Tesis Upm - Juan - Corcoles - Ortega

UNIVERSIDAD POLITÉCNICA DE MADRID
ESCUELA TÉCNICA SUPERIOR DE INGENIEROS DE TELECOMUNICACIÓN
ADVANCED TECHNIQUES OF OPTIMIZATION AND

ELECTROMAGNETIC DESIGN BASED ON A HYBRID
FINITE ELEMENT – MODAL ANALYSIS METHOD
FOR ANTENNAS AND ANTENNA ARRAYS
TÉCNICAS AVANZADAS DE OPTIMIZACIÓN Y DISEÑO

ELECTROMAGNÉTICO BASADO EN UN MÉTODO
HÍBRIDO ELEMENTOS FINITOS – ANÁLISIS MODAL
PARA ANTENAS Y AGRUPACIONES DE ANTENAS
TESIS DOCTORAL
Juan Córcoles Ortega

Ingeniero de Telecomunicación
Madrid, 2009
DEPARTAMENTO DE
ELECTROMAGNETISMO Y TEORÍA DE CIRCUITOS
ESCUELA TÉCNICA SUPERIOR DE INGENIEROS DE TELECOMUNICACIÓN
ADVANCED TECHNIQUES OF OPTIMIZATION AND

ELECTROMAGNETIC DESIGN BASED ON A HYBRID
FINITE ELEMENT – MODAL ANALYSIS METHOD
FOR ANTENNAS AND ANTENNA ARRAYS
TÉCNICAS AVANZADAS DE OPTIMIZACIÓN Y DISEÑO

ELECTROMAGNÉTICO BASADO EN UN MÉTODO
HÍBRIDO ELEMENTOS FINITOS – ANÁLISIS MODAL
PARA ANTENAS Y AGRUPACIONES DE ANTENAS
Autor:
Juan Córcoles Ortega
Director:
Miguel Ángel González de Aza
Doctor Ingeniero de Telecomunicación
Profesor Titular de Universidad
Madrid, 2009
TESIS DOCTORAL: Advanced techniques of optimization and electromagnetic de-
sign based on a hybrid Finite Element – Modal Analysis
method for antennas and antenna arrays.
(Técnicas avanzadas de optimización y diseño electro-
magnético basado en un método hı́brido Elementos Finitos
– Análisis Modal para antenas y agrupaciones de antenas).
AUTOR: Juan Córcoles Ortega

DIRECTOR: Miguel Ángel González de Aza

Doctor Ingeniero de Telecomunicación
Profesor Titular de Universidad
DEPARTAMENTO: Electromagnetismo y Teorı́a de Circuitos

Universidad Politécnica de Madrid
El Tribunal de Calificación, compuesto por:
PRESIDENTE:
Prof.
VOCALES:
Prof.
Prof.
Prof.
VOCAL SECRETARIO:
Prof.
VOCALES SUPLENTES:
Prof.
Prof.
Acuerda otorgarle la CALIFICACIÓN de:
Madrid, a de de 2009.
A mis padres Juan e Isabel.
A mi hermana Isabel.
A mis abuelos Fidel y Simón.
A mis abuelas Juana y Ascensión.
A mis tı́as Mari y Reme.
El tiempo es ese juez insobornable que da y
quita razones y pone a cada uno en su sitio.
Time is that incorruptible judge that proves people
right and wrong and puts everybody in his place.
For those who never try. . .

they are sure to fail.
Porque aquéllos que no lo intentan. . .
seguro que no lo consiguen.
Agradecimientos
Acknowledgements
El primer agradecimiento es, sin duda, para mi director de tesis, Miguel Ángel. Y no sólo
porque jamás he recibido por su parte un “estoy ocupado”, o un “ahora no puedo atender
esto”, o ni tan siquiera un “espérate un minuto”. Todo lo contrario. Me consta que
siempre que ha podido ha antepuesto las cuestiones relativas a mi tesis, tanto cientı́ficas
como administrativas, a muchas de sus tareas como Profesor Titular. Además, siempre
ha sabido dar valor a cualquier idea que haya surgido sin descartarla, por muy ilusa
que pudiera parecer, y otorgado la libertad para desarrollarla. Sin duda, su dedicación,
accesibilidad y flexibilidad han hecho que la realización de esta tesis me fuera más llevadera.
Pero por encima de todo, lo que ha hecho que esta tesis llegue a buen puerto ha sido su
confianza en mı́.
También me gustarı́a agradecer a Mari Carmen todo lo que me ha ayudado. Su
eficiencia en las tareas administrativas es encomiable. Pero hay algo mucho más importante
que aunque no forme parte de su trabajo ha sabido hacer: decirme esas palabras de ánimo
y apoyo en el momento en que justo más las necesitaba.
Darle las gracias a Juan Zapata por darme la oportunidad de entrar a formar parte
de un grupo cientı́ficamente tan reconocido, aún cuando ni tan siquiera me conocı́a, y por
ser la voz de la experiencia a través de sus valiosos consejos. Y, por supuesto, a Jesús
Rubio, por todas las horas que nos hemos pasado al teléfono hablando y debatiendo sobre
los dichosos modos esféricos. Su ayuda ha servido, sin duda, para salir de varios atolladeros
que ha deparado la investigación.
Como no, un muy especial agradecimiento a mis compañeros de doctorado que han
recorrido o están recorriendo el mismo camino, desde los primeros que conocı́ hasta los
más nuevos que han llegado por el Departamento últimamente. Pero en particular, para
Javi, Edu, Manolo y Jose, con los cuales he compartido la mayorı́a de estos 4 años, y que
han hecho que las veces que pasaba por allı́ tuviesen un corte agradable.
Finalmente, mi más sincero agradecimiento al resto de gente del Departamento.
Profesores como Joserra, Encinar, Rebo, Chemari, Mariano, Page, Manu, Jaime, Alejan-
dro o Jesús, técnicos como Fernando o Teo, además de toda la gente que he nombrado
anteriormente, hacen que se me antoje difı́cil pensar en un grupo de personas que aúnen
a la par no sólo la calidad cientı́fica sino la calidad humana del Departamento.
Esta tesis también ha sido posible gracias a la disposición del profesor Thomas Zwick,
de Alemania, quien me dio la oportunidad de realizar una estancia de investigación en su
instituto. Me gustarı́a agradecer a toda la gente de allı́ la hospitalidad que demostraron,
ya que hicieron que unos cuantos meses en un paı́s del que ni tan siquiera conocı́a el idioma
pasaran más rápido de lo que inicialmente imaginé. Además del lobby polaco, no puedo
dejar de agradecer a Juan Pontes y a su esposa su dedicación desde el primer momento
v
vi
en que pisé suelo alemán. No sólo gané un colaborador cientı́fico, sino un gran amigo.
En el plano más personal, por muy tı́pico que pueda sonar, no habrı́a llegado hasta
aquı́ sin mis padres y sin mi hermana. Por, como ya dije en su momento, haber estado
siempre ahı́, para lo bueno y para lo malo, por ayudarme, por apoyarme, por aguantarme,
por enseñarme cosas que jamás se aprenderán en ningún libro ni en ninguna escuela, y
por cien millones de motivos más. Supongo que nunca seré consciente de lo afortunado
que soy por tener la certeza de que si todo va mal, o incluso si las cosas sólo se tuercen un
poco, siempre puedo llamar a mis padres o ir a casa.
Asimismo, me gustarı́a recordar por todo lo que me han aportado a mis abuelos,
mis yayos, aunque uno se fue hace ya tiempo y los otros tres no sean conscientes de lo que
académicamente supone un Doctorado. Sin duda sé que, al igual que mis tı́as, se alegran
más por mı́ que si hubieran sido ellos mismos.
No puedo olvidar aquı́ a todos los amigos, compañeros y trabajadores del C.M.U.
San Juan Evangelista que he acumulado durante más de un tercio de mi vida, y muy en
especial a Juan Luis Cámara, de quien nunca me ha faltado el ánimo y el apoyo en esta y
otras muchas andaduras que he emprendido. Todos ellos hicieron y siguen haciendo que
el Johnny, que tanto me ha dado y tanto me ha quitado (aunque últimamente me quedo
con lo que me ha dado), sea como siempre he dicho mi segunda casa, y ellos mi segunda
familia.
Y finalmente, a los de siempre, los amigos de Albacete, que siempre anduvieron por
ahı́ para hacerme olvidar los quebraderos de cabeza que te da una tesis y otras cosas, y
que, como ya dije en su momento, me han hecho comprender el verdadero sentido de la
amistad en tantos años fuera de mi ciudad natal.
First of all, I would like to thank my Ph. D. advisor, Prof. M. A. González, for
his valuable guidance and dedication, but, above all, for his trust in me and my work. I
must admit my admiration for Prof. J. Zapata, who gave me the opportunity to carry
out this Ph. D. even when he did not know me personally, and whose experience has
always been present through his advice. Discussions with Prof. J. Rubio always led to
solutions for the emerging problems and usually resulted in useful and interesting research
lines. I would also like to thank everybody in the Departamento de Electromagnetismo y
Teorı́a de Circuitos, since I have shared these four years of work with all of them. Part
of this Ph. D. has been carried out in Germany, thanks to Prof. T. Zwick, who gave
me the opportunity to spend a research period at his institution. Once there, scientific
collaboration with Eng. J. Pontes turned out to be very fruitful. Last, but not least,
a final word of great gratitude to those who have always been in my way through the
good times and the bad times, for their encouragement and their support: parents, sister,
grandparents, aunts and friends.
Abstract
The main objective of this thesis is the development of, on the one hand, efficient analysis
methods that allow a fast and rigorous electromagnetic characterization of radiating struc-
tures, and, on the other hand, efficient design techniques that are capable of inherently
taking into account all electromagnetic effects and interactions in antennas and antenna
arrays throughout the whole design process.
To that effect, a novel computer-aided design methodology for probe-fed, cavity-
backed, stacked microstrip patch antennas is proposed. This configuration is widely used
to increase bandwidth. The methodology, which makes use of neural networks and global
optimization algorithms, incorporates the rigour of a numerical technique, such as finite
element methods. It is validated through a real design.
Moving on to antenna array theory, the generalized-scattering-matrix (GSM) anal-
ysis of finite arrays whose elements’ radiated fields can be individually described in terms
of spherical modes, such as arbitrarily shaped apertures, horns, monopoles, dielectric res-
onator antennas or cavity-backed patch antennas, is presented. It is based on rotation and
translation properties of spherical waves, which can be expressed analytically. Numerical
validation is accomplished through the comparison with measured results.
As for array design techniques, the multiobjective optimization of arbitrary pla-
nar array excitations is presented. The problem formulation is based on matrix-valued
functions computed from the GSM of an array and spherical mode expansions. It al-
lows the maximization of the directive gain subject to a maximum sidelobe level (SLL), a
maximum crosspolar level and a minimum aperture illumination efficiency, the setting of
null-pointing directions and the dynamic range control. To solve the resulting non-linear
optimization problem, a primal-dual interior point filter method specifically adapted to
the formulation is developed.
Going further into array design based on optimization theory, in this case linear
programming, another pattern synthesis technique is introduced. The procedure yields
the complex-valued excitations to achieve the minimum maximum SLL given a specified
pointing direction and mainlobe width and prescribed field nulls. Again, all interelement
coupling effects are inherently taken into account. The procedure is extended to synthesize
patterns with various main beam pointing directions with different widths and shaped-
beam patterns with controllable SLL.
This thesis also presents a flexible method, based on quadratic programming, to
compensate the interelement mutual coupling effects that may degrade the field pattern.
A closed expression for a mutual coupling compensation matrix is derived from the GSM
of an antenna array and spherical mode expansions. This matrix is used to compensate the
value of any excitations obtained with an isotropic-based synthesis. Conditions regarding
null field pattern directions can also be reinforced in this matrix.
Leaving optimization apart, a simple pattern synthesis for linear equispaced arrays
vii
viii
is introduced. It can be viewed as an extension of the classic design techniques of the

Fourier Series (FS) or the Woodward-Lawson frequency sampling method to the case in
which the individual antenna elements’ patterns and all interelement couplings are taken
into account. The design procedure is based on either the FS or the Discrete Fourier
Transform (DFT) of the spherical mode expansion of the array radiated field, as well as
on various properties associated to the FS or DFT coefficients.
As another non-optimization based design technique, a shaped-beam pattern syn-
thesis based exclusively on the handling of spherical waves is presented. Rotation and
translational addition theorems are used to establish the necessary relationships between
different spherical mode expansions corresponding to the radiated field by each antenna
and by the whole array. Therefore, no involved algorithm is required to yield the complex
feeding excitations.
Regarding array thinning, an original procedure is introduced. The resulting thinned
array is the one which, having the minimum number of active elements, achieves several
prescribed design specifications. The proposed method yields both the active elements
and their excitations. Specifications such as achieving a minimum gain, obtaining a pat-
tern with a maximum allowable SLL or synthesizing a shaped beam pattern confined into
a mask are considered. Null field directions can also be added. In order to carry out
the thinning, a genetic algorithm is used, while computing the excitations is carried out
through linear or quadratic programming. Since the presence of an element can sub-
stantially alter the array features because of mutual coupling, two types of thinning are
considered: removing elements or turning them off.
Finally, two models to characterize multiple-input multiple-output (MIMO) systems
are presented. Line-of-sight MIMO systems are analyzed through the GSM of each an-
tenna and rotation and translation coefficients of spherical modes. The resulting channel
matrix rigorously includes the exact spherical vector nature of electromagnetic propaga-
tion (which may exert a significant influence over short-range links) and mutual coupling
effects. A modal network model for MIMO systems is also introduced and applied to an
urban scenario. The model, whose core consists of a ray-tracing tool, is based on the
modal expansion of the field pattern of both receiving and transmitting antennas and al-
lows a significant simulation time reduction to yield statistical results from deterministic
propagation models.
Resumen
El principal objetivo de esta tesis es el desarrollo de, por una parte, métodos de análisis
eficientes que permitan una caracterización electromagnética de manera rápida y rigurosa
de estructuras radiantes y, por otra parte, técnicas de diseño eficientes que sean capaces
de tener inherentemente en cuenta todos los efectos e interacciones electromagnéticas en
antenas y agrupaciones de antenas (arrays) a lo largo de todo el proceso de diseño.
A tal efecto, se propone una novedosa metodologı́a de diseño asistido por ordenador
para antenas microstrip de parches apilados alimentados por sonda coaxial. Esta confi-
guración es ampliamente utilizada para incrementar el ancho de banda. La metodologı́a,
que hace uso de redes neuronales y algoritmos de optimización global, incorpora el rigor
de una técnica numérica como el método de elementos finitos. Queda validada a través de
un diseño real.
Pasando al campo de la teorı́a de arrays, se lleva a cabo el análisis modal a través de
la Matriz de Dispersión Generalizada (MDG) de arrays finitos con elementos cuyo campo
radiado se puede expresar individualmente en términos de modos esféricos, como aperturas
con forma arbitraria, bocinas, monopolos, resonadores dieléctricos o parches impresos
situados en cavidad. El método se basa en las propiedades de rotación y traslación de
ondas esféricas, que pueden ser expresadas de manera analı́tica. La validación numérica
se realiza a través de la comparación de resultados medidos.
En lo que refiere a técnicas de diseño de arrays, se presenta una optimización multi-
objetivo de arrays planos con geometrı́a arbitraria. La formulación del problema se realiza
a través de funciones matriciales calculadas a partir de la MDG de un array y expan-
siones en modos esféricos. Permite la maximización de la directividad sujeta a un máximo
Nivel de Lóbulos Secundarios (NLS), un máximo nivel de componente contrapolar y una
eficiencia mı́nima de apertura, ası́ como definir direcciones de campo nulo y el control
del margen dinámico. Para resolver el problema de optimización no lineal resultante, se
desarrolla un método de punto interior primal-dual con filtrado especı́ficamente adaptado
a la formulación.
Profundizando en el diseño de arrays basado en la teorı́a de optimización, en este
caso de programación lineal, se desarrolla otra técnica de sı́ntesis de diagramas. El pro-
cedimiento proporciona las excitaciones para alcanzar el mı́nimo NLS máximo dada una
dirección de apuntamiento con una anchura de haz y direcciones de campo nulos. De
nuevo, todos los efectos de los acoplos se consideran inherentemente. Esta técnica se
amplı́a para sintetizar diagramas con varias direcciones de apuntamientos con diferentes
anchos de haz y diagramas de haz conformado con NLS controlable.
Esta tesis también presenta un método flexible, basado en programación cuadrática,
para compensar los efectos de los acoplos que pueden degradar el diagrama de campo. Se
deriva una expresión compacta de una matriz de compensación de acoplos a partir de la
MDG de un array y expansiones de modos esféricos. Esta matriz se usa para compensar
ix
x
el valor de las excitaciones obtenidas con una sı́ntesis que asuma fuentes isotrópicas. En
esta matriz se pueden incorporar condiciones para definir direcciones de campo nulo.
Dejando a un lado la optimización, se presenta una técnica de sı́ntesis simple para
arrays lineales equiespaciados. Se puede ver como una extensión de las técnicas clásicas
de las Series de Fourier (SF) o el método de Woodward-Lawson, al caso en el que se con-
sideran los diagramas individuales de cada elemento y los acoplos entre los mismos. El
procedimiento se basa en la SF o la Transformada Discreta de Fourier (TDF) de la ex-
pansión en modos esféricos del campo radiado por el array, ası́ como en varias propiedades
asociadas a los coeficientes de la SF o la TDF.
Como otra técnica de diseño no basada en optimización, se presenta una sı́ntesis
exclusivamente basada en la manipulación de ondas esféricas. Los teoremas de rotación
y traslación se usan para establecer las relaciones necesarias entre distintas expansiones
de modos esféricos que se corresponden con el campo radiado por cada antena y por el
array. De este modo, no se requiere ningún algoritmo complejo para calcular el valor de
las excitaciones.
En lo que refiere al thinning de arrays, se presenta un procedimiento original. El
array resultante es aquél que, con el mı́nimo número de elementos activos, cumple con
varias especificaciones de diseño. El método propuesto proporciona tanto los elementos
activos como sus excitaciones. Se consideran especificaciones como alcanzar una directivi-
dad mı́nima, obtener un diagrama con un NLS máximo o sintetizar un diagrama de haz
conformado dentro de una máscara. Se pueden añadir también direcciones de campo nulo.
Para llevar a cabo el thinning, se usa un algoritmo genético, mientras que el cálculo de las
excitaciones se realiza a través de programación lineal o cuadrática. Dado que la presen-
cia de un elemento puede alterar las caracterı́sticas del array debido al acoplo mutuo, se
consideran dos tipos de thinning: retirar elementos o apagarlos.
Finalmente, se presentan dos modelos para caracterizar sistemas MIMO con antenas.
Los sistemas con visión directa (sin obstáculos) entre antenas se caracterizan a través de
la MDG de cada antena y coeficientes de rotación y traslación de modos esféricos. La
matriz de canal resultante incluye rigurosamente la naturaleza vectorial esférica de la
propagación electromagnética (que puede ejercer una influencia significativa en enlaces de
corto alcance) y los efectos de los acoplos. Se desarrolla también un modelo de red modal
para sistemas MIMO que se aplica a un escenario urbano. El modelo, cuyo núcleo consiste
en una herramienta de trazado de rayos, se basa en la expansión modal del diagrama de
radiación tanto de las antenas transmisoras como de las receptoras y permite una reducción
significativa del tiempo de simulación para proporcionar resultados estadı́sticos de modelos
de propagación deterministas.
Contents
Agradecimientos/Acknowledgements v
Abstract vii
Resumen ix
Contents xi
List of Figures xv
List of Tables xxi
Glossary of symbols xxiii
Glossary of acronyms xxv
1 Introduction 1
1.1 Motivation and objective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Text organization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
2 Design methodology for stacked microstrip patch antennas 3

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.2 Hybrid Finite Element - Modal Analysis characterization of an antenna . . 4
2.2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.2.2 Finite Element formulation . . . . . . . . . . . . . . . . . . . . . . . 6
2.2.3 Multimode multiport network characterization of an antenna . . . . 7
2.2.4 Example: Design of a broadband U-slot microstrip patch antenna . . 8
2.3 Inclusion of artificial ports: multipurpose admittance matrices . . . . . . . . 10
2.4 Neural network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.4.1 Normalizing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.4.2 Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.5 Global optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.5.1 Genetic algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.5.2 Simulated annealing . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.6 Application: broadband stacked circular patches for DCS and UMTS bands 20
2.6.1 Finite element simulations . . . . . . . . . . . . . . . . . . . . . . . . 21
2.6.2 Neural network training . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.6.3 Optimization phase . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.6.4 Final design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
xi
xii CONTENTS
2.6.5 Other designs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
3 Generalized-Scattering-Matrix analysis of antenna arrays 29

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.2 GSM analysis of an array from rotation and translation of spherical waves . 30
3.3 Multimode Multiport network characterization of an antenna array . . . . . 33
3.4 Numerical validation and examples . . . . . . . . . . . . . . . . . . . . . . . 35
3.4.1 Comparison with measured results . . . . . . . . . . . . . . . . . . . 35
3.4.2 Analysis of an array scanning performance . . . . . . . . . . . . . . 36
3.5 Inclusion of the feeding network effects: waveguide-fed DRA array . . . . . 42
4 Multiobjective optimization of an array through an IPM with filtering 47

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
4.2 Non-linear constrained optimization problem formulation . . . . . . . . . . 48
4.2.1 Matrix formulation of the power and the radiated field . . . . . . . . 48
4.2.2 Direct maximization of the gain . . . . . . . . . . . . . . . . . . . . . 49
4.2.3 Inclusion of constraints . . . . . . . . . . . . . . . . . . . . . . . . . 51
4.2.4 Equivalent real-valued constrained optimization problem . . . . . . . 52
4.3 Primal-dual Interior Point Method with Filtering . . . . . . . . . . . . . . . 54
4.3.1 The logarithmic barrier method and the primal-dual approach . . . 54
4.3.2 Modified Newton’s step: Hessian, Lagrangian and gradient . . . . . 54
4.3.3 Filtering with backtracking . . . . . . . . . . . . . . . . . . . . . . . 55
4.4 Numerical Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
4.4.1 Computational aspects . . . . . . . . . . . . . . . . . . . . . . . . . . 56
4.4.2 Optimization of linear arrays . . . . . . . . . . . . . . . . . . . . . . 57
4.4.3 Optimization of circular arrays . . . . . . . . . . . . . . . . . . . . . 63
4.4.4 Optimization of planar arrays . . . . . . . . . . . . . . . . . . . . . . 65
5 Linear and Quadratic Programming for array pattern synthesis 69

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
5.2 LP-based techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
5.2.1 Synthesis of minimum sidelobe level patterns with prescribed nulls . 69
5.2.1.1 Minimax problem . . . . . . . . . . . . . . . . . . . . . . . 69
5.2.1.2 Real-valued equivalent LP problem . . . . . . . . . . . . . 70
5.2.2 Multibeam pattern synthesis with minimum sidelobe level . . . . . . 73
5.2.3 Shaped-beam pattern synthesis with controllable sidelobe level, crosspo-
lar level and nulls . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
5.3 QP-based methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
5.3.1 Mutual coupling compensation . . . . . . . . . . . . . . . . . . . . . 77
5.3.2 Null-constrained mutual coupling compensation . . . . . . . . . . . . 81
6 Non-optimization based field synthesis techniques 85

6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
6.2 Fourier synthesis of linear arrays . . . . . . . . . . . . . . . . . . . . . . . . 85
6.2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
6.2.2 Linear equispaced array characterization . . . . . . . . . . . . . . . . 86
6.2.3 Inclusion of the isolated element pattern . . . . . . . . . . . . . . . . 87
6.2.4 Fourier Series analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 87
6.2.5 Discrete Fourier Transform analysis . . . . . . . . . . . . . . . . . . 88
CONTENTS xiii
6.2.6 Fourier-based design procedure . . . . . . . . . . . . . . . . . . . . . 89

6.2.7 Numerical examples . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
6.3 Spherical-wave based shaped-beam synthesis . . . . . . . . . . . . . . . . . . 95
6.3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
6.3.2 Spherical wave expansion of the shaped-beam field . . . . . . . . . . 96
6.3.3 Relationship between spherical modes and feeding . . . . . . . . . . 97
6.3.4 Numerical results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
7 Array thinning with a nested optimization scheme 107

7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
7.2 Thinned arrays using the GSM analysis . . . . . . . . . . . . . . . . . . . . 108
7.3 Nested optimization scheme for thinning . . . . . . . . . . . . . . . . . . . . 109
7.4 Fitness function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
7.4.1 Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
7.4.2 Design for a minimum gain . . . . . . . . . . . . . . . . . . . . . . . 112
7.4.3 Design for a maximum SLL . . . . . . . . . . . . . . . . . . . . . . . 113
7.4.4 Design for a shaped beam pattern . . . . . . . . . . . . . . . . . . . 114
7.4.5 Flowchart of the fitness function . . . . . . . . . . . . . . . . . . . . 115
7.5 Numerical examples and results . . . . . . . . . . . . . . . . . . . . . . . . . 115
7.5.1 Computational aspects . . . . . . . . . . . . . . . . . . . . . . . . . . 115
7.5.2 Isotropic array thinning . . . . . . . . . . . . . . . . . . . . . . . . . 117
7.5.3 Linear array thinning . . . . . . . . . . . . . . . . . . . . . . . . . . 119
7.5.4 Planar array thinning . . . . . . . . . . . . . . . . . . . . . . . . . . 124
8 Modelling MIMO systems with a hybrid spherical-mode-based approach127

8.1 Line-of-sight SISO and MIMO systems . . . . . . . . . . . . . . . . . . . . . 127
8.1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
8.1.2 Single-Input-Single-Output system . . . . . . . . . . . . . . . . . . . 127
8.1.3 Multiple-Input-Multiple-Output system . . . . . . . . . . . . . . . . 128
8.2 Modal network model for MIMO systems . . . . . . . . . . . . . . . . . . . 130
8.2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
8.2.2 Generalized network model . . . . . . . . . . . . . . . . . . . . . . . 131
8.2.3 Modal network model . . . . . . . . . . . . . . . . . . . . . . . . . . 134
8.2.4 Characterization of the mobile station . . . . . . . . . . . . . . . . . 135
8.2.5 Base station characterization . . . . . . . . . . . . . . . . . . . . . . 136
9 Conclusion 143
9.1 Original contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
9.2 Future research lines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
9.3 Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
9.4 Publications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
9.4.1 Journal articles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
9.4.2 Conference proceedings . . . . . . . . . . . . . . . . . . . . . . . . . 147
9.4.2.1 International . . . . . . . . . . . . . . . . . . . . . . . . . . 147
9.4.2.2 National . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
Bibliography 149
xiv CONTENTS
List of Figures
2.1 Canonical geometry of the antenna to be designed. . . . . . . . . . . . . . . 4

2.2 GSM of an antenna i . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.3 Top and side view of the geometry of the probe-fed U-slot rectangular patch
embedded in a cylindrical cavity. Coaxial feed (SMA connector): εr =1.95,
do =0.41 cm, di =0.128 cm. Antenna parameters from (Tong et al., 2000):
εr =2.33, a=1.4 cm, b=1.8 cm, c=0.2 cm, g=0.7 cm, L=3.6 cm, W=2.6 cm,
h=0.8 cm. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.4 Port diagram of the antenna shown in Figure 2.3 . . . . . . . . . . . . . . . 10
2.5 Measured VSWR of the U-slot patch antenna on a finite microwave sub-
strate (antenna C in (Tong et al., 2000)) and computed VSWR of the same
antenna enclosed in a cylindrical cavity (see Figure 2.3) for two different
radii, R=5 cm and R=2.34 cm. . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.6 Computed VSWR of the U-slot patch antenna in Figure 2.3 with b=1.9 cm,
h=0.9 cm enclosed in a cavity of radius R=2.34 cm., compared to Ansoft
HFSS results. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.7 Example for two typical patch geometries of the possibilities given by the
artificial ports. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.8 Graphical description of the first step of the CAD methodology: application
of MAM. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.9 MLP feedforward NN used in the methodology. . . . . . . . . . . . . . . . . 17
2.10 Hidden neuron in the NN of Figure 2.9 . . . . . . . . . . . . . . . . . . . . . 17
2.11 Geometry and parameters of the proposed antenna . . . . . . . . . . . . . . 21
2.12 Photograph of the constructed and measured antenna. The superstrate has
been removed and the upper patch is the one shown at the bottom. . . . . . 25
2.13 Measured and computed VSWR for the designed antenna showing a 26%
relative bandwidth at VSWR=2. . . . . . . . . . . . . . . . . . . . . . . . . 25
2.14 Computed E-plane pattern at different frequencies. ECP : copolar component. 26
2.15 Computed H-plane pattern at different frequencies. ECP : copolar compo-
nent. EXP : crosspolar component. . . . . . . . . . . . . . . . . . . . . . . . 26
2.16 Return loss for the two different designs carried out with the methodology
according to the topology of Figure 2.11. . . . . . . . . . . . . . . . . . . . . 27
3.1 Scheme of the hybrid methodology used to obtain the GSM of the whole
array from the GSM(s) of the individual antenna element(s) considered as
isolated. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.2 Multiport network characterization of an antenna array. . . . . . . . . . . 34
xv
xvi LIST OF FIGURES
3.3 Coupling coefficient magnitude between the fundamental modes of two

WR90 rectangular apertures on an infinite ground plane versus the aperture
rotation angle α, for two aperture separations at 9.2 GHz. . . . . . . . . . . 35
3.4 Geometry of the hemispherical dielectric resonator antenna used as an array
antenna element. R = 12.7 mm, ǫr = 9.5, s = 6.4 mm, h = 6.5 mm. Coaxial
feed (50-ohm): ri = 0.5 mm, ro = 1.505 mm. . . . . . . . . . . . . . . . . . 36
3.5 E- and H-plane mutual coupling coefficient magnitude versus dielectric
resonator spacing at 3.68 GHz. . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.6 E-plane broadside-matched active reflection coefficient magnitude |Γ| versus
scan angle at 3.2 GHz of the central element for arrays of various sizes with
(a) d=0.5λ0 (b) d=0.6λ0 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.7 E-plane active element field pattern of the central element at 3.2 GHz for
arrays of various sizes with d=0.6λ0 . . . . . . . . . . . . . . . . . . . . . . . 39
3.8 H-plane active element field pattern of the central element at 3.2 GHz for
arrays of various sizes with d=0.5λ0 . . . . . . . . . . . . . . . . . . . . . . . 39
3.9 Computed active reflection coefficient magnitude |ρ| at 3.2 GHz for arrays
of various sizes with d=0.5λ0 for the elements across (a) the H-plane when
the array is at broadside scan, and across (b) the E-plane when the array
is scanned at 70◦ in the E-plane. . . . . . . . . . . . . . . . . . . . . . . . . 40
3.10 Computed active reflection coefficient magnitude |ρ| at different frequencies
for a 35×35 array with d=0.5λ0 at 3.2 GHz for the elements across the E-
plane when the array is scanned at (a) Broadside (b) 60◦ in the E-plane,
and (c) for the elements across the H-plane when the array is scanned at
45◦ in the H-plane. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.11 Finite array of stacked DRAs on an infinite ground plane fed with coax
probes through a rectangular waveguide. (a) Top view. (b) Longitudinal
view including the feeding network, the inner ports, the input ports, the
coax excitation ports and the hemispherical ports. . . . . . . . . . . . . . . 43
3.12 Scattering parameters versus frequency with two interelement spacings com-
pared to the results in (Eshrah et al., 2006). (a) d=20 mm. (b) d=15 mm. . 44
3.13 Radiation patterns compared to (Eshrah et al., 2006). (a) d=20 mm, f =8
GHz. (b) d=15 mm, f =8 GHz. (c) d=15 mm, f =10 GHz. . . . . . . . . . . 45
4.1 Top and side view of the cavity-backed circular microstrip patch used as an
array antenna element. R1 = 2 cm, R2 = 3 cm, c1 = c2 = 0.6 cm, xp = 0.62
cm, ǫr1 = 2.5, ǫr2 = 20. Coaxial feed (SMA connector): ǫr = 1.95, ri = 0.65
mm, ro = 2.05 mm. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
4.2 Field patterns of the E-plane 8 cavity-backed patch linear array for the ex-
citations given by the two optimization cases: when interelement couplings
are included in and excluded from the optimization procedure. . . . . . . . 58
4.3 Field pattern of the optimized 32 patch linear array along the H-plane.
Interferences at −70◦ , −40◦ and −20◦ are rejected and the maximum SLL
between −30◦ and −10◦ is set to -40 dB. . . . . . . . . . . . . . . . . . . . . 59
4.4 Field pattern of the optimized 32 patch linear array along the E-plane. . . . 60
4.5 Optimized 32-element linear patch array field pattern for different maximum
SLL regions and null directions. . . . . . . . . . . . . . . . . . . . . . . . . . 61
4.6 RDC and optimized (to ensure -20 dB of maximum SLL) field patterns of
the 17 linear DRA array along the H-plane. . . . . . . . . . . . . . . . . . . 62
LIST OF FIGURES xvii
4.7 Field pattern of the H-plane 17 linear DRA array with optimized excitations
to ensure -75 dB crosspolar level at θ0 = 30◦ . . . . . . . . . . . . . . . . . . 63
4.8 Optimized 21-element linear DRA array field pattern for two different max-
imum SLL regions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
4.9 Optimized 21-element linear DRA array field pattern for different maximum
SLL regions and two null locations. . . . . . . . . . . . . . . . . . . . . . . . 64
4.10 Field pattern of the 30 DRA circular array with optimized excitations to
ensure -30 dB of maximum SLL and crosspolar level. . . . . . . . . . . . . . 66
4.11 Optimized 9×9 patch array field pattern for -30 dB maximum SLL. . . . . . 68
4.12 Optimized 9×9 patch array field pattern for different regions of maximum
SLL. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
5.1 Field patterns of the uniform 21 DRA array for different excitations. ··· case
1: ideal isotropic pattern from RDC excitations. - - - case 2: real pattern
from RDC excitations. —– case 3: real pattern from proposed-technique
excitations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
5.2 Optimized field patterns of the 21 DRA array for different mainlobe widths.
- - - case 1: maximum SLL: -7.3 dB, 3-dB beamwidth: 5.8◦ . —– case 2:
maximum SLL: -16.9 dB, 3-dB beamwidth: 7.4◦ . . . . . . . . . . . . . . . . 72
5.3 Field patterns of the nonuniform 22 DRA array with a SLL region from
±5◦ to ±90◦ . - - - case 1: technique in (Holm and Elgetun, 1997) for sym-
metric isotropic arrays. —– case 2: proposed technique. . . . . . . . . . . . 73
5.4 Optimized 21 DRA array patterns with two main beams for two different
pairs of values of p−30◦ and p60◦ . . . . . . . . . . . . . . . . . . . . . . . . . 74
5.5 Two optimized 21 DRA array patterns with three main beams. In the
second one, the central mainbeam is broader than the other two mainbeams. 74
5.6 Optimized 21 DRA array patterns for a shaped-beam and a sidelobe region.
· · · case 1: fSB /fSL = 4. - - - case 2: fSB /fSL = 2. —– case 3: fSB /fSL = 1. 78
5.7 Zoom of the shaped-beam region of the optimized 21 DRA array patterns
for a shaped-beam and a sidelobe region. · · · case 1: fSB /fSL = 4. - - -
case 2: fSB /fSL = 2. —– case 3: fSB /fSL = 1. . . . . . . . . . . . . . . . . 78
5.8 Two-beam pattern of the 21-element DRA array obtained with the original
and the MC compensated excitations. . . . . . . . . . . . . . . . . . . . . . 80
5.9 csc2 pattern of the 21-element DRA array obtained with the original Woodward-
Lawson excitations and the MC compensated excitations. . . . . . . . . . . 81
5.10 Tseng-Cheng pattern of the 8×8 DRA array obtained with the original and
the MC compensated excitations. . . . . . . . . . . . . . . . . . . . . . . . . 82
5.11 Schelkunoff E-plane field pattern of the 8×8 DRA array obtained with the
original and the two different MC compensated excitations. . . . . . . . . . 83
6.1 Sector field patterns for the 81 DRA array obtained with the excitations
from the proposed FS procedure (coincident with the isotropic array pattern
from the classic isotropic FS method) and from the classic isotropic FS
method, compared to the ideal response. . . . . . . . . . . . . . . . . . . . 90
6.2 Sector field patterns for the 81 DRA array obtained with the excitations,
with a Hamming window, from the proposed FS procedure (coincident with
the isotropic array pattern from the classic isotropic FS method) and from
the classic isotropic FS method, compared to the ideal response. . . . . . . 91
xviii LIST OF FIGURES
6.3 Multisector field patterns for the 81 DRA arrays obtained with the excita-
tions, with a Hamming window, from the two classic WL methods and from
the proposed DFT procedure (coincident with the isotropic array pattern
from the classic isotropic WL design), compared to the ideal response. . . 92
6.4 csc2 field patterns for the 81 DRA array obtained with the excitations from
the proposed DFT procedure and from the two classic WL methods, com-
pared to the isotropic array pattern from the classic isotropic WL design
and the ideal response. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
6.5 Multisector field patterns, compared to the ideal response, obtained with
the excitations (with a Hamming window) from the two classic WL meth-
ods and from the proposed DFT procedure, coincident with the isotropic
array patterns from the classic isotropic WL design, for the two electrically
identical, but with different antennas, 49 element arrays. . . . . . . . . . . 94
6.6 Scheme of the relationships between feeding with/and spherical modes used
in the proposed spherical-wave-based pattern synthesis procedure . . . . . . 97
6.7 Layout of the 325-element array used for the spherical-wave-based synthesis
of a flat-top beam pattern . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
6.8 Three-dimensional normalized copolar field pattern of the 325-element array
of microstrip patches with 1.215 GHz of resonant frequency for a desired
flat-top beam objective copolar pattern. . . . . . . . . . . . . . . . . . . . . 101
6.9 Three-dimensional normalized crosspolar field pattern of the 325-element
array of microstrip patches with 1.215 GHz of resonant frequency for a
desired flat-top beam objective copolar pattern. . . . . . . . . . . . . . . . . 101
6.10 Normalized copolar (black) and crosspolar (grey) field patterns for all ϕ-
cuts of 1◦ of the 325-element array of microstrip patches with 1.215 GHz of
resonant frequency for a desired flat-top beam objective copolar pattern. . . 102
of microstrip patches with 2.5 GHz of resonant frequency for a desired flat-
top beam objective copolar pattern. . . . . . . . . . . . . . . . . . . . . . . 103
array of microstrip patches with 2.5 GHz of resonant frequency for a desired
flat-top beam objective copolar pattern. . . . . . . . . . . . . . . . . . . . . 103
resonant frequency for a desired flat-top beam objective copolar pattern. . . 104
of microstrip patches with 1.215 GHz of resonant frequency for a desired
sectorized csc2 objective copolar pattern. . . . . . . . . . . . . . . . . . . . 104
array of microstrip patches with 1.215 GHz of resonant frequency for a
desired sectorized csc2 objective copolar pattern. . . . . . . . . . . . . . . . 105
resonant frequency for a desired sectorized csc2 objective copolar pattern. 105
7.1 Proposed nested optimization scheme. . . . . . . . . . . . . . . . . . . . . . 111

7.2 Sketch of the value of fitness function (specified at each point) vs. the
number of active elements in a thinned array. . . . . . . . . . . . . . . . . . 112
LIST OF FIGURES xix
7.3 Flowchart of the computation of an individual’s fitness function. . . . . . . 116

