2.10 and 2.11 show the spatial envelope correlation ρs = |Rs(m, n)| for θ = 0◦
and θ = 30◦, respectively, and for various angle spreads � based on the expressions in
(2.2.18) and (2.2.19). Figure 2.10 shows that, as � decreases, the first zero in the correla-
tion occurs at a larger antenna spacing. Specifically, the first zero occurs for d/λ ≈ 30/�.
When the signal arrives from a direction other than broadside, as in Figure 2.11, the antenna
spacing for low correlation increases and the envelope correlation is never zero for all val-
ues of θ �= 0◦ and � < 180◦ ( ρs is zero when Re{Rs(m, n)} and Im{Rs(m, n)} have zero
crossings at exactly the same spacing).
The space-time array correlation matrix Ra(ω, ν) provides a full characterization of the
Chapter 2. Description and Modeling of Wireless Channels 36
Antenna Separation (d / λ)
0 1 2 3 4 5
Spat
ial E
nvel
ope
Cor
rela
tion
ρs
0 .0
0.2
0.4
0.6
0.8
1.0
1.2
∆ = 0ο
∆ = 5ο
∆ = 15ο
∆ = 60ο
θ = 0ο
Figure 2.10: Spatial envelope correlation vs. antenna spacing: mean AOA θ = 0◦
dynamics of the channel vector a(t). First, we rewrite Ra(ω, ν) as follows
Ra(ω, ν) = ρ(ν) · Rs (2.2.20)
where ρ(ν) is the time correlation part, which under the assumptions stated above, was
shown to be Jo(ωdν). As time progresses, the channel vector a(t) amplitude and direction
in C K will change. While the time correlation part of Ra describes how fast a(t) changes,
with the rate of change being proportional to ωd, the spatial correlation part Rs character-
izes the complexity of change of a(t) (in terms of the dimension of the subspace spanned
by a(t)) in C K . This complexity is proportional to the angle spread �. By (2.2.11), a(t) is
a linear combination of the path vectors v(θ1), v(θ2), · · · ,v(θL), and hence it tends to lie
in the subspace spanned by them. For example, if the angle spread is small, all path vectors
Chapter 2. Description and Modeling of Wireless Channels 37
Antenna Separation (d / λ)
0 1 2 3 4 5
Spat
ial E
nvel
ope
Cor
rela
tion
ρs
0 .0
0.2
0.4
0.6
0.8
1.0
1.2
∆ = 0ο
∆ = 5ο
∆ = 15ο
∆ = 60ο
θ = 30ο
Figure 2.11: Spatial envelope correlation vs. antenna spacing: mean AOA θ = 30◦
will point approximately to the same direction. Therefore, the fluctuations in a(t) will be
mostly in magnitude only and for a large percentage of time it will be in a certain preferred
direction. In this case Rs will have off-diagonal correlation terms that are a large fraction
of the diagonal entries. In this case, we have nonspace selective fading. For a large angle
spread, the path vectors will span the full space and therefore the a(t) does not exhibit any
preferred direction and the changes of a(t) in C K will be in both amplitude and direction.
The off-diagonal entries in Rs will become very small compared to the diagonal elements.
In this case, we have space selective fading.
Chapter 2. Description and Modeling of Wireless Channels 38
where Rss,1 is the covariance of s1(n) signal component defined as
Rss = 12Tc
E{s1(n)s∗1(n)} (3.2.25)
The goal is to find w1 that will maximize γ1(w), that is
w1 = maxw
γ1(w) = maxw
w∗Rss,1ww∗Ruu,1w
(3.2.26)
Let η1 = γ1(w1), so that
η1 = w∗1Rss,1w1
w∗1Ruu,1w1
(3.2.27)
and
w∗1(Rss,1 − ηRuu,1)w1 = 0 (3.2.28)
Chapter 3. Adaptive Beamforming with Antenna Arrays 64
It is clear from (3.2.28) that the pair (η,w1) is the solution corresponding to the largest
eigenvalue of the generalized eigenvalue problem
Rss,1w = ηRuu,1w (3.2.29)
Now, we note that
Ryy,1 = Rss,1 + Ruu,1 (3.2.30)
and the problem
(Rss,1 + Ruu,1)w = ηRuu,1w (3.2.31)
is equivalent to
Rss,1w = (η− 1)Ruu,1w (3.2.32)
Therefore, we can find the weight vector w1 directly by solving for the eigenvector corre-
sponding to the largest eigenvalue of the generalized eigenvalue problem
Ryy,1w = ηRuu,1w (3.2.33)
The proposed code-filtering approach for adaptive beamforming can be summarized as
follows
• Perform code filtering at each array element for each user in the cell.
• Given the pre- and post-correlation vectors’ samples x(1), x(2), · · ·, x(S) and y(1),
y(2), · · ·, y(S), estimate the array covariance matrices Rxx and Ryy,1, assuming that
we have enough samples (at least 2K uncorrelated snapshots [91]) to obtain quasi-
asymptotic covariance estimates.
• Estimate the channel vector a1 as the generalized eigenvector corresponding to the
largest generalized eigenvalue of the matrix pair (Ryy,1,Rxx).
• Estimate the interference-plus-noise covariance Ruu,1 according to (3.2.19).
• The optimum weight vector can be found according to (3.2.20).
Chapter 3. Adaptive Beamforming with Antenna Arrays 65
despreading filter(correlator)
channel matchedfilter
(chip rate)
channelsounder
channel impulseresponse
decisioninput
Figure 3.5: RAKE receiver block diagram
3.3 CDMA Beamforming with Multipath
In this section, we will extend the code-filtering approach for adaptive beamforming de-
scribed above to the case where we have multipath propagation. CDMA is usually applied
to multipath channels where the delay spread of the channel is significantly larger than a
symbol period. The wider the bandwidth of the spread signal, the more resolvable the in-
dividual path components are in time. By combining different components, signal strength
fluctuation due to Rayleigh fading is reduced. On such channels intersymbol interference
(ISI) is not a problem after matched filtering and despreading. In the presence of multipath
and additive white Gaussian noise, the optimal receiver for CDMA modulation is the famil-
iar RAKE receiver, originally proposed by Price and Green in [51] and studied in [92-95].
A simplified block diagram for the RAKE receiver shown in Figure 3.5. First, the input is
despread by a filter matched to the spreading sequence, creating a response that has peaks
where the channel response has peaks. Then the following transversal filter, matched to the
channel by a channel sounder, reinforces the contributions from various peaks. If nonco-
herent detection is required due to the lack of a phase reference or a pilot signal, then the
channel sounder does not give phase information, and the transversal filter can implement
a different combining law. For coherent detection, the RAKE receiver is effectively a filter
matched to the scattered (multipath) signal, at the spreading sequence (chip) rate.
Chapter 3. Adaptive Beamforming with Antenna Arrays 66
τ1 ∫
τ2 ∫
c (t)
c (t)
c (t)
τL ∫
scanningcorrelator
select delays
Combinerdecision
input
~*
~*
~*
Figure 3.6: Multipath CDMA RAKE receiver
The RAKE receiver can be simplified by implementing only a small number of taps in
the chip-rate matched filter, as shown in Figure 3.6. A scanning correlator identifies the
delays at which the most significant peaks occur, and only at those time delays is the received
signal despread and combined. The performance of this structure is optimum for the case of
discrete multipath when the number of paths equals to the number of correlators. In the case
of a diffuse multipath, this structure performs significantly worse than the RAKE receiver
in Figure 3.5, because then a significant fraction of the transmitted energy is lost unless a
large number of correlators is used [96].
Thus, in a conventional multipath receiver for CDMA, temporal code filtering is used
to resolve paths that are separated in time by more than a chip interval Tc and to estimate
the multipath structure of the channel (in terms of the time delay and the complex path
strength of each resolvable path). Then, the signal is passed through a RAKE correlator
that is matched to the channel output and that coherently/incoherently combines correla-
tors outputs. In this way the temporal structure of the multipath received signal is exploited
Chapter 3. Adaptive Beamforming with Antenna Arrays 67
efficiently. We refer to this as a 1-D or temporal RAKE since only temporal structure is
exploited.
In general, different multipath components arrive at the receiver not only with different
time delays, but also from different directions and therefore each multipath component has
a different channel vector al,1. A single antenna receiver can not exploit the spatial structure
of the received signal. With antenna arrays, the spatial dimension can be used to efficiently
resolve and combine different multipath components. A new approach that we describe be-
low is to identify both the temporal and spatial structure of the individual paths arriving at
the receiver and then construct a space-time receiver that we call a Beamformer-RAKE. This
receiver is matched to the desired signal while maximally rejecting the interfering signals.
3.3.1 Space-Time Matched Filter
Consider the output of the antenna array for a single user case and additive white Gaussian
noise:
x(t) =L∑
l=1
√Plb(t − τl)c(t − τl)e
jφl al + n(t) (3.3.1)
In order to derive the optimal space-time matched filter for the received signal, we consider
the likelihood function of the received signal, conditioned on the knowledge of all param-
eters. First, let sl(t) =√
Plb(t − τl)c(t − τl)e jφl al. The likelihood function of the received
multipath signal can be written via the Cameron-Martin formula [97] as
L({x(t);−∞ < t < ∞}) = C · exp{!(b(t))/σ2
n
}(3.3.2)
where C is an arbitrary constant and !(b(t)) is defined as
!(b(t)) = 2 · Re
{∫ ∞
−∞
L∑l=1
s∗l (t)x(t)dt
}−∫ ∞
−∞
∣∣∣∣∣L∑
l=1
sl(t)
∣∣∣∣∣2
dt (3.3.3)
The objective is to select the bits b(t) that will maximize ! and therefore will maximize
the likelihood function. Assuming that the user sends M information bits, the first integral
Chapter 3. Adaptive Beamforming with Antenna Arrays 68
yields
∫ ∞
−∞
L∑l=1
s∗l (t)x(t)dt =L∑
l=1
√Ple
− jφl
∫ ∞
−∞b(t − τl)c
∗(t − τl)a∗l x(t) dt (3.3.4)
=M∑
n=1
b(n)L∑
l=1
√Ple
− jφl zl(n) (3.3.5)
where z1(n), z2(n), · · · , zL(n) are the matched filter outputs synchronously sampled with
respect to each path signal and
zl(n) =∫ nTb+τl
(n−1)Tb+τl
c∗(t − τl)a∗l x(t) dt , l = 1, · · · , L (3.3.6)
The second integral in ! in (3.3.3) does not depend on the received signal at the array.
Therefore, the sufficient statistic for the detection of bits b(n), n = 1 · · · M, is zl(n), l =1 · · · L, obtained by a linear operation on the received signal vector x(t).
Equation (3.3.6) means that the sufficient statistic is obtained by passing the noisy re-
ceived signal vector through a spatial matched filter A∗ = [a1 a2 · · ·aL]∗, which is equiva-
lent to conventional beamforming [91], followed by a bank of L matched filters. The lth
matched filter is matched to the code waveform in the lth path c(t − τl). In general, we
can use a more selective spatial matched filter. This is important if the noise is not spatially
white and if interference is present. In this case, the lth sufficient statistic is given by
zl(n) =∫ nTb+τl
(n−1)Tb+τl
c∗(t − τl)w∗l x(t) dt , l = 1, · · · , L (3.3.7)
where the spatial matched filter, or the beamformer weight vector, wl for the lth path is cho-
sen so that w∗l x(t) is a good estimate of the signal received in the lth path. In the previous
section, we showed that wl is given by
wl = ζR−1uu,lal (3.3.8)
where Ruu,l is the covariance matrix of the sum of all of the undesired signals.
Chapter 3. Adaptive Beamforming with Antenna Arrays 69
τ
τ
τ
1
2
L
∫
∫
c (t- )
c (t- )
c (t- )
∫
~*
~*
~*
w1*
beamformer
w2*
beamformer
wL*
beamformer
spatialmatched filter
temporalmatched filter
1z (n)
2z (n)
Lz (n)
x
x
x
(t)
(t)
(t)
Figure 3.7: Space-Time Matched Filter
3.3.2 Beamformer-RAKE Receiver Structure
The outputs of the space-time matched filter shown in Figure 3.7 can be combined (coher-
ently or incoherently depending on whether a pilot signal, from which a phase reference
can be obtained, is available or not) and the output of the combiner is passed to a decision
device. This overall receiver structure is shown in Figure 3.8 and is called Beamformer-
RAKE, since the resulting structure is equivalent to a beamformer front-end followed by a
conventional temporal RAKE receiver.
Although the idea of the RAKE receiver was first proposed for single access channels
Chapter 3. Adaptive Beamforming with Antenna Arrays 70
Spatial
Beamformer RAKE
"Beamformer-RAKE"
Mobile
Temporal
Combiner
Figure 3.8: Beamformer-RAKE receiver structure
with additive white Gaussian noise, with orthogonal codes or codes with low crosscorrela-
tions they can be used in CDMA multiple access systems [8, 98]. However, in this case the
receiver structure, although simple to implement, is not optimum and other structures based
on multiuser detection have been proposed [99-102].
