Sidelobe blanking in the presence

of noise-like interference

Master’s thesis in Applied Physics

Stina Wahlgren

Department of Space, Earth and Environment

C HALMERS U NIVERSITY OF T ECHNOLOGY
Gothenburg, Sweden 2018
 Master’s thesis 2018

Sidelobe blanking in the presence of noise-like

interference

STINA WAHLGREN ∗

Department of Space, Earth and Environment

Division of Microwave and Optical Remote Sensing
Chalmers University of Technology
Gothenburg, Sweden 2018

∗
Contact: [email protected]
Sidelobe blanking in the presence of noise-like interference
STINA WAHLGREN

© STINA WAHLGREN, 2018.

Supervisor: Björn Hallberg, Saab Surveillance

Supervisor: Erik Blomberg, Department of Space, Earth and Environment
Examiner: Lars Ulander, Department of Space, Earth and Environment

Master’s Thesis 2018

Department of Space, Earth and Environment
Division of Microwave and Optical Remote Sensing
Chalmers University of Technology
SE-412 96 Gothenburg
Telephone +46 31 772 1000

Typeset in LATEX
Gothenburg, Sweden 2018

Sidelobe blanking in the presence of noise-like interference

STINA WAHLGREN
Department of Space, Earth and Environment
Chalmers University of Technology

iv
Abstract
Interference suppression in antenna arrays is an important topic for target detection
in radar systems. This thesis investigates the performance of two different algo-
rithms for determining whether a detection lies in the sidelobe or mainlobe of an
antenna in the presence of noise-like interference (NLI). Both algorithms are based
on a combination of sidelobe blanking (SLB) and sidelobe cancellation (SLC). The
first algorithm is a simple cascading of SLC and SLB, while the second algorithm is
a modification of this where multiple guard channels are used. The performance is
evaluated with Monte Carlo simulations, where the antenna is modelled as a linear
phased-array antenna and interference and target signals are modelled as uncorre-
lated, single-frequency plane waves. The influence of covariance estimation error,
number of NLI sources, placement of auxiliary antennas, signal-to-noise ratio (SNR)
and jammer-to-noise ratio (JNR) on performance and suitable blanking threshold is
investigated. It is shown that the second algorithm performs better than the first
when the number of NLI sources are few relative to the number of auxiliary anten-
nas, and reduces to the same performance as the first algorithm when this is not
the case.

Keywords: antenna-related ECCM, sidelobe blanking, sidelobe cancellation, side-

lobe jamming.

v
Sidelobe blanking in the presence of noise-like interference
STINA WAHLGREN
Department of Space, Earth and Environment
Chalmers University of Technology

Sammanfattning
Störundertryckning i antenner är viktigt för att kunna detektera mål korrekt i
radarsystem. I det här projektet har prestandan hos två olika algoritmer under-
sökts, med syfte att avgöra om en detektion kommer från en antenns huvudlob eller
sidlob, då antennen belyses av störare. Båda algoritmerna bygger på en kombination
av sidlobsblankning (SLB) och adaptiv sidlobsundertryckning (ASLU). Den första
algoritmen är en enkel kaskadkoppling av SLB och ASLU, medan den andra algo-
ritmen är en modifikation av den förra med multipla vaktkanaler. Prestandan har
utvärderas med hjälp av Monte Carlo-simuleringar, där antennen modellerats som
en linjär gruppantenn bestående av isotropa antennelement, medan stör- och målsig-
naler modellerats som okorrelerade, monokroma, plana vågor. Effekten av kovar-
iansskattningsfel, antal störare, placering av hjälpantenner, signal-brusförhållande
och stör-brusförhållande på prestanda och lämpliga tröskelvärden för SLB har un-
dersökts. Den andra algoritmen presterar bättre än den första då antalet störare
är få relativt antalet hjälpantenner. Skillnaden i prestanda mellan algoritmerna
minskar då antalet störare ökar.

Keywords: störskydd, sidlobsblanking, adaptiv sidlobsundertryckning, sidlobsstörn-

ing.

vi
Acknowledgements
This work has been carried out with support from Saab Surveillance and the Depart-
ment of Space, Earth and Environment at Chalmers University of Technology. Sin-
cere thanks goes to my supervisor Björn Hallberg and the recruiter Tomas Berling
at Saab, as well as the examiner Lars Ulander and supervisor Erik Blomberg at
Chalmers. I would also like to thank Anders Hellman and Beng-Erik Mellander, for
making this thesis possible.

Stina Wahlgren, Gothenburg, March 2018

vii
 Contents

1 Introduction 1
1.1 Phased-array antennas . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.1.1 Phase shift between two antenna elements . . . . . . . . . . . 3
1.2 Electronic counter-countermeasures . . . . . . . . . . . . . . . . . . . 3
1.2.1 Sidelobe Blanking . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2.2 Sidelobe cancellation . . . . . . . . . . . . . . . . . . . . . . . 5
1.2.2.1 On the number of possible nulls . . . . . . . . . . . 6
1.2.3 SLC in guard channel . . . . . . . . . . . . . . . . . . . . . . . 6
1.3 Aim . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.4 Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

2 Methods 9
2.1 Tested guard functions . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2 Evaluation of performance . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2.1 Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2.1.1 Signal-to-noise ratio . . . . . . . . . . . . . . . . . . 12
2.2.2 The covariance matrix . . . . . . . . . . . . . . . . . . . . . . 12
2.2.2.1 Estimation error model . . . . . . . . . . . . . . . . 14
2.2.3 Simulation method . . . . . . . . . . . . . . . . . . . . . . . . 14

3 Results 17
3.1 The undisturbed case . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.2 Effect of auxiliary antenna position . . . . . . . . . . . . . . . . . . . 17
3.3 Comparison between methods . . . . . . . . . . . . . . . . . . . . . . 21
3.4 Effect of JNR and SNRsum . . . . . . . . . . . . . . . . . . . . . . . . 22
3.5 Effect of covariance estimation error . . . . . . . . . . . . . . . . . . . 22
3.6 More than one jammer . . . . . . . . . . . . . . . . . . . . . . . . . . 25

