1441957561

Efficient Modeling and Control
of Large-Scale Systems
Javad Mohammadpour
Karolos M. Grigoriadis
Editors
Efficient Modeling
and Control of
Large-Scale Systems
123
Editors
Javad Mohammadpour
University of Houston
Dept. Mechanical Engineering
Calhoun Road 4800
77204-4006 Houston Texas
N204, Engineering Bldg. 1
USA
jmohammadpour@uh.edu
University of Houston
Dept. Mechanical Engineering
Calhoun Road 4800
77204-4006 Houston Texas
N204, Engineering Bldg. 1
USA
karolos@uh.edu
ISBN 978-1-4419-5756-6
e-ISBN 978-1-4419-5757-3
DOI 10.1007/978-1-4419-5757-3
Springer New York Dordrecht Heidelberg London
Library of Congress Control Number: 2010929392
c Springer Science+Business Media, LLC 2010

All rights reserved. This work may not be translated or copied in whole or in part without the written
permission of the publisher (Springer Science+Business Media, LLC, 233 Spring Street, New York,
NY 10013, USA), except for brief excerpts in connection with reviews or scholarly analysis. Use in
connection with any form of information storage and retrieval, electronic adaptation, computer software,
or by similar or dissimilar methodology now known or hereafter developed is forbidden.
The use in this publication of trade names, trademarks, service marks, and similar terms, even if they are
not identified as such, is not to be taken as an expression of opinion as to whether or not they are subject
to proprietary rights.
Printed on acid-free paper
Springer is part of Springer Science+Business Media (www.springer.com)
Foreword
Almost five decades have passed from the organized formulation of Large-Scale
Systems (LSS) engineering as a branch of control and systems engineering. In the
1960s modeling, analysis and synthesis of complex systems were being pursued by
a number of research teams in the US and overseas. It is impossible to identify all
of them here. Among those there were the MIT team headed by Professor Michael
Athans and many of his colleagues and his former doctoral students; the other was at
the University of Illinois, headed by Professor Joe Cruz, Jr. with able peers like Petar
Kokotovic and Bill Perkins and their many doctoral students; another was the work
done by Ted Davison and his peers at the University of Toronto in Canada and Andy
Sage at the University of Virginia and later on at George Mason University. At the
time I was a doctoral student of Petar Kokotovic at the University of Illinois. These
were not the only ones of course, but they were very visible in the annual IEEE
CDC and ACC meetings. The approaches of LSS theory at Illinois were based on
sensitivity of system with respect to inherent models parameters and by determining
the sensitivity of inputs and outputs and thereby deal with many attributes of an LSS
be it nonlinearity, dimension, time delays, etc. Others were concerned with robust
control or multi-variable control approaches. Then it was in 1978 that a special issue
of IEEE Transaction on Automatic Control (Volume AC-23, 1978) guest edited by
Mike Athans, Nil Sandell, Pravin Varaiya and Mike Safonov, the control of complex
systems were designated Large-Scale Control Systems.
Three definitions of LSS can be documented, as this author enumerated in his
1983 textbook on Large-Scale Systems Modeling and Control (North-Holland
Publishers, New York, 1983). First a system is considered an LSS with a hierarchical structure so that behavior of subsystems can be coordinated from above
and feasibility and optimality of the overall system can be guaranteed by iterative processes based on optimization principles (Minimum Principle, Dynamic
Programming or Lagrangian optimization). This definition led to hierarchical control, which was pursued extensively by M.D. Mesraovic et al. at the Case Western
Reserve University. The second definition is that a system whose components or
subsystems are not centralized in one location is an LSS. This approach was researched by many, including the Davison group in Toronto. The third one is that a
vi
Foreword
system whose attributes (nonlinear, stochastic, time-varying, delays, dimensionality,

etc.) are such that standard theories and approaches of classical or modern control
theories need to be re-invented. This definition led to new paradigms for modeling
and control of LSS which was based on sensitivity functional techniques and tools
that were pursued at Illinois.
Now, nearly 50 years later, at the beginning of the twenty-first century, systems
and control engineering fields are being challenged once again to deal with integration and interoperability of a group of legacy or other systems to make the full use of
our global wireless flat world to quote Thomas Friedman of New York Times. This
new observation of complex systems is called either System of Systems (SoS) or
Cyber-Physical Systems (CPS). Interoperability or system integration adds another
complexity dimension to an already complex system and for the first time control
engineers need to design a feedback control paradigm where a wireless Ad-hoc network lies in the feedback loop. Here one has a marriage between the Cyberspace
and Physical Space. This new class of complex systems is really not new and does
in fact follow the 2nd definition of LSS of decentralize nature of many complex systems. The added complexity is the communication protocol among SoS constituents
and Internet (network) induced time delays and their potential adverse effect on
the system performance. These time-delays are not necessarily even Gaussian and
are often best estimated by using alpha-stable distribution like a non-Gaussian
noise.
This volume, Efficient Modeling and Control of Large-Scale Systems, edited by
Javad Mohammadpour and Karolos Grigoriadis, Springer, 2010 is a welcome new
addition to a collection of published books on LSS. Scanning the chapters of the
book, it serves, in a way, as a bridge from LSS to SoS class of complex systems.
Topics covered are modeling, model reduction, and distributed (decentralized) control, network, among others. In the rest of this century, system integration will appear in all aspects of our modern lives. Our iPhoneTM , laptop, BlackberryT M , a
smart-grid distributed power generation systems, etc. are all examples of twentyfirst century cyber-physical (or SoS) systems. One may say that LSS is moving
into a new phase of universal integration systems thanks to the advances in wireless communication. The authors and editors of this volume deserve our thanks for
adding a volume to this re-invigorated branch of control and systems engineering.
San Antonio, TX
November 2009
Prof. Mo Jamshidi
Preface
The complexity and dynamic order of controlled engineering systems is constantly

increasing. Complex large scale dynamic systems where large reflects the systems order and not necessarily its physical size appear in many engineering fields,
such as, aerospace, micro-electro-mechanics, manufacturing, civil engineering and
power systems. Modeling of these systems often results in high-order models imposing great challenges to their analysis, design, and feedback control. This edited
volume is aimed at providing comprehensive contributions on recent analytical and
computational methods for addressing the model reduction, performance analysis,
and feedback control design of such systems. The book presents new theoretical
developments, computational approaches, and illustrative applications to various
fields. The scope of the book is intended to be interdisciplinary emphasizing the
commonality and applicability of the methods to various fields.
The book comprises 12 chapters.
The first chapter, by A.C. Antoulas, C.A. Beattie, and S. Gugercin, presents a detailed and thorough exposition of interpolatory model order reduction methods for
large-scale systems. Roughly speaking, these methods seek to obtain reduced-order
models whose transfer function interpolates that of the original system at selected
interpolation points in the frequency domain along selected input or output tangent
directions. The question of the selection of interpolation points and tangential directions is examined in the context of meeting desired approximation goals. Finally,
the methods are applied to systems with a generalized coprime factorization (such
as second-order systems and systems with delays) and systems with a structured
dependence on parameters that need to be retained in the reduced-order model. Examples from various disciplines conclude the chapter.
The chapter by R. Colgren examines the efficient model reduction of large-scale
systems for control design. The motivation is based on meeting computational requirements for control and meeting practical implementation constrains of full-order
dynamic controllers. A summary of the evolution of model reduction methods for
control is presented with an emphasis on approaches that are easily applicable to
large generalized models obtained by Computer Aided Design (CAD) and Finite
Element Analysis (FEA) tools. A large-scale model of an air vehicle with significant aero-servo-elastic coupling is used to illustrate the methods.
vii
viii
Preface
The chapter by M.C. de Oliveira and A.S. Wroldsen presents a thorough study of
the dynamic modeling and simulation of large tensegrity structural systems. Tensegrity structures are composed of just rods and strings, so that the rods are in compression and the strings in tension. Such large tensegrity systems can find applications
as lightweight, deployable, and shape controllable structures. The authors derive the
dynamic models for a general Class k tensegrity system and examine the implementation of these models in a computer simulation environment.
The chapter by X. Dai, T. Breikin, H. Wang, G. Kulikov, and V. Arkov investigates the data-driven reduced-order modeling of gas turbine engines for the purpose of on-board condition monitoring and fault detection. Model-based condition
monitoring relies on the long-term predictive capabilities of reduced-order models
that are compatible with limited on-board computational resources. A data-driven
reduced-order model development is presented for gas turbine systems based on
dynamic nonlinear least-squares identification. Methods to accelerate the speed of
the identification algorithm are proposed using iterative calculation of the gradient
and Hessian approximation. Data gathered in an aero-engine test bed are used to
illustrate the reduced-order modeling approach.
The chapter by J. Lavaei, S. Sojoudi, and A.G. Aghdam examines the robust controllability and observability of a large-scale uncertain system with the objective of
selecting dominant inputs and outputs to obtain a simplified control structure. The
problem is formulated as the minimization of the smallest singular value of the corresponding Gramian in a polynomial uncertain system. It is shown that for such a
system the Gramian is a rational function that can be approximated by a matrix polynomial. Subsequently, sum-of-squares (SoS) optimization techniques are employed
for efficient computational solution, and simulation studies are used to validate the
effectiveness of the proposed results.
The chapter by P. Krishnamurthy and F. Khorrami addresses the decentralized
output feedback control of a class of interconnected nonlinear uncertain large-scale
systems using a dynamic high-gain scaling approach. The method provides a unified
framework for observer/controller design based on the solution of coupled statedependent Lyapunov inequalities. Stability and disturbance attenuation of the proposed decentralized control solution are discussed in the input-to-output practical
stability and integral-input-to-output practical stability frameworks.
The chapter by J. Xiong, V.A. Ugrinovskii, and I.R. Petersen examines the
guaranteed-cost output feedback control problem for a class of large-scale uncertain stochastic systems with random parameters. It is assumed that the system uncertainties satisfy integral quadratic constraints and the random parameters follow a
Markov process. Sufficient conditions are derived for decentralized controller synthesis that guarantees stability and a suboptimal level of quadratic performance.
The control law uses local, in general non-Markovian, subsystem outputs and local
subsystem operation modes to produce local subsystem control actions. Design conditions are provided in terms of rank constrained linear matrix inequalities (LMIs).
The chapter by M.S. Stankovic, D.M. Stipanovic, and S.S. Stankovic presents
novel algorithms for decentralized overlapping control of large-scale complex systems. The approach is based on multiagent networks in which the agents utilize
Preface
ix
dynamic consensus strategies to reach the agreement upon their actions. Different
decentralized control structures are proposed and different algorithms are derived
depending on the local control laws implemented by the agents. Properties and performance of the algorithms are discussed and an application is presented to the
decentralized control of formations of unmanned aerial vehicles (UAVs).
The chapter by D. Zelazo and M. Mesbahi develops a network-centric analysis
and synthesis framework for certain classes of large-scale interconnected systems.
The systems under consideration involve linear dynamic subsystems that interact
with other subsystems via an interconnection topology. The chapter examines the
controllability and observability of such networked systems and investigates the network performance with respect to its H2 system norm. The effect of the structural
properties of the network interconnection on its controllability, observability, and
H2 norm performance are delineated. An algorithm for synthesizing optimal networks in the H2 system norm setting is presented.
The chapter by R.R. Negenborn, G. Hug-Glanzmann, B. De Schutter, and
G. Andersson, illustrates a novel coordination strategy for coordinating multiple
control agents that control overlapping subnetworks in a network. The motivation
stems from the distributed control of large-scale power networks. A simulation study
on an adjusted IEEE 57-bus power network with Flexible Alternating Current Transmission Systems (FACTS) devices as controlled entities is used to validate the coordination strategy.
The chapter by J. Rice, P. Massioni, T. Keviczky, and M. Verhaegen, examines
recently developed distributed control techniques for designing controllers for largescale systems with sparse structure. The methods rely on the structural properties of
the system and the associated control problem by assuming a sequentially semiseparable, decomposable, or identical subsystem architecture. A benchmark problem
of the control of an infinite-dimensional car platoon in an H2 norm setting is used
for a comparative study of the proposed methods.
The chapter by M. Meisami-Azad, J. Mohammadpour-Velni, K. Hiramoto, and
K.M. Grigoriadis, investigates the integrated plant parameter and control parameter
design in controlled large-scale structural systems. The objective is to integrate the
structural parameter and the controller gain design steps to achieve an improved final design. To this end, an explicit upper bound of the H2 /H norm of a collocated
structural system is developed along with a parameterization of output feedback
gains that guarantee such bounds. The proposed bounds are subsequently used in a
convex optimization framework to design structural damping parameters and feedback control gains that meet closed-loop performance specifications.
We thank Steven Elliot, Senior Editor Engineering in Springer for his assistance
and most importantly the authors of the individual chapters for their contribution to
this effort.
Houston, TX
May 2010
Javad Mohammadpour
Contents
Part I Model Reduction, Large-Scale System Modeling

and Applications
Interpolatory Model Reduction of Large-Scale Dynamical Systems . . . . . .
Athanasios C. Antoulas, Christopher A. Beattie, and Serkan Gugercin
1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2
Problem Setting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.1
The General Interpolation Framework . . . . . . . . . . . . . . . .
2.2
Model Reduction via Projection . . . . . . . . . . . . . . . . . . . . .
2.3
Interpolatory Projections . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.4
Error Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3
Interpolatory Optimal H2 Approximation . . . . . . . . . . . . . . . . . . . . .
3.1
An Algorithm for Interpolatory Optimal H2 Model
Reduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2
Numerical Results for IRKA . . . . . . . . . . . . . . . . . . . . . . . .
4
Interpolatory Passivity Preserving Model Reduction . . . . . . . . . . . .
4.1
An Example of Passivity Preserving Model Reduction . . .
5
Structure-Preserving Model Reduction Using Generalized
Coprime Factorizations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.1
A Numerical Example: Driven Cavity Flow . . . . . . . . . . . .
5.2
Second-Order Dynamical Systems . . . . . . . . . . . . . . . . . . .
6
Model Reduction of Parametric Systems . . . . . . . . . . . . . . . . . . . . . .
6.1
Numerical Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7
Model Reduction from Measurements . . . . . . . . . . . . . . . . . . . . . . . .
7.1
Motivation: S-Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.2
The Loewner Matrix Pair and Construction
of Interpolants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.3
Loewner and Pick Matrices . . . . . . . . . . . . . . . . . . . . . . . . .
7.4
Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8
Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3
3
4
6
8
10
14
18
21
22
25
27
30
33
35
37
40
41
41
42
46
47
53
53
xi
xii
Contents
Efficient Model Reduction for the Control of Large-Scale Systems . . . . . .

Richard Colgren
1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2
Spectral Decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3
Simultaneous Gradient Error Reduction . . . . . . . . . . . . . . . . . . . . . . .
4
Balancing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.1
Techniques Not Requiring Balancing . . . . . . . . . . . . . . . . .
4.2
Balancing Over A Disk . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5
Example: Large-Scale System Application . . . . . . . . . . . . . . . . . . . .
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Dynamics of Tensegrity Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Maurcio C. de Oliveira and Anders S. Wroldsen
1
Introduction and Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2
Dynamics of a Single Rigid Rod . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.1
Nodes as Functions of the Configuration . . . . . . . . . . . . . .
2.2
String Forces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3
Generalized Forces and Torques . . . . . . . . . . . . . . . . . . . . .
2.4
Equations of Motion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3
Class 1 Tensegrity Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4
Class k Tensegrity Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.1
A Class 2 Tensegrity Cable Model . . . . . . . . . . . . . . . . . . .
5
Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
59
59
60
62
64
66
68
69
71
73
73
74
76
77
78
78
79
82
82
87
87
Modeling a Complex Aero-Engine Using Reduced Order Models . . . . . . . 89

Xuewu Dai, Timofei Breikin, Hong Wang, Gennady Kulikov,
and Valentin Arkov
1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
2
Gas Turbine System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
2.1
Nonlinear Static Model of Gas Turbine . . . . . . . . . . . . . . . 91
2.2
Nonlinear Dynamic Model of Gas Turbine . . . . . . . . . . . . 92
3
Problem Formulation for Reduced Order Data Driven Modeling . . 93
3.1
Criterion Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
3.2
Model Selection: EE vs. OE . . . . . . . . . . . . . . . . . . . . . . . . . 96
4
NLS for OE Parameter Identification . . . . . . . . . . . . . . . . . . . . . . . . . 99
4.1
Calculation of V (k )/ and the Jacobian . . . . . . . . . . . 100
4.2
Approximation of R(k) and the Hessian . . . . . . . . . . . . . . . 101
5
Application and Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
5.1
First-Order Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
5.2
Second-Order Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
6
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
Contents
xiii
Part II Large-Scale Systems Control and Applications

Robust Control of Large-Scale Systems: Efficient Selection
of Inputs and Outputs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
Javad Lavaei, Somayeh Sojoudi, and Amir G. Aghdam
1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
2
Preliminaries and Problem Formulation . . . . . . . . . . . . . . . . . . . . . . . 117
2.1
Background on Sum-of-Squares . . . . . . . . . . . . . . . . . . . . . 118
3
Robust Controllability Degree . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
3.1
Special Case: A Polytopic Region . . . . . . . . . . . . . . . . . . . . 124
3.2
Comparison with Existing Results . . . . . . . . . . . . . . . . . . . 128
4
Numerical Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
5
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
Decentralized Output-Feedback Control of Large-Scale Interconnected
Systems via Dynamic High-Gain Scaling . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
P. Krishnamurthy and F. Khorrami
1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
2
Decentralized Control Based on The Adaptive Dual Dynamic
High-Gain Scaling Paradigm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
2.1
Assumptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
2.2
Observer and Controller Designs . . . . . . . . . . . . . . . . . . . . . 140
2.3
Stability Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
3
Generalized Scaling: Application to Decentralized Control . . . . . . . 151
3.1
Assumptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
3.2
Observer Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
3.3
Controller Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
3.4
Stability Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
Decentralized Output Feedback Guaranteed Cost Control of Uncertain
Markovian Jump Large-Scale Systems: Local Mode Dependent Control
Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
Junlin Xiong, Valery A. Ugrinovskii, and Ian R. Petersen
1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
2
Problem Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170
3
Guaranteed Cost Controller Design . . . . . . . . . . . . . . . . . . . . . . . . . . 173
3.1
Design Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
3.2
Design of Global Mode Dependent Controllers . . . . . . . . . 176
3.3
The Main Result: Design of Local Mode Dependent
Controllers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180
3.4
Design Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
4
An Illustrative Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
5
Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186
xiv
Contents
Appendix 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
Appendix 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
Consensus Based Multi-Agent Control Algorithms . . . . . . . . . . . . . . . . . . . 197
Milos S. Stankovic, Dusan M. Stipanovic, and Srdjan S. Stankovic
1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
2
Problem Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
3
Consensus at the Control Input Level . . . . . . . . . . . . . . . . . . . . . . . . . 200
3.1
Algorithms Derived from the Local Dynamic Output
Feedback Control Laws . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200
3.2
Algorithms Derived from the Local Static Feedback
Control Laws . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
4
Consensus at the State Estimation Level . . . . . . . . . . . . . . . . . . . . . . 207
5
Consensus Based Decentralized Control of UAV Formations . . . . . 210
5.1
Formation Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210
5.2
Global LQ Optimal State Feedback . . . . . . . . . . . . . . . . . . . 212
5.3
Decentralized State Estimation . . . . . . . . . . . . . . . . . . . . . . 213
5.4
Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
Graph-Theoretic Methods for Networked Dynamic Systems:
Heterogeneity and H2 Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219
Daniel Zelazo and Mehran Mesbahi
1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219
1.1
Preliminaries and Notations . . . . . . . . . . . . . . . . . . . . . . . . . 221
2
Canonical Models of Networked Dynamic Systems . . . . . . . . . . . . . 224
3
Analysis and Graph-Theoretic Performance Bounds . . . . . . . . . . . . 229
3.1
Observability and Controllability of NDS . . . . . . . . . . . . . 229
3.2
Graph-Theoretic Bounds on NDS Performance . . . . . . . . . 232
4
Topology Design for NDS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241
4.1
H2 Topology Design for NDS Coupled at the Output . . . . 242
4.2
Sensor Placement with H2 Performance
for NDS Coupled at the State . . . . . . . . . . . . . . . . . . . . . . . . 244
5
Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 246
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247
A Novel Coordination Strategy for Multi-Agent Control Using
Overlapping Subnetworks with Application to Power Systems . . . . . . . . . . 251
R.R. Negenborn, G. Hug-Glanzmann, B. De Schutter, and G. Andersson
1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251
1.1
Multi-Agent Control of Power Networks . . . . . . . . . . . . . . 252
1.2
Control of Subnetworks . . . . . . . . . . . . . . . . . . . . . . . . . . . . 253
1.3
Optimal Power Flow Control . . . . . . . . . . . . . . . . . . . . . . . . 255
1.4
Goal and Outline of This Chapter . . . . . . . . . . . . . . . . . . . . 256
Contents
xv
Modeling of Network Characteristics and Control Objectives . . . . . 256

2.1
Network Characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . 256
2.2
Control Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257
2.3
Definition of Subnetworks . . . . . . . . . . . . . . . . . . . . . . . . . . 257
3
Multi-Agent Control of Touching Subnetworks . . . . . . . . . . . . . . . . 258
3.1
Internal and External Nodes . . . . . . . . . . . . . . . . . . . . . . . . . 258
3.2
Control Problem Formulation for One Agent . . . . . . . . . . . 259
3.3
Control Scheme for Multiple Agents . . . . . . . . . . . . . . . . . 262
4
Multi-Agent Control for Overlapping Subnetworks . . . . . . . . . . . . . 263
4.1
Common Nodes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263
4.2
Control Problem Formulation for One Agent . . . . . . . . . . . 264
4.3
Control Scheme for Multiple Agents . . . . . . . . . . . . . . . . . 267
5
Application: Optimal Flow Control in Power Networks . . . . . . . . . . 268
5.1
Parameters of the Power Network . . . . . . . . . . . . . . . . . . . . 268
5.2
Steady-state Characteristics of Power Networks . . . . . . . . 268
5.3
Control Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273
5.4
Setting Up the Control Problems . . . . . . . . . . . . . . . . . . . . . 273
5.5
Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 274
6
Conclusions and Future Research . . . . . . . . . . . . . . . . . . . . . . . . . . . . 277
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 277
Distributed Control Methods for Structured Large-Scale Systems . . . . . . . 279
Justin Rice, Paolo Massioni, Tamas Keviczky, and Michel Verhaegen
1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279
2
Problem Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 280
2.1
H2 Problem and Exact Solution . . . . . . . . . . . . . . . . . . . . . 281
3
A Rational Laurent Operator Structure Preserving
Iterative Approach to Distributed Control . . . . . . . . . . . . . . . . . . . . . 282
3.1
L-Operator Sign Function . . . . . . . . . . . . . . . . . . . . . . . . . . . 284
3.2
Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 285
3.3
Convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 285
3.4
Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 286
3.5
Numerical Difficulties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 287
3.6
Application to the Example Problem . . . . . . . . . . . . . . . . . 288
4
Distributed Control Design for Decomposable Systems . . . . . . . . . . 288
4.1
General Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 289
4.2
5
Distributed LQR of Identical Systems . . . . . . . . . . . . . . . . . . . . . . . . 294
5.1
Special Properties of LQR for Dynamically Decoupled
Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 295
5.2
6
Numerical Results of the Car Platoon Benchmark Problem . . . . . . . 299
7
Conclusions and Open Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 301
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 302
xvi
Contents
Integrated Design of Large-Scale Collocated Structural System

and Control Parameters Using a Norm Upper Bound Approach . . . . . . . . 305
Mona Meisami-Azad, Javad Mohammadpour Velni, Kazuhiko Hiramoto,
and Karolos M. Grigoriadis
1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 305
2
Symmetric Output Feedback Control of Collocated Systems . . . . . 307
3
Upper Bounds on Collocated Structural System Norms . . . . . . . . . . 308
4
Integrated Damping and Control Design Using the Analytical
Bound Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 311
4.1
Integrated Design Based on an H Specification . . . . . . . 311
4.2
Integrated Design Based on an H2 Specification . . . . . . . . 312
4.3
Integrated Design Based on a Mixed H2 /H
Specification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 314
4.4
Decentralized Control Using the Norm Upper Bound
Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 315
4.5
Additional Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 317
5
Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 317
6
Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 326
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 326
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 329
List of Contributors
Amir G. Aghdam
Department of Electrical and Computer Engineering, Concordia University,
Montreal, Canada, H3G 2W1, e-mail: aghdam@ece.concordia.ca
Goran Andersson
Power Systems Laboratory, ETH Zurich, Physikstrasse 3, 8092 Zurich, Switzerland,
e-mail: andersson@eeh.ee.ethz.ch
Athanasios C. Antoulas
Department of Electrical and Computer Engineering, Rice University, Houston,
TX 77005-1892, USA, e-mail: aca@rice.edu
Valentin Arkov
Department of Automated Control Systems, Ufa State Aviation Technical
University, K. Marx Street 12, Ufa, 450000, Russia,
e-mail: t.breikin@manchester.ac.uk
Christopher A. Beattie
Department of Mathematics, Virginia Tech, Blacksburg, VA 24061-0123, USA,
e-mail: beattie@vt.edu
Timofei Breikin
Control Systems Center, School of Electrical and Electronic Engineering,
University of Manchester, Manchester, UK, M60 1QD,
Richard D. Colgren
Viking Aerospace LLC, 100 Riverfront Rd., Suite B, Lawrence, KS 66044, USA,
e-mail: rcolgren@gmail.com
Xuewu Dai
School of Electronic and Information Engineering, Southwest University,
Chongqing, China, 400715, e-mail: dxw.dai@gmail.com
xvii
xviii
Maurcio C. de Oliveira
Department of Mechanical and Aerospace Engineering, University
of California San Diego, 9500 Gilman Dr., La Jolla CA, 92093-0411, USA,
e-mail: mauricio@ucsd.edu
Bart De Schutter
Delft Center for Systems and Control & Department of Marine and Transport
Technology, Delft University of Technology, Mekelweg 2, 2628 CD Delft,
The Netherlands, e-mail: b@deschutter.info
Department of Mechanical Engineering, University of Houston, Houston,
TX 77204, USA, e-mail: karolos@uh.edu
Serkan Gugercin
Department of Mathematics, Virginia Tech, Blacksburg, VA 24061-0123, USA,
e-mail: gugercin@math.vt.edu
Kazuhiko Hiramoto
Department of Mechanical and Production Engineering, Niigata University,
8050 Ikarashi-2-no-cho, Nishi-ku, Niigata 950-2181, Japan,
e-mail: hiramoto@eng.niigata-u.ac.jp
Gabriela Hug-Glanzmann
Department of Electrical and Computer Engineering, Carnegie Mellon University,
5000 Forbes Avenue, Pittsburgh, PA, USA, e-mail: ghug@andrew.cmu.edu
Tamas Keviczky
Delft Center for Systems and Control, Delft University of Technology, 2628 CD,
Delft, The Netherlands, e-mail: t.keviczky@tudelft.nl
Farshad Khorrami
Control/Robotics Research Laboratory (CRRL), Department of Electrical
and Computer Engineering, Polytechnic Institute of NYU, Brooklyn, NY 11201,
USA, e-mail: khorrami@smart.poly.edu
Prashanth Krishnamurthy
Control/Robotics Research Laboratory (CRRL), Department of Electrical and
Computer Engineering, Polytechnic Institute of NYU, Brooklyn, NY 11201,
USA, e-mail: pk@crrl.poly.edu
Gennady Kulikov
Department of Automated Control Systems, Ufa State Aviation Technical
University, K. Marx Street 12, Ufa, 450000, Russia,
Javad Lavaei
Control and Dynamical Systems, California Institute of Technology, Pasadena,
CA 91125, USA, e-mail: lavaei@cds.caltech.edu
xix
Paolo Massioni
Delft, The Netherlands, e-mail: p.massioni@tudelft.nl
Mona Meisami-Azad
TX 77204, USA, e-mail: mmeisami@mail.uh.edu
Mehran Mesbahi
Department of Aeronautics and Astronautics, University of Washington,
Box 352400, Seattle, WA 98195, USA, e-mail: mesbahi@aa.washington.edu
Javad Mohammadpour
TX 77204, USA, e-mail: jmohammadpour@uh.edu
Rudy R. Negenborn
Delft Center for Systems and Control, Delft University of Technology, Mekelweg
2, 2628 CD Delft, The Netherlands, e-mail: r.r.negenborn@tudelft.nl
Ian R. Petersen
School of Engineering and Information Technology, University of New South
Wales at the Australian Defence Force Academy, Northcott Drive, Canberra,
ACT 2600, Australia, e-mail: i.r.petersen@gmail.com
Justin Rice
Delft, The Netherlands, e-mail: j.k.rice@tudelft.nl
Somayeh Sojoudi
Control and Dynamical Systems, California Institute of Technology, Pasadena,
CA 91125, USA, e-mail: sojoudi@cds.caltech.edu
Milos S. Stankovic
ACCESS Linnaeus Center, School of Electrical Engineering, Royal Institute
of Technology, 100-44 Stockholm, Sweden, e-mail: milsta@kth.se
Srdjan S. Stankovic
School of Electrical Engineering, University of Belgrade, Serbia,
e-mail: stankovic@etf.rs
Dusan M. Stipanovic
Department of Industrial and Enterprise Systems Engineering and the Coordinated
Science Laboratory, University of Illinois at Urbana-Champaign, IL, USA,
e-mail: dusan@illinois.edu
Valery A. Ugrinovskii
School of Engineering and Information Technology, University of New South
Wales at the Australian Defence Force Academy, Northcott Drive, Canberra,
ACT 2600, Australia, e-mail: v.ugrinovskii@gmail.com
xx
Michel Verhaegen
Delft, The Netherlands, e-mail: m.verhaegen@moesp.org
Hong Wang
Control Systems Center, School of Electrical and Electronic Engineering,
University of Manchester, Manchester M60 1QD, UK,
e-mail: hong.wang@manchester.ac.uk
Anders S. Wroldsen
Center for Ships and Ocean Structures, Norwegian University of Science
and Technology, Trondheim, Norway, e-mail: wroldsen@ntnu.no
Junlin Xiong
Department of Automation, University of Science and Technology of China,
Hefei 230026, China, e-mail: junlin.xiong@gmail.com
Daniel Zelazo
Institute for Systems Theory and Automatic Control, University of Stuttgart,
Pfaffenwaldring 9, 70550 Stuttgart, Germany,
e-mail: Daniel.Zelazo@ist.uni-stuttgart.de
Part I
Model Reduction, Large-Scale System

Modeling and Applications
Interpolatory Model Reduction of Large-Scale

Dynamical Systems
Athanasios C. Antoulas, Christopher A. Beattie, and Serkan Gugercin
1 Introduction
Large scale dynamical systems are a common framework for the modeling and
control of many complex phenomena of scientific interest and industrial value,
with examples of diverse origin that include signal propagation and interference
in electric circuits, storm surge prediction before an advancing hurricane, vibration suppression in large structures, temperature control in various media, neurotransmission in the nervous system, and behavior of micro-electro-mechanical
systems. Direct numerical simulation of underlying mathematical models is one of
few available means for accurate prediction and control of these complex phenomena. The need for ever greater accuracy compels inclusion of greater detail in the
model and potential coupling to other complex systems leading inevitably to very
large-scale and complex dynamical models. Simulations in such large-scale settings can make untenable demands on computational resources and efficient model
utilization becomes necessary. Model reduction is one response to this challenge,
wherein one seeks a simpler (typically lower order) model that nearly replicates the
behavior of the original model. When high fidelity is achieved with a reduced-order
model, it can then be used reliably as an efficient surrogate to the original, perhaps
replacing it as a component in larger simulations or in allied contexts such as development of simpler, faster controllers suitable for real time applications.
A.C. Antoulas
Department of Electrical and Computer Engineering, Rice University, Houston,
TX 77005-1892, USA
e-mail: aca@rice.edu
C.A. Beattie
Department of Mathematics, Virginia Tech, Blacksburg, VA 24061-0123, USA
e-mail: beattie@vt.edu
S. Gugercin
Department of Mathematics, Virginia Tech, Blacksburg, VA 24061-0123, USA
e-mail: gugercin@math.vt.edu
J. Mohammadpour and K.M. Grigoriadis, Efficient Modeling and Control,
of Large-Scale Systems, DOI 10.1007/978-1-4419-5757-3 1,
A.C. Antoulas et al.
Interpolatory model reduction methods have emerged as effective strategies for

approximation of large-scale linear dynamical systems. These methods produce
reduced models whose transfer function interpolates the original system transfer
function at selected interpolation points. They are closely related to rational Krylovbased methods and indeed, are precisely that for single input/single output systems.
One main reason why interpolatory model reduction has become the method of
choice in true large-scale settings is that this class of methods is numerically stable
and well-suited to large scale computations; the model reduction process does not
require any dense matrix transformations such as singular value decompositions.
Moreover, the approximation theory behind this approach is similar to that of rational interpolants used to approximate meromorphic functions. The goal of this
contribution is to present a survey of model reduction by interpolation and provide
a selection of the most recent developments in this direction.
2 Problem Setting
Linear dynamical systems are principally characterized through their inputoutput
map S : u y, mapping inputs u to outputs y via a state-space realization given as:

Ex(t) = Ax(t) + Bu(t)
S:
with x(0) = 0,
(1)
y(t) = Cx(t) + Du(t)
where A, E Rnn , B Rnm , C R pn , and D R pm are constant matrices. In
(1), x(t) Rn , u(t) Rm and y(t) R p are, respectively, an internal variable (the
state if E is non-singular), the input and the output of the system S. We refer to S
as a single-input/single-output (SISO) system when m = p = 1 (scalar-valued input,
scalar-valued output) and as a multi-input/multi-output (MIMO) system otherwise.
Systems of the form (1) with extremely large state-space dimension n arise in
many disciplines; see [2] and [57] for a collection of such examples. Despite large
state-space dimension, in most cases, the state space trajectories, x(t), hew closely to
subspaces with substantially lower dimensions and evolve in ways that do not fully
occupy the state space. The original model S behaves nearly as if it had many fewer
internal degrees-of-freedom, in effect, much lower state-space dimension. Model reduction seeks to produce a surrogate dynamical system that evolves in a much lower
dimensional state space (say, dimension r n), yet is able to mimic the original dynamical system, recovering very nearly the original inputoutput map. We want the
reduced inputoutput map, Sr : u yr , to be close to S in an appropriate sense.
Being a smaller version of the original dynamical model, the inputoutput map
Sr is described by the reduced system in state-space form as:

Er x r (t) = Ar xr (t) + Br u(t)
Sr :
(2)
with xr (0) = 0,
yr (t) = Cr xr (t) + Dr u(t)
Interpolatory Model Reduction of Large-Scale Dynamical Systems
where Ar , Er Rrr , Br Rrm , Cr R pr and Dr R pm . Note that the number

of inputs, m, and the number of outputs, p, are the same for both the original and
reduced models; only the internal state-space dimensions differ: r n. A successful
reduced-order model should meet the following criteria:
Goals for Reduced Order Models
1. The reduced inputoutput map Sr should be uniformly close to S in an
appropriate sense. That is, when presented with the same inputs, u(t),
the difference between full and reduced system outputs, y yr , should
be small with respect to a physically relevant norm over a wide range of
system inputs, such as over all u with bounded energy (e.g., in the unit
ball of L2 ([0, ), Rm ) ).
2. Critical system features and structure should be preserved in the reduced
order system. This can include passivity, Hamiltonian structure, subsystem interconnectivity, or second-order structure.
3. Strategies for obtaining Er , Ar , Br , Cr , and Dr should lead to robust, numerically stable algorithms and furthermore require minimal applicationspecific tuning with little to no expert intervention. It is important that
model reduction methods be computationally efficient and reliable so that
very large problems remain tractable; they should be robust and largely
automatic to allow the broadest level of flexibility and applicability in
complex multiphysics settings.
Whenever the input u(t) is exponentially bounded that is, when there is a fixed
R such that u(t) O(e t ), then x(t) and y(t) from (1) and xr (t) and yr (t) from
(2) will also be exponentially bounded and the Laplace transform can be applied to
(1) and (2) to obtain

y(s) = C (sE A)1 B + D
u(s),
(3)

yr (s) = Cr (sEr Ar )1 Br + Dr
u(s),
(4)
where we have denoted Laplace transformed quantities with . We define transfer

functions accordingly from (3) and (4)
H(s) =C (sE A)1 B + D
Hr (s) =Cr (sEr Ar )1 Br + Dr
so that

u(s)
y(s)
yr (s) = [H(s) Hr (s)]
(5)
(6)
For a given input, the loss in fidelity in passing from the full model to a reduced
model can be associated with the difference between respective system transfer
functions over a range of inputs. This is evaluated more precisely in Sect. 2.4.
It is a fact of life that many dynamical systems of critical interest are nonlinear.
By contrast, we will consider here only linear dynamical systems and the methods we discuss fundamentally exploit this linearity. We do this unapologetically but
note that by narrowing the scope of the problems that one considers in this way, one
is able to be far more ambitious and demanding in the quality of outcomes that are
achieved, and indeed, the techniques we discuss here are often dramatically successful. To the extent that many methods for approaching nonlinear systems build upon
the analysis of carefully chosen linearizations of these systems, the methods we describe here can be expected to play a role in evolving strategies for the reduction of
large scale nonlinear dynamical systems as well. See, for instance, [12].
2.1 The General Interpolation Framework

Interpolation is a very simple and yet effective approach that is used ubiquitously
for the general approximation of complex functions using simpler ones. The construction of such approximations is easy: indeed, school children learn how to construct piecewise linear approximations and Taylor polynomial approximations to
functions through interpolation. The accuracy of the resulting approximations and
the connections with strategic placement of interpolating points has been studied in
many broad contexts indeed, in the case of interpolation of meromorphic functions
by polynomials or rational functions, the associated error analysis constitutes a large
body of work tied in closely with potential theory and classical complex analysis.
Many of these recognized advantages of approximation via interpolation translate
immediately into boons in our setting as well as we seek to approximate a difficult
transfer function, H(s), with a simpler one, Hr (s).
Simply stated, our overarching goal is to produce a low order transfer function,
Hr (s), that approximates a large order transfer function, H(s) with high fidelity:
for order r n, we want Hr (s) H(s) in a sense that will be made precise later.
Interpolation is the primary vehicle with which we approach this problem. Naively,
we could select a set of points {i }ri=1 C and then seek a reduced order transfer
function, Hr (s), such that Hr (i ) = H(i ) for i = 1, . . . , r. This is a good starting
place for SISO systems but turns out to be overly restrictive for MIMO systems,
since the condition Hr (i ) = H(i ) in effect imposes m p scalar conditions at each
interpolation point. This may not be realizable.
A far more advantageous formulation is to consider interpolation conditions that
are imposed at specified interpolation points as above but that only act in specified
directions, that is, tangential interpolation. We state this formally as part of our first
problem of focus:
Problem 1: Model Reduction Given State Space System Data

Given a full-order model (1) specified through knowledge of the system matrices, E, A, B, C, and D and given
left interpolation points:
{i }qi=1 C,
with corresponding
left tangent directions:
{ci }qi=1 C p ,
and
right interpolation points:

{i }ri=1 C
with corresponding
right tangent directions:
{b i }ri=1 Cm .
Find a reduced-order model (2) through specification of reduced system matrices Er , Ar , Br , Cr , and Dr such that the associated transfer function, Hr (s),
in (6) is a tangential interpolant to H(s) in (5):
c Ti Hr (i ) = c Ti H(i )
for i = 1, . . . , q,
and
Hr ( j )b j = H( j )b j ,
for j = 1, . . . , r,
(7)
Interpolation points and tangent directions are selected to realize the model
reduction goals described on page 5.
It is of central concern to determine effective choices for interpolation points

and tangent directions. In Sect. 3, we discuss the selection of interpolation points
and tangent directions that can yield optimal reduced order models with respect to
H2 quality measures of system approximation. In Sect. 4, we describe selection of
interpolation points and tangent directions that guarantee passivity is preserved in
the reduced-order model. Preservation of other types of system structure such as
second-order systems, systems that involve delays or memory terms, and systems
having a structured dependence on parameters is discussed in Sects. 5 and 6.
Tangential interpolation is a remarkably flexible framework within which to consider model reduction when one considers the significance of the interpolation data.
Note that if y = H( )b then e t y is precisely the response of the full order system
so the tangential interpolation conditions that
to a pure input given by u(t) = e t b,
characterize Hr (s) could (at least in principle) be obtained from measured input
output data drawn directly from observations on the original system. For example,
if = 0 , one is observing in y , the sinusoidal response of the system to a pure
tone input of frequency 0 . Similarly, if the dual dynamical system were driven by
an input given by e t c producing an output e t z then c T H( ) = z T . This creates
the following alternative problem setting that we consider based entirely on observed inputoutput response data and requiring no other a priori information about
system-related quantities.
Problem 2: Model Reduction Given InputOutput Data

Given a set of inputoutput response measurements on the full-order system
specified by
left driving frequencies:
q
{i }i=1 C,
using left input directions:
{ci }qi=1 C p ,
producing left responses:
{zi }qi=1 Cm ,
and
right driving frequencies:

{i }ri=1 C
using right input directions:
{b i }ri=1 Cm
producing right responses:
{yi }ri=1 C p
Find a dynamical system model (2) by specifying (reduced) system matrices

Er , Ar , Br , Cr , and Dr such that the associated transfer function, Hr (s), in (6)
is a tangential interpolant to the given data:
c Ti Hr (i ) = z Ti
for i = 1, . . . , q,
and
Hr ( j )b j = y j ,
for j = 1, . . . , r,
(8)
Interpolation points and tangent directions are determined (typically) by the

availability of experimental data.
The structure of Problem 2 introduces difficulties involved in the requirement that
noisy and inconsistent data must be accommodated appropriately in this case, we
describe how to construct a reduced order system described by the system matrices
Er , Ar , Br , Cr , and Dr so that experimentally observed response data (reformulated as approximate interpolation conditions) are matched at least approximately:
Hr (i )b i y i and c Tj H( j ) z Tj .
Numerous issues arise spontaneously for both problem types and we discuss
here what we feel is a representative subset. Note for both problems, it is necessary to have a computationally stable method for constructing the (reduced) system matrices Er , Ar , Br , Cr , and Dr that produces an associated transfer function,
Hr (s), satisfying the interpolation conditions. We discuss this in Sects. 2.2 and 2.3
for Problem 1 and Sect. 7.2 for Problem 2.
2.2 Model Reduction via Projection

When the full-order dynamical system S described in (1) has been specified via
the state-space matrices E, A, B, C, and D, most model reduction methods proceed
with some variation of a Petrov-Galerkin projective approximation to construct a
reduced-order model Sr . We motivate this approach by describing the evolution of
the full order model (1) in an indirect way although the systems described in
(1) and (2) typically are real-valued, we find it convenient to allow all quantities
involved to be complex-valued:
Find x(t) contained in Cn such that
Ex(t) Ax(t) B u(t)
Cn (i.e., = 0).
Then the associated output is y(t) = Cx(t) + Du(t).

The Petrov-Galerkin projective approximation proceeds by replacing Cn in the above
description with (nominally small) subspaces of Cn : we choose an r-dimensional
trial subspace, the right modeling subspace, Vr Cn , and an r-dimensional test
subspace, the left modeling subspace, Wr Cn and then we describe the evolution
of a reduced order model in this way:
Find v(t) contained in Vr such that
Ev(t) Av(t) B u(t) Wr .
Then the associated output is yr (t) = Cv(t) + Du(t).
(9)
The dynamics described by (9) can be represented as a dynamical system evolving in a reduced order state-space as in (2) once bases are chosen for the two subspaces Vr and Wr . Let Ran(R) denote the range of a matrix R. Let Vr Cnr and
Wr Cnr be matrices defined so that Vr = Ran(Vr ) and Wr = Ran(Wr ). We can
represent the reduced system trajectories as v(t) = Vr xr (t) with xr (t) Cr for each
t and the Petrov-Galerkin approximation (9) can be rewritten as
WrT (EVr x r (t) AVr xr (t) B u(t)) = 0 and yr (t) = CVr xr (t) + Du(t),
leading to the reduced order state-space representation (2) with
Er = WTr EVr ,
Br = WTr B,
Ar = WTr AVr ,
Cr = CVr ,
and
Dr = D.
(10)
We note that the definitions of reduced-order quantities in (10) are invariant under
change of basis for the original state space, so the quality of reduced approximations
evidently will depend only on effective choices for the right modeling space Vr =
Ran(Vr ) and the left modeling space Wr = Ran(Wr ). We choose the modeling
subspaces to enforce interpolation which allows us to shift our focus to how best to
choose effective interpolation points and tangent directions.
Since D R pm and both p and m are typically of only modest size, usually
D does not play a significant role in the cost of simulation and Dr = D is both a
common choice and a natural one arising in the context of a Petrov-Galerkin approximation as described in (10). Note that if E and Er are both nonsingular then
the choice Dr = D also enforces interpolation at infinity: lim H(s) = lim Hr (s) = D,
s
s
facilitating, in effect, a good match between true and reduced system outputs for sufficiently high frequency inputs. The case that E is singular and other performance
goals can motivate choices for Dr = D. This is discussed in some detail on p. 13.
10
2.3 Interpolatory Projections

Given the full-order dynamical system H(s), a set of r right interpolation points
{i }ri=1 C, with right directions {bi }ri=1 Cm , and a set of q left interpolation
points {i }qi=1 C, with left-tangential directions {ci }qi=1 C p , interpolatory model
reduction involves finding a reduced-order model Hr (s) that tangentially interpolates H(s) using this given data: find Hr (s) so that
H(i )bi = Hr (i )bi ,
cTj H( j ) = cTj Hr ( j ),
for i = 1, . . . , r,
for j = 1, . . . , r,
(11)
We wish to interpolate H(s) without ever computing the quantities to be matched

since these numbers are numerically ill-conditioned, as Feldman and Freund [33]
illustrated for the special case of single-input/single-output dynamical systems. Remarkably, this can be achieved by employing Petrov-Galerkin projections with carefully chosen test and trial subspaces.
Interpolatory projections for model reduction were initially introduced by Skelton et. al. in [30, 90, 91]. Grimme [42], later, modified this approach into a numerically efficient framework by utilizing the rational Krylov subspace method of Ruhe
[79]. The projection framework for the problem setting we are interested in, that is,
for rational tangential interpolation of MIMO dynamical systems, has been recently
developed by Gallivan et al. [39] for the case of Problem 1 and by Mayo and Antoulas [66] for the case of Problems 2 as laid out in 2.1. We will be using these
frameworks throughout the manuscript.
We first illustrate and prove how to solve the rational tangential interpolation
problem in (11) by projection. The general case for matching derivatives of the
transfer function (in effect, generalized Hermite interpolation) will be presented in
Theorem 1.2. As we point out above, the Petrov-Galerkin approximation described
in (10) leads to the choice of Dr = D in a natural way, but other choices are possible. We will first assume Dr = D, so that Hr (s) in Theorems 1.1 and 1.2 below is
assumed to be obtained as in (10).
Theorem 1.1. Let , C be such that s E A and s Er Ar are invertible for
s = , . Also let Vr , Wr Cnr in (10) have full-rank. If b Cm and c C are
fixed nontrivial vectors then
(a) if ( E A)1 Bb Ran(Vr ), then H( )b = Hr ( )b;

T
(b) if cT C ( E A)1 Ran(Wr ), then cT H( ) = cT Hr ( ); and
(c) if both (a) and (b) hold, and = , then cT H ( )b = cT Hr ( )b as well.
Proof. We follow the recent proofs provided in [49] and [16]. Define
Pr (z) = Vr (zEr Ar )1 WrT (zEr A)
Qr (z) = (zE A)Pr (z)(zE A)
and
= (zE A)Vr (zEr Ar )1 WrT .
11
Both Pr (z) and Qr (z) are analytic matrix-valued functions in neighborhoods of

z = and z = . It is easy to verify that both Pr (z) and Qr (z) are projectors,
i.e. P2r (z) = Pr (z) and Q2r (z) = Qr (z). Moreover, for all z in a neighborhood of ,
Vr = Ran(Pr (z)) = Ker (I Pr (z)) and W
r = Ker(Qr (z)) = Ran (I Qr (z)) where
Ran(R) denotes the kernel of a matrix R. First observe that

H(z) Hr (z) = C(zE A)1 (I Qr (z)) (zE A) I Pr (z) (zI A)1 B (12)
Evaluating this expression at z = and post-multiplying by b yields the first assertion; evaluating (12) at z = and pre-multiplying by cT yields the second. Note
(( + )E A)1 = ( E A)1 ( E A)1E( E A)1 + O( 2 )
(( + )Er Ar )1 = ( Er Ar )1 ( Er Ar )1 Er ( Er Ar )1 + O( 2 )
so evaluating (12) at z = + , premultiplying by cT, and postmultiplying by b
under the hypotheses of the third assertion yields
cT H( + )b cT Hr ( + )b = O( 2 ).
Now observe that since cT H( )b = cT Hr ( )b,
1

1 T
c H( + )b cT H( )b cT Hr ( + )b cT Hr ( )b 0,
as 0, which proves the third assertion.

Theorem 1.1 shows that given a set of distinct right interpolation points (right
shifts) {i }ri=1 , a set of distinct interpolation left points (left shifts) { j }rj=1 , left
tangential directions {ck }rk=1 C p , and right tangential directions {bk }rk=1 Cm , the
solution of the tangential rational Hermite interpolation problem is straightforward
and constructive. One simply computes matrices Vr and Wr such that

(13)
Vr = (1 E A)1 Bb1 , . . . , (r E A)1 Bbr and
cT1 C(1 E A)1
..
WrT =
.
.
(14)
cTr C(r E A)1
The reduced order system Hr (s) = Cr (sEr Ar )1 Br defined by (10) then solves
the tangential Hermite interpolation problem provided that i Er Ar and i Er Ar
are nonsingular for each i = 1, . . . , r.
For completeness, we describe how to solve the analogous generalized Hermite
interpolation problem, which involves matching transfer function values and higher
derivative values. The mth derivative of H(s) with respect to s evaluated at s = ,
will be denoted by H(m) ( ).
12
Theorem 1.2. Let , C be such that s E A and s Er Ar are invertible for

s = , . If b Cm and c C are fixed nontrivial vectors then
j1

( E A)1 Bb Ran(Vr ) for j = 1, ., N
(a) if ( E A)1 E
()
then H() ( )b = Hr ( )b f or = 0, 1, . . . , N 1
j1

( E A)T CT c Ran(Wr ) for j = 1, ., M,
(b) if ( E A)T ET
()
then cT H() ( ) = cT Hr ( )b f or = 0, 1, . . . , M 1;
()
(c) if both (a) and (b) hold, and if = , then cT H() ( )b = cT Hr ( )b,
for = 1, . . . , M + N + 1
Theorem 1.2 solves the rational tangential interpolation problem via projection.
All one has to do is to construct the matrices Vr and Wr according to the theorem,
and the reduced order model is guaranteed to satisfy the interpolation conditions.
Note that Theorems 1.1 and 1.2 solve the interpolation problem without ever explicitly computing the values that are interpolated, since that computation is known to
be prone to poor numerical conditioning.
Model reduction by interpolation (also called Krylov-based model reduction
or moment matching in the single-input/single-output case) has been recognized
previously, especially in the circuit simulation community; see, for example,
[911, 33, 37, 73] and the references therein. [74, 76] also investigated the interpolation problem for efficient circuit analysis, however required explicit computation of
the transfer function values and derivatives to be matched.
Despite their computational efficiency compared to Gramian-based model reduction techniques such as balanced-truncation [70, 72] and optimal Hankel norm
approximation [40], interpolatory model reduction used to have the main drawback
of depending on an ad hoc selection of interpolation points and tangential directions, i.e. how to choose the interpolation points and tangential directions optimally
or at least effectively were not known until very recently. This has raised some criticism of interpolatory methods for not guaranteeing stability and not providing error
bounds or error estimates for the resulting reduced-order model. To solve the stability issue, an implicit restart technique was proposed by Grimme et al. [43] which
removed the unstable poles via an implicitly restarted Arnoldi process; however
the reduced-model no longer interpolated the original one, but rather a nearby one.
Gugercin and Antoulas [47] proposed combining interpolatory projection methods with Gramian-based approaches that guaranteed stability of the reduced-order
model. Recently, reformulation of the Fourier model reduction approach of Willcox and Megretzki [87] in the interpolation framework by Gugercin and Willcox
[50] has shown this method is indeed a stability preserving interpolatory model reduction method with an H upper bound, not necessarily easily computable. Bai
[93], and Gugercin and Antoulas [46] have also focused on the error estimates and
provided error expressions (not error bounds) for the interpolatory model reduction. However, despite these efforts, the optimal interpolation points and tangential
direction selection remained unresolved until very recently.
13
This optimal point selection strategy together with some other very recent
developments in interpolation-based model reduction will be presented in the remaining of the manuscript: Sect. 3 will illustrate how to choose the interpolation
points to minimize the H2 norm of the error systems. Passivity of the reduced-model
is vital when dealing with model coming from, especially, circuit theory. Section 4
will show how to construct high-quality passivity-preserving reduced-order models
using interpolation. Obtaining reduced-order models solely from input-output measurements without the knowledge of the original state-space quantities has been the
focus of recent research. In Sect. 7, we will illustrate how to achieve this in the interpolation framework. Even though the state-space formulation of the original system
in (1) is quite general, and indeed, will include, with suitable reformulation, the majority of linear dynamical systems, some dynamical systems might have a natural
description that takes a form quite different from this standard realization; such as
systems with internal delays, memory terms, etc. Section 5 will extend the interpolatory model reduction theory described here to settings where the transfer functions
H(s) take a much more general form than that of (5), allowing interpolatory model
reduction for a much richer category of dynamical systems.
Constructing Interpolants with Dr = D. Certain circumstances in model reduction require that a reduced-order model have a Dr term different than D. One such
case is that of a singular E with a non-defective eigenvalue at 0. This occurs when
the internal system dynamics has an auxiliary algebraic constraint that must always
be satisfied (perhaps representing rigid body translations or rotations or fluid incompressibility, for example). The dynamical system description given in (1) is then a
differential algebraic equation (DAE) and our condition that E have a non-defective
eigenvalue at 0 amounts to the requirement that the DAE be of index 1 (see [60]).
In this case, lim H(s) = D. If we want Hr (s) to match H(s) asymptotically at high
s
frequencies then we should choose

Dr = lim H(s) Cr (sEr Ar )1 Br
s
If r < rank(E) then it will often happen that Er will be nonsingular, in which case
then lim Hr (s) = Dr and we may assign Dr = lims H(s) if we wish Hr (s) to
s
match H(s) asymptotically at high frequencies. See, for example, [21] where a
reduced-order model with Dr = D constructed in case of a singular E to match
H(s) and its higher derivatives around s = .
One may be less concerned with Hr (s) matching H(s) well at high frequencies
but instead may wish to minimize the maximum mismatch between H(s) and Hr (s)
over the imaginary axis (best H approximation). Flexibility in choosing the Dr
term is necessary in this case as well. How to incorporate arbitrary choices of Dr
without losing interpolation properties, is described in the next theorem, first presented in [66] and later generalized in [16]. For simplicity, we assume D = 0, i.e.
H(s) = C(sE A)1 B.
The general case with D = 0 is recovered by replacing Dr with Dr D.
(15)
14
Theorem 1.3. Given H(s) as in (15), 2r distinct points, {i }ri=1 { j }rj=1 , in the
right halfplane, together with 2r nontrivial vectors, {ci }ri=1 C p and {b j }rj=1
and
Cm , let Vr Cnr and Wr Cnr be as in (13) and (14), respectively. Define B
as
C
= [b1 , b2 , ..., br ] and C
T = [c1 , c2 , . . . , cr ]T
B
For any Dr C pm , define
Er = WTr EVr ,
T Dr B,

Ar = WTr AVr + C
T Dr ,
Br = WTr B C

and Cr = CVr Dr B
(16)
Then the reduced-order model Hr (s) = Cr (sEr Ar )1 Br + Dr satisfies

H(i )bi = Hr (i )bi
and cTi H(i ) = cTi Hr (i )
for i = 1, ..., r.
Theorem 1.3 describes how to construct an interpolant with an arbitrary Dr term;

hence the reduced-order model Hr solves the rational interpolation problem for any
Dr that can be chosen to satisfy specific design goals.
It may happen that the state space representation of the reduced system is not
minimal (for an example see [31]), in which case interpolation of some of the data
is lost. Theorem 1.3 provides a remedy: there always exists Dr , such that the realization given by (16) is minimal, i.e., both controllable and observable, and hence
interpolation of all the data is guaranteed.
2.4 Error Measures

The loss in accuracy in replacing a full-order model with a reduced model is measured by the difference between respective system outputs over a range of inputs,
which is best understood in the frequency domain. For the full-order system described by the transfer function H(s) as in (5) and the reduced-order system by
Hr (s) as in (6), the error system, in the time domain, is defined by Serror : u(t)
[y yr ](t). Equivalently, we can represent the error in the frequency domain by

y(s)
yr (s) = [H(s) Hr (s)]
u(s); hence the error system can be considered having the transfer function H(s) Hr (s).
We consider the reduced-order system to be a good approximation of the fullorder system if the error system Serror is small in a sense we wish to make precise.
Ultimately, we want to produce a reduced order system that will yield outputs, yr (t),
such that yr (t) y(t) uniformly well over a large class of inputs u(t). Different
measures of approximation and different choices of input classes lead to different
model reduction goals. Indeed, the size of linear systems can be measured in
various ways the two most common metrics are the so-called H and H2 norms.
15
The H Norm The H norm of a stable linear system associated with a transfer
function, H(s), is defined as
HH = max H( )2 ,
R
where R2 here denotes the usual induced 2-norm of a matrix R. If the matrix
E is singular, we must require in addition that 0 is a nondefective eigenvalue
of E so that H(s) is bounded as s . Suppose one wants to ensure that the
output error y(t) yr (t) is small in a root mean square sense for t > 0 (that

1/2
is, we want ( 0 y(t) yr (t)
to be small) uniformly over all inputs, u(t),
2 dt)
having bounded energy, 0 u(t)22 dt 1. Observe first that
y(s)
yr (s) =
[H(s) Hr (s)]
u(s) so then by the Parseval relation:

0
1

y( )
yr ( )22 d
2

1
H( ) Hr ( )22
u( )22 d
2

1/2
1
2
2
max H( ) Hr ( )2

u( )2 d
y(t) yr (t)22 dt =
max H( ) Hr ( )22
=
def
H Hr2H
Hence, proving that the HH norm is the L2 induced operator norm of the associated system mapping, S : u y.
The H2 Norm. The H2 norm of a linear system associated with a transfer function,
H(s), is defined as

1/2
1
2
HH2 :=
H( )F
(17)
2
where now R2F = trace(R RT ) denotes the Frobenius norm of a complex matrix
R. Note that we must require in addition that if E is singular then 0 is a nondefective
eigenvalue of E and that lims H(s) = 0 in order that the H2 norm of the system
to be finite. For more details, see [2] and [94].
Alternatively, suppose one wants to ensure that each component of the output
error y(t) yr (t) remains small for all t > 0 (that is, wewant maxt>0 y(t) yr(t)
to be small) again uniformly over all inputs, u(t) with 0 u(t)22 dt 1.

1

t

(
y( )
yr ( )) e d
max y(t) yr (t) = max

t>0
t>0 2

1

y( )
yr ( ) d
2

1
H( ) Hr ( )F
u( )2 d
2
16

1/2

1

u( )22 d
2
+
1/2
def
2
H( ) Hr ( )F d
= H Hr H2
H( ) Hr ( )2F d
1/2
It should be noted that in the single-input single-output case, the above relation
holds with equality sign, because the H2 norm is equal to the (2, ) induced norm
of the convolution operator (see [2] for details).
H2 denotes the set of matrix-valued functions, G(z), with components that are
analytic for z in the open right half plane, Re(z) > 0, such that for Re(z) = x > 0,
G(x + y) is square integrable as a function of y (, ):

sup
x>0
G(x + y)2F dy < .
H2 is a Hilbert space and holds our interest because transfer functions associated
with stable finite dimensional dynamical systems are elements of H2 . Indeed, if
G(s) and H(s) are transfer functions associated with stable dynamical systems having the same input and output dimensions, the H2 -inner product can be defined as
G, HH2 =
def
1
2

Tr G( )H( )T d =

Tr G( )H( )T d ,
(18)
with a norm defined as

GH2 =
def
1
2
+
G( )2F d
1/2
.
(19)
where Tr(M) and MF denote the trace and Frobenius norm of M, respectively.
Notice in particular that if G(s) and H(s) represent real dynamical systems then
G, HH2 = H, GH2 so that G, HH2 must be real.
Suppose f (s) is a meromorphic function (analytic everywhere except at isolated poles of finite order). is a simple pole of f (s) if lims (s ) f (s) = 0
for 2 and the residue is nontrivial: res[ f (s), ] = lims (s ) f (s) = 0. For
matrix-valued meromorphic functions, F(s), we say that is a simple pole of F(s)
if lims (s ) F(s) = 0 for 2 and res[F(s), ] = lims (s ) f (s) has
rank 1. is a semi-simple pole of F(s) if lims (s ) F(s) = 0 for 2 and
res[F(s), ] = lims (s )F(s) has rank larger than 1.
If F(s)2 remains bounded as s then F(s) only has a finite number of
poles. If all these poles are either simple or semi-simple then we define the order
or dimension of F(s) by dim F = rank (res[F(s), ]) where the sum is taken over
all poles . In this case, we can represent F(s) as
F(s) =
dim F
i=1
1
c i bT ,
s i i
where i are indexed according to multiplicity as indicated by rank(res[F(s), i ]).
17
Lemma 1.1. Suppose that G(s) and H(s) are stable (poles contained in the open
left halfplane) and suppose that H(s) has poles at 1 , 2 , . . . m . Then
G, HH2 =

G(s)H(s)T , k ].
res[Tr
(20)
k=1
In particular, if H(s) has only simple or semi-simple poles at 1 , 2 , . . . m and

1
T
m = dim H then H(s) = m
i=1 si ci bi and
m
cTk G(k )bk
G, HH2 =
k=1
and
1/2
HH2 =
cTk H( k )bk
k=1

Proof. Notice that the function Tr G(s)H(s)T has singularities in the left half
plane only at 1 , 2 , . . . m . For any R > 0, define the semicircular contour in the
left halfplane:

3

R = {z |z = with [R, R] } z z = R e with [ , ] .
2 2
R bounds a region that for sufficiently large R contains all the system poles of H(s)
and so, by the residue theorem
G, HH2 =
1
2

Tr G( )H( )T d
1
R 2
= lim

R

Tr G(s)H(s)T ds

= res[Tr G(s)H(s)T , k ].
m
k=1
The remaining assertions follow from the definition.

Lemma 1.1 immediately yields a new expression, recently introduced by Beattie
and Gugercin in [17], for the H2 error norm for MIMO dynamical systems based
on the poles and residues of the transfer function. A similar expression for SISO
systems was first introduced by Krajewski et al. [58], and rediscovered later by
Gugercin and Antoulas [2, 44, 46]:
Theorem 1.4. Given a full-order real system, H(s) and a reduced model, Hr (s),
having the form Hr (s) = ri=1 1 ci bi T (Hr has simple poles at 1 , 2 , . . . r and
si
rank-1 residues c1 bT1 , . . . , cr bTr .), the H2 norm of the error system is given by
r
H Hr 2H2 = H2H2 2 cTk H( k )bk +

k=1
cTk c bT bk
k,=1 k
r
(21)
18
3 Interpolatory Optimal H2 Approximation

Given a full-order system H(s), the optimal H2 model reduction problem seeks to
find a reduced-order model Hr (s) that minimizes the H2 error; i.e.

r
H Hr H2 =
min
(22)
H H
.
r) = r
dim(H
r : stable
H
H2
This optimization problem is nonconvex obtaining a global minimizer can be

intractable and is at best a hard task. The common approach is to find reduced
order models that satisfy first-order necessary optimality conditions. Many researchers have worked on this problem. These efforts can be grouped into two categories: Lyapunov-based optimal H2 methods such as [53, 54, 81, 88, 89, 95]; and
interpolation-based optimal H2 methods such as [15,17,25,45,48,49,59,68,83]. The
main drawback of most of the Lyapunov-based methods is that they require solving a series of Lyapunov equations and so rapidly become infeasible as dimension
increases. We note that even though recent developments have made approximate
solution of large order Lyapunov equations possible, solving a sequence of them
through an iteration remains daunting and the effect of only having approximate
solutions throughout the iteration has not yet been assessed. On the other hand, interpolatory approaches can take advantage of the sparsity structure of E and A and
have proved numerically very effective. Moreover, both problem frameworks are
equivalent, as shown by Gugercin et al. [49], and this has further motivated expansion of interpolatory approaches to optimal H2 approximation.
Interpolation-based H2 -optimality conditions were originally developed by Meier
and Luenberger [68] for SISO systems. Based on these conditions, an effective numerical algorithm for interpolatory optimal H2 approximation, called the Iterative
Rational Krylov Algorithm (IRKA), was introduced in [45, 48]. Analogous H2 optimality conditions for MIMO systems within a tangential interpolation framework have recently been developed by [25, 49, 83] leading to an analogous algorithm for the MIMO case; descriptions can be found in [25, 49]. Below, we present
a new proof of the interpolatory first-order conditions for the optimal H2 model reduction problem. These conditions are the starting point for the numerical algorithm
presented in the next section.
Theorem 1.5. Suppose H(s) is a real stable dynamical system and that Hr (s) =
ri=1 s1 ci bTi is a real dynamical system that is the best stable rth order approxii
mation of H with respect to the H2 norm. (Hr has simple poles at 1 , 2 , . . . r and
rank-1 residues c1 bT1 , . . . , cr bTr .) Then
(a) H( k )bk = Hr ( k )bk ,
and
(b) cTk H( k ) = cTk Hr ( k ),
(c) cTk H ( k )bk = cTk Hr ( k )bk
for k = 1, 2, ..., r.
(23)
19
r (s) is a transfer function associated with a stable rth order dyProof. Suppose H
namical system. Then
r 2H = H Hr + Hr H
r 2H
H Hr 2H2 H H
2
2
r H + Hr H
r 2H
= H Hr2H2 + 2 Re H Hr , Hr H
2
2
so that
r H + Hr H
r 2H
0 2 Re H Hr , Hr H
2
2
(24)
r is made to differ from Hr , we arrive at

By making judicious choices in how H
tangential interpolation conditions via Lemma 1.1.
m
Toward this end, pick
an arbitrary unit vector C , > 0, and for some ,
T
define = arg H( ) Hr ( ) b , and so that
r (s) = e bT
Hr (s) H
s

r H = | T H( ) Hr ( ) b |.
and using Lemma 1.1, H Hr , Hr H
2
Now (24) leads to

0 | T H( ) Hr ( ) b |
b 22
2Re( )
which by taking small implies first that

T H( ) Hr ( ) b = 0
but then since was chosen arbitrarily, we must have that

H( ) Hr ( ) b = 0.
A similar argument yields (23b).
If (23c) did not hold then cT H ( )b = cT Hr ( )b and we may pick 0 <
< |Re( )| and = argcT H ( ) Hr ( ) b in such a way that =
+ e does not coincide with any reduced order poles 1 , 2 , . . . , r . Note that
r (s) so that
Re( ) < 0. Define H

1
1
r (s) =
Hr (s) H
c bT .
s s
r (s) has the same poles and residues as Hr (s) aside from which replaces as
H
r (s) with no change in the associated residue. Because of (23a-b), we
a pole in H
calculate
20
r H = cT
H Hr , Hr H
2

H( ) Hr ( ) b cT (H( ) Hr( )) b
= cT (H( ) Hr ( )) b
Then (24) leads to
| |2 Re( + )
c 22 b 22
2
| + | 2Re( )Re( )
(25)
Now, easy manipulations yield first a resolvent identity
0 2Re(cT (H( ) Hr ( )) b )
( E A)1 =( E A)1 + ( )( E A)1 E( E A)1

and then resubstituting,
( E A)1 =( E A)1 + ( )( E A)1 E( E A)1
+( )2 ( E A)1 E( E A)1 E( E A)1
(26)
Premultiplying by C and postmultiplying by B, (26) implies

H( ) = H( ) + ( )H ( )
+ ( )2 C( E A)1E( E A)1 E( E A)1 B
and analogous arguments yield
Hr ( ) = Hr ( ) + ( )Hr ( )
+( )2 Cr ( Er Ar )1 Er ( Er Ar )1 Er ( Er Ar )1 Br
Using these expressions in (25) yields

0 2 |cT H ( ) Hr ( ) b | + O( 2 ).

As 0 a contradiction occurs unless |cT H ( ) Hr ( ) b | = 0.
One byproduct of the proof of Theorem 1.5 is that if Hr (s) satisfies the necessary
conditions (23), then Hr (s) is guaranteed to be an optimal approximation of H(s)
relative to the H2 norm among all reduced order systems having the same reduced
system poles { i }ri=1 . This follows from the observation that the set of all systems
having the all same poles { i }ri=1 comprise a subspace of H2 and the projection theorem is both necessary and sufficient for Hr (s) to be a minimizer out of a subspace
of candidate minimizers.
Conversely, the set of all stable rth order dynamical systems is not only not a
subspace but it is not even convex. Indeed, the original problem (22) may have
21
multiple minimizers and there may be local minimizers that do not solve (22) yet
do satisfy the first order conditions (23). To clarify, a reduced order system Hr , is
said to be a local minimizer for (22), if for all > 0 sufficiently small,
( )
r H ,
H Hr H2 H H
2
(27)
r( ) such that both dim(H

r( ) ) = r and Hr
for all stable dynamical systems, H
r( ) H C for some constant C.
H
2
As a practical matter, global minimizers that solve (22) are difficult to obtain reliably and with certainty; current approaches favor seeking reduced order models that
satisfy local (first-order) necessary conditions for optimality and so may produce
models that are only certain to be local minimizers. Even though such strategies do
not guarantee global minimization, they often produce very effective reduced order
models, nonetheless.
3.1 An Algorithm for Interpolatory Optimal H2 Model Reduction

Theorem 1.5 states that first-order conditions for H2 optimality are tangential interpolation conditions at mirror images of reduced-order poles. However, one cannot simply construct the corresponding matrices Vr and Wr since the interpolation
points and the tangential directions depend on the reduced-model; hence are not
known a priori. The Iterative Rational Krylov Algorithm (IRKA) introduced in [49]
resolves this problem by iteratively correcting the interpolation points and the directions as outlined below.
Let Y and X denote, respectively, the left and right eigenvectors for Er Ar
so that Y Ar X = diag( i ) and Y Er X = Ir . Denote the columns of Cr X as c i and
the rows of Y Br as b Ti . Then in IRKA interpolation points used in the next step are
chosen to be the mirror images of the current reduced order poles, i.e., the eigenvalues, (Ar , Er ), of the pencil Er Ar in the current step. The tangent directions
are corrected in a similar way, using the residues of the previous reduced-ordered
model until (23) is satisfied. A brief sketch of IRKA is described in Algorithm 1:
Upon convergence, {i } will be the mirror images of the eigenvalues of the reduced pencil Er Ar ; and b Ti = eTi Y Br and c i = Cr Xei for i = 1, . . . , r. Consequently from Steps 4.(d) and 4.(e), the reduced-order transfer function satisfies
(23), first-order conditions for H2 optimality. The main computational cost of this
algorithm involves solving 2r linear systems to generate Vr and Wr . Computing the
eigenvectors Y and X, and the eigenvalues of the reduced pencil Er Ar are cheap
since the dimension r is small.
IRKA has been remarkably successful in producing high fidelity reduced-order
approximations; it is numerically effective and has been successfully applied to finding H2 -optimal reduced models for systems of high order, n > 160, 000, see [56].
22

MIMO H2 Optimal Tangential Interpolation Method
1. Make an initial r-fold shift selection: {1 , . . ., r } that is closed under conjugation
(i.e., {1 , . . ., r } {1 , . . ., r } viewed as sets)
and initial
tangent directions b 1 , . . ., b r and c 1 , . . ., c r , also closed under conjugation.

2. Vr = (1 E A)1 Bb 1 (r E A)1 Bb r

Wr = (1 E AT )1 CT c 1 (r E AT )1 CT c 1 .
3. while (not converged)
a. Ar = WTr AVr , Er = WTr EVr , Br = WTr B, and Cr = CVr
b. Compute Y Ar X = diag( i ) and Y Er X = Ir where Y and X are
the left and right eigenvectors of Er Ar .
c. i i (Ar , Er ) for i = 1, . . ., r, b i eTi Y

Br and c i Cr Xei .
d. Vr = (1 E A)1 Bb 1 (r E A)1 Bb r

e. Wr = (1 E AT )1 CT c 1 (r E AT )1 CT c 1 .
4. Ar = WTr AVr , Er = WTr EVr , Br = WTr B, Cr = CVr
Remark 1.1. Note that the H2 error formulae in Theorem 1.4 expresses the error
as a function of reduced-order system poles and residues. Based on this expression,
Beattie and Gugercin [17] recently developed a trust-region based descent algorithm
where the poles and residues of Hr (s) are the optimization parameters as opposed
to the interpolation points and tangential directions in IRKA. The algorithm reduces
the H2 error at each step of the iteration and globally converges to a local minimum.
Even though the method does not use the interpolation point as the parameters, it
is worth noting that the resulting reduced-order model indeed is an interpolatory
approximation satisfying the interpolatory H2 conditions of Theorem 1.5. Hence,
the method of [17] achieves the interpolation conditions while using the reducedorder poles and residues as the variables.
3.2 Numerical Results for IRKA

This problem arises during a cooling process in a rolling mill and is modeled as
boundary control of a two dimensional heat equation. A finite element discretization
results in a descriptor system state-dimension n = 79, 841, i.e., A, E R7984179841,
B R798417, C R679841 . For details regarding this model, see [18, 20].
Using IRKA, we reduce the order of the full-order system, H(s), to r = 20 to obtain the H2 optimal reduced model, HIRKA (s). Figure 1 illustrates the convergence
H HIRKAH
,
behavior of IRKA through the history of the relative H error,
HH
as it evolves through the course of the iteration. Even though the figure shows the
method converging after approximately 11-12 iterations, one may also see that it
nearly achieves the final error value much earlier in the iteration; already around the
sixth or seventh step.
23
Fig. 1 Evolution of IRKA
Starting from a large initial relative error (nearly 100%), the method automatically and without any user intervention corrects both interpolation points and
tangent directions, reaching within a few iterations a near-optimal solution
with a relative error of around 3 102.
We compare our approach with other commonly used model reduction techniques.
1. Modal Approximation: We reduce the order r = 20 using 20 dominant modes
of H(s). The reduced model is denoted by Hmodal .
2. Interpolation using points on the imaginary axis: Based on the bode plot of
H(s), we have placed interpolation points on the imaginary axis where H( )
is dominant. This was a common approach for choosing interpolation points before the H2 -optimal shift selection strategy was developed. The reduced model
is denoted by H .
3. 20 real interpolation points are chosen as the mirror images of the poles of H(s).
This selection is a good initialization for IRKA in the SISO case, see [49]. The
reduced model is denoted by Hreal .
In our experiments, we found that in order to be able to find reasonable interpolation points and directions for H and Hreal , we needed first to run several experiments. Many of them either resulted in unstable systems or very poor performance
24

Table 1 Error norms
Relative H error
HIRKA
3.03 102
Hmodal
1.03 101
H
5.42 101
Hreal
2.47 101
Fig. 2 Comparison of reduced-order models. (a) Error system bode plots. (b) Reduced model bode
plots for u3 and y5
results. Here we are presenting the best selection we were able to find. This is the
precise reason why IRKA is superior. We initiate it once, randomly in this case,
and the algorithm automatically finds the optimal points and directions. There is
no need for an ad hoc search. Table 1 shows the relative H error norms for each
reduced model. Clearly, IRKA is the best one, yielding an error one order of magnitude smaller than those the other four approaches yield. Note from Fig. 1 that the
initial guess for IRKA has a higher error than all the other methods. However, even
after only two steps of the iteration long before convergence, the IRKA iterate has
already a smaller error norm than all other three approaches. Note that H(s) has 7
inputs and 6 outputs; hence there are 42 inputoutput channels. HIRKA with order
r = 20, less than the total number of input-output channels, is able to replicate these
behaviors with a relative accuracy of order 102 . Even though IRKA is an H2 -based
approach, superior H performance is also observed and is not surprising. It is an
efficient general purpose H2 and H model reduction method, not only H2 optimal. Figure 2a depicts the bode plots of the error systems. It is clear from the error
plots that HIRKA outperforms the rest of the methods. To give an example how the
reduced models match the full-order model for a specific input/output channel, we
show the bode plots for the transfer function between the third input u3 and the fifth
output y5 . Clearly, HIRKA yields the best match. We note that while Table 1 lists the
relative error norms, the error norms in Fig. 2 shown are the absolute error values.
Since IRKA yielded a much lower error value, we have checked what lowest order
model from IRKA would yield similar error norms as the other approaches. We have
found that IRKA for order r = 2 yields a relative H error of 2.26 101; already
better than 20th order H and Hreal . For r = 6, for example, IRKA yielded a relative
H error of 1.64 101, a number close to the one obtained by 20th order Hmodal .
These numbers vividly illustrate the strength of the optimal H2 shift selection.
25
4 Interpolatory Passivity Preserving Model Reduction

Passive systems are an important class of dynamical systems that can only absorb
power for which the work done by the inputs in producing outputs must always
be positive. A system is called passive if
Re
t
u( ) y( )d 0,
for all t R and for all u L2 (R). In this section, we will show how to create interpolatory projections so that passivity of reduced models is preserved. For further
details we refer to [3].
A rational (square) matrix function H(s) is called positive real if:
1. H(s), is analytic for Re(s) > 0,
2. H(s) = H(s) for all s C, and
3. the Hermitian part of H(s), i.e., H(s) + HT (s), is positive semi-definite for all
Re(s) 0.
Dynamical systems as given in (1) are passive if and only if the associated transfer
function H(s) = C(sE A)1 B + D, is positive real. An important consequence
of positive realness is the existence of a spectral factorization: a matrix function
(s) for which H(s) + HT (s) = (s) T (s), and the poles as well as the (finite)
zeros of (s) are all stable. The spectral zeros of the system represented by H(s)
are defined to be those values, , for which ( ) (and hence H( ) + HT ( ))
loses rank.
Assume for simplicity that is a spectral zero with multiplicity one (so that
nullity( ( )) = 1). Then there is a right spectral zero direction, z, such that
(H( ) + HT ( ))z = 0. Evidently, if ( , z) is a right spectral zero pair for the
system represented by H(s), then ( , z ) is a left spectral zero pair: z (H( ) +
HT ( )) = 0.
The key result that is significant here was shown in [3] that if interpolation points
are chosen as spectral zeros, passivity is preserved.
Theorem 1.6. Suppose the dynamical system given in (1) and represented by the
transfer function H(s) = C(sE A)1 B + D is stable and passive. Suppose that for
some index r 1, 1 , . . . , r are stable spectral zeros of H with corresponding right
spectral zero directions z1 , . . . , zr .
If a reduced order system Hr (s) tangentially interpolates H(s) as in (11) with
i = i , bi = zi , i = i , and cTi = z for i = 1, . . . , r, then Hr (s) is stable and
passive.
As a practical matter, the first task that we face is computation of the spectral
zeros of the system represented by H(s). This can be formulated as a structured
26
eigenvalue problem. Following the development of [80], we define the following

pair of matrices which has Hamiltonian structure:
E 0 0
A
0
B
H = 0 AT CT and E = 0 ET 0 .
0 0 0
C
BT D + DT
The spectral zeros of the system are the generalized eigenvalues of the pencil: Hxi =
i Exi . To see this, partition the eigenvector xi into components vi , wi , and zi such
that
A
0
B
vi
vi
E 0 0
0 AT CT wi = i 0 ET 0 wi ,
0 0 0
C
BT D + DT
zi
zi
Then note that as a consequence,
vi = (i E A)1Bzi ,
and
wTi = zi C(i E A)1,
(H(i )+HT (i ))zi = 0.
Thus, i are spectral zeros of H(s) associated with the right spectral zero directions, zi , for i = 1, . . . , r. Furthermore, basis vectors for the right and left modeling subspaces used to impose the tangential interpolation conditions required by
Theorem 1.6, are determined immediately by the remaining two components of xi :
Vr = [v1 , v2 , . . . , vr ] and Wr = [w1 , w2 , . . . , wr ].
Since H and E are real, the eigenvalues ofHx =
Ex occur in complex con0 I 0
jugate pairs. Furthermore, if we define J = I 0 0 Then (JH) = JH and
0 0 I
(JE) = JE and it is easy to see that if xT = [vT wT zT ] is a (right) eigenvector associated with : Hx = Ex, then y = [wT v z ] is a left eigenvector of the pencil
associated with : y H = y E. Thus, if is an eigenvalue of Hx = Ex then
so is each of , , and .
Thus, the spectral zeros, associated spectral zero directions, and bases for the left
and right modeling subspaces can be obtained by means of the above Hamiltonian
eigenvalue problem.
The question remains of which r spectral zeros to choose. The concept of dominance arising in modal approximation is useful in distinguishing effective choices
of spectral zero sets. For details we refer to [55].
Assume for simplicity that D + DT is invertible, take = (D + DT )1 and define
B
B = CT , C = [C, BT , 0],
0
It can be checked (with some effort) that
G(s) = [H(s) + HT (s)]1 = + C(sE H)1 B.
27
Let the partial fraction expansion of G(s) be

G(s) =
2n
Rj
sj ,
with R j =
j=1
1
C x j yj B,
y j Ex j
where i are the spectral zeros of the original system (poles of the associated Hamiltonian system) and R j are the residues. The left and right eigenvectors of y j , x j , are
computed from Hx j = j Ex j and yj H = j yj E.
A spectral zero i is dominant over another spectral zero j , if
R j 2
Ri 2
>
.
|Re(i )| |Re( j )|
To efficiently compute the r most dominant spectral zeros of a dynamical system
represented by H(s), we use an algorithm introduced in [77], and referred to as
SADPA (Subspace Accelerated Dominant Pole Algorithm).
To summarize, the benefits of the spectral zero method are:
(1) The dominant spectral zeros j can be computed automatically
(2) The resulting eigenvectors are readily available to construct projection matrices,
Vr and Wr , for the right and left modeling subspaces
(3) The dominance criterion is flexible and easily adjustable and
(4) Iterative methods are suitable for large scale applications For a detailed description of the resulting algorithm and all the missing details see [55].
4.1 An Example of Passivity Preserving Model Reduction

We conclude this section with a simple example which illustrates the concepts discussed above. Consider the following RLC circuit:
R1
-
R2
L1
L2
C1
?
C2
?
Using the voltages across C1 , C2 , and the currents through L1 , L2 , as state variables,
xi , = 1, 2, 3, 4, respectively, we end up with equations of the form Ex(t) = Ax(t) +
Bu(t), y(t) = Cx(t) + Du(t), where
28
C1 0 G1 L1
0 C2
E=
0 0
0 0
0
L1
0
G2 L2
0
G2 L2
0
, A =
0
0
1
1 1
L2
1 1
0
0
0
,
0

0
0
B=
1 , C = [G1 , 0, 1, 0], D = G1 ,
0
and Gi = R1i , i = 1, 2, are the corresponding conductances. By construction, the system is passive (for non-negative values of the parameters), and it is easy to see that
its transfer function has a zero at s = 0. Hence the system has a double spectral zero
at s = 0. According to the definition of dominance mentioned above, among all finite
spectral zeros, those on the imaginary axis are dominant. Hence we will compute a
second order reduced system by using the the eigenpairs of (H, E), corresponding
to the double zero eigenvalue. The Hamiltonian pair is:
0
0 1 1 0 0 0 0 0
0
0 0 1 0 0 0 0 0
1 0 0 0 0 0 0 0 1
1 1 0 0 0 0 0 0 0
0 0 0 0 0 1 1 G1
H=
0
,
0
0
0
0
0
0
0
1
0
0
0
0
1
0
0
0
1
0
0 0 0 1 1 0 0 0
G1 0 1 0 0 0 1 0 2G1
and
C1
0
E=
0
0
0
0
0 G1 L1 G2 L2
0
0
C2
0
G2 L2
0
0
0
L1
0
0
0
0
0
L2
0
0
0
0
0
C1
0
0
0
0
0
C2
0
0
0
G1 L1
0
0
0
0
G2 L2 G2 L2
0
0
0
0
0
0
0
0
0
0
0
L1
0
0
0
0
0
0
0
0
0
L2
0
0
0
0
.
0
0
0
It follows that although the algebraic multiplicity of this eigenvalue is two, its geometric multiplicity is only one. Hence we need a Jordan chain of eigenvectors x1 ,
x2 , corresponding to this eigenvalue. In particular, x1 satisfies Hx1 = 0, while x2 ,
satisfies Hx2 = Ex1 . These eigenvectors are:
1
1
[x1 , x2 ] = 1
G1
0
1
0
0
1
29
C2
C1 +C2
0 .
1
C2
C1 +C2
0
Thus the projection is defined by means of Vr = [x1 (1 : 4, :), x2 (1 : 4, :)] and Wr =

[x1 (5 : 8, :), x2 (5 : 8, :)], that is
1
1
Vr =
0
0
0
0
1
C2
C1 +C2
1
1
, Wr =
G1
0
0
1
C2
C1 +C2
Therefore by (10) the reduced quantities are:

0
C1 +C2
G1 1
2
Er = Wr EVr =
,
, Ar = Wr AVr =
2
1 0
0 L1 + (C C+C
2 L2
1
2)

G1
, Cr = CVr = G1 1 , Dr = D.
Br = Wr B =
1
The corresponding transfer function Hr (s) = D + Cr (sEr Ar )1 Br , can be expressed as follows:
H1
r (s) =
1
1
+
s(C1 + C2 ) G +
1
s
where =
L1 L2 (C1 + C2 )2
.
L1C2 2 + L2 (C1 + C2 )2
From this expression or from the state space matrices, we can read-off an RLC
realization. The reduced order circuit contains namely, a capacitor C = C1 + C2 ,
2
2
an inductor L = 1 = L1 + (C C+C
2 L2 , and a resistor of value R = R1 . Thereby the
1
2)
capacitor is in series with a parallel connection of the inductor and the resistor as
shown below.
y
L
C
30
Hence in this particular case, after reduction besides passivity, the structure (topology)
of the circuit using the spectral zero reduction method, is preserved.
5 Structure-Preserving Model Reduction Using Generalized

Coprime Factorizations
The model reduction framework we have described for reducing an original system
of the form (1) to a reduced-order model of similar form (2) is quite general. With
suitable generalizations and reformulations, this framework will encompass many
linear dynamical systems of practical importance. However, there exist many linear
dynamical systems whose natural descriptions take a form rather different than the
standard canonical form (1).
We motivate the by considering a model of the dynamic response of a viscoelastic body. One possible continuum model for forced vibration of an isotropic incompressible viscoelastic solid could be described as
tt w(x,t) w(x,t)
t
0
(t ) w(x, ) d + (x,t) = b(x) u(t),
w(x,t) = 0 determining y(t) = [w(x1 ,t), . . . w(x p ,t)]T ,
(28)
where w(x,t) is the displacement field for the body, (x,t) is the associated pressure field and w = 0 represents the incompressibility constraint (see e.g., [63]).
Here,
( ) is a presumed known relaxation function satisfying ( ) 0 and

) d < ; > 0 is a known constant associated with the initial elastic re0
sponse. The term b(x) u(t) = m
i=1 bi (x) ui (t) is a superposition of the m inputs
u(t) = {u1 (t), . . . , um (t)}T Rm . Displacements at x1 , . . . , x p are the outputs. Semidiscretization of (28) with respect to space produces a large order linear dynamical
system of the form:
M x (t) + K x(t) +
DT x(t) = 0,
t
0
(t ) K x( ) d + D (t) = B u(t),
which determines y(t) = C x(t).
(29)
where x Rn1 is the discretization of the continuous displacement field w;

Rn2 is the discretization of the continuous pressure field . The matrices M and
K are n1 n1 real, symmetric, positive-definite matrices, D is an n1 n2 matrix,
B Rn1 m , and C R pn2 . The state space dimension is n = n1 + n2 , an aggregate
of x(t) and (t).
Applying the Laplace transform to (29) yields
s2 M + (
(s) + ) K

y(s) = [C 0]
T
D
D
0
!1
!
B
(s) = H(s)
u
u(s),
0
(30)
31
defining the transfer function, H(s). The form of the transfer function is clearly
different from the standard form (5). The system described in (29) is a descriptor
system described by differential-algebraic equations with a hereditary damping; and
a reformulation of (29) into the standard realization as in (1) is generally not possible
(unless (s) has a simple form) and even when possible may not be desirable.
An effective reduced model for (29) should take into account the structure associated with distributed material properties. Therefore, we wish to consider reduced
models similar to (29):
Mr x r (t) + Kr xr (t) +
DTr xr (t) = 0
t
0
(t ) Kr xr ( ) d + Dr r = Br u(t)
which determines yr (t) = Cr xr (t)
(31)
where Mr and Kr are now r1 r1 real, symmetric, positive-definite matrices, Dr

is an r1 r2 matrix, Br Rr1 m and Cr R pr2 with r1 n1 and r2 n2 and
the reduced state space dimension is r = r1 + r2 . Note that the reduced-model (31)
has the same structure with the same hereditary damping as the original model.
Systems with such structure may be expected to produce dynamic responses that
are physically plausible and consistent with behaviors of viscoelastic bodies.
We proceed to construct reduced-order matrices in a similar way as we did in the
standard case: We define matrices of trial vectors Ur Rn1 r1 and Zr Rn2 r2 ; use
the ansatz x(t) Ur xr (t) and (t) Zr r (t); and force a Petrov-Galerkin orthogonality condition on the reduced state-space trajectory residuals to obtain finally
reduced coefficient matrices for (31) as
Mr = UTr MUr ,
Br =
UTr B,
Kr =UTr KUr ,
Dr = UTr DZr ,
(32)
and Cr = CUr .
Note that this construction prevents mixing displacement state variables and pressure state variables. Also notice that both symmetry and positive-definiteness of Mr
and Kr are preserved automatically as long as Ur has full column rank. The transfer
function Hr (s) for the reduced-model can be obtained similarly as in (30). Now, in
addition to preserving the structure we want to choose Ur and Zr so that the reduced
model Hr (s) interpolates H(s) as in the generic case. The ideas of Theorem 1.2 of
2.3 must be extended to this more general setting.
To proceed, we follow the discussion of Beattie and Gugercin [16] and consider
MIMO systems with the following state space description in the Laplace transform
domain (which includes the form of (30)):
Find
v(s) such that
K(s)
v(s) = B(s)
u(s)
v(s) + D
u(s)
then
y(s) = C(s)
def
yielding a transfer function: H(s) = C(s)K(s)1 B(s) + D,
(33)
32
where D R pm , both C(s) C pn and B(s) Cnm are analytic in the right half
plane; and K(s) Cnn is analytic and full rank throughout the right half plane. The
goal is to find a reduced transfer function of the same form using a Petrov-Galerkin
projection:
(34)
Hr (s) = Cr (s)Kr (s)1 Br (s) + Dr
where Wr , Vr Cnr , Cr (s) = C(s)Vr C pr ; Br (s) = WTr B(s) Crm , and
Kr (s) = WTr K(s)Vr Cnn .
For simplicity we only consider the case D = Dr , which is equivalent with respect
to the interpolation conditions to D = Dr = 0. The general case with D = Dr (similar
to Theorem 1.3) can be found in the original source [16]. Following the notation
in [16], we write D f to denote the th derivative of the univariate function f (s)
evaluated at s = with the usual convention for = 0, D0 f = f ( ).
Theorem 1.7. Suppose that B(s), C(s), and K(s) are analytic at C and C.
Also let K( ), K( ), Kr ( ) = WTr K( )Vr , and Kr ( ) = WTr K( )Vr have full
rank. Let nonnegative integers M and N be given as well as nontrivial vectors,
b Rm and c R p .
(a) If Di [K(s)1 B(s)]b Ran(Vr ) for i = 0, . . . , N
()
then H() ( )b = Hr ( )b for = 0, . . . , N.