In order to construct the Beamformer-RAKE receiver, we use the code filtering approach
described in the previous section for each resolvable multipath component as follows. First,
we recall the total received signal vector as
x(t) =N∑
i=1
L∑l=1
√Pl,ibi(t − τl,i)ci(t − τl,i)e
jφl,i al,i + n(t) (3.3.9)
Without loss of generality, we consider the 1st user and assume that the time delays τl,1, l =1 · · · L are perfectly known. Later in this section, we will show how to use the code filtering
approach to estimate the time delays τl,1. For the nth bit, the post-correlation signal vector
for the lth multipath component is given by
yl,1(n) = 1√Tb
∫ nT+τl,1
(n−1)T+τl,1
x(t)c∗1(t − τl,1)dt (3.3.10)
yl,1(n) = 2√
Tb Pl,1b1(n)ejφl,1 al,1 + il,1 + nl,1 (3.3.11)
Chapter 3. Adaptive Beamforming with Antenna Arrays 71
where
nl,1 = 1√Tb
∫ nT+τl,1
(n−1)T+τl,1
n(t)c∗1(t − τl,1) dt (3.3.12)
is the undesired component due to thermal noise and
il,1 =L∑
k=1k �=l
√Pk,1 Il,1,k,1e
jφk,1 ak,1 +N∑
i=2
L∑k=1
√Pk,i Il,1,k,ie
jφk,i ak,i (3.3.13)
is the undesired component due to multiple access interference (MAI) plus self interference
(SI), where Il,1,k,i is defined as
Il,1,k,i = 1√Tb
∫ nT+τ1
(n−1)T+τ1
bi(t − τk,i)ci(t − τk,i)c∗1(t − τl,1) dt (3.3.14)
Let ul,1 = il,1 + nl,1 be the total undesired (all array signals other than 1st user’s lth path
signal) component. We can easily show that Rxx, Ryy,l,1, and Ruu,l,1 are given by
Rxx = 2Pl,1al,1a∗l,1 + Ruu,l,1 (3.3.15)
Ryy,l,1 = 2GPl,1al,1a∗l,1 + Ruu,l,1 (3.3.16)
Ruu,l,1 =L∑
k=1k �=l
2Pk,1ak,1a∗k,1 +
N∑i=2
L∑l=1
2Pl,ial,ia∗l,i + σ2
nI (3.3.17)
Hence, we can use the same approach developed in the previous section to find wl,1 as
wl,1 = ζR−1uu,l,1al,1 (3.3.18)
Therefore, the corresponding beamformer output, or decision statistic, for the lth path is
zl,1(n) = w∗l,1yl,1(n) = 2
√Tb Pl,1b1(n)e
jφl,1 w∗l,1al,1 + w∗
l,1ul,1 (3.3.19)
and the corresponding path signal to interference-plus-noise ratio is
γl,1 = 2GPl,1 · a∗l,1R−1
uu,l,1al,1 (3.3.20)
It remains now to show how to estimate the time delays τl,1, l = 1 · · · L. We recall that
in order to find the channel vector al,1 we solve for the principal eigenvector of the ma-
trix pencil Ryy,l,1al,1 − ηRxxal,1 and the corresponding eigenvalue ηmax is a good estimate
Chapter 3. Adaptive Beamforming with Antenna Arrays 72
for the average signal power Pl,1 in that path. Based on this observation, we can estimate
the path amplitudes and time delays by computing the maximum eigenvalue ηmax(τ) of the
matrix pencil Ryy,1(τ)al,1 − ηRxxal,1 as τ changes over the range of possible time delays
[τmin, τmax]. The values of τ that correspond to the local maxima of ηmax(τ) are good esti-
mates of the time delays τl,1, l = 1 · · · L. The value of each of those local maxima is also a
good estimate of the corresponding average path strength. In practice, each local maximum
is compared to a preset threshold to determine whether this local maximum corresponds to
a signal path or not and also to determine the number of existing signal paths L that will be
used in the subsequent detection stage. The threshold level needs to be optimized in order to
maximize system performance [94]. If the threshold level is smaller than the optimum level,
then some time delays corresponding to noise-plus-interference only will be taken as time
delays corresponding to resovable paths which will lead to performance degradation [33].
On the other hand, if the threshold level is larger than the optimum level, the performance
will also degrade as thresholding becomes more likely to reject some time delays corre-
sponding to resolvable paths as well as time delays corresponding to noise-plus-interference
only.
3.3.3 A Simulation Example
In order to evaluate the performance of the code-filtering approach for beamforming, we
considered the following simulation experiment. We considered a base station with a ULA
of 5 omni-directional antennas with antenna spacing λ/2. We assumed DQPSK spreading
with processing gain G of 128. We considered 25 mobiles randomly distributed in azimuth
around the base station with a uniform distribution over [0,2π]. We also assumed that 3
propagation paths are received from each mobile and that the total received power from
each user is the same for all users. Each path was modeled as a single planar wavefront
with average strength different for each. In this case, the pre-correlation average SINR is
Chapter 3. Adaptive Beamforming with Antenna Arrays 73
−18.7 dB. Also, the channel vector al,1 is
al,1 = αl,1
1
e− j2πd sin θl,1/λ
e− j4πd sin θl,1/λ
e− j6πd sin θl,1/λ
e− j8πd sin θl,1/λ
(3.3.21)
where αl,1 is the average path strength. Although this simulation scenario is simple and
does not fully simulate the vector channel model described in Chapter 2, it enables us to see
the ability of the approach to estimate different parameters of the multipath vector channel
accurately where other approaches would fail.
Path # 1 2 3θl,1 90◦ 45◦ 135◦
τl,1 0 4Tc 8Tc
αl,1 1 0.65 0.8
θ 89.43◦ ± 1.28 42.71◦ ± 1.88 137.88◦ ± 2.35
α 1 0.66 ± 0.029 0.81 ± 0.027
Table 3.2: Estimated multipath parameters
For the first user, for each multipath component we considered the values of the angle of
arrival θl,1, relative time delay τl,1, and average path strength αl,1 as shown in Table 3.2. The
pre and post-correlation signal vectors x and yl,1 corresponding to 10 transmitted bits were
used to estimate the array covariances Rxx and Ryy,l,1. Using Rxx and Ryy,l,1, we estimated
the channel vector al,1. From this estimate, the angle of arrival θl,1 can be estimated as
θl,1 = sin−1
(λ
2πd� al,1(2)
)(3.3.22)
where al,1(2) is the second element in al,1. The mean and standard deviation of the esti-
mated angles of arrival over 500 runs are also shown in Table 3.2. The Table also shows
Chapter 3. Adaptive Beamforming with Antenna Arrays 74
Relative delay τ (chip periodTc)
0 2 4 6 8
η max
(τ)
0 .0
0.2
0.4
0.6
0.8
1.0
1.2
mean {ηmax}
mean {ηmax} + σ
mean {ηmax} - σ
Figure 3.9: Estimated multipath profile
the mean and standard deviation of the estimated average path strength. These numbers
show that with the code-filtering approach, we are able to estimate the channel vector and
the multipath parameters accurately. Note that under this scenario with 25× 3 different ar-
riving paths, estimating different parameters for each arriving signal would be impossible if
we use any of the above mentioned DF techniques. In addition, the multipath delay profile
for the 1st user was estimated using the approach described above. The estimated profile is
shown in Figure 3.9. As we can see, the peaks of the estimated profile occur at values of τ
that correspond to the actual delays.
Chapter 3. Adaptive Beamforming with Antenna Arrays 75
τ
τ
τ
1
2
L
∫
∫
c (t- )
c (t- )
c (t- )
∫
~*
~*
~*
w1*
beamformer
w2*
beamformer
wL*
beamformer
x
x
x
(t)
(t)
(t)
decisiondevice
DPSKdemod.
DPSKdemod.
DPSKdemod.
Figure 3.10: Balanced DQPSK Beamformer-RAKE receiver with incoherent combining
3.4 ”Beamformer-RAKE” Receiver Examples
In this section we describe two ”Beamformer-RAKE” receiver examples and discuss their
average probability of error performance. The first one is for balanced DQPSK spreading
with incoherent combining. The second example is for balanced QPSK spreading with co-
herent combining.
Figure 3.10 shows a Beamformer-RAKE receiver structure for balanced DQPSK with
post detection equal gain diversity combining. After the space-time matched filter, each
signal path decision statistic zl,1(n) is fed into a DPSK demodulator. The decision variable
is the sum of all demodulator outputs. Analysis results for CDMA communication systems
employing DPSK modulation are reported in [98, 103-105]. If we assume that the multiple
access interference is Gaussian, then the bit error probability as a function of the SINR per
Chapter 3. Adaptive Beamforming with Antenna Arrays 76
bit γb is given by [33]
Pb(γb) = 122L−1
e−γb
L−1∑l=0
Cl · γ lb (3.4.1)
where for the receiver structure in Figure 3.10 γb is given by
γb =L∑
l=1
γl,1 = 2GL∑
l=1
Pl,1 · a∗l,1R−1
uu,l,1al,1 (3.4.2)
and
Cl = 1l!
L−1−l∑k=0
(2L − 1
k
)(3.4.3)
The probability of error derived above is a conditional probability since it depends on
γb. Therefore, to get the average bit error probability, we need to average this conditional
probability over the statistics of γb. First, let us assume that the average signal power per
path is the same for all paths. That is, we assume Pl,1 = P, l = 1 · · · L. If the total unde-
sired signal vector ul,1 can be modeled as a spatially white Gaussian random vector2, then
Ruu,l,1 = σ2I where σ2 is the total interference-plus-noise power. In this case we have
γb = 2GPσ2
L∑l=1
|al,1|2 (3.4.4)
In Chapter 6 we show that for zero angle spread, i.e. �= 0, γb has a χ2 probability density
function (pdf) with 2L degrees of freedom
fγb(γ) =γL−1
(γK)L(L − 1)!e−γ/γK (3.4.5)
where γ = 2GP/σ2 is the average SINR per path per antenna and K is the number of an-
tennas in the array. In this case, we can easily show that the average bit error probability Pe
is given by
Pe =∫ ∞
0Pb(γ) fγb(γ) dγ (3.4.6)
2Simulation results in Chapter 6 show that this assumption holds if the code length G is large and if thetotal number of path signals N · L is large, with a uniform angle of arrival distribution.
Chapter 3. Adaptive Beamforming with Antenna Arrays 77
Lγ, Average Antenna SINR/bit, dB
5 10 15 20 25 30
Pe
, P
roba
bilit
y of
Bit
Err
or
10-7
10-6
10-5
10-4
10-3
10-2
10-1
K =1 K=2K=4
Figure 3.11: Pe for balanced DQPSK with incoherent combining and � = 0◦
Pe =L−1∑l=0
Cl
22L−1(L − 1)!(γK)L
∫ ∞
0e−γ
(γK+1γK
)γL−1+l dγ (3.4.7)
= 122L−1(L − 1)!(1 + γK)L
L−1∑l=0
Cl · (L + l − 1)!
(γK
1 + γK
)l
(3.4.8)
Also, in Chapter 6 we show that for large angle spread γb has a χ2 pdf with 2K L degrees of
freedom
fγb(γ) =γK L−1
(γ )K L(K L − 1)!e−γ/γ (3.4.9)
Therefore, we can easily show that in this case the average bit error probability Pe is
Chapter 3. Adaptive Beamforming with Antenna Arrays 78
Lγ, Average Antenna SINR/bit, dB
5 10 15 20 25 30
Pe
, P
roba
bilit
y of
Bit
Err
or
10-10
10-9
10-8
10-7
10-6
10-5
10-4
10-3
10-2
10-1
K =1 K=2K=4
Figure 3.12: Pe for balanced DQPSK with incoherent combining and large �
given by
Pe =∫ ∞
0Pb(γ) fγb(γ) dγ (3.4.10)
=L−1∑l=0
Cl
22L−1(K L − 1)!(γ )K L
∫ ∞
0e−γ
(γ+1γ
)γK L−1+l dγ (3.4.11)
= 122L−1(K L − 1)!(1 + γ )K L
L−1∑l=0
Cl · (K L + l − 1)!
(γ
1 + γ
)l
(3.4.12)
The probability of bit error performance of the Beamformer-RAKE receiver with bal-
anced DQPSK and incoherent combining is shown in Figures 3.11 and 3.12 for angle spread
� = 0◦ and large � respectively. The probability of bit error Pe in (3.4.8) and (3.4.12) is
plotted as a function of the average antenna SINR/bit, defined as Lγ, for K = 1, 2, and 4.
Chapter 3. Adaptive Beamforming with Antenna Arrays 79
τ
τ
τ
1
2
L
∫
∫
c (t- )
c (t- )
c (t- )
∫
~*
~*
~*
w1*
beamformer
w2*
beamformer
wL*
beamformer
x
x
x
(t)
(t)
(t)
P e1-jφ
1½
P e2-jφ
2½
P eL-jφ
L½
decisiondevice
Figure 3.13: Balanced QPSK Beamformer-RAKE receiver with coherent combining
We assumed that L = 2. For the zero angle spread case in Figure 3.11, we notice the reduc-
tion in the required antenna SINR to achieve certain probability of bit error performance as
the number of antennas increases. This reduction is proportional to the number of sensors
and is due to the beamforming operation only. This is due to the fact that with zero angle
spread, the same path signal will undergo the same fading at all antennas and, therefore, the
array will not provide any space diversity. For the case of large angle spread, the signal fad-
ing at different antennas (that are sufficiently separated) will be independent. Therefore, in
addition to the beamforming gain, the array will provide a space diversity gain. The gain
due to space diversity can be seen from the Pe plot in Figure 3.12. As we can see, with 2
antennas, an additional gain of 5.5 dB due to space diversity can be obtained.