4 Discussion 29
4.1 Different approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

5 Conclusion 31
5.1 Future work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

Bibliography 33

ix
Contents

x
 Nomenclature

Abbreviations
ECCM Electronic counter-countermeasures
ECM Electronic countermeasures
JNR Jammer-to-noise ratio
NLI Noise-like interference
SLB Sidelobe blanking
SLC Sidelobe cancellation
SNR Signal-to-noise ratio
Operators
(·) Complex conjugate
Cov[ · ] Covariance
(·)H Hermitian transpose
(·)T Transpose
E[ · ] Expected value
diag(v) Diagonal matrix with diagonal entries v
Variables
θ Angle of incidence
J Number of NLI-jammers
M Number of auxiliary antennas
N Number of antenna elements
u Sinus of angle of incidence
Y Youden’s index
εcal Calibration errors
Q Covariance matrix

xi
Contents

R Covariance vector
SNRsum SNR in main channel
σcal Standard deviation of calibration error
σest Standard deviation of covariance estimation error
σth Standard deviation of thermal noise
Va Signals from auxiliary antennas
Vg Guard channel signal
Ving Signal from guard antenna
Vin Signal from main antenna before SLC
Vm Signal from main antenna after SLC
wtap Tapering weights

xii
 1
Introduction

A radar system uses electromagnetic energy in the radio- or microwave range to

measure range and direction to targets. The basic principle is that energy pulses are
transmitted in a certain direction and the time between transmission and detection
of reflected waves is used to measure the distance to the reflecting objects.
The first radar like invention, the telemobiloscope, was patented in 1904 by the
German physicist Christian Hülsmeyer, but it would take until shortly before the
second world war until the first effective radar systems were developed. The first
radar systems were primarily a British invention, aimed as an early warning sys-
tem for detecting bombers [1]. The first systems used enormous antennas, with
wavelengths of around 10 metres. During the war, new inventions made it possible
to use shorter wavelengths (centimetres), which opened up new military applica-
tions. Radar technology has continued to evolve. Today radars are used in a variety
of fields, both military applications, such as surveillance, missile defence systems,
missile launching systems, and civil applications, for instance flight radar, weather
radar or radars in modern cars. Other applications, such as biomass measurements
of forests [2] or breast tumour detection [3], are still at the research stage.
A highly simplified block diagram of the general principles of a modern radar
system is shown in Figure 1.1. The general components of a radar system can be
divided into: a transmitter, a transmitting antenna, a receiving antenna, a receiver,
a signal processing unit and in general a data processing unit [4]. The purpose
of the transmitter is to generate suitable waveforms that the transmitting antenna
propagates in the desired direction. A receiving antenna collects the reflected energy
and passes the signals to the receiver. In monostatic radars the same antenna is used
for both transmission and reception. In the signal processing step, data, such as if
a target was detected and if so at what range the target lies, is extracted from the
received signals. The data processing step includes for example target tracking.
This thesis is focused at the signal processing step, or more precisely at the task of

antenna transmitter

antenna receiver signal processing data processing

Figure 1.1: General design of a modern radar system.

1
1. Introduction

40
V1 V2 V3 V4 V5 V6
20

gain (dB)
w1 w2 w3 w4 w5 w6 0

−20
P
−40
−1 −0.5 0 0.5 1
u
Vin
(a) (b)

Figure 1.2: Block diagram of a uniform linear antenna array with 6 antenna
elements (a) and comparison between uniform tapering (blue) and Taylor tapering
(black) (b). u is sinus of the incidence angle, i.e. the angle between the plane wave
and the linear array.

detection decision.

1.1 Phased-array antennas

A common antenna type used in radar technology is the phased-array antenna. It
consists of many smaller antennas, called antenna elements, where the phase and
amplitude of each element can be individually controlled [5]. The most common
antenna elements are slots, dipoles, open-ended waveguides or patches [5]. In this
project they are treated as points with an isotropic antenna pattern, that is they have
the same gain in all directions.1 The main advantage of phased-array antennas is
that the beam can be steered electronically, which reduces the time it takes to change
beam direction to a fraction of the time it would take to change the orientation of
the antenna mechanically. This makes the antenna much more flexible [5]. Phased-
array antennas come in many forms. Depending on the application, the antenna
elements can be arranged in one or two dimensions, the surface flat or curved and
the spacing between elements uniform or irregular. In this thesis, a one-dimensional,
linear, uniform antenna array is used.
A schematic illustration of a phased-array antenna is shown in Figure 1.2a. The
signals received by the antenna elements are combined to a main channel, with a
higher directivity. Steering of the beam is achieved by phase shifting the signals
from the antenna elements in such a way that signals from the desired direction
experience constructive interference. Uniqueness is achieved by ensuring that the
antenna elements are at most half a wavelength apart.
An ideal radar beam is as narrow as possible, for a high resolution in direction,
and with as low sidelobe levels as possible. In order to reduce the sidelobe levels,
1
Note that like spherical cows, such antennas do not exist in the real world. In contrast to
spherical cows however, this simplification is in general considered useful.

2
 1. Introduction

tapering can be applied to phased-array antennas. Tapering is simply to change the

amplitude of the signals from the antenna elements before combining them to the
main channel. Tapering always comes with the cost of broadening of the mainlobe.
A common tapering in radar technology is Taylor tapering, which provides a strong
sidelobe suppression with a minimum broadening of the mainlobe [6]. It is defined
by the following three parameters: (i) the number of antenna elements, N , (ii)
the desired maximum sidelobe level, SLL, and (iii) the number of nearly constant
sidelobes adjacent to the mainlobe, n̄. Figure 1.2b shows the effect of Taylor tapering
on a typical phased-array antenna pattern.
In theory, each antenna element can have its own receiver, which opens up a lot of
possibilities, such as multiple simultaneous beams. Due to practical limitations, this
is rarely the case in real radar systems. In general, beamforming is made before the
receiver, which means that only combined channels are accessible for digital signal
processing.

1.1.1 Phase shift between two antenna elements

The phase shift of an incident plane wave between
two antenna elements is used later in this report, so
we make a quick derivation of it here. Suppose the t
fron
incoming wave has an incidence angle θ and wave- a ve d˜sin θ
length λ, as illustrated in Figure 1.3. If the distance w
between the antenna elements is d, ˜ the phase shift, θ
ϕ, between the signals from the two antenna ele- d˜
ments is
2π ˜ Figure 1.3: Schematic illus-
ϕ= d sin(θ)
λ tration of an incident plane
If the distance between adjacent antenna elements wave. The black dots represent
is λ/2, the equation above simplifies to two antenna elements.
ϕ = πd sin(θ), (1.1)

where the integer d is the number of half wavelengths between the two antenna
elements. Due to the fact that the phase shift depends on sinus of angle of incidence,
it is common to introduce the variable u = sin θ.

