Adaptive Radar Detection Modelbased Datadriven And Hybrid Approaches Angelo Coluccia

zeilonmocny 3 views 88 slides May 19, 2025
Slide 1
Slide 1 of 88
Slide 1
1
Slide 2
2
Slide 3
3
Slide 4
4
Slide 5
5
Slide 6
6
Slide 7
7
Slide 8
8
Slide 9
9
Slide 10
10
Slide 11
11
Slide 12
12
Slide 13
13
Slide 14
14
Slide 15
15
Slide 16
16
Slide 17
17
Slide 18
18
Slide 19
19
Slide 20
20
Slide 21
21
Slide 22
22
Slide 23
23
Slide 24
24
Slide 25
25
Slide 26
26
Slide 27
27
Slide 28
28
Slide 29
29
Slide 30
30
Slide 31
31
Slide 32
32
Slide 33
33
Slide 34
34
Slide 35
35
Slide 36
36
Slide 37
37
Slide 38
38
Slide 39
39
Slide 40
40
Slide 41
41
Slide 42
42
Slide 43
43
Slide 44
44
Slide 45
45
Slide 46
46
Slide 47
47
Slide 48
48
Slide 49
49
Slide 50
50
Slide 51
51
Slide 52
52
Slide 53
53
Slide 54
54
Slide 55
55
Slide 56
56
Slide 57
57
Slide 58
58
Slide 59
59
Slide 60
60
Slide 61
61
Slide 62
62
Slide 63
63
Slide 64
64
Slide 65
65
Slide 66
66
Slide 67
67
Slide 68
68
Slide 69
69
Slide 70
70
Slide 71
71
Slide 72
72
Slide 73
73
Slide 74
74
Slide 75
75
Slide 76
76
Slide 77
77
Slide 78
78
Slide 79
79
Slide 80
80
Slide 81
81
Slide 82
82
Slide 83
83
Slide 84
84
Slide 85
85
Slide 86
86
Slide 87
87
Slide 88
88

About This Presentation

Adaptive Radar Detection Modelbased Datadriven And Hybrid Approaches Angelo Coluccia
Adaptive Radar Detection Modelbased Datadriven And Hybrid Approaches Angelo Coluccia
Adaptive Radar Detection Modelbased Datadriven And Hybrid Approaches Angelo Coluccia


Slide Content

Adaptive Radar Detection Modelbased Datadriven
And Hybrid Approaches Angelo Coluccia download
https://ebookbell.com/product/adaptive-radar-detection-
modelbased-datadriven-and-hybrid-approaches-angelo-
coluccia-49161382
Explore and download more ebooks at ebookbell.com

Here are some recommended products that we believe you will be
interested in. You can click the link to download.
Advances In Adaptive Radar Detection And Range Estimation 1st Edition
Chengpeng Hao
https://ebookbell.com/product/advances-in-adaptive-radar-detection-
and-range-estimation-1st-edition-chengpeng-hao-36539876
Adaptive Radar Signal Processing 1st Edition Simon Haykin
https://ebookbell.com/product/adaptive-radar-signal-processing-1st-
edition-simon-haykin-2322566
Adaptive Radar Resource Management 1st Edition Ding Zhen Moo
https://ebookbell.com/product/adaptive-radar-resource-management-1st-
edition-ding-zhen-moo-5432776
Spacetime Adaptive Processing For Radar 2nd Edition J R Guerci
https://ebookbell.com/product/spacetime-adaptive-processing-for-
radar-2nd-edition-j-r-guerci-6810434

Spacetime Adaptive Processing For Radar J R Guerci
https://ebookbell.com/product/spacetime-adaptive-processing-for-radar-
j-r-guerci-1078686
Signal Processing For Multistatic Radar Systems Adaptive Waveform
Selection Optimal Geometries And Pseudolinear Tracking Algorithms Dr
Ngoc Hung Nguyen
https://ebookbell.com/product/signal-processing-for-multistatic-radar-
systems-adaptive-waveform-selection-optimal-geometries-and-
pseudolinear-tracking-algorithms-dr-ngoc-hung-nguyen-11059798
Principles Of Spacetime Adaptive Processing Iet Radar Sonar Navigation
And Avionics 3rd Edition Richard Klemm
https://ebookbell.com/product/principles-of-spacetime-adaptive-
processing-iet-radar-sonar-navigation-and-avionics-3rd-edition-
richard-klemm-1993304
Adaptive Antennas And Phased Arrays For Radar And Communications Alan
J Fenn
https://ebookbell.com/product/adaptive-antennas-and-phased-arrays-for-
radar-and-communications-alan-j-fenn-1199786
Cognitive Radar The Knowledgeaided Fully Adaptive Approach 1st Edition
J R Guerci
https://ebookbell.com/product/cognitive-radar-the-knowledgeaided-
fully-adaptive-approach-1st-edition-j-r-guerci-1668238

Coluccia: “fm_v2” — 2022/10/7 — 13:05 — pagei—#1
AdaptiveRadarDetection
Model-Based, Data-Driven, and Hybrid Approaches

For a complete listing of titles in the
Artech House Radar Series,
turn to the back of this book.
Untitled-1 iiUntitled-1 ii 10/7/2022 9:43:53 AM10/7/2022 9:43:53 AM

Coluccia:“fm_v2”—2022/10/7—13:05—pageiii—#3
AdaptiveRadarDetection
Model-Based,Data-Driven,andHybridApproaches
AngeloColuccia

Coluccia: “fm_v2” — 2022/10/7 — 13:05 — page iv — #4
Library of Congress Cataloging-in-Publication Data
A catalog record for this book is available from the U.S. Library of Congress.
British Library Cataloguing in Publication Data
A catalogue record for this book is available from the British Library.
Cover design by Andy Meaden
ISBN 13: 978-1-63081-900-2
© 2023 ARTECH HOUSE
685 Canton Street
Norwood, MA 02062
All rights reserved. Printed and bound in the United States of America. No part of this book
may be reproduced or utilized in any form or by any means, electronic or mechanical, including
photocopying, recording, or by any information storage and retrieval system, without permission
in writing from the publisher.
All terms mentioned in this book that are known to be trademarks or service marks have been
appropriately capitalized. Artech House cannot attest to the accuracy of this information. Use of a
term in this book should not be regarded as affecting the validity of any trademark or service mark.
10987654321

Coluccia:“fm_v2”—2022/10/7—13:05—pagev—#5
Contents
Preface ix
Acknowledgments xv
1 Model-BasedAdaptiveRadarDetection 1
1.1IntroductiontoRadarProcessing 1
1.1.1GeneralitiesandBasicTerminologyofCoherent
Radars 2
1.1.2ArrayProcessingandSpace-TimeAdaptiveProcessing5
1.1.3TargetDetectionandPerformanceMetrics 8
1.2UnstructuredSignalinWhiteNoise 9
1.2.1OldbutGold:BasicSignalDetectionandthe
EnergyDetector 9
1.2.2TheNeyman–PearsonApproach 11
1.2.3AdaptiveCFARDetection 13
1.2.4CorrelatedSignalModelinWhiteNoise 15
1.3StructuredSignalinWhiteNoise 18
1.3.1DetectionofaStructuredSignalinWhiteNoise
andMatchedFilter 18
1.3.2GeneralizedLikelihoodRatioTest 20
1.3.3DetectionofanUnknownRank-OneSignalin
WhiteNoise 24
v

Coluccia: “fm_v2” — 2022/10/7 — 13:05 — page vi — #6
vi Adaptive Radar Detection: Model-Based, Data-Driven, and Hybrid Approaches
1.3.4 Steering Vector Known up to a Parameter and
Doppler Processing 25
1.4 Adaptive Detection in Colored Noise 25
1.4.1 One-Step, Two-Step, and Decoupled Processing 27
1.4.2 General Hypothesis Testing Problem via GLRT: A
Comparison 28
1.4.3 Behavior under Mismatched Conditions:
Robustness vs Selectivity 31
1.4.4 Model-Based Design of Adaptive Detectors 33
1.5 Summary 42
References 43
2 Classification Problems and Data-Driven Tools 49
2.1 General Decision Problems and Classification 49
2.1.1M-ary Decision Problems 50
2.1.2 Classifiers and Decision Regions 55
2.1.3 Binary Classification vs Radar Detection 61
2.1.4 Signal Representation and Universal
Approximation 64
2.2 Learning Approaches and Classification
Algorithms 66
2.2.1 Statistical Learning 66
2.2.2 Bias-Variance Trade-Off 71
2.3 Data-Driven Classifiers 72
2.3.1k-Nearest Neighbors 73
2.3.2 Linear Methods for Dimensionality Reduction and
Classification 75
2.3.3 Support Vector Machine and Kernel Methods 77
2.3.4 Decision Trees and Random Forests 81
2.3.5 Other Machine Learning Tools 84
2.4 Neural Networks and Deep Learning 85
2.4.1 Multilayer Perceptron 86
2.4.2 Feature Engineering vs Feature Learning 88
2.4.3 Deep Learning 89
2.5 Summary 93
References 93

Coluccia: “fm_v2” — 2022/10/7 — 13:05 — page vii — #7
Contents vii
3 Radar Applications of Machine Learning 97
3.1 Data-Driven Radar Applications 97
3.2 Classification of Communication and
Radar Signals 100
3.2.1 Automatic Modulation Recognition and
Physical-Layer Applications 100
3.2.2 Datasets and Experimentation 102
3.2.3 Classification of Radar Signals and Radiation Sources 107
3.3 Detection Based on Supervised Machine Learning 109
3.3.1 SVM-Based Detection with ControlledP
FA 110
3.3.2 Decision Tree-Based Detection with ControlledP
FA111
3.3.3 Revisiting the Neyman–Pearson Approach 112
3.3.4 SVM and NN for CFAR Processing 114
3.3.5 Feature Spaces with (Generalized) CFAR Property 117
3.3.6 Deep Learning Based Detection 120
3.4 Other Approaches 123
3.4.1 Unsupervised Learning and Anomaly Detection 123
3.4.2 Reinforcement Learning 125
3.5 Summary 126
References 127
4 Hybrid Model-Based and Data-Driven Detection 137
4.1 Concept Drift, Retraining, and Adaptiveness 137
4.2 Hybridization Approaches 139
4.2.1 Different Dimensions of Hybridization 139
4.2.2 Hybrid Model-Based and Data-Driven Ideas in
Signal Processing and Communications 140
4.3 Feature Spaces Based on Well-Known Statistics or
Raw Data 142
4.3.1 Nonparametric Learning:k-Nearest Neighbors 142
4.3.2 Quasi-Whitened Raw Data as Feature Vector 144
4.3.3 Well-Known CFAR Statistics as a Feature Vector 147
4.4 Rethinking Model-Based Detection in a CFAR
Feature Space 151
4.4.1 Maximal Invariant Feature Space 151

Coluccia: “fm_v2” — 2022/10/7 — 13:05 — page viii — #8
viii Adaptive Radar Detection: Model-Based, Data-Driven, and Hybrid Approaches
4.4.2 Characterizing Model-Based Detectors in CFAR-FP 153
4.4.3 Design Strategies in the CFAR-FP 158
4.5 Summary 159
References 160
5 Theories, Interpretability, and Other Open Issues 165
5.1 Challenges in Machine Learning 165
5.2 Theories for (Deep) Neural Networks 167
5.2.1 Network Structures and Unrolling 168
5.2.2 Information Theory, Coding, and Sparse
Representation 171
5.2.3 Universal Mapping, Expressiveness, and
Generalization 172
5.2.4 Overparametrized Interpolation, Reproducing
Kernel Hilbert Spaces, and Double Descent 176
5.2.5 Mathematics of Deep Learning, Statistical
Mechanics, and Signal Processing 180
5.3 Open Issues 181
5.3.1 Adversarial Attacks 181
5.3.2 Stability, Efficiency, and Interpretability 182
5.3.3 Visualization 184
5.3.4 Sustainability, Marginal Return, and Patentability 185
5.4 Summary 187
References 188
List of Acronyms 195
List of Symbols 199
About the Author 203
Index 205

Coluccia: “fm_v2” — 2022/10/7 — 13:05 — page ix — #9
Preface
Radar systems have already more than one century of history and underwent
a significant evolution over decades, extending from the original military
motivations to numerous applications in civil contexts, for instance in the
automotive field. Although modern systems have multiple functionalities, the
foundational goal of a radar certainly remains that of detecting the possible
presence of a target by making an automatic decision based on the noisy signal
observed back at the receiver. Despite its arguably seeming simplicity, such a task
bears a number of formidable challenges; in the end, telling apart useful signal
from noise background is a radical ontological question that goes beyond practical
engineering, and has connections with information-theoretic and theoretical-
physics fundamental principles. In the specific case of radar systems, controllable
performance is required, especially in terms of number of false alarms, but
properties such as robustness are also desirable in practice. At the same time, the
detector must adapt to a temporally and spatially varying disturbance background
possibly made of noise, clutter, interference, and/or jamming.
In recent years, the role of data-driven techniques has grown enormously,
coming out of the original computer science, pattern recognition, and artificial
intelligence fields to reach into many disciplines in engineering. This has been
driven by significant advances in automatic classification and recognition tasks.
Key factors in this revolution have been the huge increase in computational
power due to hardware advances as well as the availability of huge datasets and
ready-to-use software libraries. Multimedia (speech, image) and online-collected
data (from platforms and apps providing social network, messaging, shopping,
streaming, file sharing, and other services) are in fact abundant and easy to gather
today, and tech companies are pushing for developing applications of their own
data-driven frameworks in more and more contexts.
ix

Coluccia: “fm_v2” — 2022/10/7 — 13:05 — pagex—#10
x Adaptive Radar Detection: Model-Based, Data-Driven, and Hybrid Approaches
Also in the signal processing and communications areas, tools from machine
learning and (shallow or deep) neural networks are increasingly adopted to attack
problems traditionally addressed by model-based approaches. This trend has also
reached the radar field, where, however, peculiar challenges are found. Data-
driven methods are in fact typically used in a black-box fashion, exploiting the
availability of a large amount of data to feed the box (e.g., a neural network),
tuning its parameters by trial and error. This approach is able to capture the
richness of real data that cannot be incorporated in a parametric model, hence
often yielding impressive results, but also exhibiting several limitations: besides
the abovementioned need for large training datasets, it lacks interpretability and
control on performance and properties, it is not robust to adversarial attacks,
and it exhibits misspecification and bias toward the training dataset. All such
issues are critical for artificial intelligence at large. In the case of radars, serious
concerns arise due to their adoption in security and safety contexts. Moving
from these premises, the book explores the possibility to adopt data-driven
techniques for the problem of target (signal) detection, alone or in combination
with model-based approaches. In fact, as mentioned, purely data-driven solutions
can provide remarkable results, but at the price of losing some desirable properties.
At the same time, domain-specific modeling has been proving its effectiveness
for decades, providing control and interpretability. An interesting possibility is
thus the combination of both paradigms, a compromise that can be termed
“hybrid.” While several excellent books are already available on the topic of (space-
time) adaptive detection—traditionally addressed by means of model-based
statistical signal processing techniques only—and books reviewing data-driven
techniques for different radar tasks and applications have also recently started to
appear, a book unifying the traditional adaptive radar detection with the modern
data-driven algorithmic view was indeed still missing.
The distinguishing feature of this book is that it is written with the aim
of truly bridging the two worlds (or two cultures) represented by the model-
based and data-driven paradigms. This is accomplished by mixing rigorous
mathematical modeling and theoretical discussion with a practical description
of algorithmic tools and data-driven intuition, along the fil rouge of statistical
learning. A by-product of the learning perspective is also to take a new look at old
work, in the hope that this unifying effort can be inspiring for engineers trained in
radar systems who seek to evolve their knowledge toward data-driven solutions,
but also for researchers who want to enter the specific topic of adaptive radar
detection in a modern and comprehensive way. The ultimate aim is to foster
a critical attitude toward data-driven tools, and also to pave the way for their
judicious integration into the well-established field of adaptive radar detection.
Overall, this methodological discussion and critical spin represents a timely and
stimulating treatment of a topic that is expected to grow significantly in the
near future.