Figure 3.13 shows a Beamformer-RAKE receiver structure for balanced QPSK spread-
ing with maximal ratio diversity combining and coherent detection. After the space-time
matched filter, each path signal is weighted by the complex (conjugate) amplitude of that
path√
Pl,1e− jφl,1 . The weighted path signals are then added together and fed to a decision
Chapter 3. Adaptive Beamforming with Antenna Arrays 80
device. For coherent detection with maximal ratio combining, in addition to the beamformer
weight vector wl,1 the receiver needs also an estimate of the path amplitude√
Pl,1 and the
phase e jφl,1 . The path amplitudes can be estimated using the approach described above. The
phase information e jφl,1 is lost in the process of computing Ryy,l,1. Therefore, for coherent
detection we assume that a pilot sequence is available. In this case, the method described
in [89] can be used to estimate the phase e jφl,1 . Analysis of CDMA communication systems
with QPSK spreading are reported in [33, 89, 106]. The bit error probability as a function
of γb is [33]
Pb(γb) = 12
erfc(√
γb
)(3.4.13)
where
erfc(x) = 2√π
∫ ∞
xe−t2
dt (3.4.14)
and γb is given by (3.4.2). If we make the same assumption that the total undesired com-
ponent is white then for zero angle spread γb has the probability density function in (3.4.5)
and the average bit error probability Pe is [33]
Pe =∫ ∞
0Pb(γ) fγb(γ) dγ (3.4.15)
=∫ ∞
0
γL−1
(γK)L(L − 1)!e−γ/γK 1√
π
∫ ∞
√γ
e−x2dx dγ (3.4.16)
= 1√π
∫ ∞
0e−x2
∫ x2
0
γL−1
(γK)L(L − 1)!e−γ/γK dγ dx (3.4.17)
=(
1 −µ1
2
)L L−1∑l=0
(L − 1 + l
l
)(1 +µ1
2
)l
(3.4.18)
where µ1 =√
γK1+γK . For large angle spread, γb has the pdf in (3.4.9). In this case, the aver-
age bit error probability Pe is
Pe =∫ ∞
0Pb(γ) fγb(γ) dγ (3.4.19)
=∫ ∞
0
γK L−1
(γ )K L(K L − 1)!e−γ/γ 1√
π
∫ ∞
√γ
e−x2dx dγ (3.4.20)
Chapter 3. Adaptive Beamforming with Antenna Arrays 81
Lγ, Average Antenna SINR/bit, dB
5 10 15 20 25 30
Pe
, P
roba
bilit
y of
Bit
Err
or
10-8
10-7
10-6
10-5
10-4
10-3
10-2
10-1
K =1 K=2K=4
Figure 3.14: Pe for balanced QPSK with coherent combining and � = 0◦
Pe = 1√π
∫ ∞
0e−x2
∫ x2
0
γK L−1
(γ )K L(K L − 1)!e−γ/γ dγ dx (3.4.21)
=(
1 −µ2
2
)K L K L−1∑l=0
(K L − 1 + l
l
)(1 +µ2
2
)l
(3.4.22)
where µ2 =√
γ
1+γ.
The probability of bit error performance of the Beamformer-RAKE receiver with bal-
anced QPSK and coherent combining with L = 2 is shown in Figures 3.14 and 3.15. We also
plot Pe for both angle spread � = 0◦ and large � and as a function of the average antenna
SINR/bit Lγ. These figures also show the performance improvement due to both beam-
forming gain and space diversity gain provided by the array.
Chapter 3. Adaptive Beamforming with Antenna Arrays 82
Lγ, Average Antenna SINR/bit, dB
5 10 15 20 25 30
Pe
, P
roba
bilit
y of
Bit
Err
or
10-10
10-9
10-8
10-7
10-6
10-5
10-4
10-3
10-2
10-1
K =1 K=2K=4
Figure 3.15: Pe for balanced QPSK with coherent combining and large �
3.5 Summary
In this chapter, we reviewed different beamforming techniques and the conditions under
which they are applicable. We also discussed the reasons why those techniques are not ap-
plicable for wireless CDMA systems. We then introduced our space-time approach for es-
timating the channel vector and optimum beamforming weight vector. We derived a space-
time receiver that we call a ”Beamformer-RAKE” to exploit both temporal and spatial struc-
ture of the received multipath signal to maximize performance. In Chapter 4, we will look
at techniques to implement the beamforming approach described above.
Chapter 4
Beamforming for Time-Variant
Channels
In the previous chapter, we introduced the code-filtering approach for adaptive beamform-
ing for CDMA signals. We assumed that the channel remains constant or varies very slowly
over several symbol periods so that we have enough samples to obtain quasi-asymptotic ar-
ray covariance estimates. However, as the mobile moves in the propagation environment,
the direction of the mobile with respect to the base station will change. Also, the effective
scatterers contributing to the received signal at the base station antenna array will change.
This will be also true for other interfering users. This means that both the array covariances
and the channel vector for different signals will change with time. This motivates the need
for recursive method for computing and tracking the optimal weight vector solution that is
easy to implement and is adaptive to the time-varying nature of the channel. In this chapter,
we will introduce a method for recursively tracking the optimum weight vector.
Tracking the optimum beamforming weight vector solution obtained earlier involves
tracking the principal eigenvector of the pre- and post-correlation array covariance matrix
pair and tracking the undesired signal vector covariance. The problem of tracking eigenvec-
tors when the matrices are slowly varying with time, turns out to be central to many adaptive
83
Chapter 4. Beamforming for Time-Variant Channels 84
antenna array processing techniques and some numerical methods have already been pro-
posed [107-109]. Some methods take into account the fact that the array covariance matrix
varies by a rank one update from one time step to the next in order to speed up computa-
tional time [90, 110]. In addition to this fact, in the method that we will describe shortly, we
also make use of the inherent time correlation structure to further speed up computation.
4.1 Recursive Adaptive Beamforming
As we mentioned earlier, when the signal environment frequently changes because of the
nonstationary desired and undesired signals, the adaptive beamformer must continuously
update the weight vector to match the changing environment. The approach that we describe
below employs direct implementation of the optimum weight vector solution in (3.3.18), i.e.
estimating the unknown undesired signal array covariance Ruu,l,1 and the channel vector
at each sampling instant and implementing the optimal solution directly. This approach of
direct implementation of the optimum solution, although more computationally expensive
than other approaches, has better convergence behavior [91, 85].
We begin by recalling the optimum weight vector solution in (3.3.18)
wl,1 = ζR−1uu,l,1al,1 . (4.1.1)
We have shown that the undesired signal covariance can be estimated as
Ruu,l,1 = GG − 1
(Rxx − 1
GRyy,l,1
)(4.1.2)
and that the channel vector al,1 can be also estimated as the principal eigenvector of the
generalized eigenvalue problem
Ryy,l,1al,1 = ηRxxal,1 (4.1.3)
Chapter 4. Beamforming for Time-Variant Channels 85
The constant ζ does not affect the beamformer output SINR and, therefore, we will not con-
sider its value. Therefore, to find the optimum weight vector at each sampling instant we
need an estimate for the channel vector al,1(n) and the undesired signal covariance Ruu,l,1.
4.1.1 Recursive Estimation of the Channel Vector
We now describe a recursive procedure to estimate the channel vector ak,1. Consider the
problem defined in (4.1.3). First, let us assume that the pre-correlation signal vector covari-
ance Rxx is equal to I. In this case, power method recursion [90]
cm+1 = Ryy,l,1cm
|Ryy,l,1cm| (4.1.4)
is known to converge so that
limm→∞
|Ryy,l,1cm| = ηmax (4.1.5)
limm→∞
cm = al,1 (4.1.6)
and
Ryy,l,1al,1 = ηmaxal,1 (4.1.7)
where ηmax is the maximum eigenvalue of Ryy,l,1. Since Rxx can be any arbitrary positive
definite matrix, the idea is to decompose Rxx into the product of two matrices
Rxx = R∗/2xx R1/2
xx (4.1.8)
where R1/2xx is called the square root of the matrix Rxx. Using (4.1.3) and (4.1.4), we have
R−∗/2xx Ryy,l,1R−1/2
xx R1/2xx al,1 = ηR1/2
xx al,1 (4.1.9)
Defining the matrix Bl,1�= R−∗/2
xx Ryy,l,1R−1/2xx and the vector el,1
�= R1/2xx ak,1, we can rewrite
(4.1.9) as
Bl,1el,1 = ηel,1 (4.1.10)
Chapter 4. Beamforming for Time-Variant Channels 86
el,1(n) = POWER(Bl,1(n), el,1(n − 1))f
• Initialization:
1. co(n) = el,1(n − 1)
2. c1(n) = Bl,1(n)co(n)/|Bl,1(n)co(n)|3. m = 1
• Computation:
1. while (|Bl,1(n)cm(n)− Bl,1(n)cm−1(n)|/|Bl,1(n)cm(n)| ≥ ε) dof
(a) cm+1(n) = Bl,1(n)cm(n)/|Bl,1(n)cm(n)|(b) m → m + 1
g
2. el,1(n) = cm(n)
g
Table 4.1: Power recursion for estimating principal eigenvector
and hence the recursion defined in (4.1.4) can be used to estimate (ηmax, el,1,max). Then, the
array response vector al,1 is given by
al,1 = R−1/2xx el,1,max (4.1.11)
As we mentioned earlier, the array covariance matrices are functions of time and need to
be estimated for every symbol n from samples of the pre- and post-correlation signal vectors
x(n) and yyy,l,1(n), respectively, where yyy,l,1(n) is defined as
yl,1(n) = 1√2Ts
yl,1(n) (4.1.12)
Therefore, for every symbol the estimates Ryy,l,1(n) and Rxx(n) are used in place of Ryy,l,1
and Rxx.
Chapter 4. Beamforming for Time-Variant Channels 87
The above recursion for estimating the channel vector al,1 for each symbol can be sum-
marized as in Table 4.1. The convergence behavior of the above recursion depends on the
initial guess of the eigenvector co(n) and on the ratio |ηmax/η2|where η2 is the second largest
eigenvalue [90]. Due to the inherent time correlation of the channel vector, the previously
estimated eigenvector el,1(n − 1) is a close estimate to el,1(n) and, therefore, is used as
the initial guess to speed up convergence. For example, for a mobile moving at 75 mph
and symbol rate of 4800 symbols per second, the time correlation between the elements of
the channel vector of two consecutive symbols is 0.995. Also, the ratio |ηmax/η2| is lower
bounded by the average SINR per path1. In the current CDMA IS-95 standard, the closed
loop power control holds the SINR at the output of the RAKE combiner such that 1% FER
( frame error rate) is achieved. This translates to average SINR per path of 10.5 dB, 5.7
dB, and 2.8 dB for 1, 2, and 4 independent paths used for combining, respectively [111]. In
general, the lower the average SINR per path the more the number of iterations required for
convergence.
4.1.2 Time-Update of Covariance Estimates
It remains now to obtain the time updates for the estimates of the array covariance matrices
R−1/2xx (n), Ryy,l,1(n), and R−1
uu,l,1(n). Our goal is to track the time varying channel so that the
covariance estimates at each sampling instant closely approximate the second order statis-
tics of the observed channel process. When the channel process is slowly time-varying with
respect to the symbol rate, a fading memory update can be used to estimate the covariance
matrices at each sampling instant [112]. In this case, given the pre- and post-correlation sig-
nal vector x(n) and yl,1(n), the time-update equations for Rxx(n), Ryy,l,1(n), and Ruu,l,1(n)
can be written as
Rxx(n) = µRxx(n − 1)+ x(n)x∗(n) (4.1.13)
1In fact, if we assume that the total undesired signal is spatially white, we can easily show that |ηmax/η2| =1 + SINR.
Chapter 4. Beamforming for Time-Variant Channels 88
which is a good measure of how fast the channel vector is changing with time.
4.1.3 Algorithm Summary
In Table 4.3, we present a summary of the adaptive algorithm for estimating the channel
vector al,1(n) and the corresponding optimum weight vector wl,1(n) given the pre- and post-
correlation signal vectors x(n) and yl,1(n), respectively, including initial conditions and
the recursions that are involved in the computation. The two building blocks for the algo-
rithm are the recursive procedure for estimating the channel vector POWER and the inverse
square root time-update procedure SR UPDATE.
For the initial value of the covariance matrices, we use an initialization similar to that
used for the recursive least squares algorithm (RLS) [112]. In this case, we set the initial
value of the covariance matrices to
Rxx(0) = Ryy,l,1(0) = Ruu,l,1(0) = δI (4.1.34)
where δ is a small positive constant. Correspondingly, the initial value for R−1/2xx (0) and
Chapter 4. Beamforming for Time-Variant Channels 93
Step # Number of Flops
1 3K2 4K2 + 4K3 4K2 + 4K4 3K2 + 3K
5+6 11K2 + 2K7 O(K2)
8 K2
9 K2
Table 4.4: Floating point operations count
R−1/2uu,l,1(0)
R−1/2xx (0) = R−1/2
uu,l,1(0) = δ−1/2I (4.1.35)
The initial value of the principal eigenvector el,1(0) is set to 1, where 1 is a K × 1 vector
whose elements are all equal to 1. This choice ensures that the initial guess used in the power
recursion has a component in the direction of el,1(1) and therefore will ensure convergence.
Table 4.4 gives the number of floating point operations (flops) for each step of the algo-
rithm shown in Table 4.3. From this Table we can easily see that the overall complexity of
the algorithm is O(K2).
The number of flops is taken to be the number of multiplication operation involved. The
reason is that the instruction set of many of the current generation digital signal processors,
such as the TMS320C30 by Texas Instruments, include a multiply-and-accumulate instruc-
tion that will implement calculations such as a · b + c in one operation. For the sake of
simplicity, we made no distinction between real and complex numbers. For instance, the
product of a complex K × 1 vector and a scalar (either real or complex) will be said to
require K (complex) flops. In calculating the number of flops required for step 5 and 6,
Chapter 4. Beamforming for Time-Variant Channels 94
Maximum Doppler Frequency fd (Hz)
0 20 40 60 80 100
Mis
mat
ch L
oss
ρ (d
B)
-1.8
-1.6
-1.4
-1.2
-0.8
-0.6
-0.4
-0.2
-2.0
-1.0
0.0
∆ = 0ο
∆ = 5ο
∆ = 10ο
∆ = 20ο
∆ = 40ο
∆ = 60ο
Forgetting Factor µ = 0.98
Figure 4.1: Mismatch loss as a function of fd and � for µ = 0.98
we considered the rank-1 update structure for both R−1/2xx (n) and Ryy,l,1(n) in (4.1.14) and
(4.1.26). If we were to directly implement steps 5 and 6, then the number of flops required
is 2K3 + 2K2. By considering the rank-1 update structure, the number of flops required is
11K2 + 2K, which implies a reduction in the required number of flops for K ≥ 5. For step
7, the number of required flops will depend on the number of iterations required for con-
vergence. The use of el,1(n − 1) as an initial guess for el,1(n) will minimize the number of
iterations required and, therefore, will speed up convergence.