1.2 Electronic counter-countermeasures

Electronic counter-countermeasures (ECCM) is an umbrella term for techniques
developed in order to avoid or reduce the effect of electronic countermeasures (ECM).
The purpose of ECM techniques is to deny the victim radar information or supply
it with misleading information [7][8]. There are numerous different ECM methods,
ranging from releasing a cloud of thin metal foils (chaff) in order to screen the target
or otherwise confuse the radar system, to active transmission of signals. One class of
ECM is the so called jamming, which can be defined as the intentional and deliberate
transmission or retransmission of signals for the purpose of disturbing or deceiving
radar systems [8].

3
1. Introduction

One problem that arises in radar technology is that the antenna typically has
sidelobes, which means that energy arriving at the antenna from a different angle
than the mainlobe direction can enter the receiver, although with a lower gain
than energy entering via the mainlobe. This phenomenon is exploited in sidelobe
jamming. This thesis is concerned with two types of sidelobe jamming, namely
false target jamming and noise-like interference (NLI). In false target jamming the
jammer transmits (or retransmits) energy pulses, which if they enter the receiver
via antenna sidelobes can be misinterpreted as a weaker signal entering via the
mainlobe. Thus, the effect of false target jamming is to introduce false detections.
NLI on the other hand is continuous signals, which if they enter the antenna via
sidelobes cause an overall increase of the noise level, i.e. decrease in SNR [9]. Thus,
the effect of NLI is to make it harder to detect targets.
Two ways to handle the types of sidelobe jamming mentioned above is sidelobe
blanking (SLB), aimed at false target jamming, and sidelobe cancellation (SLC),
aimed at NLI [7]. Both methods belong to the class of antenna-related ECCM, that
is they use additional antennas in order to reduce the effect of jamming [8]. The
additional antennas are often called guard antennas, used for SLB, and auxiliary
antennas, used for SLC. In the rest of the section, both methods are treated in some
detail.
Although the discussion here regards deliberate interference of the radar, the
same principles applies to non-malicious sidelobe interference, such as other radars,
strong scatterers in sidelobe directions, or even radio frequency interference from for
instance satellites or the sun, which can be a problem in radio astronomy [10].

1.2.1 Sidelobe Blanking

SLB is a well established technique in radar for cancelling (intentional or uninten-
tional) impulsive interference, for instance false target jammers [11]. In SLB an
auxiliary antenna, the guard antenna, with a more isotropic antenna pattern is used
in addition to the main antenna [8]. The amplitude of the received signal from the
main antenna can then be compared with the amplitude of the received signals from
the guard antenna, which makes it possible to determine if the signal entered the
main antenna via a sidelobe or the mainlobe. More precisely, if Vm is the signal
from the main antenna and Vg is the signal from the guard antenna, the blanking
logic scheme is simply

|Vm |
≥ T no blanking



 |Vg |



(1.2)
 |Vm |



< T blanking,



|Vg |

where T is the blanking threshold. An illustration of schematic antenna pattern for

the main- and guard antenna is shown in Figure 1.4a.
In the absence of NLI, SLB has been shown to be effective against false targets
with small loss in true detections [12].

4
 1. Introduction

40 40

20 20
gain (dB)

gain (dB)
0 0

−20 −20

−40 −40
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
u u

(a) Illustration of the antenna (b) Illustration of the antenna

pattern of a main antenna (black) pattern before (blue) and after
and a guard antenna (dashed) used (black) SLC introducing a null in the
for SLB. marked direction.

Figure 1.4: Schematic illustrations of SLB and SLC.

1.2.2 Sidelobe cancellation

The idea behind SLC is to use signals from auxiliary antennas to adaptively cancel
out signals from the directions of the jammers, thus introducing nulls in the antenna
pattern in those directions [7]. This will distort the antenna pattern in other direc-
tions as well, but the distortion of the mainlobe is in general negligible assuming
the gain of the sidelobes is low. An illustration of the effect of SLC in an antenna
pattern is shown in Figure 1.4b.
The challenge of SLC is to find the complex weights to the signals from the
auxiliary antennas so that jammer signal is nullified. This is accomplished by finding
the weights that minimise the output power, Pm , of the main channel when no target
signal is present [13]. h i
Pm = E V m V m ,
where Vm is the signal of the main channel after SLC. Vm is given by

Vm = Vin − w · Va ,

where Vin is the signal from the main antenna, w is the weights and Va is the signals
from the auxiliary antennas. Under the assumption that the jammer signals and
target signal are statistically independent, the weights w that minimise Pm is the
weights that will be best at cancelling out jammer signals, in the sense that they
will cause the maximum SNR.
One set of weights that minimise Pm is

w = Q-1 R, (1.3)

where Q is the covariance matrix of the (noise and jammer only) signals from the
auxiliary antennas, with elements

Qnm = Cov[Van , Vam ] ,

5
1. Introduction

and R is the covariance vector, with elements

Rn = Cov[Vin , Van ] .

For a derivation of this result, see for example [13].

1.2.2.1 On the number of possible nulls

Suppose a system with M auxiliary antennas is illuminated with NLI jammers from
J different directions. The problem of finding weights w for the signals Va that null
the signal Vm in the directions {ui } of the jammers can be described as a system of
linear equations


 Vin (u1 ) − w1 V1 (u1 ) − w2 V2 (u1 ) − · · · − wM VM (u1 ) =0
 Vin (u2 ) − w1 V1 (u2 ) − w2 V2 (u2 ) − · · · − wM VM (u2 )


=0
..




.
Vin (uJ ) − w1 V1 (uJ ) − w2 V2 (uJ ) − · · · − wM VM (uJ ) = 0


or equivalently     
V1 (u1 ) . . . VM (u1 ) w1 Vin (u1 )
.. ... ..    ..  =  . 
.   .   .. 
   

 .  (1.4)
V1 (uJ ) . . . VM (uJ ) wM Vin (uJ )
Call the matrix in the latter equality A. It is clear that the problem is solvable if
rank(A) ≥ J, and that if rank(A) > J the weights w can be chosen in infinitely
many ways. Note that A depends on both placement of auxiliary antennas and
jammer directions.