7.4 Field patterns, in ϕ=0◦ , of the isotropic 20×10 thinned arrays achieving
-20 dB of maximum SLL. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
7.5 Field patterns, in ϕ=90◦ , of the isotropic 20×10 thinned arrays achieving
-20 dB of maximum SLL. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
7.6 Layouts of the active elements needed to achieve the patterns in Figs. 7.4
and 7.5 of the isotropic 20×10 thinned array. . . . . . . . . . . . . . . . . . 118
7.7 Field pattern of the isotropic 20×10 array thinned to achieve -30 dB of
maximum SLL. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
7.8 Layout of the active elements needed to achieve the pattern in Figure 7.7
of the isotropic 20×10 thinned array compared to the layout of a directly
sinthesized sparse array of 108 isotropic elements in (Chen et al., 2007). . . 120
7.9 Field patterns of the 31 DRA array thinned to achieve different maximum
SLL and mainbeam width restrictions with null field directions. . . . . . . . 121
7.10 Field patterns, for two different transition regions, of the 31 DRA array
thinned to synthesize a shaped flat-top beam response with restrictions. . . 121
7.11 Detail of the field pattern of Figure 7.10 for the flat-top beam response with
a transition region of θ ∈ (±15◦ ,±25◦ ). . . . . . . . . . . . . . . . . . . . . . 122
7.12 Detail of the field pattern of Figure 7.10 for the flat-top beam response with
a transition region of θ ∈ (±17◦ ,±23◦ ). . . . . . . . . . . . . . . . . . . . . . 122
7.13 Field patterns of a 17-DRA array thinned with both types of thinning to
achieve a -15 dB maximum SLL with θ0 =30◦ (11 active elements). . . . . . 124
7.14 Layout of the active elements needed to achieve a broadside gain of 16 dBi
by a thinned 6×6 DRA array with the two thinning options. . . . . . . . . 125
7.15 Number of active elements required to achieve a specified directive gain at
different mainbeam directions θ0 , ϕ0 of the 9×9 cavity-backed patch array. . 126
7.16 Field pattern of the 9×9 cavity-backed patch array thinned to achieve a
broadside pattern with a maximum -20 dB SLL in the whole space outside
the main beam (34 active elements). . . . . . . . . . . . . . . . . . . . . . . 126
8.1 Channel matrix value for a SISO system of two ideal dipoles. - - - case 1:
Classic transmission equation. —– case 2: Proposed model. . . . . . . . . . 128
8.2 Open-loop capacity for a 20 dB SNR of various-sized 4×4 MIMO system
of ideal dipoles. - - - case 1: Spherical-wave model in (Jiang and Ingram,
2005). —– case 2: Proposed model. . . . . . . . . . . . . . . . . . . . . . . 129
8.3 Waterfilling capacity for a 20 dB SNR of a 2×2 dipole MIMO system. Link
distance: 100λ0 . - - - case 1: Spherical-wave model in (Jiang and Ingram,
2005). —– case 2: Proposed model. . . . . . . . . . . . . . . . . . . . . . . 129
8.4 Open-loop capacity for a 20 dB SNR of a dipole MIMO system with fixed-
length arrays (5λ0 ). Link distance: 100λ0 . - - - case 1: Spherical-wave
model in (Jiang and Ingram, 2005). —– case 2: Proposed model. . . . . . 130
8.5 Signal Flow Graph of Generalized Network Model. . . . . . . . . . . . . . . 131
8.6 Signal Flow Graph of Generalized Network Model in its compact form, i.e.
transmit antennas, physical channel and receive antennas merged into one
network. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
8.7 Equivalent circuit of transmit and receive antennas. . . . . . . . . . . . . . 133
xx LIST OF FIGURES
8.8 Relative Ergodic Capacity Gain vs horizontal antenna inter element spacing
at the base station for two 742445 vertically polarized Kathrein antennas
and two vertically oriented dipoles . . . . . . . . . . . . . . . . . . . . . . . 137
8.9 Mobile station antenna setups. (a) Two parallel dipoles rotated with respect
to the z-axis and (b) two perpendicular dipoles rotated with respect to their
starting position as vertically and horizontally polarized dipoles. . . . . . . 138
8.10 Relative capacity gain of 742445 vertically polarized Kathrein antenna for
two parallel dipoles with varying rotation angle α . . . . . . . . . . . . . . . 139
8.11 Relative capacity gain of 742445 vertically polarized Kathrein antenna for
two perpendicular dipoles with varying rotation angle α. . . . . . . . . . . . 140
8.12 Comparison of the synthesized radiation pattern (—–) and the 742445 ver-
tically polarized Kathrein radiation pattern (- - -) in (a) azimuth and (b)
elevation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
8.13 Relative ergodic capacity gain vs horizontal antenna inter element spacing
at the base station for two vertically polarized synthesized antennas and
two vertically oriented dipoles . . . . . . . . . . . . . . . . . . . . . . . . . . 142
List of Tables
2.1 Values of the parameters of antenna in Figure 2.11 . . . . . . . . . . . . . . 22

2.2 Results for the different optimization algorithms . . . . . . . . . . . . . . . 24
4.1 Computational aspects of linear patch arrays optimizations represented in

Figs. 4.2, 4.3 and 4.4. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
4.2 Results of the optimizations of the 17 linear DRA array. JSL = 162, MXP =
1 (=0 in the first optimization). . . . . . . . . . . . . . . . . . . . . . . . . . 62
4.3 Results of the optimizations of circular arrays of patches. . . . . . . . . . . 65
4.4 Optimized excitations (amplitude and phase) for maximum directive gain
of a circular array with N = 8 and d = 0.9λ0 . . . . . . . . . . . . . . . . . . 65
4.5 Optimizations of the 4×4 patch array for DR control with =-20 DB in the
H- and E- plane. JSL = 204, NDR = 16. . . . . . . . . . . . . . . . . . . . . 67
4.6 Results of the optimization of the 9×9 broadside patch array. . . . . . . . . 67
7.1 Optimized excitations of the thinned array realizations of two different 17-
element arrays for both thinning options X: element removed, 0: element
turned off . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
xxi
xxii LIST OF TABLES
Glossary of symbols
{a} Complex coefficients of incident spherical 32

modes of an array
{ai } Complex coefficients of incident spherical 30
modes of an antenna i not coming from other
antennas in an array environment
{ainc
i } Complex coefficients of incident spherical 8
modes of an antenna i
{b} Complex coefficients of scattered spherical 32

modes of an array
{bi } Complex coefficients of scattered spherical 30
{bout
i } Complex coefficients of outgoing spherical 8
~em (θ, ϕ) mth spherical far field vector mode 8
[Γ] Reflection matrix of an array 33
[R] Reception matrix of an array 33

[ρi] Reflection matrix of an antenna i 8
[ri ] Reception matrix of an antenna i 8
[S] Scattering matrix of an array 33

[si ] Scattering matrix of an antenna i 8
[T] Transmission matrix of an array 33

[ti ] Transmission matrix of an antenna i 8
{v} Complex coefficients of incident feeding or ex- 32

citation modes of an array
{vi } Complex coefficients of incident feeding or ex- 8
citation modes of an antenna i
{w} Complex coefficients of reflected feeding or ex- 32

citation modes of an array
xxiii
xxiv Glossary of symbols
{wi } Complex coefficients of reflected feeding or ex- 8

citation modes of an antenna i
Glossary of acronyms
BVP Boundary Value Problem 6
CAD Computer-Aided Design 1, 3,

4, 14,
15, 18,
19, 24,
143,
145
DFT Discrete Fourier Transform 85, 86,

88, 91,
93–95,
144
DR Dynamic Range 48, 50–
53, 55,
61–63,
65–67,
143
DRA Dielectric Resonator Antenna 30, 35,
42, 44,
59, 65,
72, 73,
77, 79,
80, 95,
143–
145
FE Finite Element 1, 3–6,

8, 9,
13–15,
18, 21,
23, 29,
42, 56,
57, 143
xxv
xxvi Glossary of acronyms
FF Fitness Function 23,

110,
112,
115,
117,
120
FS Fourier Series 85–91,
95, 144
GA Genetic Algorithm 1, 2,
19, 20,
23, 24,
107,
108,
110,
115,
144
GAM Generalized-Admittance-Matrix 5, 7,
13–15
GSM Generalized-Scattering-Matrix 1, 2,
5, 7–9,
15, 29,
30, 33,
42, 85,
95, 98,
108,
109,
127,
128,
143–
145
GTM General Translation Matrix 30, 98,
128
IPM Interior-Point Method 48, 53–

57, 59,
61, 63,
65–67,
143
KKT Karush-Kuhn-Tucker 53, 54

Glossary of acronyms xxvii
LP Linear Programming 2, 48,

69,
71–73,
75, 76,
108,
114,
115,
144
MA Modal Analysis 1, 4, 5,
9, 29,
42, 57,
143
MAM Multipurpose-Admittance-Matrix 10, 13–
15, 21,
23
MC Mutual Coupling 36, 77,
79–81,
83, 96,
100,
128,
144,
145
MIMO Multiple-Input-Multiple-Output 2, 127–
131,
135,
136,
138,
145
MLP Multilayer Perceptron 15
MSE Mean Square Error 18, 23,
69, 79,
81, 114
NN Neural Network 1, 3,
15, 18
QP Quadratic Programming 2, 69,

108,
113,
114,
144
RDC Riblet-Dolph-Chebychev 59, 61,

62, 72
SA Simulated Annealing 1, 20,

23, 24
xxviii Glossary of acronyms
SISO Single-Input-Single-Output 127,

130,
136
SLL Sidelobe Level 47, 48,
50, 51,
57, 59,
61, 62,
65–67,
69,
72, 74,
75, 77,
108,
112,
113,
115,
117,
143,
144,
147
SNR Signal-to-Noise Ratio 130,
136
WL Woodward-Lawson 85, 87,

89, 91,
93, 94,
144
Chapter 1
Introduction
1.1 Motivation and objective

Most current and future communication systems are being deployed in the microwave and
millimeter-wave frequency bands. Systems and devices operating at these high frequencies,
especially radiating structures as antennas and antenna arrays, usually require a complex
design process.
Nowadays, there are a number of Computer-Aided Design (CAD) tools, mostly based
on analytical techniques or equivalent circuit models which are valid for a specified type
of structures, and several numerical methods to carry out the electromagnetic analysis of
more complex structures. As a general rule, the latter bear a high computational burden
which makes them be pushed to the background in the design process and eventually be
recovered for validation purposes. Notwithstanding, the various assumptions regarding ge-
ometries, antenna features, . . . or the fact of neglecting important electromagnetic effects
(such as scattering or coupling), usually made when using a fast and simplified design tech-
nique (provided there is such a tool available for the intended design), make the achieved
designs with these techniques not yield the desired response in real implementations.
Therefore, the main objective of this work is the development and application of,
on the one hand, efficient analysis methods that allow a fast and rigorous electromag-
netic characterization of radiating structures (antennas and antenna arrays), and, on the
other hand, efficient design techniques that, by usually taking advantage of some of the
aforementioned analysis methods, are capable of inherently taking into account all electro-
magnetic effects and interactions (such as mutual coupling in arrays) in complex radiating
structures throughout the whole design process.
1.2 Text organization

This Ph. D. dissertation is organized as follows:
Chapter 1 serves as an introduction and reports the main motivation and objective
of the work. In Chapter 2, a novel CAD methodology for the design of probe-fed stacked
microstrip patch antennas is proposed. It is based on a hybrid Finite Element (FE)-
Modal Analysis (MA) method, Neural Networks (NNs) and global optimization through
Genetic Algorithms (GAs) and Simulated Annealing (SA). As the hybrid FE-MA will
be used throughout the thesis to yield different antennas’ Generalized-Scattering-Matrixs
(GSMs), a thorough description of the method will be given in this Chapter 2. Again, as
1
2 Introduction
the GA will be used in a different chapter for other purposes, the details of the GA will
only be given in Chapter 2.
Chapter 3 presents the GSM-analysis of finite arrays whose elements’ radiated fields
can be individually expressed as spherical vector wave expansions. This analysis method
will be the core for the array design techniques developed in successive chapters as they will
all start by assuming the GSM of a finite array is known, therefore making electromagnetic
effects be inherently taken into account through any design technique used to compute
the necessary excitations to achieve the desired pattern specifications.
Chapters 4 and 5 report a wide variety of array design techniques where the design
problem can be formulated, after different transformations, as an optimization problem.
Specifically, in Chapter 4 the optimization procedure used to solve the resulting problem
is developed, while in Chapter 5 the stress is made in making the problem formulation
result in a Linear Programming (LP) or Quadratic Programming (QP) problem.
Chapter 6 presents two non-optimization based synthesis techniques for arrays. The
first one is based on the Fourier expansion of the array radiated field, while the second
one is based on the spherical wave expansion of the array radiated field.
Chapter 7 addresses the topic of array thinning by proposing a nested optimization
procedure to thin an array. The GA previously reported is used to decide which elements
are turned on, while other optimization techniques, mainly the ones reported in Chapter
5, are used to compute the excitation of the array elements.
Chapter 8 leaves the design problem apart and goes into the use of the GSM for-
mulation, alone or in combination with other techniques, to analyze and characterize
Multiple-Input-Multiple-Output (MIMO) systems.
Finally, Chapter 9 gives some conclusions in the form of the original contributions of
this Ph. D. dissertation and the future research lines that it has given rise to. Furthermore,
the framework in which this work has been carried out (research projects, financial support,
. . . ) are reported together with the publications this Ph. D. dissertation has generated.
Chapter 2
Design methodology for stacked

microstrip patch antennas
2.1 Introduction
Microstrip patch antennas have been focusing the attention of researchers for a long time
because they are easy to use in different configurations and because of their versatility in
achieving a wide variety of responses (Garg et al., 2001). One of the most usual configura-
tions microstrip antennas can be found in is the one where two or more patches are stacked
within several dielectric layers. These patches are intended to introduce resonances, thus
being able to increase bandwidth and have a broadband/multiband antenna (Waterhouse,
1999). Circuit models are not accurate enough to carry out a rigorous analysis of these
structures, the use of a full-wave numerical technique then being required, such as FE
methods, Finite-Difference methods, Method of Moments or hybrid methods. However,
until not so long ago the time consumed by these techniques made them prohibitive when
addressing a new design, pushing them into the background and recovering them only for
validation purposes. On the other hand, not only there have been astounding advances
in the field of computation but there have also been promising achievements towards full-
wave optimization, in the sense of newly developed clever procedures, working as novel
standalone methods or applied to the aforementioned classic numerical techniques.
In this chapter a novel CAD methodology for the design of probe-fed stacked mi-
crostrip patch antennas is proposed. The main idea consists of making the computational
effort at the beginning of the design methodology which will eventually allow a fast final
global optimization to be made. This methodology incorporates the rigour of a contrasted
numerical technique, such as the FE method. Furthermore, in order to be able to perform
an optimization in our design process, two well-known advanced mathematical techniques
have been used: a NN (Haykin, 1998) and global optimization algorithms, specifically
genetic algorithms (Goldberg, 1989) and simulated annealing (Otten and van Ginneken,
1989).
A canonical representation of the antenna being designed is shown in Figure 2.1. It
basically consists of T stacked patches within L dielectric layers, all of which are enclosed
in a cavity. The lower patch is fed via a coaxial probe. The shape and location of the
patches, placed parallel to the x-y plane as seen in Figure 2.1, can be arbitrary. In the
same way, the profile of the cavity along the z direction, as well as its cross-sectional shape
(in planes perpendicular to that direction), can also be piecewise arbitrary as also shown in
3
4 Design methodology for stacked microstrip patch antennas
Figure 2.1: Canonical geometry of the antenna to be designed.
Figure 2.1. To make our design, we must first decide on some geometrical constraints and
the parameters and their associated ranges of values involved in the design x1 , x2 , . . . , xI
must be selected according to them. This choice is critical because it requires having the
necessary knowledge of which ones really do affect the desired output of the antenna. This
output, which we will refer to as y, should be directly related to any physical response of
the antenna (directive gain, matching, input-match bandwidth, . . . ) or a combination of
them.
We have divided the CAD methodology into three steps which will be described
throughout the following sections:
1. Hybrid FE-MA with artificial ports
2. Neural Network
3. Global optimization
2.2 Hybrid Finite Element - Modal Analysis characteriza-

tion of an antenna
2.2.1 Introduction
Over the last few years the FE method has been applied to radiation and scattering
problems because of its ability to consider complex structures or inhomogeneous materials.
It is well known that, when an open-region is analyzed using the FE method, this region
must be truncated to restrict the computational domain. As a consequence, an artificial
boundary appears and a boundary condition is required. Basically, three different kinds of
2.2 Hybrid Finite Element - Modal Analysis characterization of an antenna 5
boundary conditions have been proposed in the literature. One of them is the absorbing
boundary condition, including perfectly matched layers. Another type is the boundary
condition based on boundary integral equations. A very good summary of these two types
of boundary conditions is given in (Liu and Jin, 2001), where a hybrid method with the
advantages of both kinds is developed. The third kind is based on the harmonic expansion
of the field on the boundary (Rubio et al., 2002).
Usually, when an expansion in spherical harmonics is applied, a spherical mesh is

used. Although for lengthened geometries this is computationally less efficient than a
conformal mesh adapted to the shape of the structure, the handling of spherical modes
has significant advantages. First, this boundary condition is exact and the possible inac-
curacies are solely due to the discretization in the FE method or due to computational
roundoff. Second, a purely sparse FE matrix is obtained and therefore it can be efficiently
stored and solved. Third, the spherical boundary can be situated very close to the scat-
terer or the radiator by using a suitable number of spherical modes and, on the other hand,
the computational domain where the expansion in spherical harmonics is applied can be
reduced more if a domain segmentation method is used in order to analyze the feed region
separately (Rubio et al., 2002). Fourth, when homogeneous waveguides or transmission
lines are used for the excitation of the antenna, a fast frequency-sweep analysis can be
carried out by using a reduced order model obtained by means of the symmetric matrix
Padé-Via-Lanczos algorithm, from an appropriate formulation of the three-dimensional
FE method, as explained in (Rubio et al., 2002) and (Feldmann and Freund, 1997).
Segmentation Finite Element via Lanczos-Padé (SFELP) is a methodology to an-

alyze microwave passive devices (Rubio et al., 1999, 2001). It is based on the domain
segmentation technique, the FE method and the MA through the Generalized-Admittance-
Matrix (GAM) and the GSM The procedure has been extended to the analysis of radiating
structures (Rubio et al., 2002) through the inclusion of a spherical computation domain
where a spherical modal expansion (on what is called a spherical port) is performed to
characterize the radiating region. The FE simulation is synergized by considering modal
field expansions and consequently a GAM results, i.e. a matrix valued-transfer function.
This simulation can be carried out on a frequency-by-frequency basis, or on the other hand
make use of a reduced-oder model, namely a Padé approximation scheme via the Lanczos
process.
One of the most powerful features of this technique is its capability to segment a
given structure in different regions delimited by modal ports where modal expansions of
the electromagnetic fields are used. The application of the hybrid FE-MA method to each
region provides the GAM that describes it, which relates complex modal coefficients on
the considered ports in that region. The partial GAMs can be assembled analytically, with
no additional numerical effort, by connecting the desired ports through a section with an
arbitrary length of a canonical transmission line (such as a rectangular or cicular waveg-
uide, or a coaxial line) or a transmission line which has been numerically characterized by
a two-dimensional FE method.
This feature allows the design methodology to modify any distance along the z-axis
in Figure 2.1 easily by dividing the whole structure into R regions, allowing the variation
of parameters such as dielectric thicknesses or separation between patches.
2.2.2 Finite Element formulation

The formulation is as follows: by stating the Boundary Value Problem (BVP) given by
the frequency-domain Maxwell equations in terms of the magnetic vector field H̄ we can
carry out a Galerkin approach which provides the following weak formulation that governs
the electromagnetic behaviour inside the domain Ω of the region under analysis:
Find
Z H̄ ∈ H such that
1 jk
Z
2
∇ × H̄ · ∇ × W̄ − k µr H̄ W̄ dV = − Ψ̄ · W̄ dS ∀W̄ ∈ H, (2.1)
Ω ǫr η0 ∂Ω
√ p
where k = ω µ0 ǫ0 , η0 = µ0 /ǫ0 , µr and ǫr are, respectively, the relative permeability and
permittivity of the medium, H is an admissible vector function space in which a solution
to the BVP is to be found, and Ψ̄ is the tangential electric vector field on the boundary
∂Ω.
Within the discretized finite element space, the magnetic vector field can be written
in the domain Ω as X
H̄ = hn w¯n (2.2)
n
where hn is the coefficient of the basis function w¯n .
In every modal port belonging to the boundary ∂Ω, the tangential electric vector
field can be expressed as a linear combination of modes:
X
Ψ̄ = V (m) (n̄ × et,m
¯ ) (2.3)
m
where et,m
¯ is the tangential electric vector field corresponding to the mth mode on the
modal port, V (m) is its associated coefficient and n̄ is the unit outward normal vector to
∂Ω on the modal port. By applying the natural and essential boundary conditions to the
rest of ∂Ω, we arrive at the following classic FE algebraic system:
jk
[K] − k2 [M] {h} = − [B] {V}

(2.4)
η0
where {h} and {V} are coefficient vectors of the magnetic field in Ω and the modal electric
field in the modal ports of ∂Ω, while [K] and [M] are sparse symmetric matrices, and [B]
is a sparse matrix, with elements given by:
1
Z
Knm = ∇ × w¯n · ∇ × w¯m dV (2.5)
ZΩ ǫ r
Mnm = µr w¯n w¯m dV (2.6)

ZΩ
Bnm = (n̄ × et,m
¯ ) w¯m dS (2.7)
∂Ω
where the integral of Bnm is carried out in the parts of ∂Ω where a modal port is defined.
Now, the tangential magnetic field on the modal port of ∂Ω can be set as a linear
combination of modal tangential magnetic fields ht,m¯ with associated coefficients I (m) as:
¯
X
n̄ × H̄ = n̄ × I (m) ht,m (2.8)
m
From these expressions and after different operations (Rubio et al., 1999, 2001), we
can arrive at matrix relationship between the normalized modal tangential electric and
magnetic fields on the modal ports of ∂Ω as:
{I}=[Y] {V} (2.9)

−jk −1
{I}= [B]T [K] − k2 [M] [B] {V} (2.10)
η0
Matrix [Y] is the GAM corresponding to the region Ω under analysis. By factoring
matrix [B] to remove the frequency dependency, the Padé via Lanczos algorithm can be
applied to obtain a reduced order model of this matrix transfer function.
2.2.3 Multimode multiport network characterization of an antenna

When an antenna i is going to be analyzed, a spherical port is considered in the radiation
region, as shown in Figure 2.2. In this spherical port, a modal expansion of the tangential
electric Ēt,i and magnetic H̄t,i field in terms of T E r and T M r spherical modes (Marcuvitz,
1986) is used to satisfy the boundary condition:
Mi (m)
X Vi (r)
Ēt,i = ēt,m (θ, ϕ) (2.11)
r
m=1
Mi (m)
X Ii (r)
H̄t,i = h̄t,m (θ, ϕ) (2.12)
r
m=1
Here, ēt,m and h̄t,m are the tangential electric and magnetic fields for each spherical mode
(m) (m)
m, Vi and Ii are their complex amplitudes in terms of voltages and currents, and Mi
is the number of spherical modes used in the spherical port of an antenna i.
Once the GAM is obtained, the GSM of an antenna i shown in Figure 2.2 is calcu-
lated by using the following expressions that relate complex amplitudes of voltages and
(m) (m)
currents with complex amplitudes of incoming ainc,i and outgoing bout,i spherical modes
(depending on the T E r or T M r nature of each spherical mode):
(m) (m) (m) (m)

Vi (r)⌋T E r = Ii (r)⌋T M r = ainc,i h(1) (2)
n (kr) + bout,i hn (kr) (2.13)
(1) (2)
(m) (m) (m) dhn (kr) (m) dhn (kr)
Ii (r)⌋T E r = Vi (r)⌋T M r = jainc,i + jbout,i (2.14)
d(kr) d(kr)
m = 1, . . . , Mi
(1) (2)
where hn and hn are the spherical Hankel functions of first and second kind, k is the
free-space wavenumber and r is the radius of the sphere where the spherical port is set.
The number of spherical modes Mi , is fixed by the highest degree of the spherical
Hankel functions (nmax ) in the spherical mode expansion, and by symmetry properties of
the antenna. In (Hansen, 1988) the following empirical rule to calculate this degree has
been given:
nmax = ⌊krmin ⌋ + nl (2.15)
where rmin is the radius of the smallest sphere that circumscribes the antenna and the
symbol ⌊·⌋ indicates the largest integer smaller than or equal to. This integer gives the
Figure 2.2: GSM of an antenna i
highest degree of the propagating modes. The integer nl depends on the required accuracy
as well as on krmin and on the radius where the field is evaluated (given by the FE mesh).
Fixing nl equal to 2 the approximation is good enough for antennas with a diameter close
to half a wavelength (Rubio et al., 2005). As shown in (Hansen, 1988), the analysis of an
antenna i on an infinite ground plane requires Mi = nmax (nmax + 2), without considering
other symmetry properties.
The resultant GSM for an antenna i with two ports as shown in Figure 2.2 is given
by:
[ρi ] [ri ] {vinc
i } {wi }
= (2.16)
bout

[ti ] [si ] ai i
where {vi },{wi },{ainc out

i }and {bi }are respectively the complex amplitudes of incident and
reflected modes on the feed waveguides or transmission lines, and incoming and outgoing
spherical modes on the radiation port. In the same way, [ρi],[ri ],[ti ]and [si ]are, respec-
tively, the antenna reflection, reception, transmission and scattering matrices as defined
in the classic theory (Montgomery et al., 1965). The antenna reflection matrix provides,
directly, the reflection parameters for the feed modes. Moreover, assuming no incoming
field towards the antenna, {ai } = {0}, the radiated field E ~ i (θ, ϕ) can be calculated as a
weighted sum of far field spherical vector modes ~em (θ, ϕ):
Mi
~ i (θ, ϕ) = (m)
X
E bout,i~em (θ, ϕ) (2.17)
m=1
2.2.4 Example: Design of a broadband U-slot microstrip patch antenna

The U-slot patch antenna studied in (Tong et al., 2000) has been used as the basis to design
a broadband microstrip antenna shown in Figure 2.3. It consists of a U-slot rectangular
patch antenna printed on a dielectric substrate, enclosed in a cylindrical metallic cavity
and fed in the centre via the inner conductor of a coaxial connector. The geometry
provided in (Tong et al., 2000) consists of a rectangular U-slot patch antenna on a finite
microwave substrate placed on a finite ground plane. We have embedded the antenna in a
cylindrical metallic cavity recessed in an infinite metallic plane resulting in the structure of
Figure 2.3. The dimensions of the patch, U-slot, substrate thickness, and the permittivity
of the substrate, given in this figure, are those of antenna C in (Tong et al., 2000). For
the analysis, a 50-ohm SMA connector has been chosen and all the dimensions and the
dielectric permittivities are considered in the simulations.
Figure 2.3: Top and side view of the geometry of the probe-fed U-slot rectangular patch embed-
ded in a cylindrical cavity. Coaxial feed (SMA connector): εr =1.95, do =0.41 cm, di =0.128 cm.
Antenna parameters from (Tong et al., 2000): εr =2.33, a=1.4 cm, b=1.8 cm, c=0.2 cm, g=0.7 cm,
L=3.6 cm, W=2.6 cm, h=0.8 cm.
In order to simplify the analysis process, the antenna is segmented into two regions
delimited by modal ports as shown in Figure 2.4. The first region includes the coaxial
feed and part of the cavity region, and the second region comprises the U-slot patch and
a semi-spherical computation domain where a spherical modal expansion is carried out to
characterize the radiating region. Both regions are separated by a fictitious coaxial port, as
shown in Figure 2.3, where once again an analytical modal expansion of the fields is carried
out. The application of the hybrid FE-MA method provides the GSMs that describe each
region by relating complex modal coefficients on the considered ports. Finally, these
partial GSMs are analytically assembled through a section of a fictitious coaxial line, and
the overall GSM of the isolated antenna is obtained, as shown in Figure 2.4. In this way
the thickness of the substrate may be varied easily without additional numerical effort.
Figure 2.5 shows the computed VSWR against frequency for two cavity radii, R=2.34
cm and R=5 cm, compared with the measured values in (Tong et al., 2000) corresponding
to the same patch but printed on a finite microwave substrate (without a cavity) placed
on a finite ground plane. As can be seen, a good agreement, except for a small level
discrepancy in the second resonant frequency, is obtained for a large cavity radius (5 cm),
Figure 2.4: Port diagram of the antenna shown in Figure 2.3
as is expected. For a cavity radius of 2.34 cm which situates the cavity wall very close to
the patch, there is an impedance mismatch. On the other hand, a small cavity may be
required in case a phased array is made up of these radiating elements in order to have
an interelement spacing that avoids the appearance of grating lobes in the considered
frequency band. Hence, in order to maintain a broadband characteristic in the VSWR
with R=2.34 cm, the impedance mismatch is compensated by modifying other dimensions.
In this way, the length b of the U-slot (bringing its arms nearer the upper edge of the
rectangular patch) and the substrate thickness have been modified (b=1.9 cm, h=0.9 cm)
following the guidelines detailed in (Weigand et al., 2003). The modified patch embedded
in a 2.34 cm radius cavity gives a relative bandwidth of about 20% for a VSWR<2 as
shown in Figure 2.6. This result is compared to the one provided by a commercial High-
Frequency Structure Simulator (Ansoft HFSS), both being in good agreement.
2.3 Inclusion of artificial ports: multipurpose admittance

matrices
In the Multipurpose-Admittance-Matrix (MAM) approach developed in (de la Rubia and
Zapata, 2007), a further exploitation of the matrix-valued transfer function describing
electromagnetic behaviour is considered. In addition to the inclusion of modal ports
in the analysis domain, new computational domain ports (artificial ports) are taken into
account. The aim of this new port definition is to obtain the aforementioned matrix-valued
transfer function even though some boundary field distributions are not yet established.
In other words, to be able to carry out a full-wave analysis in a given structure in spite
of certain physical shapes and dimensions not yet being established; and afterwards, to
make it possible to consider conducting strips/slots in the computational domain with
2.3 Inclusion of artificial ports: multipurpose admittance matrices 11
Figure 2.5: Measured VSWR of the U-slot patch antenna on a finite microwave substrate (antenna
C in (Tong et al., 2000)) and computed VSWR of the same antenna enclosed in a cylindrical cavity
(see Figure 2.3) for two different radii, R=5 cm and R=2.34 cm.
Figure 2.6: Computed VSWR of the U-slot patch antenna in Figure 2.3 with b=1.9 cm, h=0.9
cm enclosed in a cavity of radius R=2.34 cm., compared to Ansoft HFSS results.
2.3 Inclusion of artificial ports: multipurpose admittance matrices 13
ease by causing field discontinuities/continuities by means of simple manipulations of the

matrix-valued transfer function.
In a similar way as was done on modal ports in the previous section, we can define
the tangential electric vector field on an artificial port by taking advantage of the FE
context as an expansion of piecewise basis functions:
X
Ψ̄ = Vl (n̄ × ψ̄l ), (2.18)
l
where ψ̄l and Vl represent the piecewise function and its associated coefficient, and n̄ is
the unitary vector normal to the artificial port. Analogously, the tangential magnetic field
on such artificial port may be expressed as
X
n̄ × H̄ = n̄ × Il ψ̄l , (2.19)
l
where Il is now the associated coefficient to ψ̄l .

Following a similar scheme to the previous section, we can arrive at a matrix-valued
transfer function from problem (2.1), which stands for the relationship between the modal
fields in the P modal ports and the piecewise functions in the A artificial ports lying on
the boundary ∂Ω of our domain:

{I} −jk −1 T 2
−1 {V}
= [d] [β] [K] − k [M] [β] (2.20)
{I} η0 {V}
T
{I} = I1 T , . . . , Ip T , . . . , IP T (2.21)
T T T T

{I} = I1 , . . . , Ia , . . . , IA (2.22)
T T T T

{V} = V1 , . . . , Vp , . . . , VP (2.23)
T T
{V} = V1 T , . . . , VaT , . . . , VA

(2.24)
where Vp , Ip , Va and Ia are coefficient vectors of, respectively the tangential electric and
magnetic field on the modal port p, and the tangential electric and magnetic field on the
artifical port a, [K] and [M] are sparse symmetric matrices, [β] is a sparse matrix and [d]
is a block diagonal matrix (see (de la Rubia and Zapata, 2007) for all the details).
Hence, a full-wave admittance-type matrix, namely a MAM arises and, as a result,
we are able to deal with a multiport network where the electromagnetic complexity in
the analysis domain has been removed. This way, multiple different physical shape and
dimension analysis can be obtained straightforwardly from the MAM. So far, the full-wave
transfer functions of a variety of specific device realizations requiring only one large sparse
FE stiffnes matrix inversion-like computation is obtained.
In this full-wave network, straigthforward elementary circuit manipulation can be
carried out in the artificial ports, all with a direct electromagnetic counterpart in the
actual structure in the form of a GAM, removing all physical uncertainties. For example,
we can nullify the tangential electric field, thus rendering an artificial port short-circuit,
by simply imposing Va = 0 on the desired artifical port a in (2.20). Analogously, we may
impose Ia = 0 to have an artificial port a open-circuit. Finally, we can impose the field
continuity across two artificial ports a and b defined on the two sides of the same portion
of a surface, both sharing the same mesh, by setting Va = Vb and Ia = −Ib in (2.20).
Figure 2.7: Example for two typical patch geometries of the possibilities given by the artificial
ports.
This feature allows the design methodology to vary parameters easily directly related
to the shape and location in the x-y plane of a patch where pairs of artificial ports may
be defined on each side of the patch surface. For example, Figure 2.7 shows the different
cases for two typical geometries, a circular patch made up of two circular ports, A and B,
and two ring artificial ports, F and G; and a rectangular patch made up of two triangular
artificial ports, A and B, and two irregular quadrilateral artificial ports, F and G. As
previously stated, these two pairs of artificial ports (A and B, F and G) are defined on
the same portion of the surface but on different sides. By connecting each pair of artificial
ports or letting them short-circuit (trough the appropriate MAM manipulations), we can
easily vary the shape and dimensions of these patch geometries, having the possibilities
shown in Figure 2.7.
In this way, the first step of the CAD methodology, which is intended to bear the
majority of the computational burden, can be summarized in Figure 2.8. As can be seen,
we first calculate the MAM corresponding to the R regions we have divided our whole
structure delimited by the modal ports in Figure 2.1, which exactly requires only R FE
simulations. Within these sections, there will be a number of pairs of artificial ports used
to model the shape and location flexibly in the x-y plane of the patches inside each region
as also shown in Figure 2.1. By properly connecting these latter pairs of ports or letting
them short-circuit, we can produce Sr GAMs for region r, each one corresponding to a
different shape and/or location of the aforementioned patches, thus being able to vary
2.4 Neural network 15
the parameters associated to these shapes and locations without the need of carrying out
any more FE simulations. We can then finally assemble all combinations of the GAMs
belonging to all regions via the analytical connection of modal ports through sections of
transmission lines, which corresponds to a variation in substrate thicknesses. This way,
several GAMs (which can easily be converted into GSMs if needed), which represent the
results of carrying out a FE simulation for each different combination of different values
of design parameters belonging to discrete sets of values Xi , are easily obtained with an
affordable computational effort.
2.4 Neural network

Once we have carried out an overall MAM simulation with the advantages of the aforemen-
tioned features, we have a training set for the NN (Haykin, 1998) consisting of P patterns,
corresponding to the discrete sets of values Xi of the parameters (directly taken from the
MAM simulation). Each pattern p has an input, xp1 , xp2 , . . . , xpi , . . . , xpI , and an ouput, y p .
The purpose of this training is to obtain a fast mathematical tool that provides a value z
of the response immediately for an arbitrary value of the design parameters x1 , x2 , . . . xI ,
which would be similar to the value y given by a full-wave simulation, but without the
need to carry it out.
2.4.1 Normalizing
One of the most important steps before training a NN is to carry out a normalization both
for the input and output values of the training patterns. In our case, such a normalization
is carried out as usual by calculating first the mean and the standard deviation over the P
patterns for each input i, i.e. x̄i and σxi , and for the output, ȳ and σy , to finally normalize
every value as follows.
P P
1 X p 1 X p
x̄i = xi ȳ = y (2.25)
P P
p=1 p=1
v v
u P u P
uX p uX
u
u (xi − x̄i )2 u
u (y p − ȳ)2
t p=1 t p=1
σx i = σy = (2.26)
P −1 P −1
To finally normalize every value as follows:
xpi − x̄i y p − ȳ
x̃pi = ỹ p = (2.27)
σx i σy
2.4.2 Training
Figure 2.9 shows a simple Multilayer Perceptron (MLP) feedforward NN (Haykin, 1998),
with an input layer (with normalized inputs x̃i , i = 1 . . . I), a hidden layer (with H hidden
neurons), and an output layer with a linear neuron which directly provides output z̃.
This is the type of neural network we have implemented for our CAD methodology. The
weights connecting the input layer with the hidden layer will be referred to as wi,h , while
the weights connecting the hidden layer with the output will be referred to as vh . Note
how every input to a layer (especifically, the hidden layer and the output layer) is biased.
Figure 2.8: Graphical description of the first step of the CAD methodology: application of MAM.
2.4 Neural network 17
Figure 2.9: MLP feedforward NN used in the methodology.
Figure 2.10: Hidden neuron in the NN of Figure 2.9
The only neuron belonging to the output layer has been chosen to be implemented as a
simple sum of the weighted inputs to the layer (known as a linear neuron). On the other
hand, each neuron belonging to the hidden layer has been implemented as a perceptron
with an activation function given by a tanh(bζ), as shown in Figure 2.10, where a and b are
parameters intentionally left to be adjusted depending on the problem and the preferences
for the training. For the training, we carry out a random sequential learning based on
the error backpropagation algorithm (Haykin, 1998; Rumelhart et al., 1987; Brierley and
Batty, 1999). In sequential learning, the weights are updated for each pattern given as an
input to the network in the learning process. Once the training has completed an epoch,
that is to say, has used all the patterns as inputs and correspondingly updated the weights,
if no stopping criterion has been achieved, we proceed to another epoch containing all the
patterns, but given in a different order (chosen at random) than in the previous epoch.
For a normalized input pattern x̃pi , i = 1, . . . , I, the network produces an output z̃ p . The
error incurred by the network for that pattern with respect to the original normalized
output ỹ p is given by ep = z̃ p − ỹ p . Thus, we may find the updating rules for every weight
in the network as (Brierley and Batty, 1999; Brierley, 2006):
vh ←vh − ηout a tanh(ζh )ep (2.28)

vH+1 ←vH+1 − ηout aep (2.29)
wi,h ←wi,h − ηhid vh ab 1 − tanh (ζh ) ep x̃pi
2

(2.30)
wI+1,h ←wI+1,h − ηhid vh ab 1 − tanh2 (ζh ) ep

(2.31)
for i = 1, . . . , I h = 1, . . . , H
In (2.28)-(2.31) ηout and ηhid represent, respectively, the learning rates for the hidden-to-
output and the input-to-hidden connections. We have set ηhid = 10ηout . These learning
rates may be fixed during the whole training process or, on the other hand, have a varying
value following an adaptive scheme (Haykin, 1998). In our case, we have chosen a simple
adaptive learning rate which decays with the number of epochs according to the following
expression:
η
ηhid ← , (2.32)
1 + j/Jη
where η is the initial learning rate, j is the number of epochs, and Jη is a parameter to be
adjusted depending on the problem and the preferences for the training. However, as the
weights’ variation within each epoch has a random component, there exists the possibility
that the final updated weights after an epoch result in a network with a higher level of
errors than in the previous epoch. To overcome it, we calculate the root Mean Square
Error (MSE) of all the normalized patterns used in the learning process in each epoch.
v v
u P u P
uX uX
u
u |ep |2 u
u (z̃ p − ỹ p )2
t p=1 t p=1
e= = (2.33)
P P
If this error is higher than the error of the previous epoch, we keep the weights from the
previous epoch to initiate the following one.
To finally complete the learning process and have the NN trained, there should be
a stopping criterion. In this case, how to choose this criterion is greatly dependent on
the problem and the needs, and that choice is also crucial for an effective training, in
terms of accuracy and time. However, one must always set a maximum number of epochs.
This criterion is usually based on the aforementioned root MSE, using the denormalized
outputs to have more insight into the real physical problem in this case.
Therefore, the purpose of the second step of CAD methodology is reproducing the
response of the antenna to a given set of values of the design parameters immediately,
since this fact can be used to have a large search space (in theory, as large as desired) with
enough points to find the global maximum/minimum through the application of a global
optimization algorithm. Note that without the inclusion of this second step the search
space of the global optimization algorithm would be constrained to the discrete sets of
values (P points) given directly by the first step, which would require an increment in the
number of FE simulations in order to have a large enough search space.
2.5 Global optimization 19
2.5 Global optimization

Once we have trained the neural network, we are able to reproduce the response y = z of
the antenna to a given set of values xi , i = 1, . . . , I of the design parameters immediately
by de-normalizing as z = ȳ + σy z̃. This fact can be used to apply a global optimization
algorithm, since we are indeed able to have a large search space (in theory, as large as
desired) with enough points to find the global maximum/minimum. In our case, three
techniques have been implemented to carry out this step: a classic genetic algorithm and
a micro-genetic algorithm, and a simulated annealing scheme.
2.5.1 Genetic algorithms

GAs (Goldberg, 1989) are intended to perform the optimization by means of reproducing
the evolution via the survival of the fittest as Darwin stated. The GA implementation
used in the proposed CAD methodology can be found in (Carroll, 1999). Next, we briefly
outline the main features used in our GA implementation.
Encoding: binary encoding is the most typical and will be used here. According to
the possibilities for the values of each parameter in the search space, a specified number
of bits will be required. To make the algorithm more effective, especially with the micro
GA variant (Krishnakumar, 1990), which will be described later, it is recommended to
use a number of possibilities equal to a power of 2 for each parameter. Hence, each bit
of the string will be referred to as a chromosome. A set of chromosomes representing
the values of the parameters will form an individual. Finally, the number of individuals
chosen used in every iteration of the algorithm, denominated generation, constitutes the
population. The fitness of each individual is given by a transformation of the output z
of the neural network when the values of the input parameters are extracted from that
individual’s chromosomes. The purpose of the GA is to eventually form a population with
the fittest individuals.
Elitism: elitism is performed whenever the fittest individual of the previous gener-
ation is kept among the population of the current generation. Unless otherwise stated,
elitism will always be performed in our GA implementations.
Niching: also called sharing is the process of sharing genetic material between closely
related individuals (Goldberg and Richardson, 1987). In our methodology, niching is
always used.
Selection: among the population of each generation, the GA must select a number
of individuals to produce offspring. To make such a selection, there are several schemes
such as roulette wheel or tournament. In our case, we use tournament selection, where
a number of individuals is chosen at random and the fittest is kept as one of the parents
to perform a crossover with a new selected individual. Thus, the aim is that the fittest
individuals create fitter offspring for the next generation.
Crossover : every pair of individuals selected for reproduction produces just 1 new
individual to be incorporated into the new generation. The crossover can follow either a
single-point or a uniform approach (Carroll, 1999). For single-point crossover the chains
of chromosomes of the parents are cut at some random point and the resulting sub-chains
of chromosomes are swapped to form two new individuals. In uniform crossover, each
chromosome of the offspring is selected at random from the corresponding chromosomes
of the parents, producing a new individual. Crossover occurs with a probability that has
been set to a value of 0.5.
Mutation: After a crossover, each bit of the new individual may suffer from a jump
or creep mutation (Goldberg, 1989) according to the specified probabilities Pjump and
Pcreep . There are two types of mutation, with their corresponding probabilities: jump
mutation and creep mutation. In the former, a chromosome (bit) switches its value with a
probability Pjump . For the latter, two chromosomes in the individual swap their position
with a probability Pcreep . With these mutation techniques, the GA introduces a random
component which is claimed to make the optimization process avoid getting stuck in local
minima/maxima and keep on searching for the global minimum/maximum. Typical values
to set these probabilities can be found in (Goldberg, 1989; Carroll, 1999).
Micro Genetic Algorithm: The micro genetic algorithm (Krishnakumar, 1990; Car-
roll, 1999) is a very small population GA. Since enough diversity is introduced, mutations
in this variant are not applied.
2.5.2 Simulated annealing