Coluccia: “fm_v2” — 2022/10/7 — 13:05 — page xi — #11
Preface xi
More specifically, the classical formulation of radar detection as a hypothesis
testing problem, namely based on Neyman-Pearson or generalized likelihood
ratio test (GLRT) approaches, is revisited under the lens of the data-driven
paradigm. Moreover, fundamental concepts, schemes, and algorithms from
statistical learning, classification, and neural networks domains are reviewed,
showing how they have been adapted to deal with the problem of radar detection
in the literature. Emphasis is given to the quest for alternative design approaches
that can guarantee the constant false alarm rate (CFAR) property exhibited by
traditional model-based adaptive detection statistics. The design of detectors that
are robust to signal mismatches, or can conversely reject unwanted signals, is
also addressed. The book considers a large number of sources spanning several
communities (radar, statistical signal processing, machine learning, artificial
intelligence), enriched with original contributions, diagrams, and figures based
on experimental data. It is aimed at providing a broad yet neat coverage of the
topic, hopefully helping the interested technical audience to:
•Have a handy, concise reference for many classic (model-based) adaptive
radar detection schemes as well as the most popular machine learning
techniques (including deep neural networks);
•Identify suitable data-driven approaches for radar detection and the main
related issues, based on critical discussions on their pros and cons;
•Have a smooth introduction to advanced statistical techniques used
in adaptive detection, assisted by intuitive visual interpretations and
analogies to concepts from machine learning classification and data
science;
•Learn which aspects need to be carefully considered in the design of
hybrid solutions that can take advantages from both model-based and
data-driven approaches.
As a prerequisite, it is assumed that the reader is familiar with the necessary
mathematical, statistical, physical, and electrical engineering concepts. A basic
understanding of signal theory and radar processing is also recommended.
However, the book reviews several essential aspects and provides numerous
references to other sources where the needed know-how can be acquired. It
generally alternates between detailed explanations and concise reviews, to keep
focus while still having an eye on the big picture. Short survey sections are also
provided on certain aspects and on complementary topics or applications. Finally,
the academic community interested in the new possibilities brought by data-
driven and hybrid approaches to radar detection may find in this book an overview
on existing research directions. The book is organized into five chapters.
Chapter 1 concisely presents the fundamentals of radar and signal detection,
then reviews the most important design ideas behind classical adaptive radar

Coluccia: “fm_v2” — 2022/10/7 — 13:05 — page xii — #12
xii Adaptive Radar Detection: Model-Based, Data-Driven, and Hybrid Approaches
detectors. Emphasis is on providing insights and drawing relationships between
different model-based solutions, rather than on a detailed treatment. This is a
distinctive angle compared to the many excellent sources already available in the
literature, and ultimately provides a handy reference to popular detectors based
on traditional model-based tools, highlighting their intrinsic interpretability and
control onto the achievable performance.
Chapter 2 introduces the necessary theoretical elements and presents the
main algorithmic tools adopted in machine learning, in particular for classification
problems (of which detection is a special case). Classical hypothesis testing theory,
including the Neyman-Pearson rationale, is reviewed and put in connection with
the use of loss functions in classification problems, following the formulation
of statistical learning. Essential principles, language, and concepts of the data-
driven paradigm found in machine learning and data science are introduced and
adapted to the typical background, jargon, and needs of engineers in the radar
community. Several popular machine learning algorithms are compendiously
described to provide a self-consistent reference.
Chapter 3 details the applications of the tools introduced in Chapter 2 to
the radar field. The chapter specifically discusses how machine learning and more
generally data-driven tools have been applied to the problem of target detection in
the literature, but other applications of machine learning techniques, including
signal classification, are also reviewed.
Chapter 4 is devoted to various forms of hybridization between model-
based and data-driven concepts. Techniques for data-driven CFAR detection are
discussed, and a restricted network architecture for analysis and design of CFAR
detectors is illustrated, trying to bridge the gap between the classical model-
based approach and data-driven tools. In particular, the problem of designing
data-driven and hybrid detectors that are robust to signal mismatches or can
conversely reject unwanted signals is considered.
Chapter 5 provides an articulate discussion on some important open
issues of data-driven techniques. Cutting-edge studies aimed at understanding
the generalization capabilities and, more generally, developing foundational
theories of deep learning are reviewed, highlighting open problems regarding
explainability and interpretation. Other relevant aspects, such as adversarial
attacks, stability, sustainability, marginal return, and patentability, are also
outlined.
This book includes a carefully selected reference list for each chapter, so that
the reader is easily steered toward more detailed or complementary information
on each aspect. Lists of acronyms and recurrent symbols are also provided.
The potential unleashed by data-driven methods, especially in combination
with more principled (model-based) approaches, is likely going to produce
significant advances in the radar field at large in the near future, provided that
control is kept on the performance and interpretability of the algorithms. At

Coluccia: “fm_v2” — 2022/10/7 — 13:05 — page xiii — #13
Preface xiii
the same time, all the issues and limitations need to be carefully considered in
making design choices for mission-critical tasks such as radar detection. In this
respect, the viewpoint of the book is that a correct attitude should be neither
blind faith in data-driven methods, throwing away decades of brilliant signal
processing research, nor an a priori refusal of alternatives to classical model-based
solutions—this way progressively advancing, ideally, toward an ultimate paradigm
that provides benefits from both worlds.

Coluccia: “fm_v2” — 2022/10/7 — 13:05 — page xiv — #14

Coluccia: “fm_v2” — 2022/10/7 — 13:05 — page xv — #15
Acknowledgments
Scientific research is standing on the shoulders of giants, hence the list of people
I should acknowledge would properly need to go back to the onset of theHomo
sapienslineage, if not earlier. I would at least like to thank those who produced
the theoretical and practical knowledge that made its way into this book; they
are many as well. I especially acknowledge the coauthors of my papers, both on
and off the book’s topic: from our joint work I learned a lot. I would also like
to extend my thanks to all the scholars and community members who kindly
spent time in discussion with me at conferences or in private communications,
for unconditionally sharing their know-how and vision.
In truth, this book is a sudden outcome during my humble research journey,
not a deliberate task I was planning for. So, sincere thanks must go to Artech
House, in particular Dr. Merlin Fox (senior commissioning editor), for inviting
me to develop this project, and for encouraging me to carry it out over a short
time frame.
I also wish to thank Ms. Casey Gerard (assistant editor) for her continuous
assistance and gentle pushing, and Dr. Joseph R. Guerci for his consideration and
valuable feedback.
I am indebted to Prof. Giuseppe Ricci, who first introduced and mentored
me on the problem of radar detection over a number of years, and as a student at
the University of Salento, Italy, gifted me with an indelible imprinting through
his teaching. I would also wish to thank my colleague Dr. Alessio Fascista for his
help in preparing some of the figures, and for the nice discussions.
Last but not least, I sincerely wish to thank my wife Serena and my two
little boys Paride and Matteo, for making my life so rich and joyful: this book
could not have been written without you.
xv

Coluccia: “fm_v2” — 2022/10/7 — 13:05 — page xvi — #16

Coluccia: “chapter1_v2” — 2022/10/7 — 13:05 — page1—#1
1
Model-BasedAdaptiveRadarDetection
This chapter concisely presents the fundamental principles of radar, the basics of
signal detection, and the specific problem of (space-time) adaptive radar detection.
It is meant to introduce the necessary basic concepts, formalism, and performance
metrics (namely probability of false alarm, probability of detection, and constant
false alarm rate property) while providing a handy reference to popular detectors
based on classical model-based tools. Emphasis is put on insights and relationships
among different detection strategies, with a critical comparative discussion, rather
than on details and comprehensiveness of all radar aspects. The discussion is kept
as lightweight as possible, informal at times, but nonetheless essential formalism
and fundamental elements of mathematical rigor are progressively introduced in
order to grasp the inner logic of model-based radar detection. Specifically, the
unstructured signal case is first addressed, including the ubiquitous tool of energy
detection, the Neyman–Pearson approach, and eigenvalue-based detectors. Then,
structured signal models are reviewed, and one-step and two-step generalized
likelihood ratio test approaches are compared, from the pioneering Kelly’s detector
and adaptive matched filter to more recent detectors. The ultimate goal is to
uncover the rationale of the model-based approach to adaptive radar detection,
as it has been consolidating over several decades, highlighting its intrinsic
interpretability and control onto an achievable performance. Such aspects will
be progressively contrasted with the different characteristics of the data-driven
approach within the remaining chapters of the book.
1.1 Introduction to Radar Processing
Radar systems currently have more than one century of history and underwent
a significant evolution over the past decades. Originated in the military field
1

Coluccia: “chapter1_v2” — 2022/10/7 — 13:05 — page2—#2
2 Adaptive Radar Detection: Model-Based, Data-Driven, and Hybrid Approaches
for long-range surveillance and early detection of air and naval targets, the
applications of radar today are also very numerous in civil contexts, for instance
in the automotive field.
Radar detection involves multiple stages of processing and is itself only
one (though very fundamental) part of a modern radar system, which generally
includes several blocks and functionalities. For example, an automotive radar
needs to detect range, velocity, and direction of an object, and estimation of
such parameters clearly requires a detection algorithm but may also involve
clustering and other algorithms. Due to the technological advance, in particular
in electronics and digital processing capability, high-resolution radars are in fact
becoming widespread and detection provides point clouds that need further
processing (e.g., clustering to estimate the object contour).
1
The focus of this
book is mainly on point target detection; therefore, all the methods described
may correspond to first stages of a more advanced radar detection processing chain.
Traditional (adaptive) radar detection is a beautiful mix of engineering
attitude and flourishing mathematical creation. This is due to the importance that
models have in the processing of radar data for detecting the possible presence of
a target, under different types of noise and disturbances that may mask it. In the
following the essential prerequisite elements of this topic are introduced, but after
that the emphasis will be on providing insights and drawing relationships between
different model-based solutions, rather than on a comprehensive treatment. This
is a distinctive angle compared to the many excellent sources already available in
the literature. The reader is indeed referred to existing books for all the details
about the introduced concepts, as well as for the many aspects of radar systems that
are not covered in this book. Classical, indispensable references are [1–3], while
more modern books are [4–6]; see also [7], which includes additional topics
such as multi target tracking, data association, passive radar, radar networks,
applications, and research directions. References more centered on adaptive radar
detection are [8, 9], while cognitive radar, waveform methods, optimum resource
allocation, and radar scheduling under the knowledge-based (or knowledge-
aided) approach are discussed in [10, 11]. Specific references for space-time
adaptive processing (STAP) are [12, 13], while for a high-level introduction the
reader may refer to [14, 15].
1.1.1 Generalities and Basic Terminology of Coherent Radars
A radar is basically an engineered system employing an electromagnetic wave to
determine the presence or absence of atargetwithin a certain surveillance area,
and to determine its (nonambiguous)range, essentially based on the relationship
1. However, specific approaches also exist for the detection ofrange-spreador otherwiseextended
targetsbased on the joint processing of multiple radar cells, which do not involve clustering.

Coluccia: “chapter1_v2” — 2022/10/7 — 13:05 — page3—#3
Model-Based Adaptive Radar Detection 3
between the round-trip time-of-fly and the distance between the radar and the
target. Additionally, most modern radars can determine velocity and direction
of arrival (DOA), and in some cases even perform imaging of the target. All
such information is extracted from theechoof the transmittedwaveformreceived
by the radar. Specifically, considering the colocated transmitter and receiver
(monostatic radar), the transmitted radiofrequency (RF) waveform propagates
in the environment and may be partially reflected back to the radar by the target,
as well as by other unwanted scatterers or reflectors, collectively referred to as
clutter. The portion of reflected energy, called target echo orreturn, depends
on theradar cross sectionof the target and channel propagation factors, and
is typically weak. As a consequence, several strategies are adopted to improve
the signal-to-noise ratio (SNR), including coherent integration of severalpulses,
multiple-antenna (array) processing, and optimal filtering in both space and time
(STAP).
In a typical setting, a pulsed wave transmitter emits a sequence of pulses of
finite duration—which can be either simple uncoded pulses or include intrapulse
(and even interpulse) modulations to gain desired waveform properties—
separated by an interval calledpulse repetition time(PRT), in which the transmitter
is switched off. During this temporal period the received signal is amplified,
downconverted, and sampled at a suitable frequency after waveform-matched
filtering.
2
The resulting samples are then processed for target detection and
estimation of useful information. Specifically, considering a spherical coordinate
system with the radar located at the origin, it transmits a beam in some azimuth
and elevation angular direction and aims at determining the presence of a useful
target in range along that line. Modern radars sample the signal during thecoherent
processing intervals(CPIs) at discrete times that represent possible increments in
range, hence the namerange binorfast-timesamples. In acoherentradar, the
stability of the frequency generation is sufficient to deterministically relate the
phase of the received signals from pulse to pulse. During a CPI,N
Tpulses are
typically transmitted in burst (pulse train), and the corresponding set of range
bins for each pulse are stored in a memory as a matrix of in-phase and quadrature
(I/Q) voltage samples (baseband equivalent). The sampling rate in this dimension
is thepulse repetition frequencyPRF=
1
PRT
(number of transmit/receive cycles per
second, measured in hertz), which is much lower than the sampling rate in range
called, accordingly,slow-time. The CPI is the total amount of timeN
T·PRT
represented by the data matrix, as shown in Figure 1.1.
2. Other architectures are possible, depending on whether pulse compression is performed before
Doppler processing or different approaches are used. See the discussion later in this section and also
[8, Chapter 2] and [9, Chapter 2] for analytical derivations of radar signal models. An illustrative
example with some details on radar waveforms and matched filtering for pulse compression is
presented in Section 2.1.1.

Coluccia: “chapter1_v2” — 2022/10/7 — 13:05 — page4—#4
4 Adaptive Radar Detection: Model-Based, Data-Driven, and Hybrid Approaches
Figure 1.1Fast-time/slow-time CPI data matrix.
Considering a single pulse, the vectorrcontainingRrange bins (fast-time
samples) obtained by sampling at time instantsτ
1,...,τ Rcan be modeled as
r=





aχ(τ
1−τ,f D)
aχ(τ
2−τ,f D)
.
.
.
aχ(τ
R−τ,f D)





+





n
(1)
n
(2)
.
.
.
n
(R)





(1.1)
whereais a complex factor depending on transmit power, antenna gain and
radiation pattern, path loss, and radar cross section of the target,τis the round-
trip time between radar and target providing the time-of-fly information from
which the range can be determined ascτ/2(cdenoting the speed of light),f
Dis
the Doppler frequency shift,χ(·,·)is theambiguity function, andn
(1)
,...,n
(R)
denote noise terms (for more details, see [9, Chapter 2] and [8]). If the radar and
the target are in relative motion, in fact, the so-calledDoppler effectalso occurs: the
received wave reflected by the target will be thus shifted in frequency compared
to the wave transmitted by the radar. Denoting byvthe radial component of
the velocity vector of the target towards the radar andλthe wavelength of the
transmitted signal, the Doppler frequency shiftf
Dis approximately given by
3
fD≈
2v
λ
(1.2)
and is clearly zero hertz for a stationary target.
Target detection in range domain requires to determine whether any of
the range binsr
kinr,k=1,...,R, contains a significant return compared
to noise, which usually occurs for the correspondingτ
kclosest to the true
round-trip delayτ, where the ambiguity function takes its maximum. For better
performance, however,N
Tχ1 pulses are typically processed jointly; this also
3. Note that the Doppler frequency shift may be coupled to range, but in most cases such a
coupling is negligible.

Coluccia: “chapter1_v2” — 2022/10/7 — 13:05 — page5—#5
Model-Based Adaptive Radar Detection 5
enables estimation of the Doppler shift, from which the (radial) velocity of the
target can be determined (based on (1.2)). Indeed, for a given range binr
k, the
signal consideringN
Tpulses is expressed by aN T-dimensional complex vector
4
z=αv T+n (1.3)
where
v
T=





1
e
j2πνT
.
.
.
e
j2π(N T−1)νT






C
NT
(1.4)
is called thetemporal steering vector,ν
T=fDTis the normalized Doppler
frequency of the target (with 1/Tthe sampling frequency in fast time),α=
Ae

∈Caccounts for the amplitudeAand phaseφof the received target
echo (including ambiguity function and possible other factors), andn∈
C
NT
denotes the complex random vector representing the overall noise. The latter
may include thermal (white) noise, correlated clutter returns, and other possible
disturbances such as interference from spectrum coexistence or jamming signals.
Model (1.3)–(1.4) indicates that the detection problem can be performed in
the Doppler domain by searching for a peak frequency in the Fourier spectrum
(estimation of a complex sinusoid in noise), computed through the fast Fourier
transform (FFT) algorithm. However, we will see later that a better and more
general approach is to recast the problem in the framework of hypothesis testing.
In our notation, subscript
Tstands for “time” to distinguish the related
variables from those in “space”; that is, associated to several antennas that may be
present in the radar, for which the subscript
Swill be adopted. Indeed, the term
“steering” is more properly used in DOA processing where it intuitively refers to a
physical or digital steering of the antenna beam, as explained in the next section.
1.1.2 Array Processing and Space-Time Adaptive Processing
In advanced radar systems, the detection process can also benefit from the presence
of multiple antenna elements that coherently process one or multiple pulses, as
illustrated in Figure 1.2.The termarray processinggenerally refers to the processing
of signals acquired through a set of spatially separated sensors for a great variety
of problems and applications [16, 17]. In the radar context, an adaptive array of
N
Ssensors (antennas) can overcome the directivity and resolution limitations of a
single sensor [15]. A model of the same type as (1.3) is still valid for the coherent
signal obtained by taking complex samples from multiple channels, namely an
4. In the remainder of the chapter we omit the range bin indexkdue to its irrelevance, being that
models and tests referred to a given (generic) radar cell at time, as better explained in Section 1.1.3.

Coluccia: “chapter1_v2” — 2022/10/7 — 13:05 — page6—#6
6 Adaptive Radar Detection: Model-Based, Data-Driven, and Hybrid Approaches
Figure 1.2Visual representation of a radar with multiple antennas that processes multiple
pulses (slow time); sampling instants relative to the different range bins (fast time
samples) are also depicted, with the same shade of gray indicating a given range
bin along the slow time (within the CPI).
antenna array, in the case of a single pulse. The only difference is the definition
of the steering vector, whereinν
Tneeds to be replaced by the spatial frequency
of the targetν
S, so obtaining aN S-dimensionalspatial steering vector:
v
S=





1
e
j2πνS
.
.
.
e
j2π(N S−1)νS






C
NS
. (1.5)
The spatial steering vector depends on the array geometry, carrier frequency of
the radar, and DOA of the target echo.
5
Improved performance is finally obtained by processing several samples from
each sensor of the antenna array, in a joint multidimensional adaptive filtering
of clutter and jamming [18, 19]. In fact, traditional decoupled approaches of
beamforming followed by Doppler filtering (or vice versa) are not optimal [18],
since clutter returns may generally manifest themselves as fully two-dimensional
(2-D) (nonfactorable) structures [13]. Instead, STAP specifically refers to the
5. For example,ν S=
d
λ
sinθfor a uniform linear array (ULA) with interelement spacingd
(typicallyd=λ/2) and radar carrier frequencyf
c=c/λ, withcdenoting the speed of light. The
use of either sin or cos function depends on the convention adopted for measuring the DOAθ.

Coluccia: “chapter1_v2” — 2022/10/7 — 13:05 — page7—#7
Model-Based Adaptive Radar Detection 7
extension of adaptive antenna techniques to simultaneously combine, in a data-
adaptive way, the signals received on multiple elements of an antenna array (the
spatial domain) and from multiple pulse repetition periods (the temporal domain)
of a CPI [18]; that is, to perform multidimensional filtering that accounts for
angle-Doppler coupling [13]. In that case, the signal model at thehth antenna,
h=1,...,N
S, can be written as
z
h=αv Te
j2π(h−1)ν S
+nh. (1.6)
By stacking suchN
Svectors (each of themN T-dimensional) the resulting signal
model is formally of the same type as (1.3), withv
Treplaced by thespace-time
steering vector
v=v
T⊗vS∈C
N
(1.7)
where⊗is the Kronecker product, andN=N
TNS.
Notice that STAP is intriniscally a linear filtering technique for clutter and
jamming cancellation through optimal weighting, conceptually not very different
from minimum-variance distortionless (Capon’s) beamforming [16]. Originally
introduced in the context of airborne radar, STAP has shown its effectiveness to
maximize the output SNR hence greatly improve the visibility of the target [5,
12, 13, 18]. Still, for the detection task itself, a detection statistic is needed.
Such aspects will be discussed in more details in Section 1.4 (in particular,
Section 1.4.1).
In the following, unless differently specified, we will generally adopt the
notationv, dropping any subscript since the development applies irrespective of
the temporal, spatial, or space-time nature of the steering vector. Figure 1.3 depicts
the so-calledradar datacube, also highlighting the aforementioned different types
Figure 1.3Radar datacube with the main types of 1-D and 2-D processing highlighted.