1.2.3 SLC in guard channel

In contrast to SLC in an antenna with high directivity, SLC in a low-directivity
antenna significantly distorts the antenna pattern in other directions as well. Since
SLB is built on the comparison of signals from high- and low-directivity antennas,
it is reasonable to expect the ability to determine whether a target is located in the
mainlobe or not with SLB to change in the presence of NLI. The combination of
SLC and SLB has been studied in [14], where it is shown that cascading SLC and
SLB works well for the case of one jammer, one auxiliary antenna and no calibra-
tion or covariance estimation errors. However, we are not aware of any published
results regarding performance of combined SLC and SLB methods for more complex
scenarios.

1.3 Aim
The aim of this master’s thesis is to investigate how a number of isotropic auxiliary
antennas can be used to determine whether a target signal is located in the mainlobe
or in a sidelobe of the main antenna, when the main antenna is illuminated by one
or more NLI jammers. Such algorithms are called guard functions for the rest of
this report.

6
 1. Introduction

1.4 Limitations
The focus of this project lies on general principles of SLB in the presence of NLI,
therefore implementation details are not treated. The guard functions are evaluated
using simulations. Targets and jammers are assumed to be far away from the radar,
and jammers entering the mainlobe of the radar antenna are not considered. All
target and jammer signals are assumed to be uncorrelated and monochromatic.
Clutter is not modelled. Other signal processing than guard functions, e.g. pulse
compression or doppler processing, are not modelled.

7
1. Introduction

8
 2
Methods

2.1 Tested guard functions

Three different guard functions are evaluated in this report, here called: No SLCg,
SLCg and SLCg max. They all use sidelobe blanking logic presented in (1.2) and
are compared to the same sum channel Vm . Block diagrams of the three algorithms
are presented in Figure 2.1. The difference between the three methods is how Vg
is defined. In No SLCg Vg is simply the signal from the guard antenna. In SLCg,
Vg is the signal from the guard antenna after sidelobe cancellation using Va . SLCg
max is similar to the second algorithm, but here all auxiliary antennas and the
guard antennas are used one at a time as the guard antenna, and the other used
as auxiliary antennas in SLC. Thus M + 1 potential guard channels Vg i are created
and Vg is chosen as the maximum of those:

Vg = max{Vg i }
i

The motivation behind SLCg max is to reduce anisotropy in the guard channel in
non-jammer directions. Example of guard channel patterns are shown in Figure 2.2.
SLCg was proposed in [14] and SLCg max is a modification of another algorithm
presented in the same paper. In both SLCg and SLCg max the weights of the guard
and auxiliary antennas are normalised, so that
M
X +1
wi wi = 1. (2.1)
i=1

Note that when J = M , no more than one combination of weights that fulfils
both (1.4) and (2.1), i.e. normalised weights that both places nulls in all jammer
directions, can exist. This means that SLCg max and SLCg will produce the same
result when J = M , provided that solutions to (1.4) exists for all choices of guard
antennas.

2.2 Evaluation of performance

The purpose of the investigated methods is to determine whether a potential de-
tection, i.e. a signal in the main channel significantly higher than the noise level,
comes from the mainlobe or the sidelobes of the antenna. For the rest of the report
potential detections stemming from signals entering the antenna via the mainlobe

9
2. Methods

Vin Va Ving Vin Va Ving

SLC SLC SLC

Vg

Vm Vm Vg

SLB SLB

detection detection
(a) No SLCg (b) SLCg

Vin Va Ving

SLC
SLC
cycle guard
Vg1 ... VgM +1
Vm max
Vg

SLB

detection
(c) SLCg max

Figure 2.1: Schematic illustrations of the evaluated guard functions. Vin is the
signal from the main antenna, Va the signals from the auxiliary antennas and Ving
the signal from the guard antenna.

10
 2. Methods

40 40

20 20
gain (dB)

gain (dB)
0 0

−20 −20

−40 −40
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
u u

(a) SLCg (b) SLCg max

Figure 2.2: Example of the antenna patterns of guard channels produced by SLCg
and SLCg max. In b), the antenna pattern for each choice of guard channel is shown
in colours and the maximum of those is marked with black.

are called true targets while those stemming from signals entering the antenna via
sidelobes are called false targets.
In order to evaluate the performance of the methods, we focus on two quantities,
namely the probability of detection, Pd , and the probability of false alarm, Pfa . Pd
is defined as the probability of a true target correctly classified as a true target and
Pfa is the probability of a false target falsely classified as a true target. Using the
blanking logic in (1.2) results in the following expressions
  
|Vm |
Pd (T ) =P > T  true target

|Vg |

  
|Vm |
Pfa (T ) = P |Vg |
> T  false target

where T is the blanking threshold.

2.2.1 Model
The main antenna is modelled as a linear phased-array antenna with 100 antenna
elements, modelled as points with isotropic antenna patterns for all polarisations.
The antenna is tapered using a Taylor tapering with SLL = −40 dB and n̄ = 6. One
of the antenna elements is also used as a guard antenna and three other antenna
elements are also used as auxiliary antennas. Thus, in total the system consists
of one main antenna and four additional antennas. Target and jammer signals are
modelled as plane waves. All waves have the same wavelength λ0 , and the antenna
elements are spaced λ0 /2 apart. Each wave has a random phase drawn from a
uniform distribution on [0, 2π].
The output, Vi , from antenna element i is a complex voltage, consisting of the
(true or false) target signal si , the NLI jammer signals bi , thermal noise ni , and a
calibration error εcal
i .  
Vi = (si + bi + ni ) 1 + εcal
i

11
2. Methods

Both the thermal noise and the calibration error are modelled as zero-mean, complex,
gaussian noise, uncorrelated between antenna elements. The thermal noise has
a standard deviation σth = 1 and the calibration error has a standard deviation
σcal = 0.02. The calibration error is added in order to avoid unrealistically low
sidelobe levels.
The antenna array will always be steered at boresight, so the main channel output
(before SLC) is
N
witap Vi ,
X
Vin =
i=1

where witap is the tapering weight of antenna element i.