T
(b) If cT D j [C(s)K(s)1 ] Ran(Wr ) for j = 0, . . . , M
()
then cT H() ( ) = cT Hr ( ) for = 0, . . . , M.

(c) If both (a) and (b) hold and if = ,
()
then cT H() ( )b = cT Hr ( )b for = 0, . . . , M + N + 1.
Using this result, one can easily construct a reduced-model satisfying the desired interpolation conditions. For example, for H(s) = C(s)K(s)1 B(s), let r interpolation
points {i }ri=1 , the left-directions {ci }ri=1 and the right-directions {bi }ri=1 be given.
Construct

Vr = K(1 )1 B(1 )b1 , . . . , K(r )1 B(r )br
(35)
and
T

Wr = K(1 )T C(1 )T c1 , . . . , K(r )T C(r )T cr .
(36)
The reduced transfer function, Hr (s) = Cr (s)Kr (s)1 Br (s) satisfies bi-tangential interpolation conditions, i.e., H(i )bi = Hr (i )bi , cTi H(i ) = cTi Hr (i ) and cTi H (i )bi
= cTi Hr (i )bi for i = 1, . . . , r.
Theorem 1.7 achieves the desired result; it extends the interpolation theory of
Theorem 1.2 from the canonical first-order setting to a general setting where we have
a general coprime factorization, H(s) = C(s)K(s)1 B(s) + D and guarantees that
the reduced transfer function will have a similarly structured coprime factorization:
Hr (s) = Cr (s)Kr (s)1 Br (s) + Dr .
33
5.1 A Numerical Example: Driven Cavity Flow

We illustrate the concepts of Theorem 1.7 with a model of driven cavity flow in
two dimensions, a slightly modified version of a model in (28). Consider a square
domain = [0, 1]2 , as shown below, representing a volume filled with a viscoelastic material with boundary separated into a movable top edge (lid), 1 , and the
complement, 0 (stationary bottom, left, and right edges). The material in the
cavity is excited through shearing forces on the material caused by transverse displacement of the lid, u(t). We are interested in the displacement response of the
material, w(x,t),
at the center of , i.e. x = (0.5, 0.5).
tt w(x,t) 0 w(x,t) 1 t
w(x,t) = 0
for x
and
t
w(x, )
0
(t )
d + (x,t) = 0 for x
w(x,t) = 0 for x 0
w(x,t) = u(t) for x 1
which determines the output y(t) = w(x,t),
We use a model of a viscoelastic material given by
u(t) -
r x
w(x, t)

Bagley and Torvik in [6] and take as material constants,

0
0 = 1.05 106, 1 = 2.44 105, and = 0.519,
corresponding to experimentally derived values for the polymer butyl B252.
We discretize the continuous model using Taylor-Hood finite elements defined over
a uniform triangular mesh on . The transfer function of the discretized system is
given by H(s) = C(s)K(s)1 B(s) where
K(s) =
s2 M + (s)K
DT
!
D
,
0
!
2
"
C(s) = [C 0], and B(s) = s m + (s)k .
0
C Rn2 corresponds to measuring the horizontal and vertical displacement at x =

(0.5, 0.5), m and k are sums of the columns of the free-free mass and stiffness
matrix associated with x-displacement degrees of freedom on the top lid boundary
and (s) = 0 + 1 s . This produces a frequency dependent input-to-state map,
B(s), since the system input is a boundary displacement as opposed to a boundary
force. Note the nonlinear frequency dependency in the state-to-input map; however
Theorem 1.7 can be still applied to produce an interpolatory reduced-order model of
the same form as shown next: Suppose an interpolation point C and a direction
c C p are given. Since the input is a scalar, there is no need for a right tangential
direction. We consider bitangential Hermite interpolating conditions:
Hr ( ) = H( ), cT Hr ( ) = cT H( ), and cT Hr ( ) = cT H ( ).
(37)
34
Following Theorem 1.7, we solve the following two linear systems of equations:
!
!
!
!
!
!
F( ) D
u1
u2
N( )
F( ) D
CT c
and
=
=
.
DT 0
0
DT 0
z1
z2
0
where F( ) = 2 M + ( )K and N( ) = s2 m + ( )k. Define the matrices Ur =
[u1 , u2 ] and Zr = [z1 , z2 ] . Then the reduced system matrices defined in (32) with the
modification that Br (s) = s2 UTr m + (s)UTr k corresponds to the choice Vr = Wr =
Ur Zr and satisfies the bitangential Hermite interpolation conditions of (37).
We compare three different models:
1. Hfine , using a fine mesh Taylor-Hood FEM discretization with 51,842 displacement degrees of freedom and 6,651 pressure degrees of freedom (mesh size
1
h = 80
). We treat this as the full-order model.
2. Hcoarse , for a coarse mesh discretization with 29,282 displacement degrees of
1
freedom and 3721 pressure degrees of freedom (mesh size h = 60
);
3. H30 , a generalized interpolatory reduced order model as defined in (31)(32)
with r = 30, corresponding to 30 reduced displacement degrees of freedom and
30 reduced pressure degrees of freedom satisfying the bitangential Hermite interpolation conditions at each interpolation point.
The resulting frequency response plots shown in Figure 3:
In this example, this new framework allows interpolatory model reduction of a
descriptor system with a hereditary damping term. Moreover, the reduced model
is also a descriptor system with the same damping structure. The reduced model
Fig. 3 Bode plots of Hfine , Hcoarse and reduced models H20 and H30
35
H30 having a total of 60 degrees of freedom has an input-output response that is

virtually indistinguishable from the high fidelity model Hfine having nearly 58,500
total degrees of freedom. Moreover, H30 outperforms Hcoarse , which corresponds to
29,282 displacement and 3,721 pressure degrees of freedom (more than 500 times
the order of H30 ).
5.2 Second-Order Dynamical Systems

One important variation on the dynamical system described in (33) arises in modeling the forced vibration of an n degreeoffreedom mechanical structure:
Mx(t) + Gx(t) + Kx(t) = Bu(t),
y(t) = Cx(t),
with the transfer function H(s) = C(s2 M + sG + K)1B
(38)
(39)
where M, G, and K Rnn are positive (semi)-definite symmetric matrices describing, respectively, inertial mass distribution, energy dissipation, and elastic strain
energy (stiffness) distribution throughout the structure. The input u(t) Rm is a
time-dependent force applied along degrees-of-freedom specified in B Rnm and
y(t) R p is a vector of output measurements defined through observation matrix
C R p . Second order systems of the form (38) arise naturally in analyzing other
phenomena aside from structural vibration such as electrical circuits and microelectro-mechanical systems, as well; see [7, 8, 2628, 57, 75, 84], and references
therein.
We wish to generate, for some r n, an rth order reduced second-order system
of the same form:
Mr x r (t) + Gr x r (t) + Kr xr (t) = Br u(t),
yr (t) = Cr xr (t),
where Mr , Gr , Kr Rrr are positive (semi)-definite symmetric matrices, B Rrm ,

and C R pr . In order to preserve the symmetry and positive definiteness of M, G
and K in the course of model reduction, one-sided reduction is applied, i.e. one takes
Wr = Vr , resulting in
Mr = VTr MVr , Gr = VTr GVr , Kr = VTr KVr , Br = VTr B, , Cr = CVr .
One standard approach to model reduction of second-order systems involves the
application of standard (first-order) model reduction techniques to a standard firstorder realization of the system. However, this destroys the original second-order system structure and obscures the physical meaning of the reduced-order states. Once
the reduction is performed in the first-order framework, it will not always be possible to convert this back to a corresponding second-order system of the form (38),
see [69]. Even when this is possible, one typically cannot guarantee that structural
properties such as positive definite symmetric reduced mass, damping, and stiffness
36
matrices, will be retained. Keeping the original structure is crucial both to preserve
physical meaning of the states and to retain physically significant properties such as
stability and passivity.
There have been significant efforts in this direction. Building on the earlier work
of [82], Bai and Su [13] introduced second-order Krylov subspaces and showed
how to obtain a reduced-order system directly in a second-order framework while
still satisfying interpolation conditions at selected points for single-input/singleoutput systems. Other second-order structure preserving interpolatory reduction
techniques were introduced by Chahlaoui et al. [26] and Freund [38], though these
approaches require a first-order state-space realization.
The second-order transfer function H(s) in (39) fits perfectly in the framework
of Theorem 1.7. Indeed, due to the simple structures of C(s) = C, B(s) = B and
K(s) = s2 M + sG + K, one obtains a simple three-term recurrence for the rational
tangential interpolation of MIMO second-order systems directly in a second-order
framework. This is presented in Algorithm 4.1.
From Theorem 1.7, the resulting reduced second-order model tangentially interpolates the original model up through derivatives of order Ji 1 at the selected
(complex) frequencies i while preserving the original second-order structure. The
second-order recursion of Bai and Su [13] is a special case of this algorithm for a
single-input/single-output system using a single interpolation point .
Algorithm 4.1. Second-order Tangential MIMO Order Reduction
Given interpolation points {1 , 2 , . . . , N }, directions {b1 , b2 , . . . , bN }, and
interpolation orders {J1 , J2 , . . . , JN } (with r = Ni=1 Ji ).
1. For i = 1, ..., N
a. For each shift i and tangent direction bi , define
(i)
(i)
K0 = i2 M + i G + K, K1 = 2i M + G, and K2 = M.
(i)
b. Solve K0 fi1 = Bbi and set fi0 = 0.

c. For j = 2 : Ji
(i)
(i)
Solve K0 fi j = K1 fi, j1 K2 fi, j2 ,
2. Take Vr = [f11 , f12 , . . . , f1J1 , f21 , . . . , f2J2 , . . . , fN1 , . . . , fNJN ] and
then Mr = VTr MVr , Gr = VTr GVr , Kr = VTr KVr ,
Br = VTr B, and Cr = CVr .
Remark 1.2. The general framework presented here can handle second-order systems of the form
+ B0 u(t),
Mx(t) + Gx(t) + Kx(t) = B1 u(t)
y(t) = C1 x (t) + C0x(t),
(40)
where the input u(t) Rm is a time-dependent force or displacement, the output

y(t) R p depends potentially not only on displacements but also the velocities.
We note that second-order systems of the form (38) cannot be converted to generic
first-order form (1) in a straightforward way when B1 = 0.
37
Remark 1.3. The discussion above and the algorithm can be easily generalized to
higher order constant coefficient ordinary differential equations as well where the
system dynamics follow
A0
dx
d 1 x
+
A
+ + A x(t) = Bu(t) and y(t) = Cx(t).
1
dt
dt 1
(41)
6 Model Reduction of Parametric Systems

Frequently dynamical systems under study are varied by adjusting a comparatively
small number of parameters repeatedly during the course of a set of simulations. Parameters can enter a model representing material properties, variations in shape of
the structure or more generally of the solution domain, and strength of coupling
in various types of boundary conditions. Micro-electromechanical systems, chip
design and interconnect modeling are two of the most prominent examples where
parameterized systems naturally arise; see, e.g., [29, 36, 51, 78]. It will be of value
to have methods that will produce high fidelity reduced order models retaining the
parametric structure of the original system. This leads to the concept of parameterized model reduction. The goal is to construct a high fidelity parametric reduced
order models which recover the response of the original full order parametric system
throughout the range of variation of interest of the design parameters.
Parametric model order reduction is at an early stage of the development. For
interpolatory model reduction of parametric systems, see [14, 19, 22, 29, 32, 34
36, 51, 52, 64, 65, 71, 85, 86], and reference therein. These methods treat parametric dynamical systems in a standard first-order state-space form. Below, we extend
these results to the generalized setting similar to that of Sect. 5.
We consider a multi-input/multi-output linear dynamical system that is parameterized with q parameters p = [p1 , . . . , pq ]:
H(s, p) = C(s, p)K(s, p)1 b(s, p)
(42)
with K(s, p) Cnn and B(s, p) Cnm and C(s, p) C pn . We assume that
K(s, p) = K[0] (s) + a1 (p) K[1] (s) + . . . + a (p) K[ ] (s)
B(s, p) = B[0] (s) + b1 (p) B[1] (s) + . . . + b (p) B[ ] (s),
(43)
C(s, p) = C[0] (s) + c1 (p) C[1] (s) + . . . + c (p) C[ ] (s).

where a1 (p), a2 (p) . . . , b1 (p), ..., c (p) are scalar-valued parameter functions that
could be linear or non-linear. Equation (42) represents a system structure that we
may wish to retain that is, our goal is to to generate, for some r n, a reducedorder system with dimension r having the same parametric structure. Suppose matrices Vr Cnr and Wr Cnr are specified and consider:
Hr (s, p) = Cr (s, p)Kr (s, p)1 Br (s, p)
(44)
38
with Kr (s, p) = WTr K(s, p)Vr , Br (s, p) = WTr B(s, p), and Cr (s, p) = C(s, p)Vr . We
say that the reduced model has the same parametric structure in the sense that

K r (s, p) = WTr K [0] (s)Vr + a1 (p) WTr K [1] (s)Vr + . . . + a (p) WTr K[ ] (s)Vr

(45)
B r (s, p) = WTr B [0] (s) + b1 (p) WTr B [1] (s) + . . . + b (p) WTr B [ ] (s) ,

C r (s, p) = C[0] (s)Vr + c1 (p) C[1] (s)Vr + . . . + c (p) C[ ] (s)Vr .
with exactly the same parameter functions a1 (p), . . . , c (p) as in (43), but with
smaller coefficient matrices. Significantly, all reduced coefficient matrices can be
precomputed before the reduced model is put into service. The next result extends
Theorem 1.7 to the parameterized dynamical system setting:
Theorem 1.8. Suppose K(s, p), B(s, p), and C(s, p) are analytic with respect to s
at C and C, and are continuously differentiable with respect to p in a
neighborhood of p = [p1 , ..., p q ]. Suppose further that both K( , p ) and K( , p )
are nonsingular and matrices Vr Cnr and Wr Cnr are given such that both
Kr ( , p ) = WTr K( , p )Vr and Kr ( , p ) = WTr K( , p )Vr are also nonsingular. For
nontrivial tangential directions b Cm and c C p :
(a) If K( , p ))1 B( , p ))b Ran(Vr ) then H( , p ))b = Hr ( , p ))b

T
(b) If cT C( , p ))K( , p ))1 Ran(Wr ) then cT H( , p )) = cT Hr ( , p ))
(c) If both (a) and (b) hold and if = , then
p cT H( , p )b = p cT Hr ( , p )b
and cT H ( , p )b = cT Hr ( , p )b
Proof. We prove the first two assertions only. The proof of the third assertion is
rather technical. We refer the reader to [14] for the proof of this fact for the special
case of K(s, p) = sE(p) A(p).
Define the projections
Pr (s, p) = Vr Kr (s, p)1 WTr K(s, p) and
Qr (s, p) = K(s, p)Vr Kr (s, p)1 WTr
Also, define the vector f0 = K( , p )1 B(s, p )b. Note that the assumption of the first
assertion implies that f0 Ran(Pr ( , p )). Then, a direct computation yields
H( , p )b Hr ( , p )b = C(s, p ) (I Pr ) f0 = 0,
which proves the first assertion. Similarly, define gT0 = cT C(s, p )K( , p )1 and observe that g0 Ker(Qr ( , p )) due to the assumption of the second assumption.
Then, one can directly obtain
cT H( , p ) cT Hr ( , p ) = gT0 (I Qr ) B(s, p ) = 0,
which proves the second assertion.

39
Theorem 1.8 extends the interpolation theory for the non-parametric systems to
the parameterized ones. It can be easily extended to match higher order derivatives
as well. This result is much more general than others in the literature in the sense
that we allow transfer function to be in the generalized form; hence not constraining
it to be in standard state-space form. Also, the result allows for arbitrary linear and
nonlinear dependency in K, C, and B, which themselves can reflect the greater
generality of the structured dynamical systems setting considered earlier.
Given H(s, p) = C(s, p)K(s, p)1 B(s, p) as defined in (43), let us assume that
we want to construct a parametric reduced-model that matches H(s) at frequency
points {i }Ki=1 C and {i }Ki=1 C, the parameter points {p( j) }Lj=1 Cq along
m
the right tangential directions {bi j }K,L
i=1, j=1 C and the right tangential directions
p
{ci j }K,L
i=1, j=1 C . Define, for i = 1, . . . , K and j = 1, . . . , L,
vi j = K(i , p( j) )1 B(i , p( j) )bi, j and wi j = K(i , p( j) )T C(i , p( j) )T ci, j

Then, construct Vr and Wr such that
Vr = [v11 , . . . , v1L , v21 , . . . , v2L , . . . , vK1 , . . . , vKL ] Cn(KL)
and
Wr = [w11 , . . . , w1L , w21 , . . . , w2L , . . . , wK1 , . . . , wKL ] Cn(KL) .
Then, the resulting projection-based reduced-order parametric model of (45) satisfies the interpolation condition stated in Theorem 1.8.
As the discussion above illustrates, once the parameter points {p( j)} are selected,
the machinery for interpolatory model reduction of parametric systems is very similar to that of non-parametric systems. However, in this case one faces the problem
choosing not only the frequency points {i } but also the parameter points {p( j) }.
Hence, the question of effective parameter point selection arises. In some application, the designer specifies important parameter sets to be used in the model reduction. However, the task becomes much harder if one only has the information as the
range of parameter space without any knowledge of what parameter sets might be
important. In this case, Bui-Thanh et al. [23] proposed the so-called greedy selection
algorithm as a possible solution. Even though the method [23] proves to yield high
quality approximations, the optimization algorithm that needs to be solved at each
step could be computationally expensive. Another strategy for an efficient choice of
parameters points in a higher dimensional parameter sparse is the use of sparse grid
points [24, 41, 92]. An optimal parameter selection strategy was recently proposed
by Baur et al. [14] for the special case of the parametric single-input/single-output
dynamical systems of the form H(s) = (c0 + p1 c1 )T (sI A)1 (b0 + p2 b1 ) where
the scalars p1 and p2 are independent parameters and the system dynamic A is nonparametric. The optimal parameter set is obtained by converting the problem into an
equivalent non-parametric optimal H2 MIMO approximation.
40
6.1 Numerical Example

We illustrate the concept of parametric model reduction using a model representing
thermal conduction in a semiconductor chip as described in [61]. An efficient model
of thermal conduction should allow flexibility in specifying boundary conditions so
that designers are able to evaluate the effects of environmental changes on the temperature distribution in the chip. The problem is modeled as homogenous heat diffusion with heat exchange occurring at three device interfaces modeled with convection boundary conditions that introduce film coefficients, p1 , p2 , and p3 , describing
the heat exchange on the three device interfaces. Discretization of the underlying
partial differential equations leads to a system of ordinary differential equations
3
Ex(t) = (A + pi Ai )x(t) + Bu(t), y(t) = Cx(t),

i=1
where E R42574257 and A R42574257 are system matrices, Ai R42574257,

i = 1, . . . , 3, are diagonal matrices arising from the discretization of the convection
boundary condition on the ith interface, and B R4257 and C R74257 . We note
even though Theorem 1.8 allows parametric model reduction in a much more general setting, we use a standard state-space model to illustrate the theory.
Each parameter value varies in the range of [1, 104 ]. Important parameter set
values are listed in [61] of which we use the following two to apply model reduction:
p(1) = (104 , 104 , 1) and p(2) = (1, 1, 1)
We follow an H2 -inspired approach recently proposed in [14]: Note that Once the
parameter set p(i) for i = 1, 2 are plugged into the state-space matrices, a nonparametric standard transfer function is obtained, denoted by H(s, p(i) ). Then, we
apply the optimal H2 model reduction technique IRKA as outlined in Algorithm 1
to H(s, p(i) ) to reduce the order to ri leading to projection subspaces V(i) R4257ri
and W(i) R4257ri for i = 1, 2 with r1 = 6 and r2 = 5. We, then, concatenate these
matrices to build the final projection matrices
Vr = [V(1) , V(2) ] R425711 and Wr = [W(1) , W(2) ] R425711.
We obtain a final parameterized reduced-order model of orderr = r1 + r2 = 11:
3
WTr EVr x r (t) = (WTr AVr + pi WTr Ai Vr )xr (t) + WTr Bu(t),
i=1
yr (t) = CVr xr (t).

A high-quality parameterized reduced-order model must represent the full parameterized model with high fidelity for a wide range of parameter values. To illustrate
this is the case, we fix p3 = 1 and vary both p1 and p2 over the full parameter
range, i.e., between 1 and 104 . Then, for each mesh point (i.e. for each triple of
parameter values in this range), we compute the corresponding full-order model
41
Fig. 4 Relative H -error as p1 and p2 vary
and the reduced-order model; and compute the corresponding relative H -errors.
Figure 4 depicts the resulting mesh plot. As the figure illustrates, the reduced-model
approximate the full-model uniformly well throughout the full parameter range. The
maximum relative H -error is 3.50 102. Therefore, the parameterized reducedorder model Hr (s, p) of order r = 11 yields a relative accuracy of 102 for the
complete range of p1 and p2 .
7 Model Reduction from Measurements

In many instances, input/output measurements replace an explicit model of a tobe-simulated system. In such cases it is of great interest to be able to efficiently
construct models and reduced models from the available data. In this section we
will address this problem by means of rational interpolation. We will show that the
natural tool is the Loewner matrix pencil, which is closely related to the Hankel
matrix. For details we refer to [62, 66].
7.1 Motivation: S-Parameters

The growth in communications and networking and the demand for high data bandwidth requires streamlining of the simulation of entire complex systems from chips
42
Fig. 5 VNA (Vector Network Analyzer) and VNA screen showing the magnitude of the
S-parameters for a two port
to packages to boards. Moreover, in circuit simulation, signal intergrity (lack of

signal distortion) of high speed electronic components requires that interconnect
models be valid over a wide bandwidth.
An important tool for dealing with such problems is the S- or scattering-parameter
system representation. The S-parameters represent a system as a black box. An important advantage is that these parameters can be measured using VNAs (Vector
Network Analyzers) (see Fig. 5). Also, their norm for passive components, is not
greater than one, and in high frequencies S-parameters are important because wave
phenomena become dominant.
We recall that given a system in input/output representation: y (s) = H(s)u(s),

the associated S-parameter representation is
y s = [H(s) + I][H(s) I]1 u s (s),
#
$%
&
S(s)
are the transmitted waves and, u s = 12 (y u ) are the reflected

where y s = 12 (y + u)
waves. Thus the S-parameter measurements S( jk ), are samples of the frequency
response of the S-parameter system representation.
In the sections that follow we will describe a framework which applies also to
the efficient modeling from S-parameter measurements. An example is provided in
Sect. 7.4.3. For more details and examples see [62].
7.2 The Loewner Matrix Pair and Construction of Interpolants

Suppose we have observed response data as described in Problem 2 in Sect. 2.1.
We are given r (right) driving frequencies: {i }ri=1 C that use input directions
{b i }ri=1 Cm , to produce system responses, {yi }ri=1 C p and q (left) driving fre-
43
quencies: {i }i=1 C that use dual (left) input directions {ci }i=1 C p , to produce
q
dual (left) responses, {zi }i=1 Cm . We assume that there is an underlying dynamical system as defined in (1) so that
q
c Ti H(i ) = z Ti
for i = 1, . . . , q,
H( j )b j = y j ,
for j = 1, . . . , r.
and
yet we are given access only to the r + q response observations listed above and
have no other information about the underlying system H(s). In this section we will
sketch the solution of Problem 2 described in Sect. 2.1. Towards this goal we will
introduce the Loewner matrix pair in the tangential interpolation case. The Loewner
matrix is defined as follows:
T
z T1 b r c T1 y r
z 1 b 1 c T1 y 1

1 r
1
1
.
..
.
.
.
Cqr .
.
L=
.
.
z q b1 c Tq y 1
z Tq b r c Tq y r
q 1
q r
If we define matrices associated with the system observations as:
.. ..
..
.. ..
..
.
.
. .
. .
= b 1 b 2 . . . b r
= y 1 y 2 . . . y r
B
Y
.. ..
..
.. ..
..
. .
.
. .
.
. . . z T1 . . .
. . . z T2 . . .
T =
Z
..
.
. . . z Tq . . .
. . . c T1 . . .
. . . c T2 . . .
T =
C
..
.
. . . c Tq . . .
L satisfies the Sylvester equation

Z
T B.

T Y
L ML = C
(46)
where = diag(1 , 2 , . . . , r ) Crr and M = diag(1 , 2 , . . . , q ) Cqq Suppose that state space data (E, A, B, C, D), of minimal degree n are given such that
H(s) = C(sE A)1 B+D. If the generalized eigenvalues of (A, E) are distinct from
i and j , we define Vr and WTq as

Vr = (1 E A)1Bb 1 , , (r E A)1 Bb r Cnr and
cT1 C(1 E A)1

cT2 C(2 E A)1
WTq =
Cqn .
..
.
cTq C(q E A)1
44
It follows that
L = WTq EVr
and we call Vr and WTq generalized tangential reachability and observability matrices, respectively.
Next we introduce a new object which is pivotal in our approach. This is the
shifted Loewner matrix, defined as follows:
1 z T1 b 1 1 c T1 y 1
1 z T1 b r r c T1 y r
1 1
1 r
.
.
..
..
..
Cqr
.
M=
q z Tq b 1 1 c Tq y 1
q z Tq b r r c Tq y r
q 1
q r
M satisfies the Sylvester equation
T Y
T B.
MZ

M MM = C
(47)
If an interpolant H(s) is associated with the interpolation data, the shifted Loewner
matrix is the Loewner matrix associated to sH(s). If a state space representation
is available, then like for the Loewner matrix, the shifted Loewner matrix can be
factored as
M = WTq AVr .
It therefore becomes apparent that L contains information about E while M contains
information about A. These observations are formalized in one of the main results
of this section which shows how straightforward the solution of the interpolation
problem becomes, in the Loewner matrix framework.
Theorem 1.9. Assume that r = q and that i = j for all i, j = 1, . . . , r. Suppose
that M s L is invertible for all s {i } { j }. Then, with
Er = L,
Ar = M,
T ,
Br = Z

Cr = Y,
Dr = 0,
T (M s L)1 Y

Hr (s) = Cr (sEr Ar )1 Br = Z
interpolates the data and furthermore is a minimal realization.
Next we will outline two proofs of this important result. They are both straightforward and hence reveal the main attributes of this approach.
Proof. Multiplying (46) by s and subtracting it from (47) we get
sI) (M sI)Z

T Y(
T B.
(M sL) M(M sL) = C
(48)
Multiplying (48) by e j (the jth unit vector) on the right and setting s = j , we obtain
T
bj
( j I M)(M j L)e j = ( j I M)Z
T
T b j

j = Y(M

b j Ye
(M i L)e j = Z
j L)1 Z
45
Therefore H( j )b j = y j . This proves the right tangential interpolation property. To

prove the left tangential interpolation property, we multiply (48) by eTi (the transpose
of the ith unit vector) on the left and set s = i :
i I)
eTi (M i L)( i I) = cTi Y(
T
T
T
e Z
T =

T .
cT Y(M
e (M i L) = c Y
i L)1 Z
i
Therefore c Ti H(i ) = z Ti , which completes the proof.

Proof (Alternate version). For this proof we assume that the data have been obtained
by sampling a known high order system in state space form (A, E, B, C). Direct
calculation verifies that
H(i ) H( j ) = ( j i ) C(i E A)1 E( j E A)1 B.
Pre-multiplication of this expression by c Ti and post-multiplication by b j immediately yields
L = WTq EVr .
One may directly calculate first,
j ( j E A)1 = j (i E A)1(i E A)( j E A)1

= i j (i E A)1 E(i E A)1 j (i E A)1 A( j E A)1
and
i (i E A)1 = i (i E A)1 ( j E A)( j E A)1

= i j (i E A)1 E(i E A)1 i (i E A)1 A( j E A)1
j H( j ) = i j C(i E A)1 E( j E A)1 B j C(i E A)1 A( j E A)1 B
i H(i ) = i j C(i E A)1 E( j E A)1 B i C(i E A)1 A( j E A)1 B
which in turn implies that
i H(i ) j H( j ) = ( j i ) C(i E A)1 A( j E A)1B.

Pre-multiplication of this expression by c Ti and post-multiplication by b j immediately yields
M = WTq AVr .
T = WT B and Y
= CVr . Thus Hr (s) = Y(M

T is a
Finally, note that Z
s L)1 Z
q
tangential interpolant to H(s).
Remark 1.4. The Loewner matrix was introduced in [5]. As shown therein, its usefulness derives from the fact that its rank is equal to the McMillan degree of the
46
underlying interpolant. It also turns out that Loewner matrices for confluent interpolation problems (i.e., problems where the values of a certain number of derivatives
are provided as well) have Hankel structure. Thus the Loewner matrix provides a direct link with classical realization theory. For an overview of these results we refer
to Sects. 4.4 and 4.5 in [2]. In [1], one way of constructing state space realizations
based on the Loewner matrix was presented. But the shifted Loewner matrix, first
introduced in [66], was missing and consequently the resulting procedure is only
applicable to proper rational interpolants.
7.2.1 The General Case

If the assumption of the above theorem is not satisfied, one needs to project onto
the column span and onto the row span of a linear combination of the two Loewner
matrices. More precisely, let the following assumption be satisfied:

L
, for all s {i } { j }.
rank(sL M) = rank L M = rank
M
is a truncation index that is assumed to be no larger than ranksL M. q, r and

choosing to be the numerical rank of sL M is convenient. The best tool for determining the numerical rank of sL M, is the SVD (Singular Value Decomposition).
To that end, suppose
sL M = Y X ,
is the SVD of sL M for some choice of s {i } { j } and consider a truncated
SVD as Y Cq , X Cr .
Theorem 1.10. A realization [E , A , B , C ], of a minimal solution is given as follows:
C = Z
T X .
E = Y LX , A = Y MX , B = Y Y,
Depending on whether is the exact or approximate rank, we obtain either an
interpolant or an approximate interpolant of the data, respectively.
7.3 Loewner and Pick Matrices

The positive real interpolation problem can be formulated as follows. Given triples
(i , b i , y i ), i = 1, . . . , q, where i are distinct complex numbers in the right-half of
the complex plane, b i and y i are in Cr , we seek a rational matrix function H(s) of
size r r, such that H(i )b i = y i , i = 1, . . . , q, and in addition H is positive real.
This problem does not always have a solution. It is well known that the necessary
and sufficient condition for its solution is that the associated Pick matrix
y 1 b q + b 1 y q
y 1 b 1 + b 1 y 1
+
1 + q
1
1
..
..
..
=
.
.
.

y q b1 + bq y 1
y q bq + b q y q
q + 1
q + q
47
Cqq ,
be positive semi-definite, that is = 0. By comparing with the Loewner

matrix L defined in Sect. 7.2, we conclude that if the right (column) array for the
former is taken as (i , b i , y i ), i = 1, . . . , q, and the left (row) array as (i , b i , yi ),
i = 1, . . . , q, then
= L.
The left array is then called the mirror-image array. Thus for this choice of interpolation data the Pick matrix is the same as the Loewner matrix. This shows the
importance of the Loewner matrix as a tool for studying rational interpolation.
Remark 1.5. (a) The above considerations provide an algebraization of the positive
real interpolation problem. If namely, 0, the minimal-degree rational functions
which interpolate simultaneously the original array and its mirror image array, are
automatically positive real and hence stable as well. The data in the model reduction
problem that we studied in Sect. 4, automatically satisfy this positive definiteness
constraint, and therefore the reduced system is positive real.
(b) It readily follows that interpolants of the original and the mirror-image arrays
constructed by means of the Loewner matrix, satisfy

H(i ) + HT (i ) b i = 0.
In general the zeros i of H(s) + HT (s) are called spectral zeros, and b i are the
corresponding (right) zero directions. Thus the construction of positive real interpolants by means of the Loewner (Pick) matrix, forces these interpolants to have the
given interpolation points as spectral zeros.
(c) The observation that the Pick matrix is a special case of the Loewner matrix,
the algebraization discussed above, and passivity preservation by means of spectral
zero interpolation, first appeared in [3]. See also [4].
7.4 Examples
7.4.1 A Simple Low-order Example
First we will illustrate the above results by means of a simple example. Consider a
2 2 rational function with minimal realization:
!
1 0 0
0 1 0
0 1
1 0 0
.
A = 0 1 0 , E = 0 0 0 , B = 1 0 , C =
0 1 1
0 0 0
0 0 1
0 1
48
Thus the transfer function is

'
H(s) = C(sE A)1 B =
s 1
1
1
s
(
.
Since rank E = 2, the McMillan degree of H is 2. Our goal is to recover this function
through interpolation. The data will be chosen in two different ways.
First, we will choose matrix data, that is the values of the whole matrix are
available at each interpolation point:
1 = 1,
2 = 1,
3 = 2,
4 = 2,

b1 = 1 , b 2 = 0 , b 3 = 1 , b 4 = 0
0
1
0
1

1
1
1
2
, y 2 =
, y 3 =
, y 4 = 1
y 1 =
1
1
1
2
1 = 1,
2 = 1,
3 = 2,
4 = 2,

c T1 = 1, 0 , c T2 = 0, 1 , c T3 = 1, 0 , c T4 = 0, 1

z T1 = 1, 1 , z T2 = 1, 1 , z T3 = 2, 1 , z T4 = 1, 12
The associated (block) Loewner and shifted Loewner matrices turn out to be:
1 0 1 0
0 1 1 1
1 0 1 0
0 1 0 12
L=
, M = 1 1 0 1
1 0 1 0
1
1
1 0 1 0
0 2 0 4
Notice that the rank of both Loewner matrices is 2 while the rank of xi L M is
3, for all x equal to a i or i . It can be readily verified that the column span of
1 L M = L M is the same as that of , where
1 1 2
1 0 0
=
0 1 0 .
0 0 1
Furthermore the row span of L M is the same as that of Thus
3
2 3
1
2 2
2
= M = 1
0 4 , E = L = 2 4
4,
A
3
17
1
0
4
4
2
4
49

0
0
T = 3
= 2 3 13 ,
=Y
= Z
2 , C
B
2 2 2
3 52
E A)
B,
which shows that a (second) minimal realization of
satisfy H(s) = C(s
H has been obtained.
The second experiment involves tangential data, that is, at each interpolation
point only values along certain directions are available.
1 = 1,
2 = 2,
3 = 3,

1
0
1
, b 2 =
, b 3 =
b 1 =
0
1
1

1
4
1
y 1 =
, y 2 = 1 , y 3 = 4
1
2
3
1 = 1,
2 = 2,
3 = 3,

c T1 = 1, 0 , c T2 = 0, 1 , c T3 = 1, 1

z T1 = 1, 1 , z T2 = 1, 12 , z T3 = 2, 23 .
Thus the associated Loewner and shifted Loewner matrices are:
1 0 1
0 1 3
1
1
L=
0 4 6 , M = 1 0 1
1 1 2
1 16 10
9
It readily follows that the conditions of theorem 1.9 are satisfied and hence the
T , Y),
provides a (third) minimal realization of the original
quadruple (M, L, Z
[sL M]1 Z
T .
rational function: H(s) = Y
7.4.2 Coupled Mechanical System

Figure 6 depicts a constrained mechanical system described in [67]. There are g
masses in total; the ith mass of weight mi , is connected to the (i + 1)st mass by a
spring and a damper with constants ki and di , respectively, and also to the ground
by a spring and a damper with constants i and i , respectively. Additionally, the
first mass is connected to the last one by a rigid bar (holonomic constraint) and it is
influenced by the control u(t). The vibration of this constrained system is described
in generalized state space form as:
Ex(t) = Ax(t) + Bu(t), y(t) = Cx(t),
50
k1
m1
u
1
ki+1
ki
kg1
mg
mi
d1
di
i
di+1
dg1
g
Fig. 6 Constrained mechanical system
Fig. 7 The singular values of the Loewner matrix
where x contains the positions and velocities of the masses,
0
I 0 0
0 I 0
E = 0 M 0 , A = K D G , B = B2 , C = [C1 , C2 , C3 ];
G 0 0
0
0 0 0
furthermore M is the mass matrix (g g, diagonal, positive definite), K is the stiffness matrix (g g, tri-diagonal), D is the damping matrix (g g, tri-diagonal),
G = [1, 0, , 0, 1], is the 1 g constraint matrix.
In [67], balanced truncation methods for descriptor systems are used to reduce
this system. Here we will reduce this system by means of the Loewner framework.
Towards this goal, we compute 200 frequency response data, that is H(ii ), where
i [2, +2]. Figure 7 shows the singular values of the Loewner matrix pair, which
indicate that a system of order 20 will have an approximate error 103 (60 dB).

20
51
Bode plots of original system and three approximants
10
data1
data2
data3
data4
0
10
20
30
40
50
101
100
101
Error Bode plots

20
10
data1
data2
data3
0
10
20
30
40
50
60
70
80
101
100
101
Fig. 8 Upper pane: Frequency responses of original system and approximants (orders 2, 10, 18).
Lower pane: Frequency responses of error systems (orders 2,10,18)
Figure 8 shows that (for the chosen values of the parameters) the frequency response has about seven peaks. A second order approximant reproduces (approximately) the highest peak, a tenth order system reproduces (approximately) five
peaks, while a system of order 18 provides a good approximation of the whole
frequency response (see in particular the error plots lower pane of Figure 8).
52
7.4.3 Four-pole Band-pass Filter

In this case 1,000 frequency response measurements are given, of the 2 2 Sparameters of a semi-conductor device which is meant to be a band-pass filter. There
is no a priori model available. The range of frequencies is between 40 and 120 GHz;
We will use the Loewner matrix procedure applied to the S-parameters. This yields
L, M C2,0002,000.
In the upper left-hand plot of Fig. 9, the singular values of the Loewner matrix
corresponding to the two-port system (upper curve) is compared with the singular
values of two one-port subsystems (lower curves). As the decay of all curves is fast,
an approximant of order around 20, is expected to provide a good fit. Indeed, as
the upper right-hand plot shows, a 21st order approximant provides fits with error
less than 60 dB. For comparison the fit of a 15th order model is shown in the
lower left-hand plot. Sometimes in practical applications, the entries of the two-port
S-parameters are modeled separately. In our case 14th order models are sufficient,
but the McMillan degree of the two-port is 28 or higher (depending on the symmetries involved, e.g. S11 = S22 , S12 = S21 ).
10
20
30
40
50
60
70
80
Fig. 9 Upper row, left pane: The singular values of xL M, for the two-port and for two one-ports.
Upper row, right pane: The S(1, 1) and S(1, 2) parameter data for a 21st order model. Lower row,
left pane: Fitting S(1, 1), S(1, 2) jointly with a 15th order approximant. Lower row, right pane:
Fitting S(1, 1), S(1, 2) separately with 14th order approximants
53
8 Conclusions
In this chapter, we have surveyed two projection-based model reduction frameworks
both of which make fundamental use of tangential rational interpolation. This approach is extremely flexible in applications; scales well to handle extremely large
problems; and is capable of producing very high fidelity reduced order models very
efficiently. Throughout, examples are given that illustrate the theoretical concepts
discussed.
The first framework we consider assumes the availability of a high-order dynamical system model in state space form (often these are acquired and assembled
through the independent discretization and coupling of high resolution distributed
parameter models). The retention of high model fidelity is directly recast as the problem of choosing appropriate interpolation points and associated tangent directions.
We address this with a detailed discussion of how interpolation points and tangent
directions can be be chosen to obtain reduced order models which are optimal with
respect to H2 error measures (Sect. 3), or so that reduced order models are obtained
which retain the property of passivity (Sect. 4).
The flexibility of our approach is demonstrated in Sects. 5 and 6 where we explore significant generalizations to the basic problem setting. We consider how to
preserve second-order system structure; reduction of systems that involve delays
or memory terms; and systems having a structured dependence on parameters that
must be retained in the reduced systems.
The second major framework we explore allows for the original (high-order)
dynamical system model to be inaccessible, and assumes that only system response
data (e.g. frequency response measurements) are available. We describe an approach
using the Loewner matrix pencil and place it in the context of interpolatory model
reduction methods. We show that the Loewner matrix pencil constitutes an effective
tool in obtaining minimal state-space realizations of reduced order models directly
from measured data.
Acknowledgments The work of the A.C. Antoulas was supported in part by the NSF through
Grant CCF-0634902. The work of C. Beattie and S. Gugercin has been supported in part by NSF
Grants DMS-0505971 and DMS-0645347.
References
1. Anderson, B., Antoulas, A.: Rational interpolation and state variable realizations. In: Decision
and Control, 1990. Proceedings of the 29th IEEE Conference on pp. 18651870 (1990)
2. Antoulas, A.: Approximation of Large-Scale Dynamical Systems (Advances in Design and
Control). Society for Industrial and Applied Mathematics, Philadelphia, PA, USA (2005)
3. Antoulas, A.: A new result on passivity preserving model reduction. Systems and Control
Letters 54, 361374 (2005)
4. Antoulas, A.: On the construction of passive models from frequency response data. Automatisierungstechnik 56, 447452 (2008)
54
5. Antoulas, A., Anderson, B.: On the scalar rational interpolation problem. IMA Journal of
Mathematics Control and Information 3, 6168 (1986)
6. Bagley, R., Torvik, P.: A theoretical basis for the application of fractional calculus to viscoelasticity. Journal of Rheology 27, 201210 (1983)
7. Bai, Z.: Krylov subspace techniques for reduced-order modeling of large-scale dynamical
systems. Applied Numerical Mathematics 43(1-2), 944 (2002)
8. Bai, Z., Bindel, D., Clark, J., Demmel, J., Pister, K., Zhou, N.: New numerical techniques and
tools in SUGAR for 3D MEMS simulation. In: Technical Proceedings of the Fourth International Conference on Modeling and Simulation of Microsystems, pp. 3134 (2000)
9. Bai, Z., Feldmann, P., Freund, R.: Stable and passive reduced-order models based on partial
Pade approximation via the Lanczos process. Numerical Analysis Manuscript 97, 310 (1997)
10. Bai, Z., Feldmann, P., Freund, R.: How to make theoretically passive reduced-order models
passive inpractice. In: Custom Integrated Circuits Conference, 1998. Proceedings of the IEEE
1998, pp. 207210 (1998)
11. Bai, Z., Freund, R.: A partial Pade-via-Lanczos method for reduced-order modeling. Linear
Algebra and its Applications 332(334), 141166 (2001)
12. Bai, Z., Skoogh, D.: A projection method for model reduction of bilinear dynamical systems.
Linear Algebra and Its Applications 415, 406425 (2006)
13. Bai, Z., Su, Y.: Dimension reduction of second order dynamical systems via a second-order
Arnoldi method. SIAM Journal on Scientific Computing 5, 16921709 (2005)
14. Baur, U., Beattie, C., Benner, P., Gugercin, S.: Interpolatory projection methods for parameterized model reduction. Chemnitz Scientific Computing Preprints 09-08, TU Chemntiz, (ISSN
1864-0087) November 2009
15. Beattie, C., Gugercin, S.: Krylov-based minimization for optimal H2 model reduction. In:
Proceedings of the 46th IEEE Conference on Decision and Control pp. 43854390 (2007)
16. Beattie, C., Gugercin, S.: Interpolatory projection methods for structure-preserving model reduction. Systems and Control Letters 58(3), 225232 (2009)
17. Beattie, C., Gugercin, S.: A trust region method for optimal H2 model reduction. In: Proceedings of the 48th IEEE Conference on Decision and Control (2009)
18. Benner, P.: Solving large-scale control problems. IEEE Control Systems Magazine 24(1), 44
59 (2004)
19. Benner, P., Feng, L.: A robust algorithm for parametric model order reduction based on implicit moment matching. In B. Lohmann and A. Kugi (eds.) Tagungsband GMA-FA 1.30
Modellbildung, Identifizierung und Simulation in der Automatisierungstechnik, Workshop
in Anif, 26.28.9.2007, pp. 3447 (2007)
20. Benner, P., Saak, J.: Efficient numerical solution of the LQR-problem for the heat equation.
Proceedings in Applied Mathematics and Mechanics 4(1), 648649 (2004)
21. Benner, P., Sokolov, V.: Partial realization of descriptor systems. Systems and Control Letters
55(11), 929938 (2006)
22. Bond, B., Daniel, L.: Parameterized model order reduction of nonlinear dynamical systems.
In: IEEE/ACM International Conference on Computer-Aided Design, 2005. ICCAD-2005,
pp. 487494 (2005)
23. Bui-Thanh, T., Willcox, K., Ghattas, O.: Model reduction for large-scale systems with highdimensional parametric input space. SIAM Journal on Scientific Computing 30(6), 32703288
(2008)
24. Bungartz, H.: Dunne Gitter und deren Anwendung bei der adaptiven Losung der dreidimensionalen Poisson-Gleichung. Dissertation, Institut fur Informatik, TU Munchen (1992)
25. Bunse-Gerstner, A., Kubalinska, D., Vossen, G., Wilczek, D.: H2 -optimal model reduction
for large scale discrete dynamical MIMO systems. Journal of Computational and Applied
Mathematics (2009). Doi:10.1016/j.cam.2008.12.029
26. Chahlaoui, V., Gallivan, K.A., Vandendorpe, A., Van Dooren, P.: Model reduction of secondorder system. In: P. Benner, V. Mehrmann, D. Sorensen (eds.) Dimension Reduction of
Large-Scale Systems, Lecture Notes in Computational Science and Engineering, vol. 45,
pp. 149172. Springer, Berlin/Heidelberg, Germany (2005)
55
27. Clark, J.V., Zhou, N., Bindel, D., Schenato, L., Wu, W., Demmel, J., Pister, K.S.J.: 3D MEMS
simulation using modified nodal analysis. In: Proceedings of Microscale Systems: Mechanics
and Measurements Symposium, p. 6875 (2000)
28. Craig Jr., R.R.: Structral Dynamics: An Introduction to Computer Methods. Wiley, New York
(1981)
29. Daniel, L., Siong, O., Chay, L., Lee, K., White, J.: A Multiparameter moment-matching
model-reduction approach for generating geometrically parameterized interconnect performance models. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems 23(5), 678693 (2004)
30. De Villemagne, C., Skelton, R.: Model reductions using a projection formulation. International
Journal of Control 46(6), 21412169 (1987)
31. Fanizza, G., Karlsson, J., Lindquist, A., Nagamune, R.: Passivity-preserving model reduction
by analytic interpolation. Linear Algebra and Its Applications 425(2-3), 608633 (2007)
32. Farle, O., Hill, V., Ingelstrom, P., Dyczij-Edlinger, R.: Multi-parameter polynomial order
reduction of linear finite element models. Mathematical and Computational Modeling of
Dynamics Systems 14(5), 421434 (2008)
33. Feldmann, P., Freund, R.: Efficient linear circuit analysis by Pade approximation via the
Lanczos process. IEEE Transactions on Computer-Aided Design of Integrated Circuits and
Systems 14(5), 639649 (1995)
34. Feng, L.: Parameter independent model order reduction. Mathematics and Computers in Simulations 68(3), 221234 (2005)
35. Feng, L., Benner, P.: A robust algorithm for parametric model order reduction based on implicit moment matching. PAMM Proceedings of the Applied Mathematics and Mechanism
7(1), 1021,5011021,502 (2008)
36. Feng, L., Rudnyi, E., Korvink, J.: Preserving the film coefficient as a parameter in the compact thermal model for fast electrothermal simulation. IEEE Transactions on Computer-Aided
Design of Integrated Circuits and Systems 24(12), 18381847 (2005)
37. Freund, R., Feldmann, P.: Efficient small-signal circuit analysis and sensitivity computations
with the PVL algorithm. In: Proceedings of the 1994 IEEE/ACM International Conference on
Computer-aided Design, pp. 404411. IEEE Computer Society Press Los Alamitos, CA, USA
(1994)
38. Freund, R.W.: Pade-type model reduction of second-order and higher-order linear dynamical systems. In: P. Benner, V. Mehrmann, D.C. Sorensen (eds.) Dimension Reduction of
Large-Scale Systems, Lecture Notes in Computational Science and Engineering, vol. 45,
pp. 191223. Springer, Berlin/Heidelberg (2005)
39. Gallivan, K., Vandendorpe, A., Dooren, P.: Model reduction of MIMO systems via tangential
interpolation. SIAM Journal on Matrix Analysis and Applications 26(2), 328349 (2005)
40. Glover, K.: All optimal Hankel-norm approximations of linear multivariable systems and their
L -error bounds. International Journal of Control 39(6), 11151193 (1984)
41. Griebel, M.: A parallelizable and vectorizable multi-level algorithm on sparse grids. In: Parallel Algorithms for Partial Differential Equations (Kiel, 1990), Notes in Numerical Fluid
Mechanics, vol. 31, pp. 94100. Vieweg, Braunschweig (1991)
42. Grimme, E.: Krylov projection methods for model reduction. Ph.D. thesis, CoordinatedScience Laboratory, University of Illinois at Urbana-Champaign (1997)
43. Grimme, E., Sorensen, D., Dooren, P.: Model reduction of state space systems via an implicitly
restarted Lanczos method. Numerical Algorithms 12(1), 131 (1996)
44. Gugercin, S.: Projection methods for model reduction of large-scale dynamical systems. Ph.D.
thesis, Ph.D. Dissertation, ECE Department, Rice University, December 2002 (2002)
45. Gugercin, S.: An iterative rational Krylov algorithm (IRKA) for optimal H2 model reduction.
In: Householder Symposium XVI. Seven Springs Mountain Resort, PA, USA (2005)
46. Gugercin, S., Antoulas, A.: An 2 error expression for the Lanczos procedure. In: Proceedings
of the 42nd IEEE Conference on Decision and Control (2003)
47. Gugercin, S., Antoulas, A.: Model reduction of large-scale systems by least squares. Linear
Algebra and its Applications 415(2-3), 290321 (2006)
56
48. Gugercin, S., Antoulas, A., Beattie, C.: A rational Krylov iteration for optimal H2 model
reduction. In: Proceedings of MTNS, vol. 2006 (2006)
49. Gugercin, S., Antoulas, A., Beattie, C.: H2 model reduction for large-scale linear dynamical
systems. SIAM Journal on Matrix Analysis and Applications 30(2), 609638 (2008)
50. Gugercin, S., Willcox, K.: Krylov projection framework for fourier model reduction. Automatica 44(1), 209215 (2008)
51. Gunupudi, P., Khazaka, R., Nakhla, M.: Analysis of transmission line circuits using multidimensional model reduction techniques. IEEE Transactions on Advanced Packaging 25(2),
174180 (2002)
52. Gunupudi, P., Khazaka, R., Nakhla, M., Smy, T., Celo, D.: Passive parameterized time-domain
macromodels for high-speed transmission-line networks. IEEE Transactions on Microwave
Theory and Techniques 51(12), 23472354 (2003)
53. Halevi, Y.: Frequency weighted model reduction via optimal projection. IEEE Transactions
on Automatic Control 37(10), 15371542 (1992)
54. Hyland, D., Bernstein, D.: The optimal projection equations for model reduction and the relationships among the methods of Wilson, Skelton, and Moore. IEEE Transactions on Automatic
Control 30(12), 12011211 (1985)
55. Ionutiu, R., Rommes, J., Antoulas, A.: Passivity preserving model reduction using dominant
spectral zero interpolation. IEEE Transactions on CAD (Computer-Aided Design of Integrated
Circuits and Systems) 27, 22502263 (2008)
56. Kellems, A., Roos, D., Xiao, N., Cox, S.J.: Low-dimensional morphologically accurate models
of subthreshold membrane potential. Journal of Computational Neuroscience, 27(2), 161176
(2009)
57. Korvink, J., Rudnyi, E.: Oberwolfach benchmark collection. In: P. Benner, V. Mehrmann, D.C.
Sorensen (eds.) Dimension Reduction of Large-Scale Systems, Lecture Notes in Computational Science and Engineering, vol. 45, pp. 311315. Springer, Berlin/Heidelberg, Germany
(2005)
58. Krajewski, W., Lepschy, A., Redivo-Zaglia, M., Viaro, U.: A program for solving the L2
reduced-order model problem with fixed denominator degree. Numerical Algorithms 9(2),
355377 (1995)
59. Kubalinska, D., Bunse-Gerstner, A., Vossen, G., Wilczek, D.: H2 -optimal interpolation based
model reduction for large-scale systems. In: Proceedings of the 16th International Conference
on System Science, Poland (2007)
60. Kunkel, P.: Differential-Algebraic Equations: Analysis and Numerical Solution. European
Mathematical Society, Amsterdam (2006)
61. Lasance, C.: Two benchmarks to facilitate the study of compact thermal modelingphenomena. IEEE Transactions on Components and Packaging Technologies [see also IEEE Transactions on Packaging and Manufacturing Technology, Part A: Packaging Technologies] 24(4),
559565 (2001)
62. Lefteriu, S., Antoulas, A.: A new approach to modeling multi-port systems from frequency
domain data. IEEE Transactions on CAD (Computer-Aided Design of Integrated Circuits and
Systems) 29, 1427 (2010)
63. Leitman, M., Fisher, G.: The linear theory of viscoelasticity. Handbuch der Physik 6, 1031
(1973)
64. Leung, A.M., Khazaka, R.: Parametric model order reduction technique for design optimization. In: IEEE International Symposium on Circuits and Systems, 2005. ISCAS 2005, vol. 2,
pp. 12901293 (2005)
65. Ma, M., Leung, A.M., Khazaka, R.: Sparse Macromodels for Parametric Networks. In: Proceedings of the IEEE International Symposium on Circuits and Systems (2006)
66. Mayo, A., Antoulas, A.: A framework for the solution of the generalized realization problem.
Linear Algebra and Its Applications 425(2-3), 634662 (2007)
67. Mehrmann, V., Stykel, T.: Balanced truncation model reduction for large-scale systems in
descriptor form. In: P. Benner, V. Mehrmann, D. Sorensen (eds.) Dimension Reduction of
Large-Scale Systems, pp. 83115. Springer, Belin (2005)
57
68. Meier III, L., Luenberger, D.: Approximation of linear constant systems. IEEE Transactions
69. Meyer, D., Srinivasan, S.: Balancing and model reduction for second-order form linear
systems. IEEE Transactions on Automatic Control 41(11), 16321644 (1996)
70. Moore, B.: Principal component analysis in linear systems: Controllability, observability, and
model reduction. IEEE Transactions on Automatic Control 26(1), 1732 (1981)
71. Moosmann, K., Korvink, J.: Automatic parametric mor for mems design. In: B. Lohmann,
A. Kugi (eds.) Tagungsband GMA-FA 1.30 Modellbildung, Identifikation und Simulation in
der Automatisierungstechnik, Workshop am Bostalsee, 27.29.9.2006, pp. 8999 (2006)
72. Mullis, C., Roberts, R.: Synthesis of minimum roundoff noise fixed point digital filters. IEEE
Transactions on Circuits and Systems 23(9), 551562 (1976)
73. Odabasioglu, A., et al.: PRIMA. IEEE Transaction on Computer-Aided Design of Integrated
Circuits and Systems 17(8) (1998)
74. Pillage, L., Rohrer, R.: Asymptotic waveform evaluation for timing analysis. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems 9(4), 352366 (1990)
75. Preumont, A.: Vibration Control of Active Structures: An Introduction. Springer, Berlin
(2002)
76. Raghavan, V., Rohrer, R., Pillage, L., Lee, J., Bracken, J., Alaybeyi, M.: AWE-inspired. In:
Custom Integrated Circuits Conference, 1993. Proceedings of the IEEE 1993 (1993)
77. Rommes, J., Martins, M.: Efficient computation of multivariable transfer function dominant
poles using subspace acceleration. IEEE Transactions on Power Systems 21, 14711483
(2006)
78. Rudnyi, E., Moosmann, C., Greiner, A., Bechtold, T., Korvink, J.: Parameter Preserving Model
Reduction for MEMS System-level Simulation and Design. In: Fifth MATHMOD Proceedings 1 (2006)
79. Ruhe, A.: Rational Krylov algorithms for nonsymmetric eigenvalue problems. II: matrix pair.
Linear Algebra and Its Applications Applications, pp. 282295 (1984)
80. Sorensen, D.: Passivity preserving model reduction via interpolation of spectral zeros. Systems
and Control Letters 54, 354360 (2005)
81. Spanos, J., Milman, M., Mingori, D.: A new algorithm for L2 optimal model reduction. Automatica (Journal of IFAC) 28(5), 897909 (1992)
82. Su, T.J., Jr., R.C.: Model reduction and control of flexible structures using Krylov vectors.
Journal of Guid. Control Dyn. 14, 260267 (1991)
83. Van Dooren, P., Gallivan, K., Absil, P.: H2 -optimal model reduction of MIMO systems.
Applied Mathematics Letters 21(12), 12671273 (2008)
84. Weaver, W., Johnston, P.: Structural dynamics by finite elements. Prentice-Hall, Upper Saddle
River (1987)
85. Weile, D., Michielssen, E., Grimme, E., Gallivan, K.: A method for generating rational interpolant reduced order models of two-parameter linear systems. Applied Mathematics Letters
12(5), 93102 (1999)
86. Weile, D.S., Michielssen, E., Grimme, E., Gallivan, K.: A method for generating rational interpolant reduced order models of two-parameter linear systems. Applied Mathematics Letters
12(5), 93102 (1999)
87. Willcox, K., Megretski, A.: Fourier series for accurate, stable, reduced-order models in largescale applications. SIAM Journal of Scientific Computing 26(3), 944962 (2005)
88. Wilson, D.: Optimum solution of model-reduction problem. Proceedings of IEE 117(6),
11611165 (1970)
89. Yan, W., Lam, J.: An approximate approach to h2 optimal model reduction. IEEE Transactions
90. Yousuff, A., Skelton, R.: Covariance equivalent realizations with applications to model reduction of large-scale systems. Control and Dynamic Systems 22, 273348 (1985)
91. Yousuff, A., Wagie, D., Skelton, R.: Linear system approximation via covariance equivalent
realizations. Journal of Mathematical Analysis and Applications 106(1), 91115 (1985)
58
92. Zenger, C.: Sparse grids. In: Parallel Algorithms for Partial Differential Equations (Kiel,
1990), Notes Numerical Fluid Mechanics, vol. 31, pp. 241251. Vieweg, Braunschweig (1991)
93. Zhaojun, B., Qiang, Y.: Error estimation of the Pade approximation of transfer functions via
the Lanczos process. Electronic Transactions on Numerical Analysis 7, 117 (1998)
94. Zhou, K., Doyle, J., Glover, K.: Robust and Optimal Control. Prentice-Hall, Upper Saddle
River, NJ (1996)
95. Zigic, D., Watson, L., Beattie, C.: Contragredient transformations applied to the optimal projection equations. Linear Algebra and Its Applications 188, 665676 (1993)
Efficient Model Reduction for the Control

of Large-Scale Systems
Richard Colgren
1 Introduction
When analyzing and controlling large-scale systems, it is extremely important to
develop efficient modeling processes. The key dynamic elements must be identified
and spurious dynamic elements eliminated. This allows the controls engineer to
implement the optimal control strategy for the problem at hand. Model reduction
techniques provide an extremely effective way to address this requirement.
In this chapter the evolution of model reduction techniques for designing control systems for large-scale systems is summarized. These start with simple approaches such as spectral decomposition and simultaneous gradient error reduction
and then progresses through a variety of balanced and related model reduction approaches. Motivations for the use of these methods are given. Emphasis is given
to approaches which are easily applied to the large, generalized models which are
created by Computer Aided Design (CAD) tools. A U-2S example application is
provided and described.
The provided example of a large-scale system of very high order is a model of
an air vehicle with significant aeroservoelastic coupling. Future unstable vehicles, or
those employing features such as relaxed static stability, can display aeroservoelastic
coupling. Evaluation of the interactions between the dynamic modes must be accomplished using an efficient, integrated modeling approach. The extensive use of
composite materials has also resulted in greater aeroservoelastic coupling. Finite element models of complex systems are sparse and are also intrinsically of very high
order. Such large-scale systems must be subjected to comprehensive analysis.
The methods discussed in this chapter preserve the frequency response characteristics of the system model being examined while reducing its size to one practical for
direct controls design. They support a variety of optimal and robust control system
R. Colgren
Vice President and Chief Scientist, Viking Aerospace LLC, 100 Riverfront Road, Suite B,
Lawrence, KS 66044, USA
e-mail: rcolgren@gmail.com
59
60
R. Colgren
design approaches. Some control system design methods require the use of all of the
systems states. Without the use of model reduction techniques the full state control
system design would be too large to practically implement. States that would be
included within the controller might also be outside the bandwidth of the servos or
actuators. This might generate an ill-conditioned problem. The use of model reduction methods provides a way to generate full state controller solutions of reduced
size, and to further simplify these full state feedback solutions once they have been
generated.
2 Spectral Decomposition
This method is an efficient way to generate a reduced order model of a large-scale
system when all of its subsystems are decoupled or are at most weakly coupled.
It was initially developed by Chiodi and Davis at the Lockheed Corporation [1].
It depends on systems having distinct eigenvalues within well separated frequency
responses as shown in Fig. 1. These modal groupings are then decoupled using the
eigenvector matrix.
The starting point for this method is the standard nth order state space representation given in (1).
dx/dt = Ax + Bu; y = Cx
(1)
Fig. 1 Distinct frequency range groupings of a models eigenvalues
Efficient Model Reduction for the Control of Large-Scale Systems
61
The system matrix A is then represented using the systems eigenvalue and
eigenvector matrices as shown in (2).
dx/dt = V V 1 x + Bu; y = Cx
(2)
This system description can then be expanded into the following series representation as provided in (3). Note that the sum is initially computed using the full
number of states within the model, but that for control system design it is computed
using the eigenvalues to be obtained. Also, please note that ViVi1 is the residue of
the eigenvalue i in (sI A)1 .
dx/dt = [ ViVi1 i ]x + Bu; y = Cx
(3)
From this representation the system can be reduced into its individual elements as
is done in (4)
dxi /dt = Ai x + Bi u
(4)
The spectral decomposition process is described in Fig. 2. The system eigenvalues and eigenvectors are next calculated. For each eigenvalue to be retained within
the reduced order model the eigenvector and its inverse are multiplied to generate
ViVi1 . Each matrix element is multiplied by its corresponding eigenvalue i . The
transformed state matrix is then constructed by solving for the sum as given in (5)
for all the retained eigenvalues.
A = ViVi1 i
Fig. 2 Spectral decomposition process
(5)
62
R. Colgren
To calculate B = T B and C = CT the transformation matrix T must also be calculated. This transformation is provided as in (6).
T = ViVi1
(6)
After spectral decomposition the system is transformed into that provided as in

(7)(9).
dx1 /dt = A11 x1 + A12 x2 + B1 u
(7)
dx2 /dt = A21 x1 + A22 x2 + B2 u
(8)
y = C1 x1 + C2 x2
(9)
The reduced system model as described by the states to be used for the design of
the control system is given as in (10) and (11).
dx1 /dt = A11 x1 + B1 u
(10)
y = C 1 x1
(11)
3 Simultaneous Gradient Error Reduction

This method provides an efficient way to generate a reduced order model of a largescale system whether all of its subsystems and their outputs are decoupled or not. It
does require that all of its inputs are no more than weakly coupled to each other. For
flying qualities prediction and for aeroservodynamic modeling [2,3] the author originally developed this method for simultaneously fitting several frequency responses
depicting high order systems to low order transfer functions including time delays
[4, 5]. This method has been applied to the design the control systems for several
large-scale aerospace systems.
Simultaneous gradient error reduction is applied over a finite bandwidth. Controls analysis and design are always accomplished over a finite bandwidth. This
allows modes outside of the control bandwidth to be neglected without impacting
the responsiveness or the robustness of the closed loop system. As an example, for
representative handling qualities it is believed that the low and the high order system
models must have sufficiently similar frequency responses between approximately
1 to 10 rps. This means that, in the case of such a fit, that the pilot will judge the
dynamics to be equivalent between the high order and the reduced order system
models.
This procedure uses a conjugate gradient search routine to adjust the parameters of transfer functions including pure time delays with the desired form until
their frequency responses are as close as possible to the frequency responses from
the high order dynamics model. This is a multi-variable approach whereby several
output frequency responses to the same input can be analyzed simultaneously. The
63
Fig. 3 Simultaneous gradient error reduction
user is allowed to fix, free, or simultaneously fit the various modes, including time
delays, after selecting the order of the transfer function [5] This procedure is outlined in Fig. 3.
The desired result is to match a transfer function to each of the given plants
magnitude A( ) and phase frequency response data over the control bandwidth. Expressed in the form of a complex number A( )e j ( ) , these plant transfer
functions should respond in an equivalent manner to the full order plant model over
the desired frequency range for the given pair of inputs and outputs. The reduced
order transfer function representation is given in factored form and incorporates a
pure time delay. All of the output responses to a single large-scale system input can
be simultaneously fit using this approach. The output response models to the other
large-scale systems inputs are generated in subsequent analysis.
The cost function J represents the fit error between the plant data and the reduced
order model. It is formulated as the integral of the absolute value of the error squared
between the data and the reduced order model as shown in (12)
J=

0
|G( j ) A( )e j ( )|2 d
(12)
The fit error is weighted over each infinitesimal frequency interval. This fit error can
be assigned a relative importance as a function of frequency. A conjugate gradient
routine is used for error reduction because of its simplicity. Parameters are constrained to maintain the same sign during minimization. This is done to maintain
stability characteristics according to the Nyquist Stability Criteria. The gradient J
can be expressed analytically by differentiating the cost function given in (13).
J = Re [G( ji ) A(i ))e j i ] G( j )d
(13)
64
R. Colgren
G( j ) is defined in (14). In this equation Z( j ) is composed of the partial derivatives of G( j ) divided by their corresponding factors.
G( j ) = G( j )Z( j )
(14)
Gradient search techniques generally require parameter scaling to efficiently converge. An ideal scaling uses a cost function equally sensitive to each parameter.
The law to accomplish this scales each parameter by its own magnitude as in (15).
Re-scaling is done after completing each separate iteration.
Pi = bi /|bi |
(15)
The elements in the scaled gradient J are formed using (16).
J/ pi = bi / pi J/ bi = |bi | J/ bi
(16)
Simultaneously fit parameters are constrained within bounds derived by initially

allowing each transfer function to converge to an independent solution. An unweighted average is then computed, and is recomputed after each search. This is
continued until all reduced order transfer functions have converged, the simultaneously fit parameters have converged, or the maximum number of iterations have
been exceeded. In studies it was shown that the use of weighted averages did not
improve the cost function of the final result. Its use also created convergence stability problems. If the problem as formulated does not have a true joint minimum, the
gradient search routine will give the best estimated joint minimum.
4 Balancing
This method offers an efficient way to generate a reduced order model of a largescale system whether all of its subsystems and their inputs and outputs are decoupled or not. This method for reducing the order of large-scale system models was
originally developed by Enns while he was at Stanford [6]. An internally balanced
system representation has input and output grammians that are equal and diagonal
[7]. The magnitude of each diagonal element provides a measure of the controllability and observability of the corresponding state [8]. This is used as a guide in
selecting those states which are the least controllable or observable for elimination
[9]. The sum of the diagonal elements of the grammian corresponding to the states
eliminated provides an error bound between the high order representation and the
reduced order model.
Balancing relies on the notion of a system as a mapping from the inputs to the
servos or actuators and then to the sensor outputs [10]. This mapping is viewed as a
combination of the reachability mapping from the actuator input signals to the state
vector and an observability mapping from the state to the sensor output signals. It
follows from this mapping concept that the state is an intermediate quantity between
65
the inputs in the past and the sensor outputs in the future. The minimum amount of
control energy required to reach the desired state vector is inversely proportional to
the reachability grammian. Similarly, the amount of sensor output energy generated
by the state vector is proportional to the observability grammian. Balancing examines the amount each state component participates in the mapping from the input
to the output. The model reduction problem is now reduced into one of truncating
small terms from the partial fraction decomposition.
The following section pertains to a system again formulated as in (1), where A
and B are controllable and A and C are observable. Given these assumptions, the
controllability grammian U and the observability grammian Y are the solutions to
the Lyapunov equations given as in (17) and (18)
dx/dt = Ax + Bu = TAT 1 x + T 1 Bu
(17)
y = Cx = CT x
(18)
The transformation matrix T also relates the grammians through equations (19)
and (20) using an algorithm first developed by Laub [11, 12]. These contragedient transformations (where both U and Y are diagonal) can be calculated either
as the best conditioned contragedient transformation or as an internally balanced
transformation.
U = T 1UT 1
(19)
Y = T Y T
(20)
For an internally balanced transformation in (21) applies. In this representation the

columns of T are the eigenvectors of the product UY .
UY = T 1 T
(21)
To compute the balanced representation of the original large-scale system model,

it first must be decoupled into stable and unstable subsystem models. Balancing is
used for this step to generate a transformation based on the systems eigenvalues. For
simplicity, neutrally stable modes can be slightly perturbed to make them marginally
stable. If the conditionality of the resulting stable projection is a concern, all neutrally stable modes can also be preserved as a part of the unstable system. This can
lead to a higher order final reduced system model. The resulting system is given
in (22). In this equation G+ (s) is the stable subsystem and G (s) is the unstable
subsystem
G(s) = C(sI A)1 B = G+ (s) + G (s)
(22)
Without input or output weighting the stable subsystem is transformed into a
real Schur form. Next the Cholesky factors Lu and Ly of U and Y are computed. A
singular value decomposition of the product Ly Lu is then performed. The singular
values and the corresponding vectors are arranged in order of decreasing singular values. The transformation T is then computed in a final Schur transformation.
66
R. Colgren
Fig. 4 Balancing process
States which have relatively small singular values and thus have a relatively small
effect on the response of the large-scale system are then truncated from the model
of the stable subsystem. The unstable and ill conditioned subsystem is completely
preserved throughout this process. It is next recombined with the reduced order representation of the stable subsystem. This process is shown in Fig. 4.
4.1 Techniques Not Requiring Balancing

Other model reduction methods with an equivalent range of applicability have been
designed to avoid the balancing process. One such method uses the Hankel Minimum Degree Approximate algorithm as described in [12]. This method is similar to balanced additive model reduction routines but can produce a reduced order
model more reliably when the desired reduced model has nearly controllable and/or
observable states [13]. These conditions are equivalent to having Hankel singular
values very close to the machine accuracy.
For a stable system the Hankel singular values indicate the relative energy of each
state with respect to the state energy of the entire system [14]. The reduced order
system is directly determined by examining the systems Hankel singular values.
An optimal reduced order system is selected using this algorithm to satisfy the error
bound criterion regardless of the order selected at the beginning of the process [12].
Given the state space representation of the system as shown in (23), along with
the value selected for k (the desired reduced order), the reduced order model is
generated.
dx/dt = Ax + Bu; y = Cx
(23)
67
Fig. 5 Hankel minimum degree approximate process
The following steps, as shown in Fig. 5, produce a similarity transformation to

truncate the original state space system into the kth order reduced model. Weights
on the original models inputs and/or its outputs can make the model reduction algorithm focus on the specific frequency range of interest. These weights are required
to be stable, minimum phase, and invertible. Note that as in the previous section the
unstable subsystem must be combined with the reduced order stable subsystem to
create the final reduced order model.
In more detail the steps to accomplish model reduction without balancing using
the Hankel Minimum Degree Approximate algorithm are given as follows. First,
the original large-scale system model must be decoupled into stable and unstable
subsystem models. Next, the controllability and observability grammians P and Q
must be generated. Using these grammians the descriptor given in (24) is generated
E = QP( j ) 2 I
(24)
where k > k+1 .

A singular value decomposition is then accomplished on the descriptor. The system is next transformed to generate the system representation given in (25) and (26).
dx dt = U ( 2 A + QAP)V x + U (QB C )u
(25)
y = [CP B ]Vu
(26)
The resulting system is then partitioned and truncated to become a kth order system.
The final kth order Hankel Minimum Degree Approximate is the stable part of the
state space realization. Its unstable part must be recombined with the reduced order
model of the stable part.
68
R. Colgren
The nature of the resulting error between the original system G( j ) and the final reduced order model is discussed in [14]. This error is described by an all-pass
function. A detailed description of the Hankel Minimum Degree Approximate algorithm can be found in [13].
4.2 Balancing Over A Disk

This procedure is a further development of the balancing technique. It was originally developed by Jonckheere at the University of Southern California [15, 16].
Its advantages are that it gives precise frequency response bounds over a desired
bandwidth. Elimination of eigenvalues by inspection makes the balancing better
conditioned and computationally efficient as well as further decreasing the final size
of the reduced order system.
As was done using the previous two approaches, the unstable subsystem is removed and only the stable subsystem is processed and reduced. Note that with
proper sign changes the unstable subsystem could be balanced over a disk. Because of the Nyquist Stability Criteria the number of open loop poles should not be
reduced. In actual application work the number of unstable poles represents a very
small percentage of the total number of poles. Thus reducing the number of unstable
poles would have little effect on the dimension of the final reduced order system.
Standard balancing has infinite bandwidth. It thus does not necessarily provide
the smallest possible H error. What is desired is the smallest frequency response
error over the selected bandwidth. An infinite bandwidth is not required. From the
Hankel Singular Values for each pole the error bound is generated as in (27).
sup|G(s) G(s)| =
(27)
To develop a reduced order model optimized over a finite bandwidth and to improve
the conditionality of the stable subsystem to be balanced, the subsystem is balanced
over a disk. This is accomplished over a region as shown in Fig. 6. The disk is placed
in the complex plane based on two considerations. First, the full order model G(s)
must be analytic in the disk [10, 17]. For the targeted error bound the disk must be
placed to exclude any eigenvalues of G( j ). Second, the disk should cover the interval of real frequencies ( < < ). This bandwidth limit is usually based
on the response limitations imposed by the control servos or actuators. This guarantees that the error bound only includes this critical frequency range. A good rule
is to use = 12 (the largest real part of the poles of G( j )). This is done using a
bilinear mapping as shown in Fig. 7. The error bound becomes that given in (28)
sup|G(s) GDISK(s) | = DISK
(28)
Note that DISK . Those eigenvalues that have small singular values are
deleted from the stable subsystem. The reduced order stable subsystem is recombined
with the unstable subsystem. The overall process for reducing the order of a largescale system over a disk is presented in Fig. 8.
69
Fig. 6 Balancing over a Disk
Fig. 7 Bilinear mapping
5 Example: Large-Scale System Application

The model reduction techniques discussed in the previous sections are based on
minimizing the error between the high order and the low order models as measured
within the frequency domain. Using these approaches, a large-scale 140th order
multiple input, multiple output full dynamics model of the U-2S including rigid
body and structural dynamics modes was reduced in order [4,18]. Within this section
90th, 80th and 40th order reduced models are documented. The response of the 90th
order model is, by visual inspection, identical to the 140th order large-scale systems
response. The response of the 80th order model begins to show differences from that
of the large-scale system. The response shown in Fig. 9 is that of the roll rate due
to a rudder input. This difference results from the migration of two extremely high
frequency, lightly damped zeros from the left half into the right half of the complex
plane. The large-scale system has ten non-minimum phase zeros while the 80th
order model has 12 non-minimum phase zeros. Although the frequency response
70
R. Colgren
Fig. 8 Asymptotic Balancing Process
Fig. 9 Effect of non-minimum phase zeros
match shows very little variation between these two models responses up to the
frequency of this complex zero, the difference between the transient responses must
be noted. In practice this difference is so far outside the control bandwidth of the
servos that it will not have an impact on the design of the control laws.
71
Fig. 10 Effect of Residualization
When the model is further reduced to 40th order a steady state offset error appears
in the pitch rate response due to an elevator input. This offset is seen in Fig. 10.
Residualization is easily used to shift the low order response to nearly overlay the
140th order large-scale systems equivalent response. Only minor differences in the
peaks of the first 5 oscillations are seen after residualization. Besides this minor
error the residualized 40th order systems transient response overlays that of the
140th order large-scale system. This residualized systems time response is omitted
from Fig. 10 for clarity.
References
1. Davis W. J. and Chiodi O. A. A method for the decoupling of equations by spectral decomposition. Lockheed Report LR 27479, Dec 1975
2. Socolinsky D. A. Wolff L. B., Neuheisel J. D., and Eveland C. K. Background information
and user guide for mil-f-8785c, military specification, flying qualities of piloted airplanes.
AFWAL-TR-81-3109, Sept 1981
3. Seidel R. C. Transfer-function-parameter estimation from frequency response data - a fortran
program. NASA TM X-3286, Sept 1975
4. Colgren R. D. Methods for model reduction. AIAA-88-4144, pages 1517, Aug 1998
5. Colgren R. D. Simultaneous Fit Equivalent Systems Program Users Guide. Lockheed
California Company, Burbank, CA, 1985
6. Enns D. Model Reduction for Control System Design. PhD thesis, Department of Aeronautics
and Astronautics Stanford University, 1984
72
R. Colgren
7. Laub A. J., Heath M. T., Paige C. C., and Ward R. C. Computation of system balancing transformations and other applications of simultaneous diagonalization algorithms. IEEE Transactions on Automatic Control, Feb 1987
8. Kailath T. Linear Systems. Prentice-Hall, Upper Saddle River, 1980
9. Safonov M. G. and Chiang R. Y. A schur method for balanced model reduction. IEEE Trans.
on Automat. Contr., 34(7):729733, july 1989
10. Colgren R. D. and Jonckheere E. A. Balanced model reduction for large aircraft models.
MTNS91, June 1991
11. Anderson E. Bai Z., Bischof C., Blackford S., Demmel J., Dongarra J., Du Croz J.,
Greenbaum A., Hammarling S., McKenney A., and Sorensen D. Lapack Users Guide, 1999
12. Balas G., Chiang R., Packard A., and Safonov M. Robust control toolbox 3 users guide.
Revision for Version 3.3.3 (Release 2009a), Natick, MA, March 2009., March 2009
13. Safonov M. G., Chiang R. Y., and Limebeer D. J. N. Optimal hankel model reduction for
nonminimal systems. IEEE Transactions on Automatic and Control, 35(4):496502, April
1990
14. Glover K. All optimal hankel norm approximation of linear multivariable systems, and their l
infinity error bounds. International Journal of Control, 39(6):11451193, 1984
15. Jonckheere, E. A. and Silverman, L. M. A new set of invariants for linear systems application
to reduced order controller design. IEEE Transactions on Automatic Control, AC-28:953964,
1983
16. Jonckheere, E. A. and Silverman, L. M. A new set of invariants for linear systems application and approximation. In International Symposium on Theory, Networks and Systems, Santa
Monica, CA, 1981
17. Colgren and R. D. Finite bandwidth model reduction applied to the advanced supersonic transport. COMCON3, Victoria, BC, Canada, Oct 1991
18. Dudginski R. J., and Colgren R. D. Time domain effects of model order reduction. In 3rd
IEEE International Symposium on Intelligent Control, pp. 2426, Arlington, VA, Aug 1988
Dynamics of Tensegrity Systems

Maurcio C. de Oliveira and Anders S. Wroldsen
1 Introduction and Motivation

Buckminister Fuller [1] coined the word tensegrity as a conjunction of the two words
tension and integrity [3, 6, 10]. The artist Kenneth Snelson was the first to build a
three-dimensional tensegrity structure [9], which he perfected throughout the years,
such as the sculptures in Fig. 1. Tensegrity components are very simple elements,
often just sticks and strings. Our interest is in engineering structures. The structures
of interest have a large number of compressive parts (rods) and tensile parts (strings).
On a class 1 tensegrity system, such as the sculptures in Fig. 1, there is no contact
between the rigid bodies (rods). Integrity, or as we would say, stability, is provided
by the network of tensile members (strings). No individual member is subject to
torques even when the complete structure bends. In a class k tensegrity system, we
allow as many as k rigid bodies (rods) to be in contact through ball joints. This last
requirement preserves the ability of class k tensegrity systems to bend without any
member experiencing torque.
The purpose of this chapter is to derive dynamic models for Tensegrity Systems.
We will produce models that can cope with the large number of members often
present in Tensegrity Systems. The main concerns for the mathematical models are
that they should produce efficient computer simulations and that they may be used
as a basis for control design. In this chapter we focus primarily on the first aspect,
M.C. de Oliveira
Department of Mechanical and Aerospace Engineering, University of California San Diego,
9500 Gilman Dr., La Jolla CA, 92093-0411, USA
e-mail: mauricio@ucsd.edu
A.S. Wroldsen
Centre for Ships and Ocean Structures, Norwegian University of Science and Technology,
Trondheim, Norway
e-mail: wroldsen@ntnu.no
73
74
M.C. de Oliveira and A.S. Wroldsen
Fig. 1 Class 1 tensegrity towers by Kenneth Snelson
dynamic modeling. The interested reader is referred to [12] for an application of

such models to control of Tensegrity Systems and to the book [7] as a general reference on tensegrity systems.
We will construct dynamic models for Tensegrity Systems made of sticks and
strings in the form of ordinary differential equations under the following assumptions: (a) rods are rigid, thin and long so rotational motion about the longitudinal
axis can be neglected; (b) strings are massless elastic elements with Hookean (linear) behavior only when in tension; (c) the connectivity of the structure is fixed.
These assumptions reflect tensegrity structures where the rods are massive and
stiff, here approximated as rigid, as compared with a network of lightweight, elastic
strings. Herein we are motivated by the network approach in [8]. We first study the
dynamics of a single rod, in Sect. 2. These will be used to derive dynamic models
for class 1 tensegrity systems in Sect. 3. These models can be applied to any type
of class 1 tensegrity system. In Sect. 4 we turn our attention to the much more
complex problem of deriving dynamic models for class k tensegrity systems. Instead of providing a general solution, we study a particular example of a class 2
tensegrity system, a simple cable, which will highlight the inherent difficulties to
obtain dynamic models for class k tensegrity systems and illustrate the benefits of
our proposed approach. The chapter is concluded with some general remarks on the
implementation of these models in a computer simulation environment.
2 Dynamics of a Single Rigid Rod

We start by defining some important quantities associated with the dynamics of the
single rigid rod in three-dimensional space as illustrated in Fig. 2. This rod has mass
m > 0 and length > 0 with extreme points n j , ni R3 , hence n j ni = . We
often make use of the normalized rod vector
b = 1 (n j ni ),
b = 1.
(1)
75
nj
b
ni
r
y
x
Fig. 2 Illustration of a rigid rod with and its configuration components (r, b) in R3
Any point in the rod can be located by the vector

v( ) = n j + (1 )ni,
where [0, 1]. Let ( ) 0 be a density function defined on the interval [0, 1]
which describe the mass density along the rod, that is
m=
1
0
( ) d > 0.
We describe the position of the rod by means of the configuration matrix

Q = r b R32
(2)
where r = v( ), [0, 1], is any fixed point in the rod. Whenever possible r will
be made to coincide with the center of mass of the rod. Any point in the rod can be
equivalently described as a linear function of the configuration vector
!
1
v( ) = Q
[ , (1 )].
,
Note that and are related by = + /. It is using that we can compute
higher order mass moments around r, the next two of which are
f ( ) = m1 ( ),
J( ) = m2 ( ),
mk ( ) :=
1
(1 )

( + /) k d ,
Such moments are associated with an important quantity, the kinetic energy of the
rod which is given by the formula
T=
=
1
2
(1 )

1
trace
2
( + /) v ( )T v ( ) d

(1 )

( + /)

!