Coluccia: “chapter1_v2” — 2022/10/7 — 13:05 — page8—#8
8 Adaptive Radar Detection: Model-Based, Data-Driven, and Hybrid Approaches
of processing that are within the scope of this book, in particular for the problem
of adaptive target detection. For further important radar topics, including 2-D
imaging in slow and fast time (synthetic aperture radar (SAR)), the reader is
referred to the already cited general books (in particular [4–8]) in addition to
[20–27].
1.1.3 Target Detection and Performance Metrics
The focus of this book is on adaptive target detection against a background of
disturbances consisting of clutter, interference, possible noise-like jammers, and
(thermal) noise. Several different models for both the target echo and background
disturbance are discussed; such a variety of models is motivated by the different
types of radars as well as operational contexts, which can be as different as
fixed long-range surveillance radars, airborne radars for aerial and/or terrestrial
target detection, naval radars for different applications, and short-to-medium-
range radars for automotive applications, among others. Within this context,
the theoretical framework for optimal detection is given by the powerful tool of
statistical hypothesis testing, which is briefly recalled in the following.
Consider a coherent radar and a given cell under test (CUT) in range-
Doppler, range-azimuth (DOA) or range-Doppler-azimuth, according to the
considered case of temporal, spatial, or space-time steering vector. Radar detection
is the goal of deciding whether a target return (useful signal) is present against
a background of disturbance (noise signal). This can be formulated as a binary
hypothesis test, where the null hypothesisH
0means signal absence (noise only)
while the alternative hypothesisH
1means signal presence (plus noise); that is:

H
0:z=n
H
1:z=αv+n
(1.8)
wherevis known whileαis unknown, being the latter is generally dependent on
the transmit antenna gain, radiation pattern of the array, two-way path loss, and
radar cross section of the target (assumed to be slowly fluctuating).
The hypothesis testing problem (1.8) can be solved in different ways, thus
obtaining adecision statistic t(z)which is a function of the data (also called
observations)zand typically produces a decision according to the detection rule
t(z)
H1

H0
η. (1.9)
The resultingdetector, whenevert(z)exceeds the thresholdη, will produce either
acorrect detection(true positive) or afalse alarm(false negative) according to the
hypothesis actually in force (eitherH
1orH0, respectively); likewise, whenever
t(z)is belowη, the decision will be amissor conversely a true negative depending
on the actual hypothesis. Natural metrics for assessing the performance of a

Coluccia: “chapter1_v2” — 2022/10/7 — 13:05 — page9—#9
Model-Based Adaptive Radar Detection 9
detector are therefore theprobability of detection P Dand theprobability of false
alarm P
FA.
In practice it is of utmost importance to keep the false alarm rate under an
acceptablelevel, whichmeansguaranteeingaprefixedvalueofP
FAduringtheradar
operation.The thresholdηis therefore set accordingly, but since the noise level and
characteristicsmaychangeovertime, itisalsoimportanttoensurethatP
FAremains
constant irrespective of such changes. This leads to an additional performance
metric, whichisthesensitivityofadetector’sP
FAtovariationsofthenoisestatistics.
Morespecifically, adesirablepropertyistheconstantfalsealarmrate(CFAR),which
is obtained when the statistical distribution of the test statisticst(z)underH
0is
independent of any unknown parameter, as further discussed below.
1.2 Unstructured Signal in White Noise
In order to grasp, in a progressive manner, the subtleties by which theoreti-
cal hypothesis testing formulations can ultimately end up in reasonable,
implementable, and effective adaptive detectors, with specific characteristics that
reflect the adopted model-based formalization, it is instructive to first focus on
a simpler problem than (1.8). This would be the detection of an unstructured
signal—that is, a fully generic signal for which no information is available—in
white noise.
6
Along the path, we will review the general-purpose tool of energy
detection and highlight its relationship with the fundamental Neyman–Pearson
approach.
1.2.1 Old but Gold: Basic Signal Detection and the Energy Detector
Detecting the possible presence of a signal in noise is not unique of the radar
domain, being rather a recurrent problem in engineering as well as in biology,
finance, seismology, and other fields. Intuitively, one may expect that the presence
of a useful signals(t)tends to increase the magnitude of the measured signalz(t)
compared to the noisen(t), assumed to be wide-sense stationary and white;
therefore, if the overall signal level exceeds a certain threshold, this might likely
indicate a deviation from the noise-only condition. It is thus rather natural to select
as possible decision statistic the energy of the signal, which means obtaining the
most basic detection tool, theenergy detector(ED). The ED measures the received
energyEin a time interval and compares it to a threshold. By the independence
betweens(t)andn(t), this informally means exploiting the relationships
E(z)=

E(n) if the signal is absent(H
0)
E(s)+E(n)if the signal is present(H
1)
(1.10)
6. The most relevant case of structured signal in colored noise will be addressed in Section 1.4,
passing through intermediate formulations of structured signal in white noise in Section 1.3.

Coluccia: “chapter1_v2” — 2022/10/7 — 13:05 — page 10 — #10
10 Adaptive Radar Detection: Model-Based, Data-Driven, and Hybrid Approaches
where the general definition of energy for an analog (continuous-time) signalz(t)
is given byE(z)=

+∞
−∞
|z(t)|
2
dt. In particular, assuming thatNsamples are
collected with sampling timeT, that isz(iT),i=1,...,N, we have that
z(iT)=

n(iT) if the signal is absent(H
0)
s(iT)+n(iT)if the signal is present(H
1)
i=1,...,N
(1.11)
which can be restated in vector form as (cf. (1.8))
z=

n if the signal is absent(H
0)
s+nif the signal is present(H
1)
. (1.12)
Thus, the ED considers as statistic the energy of the discrete-time version of the
signal (i.e.,

N
i=1
|zi|
2
=νzν
2
), according to the decision rule
νzν
2
=z

z
H1

H0
η. (1.13)
Notice that, while (1.10) is valid irrespective of the assumptions abouts
andn, settingηto guarantee a desiredP
FAin (1.13) requires the statistical
characterization ofνzν
2
underH 0. This, in turn, entails the specification of
a probabilistic model for the noise, as well as to introduce some assumptions
on the useful signal.
7
The zero-mean Gaussian assumption for the noise is
a widely adopted one, and can be justified on the basis of the central limit
theorem.
8
In addition, in some contextncan be assumed to have uncorrelated
entries, that is, with a white covariance matrix
E[nn

]=σ
2
I, whereσ
2
is
the noise power (equal to the level of the power spectral density ofn(t)). The
correlated (colored) noise model is however of greatest interest for many radar
scenarios. Similarly, while in coherent radar processing the signalsis known up
to a complex factor, as highlighted in (1.8), in other scenarios an unstructured
model needs to be adopted. Putting things together, we can identify the minimal
(simplest) setup for a statistical signal detection problem: deciding about the
possible presence of a noncoherent signal in white complex Gaussian noise; that
is,n∼
CN(0,σ
2
I), in particular by assuming also forsa white complex Gaussian
7. Curiously, the squelch circuit used, for example, in devices such as old-fashioned citizens band
(CB) radio and radio microphone equipment, adopts an energy detection scheme to mute the
output speaker during time intervals with too low audio input to avoid listening to uncomfortable
white noise only. In those cases the threshold level (i.e.,η) is typically adjusted by tuning the squelch
knob. Most radar contexts require a guaranteed false alarm rate instead due to their mission-critical
nature.
8. However, situations exist where the Gaussian model is not a good fit; for instance, sea clutter
in high-resolution radars at low grazing angles. We will come back to this point later in Chapters 3
and 4.

Coluccia: “chapter1_v2” — 2022/10/7 — 13:05 — page 11 — #11
Model-Based Adaptive Radar Detection 11
model (i.e.,s∼ CN(0,σ
2
s
I)withσ
2
s
the power of the useful signal). Under such
assumptions, (1.12) is tantamount to testing

H
0:z∼ CN(0,σ
2
I)
H
1:z∼ CN(0,(σ
2

2
s
)I)
. (1.14)
1.2.2 The Neyman–Pearson Approach
The optimal test for the minimal model (1.14) can be derived by following
the Neyman–Pearson (NP) approach [28–30], which guarantees thatP
Dis
maximized under a constraint onP
FA. Specifically, the NP lemma states that
the uniformly most powerful (UMP) detector, if it exists, is given by the ratio
between the probability density function (PDF) of the data underH
1, denoted
byf(z|H
1), and the corresponding PDF underH 0, denoted byf(z|H 0). Such a
decision statistics, referred to aslikelihood ratio(LR), for the case at hand can be
easily computed as
f(z|H
1)
f(z|H 0)
=
1
π
N
det((σ
2
s

2
)I)
e
−z

((σ
2
s

2
)I)
−1
z
1
π
N
det(σ
2
I)
e
−z


2
I)
−1z
=

σ
2
σ
2
s

2
N
e
σ
2
s
σ
2

2
s+σ
2
)
z

z
(1.15)
provided thatσ
2
andσ
2
s
are known. Notably, testing (1.15) against a threshold
is equivalent, up to a monotonic transformation and absorbing known quantities
into the thresholdη, to the ED test given in (1.13). Thus, interestingly, the
ED rule we heuristically introduced in Section 1.2.1 actually coincides, under
the suitable assumptions above, with thelikelihood ratio test(LRT); that is, the
optimal NP test. A first takeaway is therefore that intuition and rigor can fruitfully
complement each other.
9
Under the assumptions above, the ED, also known assquare-law detector
(coupled with linear integrator, so overall producing a noncoherent integration
of the signal), admits a simple characterization. Specifically, it can be equivalently
rewritten as
νzν
2
σ
2
H
1

H0
η (1.16)
where for simplicity (and a slight abuse of notation, hereafter)ηis still used
to denote a modification of the original threshold. It then turns out that the
lefthand side of (1.16) is the sum of 2Nstandard (i.e., with zero mean and unit
variance) independent and identically distributed Gaussian random variables, due
to the independence and equivariance of real and imaginary parts in (circularly-
symmetric) complex Gaussian variables. This implies that, underH
0, a central
9. This mantra is in fact one of the guiding principles of this book.

Coluccia: “chapter1_v2” — 2022/10/7 — 13:05 — page 12 — #12
12 Adaptive Radar Detection: Model-Based, Data-Driven, and Hybrid Approaches
chi-square distribution with 2Ndegrees of freedom is obtained (i.e.,
νzν
2
σ
2∼
χ
2
2N
), hence
P
FA=P

νzν
2
σ
2



H 0

=1−F
χ
2
2N(η). (1.17)
From (1.17) the threshold that guarantees a chosen value ofP
FAcan be obtained,
by means of the cumulative distribution function (CDF) inverse, as
η=F
−1
χ
2
2N(1−P FA). (1.18)
Similarly, a characterization of the ED underH
1can be obtained by
multiplying and dividing (1.16) byσ
2

2
s
, so obtaining a statistic distributed as
a noncentral chi-square with 2Ndegrees of freedom and noncentrality parameter
σ
2
η
σ
2

2
s. As a consequence,
P
D=P

νzν
2
σ
2



H 1

=1−F
χ
2
2N

σ 2
η
σ
2

2
s

=1−F
χ
2
2N

η
1+SNR

(1.19)
where SNR=σ
2
s

2
is indeed interpretable as a signal-to-noise ratio.
The ED example is very instructive, since despite its simplicity it nonetheless
touches many important aspects that are encountered,mutatis mutandis, in more
advanced detectors. They are summarized as follows:
•Equations (1.17) and (1.19) show that bothP
FAandP Dare directly
related to the thresholdη, so improving the performance on one of such
metrics is a tradeoff with the performance on the other metric.
•The distribution of the test underH
0, hence, in turn,P FAis independent
of any unknown parameter, therefore the ED exhibits the CFAR property.
•The detection performance ultimately depends, besides obviously upon
the thresholdη, only upon one parameter: specifically, (1.19) reveals
thatP
Dnonlinearly increases with the SNR, and does not depend on
the signal and noise power individually. In other cases,P
Dmay also
depend on additional parameters. Generally speaking, knowing which
fundamental parameters the performance depends on is very important
to optimize the system design.
•Some of the parameters required for the implementation of the NP
detector may be unknown in practice. For the ED, this is typically the
case ofσ
2
in (1.16), which is required to make the detector CFAR; by
comparison, detector (1.13) is not CFAR since its distribution underH
0
depends onσ
2
. Fortunately, the detector can still be madeadaptiveby
replacing the unknownσ
2
with an estimateˆσ
2
, as illustrated in the next
section.

Coluccia: “chapter1_v2” — 2022/10/7 — 13:05 — page 13 — #13
Model-Based Adaptive Radar Detection 13
1.2.3 Adaptive CFAR Detection
The last point in the list in Section 1.2.2 is particularly important and deserves
further development. Figure 1.4 depicts the conceptual scheme of an adaptive
ED that uses an estimateˆσ
2
in place of the unknownσ
2
. Specifically,ˆσ
2
can be
obtained as sample variance (or other estimator, as discussed in Section 1.2.4) from
a set ofKindependent and identically distributedsecondary data z
i,i=1,...,K,
representative of the noise-only condition (i.e., sharing the same varianceσ
2
as
the CUT underH
0). Such data can be obtained from selected range bins in the
neighborhood of the CUT, possibly leaving some guard periods, as better discussed
later in this section. Setting the threshold for the resulting adaptive detector
10
:
νzν
2
ˆσ
2

νzν
2

K
i=1
|zi|
2
H
1

H0
η (1.20)
requires accounting for the additional statistical variability introduced byˆσ
2
.By
multiplying and dividing byσ
2
, the left-hand side of (1.20) turns out to be
the ratio of two sums of independent standard Gaussian variables. In particular,
underH
0, it is distributed as the ratio of two independent central chi-square
distributions with 2Nand 2Kdegrees of freedom, respectively, so it has a central
Fisher-Snedecor’s F distribution. Again, this means that the adaptive ED exhibits
the CFAR property; that is, setting a value forηyields the sameP
FAirrespective
of the actual value ofσ
2
.
More in general, all the considerations made in Section 1.2.2 still apply. In
particular, the distribution of the adaptive ED underH
1(which is equivalent to
a noncentral F distribution) depends only upon the SNR (besides obviouslyη),
hence so does the detection performance in terms ofP
D.
The NP-based approach made adaptive by the use of sample estimates in
place of unknown parameters paves the way for the generalized likelihood ratio
test (GLRT) approach that will be discussed in Section 1.3. Beforehand, we will
exploit once more the adaptive ED example to show how a theoretical derivation
Figure 1.4Conceptual scheme of an adaptive radar receiver based on energy detector.
10. The factor
1
K
appearing inˆσ
2
, being a constant, has been absorbed into the thresholdη.

Coluccia: “chapter1_v2” — 2022/10/7 — 13:05 — page 14 — #14
14 Adaptive Radar Detection: Model-Based, Data-Driven, and Hybrid Approaches
can be used to justify a practical scheme. We already noticed that the ED scheme,
in spite of its intuitive origin, is actually the optimal NP test under suitable
assumptions. Likewise, the adaptive ED in (1.20) can be recast as
νzν
2


nonadaptive statistic
H
1

H0
ηˆσ
2

adaptive threshold
(1.21)
that is, a nonadaptive ED followed by a data-adaptive threshold aimed at
recovering the CFAR property. The latter is usually referred to asCFAR processor
and represents a traditional processing technique in many practical radar systems.
It will be briefly discussed below.
The process of estimating the noise powerσ
2
—more generally, the whole
covariance matrix in case of colored noise—is usually performed by processing
selected radar cells in the neighborhood of the CUT, namely by averaging output
samples of a square-law detector in lagging and leading fast-time windows, as
shown in Figure 1.5. The assumption is that the latter contain only independent
realizations of the noise process, sharing the same statistical properties of the
CUT (homogeneous environment). This processing, known as cell-averaging
CFAR (CA-CFAR), also admits multiple robust variants, namely based on ordered
statistics such as the OS-CFAR, GO-CFAR, and several other schemes [4] which
try to limit the impact of clutter edges, possible target return contamination,
and other deviations from the homogeneity assumption (nonhomogeneous envi-
ronment).
11
CFAR processors are very popular in conventional radar receivers,
Figure 1.5Cell-averaging CFAR scheme.
11. Additional discussion on CFAR processors will be provided in Section 3.3.4.

Coluccia: “chapter1_v2” — 2022/10/7 — 13:05 — page 15 — #15
Model-Based Adaptive Radar Detection 15
as they are able to adjust the threshold in a dynamic way while preserving good
detection capabilities. An illustration of the typical behavior, also in comparison
to the NP based thresholding setting (i.e., with perfect knowledge of the noise
varianceσ
2
) is given in Figure 1.6.
However, the equivalence of the two processing approaches; that is, adaptive
detection followed by nonadaptive thresholding as in (1.20) and nonadaptive
detection followed by adaptive thresholding as in (1.21), does not hold true
in general. The ED case is rather an exception due to the lack of structure in
the signalsand white noise assumption. Adaptive detectors can better exploit
secondary data in presence of clutter and/or interference, as will be discussed in
Section 1.4. Before addressing such a case, we discuss the (kind of intermediate)
case of correlated signal in white noise.
1.2.4 Correlated Signal Model in White Noise
Detecting a correlated signal in white noise is a generalization of the noncoherent
problem (in Section 1.2.1), where instead the signalswas modeled as random
and white (uncorrelated). This intermediate case has found application in array
processing (e.g., DOA estimation and adaptive beamforming) as well as in wireless
communications. In the latter context, spectrum sensing is becoming increasingly
more important to overcome spectrum scarcity as part of thecognitive radio
Figure 1.6Illustration of the adaptive behavior of the threshold in CA-CFAR.

Coluccia: “chapter1_v2” — 2022/10/7 — 13:05 — page 16 — #16
16 Adaptive Radar Detection: Model-Based, Data-Driven, and Hybrid Approaches
paradigm: this requires uncooperative detection, through overhearing (sensing)
of the channel, of transmissions that can be originated by multiple sources (users).
The similarity of this problem with that of noncoherent radar detection is striking,
and in fact ED is a basic spectrum sensing tool as well. However, ED does
not exploit the fact that communication waveforms have a certain statistical
correlation over time (autocorrelation). Indeed, the assumption of white signal
forsintroduced in the NP-based derivation of the ED is a special case of a more
general, correlated signal model that is colored instead of white, which leads to
eigenvalue (EV)-based detectors.
Specifically, problem (1.12) can be restated by exploiting the structure of
the covariance matrixR
zofz, given by
R
z=

σ
2
I underH 0
Rs+σ
2
IunderH 1
(1.22)
whereR
sis the covariance matrix of the signalsandσ
2
Iis the covariance matrix of
the white noise (as before). Model (1.22) suggests interesting strategies for signal
detection: by noticing the relationship between the maximum and minimum
eigenvalues of the different covariance matrices; that is:

max(Rz),λmin(Rz))=
Ł

2

2
) underH 0
(λmax(Rs)+σ
2
,λmin(Rs)+σ
2
)underH 1
(1.23)
a detection statistics can be intuitively identified in their ratio, since
λmax(Rz)
λmin(Rz)
=1
underH
0while
λmax(Rz)
λmin(Rz)
>1 underH 1[31]. This leads to themaximum-
minimum eigenvalue detector(MMED):
λ
max(Rz)
λmin(Rz)
H1

H0
η. (1.24)
A simpler, alternative solution is to use as a detection statistic the ratio between the
energy of the signal and the minimum eigenvalue, thus obtaining theminimum
eigenvalue detector(MED)
νzν
2
λmin(Rz)
H1

H0
η. (1.25)
Other possibilities also exist, for instance the ratio between the sum of the
amplitude (modulus) of diagonal and off-diagonal entries ofR
z, which as for
the MMED is equal to 1 underH
0and greater than 1 underH 1, due to the fact
that the off-diagonal entries are expected non-zero only underH
1[32].