Chapter 4. Beamforming for Time-Variant Channels 95
Maximum Doppler Frequency fd (Hz)
0 20 40 60 80 100
Mis
mat
ch L
oss
ρ (d
B)
-1.2
-0.8
-0.6
-0.4
-0.2
-1.0
∆ = 0ο
∆ = 5ο
∆ = 10ο
∆ = 20ο
∆ = 40ο
∆ = 60ο
Forgetting Factor µ = 0.95
Figure 4.2: Mismatch loss as a function of fd and � for µ = 0.95
4.2 Simulation Results
To study the performance of the above algorithm, we considered a base station with a 5
element uniform linear array with sensor spacing 0.5λ. The multiple access plus self inter-
ference signal vector il,1 is modeled as an additive Gaussian noise that is both temporally
and spatially white. Simulation results in Chapter 6 show that this assumption holds if the
code length G is large and if the total number of path signals N · L is large with uniform an-
gle of arrival distribution. In this case Ruu,l,1 = σ2I, and the optimum beamforming weight
vector is wl,1 = al,1.
We considered one multipath component arriving at the array with mean angle of arrival
Chapter 4. Beamforming for Time-Variant Channels 96
-0.8
-0.7
-0.6
-0.5
-0.4
-0.3
-0.2
-0.1
0.80
0.85
0.90
0.950
2040
6080
100
Mism
atch Loss ρ (dB
)
Forgetting Factor µ
Maximum Doppler Frequency f d
Angle Spread ∆ = 5o
Figure 4.3: Mismatch loss as a function of fd and µ for � = 5◦
of θ = 90◦ and average SNR = −18dB. The processing gain G is 128. We assumed bal-
anced DQPSK spreading with the same information bits in both the in-phase and quadrature
channels. The bit rate was assumed to be 9.6 kbps. In each simulation run, a block of 100
bits was generated and differentially encoded. The differentially encoded data was spread
in the in-phase and quadrature channels using two different spreading codes. The channel
was generated using the model described in Chapter 2. We considered different cases with
different values of angle spread �, maximum doppler frequency fd, and forgetting factor
µ. The algorithm in Table 4.3 was used to estimate the channel vector al,1.
To study the performance of the algorithm, we considered the mean beamformer output
Chapter 4. Beamforming for Time-Variant Channels 97
Maximum Doppler Frequency fd (Hz)
0 20 40 60 80 100
Forg
etti
ng F
acto
r µ
0.75
0.80
0.85
0.90
0.95
1.00
1.05
∆ = 0ο
∆ = 5ο
∆ = 10ο
∆ = 20ο
∆ = 40ο
∆ = 60ο
Figure 4.4: Optimum forgetting factor µ as a function of fd and �
SINR normalized to the maximum output SINR (output SINR when al,1 = al,1, i.e. no errors
in estimating the channel vector). This mean normalized SINR is given by
ρ = E
{20 log10
|a∗l,1al,1|
|al,1||al,1|}
(4.2.1)
This quantity represents the loss in the average beamformer output SINR due to errors in
the channel vector estimate al,1 and is therefore called the mismatch loss.
Some of the simulation results are shown in Figures 4.1-4.5. Figure 4.1 shows the mis-
match loss as function of fd and � for µ = 0.98. We notice that for low doppler frequency,
the loss at low angle spread is higher than the loss at high angle spread, which can be ex-
plained by the reduction in space diversity gain provided by the array. At high Doppler fre-
quency, the situation is reversed. The reason for this is that at high Doppler frequency and
Chapter 4. Beamforming for Time-Variant Channels 98
Maximum Doppler Frequency fd (Hz)
0 20 40 60 80 100
Mis
mat
ch L
oss
ρ (d
B)
-0.6
-0.5
-0.4
-0.3
-0.2
-0.1
0.0
∆ = 0ο
∆ = 5ο
∆ = 10ο
∆ = 20ο
∆ = 40ο
∆ = 60ο
Figure 4.5: Mismatch loss as a function of fd and � with optimum forgetting factor
high angle spread, the channel variations are much faster than the rate at which array sample
covariances are updated (which is a function of the forgetting factor µ). In this case more
and more incorrect samples will be used in estimating the array sample covariances which
will lead to errors in the estimated channel vector. On the other hand, at high Doppler fre-
quency and low angle spread, the channel vector tends to fluctuate only in magnitude, and
the fluctuations in direction (phase) are very small and hence the errors in the estimated array
response vector will be reflected mainly in its magnitude, which will not affect the beam-
former output SINR. Figure 4.2 is similar to Figure 4.1 but with forgetting factor µ = 0.95.
We notice that there is a slight increase in the mismatch loss at low Doppler frequencies
while at high Doppler frequencies we notice a reduction in the mismatch loss.
Chapter 4. Beamforming for Time-Variant Channels 99
As we mentioned earlier, the forgetting factor µ that will give the best tracking perfor-
mance is a function of fd and �. With this optimum value of the forgetting factor µ, the
mismatch loss will be minimized. This is clearly illustrated in Figure 4.3 which shows the
mismatch loss as a function of fd and µ for � = 5◦. The optimum forgetting factor µ for
different values of fd and � was obtained via simulations and is shown in Figure 4.4. The
corresponding loss is shown in Figure 4.5.
4.3 Summary
In this chapter, we presented an adaptive algorithm for estimating the channel vector and the
corresponding weight vector. The algorithm is based on the code-filtering approach pre-
sented in the previous chapter. The algorithm consists of two basic building blocks. The
first is a recursive procedure ”POWER” for tracking and estimating the channel vector.
The second block is a procedure ”SR UPDATE” for time-updating the inverse square roots
R−1/2xx (n) and R−1/2
uu,l,1(n). The overall algorithm complexity is O(K2). Simulation results
show that with optimum time-update (i.e. if we use optimum forgetting factor), we can track
the channel vector within 0.5 dB of the true value even in severe propagation environments.
Chapter 5
Overview of the IS-95 CDMA Standard
CDMA is a modulation and multiple access scheme based on spread spectrum communi-
cations. Proponents of the CDMA technology cite several potential advantages over the
traditional FDMA AMPS and the IS-54 TDMA approaches [113-115]. First, statistics of
telephone conversations suggest that in a typical full-duplex two-way voice conversation,
the duty cycle of each voice is less than 35%. Exploiting the voice activity in either FDMA
or TDMA systems might be hard to implement because of the time delays associated with
reassigning channel resources during speech pauses. In CDMA, it is possible to reduce the
transmission rate when there is no speech, and thereby substantially reduce interference to
other users. This reduction in interference power can be transformed to either an increase
in system capacity or a reduction in average mobile transmit power [113]. However, we
should note that a statistical multiplexing protocol based on packet reservation multiple ac-
cess (PRMA) has been proposed to make use of the speech activity in TDMA systems [116,
117].
Another potential advantage for CDMA is that special frequency reuse plans are not nec-
essary. Since users are separated by using different codes on the entire system bandwidth
100
Chapter 5. Overview of the IS-95 CDMA Standard 101
simultaneously, there is no need to avoid co-channel interference by requiring different fre-
quency usage in neighboring cells. In fact the concept of a cell in terms of frequency plan-
ning is no longer necessary.
The CDMA IS-95 standard uses time and path diversity to mitigate the effect of fre-
quency selective multipath fading. Time diversity is obtained by the use of forward error
correction (FEC) and interleaving. Path diversity is inherently provided by the CDMA ap-
proach by spreading the signal over a wide bandwidth (1.25 MHz). Such a signal with wide
bandwidth will resolve the multipath components and, thus, provide the receiver with sev-
eral independently fading signal paths. This path diversity is exploited by the use of a RAKE
receiver to combine different multipath components. The use of a wideband signal for trans-
mission may be also viewed as just another method for obtaining frequency diversity [33].
In a CDMA system every user is a source of interference to every other user, which
makes mobile station transmit power control a key element in its current implementation.
The current CDMA standard uses several power control techniques (open loop and closed
loop power control) to optimize the system performance. In addition, the base station uses
three sectored antennas (each covers 120◦ of the azimuth) to reduce the multiple access in-
terference and thereby increase system capacity.
Field test results of the IS-95 cellular system are reported in [118, 111]. This test was
conducted in August 1993 in the San Diego area and included four base stations for a total
of eight sectors and 86 class I mobile units. Capacity estimates from this field test show
that the IS-95 radio capacity/sector is 11.4 times better than that of AMPS. This translates
to Erlang capacity/sector 16.4 times higher than that of AMPS. However, we should note
that in [119], in commenting on a similar previous field test experiment, Cox cites several
potential capacity-reducing factors that were not exercised in this experiment.
In the following sections we will briefly describe the different signal processing func-
tions used for forward and reverse link modems.
Chapter 5. Overview of the IS-95 CDMA Standard 102
ConvolutionalEncoder
and Repetition
Interleaver
Long CodeGenerator
User LongCode Mask
× × +
User WalshCode
Pilot WalshCode
SpeechCoder
I-ChannelShort Code
Q-ChannelShort Code
× ×
×
+
×
BasebandFilter
BasebandFilter
cos( t)ωc
sin( t)ωc
I(t)I
Q(t)Q
a (t)(I)
a (t)(Q)
c (t)i
W (t)i
W (t)p
Figure 5.1: CDMA IS-95 forward link waveform generation
5.1 CDMA Forward Link
In the proposed CDMA standard the forward link uses a combination of frequency division,
pseudorandom code division, and orthogonal signal multiple access techniques. Frequency
division is employed by dividing the available cellular spectrum into nominal 1.25 MHz
bandwidth channels.
A simplified description of the signal processing functions for transmission from the
base station to the mobile unit is shown in Figure 5.1. The underlying data rate for the sys-
tem is 9600 bits/sec which represents the speech coder rate of 8550 bits/sec augmented by
error correction coding which is tailored to the speech-coding technique used. (The speech
coder actually detects speech activity and changes data rate to lower values during quiet
intervals, but 9600 bits/sec is the maximum error-protected data rate). The 9600 bits/sec
bit stream is segmented into 20-msec blocks and then further convolutionally encoded to
provide the capability of error correction and detection at the receiver. The convolutional
Chapter 5. Overview of the IS-95 CDMA Standard 103
encoder used has a constraint length k = 9 and a code rate r of one-half. This will bring the
data rate to 19.2 kbits/sec. The convolutional encoding is followed by interleaving over a
20-msec interval for burst error protection due to fast fading in the radio channel. The 19
kbits/sec output of the interleaver is then modified by the use of the so-called long code,
which serves as a privacy mask.
The modified stream is then encoded for spread-spectrum transmission using binary or-
thogonal Walsh codes [33] of dimension 64. This will produce 64-fold spreading of the data
stream, resulting in a transmission rate of 1.2288 Mchips/sec. The structure of the Walsh
code provides 64 orthogonal sequences, and one of the 64 sequences is assigned to a mo-
bile unit during call setup. In this way, 64 orthogonal ”channels ”can be established on the
forward link by the CDMA encoding on the same carrier frequency. After Walsh encoding
, the spread data stream is separated into I and Q streams, each of which is modified by a
unique ”short code” of length 32768. The resulting spread spectrum stream is carried over
the air interface with filtered QPSK modulation.
All signals transmitted from a base station in a particular CDMA radio channel share
the same set of 64 Walsh codes and the same pair of short codes. However, signals from
different base stations are distinguished by time offsets from the basic short code which
allows CDMA signals from each base station to be uniquely identified. Different signals
transmitted from a given base station in a particular CDMA radio channel are distinguished
at the mobile receiver by the orthogonal Walsh code. The orthogonality of the Walsh codes
provides near perfect isolation between the multiple signals transmitted by the same base
station. However, the presence of multipath propagation will partially destroy this orthog-
onality.
An important aspect of the forward link waveform generated at the base station is the use
of a pilot signal that is transmitted by each base station and is used as a coherent carrier ref-
erence for demodulation by all mobile receivers. The pilot is transmitted a relatively higher
power level (approximately 20% of total forward link power budget [113, 120]) than other
Chapter 5. Overview of the IS-95 CDMA Standard 104
types of signals which allows for tracking of the carrier phase. The pilot channel signal is
unmodulated by information and uses the zeros Walsh function. Thus the pilot signal simply
consists of the quadrature pair of short codes. The mobile receiver can obtain synchroniza-
tion with the nearest base station without prior knowledge of the identity of the base station
by searching out the entire length of the short code. The strongest signal’s time offset cor-
responds to the time offset of the short code of the base station to which the mobile has the
best propagation channel (which is often the nearest base station). After synchronization,
the pilot signal is used as a coherent carrier phase reference for demodulation of the other
signals from this base station.
The mobile receiver uses a limited number of correlators in a RAKE structure for de-
modulating the signals from the base station. The receiver uses the pilot signal to estimate
the phase and the amplitude for each of the tracked multipath components and to coherently
combine them.
5.2 CDMA Reverse Link
The CDMA reverse link also employs PN spread spectrum modulation using the same short
code as that used for the forward link. Here, however, all mobiles use the same code phase
offset. Signals from different mobiles are distinguished at the base station by the use of a
very long (242 − 1) PN sequence with a user address determined time offset. Because every
possible time offset is a valid address, an extremely large address space is provided.
The data rate of the system is also 9600 bits/sec. The transmitted digital information
stream is segmented into 20-msec blocks and then further convolutionally encoded using a
code of rate 1/3 and constraint length k = 9. This will bring the data rate to 28.8 kbits/sec.
