Waveform and Receiver Filter Selection

for Wideband Radar Applications

Marie Ström

Department of Signals and Systems

chalmers university of technology
Göteborg, Sweden 2015
 Thesis for the degree of Doctor of Philosophy

Waveform and Receiver Filter Selection

for Wideband Radar Applications

by

Marie Ström

Department of Signals and Systems

Chalmers University of Technology
Göteborg, Sweden 2015
Waveform and Receiver Filter Selection for Wideband Radar Ap-
plications
Marie Ström
ISBN 978-91-7597-157-5

This thesis has been prepared using LATEX.

Copyright c Marie Ström, 2015.

All rights reserved.

Doktorsavhandlingar vid Chalmers Tekniska Högskola

Ny serie nr 3838
ISSN 0346-718X

Department of Signals and Systems

Signal Processing Group
Chalmers University of Technology
SE-412 96 Göteborg, Sweden

Phone: +46 (0)31 772 3713

Fax: +46 (0)31 772 1748
E-mail: [email protected]

Printed by Chalmers Reproservice

Göteborg, Sweden, September 2015
To my family
Waveform and Receiver Filter Selection for Wideband
Radar Applications
Marie Ström
Department of Signals and Systems
Chalmers University of Technology

Abstract
This thesis concerns the design of transmitter-receiver chains for wideband
radar systems. The transmitter side employs one, or several, highly flexible
signal generators, which are able to generate signals with a large bandwidth.
At the receiver side, when we are are able to select receiver filters, we have
the freedom to optimize also the receiver filters.
Herein, the transmit waveforms and receiver filters are designed to fulfil
user-defined criteria. In general, a high probability of target detection, while
maintaining a low false alarm rate, is desired. For a scenario in which interfer-
ence is present, this means to achieve a high Signal-to-Interference-and-Noise
Ratio.
When advanced transmitter-receiver technology is implemented, the pos-
sibility to adapt the system through a feedback loop arises. Information
about the the radar operating environment is provided by signal processing
techniques. We propose a Kalman filter to follow a time-evolving clutter-
map, based on the complex received signal samples. The estimates of the
complex clutter reflections are utilized to determine parameters of the clutter
distribution.
The system should, in addition, experience a robust target detection prop-
erty. This is important when targets are not confined on a user-specified grid
of time-delays and time-scalings. We derive an algorithm where the mainlobe
width of the correlation function is adapted according to a desired resolution.
The thesis also deals with hardware restrictions. A study on how to syn-
thesize time domain signals from achieved power spectra is performed. We
synthesize signals with given spectral properties that experience a low peak-
to-average-power ratio. A signal with constant envelope is also achievable by
allowing the power spectrum to deviate somewhat from its desired shape.

Keywords: Wideband radar, detection, waveform design, receiver filter de-

sign, performance evaluation, interference suppression.

i
ii
Acknowledgements

I would like to take the opportunity to thank some very special people,
without their help and support this thesis would never have been written.
First, I would like to thank my supervisor Mats Viberg, for accepting me as
a Ph.D. student, for always believing in me, and for all the support you have
given me during these years.
Thanks to my co-supervisor Kent Falk at Saab EDS, for much needed
guidance in theoretical and practical knowledge of radar systems. I have
enjoyed both our discussion on research and on exploring the world. A special
thanks to Johan Carlert, who accepted me as an Industrial Ph.D student at
Saab EDS. I look forward spending many years working with you.
There are a number of people at the Department of Signals and Systems
who deserve a special thanks. The administrate staff for helping out with all
non-technical issues, and Lars for computer related support. Also, everyone
at the Signal Processing group is acknowledged for giving me an enjoyable
working environment. Thanks to Ashkan for many interesting discussions,
and to Daniel for all the help with grammatical issues.
Thanks to Dr Jian Li, for accepting my research visit at University of
Florida in Gainesville, USA. Also, thanks to my colleagues and friends in
Gainesville for making my visit a great adventure.
All of my love to my friends: Lina, Jenni, Eija, Irena, Charlotta, Louise,
and Ulrika. Not to forget, all of my climbing friends for giving me something
else to think about when I have been overloaded with work.
Finally, I would like to express my deepest love to my mother Gunilla,
my father Gunnar, and their partners Stefan and Aina. To my step sister
Elsa, her husband Fredrik, and to my step brother Nils. Foremost, I would
like to thank my sister Anna for all the great support and encouragement
during many years. Also, for making me realize the importance of family.

iii
Marie Ström
Göteborg, February 19, 2015

iv
List of Publications

This thesis is based on the following publications:

Paper I
M. Ström, M. Viberg, and K. Falk. Robust Transceiver Design for Wideband
MIMO Radar utilizing a Subarray Antenna Structure. In Special Issue on
Advances in Sensor Array Processing EURASIP Signal Processing Journal,
vol. 93, 2013.

Paper II
M. Ström and M. Viberg, Low PAPR Waveform Synthesis with Application
to Wideband MIMO Radar. In Proc. of the 4th International Workshop on
Computational Advances in Multi-Sensor Adaptive Processing , December
2011, San Juan, Puerto Rico.

Paper III
A. Panahi, M. Ström and M. Viberg, Wideband Waveform Design for Ro-
bust Target Detection. To be published in IEEE International Conference
on Acoustics, Speech, and Signal Processing , April 2015, Brisbane, Australia.

Paper IV
M. Ström, M. Viberg, and K. Falk. Wideband Waveform and Receiver Filter
Bank Design for Clutter Suppression. Submitted to JSTSP SI on Advanced
Signal Processing Techniques for Radar Applications.

Paper V
M. Ström, D. Svensson, and M. Viberg. Waveform and Receiver Filter Se-
lection for Clutter-Map Estimation Based on an IMM Kalman Filter. To be
submitted to IEEE Transactions on Aerospace and Electronic Systems.

v
vi
Other publications by this author, but omitted in the thesis:

• M. Ström, A. Panahi, M. Viberg, K. Falk. Wideband Waveform De-

sign for Clutter Suppression. In In 8th IEEE International workshop
on Sensor array and multichannel signal processing, June 2014, La
Coruna, Spain.

• A. Panahi, M. Ström, M. Viberg. Basis Pursuit over Continuum Ap-

plied to Range-Doppler Estimation Problem. In In 8th IEEE Inter-
national workshop on Sensor array and multichannel signal processing,
June 2014, La Coruna, Spain.

• W. Rowe, J. Li, M. Ström. Frequency division orthogonal waveforms

for narrowband MIMO arrays. In IEEE International Symposium on
Phased Array Systems and Technology, October 2013, Waltham, USA.

• W. Rowe, M. Ström, J. Li, P. Stoica. Robust adaptive beamforming for

MIMO monopulse radar. In Proceedings of SPIE - The International
Society for Optical Engineering, April 2013, Baltimore, USA.

• M. Ström, M. Viberg, and K. Falk. Transmit and Receive Filter Opti-

mization for Wideband MIMO Radar. In Proc. of the 4th International
Workshop on Computational Advances in Multi-Sensor Adaptive Pro-
cessing , December 2011, San Juan, Puerto Rico.

• E. Johansson, M. Ström, L. Svensson and M. Viberg. Interpolation

based on stationary and adaptive AR(1) modeling. In IEEE Interna-
tional Conference on Acoustics, Speech, and Signal Processing, May
2011, Prague, Czech Republic.

• M. Ström, E. Johansson, and D. Stork. Mapping Colors from Paintings

to Tapestries: Rejuvenating the Faded Colors in Tapestries based on
Colors in Reference Paintings. In SPIE Electronic Imaging: Human Vi-
sion and Electronic Imaging XVII, January 2012, San Fransisco, USA.

• M. Ström and M. Viberg. Constant Modulus Waveform Synthesis with

Application to Wideband MIMO Radar. In Conference Presentation at
Swedish Radio and Microwave Days, March 2012, Stockholm, Sweden

vii
Contents

Abstract i

Acknowledgments iii

List of Publications v

Contents viii

Part I: Introduction 1
1 Introduction 3
1.1 Research Project and Support . . . . . . . . . . . . . . . . . . 5
1.2 Outline of the Thesis . . . . . . . . . . . . . . . . . . . . . . . 5

2 Introduction to Radar 7
2.1 Radar Signal Modeling . . . . . . . . . . . . . . . . . . . . . . 7
2.2 Radar Detection Fundamentals . . . . . . . . . . . . . . . . . 10
2.3 Narrowband Ambiguity Function . . . . . . . . . . . . . . . . 14
2.4 Estimation Fundamentals . . . . . . . . . . . . . . . . . . . . 16
2.4.1 Maximum-Likelihood Estimation . . . . . . . . . . . . 17
2.4.2 Least-Squares Estimation . . . . . . . . . . . . . . . . 18
2.4.3 Minimum-Mean-Squared-Error Estimation . . . . . . . 19
2.4.4 Wiener Filter . . . . . . . . . . . . . . . . . . . . . . . 20
2.4.5 Kalman Filter . . . . . . . . . . . . . . . . . . . . . . . 21
2.5 Antenna Array Beamforming . . . . . . . . . . . . . . . . . . 22
2.6 Radar Operating Environment . . . . . . . . . . . . . . . . . . 25
2.6.1 Clutter Interference . . . . . . . . . . . . . . . . . . . . 26
2.6.2 Jamming Interference . . . . . . . . . . . . . . . . . . . 27

viii
3 Wideband Models 29
3.1 Wideband Ambiguity Function . . . . . . . . . . . . . . . . . 29
3.2 Wideband Correlation Processing . . . . . . . . . . . . . . . . 30
3.2.1 Wavelet Correlation Processing . . . . . . . . . . . . . 31
3.2.2 Mismatched Filter Bank . . . . . . . . . . . . . . . . . 32
3.3 Wideband Antenna Array Beamforming . . . . . . . . . . . . 33

4 Waveform and Receiver Filter Design 37

4.1 Waveform Diversity . . . . . . . . . . . . . . . . . . . . . . . . 38
4.1.1 Waveform Design Utilizing Spatial Properties of the
Transmitted Signals . . . . . . . . . . . . . . . . . . . . 39
4.1.2 Waveform Design Utilizing Temporal Properties of the
Transmitted Signals . . . . . . . . . . . . . . . . . . . . 40
4.2 Optimization Methods for Waveform and Receiver Filter Design 41
4.2.1 Semidefinite Relaxation . . . . . . . . . . . . . . . . . 42
4.2.2 Bisection Method . . . . . . . . . . . . . . . . . . . . . 43
4.2.3 Gauss-Newton Algorithm . . . . . . . . . . . . . . . . . 44
4.2.4 Second-Order-Cone Program . . . . . . . . . . . . . . . 45
4.2.5 Robust Design . . . . . . . . . . . . . . . . . . . . . . . 45

5 Waveform Synthesis 47
5.1 Waveforms Synthesis to Minimize Peak-to-Average-Power Ratio 47
5.2 Practical Implementation . . . . . . . . . . . . . . . . . . . . . 51

6 Contributions and Future Work 53

6.1 Summary of the Appended Papers . . . . . . . . . . . . . . . 53
6.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56

Bibliography 58

Part II: Publications 65

Paper I: Robust Transceiver Design for Wideband MIMO Radar
utilizing a Subarray Antenna Structure 69

Paper II: Low PAPR Waveform Synthesis with Application to

Wideband MIMO Radar 99

Paper III: Wideband Waveform Design for Robust Target De-

tection 112

ix
Paper IV: Wideband Waveform and Receiver Filter Bank De-
sign for Clutter Suppression 127

Paper V: Waveform and Receiver Filter Selection for Clutter-

Map Estimation Based on an IMM Kalman Filter 153

x
 Part I
Introduction

1
Chapter 1
Introduction

T
he strive of mankind to develop new technologies most certainly
started the day that humans set foot on Earth. Now, eons later, our
society still advocate a necessity for new products and technologies.
In this thesis, we discuss a rather new technology, nowadays so acknowl-
edged that the abbreviation is a commonly known word, namely radar, or
radio detection and ranging. One might ask what makes this technology
so popular? – Probably its usability in various applications. We encounter
radar technology in systems ranging from active-safety systems for cars and
trucks, through medical applications, such as cancer treatments, to military
and civil surveillance. The last two application are the ones that are focused
on in this thesis.
From a historical perspective, demonstration of the similarity between
radio waves and light, conducted by Heinrich Hertz in the late 19th century,
is generally seen as the start of the great advances in the area of remote
sensing. Hertz provided the world with the knowledge of reflection on metallic
surfaces, as well as refraction of radio waves in dielectric prisms. Hertz’
research was advanced by Christian Hülsmeyer, who in 1904 obtained the
first patent for a radar system that detected ships. However, mankind was
not ready for such a new, and advanced, technology, so it slowly faded into
people’s memories. However, in the 1920s, Gugliermo Marconi advocated
these ideas, and his speech delivered before the Institute of Radio Engineers
might be seen as the start-up of great developments in radar technology. The
research accelerated and spread throughout the world during the rest of the
20th century, mostly due to its use in military operations [1, 2].
It is probably impossible, and not fair at all, to sort the importance of
developed radar techniques. However, there are four great advances that
significantly improved the radar system performance, namely

3
 Chapter 1. Introduction

• Invention of the high-power microwave magnetron.

• Use of the Doppler effect.

• Technology of pulse compression.

• Electronically steered phased array.