Coluccia: “chapter1_v2” — 2022/10/7 — 13:05 — page 17 — #17
Model-Based Adaptive Radar Detection 17
In practice,R zis unknown, hence such detectors are made adaptive by
replacing it with a covariance matrix estimate based on a sufficiently large number
of samples. In any case, the detector can be implemented without requiring
knowledge of the noise powerσ
2
, and for such a reason it is referred to as
blind detectionin spectrum sensing jargon. This corresponds to having the CFAR
property being the distribution underH
0independent of unknown parameters,
and hence the detection thresholdηcan be set offline irrespective of the noise
power (and signal covariance).
The reason for discussing the correlated signal model in white noise is that
it lends itself well to bridge the fully unstructured, uncorrelated case of the ED
to the structured, correlated case of interest in radar detection. Indeed, the MED
(1.25) is remarkably similar to the ED (1.16), the only difference being the
way the normalization of the energy is performed to gain the CFAR property
(detector (1.13), in fact, has a distribution underH
0that depends onσ
2
). In
the adaptive versions of the MMED/MED, this translates into havingλ
min(Rz)
replaced byλ
min(ˆRz), paralleling the substitution ofσ
2
withˆσ
2
in (1.20), as
a different estimator of the noise power. Thus, the same scheme of Figure 1.4
applies, although with a different definition ofˆσ
2
.
A natural question thus ultimately arises: What is the best estimator to
make the ED adaptive, in particular in the radar context? This issue is actually
much broader, as thetwo-step approachfollowed here—(i.e., derivation of test
statistics based on known noise statistics, then replacement of the unknown
parameters with estimates)—is generally suboptimal. Moreover, this should be
paired with the discussion at the end of Section 1.2.3, regarding the general
suboptimality of a nonadaptive detector followed by a CFAR processor (adaptive
thresholding) compared to the fully adaptive detection (which uses a nonadaptive
thresholding). This important difference and its remarkable implications will be
evident in Section 1.4.
To sum up this Section 1.2, a few observations are in order. Despite the
unstructured model bringing less information about the target compared to the
structured model that will be discussed in Section 1.3, from the development
above, one key advantage of model-based detectors can be already observed:
they typically provide understanding and performance guarantees. In particular,
knowing that a detector is the UMP test for the considered assumptions (thanks to
its NP-based derivation) means that there is no need to look for a better detector
for such a case provided that noise statistics are known. But the latter does not
hold true in practice, hence an intuitive (and quite effective) adaptation strategy
is to plug estimates of unknown parameters into the detection statistic. This is
nonetheless suboptimal, hence there is theoretically backed room for performance
improvement through more sophisticated approaches. Still, the CFAR property
of the detector properly made adaptive ensures that the threshold can be set offline
to guarantee a chosenP
FA, irrespective of the noise statistics actually encountered

Coluccia: “chapter1_v2” — 2022/10/7 — 13:05 — page 18 — #18
18 Adaptive Radar Detection: Model-Based, Data-Driven, and Hybrid Approaches
online (during the radar operation). However, as a matter of fact, extending
this intuition to more complex cases requires more sophisticated analytical tools.
Such a merging of intuitive reasoning and insights with an analytical, rigorous
development will be further developed later in Chapter 4, in particular by
discussing a feature-space interpretation of CFAR adaptive detection inspired
by a data-driven approach.
1.3 Structured Signal in White Noise
The aim of this section is to discuss several approaches to solve the detection
problem in which the signal of interest in (1.12) has the structures=αv, with
va known steering vector as introduced in (1.8). This problem is of greatest
interest in the context of coherent radars, since it makes better use of the available
information by coherent integration of the data, reflected in the presence of a
known (temporal, spatial, or space-time) steering vectorv. The literature over
several decades has been enriching, with a number of ideas and approaches that
rely on different modeling assumptions and background hypotheses. A number
of the most important ones will be reviewed throughout the rest of this chapter,
in particular based on the one-step and two-step GLRT methodologies, from the
pioneering Kelly’s detector and adaptive matched filter (AMF) to more recent
detectors. As a preparatory step, the structured signal model is addressed below
under multiple facets in the preliminary, simpler case of white noise.
1.3.1 Detection of a Structured Signal in White Noise and Matched Filter
Let us first consider the solution to problem (1.8) for white noisen∼
CN(0,σ
2
I)
with knownσ
2
. In general, according to the assumptions on the complex factor
α=Ae

, different derivations of the NP test are obtained. For known signal
phaseφand unknown signal amplitudeA=|α|, the hypothesis testing problem
reduces to testingA=0vsA>0. The LRT is thus given by
1
(πσ
2
)
Ne

(z−αv)

(z−αv)
σ
2
1
(πσ
2
)
Ne

z
†z
σ
2
=
e

νz−Ae
jφvν
2
σ
2
e

νzν
2
σ
2
H1

H0
η (1.26)
which after a monotonic transformation can be rewritten as
{e
−jφ
v

z}
H1

H0
η. (1.27)
The interpretation of this detector includes a filter matched tov, followed by a
phase compensation by the valueφ. In practice the phase of the target is hardly
known, hence a more reasonable option is to model it as a random variable

Coluccia: “chapter1_v2” — 2022/10/7 — 13:05 — page 19 — #19
Model-Based Adaptive Radar Detection 19
uniformly distributed in[0, 2π). To obtain the NP test for this case, which is
also known asaverage likelihood ratio(ALR), the PDF ofz(givenφ) must be
multiplied by the PDF ofφand integrated, to marginalize the phase according
to the law of total probability, so obtaining the (unconditional) PDF ofz,as
follows:
1
(πσ
2
)
N
1



0
e

νz−Ae
jφvν
2
σ
2dφ
1
(πσ
2
)
N
1



0
e

νzν

2dφ
∝I
0

2A|v

z|
σ
2

H1

H0
η (1.28)
whereI
0(·)is the modified Bessel function of order zero. Being the latter, a strictly
monotonic (increasing) function, an equivalent test is given by
|v

z|
2
H
1

H0
η (1.29)
which can be interpreted as the (squared) modulus of the output of the same
matched filter appearing in (1.27). This is reasonable, since the phase is unknown
hence it cannot be exploited by the detector. Notice that from (1.29) (but also
(1.27)) it is apparent the coherent integration of the data through the relationship
captured by the steering vectorv, while the ED (1.13) performs a noncoherent
processing. In particular, forv=1 (which corresponds to zero normalized
frequency, namely 0-Hz Doppler shift or 0

DOA) the ED in (1.13) computes
the sum of the squared modulus of the components ofz, while detector (1.29)
computes the modulus after complex-valued summation of I/Q samples. More
in general,|v

z|
2
can be interpreted as the energy of the output of a linear filter
with weightsv—indeed a matched filter in the temporal case, a beamformer in
the spatial case, or a vectorized 2-D filter in the space-time (STAP) case.
12
Notice
though that the MF (1.29) is not CFAR; again, it may be made adaptive by
dividing it by an estimate of the noise power based on secondary data, as done
for the ED to obtain (1.20), so producing
|v

z|
2
ˆσ
2
H
1

H0
η (1.30)
which instead possesses the CFAR property and may be implemented through a
CFAR processor as generally discussed in Section 1.2.3.
Clearly, the approach of introducing random models is a general one, and
several traditional radar detection schemes (sometimes based on a single pulse)
also consider a PDF for the target amplitude [4]. The drawback of this approach
is that quite strong assumptions on the target are made with no guarantee that
they will be met under operational conditions. Although one might partially
12. This interpretation will be further discussed in Section 1.4.2.

Coluccia: “chapter1_v2” — 2022/10/7 — 13:05 — page 20 — #20
20 Adaptive Radar Detection: Model-Based, Data-Driven, and Hybrid Approaches
extend this criticism to all model-based approaches, assumptions on other aspects
(namely, noise distribution) can be checked in a calibration phase and at runtime
based on secondary data, while by their nature target characteristics continuously
change and no additional data is typically assumed to be available for them.
13
Moreover, the integrals involved in the ALR are often analytically intractable.
Such motivations support the choice, in several advanced adaptive schemes, of
modeling unknown parameters as deterministic rather than random.
14
1.3.2 Generalized Likelihood Ratio Test
Unfortunately, UMP tests do not often exist because in practice more than
one parameter is unknown [33]. In other words, the hypotheses to be tested
becomecomposite(instead ofsimple), making the NP approach inapplicable.
This is already the case of unknownα∈
Cin white noise. Still, following the
same rationale adopted in Section 1.2.3 for making NP-based detectors adaptive
with respect to the noise power, one may generally consider the replacement of
other unknown parameters with corresponding proper estimates; in general, the
nonexistence of UMP tests explains the need to always develop new detectors
with improved performance in different situations.
A difficulty is however that, as mentioned at the end of the previous section
(Section 1.3.1), secondary data bring information only about noise, not useful
signal, since the presence of the target and its characteristics are unknown. Still,
the logic of the argument remains, and in fact one may consider a generalization
of the LRT where unknown parameters are conceptually replaced by estimates
as if the corresponding hypothesis were true. Indeed, a further rewriting of
the ED is
1
N
νzν
2
σ
2
H
1

H0
η (1.31)
which admits an intuitive interpretation as the ratio between the estimated signal
power and the true noise power (or its estimate based on secondary data, in case
of adaptive ED), aimed at solving a hypothesis test on the signal variance. In
particular,
1
N
νzν
2
is the sample variance of the data, which is a consistent and
unbiased estimator of the overall signal variance. More precisely, it is the maximum
likelihood (ML) estimator under the Gaussian assumption. Thus, the numerator
in the left-hand side of (1.31) accounts for the signal-plus-noise variance (under
13. This will be indeed one of the discriminant points for the possible adoption of data-driven
techniques, as discussed in Chapter 2, and more specifically, in Chapter 3.
14. See also [8, Chapter 2] for additional NP detectors under different hypotheses, in the more
general colored noise scenario (discussed later in Section 1.4), and for the GLRT framework
discussed next in Section 1.3.2.

Coluccia: “chapter1_v2” — 2022/10/7 — 13:05 — page 21 — #21
Model-Based Adaptive Radar Detection 21
H1) or noise-only variance (underH 0). As a consequence, the ratio in (1.31) is
reminiscent of the rationale underlying EV-based detectors in Section 1.2.4, and
can be adopted to test the estimated overall variance (noise plus signal, if any)
versus noise-only variance.
15
For the problem at hand (1.8), lacking any prior knowledge onα, its best
estimate can be found indeed by resorting to the ML approach, which guarantees
asymptotic unbiasedness and consistency with the sizeN(sample size) of the
observation vectorz[28]. The ML estimatorˆαis obtained by maximizing the
likelihood function with respect toα, so obtaining an estimate of the target
amplitude
ˆα=argmax
α
1
(πσ
2
)
N
e

νz−αvν
2
σ
2=argmin
α
νz−αvν
2
=
v

z
v

v
(1.32)
to be plugged into the likelihood ratio test (1.26). The resulting detector, GLRT,
is given by
e

νz−ˆαvν
2
σ
2
e

νzν
2
σ
2
=e

νP

vz
ν
2
?zν
2
σ
2
H1

H0
η (1.33)
whereP
M
=M(M

M)
−1
M

is the projector onto the subspace spanned by the
columns of the matrixM, andP

M
=I−P
M
the projector onto the corresponding
orthogonal complement. It follows thatP

v
z=(I−
vv

νvν
2)z=z−
v

z
νvν
2vand
νP

v

2
=z

P

v
z=νzν
2
−|v

z|
2
/νvν
2
, finally leading to
|v

z|
2
νvν
2
H
1

H0
η. (1.34)
Being the constantνvν
2
irrelevant,
16
test (1.34) is equivalent to the matched filter
(1.29), which was derived via the NP procedure assuming a random phase. This
intuitively confirms the soundness of the GLRT procedure, which has no general
optimality guarantees but can also lead to meaningful detection schemes in cases
where the NP approach is inapplicable. More specifically, the UMP test might not
exist but several reasonable schemes can be found through the GLRT approach,
as discussed in detail in Section 1.4 for the general colored noise scenario.
15. In fact, in the limitN→∞,
1
N
νzν
2
σ
2goes to 1 underH 0or to 1+SNR underH 1, similar
to the MED and related EV-based detectors.
16. It can be absorbed intoηas usual, or alternatively the steering vector can be normalized to
unitary norm by including in its definition a factor 1/

N.

Coluccia: “chapter1_v2” — 2022/10/7 — 13:05 — page 22 — #22
22 Adaptive Radar Detection: Model-Based, Data-Driven, and Hybrid Approaches
The GLRT approach can be applied analogously also in presence of several
unknown parameters, thus in general it can be written as
max
θ∈χ1
f(z|θ,H 1)
max
θ∈χ0
f(z|θ,H 0)
H1

H0
η (1.35)
wheref(z|θ,H
0)andf(z|θ,H 1)denote the PDF ofzunder hypothesis
H
0andH 1, respectively, given a vector of parametersθ, andχ idenotes
the corresponding set of admissible values. Since the two maximizations are
performed independently for the numerator and denominator, this exactly
corresponds to replacingθwith different ML estimatesˆθ
1andˆθ 0under the
two hypothesesH
1andH 0, respectively.
As a final remark, it is worth anticipating that additional data might be
needed for the two maximizations to be well-defined, in particular under colored
noise with unknown statistics. For the white noise case, instead, the detection
problem can be solved in a fully adaptive way by estimating alsoσ
2
under the
GLRT procedure; that is:
max
α,σ
2
1
(πσ
2
)
N
e

νz−αvν
2
σ
2
max
σ
2
1
(πσ
2
)
N
e

νzν
2
σ
2
H1

H0
η. (1.36)
Derivative-based maximization with respect toσ
2
returns the ML estimators
ˆσ
2
1
=
1
N
νz−αvν
2
andˆσ
2
0
=
1N
νzν
2
underH 1andH 0, respectively, which
substituted back in (1.36) yields the equivalent test
max
α
ˆσ
2
0
ˆσ
2
1
=
νzν
2
min
α
νz−αvν
2
=
νzν
2
νP

v

2
H
1

H0
η. (1.37)
This should be immediately contrasted with (1.33); in particular, the latter
computes the difference of the same quantities appearing as a ratio in (1.37),
resulting in a different behavior on the resulting statistics. Indeed, noticing that
νzν
2
=ν(Pv+P

v
)zν
2
=νPvzν
2
+νP

v

2
, it turns out that for the test
(1.37) the following chain of equivalences holds true:
νzν
2
νP

v

2

νPvzν
2
νP

v

2

|v

z|
2
1
N−1
νP

v

2
(1.38)
sinceνPvzν
2
=z

Pvz=|v

z|
2
/νvν
2
. The denominator of the right-most
expression in (1.38) is an unbiased estimator of the noise power, sinceP

v
z
does not contain any coherent signal anymore, as only noise components are

Coluccia: “chapter1_v2” — 2022/10/7 — 13:05 — page 23 — #23
Model-Based Adaptive Radar Detection 23
left after the orthogonal projection; indeed, simple calculations show that
17
E[
1
N−1
νP

v

2
]=
1
N−1
E[νP

v

2
]=σ
2
. As a result, (1.37) can be shown to
be CFAR thanks to the denominator that provides adaptivity to the noise power,
similarly to what was observed in Section 1.2.3 and Section 1.2.4 (particularly
regarding the adaptive ED (1.20) and MED (1.25)).
Besides the different way the noise power is estimated, the main difference
between (1.37) and (1.20) is that the former does not use secondary data, as
it estimatesσ
2
from the noise subspace (orthogonal to the signal subspace,
spanned byv); such a decomposition is depicted in Figure 1.7. Indeed,v

zis
the scalar product between the data vectorzand the steering vectorv, which
after normalization is equal to the estimated target amplitude (1.32). Thus,
the second expression in (1.38) can be geometrically interpreted as the ratio
of the (squared) lengths of the data vector projections onto the signal and noise
subspaces, respectively, which in fact increases as a stronger target echo component
is present inz. The colored noise case, addressed in Section 1.4, will require
instead a set ofK≥Nsecondary data to obtain the GLRT under both one-step
or two-step approaches.
Figure 1.7Geometrical representation of noise subspace (spanned by the steering vectorν)
and signal subspace (orthogonal complement), and examples of decomposition of
the data vectorzunder the two hypotheses (left:H
0, right:H
1).
17. In fact:
1
N−1
E[νP

v

2
]=
1
N−1
E[νnν
2
]−
1
N(N−1)
E[|n

v|
2
]
=
N
N−1
σ
2

1
N(N−1)
v

E[nn

]v
=
ł
N
N−1

νvν
2
N(N−1)

σ
2

2
sinceνvν
2
=N.

