Communications and Control Engineering

Springer-Verlag London Ltd.

Published titles include:

Passivity-based Control ofEuler-Lagrange Systems

Romeo Ortega , Antonio Loria, Per Iohan Nicklasson and Hebertt Sira-Ramirez
Stability and Stabilization ofInfinite DimensionalSystems with Applications
Zheng-Hua Luo, Bao-Zhu Guo and Orner Morgul
Nonsmooth Mechanics (2nd edition)
Bernard Brogliato
Nonlinear ControlSystemsII
Alberto Isidori
LrGain and PassivityTechniques in nonlinear Control
Arjan van der Schaft
ControlofLinear Systems with Regulationand Input Constraints
Ali Saberi, Anton A. Stoorvogel and Peddapullaiah Sannuti

Robust and He: Control

Ben M. Chen
Computer Controlled Systems
Efim N. Rosenwasser and Bernhard P. Lampe
Dissipative SystemsAnalysis and Control
Rogelio Lozano, Bernard Brogliato, Olav Egeland and Bernhard Maschke
ControlofComplex and Uncertain Systems
Stanislav V. Emelyanov and Sergey K. Korovin

Robust ControlDesign Using H" Methods

Ian R.Petersen, Valery A. Ugrinovski and Andrey V.Savkin
Model Reductionfor Control System Design
Goro Obinata and Brian D. O. Anderson
Control Theoryfor Linear Systems
Harry 1. Trentelman, Anton Stoorvogel and Malo Hautus
FunctionalAdaptive Control
Simon G. Fabri and Visakan Kadirkamanathan
Positive 1Dand 2D Systems
Tadeusz Kaczorek
Identification and Control Using Volterra Models
F.I. Doyle III, R.K. Pearson and B.A. Ogunnaike
Non-linear Control for Underactuated Mechanical Systems
Isabelle Fantoni and RogelioLozano
Jiirgen Ackermann
In cooperation with Paul Blue, Tilman Bunte, Levent Guvenc,
Dieter Kaesbauer, Michael Kordt, Michael Muhler
and Dirk Odenthal

Robust Control
The Parameter Space Approach

Second edition

With 321 Figures

i Springer
Professor Jürgen Ackermann
Deutsches Zentrum fur Luft-und Raumfahrt, Institut fur Robotik und Mechatronik,
Oberpfaffenhofen 82230 Wessling, Germany
Se ries Editors
E.D. Sontag • M. Thoma

ISSN 0178-5354
ISBN 978-1-4471-1099-6 ISBN 978-1-4471-0207-6 (eBook)
DOI 10.1007/978-1-4471-0207-6
British Library Cataloguing in Publication Data
Ackermann, Jurgen
Robust control : the parameter space approach. - 2nd ed. -
(Communications and control engineering)
I.Robust control
I.Titie
629.8'312
ISBN 978-1-4471-1099-6

Library of Congress Cataloging-in-Publicalion Data

Ackermann, J. (Jürgen)
Robust control: the parameter space approach I JÜfgen Ackermann.
p. cm. -- (Communications and control engineering)
Includes bibliographical references.
ISBN 978-1-4471-1099-6 (alk. paper)
1. Robust contro!. I. Title. 11. Series.
TJ217.2 .A25 2002
629.8--dc21 2001032821
Apart from any fair dealing for the purposes of research or private study, or criticism or review, as permitted
under the Copyright, Designs and Patents Act 1988, this publication may only be reproduced, stored or
transmitted, in any form or by any means, with the prior permission in writing of the publishers, or in the case
of reprographie reproduction in accordance with the terms of lieenees issued by the Copyright Lieensing
Ageney. Enquiries eoneerning reproduetion outside those terms should be sent to the publishers.
© Springer-Verlag London 2002
Originally published by Springer-Verlag London Berlin Heidelberg in 2002
Softcover reprint of the hardcover 2nd edition 2002

MATLAB" and SIMULINK" are the registered trademarks ofThe MathWorks Ine., 3 Apple Hill Drive Natick,
MA 01760-2098, U.S.A. hllp:/lwww.mathworks.ellm
Other registered trademarks used in this book are: BMW, Daimler-Benz, MAN, Airbus.

The use of registered names, trademarks, ete. in this publication does not imply, even in the absence of a speeifie
statement, that such names are exempt from the relevant laws and regulations and therefore free for general
lIse.
The publisher makes no representation, express or implied, with regard to the aceuracy of the information
eontained in this book and cannot aeeept any legal responsibility or liability for any errors or omissions that
maybemade.
Typesetting: Camera ready by author
69/3830-543210 Printed on acid-free paper SPIN 10834574
Preface

Linear control systems with known parameter values are usually described by state
space or transfer funct ion models . Deviations from th e nominal param eter values are
th en frequently modelled in an assumed generic structure. Examples are multiplicative
perturbations with arbitrary phas e and bounded gain or norm bounded perturbations of
state space models. The advantage of these approaches is th at one can develop generic
design procedures and software without restrictions to specific classes of plants.
In many applications, however, more specific knowledge is available on how uncer-
t ain real physical parameters (e.g. mass, velocity, friction coefficient , geometry etc. )
enter a well-known model structure. Even a large uncertainty of a few such essential
par ameters can frequentl y be tolerated by a fixed-gain or gain-scheduled cont roller, if
it is tailored to the structured parameters. If they are embedded, however , in a larger
number of generically st ruct ured or complex-valued perturbations, th en only small ad-
missible paramet er uncert ainti es result . In order to get non-conservative results, we
focus on essent ial parameters entering a known structure. Add itionally, high-frequency
unstructured model uncertainty and sensor noise are considered by frequency domain
specificati ons on Bode magnitude plots of different sensiti vity functions .
The approach of the book is graphics-oriented, and tak es advantage of the fact that
the engineer nowadays has increasingly powerful computer gra phics on his or her desk.
Optimizati on-based approaches are not covered. We believe t hat opt imizat ion should be
done at a higher system level where many trade-offs between different specifications for
th e overall system are required. At the low robust feedback design level, the flexibility
for such trade-offs should be preserved, e.g. by admissible solut ion sets.
In thi s second edition of Robust Control, the material is arra nged such that the
reader is first introduced to the parameter space approach. Once the limitation to a
few uncertain plant parameters or a few free controller par amet ers in a design step is
accept ed, the reader is rewarded by easy-t o-interpret figures that make design conflicts
transpar ent, and by non-conservative mapping of specificati ons and parameter ranges.
Analysis methods for many uncertain parameters in specifically rest ricted structures
(affine, tree-structured) are postponed to later chapters. New results for the design
of PID-controllers and frequency domain specifications are included as well as many
new applic ations in the case study chapters. Compared to the first edition of 1993, the
new edit ion is essent ially completely rewritten and augmen ted by contributions of new
coaut hors.
vi