Today, there is a hope among researchers that the possibility to arbi-

trary generate signals in combination with a diversity based transmission
scheme, will be the next item on the list of great advances. Diversity can,
for example, be introduced with a Multiple-Input-Multiple-Output (MIMO)
transmitter-receiver chain. The MIMO configuration first appeared for com-
munication applications, for which the system performance was dramatically
improved [3]. However, the underlying problems and objectives are quite dif-
ferent in communication and radar. Nonetheless, research so far have shown
that a diversity based antenna configuration, for radar, can improve, for in-
stance, target identifiability, and target location resolution [4, 5]. It is also
anticipated that MIMO radar, compared with traditional radar, will experi-
ence an improvement in difficult environments, which involve strong clutter,
jamming, and small targets.
Research on wideband systems has been a trend in hardware design for
several decades. However, traditional radar detection theory has to a great
extent focused on narrowband systems. The situation that arises is that
highly flexible wideband transmitters are available, but it is not well under-
stood how they should be used. When transmitting multiple arbitrary wide-
band waveforms, electronic surveillance equipment with classical libraries of
frequencies, pulse repetition intervals, and pulse lengths might become obso-
lete. Moreover, the possibility to design waveforms that improve the stealth
properties of the radar arises, meaning that a transmission unit will be more
difficult to discover, compared with the waveforms used nowadays.
In this thesis, we advocate employing multiple wideband waveforms in
combination with a receiver filter bank, which is possible to optimize. The
waveforms are simultaneously transmitted from different antennas, or groups
of antennas. Three different criteria, where an optimal selection of waveforms
and receive filters, which increase performance are introduced and evaluated.
First, robust waveforms and receiver filters are designed to maximize the
Signal-to-Interference-and-Noise Ratio (SINR) in a jamming environment.
Second, a new algorithm to design for robust target detection is proposed.
Third, we show how to design waveforms and receive filters with clutter
suppression capabilities, and how to adapt the design to a time-evolving

4
1.1 Research Project and Support

scenario. The adaptation is based on estimates of target and clutter charac-

teristics. The estimation of the clutter environment is further improved with
an interacting-multiple-model Kalman Filter. From the estimated complex
clutter reflection coefficients, the clutter distribution parameters are followed
in time.
In the first problem, maximization of the SINR is performed in the fre-
quency domain, and as several time domain waveforms experience the same
spectral properties, we introduce a method to synthesize waveforms that, in
addition, experience desirable time domain properties. This is an important
and fundamental interest in a broad variety of applications, and in this thesis
we focus on the design of time domain signals with a low peak-to-average-
power ratio, or even a constant envelope.

1.1 Research Project and Support

The research that has led up to the writing of this thesis has mainly been
performed within two consecutive research projects. The first part is within
the Swedish research program “Chalmers Antenna Systems Excellence cen-
ter” (CHASE) – a program financed by Vinnova (The Swedish Governmen-
tal Agency for Innovation Systems). The second project, called “Waveform
Diversity in Wideband MIMO Radar” – is a project financed by Veten-
skapsrådet (The Swedish Research Council). This project is a collaboration
between Chalmers University of Technology and Saab AB.

1.2 Outline of the Thesis

This thesis is divided into two parts: In the first part, the theoretical back-
ground of the thesis work is presented, with the purpose of introducing the
topic and of preparing the reader for its second part. In the second part,
the contributions of the author to the field of waveforms and receiver filters
selection are presented in the form of five appended papers.
The first part of the thesis is structured as follows. In Chapter 2, the
concepts of radar signal modeling, detection and estimation fundamentals,
including the narrowband ambiguity function, beamforming, and the radar
operating environment are introduced. Chapter 3 considers the receiver func-
tion in a wideband radar system and wideband beamforming. In Chapter 4,
diversity based design of waveforms and receiver filters is introduced. This
chapter also introduces different optimization criteria, and solution methods
which are utilized in the appended papers. Next, Chapter 5 describes meth-

5
 Chapter 1. Introduction

ods to synthesize time domain signals, and a practical experiment conducted

at Saab AB is summarized. Finally, in Chapter 6 the contributions of this
thesis are presented, together with suggestions for future work and directions
within the area of waveforms and receiver filters design.

6
Chapter 2
Introduction to Radar

T
his chapter provides a short introduction to radar fundamentals from
a signal processing point of view. For a comprehensive study see, for
example, [1, 2, 6]. The basic concept of a radar system, or a sonar
system if sound waves are used, is to transmit a known signal, and to process
its return to locate and identify objects in a surrounding area. If the echoes
are correctly processed, accurate information, necessary to characterize ob-
jects in the environment, is retrieved. This information commonly includes
the object’s range, velocity, reflectivity, and spatial position, i.e., angle of
azimuth and elevation.
In this chapter, we briefly discuss upon radar signal modeling, detection
and estimation fundamentals, including the narrowband ambiguity function,
beamforming, and the radar operating environment.

2.1 Radar Signal Modeling

Let the transmitter emit a known signal x(t), the signal is reflected by an
object and sampled at a receiver. The received signal is denoted by y(t).
This process, as an easy principle, is illustrated in Figure 2.1.
The distance to a scatter, R, is proportional to the time it takes for a
signal to propagate from the transmitter to the receiver. This time is called
time-delay and is, for a monostatic radar, expressed by
2R
τ= , (2.1)
c
where c is the speed of the wave in the medium of propagation.
Equation (2.1) is valid when the object and the radar system are sta-
tionary. If the object and/or the system are/is in motion, the time-delay is

7
 Chapter 2. Introduction to Radar

Transmitter
Echo signal
Transmitted signal
Receiver

Range

Figure 2.1: Basic geometry of a radar system.

instead expressed with a time-varying variable, τ (t).

When the distance depends on time, at the time instance t − τ (t)/2, the
object is at the distance R(t − τ (t)/2), for which (2.1) is instead
2
τ (t) = R(t − τ (t)/2). (2.2)
c
If the object moves with a constant velocity, which is assumed in this thesis,
with respect to the radar system, say v0 , (2.2) is, through a Taylor series
expansion, described by
2v0
τ (t) = τ0 + (t − τ0 ). (2.3)
c + v0
Here, τ0 = 2R0 /c. The received signal is given by
 
c − v0
y(t) = x(t − τ (t)) = x (t − τ0 ) . (2.4)
c + v0

Normally the transmitted signal consists of a baseband envelope x̃(t) mod-

ulated on a carrier ejωc t , where ωc = 2πfc and fc is the carrier frequency.
Then, (2.4) is
 
c − v0
 
c−v
jωc c+v0 (t−τ0 )
y(t) = x̃ (t − τ0 ) e 0 . (2.5)
c + v0

Let µ0 = c−v 0
c+v0
, where µ0 describes a so-called time-scaling of the signal. Thus,
the received signal is mathematically expressed by
√
y(t) = µ0 x̃(µ0 (t − τ0 ))ejωc µ0 (t−τ0 ) . (2.6)

This signal model is called a wideband model [7–10]. Note that, in (2.6), no
attenuation of the transmitted signal is accounted for, and the normalization
√
term, µ0 , is introduced as an energy normalization between the transmitted

8
2.1 Radar Signal Modeling

µ =1 µ=2 µ = 0.5

Figure 2.2: Effect of the time-scaling, µ on a signal x(µt).

and the received signal. The effect of the time-scaling on a signal x(µt) is
visualized in Figure 2.2.
For many applications the wideband model is unnecessary complicated,
and the signal can instead be resembled by a narrowband model. To under-
stand this, assume that |v0 /c| < 1, then µ0 is expanded by

2v0  v 2
0
µ0 ≈ 1 − +2 + .... (2.7)
c c
Further, if |v0 /c|  1, then µ0 ≈ 1 − 2vc0 . For a signal with angular carrier
frequency ωc , the time-scaling is approximated with a Doppler shift, ωd0 =
− 2vc0 ωc . Thus, (2.4) is

y(t) ≈ x(t − τ0 )ejωd0 (t−τ0 ) , (2.8)

which imposes that all frequencies are equally shifted over the bandwidth.
The approximation in (2.8) has an error of order (v0 /c)2 . The narrowband
model is computationally efficient, as estimation of velocity is calculated from
a series of pulses with a Fast-Fourier Transform (FFT).
There are two occasions when the narrowband model fails [10]. First, if a
signal experiences a large fractional bandwidth, that is, when B/fc is large,
B being the bandwidth of the signal. Second, when an object significantly
changes position during the pulse duration, T . The second statement gives
the narrowband condition, i.e.,
2v0 1
 . (2.9)
c TB
This condition is either violated if the velocity of an object is large compared
to the propagation speed of the wave, or if the time-bandwidth product, T B,
is large.
Even though the narrowband model usually is valid, it is expected that
future radar systems operate at a higher bandwidth. Thus, care has to be
taken to guarantee correct receive signal-modeling and parameter estimation.

9
 Chapter 2. Introduction to Radar

cos(2πf0t)

Antenna
yI (t)
Lowpass A/D
filter converter

Bandpass y(t) y[n] Matched g[n] Threshold

filter filter detector

Lowpass A/D
filter yQ(t) converter λ

-sin(2πf0t)

Figure 2.3: Block diagram of a classical coherent receiver utilizing a threshold

detector.

2.2 Radar Detection Fundamentals

After the signal is received, various signal processing techniques are applied.
One of these techniques concerns detecting if targets are present or absent.
We will briefly discuss detection performance and fundamentals for a classi-
cal coherent receiver. This receiver uses coherent integration, which means
that the received signal is deterministic, and therefore, phase information is
retained by integrating over samples.
The structure of a classical coherent receiver employing a threshold de-
tector is illustrated in Figure 2.3.
As seen, the incoming signal is passed through a bandpass filter and
a Low-Noise-Amplifier (LNA). These two steps increase the received sig-
nal amplitude, and suppress noise contributions outside of the signal band.
Throughout this section, it is assumed that the target echo is not corrupted
by interference. Thus, noise is only generated at the receiver side.
After these steps, the signal is down-modulated to baseband, passed
through a lowpass filter, and digitally sampled. From the Nyquist-Shannon
theorem, the signal (2.8) should be sampled at t = nTs , where Ts ≤ B1 . This
results in the digital, complex signal

y[n] = yI (nTs ) + jyQ (nTs ), (2.10)

where yI and yQ are the in-phase (I) and the quadrature (Q) signal compo-
nents, respectively. These components are given by

yI (nTs ) = r(nTs ) cos [φ(nTs )] + zQ (nTs ) (2.11)

yQ (nTs ) = r(nTs ) sin [φ(nTs )] + zI (nTs ).

In (2.11), φ(nTs ) = arctan(yQ (nTs )/yI (nTs )), and r(nTs ) is the amplified
received signal amplitude. The receiver noise, z(nTs ) = zI (nTs )+jzQ (nTs ), is

10
2.2 Radar Detection Fundamentals

assumed to be generated from a complex zero-mean circular random process

with variance σz2 .
The goal is to detect if a target is present (or not) from the received
signal vector, y = [y[0], y[1], . . . , y[N − 1]]T , where N is the number of signal
samples. To do this, formulate two hypotheses:
H1 : y = s + z ∼ N (r, σz2 I)
H0 : y = z ∼ N (0, σz2 I),
where, for an ideal detector, we always choose H1 if a target is present, and
H0 if a target is absent. However, this is not the case in reality.
There exist two errors that can occur. First, deciding that a target is
present when it is not, that is, a false alarm, and second, deciding that a
signal is absent when it is not, that is, a missed detection. In general, a value
for the probability of a false alarm, Pfa , is given by an acceptable error rate,
and we seek to maximize the probability of detection, Pd . This formulation
results in the so-called Neyman-Pearson detector.
The detector, formulated with a Likelihood Ratio Test (LRT), or equiv-
alently a log LRT, evaluates the ratio between the conditional probability
density functions (pdfs), p(y|H1 ) and p(y|H0 ) by

p(y|H1 ) H1
log L(y) = log R log λ, (2.12)
p(y|H0 ) H0

where λ is a threshold, and the conditional pdfs for N complex samples are
(y−r)H (y−r)
1 − 2
p(y|H1 ) = e σz
π N σz2N
H
(2.13)
1 −y y
p(y|H0 ) = N 2N e σz2 .
π σz
Inserting (2.13) into (2.12) gives the log LRT
 
1 H H
log L(y) = 2 2<(r y) − r r , (2.14)
σz
where <(·) denotes the real part.
The ratio is a function of the observed signal only through <(rH y). This
data dependent term, which determines the outcome of the test, is called a
sufficient statistic, for which we rewrite the log LRT as
H1
Υ(y) = <(rH y) R γ. (2.15)
H0

11
 Chapter 2. Introduction to Radar

Here, γ is a threshold for the sufficient statistic, Υ(y). Worth noting is that
the sufficient statistic is a multiplication between the transmitted signal, r, up
to a constant, and the measured signal, y. Thus, when the noise is Gaussian
distributed, optimal detection is obtained by correlating the received signal
with a so-called matched filter [11].
Let g = rH y, which is a complex Gaussian random variable. Under the
hypothesis H0 , where no signal is present, the random variable is distributed
2
as g ∼ CN (0, N A2 σz2 ), where A2 = ||r|| N
is the energy of one sample of r.
For the hypothesis H1 , assuming a non-fluctuating target, g is instead g ∼
CN (N A2 , N A2 σz2 ).
The sufficient statistic (2.15) is the real part of g. Thus,

σz2
H1 : Υ(y) ∼ N (N A2 , N A2 )
2 (2.16)
2
2 σz
H0 : Υ(y) ∼ N (0, N A ).
2

To calculate the Pfa , note that a false alarm occurs if Υ(y) ≥ γ under the
hypothesis H0 . Hence, the Pfa is
Z ∞ Z ∞ 2
1 − Υ
Pfa = p(Υ|H0 )dΥ = p e N A2 σz2 dΥ = (2.17)
γ γ πN A2 σz2
  
1 γ
= 1 − erf p .
2 N A2 σz2

In (2.17), erf(·) is the error function, and its definition is found in [12].
Rearranging (2.17) yields the threshold

N A2 σz2 erf −1 (1 − 2Pfa )

p
γ= (2.18)

that achieves a predefined Pfa .