Coluccia: “chapter1_v2” — 2022/10/7 — 13:05 — page 24 — #24
24 Adaptive Radar Detection: Model-Based, Data-Driven, and Hybrid Approaches
Finally, the left-most form of the detector in (1.37) may be also interpreted,
in the light of the insights from (1.38), as the adaptive version of detector (1.31),
thus confirming the capability of the GLRT procedure to deliver meaningful
detection schemes, despite its lack of theoretical optimality.
1.3.3 Detection of an Unknown Rank-One Signal in White Noise
Still considering white noise, a variation of the detection problem addressed above
is worth a brief discussion. If one relaxes the structure of the steering vector, a
situation in between the unstructured random models(white or correlated) and
the structured modelαvis obtained. In particular, the datazcan be arranged
in aN
b×NmatrixX, withN bdenoting the number of beams used to cover a
spatial sector or the number of array sensors, so obtaining the hypothesis testing
problem

H
0:X=N
H
1:X=a Sa

T
+N
(1.39)
wherea
Sanda Tare unknown arbitrary spatial and temporal steering vectors,
respectively, andNis the corresponding white noise matrix [34]. The GLRT in
this case is given by
λ
max(XX

)
Tr{XX

}
H1

H0
η. (1.40)
This decision statistic shows a similarity with that of the MMED in (1.24), being
XX

proportional to the sample covariance estimate ofR z. In particular, under
H
0,Tr{XX

}is proportional to the sample variance estimate ofσ
2
. Moreover,
under both hypothesesλ
max(XX

)accounts for the energy of the useful signal.
As a consequence, the ratio in the left-hand side of (1.40) is independent of
any unknown parameter underH
0, hence the detector is CFAR. Moreover, its
detection performance (equivalently, its distribution underH
1) depends only on
the SNR=νa

2
νaTν
2

2
[34].
Notice that treating the steering vectors as unknown and arbitrary
quantities may be meaningful when uncertainties exist, for instance because of
miscalibration or other mismatches. In such cases, in fact, looking for a specific
knownvmight counterproductively result in signal rejection, as it will be better
discussed in Section 1.4.3. The detection scheme (1.40) may be viewed as a
fast and simple preprocessing step, which upon detection of a possible target is
followed by a more sophisticated detector to confirm or conversely invalidate the
decision.
18
Remarkably, ifa Sis modeled as a random vector according to the
same model
CN(0,σ
2
s
I)adopted in Section 1.2 (instead of an unknown vector),
18. It has been observed in [34] that (1.39) is also relevant when a monopulse radar is used,
forN
b=2 and the rows ofXcorresponding to the outputs of the sum and difference channels,
respectively.

Coluccia: “chapter1_v2” — 2022/10/7 — 13:05 — page 25 — #25
Model-Based Adaptive Radar Detection 25
the GLRT is still given by (1.40). Finally, forN b=2 this is the UMP invariant
(UMPI) test (i.e., the UMP among a family of detectors that are invariant under
a suitable group of transformations). The importance of the maximal invariant
theory will be discussed later in Chapter 4. Nonetheless, in many practical radar
detection schemes the steering vector is considered with known structure, up to
one parameter, typically related to the Doppler or DOA; the analytical justification
for this approach is discussed below.
1.3.4 Steering Vector Known up to a Parameter and Doppler Processing
Testing a CUT actually means that the surveillance region is scanned in range and
azimuth according to a certain policy, while different Doppler shifts are tested.
The Doppler domain is in fact quantized intoQvalues, and the resulting specific
steering vectors are individually used for target detection. The same applies to
azimuth scanning through an active electronically scanned array (AESA).
This intuitive approach, which is usually interpreted as apulse-Doppler
matched filterbank, can be actually formalized by saying thatvis known up to one
parameter, namely the normalized Doppler frequency
19
νT. For such a problem,
the GLRT derivation would simply include a further, outer maximization with
respect toν
T, so justifying the aforementioned matched filterbank detector,
max
q=1,...,Q
|v


(q)
T
)z|
2
νv(ν
(q)
T

2
H
1

H0
η (1.41)
as depicted in Figure 1.8. As a by-product of this processing, an estimate of the
Doppler frequency shiftf
D, hence of the radial velocity of the target, is obtained.
Interestingly, recalling the structure of a steering vector (see Section 1.1.2), the
processing above can be interpreted as a peak finding on the frequency spectrum
obtained via discrete Fourier transform (i.e., using FFT).This justifies the practical
approach of performing Doppler processing followed by a conventional CFAR
processor. However, inthemilestonepaperbyBrennanandReed[35](space-time)
adaptive detection was introduced, which paved the way to a number of adaptive
detectionschemes—inelement-spaceorbeamspace, pre-Dopplerorpost-Doppler
(see [18])—as well as advanced detection techniques, discussed next.
1.4 Adaptive Detection in Colored Noise
Capitalizing on the comparison of the many different approaches to adaptive
target detection in white noise performed in the previous sections, it is now easier
19. A similar formalism applies to DOA using the normalized spatial frequencyν S. Extension to
more than one parameter is also straightforward.

Coluccia: “chapter1_v2” — 2022/10/7 — 13:05 — page 26 — #26
26 Adaptive Radar Detection: Model-Based, Data-Driven, and Hybrid Approaches
Figure 1.8Pulse-Doppler processing via matched filterbank.
to finally address the general problem (1.8) where the covariance matrix of the
noiseR=
E[nn

]is generally nondiagonal.
First of all, consider that the detectors presented so far in this chapter,
although derived assuming white noise, can be straightforwardly generalized to
colored noise ifRis known. In that case, in fact, a whitening preprocessing can
be performed through the square root of the matrix inverse, hence all equations
for white noise still apply under the substitution ofz,s,vwith˜z=R
−1/2
z,
˜s=R
−1/2
s,˜v=R
−1/2
v, since the transformed noiseR
−1/2
ngives back the
white noise condition
E[R
−1/2
n(R
−1/2
n)

]=R
−1/2
E[nn

]R
−1/2
=I.
In practice, however,Ris seldom known and hence must be estimated.
To this end, a set of secondary data is necessary, as for the estimation ofσ
2
in white noise discussed in the adaptive ED (1.20) or adaptive MF (1.30). We
recall that other schemes discussed above, instead, do not require secondary data,
in particular the MMED (1.24) and MED (1.25), as well as the GLRT-based
coherent detector (1.37) and rank-one signal detector (1.40). Unfortunately,
this desirable feature cannot be paralleled, in general, by the more general case
of unknownR. As a confirmation, notice that not even the noise power can
be unbiasedly estimated from the noise subspace ifRis unknown. In fact, the
denominator of the GLRT-based detector (1.37), which as discussed provides
adaptivity to the detection statistic, for colored noise has expected value
E[νP

v

2
]=Tr{R}−
v

Rv
νvν
2
.
By separating the power level from the structure of the covariance matrix, for
example,R=σ
2
Cwhere diag{C}=1(for instance, an exponentially-correlated,

Coluccia: “chapter1_v2” — 2022/10/7 — 13:05 — page 27 — #27
Model-Based Adaptive Radar Detection 27
first-order autoregressive structure is a Toeplitz matrix withρ
|k−h|
elements,
k,h=1,...,Nandρthe one-lag correlation coefficient), it turns out that
E[νP

v

2
]=σ
2

Tr{C}−
v

Cvνvν
2


2

N−
v

Cv
N

so even unbiased estimation (based on the noise subspace) of the soleσ
2
would
require knowledge of the structureCof the covariance matrix. The usualN−1
normalization factor is obviously retrieved in the white caseR=σ
2
I, whereas the
estimator can be made unconditionally unbiased, due to the fact that the number
of parameters to be estimated certainly exceeds the numberNof data samples in
z. Covariance matrix estimation involves the determination ofN
2
parameters.
This number can be lowered by considering special symmetries and estimating
the sole independent parameters. Examples are the Hermitian structure with
respect to the main diagonal, the persymmetry also with respect to the cross
diagonal, or the low-rank structure in which only a few eigenvalues ofRare
significantly different from zero. Still, the number of parameters to be estimated
often exceedsN. Moreover, in practical settings such special symmetries may be
violated due to non-idealities, hence estimating the full covariance matrix may
represent in this respect a more robust approach. Secondary data is thus typically
required for high-performance adaptive detection in general scenarios, but still
two different strategies are possible; that is, eitherone-step(1S) andtwo-step(2S)
approaches, which lead to detectors with diverse characteristics. The well-known
Reed–Mallett–Brenann rule (RMB) [36] specifies that, compared to the optimum
case of known covariance matrix,K=2Nindependent identically distributed
data produce an SNR loss of 3 dB.
1.4.1 One-Step, Two-Step, and Decoupled Processing
In 2S approaches the detector is devised, by means of GLRT or alternative
procedures, assuming thatRis known. Then, in a second step, a covariance
matrix estimateˆRobtained from secondary data is used in place ofR. This is
conceptually the same as for the white noise case (see Section 1.2.3), however
with two differences:
1.N-dimensional vectors are needed as secondary data, instead of scalar
values;
2. Making the detection statistics (not only adaptive but also) CFAR entails
additional complications.
Regarding point (1), such data can be typically obtained from neighboring range
cells in fast-time (range bins), using the same temporal, spatial, or space-time data
arrangement as for the CUT samples. Regarding point (2), the CFAR property can
be embedded into an adaptive statistics provided that the statistical distribution

Coluccia: “chapter1_v2” — 2022/10/7 — 13:05 — page 28 — #28
28 Adaptive Radar Detection: Model-Based, Data-Driven, and Hybrid Approaches
of the detector can be made independent not only of the noise power, but also
of the structure of the true covariance matrix. This feature, often referred to as
generalized CFAR property, must be checked on a case-by-case basis, and is clearly
a significant point of merit for an adaptive detector.
A simplifying approach is instead to decouple filtering and detection,
considering them as two separate processing stages. Specifically, multidimensional
(multichannel) data from the CUT are first filtered to suppress clutter/jamming
as much as possible, via heuristic (nonadaptive) clutter cancellers such asmoving
target indicator(MTI), pulse-Doppler matched filterbank (see Section 1.3.4), or
adaptive processing (beamforming or STAP; see Section 1.1.2), so compressing
theN-dimensional data vector to a scalar complex value;
20
then, a decision
is taken based on a proper (real-valued) statistic of the filtered output, and
an adaptive threshold provided by a CFAR processor. This scheme is classical
and in fact reasonable, sinceP
Ddepends on SNR hence improving the latter
via preliminary filtering for disturbance suppression is always beneficial to the
detection performance. Moreover, the use of a CFAR processor guarantees the
desiredP
FAby adjusting the threshold to changes in the noise power level.
However, best filtering capabilities that can significantly enhance the visibility
of targets buried in noise require optimal weighting based on the true covariance
matrixRof the overall disturbance, hence can be implemented in practice only
suboptimally using an estimateˆR. More advanced schemes can be devised by
trying to estimate at once all the unknowns while at the same time adapting the
statistic for CFARness (i.e., a 1S approach). In doing so, filtering, detection, and
CFAR-processing are all embedded into a fully adaptive CFAR detection statistic
(followed by nonadaptive thresholding).
From the discussion above, it is clear that covariance matrix estimation is a
common aspect to both adaptive filtering followed by a CFAR processor and 2S
adaptive detection; the latter may also embed CFAR capability into its statistic,
or not, as discussed below. In 1S fully adaptive approaches, instead, covariance
matrix estimation still plays a role but it cannot be disentangled, in general, from
the detection structure of the statistic, thus providing better performance due to
the joint processing.
1.4.2 General Hypothesis Testing Problem via GLRT: A Comparison
The generic detection problem (1.8) can be restated more explicitly by assuming
complex Gaussian colored noise with unknown covariance matrixRand a
20. As discussed in Section 1.1.2, a further decoupling is often introduced in traditional radar
schemes in which beamforming and Doppler filtering are performed separately. This is suboptimal
since clutter returns may generally manifest themselves as fully 2-D (nonfactorable) structures
[13]. Indeed, STAP combines the signals in space and time, thereby accounting for angle-Doppler
coupling [13, 18].

Coluccia: “chapter1_v2” — 2022/10/7 — 13:05 — page 29 — #29
Model-Based Adaptive Radar Detection 29
coherent target with unknown deterministicα, as follows:

H
0:z∼CN(0,R)
H
1:z∼CN(αv,R)
. (1.42)
A straightforward solution to this problem can be found by following a 2S-GLRT
approach, so obtaining the colored-noise counterpart of the MF discussed in
Section 1.3. To this aim, the matrixRis first assumed known, then the whitening
transformation˜z=R
−1/2
zdirectly yields to the same problem of Section 1.3.2.
Maximization with respect toαyields the ML estimate of the target amplitude as
ˆα=
v

R
−1
z
v

R
−1
v
(1.43)
which generalizes (1.32) to the case of colored noise, and finally leads to the
generalization of detector (1.34); that is:
t
MF(z)=
|˜v

˜z|
2
˜vν
2
=
|v

R
−1
z|
2
v

R
−1
v
H1

H0
η. (1.44)
The processing described byv

R
−1
zhas the meaning of a scalar product between
the whitened data˜z=R
−1/2
zand the whitened steering vector˜v=R
−1/2
v,
and in fact it is a generalization of the MF for white noise (1.34). Detector (1.44)
involves the cascade of a (whitening) transformation aimed at suppressing the
clutter and a projectionP
˜v=
˜v˜v

ν˜vν
2aimed at performing coherent integration of
the useful signal; the final statistic can be regarded as a square-law detector on
the output of the cascade filter (i.e.,νP
˜v˜zν
2
=˜z

P˜v˜z=
|v

R
−1
z|
2
v
†R
−1
v
), which can
be interpreted geometrically in a way similar to Figure 1.7.
The colored noise version of the ED can be also obtained, for example:
t
ED(z)=νR
−1/2

2
=z

R
−1
z
H1

H0
η (1.45)
which, as usual, does not exploit any coherence in the detection task. Both MF
and ED statistics for white noise are obviously retrieved forR=σ
2
I.
To implement the detectors above, an estimate ofRis needed. This is
obtained from a set ofKsecondary data, denoted asz
1,...,z K, withK≥N.
For convenience, the matrix notationZ=[z
1...zK]is also used, which yields
a simple expression for thesample covariance matrix(SCM):
ˆR
SCM=
1
K
K

i=1
ziz

i
=
1
K
ZZ

=
1
K
S (1.46)

Coluccia: “chapter1_v2” — 2022/10/7 — 13:05 — page 30 — #30
30 Adaptive Radar Detection: Model-Based, Data-Driven, and Hybrid Approaches
whereS=ZZ

is often referred to asscatter matrix. In particular, by substituting
ˆR
SCMin (1.44), the following equivalent detector is obtained:
t
AMF(z,S)=
|v

S
−1
z|
2
v

S
−1
v
H1

H0
η. (1.47)
It can be shown that the distribution oft
AMFdoes not depend on any unknown
parameter underH
0[37]; that is, it has the generalized CFAR property, where, as
already mentioned, the adjective “generalized” is sometimes used to denote that
CFARness holds true with respect to the whole covariance matrixR(not only to
its powerσ
2
=Tr{R}/N). Conversely, the adaptive version of the ED (1.45),
that is:
t
AED(z,S)=νS
−1/2

2
=z

S
−1
z
H1

H0
η (1.48)
does not have either the CFAR property (with respect toσ
2
) or the generalized
CFAR property (with respect toR). Remarkably, while the former can be gained
by using an estimatorˆσ
2
as normalization factor (as shown in Section 1.2.3), the
latter remains unsatisfied; at the same time, it can be expected that such a detector
can provide betterP
Dcompared to CFAR detectors, being less constrained. We
will analyze later the different performance trade-offs between adaptive detectors.
In his 1986 pioneering paper [38], Kelly derived a GLRT for problem
(1.42) assuming unknown (Hermitian) positive definite covariance matrixR,
includingK≥Nindependent and identically distributed secondary data
z
1,...,z K(independent ofz, free of target echoes, and sharing with the CUT the
statistical characteristics of the noise) directly into the hypothesis testing problem
formulation; that is:







H
0:z∼CN(0,R)
z
i∼CN(0,R),i=1,...,K
H
1:z∼CN(αv,R)
z
i∼CN(0,R),i=1,...,K
. (1.49)
With a slight abuse of notation, the resulting GLR can be thus written as
max
α,R
f(z,Z|H
1)
max
R
f(z,Z|H 0)
=
max
α
max
R
Ł CN(αv,R)×
K
˝
i=1
CN(0,R)
˛
max
R
Ł CN(0,R)×
K
˝
i=1
CN(0,R)
˛. (1.50)
The presence of the secondary data distribution makes well-defined the
additional maximization with respect toR, necessary in both the numerator and

Coluccia: “chapter1_v2” — 2022/10/7 — 13:05 — page 31 — #31
Model-Based Adaptive Radar Detection 31
denominator, which can be solved in closed form producing the ML estimates
ˆR
1=
1
K+1
[S+(z−αv)(z−αv)

] (1.51)
and
ˆR
0=
1
K
S (1.52)
underH
1andH 0, respectively. The maximization with respect toαto obtain the
1S-GLRT detector follows straightforwardly, as it turns out that the ML estimate
ofαis simply the adaptive version of (1.43) performed in 2S-GLRT; that is
21
ˆα=
v

S
−1
z
v

S
−1
v
. (1.53)
Substituting into the LR, the resulting detector can be equivalently written as
t
Kelly(z,S)=
|v

S
−1
z|
2
v

S
−1
v(1+z

S
−1
z)
H1

H0
η. (1.54)
The statistic in (1.54) is surprisingly similar tot
AMFin (1.47), but for the
presence of the additional term between brackets in the denominator, which
is equal to 1+t
AED. Notice that, while the AMF statistic can be expressed
as the output of a filter followed by a detector, Kelly’s statistic cannot. More in
general, this highlights the suboptimality of the decoupled approach, as discussed
in Section 1.4.1, and shows the capability of the 1S-GLRT to capture the
intrinsic coupling in the problem, compared to the 2S-GLRT. This is due to
the better use of secondary data, included in the hypothesis testing formulation
since the beginning, and can be shown to yield higherP
Dthan AMF, for the
sameP
FA,
22
SNR, and keeping the generalized CFAR property. In summary,
Kelly’s detector represents a paradigmatic example of radar detector embedding
the three functions of clutter suppression (filtering), coherent signal integration
with optimal detection, and CFAR processing, into a unique adaptive statistic.
As such, it is considered as a benchmark for advanced adaptive radar detectors.
1.4.3 Behavior under Mismatched Conditions: Robustness vs Selectivity
What is unexpected, at first glance, in the different reaction of Kelly’s detector and
AMF to mismatches in the signal model, given their otherwise similar behavior
21. Note that such an estimate of the target amplitude is also reasonable in non-Gaussian
environments, where it corresponds to the least-squares (LS) estimator (not the ML). Examples of
radar detection in non-Gaussian disturbance will be given in Chapters 3 and 4.
22. For Kelly’s detector,P
FAcan be also expressed in a simple closed form asP FA(η)=
1(1+η)
N−K+1
, which can be easily inverted to set the threshold that guarantees the desired false
alarm rate.