The encoded information bits are then interleaved over the 20 ms block. The interleaved
information is then grouped into symbol groups (or code words) of 6 bits. These code words
Chapter 5. Overview of the IS-95 CDMA Standard 105
UserLong Code
M-ary OrthogonalWalsh Modulator
W(t)Interleaver
ConvolutionalEncoder
andRepetition
SpeechCoder
Long CodeGenerator
Long CodeMask c (t)i
Power Control
ωcsin( t)
ωccos( t)
+
Q(t)
I(t)
(Q)a (t)
(I)a (t)
cT / 2Delay
I-ChannelShort Code
Q-ChannelShort Code
Q
I BaseBandFilter
BaseBandFilter
S/PbitsJ
Figure 5.2: CDMA IS-95 reverse link waveform generation
are used to select one of 64 different orthogonal Walsh functions for transmission. At the
output of the Walsh modulator, the chip rate is 307.2 kchips/sec. The final signal processing
elements perform the direct-sequence spreading functions. First, the modulation symbols ,
or Walsh functions, are spread by using the mobile-specific long code at a rate of 1.2288
Mchips/sec, i.e., 256 chips per modulation symbol. The data stream is then split into I and
Q streams where it is modified with the short code pair. The resulting spread spectrum signal
is then carried over the air interface with a filtered O-QPSK (Offset-QPSK) modulation.
Note that the use of the Walsh function in generating the reverse link waveform is dif-
ferent from that on the forward link. On the forward link, the Walsh function is determined
by the channel that is assigned to each mobile while on the reverse link the Walsh function
is determined by the information being transmitted. The use of Walsh function modulation
on the reverse link is a simple way of obtaining 64-ary modulation. This is the best way
of providing a high quality link over a fading channel with low SINR where a pilot phase
reference cannot be provided [89].
Figure 5.3 shows a block diagram for the base station receiver. The base station uses
Chapter 5. Overview of the IS-95 CDMA Standard 106
ΣSearchers
Incoherent RAKE
1
L
1
2
6AntennaOutputs
M WalshSequence
Correlators
M WalshSequence
Correlators
Figure 5.3: CDMA Base station receiver: IS-95 approach
dual antenna diversity in each sector to provide path diversity. The antennas are widely
separated such that independent fading at each antenna is obtained. For any given mobile,
this structure implements a noncoherent RAKE receiver with a limited number of corre-
lators for demodulating the L strongest multipath components received on all base station
antennas. Noncoherent reception is used since, as we mentioned earlier, a pilot signal is not
available on the reverse link. Time delay information of the L strongest multipath compo-
nents is obtained by the front-end searcher. The front-end searcher has a larger number of
scanning correlators that continuously scans all six antenna outputs for multipath signals
received from the mobile. Tracked multipath components are rank ordered based on their
energy and the L strongest are used for combining (current implementation of cell modem
ASICs searches for 8 paths and uses the strongest 4 for combining [8, 111]).
Chapter 5. Overview of the IS-95 CDMA Standard 107
5.3 Summary
This chapter provides a brief overview of the CDMA IS-95 cellular standard. We briefly
described different signal processing functions used for both the forward and reverse link
modems. In the next chapter, we will apply the adaptive beamforming and ”Beamformer-
RAKE” ideas presented in Chapter 3 to CDMA wireless systems that are based on the IS-95
standard and propose an antenna array-based base station receiver architecture and study its
performance.
Chapter 6
CDMA Base Station Receiver with
Antenna Arrays
In Chapter 3 we derived a space-time ”Beamformer-RAKE” receiver that exploits the spa-
tial structure in the received multipath signal in addition to the temporal structure to pro-
vide a more efficient combining of paths. We showed that for coherent detection, a pilot
signal is still needed for the ”Beamformer-RAKE” to estimate the phase information for
each multipath signal. However, transmitting a pilot in each mobile’s signal, whose power
is greater than the data-modulated portion of the signal, reduces efficiency to less than 50%
[89]. Instead, either differential phase shift keying (DPSK) which does not require phase co-
herence or M-ary orthogonal modulation with noncoherent reception should be used. For
M > 8, where M is the number of orthogonal signals, orthogonal modulation is better than
DPSK [33, 89]. Therefore, the CDMA IS-95 standard that was briefly described in the pre-
vious chapter uses 64-ary orthogonal modulation for the reverse link. We also mentioned
in the previous chapter that in the current implementation of the CDMA standard dual an-
tenna diversity is used at the base station to provide path diversity and that the L strongest
paths are incoherently combined. In this approach, the base station treats all diversity el-
ements equally. In other words, in selecting the L strongest paths for combining, the base
108
Chapter 6. CDMA Base Station Receiver with Antenna Arrays 109
ΣIncoherent RAKE
1
L
Beamformerfor 1st path
Beamformerfor L-th path
Searcher
Beamformer-RAKE Receiver
Base Station Antenna Array
M WalshSequence
Correlators
M WalshSequence
Correlators
AntennaArrayOutput
Figure 6.1: CDMA Base station receiver: beamforming approach
station receiver does not make any distinction as to whether different paths came from the
same antenna or different antennas. Also, signals received at two or more different antennas
due to the same multipath component wavefront are considered different multipath signals.
Clearly, this approach does not utilize any spatial structure in the received multipath signal.
This spatial structure can be utilized by the use of the ”Beamformer-RAKE” receiver.
In this case, the base station will have an antenna array of K elements in each sector. The
proposed base station receiver structure is similar to that of the IS-95 system except that the
front-end searcher now searches for vector multipath components (by using the approach
outlined in Chapter 3). In addition, each multipath demodulator will be preceded by a beam-
former. A simplified overview of the ”Beamformer-RAKE” base station receiver is shown
in Figure 6.1.
In this chapter, we will present an overall base station receiver architecture for CDMA
with M-ary orthogonal modulation based on the ”Beamformer-RAKE” structure and study
its performance in terms of the uncoded bit error rate as a function of number of users and
Chapter 6. CDMA Base Station Receiver with Antenna Arrays 110
number of antennas. In this analysis, we shall make use of the power control performance
results that will be presented later in Chapter 7. We start by developing a model for the
received signal vector at the base station array. Then, we will describe the proposed re-
ceiver structure and its different signal processing blocks. This will be followed by a de-
tailed performance analysis for the proposed receiver and numerical and simulation results.
The analysis in this chapter makes use of analysis results for DS/CDMA with M-ary orthog-
onal modulation but without antenna arrays that appeared in [121-125].
6.1 Received Signal Vector Model
The mobile transmitter block diagram is shown in Figure 5.2. The binary data at the output
of the interleaver are grouped into groups of J = log2 M bits. Each group is mapped into one
of M orthogonal Walsh sequences W(t). The resulting signal is then spread using the user’s
long PN code ci(t). The signal is further multiplied in both I and Q channels by the short
PN codes a( I)(t) and a(Q)(t) respectively. The PN modulated Q channel signal is delayed
by half a chip period Tc/2. The two spread signals are up-converted to radio frequency for
transmission. The power of the transmitted signal is adjusted according to both the open
and closed loop power control mechanisms (see Chapter 7). Then we can write the signal
transmitted by the i-th mobile as
si(t) = ψi
√Pi
(W (h)(t)ci(t)a
( I)(t) cos(ωct)+W (h)(t − To)ci(t − To)a
(Q)(t − To) sin(ωct))
0 ≤ t ≤ Tw (6.1.1)
where Pi is the transmitted power per symbol per dimension, Tw is the symbol period, ωc
is the carrier angular frequency, To = Tc/2 is the time offset between the I and Q channels,
and finally ψi is a Bernoulli random variable that models the voice activity of the ith user
(we assume that a user will be on with probability ν and will be off with probability 1− ν).
Chapter 6. CDMA Base Station Receiver with Antenna Arrays 111
W (h)(t) is the hth orthogonal Walsh function, h = 1, · · · , M. Let the processing gain be
G = Tw/Tc . For simplicity of notation, we shall denote the product of the user’s PN code
and the I or Q channel PN code as
c( I)i (t) = ci(t)a
( I)(t) and c(Q)i (t) = ci(t)a
(Q)(t)
To simplify our analysis, the PN codes c( I)i (t) and c(Q)
i (t) are represented by [125, 106]
c( I)i (t) =
∞∑r=−∞
c( I)i,r p(t − kTc) (6.1.2)
c(Q)i (t) =
∞∑r=−∞
c(Q)i,r p(t − kTc) (6.1.3)
where c( I)i,r and c(Q)
i,r are assumed to be i.i.d. random variables taking values ±1 with equal
probability, and p(t) is the chip pulse shape, which can be any time-limited waveform. Here
we assume that p(t) is rectangular although our results can be easily extended for any time-
limited waveform.
We assume the multipath vector channel model in (2.2.25). To simplify our analysis, we
assume that we have a constant deterministic power-delay profile and that the log-normal
slow fading is the same for all multipath components. We also assume that the channel pa-
rameters vary slowly as compared to the symbol duration Tw so that they are constant over
several symbol durations. Therefore, after downconverting to baseband, we can write the
K × 1 complex baseband received signal vector for the i-th user as
xi(t) =√
Si Piψi
Li∑l=1
c(h)i (t − τl,i)ejφl,i al,i (6.1.4)
where Si represents the log-normal shadowing experienced by the ith user, φl,i = ωcτl,i and
Li is the number of multipath component for the ith user, al,i is the K × 1 channel vector of
the base station antenna array to signals in the l-th path from the ith mobile, and c(h)i (t) is
defined as
c(h)i (t) = W (h)(t)c( I)i (t)+ jW (h)(t − To)c
(Q)i (t − To) (6.1.5)
Chapter 6. CDMA Base Station Receiver with Antenna Arrays 112
Let N be the number of cochannel mobiles. The total received signal at the cell site is
the sum of all users’ signals plus noise and is given by
x(t) =N∑
i=1
Li∑l=1
xl,i(t)+ n(t) (6.1.6)
The vector n(t)= nc(t)+ jns(t) is the K ×1 additive Gaussian noise vector with zero mean
and covariance
E{n(t1)n∗(t2)} = σ2nI · δ(t1 − t2) (6.1.7)
where σ2n is the noise variance per antenna.
6.2 Receiver Model
The block diagram of the base station antenna array receiver is shown in Figure 6.2. It has
a ”Beamformer-RAKE” structure where several multipath components are tracked in both
time and space. After down-converting to baseband, the outputs of the LPF are fed into a
bank of M Walsh correlators shown in Figure 6.3. Assuming that the hth Walsh symbol
was transmitted, where h = 1, · · · , M, the pre-correlation and post-correlation signal vec-
tors x(t) and y(h)l,i are used to estimate the channel vector al,i and the corresponding optimum
beamforming weight vector wl,i for the lth multipath component of the ith mobile from the
pre-correlation and post-correlation array covariances Rxx and Ryy,l,i1 using the code filter-
ing approach derived in Chapter 3.
1As we have seen earlier, all matrices and vectors are functions of time. However, for simplicity of nota-tion, we will drop the time index in the rest of this chapter.
Chapter
6.C
DM
AB
aseStation
Receiver
with
Antenna
Arrays
113
x
z
w , w , ... , w1 2 L
z1
(M)
z1
(1)
ET
K
1
MWalsh Sequence
Correlators
MWalsh Sequence
Correlators
MWalsh Sequence
Correlators
MWalsh Sequence
Correlators
L parallel demodulators
L parallel demodulators
Optimum Beamformingand Incoherent
RAKE Combining
Optimum Beamformingand Incoherent
RAKE Combining
Weight VectorEstimation
weight vectors
select post-correlationsignal vector
M-ary Decoder,Deinterleaver,
andViterbi Decoder
Measure FrameError Rate (FER)
Closed LoopPower Control
Algorithm
Up/DownPower Control
Command
Threshold
Select Indexof Maximum
data
pre-correlationsignal vector
post-correlationsignal vector
Down-converterand LPF
Down-converterand LPF
Figure 6.2: Base station receiver block diagram
Chapter
6.C
DM
AB
aseStation
Receiver
with
Antenna
Arrays
114
c (t)1
a (I)
a (Q)
a (Q)
a (I)
WM
W1
W1
WM
LPF
cos( t)
sin( t)
ω
ω
c
c
-
+
+
+
z (m)(1)
1,1
z (m)(1)
L,1
z (m)(1)
k,1
z (m)(M)
k,1
z (m)(M)
1,1
z (m)(M)
L,1
LPF
∫
∫
∫
∫
Figure 6.3: Correlators
Chapter 6. CDMA Base Station Receiver with Antenna Arrays 115
For each multipath component, we have M different post-correlation signal vectors y(h)l,i ,
h = 1, · · · , M. The vectors y(h)l,i are fed to an optimum beamformer. The outputs of the
L1 beamformers for the hth Walsh function w∗l,iy
(h)l,i are then fed into an incoherent RAKE
combiner. The output of the incoherent RAKE combiner z(h)i is the decision variable for the
hth Walsh function. The beamformer and the incoherent RAKE combiner for the hth Walsh
function are shown in Figure 6.4.
In order to update the post-correlation array covariance Ryy,l,i (that will be used in es-
timating al,i and wl,i), the receiver needs the post-correlation vector y(h)l,i corresponding to
the true transmitted Walsh symbol W (h)(t). However, at this stage the receiver has no prior
knowledge of which post-correlation vector y(n)l,i is the right one. Here, the receiver relies on
the inherent correlation of the multipath vector channel and the assumption that the chan-
nel remains almost constant over several symbol periods. In this case, the receiver uses a
delayed update of Ryy,l,i (and hence delayed estimation of the channel vector and the opti-
mum beamforming weight vector). This is done by using the decision on the current Walsh
symbol h to select the post-correlation vector y(h)l,i to update Ryy,l,i and obtain the optimum
weight vector wl,i. This weight vector wl,i will be used for beamforming for the next sym-
bol.