2.2.1.1 Signal-to-noise ratio

In this report, signal-to-noise-ratio (SNR) and jammer-to-noise-ratio (JNR) refer to
the ratio for one antenna element. They are defined as
!
at 2
SNR = 10 log10 2
dB
σth
!
aj 2
JNR = 10 log10 2
dB,
σth
2
where at and aj are the amplitudes of the target and jammer respectively and σth
is the variance of the thermal noise in that antenna element. Note that SNR in
the main channel, SNRsum , differs from SNR for one antenna element. SNRsum is
highly dependent on target direction relative to mainlobe direction. If the antenna
is steered at boresight, SNRsum is given by
 P 2 
 N jπut i 
 at  2w
i=1 i e  
SNRsum = 10 log10  2 PN 2 dB,
σth

i=1 |wi |

where wi is the weights of the antenna elements in the main channel.

2.2.2 The covariance matrix

As described in Section 1.2.2, a fundamental part of SLC is knowledge of the co-
variance of the unwanted part (i.e. inteference and noise) of the signals received by
the auxiliary antennas. Due to its importance in radar signal processing, the task
of estimating the covariance matrix has been thoroughly studied. However, covari-
ance estimation is outside the scope of this thesis, and analytical expressions for the
covariance matrix for the model problem are therefore derived in this section. In
order to study the effect of non-perfect estimations, a simple model for covariance
matrix estimation error is also introduced.
Let q = [q1 . . . qM ]T be the signals of jammers and thermal noise received by the
M auxiliary antennas. Suppose there are J NLI-jammers. The i:th element in q is
then given by
J
ak ej(ϕi (uk )+ξk ) + ni ,
X
qi = (2.2)
k=1

12
 2. Methods

where ak is the amplitude and ξk is the random phase of the k:th jammer signal and
ϕi (uk ) is the phase shift between the antenna element in question and a reference
point for a wave from direction uk . The expression for ϕi (uk ) is given in (1.1).
The covariance matrix Q is defined as:
 
Cov[q1 , q1 ] . . . Cov[q1 , qM ]

Q= .. .. .. 
,
 . . . 
Cov[qM , q1 ] . . . Cov[qM , qM ]

Since the elements in q have zero mean (all phases are equally probable), this sim-
plifies to
 
E[q 1 q 1 ] . . . E[q 1 q M ]
.. ... ..
h i
Q = E qqH = 
 
.
 . . 
E[qM q1 ] . . . E[qM qM ]
Using the expression for q defined in (2.2) and assuming uncorrelated jammers, the
elements in Q is given by

J
ak 2 ejπ(dn −dm )uk + E[nn nm ] ,
X
Qnm = E[qn qm ] =
k=1

where di is the distance in antenna elements between a reference point and auxiliary
antenna i. Further, due to the assumption that the thermal noise is zero-mean and
uncorrelated between antenna elements,

0 if n 6= m
E[nn nm ] =
σ 2 if n = m
th

Thus,

 J
ak 2 2
X
+ σth , n=m





Qnm = k=1
J
ak 2 ejπ(dm −dn )uk , n 6= m
X





k=1

Analogous reasoning for the cross-covariance vector R defined as

 
Cov[qin , q1 ]
R=
 ... 

Cov[qin , qM ]

yields

J
X
R= Rk + n,
k=1

13
2. Methods

2 2 2 T
where Rk is the covariance vector for one NLI-jammer and n = [σth , σth , . . . , σth ] .
Rk is given by
 −jπd u 
e i1 k

 −jπdi2 uk N 
e
 
 |ak |2 wi ejπdi uk ,
 X
Rk =  ..
.
 
  i=1
e−jπdiM uk

where i1 , . . . , iM are the indices of the auxiliary antennas.

2.2.2.1 Estimation error model

For modelling estimation error of the covariance matrix Q, the following model is
used
Q̂ = diag(1 + εest )Q diag(1 + εest )H , (2.3)
where εest is a vector with uncorrelated, complex disturbances drawn from a com-
plex, gaussian distribution with zero mean and standard deviation σest . diag(v)
represents a quadratic matrix with the elements in v as its diagonal elements and
with all other elements 0. The model is chosen because it is simple and preserves
the hermitian property of Q. Analogous reasoning for the cross-covariance vector
R yields
R̂ = (1 + εest )R diag(1 + εest )H ,
where εest is a scalar drawn from the same distribution.

2.2.3 Simulation method

In order to determine Pd and Pfa , the mainlobe and sidelobe region must first be
defined. Since the antenna pattern is continuous, it is not obvious where to draw the
line and multiple definitions exists. We have chosen to define the mainlobe region
as the half power beam width of the undisturbed antenna pattern, while the sidelobe
regions are defined as the region outside the first nulls. This results in a undefined
region between the half power beam width and the first null, which is not considered.
With those definitions and the tapering used, the mainlobe region is defined as

|u| ≤ 0.0125,

corresponding to a width of 1.4°, and the sidelobe region is defined as

|u| ≥ 0.0367,

corresponding to a gap of 4.2°.

Pd and Pfa are computed using Monte Carlo-simulations. In each simulation
SNRsum , JNR, σest , J and the placement of auxiliary antennas are held constant.
For each data point, uj and ut are drawn from uniform distributions on the side-
lobe/mainlobe regions. The amplitude of the target signal is then adjusted to get
the desired SNRsum in the main channel. Each simulation uses 5000 data points
with ut in mainlobe and 10 000 data points with ut in sidelobe region.

14
 2. Methods

We noted a tendency for the sidelobe level in the main channel to increase dra-
matically when covariance estimation errors are introduced for M > J. In order to
avoid this, only the J first auxiliary antennas are used for SLC in the main channel.
In a situation where the number of jammers is not known beforehand, this could for
instance be accomplished by Vandermonde decomposition of the covariance matrix
in order to determine J, or simply by computing Vm using different number of auxil-
iary antennas and choose the smallest value. The latter method should work because
the best weights are those that cancel out most of the jammer effect, hence would
result in the lowest Vm . When SLC is applied to the guard channel, all auxiliary
antennas are used.

15
2. Methods

16
 3
Results

In this chapter, the results of the simulations are presented. The performance is
presented with ROC-curves, which are plots of Pd (T ) versus Pfa (T ), parameterized
with T . For illustration purposes, plots with Youden’s index, Y , are also presented.
Youden’s index is defined as

Y (T ) = Pd (T ) − Pfa (T ).