1
1
T
TQ
, (3)
Q
Q
= trace J( ) Q
d
2
76
where J( ) is the positive semidefinite matrix of moments

!
m f ( )
0.
J( ) =
f ( ) J( )

Note that if we choose = 01 ( ) d so that r coincides with the center of

mass of the rod then f ( ) = 0. This leads to the well known decoupling of the
kinetic energy in translational and rotational components in a rigid body described
by its center of mass. One should choose to describe a rod by its center of mass
whenever possible, with the main exception being the case when constraints are
present in points of the rod other than the center of mass. We will illustrate this case
later in Sect. 4. In the present chapter we confine our discussion to the particular
case of uniform mass distributions (see [7] for a discussion on some nonuniform
mass distributions). For rods with mass uniformly distributed along the rod we have
( ) = m 1 . In this case the mass moments f and J are
1
f ( ) = m (1 2 ),
2

1
J( ) = m2 1 3 + 3 2
3
which are functions of , hence depends on the choice of the fixed point r. Indeed,
when the center of mass is the center of the bar, i.e. = 1/2, f and J are the familiar
f (1/2) = 0,
J(1/2) =
1
m2 .
12
Another familiar choice is when r coincides with one of the extreme points of
the rod, say r = ni ( = 0) so that
1
f (0) = m,
2
1
J(0) = m2 .
3
Interestingly, for any rod with uniform mass distribution, matrix J( ) is positive
definite, i.e. J( ) 0, regardless of (see [7]).
2.1 Nodes as Functions of the Configuration

In dynamics, the node vectors must be expressed as a function of the configuration,
in our case the matrix Q. One of the major advantages of our approach is that the
relationship between the configuration matrix and nodes is linear, as illustrated next.
Example 3.1. Consider the single rod pinned at one end and with three strings on
the other end as illustrated in Fig. 3. Let Q be as in (2). The end nodes of the rod can
be computed as
!
1
,
ni = Q
1 = 0,
2 = .
i
77
Fig. 3 A single rod with three strings
so that
!

11000
N = n1 n5 = Q
+ 0 0 n3 n4 n5 ,
0000
where nodes ni , i = 3, . . . , 5 are the attachment points of the three strings that are
not on the rod.
In general we should have
N = Q T + Y,
N, Y R3n ,
Rn2 .
(4)
where Rn2 and Y R3n are constant. The above expression is valid even
when more than one rod is considered (see Sect. 3).
2.2 String Forces

In structural tensegrity systems, forces on the rod are due to the elongation of strings
and ground reactions. See Sect. 4 for a treatment of gravitational forces. For simplicity, we assume that the strings are Hookean and that they are firmly attached to nodes
on the rods or on fixed space coordinates (see [7] for more details). That is, strings
are linear force elements with rest-length li0 and stiffness ki . The force vector of the
ith string is
)
0,
si < li0 ,
ti :=
i (si li0 )(si /si ), si li0 .
where si is a vector in the direction of the ith string. String vectors are linear functions of the nodes of the structure. Assembling a matrix of string vectors and nodes

N = n1 nn R3n ,
S = NCTS ,
S = s1 sm R3m ,
the vector nk denotes the kth node in the structure and matrix CS Rmn is the
the string connectivity matrix. The connectivity matrix is made of ones, zeros and
78
minus ones representing the string/rod connections within the structure (see [7] and
the next illustrative example). It follows that

F = f1 fn = TCS = NCTS CS
T = S ,
where we made use of the diagonal matrix which contains the force densities
i := max{0,
i (si li0)/si }
on its diagonal. The matrix F is the matrix of nodal forces.
2.3 Generalized Forces and Torques

Equations of motion must be written in terms of the configuration matrix, whereas
the forces in the previous section are expressed at the nodes. Hence, we need to
express forces in term of the configuration matrix coordinates. That is, to compute
generalized forces. Because of linearity of (4) the matrix of generalized forces is
computed as (see [7] for details)
FQ = (Q T + Y) CTS CS .
(5)
A closer look at (5) reveals that

FQ = fr fb ,
fr = fi ,
i=1
fb = i fi .
i=1
where fr is simply the sum of all forces applied to the rods and fb is related to the
sum of the torques on the rod [7].
2.4 Equations of Motion

In Sect. 4 we will deal with tensegrity structures in which some or all the rods may
have kinematic constraints. In such structures it may be advantageous to make r to
not coincide with the location of the center of mass of the rod. For this reason we
derive the equations of motion using Lagrangian methods, whose full potential we
explore in Sect. 4 (see [7] for an alternative Newtonian derivation). In this section
we consider the simpler case in which r is assumed to coincide with the center of
mass. This assumption will be removed later.
Consider the Lagrangian function
L = T V

J T
b b1 ,
2
(6)
where is the Lagrange multiplier responsible for enforcing the constraint that vector b must remain unitary (1) and where V is some appropriately defined potential
79
function. Assume that r coincides with the location of the center of mass of the rod,
i.e. f = 0. Following standard derivations as shown in [7] we arrive at the equations
of motion
m r = fr ,
J b = fb J b,
bT b 1 = 0.
(7)
where fr and fb are the vector of generalized forces acting on the rod written in the
coordinates q (see Sect. 2.3). We can avoid the explicit calculation of the Lagrange
multiplier by using the constraint (1). As shown in [7],
2
= (b/b)
+ J 1 bT fb /b2 .
Substituting on (7) produces the rotational equations of motion

2
b = J 1 I (bbT )/b2 fb (b/b)

b.
(8)
These equations are the basis for the dynamics of class 1 tensegrity systems to be
discussed in Sect. 3.
3 Class 1 Tensegrity Systems

A tensegrity system where no rigid rod is in contact with any other rod is said to
be a Class 1 Tensegrity System. Such systems are the ones often identified with the
concept of tensegrity as originally defined by Kenneth Snelson and Buckminister
Fuller [3, 6]. The equations of motion developed in the previous section can be
extended to cope with general class 1 tensegrity systems in a straightforward way.
Instead of presenting a lengthy derivation of the equations of motion for general
class 1 tensegrity systems, we shall limit ourselves to indicate what are the steps
needed to undertake such generalization based on the material presented so far.
Because in class 1 tensegrity systems no rods touch each other, there are no extra
constraints to be taken into consideration beyond the ones already dealt with in
Sect. 2. In fact, using the Lagrangian approach, all that is needed to derive equations
of motion for a class 1 tensegrity System with N rods is to define the combined
Lagrangian
L=
Lj
(9)
j=1
where each L j is a Lagrangian function, as in (6), written for each rod j = 1, . . . , N,

and following the procedure outlined in Sect. 2.4 for enforcing the individual rod
constraints and deriving the equations of motion. With that in mind define the configuration matrix

(10)
Q = R B R32N
80
where

R = r1 rN ,

B = b1 bN R3N .
(11)
Note that, in the absence of constraints, the (4) is still valid provided an appropriate
matrix Rn2N is constructed. Likewise, generalized forces are easily computed
using (5). A surprisingly compact matrix expression for the resulting equations of
motion is possible by combining (7) and (8) as in the following theorem. See [7] for
a detailed proof.
Theorem 3.1. Consider a Class 1 Tensegrity System with N rigid fixed length rods.
Define the configuration matrix (10)

Q = R B R32N
where the columns of R describe the center of mass of the N rods and the columns
of B describe the rod vectors. Let Rn2N and Y R3n be constant matrices that relate the n 2N nodes of the structure with the configuration matrix
through (4)
N = Q T + Y,
N, Y R3n ,
Rn2 .
The dynamics of such class 1 tensegrity system satisfy

+ Q ) M = FQ
(Q
(12)
where
M = diag[m1 , , mN , J1 , , JN ],
= diag[0, , 0, 1 , , N ].
The Lagrange multipliers i , i = 1, . . . , N are computed by
i = (b i /bi )2 + Ji1bTi fbi /bi 2
(13)
where fbi are columns of the matrix FB which is part of the matrix of generalized
forces

FQ = FR FB R32N
which is computed by (5)
FQ = [W (Q T + Y) CTS CS ]
where CS Rmn is the string connectivity matrix, and the columns of matrix

W = w1 w2 w2N
are external forces where wi acts on node ni .
81
Fig. 4 A class 1 tensegrity prism with three rods and 12 strings
The following example illustrates the use of the expressions introduced in the
theorem above.
Example 3.2. Consider the tensegrity prism depicted in Fig. 4. This structure has six
nodes, three rods and 12 strings. Let the node matrix be

N = n1 n2 n3 n4 n5 n6 ,
where each pair of nodes is a pair of bottom and top nodes on a rod. That is

1
1
B = 1
1 (n1 n2 ) 2 (n3 n4 ) 3 (n5 n6 ) .
Assuming that the mass m j of the jth rods is uniformly distributed then the center
of each rods is its center of mass
R=

1
n1 + n2 n3 + n4 n5 + n6 .
2

The nodes can be retrieved from the configuration matrix Q = R B through (4)
with
2 0 0 1 0 0
2 0 0 1 0 0
1
0 2 0 0 2 0
=
2 0 2 0 0 2 0
0 0 2 0 0 3
0 0 2 0 0 3
and, because of the uniform mass distribution and the choice of R, we have that
f j = 0 and
Jj =
1
m j 2j ,
12
j = {1, 2, 3}.
82
The string connectivity is
1 0 1 0 0 0
0 1 0 1 0 0
0 0 1 0 1 0
0 0 0 1 0 1
1 0 0 0 1 0
0 1 0 0 0 1
CS =
0 1 1 0 0 0 .
0 0 0 1 1 0
1 0 0 0 0 1
0 1 0 0 1 0
1 0 0 1 0 0
0 0 1 0 0 1
With this information one can write the equations of motion (12).
4 Class k Tensegrity Systems

So far we have focused on class 1 tensegrity systems, where the rigid rods are never
in contact. Other systems of rods and strings in which the rods may be in contact
are also tensegrity systems (more generally, a system of rigid bodies and tensile
elements). A tensegrity system in which as many as k 1 rigid rods are in contact
is a class k tensegrity system. See [7] for a detailed discussion. Dynamic models for
class k tensegrity systems are a lot more complicated than the ones we presented
for Class 1 tensegrity Systems. Yet, such systems still present relevant structure that
may help simplify the derivation and computation of the equations of motion. The
main difficulty arises now from the fact that rods may be in contact (through ball
joints) and kinematic constraints other than the preservation of length of the bar
vector have to be enforced on the equations of motion. These constraints will vary
from system to system, which makes difficult the obtention of a general formula.
Instead, we shall discuss perhaps the simplest class k tensegrity system, a class 2
cable as illustrated in Fig. 5. The intent is to illustrate the difficulties of such models
and the benefits that can be obtained by exploring the problem structure by using
the approach of the previous sections.
4.1 A Class 2 Tensegrity Cable Model

Our class 2 tensegrity cable is a chain system composed of N interconnected rigid
rods used to model the behavior of a long and heavy cable. We assume, as before
for simplicity, that the mass of each rigid rod is distributed uniformly. For reasons
that will become clear in the sequel, we choose the coordinate ri to be the near end
83
b2
b1
z
r1
r2
r3
rN1
bN1
rN
bN
Fig. 5 Illustration of a chain system of N interconnected rigid rods
of each rod, which implies fi = mi i /2, Ji = mi 2i /3, for i = 1, . . . , N. This setup
is illustrated in Fig. 5. The vector r1 describes the motion at the first node of the
first rod, that is the starting end of the cable. In this section we assume that the
acceleration of such node is to be prescribed. That is, we assume that r1 and all its
derivatives are known a priori and are independent of the motion of the cable. The
cable can then be though as being suspended by r1 , the problem of interest being
then to describe the dynamics of the remaining elements of the cable. We assume
that the cable is immersed in a constant gravitational field. This setup reflects, for
instance, a cable that is being towed by a massive boat in a marine application or by
an excitation system in the laboratory [11].
With the above discussion in mind we choose as configuration the matrix

Q = b1 bN .
(14)
Vector r1 is not present because it is assumed to be known a priori. This assumption
can be removed without difficulty. Clearly the challenge in modeling this class 2
tensegrity system is taking into account the constraints that the rods are connected
(by ball joints in our present model). Because of our choice of configuration Q,
such constraints are linear. Enforcing the constraint that the far and near ends of
successive rods are connect amounts at enforcing
ri+1 = ri + i bi ,
i = 1, , N 1.
which are clearly linear functions of Q. Furthermore, these equations can be solved
recursively yielding
i1
r i = r1 + j b j ,
i = 1, , N.
(15)
j=1
For this reason it will be possible to explicitly enforce these constraints when writing
the Lagrangian function, as opposed to using Lagrange multipliers. This is the route
taken for the rest of this session. In systems where loops (closed kinematic chain) are
84
present the situation is more complicated because it may not be possible to explicitly
compute a reduced configuration vector as done above.
In order to derive the equations of motion we assemble a combined Lagrangian
function as in (9) where each Li is given by (6). The terms of Li are computed by
adding the kinetic energy of each rigid rod, from (3),
1
1
Ti = mi r Ti r i + fi r Ti b i + Ji b Ti b i ,
2
2

T
'

(T
i1
i1
i1
1
1
r 1 + j b j + fi r 1 + j b j + Ji b i b i ,
= mi r 1 + j b j
2
2
j=1
j=1
j=1
where we made use of (15) explicitly, as anticipated. In this derivation we chose to
leave the expression Ti in terms of the individual vector components as opposed to
the configuration matrix Q. This will be reversed later.
In order to account for the constant gravitational field in which the cable is immersed, the potential energy of the ith rod is computed as
Vi = g
1
0
i ( )vi ( )T ez d
= mi g/i
i
0
i1
1
(ri + bi ) ez d = mi g r1 + j b j + i bi
2
j=1
T
T
ez , (16)
where ez R3 is a unit vector in the positive z-direction. The last two expressions
follow from the assumption of uniform mass distribution and the choice of configuration. The complete Lagrangian function is then assembled

N
Ji i T
(bi bi 1)
L = Ti Vi
2
i=1
and the equations of motion are derived by computing the EulerLagrange equation
which, for convenience, will have the left and right hand sides split as
M + r 1 hT = d L = L = FQ Q J,
Q
dt Q
Q
(17)
where
J = diag[J1 , . . . , JN ],
FQ :=
Vj
Q = ez gT .
j=1
As in class 1 tensegrity systems, the mass matrix M is a constant N N matrix, with

entries to be defined in (20). Vectors g and h are N-dimensional with entries to be
defined in (22). Compare how (17) is similar to (12). The main differences are that
M is now a full matrix, as opposed to diagonal, and, of course, we should expect that
85
the calculation of will be much more complex than (13). It is to the calculation of
that we now turn our attention.
In order to compute , using the properties of Kronecker1 products [2], let us
temporarily rearrange the equations of motion (17) in vector form
!
Y q
u
M
,
(18)
=
y
YT 0
where
q = vec(Q),
= M I3 ,
M
u = vec(U),
U := FQ r 1 hT = ez gT r 1hT
and
b T1 b 1
y = ... .
b T b N

Y = J1 e1 b1 JN eN bN ,
The last N equations of (18) come from differentiating the bar length constraint
bi 2 = 1 twice with respect to time. As in the case of Class 1 Tensegrity systems,
we can solve for explicitly as follows
1 Y)1 (YT M
1 u + y).
= (YT M
Online computation of is then dominated by the factorization of the N N positive
1 Y, because M (hence M)
is constant and can be factored
definite matrix YT M
offline. Evaluation of all terms involved in this computation can be further sped up
by exploring the structure of the involved matrices. Indeed, compute

V = U M1 = ez (M1 g)T r 1(M1 h)T ,
Z = QJ = J1 b1 JN bN
One can verify that
1 u = (M1 I3 )u = vec(v),
M
zT1 v1
.
1 u =
YT M
..
zTN vN
where vi (or zi ) denotes the ith column of matrix V (or Z). By denoting by 1 a vector
with all entries equal to 1 and the entrywise product of two matrices by the symbol
(known as the Hadamard product) we have that
1 u + y = (Z V)T 1 + y.
YT M
1 The Kronecker product of two n n matrices A and B, i.e. A B, is an n2 n2 matrix composed

of n n blocks of matrices of the type Ai j B.
86
Using similar ideas one can show that

1 Y = M1 (ZT Z).
YT M
The final result is summarized in the following theorem.
Theorem 3.2. Consider the Class 2 Tensegrity Systems Cable Model with N rigid
fixed length rods connected as in Figure 5, immersed in a constant gravitational
field in the ez direction. Let the trajectory of the node r1 be prescribed a priori,
independently of the rest of the system. Define the configuration matrix Q = B
R3N , as in (14), where the columns of B describe the rod vectors.
The dynamics of this class 2 tensegrity systems cable model satisfy
M + Q J = FQ r 1 hT ,
Q
(19)
where M = (Mi j ) is a constant N N matrix with entries
j ( fi + i k=i+1 mk ), i > j,
N
2
Mi j = Ji + i k=i+1 mk ,
i = j,
M ji ,
i < j.
(20)
The generalized force FQ is given by

N
FQ :=
Vj
Q = ez gT .
(21)
j=1
The entries of the N-dimensional vectors g = (gi ) and h = (hi ) are

N
N
1
mi + m j ,
hi := fi + i mk .
gi = g i
2
j=i+1
k=i+1
(22)
and J and are the diagonal matrices

J = diag[J1 , . . . , JN ],
= diag[1 , , N ].
The Lagrange multipliers i , i = 1, . . . , N are computed by

1

(Z V)T 1 + y ,
= M1 (ZT Z)
(23)
where the matrices V and Z are given by

V = ez (M1 g)T r 1 (M1 h)T ,
Z = QJ
(24)
and y = (yi ) is a N-dimensional vector with entries yi := b i 2 .

The next example illustrates the theorem in the case of a double pendulum.
87
Example 3.3. Let N = 2, m1 = m2 = 1 and 1 = 2 = . Hence for uniformly distributed mass rods f1 = f2 = /2 and J1 = J2 = 2 /3. Then
4 1
3 2
M = 2
,
1 1
2 3
3
2
g = g ,
1
2
3
2
h = .
1
2
5 Concluding Remarks
The dynamic models developed in this chapter, i.e. (12) and (19), have remarkable
structure. First they have compact matrix forms that facilitate computation, without the need to recourse to Kronecker products, as in traditional dynamic models.
Second, they are ordinary differential equations even though the systems are not described in minimal coordinates. By parametrizing the configuration matrix directly
in terms of the components of the rod vectors, the usual transcendental nonlinearities
involved with the use of angles, angular velocities and coordinate transformations
are avoided. Indeed, the absence of trigonometric functions in this formulation leads
to a simplicity in the analytical form of the dynamics. The actual number of degrees
of freedom for each rod is 5, whereas, the model (12) has as many equations as
required for 6 degrees of freedom for each rod. That is, the equations are a nonminimal realization of the dynamics. The mathematical structure of the equations
are simple, however. This allows much easier integration of structure and control
design, since the control variables (string force densities) appear linearly, and the
simple structure of the nonlinearities can be exploited when controlling the system.
The interested reader is referred to [7, 12] for further discussions on these topics.
The proposed structure also facilitates computation. Indeed, the differential equations (12) and (19) can be integrated using standard explicit methods, e.g. Runge
Kutta. In [4] we provide a detailed analysis in which the numerical stability of such
methods is investigated in the case of a single rod. In a nut shell, the conclusion
is that with a minor modification, namely the periodic normalization of the unitary
rod vectors in B, the numerical integration is very reliable and accurate. Indeed,
we have very successfully implement numerical code that can integrate tensegrity
systems with tens of thousands of rods. This code has been used, for instance to
integrate a Class 1 Tensegrity model of the dynamic behavior of the cytoskeleton of
erythrocytes or red-blood cells (see [5] for details).
References
1. Fuller, R.B.: US patent 3 063 521 Tensile Integrity Structures. United States Patent Office
(1959)
2. Horn, R.A., Johnson, C.R.: Matrix Analysis. Cambridge University Press, Cambridge, UK
(1985)
88
3. Lalvani, H.: Origins of tensegrity: views of Emmerich, Fuller and Snelson. International
Journal of Space Structures 11, 2755 (1996)
4. de Oliveira, M.C.: Dynamics of systems with rods. In: Proceedings of the 45th IEEE Conference on Decision and Control Conference, pp. 23262331. San Diego, CA (2006)
5. de Oliveira, M.C., Vera, C., Valdez, P., Sharma, Y., Skelton, R.E., Sung, L.A.: Network nanomechanics of the erythrocyte membrane skeleton in equibiaxial deformation. To appear in the
Biophysical Journal
6. Sadao, S.: Fuller on tensegrity. International Journal of Space Structures 11, 3742 (1996)
7. Skelton, R.E., de Oliveira, M.C.: Tensegrity Systems. Springer, Berlin (2009). ISBN: 9780387742410
8. Skelton, R.E., Pinaud, J.P., Mingori, D.L.: Dynamics of the shell class of tensegrity structures.
Journal of The Franklin Institute-Engineering And Applied Mathematics 338(2-3), 255320
(2001)
9. Snelson, K.: US patent 3 169 611 Continuous Tension Discontinous Compression Structures.
United States Patent Office (1965)
10. Snelson, K.: Letter to R. Motro. originally published in International Journal of Space Structures (1990). URL http://www.grunch.net/snelson/rmoto.html
11. Wroldsen, A.S.: Modelling and control of tensegrity structures. Ph.D. thesis, Department of
Marine Technology, Norwegian University of Science and Technology (2007)
12. Wroldsen, A.S., de Oliveira, M.C., Skelton, R.E.: Modelling and control of non-minimal nonlinear realisations of tensegrity systems. International Journal of Control 82(3), 389407
(2009). DOI {10.1080/00207170801953094}
Modeling a Complex Aero-Engine

Using Reduced Order Models
Xuewu Dai, Timofei Breikin, Hong Wang, Gennady Kulikov, and Valentin Arkov
1 Introduction
Gas turbine engines are widely used in many industrial applications and engine
condition monitoring is a vital issue for the aircraft in-service use and flight safety.
From the variety of condition monitoring methods, the model-based approach is
perhaps the most promising for real-time condition monitoring. This approach can
predict the engine characteristics at the expense of algorithmic redundancy and
requires real-time simulation. The main obstacles for using full thermodynamic
models in the engine condition monitoring schemes are high computing load, and
inability to incorporate unforeseen changes.
It is clear that in the on-board condition monitoring of gas turbine engines, a reduced order model is required due to the limited on-board computation resources.
This model should be individual for each particular engine. This type of models
cannot be obtained based only on full thermodynamic model of the engine. Identification of each individual engine model parameters based on experimental data is required. Simplification of the model, the model order reduction, introduces additional
X. Dai
School of Electronic and Information Engineering, Southwest University, Chongqing 400715,
China
and
He was with Control Systems Centre, School of of Electrical and Electronic Engineering,
University of Manchester, Manchester M60 1QD, UK
e-mail: dxw.dai@gmail.com
T. Breikin and H. Wang
Control Systems Centre, School of Electrical and Electronic Engineering,
University of Manchester, Manchester, UK, M60 1QD
e-mail: t.breikin@manchester.ac.uk; hong.wang@manchester.ac.uk
G. Kulikov and V. Arkov
Department of Automated Control Systems, Ufa State Aviation Technical University, K. Marx
Street 12, Ufa 450000, Russia
e-mail: kulikov@asu.ugatu.ac.ru; arkov@asu.ugatu.ac.ru
89
90
X. Dai et al.
problems during the model parameters identification from real engine data. Furthermore, in order to retain the fault information in the residual, the models long term
prediction (simulation) performance is of the main interest, rather than the one-step
a-head prediction. It is particularly important for detecting incipient faults. Although
Output Error (OE) model shows better simulation performance than other models
(e.g., ARX, ARMAX, etc.), the dependency within the output errors presents a number of challenges. This dependency makes LSE method biased leading to a poor
long-term prediction. As described literature, even if the model itself is linear, the
objective function of OE model is highly nonlinear and some kind of iterative nonlinear optimization is needed [15].
This chapter describes the modeling phase of the engine model-based condition
monitoring. The aim is to find a fast OE model identification algorithm suitable for
the limited on-board computation facilities of the on-board condition monitoring
system. The chapter is organized as follows. Section 2 provides an introduction to
Gas Turbine System. Section 3 discusses the performance criterion and structure
selection during data driven reduced order modeling. Section 4 defines the objective function for OE modeling and presents the development of DNLS (Dynamic
Nonlinear Least Squares) identification algorithm. In order to accelerate the identification speed, two techniques are employed. The first one is Iterative calculation
of the gradient (the Jacobian). The dependency within output errors is taken into
account by adding an weighted sum of past gradients to the current Jacobian matrix.
Secondly Hessian approximation is used. In order to further accelerate the convergence speed at a relatively low computation cost, the second order Hessian matrix
is used and approximated by the first order information. Finally, the potential of
the proposed DNLS for reduced order modeling of a gas turbine engine is illustrated in Sect. 5. Data gathered at an aero engine test-bed serves as the test vehicle
to demonstrate the improvements in convergence speed and the reduction of computation cost.
2 Gas Turbine System

A generalized schematic of a typical reheat bypass turbojet engine is shown in Fig. 1,
where the main units of the flowing part of the engine are as follows: I fan or low
pressure (LP) compressor; II high-pressure (HP) compressor; III bypass duct; IV
main combustion chamber; V mixing chamber; VI afterburner; VII variable
jet nozzle.
The possible factors shown in Fig. 1 that affects the engine operation are: 1
variable stator vanes of fan (V SV. f ); 2 air bypass from first stages of compressor
into bypass duct (m); 3 variable stator vanes of compressor (V SV.c ); 4 variable
radial clearance in final stages of compressor ( rc ); 5 fuel supply in combustion
chamber (W f ); 6 fuel supply in combustion chamber of bypass duct (W f .II ); 7
variable radial clearance in HP turbine ( rt.HP ); 8 variable stator vanes of fan turbine (V SV. f t ); 9 variable radial clearance in fan turbine ( r f t ); 10 mixing area
Modeling a Complex Aero-Engine Using Reduced Order Models
91
Fig. 1 Generalized schematic of reheat bypass twin-shaft turbo jet with variable working process
(Amix ); 11 system of reverse thrust (Frev ); 12 fuel supply in afterburner (W f .ab);

13 air supply in ejector nozzle (ne j ); 14 critical nozzle area (Acr ); 15 output
nozzle area (An ). The main control factors here are fuel flow into the main combustion chamber W f and afterburner W f .ab , critical nozzle area Acr and output nozzle
area An . In the following subsections nonlinear static and dynamic models of the engine are presented to demonstrate their complexity. Detailed description of the models parameters and variables as well as the nomenclature used can be found in [11].
2.1 Nonlinear Static Model of Gas Turbine

Under some assumptions [11], nonlinear equations describing operation of a gas
turbine engine can be presented in the following form

TH = 288 0.0065H
, if H < 11,000 m
pH = p0 (TH /T0 )5.25

TH = 216.5K
, if H > 11,000 m
pH = p11,000 exp( 11,000H
6,318 )
Tin
pin
nred
pc
= TH (1 + 0.2M 2)
2 3.5
= .
in pH (1 + 0.2M )
= n 288/TH
= pg /comb
c = pc /pin
Wa.red = f (c , nred
Wa = f (Wa.red , Tin , pin )
c = f (c , red )
Tc = f (c , c , Tin )
pt = pnoz /noz
t = pg /pt/
t.red = n/ 288/Tg
92
X. Dai et al.
Wg.red = f (t , nr.red )
Wg = f (Wg.red , Tg , pg )
t = f (t , nt.red )
Tt = f (Tg , t , t )
Wn = f (An , pn , Tn )
Nt = f (Wg , Tg , Tt )
Nc = f (Wa , Tin , Tc )
In addition, the following balance relationships, representing air/gas flow balance, thermal balance in the combustion chamber and jet nozzle and power balance
between the compressor and turbine, should be taken into account.
Wg (Wa + W f ) = 0
c pa TcWa + Hu gW f c pg TgWg = 0
Wn Wg = 0
c pg TcWg c pg TnWn = 0
mech Nt Nc = 0
2.2 Nonlinear Dynamic Model of Gas Turbine

For dynamic modeling, the system given in Sect. 2.1 should be complemented by
the differential equation describing dynamics of the shaft rotation and differential
equations accounting the accumulation of the gas energy and mass within major
volumes of the flowing part of the engine [11]:
Tg =
pg
RTg
Tg Tg + Vcomb (Wa + Wt Wg )
1
cv mc [(c pa Tc Wa + Hu gW f c pg Tt Wg ) cv Tg (Wa + Wt
pn =
pn
Tn
Tn =
cv mn [(c pg Tt Wg c pg Tn Wn ) cV Tn (Wg Wn)]

mech Nt Nc
Jn( /30)2
pg =
n =
Tn +
RTn
Vn
(Wg W n)
Wg )]
(1)
Thus, the dynamic model of a gas turbine engine can be presented in state space
form:
= Fx (X, U, V)
X
(2)
= Fy (X, U, V)
Y
where X is the state vector, U is control vector, V is vector of flight conditions,
Y is vector of output observable coordinates, Fx and Fy are nonlinear operators.
This model is a detailed Nonlinear Dynamic Model making it possible to study the
dynamic properties of a turbine engine.
93
The programs for gas turbine engine control are determined with respect to the
aircraft requirements. For example, a controller may use programs for maintenance
of maximum maneuverability of the aircraft, or for maximum efficiency, or specific
programs for take-off and landing. Each program is determined via optimization of
some criterion or a group of criteria. To implement model-based condition monitoring during these programs execution either very computationally extensive model
(which still would not account specific characteristics of the engine) should be employed or alternatively on-line identification of the simplified reduced order models
(specific for the engine and its operating conditions) can be used. The later one has
obvious advantageous accounting specific characteristics of an individual engine,
degradation of the engine parameters, and other uncertainties as external operating
conditions, fuel quality, etc.
3 Problem Formulation for Reduced Order Data

Driven Modeling
Although the detailed nonlinear dynamic model (1) is the best mathematical model
to represent the gas turbine engines, it involves nonlinear equations and it is impractical to implement it in on-board condition monitoring due to the limited computation resources [11]. In addition, the model should be individual for each particular
engine.
When the engine works around some operating point, the nozzle area and inlet
guide vanes are fixed and the engine is controlled by the fuel flow. In this context,
the reheat system is assumed inoperative, the compressor bleed valve closed. Due to
the very small time constants in the thermodynamic processes and relatively large
time constants in shaft speeds, the higher order nonlinear model can be reduced
to a state model with the states of shaft speeds [5]. Hence, for on-board condition
monitoring, the shaft speeds are regarded as the primary outputs, from which the
internal engine pressures and thrust can be calculated. Because the engine performance is linked with the shaft speeds closely, in condition monitoring, much attention is paid to modeling the dynamic relationship between these shaft speeds and
the fuel flow.
Furthermore, the reduced order model can be mathematically described by a set
of SISO (Single-Input/Single-Output) transfer functions relating the shaft speeds n
to fuel flow W f . In the discrete time domain, the corresponding discrete-time specification of such a SISO transfer function is the difference equation:
y (t) = f ( , y (t 1), y (t 2), . . .y (t n), u (t 1), . . . , u (t m)).
(3)
where u (t) and y (t) are the unmeasurable actual system input and output, respectively. And is the actual parameter vector associated with f ().
94
X. Dai et al.
Consider a system corrupted by input noise du (t) and output noise dy (t), as shown
in Fig. 2 where du (t) and dy (t) denote either the measurement noise or the external
disturbance, the measured input u(t) and output y(t) can be expressed as
u(t) = u (t) + du (t)
(4)
y(t) = y (t) + dy(t)
(5)
In the case of modeling aero engines, u(t) and y(t) denotes the fuel flow W f and
the shaft speeds n of the engine model (1), respectively. Note that function f ( ) is
a mathematical description of how the engines fuel flow and shaft speed variables
relate to each other.
A model is used to approximate the function f ( ) by f( ) and estimate the
future output according the observed data sequence [u(t), y(t)] up to time t. The
system identification is such a technique to build the function structure of f( ) and
estimate the parameters .
Not that the criterion on what is a good model is highly problem-dependent. It
is useful to define and clarify what is a good model for condition monitoring, and
select the right model structure for parameter identification. In terms of condition
monitoring, the most important objective is to find a good reduced order model
whose output error is robust to the disturbances du , dy and sensitive to the faults fa ,
fs . Here, du and dy denotes the input and output disturbances(noises), respectively.
fa denotes the actuator fault, and fs denotes the sensor faults. They can be seen
in Fig. 2.
3.1 Criterion Selection

In the initial stage, it was recognized that in the context of condition monitoring, the
long-term prediction performance is more of interest than the one-step-ahead prediction. It is particularly true for detecting incipient faults. This section discusses
why the long-term prediction performance is selected as the main criterion for condition monitoring.
Basically, there are two connection modes used for condition monitoring [8], [2]:
(1) parallel connection for long-term prediction (which corresponds to the Output Error model or Infinite Impulse Response filter); (2) parallel-series connection for one-step ahead prediction (which corresponds to the Equation Error
model).
In the parallel connection mode, as shown in Fig. 2a, the model runs in parallel
to the system (thereby, forming the so-called parallel model or IIR filter) and the
error is termed as long-term prediction error, because the prediction y(t)
is calculated from the previous input u alone and has no direct link to the previous system
output y. This error is also called simulation error [13], as the model is often used
to simulate the actual plant.
a
input
noise
du
u
parallel connection mode

output
noise
actuator
fault
sensor
fault
dy
fa
u*
Plant
f(u*, y*)
y = f (u)
Parallel model
du
y*
parallel-series connection mode

input
noise
fs
95
actuator
fault
fa
u*
output
noise
Plant
f(u*,y*)
fB (u)
dy
sensor
fault
fs
y*
fA (y)
r
y
Parallel-series model
Fig. 2 Comparison of long-step prediction and one-step-ahead prediction
Figure 2b shows a block diagram of the parallel-series connection mode, where

the model consists of two FIR (Finite Impulse Response) filters fA and fB . fA is the
series part running in series to the plant, and fB the parallel part running in parallel
to the plant. Compared to the parallel connection mode, the obvious difference is the
feeding back of actual system output y(t) to the model. Because y(t)
depends on the
last system outputs y(t 1), y(t 2), . . . up to time t 1, it is called one-step-ahead
prediction and the corresponding error is termed as one-step-ahead prediction error,
or simply prediction error [13].
Although the parallel-series model is able to represent the dynamic characteristics of the system to some extent, it is not an exact duplication of the plant, because
the plant output y(t i) is used to correct the model prediction y(t).
Indeed, the
parallel-series model is designed for fitting the output rather than simulating the
plant.
On the other hand, the parallel model repeats the dynamic behavior of the plant
in the same manner as the plant (that is to simulate the plant). Hence, it is easy to
understand that the one-step-ahead predictor may not give the best fault detection
performance, and the long-term prediction works better in terms of fault detection.
For illustrating this, both models are used to detect the actuator faults happening
at 10 s. As shown in Fig. 3a1, b1, the abrupt and incipient faults are simulated and
added to the input signal, respectively. Figure 3a2, b2 shows the residuals given by
the parallel-series model, where no obvious changes can be seen when faults happening. However, the magnitude changes can be easily detected from the residuals
of the parallel model, as shown in Fig. 3a3, b3. Moreover, the shapes of the residuals
also reflect the patten of faults, which may be useful for fault estimation
These phenomena can be interpreted as that, in the parallel-series connection, the
system output is fed into the model to correct the model prediction. When some fault
emerges in the system, the faulty information contained in the system output is also
passed to the model, and the model prediction is adjusted to match the faulty output.
Thus, the effects of faults are cancelled in the residual due to the minus operation.
Hence, it fails to detected the fault by checking the residual. The residuals in the
parallel connection, however, depend on the input alone and are unaffected by the
faulty output. Thus the fault information remains in the residual.
96
X. Dai et al.
(a1) abrupt fault
(b1) incipient fault
Fault fa
0.1
0.2
0.05
0.1
0
10
15
Residual
(a2) parallelseries mode

10
5
0
10
15
10
(a3) parallel mode
15
10
15
(b3) parallel mode
20
Residual
10
(b2) parallelseries mode
10
10
30
20
10
10
0
10
0
0
5
10
Time (second)
15
10
5
10
Time (second)
15
Fig. 3 Residuals of long-term prediction and one-step ahead prediction
Therefore, in the context of condition monitoring, the parallel model is preferred.

As a consequence, the long-term prediction performance of the parallel model is
selected as the main criterion in this study.
3.2 Model Selection: EE vs. OE

From the viewpoint of system identification [2], there are two basic error concepts:
(1) output error and (2) equation error. Figure 4 gives an interpretation of these
concepts in terms of a block diagram.

(b) Equation Error
input actuator
output sensor
fault
noise
noise fault
du
fa
dy
fs
Plant
y
u* f(u*,y*) y*
u
(a) Output Error

input actuator
output sensor
noise
fault
noise fault
dy
fs
fa
du
Plant
y
u
u* f(u*, y*) y*
B(z)
F(z)
1-A(z)
r
Output
Error
y OE = B(z) u
F(z)
97
B(z)
yEE = [1-A(z)]y + B(z)u
r
Equation
Error
Fig. 4 Definitions of output error and equation error
3.2.1 Equation Error (EE) Model

The most simple way to describe the inputoutput relation f () is to represent it as
a difference equation in the discrete time domain:
y(t) = T (t) + e(t)
(6)
= [a1 . . . ana b1 . . . bnb ]T
(7)
(t) = [y(t 1) . . . y(t na ) u(t 1) . . . u(t nb )]T
(8)
where
is a parameter vector, and
an observation vector at time t, and e(t) the noise term acting as a direct error in the
difference equation. The model (6) is often called as equation error model.
By introducing the backward shift operator z1 , the model prediction can be
written in transfer function form:
where
yEE (t) = [1 A(z)]y(t) + B(z)u(t)
(9)
A(z) = 1 + a1 z1 + . . . + ana zna
(10)
B(z) = b1 z1 + . . . + anb znb .
(11)
and
Alternatively, it can be written in the vector form

yEE (t) = T (t)
(12)
And the quantities

rEE (t) = y(t) yEE (t)
= y(t) T (t)
are called equation errors.
(13)
98
X. Dai et al.
Recalling the parallel-series model (see Fig. 2), one can see that the EE model in
fact is some kind of parallel-series model. 1 A(z) corresponds to the series part fA ,
B(z) (11) is the counterpart of f B and yEE (t) in (9) is the one-step ahead prediction.
The model (6) is also called an ARX (Auto-Regressive eXogeneous) model, where
AR refers to the autoregressive part A(z)y(t), X to the extra input B(z)u(t), and (t)
to the regression vector at time t.
3.2.2 Output Error (OE) Model

The EE model separates the relation between the input and output into two transfer
functions 1 A(z) (relating y to y) and B(z) (relating y to u). From a physical point of
view, it may seem more natural to present f () in one transfer function. An immediate
way of presenting a transfer function is to parameterize f () as a rational function.
Thus the model is given as
yOE (t) =
B(z)
u(t)
F(z)
(14)
with
B(z) = b1 z1 + . . . + anb znb
(15)
F(z) = 1 + f1 z1 + . . . + fn f zn f
(16)
and
The parameter vector to be identified is
= [ f1 f2 . . . fn f b1 ...bnb ]T
(17)
(i =
By multiplying both sides of (14) with F(z) and moving terms fi zi y(t),
1, . . . , n f ) to the right side, the prediction yOE (t) can be rewritten in the vector form:
yOE (t) = T (t)
(18)
with a pseudo-regression vector [13]
(t) = [y(t
1) . . . y(t
n f ) u(t 1) . . . u(t nb)]T
(19)
The prediction error of (18)

rOE (t) = y(t) yOE (t)
= y(t) T (t)
(20)
is called Output Error.

It is worth noting that the pseudo-regression vector (19) in OE model differs from
the regressor (8) in EE model. Compared to y(t j), j = 1, 2, . . . na in the regressor
99
(8) (which are the measured plant outputs), y(t

i) in the pseudo-regression vector
(19) are not observed. Instead, y(t
i) are the previous model predictions.
One can find that the model connections are closely linked with the error concepts. The parallel connection corresponds to the output error, and the parallel-series
connection is associated with the equation error.
As discussed earlier, a good model for condition monitoring is the parallel model,
and a good criterion is the long-term prediction performance. It follows that OE
model is better than EE model in terms of condition monitoring, and Sect. 4 will
solve the problem of identifying the parameters of OE model.
4 NLS for OE Parameter Identification

The least-squares estimate (LSE) algorithm has been proved to be very effective
for the ARX model identification. The objective of LSE is to minimize the mean
squared equation error (MSEE):
VEE ( ) =
1 N 1
2 [y(t) T (t) ]2
N t=1
(21)
Here, VEE ( ) is a quadratic function of the parameter . It has been proved in literatures that the LSE can be successfully applied to the ARX model and gives the
best parameters estimation in terms of minimizing VEE ( ). However, the estimates
of {ai } given by LSE may be biased if the residuals are correlated ([2, 12], p. 207,
p. 256 in [7, 13]).
Due to the dependence within the residual, the LSE method failing to give unbiased parameter estimates results in a relatively poor long-term prediction performance. It is of interest to note that: the biased parameter estimates benefit to the
smaller one-step prediction errors. This agrees with the point that the aim of LSE is
to minimize the one-step-ahead prediction errors (21).
On the other hand, the objective of identifying OE model is to minimize the mean
squared output error (MSOE):
VOE ( ) =
1 N 1
2 [y(t) T (t) ]2
N t=1
(22)
Note that, although they have the similar form in appearance, VOE ( ) differs VEE ( )
a lot due to the difference between (t) (8) and (t) (19). Hence, (21) is a quadratic
function, but (22) is highly nonlinear with respect to the parameter [1, 14].
Because of the high nonlinearity of OE objective function, the OE model identification is not an easy task, even if the model is linear. Some iterative optimization
algorithm is inevitable, and the identification algorithm is generally expressed as
k+1 = k [R(k)]1
V (k )
(23)
100
X. Dai et al.
where the subscript k gives the number of the iteration, is a series of positive
scalars (in term of step size) tending to zero or a small value, and R(k) a n-by-n (n =
dim ) positive definite matrix to modify the search direction V(k ) . Here, V(k )
denotes the derivatives of V ( ) with respect to at kth iteration.
The following sections will give the technical details of the proposed fast identification algorithm. For ease of notation, the subscript OE is omitted and the terms
r(t), V ( ) will denote the output error rOE (t), the OE objective function VOE ( ),
respectively, in the following sections otherwise specified.
4.1 Calculation of V ( k )/ and the Jacobian

It is important to keep in mind that (t) in (22) involves model prediction y(t
i),
i = 1 n f , and y(t
i) is a function of parameters . It yields that (t) depends
on too. For explicitly expressing the links between y and , (t, k ) is adopted to
denote (t) at kth optimization step:
(t, k ) = [y(t
1, k )...y(t
n f , k ) u(t 1)...u(t nb)]T
(24)
Let gk (t) denote the individual local gradient information at time t during the kth
iteration, that is
r(t)
,
(25)
(y(t) y(t,
k ))
(y(t,
k ))
=
.
(26)
(y(t,
k )) ( T (t, k ) k )
=
(t, k )
= ( (t, k ) +
k )),
(27)
gk (t) =
followed by
gk (t) =
Since
it follows that
gk (t) = [ (t, k ) +
(t, k )
k ].
According to the definition of (t, k ) and (26),
(t,k )
(28)
is
y(t
n f , k ) u(t 1)
(t, k ) y(t
1, k )
u(t nb )
=
= [gk (t 1) gk (t n f ) 0 0]
(29)
Substituting (29) into (28) gives

gk (t) = (t, k ) [gk (t 1) gk (t n f ) 0 0] k
101
(30)
Equation (30) indicates that gk (t) at time instant t not only depends on current observation vector (t, ) but also depends on the previous gradients {gk ( )}, ( < t).
Therefore, the iterative calculation of gk (t) (30) represents the dependency within
prediction errors {r(t)}.
The gradient on all the observed data is calculated by summing all values of local
gradients.
N
V (k ) 1 N r2 (t)
=
= r(t) gk (t)
2 t=1
t=1
(31)
Compared with ARX model (where the gradient is (t) alone), the calculation of
gradient as (30) differs at the additional sum terms [gk (t 1)...gk (t n f ) 0...0] k .
The added terms enable the derivative of V ( ) reflect the actual gradient of objective
function (22) correctly.
Note that gk (t) is calculated at every time t. As a result, the Jacobian of objective function (22) is achieved simply by transforming the sequence {gk (t)} into a
matrix form:
J = [gk (1) gk (2) gk (N)]T
(32)
where J is N-by-(n f + nb ) matrix.

Let a column vector to denote the long-term prediction error sequence =
[r(1) r(2) r(N)]T , then (31) can be rewritten as
V (k )
= JT
(33)
4.2 Approximation of R(k) and the Hessian

As shown in many literatures, for minimizing a nonlinear function, it is may be
inefficient to use the first-order alone as the search direction [6, 10]. In most
cases, the inverse of gradient does not point to the minimum point straightaway.
It is particularly true when the objective function surface has a valley, e.g., the
Rosenbrock function. Thus a matrix R(k) is employed to adjust the optimization direction from V ( ) . With the aid of Taylor expansion, it has been shown
that using a quadratic model to approximate the high nonlinear function (22)
is beneficial for improving the optimization convergence. A good selection of
R(k) in (23) is the Hessian matrix [10]. However, the calculation of the Hessian matrix is computation-consuming. In order to reduce the computation costs
for the on-board application, the Hessian needs to be approximated properly at a
lower cost.
102
X. Dai et al.
One of the solution is to make use of the Jacobian to approximate the Hessian.
Consider the second-order Taylor expansion of function (22) around some parameter

!T
V ( )
V ( ) V ( ) +
( )
=
1
+ ( )T H( )( )
(34)
2
where H( ) is the Hessian matrix evaluated at . Note that this approximation is
valid when is in the neighborhood of .

For compact notation, V ( ) = is replaced by V( ) . From expansion (34),

the condition of minimum of V ( ) can be expressed as
V ( )
+ H( )( ) 0
(35)
followed by
V ( )
(36)
Comparing (36) with (23), one can easily find that the search direction should be
modified by the inverse Hessian matrix. It follows that the R(k) can be replaced by
the Hessian matrix.
The problem turns into an approximation of the Hessian matrix. According to
Gauss-Newton methods, an easy way to form Hessian estimate is to make use of the
first derivative information, as shown below.
H( )1
2V ( )
T
N
2 (r(t))
=
T
t=1
H( )
r(t) r(t)
2 r(t)
=
+ r(t)
T
T
t=1
N
= JT J + r(t)
(37)
2 r(t)
T
It is therefore convenient to ignore the second term on the right-hand side and
approximate H( ) by JT J. It follows that
R(k) = JT J
(38)
In terms of the identification convergence, this approximation approach potentially offer the best trade-off of two worlds: First, by using the Hessian matrix as
R(k) (23), the search direction is modified from inverse gradient direction to point
103
toward to the minimum more straightaway. Secondly, it only needs to compute the
first-order derivatives that have been available in the computation of local gradients
gk (t). Thus, the calculation of the Hessian does not introduce too much more extra
computation burden, and results a better parameter updating direction.
As a result, the improved NLS algorithm for reduced order OE model is named
dynamic nonlinear least squares (DNLS) and given as follows:
Dynamic Nonlinear Least Squares Algorithm

Let k = 0, gk (1) . . . gk (n f ) = 0, and set the initial values of k according
to a priori knowledge
(2) At each time instant t, (t = n f + 1, . . . , N), compute the local gradient
gk (t) by gk (t) = [ (t, k ) [gk (t 1) gk (t n f ) 0 . . . 0] k ]
(3) At time instant N, rearrange {gk (t)} to form the Jacobian J =
[gk (1) gk (2) . . . gk (N)]T
(4) Update the parameter by k+1 = k (JT J)1 JT , where =
[r(1) r(2) r(N)]T and is a fixed or adjustable step-size.
5) Stop condition check: check whether the maximum iteration has been
achieved or the updating of k has been very small. If not, set k = k + 1 and
go back Step 2.
(1)
Remark 4.1. Since the Hessian matrix is approximated by JT J, the approximation

is only accurate enough when the parameters is close enough to the minimum,
as shown in the second-order Taylor expansion (34). In this case, the step size
is equal to 1. However, in the earlier stage of search where may be far from the
minimum, is not equal to 1 and a line search method is adopted to select the
optimal value of in each search iteration.
Remark 4.2. The dependency within long-term prediction errors is solved by computing the gradient in an iterative manner, as shown in (30), leading to a better
gradient calculation.
Remark 4.3. In the proposed DNLS, the objective function is approximated by a
quadratic function that is more accurate than the first-order approximation, thus the
search direction is better than those common deepest descent methods. Furthermore,
the Jacobian J is just a rearrangement of sequence of gradients g(t) (t = 1 N)
without additional computation and the approximation of the Hessian does not need
more computation than common deepest descent methods. Therefore, one of the
benefits is that the improvement on performance is only paid by a small increase of
computation expense, which make this method suitable for on-board modeling and
condition monitoring.
104
X. Dai et al.
5 Application and Results

In this section, the proposed DNLS algorithm is employed to identify the parameters of the reduced order OE model of a two shaft gas turbine engine. Real engine
data gathered from normal engine operation at the engine test-bed are used [3].
In the duration of this test, the angle of the VGVs (Variable Guide Vanes) of the
low pressure compressor and the reheat nozzle area were fixed to their low speed
positions, the engine fuel flow W f (t) is the control input and the high pressure shaft
speed NHP (t) is the primary output.
Because the engine runs at the low speed operating point, the input/output data
are first preprocessed by subtracting the physical equilibrium. Therefore, the data
used indeed for dynamic modeling are W f (t) and NHP (t). For simplicity of notation, let u(t), y(t) denote the input W f (t) and output NHP (t), respectively. 1,500
data pairs were collected in total. The first 750 pairs compose the training data set
for parameters estimation and the remaining 750 pairs make up the validation data
set for model validation.
Figure 5 illustrates the distribution of the training data set and validation data set,
where * represents the training data and the test data. It is obvious that these
two data sets cover a little different dynamics of the gas turbine, although they are
overlapped around the center which represents the current operating point.
Distribution of training data set and validation data set
30
* training data
o validation data
20
10
y(t)
10
20
30
40
40
20
0
u(t1)
Fig. 5 Scatter plot of the output y(t) versus the input u(t 1)
20
40
105
The aim of dynamic modeling is to obtain a reduced order model whose longterm prediction errors are minimized. To measure the algorithm performances, two
criteria are examined: (1) prediction accuracy and (2) computation costs. The prediction accuracy is measured by MSOE (Mean Square Output Error)
MSOE =
1 N 1
2 [y(t) yOE (t)]2
N t=1
(39)
where N = 750 is the length of the training or validation data set. Note that some
methods (e.g., LSE) are for minimizing one-step ahead prediction errors. For such
algorithms, the MSEE (Mean Square Equation Error) on training data set is used for
comparison.
1 N 1
MSEE = [y(t) yEE (t)]2 .
(40)
N t=1 2
The standard deviation of the errors is also adopted
0
1 N
ST D =
(r(t) E(r(t))2
N 1 t=1
(41)
where r(t) is either the OE errors or EE errors, E denotes the mathematical

expectation.
The computation cost is measured by how many evaluations of the objective
function are carried out during the identification, because the main computation
burden in most iterative optimization algorithms is the calculation of the long-term
prediction sequence yOE (t),t = 1 N.
5.1 First-Order Model

Here, the first order model of the gas turbine engine is considered, where (t) =
[y(t
1) u(t 1)]T and = [ f1 b1 ]T . In these experiments, although the model is
a linear model, the objective function (22) is a high order nonlinear surface with a
narrow valley around the minimum point as shown in Fig. 6.
For comparison purposes, five different algorithms are presented here. The LSE
and ARX approaches provided by System Identification toolbox do not contain iterative search, hence their computation costs are set 1. All the rest algorithms are
iterative search methods. The standard OE method and the RIV (Refined Instrumental Variable) [16,17] are also included for comparison. A fuzzy neural networks
ANFIS (Adaptive Networks-based Fuzzy Inference System) [9] is also used. The
results show that such neural networks are not suitable for long-term prediction.
The exhaustive search method examines each possible parameter value in order
to find the possible global solution. In our experiments, it runs in two main steps
and the total number of objective function evaluation is 12000.
106
X. Dai et al.
Mean Squared Output Error (MSOE)
Objective Function of Longterm Prediction

400
350
300
250
200
150
100
50
0.8
0.6
0.4
0.2
b(101)
0.3
0.4
0.5
0.6
0.7
0.8
0.9
Fig. 6 Objective function of the first-order model
DGD Search Route (approximating the Hessian by BFGS)
MSOE
0.7
0.6
b(10 1)
b 400
300
200
100
0
0
0.5
0.4
Step Size
0.3
0.2
0.1
0.4
Trainning Error of DGD (using BFGS)
0.6
0.7
0.8
0.9
20
30
40 50 60
Iteration
70
80
90 100
Trainning Step Size of DGD (using BFGS)
0.4
0.3
0.2
0.1
0
0.5
10
10
20
30
40 50 60
Iteration
70
80
90 100
Fig. 7 The DGD algorithm: (a) Searching route of DGD, (b) Training errors and (c) step size
In all the gradient-based iterative algorithms, [ f1 , b1 ] = [0.5 0.05] is set as the

default initial values of parameters and the initial step size is 0.1. The gradient
search method uses the inverse gradient as search direction and set R(k) in (23)
as an identity matrix. In DGD (Dynamic Gradient Descent) approach [4], the BFGS
(BroydenFletcherGoldfardShanno) [13], [6] are adopted to modify the search
direction from inverse gradient direction. The detailed search progress of DGD
is presented in Fig. 7. The DGDalgorithm needs about 100 objective function
107
Search route of DNLS1

0.8
b 150
MSOE
0.7
b (101)
0.6
c
Step Size
0.4
0.3
0.2
0.4
100
50
0
1
0.5
0.5
0.6
0.7
a
0.8
0.9
Training Residuals of DNLS
4
5
Search Step
Training Stepsize of DNLS
0.8
0.6
0.4
0.2
0
4
5
Search Step
Fig. 8 The DNLS algorithm: (a) Searching route, (b) training errors and (c) training stepsize
evaluations to approach to the optimum point, and a distance from the minimum
still can be seen. One reason is that the search direction does not point to the minimum at the beginning stage. Further reason is that the search route continually
zigzags from one side of the valley to another after entering the valley as shown
in Fig. 7.
In the proposed DNLS approaches, because of line search, two different stop conditions are tested in DNLS1 and DNLS2, respectively. The search process of DNLS1
is shown in Fig. 8. Apparently, the solution point given by the DNLS is closer to the
minimum and the convergence speed is accelerated, as depicted in Fig. 8. The search
direction of DNLS points a more straightforward way to the minimum. Even in the
valley, it still looks better. More clearly, from Fig. 8b, c, it can be seen that only 7
steps are involved to arrive the optimum point and the value of objective function
drops dramatically and steadily. Compared to DGD, this improvement is benefited
from a better Hessian approximation in (38). Figure 9 shows the DNLS results of
long-term prediction on validation data with MSOE of 9.131786.
The comparison of different methods is shown in Table 1 (on training data) and
Table 2 (on validation data). It can be concluded from these tables that: (1) In terms
of one-step-ahead prediction (equation error), LSE, ANFIS, ARX achieved slightly
smaller equation errors than those methods for long-term prediction. However, LSE,
ANFIS, ARX failed on long-term prediction on both training data set and validation
data set. (2) In terms of long-term prediction (output error), the methods OE, RIV,
exhaustive search, Gradient Descent, DGD and DNLS obtain similar results. Their
performances on long-term prediction are better than those methods designed for
one-step-prediction. (3) In terms of computation costs, the DNLS is advantageous
over the exhaustive search, gradient descent, DGD approaches. The main contribution of DNLS is the reduction of computation costs that makes this approach better
for on-board identification.
One thing interesting is that the neural network ANFIS achieves the smallest
MSEE 0.18728 on training data, but it gives the worst MSOE performance on
both training data and validation data. Although static feed-forward neural networks
108
X. Dai et al.
a
40
y(t) (solid) and model preditcion y(t) (dashed)
NHP
20
0
20
40
750 800
Residuals
10
900
1000
1100
1200
1300
1400
Sampling time
Long-term prediction error r(t) on validation data
900
1000
1500
5
0
5
10
750 800
1100
1200
Sampling time
1300
1400
1500
Fig. 9 Long-term prediction of DNLS for the first-order model on the validation data
Table 1 Comparison on training data set (first order model)
Methods
Computation
Costs
LSE
ARX
OE
RIV
ANFIS
Exhaustive search
Gradient descenta
DGDb
DNLS1c
DNLS2c
1
1
N/A
N/A
250
12,000
1,000
101
59
80
1-Step-ahead prediction
Long-term prediction
STD
MSEE
STD
MSOE
0.46133
0.46169
0.54868
0.54713
0.43276
0.48241
0.48270
0.48239
0.48238
0.48237
0.21284
0.21322
0.30120
0.29965
0.18728
0.23280
0.23312
0.23281
0.23280
0.23280
4.4273
5.1073
2.8465
2.7479
5.2745
2.7470
2.7471
2.7470
2.7470
2.7470
19.6359
26.1860
8.12804
7.6628
27.8212
7.63754
7.63788
7.63754
7.63755
7.63754
a The gradient search approach uses inverse gradient as search direction b The DGD approach uses
BFGS algorithm to adjust search direction c DNLS1 and DNLS2 stops after 59 and 80 objective
function evaluations respectively because of different stop conditions
(such as ANFIS) have good ability in approximating static mathematical functions,

they may be not suitable for dynamic system simulation (in terms of long-term prediction). The smallest MSEE given by the ANFIS on the training data set may be
caused by the overfit during the training stage. Hence, the ANFIS interpolates the
training data set very well, but it fails to extrapolate the validation data.
109
Table 2 Comparison on validation data set (first order

model)
Methods
STD
MSOE
LSE
ARX
OE
RIV
ANFIS
Exhaustive search
Gradient descent
DGD
DNLS1
DNLS2
3.
4.
2.
3.
3.
3.
8.
3.
3.
3.
7796
0895
9301
2490
0258103
1807
1806
1807
1811
1807
16.
19.
10.
12.
1.
11.
11.
11.
11.
11.
8985
5996
2658
2963
1908107
8369
8311
8369
8398
8368
y(t) (solid) and model preditcion y(t)

(dashed)
40
NHP
20
0
20
40
750 800
Residuals
10
900
1000
1100
1200
1300
1400
1500
Sampling time
Long-term prediciton error r(t)on the validation data det
900
1000
5
0
5
750 800
1100
1200
Sampling time
1300
1400
1500
Fig. 10 Long-term prediction of DNLS for the second-order model on the validation data
5.2 Second-Order Model

The second-order model also is identified to test the ability of DNLS, where (t) =
[y(t
1) y(t
2) u(t 1)u(t 2)]T and = [ f1 f2 b1 b2 ]T . The MSE drops to
3.31867176 and the computation cost is 60, as shown in Fig. 10. The parameters of
such a model are = [1.8604765 0.8641699 0.070453335 0.0074745994]T .
The comparison against other approaches is listed in Table 3. It can be seen that
the computation cost of the proposed DNLS is just 103.
110
X. Dai et al.
Table 3 Comparison for second order model

on Training data
Computation
Methods
Costs
STD
MSOE
LSE
ARX
OE
RIV
Exhaustive search
Gradient descent
DGD
DNLS
a
1
1
N/A
N/A
106
20,000
1,000
103
3.41291
3.93632
1.32140
1.25340
1.25624
2.30235
1.25625
1.25624
1.16613a
1.55375a
1.89871
1.81345
1.78773
5.39615
1.78773
1.78773
on Validation data
STD
MSOE
2.87529
3.05683
1.85980
1.97333
1.95562
2.74971
1.95560
1.95563
10.3296
11.5071
4.76809
5.19088
5.16392
9.22707
5.16380
5.16391
These two are MSEEs, because LSE, ARX are for EE model only
6 Summary
In this chapter, by a comparative study, the OE model is selected as the model for
modeling gas turbine engines. The discussion on parallel/serial model connection
and the corresponding equationg/output error shed light on the way to unbiased
parameter identification. It has been proved that LSE gives biased parameter estimation in the context of reduced order modeling. In the identification algorithm
design, the nonlinear least-squares method is used and problem of correlated prediction errors in the OE model is resolved by calculating the gradient in an iterative
manner. In order to accelerate the identification speed and meet the limitation on the
on-board computation resources, the second-order Hessian matrix is approximated
by the first-order Jacobian matrix.
The main contribution of this study is to propose an iterative calculation of the
gradient to deal with the correlated prediction errors. Furthermore, by approximating the Hessian with the the Jacobian, the identification speed is accelerated with a
minor increase of computation costs. The experiment results on modeling a gas turbine engine have shown that the proposed approach achieves a faster convergence
speed at relatively lower computational costs.
Acknowledgement This work was supported by the UK Leverhulme Trust (F/00 120/BC) and the
China National Science Foundation under Grants (60828007 and 60974029), 973 National Basic
Research Program of China under Grant (2009CB320601), the 111 project under (B08015) and
the Fundamental Research Funds for the Central Universities (XDJK2009C024).
References
1. B. Anderson and C. Johnson, Jr. On reduced-order adaptive output error identification and
adaptive IIR filtering. IEEE Transactions on Automatic Control, 27(4):927933, 1982
2. K. J. Astrom and P. Eykhoff. System identification a survey. Automatica, 7(2):123162,
March 1971
111
3. T.V. Breikin, G. G. Kulikov, V. Y. Arkov, and Peter. J. Fleming. Dynamic modelling for
condition monitoring of gas turbines: Genetic algorithms approach. In Proceedings of 16th
IFAC World Congress, 2005
4. X. Dai, T. Breikin, and H. Wang. An algorithm for identification of reduced-order dynamic
models of gas turbines. In Proceedings of 1st International Conference on Innovative Computing, Information and Control, volume 1, pages 134137, 2006
5. C. Evans. Testing and modelling aircraft gas turbines: An introduction and overview. In
UKACC International Conference on Control98, pages 13611366, 1998
6. R. Fletcher. A new approach to variable metric algorithms. The Computer Journal, 13:
317322, 1970
7. M. Hong, T. Soderstrom, and W. X. Zheng. A simplified form of the bias-eliminating least
squares method for errors-in-variables identification. IEEE Transactions on Automatic Control, 52(9):17541756, September 2007
8. R. Isermann. Model-based fault detection and diagnosis status and applications. In Proceeding of 16th IFAC Symposium on Automatic Control in Aerospace, 2004
9. J.-S.R. Jang, C.-T. Sun, and E. Mizutani. Neuro-Fuzzy and Soft Computing. Prentice-Hall,
Englewood Cliffs, 1997
10. D. Kahaner, C. B. Moler, and S. Nash. Numerical Methods and Software. Prentice-Hall,
Englewood Cliffs, 1989
11. G. G. Kulikov and H. A. Thompson. Dynamic Modelling of Gas Trubines: Identification,
Simulation, Condition Monitoring and Optimal Control. Springer, London, 2004
12. I. Landau. Unbiased recursive identification using model reference adaptive techniques. IEEE
Transactions on Automatic Control, 21(2):194202, April 1976
13. L. Ljung. System Identification: Theory for the User. Prentice Hall, London, 2nd edition,
1999
14. S. L. Netto, P. S. R. Diniz, and P. Agathoklis. Adaptive IIR filtering algorithms for system
identification: A general framework. IEEE Transactions on Education, 38(1):5466, February
1995
15. A. Wills and B. Ninness. On gradient-based search for multivariable system estimates. IEEE
Transactions on Automatic Control, 53:298306, 2008
16. P. C. Young. Recursive Estimation and Time-Series Analysis. Springer, Berlin, 1984
17. P. C. Young, H. Garnier, and M. Gilson. An optimal instrumental variable approach for identifying hybrid Box-Jenkins models. In 14th IFAC Symposium on System Identification SYSID06,
pages 225230, Newcastle, NSW, Australia, 2006
Part II
Large-Scale Systems Control

and Applications
Robust Control of Large-Scale

Systems: Efficient Selection
of Inputs and Outputs
Javad Lavaei, Somayeh Sojoudi, and Amir G. Aghdam
1 Introduction
There has been a growing interest in recent years in robust control of systems with
parametric uncertainty [4,9,13,16,18]. The dynamic behavior of this type of systems
is typically governed by a set of differential equations whose coefficients belong to
fairly-known uncertainty regions. Although there are several methods to capture
the uncertain nature of a real-world system (e.g., by modeling it as a structured or
unstructured uncertainty [6]), it turns out that the most realistic means of describing
uncertainty is to parameterize it and then specify its domain of variation.
Robust stability is an important requirement in the control of a system with parametric uncertainty. This problem has been extensively studied in the case of linear
time-invariant (LTI) control systems with specific types of uncertainty regions. for
instance, sum-of-squares (SOS) relaxations are numerically efficient techniques introduced in [18] and [4] for checking the robust stability of polynomially uncertain
systems. Moreover, a necessary and sufficient condition is proposed in [13] for the
robust stability verification of this class of uncertain systems, by solving a hierarchy
of semi-definite programming (SDP) problems.
The concepts of controllability and observability were introduced in the literature,
and it was shown that they play a key role in various feedback control analysis and
design problems such as model reduction, optimal control, state estimation, etc.
[6]. Several techniques are provided in the literature to verify the controllability
J. Lavaei
Control and Dynamical Systems, California Institute of Technology, Pasadena, USA
e-mail: lavaei@cds.caltech.edu
S. Sojoudi
Control and Dynamical Systems, California Institute of Technology, Pasadena, USA
e-mail: sojoudi@cds.caltech.edu
A.G. Aghdam
Department of Electrical and Computer Engineering, Concordia University, Montreal, Canada
e-mail: aghdam@ece.concordia.ca
115
116
J. Lavaei et al.
and/or observability of a system. However, in many applications it is important to

know how much controllable or observable a system is. Gramian matrices were
introduced to address this issue by providing a quantitative measure for controllability and observability [6]. While these notions were originally introduced for fixed
known systems, they have been investigated thoroughly in the past two decades for
the case of uncertain systems [21, 22, 27].
On the other hand, real-world systems are often composed of multiple interacting
components, and hence possess sophisticated structures. They are typically modeled
as large-scale interconnected systems, for which classical control analysis and design techniques are usually inefficient. Several results are reported in the literature
for structurally constrained control of large-scale systems in the contexts of decentralized and overlapping control, to address the shortcomings of the traditional
control techniques [5, 11, 12, 24, 25].
This work aims to measure the minimum of the smallest singular value for the
controllability and observability Gramians of parametric systems, over a given uncertainty region. Given a polynomially uncertain LTI system with uncertain parameters defined on a semi-algebraic set, it is asserted that the controllability (observability) Gramian is a rational matrix in the corresponding parameters. It is desired
to attain the minimum singular value of this matrix over the uncertainty region, but
due to the rational structure of the matrix one cannot take advantage of the efficient
techniques such as SOS tools. To bypass this obstacle, it is shown that this rational
matrix can be replaced by a polynomial approximation which satisfies an important relation. An SOS formula is then obtained to find the underlying infimum. The
special case of a polytopic uncertainty region is also investigated, due to its importance in practice. An alternative approach is proposed for this special case, with a
substantially reduced computational burden.
Two primary applications of this work are as follows:
To achieve robust closed-loop performance for a system subject to perturbation,
it is very important to know the minimum energy required to control the system
for any possible values of the parameters in the uncertainty region. This energy
is known to be proportional to the inverse of the infimum of the singular values sought in this work. For instance, this infimum would determine if a system
which is controllable for the nominal parameters, is also controllable with a sufficiently safe margin in a practical environment, where the parameters are subject
to variation around the nominal values.
In a real-world system with several interacting subsystems (which may be geographically distributed), it may not be feasible to establish information flow
between all control agents. In other words, for such systems it is more desirable to have some form of decentralization, where each control input is constructed in terms of only those outputs which are available to the corresponding
local controller due to the communication and computation limitations. To determine which inputs (or outputs) are most effective in the overall control operation
and which ones are negligible, one can obtain the minimum input energy for
different information flow structures, and choose the best structure by comparing
Robust Control of Large-Scale Systems: Efficient Selection of Inputs and Outputs
117
the resultant values [28]. The results of the present work can be used to measure
the required energy for different control structures systematically (this will be
clarified in a numerical example).
This chapter is organized as follows. The problem is formulated in Sect. 2, where
some important background results are provided. The main results of the chapter are
developed in Sect. 3 for systems with polynomial uncertainty, and the special case
of a polytopic region is also addressed in detail. The results are illustrated in Sect. 4
through a numerical example, and finally the concluding remarks are summarized
in Sect. 5.
2 Preliminaries and Problem Formulation

Consider an uncertain large-scale interconnected system with the following
representation:
x(t)
= A( )x(t) + B( )u(t)
y(t) = C( )x(t) + D( )u(t)
(1)
where
x(t) Rn , u(t) Rm and y(t) Rr are the state, input and output of the system,
respectively.
:= [1 , 2 , . . . , k ] denotes the vector of unknown, fixed uncertain parameters
of the system.
A( ), B( ),C( ) and D( ) are matrix polynomials in the variable .
The system (1) is referred to as a polynomially uncertain system. Assume that
the uncertainty vector belongs to a given semi-algebraic set D characterized as
follows:
D = { Rk | f1 ( ) 0, ..., fz ( ) 0}
(2)
where f1 ( ), ..., f z ( ) are known scalar polynomials. Note that many practical uncertainty regions can be expressed either exactly or approximately in the above form.
Due to the large-scale nature of the system, the vectors u(t) and y(t) may contain
several entries, and this has important implications in controller design, in general.
Hence, assume that only a subset of the output vector, denoted by y(t),
and a subset
of the input vector, denoted by u(t),
are desired to be used in the control structure,

for the sake of simplifying the control operation. Let S( ) be the system with the
reduced input and output size, and denote its state-space representation as:
)u(t)
x(t)
= A( )x(t) + B(
)u(t)
y(t)
= C( )x(t) + D(
(3)
The objective is to evaluate the controllability/observability degradation resulted

from reducing the size of the input and output of the system (1). In other words, it is
118
J. Lavaei et al.
intended to measure the controllability/observability degree of the system S( ) in

comparison to that of the original system (1). Addressing the above point is central
to this chapter.
Suppose that the open-loop system S( ) is robustly stable over the region D, i.e.,
all eigenvalues of the matrix A( ) lie in the open left-half s-plane for every D.
This assumption is required for defining the infinite-horizon controllability Gramian
of the system. It is noteworthy that the verification of the open-loop robust stability
of the system S( ) can be carried out systematically, using existing methods in the
literature such as the SOS technique given in [13] or the SDP method proposed in
[4]. Denote the controllability and observability Gramians of the system S( ) with
Wc ( ) and Wo ( ), respectively, which are defined as follows:
Wc ( ) =
Wo ( ) =

0
0
)T eA( )T t dt
)B(
eA( )t B(
(4a)
T
)T C(
)eA( )t dt
eA( ) t C(
(4b)
In this chapter, the system S( ) is said to be robustly controllable (observable) if it

is controllable (observable) for all D. Note that the robust controllability of the
system S( ) is equivalent to the positive-definiteness of the matrix Wc ( ) for all
D. The Gramian matrix Wc ( ) provides a measure for the degree of controllability
(rather than just controllability/uncontrollability). Indeed, it is known that the input
energy required for controlling the system is, roughly speaking, proportional to the
inverse of the matrix Wc ( ), and more specifically, proportional to the inverse of
its smallest singular value (Proposition 4.5 in [6]). Hence, the robust controllability
degree of the system S( ) can be assessed in terms the minimum of the smallest
singular value of the matrix Wc ( ) over the region D. This chapter is concerned
with the computation of this minimum using an efficient method. Note that although
this work focuses on the robust controllability problem, the results can be easily
applied to the robust observability problem, as well (due to the existence of a natural
duality between the two problems). Since the main development of this chapter is
contingent upon the notion of SOS, some relevant background material is provided
in Sect. 2.1
2.1 Background on Sum-of-Squares

This section aims to present important results in the area of sum-of-squares (SOS).
A scalar polynomial p( ) is said to be SOS if there exists a set of scalar polynomials
p1 ( ), p2 ( ), ..., pq ( ) such that:
p( ) = p1 ( )2 + p2 ( )2 + + pq ( )2 ,
Rk
(5)
It is evident that an SOS polynomial is always nonnegative, but the converse statement is not necessarily true. Indeed, a nonnegative polynomial cannot always be
119
written as a sum of squared polynomials [1]. It is to be noted that checking whether

a given polynomial is SOS amounts to solving a convex optimization problem [19].
Sparked by Polyas work, a great deal of effort has been made in the literature to
make a connection between positive (nonnegative) polynomials and SOS polynomials. For instance, assume that p( ) is a homogeneous scalar polynomial which
is strictly positive over the region Rk \{0}. Polyas theorem states that there exists a
natural number such that the coefficients of the polynomial (1 + + k ) p( )
are all nonnegative
[7]. Hence,
it is easy to show that (12 + + k2 ) p( 2 ) is SOS,

where 2 := 12 k2 . Now, assume that p( ) is an arbitrary polynomial (not
necessarily a homogeneous one). Obviously, if there exists a set of SOS polynomials
p0 ( ), p1 ( ), ..., pk ( ) such that:
p( ) = p0 ( ) + p1 ( ) f1 ( ) + + pk ( ) fk ( )
(6)
then the polynomial p( ) is nonnegative over the region D (due to the fact that
p0 ( ), p1 ( ), ..., pk ( ) are always nonnegative, and f1 ( ), ..., fk ( ) are also nonnegative over the region D). The question arises as to whether the converse statement is true: given a polynomial p( ) that is nonnegative over the region D, can it
be written in the form (6) for some SOS polynomials p0 ( ), p1 ( ), ..., pk ( )? The
answer to this question is negative. However, Putinars theorem states that if the
polynomial p( ) is strictly positive over the region D, then it can be expressed as
(6), provided the region D satisfies some mild conditions [20].
The above results will be used in the present chapter, but in a more general case
when p( ) is a matrix polynomial.
3 Robust Controllability Degree

The following lemma is a generalized form of Lemma 1 in [4], and presents some
interesting properties of the controllability Gramian.
Lemma 1 There exist a matrix polynomial H( ) and a scalar polynomial h( )
)
such that the Gramian Wc ( ) can be written as H(
h( ) , where:
H( ) is positive semi-definite for all D.
h( ) is strictly positive for all D.
Proof. It is known that the Gramian matrix Wc ( ), D, is the unique solution of
the following continuous-time Lyapunov equation:
)B(
)T
A( )Wc ( ) + Wc ( )A( )T = B(
(7)
Hence, one can write:

1
2
)B(
)T
[I A( ) + A( ) I]vec{Wc ( )} = vec B(
(8)
120
J. Lavaei et al.
where denotes the Kronecker product, and vec{} is an operator which takes a
matrix and converts it to a vector by stacking its columns on top of one another.
Define now:
h( ) := (1)n det{I A( ) + A( ) I},
H( ) := h( )Wc ( ), D
Rk
(9)
Note that although h( ) is defined over the entire k-dimensional space, H( ) is

defined only over the uncertainty region due to the fact that the Gramian Wc ( ) used
in its definition may not be well-defined outside the uncertainty region. It can be
concluded from (8) and (9) that H( ) and h( ) are matrix and scalar polynomials,
respectively, over the region D. The domain of definition of H( ) can be easily
extended to the whole space Rk . Therefore, assume that H( ) and h( ) are two
polynomials defined over the entire space, which satisfy the relations in (9) for every
D. To prove that h( ) is strictly positive over the uncertainty region, first define
1 ( ), 2 ( ), ..., n ( ) as the eigenvalues of the matrix A( ). Due to a well-known
property of the Kronecker product, one can write:
n
h( ) = (1)n (i ( ) + j ( ))
(10)
i=1 j=1
Given a fixed D, assume with no loss of generality that 1 ( ), ..., q ( ) are

real numbers and that the remaining eigenvalues q+1 ( ), ..., n ( ) are non-real
complex numbers. A few observations can be made as follows:
For every i, j {1, 2, ..., n} such that either i > q or j > q, both of the terms
i ( ) + j ( ) and i ( ) + j ( ) appear in the product given in the right side
of (10), which makes the product strictly positive.
For every i, j {1, 2, ..., q}, the term i ( ) + j ( ) is strictly negative (as the
system is robustly stable). Moreover, there are q2 of such terms in (10).
Since the real-valued matrix A( ) has an even number of non-real complex
eigenvalues, n q is an even number.
These facts lead to the conclusion that the scalar polynomial h( ) is strictly positive
for all D. On the other hand, the definition of the Gramian Wc ( ) implies that
it is positive-semidefinite over the uncertainty region. As a result, the definition of
H( ) in (9) yields that this matrix is positive semi-definite for all D.

In light of Lemma 1, the Gramian matrix Wc ( ) is normally a non-polynomial
rational function. This property impedes the use of the available SOS techniques,
as they merely deal with polynomials. Hence, it is desirable to approximate Wc ( )
by a polynomial. One naive way to do so is to replace the exponential terms in
(4a) with some truncated Taylor series. However, the resultant polynomial approximation would not necessarily satisfy any important properties (namely the ones
that will be introduced in Theorem 1 and are essential to the development of this
chapter). Hence, a more advanced approximation technique will be provided in the
sequel.
121
Notation 1 Given a matrix M, {M} denotes its minimum singular value.

Notation 2 Given a vector = [1 , 2 , . . . , k ], let 2 be defined as 2 = [12 ,
22 , . . . , k2 ].
Assumption 1 The set D is compact, and there exist SOS scalar polynomials
w0 ( ), w1 ( ), ..., wz ( ) such that all vectors satisfying the inequality:
w0 ( ) + w1 ( ) f1 ( ) + + wz ( ) fz ( ) 0
(11)
form a compact set.

There are two important points concerning Assumption 1. First, the validity of
this assumption can be checked by solving a proper SOS problem. Furthermore,
if the assumption does not hold, then the results of this chapter will become only
sufficient, as opposed to both necessary and sufficient.
Consider H( ) and h( ) introduced in Lemma 1. It follows from the positiveness
of h( ) and the compactness of D that there exist reals 1 and 2 such that:
0 < 1 < h( ) < 2 ,
Definition 1 Define Pi ( ), D, to be:

h( )
Pi ( ) := Wc ( ) 1 1
2
(12)
2i
,
iN
(13)
The next theorem provides a means to approximate the rational controllability

Gramian by a polynomial with any arbitrary precision.
Theorem 1 The following statements are true:
(i) For every i N, Pi ( ) is a positive semi-definite matrix polynomial over the
region D that satisfies the matrix inequality:
)B(
)T 0
A( )Pi ( ) + Pi( )A( )T + B(
(14)
(ii) Given D, the sequence {Pi ( )}

1 converges to Wc ( ) from below monotonically, i.e. P1 ( ) P2 ( ) P3 ( ) and limi Pi ( ) Wc ( ) = 0.
Moreover:

1
1 2i
Pi ( )
(15)
Pi ( ) Wc ( ) 1 1
2
Proof of Part (i). The function Pi ( ) being matrix polynomial is a consequence of
the following facts:
In light of Lemma 1, Wc ( ) can be written as
H( )
h( ) .
122
J. Lavaei et al.
The polynomial:

h( ) 2i
1 1
2
(16)
is divisible by 1 (1 h(2 ) ), and hence is divisible by h( ) as well.

On the other hand, one can write:

)B(
)T =
A( )Pi ( ) + Pi( )A( )T + B(

h( ) 2i
1 1
2

)T
)B(
A( )Wc ( ) + Wc ( )A( )T + B(

2i
)B(
)T 1 h( )
= B(
0
2
(17)
This completes the proof.

Proof of Part (ii). The proof of this part follows directly from Definition 1 and the
inequality (12).