Contents of the Chapters

Chapter 1 introduces an idealized crane example to illustrate analytical state space and
transfer function modelling of a plant with uncertain physical parameters q bounded
by an operating domain Q. Further, it reviews the notions of robust controllability
and observability and shows that for robust observability of the crane a sensor for
the rope angle is necessary in addition to a position sensor . The relations between
open-loop and closed-loop parametric characteristic polynomials via state and output
feedback are established. A parametric Hurwitz-stability analysis for the crane yields a
robust transparent linear controller structure with only three free controller parameters
k1 , k 2 , k3 for pendulum damping and crab motion stabilization. The crane is a represen-
tative for two-mass systems like the inverted pendulum, two coupled vehicles, a robot
joint, an aircraft rudder deflection . At the end of Chapter 1, we have a controller struc-
ture and a characteristic polynomial p(s, q, k) with uncertain plant parameters q and
undetermined controller parameters k . For Chapters 2 to 4 and 8 to 11, this parametric
polynomial serves as the primary interface between the engineering problem and the
mathematical problem of root locations of parametric polynomials.
Chapter 2 introduces the concepts of critical stability conditions, fictitiuous bound-
aries, active boundaries and non-active boundaries, all arising from the last Hurwitz
determinant: Also, the distinction between real root boundaries (RRB), complex root
boundaries (CRB) and infinite root boundaries (IRB) is introduced. The boundary
crossing concept of Frazer and Duncan then leads to the more transparent parameter
sp~e approach. The pole placement technique allows a sequential shifting of pole pairs
in invariance planes as a strategy element for stepwise improvement of eigenvalue pat-
terns. Next, the case for the parameter space mapping equations being singular, giving
rise to singular frequencies, is discussed. The singular frequencies provide boundaries
between parameter ranges in which the resulting stability regions are topographically
similar . The properties of singular frequencies are exploited for the design of PID-
controllers, where for fixed proportional gain the stability regions are convex polygons.
Chapter 3 introduces the basic ideas of the parameter space approach for design
and analysis of robust control systems based on eigenvalue specifications. First, some
relations between pole and zero locations and time responses are reviewed. The de-
sirable I'-region for the closed-loop poles is used to define f-stability with a boundary
af(a), a E [a- ; o"] . The crane example illustrates that gridding of admissible parame-
ter regions may be useful - although computationally inefficient - for analysis of a given
polynomial family. It does not give directions, however, for design steps that improve
the system. Design for two controller parameters at a time can be done by mapping
af for some representative operating conditions, e.g. the vertices of the operating do-
main Q. The set of all simultaneously stabilizing regions for the representatives is then
the intersection of the stabilizing regions for the representatives. A controller from
this intersection is chosen and the resulting closed-loop system is analyzed by mapping
I' into a plant parameter plane (e.g. load mass and rope length of the crane) . This
analysis shows if the entire operating domain Q is f-stabilized, not only the represen-
tatives. Also, a gain-scheduling approach for measured rope length is illustrated. The
migration of eigenvalues under continuous controller parameter changes is discussed. It
yields a one-to-one correspondence of open-loop and closed-loop eigenvalues as long as
Preface vii

no branching points have been crossed. Finally, a simple example shows th at controlla-
bility and observability of each plant representative does not guar ant ee that there exists
a simult aneously stabilizing controll er. This is a fundam ental difference to controller
design for a nominal plant .
While the second and third chapters give the motivation for st udying the parameter
space approach, Chapter 4 deals with the mathematical generation of the mapping
equations. First, the boundary of the f -stable region is parameterized as ar(O') , 0' E
[0'-; 0'+], where the scalar 0' plays the same role as the frequency w for Hurwitz-stability
boundaries . The polynomial p(s, q) has a root at 0' = 0'* on ar = a(O') + jw(O') if
and only if both the real and the imaginary parts, of p(a(O'*) + jw(O'*) , q) vanish for
some admissible q E Q. Instead of real and imaginary part two linear combinations
thereof are used that can be generated by a simple recursion formula . The resulting
equation has the form D(O')a(q) = 0, where D(O') is a 2 x (n + 1) matrix and a(q)
is the coefficient vector of the closed-loop characteristic polynomial. If 0' is eliminated
from these two equations, then a Hurwitz-type algebraic criterion for r-stability is
obtained. It is very complicated, however. For the parameter space mapping, th e two
equations are solved, e.g. for ql (0'), q2 (0') that generate the boundary in the (ql ' q2)-
plane for a sweep over O'. The concept of singular frequencies is generalized to the
boundary ar and is used for the design of I'-stabilizing PID-controllers yielding closed-
loop poles in a shifted left half plane or in a circle. An example with bilinear coefficient
function a(q) introduces us to the treatment of non-linear par ameter dependencies by
the parameter space approach. An example with complicated polynom ial coefficient
function s shows this particular strength of the parameter space approach. The bilinear
example illustrates t hat for non-linear coefficient functi ons in a(q), th e worst case
operat ing condition is not necessarily on the boundary of the operat ing doma in, but
may be an interior point . A Jacobian condition is introduced to ident ify the interior
candidates for th e worst case. The resultant method th at is used here is described in
Appendix A. A short discussion of ext ensions to higher dimensional par ameter spaces
concludes the chapter.
Chapter 5 now extends the scope beyond the characteristic polynomial as the in-
terface between engineering systems and their mathematical robustness anal ysis and
design. An introductory example illustrates that r-stability does not always imply
good gain and phase margins. Therefore, e-stability is defined to guarantee safety
margins for Nyquist plots from the critical point - 1 and from negative-inverse describ-
ing functions. Tangent and point conditions yield the boundaries in q-space for which
8 -boundaries are crossed. Similar conditions arise when a Popov plot does not admit
a desired sector for the non-linearity. B-stability refers to bounds on Bode magnitude
plots . Again , boundaries occur for tangent and point conditions. Mathematically, these
boundary conditions lead to two equations PI (w, ql , q2) = 0, P2(W , ql , q2) = 0 that ar e
similar to the real and imaginary part of a polynomial . Finally, in Chapter 5 the multi-
input, multi-output (MIMO) case is reduced to the same typ e of mapping condit ion.
In par ticular, the H 2 and H oo condit ions are treated.
Chapters 6 and 7 illustrate the combined use of the tools in case studies. Chapter 6
introduces us to the lateral, yaw and roll dynamics of car st eering. The essential un-
certainty is th e lateral force between tire and road . The structure of t he mathematical
viii

model allows a robust unilateral decoupling such that the lateral acceleration becomes
independent of the yaw rate. In driver assistance systems, control of the lateral accel-
eration for track following is left to the driver, while the yaw motion is automatically
controlled. Automatic car steering systems take over both tasks . The ideal concept of
skidding avoidance is based on additional rear wheel steering . If that is not available,
then several trade-offs must be made. In rollover avoidance, the uncertain height of
the center of gravity poses a robustness problem. Chapter 7 contains three case studies
in flight control. The first one directly transfers the above robust decoupling concept
to an aircraft in an engine fault situation. Common to both problems is that the yaw
disturbance causes large yaw motions followed by overreaction of the driver or pilot. In
the case of the aircraft, the robust automatic control system allows a significant weight
reduction of the vertical fin. The other two case studies deal with aircraft stabilization
and control of the short-period longitudinal mode. First, r -specifications are used, then
B-stability is specified and achieved over a large operating domain .
In Chapter 8, the value set approach is presented that is particularly suited to the
robustness analysis of polynomials with many uncertain parameters. However, there are
restrictions on the kind of coefficient functions that can be handled. A common basis
is the Mikhailov plot pOw) for Hurwitz-stability, which blows up to a family of plots
for uncertain q E Q, where the operating domain Q is a hyperrectangle (Q-box) . This
Mikhailov set must contain one stable polynomial and it must not contain the origin of
the complex pOw)- plane. The zero exclusion condition is used in Chapter 8 for proofs
of the Kharitonov and edge theorems . The Kharitonov theorem applies to interval
polynomials and requires stability of, at most, four extremal polynomials. The edge
theorem applies to polynomials with affine coefficient functions and requires stability
tests for edges of the Q-box. An edge may be checked for stability by the Bialas test.
Also, the singularity of value sets is analyzed. The construction of Mikhailov sets is done
for fixed frequencies, i.e. only the scalar w is gridded and not the higher dimensional
Q-box . For linear coefficient functions, the value set for fixed frequency is a parpolygon
(polygon with pairwise parallel edges) and the origin can enter only through the edges.
The edge result also holds for r-stability. In this case, it is more efficient to plot a root
locus for each edge rather than generalizing the Bialas test.
In Chapter 9, the value set approach is applied to non-linear coefficient functions. A
warning example at the beginning shows that unstable islands, even isolated unstable
points, can occur due to a bilinear term in one of the coefficients. Thus , simple general
results, as in the linear case, cannot be expected in the non-linear case. However, useful
tools for the computer-aided rendering of value sets are developed. For multilinear co-
efficient functions, the mapping theorem by Desoer provides a simple sufficient stability
condition. It says th at the convex hull of the value set is generated by the images of
the vertices of the Q-box. It restricts the frequency band in which the actual value
set must be constructed. Necessary and sufficient stability tests for multilinear (and
some polynomial) coefficient functions can be performed if the uncertain parameters
enter in form of a tree-structure into the characteristic polynomial p(s, q), i.e. p(s, q)
may be expressed as a sum of subpolynomials, which in turn may be products of sub-
subpolynomials, etc. until basic polynomials are reached. The sequential construction
of the value set becomes possible, even for many uncertain parameters , if each uncer-
P reface ix