An adaptive version of the SA algorithm described in (Otten and van Ginneken, 1989) and
used in (Monge et al., 2005, 2006; Gil et al., 2006) to perform several circuit and antenna
designs, has also been incorporated into the methodology as a means of having another
contrasted technique to perform the final optimization.
The concept is based on the manner in which liquids freeze or metals recrystal-
lize in the process of annealing. The algorithm finds the minimum by moving from one
point of the search space to another with some probability if the evaluation of a desired
cost function results in a higher value. This probability decreases following a Boltzmann
scheme, based on the decrease in the temperature parameter. The initial value of this
parameter, the rate of its decrease and the number of iterations are evaluated dynamically
by collecting the statistical data of the problem.
2.6 Application: broadband stacked circular patches for DCS

and UMTS bands
To validate our methodology, we have carried out the design of a broadband antenna
covering both the DCS and UMTS bands, that is to say, from 1710 MHz to 2170 MHz.
The geometry of this antenna is shown in Figure 2.11, and consists of two stacked circular
patches embedded in a circular cylindrical cavity. The lower patch lies on a substrate and
it is excited via the inner conductor of a coaxial connector situated at an offset xc from
the center. The upper patch is coupled through air, since it is printed at the bottom of a
superstrate which in turn lies on top of the cavity, whose profile has been lowered to that
effect. This geometry was intentionally conceived to facilitate its physical implementation.
A good starting point for a similar design (with no superstrate) can be found in
(Waterhouse, 1999). It is based on how each antenna parameter affects the impedance
behaviour, given by a loop on the Smith chart due to the presence of two resonances
corresponding to the two stacked patches. Following these rules, we have chosen the
parameters involved in the design to be the radii of the lower and upper patches R1
and R2 , and the distance between both patches d2 , since the combination of these three
parameters is enough to have a wide range in controlling the input impedance in the
desired band. In addition, the values for the substrate permittivity and thickness have
been adjusted to the commercial ones closest to those given in (Waterhouse, 1999). Given
Figure 2.11: Geometry and parameters of the proposed antenna
the fact that these initial values do not account for the presence of a superstrate, used
intentionally in our design for a matter of ease in the construction of the antenna, as well
as a possible cover, we have chosen the values of the permittivity and thickness of this
superstrate to be as small as possible at the considered frequencies, taking into account
mechanical issues, within the commercial range of values. The coaxial feed has been chosen
to be a 50-Ω N-type connector. In this way, we have set the parameters of the antenna
in Figure 2.11 to the ones shown in Table 2.1, where λ0 is the free-space wavelength and
√
λg = λ0 / ǫr1 .
2.6.1 Finite element simulations

The structure of Figure 2.11 is segmented into two regions, which are separated by a
circular port, as shown in Figure 2.11, where a modal expansion with 75 modes has been
used. One region contains the coaxial port, the substrate, the lower patch and part of the
cavity, and the other one comprises the rest of the cavity, the upper patch, the superstrate
and a semi-spherical port to model the radiating characteristics, since the cavity is recessed
in an infinite metallic plane. Thus, the distance between patches can be easily varied by
the analytical procedure of connecting both regions through a circular waveguide of the
desired length.
When calculating the MAM of each region, we have used 10 pairs of concentric
artificial ports in each patch surface, which means that 11 different values, ranging uni-
formly (in steps of 0.15 cm) from 2.5 cm to 4.0 cm, for each radius R1 , R2 will be available
according to different combinations of artificial port connections and short-circuits. All
dimensions and dielectric permittivities shown in Table 2.1 for Figure 2.11 are considered
in the simulation. The time used by this simulation (with an old version of a FE solver)
amounts to approximately half an hour per point of frequency in the computer used for
Table 2.1: Values of the parameters of antenna in Figure 2.11
Parameter Value (cm) Idea according to (Waterhouse, 1999)

ǫr1 2.5 2.2
ǫr2 1 1.07
ǫr3 2.17 -
ǫrc 1.8998 -
R1 to be designed 0.25λg at central frequency = 2.61 cm
R2 to be designed 0.25λg at central frequency = 2.61 cm
RC 4.5 -
Rs 4.7 -
d1 0.635 0.04λ0 at central frequency = 0.619 cm
d2 to be designed 0.06λ0 at central frequency = 0.928 cm
d3 0.1524 -
dc 0.3 -
di 0.31 -
do 0.98 -
xc 1.9 to yield 250Ω at resonance in the
lower edge of the band w/o upper patch
this design (Pentium IV @ 3.2 GHz with 2 GB of RAM). Note that the total time can be
considerably reduced since we have chosen a frequency-by-frequency scheme covering the
desired band and thereabouts, instead of making use of the reduced-oder model (Padé via
Lanczos).
In order to complete the training set, we have chosen to vary the distance between
patches d2 in steps of 0.125 cm from 0.75 to 2.00 cm, which provides 11 different values.
The time used to carry out the necessary matrix manipulations (artificial port interconnec-
tions and short-circuits, as well as modal interconnections) to make up the whole training
set totals to approximately 3 minutes per point of frequency. Note that our training set
consists of 1331 patterns per frequency, which with a classic FE approach would require
such a number of full-wave simulations, thus resulting in a prohibitive time. In our case,
only 2 FE numerical simulations have been carried out (one per region), and the rest of
the computational effort is reduced to 11 artificial port matrix manipulations per region
and their corresponding modal interconnections.
2.6.2 Neural network training

The output of the neural network has been chosen to be the magnitude of the reflection
coefficient |S11 | directly. We have placed 45 hidden neurons in the network, and set up
typical values a = 1 and b = 2 in (2.28)-(2.31), and the initial learning rate η to 0.3 and
Jη = 300 in (2.32). The neural network has been trained for each point of frequency
analyzed in the MAM simulation. An automatic stop criterion has been defined when
the root MSE in |S11 | changes in less than 0.01 for 500 consecutive epochs. Otherwise
the training continues until reaching a maximum of 10000 epochs. The neural network
training takes an average number of 1900 epochs per point of frequency. In the computer
used for this training (Pentium 2.8 GHz), this corresponds to less than 20 min for the
whole frequency band.
2.6.3 Optimization phase

Once the neural network has been trained, we proceed to the optimization phase. In this
case, we set the possibilities for each radius R1 , R2 to 128 (from 2.49 cm to 3.76 cm in
steps of 0.01 cm) and the variations for the distance between patches d2 also to 128 (from
0.75 cm to 2.02 cm in steps of 0.01 cm). This makes a search space of 1283 = 2097152
possibilities.
For the GA and micro-GA the Fitness Function (FF) has been chosen as the max-
imum value of the reflection coefficient in the desired band, maxf |S11 (f )|. Analogously,
for the SA the cost function has been set to 1 − maxf |S11 (f )|. For the classic GA, we
have chosen a population size of 50 individuals for a maximum number of generations of
51, with uniform crossover and we have set typical values Pjump = 0.02 and Pcreep = 0.14.
For the micro-GA we have used a population size of 5 individuals for a maximum number
of generations of 200 with single-point crossover.
The results for the three global optimization algorithms are shown in Table 2.2. The
similarity of the solutions reached by all of them can be seen, especially between the classic
GA and SA. One has to take into account that SA has a sophisticated automatic stop
criterion, which makes it require many more evaluations of the cost function than the GA
or micro-GA do, since they both stop simply when a maximum number of iterations is
reached. In any case, since the evaluation of the cost function for any combination of the
design parameters just requires their introduction in the neural network, whose output is
Table 2.2: Results for the different optimization algorithms
SA GA micro-GA
R1 (cm.) 2.88 2.88 2.88
R2 (cm.) 2.84 2.84 2.80
d2 (cm.) 1.49 1.47 1.48
maxf |S11 (f )| 0.2964 0.3057 0.3086
Evaluations 13723 2550 1000
instantaneously obtained, the time used in the optimization can be said to be ridiculous
(just a few seconds in the case of SA, less than one second for the GAs) in comparison
with the two previous phases of the proposed CAD methodology. Thus, we can state that
the first idea of making the computational effort in the beginning has been accomplished.
2.6.4 Final design

We have constructed the design given by the classic GA using a finite circular metallic
plane with 36 cm of diameter. Figure 2.12 shows a photo of the antenna where the
superstrate has been removed from its original position in order to see the lower patch
and the cavity. The upper patch is the one shown at the bottom of the figure.
Figure 2.13 shows the measured VSWR of the antenna against the results given by
the neural network, both being in agreement. As can be seen, a 26% relative bandwidth
at VSWR=2 covering the desired bands has been achieved. Figs. 2.14 and 2.15 show
respectively the E- and H-plane pattern at different frequencies covering the desired band
computed from a full-wave simulation of the design. In the E-plane, the crosspolar is
negligible and it can be seen how the pattern is not symmetrical due to the offset in the
feed. In the H-plane, the pattern is symmetrical and there is a crosspolar component
which becomes more important as the frequency is increased. In both planes, the pattern
shows a good uniformity throughout the frequency band.
2.6.5 Other designs

As has been previously said, once the network is trained for this specified topology, the
optimization phase is just a matter of one second for the classic GA. Thus, changing the
fitness function so that it covers just the UMTS band, we have obtained another design
for this band (1900 to 2170 MHz), whose design values are R1 =2.74 cm, R2 =2.49 cm
and d2 =1.86 cm. The return loss according to the neural network of this last design can
be seen against the design covering both bands in Figure 2.16, which shows how a better
matching can be achieved by reducing the band, moving from a design of approximately
26% of relative bandwidth at -10 dB to a one with approximately 13% of relative bandwidth
at -20 dB.
Figure 2.12: Photograph of the constructed and measured antenna. The superstrate has been
removed and the upper patch is the one shown at the bottom.
Figure 2.13: Measured and computed VSWR for the designed antenna showing a 26% relative
bandwidth at VSWR=2.
Figure 2.14: Computed E-plane pattern at different frequencies. ECP : copolar component.
Figure 2.15: Computed H-plane pattern at different frequencies. ECP : copolar component.
EXP : crosspolar component.
Figure 2.16: Return loss for the two different designs carried out with the methodology according
to the topology of Figure 2.11.
Chapter 3
Generalized-Scattering-Matrix
analysis of antenna arrays
3.1 Introduction
The use of entire-domain basis functions in the full-wave analysis of finite arrays is well-
known since the end of the 1960s when Mailloux studied two collinear coupled slots by
means of a finite expansion of waveguide modes (Mailloux, 1969). Since then, many
results of external mutual coupling, mostly based on waveguide mode expansions, have
been presented for different types of apertures on a ground plane, see, e.g., (Bird, 2004)
and references therein. The efficient analysis of finite arrays of more complex elements
was tackled a few years ago by using the so called Macro Basis, Synthetic Basis and
Characteristic Basis functions, all of which require a numerical generation process for
each array structure (Suter and Mosig, 2000), (Matekovits et al., 2003), (Prakash and
Mittra, 2003).
Recently, the use of spherical modes on a hemispherical port in each element as
entire-domain functions for the analysis of the external mutual coupling in finite arrays
has been proposed (Rubio et al., 2003, 2005). When spherical modes are used, it is possible
to study a finite array of antennas or scatterers including the coupling effects, by using the
rotation and translation theorems for spherical modes. Translational addition theorems for
spherical modes were introduced by Stein (Stein, 1961) and Cruzan (Cruzan, 1962), and
they have been widely applied to different problems such as the scattering from multiple
conducting (Bruning and Lo, 1971) or dielectric spheres (Hamid et al., 1990), or mea-
surements with probe correction in spherical near-field antenna measurements (Hansen,
1988).
In (Rubio et al., 2003), the reflection and coupling characteristics of two-element
arrays of cavity-backed microstrip antennas are obtained with a procedure which consists
of two steps. First, the GSM of each single antenna, in terms of spherical modes and
feed modes, is obtained from the hybrid FE-MA method described in Chapter 2. Second,
the reflection and coupling characteristics of a two-element array are obtained by using
the matrix expressions given in (Hansen, 1988) (pp. 85-86) for this particular case. In
this chapter, this full-wave procedure is extended in order to completely characterize finite
arrays of arbitrary size.
Although the method can be applied only to finite arrays whose elements’ radiated
fields can be individually described in terms of spherical modes, such as arbitrarily shaped
29
30 Generalized-Scattering-Matrix analysis of antenna arrays
apertures, horns, monopoles, Dielectric Resonator Antennas (DRAs) or cavity-backed

patch antennas, it offers several advantages. Firstly, spherical modes are well-known
analytical basis functions, and the maximum number of functions required to characterize
a radiating element of moderate size is low and can be easily determined (Hansen, 1988),
(Jensen and Frandsen, 2004). Secondly, the general translation matrix which allows the
interconnection of hemispherical ports is obtained analytically, so that the fill-in of the
matrix system can be achieved very fast. Thirdly, it provides the reflection and mutual
coupling parameters directly, and the radiation pattern computation requires a very sim-
ple analytical post processing (Rubio et al., 2005). Finally, as pointed out in (Mongiardo
et al., 2007), it allows the rigorous derivation of an equivalent network for the finite ar-
ray and it provides an exact boundary condition for numerical methods based on space
discretization.
3.2 GSM analysis of an array from rotation and translation

of spherical waves
In this section, the overall GSM of an antenna array is analytically calculated, by us-
ing rotation and translation of spherical modes, starting from the GSMs of the isolated
antennas, as shown in the scheme of the method depicted in Figure 3.1. The procedure
provides, in terms of a multimode scattering matrix, and in a rigorous way, the impedance,
coupling, scattering and radiation characteristics of the aforementioned array.
As discussed in (Hansen, 1988), the GSM of an antenna i of the array, also given in
(2.16), is best described as:

[ρi ] [ri ] {vinc
i } {wi }
= (3.1)
[ti ] [si ] − [Ii ] ai {bi }
where {bi } = bout
inc
i − ai are the complex amplitudes of the spherical
modes in which
the field scattered by the antenna is expanded, while ([si ] − [Ii ]) ainc

i is the part of the
received waves scattered directly off from the antenna, with [Ii ] being the identity matrix.
When N antennas are considered together, the incident field in each antenna i can
be considered as the superposition of the direct field coming from outside the array and
the contribution of the field scattered by the other antennas translated to the position of
antenna i. If these fields are expressed in terms of incoming spherical modes with weighting
complex amplitudes, the same relationship can be established for these coefficients as:
N
X n o
ainc ahi

i = {ai } + (3.2)
h=1
h 6= i
Each one of the column vectors ahi is obtained from the field scattered by each

antenna h given in terms of the complex amplitudes associated to scattered spherical

modes, {bh }, by means of the following relationship:
n o
ahi = [Gi,h ] {bh } (3.3)
where [Gi,h ] is known as the General Translation Matrix (GTM). This matrix is obtained
by using rotation and translation of spherical modes and it includes, in the most general
3.2 GSM analysis of an array from rotation and translation of spherical waves31
Figure 3.1: Scheme of the hybrid methodology used to obtain the GSM of the whole array from
the GSM(s) of the individual antenna element(s) considered as isolated.
case, rotations, axial translations and inverse rotations of spherical waves. Substituting
(3.3) in (3.2) we get:
XN
inc
ai = {ai } + [Gi,h ] {bh } (3.4)
h=1
h 6= i
Now, by using (3.4) we can express (3.1) for each antenna i as follows:
N
X
[ρi ] {vi } + [ri ] {ai } + [ri ] [Gi,h ] {bh }={wi } (3.5)
h=1
h 6= i
N
X
[ti ] {vi } + ([si ] − [Ii ]) {ai } + ([si ] − [Ii ]) [Gi,h ] {bh }={bi } (3.6)
h=1
h 6= i
If we repeat these equations for each one of the N antennas and we arrange them,
the following matrix equations are found for the array (Rubio et al., 2005):
[ρ] {v} + [r] {a} + [r] [G] {b}={w} (3.7)

[t] {v} + ([s] − [I]) {a} + ([s] − [I]) [G] {b}={b} (3.8)
where [ρ], [r], [t] and ([s] − [I]) are diagonal block-matrices,
[ρ]=diag([ρi ]) (3.9)
[r]=diag([ri ]) (3.10)
[t]=diag([ti ]) (3.11)
[s] − [I]=diag([si ] − [Ii ]) (3.12)
and [G], {v},{w},{a}and {b}are given by

 
0 [G1,2 ] · · · ··· [G1,N ]
 ..  ..
 [G2,1 ] 0 . [Gi,h ]  .
. ..
 
[G] = 
 .. . .. . .. .. 
(3.13)
.  .
..
 
..
.
 
 . [Gh,i ] 0 [GN−1,N ] 
[GN,1 ] ··· ··· [GN,N−1 ] 0
       

 {v1 }  
 {w1 } 
 
 {a1 }  
 {b1 } 

{v } {w2 } {a } {b2 }
       
2 2

 
 
 
 
 
 
 

 ..  ..  ..  ..

 
 
 
 
 
 
 

       
. . . .
   
{v} = ,{w} = ,{a} = ,{b} = (3.14)

 {vi }  
 {wi } 
 
 {ai }  
 {bi } 


 .
.

 
 .. 
 
 .
.

 
 .. 

. . . .

 
 
 
 
 
 
 


 
 
 
 
 
 
 

{vN } {wN } {aN } {bN }
       
3.3 Multimode Multiport network characterization of an antenna array 33
In order to obtain the overall GSM of the array, which includes the coupling effects,
we will use (3.7) and (3.8), as well as the definition of the overall GSM given by:

[Γ] [R] {v} {w}
= (3.15)
[T] ([S] − [I]) {a} {b}
First, we solve the implicit equation (3.8) to find {b}:
[[I] − ([s] − [I]) [G]]−1 [t] {v} + [[I] − ([s] − [I]) [G]]−1 ([s] − [I]) {a} = {b} (3.16)
and then we insert (3.16) into (3.7) to obtain

[ρ] + [r] [G] [[I] − ([s] − [I]) [G]]−1 [t] {v} +

+ [r] + [r] [G] [[I] − ([s] − [I]) [G]]−1 ([s] − [I]) {a} = {w} (3.17)
Equations (3.16) and (3.17) provide the overall GSM as can be seen by comparing
them with the definition in (3.15). Therefore, each one of the matrices which make up the
overall GSM is given by:
[Γ]=[ρ] + [r] [G] [[I] − ([s] − [I]) [G]]−1 [t] (3.18)

−1
[R]=[r] + [r] [G] [[I] − ([s] − [I]) [G]] ([s] − [I]) (3.19)
−1
[T]=[[I] − ([s] − [I]) [G]] [t] (3.20)
−1
([S] − [I])=[[I] − ([s] − [I]) [G]] ([s] − [I]) (3.21)
As can be seen, in order to obtain the overall GSM it is only necessary to invert
one matrix. This is in general a full matrix of dimension given by the total number of
spherical modes used to describe the tangential field on the spherical ports. If the antennas
in the array are not very close, only the propagating modes of each antenna need to be
retained. For moderate-sized arrays, it is possible to carry out a fast direct evaluation
of the system given by (3.18)-(3.21). For very large arrays, this system can be efficiently
solved by using the Compressed Block Decomposition Solver (Heldring et al., 2007) as
shown in (Rubio et al., 2007). Moreover, for regularly spaced arrays of identical elements,
a complex Block-Toeplitz linear system solver can be used to save CPU time and memory
requirements (Rubio et al., 2004a).
3.3 Multimode Multiport network characterization of an

antenna array
Once the overall GSM (3.15) has been obtained, an antenna array made up of N elements
can be viewed as the multiport network shown in Figure 3.2. It has N feeding ports, each
of them with its corresponding incident {vi } and reflected {wi } excitation modes, and
N radiating ports, each of them with its corresponding incoming {ai } and scattered {bi }
radiating modes.
Assuming no external incident field on the array (the incident field on each antenna
comes exclusively from the other antennas in the array), so that {a} = {0}, the trans-
mission matrix [T] of the GSM of the array relates complex coefficients of incident modes
Radiating region
{a 1} {b 1} {a i} {b i} {a N} {b N}
Multiport Network characterized by

a Generalized Scattering Matrix
{v 1} {w 1} {v i} {w i} {v N} {w N}
Feeding ports of the antennas in the array
Figure 3.2: Multiport network characterization of an antenna array.
{v} = h{v1 }T , . . . , {vN }T iT on the excitation ports of all the elements of the array with
scattered spherical modes {b} = h{b1 }T , . . . , {bN }T iT on the radiating ports of all the
elements of the array, as {b} = [T] {v}. This matrix rigorously provides the radiating
characteristics for any selected excitations including both the element pattern correspond-
ing to every antenna and all the interelement mutual coupling effects. The radiated field
of a planar array is then obtained as a superposition of shifted spherical modes, weighted
(m)
with their complex amplitudes bi , computed from {b} = [T] {v} by imposing a value
for the excitation modes, in the following way:
N Mi
!
(m)
X X
~ ϕ) =
E(θ, bi ~em (θ, ϕ) exp jk sin θ(xi cos ϕ + yi sin ϕ) (3.22)
i=1 m=1
where (xi , yi ) are the coordinates of each antenna in the array, considering a planar array
of N elements in the x − y plane; k is the free-space wavenumber; and ~em (θ, ϕ) are the
spherical mode far field vectors.
Analogously, in the case where there is no incident field coming from outside the
array, the overall reflection matrix of the array [Γ] provides the matching characteristics
for any selected excitations including the mutual coupling effects through {w} = [Γ]{v},
where {w} = h{w1 }T , . . . , {wN }T iT . This information can be used, for example, to set
the specifications of a feeding network which achieves the desired distribution of incident
modes {v} by taking into account the value of the reflected modes {w}.
Finally, it can be said that excitation modes {v} correspond to the modes of the
feeding ports of the antennas in the array, such as the TEM mode in a coaxial line, or the
fundamental modes of circular or rectangular waveguides.
3.4 Numerical validation and examples 35
Figure 3.3: Coupling coefficient magnitude between the fundamental modes of two WR90 rectan-
gular apertures on an infinite ground plane versus the aperture rotation angle α, for two aperture
separations at 9.2 GHz.
3.4 Numerical validation and examples

3.4.1 Comparison with measured results
As a first example, we consider a two-element array of standard WR90 rectangular waveg-
uide apertures with different orientations as studied in (Bird and Bateman, 1994). The
array is analyzed for a range of relative rotation angles α and two different aperture sepa-
rations. The results in Figure 3.3 show the magnitude of the coupling coefficient between
the fundamental modes in the apertures at 9.2 GHz compared with measured data given
in (Bird and Bateman, 1994). A very good agreement is observed between experimental
and numerical results for both cases.
Next, the method is verified with a more complex structure. A two-element array
of hemispherical DRAs situated on a ground plane and fed by coaxial probes, reported in
(Luk et al., 1994), is considered. The geometry and dimensions of the array element are
detailed in Figure 3.4. A different theoretical and measured resonant frequency, 3.68 GHz
and 3.84 GHz respectively, were obtained and used to compute the mutual coupling in the
reference. This fact is mainly a result of the different probe radius used in the experiments,
0.75 mm, because of practical difficulties in drilling the holes for the probes. In this work a
50-ohm coaxial feed with the theoretical inner radius of 0.5 mm and outer radius of 1.505
Figure 3.4: Geometry of the hemispherical dielectric resonator antenna used as an array antenna
element. R = 12.7 mm, ǫr = 9.5, s = 6.4 mm, h = 6.5 mm. Coaxial feed (50-ohm): ri = 0.5 mm,
ro = 1.505 mm.
mm is used. Figure 3.5 shows the variation of the coupling coefficient magnitude with
the center-to-center separation d, between the array elements, in free-space wavelengths
at 3.68 GHz. Results for both E- and H-plane coupling configurations are presented and
compared with theoretical and measurements given in (Luk et al., 1994). Again, a good
agreement is observed.
3.4.2 Analysis of an array scanning performance

The performance characteristics of finite phased arrays made up of the broadband cavity-
backed U-slot microstrip antenna designed in Chapter 2 (corresponding to Figures 2.3 and
2.6) are analyzed. Cavity enclosed microstrip antennas have been proposed to prevent
surface wave excitation providing advantages with respect to the conventional configura-
tion in continuous dielectric layers. Mutual Coupling (MC) between elements owing to
the surface waves may affect the scanning performance by deteriorating the impedance
matching, giving rise to scan blindness effect in large arrays. In phased-array antennas
this geometry provides larger scan volumes improving the impedance matching (Zavosh
and Aberle, 1994).
In our case, phased arrays of U-slot and probe-fed microstrip patch antennas enclosed
in metallic cavities are analyzed. Investigation into this configuration is of interest since
the U-slot microstrip antenna is known to have a broadband characteristic. Furthermore,
enclosing this antenna in a cavity will give the added advantage of eliminating the surface
waves. The input impedance, radiating characteristics and scan performance of finite
phased arrays of these antennas will be analyzed. Arrays with a rectangular grid and
with the same interelement spacing and number of antennas in the x and y directions are
considered (see array scheme in Figure 3.6). In order to analyze the scan performance, the
E-plane broadside-matched active reflection coefficient magnitude versus scan angle of the
central element of different sized arrays at 3.2 GHz (corresponding to the second resonant
frequency according to Figure 2.6) has been obtained. Two different element spacings
d have been considered. Figure 3.6(a) shows the results for d = 0.5λ0 at the considered
frequency and for different sized arrays. It is observed how as the size of the array increases
the curves show more ripples converging to the smooth curve obtained from the infinite-
array model that has been computed with the procedure developed in (González et al.,
2000). As for conventional patches placed between metal walls (Zavosh and Aberle, 1994),
the cavity enclosure of U-slot patches provides an improvement in scan performance. For
the same patch arrays with a conventional continuous substrate, a substantial reduction in
Figure 3.5: E- and H-plane mutual coupling coefficient magnitude versus dielectric resonator
spacing at 3.68 GHz.
the scan coverage may be expected owing to the surface wave excitation. The lower order
surface wave propagation would yield a blind-spot at approximately 69◦ calculated from
the infinite array theory (Pozar and Schaubert, 1984). Figure 3.6(b) shows the E-plane
broadside-matched active reflection coefficient magnitude for the central element of the
same arrays considering an interelement spacing larger than a half wavelength, 0.6λ0 at 3.2
GHz. In this case, the simulations show a tendency towards a blind spot as the number of
elements of the array increases, whose value, 41.8◦ , predicted by the grating lobe theory,
coincides with the calculated with the infinite-array model (González et al., 2000).
Figure 3.7 shows the E-plane active element field pattern at 3.2 GHz of the central
element for an interelement spacing of 0.6λ0 . As previously the presence of a blind spot is
observed. The asymmetry of the patterns in this plane (independent of the interelement
spacing) is introduced by the U-slot in the patch, which in the case of the array of 0.6λ0
spacing renders the tendency to the blind angle much more pronounced in one semiplane
than in the other one. The E-plane cross-polarization component is negligible in all the
considered cases. Figure 3.8 shows the H-plane active element field pattern of the central
element for the case of 0.5λ0 spacing. Now the pattern is symmetric and there exists
a crosspolar component which becomes much more significant as the array is scanned
towards endfire.
The different behaviour of the input impedance with the element position in the array
and with the array size is illustrated in Figure 3.9. It shows the active reflection coefficient
magnitude, at 3.2 GHz in arrays of various sizes and d = 0.5λ0 , versus the element position
across (a) the H-plane (central row in the x-direction of the array scheme in Figure 3.6)
when the array is scanned at broadside, and across (b) the E-plane (central column in
Figure 3.6: E-plane broadside-matched active reflection coefficient magnitude |Γ| versus scan
angle at 3.2 GHz of the central element for arrays of various sizes with (a) d=0.5λ0 (b) d=0.6λ0 .
Figure 3.7: E-plane active element field pattern of the central element at 3.2 GHz for arrays of
various sizes with d=0.6λ0 .
Figure 3.8: H-plane active element field pattern of the central element at 3.2 GHz for arrays of
various sizes with d=0.5λ0 .
Figure 3.9: Computed active reflection coefficient magnitude |ρ| at 3.2 GHz for arrays of various
sizes with d=0.5λ0 for the elements across (a) the H-plane when the array is at broadside scan,
and across (b) the E-plane when the array is scanned at 70◦ in the E-plane.
the y-direction of the array scheme in Figure 3.6) when the array is scanned at 70◦ in
the E-plane. As it is observed, as the size of the array increases, the response gets more
uniform. The theory predicts the appearance of a blind spot at 69◦ in an infinite array on
a continuous dielectric layer when scanning in the E-plane (Pozar and Schaubert, 1984),
which for a finite array would mean a mismatch increment as the number of elements gets
higher. However, Figure 3.9(b) shows how for the cavity-backed patch array this effect
does not appear and, on the contrary, a substantial improvement is achieved.
Finally, in order to show the combination of broadband impedance and wide scan
coverage characteristics in the proposed phased array configuration we have examined
the active reflection coefficient magnitude of a 35×35 array at four different frequencies
covering the input-match frequency band, for three different scan angles. The results are
depicted in Figure 3.10 as a function of the element position across the E-plane when the
array is scanned at broadside (Figure 3.10(a)), and when it is scanned in the E-plane at
60◦ (Figure 3.10(b)); and as a function of the element position across the H-plane when
the array is scanned in the H-plane at 45◦ (Figure 3.10(c)). It is observed that the VSWR
for all the considered elements remains below 2 for the defined scan volume and bandwidth
except for some border elements. The curves of the broadside case are asymmetrical owing
to the U-slots in the patches but are quite uniform, mainly in the central elements, since a
large array is being considered. The results in Figure 3.10(b) show once again the improved
scan performance of the proposed structure with respect to the same U-slot patch array
on a continuous dielectric layer. In this case, owing to the surface wave excitation, the
infinite array theory (Pozar and Schaubert, 1984) predicts a blind spot at 69◦ and 57◦ in
the E-plane for 3.2 GHz and 3.35 GHz, respectively, which would lead to an undesirable
Figure 3.10: Computed active reflection coefficient magnitude |ρ| at different frequencies for
a 35×35 array with d=0.5λ0 at 3.2 GHz for the elements across the E-plane when the array is
scanned at (a) Broadside (b) 60◦ in the E-plane, and (c) for the elements across the H-plane when
the array is scanned at 45◦ in the H-plane.
mismatch if the considered finite array on a continuous dielectric layer were scanned at
60◦ in the E-plane at these frequencies.
As can be seen, the analyzed configuration provides a broadband impedance band-
width accompanied by a large scan volume preserving the attractive single-layer single-
patch characteristics of conventional patch arrays.
3.5 Inclusion of the feeding network effects: waveguide-fed

DRA array
In previous sections, the GSM formulation in terms of spherical modes in order to calculate
the external mutual coupling effects in a finite array has been developed. So far, internally
uncoupled excitation ports in the finite array fed by matched generators have been assumed
as a starting point. However, the internal mutual coupling and the mismatching caused by
the feeding network contribute to the reradiated field. Therefore the effects of the internal
coupling cannot be neglected in many array designs such as waveguide-fed arrays. In this
section, it is shown how the assumption of internally unconnected excitation ports fed by
matched generators can be overcome, so the GSM formulation is extended to consider the
effects of the feeding network in the calculation of the external mutual coupling in finite
arrays.
The capabilities of the proposed formulation will be shown by analyzing a 10-element
waveguide-fed array of stacked DRAs, depicted in Figure 3.11, for two different inter-
element spacings. A design procedure for this kind of arrays has been provided in (Eshrah
et al., 2006), whose results will be used here for comparison purposes. A Taylor’s synthesis
method was there used to determine the relative excitation in the coaxial ports, achieved
by changing the probe lengths in a standard X-band rectangular waveguide (22.86 mm ×
10.16 mm). The stacked DRA has a height, radius and dielectric constant of 4 mm, 5.5
mm and 4.1, respectively, for the lower DRA disc, and 3 mm, 4 mm and 12.3 for the upper
DRA disc. The probe has a radius of 0.25 mm, length inside the waveguide of 7.3 mm,
and offset from the waveguide centerline of 7 mm. In order to show the influence of the
mismatching and internal coupling effects because of the feeding network in the external
mutual coupling, results without the feeding network will also be supplied.
The first step in the methodology consists of computing the GSM of the finite array
including the feeding network and assuming that the stacked DRAs are externally uncou-
pled. Note that this fact implies the direct computation of matrices (3.9)-(3.12), which will
no longer be diagonal-block matrices but full matrices. To this aim, the full-wave domain-
decomposition technique reported in Chapter 2 will be used in the whole array. As shown
in Figure 3.11(b), the array is first segmented into different regions (types A, B or C) by
means of coaxial and rectangular ports. Then, the GSM of each region is obtained by using
modal expansions on each port. The connection of these GSMs will provide the GSM of the
array without taking into consideration the external coupling between the DRAs, whilst
the internal mutual coupling is taken into account with the analytical connection of the
GSMs of A-type regions through sections of rectangular waveguides (B-type regions). The
A-type regions correspond to waveguide-to-coax transitions. In this work the multimode
GSM of these regions has been obtained by using the aforementioned full-wave hybrid
FE-MA method, although, for this particular case, a circuit model which is more suitable
for design purposes is available (Eshrah et al., 2006). Each one of the C-type regions
corresponds to a stacked DRA delimited by a coaxial excitation port and a hemispherical
3.5 Inclusion of the feeding network effects: waveguide-fed DRA array 43
Figure 3.11: Finite array of stacked DRAs on an infinite ground plane fed with coax probes
through a rectangular waveguide. (a) Top view. (b) Longitudinal view including the feeding
network, the inner ports, the input ports, the coax excitation ports and the hemispherical ports.
Figure 3.12: Scattering parameters versus frequency with two interelement spacings compared
to the results in (Eshrah et al., 2006). (a) d=20 mm. (b) d=15 mm.
port to describe free-space propagation. The array is made up of ten identical probe-fed
stacked DRAs with alternating probe offsets. Thus, it suffices to compute just the GSM
of one isolated array element since element rotations are analytically incorporated in the
GSM formulation of the finite array (Rubio et al., 2005). It should be noted that any other
validated analytical or numerical method could be used to obtain the externally uncou-
pled GSM. Thus, e.g., an equivalent generalized network matrix can be quickly obtained
for waveguides radiating into a spherically stratified dielectric by using modal techniques
(Mongiardo et al., 2007). Once the initial GSM of the whole array has been computed
without taking into consideration the external coupling, equations (3.18) and (3.20) are
solved to obtain the overall reflection matrix and the overall transmission matrix of the
array. The radiated field of each stacked DRA has been characterized with a maximum
of 24 spherical modes at the highest frequencies of the analysis band. Therefore, in the
worst case, the dimension of the equation system is 240. The variation of the scattering
parameters with the frequency including the feeding network has been calculated for an
inter-element spacing d of 20 mm and 15 mm (Figure 3.12). A very good agreement with
the results obtained in (Eshrah et al., 2006) with a different method has been found. It has
been noted how the influence of the external mutual coupling is more noticeable for d=15
mm. From the overall transmission matrix, which takes into account the feeding network,
different radiation patterns have been obtained analytically. A very good agreement with
the reference is found again.
Figure 3.13(a) shows the radiation pattern for d=20 mm at 8 GHz. Figures 3.13(b)
3.5 Inclusion of the feeding network effects: waveguide-fed DRA array 45
Figure 3.13: Radiation patterns compared to (Eshrah et al., 2006). (a) d=20 mm, f =8 GHz.
(b) d=15 mm, f =8 GHz. (c) d=15 mm, f =10 GHz.
and 3.13(c) show, respectively, the radiation patterns at 8 GHz and at 10 GHz for d=15
mm. It can be seen how external mutual coupling changes the scanning direction slightly
and broadens the width of the main beam at 8 GHz. It also makes the directivity decrease
by 2.5 dB in the scanning direction at 10 GHz. The influence of the feeding network with
internally coupled and unmatched coaxial ports in the external mutual coupling can be ob-
served by comparing with the analysis without the feeding network. In this case the GSM
formulation based on uncoupled excitation ports fed by matched generators is directly
applied, so now matrices (3.9)-(3.12) are diagonal-block matrices. The radiation pattern
at each frequency is obtained by imposing in the coaxial ports the relative excitations
the feeding network provides. These excitations have been obtained before considering
the external mutual coupling and they produce the same externally uncoupled radiation
pattern than the whole array when the rectangular input port is fed. Results are given
at 8 GHz and at 10 GHz for d=15 mm. They are shown in Figs. 3.13(b) and 3.13(c)
for comparison purposes. It can be seen that the inclusion of the feeding network in the
calculation of the external coupling modifies the radiation patterns, especially at 8 GHz.
All calculations were carried out on a laptop computer with a 2.20 GHz Intel Core2
Duo CPU T7500 processor and 3-GB RAM. The C-type region was analyzed in 14 seconds
per frequency and each one of the A-type regions required 5 seconds per frequency without
using fast frequency-sweep techniques. These multimode problems have been solved with
an efficient sparse direct solver, which is more suitable for small-sized problems with
multiple right-hand-side vectors. The connection of all inner ports to obtain the externally
uncoupled GSM took 0.37 seconds per frequency. Finally, the computation of the external
mutual coupling taking into account the feeding network was carried out in only 0.2 seconds
per frequency, including the fill-in of the matrix system and the calculation of the overall
reflection and transmission matrices.
Chapter 4
Multiobjective optimization of an
array through an IPM with
filtering
4.1 Introduction
Since the early beginnings of antenna arrays, most of the research effort has focused on
finding the optimum excitations for the elements in the array which maximize or minimize
a global array characteristic, i.e. the directive gain or the power gain. A more involved
problem arises when we want to establish a certain control over other array characteristics,
that is, we impose several restrictions to be fulfilled, such as nulls in certain directions or a
maximum Sidelobe Level (SLL) within a region. In optimization theory, that is equivalent
to adding constraints to the original problem of maximization/minimization. A very good
summary on optimization techniques for antenna array excitations, most of which are still
used or constitute the basis for new ones, can be found as early as 1971 in (Cheng, 1971).
Recent advances in the last decade focus on evolutionary computation, especially
genetic algorithms (Markus and Vaskelainen, 1998; Yan and Lu, 1997; Shimizu, 1994;
Ares-Pena et al., 1999; Moreno et al., 2005; Haupt, 2007). One of their major advantages
is their predisposition to incorporate results from numerical simulations or measurements
of antenna elements. However, despite their proven efficiency, they may be computation-
ally expensive for some problems and the specification of constraints (SLL, null pointing
directions) cannot usually be made explicitly thus requiring their inclusion in the cost
function (Ares-Pena et al., 1999; Moreno et al., 2005; Haupt, 2007).
More classic approaches range from the maximization of the ratio of two Hermitian
forms (Cheng, 1971; Voges and Butler, 1972) to the use of Lagrange multipliers (Kurth,
1974), Newton methods (Morini et al., 2006), quadratic programming (Einarsson, 1979;
Ng et al., 1993), biquadratic programming (Hirasawa, 1988) or L∞ optimal methods (Sh-
pak, 1996). Depending on the nature of the quantity to be maximized/minimized and the
constraints, the problem may become a constrained non-linear optimization one, thus re-
quiring the use of more involved methods for its resolution (Steykal et al., 1986; Jiao et al.,
1993; Lebret and Boyd, 1997). One of the main drawbacks of some of these techniques is
that they may require the explicit knowledge of the far-field derivatives with respect to
the excitations of the array.
In (Lebret and Boyd, 1997) a constrained non-linear optimization problem is solved
47
48 Multiobjective optimization of an array through an IPM with filtering
with the aid of an Interior-Point Method (IPM). The latter were introduced in 1984
by Karmarkar (Karmarkar, 1984) and were firstly intended to be a revolution in LP.
However, they were soon discovered to be based on a well-conditioned reformulation,
through a primal-dual approach, of the non-linear optimization barrier methods (Fiacco
and McCormick, 1968), unifying the, until then, separated theories of linear and non-
linear optimization (Gill et al., 1986). It was not until the 1990s when they began to
be widely used in non-linear optimization as an alternative to several problems presented
by sequential quadratic programming methods (Nocedal and Wright, 2000). IPMs have
been implemented using indistinctly one of the two classic approaches: linear search or
trust-region; however, they both present several drawbacks in constrained optimization,
such as finding a suitable merit function in the former. To avoid these problems, new
schemes have been proposed since the beginning of this century. One of them, initially
conceived for sequential quadratic programming, was introduced in (Fletcher and Leyffer,
1997) and is known as the filtering method; recent advances in its application to an IPM
can be found in (Benson et al., 2002; Ulbrich et al., 2004).
However, most of the developments regarding optimization in arrays have tradition-
ally been made by taking for granted several assumptions which can make the solutions
obtained using these procedures not yield the desired response in real designs. Among
these assumptions, one can highlight the use of isotropic sources, which are in fact not
physically possible, or very well-known analytical element patterns. Apart from that, cou-
pling between elements has seldom been taken into account. It has not been until recently
that optimization in arrays has been integrated with full-wave models and computed or
measured data (Moreno et al., 2005; Morini et al., 2006; Blank and Hutt, 2005).
In this chapter, we develop a methodology to optimize the excitations of an arbi-
trary array of coupled and real antenna elements to obtain the maximum gain in a desired
direction subject to restrictions on different characteristics of the array such as the max-
imum SLL in a specified region, the maximum crosspolar level in another desired region,
the aperture illumination efficiency or the Dynamic Range (DR) of the excitations, as well
as setting null pointing directions. A primal-dual IPM with a Newton step specifically
adapted to the derived formulation is implemented along with novel filtering techniques
to solve the resulting non-linear optimization problem efficiently.
4.2 Non-linear constrained optimization problem formula-

tion
4.2.1 Matrix formulation of the power and the radiated field
If we consider, without loss of generality, one single mode of excitation per antenna in the
array {vi } = vi , then we have:
N
(m) (m)
X
bi = Ti,h vh . (4.1)
h=1
(m)
where coefficientsTi,hare directly extracted from the elements of the overall transmission
matrix [T] and take into account the contribution of the excitation of antenna h to the
mth spherical mode of antenna i. Therefore, (3.22) can be rewritten as:
Mi X
N X N
~ ϕ) = (m)
X
E(θ, Ti,h vh~em (θ, ϕ) exp jk sin θ(xi cos ϕ + yi sin ϕ) (4.2)
i=1 m=1 h=1
4.2 Non-linear constrained optimization problem formulation 49
Finally, (4.2) can be easily expressed in a more useful matrix form as:
~ ϕ) = hζ(θ, ϕ)i[̥
E(θ, ~ (θ, ϕ)]{v} = h~g(θ, ϕ)i{v} (4.3)
where
 T
 exp (jk sin θ(x 1 cos ϕ + y 1 sin ϕ)) 
.
 
hζ(θ, ϕ)i= .
.
 
exp (jk sin θ(xN cos ϕ + yN sin ϕ))
 
 M M1

1
(m) (m)
X X
 T1,1 ~em (θ, ϕ) . . . T1,N ~em (θ, ϕ) 
 
 m=1 .. ..
m=1
..