To derive the Pd investigate Υ(y) ≥ γ under the hypothesis H1 . This
yields
Z ∞ Z ∞ (Υ−N A2 )2
1 −
Pd = p(Υ|H1 )dΥ =p e N A2 σz2 dΥ = (2.19)
γ γ πN A2 σz2
γ − N A2
  
1
= 1 − erf p .
2 N A2 σz2

12
2.2 Radar Detection Fundamentals

0.2

0.18
0 1
0.16
Probability density
0.14
σ 2= 2 µ= 4; σ 2= 2
0.12

0.1 Pd
0.08

0.06
Pm

0.04
Pfa
0.02

0
−10 −5 0 5 10 15
Υ
Figure 2.4: Illustration of the pdfs for the sufficient statistic, Υ(y), under the
2
hypotheses H1 and H0 , with µ = 4 = N A2 and σ 2 = 2 = N A2 σ2z .

Inserting (2.18) into (2.19) gives the Pd with respect to Pfa

N A2 σz2 erf −1 (1 − 2Pfa ) − N A2

 p 
1
Pd = 1 − erf p = (2.20)
2 N A2 σz2
   s 
1 −1 N A2
= erfc erf 1 − 2Pfa − .
2 σz2

Here, erfc(·) = 1 − erf(·) and erf −1 (·) is the inverse error function. The term
A2
σz2
defines the Signal-to-Noise Ratio (SNR), and the number of samples, N ,
introduces a coherent processing gain.
Figure 2.4 depicts the pdfs for the sufficient statistic under the hypotheses
H1 and H0 . The figure gives an illustrative interpretation of the threshold’s
impact on the Pd , the Pfa and the probability of missed detection, Pm , re-
spectively.
The performance of a detector is sometimes characterized by Receiver
Operating Characteristics (ROC) curves. These curves, where the Pd for a
given Pfa is calculated for different SNRs, are summarized in Figure 2.5.
For the calculations above, we have assumed perfect knowledge of all
parameters contained in the conditional pdfs p(Υ|H1 ) and p(Υ|H0 ). Hence,
we require perfect knowledge of p(y|H1 ) and p(y|H0 ), which, generally are
not initially known.
Even if the type of pdf is known (Gaussian, Rayleigh, et cetera), the pa-
rameters of the pdf are commonly unknown and random [2]. Specifically,

13
 Chapter 2. Introduction to Radar

1
15 dB
0.9

0.8
12 dB
Probability of detection

0.7
9 dB
6 dB
0.6
3 dB
0.5

0.4
0 dB
0.3

0.2

0.1

0 −10 −9 −8 −7 −6 −5 −4 −3 −2 −1 0
10 10 10 10 10 10 10 10 10 10 10
Probability of false alarm

Figure 2.5: ROC curves describing performance of a detector for different

SNRs. The curves are calculated for a coherent receiver.

perfect knowledge of the impulse response, r, is unrealistic. A more reason-

able assumption is knowledge of r to within an unknown phase factor, ejθ ,
where θ is a random variable [2]. For this case we can employ a Generalized
Likelihood Ratio Test (GLRT), or use a Bayesian approach, where the pdfs
are computed under the hypotheses by separately averaging the conditional
pdfs [13].
Note that, nowadays concerning the receiver structure in Figure 2.3, the
analog-to-digital conversion is immediately performed after the LNA, with
a high sampling rate. After digitalization, the imaginary part is obtained
through a Hilbert transform. The I and Q signal components are thereafter
passed through a lowpass filter, and decimated to a lower sample rate.
There also exists other receiver structures. For example, if measurements
first have to be preprocessed to align phases, a noncoherent receiver and non-
coherent integration is employed. This happens, for instance, when a target
is moving. Noncoherent integration is not in the scope of this introduction,
and the reader is instead referred to [1]. Another area concerns wideband
correlation processing, this is further discussed in Chapter 3.

2.3 Narrowband Ambiguity Function

A radar system relies on accurate estimation of range and velocity. Given
a transmission model, which is governed by the narrowband approximation
(2.8), the response of the radar, using a matched filter, can be described by

14
2.3 Narrowband Ambiguity Function

x(t) T

t
PRI
Figure 2.6: Pulsed radar system, where a series of pulses are transmitted
over time.

an ambiguity function. The corresponding response, for a mismatched filter,

is instead represented by a cross-ambiguity function.
The ambiguity function is calculated from the correlation between the
matched filter and the signal, with respect to the two dimensional parameter
space, i.e., time-delay and Doppler shift. The narrowband ambiguity function
is defined by Z ∞
χh,x (ωd , τ ) = h(t − τ )ejωd t x∗ (t)dt. (2.21)
−∞
In (2.21), x∗ (t) is the complex conjugate of the signal (2.8), and h(t) is a
receiver filter, which, in the case of matched-filtering, is given by the trans-
mitted signal itself. When evaluating the cross-ambiguity function, there is
no restriction on the receiver filter response, h(t).
The equivalent wideband ambiguity function (WAF), utilized for wide-
band conditions, is defined in Chapter 3.
The ambiguity function, in the two-dimensional space, is represented with
a delay and a Doppler axis, respectively. Assume that a series of pulses is
transmitted, see Figure 2.6. In the figure, the PRI defines the pulse repetition
interval. The reciprocal of the PRI is refereed to as the pulse repetition
frequency (PRF). The ambiguity in range describes a maximum distance,
Rmax , where the system cannot distinguish if a reflection is from the first or
the second pulse. This maximum distance is defined by
c
Rmax = PRI. (2.22)
2
In comparison, to estimate the Doppler shift, a FFT on a pulse-to-pulse basis
is performed. The Doppler sample speed is thus equal to the PRF, and the
maximum unambiguous Doppler shift is
fd,max = PRF. (2.23)

15
 Chapter 2. Introduction to Radar

As seen, the PRI and the PRF define the unambiguous range and Doppler,
respectively. In addition, the so-called resolution defines how closely targets
can appear, and still be distinguished. The range resolution, in its simplest
form, is given by
cT
∆R = . (2.24)
2
For wideband signals, the range resolution actually depends on the band-
width, B, i.e.,
c
∆R = . (2.25)
2B
For the simple square pulse, the 3dB bandwidth is approximately B ≈ 1/T ,
for which (2.24) and (2.25) are equal. For the Doppler resolution, note that
the spectrum of the pulsed waveform experiences a sinc shape in the fre-
quency domain. The 3dB width of the main lobe for the sinc shape is defined
by f3dB = 1/T for which the Doppler resolution is
1
∆f = f3dB = . (2.26)
T
As seen, a good Doppler resolution is achieved with a long pulse duration,
whereas a good range resolution instead requires a short pulse. These two
requirements are in conflict with each other. However, the technique referred
to as pulse compression can be employed to mitigate this effect, where a long
pulse with a large bandwidth is constructed [6].

2.4 Estimation Fundamentals

When the presence of a target is detected, we are interested in finding the
best possible match of parameters to the corresponding measured data. For a
radar, these parameters are, for example, the target’s range, velocity, spatial
position, and reflectivity. These parameters describe the underlying physical
conditions together with a random signal component, that is, noise contri-
butions.
Estimation theory consists of two main branches, the stochastic and the
deterministic approach. For the deterministic model, it is assumed that the
outcome is certain, up to the measurement noise, if the input is known. This
means that, even after recalculation, the same result is obtained. In contrast,
the stochastic model introduces an uncertainty also in the signal model.
Techniques that incorporate the deterministic model are, for example,
Maximum-Likelihood Estimation (MLE) and Least-Squares (LS). For the
stochastic approach, techniques such as Maximum A Posteriori (MAP) or
Minimum-Mean-Squared-Error (MMSE) Estimation are often used.

16
2.4 Estimation Fundamentals

In the following, we present MLE, LS, and MMSE. The last part of the
chapter introduces two filtering techniques, namely the Wiener filter and the
Kalman filter, which are both examples of MMSE.

2.4.1 Maximum-Likelihood Estimation

For MLE we estimate a set of parameters for which the likelihood function
is maximized. The data and statistical model are assumed to be fixed.
Let the observed data, y, be a vector of N independent observations.
The samples are drawn from an unknown probability density distribution,
but belongs to a certain family of distributions, denoted p(y|θ), θ being a
vector of distribution parameters. If the samples are independent, the joint
density is expressed by
p(y|θ) = p(y1 |θ) × p(y2 |θ) × · · · × p(yN |θ), (2.27)
where the observations are assumed to be fixed and θ its variable. In addi-
tion, the function, p(y|θ), describes the likelihood, which is given by
N
Y
L(θ; y) = p(yn |θ). (2.28)
n=1

The principle of MLE aims to find the variable, θ̂ MLE , that maximizes (2.28),
that is
θ̂ MLE = arg max L(θ; y). (2.29)
θ
For some distributions, the MLE can be explicitly derived. If no explicit so-
lution is possible, optimization methods are necessary to find the maximizing
parameter set.
As an example, consider the received signal with one target present (2.12).
Assume that the noise variance, σz2 , is known, and that we wish to estimate
r(θ), which depends on the unknown parameter θ. Let C = σz2 I, the likeli-
hood function is then
1 H −1
L(r(θ); y) = N e−(y−r(θ)) C (y−r(θ)) . (2.30)
π det(C)
Taking the logarithm and differentiating with respect to θ i leads to
∂logL(r(θ); y) ∂(y − r(θ))H −1 ∂(y − r(θ))
=− C (y − r(θ)) − (y − r(θ))H C−1
∂θ i ∂θ ∂θ i
 i H

∂(r(θ))
= −2< − C−1 (y − r(θ)) .
∂θ i
(2.31)

17
 Chapter 2. Introduction to Radar

The MLE is then found by setting (2.31) to zero, i.e.,

∂(r(θ))H −1
 
< C (y − r(θ)) = 0. (2.32)
∂θ i
If r(θ) is nonlinear in the the parameter space θ, a nonlinear search method,
e.g., a Gauss-Newton method, has to be employed [14]. However, if a linear
model or a linear approximation is possible, the solution is governed by the
linear least-squares estimator. This is presented in the following section.

2.4.2 Least-Squares Estimation

Another class of estimators, called least-squares, are used when we do not
assume any probabilistic characterization of the measured data. This is an
advantage of LS. A disadvantage is that no statistical performance can be
derived.
The idea behind LS is to minimize the squared difference between an
observed signal and its estimate. Assume that the signal, r, can be modelled
in terms of a linear matrix transformation, that is

r = Hθ, (2.33)

where θ is an unknown complex-valued parameter vector which is to be

estimated, and H is a known N × p matrix with N > p of rank p. This
matrix is referred to as an observation matrix. When the model is linear, a
solution is found by minimizing

J(θ) = (y − Hθ)H (y − Hθ) . (2.34)

Expanding (2.34) yields

J(θ) = yH y − yH Hθ − θ H HH y + θ H HH Hθ. (2.35)

Differentiating with respect to θ gives

∂J(θ)
∗
= − yH H + (HH Hθ)∗ . (2.36)
∂θ
Here, the complex derivative is evaluated in the sense of Brandwood [15],
i.e.,
∂θ ∂θ
= I, = 0. (2.37)
∂θ T ∂θ H
Setting (2.36) to zero gives the LS solution

−1 H
θ̂ LS = HH H H y. (2.38)

18
2.4 Estimation Fundamentals

The solution in (2.38) is also the solution to (2.32), i.e., θ̂ LS = θ̂ ML , when

r(θ) = Hθ and the noise is Gaussian.
This is the most basic form of a LS problem. The method can be extended
by adding constraints to the parameter, a so-called constrained LS. If this
method is preferred, the technique of Lagrangian multipliers is employed to
find a solution.
The LS is successfully extended to a weighted LS, if we know the accuracy
of the information contained in the observation [16]. The idea is to introduce
weights associated with each data point. An advantage of weighted LS is the
ability to handle cases where observations experience varying qualities.
For the above derivation we assume that the parameter, θ, acts linearly
on the observation matrix. If this is not the case, we have to resort to solving
a nonlinear LS. To solve this problem, an iterative optimization approach is
usually performed.

2.4.3 Minimum-Mean-Squared-Error Estimation

In previous sections the unknown parameters were assumed to be determin-
istic. Another approach is to assume that the parameters are random, but
with a known prior distribution. This is the so-called Bayesian approach [17].
The most common Bayesian estimator is the MMSE estimator.
The goal with the MMSE estimator is to minimize the expected mean
value of the squared error. This estimator takes into account that we have
prior knowledge about the parameters which are to be estimated.
Given the measurement of a random variable Y , and the joint probability
density of X and Y , we seek the random variable X. Denote the a posteriori
density by pX|Y (x, y). The aim is to minimize
 2 
E X − X̂ |Y = y , (2.39)

where X̂ is an estimate that depends on y. Thus, we desire to minimize

Z
(X − X̂)2 pX|Y (x, y)dY. (2.40)

Setting the derivative of (2.40) to zero yields [17]

X̂ = E {X|Y = y} , (2.41)
i.e., the conditional expectation. The variance of the MMSE is the variance
of the posteriori density. When we have multiple measurements, the estimate
is instead
X̂ = E {X|Y1 = y1 , Y2 = y2 , . . . , YN = yN } . (2.42)

19
 Chapter 2. Introduction to Radar

The conditional expectation, E {X|Y1 = y1 , Y2 = y2 , . . . , YN = yN }, is gen-

erally hard to determine. A common solution is to restrict the estimator to
be fixed and linear. This results in the Linear MMSE (LMMSE) estimator.
In the following we give two examples of LMMSE estimators, namely the
Wiener and the Kalman filter.

2.4.4 Wiener Filter

The Wiener filter is proved to be optimal in a LMMSE sense [18]. It is based
on the stochastic framework, and is a linear estimation technique. The goal
is to decrease the noise that corrupts a desired signal. To use a Wiener
filter, the spectral properties of the signal should be known. Then, a linear
time-invariant filter is constructed, which assembles the original signal.
There are three main examples of Wiener filters, namely, a non-causal
solution, a causal solution, and a Finite-Impulse-Response (FIR) solution.
Herein, only the FIR solution is discussed, as it is the most practical ap-
proach.
The FIR Wiener filter is used for discrete time series. The received signal,
y, is convolved with a filter with coefficients h[i], i = 0, . . . , NF , NF being
the order of the filter. The output at sample instance n, from such a filter,
is given by
NF
X
ỹ[n] = h[i]y[n − i]. (2.43)
i=0

To determine the optimal choice of filter coefficients, h[i], investigate the

MSE criterion (2.39). Inserting (2.43) into (2.39) yields

 = E (ỹ[n] − x[n])2 = E ỹ 2 [n] + E x2 [n] − 2E {ỹ[n]x[n]} .