Corollary 1 The minimum singular values of Wc ( ) and Pi ( ) are related to each
other by the following inequalities:

1 2i
min {Wc ( )} min {Pi ( )} min {Wc ( )} (18)
1 1
D
D
D
2
where i = 1, 2, ....
Proof. The proof is an immediate consequence of Theorem 1.
Theorem 1 shows that Wc ( ) can be approximated, with any arbitrary precision,

by a matrix polynomial satisfying a certain matrix inequality. Moreover, implicit
bounds on the smallest singular value of the controllability matrix are provided in
Corollary 1. In order to obtain a proper range of values for the degree of a polynomial which can approximate Wc ( ) satisfactorily, it is required to know the ratio
1
2 a priori. Fortunately, the relation (10) can be used to obtain this quantity, as it
indicates that this ratio is, roughly speaking, related to the minimum and maximum
eigenvalues of A( ) over the region D. Note that the ratio 12 quantifies the uncertainty degree of the open-loop system (i.e. the matrix A( )) in terms of the location
of the eigenvalues. Let an optimization problem be introduced in the sequel.
Optimization 1 Given the system S( ) and the uncertainty region D, maximize the
real-valued scalar variable subject to the constraint that there exist a symmetric
matrix polynomial P( ) and SOS matrix polynomials S0 ( ), ..., Sz ( ), S0 ( ), ...,
Sz ( ) (all n n) such that:
z
)B(
)T = S0 ( ) + Si ( ) fi ( )
A( )P( ) + P( )A( )T + B(
(19a)
i=1
P( ) = In + S0( ) + Si ( ) fi ( )
i=1
(19b)
123
where In is the n n identity matrix. Denote the solution of this optimization problem
with .
Theorem 2 The quantity min D {Wc ( )} is equal to .
Proof. Let P( ) be a matrix polynomial for which there exist a real and SOS
matrix polynomials S0 ( ), ..., Sz ( ), S0 ( ), ..., Sz ( ) satisfying the equalities given
in (19). One can write:
)B(
)T 0
A( )P( ) + P( )A( )T + B(
P( ) In
(20a)
(20b)
for all D. It follows from (7) and (20a) that:

A( ) (Wc ( ) P( )) + (Wc ( ) P( )) A( )T 0,
(21)
Therefore, P( ) Wc ( ), D. This, together with the inequality (20b), yields:
min {Wc ( )}
(22)
min {Wc ( )}
(23)
As a result:
On the other hand, notice that:

Pj ( ) > (min {Pj ( )} )In ,
D
(24)
for any j N and R+ (note that R+ represents the set of positive real numbers).
Therefore, it can be inferred from Theorem 1 and Assumption 1 of the present work,
and Theorem 2 in [23] (i.e. the general matrix form of Putinars theorem given in the
previous section) that there exist SOS matrix polynomials S0 ( ), ..., Sz ( ) such that:
z
Pj ( ) = (min {Pj ( )} )In + S0 ( ) + Si ( ) fi ( )

D
(25)
i=1
Now, recall from Theorem 1 that:

)B(
)T 0
A( )Pj ( ) + Pj ( )A( )T + B(
(26)
Since the above inequality cannot be simply relaxed to strict inequality, Putinars
theorem cannot be used directly (as explained in Sect. 2.1). To bypass this obstacle,
one can use the result of Theorem 2 in [23] that there exist SOS matrix polynomials
S0 ( ), ..., Sz ( ) such that:

z
h( ) 2 j
1
= S0 ( ) + Si ( ) fi ( )
2
i=1
(27)
124
J. Lavaei et al.
because the left side of the above equation is always strictly positive over the region
D. Hence, it is concluded from (17) that:

h( ) 2 j
T
T
A( )Pj ( ) + Pj ( )A( ) + B( )B( ) = B( )B( ) 1

2
T
)B(
)T
= S0 ( )B(
z
)B(
)T fi ( )
+ Si ( )B(
(28)
i=1
= S0 ( ) + Si ( ) fi ( )
i=1
where:
)B(
)T ,
Si ( ) := Si ( )B(
i = 0, 1, ..., k
(29)
are SOS matrix polynomials. It results from (25) and (28) that:
min {Pj ( )}
(30)
The proof is completed by taking the inequalities (23) and (30) into consideration
and letting i and go to infinity and zero, respectively, and by using Corollary 1.
Theorem 2 provides a methodology for finding the quantity min D {Wc ( )}
indirectly via solving Optimization 1. This optimization problem is in the form of
SOS, which can be handled by a semi-definite program. For this purpose, one can
use a proper software tool such as YALMIP or SOSTOOLS [14, 15]. Nevertheless,
it is first required to consider some upper bounds a priori on the degrees of the polynomials involved in the corresponding optimization problem, from which a lower
bound on the solution of Optimization 1 can be found. In other words, this optimization problem can be formulated as a hierarchy of SDP problems, whose solutions
converge asymptotically to the quantity of interest, i.e. min D {Wc ( )}, from
below.
Remark 1 Despite the fact that a high-order rational matrix (Wc ( ) in the present
problem) cannot, in general, be approximated satisfactorily by a low-order polynomial matrix P( ), it will be illustrated later in an example that a relatively low-order
polynomial typically works well here. This is due to the fact that these two functions
need to be sufficiently close only at a critical point corresponding to the solution of
the optimization problem (as opposed to everywhere in the region D).
3.1 Special Case: A Polytopic Region

Although Theorem 2 provides a numerically tractable method for measuring the robust controllability of a system, the proposed optimization problem can be simplified
125
significantly for special cases of interest. For instance, assume that D is a polytopic
region P given by:
P = { |1 + + k = 1, 1 , ..., k 0}
(31)
This type of uncertainty region is of particular interest, due to its important

applications.
) are homogeneous matrix polynomials,
Assumption 2 Assume that A( ) and B(
and let their degrees be denoted by 1 and 2 , respectively.
Note that Assumption 2 holds automatically for polytopic systems, with 1 =
2 = 1.
Theorem 3 The quantity min P Wc ( ) is equal to the maximum value of for
) satisfying the following
which there exists a homogeneous matrix polynomial P(
k
inequalities for all R :
2 ) 0,
P(
!

2 ) + ( T )3 In + P(
2 ) + ( T )3 In AT ( 2 )
A( 2 ) P(
(32a)
2 )B(
2 )T ( T )max(0,3 +1 22 ) 0
( T )max(0,22 1 3 ) + B(
(32b)
).
where 3 denotes the degree of the polynomial P(
Proof. The proof follows by refining the proofs of Theorems 1 and 2 in such a
way that the homogeneity of the system matrices as well as the polytopic structure
of the uncertainty region are both taken into account. To this end, the homogenization technique given in [4] is adopted. First of all, notice that the equation (8)
yields:
1
2
)B(
)T
vec{Wc ( )} = [I A( ) + A( ) I]1 vec B(
(33)
)B(
)T are both homogeneous polynoNote also that I A( ) + A( ) I and B(
mials. Therefore, one can conclude that the polynomials H( ) and h( ) introduced
in Lemma 1 can both be assumed to be homogeneous. Now, let Pi ( ) be defined as:
2i

2i4
4
k
k
h( )
Pi ( ) := Wc ( ) j
2
j=1
j=1
(34)
in lieu of the one defined in (13), where 4 denotes the degree of h( ). Note that
Pi ( ) given above is identical to the one defined in (13) over the region P, and its
subtle difference is that it is homogeneous. In other words, Wc ( ) is approximated
by a homogeneous polynomial here. It is easy to show that Theorem 1 holds for the
126
J. Lavaei et al.
polynomial Pi ( ) defined in (34). On the other hand, due to the polytopic nature of
the set P in (31), the inequality (14) in Theorem 1 can be rewritten as:

A( )Pi ( ) + Pi ( )A( )T

max(0,22 1 i )
3
i= j
)B(
)T
+ B(
(35)
max(0,1 + i 22 )
3
0, P
j=1
where 3i denotes the degree of Pi ( ). One can write:

Pi ( ) : = Pi ( ) min{Pi ( )}
P
3i
= Pi ( ) min{Pi ( )}In 0,
P
Hence:
Pi ( ) 0,
In
j=1
(36)
(37)
Notice now the following facts:

A matrix polynomial M( ) is nonnegative over P if and only if M( 2 ) is non defined as follows:
negative over the unit sphere D
= { | 2 + + 2 = 1}
D
1
k
(38)
Assume that N( ) is a homogeneous matrix polynomial of degree . One can

write:

, = 0
N( ) = N
(39)

Since

one can conclude that N( ) is nonnegative over D

if and only if
D,
it is nonnegative over the whole space Rk .

Thus, the proof is completed by pursuing the argument given in the proof of
Theorem 2, and using the inequalities (35) and (37) after replacing with 2 and
removing the constraint P (note that the polynomial in the left side of the inequality (35) is homogeneous).

Optimization 2 Maximize subject to the constraint that there exist a homogeneous matrix polynomial P( ) and SOS matrix polynomials S1 ( ) and S2 ( ) (all
n n) such that:
P( 2 ) = S1 ( ),
!

2
2
T 3
2
T 3
T
2
A( ) P( ) + ( ) In + P( ) + ( ) In A ( )
(40a)
2 )B(
2 )T ( T )max(0,3 +1 22 ) = S2 ( ) (40b)
( T )max(0,22 1 3 ) + B(
127
where 3 denotes the degree of the polynomial P( ). Denote the solution of this
optimization problem with .
The following lemma is required in order to delve into the properties of the optimal parameter defined above.
Lemma 2 Let M( ) be a homogeneous matrix polynomial with the property that
M( 2 ) is positive definite for every Rk \{0}. There exists a natural number c so
that ( T )c M( 2 ) is SOS.
Proof. The proof of this lemma relies heavily on the extension of Polyas theorem
[7] to the matrix case, as carried out in [23]. More precisely, since M( ) is positive
definite over the polytope P, it follows from Theorem 3 in [23] that there exists
a natural number c such that (1 + 2 + + k )c M( ) has only positive semidefinite matrix coefficients. This implies that the coefficients of ( T )c M( 2 ) are
all positive semi-definite, and in addition, its monomials are squared terms. As a
result, ( T )c M( 2 ) is SOS.

Theorem 4 The quantity min P {Wc ( )} is equal to .
Proof. The proof will be performed in two steps. First, observe that if (40) holds for
) = P( ).
some matrices P( ), S1 ( ) and S2 ( ), then (32) is satisfied for P(
Conversely, assume that a matrix polynomial P( ) satisfies the inequalities given

in (32), with the non-strict inequalities replaced by strict inequalities for every
Rk \{0} (such strict inequalities correspond to the case when the term maximum is substituted by supremum). Note that for a strict inequality in (32a), one
) and with P(
) ( T )3 and + , respectively,
would need to replace P(
for a positive infinitesimal number . Now, one can apply Lemma 2 to these inequalities to conclude that there exists a natural number c such that if the expressions in
the left sides of the inequalities (32a) and (32b) are multiplied by ( T )c , then
)( T )c
they become SOS matrix polynomials. It is enough to choose P( ) as P(
for the inequalities given in (40) to hold (for some appropriate matrices S1 ( ) and
S2 ( )).

Remark 2 Optimization 2 is obtained from Theorem 3 by replacing the nonnegativity constraints with SOS conditions. Although one may argue that this can potentially introduce some conservatism due to the known gap between the set of SOS
polynomials and the set of nonnegative polynomials, Theorem 4 shows that this is
not the case. In other words, the proposed replacement does not make the resultant
conditions conservative at all.
Note that Optimization 2 can be handled using proper software as noted in
Remark 1, provided some a priori upper bounds are set on the degrees of the relevant
polynomials being sought. The question arises as to whether choosing higher degree
polynomials would result in a tighter (less conservative) solution. This question is
addressed in the next corollary.
Corollary 2 The solution of Optimization 2 is a monotone nondecreasing function
)).
with respect to 3 (the degree of the polynomial P(
128
J. Lavaei et al.
) satisfies the conProof. The proof is a direct consequence of the fact that if P(
)(1 + + k ) also satisfies them
straints of Optimization 2 for some , then P(
for the same , but for some other suitable matrices S1 ( ) and S2 ( ).
3.2 Comparison with Existing Results

It follows immediately from the result of [2] that the Gramian matrix can be approximated by a polynomial. Nonetheless, the relationship between the order of the
corresponding polynomial and the approximation error is not investigated in [2]. In
contrast, it is shown in Theorem 1 of the present chapter that the approximation
error reduces exponentially with respect to the order of the corresponding polynomial. In particular, given a desired error bound, one can find the required degree
of the approximating polynomial by computing 1 and 2 which depend on the
smallest and largest eigenvalues of A( ) over the region P. It is worth noting that
the exponential reduction of the error is only for the approximation of the Gramian
rational matrix with a polynomial matrix. Hence, for any fixed order of the polynomial approximation, Optimization 1 is to be solved in which the degrees of the
polynomials S0 ( ), ..., Sz ( ), S0 ( ), ..., Sz ( ) may be undesirably large. This issue
arises in almost all optimization problems derived based on Polyas or Putinars
theorems.
As far as the complexity is concerned, it is easy to verify that Optimization 2
introduced in the present work is basically as complex as the optimization problem
tackled in [4]. This is partly due to the fact that there are an SOS homogeneous
polynomial and two SOS constraints of a particular form in both approaches. For a
detailed comparison between the complexities of the results given in [4], [18] and
[13], the interested reader may refer to [13].
A recent paper [17] deals with the robust analysis and synthesis of linear uncertain systems. Although this paper presents useful results in the LMI framework,
Optimization 2 is a much simpler problem compared to the ones proposed in [17].
It is noteworthy that all of the papers surveyed above deal with an inequality of the
form:
)B(
)T 0
A( )P( ) + P( )A( )T + B(
(41)
whereas this work studies the opposite inequality (see Theorem 1):
)B(
)T 0
A( )P( ) + P( )A( )T + B(
(42)
This is a consequence of the fact that it is desired to maximize the minimum singular
value. In other words, due to the special structure of the optimization considered
here, the existing results cannot be directly applied to the underlying problem.
129
4 Numerical Example
Example 1
A simple example is given here to profoundly illustrate the results of this work. Consider an LTI uncertain fourth-order system with the following state-space matrices:
0.65 0.6 1
1.5
0.75 1.4
0
2.3
0.05 0.95 1.85 0.4
1.35 1.75 0.65 1.1
A( ) =
0.7 1.6 3.05 0.2 1 + 0.8
2.2 3.25 2.5 2
1.5 0.1 2.85 0.35
0.5 1.7 0.15 0.45
1.2 0.3 0.85 1.55

0.45 1.7 0.85 0.75
0.35 1.05 1.3 1.8
1.7 0.7 0.65 0.7
B( ) =
0.25 1.2 2.5 0.8 1 + 0.55 1 0.75 1.7 2
0.3 1.45 1 0.3
0.2 0.25 1.15 0.75
(43)
where 1 and 2 are the uncertain parameters of the system, which belong to the
polytope P = {(1 , 2 )|1 + 2 = 1, 1 , 2 0}. Regard (43) as an interconnected
system with four inputs. It is desired to determine which inputs contribute weakly
to the control of the system, and hence can be ignored for the sake of cost reduction.
In other words, the objective is to find out which inputs play a vital role in controlling the system. To this end, for any given set g {1, 2, 3, 4} let Sg ( ) represent
the system S( ) after ignoring those inputs whose indices belong to g. Denote the
controllability Gramian of this system with Wcg ( ).
Given P and a final state x0 of unit norm, consider the problem of finding an
input u(t) over the time interval (, 0] with minimum L2 norm such that it drives
the state of the system Sg ( ) from x() = 0 to x(0) = x0 . As discussed in Sect. 4.3
of [6], one possible solution is given by:
uopt (t) = Bg ( )T eA( ) t Wcg ( )1 x0 ,
T
t0
(44)
where Bg ( ) is the B-matrix of the system Sg ( ). The optimal input energy can be
computed as:
uopot 2 = xT0 Wcg ( )1 x0
(45)
Define (g) to be the maximum value of this optimal input energy over all final
states x0 of unit norm and all P. The idea behind this definition is that (g)
provides an upper bound for the input energy required to drive the state of the uncertain system Sg ( ), P, from the origin at time t = to any arbitrary point
in the unit ball at time t = 0. Note that if (g) corresponds to a final state x0 and
g
an uncertain parameter , then x0 must be a unit eigenvector of Wc ( ) associated
with its smallest eigenvalue (singular value). This relationship can be expressed by:
1
= min {Wcg ( )}
(g) P
(46)
130
J. Lavaei et al.
The objective is to evaluate (g) for different choices of g. To this end, four cases
are considered as follows:
Case 1: g = {2, 3}.

Case 2: g = {1}.
Case 3: g = {4}.
Case 4: g = {}.
To obtain (g) for any of the above cases, it suffices to solve Optimization 2 with
the appropriate matrix Bg ( ). For this purpose, the order of the polynomial P( )
being sought should be chosen a priori. This optimization is treated using YALMIP
on a Dell laptop with a 1.6 GHz processor and 512 MB memory, and the results are
given in Table 1. The last column of this table gives the points (1 , 2 ) for which
the largest optimal input energy (g) is required. These points are computed by
gridding the polytope properly, and performing an exhaustive search. Therefore, the
entries of the last column are computed using a brute force technique, which will
be exploited to verify the results obtained by solving Optimization 2. The second
column of the table gives the solution of Optimization 2 at the second relaxation,
i.e., when P( ) is assumed to be a homogeneous polynomial of order 2 (with the
monomials 12 , 22 , 1 2 ). It can be verified that the solution obtained for any of the
cases 1, 2 or 3 corresponds to the minimum singular value of the Gramian matrix
evaluated at the optimal point given in the last column of the table. This proves
that the relaxation arrives at the correct solution for cases 1, 2, and 3. For case 4,
Optimization 2 is also solved at the fourth relaxation (by considering the monomials
14 , 13 2 , 12 22 , 11 23 , 24 for P( )). The corresponding solution is given in the
third column, which is, in fact, the exact optimal value. The CPU time consumed
for solving Optimization 2 at the second relaxation is given in the fourth column for
each case; these values show that the problem is solved very fast. Using the results
in columns 2 and 3 of the table as well as the equation (46), the quantity (g) is
calculated and provided in column 5. It is to be noted that the proposed approach
requires that the matrix A( ) be stable (Hurwitz) over the polytope. This can be
verified using the method developed in [4]. It is worth mentioning that the ratio 12
is equal to 0.075 in this example.
The values given in Table 1 (case 2) imply that the first input of the system is
fairly important and ignoring it in the controller design would substantially increase
the control energy required for shifting the state vector from certain points in the
state-space. In contrast, the last input may be neglected, because its contribution is
not significant as reflected by the small minimum singular value (case 3). However,
Table 1 Numerical results for Example 1

Second
Fourth
CPU time for
Case
relaxation
relaxation
second relaxation
(g)
(1 , 2 )
1
2
3
4
147.06
13.44
5.21
3.97
(0.820, 0.180)
(0.240, 0.760)
(0.225, 0.775)
(0.234, 0.766)
0.0068
0.0744
0.1920
0.2516
*
*
*
0.2531
0.88 sec
0.82 sec
0.78 sec
0.85 sec
131
if the second and third inputs are ignored concurrently (case 1), although the system
remains robustly controllable by the remaining inputs, the required control energy
would be huge.
For each of the above-mentioned cases, let the optimal input corresponding to the
worst-case scenario (i.e. the final state x0 and the uncertain variable ) be applied
to the system. Note that this input is given by (44), with = (provided in Table 1)
and x0 = x0 as defined earlier. The resulting input and state of the system are plotted
in Figs. 14. Notice that although these signals extend from t = to t = 0, they
are sketched only on the interval t [15, 0]. These figures confirm the theoretical
Fig. 1 The input and state of the system in case 1 (i.e., when inputs 2 and 3 are blocked)
Fig. 2 The input and state of the system in case 2 (i.e., when input 1 is blocked)
Fig. 3 The input and state of the system in case 3 (i.e., when input 4 is blocked)
132
J. Lavaei et al.
Fig. 4 The input and state of the system in case 4 (i.e., when none of the inputs is blocked)
results obtained in this work. For instance, one can observe that in the case when
both inputs u2 (t) and u3 (t) are blocked, the worst-case optimal input of the system
has a large overshoot occurring at t = 0 (about 22 in magnitude).
5 Summary
Given a continuous-time linear time-invariant (LTI) system which is polynomially
uncertain on a semi-algebraic region, in this chapter we obtain the minimum of the
smallest singular value of its controllability (observability) Gramian matrix. For this
purpose, it is first shown that the Gramian is a rational function which can be approximated by a matrix polynomial (with any arbitrary precision) that satisfies an
important relation. A sum-of-squares (SOS) formula is then derived for solving the
underlying problem, which can be efficiently handled using proper software. An alternative SOS method is subsequently obtained for the case when the uncertainty
region is a polytope. This allows one to measure the robust controllability (observability) degree of the system, when its parameters are subject to perturbation. The
method proposed here can be used to find a dominant subset of inputs and outputs
for any given large-scale system, by determining the effectiveness of each input and
output in the overall operation of the control system. Simulations demonstrate the
efficacy of the proposed results.
Acknowledgement The authors would like to acknowledge that this chapter is written based on
their recent work [26].
References
1. Blekherman, G.: There are significantly more nonnegative polynomials than sums of squares.
Israel Journal of Mathematics 153, 355380 (2006)
2. Bliman, P. A., Oliveira, R. C. L. F., Montagner, V. F., Peres, P. L. D.: Existence of homogeneous polynomial solutions for parameter-dependent linear matrix inequalities with parameters in the simplex. Proceedings of 45th IEEE Conference on Decision and Control. San
Diego, USA. 14861491 (2006)
133
3. Chesi, G.: On the non-conservatism of a novel LMI relaxation for robust analysis of polytopic
systems. Automatica 44, 29732976 (2008)
4. Chesi, G., Garulli, A., Tesi, A., Vicino, A.: Polynomially parameter-dependent Lyapunov
functions for robust stability of polytopic systems: an LMI approach. IEEE Transactions on
Automatic Control 50, 365370 (2005)
5. Davison, E. J., Chang, T. N.: Decentralized stabilization and pole assignment for general
proper systems. IEEE Transactions Automatic Control 35, 652664 (1990)
6. Dullerud, G. E., Paganini, F.: A course in robust control theory: a convex approach. Texts in
Applied Mathematics, Springer, Berlin (2005)
7. Hardy, G. H., Littlewood, J. E., Polya, G.: Inequalities. Cambridge University Press,
Cambridge, UK, Second edition (1952)
8. Hillar, C. J., Nie, J.: An elementary and constructive solution to Hilberts 17th Problem for
matrices. Proceedings of the American Mathematical Society 136, 7376 (2008)
9. Kau, S., Liu, Y., Hong, L., Lee, C., Fang, C., Lee, L.: A new LMI condition for robust stability
of discrete-time uncertain systems. Systems and Control Letters 54, 11951203 (2005)
10. Lavaei, J., Aghdam, A. G.: Optimal periodic feedback design for continuous-time LTI systems
with constrained control structure. International Journal of Control 80, 220230 (2007)
11. Lavaei, J., Aghdam, A. G.: A graph theoretic method to find decentralized fixed modes of LTI
systems. Automatica 43, 21292133 (2007)
12. Lavaei, J., Aghdam, A. G.: Control of continuous-time LTI systems by means of structurally
constrained controllers. Automatica 44, 141148 (2008)
13. Lavaei, J., Aghdam, A. G.: Robust stability of LTI systems over semi-algebraic sets using
sum-of-squares matrix polynomials. IEEE Transactions on Automatic Control 53, 417423
(2008)
14. Lofberg, J.: A toolbox for modeling and optimization in MATLAB. Proc. of the CACSD
Conference. Taipei, Taiwan (2004) Available via http://control.ee.ethz.ch/ joloef/yalmip.php
15. Prajna, S., Papachristodoulou, A., Seiler, P., Parrilo, P. A.: SOSTOOLS sum of squares optimization toolbox for MATLAB. Users guide (2004) Available via http://www.cds.caltech.edu/
sostools
16. Oliveira, M. C., Geromel, J. C.: A class of robust stability conditions where linear parameter dependence of the Lyapunov function is a necessary condition for arbitrary parameter
dependence. Systems and Control Letters 54, 11311134 (2005)
17. Oliveira, R. C. L. F., Oliveira, M. C., Peres, P. L. D.: Convergent LMI relaxations for robust
analysis of uncertain linear systems using lifted polynomial parameter-dependent Lyapunov
functions. Systems and Control Letters 57, 680689 (2008)
18. Oliveira, R. C. L. F., Peres, P. L. D.: LMI conditions for robust stability analysis based on polynomially parameter-dependent Lyapunov functions. Systems and Control Letters 55, 5261
(2006)
19. Parrilo, P. A.: Structured semidefinite programs and semialgebraic geometry methods in robustness and optimization. Ph.D. dissertation, California Institute of Technology (2000)
20. Putinar, M.: Positive polynomials on compact semi-algebraic sets. Indiana University Mathematics Journal 42, 969984 (1993)
21. Sastry, S. S., Desoer, C. A.: The robustness of controllability and observability of linear timevarying systems. IEEE Transactions on Automatic Control 27, 933939 (1982)
22. Savkin, A. V., Petersen, I. R.: Weak robust controllability and observability of uncertain linear
systems. IEEE Transactions on Automatic Control 44, 10371041 (1999)
23. Scherer, C. W., Hol, C. W. J.: Matrix sum-of-squares relaxations for robust semi-definite programs. Mathematical Programming 107, 189211 (2006)
24. Siljak, D. D.: Decentralized control of complex systems. Academic, Cambridge (1991)
25. Sojoudi, S., Aghdam, A. G.: Characterizing all classes of LTI stabilizing structurally constrained controllers by means of combinatorics. Proceedings 46th IEEE Conference on Decision and Control. New Orleans, USA, 44154420 (2007)
134
J. Lavaei et al.
26. Sojoudi, S., Lavaei, J., Aghdam, A. G.: Robust controllability and observability degrees of
polynomially uncertain systems. Automatica 45(11), 26402645, Nov. (2009)
27. Ugrinovskii, V. A.: Robust controllability of linear stochastic uncertain systems. Automatica
41, 807813 (2005)
28. Wang, S. H., Davison, E. J.: On the stabilization of decentralized control systems. IEEE Transactions on Automatic Control 18, 473478 (1973)
Decentralized Output-Feedback Control

of Large-Scale Interconnected Systems
via Dynamic High-Gain Scaling
1 Introduction
Large-scale systems occurring in several application domains (including, as a very
short representative list, power systems, multi-agent systems, communication and
transportation networks, supply chains, etc.) can be profitably viewed as interconnections of multiple subsystems. In this general context, the development of
control algorithms for interconnected large-scale systems has attracted considerable
research interest. Interest in decentralized control designs has been significantly renewed in recent years due to noteworthy extensions promised by the application of
new results emerging in nonlinear robust and adaptive output-feedback control. In
this vein, this chapter addresses the design of decentralized output-feedback controllers for a class of nonlinear interconnected large-scale systems based on our
recent results on the dynamic high-gain scaling control design technique. The results here follow the general direction in the decentralized literature of attempting
to generalize the form of the dynamics of subsystems and simultaneously weaken
the assumptions on subsystem interconnections. We consider a class of interconnected large-scale systems with each subsystem being of the form
z(i,m) = q(i,m) (z, x, u,t, ) , i = 1, . . . , sm + 1
x(i,m) = (i,m) (z, x, u,t, ) + (i,i+1,m) (x(1,m) )x(i+1,m) , i = 1, . . . , sm 1
x(i,m) = (i,m) (z, x, u,t, ) + (i,i+1,m) (x(1,m) )x(i+1,m)
+(ism,m) (x(1,m) )um , i = sm , . . . , nm 1
P. Krishnamurthy
Control/Robotics Research Laboratory (CRRL), Department of Electrical
and Computer Engineering, Polytechnic Institute of NYU, Brooklyn, NY 11201, USA
e-mail: pk@crrl.poly.edu
F. Khorrami
Control/Robotics Research Laboratory (CRRL), Department of Electrical and Computer
Engineering, Polytechnic Institute of NYU, Brooklyn, NY 11201, USA
e-mail: khorrami@smart.poly.edu
135
136
x(nm ,m) = (nm ,m) (z, x, u,t, ) + (nm sm ,m) (x(1,m) )um
ym = x(1,m)
(1)
where xm = [x(1,m) , . . . , x(nm ,m) ]T Rnm is the state, ym = x(1,m) R the output, um
nz(1,m) +...+nz(s +1,m)
m
R the input, and zm = [zT(1,m) , . . . , zT(sm +1,m) ]T R
the state of the
subsystem. M is the number of subsystems, x =

appended dynamics of the
[xT1 , . . . , xTM ]T , u = [u1 , . . . , uM ]T , and z = [zT1 , . . . , zTM ]T . (i,i+1,m) , i = 1, . . . , nm 1,
and (i,m) , i = 0, . . . , nm sm , are known scalar real-valued continuous functions.
q(i,m) , i = 1, . . . , sm + 1, and (i,m) , i = 1, . . . , nm , are continuous scalar real-valued
uncertain functions. sm is the relative degree of the mth subsystem. Rn is the
exogenous disturbance input.
Early results in decentralized control focused on linear systems [1, 5] and linearly bounded interconnections [8, 22]. In [25], higher order (i.e., polynomial type)
interconnections were considered for large-scale systems assuming matching conditions. Backstepping-based robust decentralized controllers were designed in [3, 4]
for systems of output-feedback canonical form including uncertain parameters and
polynomially bounded uncertainties. Using the Cascading Upper Diagonal Dominance (CUDD) based technique in [19], a decentralized output-feedback disturbance
attenuation scheme was proposed in [9] for interconnected large-scale systems with
each subsystem being in the generalized output-feedback canonical form [19] and
with nonlinear appended dynamics.
In a recent sequence of papers, we have proposed a new dynamic high-gain
scaling-based control design approach with a central ingredient of the technique
being solutions of coupled Lyapunov inequalities [10, 11]. Unlike previous control
design approaches such as backstepping [21], forwarding [24], and older high-gain
techniques [2, 7, 23], the new dynamic scaling-based approach provides a unified
framework applicable to both state-feedback and output-feedback problems for both
strict-feedback [11] and feedforward [12] systems. Furthermore, the new dynamic
scaling-based technique provides strong robustness properties and allows appended
Input-to-State Stable (ISS) dynamics driven by all states1 [11] and also crossproducts of unknown parameters and unmeasured states [13]. In this chapter, we
show that the dynamic high-gain scaling-based technique can be applied to the decentralized output-feedback stabilization and disturbance attenuation problems, thus
yielding decentralized results for a wider class of large-scale interconnected systems
than available from prior results. The output-feedback decentralized control design
based on the adaptive dual dynamic high-gain observer/controller methodology is
described in Sect. 2. Thereafter, the application of the generalized scaling technique
to eliminate the cascading dominance assumption on upper diagonal terms is described in Sect. 3.
mth
Previous results required the ISS appended dynamics to have nonzero gain only from the output y.
137
2 Decentralized Control Based on The Adaptive Dual Dynamic

High-Gain Scaling Paradigm
2.1 Assumptions
The design is carried out under the Assumptions A1A5, each of which is required
to hold for all m {1, . . . , M}.
Assumption A1. A constant m > 0 exists such that
|(i,i+1,m) (x(1,m) )| m , 1 i nm 1
| (0,m) (x(1,m) )| m
(2)
(3)
for all x(1,m) R. Furthermore, the sign of each (i,i+1,m) , i = 1, . . . , nm 1, is independent of its argument.
Assumption A2. The inverse dynamics of (1)
(i,m) = (sm +i,m) (z, x, v u,t,

) + (sm +i,sm +i+1,m) (x(1,m) )(i+1,m)
+(i,m)(x(1,m) )(vm u m ) , 1 i nm sm
(4)
are Input-to-State practically Stable (ISpS) with state m = [(1,m) , . . . , (nm sm ,m) ]T =
[x(sm +1,m) , . . . , x(nm ,m) ]T and input [z, xm , v,t, ]T where (nm sm +1,m) 0 is a dummy
variable, xm is the vector comprised of the entire state x except (x(sm +1,m) , . . . , x(nm ,m) ),
vm = um + um , um =
(sm ,sm +1,m) (x(1,m) )

(0,m) (x(1,m) ) (1,m) ,
v = [v1 , . . . , vM ]T , and u = [u1 , . . . , uM ]T .
Furthermore, an ISpS Lyapunov function Vm (m ) exists which satisfies2

sm
sm
i=2
i=2
2
Vm m (|m |) + m (|x(1,m) |) v2m + x2(i,m) + (i,m)
(|z(i,m) |)
!
M
2
+m (1,m,k)
(|z(1,k) |) + m (m ,k) (x(1,k) )x2(1,k) + (m, ) (| |) (5)
M
k=1
k=1
where m is an unknown non-negative constant. m is a known class K function. m , (i,m) , i = 2, . . . , sm , (1,m,k) and (m ,k) , k = 1, . . . , M, and (m , ) are
known continuous non-negative functions. A class K function V m exists such that
Vm (m ) V m (|m |) for all m Rnm sm . Non-negative constants m and m exist
such that |m |2 m m (|m |) and |m |2 m Vm for all m Rnm sm .
Assumption A3. The uncertain functions (i,m) satisfy inequalities (6) for 1 i sm
and (7) for sm + 1 i nm
2 To minimize notation, instead of introducing an additional positive constant in V
m as in the standard ISpS Lyapunov inequality [6], we have subsumed the effect of an additive positive constant
into (m , ) by not requiring (m , ) to vanish at the origin.
138
|(i,m) | m (x(1,m) )
!
M
i
|x
|
+
(|z
|)
+
(|z
|)
m
( j,m) ( j,m) ( j,m)
(1,m,k) (1,k)
i
j=2
j=2
k=1
+m (m,k) (x(1,k) )|x(1,k) | + (m, )(| |)
(6)
k=1
|(i,m) | m (x(1,m) )
nm
sm+1
j=2
j=2
|x( j,m) | +
( j,m) (|z( j,m) |)
+m (1,m,k) (|z(1,k) |) + m |um |

M
k=1
M
+m (m,k) (x(1,k) )|x(1,k) | + (m, )(| |)
(7)
k=1
nz
for all t 0, xm Rnm , m = 1, . . . , M, z(i,m) R (i,m) , i = 1, . . . , sm + 1, m = 1, . . . , M,

u RM , and Rn , with m being an unknown non-negative constant and m ,
( j,m) , j = 2, . . . , sm + 1, (1,m,k) , k = 1, . . . , M, (m,k) , k = 1, . . . , M, and (m, ) being
known continuous non-negative functions.
Assumption A4. Positive constants (i,m) and (i,m) exist such that
|(i,i+1,m) (x(1,m) )| (i,m) |(i1,i,m) (x(1,m) )|, i = 3, . . . , sm 1
(8)
|(i,i+1,m) (x(1,m) )| (i,m) |(i1,i,m) (x(1,m) )|, i = 3, . . . , nm 1
(9)
for all x(1,m) R.

Assumption A5. The z(i,m) , i = 1, . . . , sm + 1, subsystems are ISpS with ISpS Lyapunov functions Vz(i,m) satisfying
M
Vz(1,m) z(1,m) (|z(1,m) |) + m (z(1,m) ,k) (|x(1,k) |) + (z(1,m), ) (| |) (10)

k=1
M
Vz(i,m) z(i,m) (|z(i,m) |) + m (z(i,m) ,k) (|x(1,k) |)x2(1,k)

k=1
+z(i,m) (|x(1,m) |) x2( j,m) + (z(i,m) , ) (| |), i = 2, . . . , sm
(11)
j=2
Vz(sm +1,m) z(sm +1,m) (|z(sm +1,m) |) + m (z(sm +1,m) ,k) (|x(1,k) |)x2(1,k)
k=1
3 nm
4
+z(sm +1,m) (|x(1,m) |) x2( j,m) + m u2m + (z(sm +1,m) , ) (| |) (12)
j=2
where m is an unknown non-negative constant, z(i,m) , i = 1, . . . , sm + 1, are known

class K functions, and z(i,m) , i = 2, . . . , sm + 1, (z(i,m) ,k) , i = 1, . . . , sm + 1, k =
139
1, . . . , M, and (z(i,m) , ) , i = 1, . . . , sm +1, are known continuous non-negative functions.
A class K function V z(1,m) , exists such that Vz(1,m) (z(1,m) ) V z(1,m) (|z(1,m) |) for all
nz(1,m)
z(1,m) R
. Also, positive constants V z(i,m) , i = 2, . . . , sm + 1, exist such that
Vz(i,m) (z(i,m) ) V z(i,m) z(i,m) (|z(i,m) |) for all z(i,m) R

constants z(i,m) and z(i,m) exist such that3
nz(i,m)
. Furthermore, non-negative
2
2
(i,m)
(|z(i,m) |) + (i,m)
(|z(i,m) |) z(i,m) z(i,m) (|z(i,m) |)
(13)
2
2
(i,m)
(|z(i,m) |) + (i,m)
(|z(i,m) |)
(14)
for all z(i,m) R

0+ :
nz(i,m)
z(i,m) Vz(i,m)
, i = 2, . . . , sm + 1. The following local order estimates hold as
2
2
(a) M
k=1 [(1,k,m) ( ) + (1,k,m) ( )] = O[z(1,m) ( )]
(b) M [(1,k,m) ( ) + (1,k,m)( )] = O[ ]
k=1
2
(c) M
k=1 (z(1,k) ,m) ( ) = O[ ].
Remark 1. Under Assumptions A1 and A4, symmetric positive-definite matrices

Pom and Pcm , positive constants om , om , om , om , cm , cm , and cm , and continuous
functions g(2,m) , . . . , g(nm,m) , k(2,m) , . . . , k(sm ,m) can be found [10] such that for all
x(1,m) R,
Pom Aom (x(1,m) ) + ATom (x(1,m) )Pom om Inm 1 om |(2,3,m) (x(1,m) )|CmT Cm
om Inm 1 Pom D om + D om Pom om Inm 1
Pcm Acm (x(1,m) ) + ATcm (x(1,m) )Pcm cm |(2,3,m) (x(1,m) )|Ism 1

cm Ism 1 Pcm D cm + D cm Pcm cm Ism 1
(15)

(16)
where4
Aom = upperdiag((2,3,m) , . . . , (nm 1,nm ,m) ) [g(2,m) , . . . , g(nm ,m) ]T Cm
1
D om = Dom Inm1 ; Dom = diag(1, 2, . . . , nm 1)
2
Acm = upperdiag((2,3,m) , . . . , (sm 1,sm ,m) ) Bm[k(2,m) , . . . , k(sm ,m) ]
1
D cm = Dcm Ism 1 ; Dcm = diag(1, 2, . . . , sm 1)
2
(17)
(18)
(19)
(20)
where Cm = [1, 0, . . . , 0] is an (nm 1) dimensional row vector and Bm = [0, . . ., 0, 1]T

is an (sm 1) dimensional column vector. Furthermore, by Theorem A1 in [11], a
1
m
positive
constant
Gm
exists
such
that
(si=2
g2(i,m) ) 2
Gm |(2,3,m) (x(1,m) )|.
3
4
(i,m) for i = sm + 1 is a dummy variable identically equal to zero.

For notational clarity, we drop the arguments of functions when no confusion will result.
140
2.2 Observer and Controller Designs

A reduced-order observer with states xm = [x(2,m) , . . . , x(nm ,m) ]T is designed as
i
x(i,m) = (i,i+1,m) (x(1,m) )[x(i+1,m) + rm
fi+1 (x(1,m) )] + (ism,m) (x(1,m) )um
i1
rm
g(i,m) (x(1,m) )[x(2,m) + rm f2 (x(1,m) )]
i2
f(i,m) (x(1,m) ), 2 i nm
(i 1)rm rm
(21)
where rm is the high-gain scaling parameter introduced in the dynamic controller

for the mth subsystem and
f(i,m) (x(1,m) ) =
x
(1,m) g(i,m) ( )
(1,2,m) ( )
(22)
for 2 i nm . The observer errors e(2,m) , . . . , e(nm ,m) and scaled observer errors
(2,m) , . . . , (nm ,m) are defined as
i1
e(i,m) = x(i,m) + rm
f(i,m) (x(1,m) ) x(i,m)
e(i,m)
(i,m) = i1 , i = 2, . . . , nm .
rm
(23)
(24)
The dynamics of m = [(2,m) , . . . , (nm ,m) ]T are
m = rm Aom m
rm
Do m + m
rm m
(25)
where m = [ (2,m) , . . . , (nm ,m) ]T with
(i,m) =
(i,m) (z, x, u,t, )

i1
rm
+ g(i,m) (x(1,m) )
(1,m) (z, x, u,t, )

.
(1,2,m) (x(1,m) )
(26)
Introduce m = [(2,m) , . . . , (sm ,m) ]T with

x(2,m) + rm f(2,m) (x(1,m) ) + m(x(1,m) , m )
rm
i1 f
x(i,m) + rm
(x
(i,m) (1,m) )
=
, i = 3, . . . , sm ,
i1
rm
(2,m) =
(i,m)
(27)
m being a dynamic adaptation parameter and

m (x(1,m) , m ) = (1 + m)x(1,m) (1,m) (x(1,m) )
(28)
with (1,m) being a design freedom to be chosen later. The control law is
designed as
um =
141
x

sm
rm
(sm +1,m)
+ f(sm +1,m) (x(1,m) )
(sm,sm +1,m) (x(1,m) )
sm
(0,m) (x(1,m) )
rm
!
sm
k(i,m) (x(1,m) )(i,m) .
(29)
i=2
The dynamics of m are given by

rm
Dc m + m rm Gm (2,m)
rm m
+Hm((2,m) (2,m) ) + m
(1,m)
Gm = [g(2,m) , . . . , g(sm ,m) ]T ; m = Gm
(1,2,m)
'
(T

Hm = (1 + m){(1,m) x(1,m) + (1,m) }(1,2,m) , 0, . . . , 0
m = rm Acm m
'
1

m =
m (1,m) x(1,m) + (1 + m)[(1,m)
x(1,m) + (1,m) ]
rm
(T
[(1,m) (1 + m)(1,m) x(1,m) (1,2,m) ], 0, . . . , 0

(x(1,m) denotes
where (1,m)
(30)
(31)
(32)
(33)
d (1,m) ( )

d
=x
(1,m)
2.3 Stability Analysis

Noting that the inverse dynamics with state m = [x(sm +1,m) , . . . , x(nm ,m) ]T are of the
form (4) with
vm =
4
3
sm
rm
(sm,sm +1,m) (x(1,m) )(sm +1,m) KmT (x(1,m) )m
(0,m) (x(1,m) )
(34)
where Km (x(1,m) ) = [k(2,m) (x(1,m) ), . . . , k(sm ,m) (x(1,m) )]T , and using Assumption A2,
d
dt
Vm
2sm 1
rm
rm
3
m (|m |)
2
(2sm 1) 2sm Vm + m [1 + m]2 (1,m)
x2(1,m)
2sm 1
r
m
rm
rm
sm 2 (|z
M
1
(i,m) (i,m) |)]
2
+ 2 m
+
(1,m,k)
(|z(1,k) |)
m
m
2i3
rm
rm
i=2
k=1
!
2
2
+rm m 3 + 2 (s2 m ,sm +1,m) + 2 |Km |2 (|m |2 + |m |2 )
(0,m)
(0,m)
1
m M
(m ,k) x2(1,k) + rm (m , ) .
rm k=1
(35)
142
Consider the observer and controller Lyapunov functions defined as

V(o,m) = rm mT Pom m
(36)
V(c,m)
(37)

1
1
T
= rm m Pcm m +
x2
1+
2
rm (1,m)
where Pom and Pcm are picked as in Remark 1. The dynamics of the high-gain parameter rm and the adaptation parameter m will be designed such that rm is larger
than 1, m is positive, and rm and m are monotonically non-decreasing. (1,m) will
be designed such that (1,2,m) (1,m) (and hence x(1,m) (1,2,m) m ) is positive. Differentiating V(o,m) and V(c,m) and using (15) and (16),
2
2
2
om |m |2 rm
om |(2,3,m) |(2,m)
rm om |m |2 + 2rmmT Pom m (38)
V(o,m) rm
1 rm
2
V(c,m) rm
cm |(2,3,m) ||m |2 rm cm |m |2 x(1,m) (1,2,m) m 2 x2(1,m)
2 rm
(1,m)
2 T
+2rmmT Pcm Gm
2rm
m Pcm Gm (2,m)
(1,2,m)
+2rmmT Pcm Hm ((2,m) (2,m) ) + 2rm mT Pcm m
+2|x(1,m) (1,m) | + 2|x(1,m)(rm (2,m) rm (2,m) )(1,2,m) |.
(39)
A composite Lyapunov function for the xm component of the mth subsystem is defined as
Vxm = cmV(o,m) + V(c,m) + (cm m + 1)
Vm
2sm 1
rm
(40)
2 (P )G ]/[
om cm ]. Using (35),
where cm is any positive constant larger than5 [8max
cm
m
(38), and (39), and forming upper bounds for the terms in (38) and (39) along the
lines in [11],
2 cm om
2 cm
||m |2 rm cm om |m |2 rm cm |m |2
Vxm rm
|m |2 rm
|
4
4 (2,3,m)
1 rm
(|m |)
(1 + m)x2(1,m) (1,2,m) (1,m) 2 x2(1,m) m2sm 1
2 rm
rm
rm
(2sm 1)(cm m + 1) 2sm Vm + q(1,m) (x(1,m) )x2(1,m)
rm
M
+m q(2,m,k) (x(1,k) )x2(1,k) +

k=1
1
2
q
(x
, m )(1,m)
x2(1,m)
rm (3,m) (1,m)
3
2
+rm wm (x(1,m) , m , m )[|m |2 + |m |2 ]
max (P) with P being a square symmetric matrix denotes the maximum eigenvalue of P.
143
4
M 3
2
2
+Gm m hm (x(1,m) ) (1,m,k)
(|z(1,k) |) + (1,m,k)
(|z(1,k) |)
'
k=1
m (|x(1,m) |)
+ 1 + (cm m + 1)
2
rm
rm2
cm
+m
sm +1
2 (|z
2
[(i,m)
(i,m) |) + (i,m) (|z(i,m) |)]
2i3
rm
i=2
m2 (x(1,m) )
8cm nm 2
max (Pom ) 2
o m
(0,m) (x(1,m) )
'
3 x
4
(sm +1,m)
(sm ,sm +1,m) (x(1,m) )
+ f(sm +1,m) (x(1,m) ) + KmT (x(1,m) )m
sm
rm
(2
+(m, ) (| |)
(41)
where
m = 1 + m + m2
(42)
16nmcm 2
16
2
2
2
Gm = cm m + 4 + cm +
max (Pom )Gm +
max
(Pcm )Gm
om
m cm
(43)
and the functions q(1,m) (x(1,m) ), q(2,m,k) (x(1,k) ), q(3,m) (x(1,m) , m ), wm (x(1,m) , m , m ),
hm (x(1,m) ), and (m, ) (| |) are given by
q(1,m) = 3 + 2cm +
!
2
2
+
2
cm om cm |(2,3,m) | (1,2,m)
2
2
) + 3(m,m)
+
q(2,m,m) = 2cm (1 + (m,m)
(44)
2
16nmcm 2
2 (2,3,m) 2
max (Pom )Gm 2
om
(1,2,m) m
16 2
2 |(2,3,m) | 2
max (Pcm )Gm 2
+ (cmm + 1)(m ,m)
cm
(1,2,m) m
(45)
2
2
) + 3(m,k)
+ (cm m + 1)(m ,k) for k = m
q(2,m,k) = 2cm (1 + (m,k)
(46)
q(3,m) = 3(cm m + 1)(1 + m)2 m
(47)
wm = (cm m + 1)m 3 +
2
2
(0,m)
(s2 m ,sm +1,m) +
2
2
(0,m)
!
|Km |
2
2
+cm nm max
(Pom )m2 + cmnm max
(Pom )(1 + m)2 m2 [1 + |(1,m) |]2
144

2
+3cm nm max (Pom )m + 2cmn2m max
(Pom )m2

+3max (Pcm )(1 + m)|(1,m)
x(1,m) + (1,m) ||(1,2,m) |
)
2
(Pc ) 2m2 2 + 2(1 + m)4 [ x(1,m) + (1,m) ]2 2
+max
(1,m)
2
(1,2,m) (1,m)
(1,m)
5
+(1 + m)

[(1,m)
x(1,m) + (1,m) ]2
hm = 1 + m2 + m +
(m, ) = (m , ) +
2
(2,3,m)
(48)
m2
(49)
1
cm 2
+
.
16cm 16om (m, )
(50)
2
(1,2,m)
The terms involving z(1,k) , k = 1, . . . , M in (41) can be upper bounded as

2
2
(|z(1,k) |) + (1,m,k)
(|z(1,k) |)] (1,m,k) (|z(1,k) |)
hm (x(1,m) )[(1,m,k)
+ (2,m,k) (x(1,m) )
(51)
where
2
2
(1,m,k) (|z(1,k) |) = [(1,m,k)
(|z(1,k) |) + (1,m,k)
(|z(1,k) |)]
(2,m,k) (x(1,m) ) =
sup
sup
| ||z(1,k) |
h m ( )
2
2
[(1,m,k)
( ) + (1,m,k)
( )]hm (x(1,m) ).
0 |x(1,m) |
(52)
(53)
By Assumption A5,
2
2
(1,m,k)
( ) + (1,m,k)
( ) = O[z(1,k) ( )]
(1,m,k) ( ) + (1,m,k)( ) = O[ ]
(54)
(55)
as 0+ . Hence, (1,m,k) ( ) = O[z(1,k) ( )], (1,m,k) ( ) = O[ 2], and (2,m,k) ( ) =

O[ 2 ] as 0+ . Hence, a continuous non-negative function (2,m,k) exists such
that
(2,m,k) (x(1,m) ) x2(1,m) (2,m,k) (x(1,m) ).
(56)
Using a reasoning similar to that used in the proof of Theorem 2 in [26], it is seen
that the local order estimate (1,k,m) ( ) = O[z(1,m) ( )] as 0+ implies the exand
,
istence of a new Lyapunov function Vz , class K functions z
(1,m)
(1,m)
and continuous non-negative functions m and (z(1,m) , ) such that

M
(z(1,m) ,k)
Vz(1,m) z(1,m) (|z(1,m) |) + m (m ) (z(1,m) ,k) (|x(1,k) |) + (z(1,m), ) (| |) (57)

k=1
145
with z(1,m) ( ) = O[z(1,m) ( )] as 0+ , (z(1,m) ,k) independent of m , (z(1,m) ,k) ( )

= O[(z(1,m),k) ( )] as 0+, and with z(1,m) satisfying the inequality z(1,m) (|z(1,m) |)
nz(1,m)
M
k=1 (1,k,m) (|z(1,m) |) z(1,m) R
tion (z(1,m) ,k) exists such that
. Hence, a continuous non-negative func-
(z(1,m) ,k) (|x(1,k) |) x2(1,k) (z(1,m) ,k) (x(1,k) ).
(58)
Furthermore, it can be shown that a class K function V z(1,m) exists such that
nz
Vz(1,m) (z(1,m) ) V z(1,m) (|z(1,m) |) for all z(1,m) R (1,m) .
Using Assumption A5,

M
z(i,m) (|z(i,m) |)
1
d Vz(i,m)
+ m (z(i,m) ,k) (|x(1,k) |)x2(1,k)
5
5
2i
2i
dt
rm k=1
rm 2
rm 2
3
2
+ z(i,m) (|x(1,m) |)(1 + m)2 (1,m)
x2(1,m)
rm
+3rm z(i,m) (|x(1,m) |)[|m |2 + |m |2 ]

5 rm
1
Vz
+ (z(i,m) , ) (| |) 2i
rm
2 2i 32 (i,m)
rm
(59)
for i = 2, . . . , sm . Similarly,

M
z(sm +1,m) (|z(sm +1,m) |)
d Vz(sm +1,m)
1
+
(|x
|)x2
2sm 12
2sm 12
dt
rm k=1 (z(sm +1,m) ,k) (1,k) (1,k)
rm
rm
3
2
+ z(sm +1,m) (|x(1,m) |)(1 + m)2 (1,m)
x2(1,m)
rm
+3rm z(sm +1,m) (|x(1,m) |)[|m |2 + |m |2 ]
+z(sm +1,m) (|x(1,m) |)
|m |2
2s 1
rm m 2
1 z
(|x(1,m) |)
(sm +1,m)
3 x
(sm +1,m)
(x
)
(sm ,sm +1,m) (1,m)
sm
2
rm
(0,m)
(x(1,m) )
!2
4
T
+ f(sm +1,m) (x(1,m) ) + Km (x(1,m) )m
+m rm2

1
1 rm
(z(sm +1,m) , ) (| |) 2sm
Vz
.
rm
2 r2sm + 12 (sm +1,m)
(60)
The composite Lyapunov function for the overall system is defined as

V=
Vm
m=1
(61)
146
with
Vm = Vxm + (cm + 1)
sm +1
z(i,m) Vz(i,m)
i=2
2i 5
rm 2
1
+ (1 + Gmm )Vz(1,m) + (m )2 (62)
2
where
{(1 + m )[1 + m(m )]} =
m=1
{(2 + m + m2 )[1 + m(m )]}.
(63)
m=1
The parameter estimator dynamics are designed as

M 3
m = m m + q(2,k,m) (x(1,m) )x2(1,m) + Gk (2,m,k) (x(1,m) )
k=1
+(1 + Gk )(z(1,k) ,m) (|x(1,m) |) + (ck + 1)x2(1,m)

+
sk +1
i=2
4
z(i,k) (z(i,k),m) (|x(1,m) |)
8cm nm
2 (Po ) 2 (x(1,m) )
2
om (0,m)
(x(1,m) ) max m m
3
3 x
(sm +1,m)
(|x
|)
(x
)
z
(1,m)
(sm ,sm +1,m) (1,m)
sm
2
rm
(0,m)
(x(1,m) ) (sm +1,m)
42
4
+ f(sm +1,m) (x(1,m) ) + KmT (x(1,m) )m
(64)
1
+(cm + 1)
z(sm +1,m) rm2
where m is a nonnegative design parameter. m is initialized to be positive. By

(64), m 0 so that m remains positive for all time. Using (41), (57), (59), and
(64),
V
m=1
where
2
m = rm
cm om
2 cm
|m |2 + rm
|
(x
)||m |2
4
4 (2,3,m) (1,m)
+rm cm om |m |2 + rm cm |m |2
1 rm
+(1 + m)x2(1,m) (1,2,m) (x(1,m) )(1,m) (x(1,m) ) + 2 x2(1,m)
2 rm
rm
m
(m )2 + (2sm 1)(cmm + 1) 2sm Vm
2
rm
sm +1 z
(|z
|)
z
(i,m)
(i,m)
(i,m)
+ z(1,m) (|z(1,m) |) +
2i 52
i=2
rm
+
m (|m |)
2sm 1
rm
(65)
+(cm + 1)
147
!
sm +1 (2i 5 )z
V
r
2
(i,m) z(i,m) m
2i 23
rm
i=2
+ q(1,m) (x(1,m) )x2(1,m) + m q(2,k,m) (x(1,m) )x2(1,m)

k=1
1
|m |2
m 2
2
q(3,m) (x(1,m) , m )(1,m)
(x(1,m) )x2(1,m) + q(4,m) (x(1,m) , m )
+
2sm 12
rm
2
r
m
2
sm +1 2 (|z
(i,m) (i,m) |) + (i,m) (|z(i,m) |)
+q(5,m) (x(1,m) )
2i1
rm
i=2
2
2
+rm wm (x(1,m) , m , m )[|m | + |m | ] + (m, ) (| |)

3
2
q(2,m,k) (x(1,k) ), q(3,m) (x(1,m) , m ),

q(5,m) (x(1,m) ), wm (x(1,m) , m , m ), and (m, ) (| |) are given by
where
the
functions
(66)
q(4,m) (x(1,m) , m ),
q(2,m,k) = q(2,m,k) (x(1,k) ) + Gm (2,k,m) (x(1,k) ) + (1 + Gm) (z(1,m) ,k) (|x(1,k) |)

+(cm + 1)
sm +1
i=2
z(i,m) (z(i,m) ,k) (|x(1,k) |)
(67)
q(3,m) = q(3,m) (x(1,m) , m ) + 3(1 + m)2 (cm + 1)

)
q(4,m) = 3m
sm +1
i=2
8cm nm 2
max (Pom )m2
2
om (0,m)
+(cm + 1)z(sm +1,m)
z(sm +1,m) (|x(1,m) |)

2
(0,m)
z(i,m) z(i,m) (|x(1,m) |)
(68)
(s2 m ,sm +1,m)
+(cm + 1)z(sm +1,m) z(sm +1,m) (|x(1,m) |)

q(5,m) = (cm m + 1)m (|x(1,m) |)
(69)
(70)
wm = wm (x(1,m) , m , m ) + 3(cm + 1)
sm +1
i=2
z(i,m) z(i,m) (|x(1,m) |)
5
z(sm +1,m) (|x(1,m) |)
n
8c
m
m
2
2
+3m
max (Pom )m + (cm + 1)z(sm +1,m)
2
2
om (0,m)
(0,m)
3
42
|Km (x(1,m) )| + |(sm,sm +1,m) |
(71)
(m, ) = (m, ) (| |) + (cm + 1)
sm +1
i=2
z(i,m) (z(i,m) , ) (| |)
+(1 + Gmm )(z(1,m) , ) (| |).
(72)
148
The design freedom (1,m) is picked to dominate the functions q(1,m) and q(2,k,m) by
defining
3
1
(1,m) (x(1,m) ) =
sign((1,2,m) (0)) q(1,m) (x(1,m) )
m
4
M
(73)
+ q(2,k,m) (x(1,m) ) + m (x(1,m) )
k=1
with m being a positive function of x(1,m) lower bounded by a positive constant m .

As in [11], the dynamics of the high-gain scaling parameters rm , m = 1 . . . , M, are
designed as

(74)
rm = m Rm (x(1,m) , m , m ) rm m (rm , x(1,m) , m , m )
with m being any continuous non-negative function such that m ( ) = 1 for 0
and m ( ) = 0 for rm with rm being any positive constant. By (74), rm is
monotonically non-decreasing so that by initializing rm (0) 1, we have rm (t) 1
for all time. Rm is chosen such that if rm Rm , then, for any nonnegative value
of rm , the negative terms in (66) dominate over all the positive terms in (66) (ex2
cept for (m, ) (| |) and 2m ). The function m is chosen such that if rm m ,
then, for any value of rm not smaller than 1, the negative terms in (66) dominate
2
over all the positive terms in (66) (again, except for (m, ) (| |) and 2m ). Functions Rm (x(1,m) , m , m ) and m (rm , x(1,m) , m , m ) to satisfy these considerations are
given by
)
2
(x(1,m) ) 8wm (x(1,m) , m , m ) 2
2q(3,m) (x(1,m) , m )(1,m)
,
Rm = max
,
m (x(1,m) )
min(cm om , m cm )
5
2
(75)
[2q(4,m) (x(1,m) , m ) ] , 2q(5,m) (x(1,m) )
m
)
2
m = max 2rm q(3,m) (x(1,m) , m )(1,m)
(x(1,m) ),
q(5,m) (x(1,m) )
1
rm2 (cm + 1)

max
z(i,m)
z(i,m) (2i 52 )

|i = 2, . . . , sm + 1 ,
5
1
3
rm2 m q(4,m) (x(1,m) , m ) rm2 wm (x(1,m) , m , m )
,
.
(2sm 1)(cm m + 1)
min(cm om , cm )
(76)
From the dynamics (74), it is seen that rm remains bounded if x(1,m) , m , and m
remain bounded. From (64), m is a function of rm , x(1,m) , x(sm +1,m) , and m that
1
2s 1
is guaranteed to be bounded if x(1,m) , rm2 |m |2 , x2(sm +1,m) /rm m 2 , and rm2 s2m +1 are
bounded. Hence, boundedness of m follows from boundedness of Vm implying that
149
boundedness of rm follows from boundedness of Vm . For each m {1, . . ., M} at

each time t, one of the following two cases must hold:
Case A: rm > Rm (x(1,m) , m , m ) (which implies that rm = m (rm , x(1,m) , m , m ))
Case B: rm Rm (x(1,m) , m , m ).
From (66), (73), (75), and (76), it is seen that in both cases, an upper bound for
m can be obtained as
1
cm om
2 cm
|m |2 + rm
|(2,3,m) (x(1,m) )||m |2 + x2(1,m) m (x(1,m) )
8
8
2
!
sm +1 z
(|z
(|m |)
(i,m) |)
m
(i,m) z(i,m)
2
+ m2sm 1 + z(1,m) (|z(1,m) |) +
(
+
)
m
2i 5
2
2rm
i=2
2rm 2
!
m 2
(77)
+ (m, ) (| |) +
.
2
2
m rm
Hence, for any combination of the occurrences of the Cases A and B for the M
subsystems, the following inequality is satisfied:
V
m=1
2
rm
c m o m
1
2 cm
|m |2 + rm
|
(x
)||m |2 + x2(1,m) m (x(1,m) )
8
8 (2,3,m) (1,m)
2
!
sm +1 z
(|z(i,m) |) m
m (|m |)
(i,m) z(i,m)
2
+
(m )
+ 2sm 1 + z(1,m) (|z(1,m) |) +
2i 5
2
2rm
i=2
2rm 2
!
M
m 2
.
(78)
+ (m, ) (| |) +
2
m=1
Using standard Lyapunov arguments, Input-to-Output practical Stability (IOpS)
and integral Input-to-Output practical Stability (iIOpS) properties can be inferred
from the Lyapunov inequality (78). Analogous to the analysis in [9], it is inferred
from (78) that the inequality
V H(V ) + 1 + 2
2
(79)
M
holds with 1=M
m=1 0.5 m being an uncertain constant, 2 = m=1 (m, ) (| |),
and with H being a class K function given by

V
1

m = 1, . . . , M ,
H(V ) = min min m V m
(2M + 1)(cmm + 1)

1
V
m = 1, . . . , M ,
min z(1,m) V z(1,m)
(2M + 1)(1 + Gmm )

cH
V
(80)
2M + 1
150
where
M
6
cH = min
m=1

m
o m
cm m
1
, m . (81)
,
,
,
8max (Pom ) 8max (Pcm ) 2 2(cm + 1)V z(i,m)
As in [9], (79) can be used to infer that as t , all solutions tend to a compact
set in which V H1 (lim supt [ 1 + 2 ]); in this compact set, the inequality
1
2
M
m=1 ym 2H (lim supt [ 1 + 2 ]) is satisfied. Also, it can be inferred from
(79) that the following inequalities hold:
1 M 2
ym (t) V (t0 ) +
2 m=1
1 M
m
2 m=1
t
t0
y2m ( )d V (t0 ) +
t M
(m, ) (| ( )|))d + 1 (t t0)
t0 m=1
t M
(m, ) (| ( )|))d + 1 (t t0)
(82)
t0 m=1
where t and t0 are any time instants with t > t0 .

The closed-loop stability properties are summarized in Theorems 13. Also, note
2
that from (50) and (72), (m, ) is a linear combination of (m , ) , (m,

) , (z(1,m) , ) ,
and (z(i,m) , ) , i = 2, . . . , sm + 1. This implies that if enters into each subsystem in a linear fashion, i.e., if (m, ) (| |) is linear in | | and (m , ) (| |) and
(z(i,m) , ) (| |), i = 1, . . . , sm + 1 are quadratic in | |, then it can be inferred from
(78) that L2 disturbance attenuation can be achieved by picking controller parameters appropriately.
Theorem 1. Under Assumptions A1A5, the designed dynamic compensator given
by (21), (29), (64), and (74) achieves Bounded-Input-Bounded-State (BIBS) stability
and IOpS of the closed-loop system with (x, z, r1 , . . . , rM , 1 , . . . , M , x1 , . . . , xM ) considered to be the state of the closed-loop system, the input, and (x, z, x1 , . . . , xM )
the output. Furthermore, practical regulation of x(1,m) , m = 1 . . . , m to zero is
achieved in the presence of bounded disturbances, i.e., M
m=1 |x(1,m) (t)| can be
asymptotically regulated to within as small a value as desired by appropriately tuning the controller parameters.
Theorem 2. Under Assumptions A1A5, given any initial conditions (x(0), z(0)) for
the overall plant state and (rm (0), m (0), xm (0)), m = 1, . . . , M, for the controller
states with rm (0) 1, m = 1, . . . , M, if the disturbance input terms go to zero asymptotically, i.e., if
M
m=1
3
4
sm +1
(m, ) ( (t)) + (m, ) ( (t)) + (z(i,m) , ) ( (t)) 0
i=1
as t , then the signals x(t), z(t), x1 (t), . . . , xM (t) go to zero asymptotically as

t if the controller parameters m , m = 1, . . . , M, are picked to be zero.
Theorem 3. Under Assumptions A1A5 and the additional Assumption A6 below,
given any values of the initial conditions (x(0), z(0)) for the overall plant state and
151
(rm (0), m (0), xm (0)), m = 1, . . . , M, for the controller states with rm (0) 1, m =
1, . . . , M, the designed dynamic controller achieves boundedness of all closed-loop
states.

2
Assumption A6. The values of 0 (m,
) (| (t)|)dt, 0 (m , ) (| (t)|)dt, and

0 (z(i,m) , ) (| (t)|)dt, i = 1, . . . , sm + 1 are finite for all m = 1, . . . , M.
3 Generalized Scaling: Application to Decentralized Control

In Sect. 2, the control design utilized, as a central assumption (see Assumption
A4), bidirectional cascading dominance of the upper diagonal terms (i,i+1,m) of
each subsystem. This assumption can be eliminated using the generalized scaling
technique [14, 17] as shown in this section. In this technique, the observer-context
cascading dominance is induced by choosing the high-gain scaling powers appropriately and the controller-context cascading dominance is rendered irrelevant by
resorting to a backstepping-based controller instead of a high-gain scaling-based
controller. The price paid, however, due to the use of a backstepping-based controller is that the full generality of the structure of the system uncertainties and the
coupling with the appended dynamics cannot be attained in this design, as described
further in Remark 2. For instance, it is not possible to handle the triangular structure
of state-coupled appended dynamics as in (1); instead, the appended dynamics zm
can only be materially driven (in a Lyapunov inequality sense see Assumption
A4) by the subsystem outputs.
The decentralized control design in this section addresses a class of interconnected large-scale systems wherein each subsystem is of the form:
zm = qm (z, x, u,t, )
x(i,m) = (i,m) (z, x, u,t, ) + (i,i+1,m) (x(1,m) )x(i+1,m) , i = 1, . . . , sm 1
x(i,m) = (i,m) (z, x, u,t, ) + (i,i+1,m) (x(1,m) )x(i+1,m)
+ (ism ,m) (x(1,m) )um , i = sm , . . . , nm
..
.
x(nm ,m) = (nm ,m) (z, x, u,t, ) + (nm sm ,m) (x(1,m) )um
ym = x(1,m)
(83)
where, as in Sect. 2, xm = [x(1,m) , . . . , x(nm ,m) ]T Rnm is the state, um R is the input,
ym R is the output, and zm Rnzm is the state of the appended dynamics of the mth
subsystem. M is the number of subsystems, x = [xT1 , . . . , xTM ]T , u = [u1 , . . . , uM ]T , and
z = [zT1 , . . . , zTM ]T . (i,i+1,m) , i = 1, . . . , nm 1 and (i,m) are known continuous scalar
real-valued functions. sm is the relative degree of the mth subsystem. Rn is
the exogenous disturbance input. (i,m) , i = 1, . . . , nm and qm are continuous scalar
real-valued uncertain functions.
152
The design is fundamentally based on our earlier result in [18] where a single
subsystem of form (83) (i.e., M = m = 1), but without appended dynamics z1 and
with slightly stronger assumptions than used here on bounds on functions (i,1) ,
was considered and an output-feedback controller was proposed. The design in [18]
was based on the dynamic high-gain scaling paradigm [11, 20, 23] but introduced
a multiple time scaling through the use of arbitrary (not necessarily successive)
powers of a dynamic high-gain scaling parameter r enabled through a new result
on coupled parameter-dependent Lyapunov inequalities [1618]. The utilization
of non-successive powers of the dynamic high-gain scaling parameter in [18] allowed the removal of the cascading dominance assumption on upper diagonal terms
(i.e., Assumption A4 that the ratios (i,i+1,1) /(i1,i,1) and (i1,i,1) /(i,i+1,1) for
i = 2, . . . , n1 1 are bounded) which was central in the earlier results [11,15,20,23].
The construction in [18] resulted in the cascading dominance property being induced in the scaled system when r was of an appropriate size; the dynamics of r
were then designed to achieve the required properties of the signal r(t). In contrast with [18], we consider here an appended dynamics zm , uncertain parameters
in the bounds on (i,m) , a disturbance input , and introduction of multiple subsystems with nonlinear interconnections. The decentralized extension of the technique
from [18] proves particularly challenging due to the fact that the observer gains in
this approach are designed as functions of the high-gain scaling parameter and thus
tend to amplify subsystem cross-coupling arising from (1,m) as seen in the stability analysis.
3.1 Assumptions
The design is carried out under the Assumptions A1 and A2A4, each of which is
required to hold for all m {1, . . ., M}. Assumption A1 is identically as in Sect. 2
and is not repeated here. Assumptions A2A4 are as follows.
Assumption A2. The inverse dynamics of (83) satisfies the Bounded-InputBounded-State (BIBS) condition that the system given by m = m (x(1,m) )m + vm
is BIBS stable with [x(1,m) , vTm ]T Rnm sm +1 considered the input and m Rnm sm
being the state where the (i, j)th element of the (nm sm ) (nm sm ) matrix
m (x(1,m) ) is defined as
m(i,i+1) (x(1,m) ) = (sm +i,sm +i+1,m) (x(1,m) ) , i = 1, . . . , nm sm 1

m(i,1) (x(1,m) ) =
(i,m) (x(1,m) )
(x
) , i = 1, . . . , nm sm (84)
(0,m) (x(1,m) ) (sm ,sm +1,m) (1,m)
with zeros elsewhere.

Assumption A3. Continuous functions (i,m) (t, x(1,m) , . . . , x(i,m) ) and nonnegative
functions (i, j,m) , (m,k) , k = 1, . . . , M, (m,k) , k = 1, . . . , M, and (m, ) are known
such that
153
!
M
|(1,m) (z, x, u,t, )| m (m,k) (x(1,k) )|x(1,k) |+(m,k) (|zk |) + (m, ) (| |)(85)
k=1

(i,m) (t, x(1,m) , x(2,m) , . . . , x(i,m) )
!

M

(i,m) (z, x, u,t, ) m (m,k) (x(1,k) )|x(1,k) | + (m,k) (|zk |)
k=1
i
+ (i, j,m) (x(1,m) )|x( j,m) x( j,m) |

j=2
+(m, ) (| |) , 2 i nm
(86)
for all t 0, xm Rnm , m = 1. . . . , M, zm Rnzm , m = 1, . . . , M, u RM , and Rn ,

with m being an unknown non-negative constant.
Assumption A4. The zm subsystem is ISpS with ISpS Lyapunov function Vzm
satisfying
M
Vzm zm (|zm |) + m (zm ,k) (|x(1,k) |) + (zm, ) (| |)
(87)
k=1
where m is an unknown non-negative constant, zm is a known class K function,

and (zm ,k) , k = 1, . . . , M and (zm , ) are known continuous non-negative functions.
A class K function V zm exists such that Vzm (zm ) V zm (|zm |) for all zm Rnzm . The
following local order estimates hold as 0+ :
2
(a) M
k=1 (k,m) ( ) = O[zm ( )],
(b) M
k=1 (k,m) ( ) = O[ ],
2
(c) M
k=1 (zk ,m) ( ) = O[ ].
Remark 2. Unlike in Sect. 2, the control design in this section does not require the
cascading dominance assumption on upper diagonal terms (Assumption A4). Instead, the observer-context cascading dominance will be induced through a generalized scaling. The price, however, that one must pay to relax the cascading dominance
assumption is that the assumption on the functions (i,m) needs to be stronger than in
Sect. 2 since a high-gain controller cannot be used due to the fact that the cascading
dominance of upper diagonal terms required in observer and controller contexts
are dual, i.e., the observer-context cascading dominance condition requires ratios
|(i,i+1,m) |/|(i1,i,m) | to be upper bounded while the controller-context cascading
dominance condition requires ratios |(i1,i,m) |/|(i,i+1,m) | to be upper bounded.
Hence, the high-gain observer design requires upper diagonal terms nearer to the
output to be larger while the high-gain controller design requires upper diagonal
terms closer to the input to be larger. Therefore, it is not, in general, possible to
design a high-gain observer and high-gain controller using the generalized scaling technique since either observer-context or controller-context cascading dominance can be induced by the scaling, but not both. The output-feedback design in
154
this section uses the generalized scaling technique for the observer which is then
coupled with a backstepping controller. This constrains the functions (i,m) to be
incrementally linear in unmeasured states and prevents them from having the more
general bound which can be handled in Sect. 2. Specifically, compare Assumption
A3 with Assumption A3. Also, compare Assumption A2 with Assumption A2 and
Assumption A5 with Assumption A4.
3.2 Observer Design

A reduced-order observer for the mth subsystem of the interconnected large-scale
system (83) is given by6
x(i,m) = (i,m) (t, x(1,m) , x(2,m) + f(2,m) (rm , x(1,m) ), . . . , x(i,m) + f(i,m) (rm , x(1,m) ))
+(i,i+1,m) (x(1,m) )[x(i+1,m) + f(i+1,m) (rm , x(1,m) )]
+g(i,m)(rm , x(1,m) )[x(2,m) + f(2,m) (rm , x(1,m) )]
+ (ism ,m) (x(1,m) )um rm h(i,m) (rm , x(1,m) ), 2 i nm
(88)
where rm is a dynamic high-gain scaling parameter, f(i,m) (rm , x(1,m) ) are design functions of rm and x(1,m) which will be picked during the stability analysis, and
g(i,m) (rm , x(1,m) ) = (1,2,m) (x(1,m) )
h(i,m) (rm , x(1,m) ) =
f(i,m) (rm , x(1,m) )

x(1,m)
f(i,m) (rm , x(1,m) )

.
rm
(89)
The dynamics of the high-gain scaling parameter rm will be designed to be of the

form rm = wm (rm , x(1,m) ) with wm being (sm 2)-times continuously differentiable.
rm is initialized greater than 1. The dynamics of rm designed during the stability
analysis will ensure that rm is non-decreasing. Defining the observer errors
e(i,m) = x(i,m) + f(i,m) (rm , x(1,m) ) x(i,m) , 2 i nm ,
(90)
the observer error dynamics are, 2 i nm ,

e(i,m) = (i,m) (i,m) + (i,i+1,m) (x(1,m) )e(i+1,m)
(1,m)
+ g(i,m) (rm , x(1,m) )e(2,m)
g(i,m) (rm , x(1,m) )
(1,2,m) (x(1,m) )
(91)
with e(nm+1,m) = 0 being a dummy variable where, for notational convenience, we

have introduced
For simplicity of notation, we introduce the dummy variables (nm ,nm +1,m) = x(nm +1,m) =
f (nm +1,m) = g(nm +1,m) = 0 and (i,m) 0 for i < 0.
155
(i,m) = (i,m) (t, x(1,m) , x(2,m) + f(2,m) (rm , x(1,m) ), . . . , x(i,m) + f(i,m) (rm , x(1,m) )) ,
(92)
for i = 2, . . . , nm .
Hence, the dynamics of em = [e(2,m) , . . . , e(nm ,m) ]T are
em = m + [A(o,m) + GmCm ]em
(93)
Cm = [1, 0, . . . , 0]
Gm (rm , x(1,m) ) = [g(2,m) (rm , x(1,m) ), . . . , g(nm ,m) (rm , x(1,m) )]T
m = [ (2,m) , . . . , (nm ,m) ]T
(1,m)
(i,m) = (i,m) (i,m) g(i,m) (rm , x(1,m) )
, i = 2, . . . , nm
(1,2,m)
0 (2,3,m)
0 ...
0
0 0 (3,4,m) . . .
..
..
A(o,m) = .
.
.
0
(nm 1,nm ,m)
0 0
...
0
(94)
(95)
where
(96)
(97)
3.3 Controller Design

Define
(i,m) = x(i,m) + f(i,m) (rm , x(1,m) ) , i = 2, . . . , sm .
(98)
The controller for the mth subsystem is designed through backstepping [21] using
the subsystem with states (x(1,m) , (2,m) , . . . , (sm ,m) ) whose dynamics are
x(1,m) = (1,2,m) (x(1,m) )e(2,m) + (1,m) + (1,2,m) (x(1,m) )(2,m)
(i,m) = g(i,m) (rm , x(1,m) )e(2,m) + (i,m) (t, x(1,m) , (2,m) , . . . , (i,m) )
(1,m)
g(i,m) (rm , x(1,m) )
(1,2,m) (x(1,m) )
+(i,i+1,m) (x(1,m) )(i+1,m) , i = 2, . . . , sm 1
(sm ,m) = g(sm,m) (rm , x(1,m) )e(2,m) + (sm ,m) (t, x(1,m) , (2,m) , . . . , (sm ,m) )
(1,m)
g(sm,m) (rm , x(1,m) )
(1,2,m) (x(1,m) )
3
4
+(sm ,sm +1,m) (x(1,m) ) x(sm +1,m) + f(sm+1,m) (rm , x(1,m) )
+ (0,m) (x(1,m) )um .
(99)
156
The backstepping-based controller design follows along similar lines as in [18] except for the introduction of the adaptation parameter m , and more importantly, the
introduction of a parameter m used at a key point in the stability analysis.
Step 1: The backstepping is commenced using the Lyapunov function V(1,m) =
1 2
2 (1,m) with (1,m) = x(1,m) yielding
V(1,m) = x(1,m) (1,2,m) [e(2,m) (2,m) ] + x(1,m)(1,m)
m (rm , x(1,m) )x2(1,m) + (1,2,m) (1,m) (2,m) +
e2(2,m)
2
4rm
m
1 2
4 m (1,m)
+(1,m) (m + 2m m ) + ( m m (1,m) (1,m) )(1,m)
(100)
where (1,m) is an arbitrary positive constant, (2,m) is an arbitrary nonnegative constant, m is a smooth nonnegative function to be picked during stability analysis, m
is an adaptation parameter, m is an (unknown) positive constant (depending on m )
which will be specified during stability analysis, and
(2,m) = (2,m) (2,m)

(rm , x(1,m) )
1
2
2
m rm
x(1,m) (1,2,m)
(x(1,m) )
(1,2,m) (x(1,m) )
!
+m x(1,m) + m (rm , x(1,m) )x(1,m)
(101)
(2,m)
(rm , x(1,m) ) =
(1,m) = 0
(102)
(103)
2 2
2
(1,m) = x2(1,m) + rm
x(1,m) (1,2,m)
(104)
Step i (2 i sm 1): Assume that at step (i 1), a Lyapunov function V(i1,m) has
been designed such that
i1
V(i1,m) m (rm , x(1,m) )x2(1,m) ( j,m) (2j,m) + (i1,i,m) (i1,m) (i,m)

j=2
(i 1)e2(2,m)
+
2
4rm
m
i1 2
+ ( m m (1,m) (i1,m) )(i1,m)

4 m (1,m)
+(i1,m) [m + (2,m) m ]
(105)
where, for j = 3, . . . , i,
( j,m) = ( j,m) (j,m)(t, rm , m , x(1,m) , (2,m) , . . . , ( j1,m) ),

with (j,m) being functions designed in the previous steps of backstepping. Defining
1 2
,
V(i,m) = V(i1,m) + (i,m)
2
(106)
157
and differentiating,
i
V(i,m) m (rm , x(1,m) )x2(1,m) ( j,m) (2j,m) + (i,i+1,m) (i,m) (i+1,m) +

j=2
ie2(2,m)
2
4rm
m
i 2
+ ( m m (1,m) (i,m) )(i,m) + (i,m) [m + (2,m) m ]

4 m (1,m)
where
(i+1,m) = (i+1,m) (i+1,m)

(t, rm , m , x(1,m) , (2,m) , . . . , (i,m) )
(i+1,m)
(107)
)
1
=
+ (i1,i,m) (x(1,m) )(i1,m)

(i,i+1,m) (x(1,m) ) (i,m) (i,m)
+(i,m) (t, x(1,m) , (2,m) , . . . , (i,m) )
(i,m)
x(1,m)
i1
j=2
(i,m)
(i,m)
rm
wm (rm , x(1,m) )
(1,2,m) (x(1,m) )(2,m)
(i,m)
( j,m)
( j,m) (t, x(1,m) , (2,m) , . . . , ( j,m) ) + ( j, j+1,m) (x(1,m) )( j+1,m)

1
(i,m) g(i,m) (rm , x(1,m) )
2
(1,2,m)
(x(1,m) )
!2
i1
(i,m)
(i,m)
+
(1,2,m) (x(1,m) )
g( j,m) (rm , x(1,m) )
x(1,m)
j=2 ( j,m)
5
(i,m)
(i,m)
(2,m)
m
(1,m) (i1,m)
m
m

2
+(m + (1,m) (i,m) ) rm
+
(i,m)
m

1
2
2
rm
g(i,m) (rm , x(1,m) )
= (i1,m) + (i,m)
+ 2
(1,2,m) (x(1,m) )
!2
i1
(i,m)
(i,m)
(x
)
g( j,m) (rm , x(1,m) )
+
x(1,m) (1,2,m) (1,m)
j=2 ( j,m)
(i,m) = (i1,m) (1,m)

(i,m)
(108)
(109)
(110)
where (i,m) is any positive constant.