~ (θ, ϕ) =

̥


 . . . 

MN
 X MN 
(m) (m)
X
TN,1 ~em (θ, ϕ) . . . TN,N ~em (θ, ϕ)
 
m=1 m=1
In case we are only considering the copolar component of the radiated field, then
the vector notation is dropped to have:
E(θ, ϕ) = hζ(θ, ϕ)i[̥(θ, ϕ)]{v} = hg(θ, ϕ)i{v} (4.4)
Note that in the case of dealing with an ideal array made up of isotropic sources,
the formulation is adjusted straightforwardly by setting in (4.3) hg(θ, ϕ)i = hζ(θ, ϕ)i, or
equivalently [̥(θ, ϕ)] = [I], with [I] being the identity matrix.
The total radiated power is computed as an Hermitian form from:
ZZ
Prad = ~ ϕ)|2 dΩ = 4π{v}H [BR ] {v}
|E(θ, (4.5)
Ω
Z 2π Z π
4π[BR ]= h~g(θ, ϕ)iH · h~g(θ, ϕ)i sin θdθdϕ (4.6)
ϕ=0 θ=0
where · stands for the dot product and the superscript H denotes the Hermitian transpose.
The delivered power to the array can be expressed from the reflection matrix of the
array as:
N
|vi |2 − |wi |2 = 4π{v}H [BD ] {v}
X
Pdel = (4.7)
i=1
4π[BD ]=[I] − [Γ]H [Γ] (4.8)
Finally, the total incident power in the array is easily written in matrix form as:
N
|vi |2 = 4π{v}H [BI ] {v}
X
Pinc = (4.9)
i=1
4π[BI ]=[I] (4.10)
4.2.2 Direct maximization of the gain

The field intensity can be computed as an Hermitian form in the following way:
|E(θ, ϕ)|2 ={v}H [G(θ, ϕ)]{v} (4.11)

H
[G(θ, ϕ)]=h~g(θ, ϕ)i · h~g(θ, ϕ)i (4.12)
The directive gain GD , power gain GP and system gain GS of an array whose main
beam points to the direction (θ0 , ϕ0 ) are defined as:
~ 0 , ϕ0 )|2
|E(θ
GD =4π (4.13a)
Prad
~
|E(θ0 , ϕ0 )|2
GP =4π (4.13b)
Pdel
~
|E(θ0 , ϕ0 )|2
GS =4π (4.13c)
Pinc
Placing (4.11), (4.5), (4.7) and (4.9) in (4.13), these parameters can be expressed as
a ratio of two Hermitian forms:
{v}H [G0 ] {v}
GD = (4.14a)
{v}H [BR ] {v}
{v}H [G0 ] {v}
GP = (4.14b)
{v}H [BD ] {v}
{v}H [G0 ] {v}
GS = (4.14c)
{v}H [BI ] {v}
where [G0 ] = [G(θ , ϕ )].
The maximum attainable value of a quantity G (standing for either GD , GP or
GS ) expressed as in (4.14), provided [B] (standing for [BR ], [BD ] or [BI ]) is nonsigular
and positive definite, which is our case, is the largest eigenvalue λL of the generalized
eigenvalue problem [G0 ] {v} = λ [B] {v}, and its associated eigenvector {v}L gives the
excitations in the array antenna elements to achieve this maximum value of G. Moreover,
if we consider only the copolar component of the field in the numerators of (4.14), we
would have:
[G0 ]=hg(θ , ϕ )iH · hg(θ , ϕ )i (4.15)
In this case, values for λL and {v}L can be computed straightforwardly (Cheng, 1971) as:
λL =hg(θ , ϕ )i [B]−1 hg(θ , ϕ )iH (4.16)
−1 H
{v}L =[B] hg(θ , ϕ )i (4.17)
This technique provides a direct means to maximize, with no additional constraints,
the directive or power gain of an arbitrary planar array of real antenna elements by
inherently taking into account all interelement coupling effects. However, a real array
design may have to comply with other restrictions, usually related to the maximum SLL
or crosspolar level, whose inclusion in a whole problem formulation is looked at in the
next subsection. Additionally, we will make use of two common array performance indices,
namely the aperture illumination efficiency and the DR, whose maximum and minimum
values (100% and 1) are reached with a uniform cophasal excitation, thus providing a
means of comparing a given excitation to this one. These parameters are respectively
defined as:
{v}H hζ(θ0 , ϕ0 )iH hζ(θ0 , ϕ0 )i {v}
εAP = (4.18)
N {v}H {v}
max | {v} |
DR= (4.19)
min | {v} |
4.2.3 Inclusion of constraints

We can formulate a generalized problem as follows: minimize the total radiated/delivered
power while keeping a constant value κ0 of the field in the direction of the main beam
(θ0 , ϕ0 ), as well as setting the field intensity to zero in the desired null pointing directions
(θi , ϕi ), i = 1, . . . , IN L ; making the field intensity over a grid of directions (θj , ϕj ), j =
1, . . . , JSL , dense enough to include all maxima in the sidelobe region, remain below the
prescribed maximum SLL; also making the crosspolar field intensity remain below a pre-
scribed maximum level over another grid of directions (θm , ϕm ), m = 1, . . . , MXP ; making
the aperture illumination efficiency remain higher than a desired value ε; and finally mak-
ing the absolute values of some desired NDR excitations not fall below certain specified
limit κR . We cannot, unlike in (Einarsson, 1979) (where a similar approach is followed),
assume a real-valued field in all space since we are on the one hand considering real ele-
ments with interelement coupling and on the other hand making no assumptions on the
geometry of the array and the main beam direction. This fact prevents the problem
from being of quadratic programming type, requiring a more involved algorithm for its
resolution. Thus, the general problem to be solved takes the form:
minimize with {v}
{v}H [B] {v} (4.20a)
subject to
hg(θ0 , ϕ0 )i {v} = κ0 (4.20b)
hg(θi , ϕi )i {v} = 0, i = 1, . . . , IN L (4.20c)
H (j)
{v} [ΠSL ] {v} ≥ 0, j = 1, . . . , JSL (4.20d)
(m)
{v}H [ΠXP ] {v} ≥ 0, m = 1, . . . , MXP (4.20e)
{v}H [ΠAE ] {v} ≥ 0 (4.20f)
(n)
{v}H [ΠDR ] {v} ≥ κR , n = 1, . . . , NDR (4.20g)
where κ0 is a complex-valued constant, κR is a real-valued positive constant and
h i
(j)
ΠSL =10SLLj /10 hg(θ0 , ϕ0 )iH hg(θ0 , ϕ0 )i − hg(θj , ϕj )iH hg(θj , ϕj )i (4.21a)
h i
(m)
ΠXP =10XP L/10 hg(θ0 , ϕ0 )iH hg(θ0 , ϕ0 )i − hgxp (θm , ϕm )iH hgxp (θm , ϕm )i (4.21b)
[ΠAE ] =hζ(θ0 , ϕ0 )iH hζ(θ0 , ϕ0 )i − N ε[I] (4.21c)
 h (n) i
h i  ΠDR = 1
(n)
ΠDR = h (n) ih,h (4.21d)
 Π = 0 ∀(i, j) 6= (h, h)
DR
i,j
where gxp stands for the crosspolar component of ~g, SLLj is the prescribed maximum
sidelobe level in dB in each direction θj , ϕj , XP L the prescribed maximum crosspolar
level, 0 ≤ ε ≤ 1 is the desired minimum aperture illumination efficiency and h stands for
the desired excitation over which a control is performed.
Formulation (4.20) presents a general problem involving all constraints considered in
this thesis, but a real problem may only require the fulfillment of some of these restrictions
or may instead need the specification of other constraints related to typical antenna array
parameters, whose usefulness greatly depends on the problem and design objectives. In
the presented formulation, the maximum SLL, the maximum crosspolar level and the min-
imum aperture illumination efficiency can be explicitly defined. The DR can be controlled
through the adjustment of constant κR , which constitutes a lower bound for every exci-
tation amplitude. The upper bound is implicitly established by the absolute value of the
field in the direction of the main beam, given by constant κ0 , which obviously depends on
the amplitude value of the excitations. This way, constants κR and κ0 should be properly
heuristically selected to avoid an incongruent situation.
4.2.4 Equivalent real-valued constrained optimization problem

In formulation (4.20), though the quantities of the objective function and the constraints
are usually real (since κ0 is usually set as real-valued for practical purposes), the search
space for {v} must be complex to reach a solution in the great majority of the cases.
Since most non-linear optimization procedures (Nocedal and Wright, 2000) only deal with
real-valued variables, to use one of these procedures we yield an equivalent formulation by
taking the real ℜ and imaginary ℑ parts of the excitations:
minx xT hx (4.22a)
s.t. a0R x = ℜ(κ0 ) (4.22b)
a0I x = ℑ(κ0 ) (4.22c)
(i)
aR x = 0, i = 1 . . . IN L (4.22d)
(i)
aI x = 0, i = 1 . . . IN L (4.22e)
T (j)
x bSL x ≥ 0, j = 1 . . . JSL (4.22f)
(m)
xT bXP x ≥ 0, m = 1 . . . MXP (4.22g)
xT bAE x ≥ 0 (4.22h)
(n)
xT bDR x ≥ κR , n = 1 . . . NDR (4.22i)
where (x being now the vector containing the optimization variables, which are double the
number of excitations):
x =hℜ({v}T ), ℑ({v}T )iT (4.23)

ℜ([B]) −ℑ([B])
h = (4.24)
ℑ([B]) ℜ([B])
a0R =hℜ(hg(θ0 , ϕ0 )i), −ℑ(hg(θ0 , ϕ0 )i)i (4.25)
a0I =hℑ(hg(θ0 , ϕ0 )i), ℜ(hg(θ0 , ϕ0 )i)i (4.26)
(i)
aR =hℜ(hg(θi , ϕi )i), −ℑ(hg(θi , ϕi )i)i (4.27)
(i)
aI =hℑ(hg(θi , ϕi )i), ℜ(hg(θi , ϕi )i)i (4.28)
" #
(j) (j)
(j) ℜ([Π SL ]) −ℑ([Π SL ])
bSL = (j) (j) (4.29)
ℑ([ΠSL ]) ℜ([ΠSL ])
" #
(m) (m)
(m) ℜ([Π XP ]) −ℑ([Π XP ])
bXP = (m) (m) (4.30)
ℑ([ΠXP ]) ℜ([ΠXP ])

ℜ([ΠAE ]) −ℑ([ΠAE ])
bAE = (4.31)
ℑ([ΠAE ]) ℜ([ΠAE ])
" # " #
(n) (n) (n)
(n) ℜ([ΠDR ]) −ℑ([ΠDR ]) [ΠDR ] [0]
bDR = (n) (n) = (n) (4.32)
ℑ([ΠDR ]) ℜ([ΠDR ]) [0] [ΠDR ]
where the superscript T denotes the transpose operation.
A general non-linear optimization problem is given in the form (Nocedal and Wright,
2000):
minx f (x) (4.33a)
s.t. cE (x) = 0 (4.33b)
cI (x) ≥ 0 (4.33c)
l≤x≤u (4.33d)
where f (x) is known as the objective function, cE (x) as the equality constraints, cI (x) as
the inequality constraints, and l and u represent bounds on the variables under optimiza-
tion x. The whole set of constraints represent the feasible region for x within which the
algorithm must reach a solution (minimizing the objective).
By inspection of problems (4.22) and (4.33), we can arrive at:
f (x) =xT hx (4.34a)


 a0R x − ℜ(κ0 )
 a0 x − ℑ(κ0 )

I
cE (x)= (i) (4.34b)
 R x,
 a i = 1 . . . IN L
 (i)

a x, i = 1 . . . IN L
 I (j)

 xT bSL x, j = 1 . . . JSL
 T (m)

x bXP x, m = 1 . . . MXP
cI (x) = Tb (4.34c)

 x AE x
 T (n)

x bDR x − κR , n = 1 . . . NDR
The presented formulation provides an immediate way to check if the problem is a
convex optimization problem, that is to say, if we are trying to minimize a convex function
over a convex set (defined by the constraints). There are two main reasons why we would
like our problem to be convex. Firstly, and most important, it is easy to show that
any local minimum of this type of problems is a global minimum (unique if the function
to minimize is strictly convex). Secondly, first-order necessary conditions of optimality
(Nocedal and Wright, 2000) (known as Karush-Kuhn-Tucker (KKT) conditions) are also
sufficient. An optimization problem is said to be convex if:
1. the objective function is convex, or equivalently, if its Hessian is positive-semidefinite.
This condition is always fulfilled in problem (4.34) because of the nature of the Hes-
sian, which is constant and equal to h.
2. the equality constraints are linear, which is true for problem (4.34)
3. inequality constraints are concave, or equivalently, if their Hessians are negative-
(j)
semidefinite. Since in problem (4.34) these Hessians are constant and equal to bSL ,
(m) (n)
bXP , bDR and bAE , we can know in advance if we are dealing with a convex problem
Constraint (4.33d) accounts for the imposition of lower and upper bounds on the
variables of a general non-linear optimization problem (most of which do have this type of
constraints). As can be seen, it is not present in the problem formulation with real-valued
variables (4.22) (where these bounds would apply to the real and imaginary parts of the
excitations) and thus it has no direct counterpart in the original problem formulation
(4.20). This is why throughout the development of the IPM constraint (4.33d) will be
treated separately. Anyway, setting these bounds may have a mathematical use in order
to constrain the search space in x, thus making the optimization algorithm search far away
from values which are nowhere near a solution.
4.3 Primal-dual Interior Point Method with Filtering

4.3.1 The logarithmic barrier method and the primal-dual approach
With the aid of the necessary slack variables s = (sTI , sTU , sTL )T , (s ≥ 0) to transform
the inequality constraints in (4.33) into equality constraints, we can apply a logarithmic-
barrier method (Nocedal and Wright, 2000) to (4.33) to form the following succession of
subproblems where µ is known as the barrier parameter:
X X
min f (x) − µ log si ⇔ f (x) − µ log si (4.35a)
i i
cE (x) = 0
s.t. c(x, s) = 0 ⇔ (4.35b)
cI (x) − sI = 0
x − l − sL = 0
cb (x, s) = 0 ⇔ (4.35c)
−x + u − sU = 0
We now apply a primal-dual approach (Nocedal and Wright, 2000) introducing the
dual variables defined as z = µS −1 e, where S is a matrix whose diagonal contains the
elements of s and zero-valued for elements out of the diagonal, and e is a vector with all of
its elements equal to 1. After a few algebra manipulations, we can arrive at the following
expression for the KKT perturbed conditions (Nocedal and Wright, 2000):
q(x) − A(x)T λ − C1 η
 

 −C2T λ + C3T η − z 

rµ (x, s, λ, η) = rµ (w) = 
 c(x, s) =0
 (4.36)
 cb (x, s) 
Sz − µe
where λ and η contain the Lagrange multipliers associated to, respectively, c(x, s) and
cb (x, s); q(x) is the gradient of f (x); A(x) is defined as the x-Jacobian of c(x, s), namely
A(x) = ∇x c(x, s); C1 = ∇x cb (x, s), C2 = ∇s c(x, s) and C3 = −∇s cb (x, s), which are all
constant matrices; and w = (xT , sT , λT , η T )T is just a grouping of variables used in the
algorithm.
4.3.2 Modified Newton’s step: Hessian, Lagrangian and gradient

In each global iteration k of the IPM, we apply Newton’s method (Nocedal and Wright,
2000) to the KKT perturbed conditions (4.36) to yield a system of equations from which
we can obtain the direction ∆wk along which to carry out the search. After removing
∆zk from this system (it can easily be demonstrated (Nocedal and Wright, 2000) ∆zk =
−Sk−1 Zk ∆xk − zk + µSk−1 e), we arrive at:
ATk C1T qk − ATk λk − C1T ηk

    
Hk 0 ∆xk
−1
 0 S zk C T   −C T λk + C T ηk − µS −1 e 
 k 2 −C3T 

 ∆sk
= 2 3 k  (4.37)
 Ak C2 0 0   −∆λk   ck 
C1 −C3 0 0 −∆ηk cb,k
P
where Hk comes from the x-Hessian of f (x) − i λi ci (x). In the problem being studied,
formulation with real-valued variables (4.22) allows the easy computation of matrices qk ,
4.3 Primal-dual Interior Point Method with Filtering 55
Ak and Hk in the system of equations of the Newton step (4.37) as:
qk =2hxk (4.38)
a0R
 

 a0I 

(1)
aR
 
 
 (1) 
 aI 
..
 
 
 . 
 a(INL )
 

 R 
 aI(INL )
 

 2xT b(1)
 

 k SL 
 .. 
Ak =
 . 
 (4.39)
 2xT b(JSL ) 
 k SL 
 2xTk b(1)
 
XP

..
 
.
 
 
 2xT b(MXP )
 

 k XP 
 2xT b 
 k AE 
 2xTk b(1)
 
DR 
 .. 
.
 
 
T (N )
2xk bDRDR
JSL M XP
(j+2+INL ) (j) (m+2+INL +JSL ) (m)
X X
Hk =2h − 2 λk bSL −2 λk bXP − (4.40)
j=1 m=1
N DR
(1+2+INL +JSL +MXP ) (n+2+INL +JSL +MXP +1) (n)
X
−2λk bAE −2 λk bDR
n=1
(l)
where λk denotes element l of λk (lth Lagrange multiplier).
Finally, in each iterarion the barrier parameter is updated with a quasi-quadratic
rate, µk = min{0.8µk−1 , kr0 (xk , sk , λk , ηk )k1.9 } and the IPM converges when kr0 (wk )k ≤
10−8 (1 + kqk k).
4.3.3 Filtering with backtracking

Once the direction ∆wk along which to move has been obtained, we have to calculate the
step length αk , so that we can update our variables for the next iteration according to
wk+1 = wk +αk ∆wk . Choosing two quantities ϑm = ϑ(wm ), ψm = ψ(wm ) that respectively
represent a measure in the objective progress and a measure towards feasibility (Benson
et al., 2002), we say point (ϑm , ψm ) dominates point (ϑn , ψn ) if and only if ϑm ≤ ϑn and
ψm ≤ ψn . A filter is a set of points {(ϑm , ψm )}, m = 1, 2, . . . Mf , such that no point
dominates any other point in the filter. To compute αk , we apply a bactracking procedure
(0)
to an initial value αk , calculated as the maximum step to keep s strictily in the interior
(l) (l−1)
of the feasible region (Nocedal and Wright, 2000). Backtracking, such that αk = βαk ,
(l)
where β = 0.5, is applied until the point (ϑ, ψ) associated to wk + αk ∆wk enters the
(l)
filter, setting then αk = αk . A point enters the filter if it is not dominated by any of the
points in the filter. Furthermore, when a point enters the filter, all the points dominated
by the new one must be eliminated from the filter. Analogously, the dual variables will be
(0,z) (0,z)
updated according to zk+1 = zk + min{αk , αk }∆zk , where αk is the maximum step
length to keep dual variables z strictly in the interior.
We have implemented two different filters (Benson et al., 2002), (Ulbrich et al.,
2004), according to the quantities ϑ, ψ used in each of them. Filter n. 1, which we will
call BVS (after Benson-Vanderbei-Shannon (Benson et al., 2002)) uses as filter quantities:
c(x, s) 2

X
ϑ = f (x) − µ si , ψ= (4.41)
cb (x, s)
i
and filter n. 2, which we will call UUV (after Ulbrich-Ulbrich-Vicente (Ulbrich et al.,
2004)), uses:
  2

c(x, s)

ϑ = f (x), ψ = cb (x, s)
 
(4.42)
Sz − µe
One of the problems (Benson et al., 2002) to be taken into account when implement-
ing the filters is the fact that the barrier parameter µ varies for each iteration k. Apart
from the task of updating the points in each iteration, if ϑ or ψ depend on µ, there may
be points that although present in the filter for a certain value of µ, should not be in the
filter for a new value of µ.
Before applying the algorithm described in this section to the problem being studied
(4.22), it has been tested with the non-linear optimization problems proposed in (Schit-
tkowski, 2002). Not being in the scope of this text and for the sake of brevity, these
results, which are in good agreement with those in the reference, will not be presented
here, and they indeed show that the two implemented filtering techniques offer similar
performances.
4.4 Numerical Examples

4.4.1 Computational aspects
In order to evaluate the computational performance and eventually be able to determine
any existing differences between the use of the two implemented filtering techniques (BVS
and UUV) in the optimizations, we show the number of global iterations (It.) needed by the
IPM (not related to the backtracking iterations) and the number of function evaluations
(Ev.), that is to say, the number of times f (x) = xT hx has to be computed throughout the
whole optimization procedure (including backtracking iterations). We also show the time
taken by the optimization procedures, but it should be taken into account that the code has
been written using an interpreted programming language so the times presented, despite
being small, can be still reduced using a compiled version of the algorithm. The times
shown consider just the time used by the IPM to solve the given non-linear optimization
problem. They do not take into account either the time of the antenna element FE
simulation, which has to be carried out only once and takes an average time of half a
minute for each of the two elements used as examples, nor the time taken to compute
the transmission and reflection matrices of the whole array, which is, as previously shown,
4.4 Numerical Examples 57
an analytical calculation that is in the order of tenths of seconds at most for the arrays
used in the examples. The number of spherical modes used to characterize both antennas
is 24. It must also be borne in mind that the number of variables the IPM deals with
is always double the number of antenna elements in the array. In the cases where no
constraints are set in the problem, the IPM always reaches the optimum solution (equal
to that of the direct method in Section 4.2.2) in one iteration, which in turn requires
3 function evaluations, no matter what filtering technique is used. This also happens
when only equality constraints (corresponding to null pointing directions) are set, since
this consitutes a quadratic programming problem with a direct solution (Ng et al., 1993).
Unless otherwise stated (only in the 4 × 4 planar array example) the initial point for the
optimization in all the examples was chosen to be that of a uniform cophasal excitation
and no bounds l or u are set.
4.4.2 Optimization of linear arrays

To illustrate how the presented formulation inherently takes into account all interelement
coupling effects and how this feature may be critical in some designs, we optimize a
linear array of 8 patches placed along the E-plane with an interelement distance of 0.4λ0 .
The antenna element under consideration is the cavity-backed, coaxial-fed, microstrip
circular patch with superstrate shown in Figure 4.1. The circular cylindrical cavity is
recessed in an infinite metallic plane. Details of this antenna can be found in (Zavosh
and Aberle, 1995; Rubio et al., 2002). The computed resonant frequency, which will be
used throughout the examples, is 1.215 GHz. All dimensions and dielectric permittivities
shown in Figure 4.1 are considered in the hybrid FE-MA simulation. In this configuration,
a high permittivity dielectric superstrate is used to render the antenna electrically small,
at the expense of a reduction in the directive gain, whose value has been computed to
be 5.4 dBi. The array is scanned at θ0 = 20◦ while rejecting interferences (setting null
pointing directions) at −20◦ , −50◦ and −75◦ . The optimized excitations to maximize the
directive gain under these restrictions have been obtained in two situations: considering
all interelement coupling effects and without taking them into account, which is equivalent
to setting the overall transmission matrix of the array to [T] = [t], where [t] is defined in
(3.11). In this last case, the use of a numerical quadrature, namely the Lebedev-Laikov
scheme on a sphere (Lebedev and Laikov, 1999), is necessary to calculate [BD ] since setting
[Γ] = [ρ] makes [BD ] 6= [BG ], despite not considering dielectric losses. Figure 4.2 shows
the computed patterns for the excitations yielded by both optimization cases when all
interelement couplings are considered. As can be seen, the exclusion of coupling from the
formulation and thus the optimization procedure can result in a severe distortion of the
real pattern, moving the desired nulls to other locations or even making them not appear.
The crosspolar component is negligible in both cases.
As a second example, we deal with an adaptive broadside array of 32 patch elements
equally spaced at 0.5λ0 which has been optimized to reject interferences at −70◦ , −40◦
and −20◦ , while keeping a maximum SLL of -40 dB between −30◦ and −10◦ (defined
with JSL = 41 grid points), just as the one with isotropic sources optimized in (Lebret
and Boyd, 1997) for the same constraints. Figure 4.3 shows the achieved field pattern in
the case in which the elements are placed along the H-plane. A similar solution with a
negligible crosspolar component is achieved when the elements are placed along the E-
plane. As in the previous example, if we exclude coupling from the optimization procedure
the resulting pattern does not fulfill all of the conditions, especially those related to the
Figure 4.1: Top and side view of the cavity-backed circular microstrip patch used as an array
antenna element. R1 = 2 cm, R2 = 3 cm, c1 = c2 = 0.6 cm, xp = 0.62 cm, ǫr1 = 2.5, ǫr2 = 20.
Coaxial feed (SMA connector): ǫr = 1.95, ri = 0.65 mm, ro = 2.05 mm.
Figure 4.2: Field patterns of the E-plane 8 cavity-backed patch linear array for the excitations
given by the two optimization cases: when interelement couplings are included in and excluded
from the optimization procedure.
Figure 4.3: Field pattern of the optimized 32 patch linear array along the H-plane. Interferences
at −70◦, −40◦ and −20◦ are rejected and the maximum SLL between −30◦ and −10◦ is set to -40
dB.
location of the null pointing directions.

In the next example the previous array, with the elements placed along the E-plane,
has been optimized to have a minimum SLL region beside the main beam, between −20◦
and −10◦ and between 10◦ and 20◦ , represented by 12 null field equality constraints, while
keeping the maximum SLL to -40 dB in the rest of the pattern, defined with a grid of
JSL = 142 points. This is a similar example to the one found in (Ng et al., 1993) with
isotropic sources. The resulting pattern is shown in Figure 4.4. It can be seen how the
achieved solution fulfills these conditions with a maximum SLL beside the main beam
which ranges from -100 dB to -70 dB. The crosspolar component is negligible.
Table 4.1 shows some computational aspects of the optimizations carried out on
linear patch arrays for the two implemented filtering techniques (BVS and UUV), where
the performance of the IPM can be seen.
The same 32 patch linear broadside array placed along the E-plane, with an interele-
ment distance of 0.5λ0 , has been optimized to achieve a pattern which fulfils the following
restrictions: rejecting interferences at -70◦ , -40◦ , -20◦ and 50◦ , and having a maximum
SLL of -30 dB between -90◦ and -30◦ , -40 dB between -30◦ and -5◦ , -20 dB between 5◦ and
40◦ and -50 dB between 40◦ and 90◦ . The resulting pattern is shown in Figure 4.5.
As another linear array example, we will consider a linear array of 17 DRA elements
arranged in the H-plane, equally spaced with an interelement distance of 0.4λ0 , scanned
at θ0 = 30◦ , with a desired maximum sidelobe level (SLLm ) of -20 dB. The DRA element
used is the one found in Figure 3.4, and the computed resonant frequency is 3.64 GHz.
For comparison purposes, we first use a Riblet-Dolph-Chebychev (RDC) excitation scheme
Figure 4.4: Field pattern of the optimized 32 patch linear array along the E-plane.
Table 4.1: Computational aspects of linear patch arrays optimizations represented in Figs. 4.2,
4.3 and 4.4.
BVS UUV
Figure N IN L JSL
It. Ev. Time It. Ev. Time
4.2 8 3 0 1 3 0.02 s 1 3 0.02 s
4.3 32 3 41 40 87 0.64 s 40 83 0.81 s
4.4 32 12 142 42 140 3.61 s 42 89 2.81 s
Figure 4.5: Optimized 32-element linear patch array field pattern for different maximum SLL
regions and null directions.
(Riblet, 1947; Dolph, 1946) on the array, which results in a field pattern whose sidelobes
are slightly above -20 dB. Next, we carry out an optimization to ensure SLL below -20 dB
between −90◦ and 20◦ and between 40◦ and 90◦ . Both patterns are shown in Figure 4.6 and
it can be seen how the optimized case strictly fulfills the maximum SLL requirement. To
check the capabilities of the presented optimization procedure, we also impose a constraint
on the crosspolar level in the direction of the main beam. Table 4.2 shows the results for
the optimized excitations according to different values for the desired maximum crosspolar
level in the pointing direction (none, -35, -55 and -75 dB) as well as the results for the
RDC excitation. Figure 4.7 shows the pattern obtained with the optimized excitations
for a crosspolar level of -75 dB. In general, a large decrease in the crosspolar level in
the pointing direction (obtaining a similar value for the directive gain) is achieved at the
expense of a loss in the aperture illumination efficiency (expressed as a percentage in
Table 4.2) and an increase in the DR (expressed in natural units in Table 4.2). The IPM
takes for each optimization an average of 43 iterations with about 100 function evaluations,
which amounts to approximately 2.5 s per optimization.
Now, the excitations of a 21-element linear array, with DRAs placed along the
E-plane equispaced at a distance of 0.4λ0 , are optimized to synthesise a pattern which
complies with the following specifications in the E-plane: main beam pointing at θ0 =30◦ ,
maximum SLL of -30 dB between -90◦ and 20◦ , and maximum SLL of -20 dB between
40◦ and 90◦ . The resulting pattern is shown in Figure 4.8.
Finally, the same array is optimized to achieve a pattern that fulfils the following
specifications: main beam pointing at θ0 =30◦ ; maximum SLL of -30 dB between -90◦ and
-40◦ , maximum SLL of -40 dB between -40◦ and 20◦ and maximum SLL of -20 dB be-
Figure 4.6: RDC and optimized (to ensure -20 dB of maximum SLL) field patterns of the 17
linear DRA array along the H-plane.
Table 4.2: Results of the optimizations of the 17 linear DRA array. JSL = 162, MXP =
1 (=0 in the first optimization).
SLLm XP L(θ0 ) D ε DR Time

It. Ev.
(dB) (dB) (dBi) (%) (n.u.) (s)
-20∗ -23.62 14.47 75.51 4.61 43 90 2.25
-20∗ -35∗ 14.37 40.08 10.4 42 89 2.34
-20∗ -55∗ 14.30 27.77 9.71 44 100 2.39
-20∗ -75∗ 14.30 26.80 9.61 43 103 2.44
-18.30 -24.41 14.28 95.91 2.02 RDC excitation
∗ set as a restriction
Figure 4.7: Field pattern of the H-plane 17 linear DRA array with optimized excitations to
ensure -75 dB crosspolar level at θ0 = 30◦ .
tween 40◦ and 90◦ ; two null pointing directions at θ=-40◦ and θ=0◦ . The resulting pattern
is shown in Figure 4.9 and it can be seen how it strictly complies with the prescribed
specifications.
4.4.3 Optimization of circular arrays

We now consider circular arrays of equally spaced elements scanned at θ0 = 90◦ , ϕ0 = 0◦ .
No rotation of the elements is considered with respect to their original coordinate system.
This circular configuration is known to approach a superdirective behaviour as the array
diameter becomes smaller (Cheng, 1971). However, this superdirective gain is reached at
the expense of an aperture illumination efficiency decrease. We examine the possibility
of maximizing the directive gain while fixing a desired minimum aperture illumintation
efficiency in arrays of patches of different number of elements N and different array di-
ameters d. Table 4.3 shows the results of the directive gain Dm , aperture illumination
efficiency εm and dynamic range DRm achieved with the optimized excitations obtained
to maximize the directive gain with no other constraints (with the algorithm or through
the application of the direct method of Section 4.2.2), and the results of the directive gain
D75 and dynamic range DR75 obtained to maximize the directive gain with a restriction
on the minimum aperture illumination efficiency of 75% (which is always reached by the
algorithm), as well as the directive gain Du with a uniform excitation for comparison pur-
poses. It can be seen how as the aperture illumination efficiency increases the directive
gain decreases, but is always above the one with a uniform excitation. The times taken
by the IPM to reach the solution in each case are in the order of hundredths of second.
Figure 4.8: Optimized 21-element linear DRA array field pattern for two different maximum SLL
regions.
Figure 4.9: Optimized 21-element linear DRA array field pattern for different maximum SLL
regions and two null locations.
Table 4.3: Results of the optimizations of circular arrays of patches.
Du Dm εm DRm D75 DR75

N d/λ0
(dBi) (dBi) (%) (n.u.) (dBi) (n.u.)
8 0.7 11.15 13.41 34.02 3.25 12.86 2.68
8 0.9 11.91 12.90 68.89 1.57 12.89 1.53
10 0.9 12.36 14.05 50.82 2.67 13.87 2.21
Table 4.4: Optimized excitations (amplitude and phase) for maximum directive gain of a circular
array with N = 8 and d = 0.9λ0 .
v1 v2 v3 v4 v5 v6 v7 v8
Patch antenna elements. D=12.90 dBi.
0.64 0.96 1.00 0.81 0.85 0.81 1.00 0.96
25.3◦ 0.0◦ 148.4◦ -81.9◦ 23.6◦ -81.9◦ 148.4◦ 0.0◦
Dielectric resonator antenna elements. D=13.87 dBi.

0.60 1.00 0.77 0.67 0.37 0.67 0.77 1.00
-56.6◦ 0.0◦ 159.7◦ -101.5◦ -1.6◦ -101.5◦ 159.7◦ 0.0◦
Table 4.4 shows the obtained excitations (amplitude and phase) to maximize the
directive gain (no other constraints) in the case of N = 8 and d = 0.9λ0 for the two different
antenna elements under consieration. It can be seen how the achieved solution is different
when using different antenna elements, showing how the formulation implicitly takes into
account the different antenna elements and its corresponding interelement coupling.
We finally consider a circular array of 30 DRA elements with half-wave spacing
between them, scanned at θ0 = 90◦ , ϕ0 = 0◦ . A sidelobe reduction in the azimuthal pattern
up to -29.16 dB can be found in (Yan and Lu, 1997) for this same array configuration with
isotropic sources. With a similar aim, we have fixed a maximum SLL of -30 dB between
ϕ = 20◦ and ϕ = 340◦ for the optimization of the proposed DRA circular array. Although
the algorithm reached a satisfactory solution, it was found that the crosspolar level rised
above the maximum SLL level, so we added the constraints corresponding to a maximum
crosspolar level of -30 dB in the same region. The IPM with the BVS filter took, for
JSL = 321 and MXP = 321, 57 iterations with 124 evaluations (a time of 64.58 s) to
converge to a solution. The achieved pattern, which fulfills the specified requirements, is
shown in Figure 4.10.
As a last note on the optimization of circular arrays, it can be said that although
the results are mathematically correct, they might not be accurate for small finite metallic
planes since they are based on a model with an infinite metallic plane at θ = 90◦ , which
has been the direction chosen to carry out the optimizations over the azimuthal angle ϕ.
4.4.4 Optimization of planar arrays

In the case of planar arrays, we first examine a broadside 4×4 array of patches, with the
same interelement spacings in the x- and y- directions, 11 cm (approximately 0.45λ0 ). The
Figure 4.10: Field pattern of the 30 DRA circular array with optimized excitations to ensure -30
dB of maximum SLL and crosspolar level.
excitations that achieve, with no restrictions, the maximum directive gain (16.16 dBi) are
very near the uniform excitations, which give 16.12 dBi of directive gain and a maximum
SLL of -12.58 dB. The array excitations have been optimized to achieve the maximum
directive gain while maintaining a maximum SLL of -20 dB between θ = 40◦ and θ = 90◦
in the two principal planes, E- (ϕ = 0◦ , 180◦ ) and H- (ϕ = 90◦ , 270◦ ). We have carried out
several optimizations increasing the value of constant κR to establish a control on the DR,
while keeping the prescribed maximum SLL. The results, along with the computational
aspects of the IPM for the two implemented techniques, are displayed in Table 4.5. They
show how the achieved directive gain and aperture illumination efficiency decrease as the
DR is lowered down. Since the SLL restriction is only set on the two principal planes, the
IPM not having to fulfill any constraints of this type in the rest of the space, as the DR
is lowered down the maximum relative field value between θ = 40◦ and θ = 90◦ outside
the two principal planes may rise. In the worst case (last optimization to reach a DR
of 1.30) this value is approximately -10 dB. Each optimization takes as the initial point
the solution given by the previous optimization (the first one takes the uniform cophasal
excitation and so NDR = 0).
We will now deal with a broadside 9×9 array of equally spaced patches with an
interelement spacing of 0.5λ0 . We have carried out an optimization to achieve a maximum
SLL of -20 dB in the region θ ∈ (19◦ . . . 80◦ ), and another one to obtain a maximum SLL
of -30 dB in the region θ ∈ (23◦ . . . 80◦ ), both of them for eight ϕ-cuts, ϕ ∈ (0◦ , . . . , 315◦ )
in steps of 45◦ , thus having a dense enough grid to ensure SLL below the specified values.
The maximum achievable directive gain of this array (with no restrictions) is 24.10 dBi,
corresponding to a solution which is practically equal to the uniform excitation. Table 4.6
Table 4.5: Optimizations of the 4×4 patch array for DR control with =-20 DB in the H- and E-
plane. JSL = 204, NDR = 16.
D DR εAE BVS UUV

(dBi) (n.u.) (%) It. Ev. Time (s) It. Ev. Time (s)
15.79 2.37 92.31 42 95 3.20 42 112 3.45
15.73 2.20 91.16 45 92 9.88 45 100 10.02
15.22 1.89 81.94 48 151 11.59 42 97 9.34
15.15 1.73 79.00 41 85 9.09 41 91 9.09
14.87 1.54 73.78 41 85 9.08 41 84 9.02
14.57 1.38 68.71 42 90 9.36 42 85 9.23
14.50 1.34 67.52 41 91 9.19 41 84 9.05
14.43 1.30 66.35 42 98 9.47 42 87 9.27
Table 4.6: Results of the optimization of the 9×9 broadside patch array.
SLLm XP Lm D ε DR
JSL It. Ev.
(dB) (dB) (dBi) (%) (dB)
-20 ∗ -36.00 23.80 91.75 11.12 496 52 130
-30 ∗ -38.85 22.88 75.42 44.36 464 52 130
-12.97 -39.96 24.07 100 0 Uniform excitation
∗ set as a restriction
shows the main parameters of the array with the optimized and the uniform excitations,
as well as the performance aspects (nearly identical for both filtering techniques) of the
IPM, which takes an average of 75 s to converge to a solution. The crosspolar component
in the broadside direction is negligible in all cases and the maximum crosspolar level
XP Lm is always below the maximum sidelobe level SLLm in the whole space as shown
in Table 4.6. As can be seen, in the two optimized situations the maximum SLL has been
respectively lowered by more than 7 dB and 17 dB with respect to the uniform case, while
the directive gain has decreased by less than 0.3 dBi and 1.2 dBi. The price to pay for such
achievements is a larger DR and a loss in the aperture illumination efficiency. Figure 4.11
show the three dimensional field pattern given by the optimized excitations in the case of
a desired maximum SLL of -30 dB.
In the final example, the 9×9 planar broadside patch array has been optimized to
achieve the following maximum SLL restrictions between θ=20◦ and 90◦ : -30 dB in the
region ϕ=45◦ to 225◦ , and -20 dB in the rest of the space. The resulting pattern is shown
in Figure 4.12.
Figure 4.11: Optimized 9×9 patch array field pattern for -30 dB maximum SLL.
Figure 4.12: Optimized 9×9 patch array field pattern for different regions of maximum SLL.
Chapter 5
Linear and Quadratic

Programming for array pattern
synthesis
5.1 Introduction
In the preceding chapter, a general pattern synthesis problem has been formulated in
terms of a non-linear optimization problem with various constraints and solved with an
optimization method which has been specifically developed for this aim. However, there
are other synthesis problems which may not require the use of such a sophisticated opti-
mization algorithm. In optimization theory, both LP and QP are two fields which have
been widely researched and therefore provide a large number of efficient algorithms to be
used (Nocedal and Wright, 2000). In this chapter, the main objective is expressing various
synthesis problems as LP and QP problems so that they can be efficiently solved with the
aid of any LP or QP method. Therefore, the stress is now made on the formulation of
these problems rather than on the development of the optimization method, as this latter
is a well-studied research topic in LP and QP theory. Specifically, LP problems will arise
after a transformation of minimax (i. e., SLL) problems, while QP theory is more suited
to design problems in which the minimization of a MSE appears, such as the adjustment
of the radiated field to a prespecified function.
5.2 LP-based techniques

5.2.1 Synthesis of minimum sidelobe level patterns with prescribed nulls
5.2.1.1 Minimax problem
In this case, the design problem under study is stated as: minimize the maximum value
in the sidelobe region, where sampling (θi , ϕi ), i = 1 . . . I outside the region left for the
mainlobe must be dense enough to cover all possible peaks, subject to a fixed field value
in the main beam direction (θ0 , ϕ0 ) and subject to a zero-valued field in the desired null
directions (θq , ϕq ), q = 1 . . . Q. By taking the matrix-valued expression of the radiated
field (4.4), it takes the form:
69
70 Linear and Quadratic Programming for array pattern synthesis
min max |hg(θi , ϕi )i {v} |, i = 1 . . . I (5.1a)

i
s. t. |hg(θ , ϕ )i {v} | = κ (5.1b)
|hg(θq , ϕq )i {v} | = 0, q = 1 . . . Q (5.1c)
where hg(θ, ϕ)i is defined in (4.3) and (4.4).