  
(2.44)

Inserting (2.43) into (2.44) gives

 !2  (N )
NF
 X  XF
 2
=E h[i]y[n − i] + E x [n] − 2E h[i]y[n − i]x[n] .
 
i=0 i=0
(2.45)
To find the filter coefficients that minimizes (2.45) take the derivative with
respect to the h[m], and set to zero. This results in

FN
∂ X
=2 h[i]E {y[n − i]y[n − m]} − 2E {x[n]y[n − m]} = 0. (2.46)
∂h[m] i=0

20
2.4 Estimation Fundamentals

In (2.46), Ryy [m] = E {y[n]y[n − m]} and Rxy [m] = E {x[n]y[n − m]} are
known as the auto-correlation and cross-correlation, respectively. Rearrang-
ing (2.46) gives
Nf
X
h[i]Ryy [i − m] = Rxy [m], (2.47)
i=0

for which the Wiener-Hopf equations [19] are formed

    
Ryy [0] Ryy [1] . . . Ryy [NF ] h[0] Rxy [0]
.. .. .. ..   ..   ..
 .  =  , (2.48)
 
 . . . . .
Ryy [NF ] Ryy [NF − 1] . . . Ryy [0] h[Nf ] Rxy [NF ],
| {z } | {z } | {z }
Ryy h rxy

and the optimal filter coefficients are then given by

h = R−1
yy rxy . (2.49)

The matrix Ryy is the auto-correlation matrix which is related to the power-
spectral density, and rxy is the cross-correlation which relates to the cross-
spectral density.

2.4.5 Kalman Filter

The Kalman filter is a generalization of the Wiener filter, and is used to
estimate parameters evolving in time [18, 20]. The filter utilizes observed
data to produce estimates of unknown parameters. The algorithm consists
of two steps. First, a prediction of the next state is performed. After that,
when new data is observed, in the second step an update of the parameters
is produced.
For the later discussed problem in Paper IV and V, it is assumed that
the states, xn , and the measurement framework, yn , where n is a time index,
are expressed by a state-space model of the following form:

xn = Fn−1 xn−1 + vn−1 (2.50)

yn = Hn xn + wn . (2.51)

Here, the process and the measurement noise, vn−1 and wn , are assumed to
be complex circularly symmetric Gaussian distributed with

vn−1 ∼ CN (0, Qn−1 ) (2.52)

wn ∼ CN (0, Rn ). (2.53)

21
 Chapter 2. Introduction to Radar

The matrices Qn−1 and Rn are the process noise covariance matrix and
the measurement noise covariance matrix, respectively. Gaussian distributed
noise is necessary for the Kalman filter to achieve MMSE. However, if this is
not the case, proper second-order moments are necessary to achieve LMMSE.
with In addition, Fn−1 is the system matrix, and Hn is the observation
matrix.
In the prediction step, an estimate of xn|n−1 is produced from previous
data up to n − 1. The accuracy of the estimate is given by the covariance
matrix Pn|n−1 . The estimate xn|n−1 is then corrected through a measurement
update. This is performed with knowledge of the, yn , for which a posterior
estimate, xn|n , is calculated [18].
The following equations are used for the prediction step:

xn|n−1 = Fn−1 xn−1|n−1 (2.54)

Pn|n−1 = Pn−1 Pn−1|n−1 PH
n−1 + Qn . (2.55)

The measurement update step is governed by the following equations:

ỹn = yn − Hn xn|n−1 (2.56)

Sn = Hn Pn|n−1 HH
n + Rn (2.57)
H −1
Kn = Pn|n−1 Hn Sn (2.58)
xn|n = xn|n−1 + Kn ỹn (2.59)
Pn|n = (I − Kn Hn ) Pn|n−1 (I − Kn Hn )H + Kn Rn KH
n. (2.60)

The difference between the measurement, yn , and the predicted measure-

ment, Hn x̂n|n−1 , is called the innovation, and is denoted by ỹn . The covari-
ance matrix of this innovation is Sn , and Kn is the so-called Kalman gain.
To update the covariance matrix Pn , we use the formula (2.60), which is
called the Joseph form [21, 22].
The Kalman filter only handles linear models. If the model is non-linear,
either the extended Kalman filter [23], or the unscented Kalman filter [24]
are required. More advanced approximations of the MMSE estimator, such
as the particle filter [25], also exists.

2.5 Antenna Array Beamforming

Using an antenna array introduces a possibility to form directive beams,
which, if the signals are properly combined, increases the power of outgoing
and incoming signals in a radar system [26]. Herein, we introduce narrowband

22
2.5 Antenna Array Beamforming

antenna arrays. A discussion on wideband antenna arrays is presented in

Chapter 3.3.
For transmission and reception of signals, an array of sensors (antenna
elements) distributed over a surface is utilized. Purposes of an antenna array
are, for example:
• Localization of a source.

• Reception of messages from another source.

• Imaging of an area.
The placement of the sensors gives the antenna array different charac-
teristics. In the literature, there are three commonly discussed sensor con-
figurations. These are a Uniform-Linear Array (ULA), a Uniform-Planar
Array (UPA), and a Uniform-Circular Array (UCA) [27]. Herein, we only
consider the ULA antenna configuration. Thus, UPAs and ULAs are not
further discussed.
When utilizing an antenna array, the signal is transmitted and received
from multiple antennas. Thus, the signal is built up by several outputs/inputs,
and the goal is to transmit/receive a combination of signals in the best pos-
sible way.
Investigating the receiver function, the incoming signal, at each element,
is a time-delayed version of the others. To steer the antenna array, that is,
to form a directive gain in another direction than broadside, phase shifters
are mounted after each element. The setup of a ULA employing L antenna
elements is depicted in Figure 2.7.
As illustrated, the incoming signal arrives from an angle θ. This angle
is commonly called the Direction Of Arrival (DOA), and is measured with
respect to the normal of an antenna array.
The time-delay between the different sensors, denoted by τl , depends on
the inter-element spacing, d, the DOA, and the speed of propagation, and is
given by
dl sin θ
τl = . (2.61)
c
To avoid creation of grating lobes, d ≤ λ2 must hold, where λ = fcc is the
wavelength at the frequency of operation [6]. The time-delay (2.61) intro-
duces a phase shift between the received signals, and the output voltage E,
after the signals are combined is
L−1 L−1
dl
X X
jωc τl
E = E0 wl e = E0 wl ejωc c sin θ
, (2.62)
l=0 l=0

23
 Chapter 2. Introduction to Radar

l
na
sig
θ

g
in
c om
In
Antenna
elements
d
l=1 l=2 l=3 l=L
Phase
shifters

Figure 2.7: Geometry of a ULA with phase shifters for which a steering of
the energy is possible.

where E0 is an amplitude-modulated incoming signal, and wl = al ejφl is

the applied weight for the lth antenna element. The coefficients, al , are the
amplitude tapers, and φl the applied phase shifts. To maximize energy the
weights are selected as
dl
wl = e−jωc c sin α
α=θ
. (2.63)

Here, α is the so-called the steering angle. Hence, to maximize the energy,
the array is steered towards the DOA. The magnitude of the array response
describes its directivity, and is given by

K−1
dk
X
AP(α) = e−jωc c
(sin α−sin θ)
. (2.64)
k=0

Figure 2.8 illustrates the narrowband array response, in a linear scale, for
arrays comprising L = {10, 20} elements.
As seen, when increasing the number of elements a higher directivity,
or antenna gain is obtained. Moreover, the resulting main lobe becomes
narrower.
In this section, we have discussed fixed phase shifters. However, the phase
shifters can be adaptively derived, that is, adaptive beamforming, where the
weight coefficients, wl , are adapted to prevailing conditions. Employing an
adaptive beamformer configuration gives the possibility to, for example, place
a null towards the direction of a jammer or strong clutter. The readers are
referred to [27–30] for fundamentals of adaptive array signal processing.

24
2.6 Radar Operating Environment

20
10 antenna elements
20 antenna elements
Antenna pattern 15

10

0
−80 −60 −40 −20 0 20 40 60 80
Azimuth angle [Deg]

Figure 2.8: Antenna pattern for a ULA comprising 10 and 20 antenna ele-
ments.

2.6 Radar Operating Environment

The radar scenario describes the environment where the radar system is
operating. In Figure 2.9 an example of a radar scenario is depicted.
As illustrated, an incoming signal does not only contain a desired signal,
but also contributions from various disturbances, called interference. More
precisely, interference is divided into clutter, see Chapter 2.6.1, and jamming,
see Chapter 2.6.2.
When an incoming signal is corrupted by interference, the SNR is not
a sufficient measure. Instead, of interest, is the ratio between the desired
signal component, and the interference and noise. This ratio, called Signal-
to-Interference-and-Noise Ratio (SINR), impacts on the performance of the
radar system, and is defined by

Psignal
SINR = . (2.65)
Pnoise + Pinterference

Here, P denotes the average power of the signal, the receiver noise, and the
interference, respectively.
When interference is present, the optimal transmitter and receiver func-
tion differs from the case where receiver noise is the dominant disturbance.
The following two sections specifically introduce clutter and jamming mod-
eling.

25
 Chapter 2. Introduction to Radar

Clutter Target

Jammer

Clutter

Clutter

Figure 2.9: Exemplification of a radar scenario comprising a target and var-

ious interference.

2.6.1 Clutter Interference

Clutter refers to returned echoes from undesired objects that naturally ap-
pear in the environment. Examples of clutter are: buildings, rain, ground,
sea, and ionized media [31].
In general, clutter is spatially distributed with a larger extent than the
radar resolution cell. However, it can also be a return from one scatter, such
as a building or a tower. A distinguishing characteristic of clutter returns
is that they are reflections from stationary objects, whereas targets are in
general moving.
There exist various clutter models that describe the clutter’s characteris-
tics. When clutter is distributed in space, its back scattering echo is described
by a Radar-Cross-Section (RCS) density. The mean value, for surface clutter,
is defined per unit area by
σc
σ0 = , (2.66)
Ae
where Ae is the illuminated area, and σc is the clutter RCS from this area.
If volume clutter is present, the mean value is instead described by
σc
σ0 = , (2.67)
Ve
where Ve is a volume, and σc is the clutter RCS from the volume. In addition,

26
2.6 Radar Operating Environment

the back scattered echo depends on the grazing angle, δ, as

σ0 = γ sin δ. (2.68)

Here, γ describes the clutter characteristics, and the equation reaches its
maximum when δ = 90, that is, normal incidence.
The clutter is characterized by the mean value and a probability density
function, which describes a statistical fluctuation. Different clutter experi-
ence different fluctuations. Commonly used examples are: Rayleigh, log-
normal, and Weibull distributions. In this thesis, we utilize the Weibull dis-
tribution to express clutter fluctuations. The distribution is mathematically
expressed by
k  v k−1 −( λv )k
f (v; k, λ) = e . (2.69)
λ λ
Here, k is the Weibull parameter and λ the median of the distribution. For
more discussions on clutter modeling, see [1, 6, 31].

2.6.2 Jamming Interference

Compared to clutter, radar jamming, or Electronic-Counter Measure (ECM),
is constructed to interfere with a return from a desired echo. The ECM is a
part of the radar warfare equipment, which also contains Electronic-Counter-
Counter Measure (ECCM) and Electronic-Support Measure (ESM).
The ECCM is the part of the system that is designed to reduce or elimi-
nate the effect of ECM. In comparison, the ESM module retrieves information
through a passive listening process.
There are several different methods of radar jamming. These are divided
into passive and active jamming. The category passive jamming comprises
the use of confusing reflectors, such as chaff or reflecting decoys [32], whereas
an active jammer deliberately emits electromagnetic radiation to interfere
with the radar echo.
Active jamming is partitioned into noise and deceptive jamming. The
continuous noise jammer radiates a random noise with a bandwidth Bj ,
where Bj commonly is wider than the radar receiver bandwidth B. The
outcome is an increased background noise in the receiver. In comparison,
a deceptive jammer repeats the transmitted signal with a possibly altered
angle, velocity, or range [32–34]. The intention is to produce false echoes in
the radar receiver.

27
 Chapter 2. Introduction to Radar

28
Chapter 3
Wideband Models

T
his chapter introduces concepts which are related to wideband models
and wideband signal processing. These concepts are necessary when
the narrowband approximation or condition fails, see Chapter 2.1.
We first introduce the wideband ambiguity function, followed by receiver
filter layout/design, and wideband antenna arrays.

3.1 Wideband Ambiguity Function

Similar to the narrowband case, the wideband ambiguity function (WAF)
is related to parameter sensitivity. For this model the received signal is
defined by (2.6), and the WAF is calculated for the parameter space defined
by the time-scaling, µ, and the time-delay, τ . The function specifies the
correlation between a transmitted signal and its corresponding matched filter.
Whereas, the wideband cross-ambiguity function is defined with respect to a
transmitted signal and a general filter.
The WAF is mathematically described by
Z ∞
√
χh,x (µ, τ ) = µ h(µ(t − τ ))x∗ (t)dt, (3.1)
−∞

where x∗ (t) is the complex conjugate of the signal (2.4), and h(t) = x(t) is the
matched filter. In comparison, if the cross-ambiguity function is evaluated,
a general model is used for h(t).
The magnitude of the ambiguity function, |χh,x (µ, τ )|, has its peak at
(µ, τ ) = (1, 0). From this magnitude the range and velocity resolutions are
given through the 3dB mainlobe width, which is defined in the time-delay
and the time-scaling space.