tain parameter enters only into one of the basic polynomials. If such a tree-struct ure
exists, t hen it can be exploited both for Hurwitz-stability and for f -stab ility analysis.
Tree-structures arise naturally, for exampl e, in modelling of mass-spring-damper sys-
tems and t hey are preserved under feedback . Also, given polynomials may be analyzed
for tre e-structures. A useful application is the calculation of the "st ability profile" , th at
is, t he right hand boundary of the root set .
Chapter 10 introduces the stability radius as a scalar measure of vicinity to instabil-
ity. Essentiall y, a ball or box around a test point in parameter space is blown up until
it hits the stability boundary. Tsypkin and Polyak use a frequency-dependent distance
function . Also, an algebraic formulation is given for polynomial coefficient functions .
In Chapter 11, the previous results are adapted as far as possible to sampled-data
system s. A problem in the exact treatment here arises, because physical parameters
ent er exponentially into th e polynomial coefficients. A good approximation can be
achieved, however, by a Poisson series approach.
Appendix A covers some mathematical background material on polynomials, poly-
nomial equations and conic sections . Appendix B gives a short introduction to the
software tool PARADISE (PArametric Robustness Analysis and Design Interactive Soft-
ware Environment) . Th e latest release of the software can be downloaded from
www.robotic.dlr .de/control/paradise.
The bibliography shows that many foundations of the meth ods of robust control
have been laid over a long period of time by scientists in Russia and other east ern coun-
tries (Bialas, Gantmakher, Kharitonov, Mikhailov, Mitrovic, Neimark , Polyak, Popov,
Siljak, Tsypkin, Vishnegradski and many more). Only in th e last decades have these
contributions been recognized and further developed in the western literature (Ander-
son, Barmish, Bar tlett , Bhatt acharyya, Dasgupta, Desoer, Fam , Hollot , Horowitz, Jury,
Mansour, Meditch, Tempo and many more).

General Remarks
The prerequisite for the reader is an undergraduate course in feedback control systems.
We try to keep the mathematics simple. The book is suited for an advan ced under-
graduate level or for a first graduate level course in robust control. In fact , t he mat erial
was selected and used for such courses at the University of California, Irvine , at th e
Technische Universitat Miinchen and in several short courses. It was also used in con-
tinuing education courses at th e Carl-Cranz Gesellschaft (CCG) , Oberpfaffenhofen, for
participants from industry.
For the purpose of such courses, a very restrictive selection had to be made from the
large and rapidly growing literature on robust control. Therefore , many important con-
tributions and alt ernative approaches could not be mentioned . Some cross-references
are given in the form of remark s. Remarks also indicate possible generalizations , open
problems , and other suppl ements t hat are not prerequisites for understanding th e fol-
lowing sections . The beginner should ignore all remarks , th ey are intended for th e
advanced reader.
x

In the examples with physical parameters, units are given in brackets , e.g. [m]
for meter (to be distinguished from the symbol m for mass) or Is] for seconds (to be
distinguished from the complex variable s of Laplace transforms) . In calculations , the
units are omitted. We use the following units :

Physical Variable Symbol Unit

Length e meter [m]
Time t second Is]
Mass m kilogram [kg]
Moment of inertia J [kg m"]
v

Force F newton [N] = [kg - m/s 2]

Torque M newton-meter [Nm]
Velocity v [m/s]
Acceleration a [m/s 2 ]
Angles 0:, {3, . . . radian [rad]

Frequency w [rad/s]
Flight altitude h foot [ft] = 0.3048 meter
Mach number M ratio of v and velocity of sound ,
dimensionless

Acknowledgements
The authors like to thank the coauthors of the first edition of this book (1993), who are
working in other fields now. Material written by Andrew Bartlett, Wolfgang Sienel and
Reinhold Steinhauser has been integrated in the new context . Thanks also to Jessica
Laskey for typing part of the manuscript and to Stefan von Dombrowski for his technical
support.

oberpfaffenhofen Jiirgen Ackermann

June 2002 Paul Blue
Tilman Bunte
Levent Giivenc
Dieter Kaesbauer
Michael Kordt
Michael Muhler
Dirk Odenthal
Contents

1 Parametric Plants and Controllers 1

1.1 State Space Model, Linearization, Eigenvalues 2
1.2 The Leverrier-Faddejew Algorithm 7
1.3 Transfer Function . . . . . . . . . . . . . . . . 8
1.4 Robust Controllability, Observability, Feedback Structures 12
1.5 Output Feedback, Closed-loop Characteristic Polynomial 14
1.6 Hurwitz-stability, Stabilizing Controller Parameters 19
1.7 Controller Structures for Partially Known Inputs 24

2 Hurwitz-stability Boundary Crossing

and Parameter Space Approach 29
2.1 Critical Stability Conditions . . 29
2.2 The Parameter Space Approach 33
2.3 Pole Placement . . . . . 37
2.4 Sequential Pole Shifting .. . . 40
2.5 Singular Frequencies . . . . . . 44
2.6 Hurwitz-stability Regions for PID-controllers . 47
2.7 Hurwitz-stability Regions for a Compensator or Plant Subpolynomial 56

3 Eigenvalue Specifications 59
3.1 Poles, Zeros and Step Responses . 59
3.2 Root Sets, Gamma-stability . . . 66
3.3 Physical Meaning of Closed-loop Poles 73
3.4 Remarks on Existence of Robust Controllers 74
3.5 Further Potential of the Parameter Space Approach 76

4 Gamma-boundary Mapping into

Parameter Space 80
4.1 Algebraic Problem Formulation 81
4.2 Parameter Space Mapping . . . 81
4.3 Generalized Singular Frequencies 90
4.4 Gamma-stable PID-control . . . . 92
4.5 Non-linear Coefficient Functions . 96
xii

5 Frequency Domain Analysis and Design 104

5.1 Frequency Loci Specifications (Theta-stability) . . . . . . . . . 104
5.2 Mapping of Frequency Loci Margins into Parameter Space . . 121
5.3 Frequency Response Magnitude Specifications (Beta-stability) 131
5.4 Mapping of Frequency Response Magnitude Specifications into
Parameter Space 151
5.5 MIMO Systems . . . . . . . . . . . . . . . . . . . . . . . . . . 158

6 Case Studies in Car Steering 171

6.1 Tires, Braking, and Steering . . . . . . . . . . . . . . . . . . . . . . .. 171
6.2 The Two Steering Tasks . . . . . . . . . . . . . . . . . . . . . . . . . . 177
6.3 The Non-linear Single-track Model and its Robust Unilateral Decoupling 180
6.4 The Linearized Single-track Model 192
6.5 Linear Analysis of Robust Decoupling . . . . . . . . . . . 206
6.6 Skidding Avoidance Based on Robust Decoupling . . . . 215
6.7 Skidding Avoidance Based on the Disturbance Observer. 223
6.8 Rollover Avoidance . . . 238
6.9 Patents 272
6.10 Automatic Car Steering 272

7 Case Studies in Flight Control 286

7.1 Aircraft Load Alleviation in case of an Engine Out
by Robust Yaw-lateral Decoupling . . . . . . . . . . . . . . . . 287
7.2 Robust and Fault-tolerant Gamma-stabilization of an F4-E . . 303
7.3 Large Envelope Flight Control of a High Performance Aircraft 318