This problem could be tackled by minimizing a real positive quantity δ and making
the absolute value of the field outside the region left for the mainlobe remain below this
quantity, while keeping the rest of the constraints unchanged.
min δ (5.2a)
s. t. |hg(θi , ϕi )i {v} | ≤ δ, i = 1 . . . I (5.2b)
|hg(θ , ϕ )i {v} | = κ (5.2c)
|hg(θq , ϕq )i {v} | = 0, q = 1 . . . Q (5.3)
However, in the most general case we are treating, we cannot assume a real-valued
field in the whole θ, ϕ-space as in other similar works (Holm and Elgetun, 1997), since we
are on the one hand considering real radiating elements with interelement coupling and on
the other hand making no assumptions on the geometry of the array and the main beam
pointing direction (θ0 , ϕ0 ), so the removal of the norm in |hg(θi , ϕi )i {v} | cannot be made
straightforwardly.
5.2.1.2 Real-valued equivalent LP problem

By splitting the field value in a given direction in its real and imaginary parts we can
compute them from a vector containing the real and imaginary parts of the excitations:

ℜ {v}
{x} = (5.4)
ℑ {v}
where we have:
E(θi , ϕi ) = hr(θi , ϕi )i {x} + jhh(θi , ϕi )i {x} (5.5)
hr(θi , ϕi )i=hℜ (hg(θi , ϕi )i) , −ℑ (hg(θi , ϕi )i)i (5.6)

hh(θi , ϕi )i=hℑ (hg(θi , ϕi )i) , ℜ (hg(θi , ϕi )i)i (5.7)
By using the triangle inequality, we have:
|hg(θi , ϕi )i {v} | ≤ |hr(θi , ϕi )i {x} | + |hh(θi , ϕi )i {x} | (5.8)
We can therefore lower the contribution from both the real and the imaginary parts
of the field separately in all θi , ϕi directions by setting:
|hr(θi , ϕi )i {x} |≤α (5.9a)

|hh(θi , ϕi )i {x} |≤β (5.9b)
5.2 LP-based techniques 71
(where α, β are real positive quantities), and by minimizing α + β while keeping a constant
(generally complex) value of the field in the main beam direction and setting both the real
and imaginary parts of the field to zero in the null directions. The removal of the norms
in (5.9) can now be made by considering the signs of the quantities inside the norms,
since these quantities are no longer complex numbers but real numbers. In this way, the
problem is now reformulated as:
min α+β (5.10a)

s. t. hr(θi , ϕi )i {x}≤α, i = 1 . . . I (5.10b)
−hr(θi , ϕi )i {x}≤α, i = 1 . . . I (5.10c)
hh(θi , ϕi )i {x}≤β, i = 1 . . . I (5.10d)
−hh(θi , ϕi )i {x}≤β, i = 1 . . . I (5.10e)
hr(θ , ϕ )i {x}=ℜ(κ) (5.10f)
hh(θ , ϕ )i {x}=ℑ(κ) (5.10g)
hr(θq , ϕq )i {x}=0, q = 1 . . . Q (5.10h)
hh(θq , ϕq )i {x}=0, q = 1 . . . Q (5.10i)
From this formulation, we can arrive at the standard form for a LP problem given
as:
min f T y (5.11a)
s. t. Ay ≤ b (5.11b)
Cy = d (5.11c)
by defining:
 
 {x} 
y= α (5.12)
β
 
 
 {0} 
f= 1 (5.13)
1
 
b={0} (5.14)
 
 ℜ(κ) 
d= ℑ(κ) (5.15)
{0}
 
 
[RI ] {−1} {0}
 [−RI ] {−1} {0} 
A= [HI ]
 (5.16)
{0} {−1} 
[−HI ] {0} {−1}
 
hr0 i 0 0
 hh0 i 0 0 
C= [RQ ] {0} {0} 
 (5.17)
[HQ ] {0} {0}
κ being a constant to be specified, [RI ] and [HI ] are matrices whose columns are respec-
tively hr(θi , ϕi )i and hh(θi , ϕi )i for i = 1 . . . I, [RQ ] and [HQ ] are matrices whose columns
Figure 5.1: Field patterns of the uniform 21 DRA array for different excitations. · · · case 1: ideal
isotropic pattern from RDC excitations. - - - case 2: real pattern from RDC excitations. —– case
3: real pattern from proposed-technique excitations.
Figure 5.2: Optimized field patterns of the 21 DRA array for different mainlobe widths. - - -
case 1: maximum SLL: -7.3 dB, 3-dB beamwidth: 5.8◦ . —– case 2: maximum SLL: -16.9 dB, 3-dB
beamwidth: 7.4◦ .
are respectively hr(θq , ϕq )i and hh(θq , ϕq )i for q = 1 . . . Q, while hr0 i = hr(θ , ϕ )i and
hh0 i = hh(θ , ϕ )i. Analogously, {0} and {−1} stand for column vectors (with the ap-
propiate size in each case). The resulting problem can be solved with any of the existing
LP methods; in this case, an active-set algorithm has been used.
To test the design procedure, a uniform linear 21-element broadside array made up
of DRAs as the one in Figure 3.4 placed along the E-plane with an interelement distance
of 0.4λ0 at the computed resonant frequency of 3.64 GHz is optimized. The resulting
pattern, together to the one given by a RDC excitation (Riblet, 1947; Dolph, 1946) for the
same mainlobe width, is shown in Figure 5.1, where the ideal isotropic RDC pattern is also
shown for comparison purposes. As can be seen, the RDC excitations do not achieve the
minimum maximum SLL reached by the proposed technique. To check the capabilities of
the proposed procedure, Figure 5.2 shows the results obtained for the same array scanned
at -30◦ for two different mainlobe widths and with a desired null at 45◦ .
Finally, a symmetric, nonuniform random linear array of 22 DRAs is considered.
The 11 necessary interelement distances are calculated as (0.4+0.03ψx )λ0 , where ψx is
a uniformly distributed random variable (-1 to +1). Figure 5.3 compares the pattern
obtained with the excitations from the proposed technique with the one yielded by the
excitations from the procedure to minimize the maximum SLL in (Holm and Elgetun,
1997), which assumes a symmetric array with an even number of isotropic elements, with
real symmetric excitations, scanned at broadside (directly resulting in a real-valued field
problem). The excitations in this last case, as in the RDC example, yield a non optimal
pattern. The crosspolar component of the field is negligible in all considered cases.
Figure 5.3: Field patterns of the nonuniform 22 DRA array with a SLL region from ±5◦ to ±90◦.
- - - case 1: technique in (Holm and Elgetun, 1997) for symmetric isotropic arrays. —– case 2:
proposed technique.
5.2.2 Multibeam pattern synthesis with minimum sidelobe level

This synthesis may be seen as a variation of the previous case. Now, we consider the
design problem of having different main beam directions θq , ϕq , for q = 1, . . . , Q instead
of only one main beam direction θ0 , ϕ0 . For the sake of simplicity, no null directions
are considered, although they may be easily added. Therefore, the design problem can
be stated as: minimize the maximum absolute field value in the sidelobe region, where
sampling (θi , ϕi ), i = 1 . . . I must be dense enough to cover all possible sidelobe peaks,
while keeping a constant specified complex value pq in the directions of the main beams
(θq , ϕq ), q = 1 . . . Q.
min max |hg(θi , ϕi )i {v} |, i = 1, . . . , I (5.18a)
subject to |hg(θq , ϕq )i {v} | = pq , q = 1, . . . , Q (5.18b)
Following a similar approach to the one in the previous section, we can arrive at a
standar LP problem in the form (5.11), where:
 
 {x} 
y= α (5.19)
β
 
 
 {0} 
f= 1 (5.20)
1
 
b={0} (5.21)

ℜ({pQ })
d= (5.22)
ℑ({pQ })
 
[RI ] {−1} {0}
 [−RI ] {−1} {0} 
A= [HI ]
 (5.23)
{0} {−1} 
[−HI ] {0} {−1}

[RQ ] {0} {0}
C= (5.24)
[HQ ] {0} {0}
where {pQ } is a vector column containing the values of pq for q = 1, . . . , Q.
Again, a uniform linear 21-element broadside array, with DRA elements along the
E-plane with an interelement distance of 0.4λ0 at the computed resonant frequency of 3.64
Figure 5.4: Optimized 21 DRA array patterns with two main beams for two different pairs of
values of p−30◦ and p60◦ .
Figure 5.5: Two optimized 21 DRA array patterns with three main beams. In the second one,
the central mainbeam is broader than the other two mainbeams.
GHz, is optimized to achieve a pattern with minimum maximum SLL with two mainbeams
(-30◦ and 60◦ ). Furthermore, each of the two mainbeams is required to have a different
mainlobe width. The copolar field pattern is shown in Figure 5.4 for two different pairs of
values of p−30◦ and p60◦ . In general terms, by properly setting the values of pq the relative
amplitude and phase difference of the field in the directions of the main beams can be
specified in the optimization algorithm. However, as expected, the resulting patterns will
be different (especially in the sidelobe region) for every combination of values pq . This
phenomenon can be seen in Figure 5.4.
The same array is now optimized to achieve a pattern with minimum maximum SLL
and three mainbeams (0◦ and ±45◦ ). The results are shown in Figure 5.5. In this case, two
optimizations have been carried out. In the first one, all mainbeams are required to have
the same width, while in the second optimization the beams corresponding to ±45◦ are
narrower than the central one. Therefore, in the latter case, the maximum SLL is higher.
The crosspolar component of the field is negligible in all considered examples.
5.2.3 Shaped-beam pattern synthesis with controllable sidelobe level,

crosspolar level and nulls
We will now consider the more general problem of synthesizing a pattern with a shaped-
beam region and a sidelobe region. Furthermore, the crosspolar component of the radiated
field will also be taken into account. Again, null directions will also be possible in both the
copolar (L null directions) and crosspolar (K null directions) pattern. These null directions
will be set as the L + K equality constraints the optimization problem is subject to. This
optimization problem will be defined as a minimax problem whose objective function is
in turn split into three weighted subobjectives. Each of these three subobjectives has
a correspondence with the three considered regions (which are independently sampled
with I, Q and M points): the sidelobe region, the crosspolar region and the shaped-beam
region. In the first two regions the aim is obtaining the minimum SLL and crosspolar level,
respectively, so the subobjectives are defined in a similar way as the previous sections, equal
to the absolute vale of the field. However, in the third region, the shaped-beam one, the
aim is having a shaped-beam pattern given by and ideal function with sampling points
pm , m = 1, . . . , M . This way, the subobjective is defined as the distance of the actual field
in the region to the ideal field, that is to say, the absolute value of the difference between
the radiated field and the ideal function |hg(θm , ϕm )i {v} − pm |.
As has been previously said, each subobjective in the objective function is weighted
with the three corresponding factors: fSL for the sidelobe region, fXP for the crosspolar
region, and fSB for the shaped-beam region. The values of these factors must be heuris-
tically adjusted by the designer to obtain a desired pattern, as the optimized pattern
depends on them. Setting a higher value of a weight will make the field in the associated
region have a higher degree of compliance than in the other regions.
The min max optimization problem is then defined as:
min max fSL |hg(θi , ϕi )i {v} | + fXP |hgxp (θq , ϕq )i {v} |+ (5.25a)
+fSB |hg(θm , ϕm )i {v} − pm |
i = 1, . . . , I; q = 1, . . . , Q; m = 1, . . . , M
subject to |hg(θl , ϕl )i {v} | = 0, l = 1, . . . , L (5.25b)
xp
|hg (θk , ϕk )i {v} | = 0, k = 1, . . . , K (5.25c)
Again, a complex-to-real transformation is carried out. By using the triangle in-

equality, we would have for the shaped-beam region:
|hg(θm , ϕm )i {v} − pm | ≤ |hr(θm , ϕm )i {x} − ℜ(pm )| + |hh(θm , ϕm )i {x} − ℑ(pm )| (5.26)
Now, the minimax-to-LP transformation must be carried out. Defining a set of real
quantities to be minimized while making the real and imaginary parts of each of the three
subojectives remain below these quantities is again the approach which is followed. The
weights are maintained on these real quantities. This way, α1 and α2 are used for the real
and imaginary parts of the of the sidelobe region, β1 and β2 for the real and imaginary
parts of the of the crosspolar region, and γ1 and γ2 for the real and imaginary parts of the
of the shaped-beam region. Therefore, the equivalent minimization problem is expressed
as:
min fSL (α1 + α2 ) + fXP (β1 + β2 ) + fSB (γ1 + γ2 ) (5.27a)

s. t. |hr(θi , ϕi )i {x} | ≤ α1 , i = 1, . . . , I (5.27b)
|hh(θi , ϕi )i {x} | ≤ α2 , i = 1, . . . , I (5.27c)
xp
|hr (θq , ϕq )i {x} | ≤ β1 , q = 1, . . . , Q (5.27d)
xp
|hh (θq , ϕq )i {x} | ≤ β2 , q = 1, . . . , Q (5.27e)
|hr(θm , ϕm )i {x} − ℜ(pm )| ≤ γ1 , m = 1, . . . , M (5.27f)
|hh(θm , ϕm )i {x} − ℑ(pm )| ≤ γ2 , m = 1, . . . , M (5.27g)
|hr(θl , ϕl )i {x} | = 0, l = 1, . . . , L (5.27h)
|hh(θl , ϕl )i {x} | = 0, l = 1, . . . , L (5.27i)
xp
|hr (θk , ϕk )i {x} | = 0, k = 1, . . . , K (5.27j)
xp
|hh (θk , ϕk )i {x} | = 0, k = 1, . . . , K (5.27k)
Finally, it is straightforwardly shown that the associated matrices of the standard

LP problem (5.11) take the form:
 

 {x} 




 α1 



 α2 

 

y= β1 (5.28)
β2 

 

 

γ1 

 

 

γ2
 
 

 {0} 



 fSL 


 fSL 

 

f= fXP (5.29)
fXP 

 

 

f
 
SB

 

 
fSB
 
 

 {0} 

{0}

 


 

{0}

 


 

 



 {0} 






 {0} 



{0}
 
b= (5.30)

 {0} 

{0}

 


 

ℜ({p M })

 


 

−ℜ({pM })

 


 

 
 ℑ({pM })






 −ℑ({p }) 

M
d = {0} (5.31)
5.3 QP-based methods 77
[RI ] {−1} {0} {0} {0} {0} {0}

 

 [−RI ] {−1} {0} {0} {0} {0} {0} 


 [HI ] {0} {−1} {0} {0} {0} {0} 


 [−H
h Ii] {0} {−1} {0} {0} {0} {0} 

Rxp {0} {0} {−1} {0} {0} {0}
 

 Q 

h i
−Rxp {0} {0} {−1} {0} {0} {0}
 
h Qi
 
A=  (5.32)
Hxp {0} {0} {0} {−1} {0} {0}
 
h Q i
 
 
−Hxp {0} {0} {0} {−1} {0} {0}
 

 Q 


 [RM ] {0} {0} {0} {0} {−1} {0} 


 [−RM ] {0} {0} {0} {0} {−1} {0} 

 [HM ] {0} {0} {0} {0} {0} {−1} 
[−HM ] {0} {0} {0} {0} {0} {−1}
 
[RL ] {0} {0} {0} {0} {0} {0}
 [HL ] {0} {0} {0} {0} {0} {0} 
C=  Rxp {0} {0}
 (5.33)
K
xp {0} {0} {0} {0} 
HK {0} {0} {0} {0} {0} {0}
As an example, we will optimize a 21 DRA array along the E-plane with antennas
equispaced at 0.4λ0 (at 3.64 GHz, the resonant frequency). As the crosspolar component
is negligible in this plane, we will only consider a shaped-beam and a sidelobe region.
The former goes from 0◦ to 20◦ , while the latter is defined in the rest of the space. The
shaped-beam ideal function is defined as 1 over the whole shaped-beam region. To show
how the weighting factors influence the optimized pattern, three cases are considered:
fSB /fSL = 4, fSB /fSL = 2 and fSB /fSL = 1. Figure 5.6 shows the patterns obtained for
these three cases. As expected, a higher value of the shaped-beam weighting factor results
in a pattern with a higher SLL which on the other has a higher degree of compliance with
the shaped-beam ideal function. This last fact can be seen in Figure 5.7, which is a zoom
of the shaped-beam region of the optimized patterns.
5.3 QP-based methods

5.3.1 Mutual coupling compensation
Synthesis techniques that assume the isotropic nature of the radiators provide a set of
excitations {u} from which the ideal field pattern may be expressed as hζ(θ, ϕ)i{u} by
taking [̥(θ, ϕ)] = [I] in (4.4), since no real element patterns or mutual coupling effects are
being considered. If these excitations are used in a real antenna array, the copolar field
pattern is hζ(θ, ϕ)i[̥(θ, ϕ)]{u}. The presence of matrix [̥(θ, ϕ)], which accounts for the
individual elements’ patterns and the interelement MC effects, results in a real pattern
distortion with respect to the initially designed ideal isotropic pattern. It is therefore
desirable to have a set of excitations {v} that can make the real pattern a best fit of the
ideal isotropic response, so that:
hζ(θ, ϕ)i[̥(θ, ϕ)]{v} ≈ hζ(θ, ϕ)i{u} (5.34)
The row vector hζ(θ, ϕ)i and the matrix [̥(θ, ϕ)] are defined in (4.3). The main objective
of this section is the derivation of a MC compensation matrix [M] to compute {v} readily
Figure 5.6: Optimized 21 DRA array patterns for a shaped-beam and a sidelobe region. · · · case
1: fSB /fSL = 4. - - - case 2: fSB /fSL = 2. —– case 3: fSB /fSL = 1.
Figure 5.7: Zoom of the shaped-beam region of the optimized 21 DRA array patterns for a
shaped-beam and a sidelobe region. · · · case 1: fSB /fSL = 4. - - - case 2: fSB /fSL = 2. —– case
3: fSB /fSL = 1.
from {u} as:

{v} = [M]{u} (5.35)
To that effect, a MSE minimization approach is followed:
ZZ
minv |hζ(θ, ϕ)i[̥(θ, ϕ)]{v} − hζ(θ, ϕ)i{u}|2 dΩ (5.36)
where the integration must be carried out in the region(s) where the MSE is desired to be
minimized (in this case, the whole space in planar arrays or the principal plane in linear
arrays). The MSE can be expressed from (5.36) as (Ng et al., 1993):
{v}H [Q]{v} − {u}H [P]H {v} − {v}H [P]{u} + {u}H [C]{u} (5.37)
where [Q], [P] and [C] are the following N × N matrices:

ZZ
[Q]= [̥(θ, ϕ)]H hζ(θ, ϕ)iH hζ(θ, ϕ)i̥(θ, ϕ)]dΩ (5.38)
ZZ
[P]= [̥(θ, ϕ)]H hζ(θ, ϕ)iH hζ(θ, ϕ)idΩ (5.39)
ZZ
[C]= hζ(θ, ϕ)iH hζ(θ, ϕ)idΩ (5.40)
Integrations in (5.38) and (5.39) are likely to be carried out numerically, while there is
an analytical expression of [C] for certain array configurations (Cheng, 1971; Einarsson,
1979). However, as it is shown next, the computation of [C] is not necessary for our
purposes.
Since the solution vector {v} that minimizes (5.37) is [Q]− [P]{u}, a direct expres-
sion for the MC compensation matrix is then obtained:
[M] = [Q]− [P] (5.41)
As can be seen, the MC compensation matrix of an array has to be calculated only

once, since it only depends on the radiating and scattering characteristics of the antenna
elements, as well as on their location in the array. Note that the MC compensation matrix
has no (θ, ϕ)-dependence as it arises from an MSE minimization process. Moreover, in the
MSE quadrature no numerical effort is made to compute the field value in every (θ, ϕ)-
direction since the radiation pattern of the array is expressed analytically. This way,
the MC compensation matrix can be flexibly used to compensate the presence of the real
individual elements’ patterns and the interelement MC effects for any excitations obtained
with an isotropic-based pattern synthesis method.
In order to show the effectiveness of the proposed method, some examples are de-
scribed next. The antenna element used in the arrays is again the DRA shown in Fig-
ure 3.4, at the computed resonant frequency of 3.64 GHz.
A linear equispaced 21-element array with the DRA elements placed along the E-
plane, with an interelement distance of 0.365λ0 at the resonant frequency, is used to
carry out two different designs. Firstly, a two-beam pattern, with θ1 =-15◦ and θ2 =45◦ , is
considered. The original excitations are obtained by the superposition of the two uniform
cophasal excitations which scan the array at each of the pointing directions. The resulting
field patterns using the original excitations and the MC compensated excitations are shown
in Figure 5.8. As observed, with the original excitations one of the two main beams,
Figure 5.8: Two-beam pattern of the 21-element DRA array obtained with the original and the
MC compensated excitations.
θ2 =45◦ , has a lower relative level (approximately 3 dB) than the other one. The MC
compensated excitations yield an equal level for the two main beams, with a field pattern
which is superimposed to the ideal isotropic two-beam pattern.
The second linear array example consists of a cosecant-squared pattern which is
initially designed with the Woodward-Lawson technique (Woodward and Lawson, 1948)
under the assumption of the isotropic nature of the radiators. The csc2 regions are
[−90◦ ,−20◦ ] and [20◦ ,90◦ ]. The field patterns obtained with the original excitations and
with the MC compensated excitations are shown in Figure 5.9. As before, the former
does not achieve the desired requirements since it is degraded because of interelement
MC effects and the individual DRA patterns, while the latter is coincident with the ideal
isotropic pattern.
Finally, an 8×8 DRA planar array is considered. The interelement distance in the x
and y-directions is the same as before. The Tseng-Cheng excitations (Tseng and Cheng,
1968) are used to achieve a maximum SLL of -10 dB in the whole space. Figure 5.10
shows, for three different ϕ-cuts, the pattern obtained with the original Tseng-Cheng and
the MC compensated excitations, as well as the ideal Tseng-Cheng isotropic pattern for
comparison purposes. As shown in Figure 5.10(a), the pattern with the original distri-
bution complies with the SLL restriction in the the E-plane (ϕ=0◦ ). However, the main
beam has broadened with respect to the ideal isotropic pattern, perturbing the optimum
condition the isotropic-based Tseng-Cheng distribution provides for the directive gain. In
the H-plane (ϕ=90◦ ), shown in Figure 5.10(b), the SLL of the two sidelobes closest to
the main beam is higher than -10 dB in the pattern with the original excitations. In
other planes, such as ϕ=60◦ , shown in Figure 5.10(c), there are even more discrepancies
between the real and the ideal patterns using the Tseng-Cheng distribution. The pattern
obtained with the MC compensated excitations achieves a higher level of compliance with
the specifications in the whole space, and it is clearly a best fit of the ideal isotropic
response.
Figure 5.9: csc2 pattern of the 21-element DRA array obtained with the original Woodward-
Lawson excitations and the MC compensated excitations.
5.3.2 Null-constrained mutual coupling compensation

In certain applications, the specification of field pattern nulls over a set of directions has a
critical importance. It has been proven that MC has a major impact in these cases (Huang
et al., 2006), causing displacements of null locations or an increase of the null levels in
the field patterns. The minimization of the MSE may not be enough to reach the desired
requirements for the null directions. Reinforcement of these conditions can be carried out
by explicitly establishing a zero value for the field in the q null (θ, ϕ)-directions as:
{v}H [Z] = {v}H [{z1 }, . . . , {zq }] = h0i (5.42)
where [Z] is the N × q matrix formed from:
{zi }H = hζ(θi , ϕi )i[̥(θi , ϕi )], i = , . . . , q (5.43)
Minimizing (5.37) subject to (5.42) is a constrained quadratic programming problem with

a direct solution (Ng et al., 1993). From this solution, a null constrained MC compensation
matrix can be obtained as:
− H
[MN ] = [I] − [Q]− [Z] [Z]H [Q]− [Z] [Z] [M] (5.44)
where [I] is the identity matrix. In this case, the excitations {v} that best fit the
ideal isotropic pattern with the reinforced null conditions are computed from the null-
constrained MC compensation matrix as:
{v} = [MN ]{u} (5.45)
As before, for a given array, matrices [Q] and [M] need to be calculated only once from,
respectively, (5.38) and (5.41). They both provide, through (5.44), a direct means to
Figure 5.10: Tseng-Cheng pattern of the 8×8 DRA array obtained with the original and the MC
compensated excitations.
Figure 5.11: Schelkunoff E-plane field pattern of the 8×8 DRA array obtained with the original
and the two different MC compensated excitations.
compensate the MC effects in a field pattern for any isotropic-based excitation vector {u}
and any set of desired null locations specified by matrix [Z].
As an example, the previous 8×8 DRA array is considered. Schelkunoff method
(Schelkunoff, 1943) is carried out to place 7 nulls in the E-plane (ϕ=0◦ ) pattern in the
following θ-locations: -90◦ , -60◦ , -45◦ , -30◦ , 30◦ , 60◦ and 90◦ . The 8 excitations given by
Schelkunoff method are used in each row of the array (parallel to the x-axis). Figure 5.11
shows the E-plane field pattern obtained with the null constrained MC compensated ex-
citations (using matrix [MN ]) and the MC compensated excitations (using matrix [M]).
For comparison purposes, the ideal isotropic Schelkunoff pattern and the field pattern
obtained with the original Schelkunoff excitations are also shown. The latter is degraded
with respect to the ideal design because of interelement MC effects and the individual
elements’ patterns. Using the MC compensated excitations mitigates this pattern degra-
dation, but the obtained pattern is still far from the ideal isotropic one. As can be seen,
the null constrained MC compensated excitations achieve the highest degree of compliance
with the ideal isotropic Schelkunoff pattern.
Chapter 6
Non-optimization based field

synthesis techniques
6.1 Introduction
So far, all the synthesis problems which have been reported in this thesis have been
formulated and solved within the framework of optimization theory. This chapter provides
two design methods which require neither the formulation nor the resolution of the problem
be carried out in terms of any optimization-based approaches. The first one is an adaption
of classic isotropic-based techniques based on Fourier theory. The second method digs
explicitly into the GSM characterization through the use of spherical vector modes to
provide a design technique based exclusively on rotation and translation properties of
spherical waves. The common feature in these methods is that they require the expression
of the desired field as a weighted sum of basis functions.
6.2 Fourier synthesis of linear arrays

6.2.1 Introduction
The mathematical similarity between the expressions of the radiated field by a linear eq-
uispaced antenna array and several problems related to signal processing has been widely
known since the early days of antenna array design. This analogy has enabled the trans-
mission of advances made in one of these fields, usually signal processing, to the other
one. As a major example, one can find typical pattern synthesis techniques for linear
equispaced arrays such as the Fourier Series (FS) method or the Woodward-Lawson (WL)
frequency sampling method, based on the Discrete Fourier Transform (DFT), (Stutzman
and Thiele, 1981; Woodward, 1946; Woodward and Lawson, 1948). These classic tech-
niques are especially suited to synthesise a great variety of shaped beam patterns, such
as single or multiple main beam cosine patterns (Werner and Ferraro, 1989). Despite
being long-lasting and well-known techniques, the FS and the WL methods have always
drawn much attention from the antenna array research community. In recent years, sev-
eral works have reported new developments of different applications and modifications of
the WL method (Cid et al., 1999; Dürr et al., 2000; Fondevila-Gómez et al., 2006; Goudos
et al., 2007), which usually constitute an improvement or greater insight into the original
method.
85
86 Non-optimization based field synthesis techniques
This section presents a new simple procedure for the pattern synthesis of linear
equispaced arrays whose radiated field can be expanded in spherical modes from the array
GSM. It allows the inclusion in the design of real antenna element patterns and mutual
coupling effects, neglected in the aforementioned classic methods. This procedure extends
the analogy between this problem and the signal processing theory by making use of
the linearity and multiplication properties of the FS or DFT coefficients (Oppenheim
and Willsky, 1997), (Oppenheim and Schafer, 1999) of both the desired pattern and the
spherical waves to yield a linear system of equations which provides the antenna array
excitations.
6.2.2 Linear equispaced array characterization

We will start by considering that the same spherical mode expansion is used in all the
antennas in the array, Mi = M, ∀i, so that the radiated field of the array in (3.22) can be
expressed as:
M N
~ ϕ) = (m)
X X
E(θ, ~em (θ, ϕ) bi exp (jk sin θ(xi cos ϕ + yi sin ϕ)) (6.1)
m=1 i=1
Furthermore, if we consider again, without loss of generality, one single mode of

excitation per antenna in the array, so that {vi } = vi , the complex amplitudes of the
spherical modes can be calculated as:
N
(m) (m)
X
bi = Ti,h vh . (6.2)
h=1
(m)
As before, coefficients Ti,h , directly extracted from the elements of the overall transmission
matrix [T], take account of the contribution of the excitation of antenna h to the mth
spherical mode of antenna i. This way, both the element pattern corresponding to every
antenna and all the interelement coupling effects are taken into account through coefficients
(m)
Ti,h in (6.2), which for an array of real antennas depend on their radiating and scattering
characteristics and their location in the array.
For a linear equispaced array with antenna elements placed along a ϕ-cut direction
and spaced at a distance d, the copolar component of every spherical mode em (θ) is taken
from (6.1) to arrive at the following expression for the copolar field pattern in the specified
ϕ-cut for θ ∈ [−90◦ , 90◦ ]:
M N −1
(m)
X X
E(θ) = em (θ) bi exp (jikd sin θ) (6.3)
m=1 i=0
By setting ψ = kd sin θ, this expression can be written as

M
X
Ẽ(ψ) = ẽm (ψ)Hm (ψ) (6.4)
m=1
N −1
(m)
X
Hm (ψ) = bi exp (jiψ) (6.5)
i=0
6.2 Fourier synthesis of linear arrays 87
where we have defined

ψ
ẽm (ψ) = em arcsin , −kd ≤ ψ ≤ kd (6.6)
kd
so that it is ensured that
E(θ) = Ẽ(kd sin θ) = Ẽ(ψ), ∀θ ∈ [−90◦ , 90◦ ] (6.7)
6.2.3 Inclusion of the isolated element pattern

In the case where the only available information is the spherical mode expansion corre-
sponding to the field radiated by a single isolated element, (6.2) must be rewritten as:
(m) (m)
bi = ti vi (6.8)
(m)
Assuming, as is usual, that all elements in the array are equal, then ti = t(m) , ∀i. By
considering as in basic array theory that the element pattern in the array environment is
equal to the single isolated antenna pattern, (6.4) can be rewritten in the following way:
M
X N
X −1
Ẽ(ψ) = t(m) ẽm (ψ) vi exp (jiψ) (6.9)
m=1 i=0
From this expression, the application of the classic FS method or WL frequency

sampling design technique (Stutzman and Thiele, 1981), (Orfanidis, 2008), including the
element pattern, is carried out straightforwardly from the pattern multiplication theorem
by splitting the array factor from the individual element pattern as:
N −1
Ẽ(ψ) X
= vi exp (jiψ) (6.10)
Ee
i=0
M
X
Ee = t(m) ẽm (ψ) (6.11)
m=1
With this technique, an improvement in the array field pattern over the classic
isotropic FS or WL methods can be achieved by taking the isolated element pattern into
account. However, it lacks the information concerning mutual coupling, which is critical
in some designs, and whose inclusion in the proposed design procedure is looked at in the
following subsections.
6.2.4 Fourier Series analysis

Assuming N → ∞, each function Hm (ψ) is a ψ-periodic signal with period 2π and it has
(m)
an FS representation with coefficients bi which can be directly taken from (6.5):
∞
FS (m) (m)
X
Hm (ψ) ←−
−→ bi , Hm (ψ) = bi exp (jiψ) (6.12)
i=−∞
With regard to ẽm (ψ), it can be assumed to be a ψ-periodic signal with the same
(m)
period so that it has an FS with coefficients ci :
Z π
FS (m) (m)
ẽm (ψ) ←−
−→ ci , ci = ẽm (ψ) exp (−jiψ)dψ (6.13)
−π
In this case, one has to bear in mind that ẽm (ψ) must be a periodic signal with period
2π. Given the fact that the visible region in ψ is [−kd, kd], corresponding to [−90◦ , 90◦ ]
in the θ-space, if we take kd > π, or equivalently d > 0.5λ0 , the periodic nature cannot
be ensured together with condition
S (6.6). For d ≤ 0.5λ0 the analysis is valid and ẽm (ψ)
must be defined in [−π, −kd) (kd, π].
As Hm (ψ) and ẽm (ψ) are ψ-periodic signals, Ẽ(ψ) must also be ψ-periodic, and
therefore it has a FS with coefficients ui :
Z π
FS
Ẽ(ψ) ←− −→ ui , ui = Ẽ(ψ) exp (−jiψ)dψ (6.14)
−π
By applying the linearity and multiplication properties of the FS (Oppenheim and

Willsky, 1997) to (6.4) we can easily arrive at the following expression, where ∗ stands for
the linear convolution of two sequences:
M X
∞ M
(m) (m) (m) (m)
X X
ui = bl ci−l = bi ∗ ci (6.15)
m=1 l=−∞ m=1
6.2.5 Discrete Fourier Transform analysis

(m)
Firstly, an N -point sequence hi is formed by sampling each function Hm (ψ) at the N
equidistanced sampling points ψi (Orfanidis, 2008) given by (assuming N odd):
2π(i − (N − 1)/2)
ψi = , i = 0, . . . , N − 1 (6.16)
N
(m)
The sequence hi has a corresponding N -point DFT which is directly taken from (6.5):
N −1
(m) DFT (m) (m) (m)
X
hi ←−
−→ bi , hi = Hm (ψi ) = bi exp (jψi ) (6.17)
i=0
(m)
In an analogous way, the sequence ẽi , formed from the sampling of ẽm (ψ) at ψi ,
(m)
has a DFT with coefficients ci :
N −1
(m) DFT (m) (m)
X
ẽi ←− −→ ci , ci = ẽm (ψl ) exp (−jiψl ) (6.18)
l=0
As before, if d > 0.5λ0 , the visible region extends far beyond where the sampling is carried
out [−π, π], and the analysis cannot be carried out.
By taking the samples in Ẽ(ψ) to form the sequence Ẽi , its associated N DFT
coefficients ui can be calculated as:
N −1
DFT
X
Ẽi ←−
−→ ui , ui = Ẽ(ψl ) exp (−jiψl ) (6.19)
l=0
As in the case of the FS, by applying the linearity and multiplication properties of
the DFT (Oppenheim and Schafer, 1999) to (6.4), we can easily arrive at the following
expression, where N stands for the N -point circular convolution of two sequences.
M N −1 M
(m) (m) (m) (m)
X X X
ui = bl c(i−l) mod N = bi
N ci (6.20)
m=1 l=0 m=1
6.2.6 Fourier-based design procedure

It must first be noted that expressions (6.14) and (6.19) correspond to the classic isotropic
FS and WL methods. They directly provide the excitations ui needed to achieve a de-
sired response with a linear equispaced array of isotropic radiators in these classic designs
(Orfanidis, 2008). The objective of the proposed method is to synthesize the pattern that
would be obtained from ideal isotropic sources, but with an array of real and coupled
(m)
radiators. By substituting bi , given by (6.2), in the convolution of the right-hand side of
(6.15) or (6.20), we arrive at the following expressions for ui (for each of the two Fourier
techniques considered in this section):
N M
(m) (m)
X X
ui = vh ci ∗ Ti,h (6.21)
h=1 m=1
N M
(m) (m)
X X
ui = vh ci T
N i,h (6.22)
h=1 m=1
These expressions are now used to form a linear system of equations {u} = [D]{v} to
solve for the excitation vector {v}. The design procedure can be summarized as follows:
(m)
1. Compute Ti,h . The method described in Chapter 3 is used to carry out this step.
(m)
2. Compute ci from (6.13) or (6.18). In the case of the infinite FS, no more than
(m)
the 2N − 1 central coefficients ci (for N odd) are required, since in the following
step a convolution with a finite sequence of N terms will be carried out to yield
(m)
another sequence of N terms. In any case, the number of coefficients ci to be
computed can be set as an option in the procedure.
3. Compute matrix [D]. This is an N × N matrix whose element (i, h) is calculated
from (6.21) or (6.22) (depending on the Fourier technique used) as:
M M
(m) (m) (m) (m)
X X
Di,h = ci ∗ Ti,h , Di,h = ci T
N i,h
m=1 m=1
In the case of the infinite FS, only the central N terms of the sequence resulting
from the convolution are needed.
4. Calculate ui from (6.14) or (6.19) to form the vector {u}. For the infinite FS
technique, only the central N terms need to be calculated. Coefficients ui can
be computed numerically, or analytically in the case of simple or canonical field
patterns. If desired, windowing (Oppenheim and Schafer, 1999), (Orfanidis, 2008)
can be applied to {u}.
5. Solve {u} = [D]{v} for {v}.
A restriction on the maximum distance between elements for this procedure to be
valid has been derived in previous subsections (d ≤ 0.5λ0 ). However, this limitation is not
crucial for practical purposes, since the main feature of the proposed design technique is
its inherent capability of taking coupling effects into account and these effects are known
to exert a critical influence on the field patterns of arrays with an interelement distance
of d ≤ 0.5λ0 .
6.2.7 Numerical examples