It is a measure that ranges from -1 to 1, where Y = 1 means a perfect performance,

i.e. all false targets and no true targets are blanked. Youden’s index is a measure
that puts equal importance on maximizing Pd and minimizing Pfa . It is used in
this report mainly because it is easy, but in real applications maximizing Pd or
minimizing Pfa might be valued higher than the other.
Unless stated otherwise, the simulations are run using the following parameters:
SNRsum = 10 dB, JNR = 40 dB, σest = 0.1 and 1 NLI-jammer.

3.1 The undisturbed case

When no NLI-jammers are present, all three tested algorithms reduce to ordinary
SLB. For comparison, the performance of SLB when no NLI-jammers are present
is shown in Figure 3.1. As seen in Figure 3.1b, it is possible to choose a blanking
threshold T so that all false targets and no true targets are blanked. Figure 3.1b
also shows that SNRsum influences Pd but not Pfa .

3.2 Effect of auxiliary antenna position

In order to investigate how the placement of the auxiliary antennas influence the
performance of the guard functions, three different placements are tested. The
positions are presented in Table 3.1. Examples of the resulting antenna patterns
for guard and main channel are shown in Figure 3.2 and ROC-curves are shown

Table 3.1: Auxiliary antenna placements used.

guard aux
Placement 1 1 2, 3, 4
Placement 2 1 2, 3, 100
Placement 3 1 8, 37, 100

17
3. Results

1 1

0.8 0.8

0.6 0.6
Pd

Y
0.4 0.4

0.2 0.2

0 0
0 0.2 0.4 0.6 0.8 1 −40 −20 0 20 40 60
Pfa T ( dB)

(a) ROC-curve (b) Youdens index

1 1

0.8 0.8

0.6 0.6
Pfa
Pd

0.4 0.4

0.2 0.2

0 0
−40 −20 0 20 40 60 −40 −20 0 20 40 60
T ( dB) T ( dB)

(c) Probability of detection (d) Probability of false alarm

Figure 3.1: The performance of SLB when no NLI-jammers are present. SNRsum =
5 dB (black), SNRsum = 10 dB (blue) and SNRsum = 20 dB (dashed).

18
 3. Results

40 40

20 20
gain (dB)

gain (dB)
0 0

−20 −20

−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1

u u

(a) SLCg Placement 1 (b) SLCg max Placement 1

40 40

20 20
gain (dB)

gain (dB)

0 0

−20 −20

−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1

u u

(c) SLCg Placement 2 (d) SLCg max Placement 2

40 40

20 20
gain (dB)

gain (dB)

0 0

−20 −20

−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1

u u

(e) SLCg Placement 3 (f) SLCg max Placement 3

Figure 3.2: Illustration of the effect of auxiliary antenna position on main channel
pattern (blue) and guard channel pattern (black) for uj = 0.3. The positions are
defined in Table 3.1.

19
3. Results

1 1

0.8 0.8

0.6 0.6
Pd

Y
0.4 0.4

0.2 0.2

0 0
0 0.2 0.4 0.6 0.8 1 −40 −20 0 20 40
Pfa T ( dB)

(a) ROC-curve, SLCg (b) Youden’s index, SLCg

1 1

0.8 0.8

0.6 0.6
Pd

0.4 0.4

0.2 0.2

0 0
0 0.2 0.4 0.6 0.8 1 −40 −20 0 20 40
Pfa T ( dB)

(c) ROC-curve, SLCg max (d) Youden’s index, SLCg max

Figure 3.3: ROC-curves and Youden’s index (J) for Placement 1 (black), Place-
ment 2 (blue) and Placement 3 (dashed).

20
 3. Results

1 1

0.8

0.6 0.5
Pd

Y
0.4

0.2 0
0
0 0.2 0.4 0.6 0.8 1 −40 −20 0 20 40
Pfa T ( dB)

(a) ROC-curve (b) Youden’s index

1 1

0.8 0.8

0.6 0.6
Pfa
Pd

0.4 0.4

0.2 0.2

0 0
−40 −20 0 20 40 −40 −20 0 20 40
T ( dB) T ( dB)

(c) Probability of detection (d) Probability of false alarm

Figure 3.4: Comparison of the three tested guard functions No SLCg (black),
SLCg (blue) and SLCg max (dashed).

in Figure 3.3. Although the placement clearly influences the shape of the guard
antenna diagram, the effect on overall performance is small. In fact, as seen in Figure
3.3a placement has no significant effect on performance of SLCg. For SLCg max,
Placement 1 leads to a minor decrease in performance, while no significant difference
between Placement 2 and Placement 3 is observed. This is shown in Figure 3.3c.
For the rest of the simulations, Placement 3 is used. Since the placement of auxiliary
antennas does not influence the guard channel in No SLCg, it is not included here.

3.3 Comparison between methods

Figure 3.4 shows a comparison between the three tested guard functions. For those
parameters, SLCg max has a bit better performance than SLCg, while No SLCg has
a much worse performance and could in fact be considered useless. In the rest of
this chapter the effect of SNRsum , JNR, covariance estimation errors and number of
jammer on SLCg and SLCg max are presented.

21
3. Results

1 1

0.8 0.8

0.6 0.6
Pd

Y
0.4 0.4

0.2 0.2

0 0
0 0.2 0.4 0.6 0.8 1 −40 −20 0 20 40
Pfa T ( dB)

(a) ROC-curve (b) Youden’s index

1 1

0.8 0.8

0.6 0.6
Pfa
Pd

0.4 0.4

0.2 0.2

0 0
−40 −20 0 20 40 −40 −20 0 20 40
T ( dB) T ( dB)

(c) Probability of detection (d) Probability of false alarm

Figure 3.5: Effect of SNRsum on performance of SLCg(black) and SLCg max(blue)

for SNR = 5 dB (full), 10 dB (dashed) and 20 dB (dotted).

3.4 Effect of JNR and SNRsum

The effect of SNRsum and JNR on performance of SLCg and SLCg max are shown
in Figure 3.5 and Figure 3.6. The effects are similar, both influences Pd but not Pfa .
An increase in SNRsum or a decrease in JNR increases performance.

3.5 Effect of covariance estimation error

The effect of errors in the covariance estimations used for computing SLC-weights
is shown in Figure 3.7, where it is shown that estimation error influences Pd but not
Pfa .