29
 Chapter 3. Wideband Models

correlating filters detector

y (t ) h∗ ( µ1 ( t − τ 1 ) ) ∫ dt i

y (t ) h∗ ( µ 2 ( t − τ 2 ) ) ∫ dt i
select
max ( µˆ ,τˆ )
y (t ) h∗ ( µ3 ( t − τ 3 ) ) ∫ dt i

⋮
y (t ) h∗ ( µ k ( t − τ k ) ) ∫ dt i

Figure 3.1: The received signal, y(t), is correlated with replicas of the trans-
mitted signal, h(t).

The continuous wideband ambiguity function for a finite, discrete signal

can be approximated for a long signal with a sufficiently high sampling rate.
Nevertheless, the time-scaling requires some form of interpolation.

3.2 Wideband Correlation Processing

For narrowband models there exists various techniques for estimation of tar-
get parameters. Many of these techniques are not efficient when applied to
signals with a large fraction bandwidth, signals with a large time-bandwidth
product, and signals reflected from rapidly moving targets. Advantages with
such signals are noise immunity, improved resolution, and removal of con-
straints related to velocity [10]. A disadvantage, when processing wideband
signals, is that the time-delay and the time-scaling (related to velocity) can-
not be separated in time and frequency.
For correlation processing, assuming a matched filter, the received signal
is correlated with a hypothesized replica of the transmitted signal, with an
altered scale and shift, and integrated over time. This process is depicted
in Figure 3.1. In the figure, a thresholding detector is utilized, and the
maximum value gives estimates of the time-scaling and the time-delay.
Herein, we discuss upon a Wavelet based correlation processing, where
a separable parameter space is possible with the Wavelet transform. After
that, we introduce the case where the receiver filters have a more general
structure, where the filters are, instead of matched to the transmitted signals,
mismatched or optimally selected.

30
3.2 Wideband Correlation Processing

3.2.1 Wavelet Correlation Processing

The relationship between wavelet transforms and wideband correlation pro-
cessing has been discussed in [10, 35]. Wavelets are beneficial as this partic-
ular choice of waveforms imposes separation of time-delay and time-scaling
with the Wavelet transform.
A Wavelet transform involves an integration over a kernel function. This
kernel function is what commonly is called a mother wavelet, which is trans-
formed through a time-shift and a time-scaling. There exist a number of
different wavelet functions, for example, Haar, Daubechies, Morlet, which all
experience different characteristics.
Denote the mother wavelet by Ψ(t), then the Continuous-Wavelet-Transform
(CWT) [36] of a function f (t) is
 
t−b
Z
1 ∗
CWTΨ f (a, b) = p f (t)Ψ dt. (3.2)
|a| a

The parameters a and b represent a continuous set of dilations and trans-

lations, respectively. A discretization of (3.2) is commonly performed with
respect to the parameters a and b, keeping t as a continuous variable. To
discretize, let a = al0 and b = kb0 al0 , where a0 and b0 are fixed scaling and
translation parameters, respectively. This means that the mother wavelet is
parametrized by the discrete variables k and l, for which (3.2) is
Z
−l/2
f (t)Ψ∗ a−l


CWTΨ f (l, k) = a0 0 t − kb0 dt. (3.3)

For many applications a0 = 2 and b0 = 1. Then the modified kernel is

Ψl,k (t) = 2−l/2 Ψ(2−l t − k). (3.4)

From (3.4), the convolution (3.3) is sampled at the points 2−l .

Denote the finest scale by L. Then we can write the function f (t) as
L X
X X
fL (t) = wl,k Ψl,k (t) + sl0 ,k ϕl0 ,k (t). (3.5)
l=l0 k k

In (3.5), l0 is a certain level of scaling, wl,k and sl0 ,k are the wavelet and the
scaling coefficients, respectively. Finally, ϕl0 ,k is a known scaling function
[37].
The computation of the coefficients wl,k and sl0 ,k is done with a filter bank.
Generally, we only have samples of f2L k , from which the scaling coefficients
are numerically calculated [37]. Then, the computation of the coefficients

31
 Chapter 3. Wideband Models

output to
next stage output to
G ↓2 G ↓2 next stage
input signal
H ↓2 wL − 2

H ↓2 wL −1

Figure 3.2: The forward Wavelet transform as a filter bank.

is performed by sending the scaling coefficients through a highpass, H, and

lowpass, G, filter followed by a decimation. The output from the highpass
filter gives the wavelet coefficient wL−1,k , and the output from the lowpass
filter is further processed to obtain the necessary number of coefficients. The
process is depicted in Figure 3.2.
A disadvantage of Wavelet correlation is that many highpass and lowpass
filters are needed to fulfil a required scale resolution. A possible solution for
this issue is discussed in [10].

3.2.2 Mismatched Filter Bank

A matched filter bank is only optimal when the received signal is not cor-
rupted with interference. When clutter and jamming disturbances are present,
a mismatched filter bank, or an optimally selected filter bank might be a bet-
ter choice. The filter bank can then be designed for specific purposes. For
example, to null signals at specific time-delays and time-scaling, or for side-
lobe suppression [38, 39]. Moreover, if separate receiver filters are employed
after each antenna element, spatial beamforming can be included in the cor-
relation processing. This separate channel processing is discussed in Chapter
4.
The receiver filters are selected to give an enhanced performance, for
an investigated criterion, assuming knowledge of the transmitted signal. In
addition, the transmitted signals can be introduced as a parameter in the
selection, if the transmission function has the property to synthesize arbitrary
signals. Hence, we have a system for which we can design both receiver filters
and transmit signals.
This kind of system gives more degrees of freedom compared with the case
where the receiver filters are chosen as replicas of the transmission model.
Thus, increasing the computational complexity as both the transmitted wave-
form and receiver filters require optimization.

32
3.3 Wideband Antenna Array Beamforming

3.3 Wideband Antenna Array Beamforming

This chapter focuses on wideband antenna systems, and the challenges that
appear with a high system bandwidth. For a wideband antenna system it
is anticipated that the bandwidth, B, is a significant fraction of the system
operating frequency, fc . Therefore, the radar is not only operating at the
center frequency, but instead, for transmission, all frequencies contained in
the interval f = fc − B2 , fc + B2 are used. The equations herein assume
that a ULA antenna configuration is employed.
The phase-shifts between the sensors, discussed in the Chapter 2.5, will
for the wideband case, depend on frequency. Hence, the argument, φl , of the
phase shifters, wl = al e−jφl , are instead given by
dl
φl = ω sin α. (3.6)
c
In (3.6), ω = 2πf is an angular frequency, which is defined by the center
frequency and the bandwidth, d and α are, as previously defined, the spacing
between adjacent antenna elements and the steering angle, respectively. To
avoid creating grating lobes, d is instead typically selected as
c
d≤ , (3.7)
(fc + B/2)(1 + | sin αmax |)

where αmax is the maximum steering angle.

For wideband antenna arrays, compared with a narrowband antenna ar-
ray, for which only the carrier frequency is used for transmission, if the
phases, φl , in (3.6) are fixed, a change in the frequency, f , results in a differ-
ent steering angle, see Figure 3.3. This distortion results in a beam squinting.
To overcome this distortion, a linear phase filter, i.e., a filter that achieves
an approximately constant group-delay, is introduced at the transmit and
receive subapertures. This is known as a true-time-delay technology [6, 40].
A narrowband and a wideband ULA produce different array responses.
In a one-dimensional space, the wideband array response is, compared with
the narrowband array response (2.64), integrated over the bandwidth, which
results in
Z XL−1
dl 2
AP(α) = e−jω c (sin α−sin θ) dω . (3.8)
ω l=0

Figure 3.4 illustrates the normalized wideband antenna array response steered
towards α = 20 degrees, when the system operates at the carrier frequency
fc = 9 GHz with a bandwidth of B = {1, 3, 4} GHz. The array consists of
L = 10 antenna elements.

33
 Chapter 3. Wideband Models

Antenna pattern f0 f0+ Δf

α0 α0+ Δα α

Figure 3.3: Illustration of beam squinting, where, for fixed phases, φl , a

change in the frequency affects the steering angle.

The normalized narrowband and wideband antenna array responses are

depicted in Figure 3.5. The antenna array consists of L = 10 elements, the
carrier frequency is 9 GHz, and for the wideband system the bandwidth is
2 GHz. As shown, the narrowband and wideband antenna responses differ
from each other, with an increased difference at angles far away from the
steering angle. In particular, the depth of the nulls is clearly decreased for
the wideband array response.

34
3.3 Wideband Antenna Array Beamforming

0
Normalized antenna pattern [dB]
Bandwidth: 1 GHz
Bandwidth: 3 GHz
Bandwidth: 4 GHz
−5

−10

−15
−80 −60 −40 −20 0 20 40 60 80
Azimuth angle [Deg]

Figure 3.4: Normalized wideband antenna responses utilizing a system band-

width of B = {1, 3, 4} GHz, and a center frequency fc = 9 GHz.

0
Narrowband
Normalized antenna pattern [dB]

Wideband
−5

−10

−15

−20

−25

−30
−80 −60 −40 −20 0 20 40 60 80
Azimuth angle [Deg]

Figure 3.5: Normalized narrowband and wideband antenna array patterns.

The graphs are calculated for L = 10 elements, fc = 9 GHz, and
B = 2 GHz

35
 Chapter 3. Wideband Models

36
Chapter 4
Waveform and Receiver Filter Design

I
n Chapter 1 it was mentioned that the use of multiple wideband flexi-
ble transmitters, each transmitting different suitably selected waveforms,
might significantly increase performance of future radar systems. These
waveforms, and also receiver filters, can then be optimized and adapted for
prevailing environment conditions.
When radar stations are utilizing transmitters with fully adaptive wave-
forms, electronic surveillance equipment with classical libraries of frequencies,
pulse repetition intervals, and pulse lengths, may for identification become
obsolete, as these properties can be embedded in the waveforms. It is also
anticipated that optimal waveforms improve stealth properties of the radar,
i.e., the waveforms will be more difficult to discover compared with tradi-
tionally used signals.
The first part of this chapter introduces a particular transmitter and
receiver function, which is refereed to as a Multiple-Input-Multiple-Output
(MIMO) system. This structure accomplishes diversity, as each transmitter
chain can synthesize an arbitrary waveform. Diversity can be formulated in
space, in time, and in frequency. The MIMO structure has more than one
transmission and receiver chain. If one receiver chain and multiple trans-
mitter chains are employed, the system is said to have a Multiple-Inputs-
Single-Output (MISO) structure. In contrast, if one transmitter chain and
multiple receiver chains are utilized, the structure is instead called Single-
Input-Multiple-Output (SIMO).
The second part of this chapter focuses on describing different optimiza-
tion methods, which are used in the appended papers. The discussed methods
are: semidefinite relaxation, Gauss-Newton optimization, second-order-cone
programming, and the bisection method. The optimization methods are fol-
lowed by a discussion on robustness.

37
 Chapter 4. Waveform and Receiver Filter Design

Target

Combination of
x1(t) xK (t)

x1(t) xK (t) y1(t) y L (t)

Transmit array Receive array

Figure 4.1: MIMO antenna configuration, where K waveforms are transmit-

ted and L back-scattered signals received.

4.1 Waveform Diversity

The interest for MIMO radar originated from the dramatic improvement
that the technology had in communication systems [3]. There are many
similarities between MIMO communication and MIMO radar, these are well
described in [4, 41]. A main advantage for radar technology, when compared
with communication systems, is that the transmitted waveforms commonly
are known. However, a disadvantage is that multipath propagation are re-
garded as interference, and is not useful for target parameter estimation.
In addition, the performance criteria are different. For radar, commonly,
we want to maximize probability of detection for a given false alarm rate,
whereas for communication, criteria such as capacity or bit-error-rate are
important.
A MIMO radar layout, compared with traditional radar systems, see,
e.g., [6, 42], where antenna elements sample scaled and either time trans-
lated, or phase shifted versions of a single waveform, allows array elements
to transmit arbitrary selected waveforms. This technique is depicted in Fig-
ure 4.1. As shown, each antenna element possibly transmits a different wave-
form, here denoted xk (t), k = 1 . . . K. The receiver samples the signals yl (t),
l = 1 . . . L, which are combinations of the K back scattered transmitted
waveforms.
The transmitted waveforms can either be uncorrelated, fully correlated,
or partially correlated, where fully correlated signals governs the traditional
beamforming technique. However, as there is no restriction, for the trans-

38
4.1 Waveform Diversity

mission scheme, all combinations are possible. If it is necessary to create

antenna directivity, a so-called subarray architecture is utilized. This results
in that partially correlated signals are transmitted, as a group of antennas
transmit the same signal sequence.
Figure 4.1 illustrates a co-located antenna structure, i.e., the antenna
elements are closely spaced. This is the configuration that is investigated
in this thesis. However, another possibility is to employ a widely separated
antenna configuration, where, as the name suggests, antennas are positioned
far away from each other.
These two diverse antenna configurations, co-located and widely sepa-
rated, enhance different merits of performance. Investigation of widely sepa-
rated antenna elements shows an increased SNR, when exploiting a target’s
radar cross section [43]. This is due to that, different antennas experience
different angle of aspect, and as the target radar-cross-section might vary
with respect to the aspect angle, it is more likely to encounter a good as-
pect of the target. In addition, [44] reports an improved performance when
searching for slowly moving targets, and in [5] a higher resolution for target
localization, as well as a possibility to resolve targets located in the same
range cell is discussed. In comparison, a co-located antenna configuration
likewise offers a higher target resolution [4] compared with traditional radar,
and an improvement in target identifiability [45].
When we are able to utilize arbitrary waveforms, we want to design the
waveforms such that they improve the system’s performance. In the lit-
erature, two design methods are mainly investigated. The first approach
focuses on the spatial properties of the transmit signals, see, e.g., [46–49].
The second method concerns temporal properties of the transmitter–receiver
chain, see, for example, [50–53]. These two design methods result in differ-
ent performance criteria, which are described in Chapters 4.1.1 and 4.1.2,
respectively.