The decision variables z(1)i , · · · , z(M)i at the output of the incoherent RAKE are then fed
to an M-ary decoder, deinterleaver, and Viterbi convolutional decoder. Without loss of gen-
erality let us assume that the 1st user is the desired user and let τk,1 be the time delay of the
kth tracked multipath which is assumed to be estimated perfectly and k = 1, · · · , L1. Then,
we can write the post-correlation signal vector y(n)l,1 for the kth tracked multipath component
for the first user as
y(n)k,1 = 1√
Tw
∫ τk,1+Tw
τk,1
x(t)c(n)∗1 (t − τk,1)dt (6.2.1)
= d(n)k,1 + u(n)
k,1 if n = h (6.2.2)
= u(n)k,1 if n �= h (6.2.3)
Chapter 6. CDMA Base Station Receiver with Antenna Arrays 116
where
d(n)k,1 = 2
√TwS1 P1e jφk,1 ak,1 (6.2.4)
u(n)k,1 = m(n)
k,1 + s(n)k,1 + n(n)k,1 (6.2.5)
and d(n)k,1 is the desired signal vector, m(n)
k,1 is the multiple access interference (MAI) signal
vector, s(n)k,1 is the self-interference (SI) signal vector due to other multipath components of
the 1st user, and n(n)k,1 is due to the AWGN. Let Ai =
√Si Piψi. Also, for the kth tracked
multipath let the optimum beamforming weights determined using the previously estimated
Walsh symbol be wk,1. For an equal gain combining incoherent RAKE, the nth decision
variable of the 1st user corresponding to the nth Walsh symbol is given by [33]
z(n)1 =L1∑
k=1
z(n)k,1 =L1∑
k=1
|w∗k,1y(n)
k,1|2 n = 1, · · · , M (6.2.6)
Now, to select which post-correlation signal vector y(n)l,1 should be used in estimating the
post-correlation array covariance Ryy,l,i, a hard decision is made on which Walsh symbol
was transmitted
h = arg maxn=1,···,M
(z(n)1
)(6.2.7)
However, for the data a symbol-by-symbol M-ary decoder is used [89]. Both approaches
yield exactly the same decisions for the M-ary symbol and both are optimal (i.e. a maxi-
mum likelihood rule) for an AWGN channel. Since the MAI is not necessarily Gaussian,
this decision rule is actually not optimal. However, when the number of cochannel users is
large, the multiple access interference can be modeled as Gaussian noise and therefore this
decision rule can be used. The primary reason for using the symbol-by-symbol approach
for the data is to provide improved performance with error-correcting codes by using soft
decision decoding [33, 89].
Note also that we cannot use the output after the convolutional decoding and deinter-
leaving to select the post-correlation signal vector. The reason for this is that we will have
to wait for a decision to be made on the current symbol and convolutionally encode and
Chapter 6. CDMA Base Station Receiver with Antenna Arrays 117
*1,1
w
*w2,1
*wL,1
Σ
z(n)
1,1
z (n)
L,1
z (n)
Time Align
τ 1
Time Alignτ 2
Time Alignτ L
1z (n)
2,1
2| . |
2| . |
2| . |
Figure 6.4: Optimum Beamforming and Incoherent RAKE
interleave again. By the time this process is over (which is at least twice the time of one
frame of bits), the channel would have changed and the estimated channel vector and the
channel vector of the new symbol will be quite different. This will lead to a degradation in
the beamformer output SINR.
6.3 Signal Statistics
In order to derive the uncoded bit error probability, we need to derive the statistics of the
decision variables z(1)1 , z(2)1 · · · , z(M)
1 . First, we will examine the different terms in z(n)k,1, i.e.
the multiple access interference signal vector due to other cochannel users m(n)k,1, the self in-
terference signal vector due to the user own multipath components s(n)k,1, and the noise vector
n(n)k,1.
Chapter 6. CDMA Base Station Receiver with Antenna Arrays 118
6.3.1 Noise Analysis
The noise term n(n)k,1 is given by
n(n)k,1 = 1√
Tw
∫ τk,1+Tw
τk,1
n(t)c(n)∗1 (t − τk,1) dt (6.3.1)
= (n(n), I I
k,1 + n(n),QQk,1
)+ j(−n(n), I Q
k,1 + n(n),QIk,1
)(6.3.2)
where n(n), I Ik,1 , n(n),QQ
k,1 , n(n), I Qk,1 , and n(n),QI
k,1 are defined as
n(n), I Ik,1 = 1√
Tw
∫ τk,1+Tw
τk,1
nc(t)c( I)i (t − τk,1)W
(n)(t − τk,1) dt (6.3.3)
n(n),QQk,1 = 1√
Tw
∫ τk,1+Tw
τk,1
ns(t)c(Q)i (t − To − τk,1)W
(n)(t − To − τk,1) dt (6.3.4)
n(n), I Qk,1 = 1√
Tw
∫ τk,1+Tw
τk,1
nc(t)c(Q)i (t − To − τk,1)W
(n)(t − To − τk,1) dt (6.3.5)
n(n),QIk,1 = 1√
Tw
∫ τk,1+Tw
τk,1
ns(t)c( I)i (t − τk,1)W
(n)(t − τk,1) dt (6.3.6)
For n(n), I Ik,1 , we have
n(n), I Ik,1 = 1√
Tw
∫ τk,1+Tw
τk,1
nc(t)c( I)i (t − τk,1)W
(n)(t − τk,1) dt (6.3.7)
= 1√Tw
G−1∑b=0
∫ τk,1+(1+b)Tc
τk,1+bTc
nc(t)c( I)i (t − τk,1)W
(n)(t − τk,1) dt (6.3.8)
= 1√Tw
G−1∑b=0
∫ τk,1+(1+b)Tc
τk,1+bTc
±nc(t) dt (6.3.9)
We can easily show that n(n), I Ik,1 is a Gaussian random vector with zero mean and covariance
σ2n
2 I. Similarly, we can show that n(n),QQk,1 , n(n), I Q
k,1 , and n(n),QIk,1 are all uncorrelated zero mean
Gaussian random vectors with the same covariance. Hence n(n)k,1 is a zero mean Gaussian
random vector with covariance 2σ2nI.
Chapter 6. CDMA Base Station Receiver with Antenna Arrays 119
6.3.2 Self and Multiple Access Interference Analysis
The self interference due to other multipath components is given by
s(n)k,1 =L1∑l=1l �=k
1√Tw
∫ τk,1+Tw
τk,1
xl,1(t)c(n)∗1 (t − τk,1) dt (6.3.10)
= A1
L1∑l=1l �=k
I (n)k,1,l,1ejφl,1 al,1 (6.3.11)
= A1
L1∑l=1l �=k
[(I (n), I Ik,1,l,1 + I (n),QQ
k,1,l,1
)+ j(−I (n), I Q
k,1,l,1 + I (n),QIk,1,l,1
)]e jφl,1al,1 (6.3.12)
where I(n), I Ik,1,l,i, I (n),QQ
k,1,l,i , I (n), I Qk,1,l,i , and I (n),QI
k,1,l,i are defined as
I (n), I Ik,1,l,i = 1√
Tw
∫ τk,1+Tw
τk,1
W (h)(t − τl,i)c( I)i (t − τl,i)
W (n)(t − τk,1)c( I)1 (t − τk,1) dt (6.3.13)
I (n),QQk,1,l,i = 1√
Tw
∫ τk,1+Tw
τk,1
W (h)(t − To − τl,i)c(Q)i (t − To − τl,i)
W (n)(t − To − τk,1)a(Q)
1 (t − To − τk,1) dt (6.3.14)
I (n), I Qk,1,l,i = 1√
Tw
∫ τk,1+Tw
τk,1
W (h)(t − τl,i)c( I)i (t − τl,i)
W (n)(t − To − τk,1)c(Q)
1 (t − To − τk,1) dt (6.3.15)
I (n),QIk,1,l,i = 1√
Tw
∫ τk,1+Tw
τk,1
W (h)(t − To − τl,i)c(Q)i (t − To − τl,i)
W (n)(t − τk,1)c( I)1 (t − τk,1) dt (6.3.16)
Also, we can write the MAI due to other users’ signals as
m(n)k,1 =
N∑i=2
Li∑l=1
1√Tw
∫ τk,1+Tw
τk,1
xl,i(t)c(n)∗1 (t − τk,1) dt (6.3.17)
=N∑
i=2
Li∑l=1
Ai I(n)k,1,l,ie
jφl,i al,i (6.3.18)
=N∑
i=2
Li∑l=1
Ai
[(I (n), I Ik,1,l,i + I (n),QQ
k,1,l,i
)+ j(−I (n), I Q
k,1,l,i + I (n),QIk,1,l,i
)]e jφl,i al,i . (6.3.19)
Chapter 6. CDMA Base Station Receiver with Antenna Arrays 120
Let
W (h)(t − τl,i)a( I)i (t − τl,i) =
∞∑r=−∞
q( I)i,r p(t − rTc − τl,i) (6.3.20)
where q( I)i,r = W (h)
r c( I)i,r . It follows from (6.1.2) and (6.1.3) that q( I)
i,r , r = −∞, · · · ,∞ is an
i.i.d. binary random sequence taking values ±1 with equal probability. Hence, it follows
that I (n), I Ik,1,l,i is zero mean. Using Equation (6.3.20), we can rewrite I(n), I I
k,1,l,i as
I (n), I Ik,1,l,i = 1√
Tw
G−1∑b=0
q( I)1,b
∫ τk,1+(b+1)Tc
τk,1+bTc
p(t − bTc − τk,1)
∞∑r=−∞
q( I)i,r p(t − rTc − τl,i) dt
= 1√Tw
G−1∑b=0
q( I)1,b
[q( I)
i,b−1Rp(βk,1,l,i)+ q( I)i,b Rp(Tc − βk,1,l,i)
](6.3.21)
where βk,1,l,i = τl,i − τk,1 modulo-Tc and Rp(s) is the partial auto correlation of the chip
waveform defined as
Rp(s) =∫ s
op(t)p(t + Tc − s) dt 0 ≤ s ≤ Tc (6.3.22)
For rectangular pulses, Rp(s) = s. For asynchronous networks, a reasonable assumption is
that βk,1,l,i are independent and uniformly distributed over [0, Tc]. Let
Fb = q( I)i,b−1Rp(βk,1,l,i)+ q( I)
i,b Rp(Tc − βk,1,l,i) (6.3.23)
Using the results in [106], we can show that {Fb}b=0,···,G−1 are independent random vari-
ables. Hence
Var{
I (n), I Ik,1,l,i
} = 1Tw
G−1∑b=0
Var {Fb} (6.3.24)
= 1Tw
G−1∑b=0
E{
R2p(βk,1,l,i)+ R2
p(Tc − βk,1,l,i)}
(6.3.25)
= 1T2
c
∫ Tc
0
[β2
k,1,l,i + (Tc − βk,1,l,i)2]
dβk,1,l,i (6.3.26)
= 23
Tc (6.3.27)
Similarly, we can also show that I(n), I Ik,1,l,i,I
(n),QQk,1,l,i , I (n), I Q
k,1,l,i , and I (n),QIk,1,l,i are all zero mean uncor-
related random variables with the same variance given by (6.3.27).
Chapter 6. CDMA Base Station Receiver with Antenna Arrays 121
Remark: In deriving the variance of I(n), I Ik,1,l,i, we used the assumption that the chip pulse
shape is rectangular. In reality the channel is bandlimited (due to low pass filtering follow-
ing the down converter) and the received signal cannot be a square wave. Under the condi-
tion that the same amount of energy is received regardless of the channel used, the received
signal through a bandlimited channel will have a higher peak value, resulting in a higher
level of MAI interference due to larger fluctuations. In [124, 106] it was shown that if the
bandlimited channel has ideal low pass filter characteristics with a bandwidth B = 1/Tc,
then, we would have
Var{I (n), I Ik,1,l,i} = Tc (6.3.28)
Similarly, covariance of n(n)k,1 will then be
Var{n(n)
k,1
} = 2Tcσ2nI (6.3.29)
2
The total interference vector i(n)k,1 = m(n)k,1 + s(n)k,1 is modeled as a zero mean complex Gaus-
sian random vector with covariance I(n)k,1 = E{i(n)k,1i(n)∗k,1 }. Although, this assumption does not
always hold for CDMA analysis, it was shown in [106] that it is valid for large G. Moreover,
simulation results presented later in this chapter show that for large N · L, if we assume that
the angles of arrival of the multipath components are uniformly distributed over the sector,
the total interference vector i(n)k,1 will be spatially white. In this case
I(n)k,1 = 2Tcσ2I I (6.3.30)
where σ2I is given by
σ2I = C ·
[N∑
i=2
νLi E{Si Pi} + (L1 − 1)E{S1 P1}]
(6.3.31)
and C is a constant equal to 2 for a bandlimited channel and 43 for a rectangular pulse shape.
For the remainder of our analysis we will assume the case of a bandlimited channel, i.e.
Chapter 6. CDMA Base Station Receiver with Antenna Arrays 122
C = 2. The covariance of the interference-plus-noise vector u(n)k,1 is then given by
R(n)uu,k,1 = σ2I (6.3.32)
where σ2 = 2Tc(σ2I + σ2
n).
6.3.3 Decision Statistics
Consider the pre-correlation and post-correlation signal vectors x(t) and y(h)k,1. With the as-
sumption that the MAI is spatially white, the optimum beamforming weights can be shown
to be
wk,1 = ζak,1 (6.3.33)
where ζ is some arbitrary constant (that does not change the beamformer output SINR). For
simplicity of the analysis, we set ζ = 1/√
a∗k,1ak,1.