8 Robustness Analysis by Value Sets 334

8.1 Mikhailov Plot 334
8.2 Value Sets and Zero Exclusion . . . 338
8.3 Interval Polynomials, Kharitonov Theorem 341
8.4 Affine Coefficients: Edge Theorem . 349
8.5 Edge Theorem for Gamma-stability 356
8.6 Singularity of Value Sets . . . . . . 360

9 Value Sets for Non-linear Coefficient Functions 362

9.1 A Warning Example . . . . . . . . . . 362
9.2 Desoer Mapping Theorem . . . . . . . . . . . . . 367
9.3 Tree-structured Value Set Construction . . . . . . 370
9.4 Computer-aided Execution of Value Set Operations 378
9.5 Tree-structured Transfer Functions . . . . . 382
9.6 The Stability Profile . . . . . . . . . . . . . 389
9.7 Synopsis of Parametric Robustness Analysis 391

10 The Stability Radius 395

10.1 Tsypkin-Polyak Loci . . . . . . . . . . . . . . . . . . . . . . . . . 396
10.2 Affine Dependence: The Largest Hypersphere in Parameter Space 404
Contents xiii

10.3 Polynomial Dependence . 409

11 Robustness of Sampled-data Control Systems 417

11.1 Plant and Controller Discretization 418
11.2 Discrete-time Controllers . 420
11.3 Eigenvalue Specifications 421
11.4 Classical Stability Tests . 425
11.5 Edge Test . 437
11.6 Construction of Value Sets 439
11.7 Real Radius of Stability . 440
11.8 Single-loop Feedback Structures 440
11.9 Circle Stability . 444

A Polynomials and Polynomial Equations 448

B PARADISE: Parametric Robustness Analysis and Design Interactive

Software Environment 461

Bibliography 465

Index 477
1 Parametric Plants and Controllers

A mathematical model for plant dynamic s is the basis for analysis and design of control
systems. For linear time-invariant systems, we commonly have a state space model of
th e form
x(t) = Ax(t) + Bu(t) ,
(1.0.1)
ye t) = Cx(t) ,
with u the vector of input signals (manipulated variables), the state vector z , and
th e vector of output signals y (performance variables). Th e out put signals are often
comprised of those variables that are measured and hence available for feedback.
Another commonly used form of the model can be obt ained by th e Laplace trans-
format ion of (1.0.1)

sx(s) - x(t = 0) Ax(s ) + Bu(s ),

(1.0.2)
yes) Cx (s ).

Solving for x es) and premultipl ying by C gives the t ransformed out put vector

yes) = C( sI - A)-IX(t = 0) + C( sI - A )-IBu(s) (1.0.3)

with t.he transfer function matrix

G(s) := C(sI - A)-lB. (1.0.4)

For zero init ial conditions, x(t = 0) = 0 ,

yes) = G(s)u(s) . (1.0.5)

Remark 1.1

The plant may have a direct throughput term D , i.e. yet) = Cx(t) + Du(t) and
G(s) := C(sI - A)-l B + D , for example , an accelerometer output y of a system with
force input u . Realistically, however, we would have to include the low-pass dynamics
of the force generating motor and we are back to the system (1.0.1) without throughput
term. 0
2 1 Parametri c Plants and Controllers

A state space model (A , B , C) or a transfer function model G(s) can be deter-

mined by different approaches. The first possibility is called input-output modelling . In
this approach, t he model is obtained from experiments on the plant, which is treated
like a black box. During the experimentation, the measured signals u(t) and y(t) are
recorded . These data are then processed by identification algorithms to derive a system
model. The second approach is called analytical modelling. In this approach, the model
structure is derived from first principles of physics. A model typically depends on some
physical parameters; their values are uncertain between known lower and upper bounds.
In this book the second approach, analytical modelling , is assumed . Its advantage
is that the controller can be designed before the plant is fully operational and available
for experiments (i.e. cannot be drastically changed) . Thus, a "concurrent engineer-
ing" of plant and controller becomes feasible. Also, safety-critical experiments with the
uncontrolled plant are avoided. For a space vehicle, such experiments would be impos-
sible. Practically, we may combine input-output modelling and analytical modelling ,
e.g. some submodels of system components may be obtained by experiments with the
separated subsystem or by numerical approximation as in finite element models. Other
subsystem models can be obtained by analytical modelling . Eventually, all subsystems
are integrated in an overall system. Examples of system modelling can be found in [66],
[165] and also in case studies in Chapters 6 and 7 of this book .
This chapter begins with an idealized crane example to illustrate analytical state
space and transfer function modelling of a system with uncertain physical parameters.
Further, it reviews the notions of controllability and observability in the context of
uncertain parameters. For the crane example , this leads to a robust feedback struc-
ture with two sensors, Figure 1.6. The relations between open-loop and closed-loop
characteristic polynomials via state and output feedback are established. A parametric
Hurwitz-stability analysis finally yields a robust linear controll er structure with three
free controller parameters, Figure 1.8. Control system structures for individual require-
ments regarding the inputs reference signal , plant disturbance and sensor noise are
discussed in Section 1.7.

1.1 State Space Model, Linearization, Eigenvalues

Example 1.2 (Crane)

Consider the crane in Figure 1.1. The task of the crane or loading bridge is, for
example , to load containers into a ship. First, the load mass is only the empty hook.
The hook must come to rest above the container. Sometimes , someone helps to damp
the swaying pendulum and to fix the hook. In our approach, the placement of the hook
above the known position of the container and the damping of the pendulum shall be
performed automatically. After lifting the container, it is transported for some distance
to th e vicinity of the hatch of the ship. This motion can be commanded by feedforward
1.1 State Space Model, Linearization, Eigenvalues 3

me

Figure 1.1. Crane

control considering the limited motor force at the crab and safety constraints. Feedback
control is needed again in order to position the container above the hatch and to damp
t he motion. It must be almost at rest before it can be lowered into the ship. In this
iplication, the load mass varies widely between the weight of the empty hook and the
maximum load that the crane can carry. Also, the rope length may vary (even more for
a construction crane) . In contrast, the mass of the crab varies very little (e.g. depending
on the weight of the crane operator sitting in the gantry) . Many parameters may be
uncertain; it is important to look only at the essential ones in control system design.
The other ones may be analyzed later.
In common design approaches, a linearized model is used for controller design. It
models smail motions for positioning the hook with or without the container with suffi-
cient accuracy. It is important that the controller is designed such that the assumptions
made in linearization are not violated during the operation of the crane.
The input signal is the force u that accelerates the crab. The crab mass is me·
Further parameters are rope length f, load mass mL , and gravity acceleration g. The
position of the crab is Xl and its velocity Xl =: X2 , the rope angle is X3 , and the angular
velocity X3 = : X 4 '
The following assumptions are made for simplification of modelling :

• Dynamics and non-linearity of the driving servo motor are neglected . This as-
sumption makes sense only if the controller design guarantees that lui and litl are
not excessively large.
• The crab moves along the track without friction .
• The rope has no mass and no elasticity.
• There is no damping of the pendulum (e.g. from air drag).
• The parameters rope length f and load mass tni, are constant during each opera-
tion of the loading bridge.
4 1 Parametric Plants and Controllers

The two masses of cra b and load are coupled by the longitudinal rope force F . The
equations of motion are
crab, horizontal meXl = u + F sin X3,

cP(XI + f sin x3)

load , horizontal mL dt 2 - F sin X3 ,
d2 (f cos X3)
load , vertical mL dt 2 = -FCOSX3 + mLg.
Elimination of F yields

.. cP(XI + f sin X3)

meXI + mL dt 2 u,
d2(fcoSX3) . d 2(XI +fsinx3)
mL dt 2 sm X3 - mL dt 2 cos X3

where
d2 (XI + fsinx3)
= Xl + fX3 cos X3 - fi;~ sin X3,
dt 2
cP(f cos X3 )
- f X3 sin X3 - fi;~ cos X3,
dt 2
resulting in th e two non-linear second ord er differential equat ions