22
 3. Results

1 1

0.8 0.8

0.6 0.6
Pd

Y
0.4 0.4

0.2 0.2

0 0
0 0.2 0.4 0.6 0.8 1 −40 −20 0 20 40
Pfa T ( dB)

(a) ROC-curve (b) Youden’s index

1 1

0.8 0.8

0.6 0.6
Pfa
Pd

0.4 0.4

0.2 0.2

0 0
−40 −20 0 20 40 −40 −20 0 20 40
T ( dB) T ( dB)

(c) Probability of detection (d) Probability of false alarm

Figure 3.6: Effect of JNR on performance of SLCg (black) and SLCg max (blue)
for JNR = 30 dB (full), 40 dB (dashed) and 50 dB (dotted).

23
3. Results

1 1

0.8 0.8

0.6 0.6
Pd

Y
0.4 0.4

0.2 0.2

0 0
0 0.2 0.4 0.6 0.8 1 −40 −20 0 20 40
Pfa T ( dB)

(a) ROC-curve (b) Youden’s index

1 1

0.8 0.8

0.6 0.6
Pfa
Pd

0.4 0.4

0.2 0.2

0 0
−40 −20 0 20 40 −40 −20 0 20 40
T ( dB) T ( dB)

(c) Probability of detection (d) Probability of false alarm

Figure 3.7: Effect of covariance estimation error on SLCg (black) and SLCg max
(blue), for no error (full), σest = 0.1 (dashed) and σest = 0.2 (dotted).

24
 3. Results

1 1

0.8 0.8

0.6 0.6
Pd

Y
0.4 0.4

0.2 0.2

0 0
0 0.2 0.4 0.6 0.8 1 −40 −20 0 20 40
Pfa T ( dB)

(a) ROC-curve (b) Youden’s index

1 1

0.8 0.8

0.6 0.6
Pfa
Pd

0.4 0.4

0.2 0.2

0 0
−40 −20 0 20 40 −40 −20 0 20 40
T ( dB) T ( dB)

(c) Probability of detection (d) Probability of false alarm

Figure 3.8: Effect of the number of jammers on SLCg (black) and SLCg max
(blue) for one jammer (full), 2 jammers (dashed) and three jammers (dotted).

3.6 More than one jammer

The performance of SLCg and SLCg max when more than one NLI-jammer are
present are shown in Figure 3.8. The performance of the two methods is significantly
reduced for the case of more than one jammer. As expected both methods have the
same performance, when the number of jammers are equal to the number of auxiliary
antennas. In order to shine some light on why the performance is reduced, the effect
of number of jammers without covariance estimation errors are shown in Figure 3.9.
The overall performance is better in this case, but the same tendencies of reduced
performance for more than one jammer is seen. In Figure 3.10 the effect of auxiliary
antenna placement is revisited, but for the case of three jammers. Since SLCg max
reduces to SLCg when the number of jammers is the same as the number of auxiliary
antennas, only SLCg is shown.

25
3. Results

1 1

0.8 0.8

0.6 0.6
Pd

Y
0.4 0.4

0.2 0.2

0 0
0 0.2 0.4 0.6 0.8 1 −40 −20 0 20 40
Pfa T ( dB)

(a) ROC-curve (b) Youden’s index

1 1

0.8 0.8

0.6 0.6
Pfa
Pd

0.4 0.4

0.2 0.2

0 0
−40 −20 0 20 40 −40 −20 0 20 40
T ( dB) T ( dB)

(c) Probability of detection (d) Probability of false alarm

Figure 3.9: Effect on the number of jammers on SLCg (black) and SLCg max
(blue) with no error in covariance estimation. Lines correspond to one jammer
(full), 2 jammers (dashed) and three jammers (dotted).

26
 3. Results

1 1

0.8 0.8

0.6 0.6
Pd

Y
0.4 0.4

0.2 0.2

0 0
0 0.2 0.4 0.6 0.8 1 −40 −20 0 20 40
Pfa T ( dB)

(a) ROC-curve (b) Youden’s index

1 1

0.8 0.8

0.6 0.6
Pfa
Pd

0.4 0.4

0.2 0.2

0 0
−40 −20 0 20 40 −40 −20 0 20 40
T ( dB) T ( dB)

(c) Probability of detection (d) Probability of false alarm

Figure 3.10: Effect of auxiliary antenna placement on SLCg when three NLI-
jammers are present. Lines corresponds to Placement 1 (black), Placement 2 (blue)
and Placement 3 (dashed).

27
3. Results

28
 4
Discussion

Regarding the placement of guard and auxiliary antennas, it seems beneficial to

cover a large span by using antenna elements in both edges of the antenna array,
as illustrated in Figure 3.3 and Figure 3.10. We believe this is because a large span
between the antennas creates a high frequency component in the guard channel.
Otherwise the placements does not seem to affect the performance much regard-
ing the guard channel. But as shown, errors in the covariance estimation highly
affect performance. Thus it is probably more important to choose auxiliary an-
tenna placement for optimising covariance estimation than avoiding deep notches in
non-jammer directions of the guard channel.
Increasing JNR, decreasing SNR and increasing covariance estimation errors all
have similar effects, namely to decrease Pd without affecting Pfa . This can be in-
terpreted as covariance estimation errors decrease the quality of nulls, making them
more shallow so that more jammer energy leaks in. Since the effect of NLI-jammers
is an increase in noise level, it is expected that increasing JNR and decreasing SNR
has similar results.
As seen in Figure 3.8, the number of NLI-jammers greatly affects performance.
One possible explanation for this could be that the nulls in the guard channel in
jammer directions are too shallow, so that more jammer energy leaks in simply
because there are more jammers. This explanation does not seem to be the main
reason for two reasons. Firstly, the effect is large also when no covariance estimation
errors are added. Secondly, the behaviour of Pd and Pfa differs significantly from the
case of increasing JNR. Another possibility is that multiple nulls make restrictions
on the guard channel antenna pattern, which if the nulls are close enough can force
down the antenna pattern in the region between nulls. Comparison between Figure
3.3 and Figure 3.10 confirms this hypothesis, since auxiliary placement influences
the result in SLCg for 3 jammers, but not for 1 jammer.
It is worth noticing that if one of the computed guard channels produced in SLCg
max fails at introducing a null in the antenna pattern in the desired direction, this
guard channel will have the highest amplitude (due to jammer energy leaking in),
and will thus be the used guard channel. In the worst case, this will produce a similar
result as No SLCg, and thus have a very bad performance. Failure to produce a
null in the desired direction can for example happen for a unlucky combination of
jammer directions and auxiliary antenna placement, as derived in Section 1.2.2.1.
The same failure could of course happen in SLCg as well, but the probability of this
happening in SLCg max is higher. This effect has not been observed in this project,
but might pose a problem in scenarios where the risk of failing to introduce a null
is high.