In order to validate the proposed technique, a sector beam array design (from θ = 15◦ to
65◦ ) has been carried out with an 81-element array. The antenna element used in the array
Figure 6.1: Sector field patterns for the 81 DRA array obtained with the excitations from the
proposed FS procedure (coincident with the isotropic array pattern from the classic isotropic FS
method) and from the classic isotropic FS method, compared to the ideal response.
is the probe-fed hemispherical dielectric resonator antenna shown in Figure 3.4. DRA
elements are placed along the E-plane (x-direction in Figure 3.4) with an interelement
distance of 0.365λ0 at the computed resonant frequency (3.64 GHz).
Figure 6.1 shows the normalized ideal pattern and the normalized field pattern ob-
tained with the FS design procedure proposed in this section. A good level of compliance
is achieved except for the known effects of the Gibbs phenomenon (Oppenheim and Will-
sky, 1997). The normalization is carried out in a way that the value of the pattern at
the center of the sector θ = 40◦ is equal to 1. For comparison purposes, Figure 6.1 also
shows the pattern given by the DRA array (with the GSM and spherical mode expansion
characterization) with the excitations from the classic isotropic FS design technique, that
is to say, if {v} = {u} is directly set. As can be seen, this pattern is degraded as a result
of the absence of the individual elements’ patterns and interelement coupling effects in the
classic design procedure. The pattern obtained with the proposed FS technique coincides
with the one given by the same array of isotropic radiators with the excitations from the
classic isotropic FS method, as is expected from the development of the proposed method.
The crosspolar component is negligible in all considered cases.
Figure 6.2 shows the results for the same previous cases but when a Hamming
window is applied to the excitations. The Gibbs phenomenon has now been eliminated
at the expense of a larger transition region between the desired sector and the rejection
Figure 6.2: Sector field patterns for the 81 DRA array obtained with the excitations, with a
Hamming window, from the proposed FS procedure (coincident with the isotropic array pattern
from the classic isotropic FS method) and from the classic isotropic FS method, compared to the
ideal response.
region. Again a high level of compliance with the ideal response is reached by the proposed
FS design technique.
With the same array, a multisector pattern synthesis has been carried out with the
following features: main beam from θ = 30◦ to θ = 60◦ ; a -20 dB sector from θ = −90◦
to θ = −5◦ and a -30 dB sector from θ = 5◦ to θ = 30◦ . The field in the rest of the
θ-space is null. Figure 6.3 shows, together with the ideal response, the patterns obtained
with the classic isotropic WL design technique and the classic with element pattern WL
technique described in section 6.2.3, as well as the proposed DFT method. In all designs,
a Hamming window is used. As can be seen, the inclusion of the element pattern in the
design procedure achieves in overall an improvement on the classic isotropic WL method
(except for the region near θ = −90◦ ), but because of coupling effects it is still far from the
pattern obtained with the proposed DFT method, near the ideal response. Once again,
our procedure provides a direct means to achieve the pattern obtained with the same array
of isotropic radiators for the isotropic WL excitations.
A csc2 pattern design between θ = 15◦ and θ = 75◦ , with a null field pattern in
the rest of the θ-space, is now considered with the same array. With this aim, a classic
isotropic WL method, a classic with element pattern WL method and our proposed DFT
procedure are carried out. The results obtained, compared to the pattern given by the
Figure 6.3: Multisector field patterns for the 81 DRA arrays obtained with the excitations, with
a Hamming window, from the two classic WL methods and from the proposed DFT procedure
(coincident with the isotropic array pattern from the classic isotropic WL design), compared to
the ideal response.
Figure 6.4: csc2 field patterns for the 81 DRA array obtained with the excitations from the
proposed DFT procedure and from the two classic WL methods, compared to the isotropic array
pattern from the classic isotropic WL design and the ideal response.
same array with isotropic radiators under a WL excitation and the ideal desired pattern,
are shown in Figure 6.4. No windowing is applied in any case. The same effects as in
the previous multisector pattern designs can be seen. It is again demonstrated how the
proposed DFT procedure achieves a very similar pattern to the isotropic array response
with the classic isotropic WL method, especially in the region of interest.
Finally, in order to check how the proposed design procedure inherently takes into
account the different array antenna elements’ patterns as well as the different coupling
effects between them, it is used to synthesize the pattern of two different 49 element
linear array. Both arrays have the same configuration but they are made up of different
antenna elements. The first array is made up of the aforementioned DRAs placed along
the E-plane with an interelement distance of 0.365λ0 at the computed resonant frequency,
3.64 GHz (3 cm). The elements of the second array are the cavity-backed, coaxial-fed,
microstrip circular patch with superstrate shown in Figure 4.1, which has a computed
resonant frequency of 1.215 GHz, nearly 3 times smaller than the one computed for the
DRA element. The microstrip patch elements are placed along the E-plane (x-direction
in Figure 4.1) with an interelement distance of 9 cm, which at the resonant frequency
is electrically equivalent to the one used in the DRA array. A two-sector-beam pattern
has been synthetized with both arrays, whose configurations are electrically identical and
only differ in the radiating and coupling characteristics of their elements. The main sector
Figure 6.5: Multisector field patterns, compared to the ideal response, obtained with the excita-
tions (with a Hamming window) from the two classic WL methods and from the proposed DFT
procedure, coincident with the isotropic array patterns from the classic isotropic WL design, for
the two electrically identical, but with different antennas, 49 element arrays.
ranges from θ = −60◦ to θ = −20◦ , while the other one, from θ = 20◦ to θ = 60◦ , must
have a -6 dB level with respect to the main sector. The desired field pattern in the rest
of the θ-space is null. Figure 6.5 shows the normalized field patterns obtained for the two
arrays under consideration with the excitations from the proposed DFT procedure, from
the classic with element pattern WL design and from the classic isotropic WL method.
As can be seen, with the classic WL designs the patterns are degraded with respect to the
ideal response, but in a dissimilar way since the array antenna elements are different. This
degradation is more notorious in the case of DRA elements. As expected, the inclusion of
the individual element pattern in the classic WL method brings, in general, the obtained
pattern nearer the ideal response than the one given by the classic isotropic WL method.
However, as in previous examples, it is not enough to reach the pattern obtained with
the proposed DFT design technique, since coupling between elements is not taken into
account. With the proposed DFT procedure, the synthetized pattern is better fit to the
desired response in both arrays. In fact, these curves are superimposed, with a practical
identical correspondence to the pattern of the same array made up of isotropic radiators
under a classic isotropic WL excitation. In all designs, a Hamming window is applied
to the excitation vector and, as in the preceding examples, the crosspolar component is
negligible in all cases.
6.3 Spherical-wave based shaped-beam synthesis 95
As a final note, it should be reminded that the time taken by the antenna element
FE simulation, which is carried out only once, is on average half a minute for each of
the two elements used in the examples (DRA and microstrip circular patch), while the
computation of the array GSM is an analytical calculation which for the arrays presented
in this section is in the order of tenths of seconds. The time taken by the rest of the design
procedure (matrix manipulations, computation of coefficients, convolutions and resolution
of the linear system) varies from tenths of seconds to 6 s, depending on the example
considered, when using the DFT procedure. In the FS implementation, since a recursive
adaptive quadrature has been used to compute the integrals of the FS coefficients, this
time can be up to 7 times higher than with the DFT implementation. All computations
have been carried out on a standard personal computer with a Pentium D at 2.8 GHz with
2 GB of RAM.
6.3 Spherical-wave based shaped-beam synthesis

6.3.1 Introduction
Another synthesis approach is based upon the fact that any antenna radiated field can
be expressed as a discrete (truncated) series of weighted spherical vector functions. This
issue was not only addressed in pattern synthesis problems (Wood, 1971; James, 1975;
Narasimhan et al., 1986; Mehler, 1988) long ago, but also in antenna measurements
(Hansen, 1988). Specifically, the idea of synthesizing a field pattern of an antenna ar-
ray in terms of spherical modes is outlined in (James, 1975), where no interelement effects
are assumed. In (Narasimhan et al., 1986), a spherical-mode based method for pattern
synthesis is proposed for planar arrays whose elements are placed on an arbitrary contour,
although antennas were assumed to be point dipoles. To the best knowledge of the author
of this thesis, this is the first work which uses translational and rotational addition the-
orems for spherical waves (Stein, 1961; Cruzan, 1962; Bruning and Lo, 1971; Edmonds,
1974) to relate the spherical modes describing the radiated field of each array element
with its own coordinate system to the spherical modes of a reference coordinate system
belonging to the overall antenna array. This idea has also been recently used for near field
reconstruction (Gati et al., 2006).
Let us consider a set of radiating sources, each of which is inside the minimum-radius
sphere that surrounds (circumscribes) it. On these spheres, a spherical wave expansion of
the radiated field is carried out. In (Hansen, 1988) (page 355) the two cases for the general
translation of spherical vector waves between spheres with a different centre are addressed.
In the first case, the centres of the spheres are situated outside the radius of the other
spheres, while in the second case, at least one of the centres lies inside the other sphere.
Physically, with the first theorem it is possible to compute the spherical-mode expansion
coefficients of the field induced in one sphere by the field created in another sphere, which
is in turn also expressed as a spherical wave weighted sum. Therefore, in Chapter 3 the
translation theorems corresponding to this case are used to rigorously analyze the coupling
effects in arrays by circumscribing each antenna array element with a surrounding sphere
whose centre does not lie in any other elements’ spheres. If we now consider a sphere that
surrounds the whole array, then the array radiated field can be expressed as a spherical
vector mode expansion with its corresponding coefficients. These coefficients are related
to the coefficients of the spherical mode expansion carried out on every sphere surrounding
each element of the array. However, this time the centres of the antenna elements’ spheres
lie within the array sphere, requiring the use of the theorems corresponding to the second
case.
In this section, the combination of the two cases of application of the general transla-
tion theorems for spherical modes is carried out to yield a shaped beam synthesis method
for finite arrays that includes the MC effects. In contrast with the previous section, no
constraints are applicable to the array elements’ positions.
6.3.2 Spherical wave expansion of the shaped-beam field

It is a well known result that the far field of any radiating structure can be decomposed
into a weighted sum of spherical far-field vector wave functions. In this case, the radiating
~ ϕ) will be expressed
structure corresponds to an antenna array, whose radiated field E(θ,
as:
M
X
~ ϕ) =
E(θ, d(m)~em (θ, ϕ) (6.23)
m=1
where ~em (θ, ϕ) is the mth spherical far-field vector mode and d(m) is its associated complex
amplitude.
Spherical far-field wave functions satisfy the orthogonality relationship since (from
the properties in (Hansen, 1988)):
Z 2π Z π
= 0, m 6= n
~em (θ, ϕ) · ~e∗n (θ, ϕ) sin θdθdϕ (6.24)
ϕ=0 θ=0 6= 0, m=n
where ∗stands for the conjugate function and · for the vector dot product.
Therefore, by multiplying (6.23) by every spherical wave function ~e∗m (θ, ϕ), integrat-
ing in the whole space and using (6.24), a value of every expansion coefficient needed to
~ ϕ) is obtained as:
achieve a desired field E(θ,
Z 2π Z π
~ ϕ) · ~e∗m (θ, ϕ) sin θdθdϕ
E(θ,
ϕ=0 θ=0
d(m) = Z 2π Z π (6.25)
~em (θ, ϕ) · ~e∗m (θ, ϕ) sin θdθdϕ
ϕ=0 θ=0
The integral in the denominator has an analytical expression. Furthermore, in this

thesis, spherical modes also satisfy the orthonormal relationship, so there is no need to
evaluate it (as it is equal to unity). The integral in the numerator can be carried out
with the aid of a numerical quadrature routine. In our case, the Lededev-Laikov scheme
(Lebedev and Laikov, 1999) has been used. It is important to note that, for the proposed
synthesis technique, coefficients {d} = hd(1) , d(2) , . . . , d(M ) i need to be computed only once
for a given desired shaped-beam field regardless which type of antenna elements the array
is made up of, their position and even the number of elements. However, it should be born
in mind that the number of modes M needed must comply with the empirical criterion
given in Chapter 3 relating it to the radius of the sphere circumscribing, in this case, the
whole array. These coefficients are the starting point to derive a series of relationships
between different spherical wave expansions and antenna features that give rise to the
proposed synthesis procedure.
Figure 6.6: Scheme of the relationships between feeding with/and spherical modes used in the
proposed spherical-wave-based pattern synthesis procedure
6.3.3 Relationship between spherical modes and feeding

A scheme of the overall interactions established both by the antennas with/and the differ-
ent spherical wave expansions in which the field is expressed in the proposed methodology is
shown in Figure 6.6 for two antennas i and h. There is an overall spherical wave expansion
of the radiated field (scattered modes) defined by coefficients {d} = hd(1) , d(2) , . . . , d(M ) i
on a sphere surrounding the array, while the coefficients corresponding to outgoing spher-
ical waves on the sphere surrounding the ith antenna element of the array are given by
{bi }.
We will start by assuming the contribution to {d} coming from each of the N
antennas in the array is exactly known. Therefore, we could express {d} as:
N
X
{d} = {ph } (6.26)
h=1
where {ph } is the aforementioned contribution for h = 1, . . . , N (see Figure 6.6). Rewriting
this equation in matrix form would yield:
{d} = [Π] {p} (6.27)
where:
 

 {p1 }  
{p }
 
2

 

..

 

 
.
 
{p} = (6.28)

 {pi }  

 .. 

.

 


 

{pN }
 
and
[Π] = [[I] , [I] , . . . , [I] , . . . , [I]] (6.29)
According to Figure 6.6, coefficients {ph } corresponding to contributions to the

spherical modes in the overall array sphere are related to coefficients {bh } (corresponding
to the hth element’s sphere) by means of a GTM [Fh ], where the theorems of translation
for the second case reported in (Hansen, 1988), as well as the rotation theorems, are used,
in the form:
{ph } = [Fh ] {bh } (6.30)
Now, it will be assumed that the reciprocal relationship is known, although as it will
be shown at the end of the development that this matrix does not need to be calculated,
we make use of a GTM [Qh ] relating coefficients {ph } to coefficients {bh }:
{bh } = [Qh ] {ph } (6.31)
Following a similar reasoning as in Chapter 3, the contribution to coefficients of

incident spherical waves in every element’s sphere of coefficients of scattered spherical
waves of the other spheres can be established, as shown in Figure 6.6, by means of a GTM
[Gi,h ] as (same as Eq. (3.3)):
n o
ahi = [Gi,h ] {bh } (6.32)
In this case, matrix [Gi,h ] is computed by using the translational additional theorems of
the first case in (Hansen, 1988), as well as rotation theorems.
So, by substituting (6.31) in (6.32) we get:
n o
ahi = [Gi,h ] [Qh ] {ph } (6.33)
The overall coefficients of incident spherical modes of antenna i are given then by
(same as Eq. (3.2)):
XN n o
inc
ai = {ai } + ahi (6.34)
h=1
h 6= i
Substituting (6.33) in (6.34) we get:
N
X
inc
ai = {ai } + [Gi,h ] [Qh ] {ph } (6.35)
h=1
h 6= i
Now, by using (6.35) we can get from the GSM of each antenna i, given in (3.1),
the following relationship:
N
X
[ti ] {vi } + ([si ] − [Ii ]) {ai } + ([si ] − [Ii ]) [Gi,h ] [Qh ] {ph }=[Qi ] {pi } (6.36)
h=1
h 6= i
If we repeat these equations for each one of the N antennas and we arrange them,
the following matrix equation is found for the array:
[t] {v} + ([s] − [I]) {a} + ([s] − [I]) [G] [Q] {p}=[Q] {p} (6.37)
where [t], ([s] − [I]), {v}, {a}, {b} and [G] have the same definition as in Chapter 3, while
[Q] is given by:
 
[Q1 ] 0 ··· ··· 0
 .. .. 
 0 [Q 2 ] . 0 . 
 .. .. 
 
[Q] =  . .. .. .. (6.38)
. . . . 
 ..
 
..
. [QN−1 ]

 . 0 0 
0 ··· ··· 0 [QN ]
Now, by assuming as we are dealing with a radiating array that there is no incident
field coming from outside the array, {a} = 0 (6.37) is reduced to:
[t] {v}=([I] − ([s] − [I]) [G]) [Q] {p} (6.39)
From where it immediately follows:
([I] − ([s] − [I]) [G])−1 [t] {v}=[Q] {p} (6.40)
By recalling (6.30) we can multiply both sides in the previous equation by a matrix
[F] given by:
 
[F1 ] 0 ··· ··· 0
. .. 
[F2 ] . .

 0 0 . 
[F] =  ... .. 
 
 .. .. .. (6.41)
. . . . 
 ..
 
..
. [FN−1 ]

 . 0 0 
0 ··· ··· 0 [FN ]
so that
[F] ([I] − ([s] − [I]) [G])−1 [t] {v}=[F] [Q] {p} (6.42)
−1
[F] ([I] − ([s] − [I]) [G]) [t] {v}={p} (6.43)
where [F] [Q] {p} = {p} since it follows straightforwardly, by definition of (6.30) and
(6.31), that [F] [Q] = [I].
Now, by simply inserting (6.27) into (6.42) we get:
[Π] [F] ([I] − ([s] − [I]) [G])−1 [t] {v}=[Π] {p} (6.44)
−1
[Π] [F] ([I] − ([s] − [I]) [G]) [t] {v}={d} (6.45)
where coefficients {d} are known since they are given by the spherical wave expansion of
the desired field for the shaped-beam pattern. Therefore, we can express:
{d}=[M] {v} (6.46)

−1
[M]=[Π] [F] ([I] − ([s] − [I]) [G]) [t] (6.47)
Figure 6.7: Layout of the 325-element array used for the spherical-wave-based synthesis of a
flat-top beam pattern
from where it follows immediately that the required feeding coefficients to achieve the
desired shaped-beam pattern are given by:
{v} = [M]+ {d} (6.48)
with [M]+ denoting the pseudoinverse matrix of [M], defined in (6.47).
It is important to note that matrix [M] contains all the relationships reported in
Figure 6.6, both the ones corresponding to the scattering and transmission features of each
of the individual antennas in the array and the ones corresponding to the interactions
of the different spherical wave expansions used in different spheres, which gives rise to
take interelement MC effects inherently as well as making the proposed pattern synthesis
procedure possible and efficient.
6.3.4 Numerical results

A three-dimensional flat-top beam pattern between θ = 0◦ and θ = 45◦ for all ϕ is
synthesized with the proposed technique. The number of elements is 325, and they are
distributed according to the layout of Figure 6.7. The distance between rows, as well as
the distance between columns, is 0.5λ0 at the resonant frequency of the antenna element
used.
Two different antenna elements will be considered for this array. The first antenna
chosen to be the element is the microstrip patch of Figure 4.1, and the example is carried
out at the computed resonant frequency, 1.215 GHz. Figures 6.8 and 6.9 show, respec-
tively, the three-dimensional normalized copolar and crosspolar patterns achieved with
the proposed procedure. For the sake of clarity, Figure 6.10 shows the normalized copolar
and crosspolar field patterns for all ϕ-cuts in steps of 1◦ . As can be seen, the copolar
component reaches a good level of compliance with the required specifications.
Figure 6.8: Three-dimensional normalized copolar field pattern of the 325-element array of mi-
crostrip patches with 1.215 GHz of resonant frequency for a desired flat-top beam objective copolar
pattern.
Figure 6.9: Three-dimensional normalized crosspolar field pattern of the 325-element array of
microstrip patches with 1.215 GHz of resonant frequency for a desired flat-top beam objective
copolar pattern.
Figure 6.10: Normalized copolar (black) and crosspolar (grey) field patterns for all ϕ-cuts of 1◦ of
the 325-element array of microstrip patches with 1.215 GHz of resonant frequency for a desired
flat-top beam objective copolar pattern.
As the second antenna element, the same patch configuration is used, but this time
the permittivity of the superstrate in Figure 4.1 is changed to be equal to 1.0. This
implies that the resonant frequency of the antenna is 2.5 GHz, as reported in (Rubio
et al., 2004b). All dimensions in the array are now adjusted to maintain the same distance
between rows and between columns, 0.5λ0 . The achieved three-dimensional copolar and
crosspolar patterns are shown in Figures 6.11 and 6.12. Again, for the sake of clarity
Figure 6.13 shows the same patterns for every ϕ-cut in steps of 1◦ . As can be seen, there
is no appreciable difference between the achieved copolar pattern in the two different-
element arrays, which shows the efficiency and capability of the proposed technique to
handle different antenna elements. On the contrary, the crosspolar component, as is set to
0 in the design procedure, results in different patterns depending on which patch is used
as the array element.
Finally, a more complex pattern is considered. It is a sectorized csc2 pattern. The
sector comprises the region ϕ ∈ (45◦ , 135◦ ), while the csc2 shape is defined in the sector
with region θ ∈ (10◦ , 60◦ ). The antenna element used is the first one (microstrip patch
with a resonant frequency of 1.215 GHz). The synthesized three-dimensional copolar and
crosspolar patterns are shown respectively in Figures 6.14 and 6.15. Again, for the sake
of clarity Figure 6.16 shows the same patterns for every ϕ-cut in steps of 1◦ . As observed,
with the proposed technique, a high degree of compliance is reached with an achieved
pattern very close to the objective complex desired shaped-beam pattern.
Figure 6.11: Three-dimensional normalized copolar field pattern of the 325-element array of
microstrip patches with 2.5 GHz of resonant frequency for a desired flat-top beam objective copolar
pattern.
microstrip patches with 2.5 GHz of resonant frequency for a desired flat-top beam objective copolar
pattern.
Figure 6.13: Normalized copolar (black) and crosspolar (grey) field patterns for all ϕ-cuts of
1◦ of the 325-element array of microstrip patches with 2.5 GHz of resonant frequency for a desired
flat-top beam objective copolar pattern.
Figure 6.14: Three-dimensional normalized copolar field pattern of the 325-element array of
microstrip patches with 1.215 GHz of resonant frequency for a desired sectorized csc2 objective
copolar pattern.
microstrip patches with 1.215 GHz of resonant frequency for a desired sectorized csc2 objective
copolar pattern.
Figure 6.16: Normalized copolar (black) and crosspolar (grey) field patterns for all ϕ-cuts of 1◦ of
the 325-element array of microstrip patches with 1.215 GHz of resonant frequency for a desired
sectorized csc2 objective copolar pattern.
Chapter 7
Array thinning with a nested

optimization scheme
7.1 Introduction
Sparse and thinned arrays offer several advantages associated to a lower cost, weight,
power and complexity than a fully-filled antenna array. Therefore, they have been an
object of intense research since the first studies were carried out in the 1960s (see (Skolnik
et al., 1964; Lo and Lee, 1966) and references therein), and thus various developments
have been reported over the years.
Statistical thinning methods were widely used to tackle the design problem of a
thinned or sparse array (Skolnik et al., 1964; Numazaki et al., 1987; Mailloux and Co-
hen, 1991; Tomiyasu, 1991). However, the outstanding increase in both computational
capabilities and available optimization techniques over the last 20 years have made the
latter a common factor in most of the achievements regarding array thinning. Notwith-
standing, classic optimization methods (such as conjugate gradient, Newton methods and
similar) are not well suited for thinning an array (Haupt, 1994). Stochastic, probabilistic
or evolutionary optimization algorithms are by far the best suited and therefore the most
used.
The simulated annealing technique has been reported to achieve good results in
array thinning (Ruf, 1993; Murino et al., 1996; Trucco, 1999; Trucco and Murino, 1999).
Other approaches, such as multiplicative processing (Davies and Ward, 1980), amplitude
sharing (Walker, 1979), a minimum redundancy synthesis technique (Rossouw et al., 1997),
integer linear programming (Holm and Elgetun, 1997), or the effective aperture method
(Lockwood and Foster, 1996), have been proposed. The use of neural networks and similar
strategies to thin an array has also been researched (Mulholland et al., 1995). Recently,
ant colony optimization (Quevedo-Teruel and Rajo-Iglesias, 2006; Mosca and Ciattaglia,
2006), pattern search algorithms (Razavi and Forooraghi, 2008) and the immune algorithm
(Jianfeng et al., 2008) have been reported to obtain excellent results. However, the most
widely used optimization technique in array thinning is the GA (Goldberg, 1989). Its
inherent nature to represent an element which is turned on (1) or off (0) in a thinned
array makes the GA an attractive choice. In this sense, thinning a fully-filled array is
usually a simpler strategy than a direct sparse aperiodic array synthesis (Haupt, 1994).
Since they were firstly introduced in this topic in the 1990s (Haupt, 1994; O’Neill, 1994;
Buckley, 1996), countless works have relied on GAs to carry out an efficient thinned and
107
108 Array thinning with a nested optimization scheme
sparse array synthesis. In recent years, it can be found in (Chen et al., 2007; Vigano et al.,
2007; Spence and Werner, 2006; Hebib et al., 2006).
In most works, the design objective of a thinned or sparse array is related to the SLL
of the field pattern. Thinning strategies have seldom been combined with the synthesis
of shaped beam patterns (Kim and Nespor, 1996) or other objectives. Analogously, most
array thinning methods are based on the fact that all the active elements (elements which
are turned on) in the thinned array have the same excitation coefficients. Obviously,
maintaining such a simple excitation distribution can be a priority in an overall system
design to have more flexibility or ease in other design steps, i. e., in the design of the feeding
network. On the other hand, adding the possibility of varying the excitation coefficients in
the elements which are turned on may result in a configuration with fewer active elements
and therefore achieving the possible objective of a greater degree of thinning. However,
this is not an easy problem and it usually requires the use of a sophisticated strategy.
Some combined layout and excitation optimization methods to design thinned arrays can
be found in (Walker, 1979; Holm and Elgetun, 1997; Trucco, 1999).
In this chapter, an electromagnetic-based procedure to thin an antenna array is
developed. It allows the inclusion of real antenna element patterns and mutual coupling
effects. The proposed method yields the active elements from a fully-filled array, as well as
their excitations, needed to meet some prescribed design specifications. In order to carry
out the thinning a GA is used, while computing the excitations is carried out through
LP or QP within the fitness function of this GA. Thus, the resulting thinned array is the
one which complies with the requirements with the minimum number of active elements.
Specifications such as achieving a minimum gain in a given direction, obtaining a pattern
with a maximum allowable SLL or synthesizing a shaped beam pattern confined into a
mask are considered. Null field directions can also be added to any of these problems.
7.2 Thinned arrays using the GSM analysis

The N -element array characterized by the multimode multiport network shown in Fig-
ure 3.2 is now thinned in order to have Nth active elements, while the remaining N − Nth
are either turned off (not excited) or removed from the array. Thus, the first step is defin-
ing which type of thinning is desired. With a basic array theory approach from which
interelement coupling effects are not taken into account, both types of thinning are equiv-
alent, since an antenna placed in an array environment has the same characteristics as the
isolated antenna. This is the case of arrays made up of ideal isotropic sources, for which
most of the available thinning procedures are conceived. However, with the GSM analysis
of the array, the two types of thinning will yield different results not only concerning the
absence or presence of non-active elements, but also regarding the location and excitations
of the active elements, especially in arrays where coupling between elements is relevant
(i.e. with small interelement distances). Therefore, the main features of a thinned array
need to be calculated for both types of thinning. To that effect, expressions similar to the
ones in the previous section for the radiated field and power of a thinned array must be
derived so that:
E(θ, ϕ)=hg(θ, ϕ)ith {v}th (7.1)

Prad =4π{v}H
th [BR ]th {v}th (7.2)
H
Pdel =4π{v}th [BD ]th {v}th (7.3)
7.3 Nested optimization scheme for thinning 109
where {v}th are the excitations of the active elements and the subindex th refers to a
specific thinned array realization. The definitions of hg(θ, ϕ)i [BR ] and [BD ] are derived
in Chapter 4, from (4.3) to (4.8).
Turning elements off: when the thinning is achieved by turning elements off, the
number of total array elements and thus their location is always the same. Therefore,
the transmission and reflection matrices of the GSM of the array have to be computed
only once since they remain the same for any thinned array realization. So, the following
dependences, explicitly expressed in parentheses, can be established (dropping the θ, ϕ-
variation notation of hgi):
hgith = hgith ([T]) = hgith (hgi) (7.4)

[BR ]th =[BR ]th ([T])=[BR ]th ([BR ]) (7.5)
[BD ]th =[BD ]th ([R])=[BD ]th ([BD ]) (7.6)
In fact, as the case of turning elements off is equivalent to setting to 0 those excitations
in {v} which are not present in {v}th , it suffices to remove the columns and rows corre-
sponding to these non-active elements in [BR ] and [BD ] to compute [BR ]th and [BD ]th ;
analogously, removing from hgi the entries corresponding to the non-active elements yields
hgith straightforwardly.
Removing elements: in this case, every thinned array realization is a different array,
since the total number of elements and their location varies. Therefore, the GSM will also
be different for every thinned array configuration, and the aforementioned dependences
must now be written as:
hgith = hgith ([T]th ) (7.7)

[BR ]th =[BR ]th ([T]th ) (7.8)
[BD ]th =[BD ]th ([Γ]th ) (7.9)
where [T]th and [Γ]th stand respectively for the transmission and reflection matrices of
the GSM of the specific thinned array realization. Therefore, the whole process described
in Chapter 3 must be carried out. Note that this option may be computationally much
more expensive in an optimization scheme since it also involves the calculation of the array
transmission [T]th and reflection [Γ]th matrices for each thinned array considered, which
depends on the array analysis model being used (full-wave, circuital, hybrid, . . . ).
7.3 Nested optimization scheme for thinning

A flowchart of the proposed nested optimization scheme is shown in Figure 7.1. As can
be seen, this flowchart is independent of the optimization procedures which are actually
implemented. In Figure 7.1, each of the two boxes delimited by a dashed line represents
a generic optimization process, where in every iteration the values of the optimization
variables are computed and then the result associated to these values is evaluated in order
to decide whether more iterations are needed. In the proposed scheme, the optimization
corresponding to the outer box delimited by a dashed line must decide which antenna
elements should be removed or turned off (depending on which kind of thinning is being
considered). The optimization corresponding to the inner box delimited by a dashed line
must compute the excitations of the remaining active elements. Therefore, the result
which is processed in every iteration regarding the elements to be removed or turned off
depends, on the one hand, on the previous iteration, which has chosen which elements must
be active and, on the other hand, on the final result of an optimization of the excitations
of these active elements. Thus, it is of capital importance that a proper combination of
the objectives in the two optimizations be made in order to achieve the desired design
specifications and to ensure that this is accomplished with the minimum number of active
elements. This fact will be addressed carefully in the following section.
This scheme allows not only dealing with a great number of variables (antenna
elements and their excitations) but also dealing with variables which are different in type
(discrete and continuous). It should be noted that this can be a difficult task with a
direct optimization approach, requiring complex formulations and algorithms (Holm and
Elgetun, 1997; Trucco, 1999). So, with the proposed scheme, the algorithms that are best
suited to the specific nature of each of the two types of variables can be used separately.
Following this reasoning, we have chosen a GA to carry out the optimization corresponding
to the thinning, since as previously said it has already proven to be a reliable and efficient
approach. The implementation of the GA used is the same as the one in Chapter 2.
Therefore, the optimization of the excitations must be made within the evaluation of the
fitness function of each combination of active elements (each individual in the GA).
7.4 Fitness function

7.4.1 Definition
With the proposed approach, the FF of the GA in the nested optimization scheme must
have the maximum value when the required design specifications are met with the mini-
mum number of active elements. Furthermore, it is logical that if the design specifications
are not fulfilled with a given thinned array realization, the FF associated to it must always
be lower than that of any other thinned array realization that meets the design specifi-
cations. Therefore, the proposed FF is defined as the following piecewise function which
complies with the previous conditions:
if design specifications are met then

F F = 2N − Nth (7.10a)
else
F F = Nth (7.10b)
end
Figure 7.2 shows a sketch of the value of the FF versus the number of active elements
Nth in a thinned array. The minimum number of active elements for which the design
∗ . For thinned arrays with a number of active
specifications are fulfilled is denoted by Nth
elements lower than this optimum value Nth ∗ , the FF increases as the number of active
elements becomes higher, but it is always far below the maximum value 2N − Nth ∗ since
the design specifications are not met.

It should be borne in mind that there are N !/Nth !(N − Nth )! possible thinned arrays
which have Nth active elements, so the solution of the GA will be the only one(s) from the
∗ active elements that meet the design specifications. In fact, there
possibilities with Nth
may be thinned arrays with Nth ∗ or more active elements that do not meet the required
specifications (whose FF is represented with a circle in Figure 7.2). Analogously there are
7.4 Fitness function 111
Figure 7.1: Proposed nested optimization scheme.

Figure 7.2: Sketch of the value of fitness function (specified at each point) vs. the number of
active elements in a thinned array.
thinned arrays with more than Nth ∗ active elements that do fulfill the specifications (whose
FF is represented as a triangle in Figure 7.2). Their associated FF is lower than the one
for the optimum configuration with Nth ∗ active elements but higher than any thinned array
realization that does not comply with the design requirements. Therefore, the GA will
eventually form a population with the thinned arrays which hold the highest FF values
depicted in Figure 7.2.
So far, a general framework consisting of a GA has been provided for the imple-
mentation of the thinning in the nested optimization procedure. However, the inclusion
of the design specifications and their relation to the optimization of the active elements’
excitations have not yet been accomplished. This optimization must be carried out within
the evaluation of the FF in a way that it is ensured that the resulting excitations are
somehow the best so that the given thinned array realization meets the design specifi-
cations. Therefore, it is greatly dependent on the design problem. In this section, we
propose the procedure to tackle (with the minimum number of active elements) three
common design problems: achieving a specified minimum array gain in a given direction;
achieving a pattern with a specified maximum allowable SLL given a main beam direc-
tion and beamwidth; and finally, achieving a shaped beam pattern that remains within
a given mask. Constraints regarding interference rejection (null field) directions are also
incorporated into these problems. These design problems are developed next.
7.4.2 Design for a minimum gain

In this case, the design objective is achieving a minimum specified gain in the direction of
the main beam θ0 , ϕ0 with the minimum number of active elements. Therefore, it must be
ensured that the optimized excitations provide the maximum achievable gain for a given
thinned array. For a thinned array characterized by (7.1)-(7.3), the directive GD or power
7.4 Fitness function 113
GP gain can be expressed from (7.1) and, respectively, (7.2) or (7.3) as:
{v}H H
th hg0 ith hg0 ith {v}th
GD = (7.11)
{v}H th [BR ]th {v}th
{v}th hg0 iH
H
th hg0 ith {v}th
GP = H
(7.12)
{v}th [BD ]th {v}th
where g0 = g(θ , ϕ ). Maximizing the ratio of two Hermitian forms as in (7.11) or (7.12)
is a well-known optimization problem with a direct solution {v}∗th given by:
H
{v}∗th = [B]−1
th hg0 ith (7.13)
with B being either BR or BD .

Furthermore, if any set of desired directions θi , ϕi for i = 1, . . . , I wants to be
specified as interference rejection (null field) directions, the problem can then be stated
as a QP problem in the following way: minimize with {v}th the total radiated/delivered
power subject to keeping a fixed value κ0 of the radiated field in the main beam direction
and setting to zero the field in the interference rejection directions. With the formulation
used in this thesis, it takes the form:
minvth {v}H
th [B]th {v}th (7.14a)
s.t. hg0 ith {v}th = κ (7.14b)
hgi ith {v}th = 0, i = 1 . . . I (7.14c)
In this case, a direct solution for this optimization problem is derived from the QP theory
as:
H H −1
{v}∗th = [B]−1 [G][B]−1

th [G] th [G] {f } (7.15)
where [G] is a matrix where the first row is hg0 ith and the rest of the rows are hgi ith , i =
1, . . . , I, while {f } is a column vector with I + 1 entries whose first element is equal to κ0
and the rest are set to 0.
7.4.3 Design for a maximum SLL

In this case, the objective design is synthesizing, with the minimum number of active
elements, a pattern scanned at θ0 , ϕ0 which has a maximum allowable SLL outside the
mainlobe. Therefore, it must be ensured that the optimized excitations provide the min-
imum maximum SLL in this region for a given thinned array. So, the general problem,
including the null field directions θi , ϕi , is stated in (7.16) as: minimize with {v}th the
maximum absolute field value in the directions θm , ϕm , m = 1 . . . M , where this sampling
outside the mainlobe must be dense enough to cover all possible peaks, subject to a fixed
value κ0 in the main beam direction and to a zero-valued field in the desired null directions.
minvth maxm |hgm ith {v}th |, m = 1, . . . , M (7.16a)

s.t. hg0 ith {v}th = κ (7.16b)
hgi ith {v}th = 0, i = 1...I (7.16c)
The solution to this minimax problem is given in section 5.1. It requires, as a first step,
the removal of the norm in (7.16a). However, as we are dealing with complex-valued fields
and variables this cannot be carried out straightforwardly. Thus, a real-valued equivalent
minimax formulation is derived by taking the real and imaginary parts of both the excita-
tions and the field, thus doubling the number of variables. Then, a transformation into a
LP problem is carried out. The resulting real-valued LP problem can be solved with any
of the existing LP methods.
7.4.4 Design for a shaped beam pattern

In this design option, the objective is to achieve a shaped beam pattern that is a best fit of
a prespecified ideal response Eid (θ, ϕ) with the minimum number of active elements. With
this aim, the technique proposed in (Ng et al., 1993) is adapted to the present formulation.
A MSE minimization approach is followed:
ZZ
minvth |E(θ, ϕ) − Eid (θ, ϕ)|2 dΩ (7.17)
where the integration must be carried out in the region(s) where the MSE is desired to be
minimized. Considering the fully-filled N -element array, the MSE can be expressed from
(7.17) as (Ng et al., 1993):
{v}H [Q]{v} − {P}H {v} − {v}H {P} + f (7.18)
where [Q], {P} and f are computed as:

ZZ
[Q]= hg(θ, ϕ)iH hg(θ, ϕ)idΩ (7.19a)
ZZ
{P}= hg(θ, ϕ)iH Eid (θ, ϕ)dΩ (7.19b)
ZZ
∗
f= Eid (θ, ϕ)Eid (θ, ϕ)dΩ (7.19c)
Therefore, for a thinned array realization, computation of the matrices [Q]th and
{P}th depends on which of the two types of thinning is considered. In the case of turning
elements off, simply removing rows and columns corresponding to non-active elements in
matrices [Q] and {P} of the fully-filled N -element array will yield the result. If the option
is to remove elements, then the calculations in (7.19a) and (7.19b) must be carried out
starting from the transmission matrix [T]th corresponding to the specific thinned array.
Finally, as the contribution of f to the MSE is independent of {v}th , the resulting
QP problem, including the constraints regarding the null field directions θi , ϕi , is expressed
in (7.20) as: minimize with {v}th the MSE subject to setting a zero-valued field in the
interference rejection directions.
minvth {v}H H H
th [Q]th {v}th − {P}th {v}th − {v}th {P}th (7.20a)
s.t. hgi ith {v}th = 0, i = 1...I (7.20b)
From QP theory, a straightforward solution for this optimization problem can be

derived as:

H H −1
{v}∗th = [I] − [Q]−1 −1
[Q]−1

th [A] [G][Q] th [G] [G] th {P}th (7.21)
where [G] is a matrix whose rows are hgi ith , i = 1, . . . , I.

7.5 Numerical examples and results 115
7.4.5 Flowchart of the fitness function

The whole flowchart implemented to calculate the value of the FF for a given individual
of the GA (thinned array realization) is shown in Figure 7.3. The first step is to compute
the necessary matrices (needed in the design option) according to the thinning type being
used. In the case of turning elements off, the matrices corresponding to the fully-filled
N -element array have already been computed at the beginning of the GA and stored in
memory.
The second step is to carry out the optimization of the active element excitations
according to the design option. Then, the criterion of compliance is checked. For the
minimum gain design, it must be checked that the resulting gain G0 (which stands either for
the directive or power gain) is greater than or equal to the minimum specified gain Gmin . In
the case of the maximum allowable SLL design, the resulting SLL in every θm , ϕm direction
outside the the mainlobe, SLLm , must be lower or equal than the maximum specified
SLL, SLLmax . In the case of synthesizing a shaped beam pattern, then the absolute value
of the radiated field Es in every direction θs , ϕs , for s = 1, . . . , S, corresponding to the
shaped-beam region, must be confined to the mask defined by the minimum and maximum
allowable values |Emin,s | and |Emax,s |.
Finally, once the individual has passed the specific design criterion, it is checked, if
needed, that the null level N Li in every null field direction θi , ϕi is lower or equal than
a maximum prespecified level N Lmax needed to be considered as a null field to practical
effects. As previously defined in (7.10), if the set of specifications is met then the value of
the FF will be F F = 2N − Nth ; otherwise, F F = Nth .
7.5 Numerical examples and results

7.5.1 Computational aspects
All computations are carried out on a laptop computer with an Intel Core Duo at 2.26
GHz and 3 GB RAM.
The crossover probability is set to 50% in the GA implementation. Probabilities
for the jump and creep mutations are set to 1% each. When a planar array is considered
(including the isotropic one), the population size is 100 individuals and the number of
generations, unless otherwise stated, is set to 26. In the case of linear arrays, the population
size is 50 individuals and the number of generations is 51. The number of required FF
evaluations equals the population size multiplied by the number of generations.
Sampling of the θ-regions is carried out with a step of 1◦ . All times reported are
overall execution times from the initial matrix computations to the end of the GA (in-
cluding CPU time, I/O time, . . . ). It should be borne in mind that in the maximum SLL
design option an LP problem with no straightforward solution is solved in each FF evalu-
ation. This optimization problem has a number of variables which is double the number
of active elements in the individual (real and imaginary parts of the excitations), plus two
specific variables needed to carry out the minimax-to-LP problem transformation. The
number of constraints (excluding the null pointing directions) is four times the sampling
points used in the region outside the mainlobe. This usually results in larger times for this
design option, as is also the case when the thinning option is chosen to be the removal of
elements, since the transmission and reflection matrices of the thinned array have to be
calculated in each FF. Furthermore, the actual implementation of this thinning option is
Figure 7.3: Flowchart of the computation of an individual’s fitness function.

Figure 7.4: Field patterns, in ϕ=0◦ , of the isotropic 20×10 thinned arrays achieving -20 dB of
maximum SLL.
based on an intensive file handling, which requires a high hard disk I/O time.
Finally, the FE simulation carried out (only once) to yield the isolated antenna GSM
of the two antennas used as array elements in the examples takes about 30 s for each of
them, as said in Chapter 4.
7.5.2 Isotropic array thinning

In the first example, a broadside 20×10 planar array made up of isotropic sources eq-
uispaced at 0.5λ0 is considered. The thinning of this configuration and a direct sparse
array synthesis with the same number of active elements in the same space are carried out
respectively in (Haupt, 1994) and (Chen et al., 2007) with the use of GAs. The objective,
represented with the same FF in both works, is to obtain the lowest possible maximum
SLL in the principal planes (ϕ=0◦ and ϕ=90◦ ). In these works, no optimization of the
excitations is carried out since all radiators are equally excited.
In (Haupt, 1994), an approximate -20 dB maximum SLL in both principal planes
(ϕ=0 and ϕ=90◦ ) is achieved with a 46% thinning (108 active elements). Therefore, this
◦
array has been thinned with the proposed nested optimization scheme to achieve such
a maximum SLL with the minimum number of active elements. The region outside the
mainlobe has been set as θ ∈ (6◦ ,90◦ ) in ϕ=0◦ and θ ∈ (12◦ ,90◦ ) in ϕ=90◦ , so that the
mainbeam width is similar to the one obtained in the reference in both planes. As in
(Haupt, 1994), only one quarter of the array is considered because of symmetries. Figures
7.4 and 7.5 show the patterns obtained in ϕ=0◦ and ϕ=90◦ with the proposed methodology
and the ones obtained in (Haupt, 1994) for comparison purposes. The nested optimization
procedure takes approximately 6 min. Figure 7.6 shows the layout of the active elements in
both cases for the upper right quarter of the array when it is centered in x=0 and y=0. As
can be seen, with the proposed methodology, a higher level of thinning (68%, equivalent to
64 active elements) is achieved at the expense of a non-uniform set of excitations feeding
Figure 7.5: Field patterns, in ϕ=90◦ , of the isotropic 20×10 thinned arrays achieving -20 dB of
maximum SLL.
Figure 7.6: Layouts of the active elements needed to achieve the patterns in Figs. 7.4 and 7.5 of
the isotropic 20×10 thinned array.
Figure 7.7: Field pattern of the isotropic 20×10 array thinned to achieve -30 dB of maximum
SLL.
the active elements.