Define the beamformer output for the k-th multipath component of the 1st user w∗k,1y(n)
k,1
as U (n)k,1
U (n)k,1 = 2A1
√Tw|ak,1|2e jφk,1 + w∗
k,1u(n)k,1 for n = h , (6.3.34)
= w∗k,1u(n)
k,1 for n �= h (6.3.35)
where |ak,1| = √a∗
k,1ak,1. We can easily show that V (n)k,1 = w∗
k,1u(n)k,1 is a zero mean complex
Gaussian random variable with variance σ2. For simplicity of notation, let L1 = L. Then,
the decision variables for the first user are
z(n)1 =L∑
l=1
|U (n)l,1 |2 , n = 1, · · · , M (6.3.36)
From [33], and conditioned on A1 and al,1, l = 1 · · · L, we can show that for n = h, z(n)1 has
a non-central χ2 distribution with 2L degrees of freedom and non-centrality parameter
E = 4A21Tw
L∑l=1
|al,1|2 (6.3.37)
Chapter 6. CDMA Base Station Receiver with Antenna Arrays 123
The non-centrality parameter E is the symbol energy. For n �= h, z(n)1 has a χ2 distribution
with 2L degrees of freedom. Therefore we can write the conditional probability density
function of z(n)1 as
fz(n)1(z|γs) =
1σ2
(z
σ2γs
) L−12
e−γsσ2+z
σ2 IL−1
(√4γszσ2
)n = h,
1σ2L�(L)z
L−1e−z/σ2n �= h
(6.3.38)
where IL(.) is the modified Bessel function of the L-th order defined earlier in Section 2.1.2,
�(.) is the Gamma function, and
γs = Eσ2
(6.3.39)
We may recognize γs as the symbol energy to interference-plus-noise ratio.
Remark: In this analysis, we used the assumption that the channel vector remains constant
over two symbol periods. Also, we assumed that ak,1 is estimated perfectly. In reality the
channel is time varying and the array covariances are estimated using few samples. This
will lead to errors in the estimated channel vector and hence a reduction in the symbol en-
ergy γs, as shown in Chapter 4. Therefore, the analysis results obtained here can be regarded
as an upper bound on the system performance. 2
6.4 Probability of Error Analysis
In this section we derive the uncoded bit error probability with hard decision. In our anal-
ysis, we will use some of the results of the power control performance study that will be
presented in the next chapter. To derive the probability of error, without loss of generality,
let us assume that h = 1, i.e. the first Walsh symbol W1(t) is transmitted. Then the proba-
bility of symbol error is
Chapter 6. CDMA Base Station Receiver with Antenna Arrays 124
Finally, the corresponding bit error probability Pb(γs) is given by
Pb(γs) = 2J−1
2J − 1PM(γs) (6.4.6)
The symbol error probability and the corresponding bit error probability derived above
are conditional probabilities and are functions of γs, the symbol energy to interference plus
noise ratio. γs itself is a function of the channel vectors for the multipath components re-
ceived from the 1st user a1,1, · · · ,aL,1, shadowing and path loss S1, and the 1st user trans-
mitted power P1. Also note that because of power control (both open loop and closed loop),
P1, S1, and a1,1, · · · ,aL,1 are generally dependent variables. The dependency among these
variables is in general a function of the maximum Doppler shift fd of the first user. From
the simulation results in Chapter 7 and the results in [126, 127], a reasonable assumption
is that the combination of open loop and closed loop power control is perfect in eliminat-
ing the slow fading due to shadowing and path loss. Based on the mobile speed (or fd), we
consider the following different cases.
6.4.1 Low Doppler Frequency
At low Doppler frequencies ( fdTw ≤ 0.00125) and high diversity orders, fast closed loop
power control can eliminate most of the channel variation due to multipath fast fading. In
Chapter 6. CDMA Base Station Receiver with Antenna Arrays 125
the case of ideal power control, γs is a fixed quantity and is given by
γs = γ · L · K (6.4.7)
where γ is symbol energy to interference plus noise ratio per path per antenna. Also, the den-
sity function of z(n)1 for n = h given in (6.3.38) becomes an unconditional density. Therefore,
the symbol error probability is
PM = 1 −∫ ∞
0
[1 − e−z/σ2
L−1∑l=0
1l!
( zσ2
)l]M−1
· 1σ2
(z
σ2γs
) L−12
e−γsσ2+z
σ2 IL−1
(2
√γszσ2
)dz (6.4.8)
However, due to the delay in the control loop, finite step size by which the mobile can in-
crease or decrease its power, and errors on the downlink, power control can not be ideal (see
Chapter 7). Therefore the symbol error probability obtained above needs to be averaged
over the probability density function of γs, which is not known. However, an approxima-
tion to the bit error probability can be obtained as follows. First, let Cv denote the coefficient
of variation of γs, defined as
Cv =√
Var{γs}E{γs} (6.4.9)
The average symbol error probability, denoted by PM, is an expectation of a real function
of a random variable, namely γs, so that
PM = E {PM(γs)} =∫ ∞
0PM(γ) fγs (γ) dγ (6.4.10)
where PM(γs) is given by (6.4.8) and fγs(γ) is the probability density function of γs. The
objective here is to avoid carrying out the integration in computing (6.4.10), which requires
the knowledge of fγs(γ). Here, we use the results of [128] which are outlined below to
obtain PM. PM(γs) is expanded in terms of a Taylor’s series expansion, so that
PM(γs)= PM(γs)+ (γs − γs)P′M(γs)+ (γs − γs)
2
2!P′′
M(γs)+· · ·+ (γs − γs)n
n!Pn
M(γs)+ Rn
(6.4.11)
Chapter 6. CDMA Base Station Receiver with Antenna Arrays 126
where γs is mean value of γs, PnM(γs) is the nth derivative of PM(γs) (assuming they exist),
and Rn is a remainder that vanishes as n becomes large. Taking expectations of both sides
of (6.4.11) and ignoring terms beyond the second order term yields
PM ≈ PM(γs)+σ2γ
2P′′
M(γs) (6.4.12)
where σγ is the standard deviation of γ. We can easily see that σγ also represents the power
control error. However, in order to evaluate (6.4.12), we need the second derivative of
PM(γs), which is very difficult to obtain. Instead, we expand PM(γs) in terms of central
differences [37]; then taking expectation and ignoring terms beyond the second order term
yields
PM ≈ PM(γs)+σ2γ
2PM(γs + h)+ PM(γs − h)− 2PM(γs)
h2(6.4.13)
Choosing h = √3σγ is shown to give good accuracy [128]. The results in [125] show that
this approximation is accurate for low coefficient of variations Cv. Therefore, a reasonable
approximation of the symbol error probability is
PM ≈ 23
PM(γs)+ 16
PM(γs +√
3σγ )+ 16
PM(γs −√
3σγ) (6.4.14)
where γs is the mean symbol energy to interference-plus-noise ratio and σγ is the power
control error. Then, the corresponding bit error probability for small Cv (less than 0.3 [125])
is
Pb = 2J−1
2J − 1PM (6.4.15)
6.4.2 High Doppler Frequency
For high Doppler frequency and/or long loop delay, the fading statistics of the received sig-
nal after power control remain the same as those of the multipath fast fading with only per-
fect average power control (see Chapter 7). In this case we have
γs = γ
L∑l=1
|al,1|2. (6.4.16)
Chapter 6. CDMA Base Station Receiver with Antenna Arrays 127
The distribution of γs depends on the angle spread� through al,1. We consider the following
three cases.
• Small angle spread - For zero (or relatively small) angle spread�, the channel vector
of the lth multipath component can be expressed as
al,1 ≈ αl,1vl,1 (6.4.17)
where αl,1 is a zero mean complex Gaussian random variable and for a ULA vl,1 is a
Vandermonde vector [91] given by
vl,1 = [1 e− jπ sin θl,1 D/λ · · ·e− jπ sin θl,1 D(K−1)/λ]T (6.4.18)
In this case, we can show that γs has a χ2 distribution with 2L degrees of freedom.
That is
fγs(γ) =γL−1
(γK)L(L − 1)!e−γ/(γK) (6.4.19)
Hence, it can be shown that the corresponding unconditional probability density func-
tion of z(n)1 for h = n is (see Appendix B)
fz(n)1(z) = zL−1
σ2L(1 + γK)L(L − 1)!e−z/(σ2(1+γK)) (6.4.20)
• Large angle spread - For large angle spread, the elements of al,1 becomes uncorre-
lated and henceL∑
l=1|al,1|2 is a sum of K L i.i.d. random variables having a χ2 distribu-
tion with 2 degrees of freedom. Therefore, γs is distributed as a χ2 random variable
with 2K L degrees of freedom
fγs (γ) =γK L−1
(γ )K L(K L − 1)!e−γ/γ (6.4.21)
The corresponding unconditional probability density function of z(n)1 for h = n is (see
Appendix B)
fz(n)1(z) =
(K−1)L∑l=0
RlzK L−1+l
(K L − 1 + l)!e−z/(σ2(1+γ)) (6.4.22)
Chapter 6. CDMA Base Station Receiver with Antenna Arrays 128
where
Rl = 1(σ2(1 + γ ))K L−l
((K − 1)L
l
)(γ
1 + γ
)(K−1)L−l( 11 + γ
)l
(6.4.23)
• Other values of angle spread - For other values of �, we can easily show (see Ap-
pendix B) that the symbol energy to interference plus noise ratio is
γs =L∑
l=1
K∑i=1
γli|uli|2 (6.4.24)
where u11 · · · uLK are i.i.d. zero mean complex Gaussian random variables and γli =γλli,
{λl,i
}i=1,···,K are the eigenvalues of Rs,l,1, the spatial correlation matrix of the 1st
mobile’s lth multipath component defined in (2.2.17). Let {γli}l=1···L,i=1···K be equal
to {γi}i=1···K L. Also, we assume that the γli’s are distinct (this is true if the angles of
arrival are sufficiently different). Then, γs is distributed as [33]
fγs(γ) =K L∑i=1
πi
γie−γ/γi (6.4.25)
where
πk =LK∏i=1i�=k
γk
γk − γi, k = 1, · · · , LK (6.4.26)
and the corresponding unconditional probability density function of z(n)1 for h = n is
(see Appendix)
fz(n)1(z) =
K L∑i=1
πi
{e−z/(σ2(1+γi))
(σ2(1 + γi))$ g(z)
}(6.4.27)
where
g(z) = zL−2
σ2(L−1)(L − 2)!e−z/σ2
(6.4.28)
and $ denotes the convolution operation.
Using the unconditional pdf of z(n)1 for h = n in Equations (6.4.20), (6.4.22), and (6.4.27),
the average symbol error probability is given by
PM = 1 −∫ ∞
0
[1 − e−z/σ2
L−1∑l=0
1l!
( zσ2
)l]M−1
fz(1)1(z)dz (6.4.29)
and the corresponding average bit error probability is given by Equation (6.4.15).
Chapter 6. CDMA Base Station Receiver with Antenna Arrays 129
∆
∆
θ d d
d
d=8λ
Figure 6.5: Simulation scenario
6.5 Numerical and Simulation Results
First, we study the accuracy of the approximation that the MAI signal vector can be as-
sumed to be a spatially white complex Gaussian random vector. The base station receiver in
Figure 6.2 was simulated. In our simulation we assumed that the processing gain G = 256,
L = 4, N = 40, M = 64 and ν= 0.375. We also assumed ideal power control. We assumed
that the base station has 3 sectors, each with a 5 element ULA as shown in Figure 6.5. The
angle of arrivals {θk,i} were assumed to be uniform over [0,120◦]. The angle spreads {�k,i}were assumed uniform over [0,60◦]. The results of 10000 post-correlation signal vectors
were used to estimate the statistics of the MAI signal vector. Figures 6.6 and 6.7 show the
empirical PDF of both the I and Q component of the MAI at the first antenna. From both
figures we can see the validity of the Gaussian approximation. Also, the spatial correlation
Chapter 6. CDMA Base Station Receiver with Antenna Arrays 130
Interference Signal Level
-20 -10 0 10 20
PDF
0.00
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
SimulationsNormal Fitting
Figure 6.6: I-channel: first antenna interference distribution
matrix of the MAI vector R(n)uu,k,1 was estimated as
decision device followed by a deinterleaver and a Viterbi convolutional decoder. The re-
ceiver uses the output bit sequence and other information from the convolutional decoder
(such as the branch metrics) to get an estimate of the frame error rate (FER). Based on this
estimated FER, the receiver selects the appropriate threshold level to be used. Figure 7.2
shows a block diagram for the power control algorithm.
Chapter 7. Performance of Power Control in CDMA 145
Λ(1)
Λ(2)
Λ(J)
b b . . . b(2)(1) (J)
ThresholdET
(1)
(2)
(M)
M-ary
Decision
Symbol-by-Symbol. . . .
J = log M
Convolutional
Decoder
z
z
z
Measure FER
µ2 =1 Σ
j=1ΣM
max
1 max= Σj=1 M
{T
TET
Power Control
Up/Down
Command
Signal Strength Based Power Control Algorithm
T (M-1)
pT
p1
1
1
1
12µ}
z (j)
z (j)p
(n)
n=1n=n
(n)
2
Deinterleaver
Figure 7.2: Power Control Algorithm
7.4 Simulation Results
To study the CLPC performance, a single cell CDMA system was simulated. We assumed
the same base station antenna array model used in the previous chapter. Signals received at
the base station were generated using the channel and transmitted signal models described in
Chapters 2 and 6, respectively. The uncoded bit rate was assumed to be 9.6 kbits/sec. With
a rate 1/3 convolutional code and M-ary orthogonal modulation with M = 64, the symbol
rate is 4800 symbols/sec. The processing gain G is 256. We assumed that a power control
command is sent every 1.25 ms, i.e every 6 Walsh symbols (in which case Tp = 6). Unless
otherwise specified, we assumed that the total loop delay (including power measurement)
is 1 Tp, the forward link error rate π is 0.05, the power step size �p is 0.5 dB (these are the
same values used in [127, 131]), and that the angle spread � is zero.