29
4. Discussion

4.1 Different approaches

The guard functions evaluated in this report all take the approach of using some kind
of guard channel to determine whether the detection is in the mainlobe or sidelobe.
Another approach to the problem would be to estimate the the direction of the
target signal directly from the auxiliary antenna signals. This problem is analogous
to that of frequency estimation. If the antenna elements used are few and far (i.e.
more than half a wavelength) apart, the problem is sparse or undersampled. See for
example [15] for an extensive review of method for solving such problems. For the
problem treated in this report however, two major problems arises: (i) few samples
and (ii) low (possibly negative) SNR in auxiliary antennas when the target is located
in the mainlobe. There are few samples because there are few auxiliary antennas
and target signal is typically short, so multiple snapshots of the same target will in
general not be available. The problem of sparse direction-of-arrival estimations for
the combinations low SNR with many snapshots and few snapshots with high SNR
is treated in [16], but not the combination of low SNR and few snapshots.

30
 5
Conclusion

In conclusion, both SLCg and SLCg max work well in the case of one jammer, and are
definitely a much better alternative than to use the guard channel directly when NLI
is present. In the scenarios investigated here, SLCg max in general performs better
than SLCg, but as discussed in the previous section this might not always be the
case. Additional antennas should be placed on both edges of the antenna array, but
otherwise the placement of auxiliary antennas does not effect the performance of the
guard functions substantially. However, the placement can affect the performance
indirectly, by influencing SLC in general and covariance estimation errors. The
performance of SLCg and SLCg max is reduced drastically when more than one
NLI source are present.

5.1 Future work

In this project, the guard functions are evaluated using a highly simplified model.
The next step should be to investigate the effect of allowing target and jammers
to have a frequency range instead of a single frequency, to use more realistic an-
tenna models and to study the effect of clutter and correlated interference, such as
multipath echoes. This could either be done by using more complex models, or,
ideally, by evaluating the guard functions on real radar data. Influence of polarisa-
tion, especially when jammers and target have different polarisation, may also be of
interest.
As shown in Section 1.2.2.1, the SLC weights can in general be chosen in infinitely
many ways, when the number of NLI-jammers are fewer than the number of auxiliary
antennas. It may be worth investigating if there is a better way to choose the weights
than using (1.3) in order to accomplish a desirable guard channel pattern.

31
5. Conclusion

32
 Bibliography

[1] J. Agar, “Science and the second world war”, in Science in the Twentieth
Century and Beyond, Polity, 2013, pp. 268–275.
[2] S. Sinha, C. Jeganathan, and L. Sharma, “A review of radar remote sensing
for biomass estimation”, International Journal of Environmental Science and
Technology, vol. 12, no. 5, pp. 1779–1792, May 2015.
[3] M. Klemm, I. J. Craddock, J. A. Leendertz, A. Preece, and R. Benjamin,
“Radar-based breast cancer detection using a hemispherical antenna array
–experimental results”, IEEE Transactions on Antennas and Propagation,
vol. 57, no. 6, pp. 1692–1704, Jun. 2009.
[4] M. Skolnik, “An introduction and overview of radar”, in Radar Handbook, M.
Skolnik, Ed., 3rd ed., New York: McGraw-Hill, 2008.
[5] J. Frank and J. D. Richards, “Phased array radar antennas”, in Radar Hand-
book, M. Skolnik, Ed., 3rd ed., New York: McGraw-Hill, 2008.
[6] W. G. Carrara, R. S. Goodman, and R. M. Majewski, “Sidelobe control in SAR
imagery”, in Spotlight Syntethic Aperture Radar -Signal Processing Algorithms,
Boston: Artech House, 1995, pp. 507–529.
[7] F. Bandiera and A. Farina, “Detection algorithms to discriminate between
radar targets and ECM signals”, IEEE Trans. on Signal Process., vol. 59,
no. 12, pp. 5984–5993, Dec. 2010.
[8] A. Farina, “Antenna-related ECCM”, in Radar Handbook, M. Skolnik, Ed.,
3rd ed., New York: McGraw-Hill, 2008.
[9] D. Orlando, “A novel noise jamming detection algorithm for radar applica-
tions”, IEEE Signal Process. Lett, vol. 24, no. 2, pp. 206–210, Feb. 2017.
[10] J. Kocz, F. H. Briggs, and J. Reynolds, “Radio frequency interference removal
through the application of spatial filtering techniques on the parkes multibeam
receiver”, The Astronomical Journal, vol. 140, no. 6, pp. 2086–2094, 2010.
[11] U. Nickel, “Design of generalised 2d adaptive sidelobe blanking detectors using
the detection margin”, Signal Processing, vol. 90, no. 5, pp. 1357–1372, 2010.
[12] D. A. Shnidman and N. R. Shnidman, “Sidelobe blanking with expanded mod-
els”, IEEE Transactions on Aerospace and Electronic Systems, vol. 47, no. 2,
pp. 790–805, Apr. 2011.
[13] A. Farina, “Sidelobe canceller (SLC) system”, in Antenna-Based Signal Pro-
cessing Techniques for Radar Systems, Boston: Artech House, 1992, pp. 95–
103.
[14] A. Farina, L. Timmoneri, and R. Tosini, “Cascading SLB and SLC devices”,
Signal Process., vol. 45, no. 2, pp. 261–266, Aug. 1995.

33
Bibliography

[15] Z. Yang, J. Li, P. Stoica, and L. Xie, “Sparse methods for direction-of-arrival
estimation”, CoRR, vol. abs/1609.09596, 2017. arXiv: 1609.09596v2. [Online].
Available: http://arxiv.org/abs/1609.09596v2.
[16] A. Nehorai, D. Starer, and P. Stoica, “Direction-of-arrival estimation in ap-
plications with multipath and few snapshots”, Circuits, Systems and Signal
Processing, vol. 10, no. 3, pp. 327–342, Sep. 1991.

34