In (Chen et al., 2007), synthesizing a sparse array by directly considering 108 active
elements and arranging their locations in the same space resulted in a field pattern with
a maximum SLL of approximately -30 dB in ϕ=0◦ and -16 dB in ϕ=90◦ . We have chosen
the most restrictive result (-30 dB) to be fulfilled in both principal planes by a thinned
array starting from the original 20×10 isotropic array. The region outside the mainlobe
has been set as θ ∈ (10◦ ,90◦ ) in ϕ=0◦ and θ ∈ (20◦ ,90◦ ) in ϕ=90◦ . The pattern obtained
with the proposed methodology is shown in Figure 7.7. As before, the optimization takes
approximately 6 min. The layout of the resulting active elements is shown in Figure 7.8
together with the one obtained in (Chen et al., 2007). Only 56 elements need to be excited
(72% thinning). Once again, incorporating the possibility of varying the excitations of the
active elements produces a smaller number of active elements.
7.5.3 Linear array thinning

In the following example, the GSM analysis of a 31-element DRA array is integrated in
the optimization. The antenna element used in the array is the probe-fed hemispherical
DRA taken shown in Figure 3.4. DRA elements are placed along the E-plane (x-direction
in Figure 3.4) with an interlement distance of 0.4λ0 at 3.64 GHz. To show the effective-
ness of the proposed method, the array is thinned (by turning elements off) to achieve,
with the minimum number of active elements, different maximum SLL specifications with
different mainlobe widths. The resulting patterns are shown in Figure 7.9. The crosspolar
component of the field is negligible in all cases. In the first case, the mainlobe region is
set to θ ∈ (-10◦ ,10◦ ) and the desired maximum SLL outside that region to -30 dB. This
requirements are fulfilled with 14 active elements. In the second case, a more restrictive
specification regarding the maximum SLL is set (-40 dB). Therefore, more active elements
(up to 21) are needed to fulfill it. Finally, the mainlobe is widened to θ ∈ (-20◦ ,20◦ ), while
Figure 7.8: Layout of the active elements needed to achieve the pattern in Figure 7.7 of the
isotropic 20×10 thinned array compared to the layout of a directly sinthesized sparse array of 108
isotropic elements in (Chen et al., 2007).
keeping the same -40 dB maximum SLL restriction. In this last case, the number of active
elements required to meet the specifications is 13. The nested optimization procedure
takes approximately 9 min in each case.
Now, the same array is thinned (by turning elements off) to synthesize a shaped-
beam pattern with the minimum number of active elements. A flat-top beam pattern
between θ=−20◦ and θ=20◦ is considered, having two null field pointing directions at
θ=−60◦ and θ=45◦ . The mask is set so that a maximum ripple of 1 dB is allowed in the
flat-top beam region with a maximum SLL of -20 dB in the rejection region. Two different
cases for two different transition regions between the flat-top beam and the rejection region
are considered. In the first case, the transition regions are set to θ ∈ (-25◦ ,-15◦ ) and θ ∈
(15◦ ,25◦ ), while in the second case, this specification is made more restrictive by setting
them to θ ∈ (-23◦ ,-17◦ ) and θ ∈ (17◦ ,23◦ ). The resulting patterns are shown in Figure 7.10,
where it can be seen how the maximum SLL and null level (set to -100 dB) restrictions
are fulfilled. As before, the crosspolar component of the field is negligible in both cases.
In the first case, only 18 active elements (42% thinning) are required, while in the more
restrictive second case this number is 26 (16% thinning). Figures 7.11 and 7.12 show a
zoom of these two patterns, and it can be seen how they both strictly comply with the
specifications set in the mask for the flat-top beam and the transition region in each case.
As in this problem there is a straightforward solution for the optimized excitations when
computing the FF of each individual, the nested optimization procedure takes less than 2
s in each case.
We will now consider two different 17-element linear arrays. In the first array, the
antenna element used is the cavity-backed, coaxially-fed microstrip circular patch with
superstrate shown in Figure 4.1. The patches are placed along the E-plane (x-direction
in Figure 4.1) with an interelement distance of 0.6λ0 at the resonant frequency, 1.215
GHz. Both types of thinning are applied to this array to yield a field pattern scanned
Figure 7.9: Field patterns of the 31 DRA array thinned to achieve different maximum SLL and
mainbeam width restrictions with null field directions.
Figure 7.10: Field patterns, for two different transition regions, of the 31 DRA array thinned to
synthesize a shaped flat-top beam response with restrictions.
Figure 7.11: Detail of the field pattern of Figure 7.10 for the flat-top beam response with a
transition region of θ ∈ (±15◦ ,±25◦ ).
Figure 7.12: Detail of the field pattern of Figure 7.10 for the flat-top beam response with a
transition region of θ ∈ (±17◦ ,±23◦ ).
Table 7.1: Optimized excitations of the thinned array realizations of two different 17-element
arrays for both thinning options X: element removed, 0: element turned off
at θ0 =30◦ meeting the requirement of a maximum SLL of -15 dB in θ ∈ (-90◦ ,20◦ ) and
θ ∈ (40◦ ,90◦ ). With this array configuration and with this antenna element, it is expected
that interelement coupling effects play a minor role and therefore both types of thinning
achieve similar solutions. By turning elements off, the GA produces 13 individuals, with
a number of active elements equal to 10, that meet the desired requirements. However, by
ignoring mutual coupling effects based on the aforementioned assumption of weak interele-
ment coupling, these 13 individuals can be reduced to 7 different relevant thinned array
configurations, which are obtained by not taking into account the presence of consecutive
turned off elements at the beginning and at the end of the array (ignoring leading and
trailing 0’s in the individual). In the case of removing elements the proposed method
yields two different thinned array configurations which fulfill the requirements with 10
active elements. As expected, these two configurations are also present in the 7 solutions
reached when the option of turning elements off is used. Table 7.1 (top) shows which are
the active elements and their corresponding excitations for one of these configurations. It
can be seen how the required excitations are practically identical.
Now, the former DRA is used as an array element in the 17-element array configura-
tion along the E-plane. An interelement distance of 0.4λ0 is considered. In this case, the
effects of interelement coupling are expected to exert a significant influence, so the results
obtained with both thinning options should be different. With the option of turning ele-
ments off, the nested optimization procedure yields 5 different thinned array realizations
which meet the requirements having 11 active elements. None of these configurations fulfill
the specifications if the non-active elements are removed instead of turned off, even if the
excitations of the remaining active elements are optimized. On the contrary, by using the
option of removing elements, the GA finds 7 solutions, with 11 active elements, that fulfill
the requirements. Therefore, it is demonstrated how the proposed optimization method
effectively takes interelement coupling into account. Table 7.1 (bottom) shows which are
the active elements (and their corresponding excitations) for one of the configurations
obtained in the case of turning elements off, and in the most similar configuration to this
one achieved when the thinning option is removing elements. To facilitate the comparison,
the excitation of the first active element is set to 1∠0◦ , so all amplitudes and phases of
the rest of the excitations take it as a reference. As can be seen, the locations and excita-
tions of active elements are different in both thinned arrays. However, Figure 7.13 shows
the patterns obtained in these two optimized cases (non-active elements are removed or
turned off) in Table 7.1. It can be seen how both patterns strictly comply with the initial
Figure 7.13: Field patterns of a 17-DRA array thinned with both types of thinning to achieve a
-15 dB maximum SLL with θ0 =30◦ (11 active elements).
specifications. The nested optimization procedure takes approximately 30 min in the case
of removing elements and about 3 min when the thinning option is turning elements off.
7.5.4 Planar array thinning

In the following example, a planar array with the same number of elements and the
same interlement distance in the x and y-directions is considered. It is a 6×6 array
made up of DRAs like the one shown in Figure 3.4. The interlement distance is 0.4λ0
at the resonant frequency. The maximum attainable directive gain at broadside direction
of this array is 19.3 dBi. Both types of thinning (turning elements off and removing
elements) are carried out to achieve a directive gain of 16 dBi at broadside direction
with the minimum number of elements. In the case of removing elements, the nested
optimization procedure takes approximately 1 hour and yields 5 different configurations
for which the aforementioned value can be achieved with 12 active elements. On the other
hand, in the case of turning elements off the GA has been run for up to 200 generations
and the best achieved configurations have 13 active elements (two different configurations).
The time taken when turning elements off is just 5 s. Figure 7.14 shows the active elements
for the two most similar thinned arrays obtained with the two thinning techniques. It can
be seen how they are different not only in number but also in the locations of the active
elements. In this case, the thinned array obtained by removing elements yields a directive
gain of 16.2 dBi, while the one obtained by turning elements off has a directive gain of 16.01
dBi. Once again, it is demonstrated how the presence or absence of an antenna element
is crucial when carrying out the thinning if interelement coupling effects are significant.
Finally a 9×9 planar array made up of cavity-backed circular patches like the one
shown in Figure 4.1 is considered. The interelement distance (equal in the x- and y-
directions) is 0.4λ0 at the resonant frequency. We have carried out a wide range of opti-
mizations (each of which takes from 2 s to 5 s approximately) to compute the minimum
Figure 7.14: Layout of the active elements needed to achieve a broadside gain of 16 dBi by a
thinned 6×6 DRA array with the two thinning options.
number of active elements required to achieve a prescribed directive gain in a given main
beam direction. Thinning is carried out by turning elements off. The results are shown in
Figure 7.15. As expected, the number of active elements required increases as the desired
directive gain specification is raised. Each plot (corresponding to a different main beam
direction) ends when the maximum attainable of the fully excited array is reached. It is
observed how achieving a given increase of the gain near the end of the plot requires a
greater increase on the number of active elements than in the case of achieving the same
gain increment for lower directive gain values.
This array is now thinned (by turning elements off) to achieve a broadside pattern
with a maximum SLL of -20 dB in the whole space. To this effect, the sidelobe region has
been defined as θ ∈ (25◦ ,90◦ ), with a 1◦ -sampling, and ϕ ∈ (0◦ ,360◦ ), with a 10◦ -sampling.
The nested optimization takes about 14 h to reach three configurations that meet these
requirements with 34 active elements (58% thinning). The three-dimensional field pattern
of one of these thinned array realizations is shown in Figure 7.16.
Figure 7.15: Number of active elements required to achieve a specified directive gain at different
mainbeam directions θ0 , ϕ0 of the 9×9 cavity-backed patch array.
Figure 7.16: Field pattern of the 9×9 cavity-backed patch array thinned to achieve a broadside
pattern with a maximum -20 dB SLL in the whole space outside the main beam (34 active elements).
Chapter 8
Modelling
Multiple-Input-Multiple-Output
systems with a hybrid
spherical-mode-based approach
8.1 Line-of-sight SISO and MIMO systems

8.1.1 Introduction
MIMO systems have drawn the attention of the research community for the last years.
A very fruitful and interesting current line of research is focused on the effects of mutual
coupling on signal correlation and therefore, on the capacity of a wireless MIMO system.
It has been shown that interelement coupling effects at the transmitting and/or the re-
ceiving array can play a major role in MIMO communication performance (Wallace and
Jensen, 2004), (Janaswamy, 2002). Apart from that, one of the usual assumptions taken
for granted in most MIMO wireless communication models is the plane-wave nature of the
propagating electromagnetic signal. When the transmitting and receiving arrays are far
away from each other and from scatterers, this does not pose any problems. However, the
real spherical nature of the propagating field may exert a significant influence on the char-
acteristics of short-range MIMO systems. In this section, we propose an electromagnetic-
based model to rigorously characterize line-of-sight MIMO systems. Current computa-
tional techniques based on rotation and translation coefficients of spherical modes allow
now all the aforementioned effects regarding mutual coupling and spherical vector prop-
agation to be included in a MIMO system model. Furthermore, the real features from
antennas characterised by means of a GSM are also included to account for the matching,
transmitting, receiving and scattering antenna response in an array environment. Nu-
merical validation is accomplished by showing results from a Single-Input-Single-Output
(SISO) system and different MIMO systems made up of ideal dipoles.
8.1.2 Single-Input-Single-Output system

The GSM of an antenna i relates the complex amplitude of the input vi and the output
wi signal at the feeding port (assuming
out one single mode of excitation) with coefficients
inc

of incoming ai and outgoing bi spherical vector modes used to characterize the
127
128 Modelling MIMO systems with a hybrid spherical-mode-based approach
radiating region as:

wout
i
=
[ρi ] [ri ] vinc
i
(8.1)
bi [ti ] [si ] ai
As was done in Chapter 3, incoming spherical vector waves at the receiving antenna
(i=2) can be related by means of a GTM (Hansen, 1988) to scattered (outgoing minus in-
coming) spherical modes at the transmitting antenna (i=1). A reciprocal relation can also
be established. Thus, with Gi,j standing for the corresponding GTM between transmitter
and receiver and viceversa, we have:
ainc = [G2,1 ] bout

inc
ainc = [G1,2 ] bout
inc
2 1 − a1 , 1 2 − a2 (8.2)
By inserting (8.2) in (8.1) for i = 1, 2 and solving the matrix system, the 1×1
channel matrix, which relates the input signal at the transmitter with the output signal
at the receiver as w2 = HSISO v1 , can be expressed as (Hansen, 1988):
HSISO = [r2 ] ([I] − [G2,1 ] ([s1 ] − [I]) [G1,2 ] ([s2 ] − [I]))−1 [G2,1 ] [t1 ] (8.3)
To validate the model, let us consider a SISO system made up of two ideal dipoles
facing each other, both parallel to the z-axis. The analytical expression for the ideal
electric dipole GSM is given in (Hansen, 1988). The computed value of HSISO with the
proposed model is shown in Figure 8.1 vs. the link distance.
√ It is compared with the
classic transmission equation HSISO = exp (−j2πr12 /λ0 ) D1 D2 λ0 /4πr12 , where r12 is
the distance between dipoles, λ0 is the free-space wavelength and Di the directive gain, in
natural units, of dipole i (=1.5). As expected, both models are in good agreement only
for large distances.
Figure 8.1: Channel matrix value for a SISO system of two ideal dipoles. - - - case 1: Classic
transmission equation. —– case 2: Proposed model.
8.1.3 Multiple-Input-Multiple-Output system

In this case a MIMO system with M transmitting antennas and N receiving antennas is
considered. In order to take MC effects into account, it must be borne in mind that the
incoming field at any antenna i (transmitter or receiver) results from the field scattered
by all transmitting and all receiving antennas, so that:
M
X +N
bout
inc
{ai } = [Gi,j ] j − aj (8.4)
j=1
8.1 Line-of-sight SISO and MIMO systems 129
where [Gii ] = [0]. By inserting (8.4) into (8.1) for i = 1, . . . , M + N , a direct relationship
between the input {v} = hv1 , . . . , vM +N iT and output {v} = hw1 , . . . , wM +N iT signals of
every antenna in the MIMO system can be derived in as the reflection matrix of an array
made up of all the antennas composing the MIMO system (transmitters and receivers),
given by (3.18), so that {w} = [Γ] {v} By grouping the input and output signals according
to their membership to either the receiving or the transmitting array, so that
{v}=h{v}TRx , {v}TT x iT (8.5)

{w}=h{w}TRx , {w}TT x iT , (8.6)
and by setting the input signals at the receiving antennas to be null
{v}Rx = 0, (8.7)
it is demonstrated that the MIMO channel matrix [H]M IM O can be found as the upper
right N × M submatrix of this reflection matrix [Γ], since we have:

{w}Rx [Γ]RxRx [H]M IM O {v}Rx
= (8.8)
{w}T x [Γ]RxT x [Γ]T xT x {v}T x
Figure 8.2: Open-loop capacity for a 20 dB SNR of various-sized 4×4 MIMO system of ideal
dipoles. - - - case 1: Spherical-wave model in (Jiang and Ingram, 2005). —– case 2: Proposed
model.
Figure 8.3: Waterfilling capacity for a 20 dB SNR of a 2×2 dipole MIMO system. Link distance:
100λ0 . - - - case 1: Spherical-wave model in (Jiang and Ingram, 2005). —– case 2: Proposed
model.
As an example, a 4×4 MIMO system made up of ideal dipoles parallel to the z-axis
is considered. The interelement distance (between dipoles) is the same in the transmitting
Figure 8.4: Open-loop capacity for a 20 dB SNR of a dipole MIMO system with fixed-length
arrays (5λ0 ). Link distance: 100λ0 . - - - case 1: Spherical-wave model in (Jiang and Ingram,
2005). —– case 2: Proposed model.
and in the receiving array. The open-loop capacity, for a Signal-to-Noise Ratio (SNR) of 20
dB, obtained from our proposed method and from the spherical-wave model in (Jiang and
Ingram, 2005) is shown in Figure 8.2 for various-sized arrays vs. the link distance.
√ Nor-
malized channel matrices are calculated by multiplying the original ones by M N /kHkF ,
where kHkF is the Frobenius norm of the channel matrix obtained with the spherical-wave
model in (Jiang and Ingram, 2005). Three different interelement distances, and thus array
sizes, are considered. As expected, for larger interelement distances coupling is not too
relevant and both models are in good agreement. However, in the case of an interelement
distance of 0.25λ0 , mutual coupling effects in the transmitting and the receiving arrays
result in a severe reduction of the capacity. It should be noted that for very short link
distances, both models are always in disagreement as seen in the SISO example since in
this case coupling between transmitting and receiving antennas is relevant. Figure 8.3
shows the capacity, obtained with the waterfilling procedure with a 20 dB SNR, for a
2×2 MIMO dipole system with a fixed link distance of 100λ0 and various interelement
distances at the transmitting and receiving arrays (same for both). Again, it is seen how
as the interelement distance increases both models are in good agreement, since mutual
coupling becomes less relevant. For very short interelement distances, the proposed model
rigorously takes coupling effects into account to yield lower capacity values than those
predicted by the model in (Jiang and Ingram, 2005). Finally, we consider equal arrays at
the transmitter and the receiver with a fixed length of 5λ0 . Figure 8.4 shows the open-loop
capacity, for a 20 dB SNR, obtained from the proposed model and the model in (Jiang
and Ingram, 2005) for a link distance of 100λ0 when the number of antennas is increased
and therefore interelement distance becomes smaller. As stated in (Janaswamy, 2002), it
can be seen how increasing the number of antennas above a given threshold (in this case,
about 12) is futile since the capacity actually drops with respect to the case where mutual
coupling is not taken into account.
8.2 Modal network model for MIMO systems

8.2.1 Introduction
Although MIMO systems have been a subject of research for several years, there seems
to be no general agreement on the optimal antenna spacing, configuration, polarization
and other characteristics for a specific scenario. This is mainly because of the fact that
all works are done under different frameworks which in most cases are not comparable
or flexible enough. Moreover, attempts to provide results obtained with measured data,
even though significant in number, have also proven to be inflexible since in most cases
8.2 Modal network model for MIMO systems 131
these studies are limited to certain antenna configurations or scenarios. More importantly,
many measurement campaigns focus primarily on simulation validation rather than the
study of different MIMO systems and therefore are not conclusive.
A great number of antenna and propagation prediction models for MIMO wire-
less communications is currently available. Current computational techniques allow the
evaluation of complex phenomena, such as multi-path propagation. Network models for
MIMO systems have already been proposed in the literature for path-based channels
(Waldschmidt et al., 2004; Wallace and Jensen, 2004). In both cases a general frame-
work for the use of scattering parameters was presented for computation of the system
capacity focusing on mutual coupling effects. As a result, processing of all paths for each
antenna type under study was required. When considering a small number of antennas or
for propagation channels with few propagation paths, doing so poses no problem. How-
ever, in the case of multiple channel realizations (i.e. multiple links per channel), of several
thousands paths each, the computational cost of extensive antenna evaluations becomes
steep. To address this need, the inclusion of a modal expansion of the fields of the anten-
nas is proposed. In this manner multipath processing becomes significantly reduced and
study of multiple antennas and/or orientations becomes feasible.
8.2.2 Generalized network model

In this section the generalized network model of (Waldschmidt et al., 2004) will be used. It
consists of a network model of the whole transmission chain, i.e. transmitter unit (signal
source), transmit antennas, physical channel, receive antennas, and receiver unit (signal
drain) modeled as networks described by scattering matrices as seen in Figure 8.5 for an
antenna pair nm, where the S-Parameter matrix of the propagation channel SĈ is of the
form
!
SĈ (M xM ) SĈ (M xN )
SĈ = 11 12 . (8.9)
SĈ
21
(N xM ) SĈ (N xN )
22
with N and M being the number of receiving and transmitting antennas for the system
under test.
Transmitter Transmit Antenna m Physical Channel Receive Antenna n Receiver
Figure 8.5: Signal Flow Graph of Generalized Network Model.
With this model it is possible to take into account effects like signal correlation or
antenna coupling for capacity computations and, in addition, consider additional compo-
nents such as matching networks (Wallace and Jensen, 2004). However, in order to use the
previous model an accurate description of all S-Parameters needs to be given. In the case
of the antennas and the physical channel, this can only be done through measurements
or use of numerical methods that deterministically model the interaction between the an-
tennas and the scenario. This is only possible, however, with a path-based description of
the propagation, which results in a great computational cost and significant complexity.
Therefore, in order to simplify the understanding of the whole transmission chain, it makes
sense to use a joint network for the antennas and channel, as done in (Waldschmidt et al.,
2004), resulting in the extended channel of Figure 8.6.
Transmitter Channel Receiver
Figure 8.6: Signal Flow Graph of Generalized Network Model in its compact form, i.e. transmit
antennas, physical channel and receive antennas merged into one network.
This compact form of the generalized network model results from two assumptions:
(1) there is no back transmission S21 Ĉ = 0, i.e. unilateral propagation channel, and (2)
there are no scatterers in the proximity of both the transmitting and receiving antennas,
STxA
22 = 0 and SRxA
22 = 0 respectively. It thus follows that, after some algebraic transfor-
mations of the original model (Figure 8.5), the merged inner network can be expressed
as
!
C STxA
11 0
S = TxA Ĉ RxA RxA . (8.10)
S21 S21 S12 S11
Here, the term SC TxA Ĉ RxA results from the transmission coefficients be-
21 given by S21 S21 S21
tween all transmitting and receiving ports and constitutes the transmission submatrix of
the extended channel scattering matrix SC .
For the case of a signal transmitted from port m to port n under matched conditions
S21,nm yields the channel matrix H, i.e. the ratio between the voltage VnRx induced by
C
the incoming field E~ m and the excitation voltage V Tx , according to Figure 8.7, and with
m
antenna gains Gn and Gm :
s s s
Z0,m VnRx ℜ{ZnA }

C wn Z0,m λ
Hnm = S21,nm = = = G G
A} n m
γnm (8.11)
υm Z0,n VmTx Z0,n ℜ{Zm 4π
with the link factor γnm given by

Figure 8.7: Equivalent circuit of transmit and receive antennas.
P
~ Rx (θ Rx , ϕRx ) · Πp · C
~ Tx (θ Tx , ϕTx )
X
γnm = C n p,n p,n m p,m p,m (8.12)
p=1
where the sum is done over the total number of P paths of the link. In (8.12), C ~ nRx and
~m
C Tx are the normalized patterns of both the receiving antenna n and the transmitting
Rx , ϕRx and θ Tx , ϕTx are the directions of arrival and departure for
antenna m, while θp,n p,n p,m p,m
every path p and antennas n and m.
Finally, Πp in (8.12) is known as the full polarimetric transmission matrix (Wald-
schmidt et al., 2004), and it is usually a result of a numerical method (i. e. ray tracing)
or a measurement campaign.
To simplify the analysis, a plane-wave front is assumed since the receiver is supposed
far away from the transmitter and from scatterers and viceversa. In this way, (8.12) is
rewritten as:
P
~ Tx (θ Tx , ϕTx ) exp ψ(θ Rx , ϕRx )ψ(θ Tx , ϕTx )
~ Rx (θ Rx , ϕRx ) · Πp,nm · C
X
γnm = C n p,n p,n m p,m p,m p,n p,n p,m p,m
p=1
(8.13)
with:
2π
ψ(θ, ϕ) = ((x − xp ) cos ϕ sin θ + (y − yp ) sin ϕ sin θ + (z − zp ) cos θ) (8.14)
λ
where x, y, z is the position of the antenna (mth transmitter or nth receiver) and xp , yp , zp
the point where the path p departs/arrives.
Eqs. (8.11) and (8.13) shows together with Figure 8.7 that once a certain physical
Ĉ
channel S12,nm , expressed here through the full polarimetric transmission matrix Π, is
known for a certain nm communication link, the gains Gn and Gm and the radiation
patterns C ~ n and C~ m of both receive antenna n and transmit antenna m at the direction
of arrival and the direction of departure have to be included for every path p. Therefore,
for each combination of antenna types, orientations or polarizations of interest a new SC
has to be calculated. The latter greatly limits the amount of possible configurations that
can be studied and thus represents a great bottleneck common to both simulations and
measurements when dealing with multi-path propagation channels.
8.2.3 Modal network model

The main drawback of the previous modeling approach, used throughout the literature,
is that computation of the transmission matrix SC 12 for a certain configuration requires
the processing of every single path p for every communication link nm. In addition, if
it is considered that for statistic purposes in each studied scenario there are k channel
realizations, i.e. k possible antenna positions, with several thousand paths each, the
computational cost becomes restrictive.
A way in which this computational effort can be reduced is found by expressing the
antennas’ normalized radiation patterns as a weighted summation of spherical vector wave
functions (Hansen, 1988):
R
~ Rx(θ, ϕ) =
X
Cn cRx
r,n~
er (θ, ϕ) (8.15a)
r=1
Q
~ Tx (θ, ϕ) =
X
Cm cTx
r,m~
eq (θ, ϕ) (8.15b)
q=1
In this way the radiation patterns of both the transmitting and receiving antennas
in (8.11) can be rewritten according to (8.15) thus resulting in the following expression
C
for S21,nm and the link transmission coefficient γnm of the extended channel
C
S21,nm = κnm γnm (8.16)
Q
R X
X
γnm = cRx Tx
r,n cq,m γ̌rq,nm (8.17)
r=1 q=1
where γ̌rq,nm constitutes the rq subfunction of the link γnm defined as
P T
eθ,r (θpRx , ϕRx
p ) eθ,q (θpTx , ϕTx
p )
X
Rx
, ϕRx Tx Tx

γ̌rq,nm = Πp exp ψ(θp,n p,n )ψ(θp,m , ϕp,m )
p=1
eϕ,r (θpRx , ϕRx
p )
Tx Tx
eϕ,q (θp , ϕp )
(8.18)
where eθ and eϕ are the θ and ϕ polar components of the normalized spherical wave
functions, and
s s
Z0,m ℜ{ZnA }

λ
κnm = G G
A } n m 4π
(8.19)
Z0,n ℜ{Zm
is a multiplicative factor resulting from (8.11).
It follows from (8.16) that through the use of spherical wave functions, a pseudo
modal expression for the transmission coefficient of the extended channel has been found,
where the γ̌rq,nm functions build the scenario-dependent basis set. As a result, the extended
channel transmission coefficient of any desired antenna pair can now be expressed as a
weighted sum of basis functions for the specific scenario considered, assuming the number
of modes R and Q accurately describe the field of the antennas to be studied.
With this modal network model it is now possible to save valuable time in otherwise
very lengthy simulations and, more importantly, significantly increase the number of con-
figurations that can be studied. As the calculation of functions γ̌rq,nm, which is the main
computational burden, is carried out only once and then these functions are stored, every
change in the antenna configurations only requires the computation of coefficients cRx
r,n or
Tx
cr,m and then the evaluation of (8.16).
8.2.4 Characterization of the mobile station

So far, the presented model assumes there is a modal expansion of the fields in both the
transmitting and receiving antennas. However, it might be of interest to evaluate several
mobile station antennas referenced to a base station whose patter has been measured. In
this case (8.16) becomes
R
X Pl
X
C
S21,nm = κnm cRx
r,m γ̌r,nm = (8.20)
r=1 p=1
R P T
eθ,r (θpRx , ϕRx
p ) Cθ,m (θpTx , ϕTx
p )
X X
bRx Rx
, ϕRx Tx Tx

κnm r,m Rx Rx
Πp Tx Tx
exp ψ(θp,n p,n )ψ(θp,m , ϕp,m )
r=1 p=1
eϕ,r (θp , ϕp ) Cϕ,m (θp , ϕp )
where Cθ and Cϕ are the θ and ϕ components of the normalized radiation pattern of the
base station.
As an example, the effect of mobile station placement and polarization in an urban
scenario for ideal dipoles at the mobile station is investigated. In (Hansen, 1988) the
modal expansions of the radiated fields of dipoles oriented along the x, the y and the z
axis are given. Out of it, the coefficients needed to express an arbitrary oriented dipole
can be deduced to be:
√
2
c2 (σ, β) = sin(σ) (sin(β) + j cos(β)) (8.21a)
2√
2
c6 (σ, β) = −j sin(σ) (sin(β) − j cos(β)) (8.21b)
2
c4 (σ) = cos(σ) (8.21c)
ci = 0, i 6= 2, 4, 6 (8.21d)
where σ is the rotation angle with respect to the z axis, i.e. rotation angle in elevation,
and β is the rotation angle with respect to the x axis, i.e. rotation angle in azimuth.
Thus, for the ideal dipole, the nm link in (8.16) can be written as a function of σ and β as
well, i.e. γnm (σ, β), for which only three modes are needed in order to consider all possible
orientations, i.e. polarizations. Moreover, for the specific case of a vertically oriented
dipole, α = 0, only one mode is needed, as seen from (8.21).
As for the base station, the complex measured pattern a commercial antenna of the
antenna manufacturer Kathrein will be used for the investigation of optimum antenna
placement: the 742215 vertically polarized antenna. A 2x2 MIMO system is considered,
where the antennas inter element spacings at the base station dTx and mobile station dRx
are simultaneously varied between 0.5, 1, 2 and 5 λ.
In order to compute the full polarimetric transmission matrix, a double-directional
3-D ray tracing model based on a ray-tracing approach (Fugen et al., 2006) is used. The
model has been validated with wide-band channel measurements at 2 and 5.2 GHz showing
high accuracy for fixed to mobile and mobile to mobile communications in macrocell
scenarios. For the simulation, a digital description of the city of Karlsruhe by means of a
vector database will be used containing the exact position and size of buildings, trees and
other objects. It is generated based on two-dimensional plan information of the buildings
along with a digital height profile of the environment. For brevity only one base station
location is considered, which corresponds to a current location of a German provider. The
simulated area spans about 1 km in each direction. The base station antenna is positioned
30 m above street level. Mobile station locations, on the other hand, are given by a
17.7 × 17.7 m grid rotated 45◦ , enough for statistical purposes.
Three different power schemes: waterfilling, uniform power distribution and beam-
forming, will be used. In all cases the relative ergodic capacity gain with respect to a
SISO configuration will be investigated for a fixed SNR of 10 dB with channel matrices
normalized with respect to a reference MIMO system:
√
MN
Hnorm = H (8.22)
kHref kF
in order to fairly compare the different antenna setups regardless of received power. Where
kHref kF in (8.22) denotes the Frobenius Norm of the reference system, chosen here to be
the same 2x2 system under study but based on ideal omnidirectional radiating sources
instead.
Figure 8.8 shows the ergodic capacity gain in percentage vs varying dTx for all dRx
and all power schemes for two 742445 vertically polarized Kathrein antennas at the base
station with two vertically polarized dipoles at the mobile unit, i.e. only one spherical mode
is used as explained in the previous section. Out of it, considering that no mutual coupling
effects are taken into account, three main conclusions can be drawn: 1) relative capacity
gain is in general insensitive to changes in dRx , whereas bigger inter-element spacings
at the base station result in higher capacities, 2) increased base station spacing results
in greater signal decorrelation at the mobile station and is relevant only at dTx = 5λ,
2) a beamforming scheme where only the best available subchannel is used yields no
improvement regardless of dTx and dRx , whereas a uniform power distribution departs
from a beamforming-like performance at small dTx spacings towards an optimum power
allocation for increased base station spacings.
Now, the relative capacity gain plotted against rotation angle α for the base station
antennas will be shown, together with the two different mobile station antenna setups
seen in Figure 8.9: a) two parallel dipoles rotated with respect to the z-axis and b) two
perpendicular dipoles rotated with respect to their starting position as vertically and
horizontally polarized dipoles.
In Figure 8.10 the 742445 vertically polarized Kathrein antenna is studied together
with two parallel dipoles rotated around the z-axis. It is seen that there is a cosine-like
dependence on the angle α resulting in decreased capacity gain for all angles greater than
0◦ . Nevertheless, no significant influence of rotation angle on the behavior of dTx and
dRx is seen. In Figure 8.11 the same base station antenna with two perpendicular dipoles
is studied. In this case, the capacity dependence on varying dTx and dRx also remains
constant, but in addition almost no dependence with respect to rotation angle α is seen.
8.2.5 Base station characterization

So far, a measured pattern has been considered for the base station. However, it is also
desirable to have a modal expansion of the radiated field for the antennas of the base
Figure 8.8: Relative Ergodic Capacity Gain vs horizontal antenna inter element spacing at the
base station for two 742445 vertically polarized Kathrein antennas and two vertically oriented
dipoles
station in order to investigate different antenna types. This poses one major problem:
base stations are usually big antennas, so the number of spherical modes needed to have
their radiated field rigorously characterized may be prohibitive (Hansen, 1988; Rubio et al.,
2005) to be inserted into the proposed network model. In order to overcome this problem,
we can take advantage of the knowledge of some base station configurations, such as the
one of Kathrein, to expand the base station field into a number of pseudomodes. Here,
the main objective is to obtain a main beam width as close as possible to the measured
pattern. In this case, the major variation of the field from the base station will be in the
elevation plane, as it is mainly given by the array factor. In this section we will focus
on the modal representation of the field of the 742445 vertically polarized base station
antenna. Given the fact that this antenna consists of 10 dipole pairs arranged vertically
and separated approximately by 0.8λ, in front of a a metallic plane, we can carry out the
expansion of the radiated field by focusing on the radiation pattern of one dipole pair.
Therefore, our task consists of finding a modal representation for the field of a single
antenna element (dipole pair) for a later multiplication with the antenna array factor. For
this, the elevation radiation pattern of one λ/2 dipole in front of an infinite ground plane
is used
π

cos 2 cos(θ)
Cθ (θ) = , (8.23)
sin θ
whereas in the azimuthal plane, the direct measured plane from the base station (for
example, in the broadside direction) is used to synthesize the necessary coefficients. By
(a) (b)
Figure 8.9: Mobile station antenna setups. (a) Two parallel dipoles rotated with respect to the
z-axis and (b) two perpendicular dipoles rotated with respect to their starting position as vertically
and horizontally polarized dipoles.
applying a projection synthesis method it follows that the coefficients that reproduce the
radiated pattern of one antenna element of the base station can be synthesized from their
projection on the set of normalized outgoing spherical wave functions:
Z 2π Z π
~ ϕ) · ~e∗ (θ, ϕ) sin θdθdϕ
C(θ, q
ϕ=0 θ=0
cTx
q = Z 2π Z π (8.24)
~eq (θ, ϕ) · ~e∗q (θ, ϕ) sin θdθdϕ
ϕ=0 θ=0
In our case, 35 spherical modes are needed to satisfactorily represent the single antenna
configuration. At this point, to get the base station radiation pattern from these coeffi-
cients, the spherical modes must be multiplied by the array factor in order to obtain the
“pseudomodal” representation of the base station. Figure 8.12 shows the elevation and
azimuth plane patterns achieved with this procedure, compared to the measured patterns
from the vertically polarized Kathrein base station antenna. As can be seen, a good degree
of compliance is reached.
In order to see the real benefits of this approach, in Figure 8.13 the relative ergodic
capacity gain for a 2x2 MIMO configuration based on the newly synthesized antennas is
shown. Remarkably, little difference with respect to Figure 8.8 is seen. This shows the
great potential for MIMO system analysis behind the modal expansion of the radiated
fields of the antennas, even if it is not accurately described as in this case. In addition, it
opens a significant window of opportunity since different antenna sectorizations can also
be considered for improved statistics.
(a)
(b)
Figure 8.10: Relative capacity gain of 742445 vertically polarized Kathrein antenna for two
parallel dipoles with varying rotation angle α
(a)
(b)
Figure 8.11: Relative capacity gain of 742445 vertically polarized Kathrein antenna for two
perpendicular dipoles with varying rotation angle α.
(a)
(b)
Figure 8.12: Comparison of the synthesized radiation pattern (—–) and the 742445 vertically
polarized Kathrein radiation pattern (- - -) in (a) azimuth and (b) elevation.
Figure 8.13: Relative ergodic capacity gain vs horizontal antenna inter element spacing at the
base station for two vertically polarized synthesized antennas and two vertically oriented dipoles
Chapter 9
Conclusion
9.1 Original contributions

• A novel CAD methodology for the design of probe-fed, cavity-backed, stacked mi-
crostrip patch antennas has been presented. It makes use of different advanced
mathematical tools: a hybrid numerical-analytical method based on FE and MA
to characterize the electromagnetic fields inside any structure rigourously; a neural
network to be able to reproduce the results such a method would yield with speed,
ease and reliability; and a global optimization algorithm to reach a solution to the
design as close to the desired specifications as possible. Once the neural network is
trained to yield the desired response of a given topology, any design can be obtained
by changing these specifications in the optimization phase, whose computational
effort can be said to be negligible. To validate the methodology, a real design has
been carried out.
• The inclusion of the feeding network effects in the case of a distributed feeding
network has been inserted in the GSM analysis of a finite array. With this technique,
the impedance, coupling and radiation characteristics of finite arrays (with arbitrary
geometry) of coupled antennas can be efficiently obtained. It provides a closed-form
expression of the overall GSM which characterizes the array. It is based on rotation
and translation properties of spherical waves. Therefore, it is suited for arrays with
elements whose radiated field can be individually described in terms of spherical
modes, such as arbitrarily shaped apertures, horns, monopoles, DRAs or cavity-
backed patch antennas.
• The multiobjective optimization of real and coupled antenna array excitations has
been presented. The problem formulation is based on matrix-valued functions which
are directly computed from the GSM (through the transmission and reflection ma-
trices) of the whole array and spherical mode expansions. Therefore, it inherently
takes into account every array element pattern as well as all interelement couplings.
The method allows the maximization of the directive gain of an arbitrary planar
array scanned at whichever direction, subject to constraints on the maximum SLL,
the maximum crosspolar level, the minimum aperture illumination efficiency, the
setting of null pointing directions or a DR control. In order to check the capa-
bilities of the method, numerical examples with a wide variety of requirements on
these parameters have been presented. To carry out the optimization a primal-dual
IPM with filtering has been developed incorporating a Newton-step core specifically
adapted to the matrix-valued formulation, which has been demonstrated to be an
143
144 Conclusion
efficient algorithm in computational terms.