4.1.1 Waveform Design Utilizing Spatial Properties of

the Transmitted Signals
When investigating the spatial properties of the signals, the possibility to op-
timize waveforms to coincide with a specific beampattern arises. Assuming
a narrowband radar system, the waveform design problem is generally ex-
pressed as an optimization of the spatial correlations of the waveforms [46,47].
The optimization procedure involves finding a covariance matrix of the
waveforms that achieves certain desirable properties. Specifically in [47], four
design problems that invoke different design criteria are investigated. The

39
 Chapter 4. Waveform and Receiver Filter Design

investigated criteria are:

• A maximum power design for unknown target locations.

• A maximum power design for known target locations.

• A beampattern matching design.

• A minimum sidelobe beampattern design.

Obviously, there exist other possible design criteria. For a wideband radar
system, the problem is reformulated as a matching of the cross-spectral-
density matrix to a desired spatial beampattern [48]. In [49], the signals
are instead described by the Fourier transform of a spatial beampattern.
Moreover, in [49] an algorithm that performs beampattern matching and
time-domain signals synthesis is proposed.
This is important as, even if we can synthesize arbitrary signals, it is not
obvious that the waveforms are ”hardware friendly”. A common restriction
on the time-domain waveforms is a low peak-to-average-power ratio, or even a
constant modulus. Synthesizing signals that fulfil these criteria are discussed
in Chapter 5.

4.1.2 Waveform Design Utilizing Temporal Properties

of the Transmitted Signals
For the second approach, which is investigated in this thesis, a multitude of
studies have been performed in the area of MIMO communication, see, for
example, [50–52], where design of precoders and decoders are discussed.
The design of an optimal precoder is addressed in [50], and in [51] an
optimal design of space-time precoders and decoders is described. The un-
derlying problems and objectives are quite different for a communication and
a radar system. However, the two research areas can still benefit from each
other. For example, the method proposed by [52] concerning the design of
beamforming weights for complex relay networks, which exploits a predefined
power constraint, can be modified and used for a radar system.
Concerning radar technology, the design of transmit and receive filters is
discussed in [53], where an alternating method is proposed to increase the
SINR for an extended target in clutter. A similar method is discussed for
a SIMO layout [54, 55], where the SINR is maximized for a radar scenario
with target and clutter. In the referenced work, alternating algorithms are
proposed that improve the SINR in each iteration. In [56], a gradient based
method is instead proposed, where several suboptimal solutions are studied.

40
4.2 Optimization Methods for Waveform and Receiver Filter Design

The algorithm introduced in [54] can be extended to work for a MIMO sys-
tem. However, it is not guaranteed that the SINR increases in each step [53].
In contrast, in [53] a method that works for a MIMO radar system, and
which guarantees an increasing SINR in each iteration, is proposed.
The design of adaptive waveforms, where an estimate of the channel
statistics is employed to adapt the transmit signals is discussed in [57].
In [58], robust transmit waveforms and receive filters are studied based on
a minimax method. The study is performed for uncertainties related to the
target.

4.2 Optimization Methods for Waveform and

Receiver Filter Design
In this chapter we will briefly state the problems which are discussed in
the appended papers. The main purpose of this section is to introduce the
optimization methods, which are utilized in the papers.
In Paper I we continue on the work performed in [53, 54, 58], where a
transmit and receive function is optimized for different power constraints,
which are related to the transmit function. The objective is to select wave-
forms and receiver filters to maximize a SINR criterion for a wideband radar
system.
We evaluate the following four maximization problems:

1. An alternating optimization procedure with a total power constraint

for all transmit filters.

2. An alternating optimization procedure with individual power constraints

for each transmit filter.

3. A joint optimization procedure with a total power constraint for all

transmit filters.

4. A joint optimization procedure with individual power constraints for

each transmit filter.

For the alternating algorithm with a total power criterion, both transmit
and receiver filter coefficients have a closed form solution. This is not the case
for individual power constraints. Therefore, to solve for the transmit filters,
we use Semi-Definite Relaxation (SDR) together with a bisection technique.
The semidefinite relaxation and bisection concepts are described in Chapters
4.2.1 and 4.2.2, respectively.

41
 Chapter 4. Waveform and Receiver Filter Design

For the joint optimization technique we rely on a Newton type optimiza-

tion, and specifically a Gauss-Newton search. In fact, the stated optimization
is non-linear, but can efficiently be solved with such a search. The Gauss-
Newton algorithm is summarized in Chapter 4.2.3.
In the above discussed problems, it is first assumed that the objects’
positions are known. This is not a practical assumption, as there commonly
exist uncertainties in the estimates. Furthermore, in Paper III, a method
for robust target detection is investigated. Both of these concerns regard
design of robust systems. There are different ways to achieve robustness.
Two ideas are discussed in Chapter 4.2.5, namely a worst-case maximization
and a Taylor series approximation.
In Paper IV, we instead investigate waveform and receiver filter design
based on a clutter suppression criterion. The idea is evaluated for two dif-
ferent receiver filter structures, namely,

1. An optimally selected filter bank.

2. A matched filter bank.

For the design of waveforms and a matched filter bank, the SDR technique
is utilized, see Chapter 4.2.1, whereas for the optimally selected filter bank,
we exploit that the problem can be reformulated as a Second-Order-Cone
Program (SOCP). The definition of a SOCP is given in Chapter 4.2.4.

4.2.1 Semidefinite Relaxation

The SDR technique is a good tool to approximate difficult optimization prob-
lems. Specifically, it is useful when applied to nonconvex quadratically con-
strained problems [59, 60]. These type of problems are mathematically de-
scribed by
max xH Ax
x
s.t. xH Bx ≤ g (4.1)
xH Cx = f.
The matrices A, B, and C are given, and x is the unknown. The first step
is to see that
xH Ax = Tr xH Ax = Tr AxxH

xH Bx = Tr xH Bx = Tr BxxH



(4.2)
xH Cx = Tr xH Cx = Tr H



Cxx ,
where Tr(·) is the trace operator. As seen, both the objective function and
the constraints are linear in the matrix xxH . Introduce the matrix X = xxH ,

42
4.2 Optimization Methods for Waveform and Receiver Filter Design

and note that the equality holds if X is a rank-1 and a Hermitian positive
definite matrix. Thus, we obtain the equivalent optimization problem

max Tr (AX)
X
s.t. Tr (BX) ≤ g
(4.3)
Tr (CX) = f
X  0, rank(X) = 1.

In (4.3), X  0 means that X is a Hermitian positive definite matrix. The op-

timization problem in (4.3) is as difficult to solve as the original optimization
formulation. However, the only problematic constraint is the rank-1. This
constraint is removed with a so-called relaxation. Thus, the optimization
problem is instead
max Tr (AX)
X
s.t. Tr (BX) ≤ g
(4.4)
Tr (CX) = f
X  0.
This formulation is convex, and can be solved with a convex optimization
toolbox. A fundamental issue with (4.4) is that the solution might not be
rank-1. There exist many ways to find a suboptimal solution if the rank-1
constraint in not fulfilled, for example, we can choose the eigenvector corre-
sponding to the largest eigenvalue of X as a solution for the original variable
x. This is valid only if a high portion of the energy is contained in the
largest eigenvalue. Other methods, which involve randomization, are pre-
sented in [60, 61].

4.2.2 Bisection Method

The bisection technique is a simple and robust method to, in this case, lo-
cate the maximum of a function. It works by repeatedly divide an interval,
and selects a subinterval in which the maximum point must lie. As this is
sequentially done, the procedure is rather slow.
The bisection technique works in the following way [62]. Assuming that
the problem is feasible, and that the solution lies in the interval [l, u], where
l and u are a lower and an upper limit. Calculate the function value at the
midpoint t = (l + u)/2. If there exists a feasible solution at t, the interval is
updated by l = t and u = u. After that, a new midpoint is calculated, and the
solution is found at this point. If the solution is infeasible, the intervals are
instead l = l and u = t. The process continues until the interval is smaller

43
 Chapter 4. Waveform and Receiver Filter Design

than a user-defined criterion, u − l < . The algorithm is summarized in

Algorithm 1.

Algorithm 1 Bisection method.

Initialization:
Lower limit, l, upper limit, u, and .
while u − l <  do
Calculate t = (l + u)/2.
Solve the optimization problem with t.
if a feasible solution is found, set l = t.
else set u = t.
end while

4.2.3 Gauss-Newton Algorithm

The advantage of a Gauss-Newton method, over a Newton method, is that
we do not require to calculate the second derivative of the cost function,
f (x). The algorithm works as follows [14, 63]. Assume a cost function that
is a sum of squares X
f (x) = ri (x)2 , (4.5)
i

and an initial guess x0 . For the Gauss-Newton algorithm the Hessian is

approximated by
H = 2JTr Jr , (4.6)
where Jr is the Jacobian matrix, i.e.,
∂ri (xk )
(Jr )i,j = . (4.7)
∂xj
The gradient of the cost function is
∇f (x) = 2JTr r(xk ). (4.8)
At each iteration a new vector xk+1 is given by
1 −1 1
−1 T
xk+1 = xk − H ∇f (x) = xk − JTr Jr Jr r(xk ), (4.9)
µk µk
where µk is a step length. An issue with this type of algorithm is that
an initial guess is required, and if the guess is far from the true value the
convergence might be slow. In addition, there is a great chance to get stuck
in a local minimum or maximum.

44
4.2 Optimization Methods for Waveform and Receiver Filter Design

4.2.4 Second-Order-Cone Program

The SOCP is a method, where a convex optimization problem of the following
structure is solved [64, 65]

min f T x
x
s.t. ||Ai x + bi ||2 ≤ cTi x + di (4.10)
Fx = g.

As seen, the problem involves both inequalities and equality constraints. To

see that the original problem is convex, note that the inequality constraint,

||Ai x + bi ||2 ≤ cTi x + di , (4.11)

is a second-order cone constraint, which is convex. In addition, as the cri-

terion function and the equality constraint in (4.10) are linear, the SOCP is
a convex optimization. This technique is facilitated to solve for waveforms
and receiver filters in Paper IV and V.

4.2.5 Robust Design

Robust design methods are a well studied research area, and for an introduc-
tion to robust beamforming, see [30] and references therein. In particular,
robust methods for parameter estimation, waveform estimation, or beam-
forming in the presence of model uncertainties for narrowband systems are
investigated in [66–68]. For the wideband case, in [69], a robust beamformer
is derived based on approximating the steering vector by its first order Taylor
series expansion [12].
In this thesis, we investigate two methods to perform the robust design,
namely

• A Taylor series expansion.

• A worst-case maximization.

If no robust design is imposed, the performance might dramatically deteri-

orate for small parameter deviations. In Paper I, we seek a robust system
when uncertainties in target and jammer position are present. In Paper III,
a robust design considering detection probability is instead desired. This is
important due to that a wideband signal might lead to a focused ambiguity
function. Commonly this is desired, as the system will have good resolu-
tion properties. However, this results in a large set of receiver filters, each

45
 Chapter 4. Waveform and Receiver Filter Design

computed for a specific grid-point. To reduce the complexity, we therefore

propose a technique, where resolution properties are traded for robustness
and adapted for a desired resolution.
For the Taylor series expansion, the idea is to linearize and approximate
a function, f (x), with a Taylor series expansion around some value a, that
is,
f 0 (a) f 00 (a)
f (x) = f (a) + (x − a) + (x − a)2 + . . . , (4.12)
1! 2!
where f 0 (a) and f 00 (a) are the first and the second derivative evaluated at
the value a, respectively, and ! is the factorial. The number of terms gives
the accuracy of the approximation. It is anticipated that larger uncertainties
requires more terms, and this is discussed in Paper I.
For the second method, the worst-case maximization, assumes that the
objective function f (x) is defined on a set of points, i.e., f (xi ). These points
are given by the area for which a robust scheme is desired. Then, instead
of maximizing all points at the same time, we resort to maximizing the
minimum value, or the worst-case, that is

max min f (xi ). (4.13)

x i

Obviously, the maximin or equivalently minimax optimization is not only

used in robustness analysis. For example, in Paper IV, the same setup is
utilized to minimize the maximum correlation between receiver filters and
clutter interference.

46
Chapter 5
Waveform Synthesis

I
n this chapter we will discuss constraints related to hardware. Even
though we are able to generate arbitrary signals, it is not obvious that
they can be used for transmission. Specifically, at the transmission side,
both the power amplifier and the digital-to-analogue converters are designed
to operate for signals with small magnitude variations.
This chapter contains two parts. First, in Chapter 5.1 we discuss upon
synthesizing signals with a low Peak-to-Average-Power Ratio (PAPR), or a
constant envelope. Second, in Chapter 5.2, a practical experiment where we
evaluate the effect of PAPR is presented. The work was conducted at Saab
EDS and was published in [70].

5.1 Waveforms Synthesis to Minimize Peak-

to-Average-Power Ratio
In Paper I, the design of tunable filters that results in optimal spectral prop-
erties, for each transmitted waveform, is discussed. As the radar system
performance is directly linked to the time domain characteristics of the sig-
nals, we are interested in how to design the actual time domain signals.
This is possible as many time domain signals experience the same spectrum.
Therefore, we can choose the signal that exhibits the most desirable time
domain properties.
These obtained spectra, in Paper I, are used to synthesize time domain
signals, with desirable properties that coincide with predefined system re-
quirements. The requirements investigated in this thesis are:

• A time domain signal with a low PAPR.

47
 Chapter 5. Waveform Synthesis

• A time domain signal with a constant envelope.