(mL + m e) XI + mLf(x3 cOS X3 - i;~ sin x3 ) u,

(1.1.1)
mL x I cos X3 + m Lfi 3

For deriving a state space description , the scalar differential equations must be solved
for the highest derivatives Xl and X3 ' Rewr ite (1.1.1) as

(1.1.2)

The determinant of the matrix on the left hand side is mL f(me + mi. sin 2 X3). It
vanishes for mL = 0 or f = O. In these two cases, the system degenerates to a second
order system because the system has only one degree of freedom of th e single mass . It is
a standing assumption in all following discussions of crane control that mL > 0, m e > 0
and e> O. Then

(1.1.3)

with

u + (g cos X3 + fx;)mL sin X3

.- me + mLsin2 X3
u cos X3 + (9 + fx~ cos x3)mL sin X3 + gm e sin X3
2
f (m e + mL sin X3)
1.1 State Space Model , Linearization, Eigenvalues 5

With the state vector

x = ::
X3
= [ :: ] =
X3
[~::: ~:l:::~;
rope angle
], (1.1.4)

X4 X3 rope angle rate

the non-linear state space model is written as

x = f(x ,u) = (1.1.5)

X4

h(X3,X4, u)

This state space model can be used for simulations of the crane motion by numerical
integration. For this purpose, numerical values for the parameters g, l, me. and me, an
initial condition for the state x(t = 0), and an input function u(t) must be chosen.
For controller design, a linearized model is better suited, therefore the non-linear
model (1.1.5) is linearized for small deflection angle X3 , and small angular velocity X4 .
Setting
cos X3 ~ 1, sin X3 ~ X3, sin 2 X3 ~ 0, x~ ~ 0,
it is easy to derive the following linear ~tate space model:
x = Ax+bu,

010
o 0 a23
(1.1.6)
A b=
000
o 0 a43

1
me
(mL +me)g 1
b4 = - - - .
mel mel
The characteristic polynomial of A
PA(8) = det(81 - A)
2(82 (1.1.7)
8 - a43)

= 8
2
[8
2
+ (1 + mL/me) gill
yields the parameter-independent eigenvalues
81,2 = 0 (1.1.8)
and the parameter-dependent eigenvalues
83,4 = ±jyh + mL/meJ9Ti (1.1.9)
6 1 Parametric Plants and Controllers

Notation

In general, we write the open-loop characteristic polynomial as

Po ( s ) = aoo + aolS + ...+ aOn-ISn-l + aOnsn . (1.1.10)

In the evaluation of PA(S) = det(sI - A), the polynomial arises in monic form, i.e,
llon = 1. In parametric representation, it is sometimes convenient, however, to avoid
rational coefficient functions like in (1.1.7) and to write the polynomial in the form

(1.1.11)

o
A polynomial may also be interpreted as a point in its coefficient space, i.e, we write

(1.1.12)

and in the monic case

1
s
Po(S) = [aT 1] (1.1.13)

For uncertain par ameters with lower and upper bounds, the following notation is intro-
duced :

me E [mL: ;m!J,
e E [t'- ; £+].

In a general context, the parameters are called ql , q2, . . . , qt and th ey are written as a
plant param eter vector

q = ~:] (1.1.14)

qe
The plant paramet er vector is bounded by an operating domain Q, which is typically a
hyperrectangle
(1.1.15)
Figure 1.2 illustrat es th e case for e= 2.
1.2 The Leverrier-Faddejew Algorithm 7

q:;

ql
Figure 1.2. The plant parameter vector is bounded by an operating domain , i.e. q E Q

The dependency of a state space model on the uncertain constant parameter vector q
is written in the general form

x(t) = A(q)z(t) + B(q)u(t) ,

(1.1.16)
y(t) C(q)z(t),

and the notation for the q-dependency of a transfer function model is

y(s) = G(s , q) u(s) . (1.1.17)

The set
G(s , Q) = {G(s, q) I q E Q} (1.1.18)
is called a plant family.
The open-loop characteristic polynomial family is defined by

P(s , Q) = det(sI - A(q)) = den(G(s, q)) (1.1.19)

("den" stands for "denominator of') . For the stability analysis, the closed-loop char-
acteristic polynomial is of interest. It also involves controller parameters k, which are
here subsumed under the q-parameters. The root set is denoted by

Roots[P(s , Q)] = {s Ip(s, q) = 0 for any q E Q}. (1.1.20)

1.2 The Leverrier-Faddejew Algorithm

For the calculation of the parametric transfer function matrix

G(s, q) = C(q)[sI - A(q)tl B(q), (1.2.1)

8 1 Parametric Plants and Controllers

the Leuerrier-Faddejeui algorithm [94] is recommended. It calculates

D(s)
(sI - A)-I
det(sI - A)' (1.2 .2)
D(s) Do + DIs + ... + Dn_ISn-I,

aOn 1, D n- I = I,
aOn-1 = -r1 trace ADn_I, D n- 2 ADn_1 + aOn_II,

1
aOn-2 -2 trace ADn-2, D n- 3 ADn-2 + Uan-2I ,

1 trace AD l , Do AD l +aOlI ,
aOl -n-l

aoo = 1
-n; trace ADo, o., ADo + aooI = o.
The open-loop characteristic polynomial (1.1.10) is in its monic form, i.e, aOn = 1. The
trace of a matrix with elements aij is the sum of the elements aii in the main diagonal.
The last equat ion D_ I = 0 serves as a check.

Remark 1.3

The Leverrier-Faddejew algorithm is not recommended for numerical calculations

where the checking matrix is not exactly zero and backwards iterations are needed to
improve the accuracy. In the symbolic calculations considered here, the checking matrix
D_ l must be exactly zero.
o

1.3 Transfer Function

Substitute (1.2.2) into (1.2 .1) to obtain the transfer function matrix

G(s) = CDoB + CDlBs + . .. + CDn_IBsn-l

(1.3 .1)
det(sI - A)
Its ij-th element ensI - Atlbj = nij(s)/ det(sI - A) is the transfer function from
input j to output i. It has the numerator degree mij = deg nij(s) :s
n - 1 and the
relative degree is defined as
{!ij := n - mij· (1.3 .2)
1.3 Transfer Function 9

The relative degree cannot be reduced by feedback or feedforward compensation.

The relative degree is directly related to the initial response to a step input. For
{J = 1, the step response starts at zero with non-zero slope, for {J = 2 the initial slope
is zero and the initial second derivative is non-zero . For a root locus, the relative
degree of the open-loop system determines the directions of the asymptotes of the root
locus branches that go to infinity. They occur under 1800 / {J, 3 . 1800 / {J, 5 . 1800 / (J • ••
with respect to the positive (negative) real axis for positive (negative) high-frequency
loop-gain.

Example 1.4 (Crane)

For the vector of transfer functions from u to the state vector X(8) = g(8)U(8) of
the crane , we get

2
8 £+ 9
-I 1 8(8 2£+g)
g(s) = (81 - A) b = £ () (1.3.3)
me PA 8 _8 2

-8 3

Now consider three output variables : load position YL, crab position uo, and rope angle
YR, for which we write the respective transfer functions:

i. Output load position (horizontal) YL = XI + £ sin X3 ::::::: XI + £X3,

gd8) = [1 0 £ 0] g(8) =
8 me
2[ £2
8 + r mL + me )r
9
(1.3.4)

Its relative degree is {JL = 4.

ii. Output crab position Ye = XI,

82£ +9
sc( 8 ) = = + ( mL+me ) 9l
[1 0 0 0]g(8
) 2[ 0 2 (1.3.5)
, 8 me<.8

The transfer function has parameter-dependent zeros at ±j The poles and J9Ti.
zeros are shown in Figure 1.3. Note that Jl
+ mL/me > 1, therefore the zero is
closer to the origin than the pole. The relative degree is oc = 2.
iii. Output rope angle YR = X3,

9R( 8) = [0 0 1 0] g(8) -
- me£8 2 + (mL+me)g'
-1 (1.3.6)

The relative degree is {JR = 2. It is not changed by the cancellation of the double
eigenvalue at 8 = O. The transfer function is second order instead of fourth order .
The cancellation occurs because the crab subsystem is not observable from the
rope angle X3 ' This can also be seen from the state space model (1.1.6). This
10 1 Parametric Plants and Controllers

jw

s- plane

Figure 1.3. Poles and zeros of the transfer function gc(s) from force u to crab position X l

model has the canonical form for separat ion of observable and non-observable
subsystems as given by Kalman [123] and Gilbert [97]

0 1 0 0 0
0 0 a 23 0 b2
:i: x+ u,
O' 0- 0 1 0 (1.3.7)
O' O' a43 0 b4

YR = [ O· O· 1 0 ] z.