Step sm : At this step, the control input um is designed as
um = (s m +1,m) (t, rm , m , x(1,m) , (2,m) , . . . , (sm ,m) )

(sm ,sm +1,m) (x(1,m) ) x(sm +1,m) + f(sm +1,m) (rm , x(1,m) )
(111)
158
where (s m +1,m) is defined analogously to (108) with i substituted to be sm and

with (0,m) (x(1,m) ) in the denominator of the first term rather than (i,i+1,m) (x(1,m) ).
(sm ,m) and (sm ,m) are defined as in (109) and (110), respectively, with i substituted
to be sm . The Lyapunov function
V(sm,m) =
1 sm 2
(i,m)
2 i=1
(112)
satisfies
sm
V(sm ,m) m (rm , x(1,m) )x2(1,m) ( j,m) (2j,m) +

j=2
sm e2(2,m)
2
4rm
m
sm 2
4 m (1,m)
+( m m (1,m) (sm ,m) )(sm ,m) + (sm ,m) [m + (2,m) m ].
(113)
Designing the dynamics of the adaptation parameter m as
m = (2,m) m + (1,m) (sm ,m) ,
(114)
and defining
V (sm ,m) = V(sm ,m) +
1
(m m )2 ,
2(1,m)
(115)
we have
sm
V (sm ,m) m (rm , x(1,m) )x2(1,m) ( j,m) (2j,m) +

j=2
2
sm (1,m)
4 m
sm e2(2,m)
2
4rm
m
(2,m)
(2,m) 2
(m m )2 +
.
2(1,m)
2(1,m) m
(116)
The design freedoms in the controller for the mth subsystem are the function
m (rm , x(1,m) ), and the constants (2,m) , . . . , (sm ,m) , (1,m) , and (2,m) . The constants
(2,m) , . . . , (sm ,m) , and (1,m) can be picked to be arbitrary positive constants, (2,m)
can be picked to be an arbitrary nonnegative constant, and the function m must
be chosen to satisfy a lower bound to be specified during the stability analysis
in Sect. 3.4. Note that by picking the dynamics of rm to be of the form rm =
wm (rm , x(1,m) ), the functions (2,m)

, . . . , (s m +1,m) are well-defined and continuous.
3.4 Stability Analysis

Define Mm = [M(2,m) , . . . , M(nm ,m) ]T where
M(i,m) = (i,1,m) (x(1,m) ) + |g(i,m)(rm , x(1,m) )|
(1,1,m) (x(1,m) )
|(1,2,m) (x(1,m) )|
159
and define m to be the (nm 1) (nm 1) matrix with (i, j)th entry
m(i, j) = (i+1, j+1,m) , i = 1, . . . , nm 1 , j = 1, . . . , i

m(i, j) = 0 , i = 1, . . . , nm 2 , j = i + 1, . . ., nm 1.
(117)
The matrix A(o,m) + m satisfies the assumptions of Theorem 1 in [18]. Hence,

given any positive constant m , nonnegative constants q(1,m) , . . . , q(nm 1,m) , and a
positive function Rm (x(1,m) ) 1 exist such that Tm (rm )[A(o,m) + m ]Tm1 (rm ) is
w-CUDD(m ) for all rm Rm (x(1,m) ) where
q
Tm (rm ) = [diag(rm(1,m) , . . . , rm(nm 1,m) )]1 .
(118)
The w-CUDD (weak Cascading Upper Diagonal Dominance) property was defined
in [17] and shown to be central in solvability of coupled Lyapunov inequalities [16].
From the construction in the proof of Theorem 2 in [18], q(1,m) can be taken to be 1
and7 q(nm 1,m) q(nm 2,m) = 1. Using Theorem 1 from [18], a (nm 1) 1 vector
G m (rm , x(1,m) ), a symmetric positive-definite matrix P(o,m) , and positive constants
(o,m) , (o,m) , and (o,m) exist such that for all rm Rm (x(1,m) ) and all x(1,m) R,
7
8
P(o,m) Tm (rm )[A(o,m) + Q(1,m) m Q(2,m) ]Tm1 (rm ) + G mCm
8T
7
+ Tm (rm )[A(o,m) + Q(1,m) m Q(2,m) ]Tm1 (rm ) + G mCm P(o,m)
(o,m)
q(nm 2,m) q(nm 1,m) |(nm 1,nm ,m) |I
rm
(o,m) I P(o,m) D(o,m) + D(o,m) P(o,m) (o,m) I
(119)
where D(o,m) = diag(q(1,m) , . . . , q(nm 1,m) ) and Q(1,m) and Q(2,m) are arbitrary diagonal matrices of dimension (nm 1) (nm 1) with each diagonal entry +1 or 1.
By Theorem 1 in [18], the choice of G m does not need to depend on Q(1,m) and
Q(2,m) . Gm (rm , x(1,m) ) = [g(2,m) (rm , x(1,m) ), . . . , g(nm ,m) (rm , x(1,m) )]T is defined as
q(1,m)
Gm (rm , x(1,m) ) = rm
Tm1 (rm )G m (rm , x(1,m) )
(120)
so that G mCm = Tm (rm )GmCm Tm1 (rm ). f (i,m) , i = 2, . . . , nm , are obtained as

f(i,m) (rm , x(1,m) ) =
x
(1,m) g(i,m) (rm , )
0
(1,2,m) ( )
d .
(121)
The dynamics of m = Tm (rm )em are

Note that since the observer used is a reduced-order observer, A(o,m) is of dimension (nm 1)
(nm 1). Hence, by the construction in the proof of Theorem 2 in [18], q(i,m) = 1 + (nm 2)(nm
1)/2 (nm i 1)(nm i)/2, i = 1, . . ., nm 1.
7
160
m = Tm (rm ) m + Tm (rm )[A(o,m) (x(1,m) ) + Gm (rm , x(1,m) )Cm ]Tm1 (rm )m

rm
D(o,m) m .
(122)
rm
The derivative of the Lyapunov function V(o,m) = mT P(o,m) m satisfies
V(o,m) = 2mT P(o,m) Tm (rm ) m + mT {P(o,m) Tm (rm )[A(o,m) + GmCm ]Tm1 (rm )
+Tm1 (rm )[A(o,m) + GmCm ]T Tm (rm )P(o,m) }m
rm
mT [P(o,m) D(o,m) + D(o,m) P(o,m) ]m .
rm
(123)
The scaling m = Tm (rm )em which comprises a scaling with not necessarily successive powers q(1,m) , . . . , q(nm 1,m) of the scaling parameter rm essentially yields a
multiple time scaling and is the key ingredient in allowing the removal of the cascading dominance assumption by using the Theorems 1 and 2 in [18]. In common
with [11] and as in Sect. 2, the dynamics of the high-gain parameter are designed to
be of the form
rm = m (Rm rm )m (rm , x(1,m) )
(124)
with initial value rm (0) 1 and with m being an appropriately designed function. m is chosen to be any non-negative (sm 2)-times continuously differentiable
function such that m ( ) = 1 for 0 and m ( ) = 0 for rm with rm being
any positive constant. In contrast with the design in [18] where a single subsystem
of form (83) was considered without appended dynamics, the function m should
be chosen through a careful bounding of the term 2mT P(o,m) Tm (rm ) m since this
term will generate cross-products of the form g(i,m) (rm , x(1,m) )x(1,k) which cannot be
handled in the composite Lyapunov function framework. To see the origin of such
cross-products, note that8
| m |e
T
2m P(o,m) Tm (rm ) m
e m |em |e + Qm
(125)
T
1
2m P(o,m) Tm (rm )Q(1,m) m Q(2,m) Tm (rm )m
+2|mT P(o,m) |e Tm (rm )Qm
(126)
and observe that the bound on (1,m) arising from Assumption A3 involves x(1,k) ,
zk , and . In (125)-(126),
M
Q(i,m) = m [(m,k) (|x(1,k) |)|x(1,k) | + (m,k) (|zk |)]

k=1
+(m, ) (| |) +
|g(i,m) |
| , i = 2, . . . , nm
|
|(1,2,m) | (1,m)
Qm = [Q(2,m) , . . . , Q(nm ,m) ]T ,
(127)
(128)
| |e denotes a matrix of the same dimension as with each element replaced by its absolute
value. e denotes an element-wise inequality between two matrices of equal dimension.
161
and Q(1,m) and Q(2,m) are diagonal matrices with each diagonal entry +1 or 1 such
that |P(o,m) m |e = Q(1,m) P(o,m) m and |m |e = Q(2,m) m . To handle the bounding of
the term 2mT P(o,m) Tm (rm ) m , consider the two cases:
Case A: rm > Rm (x(1,m) )
Case B: rm Rm (x(1,m) ).
For each subsystem m {1, . . . , M} at each time t, one of these two cases must
hold. Under Case A, it is inferred from Theorem 1 in [18] and the construction in
the proof of Theorem 3 in [16] that a positive constant G exists such that |g(i,m) |
Grq(i1,m) |(1,2,m) | for i = 2, . . . , nm . Also, since (119) holds under Case A, it follows
that
V(o,m)
(o,m) m
2
|m |2 +
2 (P
)|Q m |2
2
(o,m) m max (o,m)
(129)
m = [Q (2,m) , . . . , Q (n ,m) ]T with

where Q
m
M
Q (i,m) = m [(m,k) (|x(1,k) |)|x(1,k) | + (m,k) (|zk |)]

k=1
+(m, ) (| |) + G|(1,m) | i = 2, . . . , nm .
(130)
Under Case B, it follows from (124) that rm = m (rm , x(1,m) ). It can be shown that
the term 2mT P(o,m) Tm (rm ) m can be bounded as
2mT P(o,m) Tm (rm ) m 2mT P(o,m) Tm (rm )Q(1,m) m Q(2,m) Tm1 (rm )m
8
7
max g2(i,m) |i = 2, . . . , nm

(o,m) m
|m |2 max
+
,
1
2
2
G 2
(1,2,m)
(o,m) m
2
max
(P(o,m) )|Q m |2 .
(131)
Hence, designing m (rm , x(1,m) ) to be
m (rm , x(1,m) )
4
rmnm 1 3
+
G
C
||
+
||
||
||A
m m
m
q
(o,m)
(o,m)
rm1
7
8
max g2(i,m) |i = 2, . . . , nm

(o,m) m
max
+
,
1
(132)
2 2
2
G (1,2,m)
rm
+ 2max (P(o,m) )
rm
being any positive constant, it follows that

with rm
V(o,m) rm
|m |2 +
2
2 (P
)|Q m |2 .
(o,m) m max (o,m)
(133)
162
Therefore, in either of the Cases A and B, the inequality

V(o,m) min
(o,m) m
|m |2 +
, rm
(o,m) m
2
max
(P(o,m) )|Q m |2
(134)
holds.
2
+
By Assumption A4, M
k=1 (k,m) ( ) = O[zm ( )] as 0 . Using a reasoning
similar to that used in the proof of Theorem 2 in [26], it is seen that this local order
estimate implies the existence of a new Lyapunov function Vzm , class K functions
zm and (zm ,k) , and continuous non-negative functions m and (zm , ) such that
M
Vzm zm (|zm |) + m (m ) (zm ,k) (|x(1,k) |) + (zm, ) (| |)
(135)
k=1
2
nzm ,
with zm ( ) = O[zm ( )] as 0+ , zm (|zm |) M
k=1 (k,m) (|zm |)zm R
(zm ,k) independent of m , and (zm ,k) ( ) = O[(zm ,k) ( )] as 0+ . Hence, a continuous non-negative function (zm ,k) exists such that
(zm ,k) (|x(1,k) |) x2(1,k) (zm ,k) (x(1,k) ).
(136)
Furthermore, it can be shown that a class K function V zm exists such that Vzm (zm )
V zm (|zm |) for all zm Rnzm .
Defining
Vxm = V (sm ,m) +
sm + 1
2 m min(
(o,m) m V(o,m) ,
, rm )
2
(137)
and using (134) and (116), we obtain

sm
1
Vxm m (rm , x(1,m) )x2(1,m) ( j,m) (2j,m) sm |m |2
2
j=2

M
(2,m)
(2,m) 2 Qm0
2
(m m )2 +
m +
(|x(1,k) |)x2(1,k)
Mm2 (m,k)
2(1,m)
2(1,m)
m
k=1

M
2
2
+M m2 (m,k)
(|zk |) + (m,
(138)
) (| |)
k=1
where Qm0 is a constant given by

Qm0 =
2 (P
3max
sm
2
(o,m) )[sm + 1]
+
m [3MG + 2(nm 1)]

4
(o,m) m min( (o,m)
,
r
)
m
2
and sm is an unknown positive constant defined to be 1/[2 m ]. Note that Qm0 does
not depend on m . At this point, we can choose the (unknown) constants 1 , . . . , M
to be where
163
= max max{Qk0 |k = 1, . . . , M},

2
1
2
1
M max Qk0 k2 |k = 1, . . . , M max k (k ), 1 }.
(139)
Note that m is a constant used only in stability analysis and does not enter anywhere
into the observer or controller equations. The overall composite Lyapunov function
of the large-scale interconnected system is picked to be
V =
!
Qk0 M k2
Vxm + 2
Vzm .
k
k=1
m=1
(140)
We obtain
V
m=1
sm
1
m (rm , x(1,m) )x2(1,m) + ( j,m) (2j,m) + sm |m |2
2
j=2
M Q M 2
(2,m)
k0
k
(m m )2 +
zm (|zm |)
2(1,m)
k
k=1
!
(2,m) 2
2
m (|x(1,m) |)x(1,m) + (| |)
2(1,m) m
+
(141)
where
m (|x(1,m) |) =
(| |) =
2
(|x(1,m) |) + (zk ,m) (|x(1,m) |)]
[(k,m)
(142)
2
[(zm, ) (| |) + (m,
) (| |)].
(143)
k=1
M
m=1
Picking the design function m to be
m (rm , x(1,m) ) = (1,m) + m (|x(1,m) |),
(144)
where (1,m) , m = 1, . . . , M, are any positive constants, (141) reduces to

V
+
sm
m=1
j=1
(2,m)
( j,m) (2j,m) + 2 sm |m |2 + 2(1,m) (m m )2
!
(2,m) 2
Qk0 M k2
+ (| |)
(|z
|)
z
m

m
2(1,m) m
k
k=1
M
(145)
By straightforward signal chasing and using the BIBS property in Assumption A2,
it can be shown that all closed-loop signals remain bounded on the maximal interval
of existence [0,t f ) implying that t f = and that solutions exist for all time. Furthermore, using standard Lyapunov arguments, Theorems 13 follow from (145).
2
Also, note that is a linear combination of (m,

) and (zm , ) . This implies that
if enters into each subsystem in a linear fashion, i.e., if (m, ) (| |) is linear in
164
| | and (zm, ) (| |) is quadratic in | |, then it can be inferred from (145) that L2

disturbance attenuation is achieved by picking controller parameters appropriately.
Theorem 1. Under Assumptions A1 and A2A4, the designed dynamic compensator achieves BIBS stability and IOpS of the closed-loop system with
(x, z, r1 , . . . , rM , 1 , . . . , M , x1 , . . . , xM ) considered to be the state of the closed-loop
system (where xm denotes [x(2,m) , . . . , x(nm ,m) ]T ), being the input of the closed-loop
system, and (x(1,1) , . . . , x(1,M) , z1 , . . . , zM ) being the output. Furthermore, practical
regulation of x(1,m) , m = 1 . . . , M, to zero is achieved in the presence of bounded disturbances, i.e., M
m=1 |x(1,m) (t)| can be asymptotically regulated to within as small a
value as desired by appropriately tuning the controller parameters.
Theorem 2. Under Assumptions A1 and A2A4, given any values of the initial
conditions (x(0), z(0)) for the overall plant state and (rm (0), m (0), xm (0)), m =
1, . . . , M, for the controller states with rm (0) 1, m
3 = 1, . . . , M, if the disturbance4 input terms go to zero asymptotically, i.e., if M
m=1 (m, ) ( (t)) + (zm , ) ( (t))
0 as t , then the signals x(t), z(t), e1 (t), . . . , eM (t) where x = [xT1 , . . . , xTN ]T
with xm = [x(1,m) , . . . , x(sm ,m) ]T go to zero asymptotically as t if the controller parameters (2,m) , m = 1, . . . , M, are picked to be zero. Furthermore, if
the BIBS Assumption A2 is strengthened to a minimum phase assumption, then
x(t), z(t), x1 (t), . . . , xM (t) go to zero asymptotically as t .
Theorem 3. Under Assumptions A1 and A2A4 and the additional Assumption
A5 below, given any values of the initial conditions (x(0), z(0)) for the overall
plant state and (rm (0), m (0), xm (0)), m = 1, . . . , M, for the controller states with
rm (0) 1, m = 1, . . . , M, the designed dynamic controller achieves boundedness of
all closed-loop states.
Assumption A5: The values of
finite for all m = 1, . . . , M.
2
(m, ) (| (t)|)dt
and

0
(zm , ) (| (t)|)dt are
Acknowledgement This work was supported in part by the NSF under grant ECS-0501539.
References
1. Siljak
DD (1978) Large-Scale Dynamic Systems: Stability and Structure. North-Holland,
New York
2. Ilchmann A (1996) High-gain adaptive control: an overview. In: IEE Colloquium on Adaptive
Control (Digest No: 1996/139), London, UK, pp 14
3. Jain S, Khorrami F (1997) Decentralized adaptive control of a class of large-scale interconnected nonlinear systems. IEEE Transactions on Automatic Control 42(2):136154
4. Jain S, Khorrami F (1997) Decentralized adaptive output feedback design for large-scale nonlinear systems. IEEE Transactions on Automatic Control 42(5):729735
5. Jamshidi M (1983) Large-Scale Systems: Modeling and Control. North-Holland, New York
6. Jiang ZP, Teel A, Praly L (1994) Small-gain theorem for ISS systems and applications. Mathematics of Control, Signals and Systems 7:95120
165
7. Khalil HK (1996) Adaptive output feedback control of nonlinear systems represented by inputoutput models. IEEE Transactions on Automatic Control 41(2):177188
8. Khalil HK, Saberi A (1982) Decentralized stabilization of nonlinear interconnected systems
using high-gain feedback. IEEE Transactions on Automatic Control 27(1):265268
9. Krishnamurthy P, Khorrami F (2003) Decentralized control and disturbance attenuation for
large-scale nonlinear systems in generalized output-feedback canonical form. Automatica
39:19231933
10. Krishnamurthy P, Khorrami F (2004) Conditions for uniform solvability of parameterdependent Lyapunov equations with applications. In: Proceedings of the American Control
Conference, Boston, MA, pp 38963901
11. Krishnamurthy P, Khorrami F (2004) Dynamic high-gain scaling: state and output feedback
with application to systems with ISS appended dynamics driven by all states. IEEE Transactions on Automatic Control 49(12):22192239
12. Krishnamurthy P, Khorrami F (2004) A high-gain scaling technique for adaptive output
feedback control of feedforward systems. IEEE Transactions on Automatic Control 49(12):
22862292
13. Krishnamurthy P, Khorrami F (2005) Adaptive output-feedback for nonlinear systems with no
a priori bounds on parameters. In: Proceedings of the American Control Conference, Portland,
OR, pp 37133718
14. Krishnamurthy P, Khorrami F (2005) Generalized state scaling-based robust control of nonlinear systems and applications to triangular systems. In: Proceedings of the American Control
Conference, Portland, OR, pp 34273432
15. Krishnamurthy P, Khorrami F (2006) Application of the dual high-gain scaling technique to
decentralized control and disturbance attenuation. In: Proceedings of the IEEE Conference on
Decision and Control, San Diego, CA
16. Krishnamurthy P, Khorrami F (2006) On uniform solvability of parameter-dependent lyapunov inequalities and applications to various problems. SIAM Journal on Control and Optimization 45(4):11471164
17. Krishnamurthy P, Khorrami F (2007) Generalized state scaling and applications to feedback,
feedforward, and non-triangular nonlinear systems. IEEE Transactions on Automatic Control
52(1):102108
18. Krishnamurthy P, Khorrami F (2007) High-gain output-feedback control for nonlinear systems
based on multiple time scaling. Systems and Control Letters 56(1):715
19. Krishnamurthy P, Khorrami F, Jiang ZP (2002) Global output feedback tracking for nonlinear systems in generalized output-feedback canonical form. IEEE Transactions on Automatic
Control 47(5):814819
20. Krishnamurthy P, Khorrami F, Chandra RS (2003) Global high-gain-based observer and backstepping controller for generalized output-feedback canonical form. IEEE Transactions on
Automatic Control 48(12):22772284
21. Krstic M, Kanellakopoulos I, Kokotovic PV (1995) Nonlinear and Adaptive Control Design.
Wiley, New York
uner U
(1979) Near optimal control of composite systems: the multi-time-scale approach.
22. Ozg
IEEE Transactions on Automatic Control 24(4):652655
23. Praly L (2003) Asymptotic stabilization via output feedback for lower triangular systems with
output dependent incremental rate. IEEE Transactions on Automatic Control 48(6):11031108
24. Sepulchre R, Jankovic M, Kokotovic PV (1997) Integrator forwarding: a new recursive nonlinear robust design. Automatica 33(5):979984
25. Shi L, Singh SK (1992) Decentralized adaptive controller design for large-scale systems with
higher order uncertainties. IEEE Transactions on Automatic Control 37(8):11061118
26. Sontag ED, Teel A (1995) Changing supply functions in input/state stable systems. IEEE
Transactions on Automatic Control 40(8):14761478
Decentralized Output Feedback Guaranteed

Cost Control of Uncertain Markovian Jump
Large-Scale Systems: Local Mode Dependent
Control Approach
Junlin Xiong, Valery A. Ugrinovskii, and Ian R. Petersen
1 Introduction
Large-scale systems provide a mathematical model to represent dynamic systems
that have a complex structure and high dimensions. Many physical systems, such
as power systems [16], can be modeled as large-scale systems. In a large-scale system setting, one often lacks the system information that is necessary to implement
a centralized controller design, or the complexity of a centralized design and cost
of implementing centralized controllers is prohibitive. In such cases, decentralized
controllers that only use locally available subsystem information provide a viable alternative to centralized solutions. In many practical problems involving large-scale
systems, the decentralized control technique can effectively deal with such issues
as information structure constraints, large dimensions and a broad range of uncertainties. However, the design of decentralized controllers is challenging for at least
two reasons. Firstly, not all the system information is available to the controllers.
Secondly, the dynamics of each subsystem in the large-scale system are affected
by other subsystems [19] via the interconnections; in many situations the effect of
those interconnections is essentially uncertain.
A large-scale system may experience abrupt changes in system parameters, due
to, for instance, random failures of its components, sudden variations of the environment, and changes of the subsystem working conditions. In this case, the Markovian
jump large-scale system model can serve as a suitable model to capture those abrupt
system parameter changes. Specifically, a Markovian jump large-scale system can
This
work was supported by the Australian Research Council.
J. Xiong
Department of Automation, University of Science and Technology of China, Hefei 230026, China
e-mail: junlin.xiong@gmail.com
V.A. Ugrinovskii and I.R. Petersen
School of Engineering and Information Technology, University of New South Wales
at the Australian Defence Force Academy, Northcott Drive, Canberra, ACT 2600, Australia
e-mail: v.ugrinovskii@gmail.com; i.r.petersen@gmail.com
167
168
J. Xiong et al.
be seen as an interconnection of a finite number of subsystems, and each subsystem

has a finite number of operation modes. The abrupt changes of system parameters
cause the subsystems to change their operation modes. Furthermore, in such a largescale system, these operation mode changes are governed by a Markov process that
takes values in a finite set. Markovian jump large-scale systems may be regarded
as a generalization of Markovian jump systems [10] and normal large-scale systems [16]. Such a system reduces to a Markovian jump system when the system is
regarded as consisting of only one subsystem. Also, such a Markovian jump largescale system turns into a regular large-scale system when each subsystem has only
one operation mode.
In the recent control literature, much attention has been given to Markovian largescale systems subject to uncertain perturbations; e.g., see [7, 9, 18, 22]. In particular,
[7, 18] considered a class of uncertain Markovian jump parameter systems in which
both local uncertain perturbations and subsystem interconnections were described in
terms of integral quadratic constraints (IQCs) [14,15]. In these references, necessary
and sufficient conditions were derived for absolute stabilization of the system using
decentralized state and output feedback controllers, respectively. An important underlying assumption required to implement the controllers proposed in many available results including the above references is that the global operation mode of the
large-scale system must be known to every controller; we refer to such controllers as
global mode dependent controllers. In a global mode dependent decentralized control design, each subsystem is equipped with its own switching feedback controller
which utilizes the subsystem dynamic state or output. Furthermore, the number of
operation modes for each such controller is equal to the number of global operation
modes of the system, and hence is greater than the number of local operation modes
of the subsystem it controls. As a result, the controller has to change its operation
mode even if the subsystem it controls does not change. Also, to implement such a
control algorithm, one needs to ensure that the global mode information is available
to every subsystem controller. Thus, to implement a global mode dependent control
scheme in practice, this information must be collected and broadcast to each subsystem in real time. Collecting and broadcasting operation modes of each subsystem
may be expensive or even impossible in some situations. Hence it is desirable to
lessen the communication overheads by reducing the dependency of the decentralized controller on the knowledge of the global operation mode information. A local
mode dependent approach has been recently developed where the mode of the decentralized controller only depends on the mode of the subsystem it controls [20].
These types of decentralized controllers are referred to as local mode dependent
controllers. The result in [20] gives a sufficient condition and an algorithm for the
design of local mode dependent stabilizing controllers under the assumption of full
state feedback.
This chapter is a further development of the local mode dependent approach proposed in [20]. It extends the results of [20] in several directions. Firstly, in this
chapter we address an output feedback stabilization problem via the approach of
local mode dependent control. Our result leads to an algorithm for the design of
dynamic output feedback controllers of full order. In [20], static state feedback controllers were considered. Secondly, unlike [20] which focused on the stabilization
Decentralized Local Mode Dependent Control of Uncertain Markovian Jump Systems
169
problem, this chapter addresses performance of the closed-loop system. Here we

consider an output feedback guaranteed cost control problem that is similar to that
in [7]. Our control design method leads to a set of local mode dependent controllers
which are suboptimal with respect to a given quadratic performance cost functional.
Thirdly, in [20], the system model was somewhat limited in that the uncertainty outputs employed in the definition of the admissible uncertainty used in that paper did
not allow for control input feed-through. This chapter overcomes this limitation and
reintroduces the control input feed-through term in the definition of the uncertainty
output. This however leads to several additional technical difficulties, which precludes us from following the approach undertaken in [20]; in particular the results
developed in [7, 18, 20] cannot be used in the derivation of local mode dependent
output feedback controllers. Hence, the control design technique used in this chapter
is different from that used in the previous work. To develop the solution, a version of
bounded real lemma [21] is adopted to tackle the control input feed-through terms,
and the projection lemma [5] is used to derive the controller.
The decentralized guaranteed cost control problem for a class of Markovian jump
large-scale systems studied in this chapter involves two classes of uncertainties
described using IQCs [15]. Local uncertainties are the uncertainties that affect the
dynamics of the subsystems. Also, we consider subsystem interconnection uncertainties which describe the interconnections between the subsystems. Our approach
proceeds from the observation that the replacement of a global mode dependent
control action with a local mode dependent control action can be interpreted as a
result of a fictitious mode mismatch uncertainty in the global mode dependent
controller; such an fictitious uncertainty acts to cancel out the information gained
from the knowledge of the global operation mode information. This idea prompts
us to propose the following two-step controller design algorithm to design the local
mode dependent output feedback controllers. Firstly, an auxiliary class of uncertain global mode dependent controllers that stabilize the class of uncertain Markovian jump large-scale systems under consideration is designed; the uncertainty in
the controller is used to represent the mode mismatch uncertainty. To design such
controllers, the uncertain system is augmented to include the possibility of the controller uncertainty. The controller uncertainty is also described using IQCs. Next,
the local mode dependent controller parameters are designed to be the limit (as time
approaches infinity) of the conditional mean value of the global mode dependent
controller parameters conditioned on the corresponding subsystem modes. A sufficient condition in terms of a set of rank constrained linear matrix inequalities is
established to show that the proposed global and local mode dependent output controllers are guaranteed cost controllers. That is, we show that the worst case value
of the performance cost functional computed along trajectories of the closed-loop
system is bounded.
The organization of the chapter is as follows. Section 2 formulates the problem
to be studied. The main results are developed in Sect. 3. An illustrative example and
simulations are given in Sect. 4, and Sect. 5 concludes the chapter.
Notation. R+ denotes the set of positive real numbers. Rn , Rnm , and S+ denote, respectively, the n-dimensional Euclidean space, the set of n m real matrices,
170
J. Xiong et al.
and the set of real symmetric positive definite matrices of compatible dimensions.
Given a matrix A Rmn with r = rank(A) < m, A R(mr)m is an orthogonal
complement matrix of the matrix A; A satisfies A A = 0 and rank(A ) = m r.
rank(A) n means that the matrix A is real symmetric positive semi-definite and
rank(A) n.
2 Problem Formulation
Consider an uncertain Markovian jump large-scale system consisting of N subsystems. The i-th subsystem is described by the equation
x (t) = Ai (i (t))xi (t) + Bi(i (t))ui (t) + Ei (i (t))i (t) + Li (i (t))ri (t),
i
(1)
Si : i (t) = Hi (i (t))xi (t) + Gi (i (t))ui (t),
yi (t) = Ci (i (t))xi (t) + Di (i (t))i (t),

where i N {1, 2, . . . , N} indicates that Si is the ith subsystem of the large-scale
system, xi (t) Rni is the system state of subsystem Si , ui (t) Rmi is the control input, yi (t) Rti is the measured output which will be used for feedback. i (t) Rqi
is the uncertainty output, i (t) R pi is the local uncertainty input, ri (t) Rsi is the
interconnection input, which describes the interconnection effect of subsystems S j
( j N, j = i) on Si due to the interconnection between subsystem Si and subsystems
S j . The interconnection input ri (t) is treated as an uncertainty input in this chapter.
The random process i (t) denotes the operation mode of subsystem Si ; it takes values in the finite state space Mi {1, 2, . . . , Mi }. The initial condition of subsystem
Si is given by xi0 Rni and i0 Mi .
For the large-scale system, a vector mode process (1 (t), 2 (t), . . . , N (t)) can be
defined to indicate the collection of operation modes of all the subsystems. Also, an
operation mode pattern set Mp is introduced as the set of all vector mode states that
can be visited by the vector mode process (1 (t), 2 (t), . . . , N (t)). The operation
mode pattern set Mp is a non-empty subset of the set M1 MN . Suppose
Mp has M elements where maxiN Mi M Ni=1 Mi , then we say that the largescale system has M global operation modes in total. Let M {1, 2, . . . , M}, then
there exists a bijective function : Mp M with = (1 , . . . , N ). Also we
can define functions i1 : M Mi with i = i1 ( ) for all i N. By virtue of
the bijective function, we say the large-scale system is in global mode j at time
t if (1 (t), . . . , N (t)) = j. Formally, the global mode process (t) is defined as
(t) (1 (t), . . . , N (t)). Now, the mechanism of mode change for the large-scale
system can be described by the global mode process (t). The random process (t)
in this chapter is assumed to be a stationary ergodic continuous-time Markov process
defined on a complete probability space ( , F, Pr). Moreover, the state transition
rate matrix of (t) is assumed to be known and given by Q = (q ) RMM , in
which q 0 if = , and q M
=1, = q .
171
Remark 7.1. Some points about the local operation modes of the subsystems in the
large-scale system and the global operation mode of the large-scale system need to
be clarified.
1. Because of the existence of the bijective function between Mp and M, the global
operation mode process (t) carries the same information as in the local operation mode vector process (1 (t), 2 (t), . . . , N (t)). Therefore, when the global
mode dependent control techniques (for example, those in [7, 18]) are used in
practical applications, all the local operation modes need to be collected and
broadcast throughout the large-scale system. However, collecting and broadcasting all the local operation modes may be expensive or even impossible
in some cases. On the other hand, if local mode dependent control techniques
(such as, for example, the one developed in [20] and the one to be developed
in this work) are used, then collecting and broadcasting the global mode information is not required. This observation motivates our research into local mode
dependent control as one approach to reducing communication overheads of the
existing decentralized control techniques for Markov jump parameter systems.
2. The assumption that the global operation mode process (t) is a Markov process is reasonable and is motivated by the fact that such a stochastic process
model can be used to describe random events in many applications [10].
3. In this work, the global operation mode process (t) is not assumed to be decomposable or nearly decomposable into smaller Markov processes. Therefore,
the large-scale system cannot be broken into several independent subsystems,
governed by independent Markov chains. On the contrary, the local mode processes i (t) are in general non-Markovian, and dependent on each other; an
illustrative example of this will be given in Sect. 4. This observation in general
prohibits the design of local mode dependent controllers by considering each
subsystem as an isolated Markovian jump system.
The uncertainties and interconnections in the large-scale system (1) are described
by equations of the form
i (t) = i (t, i ()|t0 , i ()|t0 ),

ri (t) = ir (t, 1 ()|t0 , . . . , i1 ()|t0 , i+1 ()|t0 , . . . , N ()|t0 , ()|t0 ).
These uncertainties are assumed to satisfy the following IQCs [7, 15, 18].
Definition 7.1. Given a set of matrices Si S+ , i N. A collection of local uncertainty inputs i (t), i N, is an admissible local uncertainty for the large-scale
system if there exists a sequence {tl }
l=1 such that tl , tl 0 and
t 3

4
l
(2)
E
i (t)2 i (t)2 dt | x0 , 0 xTi0 Si xi0
0
T

for all l and for all i N, where x0 xT10 xTN0 , and 0 (0). The set of the
admissible local uncertainties is denoted by .
172
J. Xiong et al.
Definition 7.2. Given a set of matrices Si S+ , i N. The large-scale system is said

to have admissible interconnections between subsystems if there exists a sequence
{tl }
l=1 such that tl , tl 0 and
(

'
N
tl

2
2
ri (t) j (t)
dt | x0 , 0 xTi0 Si xi0
(3)
E
0
j=1, j=i
for all l and for all i N. The set of the admissible interconnections is denoted
by r .
Without loss of generality, we assume that the same sequence {tl }
l=1 is employed
in both definitions.
Remark 7.2. The IQC-type descriptions in (2), (3) can capture a broad class of linear
and nonlinear, time-invariant and time-varying, static and dynamic uncertainties and
interconnections. Time delays may also be included provided the solution of (1) exists and the process (x(t), (t)) is jointly Markov. For example, the norm-bounded
uncertainty of the form i (t) = i (t)i (t) satisfying i (t) 1 belongs to because it satisfies the IQC in (2) for any Si S+ and any sequence {tl }
l=1 . Also, the
dynamic interconnections represented in the operator form as ri (s) = r,i (s)r,i (s)
belong to r if the transfer function

r,i (s) is from the Hardy space

RH such that
r,i (s) 1, where r,i (s) = 1 (s) i1 (s) i+1 (s) N (s) .
Consider a decentralized local mode dependent output feedback controller, whose
i-th component has the form
)
xK,i (t) = AK,i (i (t))xK,i (t) + BK,i (i (t))yi (t),
(4)
Ki :
ui (t) = CK,i (i (t))xK,i (t) + DK,i (i (t))yi (t),
where xK,i (t) Rni is the state of the controller. The initial controller state is set
to zero, and the initial operation mode is set to that of the i-th subsystem Si ; see
(1). The matrices AK,i (i ), BK,i (i ), CK,i (i ), DK,i (i ), i Mi , are the parameters
of the controller to be designed. It is worthwhile to emphasize that for controllers
of the form (4) under consideration, the controller parameters are determined by the
local operation mode of the subsystem while the controllers proposed in [7,18] were
dependent on the global operation mode defined by values of (t).
Definition 7.3. The closed-loop system corresponding to the uncertain system (1),
(2), (3) with the controller (4) is said to be robustly stochastically stable if there
exists a constant c1 R+ such that xi () L2 [0, ) for all i N and
N
i=1

0

xi (t) dt | x0 , 0
2
c1 xi0 2
(5)
i=1
for any initial conditions x0 , 0 , any admissible local uncertainty i (t) and any admissible interconnection ri (t), i N.
173
Associated with the large-scale system (1) is the quadratic cost functional of the
form

N

T
(6)
xi (t)Qi (i (t))xi (t) + uTi (t)Ri (i (t))ui (t) dt | x0 , 0
J E
i=1
where Qi (i ) S+ , Ri (i ) S+ , i Mi , i N, are given weighting matrices.

The objective of the chapter is to design a dynamic output feedback controller
of form (4) for the uncertain system (1), (2), (3), such that the resulting closedloop system is robustly stochastically stable and the corresponding worst case
value of the cost functional (6) subject to the constraints (2) and (3) is upper
bounded.
3 Guaranteed Cost Controller Design

This section presents the main results of the chapter. Our local mode dependent
controller design technique is based on the decentralized global mode dependent
control with controller uncertainties. The design methodology involves augmenting
the uncertain system to include effects of mismatch between the global operation
mode and the local operation mode controllers. This design methodology is described in Sect. 3.1, where we show how a local mode dependent controller can be
derived from a given global mode dependent controller; the result of that section is a
sufficient condition to ensure that such a derivation is possible. The design of a suitable auxiliary global mode dependent controller is described in Sect. 3.2. Here we
present a sufficient condition for the existence of such an auxiliary output feedback
controller. Note that the results of [7] cannot be used in the derivation since there are
controller uncertainties in the global mode dependent controllers considered in this
work, and also the controllers considered here are more general than those considered in [7]. As a result, the algorithm to design the auxiliary global output feedback
controller is different from that proposed in [7]. In Sect. 3.3 we propose a local mode
dependent controller design technique based on the auxiliary controller presented in
Sect. 3.2. Section 3.4 summaries the design procedure for local mode dependent
controllers.
3.1 Design Methodology

As the global operation mode process (t) is Markovian, we can enlarge the mode
state space of the subsystems in the large-scale system (1), and form a new class of
uncertain large-scale systems with parameter constraints. The new class of uncertain
large-scale systems is described by
174
J. Xiong et al.
x (t) = A i ( (t))xi (t) + B i ( (t))ui (t) + E i ( (t))i (t) + L i ( (t))ri (t),
i
S i : i (t) = H i ( (t))xi (t) + G i ( (t))ui (t),
y (t) = C ( (t))x (t) + D ( (t)) (t),

i
i
i
i
i
(7)
where A i ( ) = Ai (i ), B i ( ) = Bi (i ), E i ( ) = Ei (i ), L i ( ) = Li (i ), H i ( ) =
Hi (i ), G i ( ) = Gi (i ), Ci ( ) = Ci (i ), D i ( ) = Di (i ), M, i = i1 ( ) and
i1 () is the function i1 : M Mi introduced in Sect. 2. The uncertainty inputs
i (t) and ri (t) are described, respectively, by the same functions as i (t) and ri (t) in
(1). So i (t) and ri (t) r . The initial condition is given by xi0 = xi0 , i N,
and 0 . It can be seen that the system (7) and the system (1) are in fact the same.
However, the operation mode of each subsystem in this new class of systems is now
Markovian and hence the corresponding control problem is easier to deal with.
Associated with the uncertain system (7) is the following quadratic cost functional of the form

N

T
J E
(8)
xi (t)Q i ( (t))xi (t) + uTi (t)R i ( (t))ui (t) dt | x0 , 0 ,
0
i=1
where Q i ( ) = Qi (i ), R i ( ) = Ri (i ), M, i = i1 ( ) Mi for i N, and

x0 = x0 .
We now consider the problem of guaranteed cost control of the system (7) by
means of an uncertain decentralized global mode dependent output feedback controller of the form
)
xK,i (t) = A K,i ( (t))xK,i (t) + B K,i ( (t))yi (t) + 1i (t) + 2i(t),
Ki :
(9)
ui (t) = CK,i ( (t))xK,i (t) + D K,i ( (t))yi (t) + 3i (t) + 4i(t).
The initial state of the controller is zero, and the initial operation mode is the same
as that of the system (7). The controller dynamics are assumed to be subject to
controller uncertainties of the form
1i (t) = 1i (t, xK,i (t), (t)),

3i (t) = 3i (t, xK,i (t), (t)),
2i (t) = 2i (t, yi (t), (t)),

4i (t) = 4i (t, yi (t), (t)),
which satisfy the following IQCs.

Definition 7.4. Given 1i ( ), 2i ( ), 3i ( ), 4i ( ) R+ , M, i N. A collection of controller uncertainties 1i (t), 2i (t), 3i (t), 4i (t), i N, is an admissible
uncertainty input for the dynamic controller in (9) if there exists a sequence {tl }
l=1
such that tl , tl 0 and
t
!

2
l

2
2
E
(10)
1i (t) 1i ( (t)) xK,i (t) dt | x0 , 0 0,
0

2

2
2i (t) 2i2 ( (t)) yi (t) dt | x0 , 0 0,
0
!

2
tl

2
2
E
3i (t) 3i ( (t)) xK,i (t) dt | x0 , 0 0,
0
t
!

2
l

2
2
E
4i (t) 4i ( (t)) yi (t) dt | x0 , 0 0,
tl
E

175
(11)
(12)
(13)
for all l and for all i N. The set of the admissible controller uncertainties is denoted
by K .
Again, one can assume that the same sequence {tl }
l=1 is selected in Definitions
7.1, 7.2 and 7.4.
Remark 7.3. In the uncertain controller (9), we use four (instead of two) controller
uncertainty inputs. Also, we use unity matrices as the coefficients of the controller
uncertainty inputs. One reason for this choice is to facilitate the derivation of the
result in the following theorem. Another reason is that the design of controller (9) is
not the main objective of the work and only serves as an intermediate step to obtain
a local mode dependent controller of the form (4).
The following theorem demonstrates the auxiliary role of the global mode dependent controller (9) in the design of the local mode dependent controller (4). It establishes a sufficient condition for the controller in (4) to stabilize the uncertain system
(1) provided the controller (9) stabilizes the uncertain system (7). The condition is
expressed in terms of the differences between the controller parameter matrices.
Theorem 7.1. Suppose an uncertain controller (9) stochastically stabilizes the uncertain large-scale system (7) subject to constraints (2), (3), (10)(13) and leads
to the cost bound sup , r , K J < c for some c R+ . If the controller parameter
matrices in (4) are chosen so that

A K,i ( ) AK,i (i ) 1i ( ),
(14)

B K,i ( ) BK,i (i ) 2i ( ),
(15)

CK,i ( ) CK,i (i ) 3i ( ),
(16)

D K,i ( ) DK,i (i ) 4i ( ),
(17)
for all M, i N, where i = i1 ( ) Mi , then the controller in (4) stochastically stabilizes the uncertain large-scale system in (1) subject to constraints (2), (3)
and also leads to the cost bound sup , r J < c.
Proof. Define the matrices
1i ( ) A K,i ( ) AK,i (i ),
3i ( ) CK,i ( ) CK,i (i ),
2i ( ) B K,i ( ) BK,i (i ),
4i ( ) D K,i ( ) DK,i (i ),
176
J. Xiong et al.
for M, where i = i1 ( ) Mi . Then the inequalities in (14)(17) imply that

1i ( ) 1i ( ),
3i ( ) 3i ( ),
2i ( ) 2i ( ),
4i ( ) 4i ( ).
Now consider a set of particular controller uncertainties of the form
1i (t) = 1i ( (t))xK,i (t),

3i (t) = 3i ( (t))xK,i (t),
We have

2

2
1i (t) 1i2 ( (t)) xK,i (t) ,
2

2
3i (t) 3i2 ( (t)) xK,i (t) ,
2i (t) = 2i ( (t))yi (t),

4i (t) = 4i ( (t))yi (t).
(18)
(19)

2

2
2i (t) 2i2 ( (t)) yi (t) ,

2

2
4i (t) 4i2 ( (t)) yi (t) .
Hence the IQCs in (10)(13) are satisfied. Thus the uncertainty inputs 1i (t), 2i (t),
3i (t) and 4i (t) defined in (18), (19), are an admissible uncertainty for controller
(9). Also we have
A K,i ( (t))xK,i (t) + 1i(t) = AK,i (i (t))xK,i (t),
B K,i ( (t))yi (t) + 2i(t) = BK,i (i (t))yi (t),
CK,i ( (t))xK,i (t) + 3i(t) = CK,i (i (t))xK,i (t),
D K,i ( (t))yi (t) + 4i(t) = DK,i (i (t))yi (t).
This means that the controller in (4) is a special case of the controller in (9) corresponding to the particular uncertainties given in (18), (19). Also note that the uncertain system (7) is in fact the same as the system in (1). It follows that the stability of
the system (7) implies that of system (1).
Finally, the upper bound of the cost functional follows from the following observation:
sup J =
, r
sup
, r ,1i (t),2i (t),
3i (t),4i (t),given in (18) (19)
sup
, r , K
J < c.
Here the inequality holds because the uncertainty inputs 1i (t), 2i (t), 3i (t)
and 4i (t) defined in (18), (19), belong to K .

3.2 Design of Global Mode Dependent Controllers

In this section, a sufficient condition is established for the design of the uncertain guaranteed cost controllers of the form (9). This condition, together with
177
Theorem 7.1, provides a basis for the design of local mode dependent guaranteed
cost controllers of the form (4). The following notation is defined in advance to
facilitate the presentation.
For every M and i N, let 1i ( ), 2i ( ), 3i ( ), 4i ( ) be positive constants from Definition 7.4. Also, let us be given symmetric matrices Pi ( ) S+ ,
Xi ( ) S+ and a collection of positive scalars i , i , 1i , 2i , 3i , 4i , i , 1i , 2i , 3i , 4i .
Using these matrices and constants define the following matrices:
!
!
A i ( ) 0
Ini ni
T
,
Ai ( ) =
Ni =
= Ni Ai ( )Ni ,
0ni ni
0 0
!
E ( ) L i ( ) 0 0 B i ( ) B i ( )
,
B i ( ) = i
0
0 I I 0
0
1
Q i2 ( )
0
0
0
H i ( )
0
Ci ( ) =
,
0
1i ( )I
2i ( )Ci ( )
3i ( )I
0
4i ( )Ci ( )
0
0
000 0
0
1
1
0
0 0 0 R i2 ( ) R i2 ( )
0
0 0 0 G i ( ) G i ( )
,
D i ( ) =
0
000 0
0
2i ( )D i ( ) 0 0 0 0
0
0
000 0
0
4i ( )D i ( ) 0 0 0 0
0
!
!
!
0 B i ( )
0 I
0 00000
Bi ( ) =
,
Di ( ) =
,
,
Ci ( ) =
Ci ( ) 0
Di ( ) 0 0 0 0 0
I 0
V1i ( ) 0 0 V2i ( ) 0 0 0 0 0
0 0
0 I 0 0 0 0 0 0 0
12
0 Ri ( )
0 0 I 0 0 0 0 0 0
0 Gi ( )
0 0 0 0 I 0 0 0 0
E i ( ) = 0 0 ,
2i ( ) =
,
0 0 0 0 0 I 0 0 0
0 0
0 0 0 0 0 0 I 0 0
0 0
0 0 0 0 0 0 0 I 0
0 0
0 00 0 0000 I
B ( )
!
K,i ( ) B K,i ( )

1i
A
,
(20)
Ki ( ) =
V1i ( ) V2i ( ) = R i2 ( ) ,
CK,i ( ) D K,i ( )
Gi ( )
178
J. Xiong et al.
T
T
A i ( )Pi ( ) + Pi( )A i ( ) + M
=1 q Pi ( ) Pi ( )Bi ( ) Ci ( )
i ( ) =
Q2i
D Ti ( ) ,
B Ti ( )Pi ( )
Ci ( )
Di ( )
Q3i
(21)
Pi ( )Bi ( )

,
(22)
li ( ) =
ri ( ) = Ci ( ) Di ( ) 0 ,
0
E i ( )
and
Q1i ( ) = NiT Xi ( )Ni A Ti ( ) + A i ( )NiT Xi ( )Ni + q NiT Xi ( )Ni ,
Q2i = diag(i I, i I, 1i I, 2i I, 3i I, 4i I),
Q3i = diag(I, I, i I, 1i I, 2i I, 3i I, 4i I),
3
1i ( ) =
q ,1 NiT Xi ( ) q , 1 NiT Xi ( ) q , +1 NiT Xi ( )
4
q ,M NiT Xi ( ) ,
Q4i ( ) = diag(Xi (1), . . . , Xi ( 1), Xi ( + 1), . . ., Xi (M)),

Q5i ( ) = A Ti ( )NiT Pi ( )Ni + NiT Pi ( )Ni A i ( ) +
q NiT Pi ( )Ni .
=1
Now, we are in a position to state the main result of this section, which provides
a sufficient condition for the design of an output feedback controller of the form
(9) such that the resulting closed-loop system is robustly stochastically stable and
the worst case value of the cost functional in (8) is bounded. The result involves the
following coupled linear matrix inequalities with respect to the variables Pi ( ) S+ ,
Xi ( ) S+ , and i , i , 1i , 2i , 3i , 4i , i , 1i , 2i , 3i , 4i R+ , M, i N:
Q1i ( )
NiT B i ( ) NiT Xi ( )CiT ( ) 1i ( )
B T ( )Ni
Q2i
0
D Ti ( )
i
T ( ) < 0, (23)
2i ( )
Ci ( )Xi ( )Ni D i ( )
Q3i
0 2i
1iT ( )
0
0
Q4i ( )
!
Q5i ( )
NiT Pi ( )B i ( ) NiT CiT ( )
C T ( )
iT
0 T
D Ti ( )
Q2i
D i ( )
Bi ( )Pi ( )Ni
Q3i
D i ( )
Ci ( )Ni
0
I
!
CiT ( )
D Ti ( )
0
T
0
< 0,
I
(24)
with rank constraints

!

Pi ( ) I
2n,
rank
I Xi ( )
!
!

i + i 1
1i 1
1, rank
1,
rank
1 i
1 1i
!
!
!

2i 1
3i 1
4i 1
1, rank
1, rank
1,
rank
1 2i
1 3i
1 4i
179
(25)
(26)
(27)
where i Nj=1, j=i j .
Theorem 7.2. Given 1i ( ), 2i ( ), 3i ( ), 4i ( ) R+ . Suppose there exist matrices Pi ( ) S+ , Xi ( ) S+ and scalars i , i , 1i , 2i , 3i , 4i , i , 1i , 2i , 3i ,
4i R+ , M, i N, such that the coupled linear matrix inequalities (23)(27)
hold for all M, i N. Consider an uncertain decentralized global mode dependent output feedback controller of the form (9) whose matrix parameters K i ( )
defined in (20) are obtained by solving the coupled linear matrix inequalities
i ( ) + li ( )Ki ( )ri ( ) + riT ( )K iT ( )liT ( ) < 0,
(28)
where the matrices i ( ), li ( ) and ri ( ) are defined in (21), (22) using the
solutions to (23)(27). Then the closed-loop large-scale system, consisting of the
uncertain subsystems (7) and this controller and subject to (2), (3) and (10)(13), is
robustly stochastically stable. Also, the worst-case cost functional (8) computed on
the trajectories of this closed-loop system is bounded as follows:
sup
, r , K
N

J < xTi0 NiT Pi (i0 )Ni + i Si + i Si xi0 .
(29)
i=1
Proof. The proof is given in Appendix 1.

Remark 7.4. The design condition given in Theorem 7.2 takes into account the IQC
structures of the uncertainties and the interconnections in the large-scale system. In
fact, the IQC structure information is embedded into the design condition though the
positive scalars i , i , 1i , 2i , 3i and 4i . Also, the following optimization problem
N
xTi0
Pi ( ),Xi ( ),i ,i ,1i ,2i ,
inf
3i ,4i ,i ,1i ,2i ,3i ,4i ,

subject to (23)(27)
T

Ni Pi (i0 )Ni + i Si + i Si xi0
(30)
i=1
can be used to minimize the upper bound of cost functional (8). Solving this optimization problem would lead to a suboptimal guaranteed cost controller for the
auxiliary control problem.
Remark 7.5. Due to the rank constraints (25)(27), the solution set to (23)(27)
is not convex. Although in general it is difficult to solve such problems, several
180
J. Xiong et al.
numerical algorithms have been proposed for this purpose, such as the cone complementarity
linearization algorithm in [4], the tangent and lift method in [12]. In this work, we
used the algorithm from [11, 12] with good results.
In the next section, some additional conditions will be imposed on the matrices
Ki ( ) from Theorem 7.2 to obtain a local mode dependent guaranteed cost controller from K i ( ); see Theorem 7.3.
3.3 The Main Result: Design of Local Mode Dependent

Controllers
In this section, we show that the local mode dependent controller (4) can be designed from a global mode dependent controller of the form (9) using the stochastic
projection method proposed in [20]. Specifically, we show that a robust stabilizing
dynamic local mode dependent controller (4) can obtained by taking the expectation of a controller (9) conditioned on the subsystem operation modes as time
approaches infinity. In combination with the results of Theorems 7.1 and 7.2, the
result of this section yields a computational method for the design of a local mode
dependent guaranteed cost controller (4). The method is presented in Theorem 7.3
given below.
Proposition 7.1 ([20]). Given the matrices K i ( ), M, i N, defined in Theorem
7.2, let
2
1
!
M
AK,i (i ) BK,i (i )
=1 Ki ( ) Ii ( , i )
2
1
(31)

Ki (i ) =
CK,i (i ) DK,i (i )
M
=1 Ii ( , i )
for all i Mi , i N, where
)
Ii ( , i ) =
if i = i1 ( ),
otherwise,

is the -th component of the vector = e(Q + E)1 with e = 1 1 1
T

R1M and E = eT eT eT RMM . Then

Ki (i ) = lim E K i ( (t)) | i (t) = i .
t
Proof. First we observe that the vector is the steady state distribution of the
global mode process (t). The ergodic property of the Markov process (t) implies that limt eQt = [ T T ]T . Then the probability distribution of the matrixvalued random process K i ( (t)) has a limit as t , which is limt Pr(K i ( (t)) =
K i ( )) = for all M. So the expected value of K i ( (t)) conditioned on the
subsystem operation mode, as t , is given by
lim E K i ( (t)) | i (t) = i
181
=1
8
7
K i ( ) lim Pr( (t) = | i (t) = i )
t

Pr( (t) = , i (t) = i )
= Ki ( ) lim
t
Pr(i (t) = i )
=1
)
5
M
Ii ( , i )
= K i ( ) M
=1 { Ii ( , i )}
=1
M
= Ki (i ).
This completes the proof.

Note that by choosing the matrix Ki (i ) to be the limit of the conditional expectation of K i ( (t)) given the subsystem operation mode, we introduce the asymptotically most accurate local mode dependent approximation of K i ( (t)). Indeed, the
parameters of the local mode dependent controller (4) given in (31) are the minimum variance approximation of the corresponding controller parameters in (9) in

2
the sense that limt E(K i ( (t)) Ki (i (t))F | i (t)) is minimal [1, Theorem
3.1]; here F denotes the Frobenius norm.
Also, we have
2
1

M
=1, = Ii ( , i ) Ki ( ) Ki ( )
i ( ) Ki ( ) Ki (i ) =
,
(32)
M
=1 {Ii ( , i ) }
where i = i1 ( ) Mi , M. Note that i ( ) is a linear matrix function of
K i ( ), M.
A computational method for the design of guaranteed cost controller (4) is presented in the following result, which is based upon Theorem 7.1 and Theorem 7.2.
The idea behind this result is to construct a global mode dependent controller of
the form (9), using Theorem 7.2, such that the corresponding s given in (32) also
satisfy the conditions of Theorem 7.1. Then, one will be able to construct the corresponding local mode dependent controller (4) with parameters given in (31), and
concludes that such a controller stabilizes the system (1) and ensures the worst case
quadratic performance cost which does not exceed that of the underlying auxiliary
global mode dependent controller (9). Let
!
!
!
!
I
I
0
0
N1i = ni ni , N2i = ni ni , N3i = ni ti , N4i = ni mi .
0mi ni
0ti ni
Iti ti
Imi mi
Theorem 7.3. Given a set of 1i ( ), 2i ( ), 3i ( ), 4i ( ) R+ . Suppose a set of
solutions Pi ( ) S+ , Xi ( ) S+ , i , i , 1i , 2i , 3i , 4i , i , 1i , 2i , 3i , 4i R+ ,
M, i N is found for (23)(27).
182
J. Xiong et al.
) such that the following LMIs

If there exist matrices K(
!
1i ( )I N2iT iT ( )N1i
0,
N1iT i ( )N2i 1i ( )I
!
0,
N1iT i ( )N3i 2i ( )I
!
0,
N4iT i ( )N2i 3i ( )I
!
0,
N4iT i ( )N3i 4i ( )I
(33)
(34)
(35)
(36)
and the linear matrix inequalities in (28) hold for all M, i N, where i ( ) is
the linear function of K i ( ) defined in (32), then the local mode dependent controller
of the form (4) with parameters defined in (31) robustly stabilizes the uncertain
system (1) subject to the constraints (2), (3) and leads to the cost bound
N

sup J < xTi0 NiT Pi (i0 )Ni + i Si + i Si xi0 .
, r
(37)
i=1
Proof. According to Theorem 7.2, the controller of the form (9) given by the solution of (28), (33)(36), robustly stochastically stabilizes the uncertain system (7),
and leads to the cost bound (29).
In view of (32), we have that
N1iT i ( )N2i = A K,i ( ) AK,i (i ),
N1iT i ( )N3i = B K,i ( ) BK,i (i ),
N4iT i ( )N2i = CK,i ( ) CK,i (i ),
N4iT i ( )N3i = D K,i ( ) DK,i (i ).
Therefore, the LMIs in (33)(36) are equivalent to the inequalities in (14)(17),
respectively. In view of Theorem 7.1, the local mode dependent controller given
in (31) robustly stabilizes the uncertain system (1) and leads to the cost bound in
(37).

3.4 Design Procedure

The proposed controller design procedure is summarized as follows.
1. Given 1i ( ), 2i ( ), 3i ( ), 4i ( ) R+ , M, i N, find a feasible solution to the rank constrained LMI problem in (23)(27). As noted above, although there does not exist an algorithm that is guaranteed to solve this problem,
the algorithm in [12] can be applied with good results;
183
2. Using this feasible solution, find a global mode dependent controller (9) via
solving the LMIs in (28), (33)(36);
3. Construct the local mode dependent controller (4) using equation (31).
4 An Illustrative Example
In this section, we present a numerical example to illustrate the theory which has
been developed. The uncertain large-scale system in the example has 3 subsystems
and each subsystem can operate in 2 operation modes. The system data for the system (1) are as follows.
!
!

1
0
1
A1 (1) =
, B1 (1) =
,
C1 (1) = 0.6 0 ,
0.5 0.5
0
!
!

1 0
1
A1 (2) =
,
B1 (2) =
, C1 (2) = 1 0 ,
0.1 0.5
0.1
!
!

0.6 0.5
0
,
B2 (1) =
,
C2 (1) = 0.1 1 ,
A2 (1) =
0 0.5
1
!
!

1 0
0.1
A2 (2) =
,
B2 (2) =
, C2 (2) = 0 1 ,
0.5 0.5
1
!
!

1 0
0.1
A3 (1) =
,
B3 (1) =
, C3 (1) = 0 1 ,
0.1 0.1
1
!
!

0.2 0
1
A3 (2) =
, B3 (2) =
,
C3 (2) = 1 1 ,
0.1 0.2
0
!
!
!
0 0
0.01
0.1 0
Ei (1) =
,
Li (1) =
, Hi (1) =
,
0.01 0
0
0 0
!
!
!
0.01 0
0
0 0.1
Ei (2) =
,
Li (2) =
, Hi (2) =
,
0 0
0.01
0 0.1
!
!
0
0.1
,
Gi (2) =
,
Gi (1) =
0.1
0.1

Di (2) = 0 0.1 ,
i = 1, 2, 3.
Di (1) = 0 0.1 ,
Here the matrices Ei (i ), Hi (i ), Gi (i ) and Di (i ) are chosen to satisfy the assumptions in [7]. The weighting matrices in the cost functional are given by
!
0.001 0
, Ri (i ) = 0.01, i = 1, 2, i = 1, 2, 3.
Qi (i ) =
0 0.001
184
J. Xiong et al.
Table 1 Relationship between local operation modes and global operation mode
( 1 , 2 , 3 )
(1, 1, 1)
(1, 2, 2)
(2, 1, 2)
(2, 2, 1)
1
2
3
4
The initial condition of the system is assumed to be

!
!
!
5
3
1
, x20 =
, x30 =
,
x10 =
5
3
1
10 = 20 = 30 = 1.
In this example, we assume that there are constraints on the operation modes
of the subsystems: the operation mode of subsystem S1 is assumed to be dependent on the operation modes of subsystems S2 and S3 . Specifically, 1 (t) = 1 if
2 (t) = 3 (t), and 1 (t) = 2 otherwise. Therefore, the operation mode pattern set
Mp is given by {(1, 1, 1), (1, 2, 2), (2, 1, 2), (2, 2, 1)}. Hence, the large-scale system
has four global operation modes. A relationship between the local operation modes
and the global operation mode is described in Table 1. This table can also be used
to define the bijective function : Mp M, and the functions i1 : M Mi ,
i = 1, 2, 3. The mode transition rate matrix of the Markov process (t) is assumed
to be
2 0.5 0.1 1.4

0.2 0.5 0.1 0.2
Q=
0.4 0.8 1.3 0.1 .
0.1 0.3 0.2 0.6
It can be verified by direct calculations, for instance, that the local mode process
1 (t) is not Markovian and is dependent on 2 (t).
The software packages used to obtain a solution to the set of rank constrained
LMI problem (23)(27) are the Matlab LMIRank toolbox [11] with the YALMIP
interface [8] and the underlying SeDuMi solver [17]. In Step 1 of the design procedure outlined in Sect. 3.4, we let ji ( ) = 1 for i = 1, . . . , 3, j, = 1, . . . , 4, and
solve the rank constrained optimization problem
min
subject to
xTi0

NiT Pi (i0 )Ni + i Si + i Si xi0
i=1
and the rank constrained LMIs in (23)(27). A suboptimal worst-case controller of

the form (4) can be found for = 57. Surprisingly, the controller we obtained may
be approximated by a static output controller of the form
ui (t) = DK,i (i (t))yi (t)
185
since the values of BK,i (i ) and CK,i (i ) are small. Hence we only provide the values
of DK,i (i ):
DK,1 (1) = 6.9269,
DK,2 (1) = 6.1663,
DK,1 (2) = 3.7007,

DK,2 (2) = 3.8897,
DK,3 (1) = 2.9246,
DK,3 (2) = 1.7758.
Using the resulting static output feedback controller, an upper bound of the
cost functional in (6) was evaluated by solving the worst-case performance analysis problem for the corresponding closed-loop system; the bound was found to be
sup , r J 1.1594. For comparison, the algorithm given in [7] was used to design
a global mode dependent output feedback controller. Firstly, the system (1) with the
cost functional (6) was regarded as a special class of the systems (7) with the cost
functional (8), which were studied in [7].
Then, following the design procedure proposed in [7], it was found that the best
global mode dependent controller that could be constructed yields the upper bound
of the cost functional sup , r J 1.0526. It can be seen that the local mode dependent controller achieved almost the same level of performance as the global mode
dependent controller (the difference is less than 10%). However, such a performance
was achieved by a local mode dependent schedule between two operation modes
for each subsystem. In contrast, the global mode dependent controller in [7] leads
to a global mode dependent schedule between four operation modes for each subsystem. Moreover, using the local mode dependent controller eliminates collecting
and broadcasting of the subsystem mode information throughout the entire largescale system.
We now present some simulations to illustrate the properties of the static output
feedback controller which was designed using our approach. The admissible uncertainties were chosen to be of the following form
xr,i (t) = Ar,i xr,i (t) + Br,i j j (t),

i (t) = Fi i (t), and
j=1, j=i
ri (t) = Cr,i xr,i (t),

for i = 1, 2, 3, where [1, 1] and > 0 are unknown parameters. The parameters
in the uncertainties are chosen to be
!
!
!

0.1 0.9
1 1
11
Fi =
, Ar,i =
, Br,i j =
, Cr,i = 0.3 0.8 ,
0.9 0.1
1 1
11
i, j = 1, 2, 3.
The initial condition of the dynamic interconnections is set to zero. The reason for
this particular linear uncertainty choice is as follows. With these linear uncertainties
and the controller designed using our approach, the open-loop and the closed-loop
large-scale systems are linear. As a result, we were able to use the existing results
on stability of Markovian jump linear systems to check stability of the open-loop
and closed-loop systems. It turned out that the open-loop system with ui (t) 0
186
J. Xiong et al.
7
2x 10
=1
=1
1.5
system state
1
0.5
0
0.5
1
0
10
15
time (sec)
Fig. 1 Initial condition response of the open-loop system with uncertainty
6
=1
=1
system state
6
0
time (sec)
10
15
Fig. 2 Initial condition response of the closed-loop system with uncertainty
was not stochastically stable while the closed-loop system was found to be robustly
stochastically stable. This confirms the result of Theorem 7.3. To further illustrate
these findings, a random sample of (t) was generated and used in the simulations,
and the sample state trajectories of the open-loop and the closed-loop large-scale
systems are shown in Figs. 1 and 2, respectively.
5 Conclusions
This chapter has studied the decentralized output feedback guaranteed cost control
problem for a class of uncertain Markovian jump large-scale systems via a local
mode dependent approach. The controllers are entirely decentralized with respect to
187
the subsystems. They use the subsystem states and the subsystem operation modes
to produce the subsystem control inputs. A sufficient condition in terms of rank constrained linear matrix inequalities has been developed to construct the controllers.
Also, the theory has been illustrated by a numerical example and simulations.
Appendix 1: Proof of Theorem 7.2

The proof is divided into three steps. Firstly, we show that the feasibility of the
rank constrained LMIs in (23)(27), implies the solvability of the LMIs in (28).
Secondly, we show that the solvability of (28) implies that a certain augmented system is stable and has certain H performance. Thirdly, we prove that the stability of
the augmented system implies the robust stochastic stability of the closed-loop system composed of the uncertain system (7) and the controller (9) designed using the
solution of (28). Also, we show that the H performance of the augmented system
leads to the bound (29) on the cost functional (8) of the auxiliary system (7).
Step 1
Firstly, the rank constraint in (25), together with the positive definiteness of Pi ( )
and Xi ( ), implies that Xi ( ) = Pi1 ( ).
In view of the projection lemma (see, e.g. [5, Theorem 1]), the inequality in (28)
has feasible solution K i ( ) if and only if the following two inequalities hold:
T

(li ( )) i ( ) (li ( )) < 0;

T
T
ri ( ) i ( ) riT ( )
< 0.
(38)
(39)
Consider (38). We have
Pi ( ) 0 0 Bi ( )
li ( ) = 0 I 0 0 ,
0 0I
E i ( )
and
0
I
Bi ( )
0 = 0
0
E i ( )
0
0
B i ( )
0
1
R i2 ( )
.
G i ( )
0
0
188
J. Xiong et al.
An orthogonal complement matrix of li ( ) is found to be
Bi ( )
Pi ( ) 0 0
(li ( )) = 0 0
I 0
E i ( )
0
0I
V1i ( ) 0 0 0 V2i ( ) 0 0 0 0
0 0 I 0 0 0 0 0 0
0 0 0 I 0 0 0 0 0 Pi ( ) 0 0
= 0 0 0 0 0 I 0 0 0 0
I 0
0 0 0 0 0 0 I 0 0
0
0I
0 0 0 0 0 0 0 I 0
0 000 0 000I
T
Ni 0 0 0 0 0 0 0
0 I 0 0 0 0 0 0
0 0 I 0 0 0 0 0
P1 ( ) 0 0
0 0 0 I 0 0 0 0 i
= 2i ( )
0
I 0
0 0 0 0 I 0 0 0
0
0I
0 0 0 0 0 I 0 0
0 0 0 0 0 0 I 0
0 000000I
T 1
Ni Pi ( ) 0 0 0 0 0 0 0
0
I 0 0 0 0 0 0
0
0 I 0 0 0 0 0
0 0 I 0 0 0 0
= 2i ( )
,
0 0 0 I 0 0 0
0
0
0 0 0 0 I 0 0
0 0 0 0 0 I 0
0
0
000000I
where 2i ( ) is defined in (20). So, the inequality in (38) becomes
NiT B i ( ) NiT Xi ( )CiT ( )

Q 1i ( )
2iT ( ) < 0
2i ( ) B Ti ( )Ni
Q2i
D Ti
Ci ( )Xi ( )Ni
Q3i
D i
where
Q 1i ( ) = NiT Xi ( )Ni A Ti ( ) + A i ( )NiT Xi ( )Ni +
q NiT Xi( )Pi ( )Xi ( )Ni .
=1
The above inequality is further equivalent to (23) in view of Schur complement

equivalence.
189
Consider (39). We have
0 CiT ( )
I
0
0 D T ( )
i
T

Ci ( )
0 0
T
T
ri ( ) = Di ( ) = 0 0 ,
0 0
0
0 0
0 0
0 0
so an orthogonal complement matrix of riT ( ) is given by
0
0
0
I
NT
V3i ( ) V4i ( ) 0 0 0 0 0 0 i
0
0
0 I 0 0 0 0 0
0
0
0 0 I 0 0 0 0
= 0
0 0 0 I 0 0 0
0
0
0 0 0 0 I 0 0
0
0
0 0 0 0 0 I 0
0
0
0 00000 I
0
! T
T
N 00
C ( )
i
0 i
= D Ti ( )
0 I 0 ,
0 0I
0
I
V3i ( ) 0 V4i ( ) 0 0
0 0 0 I 0
0 0 0 0I
T

ri ( ) = 0 0 0 0 0
0 0 0 00
0 0 0 00
0 0 0 00
where
000
000
000
I 00
0I 0
00I
000
0
I
0
0
0
0
0
0
00
00
I 0
0I
00
00
00
00
000
000
000
000
I 00
0I 0
00I
000
0
0
0
I
!

C T ( )
.
V3i ( ) V4i ( ) = iT
Di ( )
So the inequality in (39) becomes the inequality (24).

Therefore, if the linear matrix inequalities (23), (24) with rank constraints (25)
(27) have solutions, then the matrix inequalities (38), (39) hold. According to the
projection lemma, the LMI in (28) is feasible after the solution of (23)(27) is substituted into (28). As a result, the parameter matrices of the controller (9) can be
obtained by solving (28).
190
J. Xiong et al.
Step 2
In this step, we will consider an augmented system and show that the solvability of
(28) implies that the augmented system is stable and has a certain H performance.
Firstly, the rank constraints in (26), (27) imply that i = (i + i )1 , 1i = 1i1 ,
2i = 2i1 , 3i = 3i1 and 4i = 4i1 .
Now, we consider an augmented system. Let
' 1
'
(
(
0
2 ( (t))
Q
i
zi (t) =
(40)
xi (t) + 1
ui (t),
R i2 ( (t))
0
1i (t) = 1i ( (t))xK,i (t),
(41)
2i (t) = 2i ( (t))yi (t),

3i (t) = 3i ( (t))xK,i (t),
(42)
4i (t) = 4i ( (t))yi (t).
(44)
(43)
When a controller of form (9) is applied to the uncertain system (7), augmented with
the additional outputs defined in (40)(44), and subject to the IQCs in (2), (3), (10),
(11), (12), (13), the resulting closed-loop uncertain system can be described by the
state equations:
)
xi (t) = A i ( (t))xi (t) + B i( (t))Ui w i (t),
(45)
S i :
zi (t) = ViCi ( (t))xi (t) + ViD i ( (t))Ui w i (t),
where
i (t)
i i (t)
r (t) r (t)
i i i
!

xi (t)
1i 1i (t)
1 1i (t)
, w i (t) = Ui
xi (t) =
=
,
2i (t) 2i 2i (t)
xK,i (t)

3i (t) 3i 3i (t)

4i (t)
4i 4i (t)
zi (t)
zi (t)
(t) . + (t)
i i
i i
1i (t) 1i 1i (t)
zi (t) = Vi
,
=
2i (t) 2i 2i (t)
3i (t) 3i 3i (t)
4i (t)
4i 4i (t)
!
A i ( ) + B i ( )D K,i ( )Ci ( ) B i ( )CK,i ( )
Ai ( ) =
,
BK,i ( )Ci ( )
A K,i ( )
!
E ( ) + B i ( )D K,i ( )D i ( ) L i ( ) 0 0 B i ( ) B i ( )
B i ( ) = i
,
B K,i ( )D i ( )
0 I I 0
0
191
Q i2 ( )
0
1
1
2
2
R i ( )CK,i ( )
R i ( )D K,i ( )Ci ( )
Hi ( ) + G i ( )D K,i ( )Ci ( ) G i ( )CK,i ( )
Ci ( ) =
1i ( )I ,
0
2i ( )Ci ( )
0
3i ( )I
0
4i ( )Ci ( )
0
0
000 0
0
1
1
21
Ri ( )D K,i ( )D i ( ) 0 0 0 R i2 ( ) R i2 ( )

Gi ( )D K,i ( )D i ( ) 0 0 0 G i ( ) G i ( )
D i ( ) =
,
0
000 0
0
2i ( )Di ( )
000 0
0
0
000 0
0
4i ( )D i ( )
000 0
0
1
1
1
1
1
1
Ui = diag( I, I, I, I, I, I),
i
1i
2i
3i
4i
i
1
1
1
1
1
Vi = diag(I, I, I, I, I, I).
i
1i
2i
3i
4i
The introduction of these positive scalars i , i , 1i , 2i , 3i and 4i allows us to use
the IQC uncertainty structure information in the controller design; see Step 3 of the
proof and Remark 7.4 for more explanations about these parameters.
Since the global mode process (t) is a continuous-time Markov process, the
closed-loop system in (45) is in fact a special class of Markovian jump linear systems [2, 10]. In order to use the theory which has been developed for Markovian
jump linear systems, we re-write the system (45) as
)
= A(
(t))x(t)
(t))w(t),
+ B(
x(t)
(46)
(t))x(t)
(t))w(t),
+ D(
z(t) = C(
where
x1 (t)
x(t)
= ... ,
w 1 (t)
w(t)
= ... ,
z1 (t)
z(t) = ... ,
xN (t)
w N (t)
) = diag(A 1 ( ), . . . , A N ( )),
A(
) = diag(B 1 ( )U1 , . . . , B N ( )UN ),
B(
) = diag(V1C1 ( ), . . . ,VNCN ( )),
C(
) = diag(V1 D 1 ( )U1 , . . . ,VN D N ( )UN ).
D(
zN (t)
192
J. Xiong et al.
According to the bounded real lemma (see, e.g. [21, Proposition 2]), the Markovian
jump linear system in (46) is stochastically stable and has H performance Tzw <
) = diag(P1 ( ), . . . , PN ( )), Pi ( ) S+ , M, i N,
1 if there exist matrices P(
such that the matrix inequality
T
) + M
T
) + P(
)A(
A ( )P(
=1 q P( ) P( )B( ) C ( )
)
I
D T ( ) < 0
B T ( )P(
)
)
C(
D(
I
(47)
holds for all M. Here Tzw denotes the operator from w to z defined by system (46). For similar results, we refer to [2, 3, 6, 13].
Because all the matrices in (47) are diagonal, using a congruence transformation
(see Appendix 2 for an illustrative example), the matrix inequality (47) is equivalent
to
T
A i ( )Pi ( ) + Pi ( )A i ( ) + M
=1 q Pi ( ) Pi ( )Bi ( )Ui Ci ( )Vi
Ui B Ti ( )Pi ( )
I
Ui D Ti ( )Vi < 0
ViCi ( )
Vi D i ( )Ui
I
(48)
for all M and i N. Pre- and post-multiplying both sides of the above inequality
by diag(I,Ui1 ,Vi1 ), we have the inequality
A Ti ( )Pi ( ) + Pi ( )A i ( ) + M
=1 q Pi ( ) Pi ( )Bi ( ) Ci ( )
B Ti ( )Pi ( )
Q1i
D Ti ( ) < 0.
D i ( )
Q2i
Ci ( )
(49)
After a direct computation, the above inequality is the same inequality as (28).
Therefore, if the inequality in (28) holds, then the system in (46) is stochastically
stable and has the H performance Tzw < 1.
Step 3
Because the system in (46) is stochastically stable, the closed-loop system composed of the uncertain system (7) and the controller (9) designed via solving (28) is
robustly stochastically stable as well. Next we prove that the corresponding closedloop value of the cost functional satisfies the performance given in (29).
Define the Lyapunov functional for the system in (46) as
N
V (t, x(t),
(t)) Vi (t, xi (t), (t))
i=1
where Vi (t, xi (t), (t)) = xTi (t)Pi ( (t))xi (t). Then, the stability of system (46) implies that
193
lim E (V (t, x(t),

(t)) | x0 , 0 ) = 0.
Here, the reason that we take the expectation conditioned on (x0 , 0 ) instead of
(x0 , 0 ) is that the controller state initial condition is fixed to zero.
On the other hand, it has been shown that the inequality in (28) is equivalent to
the inequality in (48). Also, the inequality in (48) is equivalent to
!
q Pi ( ) Pi ( )B i ( )Ui
A Ti ( )Pi ( ) + Pi( )A i ( ) + M
=1
Zi ( )
Ui B Ti ( )Pi ( )
I
!
T

Ci ( )Vi
+
ViCi ( ) Vi D i ( )Ui < 0.
Ui D Ti ( )Vi
Therefore, we have

LVi (t, xi (t), (t))
= xTi (t)Pi ( )xi (t) + xTi (t)Pi ( )xi (t) + xTi (t)
= zTi (t)zi (t) + w Ti (t)w i (t) +
q Pi( )
=1
xi (t)
!T
!
xi (t)
x (t)
Zi ( ) i
w i (t)
w i (t)
< zTi (t)zi (t) + w Ti (t)w i (t)

!
xi (t)
for any
= 0. So,
w i (t)
N

i=1
< E

zi (t)2 dt | x0 , 0

i=1
N
w i (t)2 dt | x0 , 0 E
i=1

LVi (t)dt | x0 , 0 .
In view of the definitions of zi (t) and w i (t) in (45), the above inequality becomes
N

0
i=1
< E

i=1

2
2 !

+ 3i 3i (t) + 4i 4i (t) dt | x0 , 0

2
2
2

zi (t)2 + i + i i (t) + 1i 1i (t) + 2i 2i (t)

2
2
2

i i (t) + i ri (t)2 + 1i 1i (t) + 2i 2i (t)

2
2 !

+ 3i 3i (t) + 4i 4i (t) dt | x0 , 0
+ xTi0 NiT Pi (i0 )Ni xi0 .

i=1
194
J. Xiong et al.
Because the uncertainties i (t), ri (t), 1i (t), 2i (t), 3i (t) and 4i (t) satisfy the IQCs
in (2), (3), (10)(13), respectively, we conclude that
sup
, r , K
J =
sup

, r , K i=1

zi (t) | x0 , 0
2
N

< xTi0 NiT Pi (i0 )Ni + i Si + i Si xi0 .
i=1
That is, the inequality in (29) holds. This completes the proof.

Appendix 2: Equivalence Between (47) and (48)

The example in this appendix illustrates that the matrix inequality in (47) is equivalent to the matrix inequality in (48) via a congruence transformation. An extension
to a general case is straightforward.
Example 7.1. Let the left hand side of (47) be
a1 0 d1 0
0 a 0 d
2
2
d1 0 b1 0
A=
0 d2 0 b 2
f 1 0 e1 0
0 f 2 0 e2
If an elementary matrix is chosen as
I 0
0 0
0 0
T =
0 I
0 0
00
then we have
a1
d
1
f
T
TAT = 1
0
0
0
00
I 0
00
00
0I
00
d1
b1
e1
0
0
0
f1
e1
c1
0
0
0
f1
0
e1
0
c1
0
0
f2
0
.
e2
0
c2
00
0 0
I 0
,
0 0
0 0
0I
0
0
0
a2
d2
f2
0
0
0
d2
b2
e2
0
0
0
.
f2
e2
c2
195
Acknowledgment This work was supported by the Australian Research Council.
References
1. Anderson, B.D.O., Moore, J.B.: Optimal Filtering. Prentice-Hall, Englewood Cliffs (1979)
2. de Farias, D.P., Geromel, J.C., do Val, J.B.R., Costa, O.L.V.: Output feedback control of
Markov jump linear systems in continuous-time. IEEE Transactions on Automatic Control
45(5), 944949 (2000)
3. de Souza, C.E., Fragoso, M.D.: H control for linear systems with Markovian jumping parameters. Control Theory and Advanced Technology 9(2), 457466 (1993)
4. El Ghaoui, L., Oustry, F., Rami, M.A.: A cone complementarity linearization algorithm for
static output-feedback and related problems. IEEE Transactions on Automatic Control 42(8),
11711176 (1997)
5. Iwasaki, T., Skelton, R.E.: All controllers for the general H control problem: LMI existence
conditions and state space formulas. Automatica 30(8), 13071317 (1994)
6. Li, L., Ugrinovskii, V.A.: On necessary and sufficient conditions for H output feedback control of Markov jump linear systems. IEEE Transactions on Automatic Control 52(7), 1287
1292 (2007)
7. Li, L., Ugrinovskii, V.A., Orsi, R.: Decentralized robust control of uncertain Markov jump
parameter systems via output feedback. Automatica 43(11), 19321944 (2007)
8. Lofberg, J.: YALMIP : A toolbox for modeling and optimization in MATLAB. In: Proceedings
of the CACSD Conference (2004). URL http://control.ee.ethz.ch/joloef/yalmip.php
9. Mahmoud, M.S., Shi, P., Shi, Y.: H and robust control of interconnected systems with Markovian jump parameters. Discrete and Continuous Dynamical Systems Series B 5(2), 365384
(2005)
10. Mariton, M.: Jump Linear Systems in Automatic Control. Marcel Dekker, New York (1990)
11. Orsi, R.: LMIRank: software for rank constrained LMI problems (2005). URL http://users.
rsise.anu.edu.au/robert/lmirank/
12. Orsi, R., Helmke, U., Moore, J.B.: A Newton-like method for solving rank constrained linear
matrix inequalities. Automatica 42(11), 18751882 (2006)
13. Pan, Z., Basar, T.: H -control of Markovian jump systems and solutions to associated
piecewise-deterministic differential games. In: G.J. Olsder (ed.) New Trends in Dynamic
Games and Applications, pp. 6194. Birkhauser, Basel (1995)
14. Petersen, I.R.: Robust H control of an uncertain system via a stable decentralized output
feedback controller. Kybernetika 45(1), 101120 (2009)
15. Petersen, I.R., Ugrinovskii, V.A., Savkin, A.V.: Robust Control Design Using H Methods.
Springer, Berlin (2000)
16. Sandell, N.R., Varaiya, P., Athans, M., Safonov, M.G.: Survey of decentralized control methods for large scale systems. IEEE Transactions on Automatic Control 23(2), 108128 (1978)
17. Sturm, J.F.: Using SeDuMi 1.02, a MATLAB toolbox for optimization over
symmetric cones. Optimization Methods and Software 1112, 625653 (1999).
URL http://sedumi.mcmaster.ca/
18. Ugrinovskii, V.A., Pota, H.R.: Decentralized control of power systems via robust control of
uncertain Markov jump parameter systems. International Journal of Control 78(9), 662677
(2005)
19. Wang, W.J., Chen, Y.H.: Decentralized robust control design with insufficient number of controllers. International Journal of Control 65(6), 10151030 (1996)
20. Xiong, J., Ugrinovskii, V.A., Petersen, I.R.: Local mode dependent decentralized control of
uncertain Markovian jump large-scale systems. In: Proceedings of the Conference on Decision
and Control, pp. 43454350 (2008) (also in IEEE-TAC 54(11))
196
J. Xiong et al.
21. Zhang, L., Huang, B., Lam, J.: H model reduction of Markovian jump linear systems. Systems and Control Letters 50(2), 103118 (2003)
22. Zhang, X., Zheng, Y., Lu, G.: An interconnected Markovian jump system and its decentralized
observer-based control with sector nonlinearities. In: International Conference on Control and
Automation, pp. 271276 (2005)
Consensus Based Multi-Agent Control

Algorithms
Milos S. Stankovic, Dusan M. Stipanovic, and Srdjan S. Stankovic
1 Introduction
Control of complex systems can be achieved via hierarchical multilayered agentbased structures benefiting from their inherent properties such as modularity, scalability, adaptability, flexibility and robustness. The agent-based structures consist of
a number of simpler subsystems (or agents), each of which addresses in a coordinated manner a specific sub-objective or sub-task so as to attain the overall design
objectives. The complexity of the behavior of such systems arises as a result of interactions between multiple agents and the environment in which they operate. More
specifically, multi-agent control systems are fundamental parts of a wide range of
safety-critical engineering systems, and are commonly found in aerospace, traffic
control, chemical processes, power generation and distribution, flexible manufacturing, robotic system design and self-assembly structures. A multi-agent system
can be considered as a loosely coupled network of problem-solver entities that work
together to find answers to problems that are beyond the individual capabilities or
knowledge of each entity, where local control law has to satisfy decentralized information structure constraints (see, e.g., [20]), and where no global system control (or supervision) is desired. Different aspects of multi-agent control systems are
M.S. Stankovic
ACCESS Linnaeus Center, School of Electrical Engineering, Royal Institute of Technology,
100-44 Stockholm, Sweden
e-mail: milsta@kth.se
D.M. Stipanovic
Department of Industrial and Enterprise Systems Engineering and the Coordinated Science
Laboratory, University of Illinois at Urbana-Champaign, Illinois, USA
e-mail: dusan@illinois.edu
S.S. Stankovic
School of Electrical Engineering, University of Belgrade, Serbia
e-mail: stankovic@etf.rs
197
198
M.S. Stankovic et al.
covered by a vast literature within the frameworks of computer science, artificial

intelligence, network and system theory; for some aspects of multi-agent control
systems and sensor networks see, e.g., [4, 7, 16, 31].
Considering methodologies for achieving agreement between the agents upon
some decisions, important results were obtained in relation with distributed iterations in parallel computation and distributed optimization as early as in the 1980s,
e.g., [13, 29, 30]. A very intensive research has been carried out recently in this
direction, including numerous applications (see, e.g., [4, 6, 7, 1214, 17, 18]). The
majority of the cited references share a common general methodology: they all use
some kind of dynamic consensus strategy.
In this chapter an attempt is made to approach the problem of overlapping decentralized control of complex systems by using a multi-agent strategy, where the
agents (subsystems) communicate in order to achieve direct or indirect agreement
upon a control action by using a dynamic consensus methodology. The aim of the
chapter is to propose several different novel control structures derived from:
(a) the choice of the variables upon which the agreement is made;
(b) basic local controller structures derived from the decentralized control laws
implemented by the agents.
The chapter is organized as follows. Section 2 deals with the problem definition,
including the subsystem models and the distribution of control tasks among the
agents. In Sect. 3 several new control structures are proposed based on the agreement between the agents upon the control variables. In the most general setting, it
is assumed that each agent is able to formulate its local feedback control law starting
from the local information structure constraints in the form of a general four-term
dynamic output controller. The subsystem inputs generated by the agents by means
of the local controllers enter the consensus process which generates the control signals to be applied to the system by some a priori specified agents. In the general
case, the consensus scheme, determining, in fact, the control law for the whole system, is constructed on the basis of an aggregation of the local dynamic controllers.
It is shown how the proposed scheme can be adapted to either static local output
feedback controllers, or static local state feedback controllers. In Sec. 4 an alternative to the approach presented in Sect. 3 is proposed, based on the introduction of a
dynamic consensus at the level of state estimation [24, 25]. Namely, it is assumed
that the agents are able to generate local estimates of parts of the overall state vector
using their own subsystem models. The dynamic consensus scheme is introduced to
provide each agent with a reliable estimate of the whole system state. The control
signal is obtained by applying the known global LQ optimal state feedback gain to
the locally available estimates. A number of selected examples illustrate the applicability of all the proposed consensus based control schemes. Section 5 is devoted to
the problem of decentralized overlapping control of a formation of unmanned aerial
vehicles (UAVs). Starting from a specific formation model, the global LQ optimal
state feedback is defined. Further, it is demonstrated that a decentralized consensus
based estimator can be formulated on the basis of extracted subsystems attached to
the vehicles. Efficiency of this approach is illustrated by a simulation example.
Consensus Based Multi-Agent Control Algorithms
199
2 Problem Formulation
Let a complex system be represented by a linear model
S:
x = Ax + Bu
y = Cx,
(1)
where x Rn , u Rm and y R are the state, input and output vectors, respectively,
while A, B and C are constant matrices of appropriate dimensions.
Assume that N agents have to control the system S according to their own resources. The agents have their local models of the parts of S
Si :
(i) = A(i) (i) + B(i) v(i)

y(i) = C(i) (i) ,
(2)
where (i) Rni , v(i) Rmi and y(i) Ri are the corresponding state, input and
output vectors, and A(i) , B(i) and C(i) constant matrices, i = 1, . . . , N. Components of
(i)
(i)
the input vectors v(i) = (v1 , . . . , vmi )T represent subsets of the global input vector
(i)
u of S, so that v j = u pi , j = 1, . . . , mi , and pij Vi , where Vi = {pi1 , . . . , pimi } is
j
(i)
the input index set defining v(i) . Similarly, for the outputs y(i) we have y j = yqi ,
j
j = 1, . . . , i , and qij Yi , where Yi = {qi1 , . . . , qii } is the output index set; according
to these sets, it is possible to find such constant i n matrices Ci that y(i) = Ci x,
i = 1, . . . , N. The state vectors (i) do not necessarily represent parts of the global
state vector x. They can be chosen, together with the matrices A(i) , B(i) and C(i) ,
according to the local criteria for modeling the inputoutput relations v(i) y(i) . In
(i)
the particular case when (i) = x(i) , x j = xri , j = 1, . . . , ni , ni n and rij Xi , where
j
Xi = {r1i , . . . , rni i } is the state index set defining x(i) . In the last case, models Si , in
general, represent overlapping subsystems of S in a more strict sense; matrices A(i) ,
B(i) and C(i) can represent in this case sub-matrices of A, B and C.
The task of the ith agent is to generate the control vector v(i) and to implement
(i)
the control action u(i) R i , satisfying u j = usi , j = 1, . . . , i , and sij Ui , where
j
Ui = {si1 , . . . , sii } is the control index set defining u(i) . It is assumed that Ui Vi
and Ui U j = 0,
/ so that Ni=1 i = m, that is, the control vector u(i) of the i-th agent
is a part of its input vector v(i) , while one and only one agent is responsible for
generation of each component of u within the considered control task. Consequently,
all agents include the entire vectors v(i) of Si in the control design considerations,
but they implement only those components of v(i) for which they are responsible.
In the case when the inputs v(i) do not overlap, the agents perform their tasks
autonomously, without interactions with each other; that is, we have the case of
decentralized control of S, when the control design is based entirely on the local
models Si . However, in the case when the model inputs v(i) overlap, more than one
model Si can be used for calculation of a particular component of the input vector u.
200
Obviously, it would be beneficial for the agent responsible for implementation of

that particular input component to use different suggestions about the control action
and to calculate the numerical values of the control signal to be implemented on
the basis of an agreement between the agents. The agents that do not implement
/ could, in this context, represent advisors to the agents
any control action (Ui = 0)
responsible for control implementation. Our aim is to propose several overlapping
decentralized feedback control structures for S based on a dynamic consensus between multiple agents.
3 Consensus at the Control Input Level

In this section algorithms based on consensus at the control input level are presented.
The first subsection deals with the algorithms derived from local dynamic feedback
controllers, while the second subsection is related to the important special case of
static local output feedback controllers.
3.1 Algorithms Derived from the Local Dynamic Output Feedback

Control Laws
We assume that each agent generates its input vector v(i) in Si using the local dynamic controller in the form
Ci :
w (i) = F (i) w(i) + G(i) y(i)

v(i) = K (i) w(i) + H (i)y(i)
(3)
where w(i) Ri represents the controller state, and matrices F (i) , G(i) , K (i) and
H (i) , i = 1, . . . , N, are constant, with appropriate dimensions. Local controllers are
designed according to the local models and local design criteria. Assuming that
the agents can communicate between each other, the goal is to generate the control
signal u for S based on a mutual agreement, starting from the inputs v(i) generated by
Ci . The idea about reaching an agreement upon the components of u stems from the
fact that the index sets V(i) are, in general, overlapping, so that the agents responsible
for control implementation according to the index sets U(i) can improve their local
control laws by getting suggestions from the other agents.
Algorithm 1. The second relation in (3) gives rise to v(i) = K (i) w (i) + H (i) y(i) ,
wherefrom we get
v(i) = K (i) [F (i) w(i) + G(i) y(i) ] + H (i)C(i) [A(i) (i) + B(i) v(i) ]
= K (i) F (i) w(i) + K (i) G(i) y(i) + H (i)C(i) A(i) (i) + H (i)C(i) B(i) v(i) .
(4)
201
Since y(i) are the available signals, and v(i) vectors to be locally generated for participation in the agreement process, we will use the following approximation
(i)
(i)
(i)
v(i) [F + H (i)C(i) B(i) ]v(i) + [K (i) G(i) + H (i) A F H (i) ]y(i) ,