The PAPR measures the largest instantaneous power of a signal compared

with its average power. For a sampled signal x, the PAPR is defined by

maxn |x[n]|2
PAPR = 1
PN −1 , (5.1)
2
N n=0 |x[n]|

where N is the number of signal samples. For a signal with a constant

envelope, we require that the PAPR is equal to one. The PAPR is of interest
as larger absolute variations of the signal require a higher dynamic range on
the digital-to-analogue converters, as well as power amplifiers with a large
linear range. Obviously, this increases both the cost and the complexity of a
radar system.
The problem was studied already in the 70s, where Schröder [71] inves-
tigated how to synthesize a waveform from a periodic signal with a given
power spectrum. He provided formulas to adjust the phase angles, a so-
called Partial Transmit Sequence (PTS) technique, of periodic signals that
yield a low PAPR, and closed form solutions were derived (for specific power
spectra). Continuing, the problem to construct multitone signals with a low
PAPR is addressed in [72–74]. Furthermore, in [75] four different PTS based
algorithms are discussed, and an extended version of the time–frequency
swapping algorithm [74] is selected as the preferred method.
In this thesis, we have investigated two different methods to synthesize
signals, where the first invokes a parametrization of the signal in the time
domain, and the second instead utilizes a parametrization in the frequency
domain. The methods produce two different outcomes:

• A signal with a perfect match of the spectrum with a low PAPR.

• A signal with an imperfect match of the spectrum but with a constant

envelope.

The first method is discussed in detail in Paper II, and the results are com-
pared with the preferred time-frequency swapping algorithm in [74]. In this
section we therefore introduce the basics for the second method [76].
To parameterize the signal, we incorporate the discrete Fourier transform
of a constant envelope signal Ax[n], where A is the amplitude and n =
0 . . . N − 1. The total energy in Ax[n] is restricted by Parseval’s theorem to
N −1 N −1
X
2 1 X
|Ax[n]| = |Xd [k]|2 . (5.2)
n=0
N k=0

48
5.1 Waveforms Synthesis to Minimize Peak-to-Average-Power Ratio

5
Spectrum 1
0 Spectrum 2

−5
P[k] [dB]

−10

−15

−20

−25

−30
0 0.2 0.4 0.6 0.8 1
Normalized frequency

Figure 5.1: Achieved spectra for two different phase dictionaries, φ1 and φ2 .

Here, |Xd [k]|2 = N Pd [k], where Pd [k] is the desired spectrum, and k =
0 . . . N − 1. The DFT of Ax[n] is
N −1
kn
X
X[k, φ] = Ax[n]ejφn e−j2π N . (5.3)
n=0

In (5.3), the introduced phases, φn ∈ [0 2π), do not change the implied con-
stant envelope constraint. However, the spectrum changes dramatically with
the choice of phases [71]. This is illustrated in Figure 5.1, where the spec-
tra for two different phase dictionaries are depicted. Note that the spectra
achieve the same constant magnitude in the time domain. By tuning the
phases, φn , we synthesize a signal with a spectrum that is close to a desired
one. Hence, we seek the phases that minimize
2
φ̂ = arg min max P [k, φ] − Pd [k] · w[k]. (5.4)
φn k
T
Here, P [k, φ] = N1 |X[k, φ]|2 , and w = w[1] . . . w[N ] is a vector with


weight coefficients. This weight function is introduced to emphasize, if nec-

essary, importance of specific frequency indices.
Figure 5.2 illustrates the desired and the obtained spectra after optimiza-
tion, where the weights are selected as the inverse of the desired spectrum.
Hence, through this normalization, the importance of the low-energy spectral
components is increased. As illustrated, the obtained spectra are approxi-
mately equal to the desired ones.
The weight function is also useful when, for example, we require the
spectrum to turn to zero at one or more frequency indices. Assume that we

49
 Chapter 5. Waveform Synthesis

0
Obtained spectrum 1
Desired spectrum 1
−2
Obtained spectrum 2
Desired spectrum 2
−4
P[k] [dB]

−6

−8

−10

−12

−14
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Normalized frequency

Figure 5.2: Desired and obtained spectra with weights selected as the inverse
of the desired power spectrum.

0
Obtained spectrum
Desired spectrum
−5
P[k] [dB]

−10

−15

−20
0 0.2 0.4 0.6 0.8 1
Normalized frequency

Figure 5.3: Desired and obtained spectra, for which the weights are selected
as the inverse of the desired power spectrum to emphasize the
importance of a notch.

are required to insert a null at the frequency index knotch = 128, with a depth
of at least Pd [knotch ] = −20 dB. The weight function is set as the inverse of
the desired spectrum, and the acquired spectrum is depicted in Figure 5.3.

50
5.2 Practical Implementation

Figure 5.4: Hardware overview of the measuring rig.

5.2 Practical Implementation

To test the impact of a high PAPR, a practical experiment was conducted
in [70]. For the experiment, the WAF was analyzed to indicate distortions
between the original signal and a signal that was measured at specific nodes
in the measurement setup.
The measuring rig is depicted in Figure 5.4, the nodes B and C indicate
where the signal was measured with an oscilloscope. The signals at these
nodes where compared with the original, A, desired signal to investigate
sensitivity of the system when using signals with different level of PAPR.
The hardware components in the figure are:

• Tektronix AWG520 baseband generator

• Agilent E4433B signal generator

• 24 dB gain pre-amplifier

• 20 dB bi-directional coupler

• 20 dB gain Power Amplifier (PA)

• 40 dB power damper unit

• Tektronix TDS7404 oscilloscope

The pre-amplifier is necessary as it allows for the PA to be driven into com-

pression.
The system was tested on three signals, each with a different PAPR of
1.2 dB, 3 dB, and 6 dB, respectively. The merit used for evaluation is the
obtained SINR for a specific scenario, see Paper I. For the investigated PA,
the SINR was decreased by 4 dB, 4.5 dB, and 12 dB for the three tested
signals, with a PAPR of 1.2 dB, 3 dB, and 6 dB, respectively. Thus, we
conclude that minimizing the PAPR of a signal is important.

51
 Chapter 5. Waveform Synthesis

52
Chapter 6
Contributions and Future Work

T
his chapter states the contributions of the thesis, and a discussion
of possible future work within the research area is held. The contri-
butions take the form of five appended papers, which consider four
different issues within transmit waveforms and receiver filters design, to in-
crease the SINR in presence of active jamming, to suppress clutter interfer-
ence, and for robust target detection. Paper II discusses waveform synthesis.
After the presentation of the contributions, an overview is given of possible
directions of future work within the research area.

6.1 Summary of the Appended Papers

The thesis is based on the following publications:

Paper I Robust Transceiver Design for Wideband MIMO Radar utilizing

a Subarray Antenna Structure
In Special Issue on Advances in Sensor Array Processing EURASIP Signal
Processing Journal, vol. 93, 2013.

In this paper, we investigate the possibility to suppress interference for wide-

band multiple-input multiple-output radar. The idea is to employ tunable
filters at the transmitter and the receiver sides, and to derive filter coef-
ficients that result in optimal transmit signals from a system performance
point of view, for a given radar scenario. The system performance is mea-
sured as the signal-to-interference-and-noise ratio (SINR) at the receiver out-
put, from which the filter properties are derived. The focus is to suppress
active jamming interference, and especially deceptive jamming interference.

53
 Chapter 6. Contributions and Future Work

We discuss two ways to derive the transmit and the receive filters. Each pro-
cedure invokes two different power constraints related to the transmit filters.
To incorporate imperfections in the given scenario a robust extension to the
design problem is proposed. Two different robust methods are evaluated,
one that utilizes a Taylor series expansion of the SINR, and one that ex-
ploits a worst-case SINR maximization. Numerical validation illustrates the
possibility to suppress interference without actually forming a spatial null in
the direction towards interference, and the necessity to design filters that are
robust to uncertainties in the given scenario.

Paper II Low PAPR Waveform Synthesis with Application to Wideband

MIMO Radar
In Proc. of the 4th International Workshop on Computational Advances in
Multi-Sensor Adaptive Processing, December 2011, San Juan, Puerto Rico.

This paper considers the problem of waveform synthesis given a desired power
spectrum. The properties of the designed waveforms are such that the overall
system performance is increased. The metric used to evaluate the optimal-
ity of the synthesized time-domain signals is the peak-to-average-power ratio
(PAPR). We discuss how to synthesize waveforms using the technique of par-
tial transmit sequence. The key point is that the gradient can be explicitly
derived from the objective function. Furthermore, the result is extended by
allowing the power spectrum to deviate from its original shape, yielding a
further reduction in the PAPR. The method is applied to derived power spec-
tra for wideband multiple-input-multiple-output radar. It is shown that the
proposed technique can achieve optimal or near optimal performance with a
PAPR below 0.5 dB.

Paper III Wideband Waveform Design for Robust Target Detection

To be published in IEEE International Conference on Acoustics, Speech, and
Signal Processing, April 2015, Brisbane, Australia.

Future radar systems are expected to use waveforms of a high bandwidth.

An advantage is an improved range resolution. Herein, a technique to design
robust wideband waveforms for target detection is developed. The context is
detection of a single object with partially unknown parameters. The wave-
forms are robust in the sense that, for a single transmission, detection capa-
bility is maintained over an interval of time-delay and time-scaling (Doppler)
parameters. A framework is derived, approximated, and formulated as an
optimization by means of basis expansion. In terms of probabilities of detec-
tion and false alarm, numerical evaluation shows the efficiency of the method

54
6.1 Summary of the Appended Papers

when compared with a chirp signal and a Gaussian pulse.

Paper IV Wideband Waveform and Receiver Filter Bank Design for

Clutter Suppression
Submitted to JSTSP SI on Advanced Signal Processing Techniques for Radar
Applications.

Modern highly flexible wideband radar systems with complex receiver tech-
nologies raise a demand for advanced signal processing techniques. In this
paper, we propose two algorithms to select transmit waveforms and receiver
filters. The techniques are based on a clutter suppression criterion. For
the first algorithm, we employ an optimized filter bank, and for the sec-
ond algorithm, we employ a matched filter bank. Clutter suppression is
achieved by minimizing correlation between receiver filters and interfering
clutter echoes. The algorithm, for the optimized filter bank, is extended to
adapt the transmission scheme and receiver filters to a time-evolving sce-
nario. Adaptation parameters are based on estimates of a clutter map and
detected target characteristics. To estimate the clutter map we propose a
Kalman filter, whereas target parameters are found through the method of
least-squares. The efficiency of the algorithms and the adaptation scheme
are visualized through a numerical simulation. It is found that the jointly
optimized transmit waveforms and receiver filter bank outperforms the other
approaches at low Signal-to-Interference-and-Noise Ratio (SINR), whereas a
match filter gives equal performance at higher SINR. The chirp waveform is
only effective at very low probability of false alarm.

Paper V Waveform and Receiver Filter Selection for Clutter-Map Esti-

mation Based on an IMM Kalman Filter
To be submitted to IEEE Transactions on Aerospace and Electronic Systems.

An important function in a radar system is to fast and correctly estimate

the clutter map, and its distribution parameters. In this paper, we first in-
troduce an algorithm where the transmit waveforms and receiver filters are
optimized for estimation of the back scattering clutter coefficients. Second,
we propose to utilize an Interacting Multiple Model (IMM) Kalman filter,
before detection, to handle abrupt changes in the characteristics of the clut-
ter. The efficiency of the optimization and the accuracy of the estimation
are visualized through a numerical simulation, where the optimized system
is compared with a traditional Linear-Frequency-Modulated pulse. The es-
timate obtained from the IMM Kalman filter is compared with a traditional

55
 Chapter 6. Contributions and Future Work

Kalman filter, and a least-squares estimator.

6.2 Future Work

Our vision is to go towards an adaptive system, where the transmitted wave-
forms and the receiver filters constantly adapt to the prevailing conditions.
This thesis provides initial steps to make this happen. However, there are
still many obstacles to overcome.
As seen in Chapter 5, the waveforms are distorted if the magnitude does
not experience a low PAPR. This constraint should be introduced when de-
signing the waveforms. In addition, when synthesizing the waveforms, other
system requirements (not only PAPR) can be considered, for instance, in-
stantaneous bandwidth and spectral purity. Furthermore, a study of the
performance degradation when only values contained in a finite alphabet
may be used as phase angles, and not any arbitrary values contained in the
interval (0, 2π] could be conducted. Waveform optimization subject to the
finite alphabet constraint is an interesting, albeit difficult future direction.
Herein, one practical experiment was performed. However, this is only
one of many experiments that should be conducted. A next step would
be to transmit the designed waveforms, and to evaluate the performance
of the transmitter and receiver chain in an experimental setup. Also, a
measured radar channel should be introduced in proposed optimization rou-
tines. This is important as the radar scenarios, in this thesis, are based on
approximations of target, clutter, and jammer characteristics. To test the
proposed adaptation algorithm introduced in Paper IV, the measured radar
data should be obtained over time.
A deep study on a comparison of traditional radar and optimally selected
waveforms is necessary. This study should involve both the possible gains
from a smart waveform selection, and the difference in complexity. The pos-
sibility to handle sophisticated waveform selection will hopefully be possible
in a future system, as the signal generators and receiver function become
more advanced. This indicates that the increased complexity, related to
both the optimization of waveforms and receiver filters, might not be that
crucial. The work can then be expanded to also include characteristics re-
lated to the spatial angles into the optimization. This will probably increase
the computational complexity even more.
In Paper IV, it is seen that the proposed Kalman filter requires a couple
of pulses to estimate the clutter-map with a satisfactory accuracy. This
time should be minimized, and a possible solution, described in Paper V,
where an interacting multiple model Kalman Filter is promising. This filter

56
6.2 Future Work

uses two models, where the first has small model errors, and the second one
instead experiences larger model errors. This results in that we rely more
on data when the state predictions are noisy. Thus, the observed data is
more important. The method could be extended to invoke more models,
which, for example, takes into account different clutter distributions, such
as, reflections from sea and precipitation.
In Paper III, a robust target detection scheme is proposed. The receiver
filters are matched to the transmitted waveforms, and as seen in Paper IV,
an optimized receiver structure gives an enhanced performance in clutter
dominated environments. This structure can be analyzed for the evaluated
problem in Paper III.
Robustness features are important when designing waveforms and receiver
filters. In this thesis, we considered uncertainties in the pointing direction.
However, this is only one example of possible sources of errors. Regarding
the antenna elements, imperfections due to phase and amplitude errors, as
well as the effect of mutual coupling are of interest, and should be evaluated.