• A pattern synthesis technique for arbitrary planar arrays which are characterized in
terms of a GSM and whose radiated field is expressed as a spherical mode expansion
has been introduced. The procedure yields the complex-valued excitations to achieve
a minimum maximum SLL given different specified pointing directions and mainlobe
widths, as well as prescribed field nulls. All interelement coupling effects coming
from complex radiating structures used as array elements are inherently taken into
account. The procedure has been extended to the synthesis of shaped-beam patterns
with controllable SLL. The final step of the method only requires the resolution of
a LP problem.
• A flexible MC compensation technique has been derived from the GSM and spherical
mode characterization of an antenna array It allows the easy computation, through
a MC compensation matrix, of a set of excitations that mitigates the MC effects in
the array field pattern, for any distribution obtained with a design technique that
assumes the isotropic nature of the radiators. Conditions regarding null field pat-
tern directions can also be reinforced in the proposed MC compensation technique
to yield a null constrained MCCM. Furthermore, the MC compensation matrix is
independent of the excitations and therefore it has to be computed only once for a
given array.
• A novel simple design procedure for the pattern synthesis of a linear equispaced
array of real and coupled antennas has been presented. It unites the rigor of the
electromagnetic characterization of the field radiated by the array through its ex-
pansion in spherical waves from its GSM and the classic approach which exploits the
mathematical analogy of the problem with the theory of signal processing through
FS or DFT computations. In this sense, the procedure can be viewed as an ex-
tension of classic design techniques, such as the FS method or the WL frequency
sampling method, to the case in which the individual antenna elements’ patterns
and all interelement couplings are considered.
• A shaped-beam pattern synthesis based exclusively on the handling of spherical
modes has been presented. Translational addition theorems for spherical waves
are used to establish the necessary relationships between different spherical mode
expansions corresponding to the radiated field by each antenna in the array and
by the whole array. Therefore, the procedure analytically takes MC effects into
account. Furthermore, no involved algorithm is required as only a pseudoinverse is
needed to yield the complex feeding amplitudes.
• An original array thinning methodology based on a nested optimization scheme that
integrates a GA to thin the array layout with LP and QP to optimize the element
excitations has been proposed and applied successfully. Design specifications such
as achieving a minimum gain in a given direction, a maximum allowable SLL, or a
shaped-beam radiation pattern (while fixing null field directions), with the minimum
number of active elements, have been implemented. Furthermore, the formulation
includes an electromagnetic-based analysis that allows complex radiating antennas
to be considered trough their GSMs. All interelement coupling effects are thus in-
herently taken into account and therefore the thinning process can be carried out
by choosing whether the array elements are removed or turned off. The procedure
has proved to be well-suited for moderate-sized arrays where the field radiated by
each antenna can be described as a spherical wave expansion. Linear and rectan-
gular planar arrays of cavity-backed microstrip antennas, DRAs and isotropic point
9.2 Future research lines 145
sources have been considered in several examples. The results evidence that different
optimized array configurations are achieved with each type of the proposed thinning
options when MC is relevant. A higher degree of thinning with respect to uniform
array excitations at the expense of a weighted distribution has been confirmed.
• A model to rigorously characterize line-of-sight MIMO systems has been introduced.
It is based on the GSM of each antenna, considered as isolated, and rotation and
translation coefficients of spherical modes. The resulting channel matrix rigorously
includes the exact spherical vector nature of electromagnetic propagation, MC ef-
fects and real antenna reflection, transmission, reception and scattering features.
Therefore, it is especially suited for short-range communications, where the spheri-
cal vector nature of the propagating field plays a major role, in the case MC effects
are significant at the transmitting and/or receiving arrays and eventually when the
antennas are in the proximity of scatterers.
• A modal network model for multiple antenna systems has been introduced and ap-
plied to an urban scenario with measured antenna data. The model is based on
the spherical mode expansion of the fields of both the receiving and the transmit-
ting antennas. In this manner a significant simulation time reduction is achieved
which allows an efficient computation of statistical results based on extensive de-
terministic propagation models. In this manner exhaustive simulations, otherwise
computationally restrictive, for 2x2 MIMO systems in urban scenarios with varying
antenna types, spacings and orientations have been done.
9.2 Future research lines
• Extensions of the CAD methodology for cavity-backed patches should be carried

out to account for other antenna types, such as DRAs, as the core mathematical
tools are not constrained to this specific type of antennas.
• Segmentation or subarraying methods must be studied and combined with the GSM
characterization of finite arrays in order to be able to reduce the computational
burden and to analyze larger antenna arrays.
• In general, different planar array layouts, such as hexagonal or triangular arrays, or
even arrays with an arbitrary geometry, can be considered for any of the presented
design techniques. Furthermore, other desired objectives could be formulated and
integrated in those techniques where the synthesis is stated as an optimization prob-
lem.
• Optimization of not only the complex feeding coefficients of the elements in the array,
but also of their positions or even of strictly the phase of the feeding coefficients for
a constant amplitude, should be studied.
• Further extensions of the line-of-sight MIMO model to account for the inclusion of
non-line-of-sight channels with scattering richness should be considered.
• The inclusion of the overall GSM formulation, not only the modal expansions of the
fields, in the generalized network model for MIMO systems must be carried out in
order to take MC effects into account.
146 Conclusion
9.3 Framework
The work reported in this Ph. D. dissertation has been carried out at the Departamento
de Electromagnetismo y Teorı́a de Circuitos, in the Escuela Técnica Superior de Ingenieros
de Telecomunicación of the Universidad Politécnica of Madrid.
Part of the work reported in this Ph. D. dissertation (specifically, Chapter 8) was car-
ried out within a research period at the following foreign institution: Institut für Hochfre-
quenztechnik und Elektronik, Universität Karlsruhe, Karlsruhe, Germany.
The research projects for which the work reported in this Ph. D. dissertation has
generated results are:
• Project title: Nuevas herramientas software para CAD de antenas. Financial in-
stitution: Plan Nacional de Investigación Cientı́fica, Desarrollo e Innovación Tec-
nológica 2004–2007. Research institutions: E.T.S.I. Telecomunicación, Universidad
Politécnica de Madrid. Duration: December 2004 - December 2007. Researcher in
chief : Juan Zapata Ferrer.
• Project title: Antenna Centre of Excellence 2 (ACE-2). Financial institution: Eu-
ropean Comission. VI Framework Programme. Research institutions: 45 European
research institutions. Duration: January 2006 - December 2007. Researcher in
chief : Bruno Casali.
• Project title: Diseño de antenas y agrupaciones de antenas mediante técnicas de seg-
mentación. Financial institution: Plan Nacional de Investigación Cientı́fica, Desar-
rollo e Innovación Tecnológica 2008–2010. Research institutions: E.T.S.I. Telecomu-
nicación, Universidad Politécnica de Madrid. Duration: January 2008 - December
2010. Researcher in chief : Juan Zapata Ferrer.
The main financial support for the development of this Ph. D. has come from a grant
of the Spanish Ministry of Education and Science. This grant was directly associated to
the first of the aforementioned research projects.
9.4 Publications
This Ph. D. dissertation has given rise to the following publications:
9.4.1 Journal articles

• J. Córcoles, M. A. González, and J. Rubio, “Fourier synthesis of linear arrays based
on the generalized scattering matrix and spherical modes,” IEEE Transactions on
Antennas and Propagation, vol. 57, no. 7, pp. 1944–1951, July 2009.
• J. Córcoles, J. Pontes, M. A. González, and T. Zwick, “Modelling line-of-sight cou-
pled MIMO systems with generalised scattering matrices and spherical wave trans-
lations,” Electronics Letters, vol. 45, no. 12, pp. 598–599, June 4, 2009.
• J. Córcoles, M. A. González, and J. Rubio, “Multiobjective optimization of real
and coupled antenna array excitations via primal-dual, interior point filter method
from spherical mode expansions,” IEEE Transactions on Antennas and Propagation,
vol. 57, no. 1, pp. 110–121, Jan. 2009.
• J. Córcoles, M. A. González, and J. Zapata, “Linear programming from generalised
scattering matrix analysis of array for minimum sidelobe level and prescribed nulls,”
Electronics Letters, vol. 45, no. 1, pp. 9–10, Jan. 1, 2009.
9.4 Publications 147
• J. Córcoles, M. A. González, and J. Rubio, “Mutual coupling compensation in arrays

using a spherical wave expansion of the radiated field,” IEEE Antennas and Wireless
Propagation Letters, vol. 8, pp. 108–111, 2009.
• J. Córcoles, M. A. González, J. Rubio, and J. Zapata, “Performance characteriza-
tion of wideband, wide-angle scan arrays of cavity-backed U-slot microstrip patch
antennas,” International Journal of RF and Microwave Computer-Aided Engineer-
ing, vol. 19, no. 3, pp. 389–396, 2009.
• J. Córcoles, M. A. González, and J. Zapata, “CAD of stacked patch antennas
through multipurpose admittance matrices from FEM and neural networks,” Mi-
crowave and Optical Technology Letters, vol. 50, no. 9, pp. 2411–2416, 2008.
• J. Rubio, J. Córcoles, and M. A. González, “Inclusion of the feeding network effects
in the generalized-scattering-matrix formulation of a finite array,” IEEE Antennas
and Wireless Propagation Letters, vol. 8, pp. 819–822, 2009.
• J. Pontes, J. Córcoles, M. A. González, and T. Zwick, “Modal network model for

MIMO antenna in-system optimization,” currently undergoing a major revision in
IEEE Transactions on Antennas and Propagation.
• J. Córcoles, and M. A. González, “Efficient array thinning for particular design
specifications with a nested optimization scheme,” currently undergoing a major
revision in IEEE Transactions on Antennas and Propagation.
9.4.2 Conference proceedings

9.4.2.1 International
• J. Córcoles and M. A. González, “Multibeam synthesis with minimum SLL through
linear programming from the GSM-analysis of an array,” in IEEE Antennas and
Propagation Society International Symposium, 2009 - APSURSI ’09, Charleston,
SC, USA, June 1–5, 2009, pp. 1–4.
• J. Córcoles, M. A. González, J. Rubio, and J. Zapata, “Array design for different SLL
and null directions with an interior-point optimization method from the generalized-
scattering-matrix and spherical modes,” in 3rd European Conference on Antennas
and Propagation - EuCAP 2009, Berlin, Germany, Mar. 23–27, 2009, pp. 1381–
1385.
• J. Córcoles, M. A. González, J. Rubio, and J. Zapata, “Optimization of coupled
antenna arrays characterized by spherical mode expansions by means of an interior-
point filtering method,” in VI Iberian Meeting on Computational Electromagnetism,
Chiclana de la Frontera, Cádiz, Spain, Oct. 22–24, 2008.
9.4.2.2 National
• J. Córcoles, M. A. González, and J. Rubio, “Sı́ntesis de Fourier para agrupaciones
lineales de antenas reales y acopladas,” in XXIV Simposium Nacional de la Unión
Cientı́fica Internacional de Radio - URSI 2009, Santander, Cantabria, Spain, Sept.
16–18, 2009.
• J. Córcoles, M. A. González, and J. Zapata, “Diseño automatizado de antenas de
parches apilados en cavidad mediante descomposición de dominio en el MEF, re-
des neuronales y optimización global,” in XXIII Simposium Nacional de la Unión
Cientı́fica Internacional de Radio - URSI 2008, Madrid, Spain, Sept. 22–24, 2008.
148 Conclusion
Bibliography
F. J. Ares-Pena, J. A. Rodrı́guez-González, E. Villanueva-López, and S. R. Rengarjan,

“Genetic algorithms in the design and optimization of antenna array patterns,” IEEE
Transactions on Antennas and Propagation, vol. 47, no. 3, pp. 506–510, Mar. 1999.
H. Y. Benson, J. Vanderbei, and D. F. Shanno, “Interior-point methods for nonconvex

nonlinear programming: filter methods and merit functions,” Computational Optimiza-
tion and Applications, vol. 23, no. 2, pp. 257–272, Nov. 2002.
T. Bird, “Mutual coupling in arrays of coaxial waveguides and horns,” IEEE Transactions
on Antennas and Propagation, vol. 52, no. 3, pp. 821–829, March 2004.
T. Bird and D. Bateman, “Mutual coupling between rotated horns in a ground plane,”
IEEE Transactions on Antennas and Propagation, vol. 42, no. 7, pp. 1000–1006, Jul
1994.
S. J. Blank and M. F. Hutt, “On the empirical optimization of antenna arrays,” IEEE
Antennas and Propagation Magazine, vol. 47, no. 2, pp. 58–67, Apr. 2005.
P. Brierley. (2006) Neural network software. [Online]. Available:

http://www.philbrierley.com/
P. Brierley and B. Batty, “Data mining with neural networks - an applied example in un-
derstanding electricity consumption patterns,” in Knowledge discovery and data mining,
M. Bramer, Ed. Stevenage, UK: Institution of Electrical Engineers, 1999, pp. 240–303.
J. Bruning and Y. Lo, “Multiple scattering of em waves by spheres part i–multipole expan-
sion and ray-optical solutions, and part ii–numerical and experimental results,” IEEE
Transactions on Antennas and Propagation, vol. 19, no. 3, pp. 378–400, May 1971.
M. J. Buckley, “Linear array synthesis using a hybrid genetic algorithm,” in IEEE An-
tennas and Propagation Soc. Int. Symp., vol. 1, Baltimore, MD, USA, Jul. 21–26, 1996,
pp. 584–587.
D. L. Carroll. (1999) Fortran genetic algorithm. [Online]. Available:

http://cuaerospace.com/carroll/ga.html
K. Chen, X. Yun, Z. He, and C. Han, “Synthesis of sparse planar arrays using modified real
genetic algorithm,” IEEE Transactions on Antennas and Propagation, vol. 55, no. 4,
pp. 1067–1073, Apr. 2007.
D. K. Cheng, “Optimization techniques for antenna arrays,” Proceedings of the IEEE,

vol. 59, no. 12, pp. 1664–1674, Dec. 1971.
149
150 BIBLIOGRAPHY
J. M. Cid, J. A. Rodrı́guez, and F. Ares, “Shaped power patterns produced equispaced lin-
ear arrays: optimized synthesis using orthogonal sin(nx)/sin(x) beams,” J. Electromagn.
Waves Appl., vol. 13, pp. 985–992, 1999.
O. R. Cruzan, “Translational addition theorem for spherical vector wave functions,” Quar-
terly Appl. Math., vol. 20, no. 1, pp. 33–40, 1962.
D. E. N. Davies and C. R. Ward, “Low sidelobe patterns from thinned arrays using
multiplicative processing,” IEE Proceedings-F Communications, Radar and Signal Pro-
cessing, vol. 127, no. 1, Feb. 1980.
V. de la Rubia and J. Zapata, “MAM-A multi-purpose admittance matrix for antenna de-
sign via the finite element method,” IEEE Transactions on Antennas and Propagation,
vol. 55, no. 8, pp. 2276–2286, 2007.
C. L. Dolph, “A current distribution for broadside arrays which optimizes the relationship
between beam width and side lobe level,” in Proc. IRE, vol. 34, Jun. 1946, pp. 335–348.
M. Dürr, A. Trastoy, and F. Ares, “Multiple-pattern linear antenna arrays with single
prefixed amplitude distributions: modified Woodward-Lawson synthesis,” Electronics
Letters, vol. 36, no. 16, pp. 1345–1346, Aug. 2000.
A. R. Edmonds, Angular Momentum in Quantum Mechanics. Princeton, NJ, USA:

Princeton University Press, 1974.
O. Einarsson, “Optimization of planar arrays,” IEEE Transactions on Antennas and Prop-

agation, vol. 27, no. 1, pp. 86–92, Jan. 1979.
I. Eshrah, A. Kishk, A. Yakovlev, and A. Glisson, “Equivalent circuit model for a waveg-
uide probe with application to dra excitation,” IEEE Transactions on Antennas and
Propagation, vol. 54, no. 5, pp. 1433–1441, May 2006.
P. Feldmann and R. W. Freund, “Interconnect-delay computation and signal-integrity

verification using the SyMPVL algorithm,” in Proc. European Circuit Society Conf.
Circuit Theory and Design, 1997, pp. 408–413.
A. V. Fiacco and G. P. McCormick, Eds., Nonlinear Programming: Sequential Uncon-

strained Minimization Techniques. New York, NY: John Wiley & Sons, 1968.
R. Fletcher and S. Leyffer, “Nonlinear programming without a penalty function,”

University of Dundee, Dundee DD1 4HN, UK, Tech. Rep., 1997. [Online]. Available:
citeseer.ist.psu.edu/fletcher00nonlinear.html
J. Fondevila-Gómez, J. A. Rodrı́guez, F. Ares, and E. Moreno, “A simple way of obtaining

optimized patterns using the Woodward-Lawson method,” in Antennas and Propagation
Society International Symposium 2006, IEEE, Albuquerque, NM, 2006, pp. 3383–3386.
T. Fugen, J. Maurer, T. Kayser, and W. Wiesbeck, “Capability of 3-d ray tracing for
defining parameter sets for the specification of future mobile communications systems,”
IEEE Transactions on Antennas and Propagation, vol. 54, pp. 3125–3137, Nov. 2006.
R. Garg, P. Bhartia, I. Bahl, and A. Ittipiboon, Microstrip antenna design handbook.

Artech House, 2001.
BIBLIOGRAPHY 151
A. Gati, Y. Adane, M. Wong, J. Wiart, and V. Fouad Hanna, “Inverse source characteriza-
tion for electromagnetic near field reconstruction and interaction with the environment,”
in IEEE Antennas and Propagation Society International Symposium 2006, July 2006,
pp. 1065–1068.
J. M. Gil, J. Monge, J. Rubio, and J. Zapata, “A CAD-oriented method to analyze

and design radiating structures based on bodies of revolution by using finite elements
and generalized scattering matrix,” IEEE Transactions on Antennas and Propagation,
vol. 54, no. 3, pp. 899–907, Mar. 2006.
P. E. Gill, W. Murray, M. A. Saunders, J. A. Tomlin, and M. H. Wright, “On projected

newton barrier methods for linear programming and an equivalence to Karmarkar’s
projective method,” Mathematical Programming, vol. 36, pp. 183–209, 1986.
D. E. Goldberg, Genetic Algorithms in Search, Optimization, and Machine Learning.

Addison-Wesley Professional, January 1989.
D. E. Goldberg and J. Richardson, “Genetic algorithms with sharing for multimodal func-
tion optimization,” in Proceedings of the Second International Conference on Genetic
Algorithms on Genetic algorithms and their application. Mahwah, NJ, USA: Lawrence
Erlbaum Associates, Inc., 1987, pp. 41–49.
M. A. González, J. A. Encinar, and J. Zapata, “Radiation pattern computation of cavity-

backed and probe-fed stacked microstrip patch arrays,” IEEE Transactions on Antennas
and Propagation, vol. 48, pp. 502–509, Apr. 2000.
S. K. Goudos, G. S. Miaris, K. Siakavara, and J. N. Sahalos, “On the orthogonal nonuni-

form synthesis from a set of uniform linear arrays,” IEEE Antennas and Wireless Prop-
agation Letters, vol. 6, pp. 313–315, 2007.
A. K. Hamid, I. R. Ciric, and M. Hamid, “Multiple scattering by an arbitrary configuration

of dielectric spheres,” Can. J. Phys., vol. 68, pp. 1419–1428, Dec 1990.
J. E. Hansen, Ed., Spherical Near-Field Antenna Measurements. London, UK: Peter

Peregrinus Ltd., 1988.
R. L. Haupt, “Thinned arrays using genetic algorithms,” IEEE Transactions on Antennas

and Propagation, vol. 42, no. 7, pp. 993–999, Jul. 1994.
——, “Antenna design with a mixed integer genetic algorithm,” IEEE Transactions on
Antennas and Propagation, vol. 55, no. 3, pp. 577–582, Mar. 2007.
S. Haykin, Neural networks: a comprehensive foundation, 2nd ed. Upper Saddle River,
NJ, USA: Prentice Hall, 1998.
S. Hebib, N. Raveu, and H. Aubert, “Cantor spiral array for the design of thinned arrays,”
IEEE Antennas Wireless Propag. Lett., vol. 5, no. 1, pp. 104–106, Dec. 2006.
A. Heldring, J. Rius, J.-M. Tamayo, J. Parron, and E. Ubeda, “Fast direct solution of
method of moments linear system,” IEEE Transactions on Antennas and Propagation,
vol. 55, no. 11, pp. 3220–3228, Nov. 2007.
152 BIBLIOGRAPHY
K. Hirasawa, “The application of a biquadratic programming method to phase-only opti-

mization of antenna arrays,” IEEE Transactions on Antennas and Propagation, vol. 36,
no. 11, pp. 1545–1550, Nov. 1988.
S. Holm and B. Elgetun, “Properties of the beampattern of weight- and layout-optimized

sparse arrays,” IEEE Transactions on Ultrasonics, Ferroelectrics and Frequency Control,
vol. 44, no. 5, pp. 983–991, Sep. 1997.
Z. Huang, C. A. Balanis, and C. R. Birtcher, “Mutual coupling compensation in UCAs:

simulations and experiment,” IEEE Transactions on Antennas and Propagation, vol. 54,
no. 11, pp. 3082–3086, Nov. 2006.
J. R. James, “Conformal antenna synthesis using spherical harmonic wavefunctions,” Proc.

IEE, vol. 122, no. 5, pp. 479–486, May 1975.
R. Janaswamy, “Effect of element mutual coupling on the capacity of fixed length linear
arrays,” Antennas and Wireless Propagation Letters, IEEE, vol. 1, pp. 157–160, 2002.
F. Jensen and A. Frandsen, “On the number of modes in spherical wave expansions,”
in 26th meeting and symposium of the Antenna Measurement Techniques Association,
Stone Mountain Park, GA, USA, Oct 2004, pp. 489–494.
Y. Jianfeng, P. Weizheng, W. Yufeng, L. Daxin, and C. Yan, “Side lobe reduction in

thinned array synthesis using immune algorithm,” in Int. Conference on Microwave
and Millimeter Wave Technology, vol. 3, Apr. 21–24, 2008, pp. 1131–1133.
Jiang and Ingram, “Spherical wave model for short-range mimo,” IEEE Transactions on
Communications, vol. 53, no. 5, pp. 905–905, May 2005.
Y.-C. Jiao, W.-Y. Wei, L.-W. Huang, and H.-S. Wu, “A new low-side-lobe pattern synthe-
sis technique for conformal arrays,” IEEE Transactions on Antennas and Propagation,
vol. 41, no. 6, pp. 824–831, Jun. 1993.
N. Karmarkar, “A new polynomial-time algorithm for linear programming,” Combina-

torics, vol. 4, no. 4, pp. 373–395, 1984.
Y. U. Kim and J. D. Nespor, “Shaped beam synthesis and conditional thinning for planar
phased array,” in IEEE Antennas and Propagation Soc. Int. Symp., vol. 2, Baltimore,
MD, USA, Jul. 21–26, 1996, pp. 802–805.
K. Krishnakumar, “Micro-genetic algorithms for stationary and non-stationary function

optimization,” in Proceedings of the SPIE, ser. Intelligent Control and Adaptive Sys-
tems, vol. 1196, Philadelphia, PA, 7-8 Nov 1990, pp. 289–296.
R. R. Kurth, “Optimization of array performance subject to multiple power constraints,”

IEEE Transactions on Antennas and Propagation, vol. 22, no. 1, pp. 103–105, Jan. 1974.
V. Lebedev and D. Laikov, “A quadrature formula for the sphere of the 131st algebraic
order of accuracy,” Doklady Mathematics, vol. 59, no. 3, pp. 477–481, 1999.
H. Lebret and S. Boyd, “Antenna array pattern synthesis via convex optimization,” IEEE
Transactions on Signal Processing, vol. 45, no. 3, pp. 526–532, Mar. 1997.
BIBLIOGRAPHY 153
J. Liu and J.-M. Jin, “A novel hybridization of higher order finite element and boundary
integral methods for electromagnetic scattering and radiation problems,” IEEE Trans-
actions on Antennas and Propagation, vol. 49, no. 12, pp. 1794–1806, Dec 2001.
Y. Lo and S. Lee, “A study of space-tapered arrays,” IEEE Trans. Antennas Propag.,
vol. 14, no. 1, pp. 22–30, Jan. 1966.
G. R. Lockwood and F. S. Foster, “Optimizing the radiation pattern of sparse periodic two-
dimensional arrays,” IEEE Trans. Ultrason. Ferroelectr. Freq. Control, vol. 43, no. 1,
pp. 15–19, Jan. 1996.
K. M. Luk, W. K. Leuk, and K. W. Leung, “Mutual impedance of hemispherical dielectric
resonator antennas,” IEEE Transactions on Antennas and Propagation, vol. 42, no. 12,
pp. 1652–1654, Dec. 1994.
R. Mailloux, “Radiation and near-field coupling between two collinear open-ended waveg-
uides,” IEEE Transactions on Antennas and Propagation, vol. 17, no. 1, pp. 49–55, Jan
1969.
R. J. Mailloux and E. Cohen, “Statistically thinned arrays with quantized element
weights,” IEEE Trans. Antennas Propag., vol. 39, no. 4, pp. 436–447, Apr. 1991.
N. Marcuvitz, Waveguide Handbook. London, UK: IEE Electromagnetic Waves Series,
No. 21, Peter Peregrinus Ltd., 1986.
K. Markus and L. Vaskelainen, “Optimisation of synthesised array excitations using ar-
ray polynome complex root swapping and genetic algorithms,” in IEE Proc.-Microw.
Antennas Propag., vol. 145, no. 6, Dec. 1998, pp. 460–464.
L. Matekovits, G. Vecchi, and M. Orefice, “Synthetic-function analysis of large printed
structures,” in IEEE Antennas and Propagation Society International Symposium,
vol. 2, June 2003, pp. 1063 vol.2–.
M. Mehler, “Spherical wave expansions for shaped beam synthesis,” IEE Proceedings Mi-
crowaves, Antennas and Propagation, vol. 135, no. 5, pp. 327–332, Oct 1988.
J. Monge, J. M. Gil, and J. Zapata, “Optimization of arbitrarily shaped H-plane passive
microwave devices through finite elements and the annealing algorithm,” Microwave
and Optical Technology Letters, vol. 46, no. 4, pp. 360–366, Aug. 2005.
J. Monge, J. M. Cid, J. M. Gil, and J. Zapata, “Optimization of arbitrarily shaped di-
electric axisymmetric horns through finite elements and the annealing algorithm,” Mi-
crowave and Optical Technology Letters, vol. 48, no. 6, pp. 1074–1079, Jun. 2006.
M. Mongiardo, C. Tomassoni, and P. Russer, “Generalized network formulation: Appli-
cation to flangemounted radiating waveguides,” IEEE Transactions on Antennas and
Propagation, vol. 55, no. 6, pp. 1667–1678, June 2007.
C. G. Montgomery, R. H. Dicke, and E. M. Purcell, Eds., Principles of Microwave Circuits.
New York, USA: Dover Publications, 1965.
C. Moreno, S. Garcı́a, M. Fernández, A. Rubio, and R. Gómez, “Microstip-patch array
design using multiobjective GA,” IEEE Antennas and Wireless Propagation Letters,
vol. 4, pp. 100–103, 2005.
154 BIBLIOGRAPHY
A. Morini, M. Farina, T. Rozzi, and G. Venanzoni, “An effective strategy for the optimiza-
tion of the phase distribution of large arrays based on a full-wave model,” in Proc. IEEE
Antennas and Propagation Society International Symposium 2006, vol. 1, Albuquerque,
NM, 2006, pp. 4205–4208.
S. Mosca and M. Ciattaglia, “Ant colony optimization to design thinned arrays,” in IEEE
Antennas and Propagation Soc. Int. Symp., 2006, pp. 4675–4678.
J. E. Mulholland, F. N. DiMeo, and A. Hoorfar, “Optimization of thinning phased ar-

rays using soft-computing algorithms,” in Intl. Symp. on Electromangetic Theory, St.
Petersburg, Russia, May 1995.
V. Murino, A. Trucco, and C. S. Regazzoni, “Synthesis of unequally spaced arrays by

simulated annealing,” IEEE Trans. Signal Process., vol. 44, no. 1, pp. 119–122, Jan.
1996.
M. Narasimhan, K. Varadarangan, and S. Christopher, “A new technique of synthesis of

the near- or far-field patterns of arrays,” IEEE Transactions on Antennas and Propa-
gation, vol. 34, no. 6, pp. 773–778, Jun 1986.
B. P. Ng, M. H. Er, and C. Kot, “A flexible array synthesis method using quadratic
programming,” IEEE Transactions on Antennas and Propagation, vol. 41, no. 11, pp.
1541–1550, Nov. 1993.
J. Nocedal and S. J. Wright, Numerical Optimization. Springer, 2000.
T. Numazaki, S. Mano, T. Katagi, and M. Mizusawa, “An improved thinning method

for density tapering of planar array antennas,” IEEE Trans. Antennas Propag., vol. 35,
no. 9, pp. 1066–1070, Sep. 1987.
D. J. O’Neill, “Element placement in thinned arrays using genetic algorithms,” in Proc.

OCEANS ’94, ’Oceans Engineering for Today’s Technology and Tomorrow’s Preserva-
tion’, vol. 2, Brest, France, Sep. 13–16, 1994.
A. V. Oppenheim and R. W. Schafer, Discrete-Time Signal Processing. Upper Saddle

River, NJ: Prentice Hall, Inc., 1999.
A. V. Oppenheim and A. S. Willsky, Signals & Systems. Upper Saddle River, NJ: Prentice
Hall, Inc., 1997.
S. J. Orfanidis, Electromagnetic Waves and Antennas. Piscataway, NJ, USA: Rutgers

University, 2008. [Online]. Available: http://www.ece.rutgers.edu/ orfanidi/ewa/
R. H. J. M. Otten and L. P. P. P. van Ginneken, The annealing algorithm. Deventer,

The Netherlands, The Netherlands: Kluwer, B.V., 1989.
D. M. Pozar and D. H. Schaubert, “Analysis of an infinite array of rectangular microstrip

patches with idealized probe feeds,” IEEE Transactions on Antennas and Propagation,
vol. 32, pp. 1101–1107, Oct. 1984.
V. V. S. Prakash and R. Mittra, “Characteristic basis function method: A new technique

for efficient solution of method of moments matrix equations,” Microwave and Optical
Technology Letters, vol. 36, no. 2, pp. 95–100, Jan 2003.
BIBLIOGRAPHY 155
O. Quevedo-Teruel and E. Rajo-Iglesias, “Ant colony optimization in thinned array syn-

thesis with minimum sidelobe level,” IEEE Antennas Wireless Propag. Lett., vol. 5,
no. 1, pp. 349–352, Dec. 2006.
A. Razavi and K. Forooraghi, “Thinned arrays using pattern search algorithms,” Progress
In Electromagnetics Research, vol. 78, pp. 61–71, 2008.
H. J. Riblet, “Discussion on ’A current distribution for broadside arrays which optimizes

the relationship between beam width and side lobe level’,” in Proc. IRE, vol. 35, 1947,
pp. 489–492.
M. J. Rossouw, J. Joubert, and D. A. McNamara, “Thinned arrays using a modified

minimum redundancy synthesis technique,” Electronics Letters, vol. 33, no. 10, pp.
826–827, May 8, 1997.
J. Rubio, M. Gonzalez, and J. Zapata, “Efficient full-wave analysis of mutual coupling

between cavity-backed microstrip patch antennas,” IEEE Antennas and Wireless Prop-
agation Letters, vol. 2, pp. 155–158, 2003.
——, “Fast and rigorous analysis of the mutual coupling effects in finite arrays,” in Pro-
ceedings JINA Symposium, Nice, France, Nov. 2004, pp. 42–43.
——, “Full wave analysis of finite arrays on an infinite ground plane including the mutual
coupling effects,” in IEEE Antennas and Propagation Society International Symposium,
2004., vol. 3, June 2004, pp. 2695–2698 Vol.3.
J. Rubio, M. A. Gonzalez, and J. Zapata, “Generalized-scattering-matrix analysis of a class

of finite arrays of coupled antennas by using 3-D FEM and spherical mode expansion,”
IEEE Transactions on Antennas and Propagation, vol. 53, no. 3, pp. 1133–1144, Mar.
2005.
J. Rubio, A. Heldring, J. Rius, M. Gonzalez, and J. Zapata, “Integration of cbd solver

in the finite array gsm code,” in Antennas and Propagation, 2007. EuCAP 2007. The
Second European Conference on, Nov. 2007, pp. 1–5.
J. Rubio, J. Arroyo, and J. Zapata, “Analysis of passive microwave circuits by using

a hybrid 2-D and 3-D finite-element mode-matching method,” IEEE Transactions on
Microwave Theory and Techniques, vol. 47, no. 9, pp. 1746–1749, Sep. 1999.
——, “SFELP-an efficient methodology for microwave circuit anlysis,” IEEE Transactions
on Microwave Theory and Techniques, vol. 49, no. 3, pp. 509–516, Mar. 2001.
J. Rubio, M. A. González, and J. Zapata, “Analysis of cavity-backed microstrip antennas

by a 3-D finite element/segmentation method and a matrix Lanczos-Padé algorithm
(SFELP),” IEEE Antennas and Wireless Propagation Letters, vol. 1, pp. 193–195, 2002.
C. S. Ruf, “Numerical annealing of low-redundancy linear arrays,” IEEE Trans. Antennas

Propag., vol. 41, no. 1, pp. 85–90, Jan. 1993.
D. E. Rumelhart, G. E. Hinton, and R. J. Williams, “Learning internal representations by

error propagation,” in Parallel Distributed Processing: Volume 1: Foundations, D. E.
Rumelhart, J. L. McClelland et al., Eds. Cambridge: MIT Press, 1987, pp. 318–362.
156 BIBLIOGRAPHY
S. A. Schelkunoff, “A mathematical theory of linear arrays,” in Bell Syst. Tech. J., vol. 22,
1943, pp. 80–107.
K. Schittkowski, “Test problems for nonlinear programming - User’s guide,” Department

of Mathematics, University of Bayreuth, Tech. Rep., 2002. [Online]. Available:
www.klaus-schittkowski.de
M. Shimizu, “Determining the excitation coefficients of an array using genetic algorithms,”

in Proc. IEEE Antennas and Propagation Society International Symposium 1994, vol. 1,
Seattle, WA, USA, Jun. 20–24, 1994, pp. 530–533.
D. J. Shpak, “A method for the optimzal pattern synthesis of linear arrays with prescribed
nulls,” IEEE Transactions on Antennas and Propagation, vol. 44, no. 3, pp. 286–294,
Mar. 1996.
M. Skolnik, I. Sherman, J., and J. Ogg, F., “Statistically designed density-tapered arrays,”
IEEE Trans. Antennas Propag., vol. 12, no. 4, pp. 408–417, Jul. 1964.
T. G. Spence and D. H. Werner, “Thinning of aperiodic antenna arrays for low side-lobe
levels and broadband operation using genetic algorithms,” in Antennas and Propagation
Society International Symposium 2006, IEEE, Jul. 9–14, 2006, pp. 2059–2062.
S. Stein, “Addition theorem for spherical wave functions,” Quarterly Appl. Math., vol. 19,
no. 1, pp. 15–24, 1961.
H. Steykal, R. A. Shore, and R. L. Haupt, “Method for null control and their effects on
the radiation pattern,” IEEE Transactions on Antennas and Propagation, vol. 34, no. 3,
pp. 404–409, Mar. 1986.
W. L. Stutzman and G. A. Thiele, Antenna Theory and Design. New York: John Wiley
& Sons, 1981.
E. Suter and J. R. Mosig, “A subdomain multilevel approach for the efficient mom analysis
of large planar antennas,” Microwave and Optical Technology Letters, vol. 26, no. 4, pp.
270–277, Aug 2000.
K. Tomiyasu, “Combined equal and unequal element spacings for low sidelobe pattern of
a symmetrical array with equal-amplitude elements,” IEEE Trans. Antennas Propag.,
vol. 39, no. 2, pp. 265–266, Feb. 1991.
K.-F. Tong, K.-M. Luk, K.-F. Lee, and R. Lee, “A broad-band u-slot rectangular patch
antenna on a microwave substrate,” IEEE Transactions on Antennas and Propagation,
vol. 48, no. 6, pp. 954–960, Jun 2000.
A. Trucco, “Thinning and weighting of large planar arrays by simulatedannealing,” IEEE

Transactions on Ultrasonics Ferroelectrics and Frequency Control, vol. 46, no. 2, pp.
347–355, Mar. 1999.
A. Trucco and V. Murino, “Stochastic optimization of linear sparse arrays,” IEEE Journal
of Oceanic Engineering, vol. 24, no. 3, pp. 291–299, Jul. 1999.
F. I. Tseng and D. K. Cheng, “Optimum scannable planar arrays with an invariant sidelobe
level,” Proc. IEEE, vol. 56, no. 11, pp. 1771–1778, Nov. 1968.
BIBLIOGRAPHY 157
M. Ulbrich, S. Ulbrich, and L. Vicente, “A globally convergent primal-dual interior point

filter method for nonconvex nonlinear programming,” Mathematical Programming, vol.
100, no. 2, pp. 379–410, Jun. 2004.
M. C. Vigano, G. Toso, S. Selleri, C. Mangenot, P. Angeletti, and G. Pelosi, “Ga optimized
thinned hexagonal arrays for satellite applications,” in IEEE Antennas and Propagation
Soc. Int. Symp., Jun. 9–15, 2007, pp. 3165–3168.
R. C. Voges and J. K. Butler, “Phase optimization of antenna array gain with constrained
amplitude excitation,” IEEE Transactions on Antennas and Propagation, vol. 20, no. 4,
pp. 432–436, Jul. 1972.
C. Waldschmidt, S. Schulteis, and W. Wiesbeck, “Complete rf system model for analysis
of compact mimo arrays,” IEEE Transactions on Vehicular Technology, vol. 53, no. 3,
pp. 579–586, May 2004.
G. H. Walker, “Thinned arrays using amplitude sharing,” in 9th European Microwave
Conference, Oct. 1979, pp. 159–163.
J. W. Wallace and M. A. Jensen, “Mutual coupling in mimo wireless systems: a rigorous
network theory analysis,” IEEE Transactions on Wireless Communications, vol. 3, no. 4,
pp. 1317–1325, Jul. 2004.
R. B. Waterhouse, “Design of probe-fed stacked patches,” IEEE Transactions on Antennas
and Propagation, vol. 47, no. 12, pp. 1780–1784, Dec. 1999.
S. Weigand, G. Huff, K. Pan, and J. Bernhard, “Analysis and design of broad-band single-
layer rectangular u-slot microstrip patch antennas,” IEEE Transactions on Antennas
and Propagation, vol. 51, no. 3, pp. 457–468, March 2003.
D. H. Werner and A. J. Ferraro, “Cosine pattern synthesis for single and multiple main
beam uniformly spaced linear arrays,” IEEE Transactions on Antennas and Propaga-
tion, vol. 37, no. 11, pp. 1480–1484, Nov. 1989.
P. J. Wood, “Spherical waves in antenna problems,” Marconi Rev., vol. 34, no. 182, pp.
149–172, 1971.
P. M. Woodward, “A method for calculating the field over a plane aperture required to
produce a given polar diagram,” J. IEE, vol. 93, pp. 1554–1558, 1946.
P. M. Woodward and J. D. Lawson, “The theoretical precision with which an arbitrary
radiation-pattern may be obtained from a source of a finite size,” J. IEE, vol. 95, pp.
363–370, Sep. 1948.
K.-K. Yan and Y. Lu, “Sidelobe reduction in array-pattern synthesis using genetic algo-
rithm,” IEEE Transactions on Antennas and Propagation, vol. 45, no. 7, pp. 1117–1122,
Jul. 1997.
F. Zavosh and J. T. Aberle, “Infinite phased arrays of cavity-backed patches,” IEEE
Transactions on Antennas and Propagation, vol. 42, no. 3, pp. 390–398, Mar. 1994.
F. Zavosh and T. Aberle, “Single and stacked circular microstrip patch antennas backed
by a circular cavity,” IEEE Transactions on Antennas and Propagation, vol. 41, no. 7,
pp. 746–750, Jul. 1995.

Tesis Upm - Juan - Corcoles - Ortega

Transféré par

Informations du document

Titre original

Copyright

Formats disponibles

Partager ce document

Partager ou intégrer le document

Options de partage

Avez-vous trouvé ce document utile ?

Ce contenu est-il inapproprié ?

Droits d'auteur :

Formats disponibles

Tesis Upm - Juan - Corcoles - Ortega

Transféré par

Droits d'auteur :

Formats disponibles

UNIVERSIDAD POLITÉCNICA DE MADRID

ESCUELA TÉCNICA SUPERIOR DE INGENIEROS DE TELECOMUNICACIÓN

ADVANCED TECHNIQUES OF OPTIMIZATION AND

TÉCNICAS AVANZADAS DE OPTIMIZACIÓN Y DISEÑO

Juan Córcoles Ortega

ADVANCED TECHNIQUES OF OPTIMIZATION AND

TÉCNICAS AVANZADAS DE OPTIMIZACIÓN Y DISEÑO

AUTOR: Juan Córcoles Ortega

DIRECTOR: Miguel Ángel González de Aza

DEPARTAMENTO: Electromagnetismo y Teorı́a de Circuitos

El Tribunal de Calificación, compuesto por:

Acuerda otorgarle la CALIFICACIÓN de:

For those who never try. . .

is introduced. It can be viewed as an extension of the classic design techniques of the

List of Tables xxi

Glossary of symbols xxiii

Glossary of acronyms xxv

2 Design methodology for stacked microstrip patch antennas 3

2.6.5 Other designs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

3 Generalized-Scattering-Matrix analysis of antenna arrays 29

4 Multiobjective optimization of an array through an IPM with filtering 47

5 Linear and Quadratic Programming for array pattern synthesis 69

6 Non-optimization based field synthesis techniques 85

6.2.6 Fourier-based design procedure . . . . . . . . . . . . . . . . . . . . . 89

7 Array thinning with a nested optimization scheme 107

8 Modelling MIMO systems with a hybrid spherical-mode-based approach127

2.1 Canonical geometry of the antenna to be designed. . . . . . . . . . . . . . . 4

3.3 Coupling coefficient magnitude between the fundamental modes of two

7.1 Proposed nested optimization scheme. . . . . . . . . . . . . . . . . . . . . . 111

7.3 Flowchart of the computation of an individual’s fitness function. . . . . . . 116

2.1 Values of the parameters of antenna in Figure 2.11 . . . . . . . . . . . . . . 22

4.1 Computational aspects of linear patch arrays optimizations represented in

{a} Complex coefficients of incident spherical 32

{b} Complex coefficients of scattered spherical 32

~em (θ, ϕ) mth spherical far field vector mode 8

[Γ] Reflection matrix of an array 33

[R] Reception matrix of an array 33

[S] Scattering matrix of an array 33

[T] Transmission matrix of an array 33

{v} Complex coefficients of incident feeding or ex- 32

{w} Complex coefficients of reflected feeding or ex- 32

{wi } Complex coefficients of reflected feeding or ex- 8

BVP Boundary Value Problem 6

CAD Computer-Aided Design 1, 3,

DFT Discrete Fourier Transform 85, 86,

FE Finite Element 1, 3–6,

FF Fitness Function 23,

IPM Interior-Point Method 48, 53–

KKT Karush-Kuhn-Tucker 53, 54

LP Linear Programming 2, 48,

QP Quadratic Programming 2, 69,

RDC Riblet-Dolph-Chebychev 59, 61,

SA Simulated Annealing 1, 20,

SISO Single-Input-Single-Output 127,

WL Woodward-Lawson 85, 87,

1.1 Motivation and objective

1.2 Text organization

Design methodology for stacked

Figure 2.1: Canonical geometry of the antenna to be designed.

2.2 Hybrid Finite Element - Modal Analysis characteriza-

Usually, when an expansion in spherical harmonics is applied, a spherical mesh is

Segmentation Finite Element via Lanczos-Padé (SFELP) is a methodology to an-

2.2.2 Finite Element formulation