Coluccia: “chapter1_v2” — 2022/10/7 — 13:05 — page 32 — #32
32 Adaptive Radar Detection: Model-Based, Data-Driven, and Hybrid Approaches
in terms ofP D, especially for largeK. Mismatches may arise due to several
uncertainties on the steering vectors and/or the waveforms, in terms of beam-
pointing errors, array miscalibration, angle/Doppler quantization with off-grid
targets, waveform and array distortions, and other reasons [39–42].
In such cases, the desired behavior of the detector depends on the
application: for instance, a selective detector is desirable for target localization;
instead, a certain level of robustness to mismatches is preferable when the radar
is working in searching mode. In [43] the performance of Kelly’s detector was
assessed when the actual steering vector, sayp, is different from the nominal one
v. The analysis showed that it is a selective detector (i.e., it tends to reject signals
not arriving from the nominal direction). The mismatch level is measured by the
squared cosine of the angleθbetween the two directions; that is:
cos
2
θ=
|p

R
−1
v|
2
p

R
−1
pv

R
−1
v
. (1.55)
Figure 1.9 reports the so-calledmesa plots, that is iso-P
Dcontour curves for
Kelly’s detector, AMF, and theadaptive coherence estimator(ACE) [44] (also called
adaptive normalized matched filter [45]) whose statistic is given by
t
ACE(z,S)=
|v

S
−1
z|
2
v

S
−1
vz

S
−1
z
. (1.56)
Mesa plots are very informative, since they highlight how the detectors behave
under matched conditions (cos
2
θ=1, corresponding to the top horizontal axis)
Figure 1.9Mesa plots of Kelly’s and AMF and ACE detectors, forN=16,K=32, and
P
FA=10
−4
.

Coluccia: “chapter1_v2” — 2022/10/7 — 13:05 — page 33 — #33
Model-Based Adaptive Radar Detection 33
as well as mismatched conditions for various levels of mismatch (until cos
2
θ=0,
which means thatvandpare orthogonal), as a function of the SNR, defined as
SNR=|α|
2
p

R
−1
p. (1.57)
From Figure 1.9 it is evident that Kelly’s detector guarantees the highestP
Dunder
matched conditions, while AMF and ACE experience a certain performance loss;
moreover, AMF is a robust detector, ACE is a quite selective one, while the
behavior of Kelly’s detector is selective but comparably more moderate. Notice
that the ACE statistics can be rewritten as
t
ACE=
t
AMF
tAED
(1.58)
which can be interpreted as the detecting the presence of a significant coherent
component (alongv) out of the overall signal energy. It is worth remarking once
again that the AED totally ignores the information on the steeringv(it is an
incoherent processing), hence when used alone as detection statistic will lead to
reducedP
Dbut strong robustness. By comparison, Kelly’s detection statistic is
conversely expressed by
t
Kelly=
t
AMF
1+tAED
(1.59)
hence Kelly’s and ACE detectors differ only for the additional unity in the
denominator, which however makes a significant difference in the performance
under both matched and mismatched conditions. The lesson learned is that
diversified behaviors may arise by minor changes in the detection statistic.
Along this line, many authors have thus addressed the problem of enhancing
either the robustness or the selectivity of adaptive detectors. However, this is a
nontrivial task, since detectors should still guarantee the CFAR property. Typical
design procedures include statistical tests with modified hypotheses, asymptotic
arguments, approximations, and ad hoc strategies, as presented next.
1.4.4 Model-Based Design of Adaptive Detectors
Several decades of work have already been dedicated to derive model-based
adaptive detectors through different approaches. One of the most interesting
ideas is to modify the hypothesis test to promote some characteristics in its
solution (i.e., to obtain a detector with desired properties). In the following,
we review tunable receivers, the subspace approach, the orthogonal rejection
(adaptive beamformer orthogonal rejection test (ABORT)) approach, the cone
acceptance/rejection approach, and second-order approaches. Other approaches
different from GLRT also exist, like Rao’s and Wald’s tests [30], which return
alternative detectors with specific characteristics. For instance, a detector based

Random documents with unrelated
content Scribd suggests to you:

– Vigye haza az én feleségemet. Mindegyiknek a palota készen
van, kiki szálljon bé a maga feleségével együtt. A legküssebb leányt,
az én feleségemet tartsa magánál, mert ingemet az ördögnek adott,
mikor a kapun keresztül mentek, és én mennem kell a pokolba. De
onnat ës kiszabadulok, és jól kapjam a feleségemet.
Azzal hazamentek a bátyjai, kiki a maga szobájába vitte a
feleségét. De mivelhogy soká maradt el a kicsi királyfi, az ő feleségét
libapásztornak tették ki a mezőbe.
Elmënën a kicsi királyfi az ördöghöz, mivelhogy akkor csak egy
ördög volt a világon, és béköszön hëzza:
– Jó napot, nagyapám.
Az ördög kést hajított utánna és azt mondja, néki:
– Mondjad: jó napot, édes apám.
Azt mondja a kicsi királyfi:
– Nem mondom addig, míg meg nem fërëdsz egy kád tébe
(tejben).
Az ördögnek tetszett a fërëdés s azt mondta, nem bánja, ő
mëgfërëdik. A királyfi megtöltött ën nagy kádat tével, és akkor
bément az ördöghöz és azt mondta az ördögnek:
– Volt magának egy lova vaj egyszer. Meg van-é a kantárja?
Azt mondja az ördög:
– Valahol a füstön megkapja.
Elment a királyfi és csakugyan megkapta a kantárt. Aval a földhöz
ütött, és a tátos ló kijött a földből. Akkor azt mondja a lónak a
királyfi:
– Fújd meg a tejet, hogy főjön a kádban.

A ló csakugyan belefújt a tébe, hogy az igazán főtt a kádban. Azt
mondta most a lónak:
– Fújd meg a tetején, hogy csak kissé legyen meleg, de alatta
főjön a kádban a té.
Akkor bément az ördöghöz és azt mondta neki, hogy mán készen
van, jőjön fërëdőre. Az ördög kiment és látta, hogy csak gyenge
meleg a té, és beleugrott a tébe, a kádba. Akkor azt mondta a
királyfi a lónak:
– Fújd meg a tejet, hogy főjön igazán.
És az ördögöt megfőzte és kiszedte a csontjait és ësszetörte
mozsárban és a szélnek eresztette. Abból minden morzsából egy
ördög lett, s azután lett a nagy pokol.
Ekkor felült a tátos lóra és hazament az édesapja házához. A kis
leány, az ő felesége, őrzötte a libákat a mezőn, és ő az ostorával egy
ütésre tizet lecsapott. A kis leány sírt, hogy azért a bátyjai meg
fogják ölni. Azt mondta a királyfi:
– Ne búsulj, mert nem lesz bántásod.
Aval hazament a leánynyal. Azt mondta a bátyjainak:
– Én palotát csináltam néktek, az én feleségemet rátok bíztam, tü
pedig libapásztornak tettétek, engemet ördögnek adtatok, de onnét
ës kiszabadultam, de nektek most pusztulni kell!
Azzal a lovával elfuvatta a palotákat, hogy csak a poruk maradt a
helyén. Az öreg király pedig királyságát neki adta, és máig ës élnek,
ha meg nem haltak.
(Fejér M. György, Pürkerecz.)

18. Ezeregy János.
Volt egyszer egy ember, Ezeregy Jánosnak hítták. Volt neki
ezeregy fia, s ő úgy akarta, hogy azokat egyszerre házasítsa meg.
Elindult az úton, hogy kapna valahol ezeregy leányt. A mint ment,
mendegélt az úton, egy vasárnap reggel, hát látta, hogy egy ember
a jegenyefa tetején cséplett kölest. S kérdette tölle:
– Mit csinálsz, mit dolgozol?
Azt felelte vissza:
– Hogyne dolgoznám, mikor ezeregy leányom vagyon. Muszáj,
hogy dolgozzam, csépeljek, mert nem győzöm őket étellel, ruhával.
Azt felelte az ember, hogy:
– Nékem ezeregy fiam van, úgy akartam, hogy kapjak valahol
ezeregy leányt, hát jó hogy összetalálkozunk. Gyerünk, nézzük meg
a leányokat.
Azonnal meg is kérte őket, mikor megnëzte. Egyik mezítlen volt,
a másik még mezítlenebb, hát gyenge állapotban voltak. Oda ës
igérték.
Aval hazament Ezeregy János. A legküssebb fia már tudta, mert a
tátos volt, s azt mondta, hogy:
– Jó, hát menjenek el és hozzák el a leányokat, addig
mindnyájoknak házat csálok. De van egy gyémánterdő s egy
aranyerdő és egy ezüsterdő. De arra menet vagy erre jövet egy ágat
se bántsanak, mert jaj lesz nekem.

El ës mentek, azt fogadták, hogy nem bántnak semmit. Ott a
többi mëgeskëttek a leányokkal, csak ő nem, mert nem volt ott. És
hogy ettek s ittak jól, megittasodtak. Mikor visszajövet a
gyémánterdőn és aranyerdőn keresztül mentek, nem bántottak
semmit, de mikor az ezüsterdőn jöttek, kezdték az ágakat tördelni le,
mert azt hitték, hogy ott már nem bántja senki. És mikor kijöttek az
erdőből, a hétfejű sárkány elejükbe állott, hogy mért rontották ëssze
az ezüsterdőt, és azt mondta, hogy egyiknek sem szabad elmenni,
még meg nem fizetik az erdőt, vagy egy fiát ott nem hagyja.
Az atyjuk azt gondolta fel magában, hogy azt a fiút, a melyik
otthon van, azt odaigéri, mert abból úgy se lesz semmi.
Mikor hazaértek, az ezeregy ház készen volt az ezeregy fiúnak.
De az otthon való fiú búsult erősen, mert tátos volt, mert tudta,
hogy mit csináltak az édes apjáék. Kérdezte a fiút:
– Mért búsulsz fiam, mert elhoztuk neked ës a feleséget, mind a
többieknek.
Azt felelte:
– Nékem jábo (hiába) hozták, mert megmondtam, ne bántsanak
semmit egy erdőn ës keresztül. Engemet mán odaigért a hétfejű
sárkánynak, mán holnap oda kell menjek és itt kell hagyjam a
feleségemet.
Arra azt felelte az édes apja:
– Azzal ne törődj semmit, mert nem kell menned, csak
kóbítottunk neki, azért ne menj el.
– De nekem mennem kell, – azt felelte, és el ës indúlt, és mikor
odaért a sárkánykirályhoz, köszöntötte:
– Adjon isten jó napot. Ammint ígérte apám, megjelentem
szolgálatjára.

– Hát akkor itt maradsz nállam addig, míg hogy három
kérdésemre meg nem felelsz. Itt az égő kemencze, ha ki nem
találod, a mit kérdek, ezer itt elégett már, te léssz az ezeregyedik.
A fiú azt felelte, hogy jó lesz, és azt kérdezte tölle (t. i. a
sárkány), hogy:
– Van nékem három leányom. És egyforma a három. Holnap
reggel, ha meg nem mondod, melyik a küssebb, fejedet veszem.
És elrekesztette a legényt egy külön szobába. És a leány is, a
legküssebb, tátos volt. Mikor az apjáék lefeküdtek, a sárkánykirály,
akkor a leány a legényhez bébútt a kócsjukon s azt kérdezte tölle:
– Nézd meg, nekünk az ujjunkon mindegyiken egyforma gyürű
van, és az én középső ujjamon küjebb lesz húzva a gyürű, mind a
többiekén s arról megtudod, hogy én vagyok a legküssebb. És édes
apám reggel körbe állít hármunkat és az ujjamon a gyürűt nëzd jól.
Reggel korán a sárkánykirály eléállította a leányokat és a tátos
legényt elévezette és kérdette:
– No most mondd meg, hogy melyik a legküsssebb leányom,
mert hanem a fejedet veszem.
A legény nëzëgette őket és egyszer azt felelte:
– Ez a legküssebb leányod, sárkánykirályom!
Akkor azt mondta a király:
– Igazad van, ha önfejedtől van, ha nem tanított senki.
Úgy még visszatette a szobába, hogy holnap reggel mutassza
meg, hogy melyik lesz a második leányom.
És úgy a második leány ës bébútt a kócsjukon és azt mondta a
legénynek, hogy az ő fejrevalója (kendője) nem lesz úgy megkötve,
mind a többieknél, meg lesz tágítva a fején.

Reggel korán megint sorba állította őket és még eléhívta Jánost,
hogy mondja meg, melyik a második leányom, mert hanem fejét
veszi. És János egy keveset gondolkozott, és mindjárt megmondta,
hogy melyik a második leánya. És azt mondta a sárkánykirály:
– Igazad van, János, ha önfejedből tudod s nem tanított meg
senki.
És így történt a harmadikkal ës mind. Mikor kijött a harmadik
próbára, akkor ës rëaüsmert, hogy melyik a nagyobbik leánya, mert
meg volt tágítva a fülbevalója. S ismét mondta a király:
– Igazad, van János, ha meg nem tanított rá senki. Hát téged
megszabadítlak, hogy nem égetlek el. De itt tartlak élted hosszáig,
itt szolgálj engemet.
Hát míg ott szolgált, szerelembe esett a küssebb leánynyal. És
mindig kéregette a sárkánykirályt, hogy adja neki a legküssebb
leányát. De a király sohasem akarta a leányt oda adni a szolga
legénynek. Hát egyszer ësszebeszéltek a leánynyal, hogy ugorjanak
meg, vessenek kecskebucsát. S lett belöllük két galamb, s úgy
elrepültek, hogy máig ës repülnek, ha meg nem állottak valahol.
(Papp A. István, Pürkerecz.)

19. A csaló vándorló legények.
Voltak egyszer valami vándorló legények hárman. Ők
vándoroltak, elmentek egy városba és nékik nem volt egy keréczár.
Hát azt számították, hogy menjenek és egyik egyet szerezzen bé, a
másik mást.
Az első azt mondja, hogy ő megy, és kenyeret szerez. Hát ő
elment. A mind a szolgáló mënyën a kenyérszërën, kérdi tölle:
– Nincs egy kis hubeszkád, hugom?
Hát azt mondja a szolgáló:
– Álljon meg egy kicsit, leteszem a kast, visszafutok haza, és
hozok én egyet magának.
No letette a szolgáló a kast, és visszament. Addig a vándorló
legény a kast csak fëlvëllintëtte, és elment oda, a hol voltak szállva.
No immán a kenyér megvolt.
Most a második elment, a hol árulták a halakat. Volt egy
cseberben egy tíz darab pisztrang, egy öreg ember árulta egy
gyermekkel. Azt mondja az öreg embernek, hogy mit kér azért a tíz
darab halért. No hát a mit kért, megigérte, s hogy ő azt viszi a
főtisztelendő úrnak, és hogy a gyermek menjen vélle, és ott kifizetik.
Elvitték egy oan helyré, a hol egy ajtón bément, és ott letette a
halakat, aztán bément a tiszteletes úrhoz és azt mondta:
– Itt van egy gyermek, a ki meg van hibázva az eszében, és hogy
imádkozna réta, hogyha tudna segíteni réta. Itt hagynám a
gyermeket a főtisztelendő úrnál.

Hát ő ott ës hagyta a gyermeket, de a halakat a nagy pap nem
látta. Ő a halakkal elment, és a gyermek ott maradt. A gyermek
hallgatott, és a pap imádkozni kezdett. És egyszer azt mondja a
gyermek, hogy:
– No mán a halakért fizessen.
Hát a nagy pap aztán látta, hogy bolond, és elkergette a
gyermeket.
No immán, kenyér volt, hal volt, no immán a harmadik
szokotáljon bort. Elment a kórcsomába, és mind ott leskelődött a
pincze torkában. Éppen a szolgáló hozta ki a bort egy nagy korsóval.
Azt mondja a szolgálónak, adjon neki egy deczi pálinkát. A szolgáló
leteszi a korsót a borral, és elmënyën a pinczébe hátra, hogy hozzon
neki egy deczi pálinkát. No ő aval felvette a korsót a borral, és
elment a szállásukra.
No most ettek, ittak, minden volt elég. Immán azt számították,
hogy hogy mesterkedjenek ők, hogy pénzt tudjanak faszolni. Hát a
három vándorló legény közül egynek jó ismerős volt egy nagy
gazdag uraság. Azt mondja a vándorló legény:
– Menjünk el, kérjük el a kocsiját s a lovait, s a legszebb öltöző
ruháját.
Elmentek ők oda és elbeszéli ez a legény, hogy ők el akarnak
menni egy másik városba, hát legyen szíves adja oda a hintót a
lovakkal, s a legszebb öltöző ruháját csak huszonnégy órára. Adnak
száz forintot, de csak akkor fizetik meg a pénzt, mikor visszajőnek.
No odaadta, felültek ők hárman. Az öltöző ruha bé volt pakolva
egy kocsiládába. Elindultak, a mint mentek a mezőben, találkoztak
egy sátoros czigánynyal. Azt mondja egy a három közül a sátoros
czigánynak:
– Dobd el azt az üstöt a hátadról, jere ide, légy király, mert né
ebbe és ebbe a városba kell király, és most megyünk, hogy
keressünk.

No a czigány eldobja az üstöt, egyszeribe levetkeztetik, de
mossák igen-igen mindennel, de úgy ës feteke volt. Felöltöztetik
abba a szép rend öltöző ruhába, felültetik a hintóba, elmennek abba
a városba, a melyikbe volt szándékuk. Elmennek egy nagy
vendéglőbe, békéredznek, kérdik a vendégfogadóst, hogy lenne-é
egy tisztességes szoba a király számára, külön.
– No igen – azt mondja a vendégfogadós.
Bévezetik a czigánykirályt egy szép terembe, ëppeg esve volt
vacsorakor, egy hat órakor. No hát viszik a királynak az ételeket,
italokat, de ő nem tudott németül, és meg volt tanítva, hogyha
kérdik, hogy az étel hogy tetszik, mondja, hogy: ja ja. Csak annyit
mondjon.
No hát aztán a czigányt esve, mikor le kellett feküdni, a királyt
levetkeztetik abból a gunyából. Ott volt egy szép vetett ágy, a
czigányt mezíttelen oda lefektetik.
No hát reggel ezek felköltek hamarébb a vándorlegények, s azt
mondják a vendégfogadósnak, adjon nekik egy ezer forintot, hogy
ők menjenek ki a piaczra egyet-mást vásárolni, míg a király ő felsége
felkel, aztán ő megfogja fizetni. Egyszeribe kiolvassa a
vendégfogadós nekik az ezer forintot, felülnek a hintóba, mennek a
vásárba, de többet vissza se mentek.
Várja a vendégfogadós, hogy vissza menjenek, vagy a király ő
felsége felkeljen, de nem mentek vissza, a király se kőlt fel, mert
csórén volt az ágyban, hogy keljen fel? Egy tíz órakor nem győzték
várni, elmënyën egy inas, megnyitja az ajtót és visszahúzza magát.
Elmënyën a másik, az ës megnyitja az ajtót, hát akkor a király
felemelkedik mezittelen. Mondja a legény a vendégfogadósnak:
– Ni a király itt van, de csupa mezíttelen.
Oda mënyën a vendégfogadós, látja minyát, hogy meg van
csúfolva. Elvëszën egy kólbácsot, úgy elverte a czigányt, hogy azt a

szép házat mind összegazolta, a mint sokat ëtt volt az este a jó
ételekből. No hát ennyi volt a dolguk, az ezer forintot elvitték.
(Id. Fazekas Fóris Mihály, Pürkerecz.)