(i)
(5)
(i)
where F = K (i) F (i) K (i)+ and A = C(i) A(i)C(i)+ are approximate solutions of the
(i)
(i)
aggregation relations K (i) F (i) = F K (i) and C(i) A(i) = A C(i) , respectively, where
+
A denotes the pseudo-inverse of a given matrix A [11, 20].
We will assume in the sequel, for the sake of presentation clarity, that all the
agents can have their suggestions for all the components of u; that is, we assume
that the vector v(i) = Ui Rm is a local version of u proposed by the ith agent to
the other agents. Furthermore, we introduce m i and m i constant matrices Ki
and Hi , obtained by taking the rows of K (i) and H (i) at the row indices defined by
the index set V(i) and leaving zeros elsewhere, and ni m matrix Bi obtained from
B(i) by taking its columns at the indices defined by Vi and leaving zeros elsewhere.
Let U = col{U1 , . . . ,UN }, Y = col{y(1) , . . . , y(N) }, K = diag{K1 , . . . , KN } and H =
diag{H1 , . . . , HN }. Similarly, let A = diag{A(1) , . . . , A(N) }, B = diag{B1 , . . . , BN },
C = diag{C(1) , . . . ,C(N) }, F = diag{F (1) , . . . , F (N) }, and G = diag{G(1) , . . . , G(N) }.
Assume that the agents communicate between each other in such a way that they
send current values of Ui to each other according to a communication strategy determined by the consensus matrix = [i j ], where i j , i, j = 1, . . . , N, i = j, are
m m diagonal matrices with positive entries and ii = Ni=1,i= j i j , i = 1, . . . , N.
Then, the algorithm for generating U, i.e., the vector containing all the agent input
vectors Ui , i = 1, . . . , N, representing the result of the overall consensus process, is
given by
U i =
j=1, j=i
i j (U j Ui ) + [Ki F (i) Ki+ + HiC(i) B(i) ]Ui
+ [Ki G(i) + HiC(i) A(i)C(i)+ Ki F (i) Ki+ Hi ]y(i) ,
(6)
i = 1, . . . , N, or, in a compact form,

+ [K G + H C AC + K F K + H]Y.
U = [ + K F K + + H C B]U
(7)
The vector U generated by (7) is used for control implementation in such a way that
the ith agent picks up the components of Ui selected by the index set U(i) and applies
them to the system S. If Q is an m mN matrix with zeros everywhere except one
place in each row, where it contains 1; for the jth row with j U(i) , 1 is placed at the
column index (i 1)m + j. Then, we have u = QU, and system (1) can be written as
x = Ax + BQU.
(8)
(i)
T
3Also, according
4 to the adopted notation, y = Ci x, so that Y = Cx, where C =
C1T CNT . Therefore, the whole closed-loop system is represented by
202
U
x
+ K F K + + H C B (K G + H C AC + 3 4
C
U
K F K + H)
x .
BQ
A
(9)
Obviously, the system is stabilized by the controller (7) if the state matrix in (9) is
asymptotically stable. In general, analysis of the stability of (9) is not an easy task.
Algorithm 2. One alternative to the above algorithm is the algorithm using explicitly
the regulator state w(i) . It has a disadvantage of being of higher order than Algorithm
1; however, it does not utilize any approximation of w(i) with v(i) . Recalling (4), we
obtain
v(i) K (i) F (i) w(i) + H (i)C(i) B(i) v(i) + [K (i) G(i) + H (i)C(i) A(i)C(i)+ ]y(i) ,
since w(i) is generated by the first relation in (3). If W = col{w(1) , . . . , w(N) }, then
we have, similarly as in the case of (7), that
+ K FW
+ [K G + H C AC + ]Y.
U = [ + H C B]U
The whole closed-loop system can be represented as

+ H C B K F (K G + H C AC + )C ' U (
U
W
=
W .
0
F
G C
x
x
BQ
0
A
(10)
(11)
Both control Algorithms 1 and 2 have the structure which reduces to the local
controllers when = 0. In the case of Algorithm 1, the local controllers are derived from Ci after aggregating (3) to one vector-matrix differential equation for v(i) ,
while in the case of Algorithm 2 the differential equation for v(i) contains explicitly
the term w(i) , generated by the local observer in Ci . The form of these controllers is
motivated by the idea to introduce a first order dynamic consensus scheme. Namely,
without the local controllers, relation U = U provides asymptotically a weighted
sum of the initial conditions Ui (t0 ), if the graphs corresponding to the particular
components of Ui have center nodes (see, e.g., [16, 25]). Combination of the two
terms provides a possibility to improve the overall performance by exploiting potential advantages of each local controller. However, the introduction of additional
dynamics, required by the consensus scheme, can contribute to deterioration of the
overall performance, and make the choice of the local controller parameters dependable upon the overall control scheme.
Example 1. An insight into the possibilities of the proposed algorithms can be obtained
in which the system S is represented by (1), with
from a simple example
' (
!
0.8 2
0
0
1
0
0
A = 2.5 5 0.3 , B = 1 and C =

. Assume that we have two
0 0 1
0
0 10 2
!
!
0.8 2
agents characterized by S1 with A(1) =
, B(1) = 0 and C(1) = [ 1 0 ],
2.5 5
1
203
!
!
5 0.3
1
(2)
(2)
,B =
and S2 with A =
and C(2) = [ 0 1 ]. Obviously, there
0
10 2
is only one control signal u. Assume that the second agent is responsible for
control implementation, so that u = u(2) = v(2) , according to the adopted notation. Assume that both agents have their own controllers C1 and C2 , obtained by
the LQG methodology, assuming
! a low measurement!noise level, so that we ob0.8502
1.6502
2.0000
(1)
, G(1) =
, K (1) = [ 0.7414 0.82231 ]
tain F =
2.4717 2.8223
0.26970
!
!
and H (1) = 0, and F (2) = 2.2361 24.3071 , G(2) = 24.2068 , K (2) =
3.1200
0.1000 1.1200
[ 0.2361 0.0003 ] and H (2) = 0. The system S with the local controller C2 is unstable. Algorithm 1 has been applied according to (7), after introducing Q = [ 0 1 ]
and 12 = 21 = 100I2. Figure 1 presents the impulse response for all three components of the state vector x for S. Algorithm 2 has then been applied according to
(11); the corresponding responses are presented in Fig. 2.
It is to be emphasized that the consensus scheme puts together two local controllers, influencing in such a way both performance and robustness. Here, the role
of the first controller is only to help the second controller in defining the control
signal. The importance of the consensus effects can be seen from Fig. 3 in which
the responses in the case when = 0 are presented for the Algorithm 1. It is obvious that the response is worse than in Fig. 1. In the case of Algorithm 2, the system
without consensus is even unstable (Fig. 4).
1.5
x1
1
x2
x3
0.5
0
0.5
1
1.5
2
2.5
3
10
Fig. 1 Impulse response for Algorithm 1
15
Time[s]
20
25
30
204

1.5
x1
x2
x3
0.5
0
0.5
1
1.5
2
2.5
10
Time[s]
15
20
1.5
x1
1
x2
x3
0.5
0
0.5
1
1.5
2
2.5
3
10
15
Time[s]
Fig. 3 Algorithm 1: local controllers without consensus
20
25
30
205
6
x1
x2
4
x3
50
100
Time[s]
150
200
Fig. 4 Algorithm 2: local controllers without consensus
The control algorithms can be made more flexible by introducing some adjustable
parameters, so that, for example, the terms K F K + in (7) and K F in (10) are multiplied by a parameter , and the term K G in both algorithms by ; it has been found
to be beneficial to have > 1 and < 1.
The problem of stabilizability of S by the proposed algorithms is, in general,
very difficult having in mind the supposed diversity of local models and dynamic
controllers. Any analytic insight from this point of view into the system matrices in
(9) and (11) seems to be very complicated. It is, however, logical to expect that the
introduction of the consensus scheme can, in general, contribute to the stabilization
of S. Selection of the elements of can, obviously, be done in accordance with the
expected performance of the local controllers and the confidence in their suggestions (see, for example, an analogous reasoning related to the estimation problem
addressed in the next section). In this sense, connectedness of the agents network
contributes, in general, to the overall control performance. The methodology of the
vector Lyapunov functions offers good possibilities for at least qualitative conclusions [19, 20].
3.2 Algorithms Derived from the Local Static Feedback

Control Laws
Algorithm 3. Assume now that we have static local output controllers, obtained
from Ci in (3) by introducing F (i) = 0, G(i) = 0 and K (i) = 0, so that we have
v(i) = H (i) y(i) . Both Algorithms 1 and 2 give in this case
206
BU
+ AC +Y ].
U = U + H C[
The closed-loop system is now given by
!3 4
!
+
U = + H C B H C AC C
U .
x
x
BQ
A
(12)
(13)
A special case of the above controller deserves particular attention. Assume in

the Algorithm 3 that C(i) = Ini and that that y(i) = x(i) and (i) = x(i) represents a part
of the vector x. In the special case when all the agents possess the knowledge about
the entire model of S, y(i) = x, and the agents can differ by their control laws and
responsibilities for control actions. Under these assumptions, Algorithm 3 becomes
BU
+ A x],
U = U + H[
(14)
where x = col{x(1) , . . . , x(N) }, dim{x}

= Ni=1 ni represents the expanded vector x,
available through measurements. Notice that it is always possible to find a full rank
Ni=1 ni n matrix V that x = V x (for a general discussion about state expansion, see
[20]). The closed-loop system is now
!
!3 4
U = + H B H AV
U .
(15)
x
x
BQ
A
Example 2. Assume that we have the same system as in Example 1, and that we have
the local stabilizing output feedback gains H (1) = 0.7414 and H (2) = 30, obtained
from the state feedback part of the local controllers formulated in Example 1, and
increasing the gain H (2) in order to have any significant effect of the feedback.
Figure 5 depicts the responses in the case when the consensus with 12 = 21 =
100I2 is applied. The system without consensus (with only one controller applied)
is unstable.
Remark 1. Properties of (14), can be somewhat clarified by the following consideration using a simple example. Assume that x = ax + bu, where x, u, a and b are
scalars, and assume that u = hx is a stabilizing feedback. Differentiating the last
relation one obtains u = hx = h(ax + bu). The obtained relation, together with the
system model, defines a system with one pole of the closed-loop system at a + hb,
and one at the origin. If one modifies the controller relation in such a way that
u = cu + h(ax + bu), one obtains a root locus with respect to c with two branches,
one of which goes from 0 to a, and the other from a + bk to . Formally, in (14),
the role of c is given to ; it comes out that the introduction of a structure modeling
+ A x generates a pole at the origin when BU
+ A x = Bu + Ax. When
x in (14) by BU
using local models, the introduced additional dynamics obviously deteriorates the
overall performance, so that the consensus matrix (together with the other controller
parameters), has as one of its duties to cope with this effect. Consequently, special
parameters can be introduced in the algorithm, such as negative diagonal terms in
respectively
, or already mentioned parameters and , multiplying B and A,
(a more rigorous analysis can be based on vector Lyapunov functions).
207
1.5
x1
1
x2
x3
0.5
0
0.5
1
1.5
2
2.5
3
10
15
20
Remark 2. The proposed multi-agent control schemes can be compared to those

overlapping decentralized control schemes for complex systems that are derived by
using the expansion/contraction paradigm and the inclusion principle (especially in
the case of Algorithm 3), e.g., [8, 9, 11, 20, 26, 28], having in mind that both approaches follow similar lines of thought (the above presented approach is, however,
much more general). From this point of view, formulation of the local controllers
connected to the agents corresponds to the controller design in the expanded space in
the case of the inclusion based design, and the application of the dynamic consensus
strategy to the contraction to the original space for control action implementation,
see, e.g., [8,9,20]. The proposed methodology offers, evidently, much more flexibility (local model structure, agreement strategy), at the expense of additional closed
loop dynamics introduced by the consensus scheme itself. Moreover, it is interesting
to notice that numerous numerical simulations show a pronounced advantage of the
proposed scheme (smoother and even faster responses). The reason could be found
in the advantage of the consensus strategy over the contraction transformation. In
Sect. 5 an application of the mentioned expansion/contraction methodology to the
control of formations of UAVs will be presented and compared to a consensus based
methodology.
4 Consensus at the State Estimation Level

The previous section was devoted to general structures for introducing consensus at
the input level in a multi agent system in which a number of agents with overlapping
resources and different competencies participate in defining the global control law.
208
In this section we will approach the problem in a different way, in which the consensus strategy is introduced at the level of state estimation. A consensus based
estimation scheme has been proposed in [25] for the continuous time case, and in
[24] for the discrete time case.
Assume that the local models are such that (i) = x(i) , so that the dynamic systems Si are overlapping subsystems of S [5, 20, 22, 23]. Starting from the model Si
and the accessible measurements y(i) , each agent is able to generate autonomously
its own local estimate x(i) of the vector x(i) using a local estimator which can be
defined in the following Luenberger form:
E i :
x(i) = A(i) x(i) + B(i)v(i) + L(i) (y(i) C(i) x(i) )
(16)
where L(i) is a constant matrix, which can be taken to be the steady state Kalman
gain, and v(i) is is the input which is supposed to be known.
The overlapping decentralized estimators (16) provide a set of overlapping estimates x(i) . However, if the final goal is to get an estimate x of the whole state vector
x of S, a consensus scheme can be introduced which would enable all the agents to
get reliable estimates of the whole state vector x on the basis of:
(1) the local estimates x(i) , and
(2) communications between the nodes based on a decentralized strategy uniform
for all the nodes.
In [25] an algorithm providing a solution to this problem has been proposed. If
Xi is an estimate of x generated by the ith agent, the following set of estimators is
attached to the agents in the network:
Ei : Xi = Ai Xi + Bi u + Nj=1 i j (X j Xi ) + Li (y(i) Ci Xi ),
(17)
j=i
i = 1, . . . , N, Ai is an n n matrix with ni ni nonzero elements being equal to

those of A(i) but being placed at the indices defined by Xi Xi while leaving zeros
elsewhere. Li is an n i matrix obtained similarly as Ai in such a way that its
nonzero elements are those of L(i) placed row by row at row-indices defined by
Xi leaving zeros elsewhere. Bi is an n m matrix obtained from Bi by putting its
rows at the indices defined by Xi while leaving zeros elsewhere, and i j , i = j, are
constant n n diagonal matrices with positive entries. The algorithm is, in fact,
based on a combination of decentralized overlapping estimators and a consensus
scheme with matrix gains i j , tending to make the local estimates Xi as close as
possible. If X = col{X1 , . . . , XN } is the vector composed of all the state estimates in
the agents network, the following model describes its global behavior:
E:
X = ( + A L C )X + B U + L Y,
(18)
where represents now an nN nN matrix composed of the blocks i j , i = j, with

ii = Ni=1,i= j i j , i = 1, . . . , N, A = diag{A1 , . . . , AN }, L = diag{L1 , . . . , LN },
C = diag{C1 , . . . ,CN }, B = diag{B1 , . . . , BN } and U = col{u, . . . , u}.
209
Moreover, we shall assume that all the agents have the a priori knowledge about
some convenient or optimal state feedback control law for S, expressed as u = K o x.
Using this knowledge and the estimation scheme (17), the agents can calculate
the corresponding inputs Ui = K o Xi on the basis of the separation principle. Accordingly, the implementation of the control signals is done according to the index
sets Ui .
Algorithm 4. The described decentralized overlapping estimation scheme with consensus, which provides state estimates of the whole state vector x to all the agents,
used in conjunction with the globally optimal state feedback control law, represents
a specific control algorithm which provides a solution to the general multi-agent
control problem of S.
Defining K o = diag{K o , . . . , K o }, we have, according to the above given notation,
that u = QK o X , so that the whole closed-loop system becomes
(
! '

3 4
K
L
C
+
A
L
+
B
X =
X ,
(19)
x
x
BQK o
A
where K = col{QK o , . . . , QK o } and L = col{L1C1 , . . . , LN CN }. A more realistic version of the above algorithm is obtained by replacing the actual input u in (17) by the
local estimates of the input vector Ui = K o Xi , having in mind the local availability
of Xi . This imposes, obviously, additional problems related to stability of the closed
loop system.
Example 3. In this example, performance of the above algorithm is demonstrated
on the same system as in the Example 1. The local estimators are performing the local state estimation using the gains L1 = [4 9]T and L2 = [2 7]T . The consensus
gains in the matrix are selected to be 12 = 21 = 100I2 . The global LQ optimal
control matrix K o is implemented by both agents. Since only the second agent implements the input u, we assume that the first one uses the estimate U1 = K o X1 in
the local state estimation algorithm. The impulse response of the proposed control
algorithm, which is shown in Fig. 6, is comparable to the the impulse response of
the globally LQ optimal controller shown in the same figure.
Stability analysis of Algorithm 4 represents in general a very complex task. It is
possible to apply the methodology of [21] under very simplifying assumptions, and
to show that the eigenvalues of (19) are composed of the eigenvalues of A L C ,
A + B K and A L C + B K modified by a term depending on the eigenvalues of
the Laplacian of the network and the consensus gain matrices. However, the underlying assumptions in [21] include the one that all the agents have the exact system
model, as well as that the control inputs are transmitted throughout the network; in
the overlapping decentralized case, which is in the focus of this work, these assumptions are violated, making the stability analysis much more complex, dependent on
the accuracy of the local models and the related estimators.
210
4
x1
2
x1opt
x2opt
x3opt
x2
x1
x2
x1opt
x3
x3
x2opt
x3opt
6
0
10
Time[s]
15
20
Fig. 6 Algorithm 4 and the globally LQ optimal controller
5 Consensus Based Decentralized Control of UAV Formations

In [27] decentralized control and estimation strategies are proposed for formations
of UAVs starting from a specific formation model and the definition of specific subsystems attached to the vehicles. The design of the corresponding control and estimation algorithms has been based on the expansion/contraction paradigm and the
inclusion principle [9, 20, 2628]. As a result, each vehicle, according to the available measurements, is able to generate its own control signal. In this section, we
shall show how the Algorithm 4 described in Sect. 4 can be applied to the control of
formations of UAVs. It will be demonstrated that superior performance can be obtained when the state estimation is based on consensus between the agents and when
the control signals are generated by using the globally LQ optimal state feedback
gain matrix.
5.1 Formation Model

Consider a set of N vehicles moving in a plane, where the ith vehicle is represented
by the linear double integrator model
!
!
022 I2
0
(20)
zi = Av zi + Bv ui =
zi + 22 ui ,
022 022
I2
211
(i = 1, . . . , N), where zi R4 and ui R2 are the state and the control input vectors,
respectively (0mn denotes the m n zero matrix, and In the n n identity matrix)
(see [28] for possible physical interpretations). The vehicle models (20) are coupled
through the control inputs. We shall assume that the ith vehicle is provided with
the information about distances with respect to the vehicles whose indices belong
to a set of indices of the sensed vehicles Si = {si1 , . . . , simi }. Accordingly, after de4T
4T
4T
3
3
3

T
z
z
z
z
composing zi as zi = zT
,
where
z
=
and
z
=
, we
z
i,1
i,2
i,1
i,2
i
i
i
i
define
xi =
ij zj zi,
xi = zi ,
(21)
jSi
3
4T
represents the distance between
where ij 0 and jSi ij = 1; xi = xi,1 xi,2
the ith vehicle and a centroid of the set of vehicles selected by Si obtained by
using a priori selected weights ij . In the case of formation leaders, when Si = 0,
/ we
have xi = zi . Therefore,
xi =
ij zj zi = ij xj xi ,
jSi
xi = ui ,
(22)
jSi
4T
4T
3
3
i = 1, . . . , N, using the fact that zi = zi , so that xi = xi,1 xi,2 = zi = zi,1 zi,2 =
4T
3
zi = zi,1 zi,2 .
The above described set of N vehicles with their sensing indices and the corresponding weights can be considered as a directed weighted graph G in which each
vertex represents a vehicle, and an arc with the weight ij leads from vertex j to vertex i if j Si . Consequently, the weighted adjacency matrix G = [Gi j ] is an N N
square matrix defined by Gi j = ij for j Si , and Gi j = 0 otherwise. We shall define
the weighted Laplacian of the graph as L = [Li j ], Li j = Gi j , i = j, Lii = j i j
(e.g., see [6]).
4T
3
T
Defining the vehicle state and control input vectors as xi = xT
and ui ,
x
i
i
i = 1, . . . , N, respectively, we obtain from (22) the following formation state model
S : x = Ax + Bu = [(G I) Av]x + [I Bv ]u,
(23)
where x and u are the formation state and control vectors defined as concatenations
of the vehicle state and control vectors, while denotes the Kroneckers product.
We shall assume3 that each
4T vehicle has the information about the reference state
T
T
, so that the control task to be considered is the task of
trajectories ri = ri ri
tracking the desired, possibly time varying, references.
212
5.2 Global LQ Optimal State Feedback

We shall attach the following quadratic criterion to (23)
J=

0
(xT Qx + uT Ru)dt,
(24)
where Q 0 and R > 0 are appropriately defined matrices. The design of the state
feedback gain minimizing J is faced with the problem that the model (23) is in
general not completely controllable. Namely, one can directly observe that the part
of the state vector of (23) which corresponds, for example, to the relative positions
3
4T
with respect to the first axis x1 = xi,1 xN,1 , satisfies the relation x1 = (I
G)p1 , where p1 is the vector of absolute vehicle positions with respect to a reference
frame. If the graph G has a spanning tree, the Laplacian L has one eigenvalue at the
origin, and the rest in the open left-half plane, e.g., [15, 18]. This means that in
the case when I G = L, we have rT x = 0, where rT is the left eigenvector of L
corresponding to the zero eigenvalue [27]. The controllability matrix does not have
full rank since rank [ B AB ] = 2(N 1) (having in mind that A2 = 0). However, it
is possible to observe that the system is in this case controllable for the admissible
initial conditions for (23) which have to satisfy rT x0 = 0 for a real formation. Notice
that in the case of no formation leader, i.e., when I G = L, the matrix I G is
nonsingular provided G has a spanning tree. A way of solving the! problem can be
T
in which W is
seen after applying to x a nonsingular transformation T = r
W
a full rank matrix such that rT is linearly independent of the rows of W . It can
be seen that S is controllable for all the admissible initial conditions provided the
formation model is controllable with respect to v = W x. Models for v in the form
Sa : v = Aa v + Ba u.
(25)
represent aggregations of S having in mind that W is a full rank matrix [20, 22, 26].
The system matrices satisfy then the aggregation conditions WA = AaW and Ba =
W B [20]. In order to take care of optimality, we shall attach to (25) the following
criterion
J =
a

0
(vT Qa v + uT Ra u)dt.
(26)
Obviously, the criterion J includes the criterion J a , i.e., J = J a , if W T QaW = Q

and R = Ra (see [10, 20] for a general discussion about the inclusion of performance indices). If one starts from J, an approximate solution to the posed
optimization problem can be found by formulating J a using the approximate relation Qa = W +T QW + (where W + denotes the pseudo-inverse of W ) and solving the minimization problem of J a taking Sa as a constraint. If K a is obtained
as the corresponding optimal feedback gain matrix, a feedback gain matrix K for
S can be found simply by applying the relation K = K aW [22], since in this case
213
the closed-loop system (Sa , K a ) is an aggregation of the closed-loop system (S, K)

and J = J a . Notice that W can be chosen in many different ways: a simple choice
0.5 0 0.5

0 1 0
is, for example, W =

0 0 0 1 , which does not substantially change
1
the structure of Qa with respect to Q. Also, for a given W , different choices of
Aa are possible; having in mind the sparsity of A, Aa can be found for adequate
choices of W by using simple linear combinations of the rows of A and deleting its
columns.
5.3 Decentralized State Estimation

We shall consider formation state estimation starting from a convenient formation
model involving all the inter-vehicle distances measured by the agents
S : x = A x + Bu,
where x =
xcT
1
...
'
{A c1 . . . A cN }, A ci =
xcT
N
4T
, xci =
A c
i
022 N
(27)
!T
(zsi zi )T (zsi zi )T (zi )T , A = blockcol
mi
1
(
in which A c is composed of N blocks in a
k=1 (mk +1)
+ 1), k =1, ..., N; these blocks are nonzero

block-row having dimensions 2mi 2(mk
I2
only for: 1) k = i, when it has the form 0 ...

and for 2) all k Si
I2 2m 2(m +1)
i
i
..
.
when they have the form

in which I2 is placed at the row index
0 I2
..
. 2mi 2(mk +1)
!
02mi 2
j satisfying sij = k, and B = blockcol {B c1 . . . B cN }, B ci =
.
I2
It is easy to verify that x = T x,
where T = blockdiag{T1 . . . TN } and Ti =
' i
(
i
si I2 si I2
mi
the system ma1
, so that S represents an aggregation of S;
I2
trices are, consequently, related by T A = AT and B = TB [9, 20].
Following the basic idea exposed in [27], we shall extract the following overlapping subsystems from S
S (i) :
x(i) = A (i) x(i) + B (i) u(i) ,
(28)
214
!T
T
T

T

T
T
i = 1, . . . , N, where x(i) = (zsi1 ) (zsim ) (zsi1 zi ) (zsim zi ) (zi )
,
i
i
02mi 2(mi +1)

!T
T
T
T
(i)
(i)
u = usi usi ui
, A =
A i
, in which A i has the form
mi
I2mi
0
I2
.
0 ..
I2
022(mi +1)
I2mi
02mi 2
and B (i) = 02mi 2mi 02mi 2 .
022mi
I2
Following the idea exposed in [27], models S (i) can be used for constructing local
observers of Luenberger type providing overlapping estimates x(i) of the subsystem
states:
E(i) :
x(i) = A (i) x(i) + B (i) u(i) + L(i) [y(i) C(i) x(i) ],
(29)
where y(i) is the local measurement vector available in the ith vehicle, C(i) a matrix
defining mapping from the local state to the local output, and L(i) is the estimator
gain (compare with (17)). It should be noticed that the formulated local observers
(29) can be implemented provided the control vectors u(i) are appropriately defined.
When we have decentralized control like in the case treated in [27], the local control law for the ith vehicle gives ui , while the remaining elements of u(i) are then
obtained by considering the sensed vehicles as leading vehicles, with their own
velocity feedbacks adapted to the given velocity reference (see, for example, the
methodology based on LQ optimization applied to the platooning problem [26]).
In order to obtain the estimates of the whole state vector x of S by all the agents,
a consensus scheme can be introduced as in Sect. 4. If xi is the estimate of the whole
formation state vector x generated by the ith agent, the following consensus based
estimator results now directly from (17)
Ei : xi = Ai xi + BiUi + Nj=1 i j (x j xi ) + Li (y(i) Ci xi ),
(30)
j=i
i = 1, . . . , N, where matrices Ai , Bi and Li are obtained from A (i) , B (i) and L(i) sim while matrices i j contain
ilarly as in the case of (17) using the model (27) for S,
the consensus parameters. The main point here is generation of the global input vector which should be introduced in (30) and which is not available to all the agents.
Vector Ui introduced in (30) represents an approximation of u achievable by the ith
agent based on the assumption that all the agents are supposed to know the global
state feedback gain K in the optimal mapping u = Kx = K aW x, where K a is obtained
by minimizing J a in (26). Therefore, we have, according to Sect. 4, that Ui = KT xi ;
this means that the implemented control signal ui is the i-th component of the Ndimensional vector Ui . The resulting closed loop system model can be obtained
directly using (19).
215
5.4 Experiments
In this section we shall illustrate the proposed method for control of formations of
UAVs. A formation of four vehicles without a formation leader has been simulated.
It has been assumed that the second and the third vehicle observe the first, the fourth
observes the second and the third, while the first vehicle observes the fourth one. The
globally LQ optimal feedback gain has been found on the basis of Sect. 5.2. The consensus based estimator, proposed in Sect. 5.3 has been implemented by each agent,
assuming the adopted information flow between the agents. The consensus gains are
all set to be the same, equal to 100. In Figs. 7 and 8 x-components of the distances
and velocities of all four vehicles in the formation are depicted, assuming step distance reference change. On the other hand, Figs. 9 and 10 represent the responses of
the same formation with the controllers designed using the expansion/contraction
paradigm and inclusion principle with local estimators [8, 20, 26, 28]. It is obvious that better performance is obtained using the consensus based control structure, at the expense of additional communications between the vehicles in the
formation.
reference
actual
1
1.5
2
0
10
15
20
25
30
10
15
20
25
30
10
15
20
25
30
10
15
Time[s]
20
25
30
1
0.5
0
1
0.5
0
1
0.5
0
Fig. 7 Distance plots: consensus based controllers
216

2
reference
actual
1
0
10
15
20
25
30
10
15
20
25
30
10
15
20
25
30
10
15
Time[s]
20
25
30
1
0.8
1
0.8
2
1
0
Fig. 8 Velocity plots: consensus based controllers

0
reference
actual
2
4
10
15
20
25
30
10
15
20
25
30
10
15
20
25
30
10
15
Time[s]
20
25
30
2
1
0
2
1
0
2
1
0
Fig. 9 Distance plots: expansion/contraction based controllers

3
2
1
0
1
217
reference
actual
10
15
20
25
30
10
15
20
25
30
10
15
20
25
30
10
15
Time[s]
20
25
30
3
2
1
0
3
2
1
0
3
2
1
0
Fig. 10 Velocity plots: expansion/contraction based controllers
References
1. B. Baran, E. Kaszkurewicz, and A. Bhaya, Parallel asynchronous team algorithms: convergence and performance analysis, IEEE Trans. Parallel Distrib. Syst. 7 (1996), 677688
2. D. P. Bertsekas and J. N. Tsitsiklis, Parallel and distributed computation: Numerical methods,
Prentice-Hall, Englewood Cliffs, 1989
3. V. D. Blondel, J. M. Hendrickx, A. Olshevsky, and J. N. Tsitsiklis, Convergence in multiagent
coordination, consensus and flocking, Proc. IEEE Conf. Decision Contr., 2005
4. C. G. Cassandras and W. Li, Sensor networks and cooperative control, Eur. J. Contr. 11 (2005),
436463
5. X. B. Chen and S. S. Stankovic, Decomposition and decentralized control of systems with
multi-overlapping structure, Automatica 41 (2005), 17651772
6. A. Fax and R.M Murray, Information flow and cooperative control of vehicle formations, IEEE
Trans. Automat. Contr. 49 (2004), 14651476
7. H. Gharavi and S. Kumar (eds.), Proceedings of the IEEE: Special Isssue on Sensor Networks
and Applications, vol. 91, August 2003
8. M. Ikeda and D. D. Siljak,

Decentralized control with overlapping information sets, J. Optim.
Theory Applic. 34 (1981), 279310
, Overlapping decentralized control with input, state and output inclusion, Contr. The9.
ory Adv. Technol. 2 (1986), 155172
10. M. Ikeda, D. D. Siljak,

and D.E. White, Decentralized control with overlapping information
sets, J. Optim. Theory Applic. 34 (1981), 279310
11.
, An inclusion principle for dynamic systems, IEEE Trans. Autom. Contr. 29 (1984),
244249
12. A. Jadbabaie, J. Lin, and A. Morse, Coordination of groups of mobile autonomous agents
using nearest neighbor rules, IEEE Trans. Automat. Contr. 48 (2003), 9881001
218
13. Z. Lin, B. Francis, and M. Maggiore, Necessary and sufficient conditions for formation control
of unicycles, IEEE Trans. Automat. Contr. 50 (2005), 121127
14. R. Olfati-Saber and R. Murray, Consensus problems in networks of agents with switching
topology and time-delays, IEEE Trans. Automat. Contr. 49 (2004), 15201533
15. W. Ren and E. Atkins, Distributed multi-vehicle coordinated control via local information
exchange, Int. J. Robust Nonlinear Contr. 17 (2007), 10021033
16. W. Ren, R. W. Beard, and E. M. Atkins, A survey of consensus problems in multi-agent coordination, Proceedings of American Control Conference, 2005, pp. 18591864
17. W. Ren, R. W. Beard, and D. B. Kingston, Multi-agent Kalman consensus with relative uncertainty, Proceedings of American Control Conference, 2005
18. W. Ren and R.W. Beard, Consensus seeking in multi-agent systems using dynamically changing interaction topologies, IEEE Trans. Autom. Contr. 50 (2005), 655661
19. D. D. Siljak,
Large scale dynamic systems: Stability and structure, North-Holland, New York
1978
20. D. D. Siljak,
Decentralized control of complex systems, Academic, New York, 1991
21. R. S. Smith and F. Y. Hadaegh, Closed-loop dynamics of cooperative vehicle formations with
parallel estimators and communication, IEEE Trans. Autom. Contr. 52 (2007), 14041414
22. S. S. Stankovic and D. D. Siljak,

Contractibility of overlapping decentralized control, Syst.
Contr. Lett. 44 (2001), 189199
23. S. S. Stankovic and D. D. Siljak,

Stabilization of fixed modes in expansions of LTI systems,
Syst. Contr. Lett. 57 (2008), 365370
24. S. S. Stankovic, M. S. Stankovic, and D. M. Stipanovic, Consensus based overlapping decentralized estimation with missing observations and communication faults, Automatica 45
(2009), 13971406
25. S. S. Stankovic, M. S. Stankovic, and D. M. Stipanovic, Consensus based overlapping decentralized estimator, IEEE Trans. Autom. Contr. 54 (2009), 410415
26. S. S. Stankovic, M. J. Stanojevic, and D. D. Siljak,

Decentralized overlapping control of a
platoon of vehicles, IEEE Trans. Contr. Syst. Technol. 8 (2000), 816832
27. S. S. Stankovic, D. M. Stipanovic, and M. S. Stankovic, Decentralized overlapping tracking
control of a formation of autonomous unmanned vehicles, Proceedings of American Control
Conference, 2009
28. D. M. Stipanovic, G. Inhalan, R. Teo, and C. Tomlin, Decentralized overlapping control of a
formation of unmanned aerial vehicles, Automatica 40 (2004), 12851296
29. J. N. Tsitsiklis, Problems in decentralized decision making and computation, Ph.D. thesis,
Dep. Electrical Eng. Comput. Sci., M.I.T., Cambridge, MA, 1984
30. J. N. Tsitsiklis, D. P. Bertsekas, and M. Athans, Distributed asynchronous deterministic and
stochastic gradient optimization algorithms, IEEE Trans. Autom. Contr. 31 (1986), 803812
31. P. Yang, R.A. Freeman, and K.M. Lynch, Multi-agent coordination by decentralized estimation and control, IEEE Trans. Autom. Contr. 53 (2008), 24802496
Graph-Theoretic Methods for Networked

Dynamic Systems: Heterogeneity
and H2 Performance
Daniel Zelazo and Mehran Mesbahi
1 Introduction
The analysis and synthesis of large-scale systems pose a range of challenges due to
the number of their subsystems and the complexity of their interactions. In the meantime, the importance of these systems has become increasingly prevalent in science
and engineering, especially in the realm of multi-agent systems such as coordination of autonomous vehicles on ground, in sea, air, and space, as well as localization
and sensor fusion, energy networks, and distributed computation [1,2,47,9,19,21].
One aspect of the complexity of large-scale systems is that their subsystems, which
we occasionally refer to as agents, may not be described by the same set of inputoutput dynamics. The difference between the dynamics of distinct subsystems may
be the result of manufacturing inconsistencies or intentional differences due to varying requirements for each subsystem in the ensemble. An important component to
the analysis of these systems, therefore, is to understand how heterogeneity in the
subsystem dynamics affects the behavior and performance of the entire system. Another facet of this complexity relates to the interconnection of different subsystems.
The underlying interconnection topology of a large-scale system may be determined
by the governing dynamic equations of each subsystem, e.g., when the interconnection is a function of some state of each subsystem, or it may be designed as
part of the engineering process. In both cases, the interconnection topology has a
profound impact on the overall system in terms of its stability, controllability, observability, and performance. Hence, it becomes crucial to understand and explicitly
D. Zelazo
Institute for Systems Theory and Automatic Control, University of Stuttgart,
Pfaffenwaldring 9, 70550 Stuttgart, Germany
e-mail: Daniel.Zelazo@ist.uni-stuttgart.de
M. Mesbahi
Department of Aeronautics and Astronautics, University of Washington,
Box 352400, Seattle, Washington, USA
e-mail: mesbahi@aa.washington.edu
219
220
D. Zelazo and M. Mesbahi
parameterize the role of the interconnection topology for both analysis and synthesis
of large-scale systems.
This chapter aims to address these issues via the development of four canonical
models of large-scale systems, those that will be collectively referred to as networked dynamic systems (NDS). A crucial tool that we employ to derive these models are algebraic representations of graphs and networks [12].
The four systems are NDS coupled at the output, NDS coupled at the input, NDS
coupled at the state, and NDS coupled at a combination of input, output, and state.
We further subdivide each class by distinguishing between NDS with homogeneous
agent dynamics and NDS with heterogeneous agent dynamics. For simplicity of presentation, we will focus on continuous linear time-invariant systems; an analogous
framework can be developed for the discrete-time case.
Graph-centric analysis of networked dynamic systems has been extensively
treated in the literature. Consensus-type problems, which fall under the category
of NDS coupled at the state, are prime examples of how notions from graph theory can be applied in a dynamic systems setting [27]. Examples of such studies
include Nyquist-based stability analysis for consensus-based feedback systems [9],
graph-centric notion of controllability in consensus problems [25], and consensus
algorithms with guaranteed H performance [18]. Works related to formation flying applications which rely on relative sensing, falling under the category of NDS
coupled at the output, also use results from algebraic and spectral graph theory
[19, 24, 31, 39].
The outline of this chapter is as follows. First, we describe the models for various
classes of NDS in Sect. 2. In Sect. 3 we proceed to present a graph-theoretic analysis
for four classes of NDS models. The goal of this section is to highlight the role of
heterogeneity and the interconnection topology on the system-theoretic properties
of NDS. In many instances, the differences between agent dynamics and the interconnection topology can be embedded into the system matrices resulting in a single
state-space representation of the system. The discussion in Sect. 3 aims to emphasize
the importance of keeping the interconnection topology explicit in the formulation
of the model. That is, for analysis purposes, we prefer to consider a quintuple representation of the system, (A, B, C, D, G), where G denotes the underlying connection
topology. Using this approach we will then examine the controllability and observability properties of different NDS models as well as a graph-theoretic characterization of their H2 performance. The results obtained from the analysis of NDS will
then motivate techniques for the design of the underlying interconnection topology
in 4. In this section, we will rely on results from combinatorial optimization and
semi-definite programming that lead to numerically tractable solutions for topology design. For NDS coupled at the output model, we show that with an appropriate representation of the corresponding network synthesis problem, the celebrated
Kruskals algorithm can be used to find an optimal topology in the H2 setting. For
NDS coupled at the state, we propose a convex relaxation for the minimum cost
sensor placement problem that leads to a semi-definite program. Finally in Sect. 5,
we present some concluding remarks regarding the implications of the framework
discussed in this chapter.
Graph-Theoretic Methods for NDS: Heterogeneity and H2 Performance
221
1.1 Preliminaries and Notations

We provide some mathematical preliminaries and notations here. Matrices are denoted by capital letters, e.g., A, and vectors by lower case letters, e.g., x. Diagonal
matrices will be written as D = diag{d1 , . . . , dn }; this notation will also be employed
for block-diagonal matrices and linear operators. A matrix and/or a vector that consists of all zero entries will be denoted by 0; whereas, 0 will simply denote the
scalar zero. Similarly, the vector 1 denotes the vector of all ones, and J = 11T . The
set of real numbers will be denoted as R, and . p denotes the p-norm of its argument, e.g., p = 2, , which is used for vector, matrix, and system norms. The adjoint
of a linear operator f is denoted by f . The notation A B denotes the Hadamard
product of the two matrices [15]. The Kronecker product of two matrices A and B
on the other hand is written as A B. The following result for Kronecker products
will be used subsequently.
Theorem 9.1 ([14]). Let A Rmn and B R pq each have a singular value decomposition of A = UA AVAT and B = UB BVBT . The singular value decomposition
of the Kronecker product of A and B is then
A B = (UA UB )(A B )(VAT VBT ).
(1)
An immediate consequence of Theorem 9.1 is the identity

A B2 = A2 B2,
(2)
for the matrix 2-norm. We also make use of the multiplication property
(A B)(C D) = (AC BD),
(3)
for Kronecker products, where the matrices are all of commensurate dimension.
Graphs and the matrices associated with them form a convenient construct for
much of the work presented in this chapter. The reader is referred to [12] for a
detailed treatment of the subject and we present here only a minimal summary of
relevant notions and results. An undirected (simple) graph G is specified by a vertex
set V and an edge set E whose elements characterize the incidence relation between
distinct pairs of V. Two vertices i and j are called adjacent (or neighbors) when
{i, j} E; we denote this by writing i j. The cardinalities of the vertex and edge
sets of G will be denoted by |V| and |E|, respectively. An orientation of an undirected
graph G is the assignment of directions to its edges, i.e., an edge ek is an ordered pair
(i, j) such that i and j are, respectively, the initial and the terminal nodes of ek . In
our discussion, we make extensive use of the |V| |E| incidence matrix, E(G), for a
graph with arbitrary orientation. The incidence matrix is a {0, 1}-matrix with rows
and columns indexed by the vertices and edges of G such that [E(G)]ik has the value
1 if node i is the initial node of edge ek , 1 if it is the terminal node, and 0 otherwise. The degree of vertex i, di , is the cardinality of the set of vertices adjacent to
it; we define the degree matrix as (G) = diag{d1 , . . . , d|V| }. The adjacency matrix
222
of an undirected graph, A(G), is the symmetric |V| |V| matrix such that [A(G)]i j
takes the value 1 if node i is connected to node j, and 0 otherwise.
A connected graph G can be written as the union of two edge-disjoint subgraphs
on the same vertex set as G = G Gc , where G is a spanning tree subgraph and Gc
contains the remaining edges that necessarily complete the cycles in G. Similarly,
the columns of the incidence matrix for the graph G can always be permuted such
that E(G) is written as

E(G) = E(G ) E(Gc ) .
(4)
The cycle edges can be constructed from linear combinations of the tree edges
via a linear transformation [28], as
where
E(G )Tc = E(Gc ),
(5)

1
E(G )T E(Gc ).
Tc = E(G )T E(G )
(6)
Using (5) we obtain the following alternative representation of the incidence matrix
of the graph

E(G) = E(G ) I Tc = E(G )R(G);
(7)
the rows of the matrix

R(G) = I Tc
(8)
T

are viewed as the basis for the cut space of G [12]. The matrix Tc I , on the
other hand, forms a basis for the flow space.
The matrix R(G) has a close connection with a number of structural properties
of the underlying network. For example, the number of spanning trees in a graph,
(G), can be determined from the cut space basis [12], as

(9)
(G) = det R(G)R(G)T .
The (graph) Laplacian of G,
L(G) := E(G)E(G)T = (G) A(G),
(10)
is a rank deficient positive semi-definite matrix. The eigenvalues of the graph Laplacian are real and will be ordered and denoted as
0 = 1 (G) 2 (G) |V| (G).
The edge Laplacian is defined as [40]
Le (G) := E(G)T E(G).
(11)
The edge Laplacian is intimately related to the graph Laplacian, as shown

through the following similarity transformation.
223
Theorem 9.2. The graph Laplacian for a connected graph L(G) containing cycles
is similar to
!
Le (G )R(G)R(G)T 0
,
0
0
where G is a spanning tree subgraph of G and the matrix R(G) is defined via (8).
Proof. We define the transformation
3
4

Sv (G) = E(G ) E(G )T E(G ) 1 1 , Sv (G)1 =
!
E(G )T
,
(1/|V|) 1T
(12)
where E(G ) is the incidence matrix of G(G ). Applying the transformation as

!

E(G )T E(G )
Sv (G)1 L(G)Sv (G) =
R(G)R(G)T I 0
0
!
Le (G )R(G)R(G)T 0
,
(13)
=
0
0
leads to the desired result.
The transformation (13) provides a transparent way to separate the zero eigenvalue
of the Laplacian for a connected graph while preserving algebraic properties of the
graph via the edge Laplacian.
In order to apply the framework developed here to specific graphs, we will work
with the complete graph and its generalization in terms of k-regular graphs, which
are defined as follows. The complete graph on n nodes, Kn , is the graph where all
possible pairs of vertices are adjacent, or equivalently, if the degree of all vertices is
| V | 1. Figure 1(a) depicts K10 , the complete graph on ten nodes. When every node
2
4
9
7
10
10
(a) K10 Graph

Fig. 1 Example of regular graphs on ten vertices
(b) A 4-regular Graph
224
in a graph with n nodes has the same degree k n 1, it is called a k-regular graph.
The k-regular graph on n nodes for k = 2 is called the cycle graph, Cn . Figure 1(b)
shows a 4-regular graph.
2 Canonical Models of Networked Dynamic Systems

In this section we develop a general linear time-invariant model for networked dynamic systems with an emphasis on the means by which the underlying connection
topology enters into the system. As alluded to in the introduction, we develop a
model which explicitly highlights how the underlying connection topology interacts
with each agent in the ensemble.
Fundamental to all NDS is the notion of a local and global dynamic system
layer. The local layer corresponds to the dynamics of each individual agent in the
ensemble. This layer captures both the dynamic behavior of each agent in addition to
local performance criteria that may or may not be related to certain global or team
objectives. For example, the formation control for a team of unmanned vehicles
may require each agent to perform a local control and estimation in order to accept
higher level navigation commands relating to the team objective. In this direction,
we identify two broad classes of NDS: (1) homogeneous, and (2) heterogeneous.
For both cases, we will work with a group of n dynamic systems, referred to as
agents, each modeled as a linear and time-invariant system of the form
xi (t) = Ai xi (t) + Bi ui (t) + i wi (t)

zw
i : zi (t) = Ciz xi (t) + Dzu
(14)
i ui (t) + Di wi (t)
y
yw
yi (t) = Ci xi (t) + Di wi (t),
where each agent is indexed by the sub-script i. Here, xi (t) Rni represents the
state, ui (t) Rmi the control, wi (t) Rri an exogenous input (e.g., disturbances
and noises), zi (t) R pi the controlled variable, and yi (t) Rbi the locally measured
output.
When working with homogeneous NDS, the subscript is dropped, as each agent
is described by the same set of linear state-space dynamics (e.g., i = j for all i, j).
It should be noted that in a heterogeneous system, the dimension of each agent need
not be the same; however, without loss of generality, we assume each agent to have
the same dimension.
The parallel interconnection of all the agents has a state-space description
x (t) = Ax(t) + Bu(t) + w(t),

z(t) = Cz x(t) + Dzu u(t) + Dzw w(t),
(15)
:
y(t) = Cy x(t) + Dyw w(t),

with x(t), u(t), w(t), z(t), and y(t) denoting, respectively, the concatenated state
vector, control vector, exogenous input vector, controlled vector, and output vector
225
of all the agents in the NDS. The bold faced matrices represent the block diagonal
aggregation of each agents state-space matrices, e.g., A = diag{A1 , . . . , An }.
Given the above model for each agent and motivated by the diverse applications
of multi-agent systems, we can begin to incorporate the role of the interconnection
topology G. To begin, we first define four canonical classes of NDS models. Such
a classification is useful for analysis purposes; we will also show in the sequel that
under certain conditions they are, in a sense, equivalent.
NDS Coupled at the Output
In this class of NDS, the underlying network topology couples each agent through
their outputs. Systems relying on relative sensing to achieve global objectives such
as formation flying fall under this classification [10, 17, 31]. The block diagram
in Fig. 2 shows how the connection topology interacts with each agent. Here we
have shown disturbances entering each agent and the global output of the entire
system. An important feature of these types of systems is the underlying connection
topology does not affect, in the open-loop, the dynamic behavior of each agent.
Motivated by applications that rely on relative sensing, we now derive a mathematical model to capture the global layer of this type of NDS. The sensed output of the system is the vector yG (t) containing relative state information of each
agent and its neighbors. The incidence matrix of a graph naturally captures differences and will be the algebraic construct used to define the relative outputs.
For example, the output sensed between agent i and agent j would be of the form
yi (t) y j (t). This can be compactly written using the incidence matrix for the entire
system as
yG (t) = (E(G)T I)y(t).
(16)
Here, G is the graph the describes the connection topology; the node set is given as
V = {1, . . . , n}.
When considering the analysis of the global layer, we are interested in studying
the map from the agents exogenous inputs, w(t), to the sensed output of the NDS,
yG (t). Therefore, for the homogeneous NDS, the system in (15) is augmented to
include the sensed output
yG (t) = (E(G)T Cy )x(t).
(17)
w(t)
A
C
B
D
Gout
yG(t)
Fig. 2 NDS coupled at the output; the feedback connection represents an upper fractional transformation [8]
226
For the heterogeneous case, the sensed output is

yG (t) = (E(G)T I)Cy x(t).
(18)
Remark 9.1. For relative sensing, the observation matrix Cy used in (17) and (18)
may in fact be different from the local observation of each agent, as described in
(14).
In the context of NDS coupled at the output, we denote by hom (G) the homogeneous system (15) with the additional sensed output (17). The heterogeneous system
will be denoted by het (G) and corresponds to the system (15) with the additional
sensed output (18).
wG
wG
For notational simplicity, we denote Thom
and Thet
as the map from the exogenous inputs to the NDS sensed output for homogeneous and heterogeneous systems
respectively.
NDS Coupled at the Input
In this class of NDS, the underlying network topology enters at the system input.
The block diagram in Fig. 3 shows a networked input being distributed to each agent
via an interconnection topology. Large physically coupled systems where actuation
affects multiple components might be modeled in this way. In fact, this class of NDS
may even be considered the dual of the output coupled NDS presented above. The
agents are therefore coupled via the inputs. To maintain a close connection with
the NDS coupled at the output model, we will assume the network input, uG (t), is
distributed to each agent via the incidence matrix. The control applied to each agent,
therefore, is the net contribution of the control applied to all the edges incident
to that agent. When the underlying graph is directed and connected, each agents
control can be written as
ui (t) =
[uG (t)](i, j)
(i, j)E
[uG (t)]( j,i) ,
(19)
( j,i)E
where [uG (t)](i, j) denotes the component of the input vector corresponding to the
directed edge (i, j). This can be compactly written using the incidence matrix and
Kronecker products to obtain a complete model for NDS coupled at the input.
uG (t)
Gin
A
C
B
D
y(t)
Fig. 3 NDS coupled at the input; the feedback connection represents an upper fractional transformation [8]
227
For the homogeneous case, we have

u(t) = (E(G) B)uG (t),
(20)
and for the heterogeneous case

u(t) = B(E(G) I)uG (t).
(21)
The parallel interconnected system (15) can be modified to include the network
input using (20) and (21) as
x (t) = (In A)x(t) + (E(G) B)uG (t) + (In )w(t),

hom (G) : z(t) = (In Cz )x(t) + (In Dzu )u(t) + (In Dzw )w(t), (22)
y(t) = (I Cy )x(t) + (I Dyw )w(t),

n
n
and
x (t) = Ax(t) + B(E(G) I)uG (t) + w(t),

Dzu u(t)
+ Dzw w(t),
het (G) : z(t) = Cz x(t) +
y(t) = Cy x(t) +
Dyw w(t).
(23)
NDS Coupled at the State

This class of NDS is perhaps one of the most studied in the systems and control
community. In this type of NDS, the underlying topology couples each agent at
the state level, resulting in an important connection between the dynamic evolution
of each agent and the underlying topology. The block diagram in Fig. 4 shows the
connection topology entering a dynamic system at the state level.
The most general way to model such systems is to simply denote the dependence
of the state-matrix on the network with the notation A(G). For our purposes, however, we will focus on a special instance of this system, known as the agreement or
consensus protocol [20, 23, 26]. Consequently, we will only focus on homogeneous
systems for this case.
Gstate
w(t)
A
C
B
D
y(t)
Fig. 4 NDS coupled at the state; the feedback connection between the plant matrices and the integrator represents an upper fractional transformation [8], whereas the feedback connection between
the integrator and the graph represents a relation such as (26)
228
The consensus model is built upon a general setup consisting of a group of n

identical single integrator units,
xi (t) = ui (t),
i = 1, . . . , n,
(24)
each connected to a fixed number of other units in the ensemble, determined by the
interconnection topology G. The interaction or coupling between units dynamics
is realized through the control input ui (t) in (24), assumed to be the sum of the
differences between states of an agent and its neighbors, i.e.,
ui (t) =
(x j (t) xi(t)).
(25)
i j
Expressing the dynamic evolution of the resulting system in a compact matrix form
one has
x (t) = L(G) x(t),
(26)
where L(G) is the graph Laplacian. This model can be extended to include exogenous inputs and controlled variables, which will be discussed in the sequel.
Extensions of this model have been extensively treated, including random networks [13, 32], switching topologies [22], and noisy networks [36].
NDS Coupled by Combinations of State, Input, and Output
A natural extension of the above models is to consider systems that have a network
coupling the agents at all component levels. Figure 5 shows a dynamic system where
there are different connection topologies at the input, output, and state level. Clearly,
this type of model represents the most complex and intricate connection between the
dynamic properties and interconnection topology of a system. It is worth noting that
although this type of system can be exhaustively studied on its own, we only present
it here for completeness and as a vehicle to illustrate how each of the previous types
can be interrelated.
Gstate
uG(t)
Gin
A
C
B
D
Gout
yG(t)
Fig. 5 NDS coupled at the state, input, and output; the feedback connection between the plant
matrices and the integrator represents an upper fractional transformation [8], whereas the feedback
connection between the integrator and the graph represents a relation such as (26)
229
3 Analysis and Graph-Theoretic Performance Bounds

The development of the NDS models in Sect. 2 allows us to examine systemstheoretic properties and performance bounds for networked system from a graphtheoretic perspective. The objective of this section, therefore, is to develop explicit
connections between control-theoretic concepts such as observability, controllability, and performance in terms of the underlying interconnection graph. In this section we will first discuss the observability properties of NDS coupled at the output,
and then proceed to highlight how duality streamlines the controllability analysis of
NDS structure coupled at the input. We then focus on characterizing the H2 performance of NDS coupled at the output and at the state via constructs from algebraic
graph theory.
3.1 Observability and Controllability of NDS

Studying the observability and controllability properties of a linear system can provide qualitative as well as quantitative insights into the design of the corresponding
controllers and estimators. In the context of NDS, we also consider how the underlying topology affects these properties in addition to examining the effects of
homogeneity and heterogeneity of the agent dynamics comprising the NDS. In this
direction, we consider simplified versions of the models presented in Sect. 2. For
example, for observability analysis, we consider the following simplified model of
an NDS coupled at the output,

x (t) =
Ax(t)
.
(27)
yG (t) = (E(G)T I)Cy x(t)
Analogously, the following simplified model for an NDS coupled at the input will
be used to study the controllability properties,
x (t) = Ax(t) + B(E(G) I)uG(t);
(28)
for both systems, we will examine the homogeneous and heterogeneous realizations.
Our observability and controllability analysis relies on the observability and controllability gramians for networked systems. Recall that the observability gramian of
a stable linear system with state matrix A and observation matrix C can be written as
Yo =
eA t CT CeAt dt.
(29)
Similarly, the controllability gramian of a stable linear system with state matrix A
and input matrix B can be written as
Xc =

0
eAt BBT eA t dt.
(30)
230
For this analysis, we will assume that each agent is stable, e.g., Ai is Hurwitz, the
y
pair (Ai ,Ci ) is observable, and (Ai , Bi ) is a controllable pair.
Observability of NDS Coupled at the Output
A natural question for this analysis is whether the initial condition of each agent in
an NDS coupled at the output can be inferred from their relative states. The answer
to this question can have profound implications for the design of estimators for such
systems.
For homogeneous NDS, the observability gramian can be written as
Yo = L(G)

0
eA t (Cy )T Cy eAt dt = L(G) Yo ,
(31)
where Yo is the gramian for an individual agent in the NDS.

Theorem 9.3. The homogeneous NDS coupled at the output in (27) is unobservable.
Proof. Using the gramian expression in (31) and Theorem 9.1 we conclude that Yo
has precisely n eigenvalues at the origin, leading to an unobservable system.

The unobservable modes of (27), in fact, correspond to the inertial position of the
entire formation; these modes lie in the subspace span{1 I}. The importance of
this result is that when each agent has identical dynamics, relative measurements
alone are insufficient to reconstruct their inertial states. If in addition to the relative
output, an additional inertial measurement is available, say one that corresponds to
the inertial position of a single agent, then the observability of the system can be
recovered.
An interesting consequence of this result highlights how the underlying connection topology influences the relative degree of observability of the observable
modes. We denote and index each singular value of Yo as i , and using the results
of Theorem 9.1 we can express the non-zero singular values of Yo as j (G)i for
j = 2, . . . , n and all i. The eigenvalues of the graph Laplacian, therefore, can amplify or attenuate the relative degree of observability of the system. For example,
the complete graph as in Fig. 1(a), has i (G) = j (G) = n for i, j 2. In this case,
the connection topology does not favor any particular modes of the system as each
is scaled by the same amount. Conversely, when the graph is disconnected with two
connected components, then 2 (G) = 0 and n additional unobservable modes are
introduced into the system.
In the heterogeneous case, the observability gramian of (27) has a non-trivial
form. We define the observability operator for an individual agent as i (x) = Ciy eAit x,

T
and its adjoint as i (y(t)) = 0 eAi t (Ciy )T y(t)dt [8]. The observability gramian of
(27) can be written as
Yo = diag{ }(L(G) I)diag{ } = (L(G) I) ,

where = 1 n . The derivation of (32) can be found in [37].
(32)
231
Theorem 9.4. The heterogeneous NDS coupled at the output in (27) is unobservable
if and only if the following conditions are met:
1. There exists an eigenvalue, , of A that is common to each Ai , and
2. One has Ciy qi = Cyj q j for all i, j with Ai qi = qi for all i.
Proof. The necessary condition is verified when all agents have identical dynamics, as shown in Theorem 9.3. For the sufficient condition, assume that there exists
that is an eigenvalue for each Ai . We can then construct an eigenvector for A
T

as q = qT1 qTn , with Ai qi = qi . By condition 2, we have that Cq = 1 r,
where r = Ci qi = 0 for all i. Using properties of the Kronecker product we then
have
(E(G)T I)Cy q = (E(G)T I)(1 r) = (E(G)T 1 r) = 0.
(33)
This shows the system is unobservable with q the corresponding unobservable

mode.

Theorem 9.4 shows that a heterogeneous NDS becomes unobservable only when
the outputs of each agent associated with a certain initial condition direction becomes indistinguishable. For general heterogeneous NDS therefore the system
is expected to be observable. This is a rather non-trivial, as it suggests that the inertial position of each agent can be reconstructed solely from relative measurements.
As in the homogeneous case, the underlying connection topology can have a profound affect on the relative degree of observability of the system. The form of (32)
is appealing in how it separates the role of the network from each agent. Although
the precise characterization of the eigenvalues of (32) is non-trivial, bounds on
those values can be derived, as presented in [15]. In particular, since both terms
in the Hadamard product are positive semi-definite matrices, we can apply Schurs
Theorem to obtain the bound,
d ( ) (Yo ) (Yo ) d ( ),
(34)
where (Yo ) and (Yo ) correspond, respectively, to the smallest and largest singular values of Yo , and
d = min [L(G) Jn ]ii , d = max [L(G) Jn ]ii ;
i
the quantities d and d correspond, respectively, to the minimum and maximum degree vertices of the underlying graph. We note that the bounds (34) become tight
when agent have homogeneous dynamics. Such observations point to interesting
connections between the degree of each agent in the ensemble and the relative observability of the modes of the system. This theme will be revisited when we study
the H2 performance of heterogeneous NDS coupled at the output.
232
Controllability of NDS Coupled at the Input

In this section, we consider whether from any initial condition, each agent can be
driven to an arbitrary final state when the input is distributed to each agent through
a network. For homogeneous NDS, the controllability gramian can be written by
inspection as
Xc = L(G)

0
eAt BBT eA t dt = L(G) Xc ,
(35)
where Xc is the controllability gramian for an individual agent in the NDS.

Theorem 9.5. The homogeneous NDS coupled at the input (28) is uncontrollable.
Proof. The gramian (35) has precisely n eigenvalues at the origin, leading to an
uncontrollable system.

Here, the dual nature of the NDS coupled at the output and input becomes immediately apparent. As in the former case, the uncontrollable mode corresponds to the
inertial position of the entire ensemble, lying in the subspace span{1 I}. We also
note that the relative degree of controllability can be inferred from the gramian,
but this analysis is omitted as it mirrors that of the observability analysis for NDS
coupled at the output.
For the heterogeneous case, we arrive at the following result.
Theorem 9.6. The heterogeneous NDS coupled at the input (28) is uncontrollable if
and only if the following conditions are met:
1. There exists an eigenvalue, , of A that is common to each Ai , and
2. One has qTi Bi = qTj B j for all i, j with qTi Ai = qTi for all i.
The proof of Theorem 9.6 follows the same procedure as for Theorem 9.5, and is
omitted. The conclusion, as expected, shows that heterogeneity in the dynamics of
each agent can lead to a fully controllable system.
The dual structure between the NDS coupled at the input and coupled at the
output should now be clear. A simple exercise will show that the controllability
gramian for (28) has a similar form to the observability gramian (32), with the role of
replaced with the controllability operator. It becomes apparent that the degree of
each agent in the ensemble can have a profound affect on the overall controllability
properties of the system.
3.2 Graph-Theoretic Bounds on NDS Performance

In this section we explore a graph-theoretic characterization of the H2 performance
of different NDS models. The main goal is to again make explicit the role of the
underlying connection topology on the system performance norms. We will assume
throughout this section that the underlying connection graph G is connected. For
analysis, we also assume each agent is stable.
233
H2 Performance of NDS Coupled at the Output

The sensed output of an NDS coupled at the output can be used to achieve a variety of objectives, including localization- for which the observability results of 3.1
apply- and formation control. It is important therefore to examine how noise entering the dynamics of each agent in the NDS propagates through the network to
the sensed output. A natural measure for quantifying this property is the H2 system
norm. This section, therefore, aims to explicitly characterize the affect of the network on the H2 norm of the system. For this analysis, we assume that each agent
is driven by a Gaussian white noise. A simplified version of the agent dynamics is
given as

xi (t) = Ai xi (t) + i wi (t)
i :
(36)
yi (t) = Ciy xi (t).
The corresponding model for the NDS coupled at the output takes the form
Ax(t)
+ w(t)
x (t) =
y(t) =
Cy x(t)
het (G)
yG (t) = (E(G)T I)Cy x(t).
(37)
The H2 -norm of a system can be calculated using the controllability and observability gramians of the system, as discussed in 3.1. The H2 norm of each agent from
the exogenous input channel to the measured output can be expressed in terms of
the gramians as
/
(38)
i 2 = trace(i T Yoii )
/

y
(39)
= trace Ci Xci (Ciy )T ,
where Yoi and Xci denote, respectively, the observability gramian and controllability
gramian of agent i.
Theorem 9.7. The H2 norm of the homogeneous NDS coupled at the output (37) is
given by

wG
(40)
Thom = E(G)F 2 .
2
Proof. The H2 norm can be written directly from (31) as

/

wG
Thom = trace((In )T (L(G) Yo )(In )).
2
Using the properties of the Kronecker product

. defined in 1.1 and the definition of
the Frobenius norm of a matrix, MF = trace(MT M), leads to the expression in
(40).

The expression in (40) gives an explicit characterization of how the network affects the the system performance. For homogeneous systems, we find that the H2
234
performance changes with the addition or removal of an edge. Recall that the Frobenius norm of a matrix can be expressed in terms of the 2-norm of each column, as

MF =
1/2
mi 22
i=1
where mi is the ith column of the matrix M. As each column of E(G) represents an
edge in G, the Frobenius norm can be expressed in terms of the number of edges in
the graph, |E|, as
E(G)F = (2 |E|)1/2 .
(41)
This highlights the importance of the number of edges as opposed to the actual structure of the graph (e.g., a star graph or k-regular graph). This makes intuitive sense,
as more edges would correspond to additional amplification of the disturbances entering the system.
If we consider only connected graphs, we arrive at the following corollaries providing lower and upper bounds on the H2 norm of the system.
Corollary 9.1. The H2 norm of the homogeneous NDS coupled at the output (37)
for an arbitrary connected graph G is lower bounded by an NDS where G is a spanning tree, as

wG 2
(42)
Thom 2 22 (n 1);
2
the lower bound is attained with equality whenever the underlying graph is a spanning tree.
It is clear from the definition of the Frobenius norm that the choice of tree is irrelevant (e.g., a star or a path).
Corollary 9.2. The H2 norm of the homogeneous NDS coupled at the output (37)
for an arbitrary connected graph G is upper bounded by an NDS where G = Kn , the
complete graph, as

wG 2
(43)
Thom 2 22 n (n 1);
2
the upper bound is attained with equality whenever the underlying graph is
complete.
For the heterogeneous case we rely on (39) to derive the H2 norm. The connection topology only couples agents at the output leading to a block diagonal description for the controllability gramian, with each block corresponding to each agents
controllability gramian.
Theorem 9.8. The H2 norm of the heterogeneous NDS coupled at the output (37) is
given as

1/2

wG
2
,
(44)
Thet = di i 2
2
where di is the degree of the ith agent in the graph.
Proof. The norm expression in (44) can be derived using (39) as,

1
21/2
wG
,
Thet = trace (E(G)T I)Cy Xc (Cy )T (E(G) I)
2
235
(45)
where Xc denotes the block diagonal aggregation of each agents controllability

gramian. First, we make the following observation,
2
1
trace Cy Xc (Cy )T =
i 22 .
i=1
Using the cyclic property of the trace operator [41] and exploiting the block
diagonal structure of the argument leads to the following identity simplification,
1
2
1
2
trace Cy Xc (Cy )T (( (G) A(G))I) = trace Ciy Xci (Ciy )T (di I)
i
= di i 22 .
(46)
This leads to the desired result.

A further examination of (44) reveals that it can be written as the Frobenius norm
of a node-weighted incidence matrix,

1 2

wG
.
..
(47)
Thet =
E(G) .

2

n 2
F
When each agent has the same dynamics, (47) reduces to the expression in (40).
This characterization paints a clear picture of how the placement of an agent within
a certain topology affects the overall system gain. In order to minimize the gain, it
is beneficial to keep systems with high norm in locations with minimum degree.
For certain graph structures, a more explicit characterization of the H2 performance can be derived, leading to the following corollaries.
Corollary 9.3. The H2 norm of the heterogeneous NDS coupled at the output (37)
when the underlying connection graph is k-regular is

wG
Thet =
2
1/2

k
i 22
(48)
where every node has degree k.

Note that having regularity in the connection topology introduces homogeneity
into the heterogeneous NDS. As in the homogeneous case, the placement of an agent
in the network will not affect the overall performance. The system norm for this
topology becomes a scaled version of the parallel connection of the n sub-systems.
236
H2 Performance of NDS Coupled at the State

The consensus model derived in Sect. 2 is an important model for an increasingly
wide range of applications. In the literature there are many variations of the model
(26) including its nonlinear extensions and those evolving over random and switching graph topologies. The common theme in all these scenarios is the focus on the
convergence of the systems state to the agreement subspace when arbitrarily initialized.
In this section, we like to consider a more system-theoretic approach to the analysis of consensus seeking systems. To this end, we will work with a consensus model
that is corrupted by noise at the process and the corresponding measurements. We
will then characterize the H2 performance of the model, which can be used to reason about how noise in the network result in the asymptotic deviation of each nodes
state from the consensus state.
For the derivation of the consensus model with noise we highlight again the relationship of the different type of NDS models derived in Sect. 2. The consensus
model with noise can be considered as an NDS with coupling at both the input and
and output with a unity feedback applied to close the loop, as shown in Fig. 6.
For this model, we will combine the NDS coupled at the input and output for
homogeneous agent dynamics, each described by a single integrator dynamics. The
process noise entering each agent, wi (t), is assumed to be a Gaussian white noise
with covariance
E[wi (t)wi (t)T ] = w2 I
and E[wi (t)w j (t)T ] = 0 when i = j.
The NDS measurement, which is the sensed relative measurement between neighboring agents, is also corrupted by a Gaussian white noise v(t) with covariance
E[v(t)v(t)T ] = v2 I. The open-loop model, therefore, can be written as

x (t) = E(G)uG (t) + w(t)
.
(49)
yG (t) = E(G)T x(t) + v(t)
When the output-feedback control uG (t) = yG (t) is applied, we obtain a closedloop consensus system driven by noise. To complete the input-output description of
the system, we recall that in consensus seeking systems, the objective is for each
v(t)
w(t)
uG(t)
z(t)
E(G)
E(G)T
Fig. 6 Consensus as an NDS with coupling at the input and output
yG(t)
237
state to agree on a common value. Therefore, we include the performance variable

z(t) that captures the difference between states.
!

w(t)
x (t) = L(G)x(t) + I E(G)

v(t) .
(G) :
(50)
z(t) = E(G)T x(t)

Note that in the absence of noise, the system is driven by its initial conditions and the
original consensus model derived in 2 is recovered. More interestingly, an input
output description of the consensus model can be considered as an NDS with coupling at the input, output, and state.
The final hurdle to overcome in an H2 performance analysis of (50) is that the
system state matrix, L(G), has an eigenvalue at the origin, which will lead to
an unbounded H2 norm. The observability and controllability results of Sect. 3.1,
however, can immediately be applied to conclude that (50) is neither controllable nor observable. The uncontrollable and unobservable modes are precisely the
modes that lie in the agreement subspace, span{1}. Therefore, we can consider
our analysis on a minimal representation of (50) instead. Theorem 9.2 can be applied as a coordinate transformation on (50) that has the desired feature of separating the uncontrollable and unobservable modes while maintaining an algebraic
representation of the underlying connection topology via the edge Laplacian (see
Sect. 1.1). In this direction, we introduce the coordinate transformation Sv xe (t) =
x(t), where Sv is defined in (12). The minimal system representation can therefore be
expressed as
x (t) = Le (G )R(G)R(G)T x (t) +w E(G )T w(t)

v Le (G )R(G)v(t)
:
(51)
z(t) =
R(G)T x (t);
the signals w(t)
and v(t)
are the normalized process and measurement noise signals. We also note that x (t) corresponds to the states on the edges of the spanning
tree sub-graph G . The performance variable, z(t), contains information on the tree
states in addition to the cycle states. Here we recall that the cycle states are a linear
combination of the tree states and we note that z(t) actually contains redundant information. This is highlighted by recognizing that the tree states converging to the
origin forces the cycle states to do the same. Consequently, we will consider the
system with cycles as well as a system containing only the tree states at the output,
which we denote as ,
x (t) = Le (G )R(G)R(G)T x (t) +w E(G )T w(t)

v Le (G )R(G)v(t)
:
(52)
z(t) =
x (t).
This distinction will subsequently be employed to quantify the effect of cycles on
the system performance.
238
The H2 norm of and can be calculated using the controllability gramian as,
22 = trace[RT X R],
and 22 = trace[X ],
(53)
where R is defined in (8) and X is the positive-definite solution to the Lyapunov

equation
Le (G )RRT X XRRT Le (G ) + w2 Le (G ) + v2 Le (G )RRT Le (G ) = 0.
(54)
The structure of (54) suggests that the solution will be dependent on certain properties of the graph. In fact, the solution can found by inspection by first noting that

w2 Le (G ) + v2 Le (G )RRT Le (G ) = Le (G ) w2 (Le (G ))1 + v2 RRT Le (G ).
The solution to (54) is therefore
X =

1 2
w (RRT )1 + v2 Le (G ) ,
2
(55)
and we arrive at the following result.

Theorem 9.9. The H2 norm of the system (51) is
22 =
w2
(n 1) + v2|E|.
2
(56)
On the other hand, the H2 norm of the system (52) is

22 =
w2
trace[(RRT )1 ] + v2 (n 1).
2
(57)
Proof. The proof follows from (55) and noting that trace[Le (G )] = 2(n 1), or
twice the number of edges in a spanning tree.

We observe that 22 is a linear function of the number of edges in the graph.
This has a clear practical relevance, as it indicates that the addition of each edge
corresponds to an amplification of the noise in the consensus-type network. Let us
consider the implications of the graph-theoretic characterization of the H2 norm for
two classes of graphs.
(a) Spanning Trees: The first case resulting in a simplification of (56) arises when
G is a spanning tree. In this case R = I and (57) simplifies to
2

w
2
2
2 = (n 1)
(58)
+ v .
2
A direct consequence of this result is that all spanning trees result in the same
H2 system performance. That is, the choice of spanning tree (e.g., a path or
239
a star) does not affect this performance metric. As expected, in this scenario
22 = 22 .
(b) k-Regular Graphs: Regular graphs also lead to a simplification of (57). In
general, any connected k-regular graph will contain cycles resulting in a nontrivial expression for matrix product RRT . The H2 norm is therefore intimately
related to the cut space of the graph.
Denote the eigenvalues of RRT by i and note that

trace (RRT )1 =
n1
n1 n1
i = (G) j ,
i=1
(59)
i=1 j=i
where (G) is the number of spanning trees in G. The quantity n1

j=i j is recognized as a first minor of the matrix RRT .
Corollary 9.4. The cycle graph Cn has n spanning trees and hence

(n 1)2
trace (R(Cn )R(Cn )T )1 =
.
n
(60)
Thereby, the H2 norm of the system when the underlying graph is the cycle
graph Cn is given as
2

w (n 1)
22 = (n 1)
+ v2 .
(61)
n
Proof. Without loss of generality, we consider a directed path graph on n nodes,
with initial node v1 and terminal node vn as the spanning tree subgraph G .
Index the edges as ei = (vi , vi+1 ). The cycle graph is formed by adding the edge
T
en = (vn , v1 ). For this graph, we have Tc =
1n1 and R(Cn )R(Cn ) = I + J.
T
From this it follows that det R(Cn )R(Cn ) = n and all its first minors have
value n 1. Combined with (56) yields the desired result.

Corollary 9.5. The complete graph Kn has nn2 spanning trees, and therefore
3
1 4 2(n 1)nn3 2(n 1)
trace R(Kn )R(Kn )T
=
.
=
nn2
n
(62)
Thereby, the H2 norm of the system when the underlying graph is the complete graph Kn is given as
2

w
2
2
2 = (n 1)
(63)
+ v .
n
Proof. Without loss of generality, we consider a star graph with center at node
v1 and all edges are of the form ek = (v1 , vk+1 ). Then the cycles in the graph are
created by adding the edges e = (vi , v j ), i, j = 1 and R(Kn )R(Kn )T = nI J. It
then follows that
240

5
4.95
trace[(RRT)1]
4.9
4.85
4.8
4.75
4.7
50
100
150
200
250
300
350
400
450
500
Fig. 7 trace[(R(G)R(G)T )1 ] for random 5-regular graphs

det R(Kn )R(Kn )T = nn2
and all the first minors have value 2nn3 . Combined with (56) yields the desired
result.

3
1 4
Figure 7 depicts the sorted values of trace RRT
for 500 randomly generated regular graphs of degree five. As this figure shows, although the degree
of each node remains constant, the actual cycle structure of each graph instance
varies, effecting the resulting H2 norm of the corresponding consensus-type
input-output system.
Using the above analysis, we now proceed to characterize how the cycle structure of the graph effects the H2 performance for the corresponding consensus-type
system. In fact, examining the ratio
(G)22
(G )22
provides an indication of how the cycles increase the H2 norm; recall that G is
in general a graph containing cycles and G G is the spanning tree subgraph.
For example, consider the cycle graph Cn and assume unit covariance for both the
process and measurement noise. Then, as the number of nodes increase, the ratio of
the two H2 norms behaves as
(Cn )22
4n 2 4
=
= lim
n (Pn )2
n 3n
3
2
lim
241
(64)
indicating that for large cycles, the H2 performance is a constant multiple of the H2
performance for the path graph Pn .
In the meantime, for the complete graph Kn we have
n+1
(Kn )22
;
=
2
3
(G )2
(65)
in this case, we see that the norm is amplified linearly as a function of the number of
vertices in the graph. It is worth mentioning here that typical performance measures
for consensus problems, such as 2 (G), would favor the complete graph over the
cycle graph. However, in terms of the H2 performance, we see that there is a penalty
to be paid for faster convergence offered by the complete graph due to its cycle
structure.
Alternatively, insight is also gained by considering the ratio
(G)22
,
(G)22
which highlights the effects of including cycles in the performance variable z(t). For
the cycle graph we have
(Cn )22
n(3n 1)
3
= lim
= ,
n (Cn )2
n 2(n 1)(2n 1)
4
2
lim
(66)
suggesting that the effect of including the cycle for performance does not vary significantly with the size of the graph.
For the complete graph, on the other hand, one has
(Kn )22
n
= ,
2
2
(Kn )2
(67)
suggesting that the inclusion of cycles results in H2 performance that increases

linearly as a function of vertices in the graph.
4 Topology Design for NDS

The analysis results of 3 points to the importance of the underlying interconnection
topology on the overall performance of NDS. In fact, these results can be used to
motivate network synthesis problems for NDS. For multi-agent systems, in addition
to designing control and estimation algorithms for each agent in the ensemble, it
should be a primary objective of the designer to also consider what the underlying
242
connection topology should look like. In its most general form, therefore, we would
like to consider the problem of finding the underlying connection topology G, such
that the resulting NDS, (G), achieves a specified performance. This problem can
be stated formally as
min (G) p
G
(68)
s.t. G is connected.
The challenges associated with this problem is to find numerically tractable algorithms for (68). The difficulty arises from the combinatorial nature of (68); the decision to allow agent i and j to be connected is binary.
In this section we present a solution to (68) for NDS coupled at the output for the
H2 norm. Our results point to an intriguing connection between results in combinatorial optimization and systems theory. We also present a variation of (68) related to
sensor placement for NDS coupled at the state.
4.1 H2 Topology Design for NDS Coupled at the Output

We now present a polynomial time algorithm for the design of the interconnection topology for heterogeneous agent dynamics in an NDS that is coupled at the
output. Recall from Sect. 3.2 that in terms of the H2 norm objective, an optimal
topology should always correspond to a spanning tree. Hence, the design problem
is to determine which spanning tree achieves the smallest H2 norm for the NDS.
Note that the design of the topology reduces to the design of the incidence matrix,
E(G). This problem is combinatorial in nature, as there are only a finite number of
graphs that can be constructed from a set of n vertices. As the number of agents in
the NDS grows, solving this problem becomes prohibitively hard [16]. However, as
we will shortly show, with an appropriate modification of the problem statement,
a celebrated algorithm in combinatorial optimization can be used for solving the
topology design problem in polynomial time. Specifically, we will show that the
minimum spanning tree (MST) problem captures the essential features of this problem. The MST can be solved using Kruskals algorithm in O(|E|log(|V|)) time. The
algorithm is given below and a proof of its correctness can be found, for example,
in [16].
In order to apply the MST to the H2 synthesis problem we must reformulate the
original problem statement. To begin, we first write the expression for the H2 norm
of the system in (37) as

n
wG 2
Thet = di trace{Cy Xi (Cy )T }
2
i
n
di Tiwy 22,
i
(69)
243
Data: A connected undirected graph G(V, E) and weights w : E R.

Result: A spanning tree Gt of minimum weight.
begin
Sort the edges such that w(e1 ) w(e2 ) w(e|E | ), where ei E
Set Gt := Gt (V, 0)
/
for i := 1 to |E| do
if Gt + ei contains no cycle then
set Gt := Gt + ei
end
end
end
Algorithm 1: Kruskals Algorithm
wy
where Ti
is the map from the exogenous input entering agent i to its position,
Cy xi (t). We reiterate here that the NDS norm description is related to the degree of
each node in the network. Using the weighted incidence graph interpretation of the
norm, as in (47), we see that the gain of each agent, Tiwy 22 , acts as a weight on the
nodes. As each agent is assumed to have fixed dynamics, the problem of minimizing the NDS H2 norm reduces to finding the degree of each agent while ensuring
the resulting topology is a spanning tree. This objective is related to properties of
the nodes of the graph. In order to use the MST framework, we must convert the
objective from weights on the nodes to weights on the edges.
To develop this transformation, consider the graph G = (V, E) with fixed weights
wi on each node i = 1, . . . , n. The node-weighted Frobenius norm of the incidence
matrix is then
W E(G)2F =
di w2i ,
(70)
where W = diag(w1 , . . . , wn ).
Next, consider the effect of adding an edge e = (i, j) to E in terms of the Frobenius norm of the augmented incidence matrix,

2
W E(G) e = dk w2 + w2i + w2j ,
(71)
F
where dk represents the degree of node k before adding the new edge e.
This shows
that each edge e = (i, j) contributes (w2i + w2j ) to the overall norm. Therefore,
weights on the edges can be constructed by adding the node weights corresponding to the nodes adjacent to each edge as
we = |E(G)T |w2n .
(72)
Using the above transformation from node weights to edge weights, we arrive at
the following result.
Theorem 9.10. The connection topology that minimizes the H2 norm of (37), can
be found using Kruskals MST algorithm with input data G, and weights
244
T1wy 2
..
w = |E(G)|T
.
.
(73)
Tnwy 22
Proof. The proof follows from (69) and the transformation from node weights to
edge weights described in (70)(72).

Remark 9.2. The choice of the input graph G may be application specific, and can
capture certain communication or sensing constraints between agents. For example, one may consider a scenario where agents are initially randomly distributed (a
geometric random graph) upon deployment and can only sense neighboring agents
within a specified range. The results of Theorem 9.10 can be used to determine the
optimal spanning tree for that initial configuration.
Remark 9.3. There are a number of distributed algorithms that solve the MST problem [3,11]. These could be used in place of the centralized version when the optimal
spanning tree topology needs to be reconfigured. This scenario can arise due to the
initialization problem discussed in Remark 9.2, or in situations when certain agents
are disabled, lost, or reallocated for different mission purposes.
If there are no initial constraints on the input graph for Theorem 9.10, then we
arrive at the following result.
Corollary 9.6. When the input graph in Theorem 9.10 is the complete graph, then
the star graph with center node corresponding to the agent with minimum norm is
the (non-unique) optimal topology.
Proof. The degree of the center node in a star graph is n 1, and all other nodes
have degree one. Assume the node weights are sorted as w1 wn , then the H2
wG 2
norm of the RSN is Thet
2 = (n 1)w1 + ni=2 wi . Any other tree can be obtained
by removing and adding a single edge, while ensuring connectivity. With each such
operation, the cost is non-decreasing, as any new edge will increase the degree of
node i > 1 and by assumption w1 wi .