57
Bibliography

[1] M. Skolnik. Radar Handbook. The McGraw-Hill Companies, New York,

NY, 1976.
[2] Mark A Richards. Fundamentals of radar signal processing. Tata
McGraw-Hill Education, 2005.
[3] J.H. Winters, J. Salz, and R.D. Gitlin. The impact of antenna diversity
on the capacity of wireless communication systems. Communications,
IEEE Transactions on, 42(234):1740 –1751, feb/mar/apr 1994.
[4] D.W. Bliss and K.W. Forsythe. Multiple-input multiple-output (MIMO)
radar and imaging: Degrees of freedom and resolution. pages 54–59,
November 2003.
[5] N.H. Lehmann, A.M. Haimovich, R.S. Blum, and L. Cimini. High reso-
lution capabilities of MIMO radar. In Signals, Systems and Computers,
2006. ACSSC ’06. Fortieth Asilomar Conference on, pages 25 –30, 29
2006-nov. 1 2006.
[6] M. Skolnik. Introduction to Radar Systems. The McGraw-Hill Compa-
nies, New York, NY, 1981.
[7] James D Taylor. Introduction to ultra-wideband radar systems. CRC
press, 1994.
[8] H.A. Khan, W.Q. Malik, D.J. Edwards, and C.J. Stevens. Ultra wide-
band multiple-input multiple-output radar. In Radar Conference, 2005
IEEE International, pages 900–904, May 2005.
[9] M.G.M. Hussain. Ultra-wideband impulse radar- an overview of the
principles. Aerospace and Electronic Systems Magazine, IEEE, 13(9):9–
14, 1998.

58
BIBLIOGRAPHY

[10] L.G. Weiss. Wavelets and wideband correlation processing. Signal Pro-
cessing Magazine, IEEE, 11(1):13–32, 1994.

[11] D Oo North. An analysis of the factors which determine signal/noise

discrimination in pulsed-carrier systems. Proceedings of the IEEE,
51(7):1016–1027, 1963.

[12] Milton Abramowitz and Irene A. Stegun. Handbook of Mathematical

Functions with Formulas, Graphs, and Mathematical Tables. Dover Pub-
lications, 1970.

[13] Steven M Kay. Fundamentals of statistical signal processing: Detection

theory, vol. 2, 1998.

[14] Ake Björck. Numerical methods for least squares problems. Siam, 1996.

[15] D.H. Brandwood. A complex gradient operator and its application in

adaptive array theory. Microwaves, Optics and Antennas, IEE Proceed-
ings H, 130(1):11–16, February 1983.

[16] David W Hosmer, Stanley Lemeshow, and Rodney X Sturdivant. Intro-

duction to the logistic regression model. Wiley Online Library, 2000.

[17] Steven M Kay. Fundamentals of statistical signal processing: Estimation

theory, vol. 1, 1993.

[18] Rudolph Emil Kalman. A new approach to linear filtering and prediction
problems. Journal of Fluids Engineering, 82(1):35–45, 1960.

[19] John G Proakis. Digital signal processing: principles algorithms and

applications. Pearson Education India, 2001.

[20] R.G. Brown and P.Y.C. Hwang. Introduction to Random Signals and
Applied Kalman Filtering with Matlab Exercises. CourseSmart Series.
Wiley, 2012.

[21] P. Kaminski, Arthur E. Bryson, and S. Schmidt. Discrete square root

filtering: A survey of current techniques. Automatic Control, IEEE
Transactions on, 16(6):727–736, Dec 1971.

[22] Peter S Maybeck. Stochastic models, estimation, and control, volume 3.

Academic press, 1982.

[23] A.H. Jazwinski. Stochastic Processes and Filtering Theory. Dover Books
on Electrical Engineering Series. Dover Publications, 2007.

59
 BIBLIOGRAPHY

[24] Simon J. Julier and Jeffrey K. Uhlmann. New extension of the kalman
filter to nonlinear systems. Proc. SPIE, 3068:182–193, 1997.

[25] Fredrik Gustafsson. Particle filter theory and practice with position-
ing applications. Aerospace and Electronic Systems Magazine, IEEE,
25(7):53–82, 2010.

[26] Robert J. Mailloux. Phased Array Antenna Handbook. Artech House

Inc., 1 edition, 1994.

[27] Prabhakar S. Naidu. Sensor Array Signal Processing. CRC Press, 2000.

[28] H. L. Van Trees. Optimum Array Processing- Part IV, Detection, Esti-
mation, and Modulation Theory. John Wiley & Sons, 2002.

[29] H. Krim and M. Viberg. Two decades of array signal processing research:
the parametric approach. Signal Processing Magazine, IEEE, 13(4):67
–94, jul 1996.

[30] Jian Li and Petre Stoica. Robust adaptive beamforming. John Wiley
and Sons, Inc., 2006.

[31] Matsuo Sekine and Yuhai Mao. Weibull Radar Clutter. Number 3. IET,
1990.

[32] Robert N. Lothes, Michael B. Szymanski, and Richard G.Wiley. Radar

vulnerability to jamming. Artech House, 1990.

[33] Li Neng-Jing and Zhang Yi-Ting. A survey of radar ECM and ECCM.
Aerospace and Electronic Systems, IEEE Transactions on, 31(3):1110
–1120, jul 1995.

[34] David Knox Barton. Modern radar system analysis. Artech House, Inc.,
1988.

[35] H. Naparst. Dense target signal processing. Information Theory, IEEE

Transactions on, 37(2):317–327, Mar 1991.

[36] R.K. Young. Wavelet Theory and Its Applications. Kluwer international
series in engineering and computer science: VLSI, computer architec-
ture, and digital signal processing. Springer US, 1993.

[37] J. Bergh, F. Ekstedt, and M. Lindberg. Wavelets. Studentlitteratur,

1999.

60
BIBLIOGRAPHY

[38] R. McAulay and J. Johnson. Optimal mismatched filter design for radar
ranging, detection, and resolution. Information Theory, IEEE Transac-
tions on, 17(6):696–701, Nov 1971.
[39] H. Rohling and W. Plagge. Mismatched-filter design for periodic binary
phased signals. Aerospace and Electronic Systems, IEEE Transactions
on, 25(6):890–897, Nov 1989.
[40] I. Frigyes and A.J. Seeds. Optically generated true-time delay in phased-
array antennas. Microwave Theory and Techniques, IEEE Transactions
on, 43(9):2378 –2386, sep 1995.
[41] E. Fishler, A. Haimovich, R. Blum, D. Chizhik, L. Cimini, and R. Valen-
zuela. MIMO radar: an idea whose time has come. In Proc. of the IEEE
Int. Conf. on Radar, Philadelphia, PA, April 2004.
[42] D. E. Vakman. Sophisticated Signals and the Uncertainty Principle in
Radar. Springer Verlag Berlin, 1968.
[43] E. Fishler, A. Haimovich, R.S. Blum, Jr. Cimini, L.J., D. Chizhik, and
R.A. Valenzuela. Spatial diversity in radars-models and detection per-
formance. Signal Processing, IEEE Transactions on, 54(3):823 – 838,
march 2006.
[44] Qian He, Nikolaus H. Lehmann, Rick S. Blum, and Alexander M.
Haimovich. MIMO radar moving target detection in homogeneous
clutter. Aerospace and Electronic Systems, IEEE Transactions on,
46(3):1290 –1301, july 2010.
[45] Jian Li, P. Stoica, Luzhou Xu, and W. Roberts. On parameter iden-
tifiability of MIMO radar. Signal Processing Letters, IEEE, 14(12):968
–971, dec. 2007.
[46] D. R. Fuhrmann and G. San Antonio. Transmit beamforming for mimo
radar systems using partial signal correlations. In 38th Asilomar Con-
ference on Signals, Syst. and Comput., Pacific Grove, CA, November
2004.
[47] J. Li, P. Stoica, and Y. Xie. On probing signal design for MIMO radar.
IEEE Trans. on Sig. Process., 55:4151–4161, August 2007.
[48] G. San Antonio and D.R. Fuhrmann. Beampattern synthesis for wide-
band mimo radar systems. In 1st IEEE Int. workshop on Comp. Ad-
vances in Multi-Sensor Adaptive process., pages 105–108, Puerto Val-
larta, Mexico, December 2005.

61
 BIBLIOGRAPHY

[49] Hao He, Petre Stoica, and Jian Li. Wideband mimo systems: signal
design for transmit beampattern synthesis. IEEE Trans. Signal Process.,
59:618–628, February 2011.
[50] L. Collin, O. Berder, P. Rostaing, and G. Burel. Optimal minimum
distance-based precoder for MIMO spatial multiplexing systems. Signal
Processing, IEEE Transactions on, 52(3):617 – 627, march 2004.
[51] A. Scaglione, P. Stoica, S. Barbarossa, G. B. Giannakis, , and H. Sam-
pat. Optimal designs for space time linear precoders and decoders. IEEE
Trans. Signal Process., 50:1051–1064, May 2002.
[52] V. Havary-Nassab, S. Shahbazpanahi, A. Grami, and Z.-Q Luo. Dis-
tributed beamforming for relay networks based on second-order statistics
of the cannel state information. IEEE Trans. Signal Process., 56:4306–
4316, September 2008.
[53] C.Y. Chen and P P. Vaidyanathan. Mimo radar waveform optimization
with prior information of extended target and clutter. IEEE Trans.
Signal Process., 57:3533–2543, September 2009.
[54] S.U. Pillai, H.S. Oh, D.C. Youla, and J.R. Guerci. Optimal transmit-
receiver design in the presence of signal-dependent interference and
channel noise. Information Theory, IEEE Transactions on, 46(2):577
–584, mar 2000.
[55] D. DeLong and E.M. Hofstetter. On the design of optimum radar wave-
forms for clutter rejection. Information Theory, IEEE Transactions on,
13(3):454–463, 1967.
[56] B. Friedlander. Waveform design for MIMO radars. Aerospace and
Electronic Systems, IEEE Transactions on, 43(3):1227 –1238, july 2007.
[57] J.R. Guerci, M.C. Wicks, J.S. Bergin, P.M. Techau, and S.U. Pillai.
Theory and application of optimum and adaptive mimo radar. In Radar
Conference, 2008. RADAR ’08. IEEE, pages 1 –6, may 2008.
[58] Bo Jiu, Hongwei Liu, Dazheng Feng, and Zheng Liu. Minimax robust
transmission waveform and receiving filter design for extended target
detection with imprecise prior knowledge. Signal Processing, 92(1):210
– 218, 2012.
[59] Zhi-Quan Luo, Wing-Kin Ma, A.M.-C. So, Yinyu Ye, and Shuzhong
Zhang. Semidefinite relaxation of quadratic optimization problems. Sig-
nal Processing Magazine, IEEE, 27(3):20–34, May 2010.

62
BIBLIOGRAPHY

[60] Shuzhong Zhang. Quadratic maximization and semidefinite relaxation.

Mathematical Programming, 87(3):453–465, 2000.

[61] Paul Tseng. Further results on approximating nonconvex quadratic op-

timization by semidefinite programming relaxation. SIAM Journal on
Optimization, 14(1):268–283, 2003.

[62] Stephen Boyd and Lieven Vandenberghe. Convex Optimization. Cam-

bridge University Press, New York, NY, USA, 2004.

[63] C. Kelley. Iterative Methods for Optimization. Society for Industrial and
Applied Mathematics, 1999.

[64] Miguel Sousa Lobo, Lieven Vandenberghe, Stephen Boyd, and Hervé
Lebret. Applications of second-order cone programming. Linear algebra
and its applications, 284(1):193–228, 1998.

[65] Farid Alizadeh and Donald Goldfarb. Second-order cone programming.

Mathematical programming, 95(1):3–51, 2003.

[66] S.A. Vorobyov, A.B. Gershman, and Zhi-Quan Luo. Robust adaptive
beamforming using worst-case performance optimization: a solution to
the signal mismatch problem. Signal Processing, IEEE Transactions on,
51(2):313 – 324, feb 2003.

[67] O. Besson, A.A. Monakov, and C. Chalus. Signal waveform estima-

tion in the presence of uncertainties about the steering vector. Signal
Processing, IEEE Transactions on, 52(9):2432 – 2440, sept. 2004.

[68] Yang Yang and R.S. Blum. Minimax robust MIMO radar waveform
design. Selected Topics in Signal Processing, IEEE Journal of, 1(1):147
–155, june 2007.

[69] M.H. Er and B.C. Ng. A robust method for broadband beamforming in
the presence of pointing error. Signal Processing, 30(1):115 – 121, 1993.

[70] John Dahl and Sebastian Holmqvist. The effects of the power amplifier
on wideband radar signals, 2013.

[71] M. R. Schroeder. Synthesis of low-peak-factor signals and binary se-

quences with low autocorrelation. IEEE Trans. Inf. Theory, IT-16:85–
89, January 1970.

[72] S. Boyd. Multitone signals with low crest factor. IEEE Trans. Circuits
and Syst., 33(10):1018 – 1022, oct 1986.

63
 BIBLIOGRAPHY

[73] A. Van Den Bos. A new method for synthesis of low-peak-factor signals.
IEEE Trans. Acoust., Speech Signal Process., 35(1):120 – 122, jan 1987.

[74] E. Van der Ouderaa, J. Schoukens, and J. Renneboog. Peak factor

minimization of input and output signals of linear systems. IEEE Trans.
Instrum. Meas., 37(2):207 –212, jun 1988.

[75] M. Friese. Multitone signals with low crest factor. IEEE Trans. Com-
mun., 45(10):1338 –1344, oct 1997.

[76] M. Ström and M. Viberg. Constant modulus waveform synthesis with

application to wideband MIMO radar. In presented at the Swedish Radio
and Microwave Days, Stockholm, Sweden, march 2012.

64