20. A három tanács.
Volt egyszer egy gazdag ember, és annak volt egy fia, de leánya
nem volt. És az anyának segítsége nem volt, egy szolgáló leányt
kellett tartani. És a leány, szegény, szép, de nagyon szép volt, és a
gazdag ember fia nagyon megszerette. A szolgáló leány
elterhesedett, és a gazdag ember nagyon megharagudott a fiára.
Úgy se engedi, a szegény szolgáló leányt hogy feleségül vegye. De a
legény azt mondta, néki az mindegy, mert ő a leányt szereti és
feleségül elveszi. Az édes apja azt mondta, van az udvar széjin egy
kicsi kunyhócska, aztot odaadja, hogy benne lakjanak, de egyebet
semmit nem fog adni.
No a legény búsult, mert nem volt miből élődjenek. Hiába az
apjának sok vagyona van, mert nékik nem ës volt mit egyenek.
Kiment egy écczoka az udvarra, felsóhajtott az úristenhez, hogy
vegye ki az ő feleségét az ő szüvéből, hogy ő menjen el vándorolni,
keressen, hogy legyen miből elődjenek.
El ës ment húsz esztendeig. A királyhoz talált menni, és ott
töltötte le a húsz évet szolgálatban. Mikor a húsz év letőlt, eszébe
jutott az ő kedves felesége. Mindennap szomoruan ült az udvaron,
nem volt semmi kedve. Mondták a többi szolgalegények a királynak,
hogy:
– A mű szolgatársunk harmadnapja, hogy mindig oan szomorú és
sír.
A király felhívta magához és felkérdezte, hogy mi oka lehet az ő
szomoruságának. És ő megmondotta, hogy húsz éve, hogy eljött az
ő feleségétől, és most eszébe jutott, és nem marad többet, hanem
hazamegy, hogy lássa, hogy az ő felesége él-é, vaj nem. A király

felhozatott egy kenyeret, azt kettőbe vágta, a belét kivájta és a
kenyérnek a héját megköttette (valószínűleg nyelvbotlás:
megtőttette helyett), egyet aranynyal, a másikat ezüsttel. S azt
mondta a szolgának, hogy:
– Menj az istállóba és hozz ki egy lovat. Nem ës a legjobbat, nem
ës a leghitványabbat.
És ő bément, kihozta a leghitványabbat. De a királynak nem
tetszett ez, és a király maga hozta a legjobbat ki az istállóból. A
kenyeret bétette az átólvetőbe, az egyiket az egyikbe, a másikat a
másikba. (T. i. az átólvetőnek, a kettős tarisznyának egy-egy felébe.)
Adott neki egy puskát és azt poroncsolta:
– No te hűséges szolgám voltál, fogadd el, ezzel én
mëgjándékozlak. És fogadd meg az én tanácsomat. Mikor mész,
menj az igaz úton, ne térj az ösvenútra. És a hol meghálsz, a hol
háboruság van az életben (a házban) az asszony az urával, ott meg
ne hálj, hanem keress a szomszédban alkalmatos szállást. Az estvéli
haragodat hadd el reggelre, és a reggelit estvére. És azzal
elbócsúzott.
A mint ment, találkozott két más utazóval. Mentek továrosságban
együtt. A két utazó ember eltért az ösvenútra, de néki eszébe jutott
a király tanácsa. Mikor mentek egy darabég, az ösvenúton, a két
utazó embert a tolvajok megfogták és meg akarták fosztani. Azok
kiábáltak, ennek híre lett, az erdőn keresztül vágott és rëok lőtt. Úgy
futottak el a tolvajok. Akkor ez a két utazó ígért kétszáz forintot az ő
megmentésükért. Ő megköszönte, de nem fogadta el.
Mentek aztán hárman együtt. Béesvéllëdëtt, mán meg kellett
szállaniok. Egy vendéglőbe mentek, ott a kórcsomáros veszekedett a
feleségével. A király szolgája nem hált meg ott és hítta ezeket ës.
Ezek azt mondták:
– Itt mán mű meghálunk.

A király szolgája a felső szomszédba ment szállásra. A
kórcsomárosnéhoz a pap járt. Nagyon örvendett a kórcsomárosné,
écczoka kimentek a pappal, elbeszélték együtt, hogy öljék meg a
kórcsomárost s tegyék a kést a két utazó átólvetőjébe s úgy azt
mondják, hogy ezek a két utazók ölték meg. De a király szolgája
mindent hallott, mert écczoka kiment az udvarra sétálni. A sziny
(kocsiszin) alatt volt ez a beszéd, s ő egy hasadékon keresztülnyúlt,
a papnak a rókját keresztülhúzta, s egy darabot kivágott. Azt
gondolta magában:
– Erre nékem szükségem lesz még.
Úgy ës lett. Reggel zaj kerekedett: a kórcsomárost megölték. Az
utazókat elfogták, hogy ők ölték meg a kórcsomárost. Elvitték a bíró
elejébe és felmondták az itéletet, hogy őket fel fogják akasztani. A
király szolgája ës felkérte, hogy őt ës megengedjék, hogy látására
odamenjen. Oda ës ment s felmondta a bírónak, a mit hallott. A bíró
odahivatta a kórcsomárosnét és a papot. A pap ëszre se vette, hogy
a rókjából ki van vágva. Mikor felhítták, ők tagadták, hogy nem igaz,
mert a két utazó ölte meg, annak találtatott az átólvetőjében a két
kés véresen. Akkor a király szolgája kivette a róknak a darabját, és a
bírósággal megnëzették, hogy éppen talál a papnak a rókjához. Aval
megszabadult onnét ës a két vándorló legény. A papot és a
kórcsomárosnét felakasztották, és ők hárman elutaztak együtt. Ott
ës igértek neki kétszáz forintot, de ő nem vette el töllik. A két utazó
megköszönte a szüvességét, és ő eltért más útra, hazafelé.
A mint hazaért, hát bé ës esvéllëdëtt. Elment haza a feleségéhez,
de nem ment bé, csak bénëzëtt az ablakon. A feleségének, terhes
volt, mikor elment, és két ikergyermeke született és mindakettő fiú.
A felesége kitaníttatta, egyiket deáknak (kántornak), a másikat
papnak. Mán másnap reggel esketni kérték fel a két ifjút. Az anyjuk
futott egyiknek, csókolta, futott a másiknak, azt ës csókolta. Ezt az
ura meglátta. Bé akart lőni az ablakon, de eszébe jutott a király
tanácsa, hogy az esvéli haragját hagyja el reggelre. S azzal
leeresztette a puskáját és elment a szomszédba. Ott kért szállást,
hogy kitudja, hogy az ő felesége micsoda fiatalokkal csókolódzik.

A szomszédban beszélték, hogy húsz éve, hogy az ő ura elment,
s a fejérnépnek ikergyermeke született. Isten tudja, hogy nevelte fel
őket apa nélkül. És most a szegény asszony azért csókolja, bólcsuzik
töllik, mert holnap esketik fel az egyiket papnak, a másikat deáknak.
Sajnálta, hogy meg kell válni töllik.
És kiváncsi volt ő ës elmenni az esketésükre. Reggel el ës mënën.
Mikor megeskették őket, hazamentek. Akkor ő ës odament hëzzik és
kérdette az asszonyt, hogy:
– Há magának hol a férje?
A nő elbeszéllëtte, hogy mióta elment a férje, semmit nem tud
felölle. A férje kérdezte:
– Megesmerné a férjét, ha meglátná?
Az asszony azt mondta:
– Meg, mert egy légy van a hátán.
Akkor az ember kirántotta az ingét a nyakába, és az asszony
mindjárt megüsmerte, hogy az az ura. Egymásra borultak, úgy sírtak
örömükben. De az egyik fél kenyeret adta a gyermekeinek, az
aranynyal, a másikkal ők élődtek. Nagy öröm lett az apjának, és
nékik ës egymással. És máig ës élnek, ha meg nem haltak.
(Fejér Györgyné, Lőrincz Kata, Pürkerecz.)

21. Palkó és a boszorkány.
Volt egy szegény embernek két fia. Azokat nevelte
szegénységgel, s mikor húsz évesek voltak, elmentek országot-
világot látni, s azt mondták az anyjuknak, hogy süssön nékik
pogácsát. Az anyjuk sütött pogácsát nékik, de mindkettőnek egy-egy
hármat. Bétarisznyálta s kikisérte a falu végéig, és sírva haza kellett
menni, elbólcsuzott töllik.
Mennek, mendegélnek, elérnek egy nagy erdőbe. Egyik fiút
Jankónak hívták, s a mást Palkónak. Mikor mennek egy kicsidég, azt
mondja Jankó Palkónak, hogy:
– Égeti valami az oldalomat, tovább nem tudom tűrni.
S azt mondja Palkó:
– Nézz hátra te.
Hátra nézett, egy nagy lompos farkas okádta a tüzet az oldalára.
S azt mondja Palkó Jankónak:
– Dobj bé egy fél pogácsát a szájába
És ő bédobott egy fél pogácsát, s a farkas úgy megpajtásult,
hogy ment utánuk örökké, nem hagyta el. Mikor mennek még egy
kicsidég, azt mondja:
– Jaj, hogy égeti most valami a bal oldalamat.
Azt mondja Palkó:
– Nézz hátra.

Hátra nézett, akkor egy nagy medve okádta a tüzet a bal
oldalára. Azt mondja:
– Dobj belé a szájába egy fél pogácsát.
Abba ës belédobott, s úgy megszelidült a medve, hogy mind
ment utánuk. Mikor még mennek egy kicsidég, azt mondja Jankó
Palkónak:
– Juj, most hogy égeti valami mind a két oldalamat.
S azt mondja Jankó Palkónak:
– Nézz hátra.
S akkor még egy nagy oroszlán hét singnyire okádta a tüzet mind
a két oldalára néki, Jankónak. S azt mondja:
– Dobj belé a szájába egy fél pogácsát.
Akkor ës dobott belé, és megszelidült, és mentek utána mind a
hárman. Mikor mentek még egy kicsidég, azt mondja Palkó:
– Juj, hogy égeti nékem ës valami az oldalamot.
És hátra nézett, és egy farkas okádta a tüzet az oldalára. S akkor
ës belédobott a szájába egy fél pogácsát és ment utána. S későbben
egy medve, s az után egy oroszlán. Mind a kettőnek bédobott a
szájukba egy fél pogácsát és mentek utánuk.
Most mindenik fiúnak volt három állatja. Az egyiknek is, s a
másiknak is. Mennek, mendegélnek ketten, az erdőben egy nagy
tisztát találtak s egy friss forrást. Tüzet csináltak, s ettek ott. Azt
mondja Palkó Jankónak:
– Itt van két keresztút, te menj jobbra, én mënëk balra.
Akármelyik visszatérjen, vaj egyik meg lesz halva, hát a forrásból vér
jőjön víz helyett. Arról tudjuk meg, hogy él-e vaj hal, mikor
visszajövünk.

Elindul Palkó a jobb úton, mënën, mendegél, elér egy országot.
Elér egy országot és látja, hogy gyászba van borítva, mindenütt
gyászlepedők vannak. Bémënën egy öreg asszonyhoz, hogy kérjen
vizet. Megkérdezte, hogy hát ez az ország mért van gyászba borítva.
S elmondta az öreg asszony, hogy az ország szélén egy nagy tó van,
a melyikben egy tizenkétfejű sárkány lakik. S minden harmadik nap
egy huszonegy éves leányt kell adni neki, addig nem ad vizet, s most
a király leányára került a sor, s azért van gyászban az ország. S
kihirdette a király, hogy a ki megváltja, a leányát a sárkánytól, hát
neki adja, s fele királyságát.
Palkó, hogy hallotta az öreg asszonytól, hát elment az országba,
és szerzett magának egy rozsdás kardot. S eljött a nap, a melyikben
a király leányát kivezessék a sárkányhoz. Akkor ő is elment a tóhoz,
három állatjával. S egyszer csak a víz kezdett a tóban hánykolódni,
hányta magát, a sárkány jött fel a tó fenekéről. S azt mondta:
– Te emberi lélek, te ingem akarsz legyőzni? Azt a királyleányt te
akarod elvenni?
Azt mondta Palkó:
– Én, ha az isten megsegít.
S kezdtek verekedni a sárkánynyal. A sárkány ës megfogta, s ők
ës négyen, de egyik se győzött. Addig verekedtek, míg elfáradtak, s
a sárkány lement s azt mondta:
– Várj egy kicsit s még újra küzsdjünk.
S mikor feljött, mégannyi ereje volt. S azt mondta a legény:
– Megállj! Ha tudtam volna, nem állottam volna meg, küzsdöttem
volna tovább.
Azt mondja akkor az állatjainak:
– Fogjátok meg állatjaim.

S azok megfogták, s felhúzta a kardját s egy vágásra mind a
tizenkét fejét levágta s a fejekből a nyelveket kivette s a táskába
bétette. Akkor a királyleány a vállára borult s örömében sírt. Azt
mondta:
– Te az enyém, én a tiéd, kasza-kapa válaszszon el egymástól.
Hazamennek, látja a király, hogy megváltotta a leányát, s
örvendett. Megmátkásodtak, lett a lakadalmuk másnap aztán. Az
asztalnál mindenki mondani kellett egy verset. Mikor erre a legényre
került a sor, hát a táskája ott volt vele, s azt mondja:
– Felséges uram, királyom, megszabadítottam a leányát a
tizenkétfejű sárkánytól, itt vannak a nyelvek.
S megmutasztotta a nyelveket. Mindenki örvendett, s a
lakodalom után éltek boldogul, s máig ës élnek, ha meg nem haltak.
Élnek egy darabig boldogul egymással. Azt mondja a legény:
– Vissza kellene menjek az öreg anyámhoz, a honnan eljöttem.
A király leányát ott hatta az apjánál s ő elindult hazafelé. S mikor
elért oda a tisztához, a hol tüzet csináltak volt, hát lájta, hogy a
forrás igen tiszta, hogy a bátyja még él. S ott futott el egy
nyulacska, s meglőtte s nyársra húzta s a tűz mellett sütötte. S a
mind ott sütötte, egy nagy fa volt ott a tűz mellett, s arra szállott
valami s kájtotta ott, hogy:
– Fázom!
Azt mondja Palkó:
– Gyere le, mert jó tűz van, ha tisztalelkű ember vagy.
S akkor azt mondja, egy boszorkány volt az öreg asszony:
– Me (nesze) ez a három hajszál. Tedd az állatjaidra, mert félek
töllik.
É

És Palkó elvette és a három állatjára tette. Akkor kőbálványnyá
váltak a három állat, s akkor nem félt a boszorkányné. Akkor
leszállott, kígyót, békát fogott egy nyársra s ott sütötte. S azt
mondta Palkónak:
– Te kígyót, békát eszel, de én a nyúlhúst.
Azt mondja Palkó:
– Adok én néked. Nem érsz meg aval, hogy ideeresztettelek a tűz
mellé, hogy melegedjél?
S ott kezdtek verekedni a nyárssal. De mivel az állatok nem
voltak meg, legyőzte a boszorkány Palkót, összevagdalta egy
tekenyőbe s feltette egy fának a tetejébe.
Hát egyszer visszatért a testvére is, Jankó. Látja, hogy a forráska
megállott a tűz mellett, látja, hogy a forráskából vér jő, nem víz. Azt
mondja:
– Az én bátyámat megölték valahol.
S ott futott el egy nyulacska, s meglőtte ő ës és nyársra húzta. S
akkor ës arra a nagy fára szállott a boszorkányné. Azt mondja:
– Huhú, hogy fázom!
Azt mondja ő ës:
– Gyere le, ha tisztalelkű ember vagy, mert jó tűz van, melegülj
meg.
S azt mondja a boszorkányné:
– Me ez a három hajszál, s tedd a három állatodra, mert félek
töllik.
Ő igaz, hogy elvette, de nem tette az állatokra, hanem a tűzbe
bévetette, s a boszorkányné nem vette számba, hogy ő a tűzbe
vetette. Leszállott, húzott egy nyárs kígyót, békát, s azt mondja néki:

– Te kígyót, békát eszel, de én nyúlhúst!
Azt mondja Jankó néki:
– Adok én neked! Nem elég, hogy idehíttalak, még nyúlhúst enni
akarsz? Csak hetvenkedj, ne hogy megverjelek.
S aval verekedni kezdtek a nyársakkal, s Jankó egyet füttyentett,
s a három állat eléjött s megfogta a boszorkánynét, s könyörgött a
boszorkányné, hogy hagyják meg az életét. Azt mondta:
– A testvéredet ës én pusztítottam el, s ha meghagyod az
életemet, akkor felélesztem.
Adott Jankónak egy pálczát. Azt mondta:
– Hágj fel a fa tetejébe s veregessed, ott van egy tekenyőbe,
mert feléled.
Azt mondja Jankó az állatoknak:
– Fogjátok meg jól.
S azok megfogták, mind a vasat, nem tudott mozdulni. S felment
a fára aval a pálczával s kezdte veregetni. Látja, hogy egyik
darabocska ës, meg a másik ës, mind úgy szërre (rendre) forrnak
össze, míg egy test lett belölle. Akkor felült s kezdett dermeszkedni
az a test. Mondja ő:
– Mióta aludtam!
Azt mondja Jankó:
– Aludnál bizony te bátya, ha én ne lettem volna, még máig
(mostanig) ës.
Kérdezte, hogy hát mért? És elbeszélte, hogy ő hogy járt. S akkor
az ő három állatja még átváltoztak, a hogy voltak, farkasé, medvéé s
oroszláné. Akkor megfogták nyolczan, hat állat s két fiú, s

ësszetépték, tűzön elégették s a hammuját a széllel elfújatták. Azért
nincsenek most boszorkányok.
S ők mennek, mendegélnek, elmennek haza az öreg anyjukhoz,
elbeszélik, hogy hogy jártak. S csudálkozott az öreg anyjuk. S azt ës
kocsiba tették s elvitték magukkal a királyhoz. Élnek öten boldogul,
ha meg nem hóltak.
(Korodi Sándor, Pürkerecz.)