Corollary 9.6 shows that if there are no restrictions on the initial configuration, the optimal topology can be obtained without the MST algorithm. The computational effort required is only to determine the agent with smallest norm. The
non-uniqueness of the star graph can occur if certain agents have identical norm,
resulting in other possible configuration with an equivalent overall cost.
4.2 Sensor Placement with H2 Performance

for NDS Coupled at the State
The inputoutput description of consensus-type systems derived in 3.2 highlight
how noise entering the system can affect the performance. A common challenge in
245
the design of engineered systems is to minimize the overall cost while achieving the
best performance. In this direction, we can formulate a variation of the topology design (68) problem that focuses on choosing sensors for a consensus-seeking system
that aims to minimize both the cost of each sensor and the H2 performance of the
system.
In this direction, consider a modification of the system in (51) in the form

Le (G )T R v(t)
x(t)
= Le (G )RRT x (t) + w E(G )T w(t)
:
(74)
T
z(t) =
R x (t),
where w(t)
and v(t)
are the normalized noise signals, and the matrix is a diagonal
matrix with elements i corresponding to the variance of the sensor on edge i. We
note that the most general version of this problem considers a finite set of p sensors
each with an associated variance,
P = {12 , 22 , . . . , p2 },
(75)
where for each element i2 P there is an associated cost c(i2 ). The cost function has the property that c(i2 ) > c( 2j ) if i2 < 2j . Using (53)(54), in order to
find the optimal placement of these sensors, one can consider the mixed-integer
program [16],
P1 :
|E|
min trace[RT XR] + c(wi )

X,W
i=1
s.t. W = diag{w1 , . . . , w|E| }, wi P,
wi ,
i
Le (G )T RRT X XRRT Le (G )T + w2 Le (G ) + Le (G )T RW RT Le (G ) = 0,
where represents a weighting on the H2 performance of the solution, and represents the maximum aggregated noise covariance. Note that in general |E| mini i2
|E| maxi i2 .
The problem P1 is combinatorial in nature, as a binary decision needs to be made
as to which sensor to use and place in the network. Although P1 can certainly be
solved by using a mixed-integer programming solvers [16], certain relaxations can
be made to convexify the resulting problem. Most notably, one approach involves
relaxing the discrete nature of the set P (75) into a box-type constraint as

(76)
P = 2 , 2 .
The cost function can now be written as a continuous map c : P R which is convex
and a strictly decreasing function. The simplest version of such a function would be
the linear map
c(i2 ) = i2
246

10
0.0781
0.0456
8
3
0.0431
0.0359
1
0.0328
0.0370
v = 0.0409
9
0.0431
0.0409
0.0583
6
0.0453
5
Fig. 8 A graph on ten nodes with optimal sensor selection; v denotes the sensor variance
for some > 0. This relaxation leads to the following modified program,
P2 :
min trace[RT XR] trace[W ]

X,W
s.t. W = diag{w1 , . . . , w|E| }, 2 wi 2 ,

Le (G ) RR X XRR Le (G )
T
wi ,
i
2
T
+ w Le (G ) + Le (G ) RW RT Le (G )
= 0.
As an example of the applicability of P2 , we considered the sensor selection for

the graph in Fig. 8. A random graph on ten nodes with an edge probability of 0.15
was generated. The resulting graph is connected and contains two independent cycles, resulting in a more general problem instance. The sensor constraints were
P = [0.001 0.1] and 2 = 0.501. Finally, the cost function weights were chosen
as = 5 and = 1.
Solving P2 then resulted in a non-trivial selection of sensors for each edge. The
sensor covariance for each edge is labeled in Figure 8; we observe that the highest
fidelity sensors tend to be concentrated around the node of highest degree. Also, the
edge with the lowest fidelity sensor is placed in low traffic areas.
The complexity of large-scale systems requires a systematic approach for their analysis and synthesis that blends constructs from system theory on one hand, and graph
theory on the other. This chapter presents a viable framework for studying these systems that highlights their structural properties.
247
By defining four canonical models of networked dynamic systems, we proceeded

to explicitly describe how interconnections effect some of the system-theoretic properties of the overall system. Furthermore, the generality of the models developed
allows for a further specification of system complexity in the realm of the notions
of homogeneity and heterogeneity.
One of the main themes of this work revolved around a characterization of the
H2 performance of NDS for both analysis and synthesis. The performance of these
systems was shown to be intimately related to the number of edges in the network,
and the degree of each agent in the ensemble. A natural extension of this work is
to consider other system norms for analysis and synthesis, such as H norm of
NDS, that is the subject of forthcoming work by the authors. Perhaps a more subtle
point of this chapter is the relationship between the different NDS models. Specifically, we noted that via appropriate transformations and the inclusion of structured
decentralized control laws, distinct types of NDS models can be transformed to
one another. This suggests, for example, that certain NDS models might be more
advantageous to use than others when considering analysis and synthesis for a particular networked system. This last comment also relates to the necessity of finding
tractable algorithms for the synthesis of NDS. The results of 4 highlighted a convenient connection between results in combinatorial optimization and systems theory.
While the combinatorial structure of NDS may seem prohibitive at first glance, a
complete understanding of how certain properties of the graph relate to the overall
performance can lead to useful relaxations and efficient algorithms for synthesis of
networks that support the operation of distributed dynamic systems.
Acknowledgement This work was supported by NSF grant ECS-0501606.
References
1. Autonomous Nanotechnology Swarm, http://ants.gsfc.nasa.gov
2. I. Akyildiz, W. Su, Y. Sankarasubramnian, and E. Cayirci, A survey on sensor networks, IEEE
Communications Magazine, 102114, August 2002
3. B. Awerbuch, Optimal distributed algorithms for minimum weight spanning tree, counting,
leader election, and related problems, Proceedings of the 19th Annual ACM Symposium on
Theory of Computing, New York, New York, 1987
4. P. Barooah, N. Machado da Silva, and J. P. Hespanha, Distributed optimal estimation from
relative measurements for localization and time synchronization, Proceedings of Distributed
Computing in Sensor Systems, 2006
5. A. Behar, J. Matthews, F. Carsey, and J. Jones, NASA/JPL Tumbleweed Polar Rover, IEEE
Aerospace Conference, Big Sky, MT, March 2004
6. D. P. Bertsekas and J. N. Tsitsiklis, Parallel and Distributed Computation, Prentice-Hall,
Englewood Cliffs, NJ, 1989
7. K. Danzmann and LISA Study Team, LISA: Laser Interferometer Space Antenna for the
Detection and Observation of Gravitational Waves, Max-Planck Institute fur Quantenoptik,
Garching, Germany, 1998
8. G. E. Dullerud and F. Paganini, A Course in Robust Control Theory: A Convex Approach,
Springer, New York, 2000
248
9. J. A. Fax and R. M. Murray, Information flow and cooperative control of vehicle formations,
IEEE Transactions on Automatic Control, (49) 9:14651476, 2004
10. C. V. M. Fridlund, Darwin-the infrared space interferometry mission, ESA Bulletin, (103):20
25, August 2000
11. R. G. Gallager, P. A. Humblet, and P. M. Spira, A distributed algorithm for minimum-weight
spanning trees, ACM TOPLAS, (5) 1:6677, 1983
12. C. Godsil and G. Royle, Algebraic Graph Theory, Springer, Berlin, 2001
13. Y. Hatano and M. Mesbahi, Agreement over random networks, IEEE Transactions on Automatic Control, (50) 11:18671872, 2005
14. R. A. Horn and C. R. Johnson, Topics in Matrix Analysis, Cambridge University Press,
Cambridge, 1991
15. C. R. Johnson, Matrix Theory and Applications, Proceedings of Symposia in Applied Mathematics (40), 1990
16. B. Korte and J. Vygen, Combinatorial Optimization: Theory and Algorithms, Springer, Berlin,
2000
17. P. R. Lawarence, The Terrestrial planet finder, Proceedings of the IEEE Aerospace Conference, March 2001
18. P. Lin, Y. Jia, and L. Li, Distributed robust H consensus control in directed networks of
agents with time-delaty, Systems and Control Letters, (57):643653, 2007
19. M. Mesbahi and F. Y. Hadaegh, Formation flying control of multiple spacecraft via graphs,
matrix inequalities, and switching, AIAA Journal of Guidance and Control, (24) 2:369377,
2001
20. M. Mesbahi and M. Egerstedt, Graph-theoretic Methods in Multi-agent Networks, Princeton
University Press, 2010
21. R. Olfati-Saber, Distributed Kalman filtering for sensor networks, IEEE Conference on
Decision and Control, December 2007
22. R. Olfati-Saber and R. M. Murray, Consensus problems in networks of agents with switching
topology and time-delays, IEEE Transactions on Automatic Control, (49) 9:15201533, 2004
23. R. Olfati-Saber, J. A. Fax, and R. M. Murray, Consensus and cooperation in networked multiagent systems, Proceedings of the IEEE, (95) 1:215233, 2007
24. G. Purcell, D. Kuang, S. Lichten, S. C. Wu, and L. Young, Autonomous formation flyer
(AFF) sensor technoloyg development, American Astronautical Society Guidance and Control Conference, February 1998
25. A. Rahmani, M. Ji, M. Mesbahi, and M. Egerstedt, Controllability of multi-agent systems
from a graph theoretic perspective, SIAM Journal of Control and Optimization, 48 (1):162
186, 2009
26. W. Ren and R. W. Beard, Distributed Consensus in Multi-vehicle Cooperative Control,
Springer, London, 2008
27. W. Ren, R. W. Beard, and E. M. Atkins, A survey of consensus problems in multi-agent
coordination, American Control Conference, June 2005
28. J. Sandhu, M. Mesbahi and T. Tsukamaki, Relative Sensing Networks: Observability, Estimation, and the Control Structure, IEEE Conference on Decision and Control, 2005
29. C. Scherer, P. Gahinet, and M. Chilali, Multiobjective output-feedback control vial LMI optimization, IEEE Transactions on Automatic Control, (42) 7:896911, 1997
30. G. Sholomitsky, O. Prilutsky, and V. Rodin, Infra-red space interferometer, International
Astronautical Federation, Paper IAF-77-68, 1977
31. R. S. Smith and F. Y. Hadaegh, Control of deep space formation flying spacecraft; relative
sensing and switched information, AIAA Journal of Guidance and Control, (28) 1:106114,
2005
32. A. Tahbaz-Salehi and A. Jadbabaie. Necessary and sufficient conditions for consensus over
random networks, IEEE Transactions on Automatic Control, (53) 3:791796, 2008
33. G. Walsh, H. Ye, and L. Bushnell, Stability analysis of networked control systems, American
Control Conference, June 1999
34. P. K. C. Wang and F. Y. Hadaegh, Coordination and control of multiple microspacecraft
moving in formation, Journal of Astronautical Sciences, 44:315355, 1996
249
35. B. Wie, Space Vehicle Dynamics and Control, American Institute of Aeronautics and Astronautics, 1998
36. L. Xiao, S. Boyd, and S. Lall, A scheme for robust distributed sensor fusion based on average
consensus, Fourth International Symposium on Information Processing in Sensor Networks,
April 2005
37. D. Zelazo and M. Mesbahi, On the observability properties of homogeneous and heterogeneous networked dynamic systems, IEEE Conference on Decision and Control, December
2008
38. D. Zelazo and M. Mesbahi, Edge agreement: Graph-theoretic performance bounds and passivity analysis, IEEE Transactions on Automatic Control 2010 (to appear)
39. D. Zelazo and M. Mesbahi, H2 Performance of relative sensing networks: Analysis and
synthesis, AIAA@Infotech Conference, April 2009
40. D. Zelazo, A. Rahmani, and M. Mesbahi, Agreement via the edge Laplacian, IEEE Conference on Decision and Control, December 2007
41. F. Zhang, Matrix Theory. Springer, Berlin, 1999
A Novel Coordination Strategy for Multi-Agent

Control Using Overlapping Subnetworks
with Application to Power Systems
R.R. Negenborn, G. Hug-Glanzmann, B. De Schutter, and G. Andersson
1 Introduction
Power networks [15, 16, 24] are one of the corner stones of our modern society. The
dynamics of a power network as a whole are the result of the interactions between
the millions of individual components. Conventionally, the power in power networks
is generated using several large power generators. This power is then transported
through the transmission and distribution network to the location where it is consumed, e.g., households and industry. Power flows are then relatively predictable,
and the number of control agents is relatively low. Due to the ongoing deregulation
in the power generation and distribution sector in the US and Europe, the number of players involved in the generation and distribution of power has increased
R.R. Negenborn
Delft Center for Systems and Control, Delft University of Technology, Mekelweg 2,
2628 CD Delft, The Netherlands
e-mail: r.r.negenborn@tudelft.nl
G. Hug-Glanzmann
Department of Electrical and Computer Engineering, Carnegie Mellon University,
5000 Forbes Avenue, Pittsburgh, Pennsylvania
e-mail: ghug@andrew.cmu.edu
B. De Schutter
Delft Center for Systems and Control, Delft University of Technology, Mekelweg 2,
2628 CD Delft, The Netherlands
and
Department of Marine and Transport Technology, Delft University of Technology,
Mekelweg 2, 2628 CD Delft, The Netherlands
e-mail: b@deschutter.info
G. Andersson
Power Systems Laboratory, ETH Zurich, Physikstrasse 3, 8092 Zurich, Switzerland
e-mail: andersson@eeh.ee.ethz.ch
251
252
R.R. Negenborn et al.
significantly. The number of source nodes of the power distribution network is

increasing even further as also large-scale industrial suppliers and small-scale individual households start to feed electricity into the network [13].
As a consequence, the structure of the power network is changing from a hierarchical top-down structure into a much more decentralized system with many
generating sources and distributing agencies. This causes that power flows become
less predictable and may actually change their conventional directions. To still guarantee basic requirements and service levels, such as voltage magnitude and frequency levels, bounds on deviations, stability, elimination of transients, etc., and to
meet the demands and requirements of the users, new infrastructure in the shape of
transmission lines and so-called Flexible Alternating Current Transmission Systems
(FACTS) [10] is installed. Transmission lines increase directly the capacity of the
network on the one hand. FACTS devices can be used to actively change the way in
which power flows over the network on the other hand. FACTS devices can change
voltage magnitudes, line impedances, and phase angles, and therefore have the potential to improve the security of the network, to increase the dynamic and transient
stability, to increase the quality of supply for sensitive industries, and to enable environmental benefits [10]. Two particular types of FACTS devices that frequently
appear in practice and that also will be used later on in this chapter are Static Var
Compensators (SVCs) and Thyristor Controlled Series Compensators (TCSCs) [5].
To optimally use and control such devices and to optimally use the existing
infrastructure, new control techniques have to be developed and implemented [18].
A major challenge in this context is that the devices in the network, such as the
various FACTS devices, are usually owned and operated by different authorities.
Despite this, the operators of the various devices have as objective to determine
their actions in such a way that the best overall network performance is obtained.
Hence, multi-agent control, in which communication and cooperation between various control authorities is explicitly taken into account, has to be employed.
1.1 Multi-Agent Control of Power Networks

The control structure of power networks can be represented as a multi-agent system
[18,2527], in which the control agents are organized in several layers as illustrated
in Fig. 1. A control agent hereby is an entity, e.g., a human, a computer, or a hardware device, that on the one hand observes the state or situation of the network and
on the other hand chooses actions to be taken in the network by changing settings
of actuators, such as the reference for the power output of generators or the reference for settings of FACTS devices. A control agent has to choose its actions in
such a way that the performance of the network in terms of safety, security, and stability, is the best possible, while respecting operational constraints and minimizing
costs. High costs hereby indicate a bad performance of the network, whereas low
costs indicate a good performance. In the control hierarchy that power networks
are controlled by, at the lower layers control agents consider faster dynamics, more
local information, smaller subnetworks, and shorter time spans. At the higher layers
A Novel Coordination Strategy for Control of Overlapping Subnetworks
253
control structure
control agent
optimizer
control agent
control agent
control agent
control agent
control agent
control agent
optimizer
optimizer
optimizer
optimizer
optimizer
optimizer
physical network
Fig. 1 Illustration of multi-layer control of large-scale networks (inspired by [18]). The control
structure consists of several layers of control agents. The control agents make measurements of the
state of the network and determine which actions to take
control agents consider slower dynamics, more global information, larger subnetworks, and longer time spans [20].
The control problem that an individual control agent in a control hierarchy faces
can be cast as an optimization problem, based on a local objective function that encodes the control goals of the agent, subject to a model of the part of the network
that the control agent controls, and additional constraints, e.g., on the range of the
inputs. The model of the part of the network that the control agent controls is referred to as its prediction model. This prediction model describes how the values of
variables of interest (such as voltage magnitudes, power flows, etc.) react to changes
in inputs and can therefore be used to predict what the effect of certain input choices
is going to be.
1.2 Control of Subnetworks

In a multi-agent system, control is distributed over several control agents. Each of
the control agents controls only its own part of the network, i.e., its own subnetwork. Let for now a network be modeled at an abstract level using a number of
nodes with arcs interconnecting the nodes. The nodes represent characteristics of the
254
components of the physical network, whereas the arcs model the direct interaction
between the nodes. E.g., one node could model the characteristics of a power generator together with a bus and a transmission line, and another node could model
the characteristics of a load and a bus. If the bus of this load is physically connected
to the transmission line, then an arc is defined between the nodes and . The subnetwork of a control agent then constitutes a number of nodes together with the arcs
connected to these nodes.1
Usually subnetworks are defined through geographical or institutional borders,
such as borders of cities, provinces, countries, the European Union, etc. Subnetworks can however also be defined differently, e.g., based on a fixed radius around
input nodes. Nodes that are reachable within a certain number of arcs from a particular node with an actuator are then included in a particular subnetwork [9]. Or,
subnetworks can be defined using an influence-based approach [8]. The idea of
influence-based subnetworks is that the subnetworks are defined based on the nodes
that a certain input and, hence, a control agent controlling that input, can influence.
Sensitivities are then used to determine which variables an input can influence, and
hence, which nodes should be considered part of a subnetwork. The fixed-radius and
the influence-based approaches have as advantage that the subnetworks are defined
taking a more actuator-centered perspective. Using the fixed-radius approach, this
definition is somewhat ad hoc and heuristic. On the contrary, the influence-based approach is more flexible and allows for a structured determination of the subnetwork
that a control agent has to consider.
When using the mentioned approaches for defining subnetworks, any pair of two
resulting subnetworks can be categorized as non-overlapping, touching, or overlapping, as illustrated in Fig. 2. If for two subnetworks, the nodes belonging to one of
A
1
5
2
5
1
(a) Non-overlapping subnetworks
(b) Touching subnetworks
5
1
(c) Overlapping subnetworks

Fig. 2 Illustration of different types of subnetworks
In Sect. 2 we define networks, nodes, arcs, and subnetworks more formally.
255
these do not coincide with the nodes belonging to the other subnetwork, and if there
are no arcs going from nodes in the one subnetwork into nodes of the other subnetwork, then the subnetworks are non-overlapping. If for two subnetworks, the nodes
belonging to one of these do not coincide with the nodes of the other subnetwork,
but if there are arcs between nodes of the one subnetwork and nodes of the other
subnetwork, then the subnetworks are touching. If for two subnetworks, the nodes
belonging to one of these partially coincide with the nodes belonging to the other
subnetwork, then the subnetworks are overlapping. In that case, a common subsubnetwork is defined consisting of those nodes and arcs that belong to both subnetworks.
If the subnetworks are non-overlapping, then the values of the variables of the
nodes that control agents can influence significantly do not overlap, so no coordination among control agents is necessary. In that case, adequate control performance
can be obtained, as illustrated in [8]. If the subnetworks are touching, coordination
can be obtained by adapting the technique of [3], as will be discussed in Sect. 3.
For subnetworks that are overlapping, no techniques have been proposed so far for
obtaining coordination. For overlapping subnetworks, the control agents will have
to find agreement on the values of variables involved in the characteristics of the
common sub-subnetworks. This topic is addressed in this chapter.
1.3 Optimal Power Flow Control

Optimal power flow control is a well known-method to optimize the operation of a
power network at higher control layers [15]. Optimal power flow control is typically
used to improve steady-state network security by improving the voltage profile, preventing lines from overloading, and minimizing active power losses. The optimal
power flow control problem is usually stated as an optimization problem in which
variables to be optimized consist of inputs or settings for generators, the objective
function encodes the control goals (such as maintaining voltage magnitudes within
desired bounds, preventing transmission lines from overloading, minimizing power
losses, etc.), and the prediction model consists of the steady-state characteristics of
the network.
To optimally make use of the FACTS devices installed in the power network we
employ optimal power flow control to determine the settings for these devices. As
mentioned, the devices in the network can be owned and controlled by different authorities. Traditional approaches for optimal power flow control in power networks
using multiple control agents assume that control agents consider at most touching, and thus not overlapping, subnetworks [14, 22]. In these cases subnetworks
are typically defined based on existing geographical borders of countries, states,
provinces, cities, etc. However, when the subnetworks are overlapping, the traditional approaches may not be suitable. Therefore, a new coordination approach for
control of overlapping subnetworks has to be developed. We have already made a
first step in this with the proposal of the approach described in [12], of which the
approach proposed in this chapter is a further elaboration and generalization.
256
1.4 Goal and Outline of This Chapter

In this chapter we propose a coordination scheme for control agents controlling
overlapping subnetworks with the aim of obtaining the best overall network performance. This chapter is organized as follows. In Sect. 2, we formalize the modeling
of networks, subnetworks, and control objectives used in this chapter. In Sect. 3,
we first discuss a recently proposed approach that can be used for the multi-agent
control of subnetworks that are not overlapping (i.e., non-overlapping or touching).
We then propose an extension of this approach to multi-agent control of subnetworks that are overlapping in Sect. 4. In Sect. 5, we apply the proposed approach
to an optimal power flow control problem from the domain of power networks. In
particular, we employ the approach to control FACTS devices in an adjusted IEEE
57-bus power network, in which each FACTS is controlled by a different control
agent. Section 6 contains conclusions and directions for future research.
2 Modeling of Network Characteristics and Control Objectives

In this section we formalize the way in which we describe the network characteristics, subnetworks, and control objectives in this chapter. An example of the application of this formalization is given in Sect. 5.
2.1 Network Characteristics

We consider the control of power networks by multiple control agents that operate
in a higher control layer. At this layer, we are interested in controlling the very slow
dynamics or the long-term behavior of the network, and therefore we can assume
that dynamics of the lower control layers and physical network can be represented
or approximated by instantaneous, steady-state characteristics.
Let a network be represented by a network model. Let the model consist of
nodes, and let , for {1, . . . , } denote a particular node. Each of the nodes in
the network model is labeled with a set of variables (e.g., voltage magnitudes and
angles) and constraints (e.g., power flow equations) used to compute the steadystate values for these variables, given values for inputs (e.g., amount of power to be
generated) and disturbances (e.g., amount of power consumed). The constraints of
a particular node involve variables of that particular node and variables of other
nodes, referred to as the neighboring nodes N = { ,1 , . . . , ,n }. To indicate
the interaction between node and its neighboring nodes in N , we define an arc
between and each node N .
Let for node {1, . . . , }, the variables z Rnz , u Rnu , and d Rnd ,
denote the (static) states,2 the input variables, and the disturbance variables
2
Sometimes the static states are also referred to as algebraic variables.
257
associated with node , respectively, and let the constraints of node be given by
g (z , u , d , z ,1 , . . . , z ,n ) = 0,
(1)
where z are the variables of neighboring node N , and g are the constraint
functions of node . These constraint function are assumed to be smooth. A steadystate model for the overall network is obtained by aggregating the constraints (1) for
all nodes {1, . . . , }, and is compactly represented as
g(z, u, d) = 0,
(2)
where z, u, and d are the state, input, and disturbance variables of the overall network, and g defines the steady-state characteristics of the network. Given the inputs
u and the disturbance variables d, the steady state in which the network settles is
determined by solving the system of equations (2).
2.2 Control Objectives

With each node objective terms can be associated. These objective terms specify
which behavior is desired by assigning costs to the values of the variables z and u
of that node. The objective terms involve the variables of node and may in addition
also involve the variables of the neighboring nodes N . The summation of the
objectives terms of all nodes in the network model gives the objective for the control
of the overall network. E.g., if a node has assigned to it constraints representing the
characteristics of a transmission line, then as objective term the costs on power
losses of that transmission line may be associated to the node. In addition, if a node
represents the characteristics of a bus, then an objective term representing costs on
a voltage magnitude violation of that bus may be associated to this node.
2.3 Definition of Subnetworks

The values of the inputs u should be adjusted in such a way that the objectives
associated with the nodes are achieved as well as possible. Let for a control agent
i the nodes that it controls define its subnetwork. The prediction model that control
agent i then uses consists of the union of the constraints of each node that is part
of its subnetwork. Let the subnetwork and the control goals of a control agent be
defined using one of the approaches mentioned in Sect. 1.2.
In the following we first discuss an approach that can be used in the case that
the subnetworks are touching. Then we extend this approach to be able to deal with
overlapping subnetworks. For the sake of simplicity we assume below that there are
no nodes that do not belong to any subnetwork.
258
3 Multi-Agent Control of Touching Subnetworks

In this section we discuss a technique for coordinating control agents that use
touching subnetworks. This technique is based on an adaptation of the ideas of
the modified Lagrange technique proposed in [3] to our network and objective formalization. The technique requires that subnetworks of any two control agents are
touching, i.e., the nodes in the subnetwork of one control agent are only taken into
account (i.e., modeled and controlled) by that control agent and not by any other. In
short, when the control agents have to determine actions, they perform a series of
iterations, in each of which the control agents perform a local optimization step and
communicate information. The local optimization problems are formulated using
local objective functions, local prediction models of the subnetworks, and local constraints. After each local optimization the control agents exchange information, reformulate their local optimization, and perform a new optimization. This continues
until a stopping condition is satisfied. Below we first introduce some terminology,
then formulate the control problem as considered by an individual control agent,
and then we discuss the scheme used by multiple control agents for coordination
and communication.
3.1 Internal and External Nodes

We define the following concepts that will be frequently used in the remainder of
this chapter:
We categorize the nodes that control agent i considers based on their location
from the point of view of control agent i. For touching subnetworks, the nodes
that control agent i considers can be internal nodes or external nodes. The internal nodes of control agent i are those nodes that belong exclusively to its subnetwork. The external nodes of control agent i are those nodes that do not belong to
its subnetwork.
Based on the distinction between internal and external nodes of control agent i,
we make a distinction between internal and external variables of control agent i.
The internal variables are those variables associated with the internal nodes of
control agent i. The external variables are those variables associated with the
external nodes of control agent i.
For control agent i, the localized constraint type of a particular constraint associated with a node that control agent i considers is formed by the combination of
the location and the types of variables involved in that constraint. The localized
constraint type of a constraint associated with a node considered by control
agent i is denoted by CVars
i,Loc , where Loc {int, ext} indicates the location of the
node to which the constraint is associated, and Vars {int, int+ext} indicates
the variables involved in the constraint. Recall that a constraint associated with a
particular node involves variables of that particular node and possibly variables
259
Table 1 Overview of the localized constraint types of constraints associated with nodes in a subnetwork that touches other subnetworks. The location indicates the location of the node from the
point of view of control agent i. The variables involved in the constraint indicate which variables
are involved in the constraint, from the point of view of control agent i
Type
Location
Variables involved in constraint
Cint
i,int
Internal
Internal
Cint+ext
i,int
Cext
i,ext
Cint+ext
i,ext
Internal
Internal+external
External
External
External
Internal+external
3
2
1
1
1
b
1
2
Fig. 3 Illustration of different localized constraint types that can be found at nodes considered by
control agent i. The number next to a node in the figure corresponds as follows to the localized
int
int+ext
constraint types of the constraints that can be associated to that node: 1: Cint
i,int ; 2: Ci,int , Ci,int ;
int+ext
ext
3: Ci,ext , Ci,ext
of neighboring nodes. The constraints associated with the nodes considered by

control agent i can therefore have the localized constraint types listed in Table 1.
Figure 3 illustrates for some nodes the localized constraint types that can be
found at these nodes.
In a similar way as we have defined localized constraint types CVars
i,Loc , we also
Vars
define localized objective term types Ji,Loc, referring to the location of the node
to which an objective term is associated and the variables that are involved in the
objective function term.
3.2 Control Problem Formulation for One Agent

The local optimization problem of control agent i consists of minimizing the local
objective function Ji , subject to the prediction model of subnetwork i and additional
constraints on inputs and outputs. Below we focus on the issues arising due to the
presence of subnetworks that touch the subnetwork of control agent i. We discuss
the issues arising with respect to the prediction model and the objective function of
260
control agent i. For the sake of simplicity of explanation we consider two control
agents, control agent i with neighboring agent j, that together control subnetworks
that cover all nodes of the network model. The generalization to more than 2 control
agents and not fully-covered networks is straightforward.
3.2.1 Prediction Model

The prediction model of control agent i consists of the constraints associated with all
its internal nodes. In order to make predictions, control agent i has to know accurate
values for all variables involved in the constraints of these nodes. The internal nodes
that do not have external neighboring nodes do not require special attention, since
the variables involved in the constraints of these internal nodes are of localized constraint type Cint
i,int and thus only involve variables of the subnetwork of control agent i.
However, the internal nodes that are connected to external nodes do require special
attention, since the constraints associated with these internal nodes can be of localized constraint type Cint+ext
i,int , and thus involve not only variables of the subnetwork
of control agent i, but also variables of the subnetwork of neighboring agent j. For
the external variables, control agent i has to coordinate with the neighboring agents
which values these variables should have. To obtain coordination on the values of
the external variables, we apply an idea that was first proposed in [3] as follows.
Below a distinction is made between constraints that are considered as hard,
and constraints that are considered as soft. The hard constraints are constraints that
have to be satisfied at all costs. The soft constraints are constraints for which it
is desirable that they are satisfied, but for which this should not be done at any
price. The hard constraints are included in the formulation of optimization problems
as explicit equality constraints; the soft constraints are included in the objective
function of optimization problems through a penalty term, weighted by a parameter
specifying the costs for violation of the soft constraint.
Recall that the control agents perform a series of iterations and that in each iteration the control agents solve a local optimization problem followed by an exchange
of information. Note that internal and external nodes of control agent i correspond
to external and internal nodes, respectively, of control agent j. Control agent i considers in its local optimization problem the constraints that are associated with its
internal nodes and that are of localized constraint type Cint+ext
as hard constraints,
i,int
using fixed values for the external variables. The values for these external variables
are obtained from the neighboring agent j. Control agent i solves its local optimization problem using these values for the external variables. The optimization yields
values for the internal variables of control agent i, and for the Lagrange multipliers that are associated with the constraints of localized constraint type Cint+ext
i,int . The
Lagrange multipliers of these constraints and the values of the internal variables
involved in these constraints are sent to neighboring agent j.
Neighboring agent j considers the constraints of the internal nodes of control
agent i that involve external variables of control agent i in its decision making by
including the associated constraints as soft constraints in its local objective function.
261
In the soft constraints of control agent j, the external variables, which correspond
to internal variables of control agent i, are fixed to the values that control agent i
has sent to control agent j. Also, the soft constraints are weighted by the Lagrange
multipliers as given by control agent i. Neighboring agent j solves its optimization
problem, yielding values for its internal variables. It sends the values of the internal
variables that appear in the soft constraints to control agent i, such that control agent
i can update its information about the corresponding external variables.
Based on this idea, Table 2 shows how control agent i deals with the different
constraints when formulating its optimization problem.
3.2.2 Objectives
The local objective function for control agent i consists of objective function terms
that are associated with the nodes in its subnetwork. Objective terms associated
with internal nodes that are only connected to internal nodes are simply included
in the local objective function. However, objective terms associated with internal
nodes that are also connected to external nodes cause problems for the same reason as constraints associated with such nodes. Coordination on the values of these
variables is achieved by obtaining the desired values for the external variables from
neighboring agents.
Table 3 summarizes how the different localized objective term types that control
agent i are considered, and how the agent deals with these types, when formulating
its optimization problem.
Table 2 Overview of the constraints that control agent i can have and how it deals with these
constraints. For the hard and soft constraints, the external variables are fixed to values obtained
from neighboring agents. For the hard constraints with external variables Lagrange multipliers
are determined. The soft constraints are weighted using the Lagrange multipliers received from
neighboring agents
Localized constraint type
Constraint
Cint
i,int
Hard
Cint+ext
i,int
Hard
Cint+ext
i,ext
Soft
Table 3 Overview of the localized objective term types that control agent i considers and how it
deals with these terms. External variables are fixed to values obtained from neighboring agents
Localized objective term type
How deal with the objective term
Jint
i,int
Include as is
Jint+ext
i,int
Include as is
262
3.3 Control Scheme for Multiple Agents

The outline of the scheme for coordination of control agents controlling touching
subnetworks, based on the scheme proposed in [3], is as follows:
1. Each control agent i measures the current values for the state variables zi and
the input variables ui that are associated with the nodes in its subnetwork. In addition, it obtains predictions of known disturbance variables di . Furthermore, it
obtains through communication from its neighbors values for the external variables and Lagrange multipliers associated with the external nodes that control
agent i considers.
2. The iteration counter s is set to 1.
(s1)
(s1)
3. Let win,i and soft,i denote the external variables and Lagrange multipliers,
respectively, of which control agent i has received the values from neighbor(s1)
(s1)
ing agents. Given win,i and soft,i , each control agent i {1, . . . , n} performs
concurrently with the other control agents the following steps:
a. Control agent i solves the local optimization problem:

(s1)
(s1) T
(s1)
+ soft,i
g soft,i zi , ui , win,i
min Ji zi , ui , win,i
zi ,ui
(3)
subject to
g hard,i (zi , ui , di ) = 0

(s1)
=0
g hard,ext,i zi , ui , di , win,i
(5)
zi,min zi zi,max
ui,min ui ui,max ,
(6)
(7)
(4)
where zi,min and zi,max are upper and lower bounds on zi , ui,min and ui,max
are upper and lower bounds on ui , g soft,i are the constraints of localized
hard,i are the constraints of localized constraint type
constraint type Cint+ext
i,ext , g
,
g
are
the
constraints
of localized constraint type Cint+ext
Cint
hard,ext,i
i,int
i,int . Solv(s)
(s)
ing this local optimization results in values for the variables zi and ui ,
(s)
as well as Lagrange multipliers hard,ext,i associated with the constraints (5)
for current iteration s. After solving this optimization problem the variables
(s)
wout,i can be determined as:
(s)
i
wout,i = K
!T
T T
(s)
(s)
T
,
(di )
zi
ui
(8)
where wout,i are the so-called interconnecting output variables, selected

i . These variables represent the variables that
using a selection matrix K
control agent i uses in its communication to neighboring agents. Selection
i has in each row only zeros, except for a single 1 in the column
matrix K
corresponding to the position of an element of
263
!T
T T
(s)
(s)
zi
, ui
, (di )T
that is an interconnecting output variable.

(s)
b. Control agent i sends the values of the Lagrange multipliers hard,ext,i of the
and the values of wout,i
hard constraints of localized constraint type Cint+ext
i,int
corresponding to internal variables of these nodes to the neighboring agents
that consider the involved external variables.
c. Control agent i receives from the neighboring agent j those Lagrange multipliers related to the localized constraint type Cint+ext
and those values of
i,ext
the internal variables of the neighboring agents that control agent i requires
in order to fix its external variables. Control agent i uses this received in(s)
(s)
formation at the next iteration as soft,i and win,i .
4. The next iteration is started by incrementing s and going back to step 3, unless a local stopping condition is satisfied for all control agents. The stopping
condition is defined as the condition that the absolute changes in the Lagrange
multipliers from iteration s 1 to s are smaller than a pre-defined small positive
constant ,term .
A shortcoming of this method is that it requires that the subnetworks are touching, since it assumes that each node in the network model is assigned to only one
of the subnetworks. However, in the case of control of overlapping subnetworks,
some of the nodes are included in more than one subnetwork and the identification
of internal and external nodes of a control agent is not straightforward any more.
Therefore, the method is not directly applicable to overlapping subnetworks. In the
following section we extend the method discussed above to control of overlapping
subnetworks.
4 Multi-Agent Control for Overlapping Subnetworks

We first propose some new definitions, next we consider the issues appearing due to
the overlap, and then we propose a way to deal with these issues. Again, for simplicity of explanation we consider two control agents, control agent i with neighboring
control agent j, that together control the subnetworks, which are assumed to cover
the full network model.
4.1 Common Nodes

In addition to internal and external nodes as defined before, for control of overlapping subnetworks we make the following definitions:
Common nodes are nodes that belong to the subnetwork of control agent i and
that also belong to the subnetwork of the control agent j. A sub-subnetwork
264
Table 4 Overview of the localized constraint types for overlapping subnetworks

Type
Location
Cint
i,int
int+com
Ci,int
Cint+ext
i,int
int+com+ext
Ci,int
Cint+com
i,com
CInt+com+ext
i,com
Ccom
i,com
Ccom+ext
i,com
Cext
i,ext
Cint+ext
i,ext
Ccom+ext
i,ext
Cint+com+ext
i,ext
Variables involved in constraint
Internal
Internal
Internal
Internal+common
Internal
Internal+external
Internal
Internal+common+external
Common
Internal+common
Common
Common
Common
Common
Common+external
External
External
External
Internal+external
External
Common+external
External
defined by the nodes common to several subnetworks is referred to as a common

sub-subnetwork.
The variables associated with the common nodes are referred to as the common
variables.
Given the definition of a common node, the number of possibilities for localized
constraint types increases. Table 4 lists the localized constraint types that can
be considered by a control agent when subnetworks can be overlapping. In total
there are 12 different localized constraint types. Figure 4 illustrates some of the
possible localized constraint types.
In addition to the extension of the localized constraint types, the localized objective term types are extended as well, by also defining localized objective term
types that are based on variables of common nodes.
4.2 Control Problem Formulation for One Agent

For multi-agent control of overlapping subnetworks an approach has to be found to
deal with the common nodes. Since the common nodes are considered by several
control agents, the constraints associated with these common nodes appear in the
subnetwork models of multiple control agents. Even though the control agents have
the same objective with respect to these nodes, combined with the objective for
their internal nodes, conflicting values for the variables of the common nodes can
be the result. Below we discuss how to extend the scheme of Sect. 3 for control of
overlapping subnetworks. Again, for the sake of simplicity of explanation we focus
on two control agents: control agent i with neighboring agent j.
11
1
9
8
6
10
1
i
265
Fig. 4 Illustration of different localized constraint types that can be found at particular nodes
considered by control agent i. The number next to a node in the figure corresponds as follows to
the localized constraint types of the constraints that can be associated to that node: 1: Cint
i,int ; 2:
int+ext ; 3: Cint+ext , Cext ; 4: Cint , Cint+com ; 5: Cint , Cint+com , Cint+ext , Cint+com+ext ; 6: Ccom ;
Cint
,
C
i,ext
i,ext
i,com
i,int
i,int
i,int
i,int
i,int
i,int
i,int
i,int
com+ext
com
int+com+ext ; 8: Ccom , Ccom+ext ; 9: Ccom , Cint+com ; 10: Cext , Cext+com ;
7: Cint+com
i,com , Ci,com , Ci,com , Ci,com
i,com
i,com
i,com
i,com
i,ext
i,ext
com+ext , Cext , Cint+com+ext
11: Cint+ext
i,ext , Ci,ext
i,ext
i,ext
4.2.1 Prediction Model

Similarly as for control of touching subnetworks, for control of overlapping subnetworks, internal nodes of control agent i that are connected to external nodes require
special attention, since the constraints associated to these nodes may involve external variables. In addition to this, common nodes of control agent i that are connected
to external nodes also require special attention. The extension of the approach for
control of touching subnetworks to the control of overlapping subnetworks involves
the following extension of the prediction model.
Control agent i considers as prediction model the constraints of all internal and
int+ext+com ,
common nodes. For the constraints of localized constraint types Cint+ext
i,int , Ci,int
com+ext
int+com+ext
Ci,com , and Ci,com
the control agent takes for the external variables values
that it has received from neighboring agent j. When control agent i has solved its
optimization problem, it sends the values of the internal and the common variables
of the constraints of these specialized constraint types to neighboring agents.
Neighboring agent j considers in its optimization problem the constraints of the
internal and common nodes of control agent i that involve external variables of
control agent i as soft constraints by including them in the objective function through
a penalty term, weighted by the Lagrange multipliers provided by control agent i,
and with fixed values for the external and common values in the soft constraints
as received from control agent i. Note that although control agent j considers fixed
values for the common variable in the soft constraints, it will not fix the values for
the common variables in the hard constraints (similarly as control agent i). Hence,
control agents i and j share the responsibility for the common variables. The result
266
Table 5 Overview of the way in which control agent i considers the constraints of particular localized constraint types in its optimization problem. For the hard constraints all external variables
are fixed to values obtained from neighboring agents. For the soft constraints all external and common variables are fixed. For the hard constraints with external variables Lagrange multipliers are
determined. The soft constraints are weighted with Lagrange multipliers obtained from neighboring agents. Note that the soft constraint part of the inclusion of constraints of type Cint+com+ext
i,com
involves fixed external and common variables and a Lagrange multiplier as obtained from neighboring agents, whereas the hard constraint part of the inclusion of constraints of type Cint+com+ext
i,com
involves only fixed external variables
Localized constraint type
Constraint
Cint
i,int
Hard
int+com
Cint+ext
i,int , Ci,int
Hard
Cint+com+ext
i,int
Cint+com
i,com
Cint+com+ext
i,com
Ccom
i,com
Ccom+ext
i,com
Cint+ext
i,ext
Cint+ext+com
i,ext
Hard
Hard and soft
Hard and soft
Hard
Hard
Soft
Soft
of solving the optimization problem of neighboring agent j therefore yields values

for the internal, common, and external variables of control agent j. The internal
variables of control agent j related to the soft constraints are sent to control agent i.
Table 5 summarizes how control agent i deals with the different localized constraint types.
4.2.2 Objectives
With the nodes that control agent i has in its subnetwork objective terms are associated. The objective function terms associated with each node can depend on
the variables associated with that node and its neighboring nodes. As before, the
objective terms involving only internal variables require no special attention. The
objective terms involving both internal and external variables can be dealt with by
fixing the external variables, as is also done for control of touching subnetworks.
However, the common variables appearing in control of overlapping subnetworks
do require special attention.
For control of overlapping subnetworks, multiple control agents will try to control the values of the common variables. To allow control agents to jointly achieve
performance comparable to the performance that an overall centralized control agent
can achieve, the responsibility for the objective terms involving only common variables, i.e., of localized objective term type Jcom
i,com , is shared equally by the control
agents. Hence, each control agent i that considers a particular common node ,
267
Table 6 Overview of the localized objective term types that control agent i considers and how it
deals with the associated objective terms. External variables are fixed. Variable N is the number
of control agents considering node as common node
Localized objective term type
How deal with the objective term
Jint
i,int
Jint+ext
i,int
Jint+com
i,int
Jcom
i,com
Jint+com
i,com
Include as is
Include as is
Include as is
Include partially by weighting it with a factor 1/N
Include as is
includes in its objective function 1/N times the objective function terms of such
nodes of localized objective term type Jcom
i,com , where N is the number of control
agents considering node as common node. Control agent i in addition includes
into its objective function the objective terms of all its internal nodes, and the objective terms of these common nodes that involve only internal and common variables,
int+ext int+com and
i.e., the objective terms of localized objective term types Jint
i,int , Ji,int , Ji,int
int+com
Ji,com .
Table 6 summarizes how control agent i deals with the different localized objective term types.
4.3 Control Scheme for Multiple Agents

We have discussed how each control agent formulates its prediction model and objective function. The scheme that we propose for multi-agent control for overlapping
subnetworks consists of the scheme proposed in Sect. 3 for touching subnetworks,
with the following changes:
Control agent i receives from the neighboring agents the following information
at initialization and after each iteration:
Lagrange multipliers with respect to the constraints of localized constraint
int+com+ext , Cint+ext , and Cint+ext+com .
type Cint+com
i,com , Ci,com
i,ext
i,ext
Values for the external variables and the common variables involved in these
constraints.
The optimization problem that each agent solves is changed accordingly to reflect
the extensions discussed in this section, i.e., to take into account the constraints
as given in Table 5 and the objective terms as given in Table 6.
The result is a control scheme that can be used by higher-layer control agents
that control subnetworks that are overlapping. The control agents hereby share the
responsibility for the common variables. In the next section we apply this scheme
on an optimal flow control problem in power networks.
268
5 Application: Optimal Flow Control in Power Networks

In this section apply the scheme for multi-agent control of overlapping subnetworks,
as discussed in Sect. 4, to the problem of optimal power flow control in power networks. A case study is carried out on the IEEE 57-bus power network [2], comprising as components generators, loads, transmission lines, and buses, with in addition
FACTS devices installed at various locations, as illustrated in Fig. 5. Two configurations are considered: in the first configuration only SVCs are included; in the second
configuration only TCSCs are present. Each of the FACTS devices is controlled by
an individual control agent.
5.1 Parameters of the Power Network

The parameters of the IEEE 57-bus base network can be obtained from the Power
Systems Test Case Archive [2]. Line limits on the apparent power flows have been
assigned to all transmission lines in such a way that no lines are overloaded. In
order to find an interesting and meaningful situation for FACTS control, the grid
was adapted by placing an additional generator at bus 30 leading to increased power
flows in the center of the grid. The parameters of this generator are as follows:
dV,gen,m = 1.03 p.u., dP,gen,m = 0.5 p.u., for m = 30. The parameters of the base IEEE
57 network used in this chapter can be found in [2].The limits on the apparent power
flows are set as listed in Table 7.
In the following we make a distinction between a bus and a node. A bus refers to
an element of the physical power network, whereas a node refers to an element of the
model of the physical power network. Since for each physical bus a corresponding
node is included in the model, references to a bus or its corresponding node can be
interchanged, except for when assigning constraints relating to two buses, such as
constraints imposed due to transmission lines, to a single node, as we will see next.
Below we formulate the steady-state models used to describe the network behavior, we assign the constraints to nodes, we set up the objective terms associated
with the nodes, we discuss the way in which the subnetworks can be determined
using the influence-based approach, and we illustrate the workings of the proposed
approach.
5.2 Steady-state Characteristics of Power Networks

As the focus lies on improving the steady-state network security, the power network
is modeled using equations describing the steady-state characteristics of the power
network. As we will see, the aspects of the steady-state security that we are interested in can be determined from the voltage magnitude and voltage angle at each
of the 57 (physical) buses in the network. We therefore define 57 nodes to model
269
56
51
9
80
40
54
36
52
53
69
67
SV 35
C
68
32
67
30
36
29
40
22
TCSC
21
41
28
39
27
38
26
20
18
37
19
52
43
42
25
24
35
29
34
31
19
44
23
45
30
33
33
46
20
22
34
32
31
21
47
53
48
49
37
50
70
76
51
78
79
38
57
44
39
49
48
62
61
60
47
46
72
58
45
73
77
C
SV
14
59
28
15
18
2
3
TCSC
55
74
75
57
63
13
41
55
42
66
14
15
56
54
43
71
10
24
11
12
25
13
17
17
16
11
27
10
23
16
26
12
65
50
64
Fig. 5 IEEE 57-bus network extended with either SVCs installed at buses 14 and 34, or with
TCSCs in lines 22 and 72
the network, and assign to each node m the voltage magnitude zV,m per unit (p.u.)
and the voltage angle z ,m (degrees) as variables. In order to determine the values for
these variables under different disturbance variables and actuator values, models for
the components and their influence on the voltage magnitude and angle are defined.
We model the transmission lines, the generators, the loads, and the FACTS devices.
270
Table 7 Line limits on the apparent power flows

Line nr.
Limit (p.u.)
Line nr.
Limit (p.u.)
Line nr.
Limit (p.u.)
Line nr.
Limit (p.u.)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
1.800
1.650
0.721
0.412
0.800
0.750
1.200
2.200
0.600
0.450
0.300
0.500
0.700
0.936
1.900
1.050
1.200
1.200
0.300
0.300
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
0.550
2.160
0.700
0.700
0.900
0.600
0.750
1.000
0.300
0.300
0.300
0.300
0.400
0.450
0.300
0.300
0.350
0.350
0.350
0.400
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
1.100
0.500
0.500
0.400
0.300
0.400
0.400
0.400
0.700
0.800
0.300
0.300
0.400
0.300
0.300
0.300
0.500
0.600
0.636
0.650
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
0.350
0.300
0.500
0.300
0.450
0.600
0.550
0.500
0.350
0.400
0.300
0.700
0.300
0.300
0.300
0.300
0.300
0.550
0.550
0.500
5.2.1 Transmission Lines

For the transmission lines the well-known -model is used [15]. The active power
zP,mn (p.u.) and the reactive power zQ,mn (p.u.) flowing from bus m over the transmission line to bus n are then given by:

R,mn
2
zP,mn = (zV,m )
(R,mn )2 + (X,mn )2

R,mn
cos(z ,m z ,n )
zV,m zV,n
(R,mn )2 + (X,mn )2

X,mn
+zV,m zV,n
sin(z ,m z ,n )
(9)
(R,mn )2 + (X,mn )2

X,mn
zQ,mn = (zV,m )2
(R,mn )2 + (X,mn )2

R,mn
zV,m zV,n
sin(z ,m z ,n )
(R,mn )2 + (X,mn )2

B,mn
X,mn
cos(z ,m z ,n ) ,
(zV,m )2
zV,m zV,n
2
(R,mn )2 + (X,mn )2
(10)
271
where B,mn (p.u.) is the shunt susceptance, R,mn (p.u.) is the resistance, and X,mn
(p.u.) is the reactance of the line between buses m and n.
The constraints for each transmission line going from bus m to bus n, for n Nm
(where Nm is the set of neighboring buses of bus m, i.e., the buses that are physically
connected to bus m through a transmission line), are assigned to node m, if m < n,
and to node n otherwise.
5.2.2 Generators
Generators are assumed to have constant active power injection and constant voltage
magnitude, and therefore
zP,gen,m = dP,gen,m
zV,m = dV,gen,m ,
(11)
(12)
where dP,gen,m is the given active power that the generator produces, and dV,gen,m
is the given voltage magnitude that the generator maintains. At most one generator
can be connected to a bus, since a generator directly controls the voltage magnitude
of that bus.
The generator connected to bus 1 is considered as a slack generator, i.e., a generator with infinite active and reactive power capacity, with fixed voltage magnitude
and angle [15]. So, for this generator we have with m = 1
zV,m = dV,gen,m
(13)
z ,m = d ,gen,m ,
(14)
where d ,gen,m is the given voltage angle ensured by the generator.

The constraints of a generator at bus m are assigned to node m.
5.2.3 Loads
The loads are constant active and constant reactive power injections, i.e.,
zP,load,m = dP,load,m
(15)
zQ,load,m = dQ,load,m ,
(16)
where dP,load,m and dQ,load,m are the given active and reactive power consumption,
respectively, of the load connected to bus m. For simplicity, only one load can be
connected to a bus. Multiple loads can easily be aggregated to obtain a single load.
The constraints of the loads at bus m are assigned to node m.
272
5.2.4 FACTS Devices

SVC
An SVC is a FACTS device that is shunt-connected to a bus m and that injects or
absorbs reactive power zQ,SVC,m to control the voltage zV,m at that bus [10]. The
SVC connected to bus m accepts as control input the effective susceptance uB,SVC,m .
The injected reactive power zQ,SVC,m of the SVC is:
zQ,SVC,m = (zV,m )2 uB,SVC,m .
(17)
The control input uB,SVC,m is limited to the domain:

uB,SVC,min,m uB,SVC,m uB,SVC,max,m ,
(18)
where the values of uB,SVC,min,m and uB,SVC,max,m are determined by the size of the
device [7].
The constraints of an SVC at bus m are assigned to the node m.
TCSC
A TCSC is a FACTS device that can control the active power flowing over a line
[10]. It can change the line reactance zX,line,mn . The TCSC is therefore considered as
a variable reactance uX,TCSC,mn connected in series with the line. If a TCSC is connected in series with a transmission line between buses m and n, the total reactance
zX,line,mn of the line including the TCSC is given by:
zX,line,mn = X,mn + uX,TCSC,mn ,
(19)
where X,mn is the reactance of the line without the TCSC installed. The reactance
uX,TCSC,mn is limited to the domain:
uX,TCSC,min,mn uX,TCSC,mn uX,TCSC,max,mn ,
(20)
where the values of uX,TCSC,min,mn and uX,TCSC,max,mn are determined by the size of
the TCSC and the characteristics of the line in which it is placed, since due to the
physics the allowed compensation rate of the line uX,TCSC,mn /X,mn is limited [7].
The constraints of the TCSC at the line between bus m and n are assigned to node
m, if m < n, and to node n otherwise.
5.2.5 Power Balance

By Kirchhoffs laws, at each bus the total incoming power and the total outgoing
power has to be equal. This yields the following additional constraints for bus m:
273
zP,load,m zP,gen,m + nNm zP,mn = 0
(21)
zQ,load,m zQ,gen,m zQ,SVC,m + nNm zQ,mn = 0.
(22)
If no generator is connected to bus m, then zP,gen,m and zQ,gen,m are zero. If no load
is connected to bus m, then zP,load,m and zQ,load,m are zero. If no SVC is connected to
bus m, then zQ,SVC,m is zero.
The constraints resulting from Kirchhoffs laws for bus m are assigned to node m.
5.3 Control Objectives

The objectives of the control are to improve the system security through minimization of the deviations of the bus voltages from given references to improve
the voltage profile, minimization of active power losses, and preventing lines from
overloading, by choosing appropriate settings for the FACTS devices. These objectives are translated into objective terms associated with the buses as follows:
To minimize the deviations of the bus voltage magnitude zV,m of bus m from a
given reference dV,ref,m , an objective term pV (zV,m dV,ref,m )2 is associated with
node m, where pV is a weighting coefficient.
To minimize the active power losses over a line between bus m and bus n, an objective term ploss (zP,mn + zP,nm ), where ploss is a weighting coefficient, is associated to node m, if m < n, and to node n otherwise. Note that the term zP,mn + zP,nm ,
which represents the power losses, is always nonnegative.
To minimize the loading of the line between buses m and n, an objective term

2
zS,mn
is associated to node m, if m < n, and to node n otherwise, as pload zS,max,mn
,
where pload is a weighting coefficient, and where zS,mn is/the apparent power
flowing over the line from bus m to bus n, defined as zS,mn = (zP,mn )2 + (zQ,mn )2 .
The relative line loading is penalized in a quadratic way such that an overloaded
line is penalized more severely than a line that is not overloaded.
The weighting coefficients pV , ploss , and pload allow to change the weight given to
each objective. In the following we take pV = 1,000, ploss = 100, and pload = 1.
5.4 Setting Up the Control Problems

Each FACTS device is controlled by a different control agent. The influence-based
subnetworks of the control agents controlling the FACTS devices can be overlapping, and therefore the control problems of the control agents are set up using the
approach discussed in Sect. 4. To solve their subproblems at each iteration the control agents use the nonlinear problem solver SNOPT v5.8 [6], as implemented in
Tomlab v5.7 [11], and accessed from Matlab v7.3 [17].
274
In the following we illustrate how the approach works for a particular assignment
of nodes to subnetworks in two representative scenarios.
5.5 Simulations
Various test scenarios with different FACTS devices and subnetworks have been
examined. Here we present two representative scenarios. The subnetworks used in
these scenarios are shown in Fig. 6. It can be seen that these subnetworks are overlapping, since there are several nodes that are included in both subnetworks.
5.5.1 Scenario 1: Control of SVCs

In the first scenario, SVCs are placed at buses 14 and 34. As the SVCs are mainly
used to influence the voltage profile, the line limits are chosen such that no line is at
the risk of being overloaded.
4
20
21
45
19
20
31
areaB
32
34
38
36
48
53
22
78
37
33
57
42
76
51
49
45
77
43
55
75
46
44
43
30
28
32
31
40
41
TCSC
29
52
53
68
48
47
34
SV
35 67
52
36
24
71
56
73
40
54
69
50
63
49
39
25
12
areaA
62
35
42
39
66
61
79
50
27
12
60
38
23
25
47
57
33
37
13
13
SV
46
44
24
59
72
21
27
14
28
14
26
17
15
TCSC
18
15
58
30
29
26
17
18
19
16
16
56
74
41
55
70
11
54
11
23
64
10
80
22
9
8
65
10
51
Fig. 6 IEEE 57-bus system with decomposition into two subnetworks. Scenario 1: SVCs at buses
14 and 34, scenario 2: TCSCs in lines 22 and 72. The dotted line indicates the borders of subnetwork 1 (light shaded); the dashed line indicates the borders of subnetwork 2 (dark shaded).
The region encapsulated both by subnetwork 1 and subnetwork 2 is the common region (medium
shaded)
0.5
zV,err,m (p.u.)
0.04
0.4
0.3
0.2
m = 19
m = 21
m = 40
0.03
0.02
0.01
0.1
0
0.1
10
12
14
16
18
20
0.02
0.2
m = 19
m = 21
m = 40
0.015
0.3
0.4
0.5
s
m = 34
m = 14
z ,err,m (deg)
uB,SVC,m (p.u.)
275
0.01
0.005
10
12
14
16
18
20
10
(a)
(b)
12
14
16
18
20
Fig. 7 (a) Convergence of the settings of the SVCs at buses 14 and 34, as a function of the iteration,
for scenario 1. (b) Convergence of the difference between the values of the voltage magnitudes
(top) and the voltage angles (bottom) as considered by both control agents for buses 19, 21, 40, as
a function of the iteration, for scenario 1
Figure 7(a) shows the convergence of the SVC settings over the iterations. As
can be seen, the settings of the SVCs converge within only a few iterations to the
final values, which in this case are equal to the values obtained from a centralized
optimization. Figure 7(b) shows the evolution of the deviations between the values determined by both subnetworks for the voltage magnitudes and angles at some
common buses. In the figure the error zV,err,m is defined as the absolute difference
between the values that control agents 1 and 2 want to give to the voltage magnitude zV,m . Similarly, the error z ,err,m is defined as the absolute difference between
the values that control agents 1 and 2 want to give to the voltage angles. As can be
seen fast convergence is obtained.
5.5.2 Scenario 2: Control of TCSCs

In the second scenario, TCSCs are installed on lines 72 and 22. Since TCSCs are
mainly used to influence active power flows and to resolve congestion, the line limits
are chosen such that lines 7 and 60 are overloaded if the FACTS devices are not
being used.
The results for the TCSC settings and the difference between the voltage magnitudes and angles for some common buses over the iterations are given in Figs. 8(a)
and 8(b), respectively. The control agent of subnetwork 1 sets the TCSC to its upper
limit at the first few iterations. But after some additional iterations, the values that
the control agents choose converge to their final values, which are again equal to the
values obtained by a centralized control agent.
276
0.04
1.5
x 10
bus19
v1minv2
subnet1
0.03
subnet2
0.02
bus21
bus40
0.5
0.01
xtcsc
1.5
x 10
10
12
iteration
14
16
18
20
0.01
th1minth2
bus19
0.02
0.03
0.04
10
iteration
15
bus40
0.5
20
bus21
(a)
10
12
iteration
14
16
18
20
(b)
Fig. 8 (a) Convergence of the settings of the TCSCs in lines 22 and 72 (i.e., the lines between buses
7 and 8, and buses 44 and 45, respectively), as a function of the iteration number, for scenario 2. (b)
Convergence of the difference between the values of the voltage magnitudes (top) and the voltage
angles (bottom) as considered by both control agents for buses 19, 21, 40, as a function of the
iteration number, for scenario 2
1.02
Smn /Smn,max
1.01
0.99
0.98
m = 46, n = 47
m = 6, n = 8
0.97
0.96
10
12
14
16
18
20
s
Fig. 9 Convergence of the relative line loadings of lines 7 and 60 (i.e., the lines between buses 6
and 8, and 46 and 47, respectively), as a function of the iteration number, for scenario 2
In Fig. 9 the line loadings of lines 7 and 60, i.e., the lines which are overloaded
without FACTS devices in operation, are shown. Line 7 is immediately brought
below its limit whereas for line 60, the loading approaches 100% in the course of
the optimization process.
277
6 Conclusions and Future Research

In this chapter we have focused on an alternative way to define subnetworks for
higher-layer multi-agent control. The higher control layer uses steady-state characteristics only. We have discussed how subnetworks can be defined based on the
influence of inputs on the variables of nodes. When such an approach is used to define subnetworks, some subnetworks could be overlapping, resulting in constraints
and objectives in common sub-subnetworks. We have proposed a method for higherlayer multi-agent control that can be used by control agents that control such overlapping subnetworks.
To illustrate the topics discussed and the proposed approach, we have defined
overlapping subnetworks for Flexible Alternating Current Transmission Systems
(FACTS) in an adjusted version of the IEEE 57-bus power network. Using the proposed control approach, we have then solved an optimal power flow control problem. The simulations illustrate that in the considered cases the proposed approach
can achieve fast convergence to actuator values that are globally optimal.
Further research will address the following issues and topics. It will be determined formally when the approach converges and what the quality of the obtained
solutions is, in particular when compared to an overall single-agent, centralized,
control scheme. This will provide more insight into the quality of the solutions and
the time required to obtain these solutions. Also, power networks are just a particular network from the general class of transportation networks. Other examples from
the class of transportation networks to which the approach discussed in this chapter
could be successfully applied in future work are traffic and transportation systems
[4], natural gas networks [23], combined electricity and gas networks [1], water networks [21], etc. The approach will also be extended to deal with dynamics using
ideas from [19].
Acknowledgement This research is supported by the BSIK project Next Generation Infrastructures (NGI), the Delft Research Center Next Generation Infrastructures, the project Coordinated
control of FACTS devices, supported by ABB Switzerland, the European STREP project Hierarchical and distributed model predictive control (HD-MPC), and the project Multi-Agent Control
of Large-Scale Hybrid Systems (DWV.6188) of the Dutch Technology Foundation STW. The
authors thank Swissgrid for their technical support.
References
1. M. Arnold, R. R. Negenborn, G. Andersson, and B. De Schutter. Multi-area predictive control
for combined electricity and natural gas systems. In Proceedings of the European Control
Conference 2009, Budapest, Hungary, August 2009
2. R. D. Christie. Power Systems Test Case Archive. URL: http://www.ee.washington.edu/
research/pstca/, 2008. Last accessed at May 23, 2008
3. A. J. Conejo, F. J. Nogales, and F. J. Prieto. A decomposition procedure based on approximate
Newton directions. Mathematical Programming, Series A, 93(3):495515, December 2002
4. C. F. Daganzo. Fundamentals of Transportation and Traffic Operations. Pergamon Press,
New York, 1997
278
5. A. Edris, R. Adapa, M. H. Baker, L. Bohmann, K. Clark, K. Habashi, L. Gyugyi, J. Lemay,

A. S. Mehraban, A. K. Meyers, J. Reeve, F. Sener, D. R. Torgerson, and R. R. Wood. Proposed
terms and definitions for flexible AC transmission system (FACTS). IEEE Transactions on
Power Delivery, 12(4):18481853, October 1997
6. P. E. Gill, W. Murray, and M. A. Saunders. SNOPT: An SQP algorithm for large-scale constrained optimization. SIAM Journal on Optimisation, 12(4):9791006, 2002
7. G. Glanzmann and G. Andersson. Using FACTS devices to resolve congestions in
transmission grids. In Proceedings of the CIGRE/IEEE PES International Symposium, San
Antonio, TX, October 2005
8. G. Glanzmann and G. Andersson. FACTS control for large power systems incorporating security aspects. In Proceedings of X SEPOPE, Florianopolis, Brazil, May 2006
9. P. Hines, L. Huaiwei, D. Jia, and S. Talukdar. Autonomous agents and cooperation for the
control of cascading failures in electric grids. In Proceedings of the 2005 IEEE International
Conference on Networking, Sensing and Control, pages 273278, Tucson, AZ, March 2005
10. N. G. Hingorani and L. Gyugyi. Understanding FACTS Concepts and Technology of Flexible
AC Transmission Systems. IEEE Press, New York, New York, 2000
11. K. Holmstrom, A. O. Goran, and M. M. Edvall. Users guide for Tomlab /SNOPT, December
2006
12. G. Hug-Glanzmann, R. R. Negenborn, G. Andersson, B. De Schutter, and J. Hellendoorn.
Multi-area control of overlapping areas in power systems for FACTS control. In Proceedings
of Power Tech 2007, Lausanne, Switzerland, July 2007. Paper 277
13. N. Jenkins, R. Allan, P. Crossley, D. Kirschen, and G. Strbac. Embedded Generation.
TJ International, Padstow, UK, 2000
14. B. H. Kim and R. Baldick. A comparison of distributed optimal power flow algorithms. IEEE
Transactions on Power Systems, 15(2):599604, May 2000
15. P. Kundur. Power System Stability and Control. McGraw-Hill, New York, New York, 1994
16. J. Machowski, J. Bialek, and J. R. Bumby. Power System Dynamics and Stability. Wiley,
New York, New York, 1997
17. Mathworks. Matlab. URL: http://www.mathworks.com/, 2007
18. R. R. Negenborn. Multi-Agent Model Predictive Control with Applications to Power Networks.
PhD thesis, Delft University of Technology, Delft, The Netherlands, December 2007
19. R. R. Negenborn, B. De Schutter, and J. Hellendoorn. Multi-agent model predictive control for
transportation networks: Serial versus parallel schemes. Engineering Applications of Artificial
Intelligence, 21(3):353366, April 2008
20. R. R. Negenborn, S. Leirens, B. De Schutter, and J. Hellendoorn. Supervisory nonlinear MPC
for emergency voltage control using pattern search. Control Engineering Practice, 17(7):
841848, July 2009
21. R. R. Negenborn, P. J. van Overloop, T. Keviczky, and B. De Schutter. Distributed model
predictive control for irrigation canals. Networks and Heterogeneous Media, 4(2):359380,
June 2009
22. F. J. Nogales, F. J. Prieto, and A. J. Conejo. Multi-area AC optimal power flow: A new decomposition approach. In Proceedings of the 13th Power Systems Control Conference (PSCC),
pages 12011206, Trondheim, Germany, 1999
23. A. J. Osiadacz. Simulation and Analysis of Gas Networks. Gulf Publishing Company, Houston,
TX, 1987
24. P. W. Sauer and M. A. Pai. Power System Dynamics and Stability. Prentice-Hall, London, UK,
1998
25. K. P. Sycara. Multiagent systems. AI Magazine, 2(19):7992, 1998
26. D. D. Siljak.
Decentralized Control of Complex Systems. Academic, Boston, Massachusetts,
1991
27. G. Weiss. Multiagent Systems: A Modern Approach to Distributed Artificial Intelligence. MIT
Press, Cambridge, Massachusetts, 2000
Distributed Control Methods for Structured

Large-Scale Systems
Justin Rice, Paolo Massioni, Tamas Keviczky, and Michel Verhaegen
1 Introduction
The development of efficient and tractable distributed control design procedures for
large-scale systems has been an active area of research in the past three decades
[1, 2]. The growing number of applications where such solutions offer increased
functionality, flexibility or efficiency has spawned a renewed interest in this topic.
The main challenges lie in the computational cost and complexity of efficient controller design and implementation. For a thorough overview of distributed and
decentralized control research, see [1], and the introduction to [2].
In this work we describe three recently developed approaches and their extensions,
and apply them to a benchmark problem involving an infinite platoon of vehicles.
For infinite arrays of spatially invariant interconnected systems, a spatial Fourier
transform may be used for solving optimal control problems. The optimal controller
is then often approximated using spatial truncation. Although there are significant
differences between finite and infinite-dimensional interconnected systems, we will
perform our numerical studies for the infinite-dimensional case, noting that all of
the discussed approaches can be extended to the finite-dimensional problem as well,
e.g. [16, 25, 30].
Our first approach to be considered is to exploit the special operator structure
induced by such interconnections. Structure preserving iterative algorithms can be
employed on these Laurent operators with rational symbols to solve a myriad of control problems. While the exact solutions to optimal control problems involving shift
invariant systems are often Laurent operators with irrational symbols, using this
J. Rice, P. Massioni, T. Keviczky, and M. Verhaegen

Delft Center for Systems and Control, Delft University of Technology, 2628 CD, Delft,
The Netherlands
e-mail: j.k.rice@tudelft.nl; p.massioni@tudelft.nl; t.keviczky@tudelft.nl;
m.verhaegen@moesp.org
279
280
J. Rice et al.
approach we can find arbitrarily close rational approximations to these solutions,

which also have the favorable property of admitting spatially discrete distributed
controller realizations.
Special structures in the problem formulation itself have also often offered possibilities to render the design problem tractable. Several approaches have been applying various types of decoupling transformations that allow significant reduction
of the problem complexity in the transformed domain [35]. The second approach
discussed in this work belongs to this class of methods and is applicable to largescale systems composed of identical linear systems interconnected with each other
according to a certain pattern. The proposed method exploits a decomposition of
the global system that converts the global synthesis problem into a set of computationally simple problems, and the use of Linear Matrix Inequalities (LMIs) allows
distributed controller synthesis under performance constraints.
If the large-scale system is composed of independent identical linear system dynamics, the system structure allows significant simplification of the control design
procedure: the synthesis of stabilizing distributed control laws can be obtained by
using a simple local LQR design. Such a local LQR controller based design procedure will be described as the third approach under consideration in this work. We
will also illustrate how stability of the overall large-scale system is related to the
robustness of local controllers and the spectrum of a Laurent matrix representing
the desired sparsity pattern. Although this design method is extremely simple to
calculate and implement, it relies heavily on the identical, dynamically decoupled
subsystem assumption. Furthermore, although the stability result does not depend
on the weighting matrices chosen in the cost function, optimality of the distributed
controller is not guaranteed.
In Sect. 2, the benchmark problem is described, while Sects. 35 provide a summary and extensions of the three discussed control design methods. Each section
contains an explanation of how the particular approach can be used to address the
benchmark problem. Section 6 summarizes our numerical results, while Sect. 7 collects our observations and possible avenues of future research.
2 Problem Statement
As a benchmark problem used for comparison, we consider the problem of controlling the absolute and relative distances in an infinite-dimensional car platoon and
assume a second order model for each vehicle [6]. The dynamics of the states is
described by the following:
!
!
!
!
!
x1i
0
0 1 x1i
0
u
+
(1)
w
=
+
g i
0 0 x2i
q i
x2i
for i = . . . + ; x1i is the position, x2i is the velocity, ui is the control input, wi
is the disturbance input, and q and g are constants which we assume to be both
1 without loss of generality. As we are going to consider the state feedback H2
Distributed Control Methods for Structured Large-Scale Systems
281
problem, the measured output will be the state itself, while as performance output zi
we choose the following:
ui
f1 x1i
(2)
zi =

1 1
1 1
1
f2 2 xi1 xi + 2 xi+1
The performance outputs include the control effort, a symmetric measure of the
relative position and absolute position (needed for the well-posedness, as explained
in [6]), weighted by f2 and f1 .
The full, infinite order system, is then described by the following equations:

= Ax(t)
+ B1 w(t)
+ B2 u(t)
x(t)
(3)
+ Du(t)
z(t) = Cx(t)
where x,
z, u,
w are in the Hilbert space l2 (, ) with the usual inner product and
represent the infinite dimensional vectors containing the stack of all xi , zi ui and wi ,
and the letters with a bar over them represent bounded Laurent operators, of which
A, B1 , B2 , and D are block diagonal, whereas C is banded, with one non-zero band
over the diagonal and one below. We will look for controllers of the form:
u(t)
= K x(t)
(4)
which minimize the

! H2 norm from the disturbance w to the output z. For some
A B
system, G = , the continuous time, discrete-space, spatiotemporal H2 norm
C0
[7] is defined as:
G22 = (
1 2
)
2
2
0
Tr[G(e j , j )G(e j , j ) ]d d
(5)
For the benchmark problem under consideration in this work, solving the problem
in a distributed fashion will imply restricting the search to operators K with only a
limited number of off-diagonal bands.
2.1 H2 Problem and Exact Solution

The Riccati equation for computing the H2 optimal controller is the following:
T
A P + PA PB2 B2 P + C C = 0
T
(6)
T
Notice that the only non block-diagonal term is C C. Let us define Q = C C. By

using a Fourier transform as in [7], we can turn this infinite-size Riccati equation
into an infinite set of parameter dependent equations:
A TP + P A P B 2, B T2,P + Q = 0
(7)
282
J. Rice et al.
with = e j , [0, 2 ]. Actually, only Q depends on (it derives from the only
non-diagonal operator). We can easily see that the equation is equivalent to:
!
!
!
01
00
00
P
P +
P + P
00
01
10
!
(8)

1 0

=0
+ f12 + f22 32 1 + 14 2 + 14 2
00
Partitioning P as:
p p
P = 1 2
p2 p 3
!
(9)
and doing the computations, we have:

2

p2 = k
1 2 1 2
2
2 3
1
2
+ +
2p p3 = 0 with k = f1 + f2
2
2
4
4
p 1 = p 2 p3
from which, the real positive definite stabilizing (and irrational) solution is:
'. (
2k k . k
P =
k
2 k
(10)
(11)
In the following three sections, we will propose different methods for the case
when the explicit solution cannot be calculated as simply as described above.
3 A Rational Laurent Operator Structure Preserving

Iterative Approach to Distributed Control
The first distributed control method that we will discuss exploits the rationally symbolled Laurent matrix structure in distributed systems composed of identical subsystems connected on a string. We will first briefly introduce the type of interconnected
systems that we will consider; see [2, 810] for more detailed discussions.
Consider the set of subsystems:

A Bp
xs
p
v Cp W p
s1
m
v = Cm Z p
s+1 1 p
zs C J
ys
C2 H p
Bm
Zm
Wm
Jm
Hm
B1
Lp
Lm
D11
D21

B2
xs
vsp
Vp
m

Vm
vs
D12 ws
us
D22
(12)
where xs is the local state, us and ys are the local inputs and outputs, zs and ws are
p
performance channel and disturbance input, and vm
s and vs are the coupling terms
to the neighboring subsystems. In this work we will investigate doubly infinite onedimensional strings of such systems (s Z), as in Fig. 1.
283
Fig. 1 String interconnection
Note that for well-posed interconnections [2], can always be transformed such
that Zm = 0, Z p = 0 [11], and from now on we will make this assumption w.l.o.g. In
p
this case the coupling terms, vm
s and vs , are superfluous, and if we resolve them we
get a lifted system:

A B 1 B 2
x
x
: z = C1 D 11 D 12 w
(13)
y
u
C2 D 12 D 22

T
with states x = . . . xT1 xT0 xT1 xT2 . . . , and similarly structured inputs u,
w and
B 1 , etc., are block Laurent operators (see e.g.,
outputs y,
z in Hilbert space l2 . A,
[12,13]) with rational symbols: doubly infinite Toeplitz matrices (see (14) below for
the matrix realization of A) that we may diagonalize using the Fourier transform
to obtain their operator symbols on the unit circle, z T:
1 := A(z) = D + Bm (zI W m )1Cm + B p(z1 I W p )1C p
FAF
assuming that (W p ), (W m ) < 1, so that the series converge. A(z), the sym is a rational transfer function on z T, but not all Laurent operabol of A,
tors have rational symbols, and so we will distinguish them using the notation
A = Lr {Bm ,W m ,Cm , D, B p ,W p ,C p }.
The Fourier transform here is unitary (FF = I) and an isomorphism, so there is
an equivalence in norm and spectrum between the block Lr -operator and its symbol [13]:
= A(z) = sup A(z)2
A
zT
= (A(T))
(A)

.
.
.
.
.
.
..
.
.
.
.
.
.
.
.
.
.
.
. .

p
p
p
p
p
p
p
p
p
p
p
x1 . . .
A
B C
B W C B W W C B W W pW pC p

m
m
p
p
p
p
p
p
p
x2 . . .
B C
A
B C
B W C
B W W pC p

m m m
m m
p p
x3 = . . .
B
W
C
B
C
A
B
C
B pW p C p

x . . . BmW mW mC m
B mW mC m
B mC m
A
B pC p
4
x . . . BmW mW mW mC m BmW mW mC m BmW mC m
B mC m
A
5

.
.
.
.
.
.
..
.
.
.
.
.
.
.
.
.
.
.
.
.
# $% & #
$%

.
. ..

. . . x1

...
x2

...
x3 +B1 w + B2 u

...
x4

...
x5

.
..
.
.
.
& # $% &
..
(14)
284
J. Rice et al.
Fig. 2 String interconnection with similarly connected controller
and also, from a system theoretic point of view, an equivalence of exponential stability and performance,!controllability, etc., [8,9,14] between the lifted system of Lau!
A(z) B(z)
A B
and the Fourier transformed family of systems:
rent operators
C(z) D(z)
CD
parameterized over the unit circle z T. Calculations for the lifted system can thus
also be reduced to finite dimensional eigenvalue problems and Lyapunov and Riccati
equations, rationally parametric on the unit circle [8,9]. However, it is impossible to
compute independently at every z T, and it is not clear how to go about sampling
T, performing local computations, and then interpolating to all T for guaranteed
stability and performance (the eigenvalues of transfer matrices may not even have
Lipschitz smoothness properties [15]), so we will use these properties solely for
proofs, and develop another strategy for computations.
Also, in the same way that the interconnected subsystems s induce a system
can be directly distributed into
of Lr operators, , a controller of Lr operators, K,
interconnected subcontrollers Ks of the same structure as the plant, as we see in
Fig. 2 (see [16] for the finite dimensional analog with explicit controller formulas).
Hence in the following we will be careful to preserve the Lr structure.
3.1 L-Operator Sign Function

As one would expect given the mixed-causal LTI system input-output mapping interpretation of our rational Laurent operators, it is possible to add, multiply, and
invert Lr operators while preserving their structure, and to calculate the operator
norm and perform model order reductions. Simple formulas for these calculations
are available in [17].
However, system analysis and controller synthesis calculations will require more
than an arithmetic, we must be able to solve eigenvalue problems and Lyapunov and
Riccati equations, which is possible using the matrix sign function, as we will now
discuss.
The matrix sign function [18] has been shown to be a very powerful tool for
finite dimensional linear systems analysis and control synthesis (see [19] for an
overview). The Newtons method for calculations also converges extremely fast
(locally quadratically), making it one of the most efficient computational techniques
for solving Riccati equations and other common control problems. In the following
285
we will extend the sign function definition, some convergence bounds, and numerical robustness calculations from the finite dimensional case to the rational Laurent
operator case.
3.2 Definition
We can define the Lr operator sign iteration and sign function as in Algorithm 2
below. As we see, the sign iteration only requires Lr -structure preserving arithmetic
Z 0 = X
1
Z k+1 = (Z k + Z k1 )
2
sign(X ) = lim Z k
for k = 0, 1, 2, ...
Algorithm 2: Sign Iteration

for its computation, and just as in the finite matrix case [18], if X has no spectrum on
the imaginary axis, then every Z k is regular. We should emphasize that for some Lr
Z = sign(X)
will be a Laurent operator, but not always with a rational
operator, X,
symbol, since the space of rational functions is not complete. However, Z k will have
arbitrarily close
a rational symbol k < and thus we can approximate sign(X)
using rational Laurent operators, and the approximation generated by the halted
in operator norm, as we will next
sign iteration will converge very fast to sign(X)
show.
3.3 Convergence
As we discussed in the introduction, the Fourier operator block-diagonalizes an
Lr operator to its symbol. Hence the sign iteration for Lr operators can be considered just as a sign iteration of the rational symbols, Zk (z), z T, or equivalently of
complex matrices Zk (z0 ) at each z0 T.
For a complex finite dimensional matrix, X, the matrix sign can be considered
basically as a sign iteration on each of the eigenvalues individually. Using this property, it can be shown ([17], using results in [20]) that the distance from the spectrum
can be upper bounded based on the locations
of some iterate Z k to that of sign(X)
of the spectrum of X:
<
Z k sign(X)
after k = supzT O(log2 ( (z))2 + log2 log2 ( 1 2 (V (z)))) iterations, where
286
J. Rice et al.
(z) =
max {1 + |R(i)| + |R(i)|1 +
i (X(z))
|Im(i )|
}
|R(i )|
2 () is the condition number in the 2-norm, and X(z) = V (z) (z)V (z)1 is an eigenvalue decomposition. So basically, assuming some bounds on the non-normality of
X(z) and on its spectral radius and the distance from its spectrum to the imaginary
axis (bound on (z)), we can bound the number of sign iterations necessary to get
in the operator norm using the sign iteration. Note that
arbitrarily close to sign(X)
the contribution from the eigenvectors is only a local one, at each V (z), since X
is block diagonalized by the unitary Fourier transform, and a large bound on nonnormality is not much of a limitation anyway, since log22 is a very severe function
(e.g. log2 log2 (10100 ) < 9).
Thus due to the diagonalizability of Lr operators, the sign iteration on them
will work essentially the same as it does on finite dimensional matrices, and we can
compute an arbitrarily close approximation Z k sign(X ), with an Lr structured Zk ,
in a reasonable number of iterations.
3.4 Applications
As discussed in [16] and references therein, in finite dimensions, the matrix sign
function is useful for many things in control analysis and design, such as checking
matrix stability (sign(X) = I (X) C ) and solving Lyapunov and Riccati
equations. As it turns out, most of these results extend directly to the operator sign
function on Lr operators in a straightforward way. In the following we will just show
for solving Riccati equations, but the technique applies to the other applications as
well.
It is well known that one can solve a continuous time finite dimensional Riccati
equation: XA + AT X + Q XRX = 0 where Q =!QT , R = RT , by applying sign itera!
S11 S12
A R
,
to
calculate
sign(H)
=
tions to the Hamiltonian matrix H =
S21 S22
Q AT
and solving the linear system of equations:
!
!
S12
S +I
X = 11
.
(15)
S21
S22 + I
The Riccati equation has a unique stabilizing solution X if and only if (15) has a
unique solution and H has no eigenvalues on the imaginary axis [21].
We now consider the Lr operator case. We can calculate:
!
!
A R
S11 S12
sign(
)
(16)
S21 S22
Q A T
to arbitrary accuracy using Algorithm 1, with Lr structured S11 , S12 , S21 , S22 . We then
solve
S + I
S12
X = 11
S22 + I
S21
287
!
(17)
since our sign calculation is only

where we are solving for some approximate X,
approximate. The Fourier transform can be used to show here that for good enough
approximations of (16), assuming exponential stabilizability and detectability of the
operator system, then X is (exponentially) stabilizing and solves
X A + A T X + Q X R X 0
The discrete time Riccati, and continuous and discrete time Sylvester and Lyapunov
equations and stability checks will work in essentially the same way, and with the
same results. Hence we can extend many finite dimensional computational techniques using the sign iteration, such as stability, and spatiotemporal H2 performance
analysis, and controller synthesis to Lr operators.
3.5 Numerical Difficulties

Of course, a key assumption in the above discussion is that X is a close enough approximation to X so as to be exponentially stabilizing and have good performance.
since our calculation of the matrix sign
There will always be some error X X
in Algorithm 1 will never be taken to k = , but will always be halted after some
threshold of convergence Z k Z k1 < . Also, repeated order-reducing approximations, must be used to prevent the complexity from blowing up [22]. Such iterative approximations if too aggressive, can cause numerical instability in the sign
iteration.
However, in linear systems analysis and control applications, a posteriori closed
loop stability and performance(e.g. H2 , H ), can be relaxed to -sub-optimal problems involving Lyapunov and Riccati inequalities. The verification of solutions to
such problems can thus be reduced to checking the positive definiteness of a Hermitian matrix [16, 22].
Similar arguments can be made to reduce the problem of the a posteriori checking
of the closed loop exponential stability and spatiotemporal H2 norm of an Lr system
to the checking of the positive definiteness of a symmetric Lr operator, which in turn
can be checked by solving a Riccati equation, as follows:
Lemma 11.1. Let A be a stable matrix and D = DT . The following statements are
equivalent:
1. X = Lr {C, A, B, D, B , A ,C } 0
2. D + C(zI A)1B + BT (z I AT )1CT 0, z T
3. P 0, unique, such that: P = APA + (B APC)(D CPC )1 (B APC )
and D CPC 0 and (A KC) < 1 where K = (B APC)(D CPC )1
288
J. Rice et al.
Proof. 1 2 is due to the isomorphism between an Lr operator and its symbol, and
2 3 is an application of the Positive Real Lemma ([23], Lemma 8.C.2)
In summary, while the sign iteration for solving Laurent operator Riccati and
Lyapunov equations might have numerical difficulties, checking to see if a solution satisfies a Laurent operator Riccati or Lyapunov inequality can be reduced to
checking the positive realness of an operator, which is a simple calculation (a finite
dimensional Riccati equation) as in Lemma 11.1 statement 3. Hence a posteriori
stability and performance analysis is numerically feasible.
3.6 Application to the Example Problem

In order to apply these methods to the platoon example, we thus simply need to
write it as a set of interconnected subsystems as in (12):
0 1 0 0 0 0
0 0 0 0 q1 gs
s
1 0 0 0 0 0
A B p B m B1 B2
C p W p Zm Lp V p
1 0 0 0 0 0
m p m m m
C Z W L V = 0 0 0 0 0 1
(18)
1 p m 11 12
C J J D D f 2 0 fs2 fs2 0 0
2 2
1s
C2 H p H m D21 D22
fs 0 0 0 0 0
1 0 0 0 0 0
0 1 0 0 0 0
then resolve the interconnection variables to get the Lr operator system, in which
we can apply our structure preserving iterative techniques. We consider the H2
optimal state feedback problem ([24] 14.8.1) and hence only have to solve one
Riccati equation, which, for our problem, reduces to:
A X + X A + C1C X B 2 B 2 X = 0
(19)
Where we will instead solve for some Lr operator X

with controller K = B 2 X.
< , such that K = B X will have a discrete distributed implesuch that X X
2
mentation as in Fig. 2.
4 Distributed Control Design for Decomposable Systems

The second distributed control method to be discussed relies on the concept of
decomposable systems. A distributed controller synthesis method has been introduced in [25] that applies to systems of finite order with a special structure in the
state-space matrices. We call this class of systems decomposable systems, as the
289
synthesis method exploits a kind of modal decomposition that is made possible by

this structure. LMIs are used in order to impose the same structure on the controller
as well.
In this section we are first going to explain the idea at the base of the method,
and then we will show how it can be extended to infinite dimensional systems and
applied to the infinite platoon problem.
4.1 General Description

We start by defining the class of systems to which the method applies.
Definition 11.1 (Decomposable systems). Let us consider the Nnth order linear
dynamical system described by:

x(t)
= Ax(t) + B1w(t) + B2 u(t)
(20)
y(t) = Cx(t) + Du(t)
where x is the state, u RNmu is the control input, w RNmw the disturbance input
and z RNr is the output. We call such systems decomposable systems iff there
exists a diagonalizable pattern matrix P RNN (with P = Z 1 Z, where is
the diagonal matrix containing the eigenvalues) such that the state-space matrices
can be parameterized as:
A
B
C
D
= IN Aa + P Ab
= IN B,a + P B,b
= IN Ca + P Cb
= IN Da + P Db
(21)
If the pattern matrix P is symmetric, then we call the system a symmetric decomposable system.
We then introduce the theorem that is at the base of the synthesis method for
this class of systems, and which explains the name. We focus on symmetric decomposable systems as the pattern matrix will have only real eigenvalues, making the
reasoning simpler.
Theorem 11.1. A symmetric decomposable system of order Nn as described in
Definition 11.1 is equivalent to N independent modal subsystems of order n. Each
of these subsystems has only mu control inputs, mw disturbance inputs and r outputs:

xi (t) = Ai xi (t) + B1,iw i (t) + B2,i ui (t)
zi (t) = Ci xi (t) + Diui (t)
(22)
for i = 1, . . . , N
where for all the matrices in bold font it holds that:
Mi = Ma + i Mb
(23)
290
J. Rice et al.
where the i are the eigenvalues of P and the matrices Ma , Mb are defined as in (21).
Conversely, it is true that all the sets of systems as in (22) for which the parameterization (23) holds are equivalent to a decomposable system.
Proof. The system is decomposed with the following change of variables:
x = (Z In )x
u = (Z Imu )u
w = (Z Imw )w
z = (Z Ir )z
(24)
Every state-space matrix M in (20) is equivalent to a matrix in (22) through the

following relation:
M = IN Ma + P Mb = (Z I)M(Z I)1
(25)
where M is a block diagonal matrix containing all the matrices Mi in the correct order in its diagonal. The idea that is exploited is that if Z diagonalizes P, then (Z I)
block-diagonalizes any matrix parameterized as in (21) thanks to the properties of
the Kronecker product. For the details of the proof see [25].
This last theorem basically says that a decomposable system (that is, a system
for which all the state-space matrices M can be parameterized as M = IN Ma +
P Mb ) is equivalent to a set of N independent systems with state-space matrices
Mi = Ma + i Mb . Conversely, if a set of N independent systems has matrices which
are parameterized as Mi = Ma + i Mb , then they are equivalent to a decomposable
system. This is of fundamental importance, because:
1. For a decomposable system, it is possible to simplify control problems by evaluating them in the framework of the N independent systems, which are of much
smaller order.
2. If the state-space matrices of the controllers of the N independent systems are
parameterized as in (23) (with the same P), then the controller in its untransformed form will be a decomposable system with the same sparsity as the plant.
From now on, we will always use the bold font to identify matrices which can be
parameterized according to (23).
As just seen, as consequence of Theorem 11.1, for decomposable systems control problems can be approached in the domain of the transformed variables,
where the system is equivalent to a set of smaller independent modal subsystems.
Once the solution has been obtained independently for each subsystem, one can retrieve the solution to the original problems through the inverse of the transformation
shown in the proof of Theorem 11.1. Notice that the fact of working with a decomposed system does not imply that the final controller will be distributed or sparse,
or decomposable as well; for this purpose, additional care will be needed, which we
will discuss next.
For example, let us now consider the problem of finding a stabilizing static state
feedback for the system in (20). The basic LMI approach for solving the problem is
to find a feasible solution to the following inequalities [26]:
X 0
AX + XAT + B2 L + LT BT2 0
291
(26)
where X = X T and L are decision variables; the static state feedback is given by the
expression K = LX 1 . In the transformed domain, the LMI above is equivalent to
the following set of smaller independent LMIs:
)
Xi 0
(Aa + i Ab )X + X(Aa + i Ab )T + (B2,a + i B2,b )Li + LTi (B2,a + i B2,b )T 0
for i = 1, . . . , N
(27)
where now Xi = XiT and Li are decision variables. If we just solve each of the N
LMIs independently, then there will be a gain Ki = Li Xi1 for each subsystem; but
if we stack all these gains in a block diagonal matrix K and perform the inverse
transformation shown in (25), that is:
I)1
K = (Z I)K(Z
(28)
then this K will not be the matrix of a decomposable system as in (21), because K
is not parameterized according to (23) (it is not bold).
This problem can be solved by introducing a constraint in the LMI optimization
that will force the Ki to have the structure that we want. This can be obtained by
introducing the following coupling constraints to (27):
Xi = X
for i = 1, . . . , N
Li = La + i Lb
and thus, the gains Ki will be parameterized according to (23):
Ki = (La + i Lb )X 1 = Ka + i Kb = Ki
thus yielding a K that is the matrix of a decomposable system, which means that
the final controller will have a sparse structure described by the pattern matrix; in
fact, K = In Ka + P Kb . This approach is similar to the so-called multiobjective
optimization [27]; some conservatism is introduced because we have set the same
X matrix for all the LMIs. Since X is associated to the Lyapunov function of the
closed loop system, this method is also called Lyapunov shaping. The approach that
has been used here for finding a stabilizing feedback can be extended to a wider
range of problems. For example, it can be used to approach the H2 problem. The
optimal state feedback law of the kind u = Kx for the system:
)
x(t)
= Ax(t) + B1 w(t) + B2 u(t)
(29)
z(t) = Cx(t) + Du(t)
292
J. Rice et al.
minimizing the H2 norm from w to z can be obtained by solving the following

problem:
T
XA + AX + LT BT2 + B2L + B1 BT1 0
!
minimize
X XCT + LT DT
( )
0
(30)
W

under:
trace(W ) < 2
The application of this formula to a decomposable system and the introduction of
the necessary constraints lead to the following theorem.
Theorem 11.2. Consider a symmetric decomposable system (Definition 11.1).
There exists a decomposable state feedback law of the kind: u = (IN Ka + P
Kb )x that yields an H2 norm from w to z smaller than if the following LMI set (for
i = 1 . . . N) is feasible:
XATi + Ai X + LTi BT2,i + B2,i Li + B1,iBT1,i 0
X XCTi + LTi DTi

0
Wi
i trace(Wi ) < 2
(31)
where x0 is the initial state at time t = 0, Ai = Aa + i Ab , Wi = Wa + iWb etc.;

X = X T , Wa = WaT , Wb = WbT , La and Lb are the decision variables. The controller
is retrieved by the relations: Ka = La X 1 , Kb = Lb X 1 . Notice that the theorem
expresses a suboptimality result (it is only a sufficient condition) due to the presence
of constraints in the LMI.