The zeros with an asterisk are crucial; they indicate that th e zero input dynamics
of the crab (states Xl and X2) with arbitrary initial conditions have no influence
on the deflection angle X 3. Physically, this means that th e pendulum motion
relative to the crab does not depend on the position or (constant) velocity of th e
crab. The transfer function (1.3.6) describes only the controllable and observable
subsystem

(1.3.8)
YR = [1
Its eigenvalues at ±jy'l + mL/mc.../9!i can be shifted by feedback of X3 ' Th e
unstable cancelled double eigenvalue at s = 0 cannot be shifted by this feedback.

Th e model structure (1.3.7) is illustrated by t he signal flow diagram in Figure 1.4. We

1.3 Transfer Function 11

XI, X2

X3, X4
Crab
....... Pendulum

Figure 1.4. Signal flow diagram of the crane

will come back to the loading bridge in later sections. There we will consider that
the load mass mL and the rope length f are parameters with large uncertainty. The
gravitational acceleration is assumed fixed at 9 = 10 [m . S-2].
In principle, the parameters mL and f can be measured before each new operation
of the crane . However, this would not be practical. In the context of robust control, the
uncertain physical parameters are treated as fixed but unknown quantities for which
only lower and upper bounds are known.
o

Remark 1.5

A particularly simple form of the state space model arises for load coordinates
XL = Xl +fX3 ,

[~o ~ ~g ~
YL
= YL u (1.3.9)
0 0 1
OOa43 0

Its signal flow diagram is shown in Figure 1.5. We do not use it here, because measuring

u
Crab

Figure 1.5. Cascade representation of crane

Xl and/or Xl is simpler than measuring XL and/or XL , and the model structure is simple
anyway. In Chapter 6 on car steering, we will use such unilateral decoupling structure
and exploit exact cancellations as in (1.3.6).
o
12 1 Parametric Plants and Controllers

1.4 Robust Controllability, Observability, Feedback

Structures

It is assumed that the reader is familiar with the notions of controllability and ob-
servability of linear state space systems [122]; a convenient reference is [120] . In the
parameter-dependent case, we are particularly interested in operating conditions q for
which such properties are lost. We review the controllability and observability condi-
tions and apply them to the crane example.
Consider a plant family in state space notation
x = A(q)x + B(q)u,
(1.4.1)
y = C(q)x,
with q E Q. The pair (A(q),B(q)) is robustly controllable if
rank [B(q) A(q)B(q) ... An-1(q)B(q)]=nforallqEQ. (1.4.2)

Example 1.6
The controllability matrix of the crane is, by (1.1.6),
0 b2 0 b4a23
2 3 b2 0 b4a23 0
[ b Ab A b A b ] = (1.4.3)
0 b4 0 b4a43
b4 0 b4a43 0
(For notational convenience, the dependency of all terms on q is not explicitly indi-
cated.) The determinant of the controllability matrix is
det [b Ab A 2b A 3b] = b~(b2a43 - b4a23?' (1.4.4)
It vanishes for
b4 = -l/mcf = 0,
and for
b2a43 - b4a23 = -g/mcf = O.
Controllability is lost only under zero gravity.
o
The pair (C(q), A(q)) is robustly observable if

C(q) ]
C(q)A(q)
rank : = n for all q E Q. (1.4.5)
[
C(q)An-1(q)
1.4 Robust Controllability, Observability, Feedback Structure 13

Example 1.7

For the crane, let

0 1 0
C= 0 0 1
[ 000

i.e. all state variables except Xl are measured .

In C , C A , and also for all further C A j, i = 2,3 , . . . , the first column is zero, i.e.
the plant is not observable. More specifically, the crab position X l is not observable and
one of th e eigenvalues at s = 0 cannot be moved by output feedback u = -KCx. For
a stabilizing feedback structure, the crab position X l must be measured and fed back.
The observability matrix for th e crane with output Yl = eT x = [1 0 0 O]x is

100
010
(1.4.6)
o 0
000

As mL --> 0, then a23 = mL9/mc --> 0 (see (1.1.6)) and the rank of th e matrix drops
from 4 to 2 at mL = O. For small mL , the system is "almost non-observable". Physically,
th is means: for th e empty hook case, the horizontal force tr ansmitted from the load
to the crab is so small that its effect is hardl y recognizable in the measur ements of Xl
and X2.
For small load mass mL , an add itional sensor is necessary for damping th e pendulum
oscillations. Is it better to use a rate sensor for X3 or an accelerat ion sensor for X4 ?

i. X3
eT = [ 0 0 1 0],
eTA [ 0 0 0 1 ].

Observability is not parameter-dependent.

ii . X4
eT [ 0 0 0 1 ],
eTA [ 0 0 a43 1 ].

Observability is parameter-dependent . From a43 = -(mL +mc) / 9 /mc £, it follows

t hat observability becomes weak for large rope length e. For small rope lengt h,
th e feedback of X4 is useful, however, as will be shown in Example 3.12.
14 1 Parametric Plants and Controllers

Xl

W
Controller ~ Crane
- .- ~

Figure 1.6. Feedback structure for robust controllability and observability

We conclude that feedback of X3 is better suited for robust control.

Robust controllability and observability is guaranteed by the feedback structure of
Figure 1.6. At this point, the controller may still be linear or non-linear (e.g. adaptive) .
We will restrict the following discussion, however, to linear controllers.
Then, the next question is: what dynamical order of the controller is needed for
stabilization? A classical approach would be to choose a second order controller th at
contains a reduced order observer for reconstructing X2 and X4 and state feedback.
Such a controller has six parameters: two observer poles and four state feedback gains.
These parameters must be chosen such that the sixth order closed-loop system is stable
in the entire operating domain of the uncertain parameters. Problems of this approach
are i) the mismatch between the actual plant and its nominal model in the observer,
such that separation of state observation and pole placement does not hold, and ii)
the state feedback design for stabilizing the six uncertain closed-loop eigenvalues is
not transparent. In Section 1.6, Hurwitz-stabilization by state and output feedback
will be treated. It leads to a simpler transparent controller structure with only three
tuning parameters corresponding to three specifications for the closed-loop location of
the crab eigenvalues and for the pendulum damping (in other words, the closed-loop
natural frequency of the pendulum is not specified).
o

1.5 Output Feedback, Closed-loop Characteristic

Polynomial

Consider a state space model of a single-input plant family

A(q)x + b(q)u, q E Q,
(1.5.1)
y C(q)x,
and the control law
u = -kTy+w (1.5.2)
1.5 Output Feedba ck , Closed-loop Characteristic Polynomial 15

(see Figure 1.7). State feedback corresponds to the case C(q) = I . The closed-loop

w u y

Figure 1.7. Output feedback system

characteristic polynomial is

p( s, q, k) = det lsI - A(q) + b(q)eC(q)] . (1.5 .3)

For given q and k ; it is easy to compute the eigenvalues of A(q) - b(q)kTC(q) numer-
ically. For calculations with general q and k, a computer algebra program may be used
to evaluate p(s, q, k) .
It is not practical to evaluate the determinant as in (1.5.3) because multilinear
terms (kjkj ,j # i) in the elements of k occur , which all cancel out , such that p(s, q, k)
is linear in k . Therefore , we first rewrite p(s , q, k) in a form that is better suited for
the evaluation, also in the case of high system order .
Consider the loop of Figure 1.7 broken at u. Then the open-loop transfer function
from input u to output k T Y is

L(s ,q,k) = kTC(q)[sI - A(q)tlb(q) , (1.5.4)

and the closed-loop characteristic polynomial may be written as

p(s, q , k) = num(l + L(s , q, k)) (1.5 .5 ,

( "num" stands for "numerator of") .