22. A disznyópásztor leánya.
Volt egyszer egy szegény disznyópásztor, annak volt egy igen
szép leánya. S kihajtják a disznyókat ők a tollóra. Abban a városban
volt egy gazdag kereskedő ember, annak ës volt egy fia s aztán eljött
mán az üdő, hogy megházasodjék. Hát ő egy écczoka azt látta
álmotta, hogy menjen ki jó reggel s talál egy szegény leányt, a ki
vizet viszen a korsóval. Kérjen vizet tölle s azt vegye nőül.
Nohát ëppen annak a szegény disznyópásztornak a leánya ment s
vitte a vizet. Megszólítja a leányt, hogy adjon neki vizet, hogy igyék.
A leány szüvesen odaadta s mondja az úrfinak, ilyen jó reggel hogy
esik jól neki a víz? Mondja:
– Azon ne csodálkozzék, kedvesem, mert az éjjel sokat mulattunk
s jól esik.
És azonnal meg ës kérte őt feleségül.
No de ő azt mondta, hogy hogy ës kéri őt, mikor ő olyan szegény
leány, tám csúfolódik vélle. De a legény csak azt mondja neki:
– Menj el haza s űzzétek el a disznyókat apáddal s kéredzz el
tölle. Mondjad apádnak, hogy valami dolgot keressz a városban s
gyere az én ruhaboltomba.
A leány úgy is tett. Elkéredzett az apjától, s rögtön elment az
úrfinak a boltjához. A legény ës rögtön bészólítja s azt mondja:
– Gyere, válaszsz magadnak egy rend gunyát, a mi a testedhez
áll.
A leány választott ës magának tetőtől talpig egy rend szép
gunyát s felöltözött beléje. Oan gyönyörű szép küsasszony lett

belölle, a milyent soha életében se látott. Most azt mondja az ő
kedvese:
– Vesd le azt a gunyát, hogy pakoljam bé. És adok még száz
darab aranyat, vidd haza, és tedd el szépen a ládába úgy, hogy senki
semmit róla ne tudjon. És apádnak készíts jó gazdag vacsorát.
És mikor az öreg hazaért, megérzette az öreg a finom jó
vacsorát. Még künn mosolygott a jó büzütől hogy:
– Most mán kapok jó vacsorát!
Béér az öreg a házba és mondja leányának, hogy:
– Még soha oan jóillatú vacsorát nem éreztem a házamban!
De csak akkor örvendett az öreg igazán, mikor a disznyósültet
elejébe rakja az ő leánya, és egy kulacs bort. Azt mondja:
– Leányom, én téged még máskor ës eleresztlek a városba, hogy
keress dolgot magadnak.
– De azt jól teszi apám – mondja a leány – mert ëssze volt
beszélve az ifjúval, hogy ő még két reggel elmënyën a más
boltjához, mert három boltja volt annak.
Másnap, mikor kihajtják megint a disznyókat, mondja az öreg:
– Édes leányom, menj el, hogy este még készíthess olyan jó
vacsorát.
El is mënyën a leánya, nem várja, hogy kecczër (kétszer)
mondják. És ismét elment az úrfinak a más boltjához és az úrfi mán
az ajtóban várta.
– Ó be jó, hogy jösz, kedvesem. Jer be és válassz magadnak
még szebb öltöző gunyát, mind a más volt. Tiszta selyemből hogy
legyen.

Oant választott a leány, hogy még hétszerte inkább megszerette
az ő kedvese.
– No most mán vesd le, hogy pakoljam be és vidd haza. És adok
hozzá még kétszáz aranyat, tedd a ládádba. Ne mondj semmit az
öreg apádnak, hanem megint készíts neki jó vacsorát.
No most oan jó vacsorát készített az ő atyjának, hogy még
messze volt és örvendett már örvendett az ő leányának, hogy olyan
jó vacsorát készített. Bémënyën az öreg és kérdi az ő leányától:
– Kaptál-é fiam jó munkát? S erre a leány igennel felelt.
– No leányom – mondá – holnap reggel, ha kiüzzük a disznyókat,
csak még menj el.
El ës ment az ő leánya harmadszor ës az ifiúr harmadik
ruhaboltjához. Itt már nagyon várta az ő kedvese s mondá:
– Most válaszsz magadnak aranyszínű gúnyát, milyent még szem
nem látott.
Kapja magát a leány és oant választ magának, hogy mikor
felvette, úgy megszerette az ő kedvese, hogy ki se akarta ereszteni.
De mégës azt mondja:
– Vetkezz le, hogy pakoljam bé s mesze, mostan adok háromszáz
darab aranyat, vidd haza és tedd bé a mások mellé a ládába, mert
én utánad fogok jőni, hogy menjünk el a bálba.
Hazajött a szegény disznyópásztor, megérzi a sok jó büzüt, azt
mondja a leányának:
– Máma este még tán jobb vacsoránk lesz?
Történt abban a városban egy nagy bál. Meg volt híva a
kereskedőnek ës a fia és ő kocsit rendel, elküldi a szegény
disznyópásztornak a házához. Azt mondja:
– Öltözz fel abb a gunyába, a kit néked elsőben adtam.

Azzal felülnek a kocsiba és elmennek a bálba. Olyan szép
küsasszony volt, hogy mindenkinek szeme-szája tátva maradt.
Húzzák a mozsikások, és tánczba kerekednek. Kézről kézre veszik,
mind a mennyi úgy szereti, hogy mán a kereskedőnek a fia kiveszi
onnan és a kocsiba ülteti és haza mennek. Azt mondja neki, hogy ha
még majd bál fog történni, még kétszerig vele menjen el.
Úgy ës történt, még lőn egy nagy bál. Akkor ës kocsira ül és
elhajtatott a szegény disznyópásztorhoz. Azt mondja a leánynak:
– Most mán vedd fel, a kit másodszor adtam gunyát.
Felveszi a leány és oan szép küsasszony lett belölle, hogy oant
még életébe sohase látott. S most mán elhajtnak. Bémennek a bálba
s minden vendégnek a szeme csak a szép küsasszonyon volt. Most
mán esznek és vígan lakomáznak, míg a mozsikások húzni kezdik, és
mindenki a szép küsasszonynyal akarnak tánczolni.
– Megállj! – gondolja magában a kereskedőnek a fia – így nem jó
ám a dolog. Menjünk haza.
Megint kocsiba ülnek és haza mennek.
Most má megint következik a harmadik bál. Most mán azt
mondja:
– Vedd fel, a kit harmadszor adtam neked, azt az aranyszínű
gunyát.
Magára veszi, felöltözködik. Oan szép küsasszony lett belölle,
hétszerte szebb, mint máskor. Most mán esznek, isznak, vígan
vannak, de mindennek a szeme csak a szép küsasszonyon volt. Ott
volt a király fia ës, úgy megkedvelte, hogy azt mondja:
– Én ës elvenném magamnak nőül.
Megsejti a kereskedőfiú, azonnal kocsiba ülteti és hazamennek.
Azt mondja:
– Csak légy türelemmel, mert holnap meg foglak kérni apádtól.

Haza mënyën a kereskedőfiú ës, és az apjáék azt mondják neki:
– Jó volna fiam, ha megházasodnál. Néztünk ës ki leányt a te
számodra egy nagy herczegi családból.
A mire a legény bús és szomorú volt, s azt mondta:
– Én má néztem ki leányt az én számomra.
A mire kérdik az ő szülői:
– Kit néztél ki édes fiam?
– Éppen a disznyópásztornak a leányát.
– No azt mán nem hagyjuk, hogy megvedd (elvedd).
De a legény azt mondta, hogy neki nem kell más. Hanem úgy ës
beléegyeztek, mert nagyon szerették a fiút. Azt mondták, hogy a
leányt vigyék el egy más rokonukhoz, városba.
Másnap este két jóbarátja a legénynek és a legény beállnak a
disznyópásztorhoz, köszöntik és hogylétéről tudakoznak, a mire az
öreg nagyon megijedt, hogy még ilyen nagy úr nem jött bé a házba.
Mostan széket vesz elé és leülteti őket és kérdi a jó öreg, hogy
miben fáradnak. A mire a kereskedőfiú kijelenti áperté, hogy az ő
leányát kérné nőül magának. Mire az öreg azt felelé, hogy tán
csúfolódnak az ő leányával. Melyre a legény kijelenti komolyan, hogy
ő már érte jött volna. Az alatt a leány felöltözött abba a a szép
gunyába, a kit elsőben adott. Olyan szép küsasszony lett belölle,
hogy az öreg is egészen elbámult, hogy hát ezt a leány honnan
kapta.
– De sebaj, ha úgy áll a dolog, édes leányom, legyetek az
egymásé.
El ës vitték. Most mán kocsiba ültették a szép leányt és elvitték
egy atyjafiához, onnét vitték a kereskedői nagy kastélyba.

Tőlt-múlt az üdő s a gazdag kereskedő meghalt, csak a fiú
maradt. A fiára maradt minden gazdagság és kereskedés. Boldogul
ës éltek az ő szeretett kedvesével, hogy mindenki irigyelte az ő jó
életüket.
Egyszer útra kelt a gazdag kereskedő, hogy portékát vásároljon
az ő boltjába. Ez alatt az ő nőjére bízza a boltokat s elindul, hogy
vásároljon. Elbúcsuzott az ő nejétől, de mikor elindult, az ő nőjét
terhesen hagyta.
Ez alatt egy hunczut zsidó a kereskedőhöz ment és mondá:
– Mibe kötelezed magadat, ha én feleségeddel élek.
Azt mondá a kereskedő, hogy:
– A három boltom és a kastélyom mind a tied legyen és én neked
legaljasabb szolgád fogok lenni.
Má most meg volt kötve azzal, a mit a kereskedő mondott, és
elindult a sok szekérrel, hogy portékát vásároljon.
Ezalatt az ő nője fiat szült, és a zsidó szövetkezett a bábával,
hogy ő egy üres ládába béfekszik, a kin likak vadnak, és a bába kérje
meg az asszonyt, hogy azt a ládát vihesse bé, a hol ő hál. Ezennel a
bába kéri a kereskedőnét, hogy engedje meg, hogy a ládát vihesse
bé, mivel ott sok portékája van és az ura részeges, könnyen
kivehetne belőle valamit, hogy ide tenné be, a mit a kereskedőné
szívesen meg ës engedett, No most mán béviszik a ládát és oan
helyré viszik, a hol ő mindent láthat.
Estére levetkezik a kereskedőné meztelenre, s meglátja, hogy a
bal csöcsü alatt vagyon egy sümölcsü (szemölcs).
No, gondolá a zsidó: Jó, hogy ezt megláttam.
No most mán lefeküdt a kereskedőné és első álmába elaludott. S
a zsidó kijött a ládából, elvette a fehér zsebkendőt és a jegygyűrűt,
és szépen eltakarodott.

De azonban érkezik meg a hír, hogy jő a gazdag kereskedő a sok
portékával. És a zsidó elejébe mënyën, azt mondja neki, ismeri-é ezt
a zsebruhát és ezt a jegygyűrűt, a melyen a kereskedőnek a neve
volt.
Egészen megváltozott má most a képében és mondá:
– E még nem elég jegy.
Mire a zsidó azt mondta, hogy még a bal csöcse alatt van egy
sümölcsü.
– No má mostan meg vagyok verve – mondá a kereskedő.
Hazament és a feleségét egy ládába bézárta és a tengeren
leeresztette. No most a tenger vitte egy dararabig, míg kifogták a
halászok. Hagyjuk a kereskedőt, hogy szolgálja a zsidót.
No má most az asszonyt a tengerből kifogták a halászok, orvost
híttak és azonnal életre hozták. S kérdi a halásztól, hogy ő hol van,
mert mán egész más országban volt. No má most megköszönte az ő
jótevőinek az ő szabadítását és elindult, hogy menjen hazafelé.
De egy mezőn kellett keresztül menjen, a hol száz tolvaj volt. No
má most ő mit tegyen, mert elfogták a tolvajok. És a tolvajok
kapitányuk megszerette. Azt mondá:
– Asszony, te itt fogsz maradni véllem!
De megengedte, hogy künn az ajtó előtt sétáljon.
Mi jutott eszébe az asszonynak? Egy faoszlopot felöltöztet,
mintha ő volna, és megindul hazafelé. A tolvajok vígan esznek,
isznak, hogy az ő kapitányuknak olyan szép asszonyt tudtak
szokotálni. De az asszony már messze elhaladt egy erdőbe.
Csodák-csodája, hát mit lát álmában? Egy öreg ősz ember
odament és azt mondja neki:

– Kelj fel, nem messze innét e fa mellett van egy forrás, kinek oly
csodás ereje van, hogy mindenféle szembetegséget meggyógyít. Itt
van egy korsó és töltsd meg, vigyed magaddal, nemsokára nagy
szerencséd vár reád.
Úgy ës tett az asszony. Felkelt s úgy találta, a mint az öreg ember
mondta. Megtalálta a forrást és merített egy korsóval, és ment, míg
elért egy nagy városba, a melyben a király lakott. És ottan a
királynak a leánya nagy szemfájós betegségben volt. A király
kihirdette, hogy a ki leányának a szemét meggyógyítja, a leányát és
fele királyságát adja neki.
Meghallja ezt az asszony és vëszën magának orvosi gunyát,
jelenti a királynak, hogy ő volna egy orvos, a ki a leányának a
szemét megjavítaná (meggyógyítaná). Erre a király nagyon megörült
és jutalmul a fele királyságát és a leánya kezét igérte.
Az asszony most a leányt egy külön szobába vezette, háromszor
abból a vízből megmosdatta és mindjárt megjavult (meggyógyult) a
királyleány szeme. A király annyira megörvendett, hogy vivátot ës
lövetett, hogy éljen a jó király. De ő ezt el nem fogadhatta, hanem
azt mondá, hogy neki van felesége, adjon a király, a mit a szüve
kíván.
Most mán a király a legvitézebb huszáraiból négyszáz
huszárlegényt rendelt, s egy kocsit aranynyal. Úgy hajtottak most a
városba, a hol az ő urának a kastélya volt, a melyben egy zsidó
lakott most.
A zsidó elejükbe ment és nagy alázatossággal béereszti a
katonákat és az orvost vagy tisztet, mert olyan, mind egy kapitány,
olyan volt, bé az első szobába. És a hogy kitekint az ablakból,
meglátja az urát nagy bocskorosan és rongyosan, hogy hányja a
ganyét az istállóból, a lovak alól. Kérdi a zsidótól, hogy ki volna az az
ember?
A zsidó azt mondja:

– Az egy hányódó rossz szolga.
– De hogy is jutott hát a kastélyhoz? – kérdi.
A mire a zsidó mondja:
– Fogadásból nyertem.
No má most azt mondá az asszony:
– Hívja be azt a szolgát nekem.
A mire a zsidó nagyon meg volt ijedve, de mégës csak bé kellett
hogy hívja azt a szolgát.
Bémënyën a szolga és kérdi most a felesége, de ő nem ismerte:
– Jó ember, ez egykor, ez a kastély, úgy hallottam, a magáé volt.
Hát hogy került ennek a zsidónak a birtokába?
A melyre ő mondá:
– Fogadást tettem ezen kereskedő zsidóval, és azt mondta, hogy
ő a feleségemmel hál és én odaigértem a boltjaimat, a kastélyomat,
minden gazdagságomat, és én neki legaljasabb szolgája leszek. Úgy
volt a fogadásunk kikötve, s én egykor megérkeztem a portékáimmal
a város végére s elémbe jött és a zsebruhát és a jegygyűrűt, a kin a
nevem volt, megmutatta, és én mondám, hogy ez még nem elég, és
ő akkor mondá, hogy a bal csecse alatt még egy sümölcsü ës van.
Akkor mán egészen hittem neki.
Akkor a felesége kibongolja az ő lájbiját s mutatja mind a
kettőnek s mondja a zsidónak:
– Hált-é velem?
Jó két pofot adott a zsidónak, s az urának ës, mire a zsidó
egészen megvallotta, hogy vele nem hált, s hogy mesterségből tette.
S úgy megharagudott az asszony, hogy felakasztatta a zsidót. Ott

minden vagyonukat kezéhez vette s boldogan élnek ma ës, ha meg
nem haltak.
(Lukács István, Pürkerecz.)

23. Az ágy alá esett gyűrűs ujj.
Volt egyszer egy király a feleségével s egy leányával. Ez a
királyleány nagyon szelid volt. Csak az az egy volt, az anyja nagyon
őrzötte, hogy ne menjen akárhova. Egykor odament két királyleány s
hivták, hogy menjen el véllik sétálni. A királyné gondolkozott, hogy
vajjon elereszsze, vagy ne ereszsze. Az a más két királyleány erősen
biztatták az anyját, hogy jó helyré mennek s visszajőnek, de
elvezették messze egy erdőbe. A mint mennek, mendegélnek,
elfáradott a királyleány. Leültek, az álom megütötte s elaludt. Az a
más kettő irigyek voltak rëa; ott hagyták aluva s eljöttek.
Mikor megébredett, hát már este felé volt. Búsult, hogy má most
ő hazamenne. Megindult s ott az erdőben látott ő egy szép házat, s
bátorkodott, s elment ahhoz a házhoz. Ott volt egy öreg szakácsné.
Mondta az az öreg szakácsné:
– Jaj fiam, hogy bátorkodtál ide jőni, mert itt tizenkét tolvaj lakik.
Akkor tanította az az öreg szakácsné, hogy menjen bé a közepső
szobába, bújjék bé az ágy alá. S ő elment, mert bátor volt. Bénézett
az első szobába ës, hogy ott mi van. Ott látott mindenféle szép
küsasszonyi gunyát, királyi gunyát, mindent. Ott volt egy
oldalszobácska ës, bénézett oda ës, hát ott látott mind
küsasszonyfejet és emberfejet. Akkor megirtózott, kijött abból a
szobából s az ágy alá bébújt. Ő gondolta magában: Isten velem!
Akkor jöttek három királyfiforma, de csak azok a tolvajok voltak.
Vittek egy küsasszonyt, aval úgy bántak el, hogy ő egészen
megijedett. Levágták a fejét, de elsőbe levették róla a szép öltözetet.
Volt egy szép arany gyűrű, az az ujjában volt s a neve rëa volt
vágva. Nem tudták kihuzni azt a gyűrűt az ujjából s úgy levágták,

Welcome to our website – the perfect destination for book lovers and
knowledge seekers. We believe that every book holds a new world,
offering opportunities for learning, discovery, and personal growth.
That’s why we are dedicated to bringing you a diverse collection of
books, ranging from classic literature and specialized publications to
self-development guides and children's books.
More than just a book-buying platform, we strive to be a bridge
connecting you with timeless cultural and intellectual values. With an
elegant, user-friendly interface and a smart search system, you can
quickly find the books that best suit your interests. Additionally,
our special promotions and home delivery services help you save time
and fully enjoy the joy of reading.
Join us on a journey of knowledge exploration, passion nurturing, and
personal growth every day!
ebookbell.com