4.2.1 Generalization to Infinite Dimensional Systems
The synthesis method in Theorem 11.2 applies to systems of finite order, but it
can be easily generalized to systems of infinite dimension like the infinite platoon
example. This is possible by replacing the pattern matrix with a bounded Laurent
operator (for which we use the notation P), and by replacing the standard H2 norm
with the spatiotemporal one introduced in [7]. In this way the same results still
hold formally, with the difference that the eigenvalues i of the pattern matrix are
replaced by the real spectrum (P) of the pattern operator, which we consider as
a function of z = e j (with [0, 2 ]). With this reasoning, the H2 suboptimal
problem would be solved by the following set of LMIs:
X AT + A X + LT BT2, + B2, L + B1, BT1, 0
X XCT + LT DT
0
for [0, 2 ]
2
j
0 trace(Wa + (e )Wb )d < 2
293
(32)
where now A = Aa + (e j )Ab , W = Wa + (e j )Wb etc.

The problem consists of an infinite number of LMIs, or better, a set of parametervarying LMIs depending either on the parameter or on the real spectrum (e j ).
In the case of Bb = 0 and Db = 0, the LMIs are all affine in the parameter , so
the convexity property of the LMIs allows the replacement of the parameter varying
inequalities with only those for the extreme (maximum and minimum) values of
the spectrum, reducing the problem from infinite to finite complexity. The result is
summarized in the following theorem.
Theorem 11.3. Consider an infinite dimensional decomposable system
(Definition 11.1), with a symmetric Laurent operator P as pattern, and with Bb = 0
and Db = 0. There exists a decomposable state feedback law of the kind: u =
(I Ka + P Kb )x that yields a spatiotemporal H2 norm from w to z smaller than
if the following LMI set is feasible:
X(Aa + Ab )T + (Aa + Ab )X + (La + Lb )T BT2,a +
+B2,a(La + Lb ) + (B1,a + B1,b )(B1,a + B1,b )T 0
X X(Ca + Cb )T +(La + Lb )T DTa
0
Wa + Wb )

X (Aa + Ab )T + (Aa + Ab )X + (La + Lb )T BT2,a +
+B2,a(La + Lb ) + (B1,a + B1,b )(B1,a + B1,b )T 0
X X(Ca + Cb )T +(La + Lb )T DTa
0
Wa + Wb )
trace(Wa +avWb ) < 22
(33)
where , and av are respectively the maximum, minimum and average value of
(av = 02 (e j )d /2 ). The decision variables are X = X T, Wa = WaT, Wb = WbT ,
La and Lb , and the controller is retrieved by the relations: Ka = La X 1 , Kb = Lb X 1 .
4.2.2 The Platoon

The tools of Theorem 11.3 can be applied directly to the platoon example, for which
we will have:
294
J. Rice et al.
Aa =
01
0
0
, B1,a =
, B2,a =
,
00
q
g
0 0
0 0
1
Ca = 0 0 , Cb = f2 0 , Da = 0
f1 0
0
0 0
(34)
and the other matrices are zero. The pattern is given by:
1
P=
2
.. .. ..
. . .
1 2 1
1 2 1
1 2 1
.. .. ..
. . .
(35)
It is easy to show that the real spectrum of this pattern operator is in the interval
[2 0], with 1 as average value. The operator that diagonalizes P is again simply
the Fourier operator F, which we can use to compute the symbol of the operator as
a function of z T:
1
1
FPF 1 = z1 1 + z = cos 1 [2 0]
2
2
(36)
and as a consequence, av = 02 (cos 1)d /2 = 1.

We remark that the extension of the results of [25] to infinite dimensional, spatially invariant systems, brings to results that are quite close to the ones shown
in the earlier paper [28]. In this reference, the reduction of the number of LMIs
from infinite to finite is done by means of application of the discrete-time Kalman
YakubovichPopov lemma [29]. In this article instead we have shown that the reduction can be done just by exploitation of the convexity properties of the LMIs,
provided that the pattern operator is symmetric (and thus it has a real spectrum).
5 Distributed LQR of Identical Systems

If stability and design simplicity of the distributed control system are more important than optimality, then the complexity of the control design task can be further
reduced. In this section we provide a conceptually simple way of generating stabilizing structured controllers for the special class of systems under consideration,
using the solution of small-sized local LQR problems. These results rely on robustness and other special properties of LQR solutions for identical dynamically
decoupled systems. They enable the construction of structured stabilizing, suboptimal controllers for finite-dimensional systems as presented in [30]. In Sect. 5.1 we
provide an overview of these properties and then extend their applicability to the
infinite-dimensional string system under study in Sect. 5.2.
295
5.1 Special Properties of LQR for Dynamically Decoupled Systems

Reference [30] studies distributed LQR design for identical dynamically decoupled
systems and proposes a structured controller design procedure based on local stabilizing LQR controllers. A similar problem setup with a special class of symmetrically interconnected systems is treated in [31]. In the following, we summarize the
main properties of the local LQR solutions in [30].
Consider a set of NL identical, decoupled linear time-invariant dynamical systems, the ith system being described by the continuous-time state equation:
xi = Axi + Bui,
xi (0) = xi0 .
(37)
where xi (t) Rn , ui (t) Rm are states and inputs of the ith system at time t, respectively. Let x(t)
RnNL and u(t)
RmNL be the vectors which collect the states and
inputs of the NL systems at time t:
x = A x + B u,
x(0)
= x0 [x10 , . . . , xNL 0 ]T ,
(38)
with
A = INL A,
B = INL B.
(39)
We consider an LQR control problem for the set of NL systems where the cost
function couples the dynamic behavior of individual systems. This cost function
contains terms which weigh the ith system states and inputs, as well as the difference
between the ith and the jth system states and can be written using the following
compact notation:
J(u(t),
x0 ) =

0

) + u(
)T Ru(
x(
)T Q x(
) d ,
(40)
where the matrices Q and R have a special structure defined next. Q and R can be
decomposed into NL2 blocks of dimension n n and m m respectively:
R 0 0
Q11 Q 12 Q 1NL
. .
..
..
.. . . .. ..
, R =
.
.
Q =
(41)
.
.
.
.
. .
.
.
Q N 1 . . . . . . Q NL NL
0 ... ... R
L
296
J. Rice et al.
with
Q ii = Qii +
NL
Qik , i {1, . . ., NL }.
k=1, k=i
Q i j = Qi j , i, j {1, . . . , NL }, i = j.
R = INL R.
(42)
where
Rii = RTii = R > 0, Qii = QTii 0 i,
Qi j = QTij = Q ji 0 i = j.
(43a)
(43b)
Using the cost function defined above, let K and xT0 P x0 be the optimal controller
and the value function corresponding to the following LQR problem:
min J(u,
x0 )
u
subj. to x = A x + Bu
x(0)
= x0
(44)
We will assume that a stabilizing solution to the LQR problem (44) with finite performance index exists and is unique (see [32], p. 52 and references therein).
It is well known that
K = R 1 B TP,
(45)
where P is the symmetric positive definite solution to the following ARE:
A T P + P A P B R 1 B T P + Q = 0
(46)
Theorem 11.4. [30] Assume the weighting matrices (42) of the LQR problem (44)
are chosen as
Qii = Q1 i {1, . . . , NL }
Qi j = Q2 i, j {1, . . . , NL }, i = j.
(47)
Let xT0 P x0 be the value function of the LQR problem (44) with weights (47), and
1)n : in, ( j 1)n : jn] with
the blocks of the matrix P be denoted by Pi j = P[(i
i, j {1, . . . , NL }.
Then,
L
(I) Nj=1
Pi j = P for all i {1, . . . , NL }, where P is the symmetric positive definite
solution of the ARE associated with a single node local problem:
AT P + PA PBR1BT P + Q1 = 0.
(48)
297
L
K i j = K for all i {1, . . . , NL }, where K = R1BT P.
(II) Nj=1
(III) Pi j = Plm P2 i = j, l = m is a symmetric negative semidefinite matrix.
Proof. See in [30].

The following corollary of Theorem 11.4 follows from the stability and the robustness of the LQR controller R1 BT (NL P2 ) for system A XP as shown in
[30].
Corollary 11.1. A XP + NL X P2 is a Hurwitz matrix for all > 12 , and = 0.
This corollary can be relaxed to any 0 for a wide class of systems as discussed
in [30].
In Sect. 5.2 we show how to construct a distributed stabilizing suboptimal controller for the infinite-dimensional application example at hand using the properties
described above.

As a key ingredient to the extension of the above results to infinite-dimensional
systems, we will establish the following lemma, which forms a basis for the results
described in this section.
Lemma 11.2. Given A,C Rmm and a Laurent operator B with scalar symbol,
bounded on l 2 (Z), consider two infinite-dimensional Laurent operators A = I A
and C = B C, where I stands for the infinite-dimensional matrix with ones on its
diagonal and zeros elsewhere. Then the spectrum of A + C can be characterized as:
= (A + (B)C),
(A + C)
(49)
is the spectrum of B.
where (B)
Proof. Let us denote the infinite-dimensional Laurent matrix under investigation as
and its symbol
X = A + C,
FX F 1 = X (z) = A + B(z)C = A + (B(z))C,
z = e j T
(50)
since B(z) = (B(z)) is a scalar, where denotes the spectrum and T is the unit
circle.
Since there is an equivalence in terms of spectrum between the Laurent operator
and its symbol [13], we can express the spectrum of X in the following way:
(X(T)) = (A + (B(T))C)
%
(X) = (A + (B)C)
(51)
This proves the lemma, which can be readily extended to block Laurent matrices
based on [13].
298
J. Rice et al.
This result can be interpreted as an extension of Proposition 2 in [30] to the case

where the identity matrix and B form infinite-dimensional pattern matrices. The
following theorem will propose a procedure to construct stabilizing structured controllers for infinite-dimensional systems using the local LQR controller as building
blocks.
Theorem 11.5. Consider the LQR problem (44) with NL and weights chosen as in
(47) and its solution K = 1 K2 INL K2 + INL K1 and K = 1 P2 + INL (P
NL P2 ). (The matrix 1 RNL NL has ones in all its entries.)
Let M be a bounded Laurent operator whose spectrum satisfies the following
property1 :
NL
.
(52)
(M) >
2
Construct a state feedback controller as the following Laurent matrix:
K = I R1BT P + M R1BT P2 .
(53)
Then, the closed loop system

A cl = I A + (I B)K
(54)
is asymptotically stable.
Proof. Consider the spectrum of the closed-loop system matrix A cl :

(X P2)
(A cl ) = I (A XP) + M
By Proposition 11.2:

(X P2) = A XP + (M)X
P2 .
I (A XP) + M
(55)
From Corollary 11.1 and from condition (52), we conclude that A XP +

P2 is Hurwitz for all (M)
thus the closed-loop system is asymptotically
(M)X
stable.
Theorem 11.5 has several main consequences:
1. The controller K in (53) will have the same sparsity structure as the Laurent
which leads to an asymptotically stable distributed controller, if M
operator M,
is sparse, banded or otherwise structured.
2. We can use one local LQR controller to compose distributed stabilizing controllers for an infinite collection of identical dynamically decoupled subsystems.
3. The first two consequences imply that we can not only find a stabilizing
distributed controller with a desired sparsity pattern (which is in general a
formidable task by itself), but it is enough to solve a small finite-dimensional
problem. This attractive feature relies on the specific problem structure and
LQR robustness properties defined in Sect. 5.1.
1
0 for a wide class of systems.

This property can be relaxed to (M)
299
4. The result is independent of the local LQR tuning. Thus Q1 , Q2 and R in (47)
can be used in order to influence the compromise between minimization of absolute and relative terms, and the control effort in the global performance.
For the platoon example, the following weighting matrices were used to solve a
3-by-3 local LQR problem (NL = 3):
'
' 2 (
(
f2
f12 0
Q1 =
, Q2 = 4 0 , R = 1.
0 0
0 0
The stabilizing controller having the same banded structure as the cost function is
built in the following way. Using real parameters b 0 and a, construct the global
controller gain matrix K as

.. .. .. .. ..
.. .. ..
.
.
.
.
.
.
.
.

(56)
K = 0 K1 0 + 0 bK2 aK2 bK2 0

.. .. ..
.. .. .. .. ..
. . .
.
.
. . .
where the gain matrices K1 and K2 are obtained from the local LQR solution as
explained in [30]. The closed-loop state matrix can be written as
A cl = I A + (I B)K.
(57)
The controller K has R1 BT (P1 aP2 ) on its diagonal entries and bR1BT P2 on
its two sub-diagonal entries. Using Theorem 11.4, K can be written as
K = I (R1BT P) + M (R1 BT P2 ).
which defines the pattern of the distributed controller has
The Laurent operator M,
NL 1 a = 2 a on its diagonal and b on its two subdiagonals. This means that
the spectrum of the symbol of M on the unit circle is 2 a be j be j . Thus M
has a purely real spectrum with elements belonging to the interval [2 a 2b, 2
a + 2b]. Following similar arguments as in [30], this leads to the sufficient stability
condition of a + 2b 12 (or a + 2b 2 for a large class of systems as explained in the
above reference). For the specific system under consideration the values of a = 1.98
and b = 0.01 were chosen.
6 Numerical Results of the Car Platoon Benchmark Problem

This section presents the numerical results of the application of the different controllers to the benchmark problem. We computed the controllers for different values
of the weighting parameter f2 while keeping f 1 = 1. A larger value of f2 indicates
300
J. Rice et al.
1.4
Centralized Optimal Controller
Truncated Optimal Controller
Decomposable System Approach
1.35
Distributed LQR
H2 norm
1.3
1.25
1.2
1.15
0.1
0.2
0.3
0.4
0.5
f2
0.6
0.7
0.8
0.9
Fig. 3 Closed-loop performance of the four different controllers
an increased penalty on the relative positions of the vehicles in the platoon, while
for f2 0 the problem turns into the control of each single vehicle independently
from the others.
The results are depicted in Fig. 3, which shows the spatiotemporal H2 norm
achieved with:
1. The optimal controller from Sect. 2.1 (identical in performance to the controller
produced with the method of Sect. 3, to within 1015);
2. The truncated version of this global controller, with only one off-diagonal band;
3. The decomposable system approach described in Sect. 4;
4. The distributed LQR controller of Sect. 5.
The centralized -suboptimal controller yields the best performance, while its
truncated version is slightly better than the controller given by the decomposition
approach, followed by the distributed LQR method, as could be expected since
performance is not explicitly considered in that technique. Note as the coupling
weight becomes small ( f2 0), the four controllers tend to converge to the same
performance.
Figure 4 illustrates the magnitude of the entries in the centralized optimal controller gain, providing some justification for the use of its truncated versions.
301
0
10
20
10
30
15
40
20
50
25
60
30
70
35
80
40
90
45
100
50
100
150
200
Fig. 4 A visualization of a section around the diagonal of the optimal centralized gain operator K.
The colors represent log10 () of the magnitude of the entries of the matrix. The exponential spatial
decay of the operator is apparent
7 Conclusions and Open Problems

We have presented three conceptually quite different approaches for distributed
control design for large-scale systems. They all utilized special properties of the
problem structure in order to arrive at various degrees of simplification for the largescale control design problem.
Although they are applicable to wider classes of problems (as stated in the corresponding sections), the three approaches were applied for solving a benchmark control example taken from the literature involving an infinite platoon of vehicles. The
resulting performance was compared with an -suboptimal centralized controller.
Despite the progress that can be observed in the development of various distributed control techniques for special classes of large-scale systems, several important challenges remain to be solved in this field. Below we list a few of them
pertaining to the three approaches discussed in this work.
Although the iterative, Laurent operator structure preserving approach in Sect. 3
provides an attractive numerical solution for the class of systems under study, it does
leave the question of designing optimal controllers with a specified width of their
banded structure open. Optimal design for a fixed rational order remains a challenging problem as well. It would be of practical interest also to study suboptimal
control design for which the sign iteration converges in only a few iterations, e.g.,
by specifying a certain region of closed-loop spectrum.
302
J. Rice et al.
Using the decomposition-based method of Sect. 4, we have seen that conservatism is introduced by assuming a common Lyapunov matrix for all the modal subsystems. An interesting research question would be whether it is possible to reduce
this conservatism by employing robust control methods other than multi-objective
optimization.
Although the local LQR solutions of Sect. 5 provide a very simple way of obtaining stabilizing solutions for a special class of systems under a wide variety of interconnections, obtaining near-optimal or guaranteed suboptimal controllers with this
procedure remains an important open challenge. Addressing dynamically coupled
systems has been attempted only in [31] for special symmetric interconnections.
The powerful robustness properties of local LQR solutions could also allow the
scheme to work with time-varying interconnections and a wider variety of weights
in the cost function.
References
1. N. R. Sandell Jr., P. Varaiya, M. Athans, and M. G. Safonov. Survey of decentralized control
methods for large scale systems. IEEE Trans. Autom. Contr., 23:108129, 1978
2. R. DAndrea and G. E. Dullerud. Distributed control design for spatially interconnected systems. IEEE Trans. Autom. Contr., 48:14781495, 2003
3. Roger W. Brockett and Jacques L. Willems. Discretized partial differential equations: Examples of control systems defined on modules. Automatica, 10:507515, 1974
4. Morten Hovd, Richard D. Braatz, and Sigurd Skogestad. SVD controllers for H2 -, H - and
-optimal control. Automatica, 33:433439, 1997
5. Jeremy G. VanAntwerp, Andrew P. Featherstone, and Richard D. Braatz. Robust crossdirectional control of large scale sheet and film processes. J. Process Contr., 11:149177,
2001
6. M. R. Jovanovic and B. Bamieh. On the ill-posedness of certain vehicular platoon control
problems. IEEE Trans. Autom. Contr., 50(9):13071321, September 2005
7. B. Bamieh, F. Paganini, and M. A. Dahleh. Distributed control of spatially invariant systems.
IEEE Trans. Autom. Contr., 47(7):10911107, 2002
8. E. W. Kamen. Multidimensional systems Theory: Stabilization of Linear Spatially-Distributed
Continuous-Time and Discrete-Time Systems, pages 101146 Norwell, MA: Kluwer, 1985
9. B. Bamieh, F. Paganini, and M. A. Dahleh. Distributed control of spatially invariant systems.
IEEE Trans. Autom. Contr., 47:10911106, 2002
10. G. A. de Castro and F. Paganini. Convex synthesis of localized controllers for spatially invariant systems. Automatica, 38:445456, 2002
11. Rufus Fraanje. Private communication
12. A. Bottcher and B. Silberman. Introduction to Large Truncated Toeplitz Matrices. Springer,
Berlin, 1991
13. I. Gohberg, S. Goldberg, and M. Kaashoek. Classes of Linear Operators Vol. II. Birkhauser,
Basel, 1993
14. R. Curtain, O. Iftime, and H. Zwart. System theoretic properties of platoon-type systems.
Proceedings of the IEEE Conference on Decision and Control, 2008
15. S. Boyd and V. Balakrishnan. A regularity result for the singular values of a transfer matrix
and a quadratically convergent algorithm for computing its L norm. Syst. Contr. Lett., 15:17,
1990
16. J. K. Rice and M. Verhaegen. Distributed control: A sequentially semi-separable approach for
heterogeneous linear systems. IEEE Trans. Autom. Contr., 54:12701283, 2009
303
17. J. Rice and M. Verhaegen. Distributed control of spatially invariant systems using fast iterative
solutions to rationally parametric matrix problems. Proceedings of IEEE Conference Decision
and Control, 2009
18. J. D. Roberts. Linear model reduction and solution of the algebraic Riccati equation by use of
the sign function. Int. J. Contr., 32:677687, 1980
19. C. S. Kenney and A. J. Laub. The matrix sign function. IEEE Trans. Autom. Contr., 40:
13301348, 1995
20. L. Grasedyck, W. Hackbusch, and B. N. Khoromskij. Solution of large scale algebraic matrix
Riccati equations by use of hierarchical matrices. Computing, 70:121165, 2003
21. C. Kenney, A. J. Laub, and E.A. Jonckheere. Positive and negative solutions of dual Riccati
equations by matrix sign function iteration. Syst. Contr. Lett., 13:109116, 1989
22. J. K. Rice and M. Verhaegen. Distributed control: A sequentially semi-separable approach.
Proceedings of IEEE Conference Decision and Control, 2008
23. T. Kailath, A. H. Sayed, and B. Hassibi. Linear Estimation. Prentice-Hall, Englewood Cliffs,
2000
24. K. Zhou, K. Glover, and J. C. Doyle. Robust and Optimal Control. Prentice-Hall, Englewood
Cliffs, 1996
25. P. Massioni and M. Verhaegen. Distributed control for identical dynamically coupled systems:
a decomposition approach. IEEE Trans. Autom. Contr., 54(1):124135, January 2009
26. C. Scherer and S. Weiland. Linear matrix inequalities in control. Online, http://www.dcsc.
tudelft.nl/cscherer/2416/lmi05.pdf, 2005, Lecture notes
27. C. Scherer, P. Gahinet, and M. Chilali. Multiobjective output-feedback control via LMI optimization. IEEE Trans. Autom. Contr., 42(7):896911, 1997
28. G.A. de Castro and F. Paganini. Convex synthesis of localized controllers for spatially invariant systems. Automatica, 38(3):445456, 2002
29. A. Rantzer. On the KalmanYakubovichPopov lemma. Syst. Contr. Lett., 28(1):710, 1996
30. F. Borrelli and T. Keviczky. Distributed LQR design for identical dynamically decoupled systems. IEEE Trans. Autom. Contr., 53(8):19011912, August 2008
31. M. K. Sundareshan and R. M. Elbanna. Qualitative analysis and decentralized controller synthesis for a class of large-scale systems with symmetrically interconnected subsystems. Automatica, 27(2):383388, 1991
32. B. D. O. Andreson and J. B. Moore. Optimal Control: Linear Quadratic Methods. PrenticeHall, Englewood Cliffs, 1990
Integrated Design of Large-Scale Collocated

Structural System and Control Parameters
Using a Norm Upper Bound Approach
Mona Meisami-Azad, Javad Mohammadpour Velni, Kazuhiko Hiramoto,
and Karolos M. Grigoriadis
1 Introduction
The traditional design optimization of system parameters and a feedback controller
usually follows a sequential strategy, where the open-loop system parameters are optimized first, followed by the controller design. However, this design strategy will
not lead to an optimal closed-loop performance due to the coupled nature of the plant
and controller optimization problems. Specifically, this two-step design methodology does not utilize the full design freedom to achieve an optimal overall system. It
has been shown that the overall system performance can be significantly improved
if the design process of the plant and the control system is integrated [17,18,2830].
The integrated design strategy corresponds to a simultaneous optimization of the design parameters of both the plant and the controller to satisfy desired set of design
specifications and to optimize the closed-loop system performance. Past research
work has demonstrated the improved performance achieved using the integrated design strategy compared to the sequential method of design. However, the general
integrated plant/controller design optimization problem turns out to be a complex
nonlinear nonconvex optimization problem and no solution is guaranteed to provide
convergence to the global optimum of the design variables [1, 13, 21, 28, 32, 33].
This makes the integrated design strategy computationally challenging. Recently,
several integrated H -based plant/controller design approaches have been proposed
using a Linear Matrix Inequality (LMI) formulation of the control problem to take
advantage of systematic approaches developed in the past two decades for robust
control design [12, 20, 27]. In principle, these formulations result in Bilinear Matrix
M. Meisami-Azad, J.M. Velni, and K.M. Grigoriadis
Department of Mechanical Engineering, University of Houston, Houston, TX 77204, USA
e-mail: mmeisami@mail.uh.edu; jmohammadpour@uh.edu; karolos@uh.edu
K. Hiramoto
Department of Mechanical and Production Engineering, Niigata University,
8050 Ikarashi-2-no-cho, Nishi-ku, Niigata 950-2181, Japan
e-mail: hiramoto@eng.niigata-u.ac.jp
305
306
M. Meisami-Azad et al.
Inequality (BMI) problems even if one assumes that the coefficient matrices of the
plant state-space representation are linear functions of the system design parameters.
In past attempts, the BMI formulation of the integrated plant/controller design has
been solved using iterative LMI-based optimization schemes. However, these iterative methods are also unable to guarantee convergence to the optimum solution, and
they are computationally intensive. Yang and Lee [38] proposed an analytical model
for structural control optimization in which LQR feedback control parameters along
with the non-collocated sensor/actuator placement were considered as independent
design variables. Iwasaki et al. [15] proposed a design framework based on a finite
frequency positive realness property and formulated the corresponding design condition in terms of LMIs. And recently Hiramoto and Grigoriadis [14] proposed an
integrated design methodology using a homotopy approach to simultaneously optimize structural system and control design parameters. The approach taken in the
latter work is an iterative one which introduces a small amount of perturbation in
the coefficient matrices of the state-space representation of the plant and the matrices associated with the controller. They show that in case of small perturbations, the
optimization problem can be cast as an LMI. However, the general integrated design
problem is in form of a BMI and the number of iterations for the convergence of the
algorithm significantly increases due to the linear approximations of the nonlinear
problem.
The control design problem for structural systems with collocated sensors and
actuators has been shown to provide great advantages from a stability, passivity, robustness and an implementation perspective [16]. For example, collocated control
can easily be achieved in a space structure when an altitude rate sensor is placed at
the same location as a torque actuator [4, 11]. Collocation of sensors and actuators
leads to externally symmetric transfer functions [5]. Several other classes of engineering systems, such as circuit systems, chemical reactors and power networks, can
be modeled as systems with symmetric transfer functions [2].
In this chapter, we present an efficient and computationally tractable design
method in order to integrate the structural parameters and control gain design in
collocated structural systems using H2 or H norm closed-loop performance specifications. Specifically, the objective is to determine the optimal values of the damping parameters of the structural system and to simultaneously design optimal output
feedback gains such that an upper-bound on the closed-loop system norm (either
in the H2 or the H setting) from the disturbance input signals to the desired outputs is minimized. The theoretical development of this chapter takes advantage of
recently developed control-oriented algebraic tools to formulate the simultaneous
damping and control gain design problem as an LMI optimization problem. LMI
optimization problems involve the minimization of a linear objective function subject to matrix inequality constraints that are linear with respect to the parameter
variables [6, 7]. An LMI optimization problem is essentially a generalization of the
linear programming (LP) problem to the matrix case with positive definite matrix
inequality constraints. The LMI problems are convex and can be efficiently solved
using interior-point optimization solvers [10, 26].
Integrated Design of Large-Scale Structural System and Control Parameters
307
A fundamental idea behind the work presented in this chapter is to demonstrate

how to take advantage of the particular structure of collocated structural systems
to develop explicit upper bound expressions for the H2 and H norms of such systems. Following this idea, we demonstrate the formulation of the integrated damping
parameters and output feedback control gain design as a convex optimization problem with respect to the unknown damping and control variables subject to practical
constraints on the damping parameters and the feedback control gains. Hence, unlike past approaches, the proposed method in this chapter for integrated design is
not based on an iterative procedure to determine the design parameters. The benefits
of the proposed integrated design become apparent for large-scale structural systems, where the use of iterative methods for integrated design becomes intractable
or even fails.
The notation used throughout this chapter is standard. The notation > (<) is used
to denote the positive (negative) definite symmetric matrices. The ith eigenvalue
of a real symmetric matrix X will be denoted by i (X) where the ordering of the
eigenvalues is defined as max (X ) = 1 (X) 2 (X) ... n (X). The maximum
singular value of a matrix Y will be denoted by max (Y ), which is also referred to
as its spectral norm ||Y ||.
2 Symmetric Output Feedback Control of Collocated Systems

Consider the following vector second-order representation of a structural system
with collocated actuators and sensors
M q(t)
+ Dq(t)
+ Kq(t) = Fu(t) + Ew(t)
y(t) = F T q(t)
T
z(t) = E q(t)
(1)
where q(t) Rn is the coordinate vector, u(t) Rm is the control input vector,
w(t) Rk is the disturbance vector, y(t) Rm is the measured output vector, and
z(t) Rk is the controlled output vector. The matrices M, D and K are assumed
to be symmetric positive definite representing the structural system mass, damping
and stiffness distribution, respectively. The above finite-dimensional representation
is often encountered in the dynamics of structural systems resulting from a finite element approximation of distributed parameter structural systems [11]. It is noted that
velocity feedback as in (1) is common in the collocated control of structural systems through a velocity sensor, a displacement sensor with a derivative controller
or an accelerometer with an integral controller [5]. In smart structures with piezoelectric sensors velocity feedback can be readily achieved through direct strain rate
feedback [8].
The symmetric static output feedback control problem is to design a symmetric
static feedback gain G such that the output feedback control law
u(t) = Gy(t)
(2)
308
renders the closed-loop system stable with appropriate closed-loop performance.

The closed-loop system has a state-space realization as follows
x = Ax + Bw
z = Cx
(3)
with
!
0
I
A=
M1 K M 1 (D + FGF T )
!

0
, C = 0 ET
B=
M 1 E
(4)
where xT = [qT qT ]. The system (3)(4) is an externally symmetric state-space

realization since its transfer function is symmetric. This implies that there exists a
nonsingular matrix T such that
AT T = TA, CT = T B
(5)
This class of systems is more general than the class of internally or state-space
symmetric systems that satisfy the symmetry conditions (5) with a positive definite
transformation matrix T . An analytical solution to the H control problem for internally symmetric systems has been presented by Tan and Grigoriadis [37].
3 Upper Bounds on Collocated Structural System Norms

Recall that the H norm of the system (3) is given by
||H|| = sup max {H( j )}
R
(6)
where H(s) is the transfer function of the system [39]. In a time-domain interpretation, the H norm corresponds to the energy (or L2 norm) gain of the system from
the input w to the output z. Hence, in this setting the H norm defines a disturbance
rejection property of the system.
The collocated H control synthesis problem is to design a symmetric control
law (2) that stabilizes the closed-loop system and guarantees an H norm less than
a prescribed bound > 0. The following result shows that for an open-loop vector
second-order realization (3)-(4) (i.e., with G = 0), an upper bound on its H norm
can be computed using a simple explicit formula [3].
Theorem 1 Consider the open-loop (u(t) = 0) vector second-order system realization (1). The system has an H norm from w(t) to z(t) that satisfies
< = max (E T D1 E)
(7)
309
The explicit bound of the above result is obtained using the Bounded Real Lemma
(BRL) characterization of the H norm of an LTI system presented in the following
lemma [6].
Lemma 1 A stable system with a state-space realization (3) has an H norm from
input w to output z less than or equal to if and only if there exists a matrix P 0
satisfying
T
A P + PA PB CT
B T P I 0 0
(8)
C
0 I
Employing the BRL condition and the choice of a block diagonal matrix for
the Lyapunov matrix P as a particular solution results in the bound of Theorem 1.
Numerical examples demonstrate the validity and computational efficiency of the
above analytical bound [3].
The H2 norm of a stable continuous-time system with transfer function H(s) =
C(sI A)1 B is defined as the root-mean-square (rms) of its impulse response [39],
or equivalently
9
1
H2 =
trace(H H ( j )H( j ))d
2
In the following, an LMI formulation for computing the H2 norm of a system using
its state-space data is recalled. This formulation enables us to use the efficient LMI
solvers to solve for the Lyapunov matrix and compute the H2 norm .
The collocated H2 control synthesis problem is to design a symmetric static feedback gain G such that the output feedback control law (2) stabilizes the closed-loop
system and guarantees an H2 norm less than a prescribed level > 0.
Lemma 2 [34] Suppose that the system (3) is asymptotically stable, and let H(s)
denote its transfer function. Then the following statements are equivalent:
H2
There exist symmetric nonnegative definite matrices P and Z such that
!
PA + AT P PB
0
BT P I
!
P CT
0
C Z
trace(Z) 2
(9)
(10)
(11)
In the following lemma, we recall the H2 norm calculation based on the solution
of a Lyapunov equation.
Lemma 3 [34] The H2 norm of the system (3) is given by
1
H2 = [trace(CPCT )] 2
(12)
310
where P is determined by solving the following Lyapunov equation.

AP + PAT + BBT = 0
(13)
To avoid the need for solving an LMI problem to determine H2 norm of a system
in the collocated form, the following result provides a simple analytical explicit
expression for an upper bound on the H2 norm of such systems [25].
Theorem 2 Consider the open-loop (u(t) = 0) collocated structural system in (1).
This system has an H2 norm from the input w(t) to the output z(t) that satisfies
the following bound
1
1
[max (E T D1 E)] 2
[trace(E T M 1 E)] 2
=
2
(14)
Numerical examples demonstrate the validity and computational efficiency of

the above analytical bound on the H2 norm of collocated systems [25]. MeisamiAzad et al. [25] also developed an explicit parametrization of the suboptimal output
feedback control gains that achieve a desired level of closed-loop H2 performance.
Next, we discuss the problem of computing an explicit expression for the mixed
H2 /H norm of the collocated structural systems. The mixed problem solves the
minimization problem of an H2 norm criterion subject to an H norm constraint.
Presented next is an explicit upper bound on the mixed H2 /H norm of the collocated structural system represented by (1). The following theorem provides such an
explicit bound resulting in a computationally efficient calculation.
Theorem 3 Consider the unforced system in (1), i.e., u(t) = 0. For any given
= max (E T D1 E), an upper bound on the H2 norm of the system, while the H
norm satisfies the condition Hzw , can be computed from the following expression
1
1
[max (F T D1 F)] 2
[trace(F T M 1 F)] 2
=
(15)
2
if 1. Otherwise, for < 1, an upper bound on the H2 norm of the system is
determined from
1
= [max( , ) trace(F T M1 F)] 2
(16)
where
1
max (F T D1 F)
2
max (F T D1 F)
=
.
2 (F T D1 F)]1/2
+ [ 2 max
(17)
The detailed proof of this result can be found in [23]. Meisami-Azad [23] also provides an explicit expression for the output feedback control gain to guarantee the H
and H2 norms of the closed-loop to be less than given bounds and , respectively.
311
4 Integrated Damping and Control Design Using the Analytical

Bound Approach
In this section, we address the integrated design problem of simultaneously designing the damping parameters and the output feedback control gain of the collocated
structural system (1) and (2) to satisfy H norm, H2 norm or mixed H2 /H norm
closed-loop specifications. For lumped parameter systems, the damping matrix D
can be expressed in terms of the elemental damping coefficients as follows
l
D = c i Ti
(18)
i=1
where ci denotes the viscous damping constant of the ith damper and Ti represents
the distribution matrix of the corresponding damper in the structural system. The
distribution matrices Ti are known symmetric matrices with elements 0, 1 and 1
that define the structural connectivity of the damping elements in the structure. Our
objective is to formulate the H2 and H integrated damping parameter and control
gain design problems as LMI optimization problems.
Practical structural system design specifications impose upper bound constraints
on the values of the damping coefficients, that is
0 ci ci max , i = 1, . . . , l
(19)
Also, often an upper bound on the total available damping resources is enforced,
that is
l
ci ccap .
(20)
i=1
Another useful constraint in the proposed integrated design method is a bound on the
norm of the feedback gain matrix. This restriction is placed to limit the amount of
control effort required by the controller. For this purpose, we include the following
constraint in the integrated design problem.
G gbound
(21)
We assume that ci max , ccap and gbound are given scalar bounds determined by the
physical constraints of the design problem.
4.1 Integrated Design Based on an H Specification

Using the above formulation, the solution of the integrated design of the damping
parameters and the symmetric output feedback controller to satisfy closed-loop H
specifications is obtained by the following result.
312
Theorem 4 Consider the collocated structural system (1) with the damping distribution (18). For a given positive scalar , the H norm of the closed-loop system
of the collocated structural system (1) and the output feedback controller (2) is less
than if the following matrix inequalities with respect to the controller gain G and
the damping coefficients ci are feasible.
l
T E
c
T
+
FGF
i i
i=1
0
(22a)
T
E
I
0 ci ci max , i = 1, . . . , l
(22b)
ci ccap
(22c)
G gbound
(22d)
i=1
Proof. Consider the closed-loop interconnection of the collocated structural system

(1) and the output feedback law (2). We consider the Lyapunov matrix P as
!
K 0
(23)
P=
0 M
along with substituting the closed-loop system matrices (4) into the BRL condition
(8) and using the Schur complement formula [34] results in the following inequality
!
D + FGF T E
0
(24)
ET
I
Substitution of the damping matrix expansion (18) results in the inequality (22a).
The constraints (22b)(22d) represent the physical constraints of the design problem
as discussed earlier.
4.2 Integrated Design Based on an H2 Specification

Based on the proposed H2 upper bound results, the integrated design of the damping
parameters of a collocated structural system and the symmetric output feedback
control law to satisfy closed-loop H2 norm specifications is formulated as follows.
Theorem 5 Consider the collocated structural system (1) with the damping distribution (18). For a given positive scalar , the H2 norm of the closed-loop system
of the collocated structural system (1) and the output feedback controller (2) is less
313
than if the following matrix inequalities with respect to the controller gain G, the
damping coefficients ci , and the positive scalar are feasible.
l
2( ci Ti + FGF T ) + FF T 0
(25a)
i=1
!
M F
0
FT Z
(25b)
trace(Z) 2
(25c)
0 ci ci max , i = 1, . . . , l
(25d)
ci ccap
(25e)
G gbound
(25f)
i=1
Proof. We now consider a Lyapunov matrix P as follows

!
K 0
P=
0 M
(26)
where is a positive scalar. Substituting the matrix P and the closed-loop system
matrices (4) into the H2 inequality conditions (9)(11) results in the following inequalities
!
2 (D + FGF T ) F
0
(27a)
FT
I
!
M F
0
(27b)
FT Z
trace(Z) 2
(27c)
The scalar in the selected Lyapunov matrix (26) is an unknown parameter that
can be used as an additional degree of freedom in our formulation in order to reduce
the conservativeness of the H2 norm bound. Note that due to the cross product of
and D in (27a), this inequality is not an LMI. However, application of the Schur
complement formula to (27a) yields
2(D + FGF T ) + FF T 0
which is an LMI with respect to , G, and D. Substitution of the damping matrix
expansion (18) completes the results.
314
4.3 Integrated Design Based on a Mixed H2 /H Specification

The solution of the integrated design problem of the damping parameters and
the symmetric output feedback controller to satisfy a closed-loop mixed H2 /H
specifications is obtained from the following result. The proof of this result follows
similar lines as the proofs of Theorems 4 and 5.
Theorem 6 Consider the collocated structural system (1) with the damping distribution (18). For given positive scalars and , the H norm of the closed-loop
system of the collocated structural system (1) and the output feedback controller (2)
is less than and the H2 norm of the closed-loop system is less than if the following matrix inequalities with respect to the damping coefficients ci , the controller
gain G, and the parameters and Z are satisfied.
l
2( ci Ti + FGF T ) + EE T 0
(28a)
i=1
!
M E
0
ET Z
(28b)
l
T EE T E
c
T
+
FGF
i
i
2
i=1
0
T
E
2
(28c)
trace(Z) 2
(28d)
0 ci ci max , i = 1, . . . , l
(28e)
ci ccap
(28f)
G gbound
(28g)
i=1
It should be noted that the integrated design formulation presented here provides
a non-iterative and computationally tractable method for simultaneous design of
structural damping parameters and control gain for collocated structural systems.
This is in contrast to the available methods in the literature, that seek to solve complex nonlinear optimization problems to address the integrated plant/control design.
It is emphasized that the proposed upper bound design approach is only capable of
optimizing the values of damping coefficients and control gains. Simultaneous design of stiffness and damping parameters and control gains as in [12,27] would result
in a non-convex optimization problem of much greater computational complexity.
315
4.4 Decentralized Control Using the Norm Upper Bound

Formulation
Complex large-scale systems often consist of smaller interconnected interacting
subsystems. A need for decentralized control naturally arises in such systems due
to communication and hardware implementation constraints. The goal of decentralized control is to design local controllers that use local measurements to generate
a local control action for each subsystem. At the same time global objectives for
stability and performance of the closed-loop large-scale system should be satisfied
[35]. Decentralized H or H2 control design problem for large-scale systems has
been shown to result in a non-convex optimization problem [40], and several iterative algorithms to achieve a locally optimal control have been proposed [22, 31, 40].
The proposed norm upper bound method of this chapter results in an LMI problem
even if the decentralized control structure is fixed. Imposing a particular structure
for the control matrix G is dictated by the structure of the subsystems in the largescale system. Using the method of this chapter, we are able to obtain the optimal
static rate-feedback gain matrices and damping parameters for externally symmetric
systems when an upper bound on the closed-loop system norm is sought. The procedure to formulate the decentralized control design problem for the systems with
collocated sensors and actuators in the form presented by (1) is described below.
It is noted that stability conditions for such systems using decentralized dynamic
displacement feedback are provided in [9].
We consider l structural subsystems with collocated sensors and actuators interconnected by flexible links which can be modeled as springs and dampers and
described by
Mi qi + Di qi + Ki qi = Fi ui + Ei wi
+
Ni j
2
KCi j (N Tji q j NiTj qi ) + DCi j (N Tji q j NiTj qi )
j=1
yi = FiT qi
zi = EiT qi ,
i = 1, . . . , l
where qi Rni , ui Rri , and yi Rri are the displacement (translational displacement and rotational angle), control input (force and torque), and the measured velocity output of the ith subsystem, respectively. Also, wi Rti and zi Rti represent
the external disturbance input, and the controlled output of the ith subsystem, respectively. The matrices Mi , Di and Ki are the mass, damping, and stiffness matrices
corresponding to the ith subsystem. Also, Ni j is a matrix representing the locations
and directions of the springs and dampers, which connect the ith subsystem with
the jth subsystem. The matrices KCi j and DCi j denote spring and damper matrices,
respectively, which are positive definite. The effects of the springs and dampers are
bilateral, and hence KCi j = KC ji and DCi j = DC ji .
316
We consider local velocity output feedback control laws as

ui = Gi yi
(29)
The overall closed-loop system is then described by

M q + Dq + Kq = Fu + Ew
y = F T q
z = E T q
(30)
where the control law is

u = Gy
(31)
with
G = diag{G1 , . . . , Gl }
and
q = [qT1 , . . . , qTl ]T , w = [wT1 , . . . , wTl ]T , z = [zT1 , . . . , zTl ]T , u = [uT1 , . . . , uTl ]T
E = diag{E1 , . . . , El }
M = diag{M1 , . . . , Ml }
l1
D = diag{C1, . . . ,Cl } +
l1
N i j DCi j N iTj
i=1 j=i+i
K = diag{K1, . . . , Kl } +
N i j KCi j N iTj
i=1 j=i+i
where
..
.
Ni j
..
.
N i j =
N ji
..
.
(32)
in which all the elements except the submatrices Ni j and N ji are zero.
The design of the decentralized control law (31) then follows a similar formulation as the prior centralized control design results with the additional constraints that
G has a block diagonal structure. Hence, such constraint can be easily incorporated
in the corresponding upper bound LMI formulation.
317
4.5 Additional Remarks

In this section, we provide some additional remarks and comments on the integrated
design method proposed in this chapter.
Remark 1 The conditions of Theorems 4 and 5 establish an LMI feasibility problem
with respect to the damping coefficients ci and the controller gain G. Given the
scalar bounds ci max , ccap and gbound , the optimum values of the damping coefficients
and the controller gain that minimize the H2 norm bound can be obtained by solving
the following LMI optimization problem

min ,ci ,G
2
(33)
sub ject to (25a) (25f)
Remark 2 The total number of unknown parameters in the LMI optimization prob+ l and m(m + 1) + l + 1, respectively. These
lems of Theorems 4 and 5 is m(m+1)
2
correspond to the independent elements of the symmetric feedback gain matrix G
and the damping parameters ci . It is also noted that the total number of LMI constrains is l + 3 and l + 5, respectively.
Remark 3 The control gain matrix norm bound condition in (22d) and (25f) can be
expressed in an LMI form as follows [6]
!
g2bound I GT
0
(34)
G
I
Remark 4 Following similar lines as above, the results of Theorems 4 and 5 can
be used to minimize the available damping resources ccap or the control gain norm
gbound subject to a given bound on the H or the H2 norm of the closed-loop system.
For example, minimization of the control gain norm gbound subject to a given bound
of the H norm of the closed-loop system can be achieved by solving the following
LMI optimization problem.

minci ,G
gbound
(35)
subject to (22a) (22d)
5 Simulation Results
In this section, we validate the proposed integrated damping parameters and control
gain design methods using the different norm specifications for a collocated structural system and the corresponding static output feedback control gain computation
by providing illustrative examples. The MATLAB Robust Control Toolbox is used
for the computational solution of the corresponding LMI optimization problems that
are involved in the integrated design procedures.
318
Example 1
As the first application example, we consider the lumped model of a ten-story
base isolated building structure as shown in Fig. 1. The mass of each floor, including
that of the base, is assumed to be 250 tons. The stiffness of the structure varies in
steps of 107 N/m between floors from 108 N/m for the first floor to 1.9 108 N/m
for the tenth floor. The design objective is to optimize the values of the damping
coefficients ci , i = 1, ..., 10, as well as, the output feedback control gain G such that
the H norm of the closed-loop system consisting of the collocated structure and
the output feedback from the disturbance forces w1 (t) and w2 (t) (acting on the first
and forth floors) to the velocities of the masses m1 and m4 is minimized.
The damping matrix D of this system is given by (18), where the elemental distribution matrices Ti are easily defined based on the system configuration. To examine
integrated design trade-offs, we consider a family of optimal designs using the result of Theorem 4. We examine two scenarios. First, we fix the upper bound on the
feedback control gain matrix norm to be gbound = 1. We consider different designs
corresponding to different values of the total damping capacity ccap ranging from 0.1
to 104 Ns/m. The results of the corresponding integrated designs (using Theorem 4)
are shown in Figs. 2, 3 and 4. Figure 2 shows the values of the H norm bound
obtained from solving the convex optimization problem for each design, as well as,
Fig. 1 Ten-DOF system
319
1
H norm bound
Actual H norm
0.9
0.8
H norm
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
101
100
101
ccap
102
103
104
Fig. 2 Profiles of the H norm upper bound and the actual H norm for the optimized structure
vs. the total damping capacity
damping coefficients (Ns/m)
3000
2500
2000
c1
c2
c3
c4
c5
1500
1000
500
0
101
100
101
ccap
102
103
104
Fig. 3 Values of the optimized structural damping coefficients vs. the total damping capacity
the exact H norm for each design as the total damping capacity ccap changes. As
we expect, the optimal closed-loop H norm of the integrated design decreases as
the overall damping allowed in the system increases. This plot also illustrates the
accuracy of the H norm bound calculated using Theorem 4. Figures 3 and 4 show
the values of the structural damping parameters and the closed-loop damping ratios
(in %) corresponding to each design, respectively. It should be noted that the damping ratios corresponding to all the floors are comparable even though the dominant
damping parameters are c1 , c2 , c3 and c4 . Indeed, the rest of the damping parameters
damping ratios (%)
320
0.03
0.02
0.01
1
2
3
4
5
0
101
100
101
102
103
104
damping ratios (%)
ccap
0.02
0.015
0.01
0.005
0
101
8
9
10
100
101
102
103
104
ccap
Fig. 4 The optimized closed-loop damping ratios vs. the total damping capacity
remain very small (close to zero) and the damping parameters corresponding to the
four lower floors become significantly larger as ccap increases. The reason for this
behavior is the location of the sensors (and actuators) on the first and forth floors. It
is indeed expected that the first four dampers will be the dominant ones to damp the
structures velocity response measured by the sensors on the first and fourth floors.
As a second design scenario, we consider a given bound for the total damping capacity ccap = 50 Ns/m, and we minimize the H norm of the closed-loop
system with respect to the damping coefficients and the control gain. We compare
different designs obtained by varying the upper bound on the controller gain norm
gbound . Figure 5 depicts the optimal H norm bound (obtained from solving the
optimization problem of Theorem 4), as well as, the exact H norm corresponding
to each design versus gbound . As expected the closed-loop H norm decreases when
the allowable magnitude of the control gain matrix norm increases. Figure 6 shows
the optimized structural damping coefficients (associated with the first five floors)
obtained from the integrated designs. Note that the damping parameters associated
with the 6th to 10th floors are close to zero. From Fig. 6, it is observed that the
damping parameters do not change significantly as gbound varies, and as expected,
c1 , c2 , c3 and c4 are the dominant damping parameters. However, the closed-loop
damping ratios vary as the allowed control gain increases (not shown here).
Figures 2 and 5 verify the accuracy of the obtained closed-loop H norm of the
system. Note that for calculating the actual H norm of the closed-loop system, the
feedback interconnection of the open-loop structure and the controller is considered,
where the structure includes the designed values of the damping parameters ci , and
the controller is constructed using the feedback control gain G obtained from the
solution to the convex optimization problems discussed in Sect. 4
321
0.35
Norm obtained from the design
Exact H norm
0.3
H norm
0.25
0.2
0.15
0.1
0.05
0
102
101
100
101
102
Bound on controller gain
103
104
Fig. 5 Profiles of the closed-loop H norm obtained by solving the LMI optimization problem
and the actual norm calculated based on the optimized structure vs. bound on feedback control
gain
damping parameters (NS/m)
15
10
c1
c2
c3
c4
c5
0
102
101
100
101
gbound
102
103
104
Fig. 6 Values of the optimized structural damping coefficients vs. bound on the norm of the feedback control gain
Example 2
Next, we consider the single-input/single-output structural model of a cantilevered beam controlled by pairs of piezoelectric patches as collocated sensors
and actuators. This is the model of a thin aluminum flexible beam with low damping
322
characteristics with properties listed in Table 1 [23]. We derive a 40-element finite

element structural model for this structural system that has the vector second-order
form (1). We seek to examine the application of the proposed integrated design
method for the damping distribution and control gain design of this large-scale
collocated structural system. For this example, we are interested in the optimum
design of the damping parameters and the output feedback control gain from the
closed-loop H2 norm performance viewpoint. We assume given bounds on the total damping capacity and the norm of the controller gain as follows: ccap = 1 Ns/m
and gbound = 20. Solving the LMI optimization problem of Theorem 5 for the unknown damping parameters and control gains results in an optimal closed-loop H2
norm bound of = 18.542. It is noted that the actual H2 norm of the system for
the designed parameters is = 18.360. The frequency responses of the undamped
open-loop system, the open-loop system damped by optimized damping parameters
only, and the optimal closed-loop system designed using the H2 upper bound approach are shown in Fig. 7. The results demonstrate that the simultaneous design of
Table 1 Aluminum beam properties

Quality
Description
L
Beam length
wb
Beam width
Beam thickness
tb
b
Beam density
E
Modulus of elasticity
Units
mm
mm
mm
kg/m3
N/m2
Value
736.5
53.1
1
2,690
7.03 1010
Bode Diagram
150
100
Magnitude (dB)
50
0
50
100
simultaneous design
designed damped system
undamped system
150
200
100
102
104
Frequency (rad/sec)
106
108
Fig. 7 Frequency responses of the undamped system, damped system constructed by designed
damping parameters, and the closed-loop system of the damped system and H2 controller, from
w1 (t) to q1 (t)
323
20
H2 norm bound
18
Actual H2 norm
16
H2 norm
14
12
10
8
6
4
2
0
100
101
102
ccap
103
104
Fig. 8 Profiles of the H2 norm obtained by solving the convex optimization problem of Theorem 5
and the actual H2 norm of the closed-loop system of the optimized structure and output feedback
control vs. the total damping capacity
damping parameters and controller provides improved disturbance rejection compared to past work that seeks to optimize only damping parameters [24]. Note that
using traditional methods for simultaneous control and damping parameter design
for this system could easily become prohibitive due to the high dimensionality of
the system.
Finally, Fig. 8 shows the H2 norm bound obtained by solving the integrated
damping parameter and control gain optimization problem presented in Theorem 5
for different values of the total damping capacity ccap and the actual H2 norm of
the structural system for each design. It is observed that the value of the H2 norm
bound and the achievable H2 norm are indeed extremely close.
Example 3
As a last example, we consider the model of an interconnected structure composed of two substructures, denoted by S1 and S2 as depicted in Fig. 9. The structural
parameters are listed in Table 2, and the two subsystems S1 and S2 are connected
with a damper dc and a spring kc at the third floor. We assume a decentralized control
scheme for each subsystem given as the follows.
1
1
u1 (t)
q1 (t)
u1 (t)
q1 (t)
2
3
ud (t) = Gyd (t), ud (t) =

u2 (t) , yd (t) = q2 (t) , G = diag(G1 , G2 ) (36)
1
1
u22 (t)
q23 (t)
324
Fig. 9 Interconnected ten-DOF system consisting of two subsystems S1 and S2
Table 2 Structural parameters of the decentralized

control problem
Structural parameters
Unit
Value
mij (i = 1, . . ., 5, j = 1, 2)
kg
6.0 105
k1j ( j = 1, 2)
N/m
7.0 108
k2j ( j = 1, 2)
N/m
7.5 108
j
k3 ( j = 1, 2)
N/m
8.0 108
j
k4 ( j = 1, 2)
N/m
8.5 108
j
k5 ( j = 1, 2)
N/m
9.0 108
Table 3 Optimized damping parameters using the proposed
decentralized control method
Optimized damping parameters (Ns/m)
Value
d11 = d12
11.765
d21 = d22
17.647
17.647
d31 = d32
d41 = d42
23.529
23.529
d51 = d52
188.24 ccap
2i=1 5j=1 d ij
Upper bound of the closed-loop H norm
0.02
0.02
Exact closed-loop H norm
Let dc = 100 Ns/m and kc = 8 107 N/m. Using the results in Theorem 4, the
output feedback gain matrix G and damping coefficients d ij (i = 1, 2, j = 1, . . . , 5)
are obtained by minimizing an upper bound on the closed-loop H norm considering gbound = 50 and ccap = 2i=1 5j=1 d ij = 200. The result is given in Table 3
with the exact closed-loop H norm obtained from the solution to the integrated
design problem. The open and closed-loop singular value plots and the impulse responses are shown in Figs. 10 and 11, respectively. It is observed that the obtained
325
Maximum singular value plot

40
Closedloop system
Openloop system
Maximum singular values [dB]
20
20
40
60
80
100
101
102
Frequnency [rad/s]
103
104
Fig. 10 Maximum singular values of the open-loop and closed-loop systems
dq11/dt [m]
Impulse response
1
0
dq23 /dt [m]
dq21/dt [m]
dq13 /dt [m]
0.2
0.4
0.6
0.8
0.4
0.6
0.8
0.6
0.8
0.6
0.8
1
0
1
0.5
Closedloop
Openloop
0
0.5
0.2
0.4
0.2
0.4
0.5
0
0.5
t [s]
Fig. 11 Impulse responses of the open-loop and closed-loop systems
326
upper bound on the closed-loop H norm is accurate and that closed-loop vibration suppression is achieved using the decentralized control scheme of this
chapter.
Presented in this chapter is an efficient computational methodology for the simultaneous design of the damping parameters and the output feedback control gain of a
collocated structural system with velocity feedback such that the closed-loop system
satisfies H2 , H or mixed H2 /H performance specifications. The proposed integrated design approach is based on an LMI formulation of the design problem that
can be efficiently solved for the design variables using available semi-definite programming optimization solvers. Despite the fact that the method is based on an
upper bound formulation of the norm performance specifications of the closed-loop
system, computational examples demonstrate that the bounds result in a close approximation of the actual norms of the system and are effective for structural parameter and control design. The integrated design method is also shown to be efficient
for design of large-scale interconnected structural systems with a large number of
subsystems and states. As demonstrated, the proposed method is especially suitable
for large-scale systems where existing nonlinear optimization approaches used as
the standard tools to determine system parameters and control gains are computationally prohibitive.
References
1. K. Adachi, K. Sakamoto, and T. Iwatsubo, Redesign of Closed-Loop System for Integrated
Design of Structure and its Vibration Control System, Proceedings IEEE International Conference on Control Applications, pp. 8085, 1999
2. B. Anderson and S. Vongpanitlerd, Network Analysis and Synthesis: A Modern Systems
Theory, Prentice-Hall, Englewood Cliffs, 1973
3. Y. Bai and K.M. Grigoriadis, H Collocated Control of Structural Systems: An Analytical
Bound Approach, AIAA Journal of Guidance, Control, and Dynamics, vol. 28, pp. 850853,
2005
4. A.V. Balakrishnan, Compensator Design for Stability Enhancement with Collocated Controllers, IEEE Transactions on Automatic Control, vol. 36, pp. 9941007, 1991
5. M.J. Balas, Direct Velocity Feedback Control of Large Space Structures, Journal of Guidance and Control, vol. 2, pp. 252253, 1979
6. S.P. Boyd, L. El Ghaoui, E. Feron, and V. Balakrishnan, Linear Matrix Inequalities in System
and Control Theory, vol. 15, Studies in Applied Mathematics, SIAM, Philadelphia, PA, 1994
7. S. Boyd and L. Vandenberghe, Convex Optimization, Cambridge University Press, Cambridge,
2004
8. E.F. Crawley and J. de Luis, Use of Piezoelectric Actuators as Elements of Intelligent Structures, AIAA Journal, vol. 25, pp. 13731385, 1987
9. Y. Fujisaki, M. Ikeda, and K. Miki, Robust Stabilization of Large Space Structures via Displacement Feedback, IEEE Transactions on Automatic Control, vol. 46, pp. 19931996, Dec.
2001
327
10. P. Gahinet, A. Nemirovskii, A.J. Laub, and M. Chilali, LMI Control Toolbox Users Guide, The
Mathematical Works, Natick, MA 1995
11. W.K. Gawronski, Advanced Structural Dynamics and Active Control of Structures, Mechanical Engineering Series, Springer, New York, 2004
12. K.M. Grigoriadis and F. Wu, Integrated H Plant/Controller Design using Linear Matrix
Inequalities, Proceedings 36th IEEE Conference on Decision and Control, pp. 789790, San
Diego, CA, Dec. 1997
13. K.M. Grigoriadis, G. Zhu, and R.E. Skelton, Optimal Redesign of Linear Systems, ASME
Transactions on Dynamic Systems, Measurements and Control, vol. 118, pp. 598605, 1996
14. K. Hiramoto and K.M. Grigoriadis, Integrated Design of Structural and Control Systems with
a Homotopy Like Iterative Method, International Journal of Control, vol. 79, pp. 10621073,
2006
15. T. Iwasaki, S. Hara, and H. Yamauchi, Structure/Control Design Integration with Finite Frequency Positive Real Property, Proceedings of American Control Conference, pp. 549553,
2000
16. S.P. Joshi, Control for Energy Dissipation in Structures, Journal of Guidance, Control, and
Dynamics, vol. 13, pp. 751753, 1990
17. N.S. Khot, Structure/Control Optimization to Improve the Dynamic Response of Space
Structures, Computational Mechanics, vol. 3, pp. 179186, 1988
18. N.S. Khot, V.B. Venkayya, H. Oz, R.V. Grandhi, and F.E. Eastep, Optimal Structural Design
with Control Gain Norm Constraint, AIAA Journal, vol. 26, pp. 604611, 1988
19. Y. Kobayashi, M. Ikeda, and Y. Fujisaki, Stability of Large Space Structures Preserved under
Failures of Local Controllers, IEEE Transactions on Automatic Control, vol. 52, pp. 318322,
Feb. 2007
20. F. Liao, K.Y. Lum, and J.L. Wang, An LMI-Based Optimization Approach for Integrated
Plant/Output-Feedback Controller Design, Proceedings of 24th American Control Conference, pp. 44804485, Portland, OR, 2005
21. J.B. Lu and R.E. Skelton, Integrating Structure and Control Design to Achieve Mixed
H2 /H Performance, International Journal of Control, vol. 73, pp. 14491462, 2000
22. C.S. Mehendale and K.M. Grigoriadis, A Double Homotopy Method for Decentralised Control Design, International Journal of Control, vol. 81, pp. 16001608, Oct. 2008
23. M. Meisami-Azad, Control Design Methods for Large-Scale Systems Using an Upper Bound
Formulation, Masters Thesis, University of Houston, TX, Aug. 2007
24. J. Mohammadpour, M. Meisami-Azad, and K.M. Grigoriadis, An Efficient Approach for
Damping Parameter Design in Collocated Structural Systems Using An H2 Upper Bound,
Proceedings of the 26th American Control Conference, pp. 30153016, New York, NY, Jul.
2007
25. M. Meisami-Azad, J. Mohammadpour, and K.M. Grigoriadis, Explicit Solutions for Collocated Structural Control with Guaranteed H2 Norm Performance Specifications, Smart Materials and Structures, 18 035004 (9pp), Mar. 2009
26. Y. Nesterov and A. Nemirovsky, Interior-point Polynomial Methods in Convex Programming,
vol. 13, Studies in Applied Mathematics, SIAM, Philadelphia, PA, 1994
27. R.J. Niewoehner and I.I. Kaminer, Integrated Aircraft-Controller Design Using Linear Matrix
Inequalities, AIAA Journal of Guidance, Control and Dynamics, vol. 19, pp. 445452, 1996
28. J. Onoda and R.T. Haftka, An Approach to Structure/Control Simultaneous Optimization for
Large Flexible Spacecraft, AIAA Journal, vol. 25, pp. 11331138, 1987
29. S.S. Rao, V.B. Venkayya, and N.S. Khot, Game Theory Approach for the Integrated Design
of Structures and Controls, AIAA Journal, vol. 26, pp. 463469, 1988
30. M. Salama, J. Garba, L. Demsetz, and F. Udwadia, Simultaneous Optimization of Controlled
Structures, Computational Mechanics, vol. 3, pp. 275282, 1988
31. G. Scoletti and G. Duc, An LMI Approach to Decentralized H Control, International Journal of Control, vol. 74, pp. 211224, Feb. 2001
32. G. Shi and R.E. Skelton, An Algorithm for Integrated Structure and Control Design with
Variance Bounds, Proceedings of 35th IEEE Conference on Decision and Control, pp. 167
172, Kobe, Japan, Dec. 1996
328
33. R.E. Skelton, Integrated Plant and Controller Design, Proceedings of American Control
Conference, Seattle, WA, Jun. 1995
34. R.E. Skelton, T. Iwasaki, and K.M. Grigoriadis, A Unified Algebraic Approach to Linear Control Design, Taylor & Francis, 1998
35. D.D. Siljak, Decentralized Control of Complex Systems, Academic, Boston, MA, 1991
36. S. Sidi-Ali-Cherif, K.M. Grigoriadis, and M. Subramaniam, Model Reduction of Large Space
Structures Using Approximate Component Cost Analysis, AIAA Journal of Guidance, Control and Dynamics, vol. 22, pp. 551558, 1999
37. K. Tan and K.M. Grigoriadis, Stabilization and H Control of Symmetric Systems: An Explicit Solution, Systems and Control Letters, vol. 44, pp. 5772, 2001
38. S.M. Yang and Y.J. Lee, Optimization of Non-collocated Sensor/actuator Location and Feedback Gain in Control Systems, Smart Materials and Structures, vol. 2, pp. 96102, 1993
39. K. Zhou, J.C. Doyle, and K. Glover, Robust and Optimal Control, Prentice-Hall, Upper Saddle
River, 1996
40. G. Zhai, M. Ikeda, and Y. Fujisaki, Decentralized H Controller Design: A Matrix Inequality
Approach Using A Homotopy Method, Automatica, vol. 37, pp. 565572, 2001
Index
A
Adaptive dual dynamic high-gain scaling
paradigm
assumptions, 137139
observer and controller designs
m dynamics, 141
observer errors and scaled observer errors,
140
stability analysis
closed-loop stability properties, 150151
composite Lyapunov function, 142, 145
observer and controller Lyapunov
functions, 142
parameter estimator dynamics, 146
Aero-engine model
application
first-order model, 105109
second-order model, 109110
gas turbine system
nonlinear dynamic model, 9293
nonlinear static model, 9192
reheat bypass turbojet engine, 9091
OE parameter identification
least-squares estimate (LSE) algorithm, 99
R(k) and Hessian approximation, 101103
V( )/ and Jacobian calculation,
100101
reduced order data driven modeling
criterion selection, 9496
model selection: EE vs. OE, 9699
Air vehicle model, 59
Auto-regressive exogeneous (ARX) model, 98
B
Balancing method, reduced order model
balancing over a disk
asymptotic balancing process, 68, 70
bilinear mapping, 6869

infinite bandwidth, 68
Cholesky factors, 65
Hankel minimum degree approximate
algorithm, 6668
observability and reachability mapping, 64
Bilinear matrix inequality (BMI), 307
Brute force technique, 130
C
Canonical models
linear and time-invariant systems, 220
linear state-space dynamics, 224
NDS
coupled at input, 226227
coupled at output, 225226
coupled at state, 227228
coupled by state, input and output, 228
Cascading upper diagonal dominance
(CUDD), 136
Closed-loop stability properties, 150151
Collocated structural system
advantages, 306
symmetric output feedback control, 307308
upper bound, 308310
Composite Lyapunov function, 142, 145
Continuous-time Lyapunov equation, 119
Control input level consensus
local dynamic output feedback control laws
aggregation relations, 201
impulse response, 200203
LQG methodology, 203
vector Lyapunov functions, 205
vector-matrix differential equation, 202
local static feedback control laws
design, 207
329
330
Coordination strategy. See Multi-agent control
CUDD. See Cascading upper diagonal
dominance
D
Decentralized control. See also Decentralized
output feedback control
local controller design, 315
overall closed-loop system, 316
UAVs formations
decentralized state estimation, 213214
design, 210
experiments, 215217
formation model, 210211
global LQ optimal state feedback, 212213
Decentralized output feedback control
adaptive dual dynamic high-gain scaling
paradigm
assumptions, 137139
observer and controller designs, 140141
stability analysis, 141151
cascading upper diagonal dominance, 136
example
admissible uncertainties, 185
local operation modes and global operation
mode, 184
uncertain large-scale system, 183
generalised scaling
assumptions, 152154
controller design, 155158
observer design, 154155
guaranteed cost controller design
design methodology, 173176
global mode dependent controllers,
176180
local mode dependent controllers,
180182
procedure, 182183
problem formulation
IQC-type descriptions, 172173
operation modes, 170171
stationary ergodic continuous-time
Markov process, 170
vector mode process, 170
Decomposable systems, distributed control
design
applications
infinite dimensional system generalization,
292293
platoon, 293294
LMIs, 280, 291
Lyapunov shaping method, 291
Index
multiobjective optimization, 291
state-space matrices, 290
static state feedback, 290
symmetric decomposable systems, 289
Directed weighted graphs, 211
Distributed control methods
Car platoon benchmark problem
controllers, 299
optimal centralized gain operator,
301302
spatiotemporal H2 norm, 300
decomposable systems
application, 292294
LMIs, 280, 291
multiobjective optimization, 291
state-space matrices, 290
static state feedback, 290
symmetric decomposable systems, 289
identical systems, LQR
closed-loop state matrix, 299
closed-loop system matrix, 298
dynamically decoupled systems, 295297
infinite-dimensional Laurent matrix, 297
Laurent operator, rational
applications, 286287
convergence, 285286
definition, 285
doubly infinite one-dimensional strings,
282283
H2 optimal state feedback problem, 288
L-operator sign function, 284285
posteriori closed loop stability, 287
string interconnection, 284
problem statement
controlling absolute and relative distances,
280
H2 problem and exact solution, 281282
structure preserving iterative algorithms,
279
DNLS. See Dynamic nonlinear least squares
Dynamic gradient descent (DGD) algorithm,
106
Dynamic nonlinear least squares (DNLS)
algorithm, 103
E
Equation error (EE) model
auto-regressive exogeneous (ARX) model,
98
definition, 9697
Equations of motion, 7879
EulerLagrange equation, 84
Index
F
FACTS. See Flexible alternating current
transmission systems
Flexible alternating current transmission
systems (FACTS), 272
G
Gas turbine system
first order model
DGD algorithm, 106
DNLS algorithm, 107
objective function, 105106
nonlinear dynamic model, 9293
nonlinear static model, 9192
reheat bypass turbojet engine, 9091
second-order model, 109
Gaussian white noise, 234, 237
GaussNewton methods, 102
Global LQ optimal state feedback, 212213
Global mode dependent controllers
definition, 168
guaranteed cost controller design, 176180
Graph-theoretic bounds and analysis
H2 performance, NDS
coupled at the output, 233235
coupled at the state, 236241
observability and controllability of NDS
coupled at input, 232
coupled at output, 230231
properties of linear system, 229
Guaranteed cost controller design
design procedure, 182183
global mode dependent controllers, 176180
local mode dependent controllers, 180182
methodology, 173176
H
Hadamard product, 221, 231
algorithm, 6668
H2 norms
graph-theoretic bounds performance
NDS coupled at output, 233235
NDS coupled at state, 236241
topology design
NDS coupled at state, sensor placement,
244246
H norm upper bound approach, 318319
H2 upper bound approach, 322
I
Integral quadratic constraints (IQCs)
controller uncertainty, 169
331
Markovian jump parameter systems, 168,
169
uncertainties and interconnections, 171
Integrated design, large-scale collocated
structural system
BMI formulation, 306
damping coefficients, 317318
decentralized control, 315316
H specification, 311312
H2 specification, 312313
mixed H2 /H specification, 314
validation
interconnected structure, 323326
lumped model, 318321
single-input/single-output structural
model, 321323
Interpolatory model reduction
advantages, 4
approximation theory, 4
coprime factorizations
driven cavity flow, 3335
isotropic incompressible viscoelastic solid,
30
second-order dynamical systems, 3537
error measurement
full-order system, 14
H norm, 15
H2 norms, 1516
matrix-valued meromorphic function, 16
transfer functions G(s) and H(s), 16, 17
framework
approximation, 6
low order transfer function, 6
system matrices, 8
tangential interpolation, 67
goals, 5
input-output map
input-output data, 8
reduced system, 5
state-space realization, 45
interpolatory projections
differential algebraic equation (DAE), 13
full-order dynamical system, 10
Hermite interpolation problem, 11
interpolant construction, 1314
optimal point selection strategy, 13
PetrovGalerkin projective approximation,
89
rational tangential interpolation problem,
12
tangential direction selection, 12
measurements
coupled mechanical system, 4951
four-pole band-pass filter, 52
332
Loewner and Pick matrices, 4647
Loewner matrix pair and construction
of interpolants, 4246
simple low-order example, 4749
S-parameter representation, 42
optimal H2 approximation
interpolation-based H2 -optimality conditions, 1824 (see also Iterative rational
Krylov algorithm)
Lyapunov-based optimal H2 method, 18
parametric systems, 3741
passive systems
rational square matrix function H(s), 25
RLC circuit, 2730
spectral zeros (), 2527
state space system data, 7
Inverse dynamics
bounded-input-bounded state (BIBS), 152
input-to-state practically stable (ISpS),
137
Iterative rational Krylov algorithm (IRKA)
bode plots of error systems, 2425
convergence behavior, 2223
first-order conditions, 1820
MIMO H2 optimal tangential interpolation
method, 22
numerical results, 2324
reduced model, error norms, 24
K
Kruskals algorithm, 243
L
Laurent operator, rational
convergence, 285286
definition, 285
doubly infinite one-dimensional strings,
282283
H2 optimal state feedback problem, 288
L-operator sign function, 284285
posteriori closed loop stability, 287
string interconnection, 284
Least-squares estimate (LSE) algorithm, 99
Linear matrix inequalities (LMIs), 280, 291
Linear time-invariant (LTI) control systems,
115
LMIs. See Linear matrix inequalities
Local dynamic output feedback control laws
aggregation relations, 201
LQG methodology, 203
vector Lyapunov functions, 205
vector-matrix differential equation, 202
Index
Local mode dependent controllers
definition, 169
guaranteed cost controller design, 180182
Local static feedback control laws
design, 207
Loewner and Pick matrices
construction of interpolants
generalized tangential reachability
and observability matrices, 4344
shifted Loewner matrix, 44
singular value decomposition (SVD), 46
coupled mechanical system, 4951
four-pole band-pass filter, 52
simple low-order example, 4749
Long-term prediction, DNLS
first order model, 107108
second order model, 109
LQR solutions
closed-loop state matrix, 299
closed-loop system matrix, 298
dynamically decoupled systems, 295297
infinite-dimensional Laurent matrix,
297
Lyapunov-based optimal H2 method, 18
M
Markovian jump large-scale system model,
167
MATLAB Robust Control Toolbox, 317
Matrix sign function, 284
Minimum spanning tree (MST) problem,
242244
Mode dependent controllers
global, 168
local, 169
Model reduction techniques
balancing method
balancing over a disk, 6869
Cholesky factors, 65
algorithm, 6668
observability and reachability mapping, 64
frequency response characteristics, 59
large-scale systems
applications, 6971
asymptotic balancing process, 70
non-minimum phase zero effect, 70
residualization effect, 71
simultaneous gradient error reduction
conjugate gradient search routine, 62
fit error, 63
schematic representation, 63
Index
spectral decomposition process
distinct frequency range groupings, 60
eigenvalue and eigenvector calculation,
6062
Multi-agent control
control input level consensus
local dynamic output feedback control
laws, 200205
local static feedback control laws, 205207
dynamic consensus methodology, 198
goals, 256
network characteristics and control
objectives, 256257
optimal power flow control, 255
overlapping subnetworks
common nodes, 263264
control problem formulation for one agent,
264267
control scheme for multiple agents, 267
parameters
control objectives, 273
setting up control problems, 273274
simulations, 274276
steady-state characteristics, 268273
problem formulation, 199200
state estimation level consensus
globally LQ optimal controller, 209, 210
Luenberger form, 208
overlapping decentralized estimators, 208
subnetworks
control, 253255
definition, 257
non-overlapping, touching/overlapping,
254
touching subnetworks
control problem formulation for one agent,
259261
control scheme for multiple agents,
262263
internal and external nodes, 258259
UAVs formations, decentralized control
design, 210
experiments, 215217
formation model, 210211
Multi-input/multi-output (MIMO) system, 4
N
NDS. See Networked dynamic systems
Networked dynamic systems (NDS)
canonical models
333
continuous linear time-invariant systems,
220
linear state-space dynamics, 224
NDS coupled at input, 226227
NDS coupled at state, 227228
NDS coupled by state, input and output,
228
graph-centric analysis, 220
graph-theoretic bounds and analysis
H2 performance, 233241
observability and controllability, 229232
H2 topology design
NDS coupled at state, sensor placement,
244246
mathematical preliminaries and notations
edge and graph Laplacian relation, 222
K10 complete graph, 223224
k-regular graph, 223224
Kronecker product, 221
singular value decomposition, 221
subsystem dynamics, 220
Non-minimum phase zeros effect, 70
Non-overlapping power subnetwork, 254
Norm upper bound approach
collocated structural system
mixed H2 /H , 310
stable continuous-time system, 309
transfer function, 308
control design problem, 306
convex optimization problem, 307
integrated damping and control design
damping matrix, 311
decentralized control, 315316
integrated design, 311314
LMI, 317
LMI-based optimization schemes, 306
simulation
aluminum beam properties, 321322
cantilevered beam, 321
closed-loop damping ratios vs. total
damping capacity, 319320
closed-loop H norm, 320321
decentralized control problem, 323324
H norm upper bound, 318319
H2 upper bound approach, 322
interconnected structure, 323
MATLAB Robust Control Toolbox, 317
open-loop and closed-loop systems,
324325
structural damping coefficients vs. bound,
320321
334
structural damping coefficients vs. total
damping capacity, 319
ten-DOF system, 318
symmetric output feedback control
closed-loop system, 308
symmetric static feedback gain, 307
Nyquist stability criteria, 63
O
Observer and controller Lyapunov functions,
142
Observer errors and scaled observer errors,
140
Open-loop system, 308310
Optimal and robust control system design,
5960
Output error (OE) model
definition, 9697
parameter identification
R(k) and Hessian approximation, 101103
V( )/ and Jacobian calculation,
100101
pseudo-regression vector, 98
Overlapping power subnetwork
coordination strategy
control, 253255
goals, 256
network characteristics and control
objectives, 256257
optimal flow control, 268
optimal power flow control, 255
parameters, 268
power networks, control structure,
252253
definition, 257
multi-agent control
overlapping, 263267
touching, 258263
P
PetrovGalerkin approximation, 89
Pick matrices. See Loewner and Pick matrices
Polyas theorem, 119
Power networks
control objectives, 273
multi agent control, 252253
parameters, 268
setting up the control problems, 273274
simulations, 274276
steady-state characteristics
FACTS devices, 272
generators, 271
loads, 271
Index
power balance, 272273
transmission lines, 270271
Putinars theorem, 119
R
Reduced order data driven model
criterion selection
long-step prediction vs. one-step-ahead
prediction, 9496
parallel-series model, 95
model selection: EE vs. OE
equation error (EE) model, 9798
output error (OE) model, 9899
Robust closed-loop performance, 116
Robust control
continuous-time Lyapunov equation, 119
controllability and observability, 116
LTI uncertain fourth-order system, 129
non-polynomial rational function, 120
polytopic region
homogeneous matrix polynomial, 125126
optimization problem, 124
positive semi-definite matrix polynomial,
121
problem formulation
polynomially uncertain system, 117
sum-of-squares, 118119
robust stability, 115
SOS matrix polynomials, 122123
S
SADPA. See Subspace accelerated dominant
pole algorithm
Scalar real-valued continuous functions, 136
Second-order tangential MIMO order
reduction, 36
Shifted Loewner matrix, 44
Simultaneous gradient error reduction
conjugate gradient search routine, 62
fit error, 63
Single-input/single-output (SISO) system, 4
Single rigid rod dynamics
configuration matrix, 75
equations of motion, 7879
generalized forces and torques, 78
illustration, 7475
kinetic energy, 75
nodes, 7677
string forces, 7778
SOS. See Sum-of-squares
Spanning trees, 222, 223, 238240, 242244
Spectral decomposition process
Index
distinct frequency range groupings, 60
eigenvalue and eigenvector calculation,
6062
State estimation level consensus
globally LQ optimal controller, 209, 210
Luenberger form, 208
overlapping decentralized estimators, 208
Static var compensators (SVC)
flexible alternating current transmission
systems, 272
simulation control, 274276
Stochastic projection method, 180
String connectivity, 82
String forces, 7778
Structural damping coefficients, 319321
Structure-preserving model reduction, coprime
factorizations
driven cavity flow, 3335
isotropic incompressible viscoelastic solid,
31
second-order dynamical systems, 3537
Subspace accelerated dominant pole algorithm
(SADPA), 27
Sum-of-squares (SOS)
Polyas theorem, 119
Putinars theorem, 119
scalar polynomial, 118
T
Tangential interpolation, 67
Ten-DOF system, 318
Tensegrity systems
class k
class 2 tensegrity cable model, 8287
N interconnected rigid rods, 8283
335
class 1 tensegrity systems
compact matrix expression, 80
Lagrangian approach, 79
prism, 81
string connectivity, 82
definition, 73
dynamic models, 74
single rigid rod dynamics
equations of motion, 7879
generalized forces and torques, 78
kinetic energy, 75
nodes, 7677
string forces, 7778
Thyristor controlled series compensators
(TCSC)
flexible alternating current transmission
systems, 272
simulation control, 275276
U
Unmanned aerial vehicles (UAVs) formations
design, 210
experiments
consensus based controllers, 215216
expansion/contraction paradigm, 215217
inclusion principle, 215
formation model
Kroneckers product, 211
linear double integrator model, 210
W
Weak cascading upper diagonal dominance
(w-CUDD), 159

1441957561

Uploaded by

Document Informationclick to expand document information

Copyright:

Available Formats

1441957561

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

1441957561

Uploaded by

Copyright:

Available Formats

Efficient Modeling and Control

system whose attributes (nonlinear, stochastic, time-varying, delays, dimensionality,

The complexity and dynamic order of controlled engineering systems is constantly

Part I Model Reduction, Large-Scale System Modeling

Efficient Model Reduction for the Control of Large-Scale Systems . . . . . .

Modeling a Complex Aero-Engine Using Reduced Order Models . . . . . . . 89

Part II Large-Scale Systems Control and Applications

Modeling of Network Characteristics and Control Objectives . . . . . 256

Integrated Design of Large-Scale Collocated Structural System

Model Reduction, Large-Scale System

Interpolatory Model Reduction of Large-Scale

A.C. Antoulas et al.

Interpolatory model reduction methods have emerged as effective strategies for

Interpolatory Model Reduction of Large-Scale Dynamical Systems

where Ar , Er Rrr , Br Rrm , Cr R pr and Dr R pm . Note that the number

where we have denoted Laplace transformed quantities with . We define transfer

A.C. Antoulas et al.

2.1 The General Interpolation Framework

Interpolatory Model Reduction of Large-Scale Dynamical Systems

Problem 1: Model Reduction Given State Space System Data

right interpolation points:

It is of central concern to determine effective choices for interpolation points

A.C. Antoulas et al.

Problem 2: Model Reduction Given InputOutput Data

right driving frequencies:

Find a dynamical system model (2) by specifying (reduced) system matrices

Interpolation points and tangent directions are determined (typically) by the

2.2 Model Reduction via Projection

Interpolatory Model Reduction of Large-Scale Dynamical Systems

Then the associated output is y(t) = Cx(t) + Du(t).

A.C. Antoulas et al.

2.3 Interpolatory Projections

We wish to interpolate H(s) without ever computing the quantities to be matched

= (zE A)Vr (zEr Ar )1 WrT .

Interpolatory Model Reduction of Large-Scale Dynamical Systems

Both Pr (z) and Qr (z) are analytic matrix-valued functions in neighborhoods of

as 0, which proves the third assertion.

cT1 C(1 E A)1

cTr C(r E A)1

A.C. Antoulas et al.

Theorem 1.2. Let , C be such that s E A and s Er Ar are invertible for

Interpolatory Model Reduction of Large-Scale Dynamical Systems

A.C. Antoulas et al.

Then the reduced-order model Hr (s) = Cr (sEr Ar )1 Br + Dr satisfies

and cTi H(i ) = cTi Hr (i )

Theorem 1.3 describes how to construct an interpolant with an arbitrary Dr term;

2.4 Error Measures

Interpolatory Model Reduction of Large-Scale Dynamical Systems

max H( ) Hr ( )22

A.C. Antoulas et al.

G(x + y)2F dy < .

with a norm defined as

where i are indexed according to multiplicity as indicated by rank(res[F(s), i ]).

Interpolatory Model Reduction of Large-Scale Dynamical Systems

In particular, if H(s) has only simple or semi-simple poles at 1 , 2 , . . . m and

cTk G(k )bk

The remaining assertions follow from the definition.

H Hr 2H2 = H2H2 2 cTk H( k )bk +

A.C. Antoulas et al.

max H( ) Hr ( )22

G(x + y)2F dy < .

H Hr 2H2 = H2H2 2 cTk H( k )bk +

(c) cTk H ( k )bk = cTk Hr ( k )bk

r is made to differ from Hr , we arrive at

r( ) such that both dim(H

c. i i (Ar , Er ) for i = 1, . . ., r, b i eTi Y