The Leverrier-Faddejew algorithm (1.2.2) yields

lsI - A(q)tl = D(s, q)/po(s, q) , (1.5 .6)

where

Po(s, q) = det[sI - A(q)] = aoo + aOiS + ...+ aOn_ls n - 1 + s" (1.5 .7)

is the open-loop characteristic polynomial. Th e closed-loop characteristi c polynomial

(1.5.5) may be written as

eC(q)D(s, q)b(q))
p( s , q, k) nurn ( 1 + ()
Po s, q (1.5 .8)
Po(s. q) + eC(q)D( s, q)b(q) .
16 1 Param etric Plants and Controllers

Th us, the feedback gain k T enters linearly into the coefficients of the closed-loop char-
acteristic polynomial.
The polynomial equation (1.5.8) may also be written in terms of its coefficient
vectors in the notation of (1.1.13). Th e coefficient vector of the monic (i.e, aOn = 1)
open-loop polynomial Po is
(1.5.9)
while for the monic (i.e. an = 1) closed-loop polynomia l p , the coefficient vector is

(1.5.10)

T hen (1.5.8) reads

1 1
s s 8

sn sn 8n - 1

(1.5.11)
By matching the coefficients of equal powers of s , we get

(1.5.12)

where
(1.5.13)
Equations (1.5.12) and (1.5.13) bring th e linear relationship between k and th e open-
and closed-loop polynomials into a compact form , where the symbolic calculations for
Ware done by the Leverrier-Faddejew algorithm.
In the MIMO (multi-input, multi-output) case, the vectors k T and b(q ) are replaced
by matrices K and B (q ) and th e charact erist ic polynomi al is

p(s , q , K ) det lsi - A(q) + B (q )KC(q )]

det lsi - A (q ) + bl (q )k iC (q ) + ... bm (q)k~C (q ) ],

where bj(q ) is the i-t h column of B and ki is the i-t h row of K . Then the closed-loop
coefficient vector is
a- T = a-T kTW kTW
o + 1 1 + ... + m rn - (1.5.14)
The matrix W i is calculated from th e pair (A , bj ) . It quantifies t he eigenvalue-shifting
effect of feedback Uj = - ki x to th e i-t h input. Eigenvalue shifts of feedback to th e
other inputs are just sup erimposed.
In the expansion of this determinant , multilinear terms kjjklm occur only if i i- f
and j i- m. Linear ity with respect to th e elements of K can be preserved by ste pwise
design, in which in each design step, feedback from several sensors to one actuator (i.e.
elements from one row of K ) or feedback from one sensor to several actu at ors (i.e.
elements from one column of K ) is assumed.
1.5 Output Feedback, Closed-loop Characteristic Polynomial 17

Ex am ple 1.8 (Crane)

Consider the motion of th e crane (1.1.6) in the neighborhood of an equilibrium

point x = [w 0 0 of, where w is a constant reference input for the crab position. The
reference input w will be further discussed in Section 1.7.
By (1.1.6), the matrices of the state space model are

[~
0 1 0 0
0 0 a23 0
A= , b=
0 0 0 1
0 0 a43 0 b4

a23 --!!!:L.
meg , b2 = Je ,
_ (mL +mc)g
a43 - - meR b4 =-rib
me '

and C = I for st ate feedback. The Leverrier-Faddejew algorit hm yields

w = [Dob D1b D 2b D 3b]

-a43b2 + a23 b4 0 b2 0
0 -a43 2 + a23
b b4 0 b2
= 0 0 b4 0
0 0 0 b4

[9Tf
0 lime 0
gl mef. 0 lime
= 0 -l/m e f. 0
0 0 -l/mef.

II
9 0 f.
1 0 9 0
(1.5.15)
mef. 0 0 -1
0 0 0

The crane is controlled by state feedback

u = k1(w - xd - k2X2 - k3X3 - k4X4 . (1.5.16)

18 1 Parametric Plants and Controllers

The closed-loop characteristic coefficient vector is, by (1.5.12),

aT = [0 0 (mL+me)9
mel
0]

+ [k l k2 ka k4 1 [~~
o
~ ~ 0 -1 0
&
f:me
(1.5.17)

o 0 0-1
= & [k
f:me 19
k29 (mL + me)g + kif - ka k2f - k4 ]. -

For f > 0 , me > 0, we can replace the monic polynomial p( s) with rational coefficient
functions by the non-monic polynomial

p(s) := fmeP(s) ao + als + a2s2 + aas3 + a4s4,

ao k19 ,
k 29,
(1.5.18)
(mL + me)g + kif - ka,
a3 = k 2f - k4 ,
fmc ·

Equation (1.5.18) shows how the gains k l to k 4 affect the coefficients of the characteristic
polynomial and thereby the eigenvalues of the closed loop.
o

Remark 1.9

The uncertain parameter mL and the gain ka only enter a2. Assuming that mL can
be measured or estimated, the feedback

(1.5.19)

makes the eigenvalues independent of mL . For a measured or estimated rope length f , it

is not so easy to see how gain-scheduled feedback improves the robustness with respect
to f .
We will come back to gain-scheduled control laws in Section 3.5. Here, we first
pursue the more ambitious goal of robust stabilization and control without knowledge
of the uncertain parameters mL and e.
o
1.6 Hurwit z-stability, St abilizing Controller Parameters 19

1.6 Hurwitz-stability, Stabilizing Controller

Parameters

This section reviews the use of algebraic meth ods for testing the stability of a known
polynomial. There are three main originators of algebraic stability tests . The first alge-
braic test was derived by Hermite [1071 in 1856. Partly due to its abstractness, Hermite 's
paper went unnoticed by most people interested in stability of engineering systems. In
1875, Routh [174) independently presented another algebraic stability test. Unaware
of Routh's work, Hurwitz [113) removed the abstract ion from Hermit e's method and
produced a more at tractive stability test in 1895. Of the three methods, this section
shall present only the Hurwitz method. This method admits th e simplest generalizati on
to uncertain polynomials.
Hurwitz has related the stability of an n-th order polynomial

an > 0, (1.6.1)

to a set of determinants ~i = det Hi . Th e determinants come from the following set

of n Hurwitz matrices:

(1.6.2)

This pattern continues until an n x n matrix is obtained. For n even, th is last matrix
H n has the form

an- l a n- 3 a n- 5 al 0 0 0 0
an a n-2 a n-4 a2 aD 0 0 0
0 an -l a n -3 a3 al 0 0 0
H n = 0 an a n- 2 a4 a2 aD 0 0

0 0 an -l a n- 3 a n- 5 a3 al 0
0 0 an a n -2 a n-4 a4 a2 aD
20 1 Parametric Plants and Cont rollers

while for n odd , it has t he form

an-l an-3 an-5 o o 0
an an-z an-4 o o 0
0 an-l an- 3 o 0
H n=
0 an an-z o 0
al 0
o
As an example, if the polynomial is of order n = 4, the largest Hurwitz matrix is
a3 al 0 0
a4 az ao 0
H 4=
0 a3 al 0
0 a4 az ao
while for n = 5, the largest Hurwitz matrix is
a4 az ao 0 0
a5 a3 al 0 0
H 5= 0 a4 az ao 0
0 a5 a3 al 0
0 0 a4 az ao
Th e relati on of th ese matrices to stability is given in th e following theorem. No proof
shall be provided for this well-known result of Hurwitz .