Chapter 7. Performance of Power Control in CDMA 146
Symbol Period Tw
0 200 400 600 800 1000
Sign
al L
evel
(dB
)
-5
0
5
10
15
Multipath Fast Fading
Power-Controlled Signal
(a) Signal waveforms
Normalized Signal Level (dB)
-4 -2 0 2 4
Pr
{Sig
nal L
evel
≤ A
bsci
ssa}
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
Power-Controlled Signal LevelMultipath Fast Fading
(b) Signal level distributions
Figure 7.3: Power-controlled received signal vs. simulated Rayleigh fading: fd = 5 Hz,K = 5, L = 4.
Chapter 7. Performance of Power Control in CDMA 147
Symbol Period Tw
0 200 400 600 800 1000
Sign
al L
evel
(dB
)
-5
0
5
10
15
Multipath Fast FadingPower-Controlled Signal
(a) Signal waveforms
Normalized Signal Level (dB)
-4 -2 0 2 4
Pr
{Sig
nal L
evel
≤ A
bsci
ssa}
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
Power-Controlled Signal LevelMultipath Fast Fading
(b) Signal level distributions
Figure 7.4: Power-controlled received signal vs. simulated Rayleigh fading: fd = 100 Hz,K = 5, L = 4.
Chapter 7. Performance of Power Control in CDMA 148
Maximum Doppler Shift fd (Hz)
20 40 60 80 100
Pow
er C
ontr
ol E
rror
σE (
dB)
0
1
2
3
4
5
6
L = 1L = 2L = 4
Figure 7.5: Power control error vs. number of paths: K = 5.
First, we study the effect of closed loop power control on the received signal level statis-
tics. Figures 7.3 and 7.4 show the RAKE output received signal level waveform and the es-
timated distribution versus the simulated multipath fast fading for fd = 5 Hz and fd = 100
Hz, respectively. From these two figures we can easily see that closed loop power control
eliminates most of the channel variations due to fast fading at low fd, while at high fd the
received signal statistics after power control remain almost the same as that of the simulated
multipath fast fading with only average power control. This result is consistent with the re-
sults in [127]. In addition, we may note that in [127, 115], it was suggested that for high fd
the signal level distribution can be approximated by a log-normal distribution.
Chapter 7. Performance of Power Control in CDMA 149
Maximum Doppler Shift fd (Hz)
20 40 60 80 100
Pow
er C
ontr
ol E
rror
σE (
dB)
0
1
2
3
4
∆p = 0.25 dB∆p = 0.5 dB∆p = 1 dB
Figure 7.6: Power control error vs. power step size: L = 2,K = 5.
Figure 7.5 shows the effect of the number of RAKE branches L on the power control
error σE. It can be seen from this figure that σE is reduced with higher diversity order. We
can also note that with high diversity order and low fd, CLPC can eliminate most of the
channel variation. Figure 7.6 shows the effect of step size �p on the power control error
σE for �p = 0.25, 0.5, and 1 dB. We can see that σE is lower with large step size �p and
high fd, while at low fd, a small step size provides more precise control. The reason is that
at high fd the fading rate is too high and a large step size is necessary to track the fading.
On the other hand, at low fd the fading rate is slow enough to allow the control loop to
track the fading with a small step size. Figure 7.7 shows the effect of loop delay kTp for
Chapter 7. Performance of Power Control in CDMA 150
Maximum Doppler Shift fd (Hz)
20 40 60 80 100
Pow
er C
ontr
ol E
rror
σE (
dB)
0
1
2
3
4
Loop Delay = 1 Tp
Loop Delay = 2 Tp
Loop Delay = 3 Tp
Figure 7.7: Power control error vs. loop delay: L = 2,K = 5.
k = 1,2, and 3. We can see from the figure that the longer the loop delay is, the poorer σE.
Figure 7.8 shows the effect of return channel error rate π on σE and it suggests the CLPC is
robust against return channel errors. Again, these results are consistent with the results in
[127, 132].
Next we look at the effect of angle spread and number of antennas in the array. As in the
previous chapter, we assumed that the array size is 8λ and the sensors are placed as a ULA.
In Figures 7.9 and 7.10 we plot the power control error σE against fd for different values of
� for K = 5 and 9. From these figures we can make the following observations. With zero
angle spread �, the number of antennas has no effect on the CLPC error σE. This is due
Chapter 7. Performance of Power Control in CDMA 151
Maximum Doppler Shift fd (Hz)
20 40 60 80 100
Pow
er C
ontr
ol E
rror
σE (
dB)
0
1
2
3
4
π = 0.05π = 0.1
Figure 7.8: Power control error vs. forward link error rate: L = 2, K = 5.
to the fact that with zero angle spread �, the received signal in any multipath component
will have the same fading at each antenna. Thus, the antenna array will not provide any
space diversity for this multipath component. Also, for a given number of antennas K, as
the angle spread � increases, the gain due to space diversity increases which will lead to a
reduction in the power control error σE (i.e. better CLPC performance).
Chapter 7. Performance of Power Control in CDMA 152
Maximum Doppler Shift fd (Hz)
20 40 60 80 100
Pow
er C
ontr
ol E
rror
σE (
dB)
0
1
2
3
∆ = 0ο
∆ = 5ο
∆ = 10ο
∆ = 40ο
∆ = 60ο
Figure 7.9: Power control error vs. angle spread: L = 4, and K = 5.
7.5 Summary
In this chapter, we studied the closed loop power control performance of the reverse link
in wireless DS/CDMA through a combination of analysis and discrete event simulation.
The closed loop power control mechanism is modeled as a first-order discrete time feed-
back control loop. Based on this model, we simulated the closed loop power control on the
communication link from a CDMA mobile to the base station. Several conclusions can be
drawn from our simulation results. Our simulation results show that with high diversity or-
der (either path or space) and low Doppler frequencies ( relative to the power control rate),
Chapter 7. Performance of Power Control in CDMA 153
Maximum Doppler Shift fd (Hz)
20 40 60 80 100
Pow
er C
ontr
ol E
rror
σE (
dB)
0
1
2
3
∆ = 0ο
∆ = 5ο
∆ = 10ο
∆ = 40ο
∆ = 60ο
Figure 7.10: Power control error vs. angle spread: L = 4, and K = 9.
closed loop power control can eliminate most of the channel variations, and with small an-
gle spread adaptive beamforming with antenna arrays has a little effect on the power control
error.
Chapter 8
Conclusions
The main focus of this thesis has been the development and application of advanced array
signal processing techniques to CDMA wireless systems that have a practical implementa-
tion complexity and achieve high performance levels. This work has been prompted by the
current thrust in wireless communication technology to look for new approaches and tech-
nologies to improve spectrum efficiency and to be able to support the projected capacity
demands with the introduction of new personal communication services.
8.1 Thesis Summary
The second half of Chapter 2 and Chapters 3, 4, 6, and 7 presented the thesis contributions.
In our work, we focused on the mobile to base station link and assumed that only the base
station uses an antenna array for transmission and reception of signals to and from the mo-
bile. The first half of Chapter 2 provides a brief overview of wireless radio channels and
statistical models for scalar (single antenna) channels. In the second half of this chapter, we
develop a statistical vector (multiple antennas) channel model based on the physical prop-
agation environment and derive its statistical properties. This is necessary in order analyze
the new array signal processing and beamforming techniques proposed later on in the thesis.
154
Chapter 8. Conclusions 155
In Chapter 3, we start by reviewing general techniques for adaptive beamforming and
discuss the reasons why such techniques are not suitable for CDMA wireless mobile sys-
tems. We then present a new space-time code-filtering approach for channel vector esti-
mation and optimum beamforming. This approach exploits both the temporal and spatial
information in the received signal. A key result in this thesis is the derivation of the space-
time matched filter receiver for multipath signals over AWGN channels. This receiver con-
sists of a front-end spatial matched filter, or beamformer, followed by a temporal matched
filter. Based on the space-time matched filter, we construct a Beamformer-RAKE receiver
for CDMA signals. Performance analysis results of the Beamformer-RAKE receiver show
improved bit error rate performance due to the exploitation of the spatial dimension in the
received signal. In Chapter 4, we consider the time-variant nature of the wireless channel
and develop a recursive algorithm for tracking the channel vector and estimating the beam-
former weight vector. The algorithm has a reasonable computational complexity of O(K2)
per path per user. Simulation results show that the proposed algorithm can track the channel
vector closely even in severe propagation conditions.
Chapters 6 and 7 look at the application of the array signal processing techniques derived
in the previous chapters to the existing CDMA IS-95 cellular standard which we briefly de-
scribe in Chapter 5. In Chapter 6 we propose an overall base station receiver architecture
based on the Beamformer-RAKE structure and describe different signal processing func-
tions. We then study the performance of the proposed base station receiver in terms of the
uncoded bit error probability as a function of loading (number of users) and number of an-
tennas. Under the propagation conditions considered, the analysis results show a perfor-
mance improvement in terms of the number of users that can be supported due to the use
of antenna arrays and the associated signal processing. These analysis results make use of
the results of Chapter 7, where we look at modeling and performance of closed loop power
control. Many conclusions can be drawn from the simulation results described in Chap-
ter 7. The most important result is that beamforming reduces the power control error when
Chapter 8. Conclusions 156
the angle spread is non-zero.
8.2 Future Directions
Although this thesis has answered several questions related to the application of array signal
processing techniques to CDMA wireless systems, several questions remain open including:
• The Beamformer-RAKE receiver in Chapter 3 is not an optimum structure when mul-
tiple access interference is present. Therefore, optimal Beamformer-RAKE structures
which include multiuser detection and interference cancellation is an interesting fu-
ture research topic.
• The performance analysis of the proposed base station receiver for wireless CDMA
did not include the effects of forward error correction and interleaving, nonuniform
power-delay profiles, and the effect of channel vector estimation errors on the overall
receiver performance. Extension of this performance analysis to include these effects
will also be another interesting future research direction.
• The base station must also beamform on the forward link in order to effectively im-
prove overall system performance. However, since both forward and reverse links
use different frequency bands and, therefore, the multipath vector channels for both
links are quite different, the base station cannot use the reverse link beamformer for
forward link transmission. Hence, forward link beamforming is also an interesting
and important future research direction.
Appendix A
Multipath Fading Correlation
A.1 Fading Correlation for Scalar Channels
In this Appendix, we derive the time-frequency correlation of the multipath fast fading for
scalar channels. First, we recall the definition of β(ω, t) from (2.1.13) as
β(ω, t) =L∑
i=1
Riejφi(t) (A.1.1)
where φi = ωd cosψit −ωτi and the φi modulo 2π are assumed to be i.i.d. and uniform over
[0,2π]. We assume that the angular position of the ith scatterer ψi is uniformly distributed
over [0,2π]. We also assume that the time delays τi τi are i.i.d. with probability density
function fT(τ), where fT(τ) is nonzero for 0 ≤ τ <∞ and zero otherwise. Then, the time-
frequency correlation of β(ω, t) can be derived as follows:
ρβ(ω1, ω2, t, t + ν) = E{β(ω1, t)β∗(ω2, t + ν)} (A.1.2)
= E
{L∑
i=1
L∑l=1
Ri Rlej(φi(ω1,t)−φl(ω2,t+ν))
}(A.1.3)
The average will vanish unless i = l. In this case φi(ω1, t)− φi(ω2, t + ν) = ωd cosψiν−�ωτi, where �ω = ω1 − ω2. Therefore
ρβ(ω1, ω2, t, t + ν) = ρβ(�ω, ν) (A.1.4)
157
Appendix A. Multipath Fading Correlation 158
ρβ(ω1, ω2, t, t + ν) =∑
i
E{R2i }e j(ωd cosψiν−�ωτi) (A.1.5)
E{R2i } represents the average fraction of incoming power in the ith path, which can be rewrit-
ten as
E{R2i } = σ2 f&(ψi) fT(τ)dψidτi (A.1.6)
where σ2 is the total radiated from the mobile, which will be assumed to equal 1, and the
term f&(ψi) fT(τi)dψidτi represents the average fraction of incoming power within dψi of
angle ψi and within dτi of the time τi. If we assume that L is large (i.e. L → ∞), then we
can express the sum in (A.1.5) with integrals, independent of i,
ρβ(�ω, ν) = 12π
∫ 2π
0
∫ ∞
0e j(ωd cosψν−�ωτ) fT (τ)dψdτ (A.1.7)
= 12π
∫ 2π
0e j(ωd cosψν)dψ ·
∫ ∞
0e− j�ωτ fT(τ)dτ (A.1.8)
= Jo(ωdν) ·∫ ∞
0e− j�ωτ fT(τ)dτ (A.1.9)
= Jo(ωdν) · FT( j�ω) (A.1.10)
where FT(s) is the characteristic function of the time delay τ defined as
FT(s) =∫ ∞
0e−sτ fT(τ) dτ (A.1.11)
We notice that FT( j�ω) is also the Fourier transform of the probability density function
fT(τ).
A.2 Fading Correlation for Vector Channels
In this Appendix, we derive the space-time-frequency correlation for the multipath vector
channel. The derivation follows the same steps and reasoning used in the preceding ap-
pendix. Again, we start by recalling the definition of the channel vector
a(ω, t) =L∑
i=1
v(θi)Riejφi(t) (A.2.1)
Appendix A. Multipath Fading Correlation 159
In addition to the assumptions used in the preceding Appendix, we also use the assumption
from (2.2.9) that the angle of arrival of the ith path θi is uniformly distributed over [−�+θ,�+ θ] where θ is the mean angle of arrival and 2� is the angle spread. Then, we can
write Ra, the space-time-frequency correlation matrix of the channel vector, as