Theorem 1.10 (Hurwitz)

An n-th order polynom ial (1.6.1) is stable if and only if
det Hi > 0 for all i = 1, 2, . . . , n . (1.6.3)
o
Note th at , for both n even and n odd , it follows from expanding det H n by its last
column th at
det H'; = aodetH n - l . (1.6.4)
Thu s, the stability condition (1.6.3) is equivalent to det H i > 0 for i = 1,2 , . . . , n - 1,
and ao > O.
Given a polynomial with known coefficients ao, ai , az, . . . , an, it is straightforward
to apply the Hurwitz- stability test. The result ing determinant s are real numbers th at
can be easily compared with zero. If the coefficients depend on uncert ain parame ters,
ao(q), al(q) , az(q), .. . , an(q), it is still possible to form the Hurwitz determinants
symbolically. However, the resulting inequaliti es are - except for some simple cases like
th e crane - not very tractable.
A useful necessary condit ion is th e positivity of all coefficients ai.
1.6 Hur witz-stability, Stabilizing Contro ller Param eters 21

Th eorem 1.11
A stable polynomial (1.6.1) of degree n satis fies

for all i = 0, 1,2 , . . . , n - 1. (1.6.5)

o
Proof
The proof of thi s th eorem follows from t he fact th at (1.6.1) is a real polynomial
and can be factored as
m
II(s- a, - jw;}(s - ai + jWi ) II
n

p(s) = an (s - ak)
i=1 k= 2m+1

m n

= an II (S2 - 2ais + a; + w;) II (s - ak)'

i= 1 k= 2m+1
where the a's and w's are real numbers . If this polynomial is stable, then the a's
are all negat ive. This implies th at all th e subpolynomials

S2 + (- 2ai)S + (a; + w?) ,

s +( - ak),
have posit ive coefficients , and hence their product p(s) must have only positive
coefficients .
o

The positive coefficient criteria of Th eorem 1.11 can be used to reduce t he number of
determ inant criteria in Theorem 1.10. This simplificat ion is due to Lienard and Chipart
[132]' for a simpler proof see [941.

Th eorem 1.12 [Lienard, Chipart)

Necessary and sufficient conditions for a polynomial

to be stable can be given in one of the four following forms:

LCI ao > 0, a2 > 0, ; ~I > 0, > 0,

~3 ,
LC2 ao > 0, a2 > 0, ; ~2 > 0, ~4 > 0, ,
(1.6.6)
LC3 ao > 0, a l > 0, a3 > 0, , ~I > 0, ~3 > 0,
LC4 ao > 0, al > 0, a3 > 0, , ~2 > 0, ~4 > 0, • •• J

where ~i = det H i.
o
22 1 Parametric Plants and Controllers

Example 1.13

For the fourth order example of the crane with a4 = f mc > 0, the forms LCI and
LC3 are convenient because only the determinant of a 3 x 3 matrix must be evaluated.
With t he coefficients of (1.5.18), form LC3 yields

aD k l9 > 0,
al = k29 > 0,

fl.1 = a3 k2f - k4 > 0,

a3 al 0
fl.3 = detH 3 a4 a2 aD > O.
0 a3 al

The evaluation gives

(1.6.7)

Remark 1.14

Form LCI yields a2 = (mL + mc)9 + ki f - k3 > 0 instead of al = k29 > 0 and
the conditions a D > 0, a3 > 0, fl.3 > 0 as above. At first glance, it is surpris ing that
the inequalities al > 0 and a2 > 0 serve the same purpose in the two forms of stability
tests. This puzzle will be solved in Section 2.1 on critical stability conditions.
o
Equation (1.6.7) shows that mL and k 3 enter only into th e combination mL9 - k 3
as contained in a2' Thus, the effect of a variation of the load mL on the eigenvalue
location can be exactly compensated by an adjustment of k 3 • However, we want to
avoid the measurement or estimation of mL . Also, we will see lat er that actuator
constraints lead to a solution that leaves the fast system (mL = mt) fast and the slow
system (mL = mi) slow, see the open-loop eigenvalues in (1.1.9) and Figure 1.3. Such
constraints would not be observed if the same eigenvalues are prescribed for all loads.
The stabilization analysis gives a new insight for the controller structure, that is,
k2 > 0 is required for stabilization. The feedback signal X2 = XI must eith er be mea-
sured or generated by controller dynamics , e.g. by low-pass filtered different iation of X l
by X 2 ( s) = I:Ts XI (s) with a small time const ant T. The time constant is chosen inde-
pendently of oth er feedback gains as a trade-off between suppression of measurement
noise in X I and maintaining the stabilizing effect of the ideal differentiator with T = O.
Regarding k3 and k4 , interest ing special cases of fl.3 arise for k4 = 0:

(1.6.8)
1.6 Hurwitz-stability, Stabilizing Controller Parameters 23

and for k3 = 0 and k4 = 0:

~3 = k~llmL > 0, (1.6.9)
i.e, the output feedb ack controller

stabilizes all cranes as modelled by (1.1.6) . The non-robust observability of X3, X 4 from
X l and X 2, however, indicates that t he above controller cannot give sufficient damping
to the pendulum in the empty hook case.
For k«, there is an upper bound k 4 < k 2e resulting from a 3 > O. Robust stability
e
for small rope length requires k 4 :::; O. For large Ik4 1, the dominating term in ~3 is
-gklk~ . Thus, k 4 = 0 is a good choice for robust satisfaction of ~3 > o. Robustness for
small loads then requires k 3 < 0, i.e. positive feedback of th e rop e angle. The resulting
controller structure is shown in Figure 1.8.

Xl
W
- GJ -
u
Crane

~
~

k2 s
l +Ts

Figure 1.8. Robust controller stru cture for the crane with k l > 0, k 2 > 0, k3 < °
The gain k l is chosen indep end ently of k 2 and k3 by a trade-off between fast posi-
tioning and a bound on lui. For a transport transition from

to

typically t he maximum of u(t ) is u(O) = k, for a good controller design . The limita-
tion of lu(t )1 will be further discussed in Sect ion 1.7. Also, k l should not be so large
that the crab eigenvalues approach the pendulum eigenvalues, as this would enforce
undesirable fast back and forth motions of the crab. The gain k 2 provides damping
for the crab motion and - for medium and large rope lengths - also for the pendulum
mot ion. A negat ive value of t he gain k 3 also provides pendulum damping for small
24 1 Parametric Plants and Controll ers

rope length e. There is a mutual influence between k2 and k3 such that these controller
par ameter s should be designed together. Finally, T is increased beginning from zero,
and its influence on the closed-loop eigenvalues is tracked.
The controller structure of Figure 1.8 is:

i. Lean: only three essential parameters must be tuned.

ii. Transparent: each controller parameter has a well-defined influence on the crane
dynamics as long as the dynamic ranges of the two subsystems are kept separate.
iii. Robust: also large parameter variations in f and mL do not destabilize the sys-
tem. The low-pass filtered differentiation is not parameter-dependent. This is in
contrast to the observer with the prob lem of a mismatched model.

The closed-loop transfer funct ion is obtained by straightforward calculation as

(1.6.10)

Note that we have not collected t he terms in t he powers of s as usual in numerical

calculations, because the symbolic expression would get more complicated. For later
investigations in Chapter 9, it is even desirable to write the transfer function such that
each parameter has only a local influence in one subpolynomial. This leads to a form
that resembles continued fraction s
1
(1.6.11)
1 [2
1 + -k s m e
k2s s2(mL9 -k3)] '
+ -1 + T s + ---'~2l:-=--':':"
j s +9
Each parameter m e , mL , f , k j , k 2, k 3, and T enters only once.

Rem ark 1.15

Feedback structures for other plants can be derived similarly by analyzing the para-
metric contro llability and observability for the worst plant parameters. The crane is
typical for two-mass systems as they arise for the inverted pend ulum , two coupled
vehicles, a robot joint, or an aircraft rudder deflection. 0

1. 7 Controller Structures for Partially Known

Inputs

In Sections 1.4 to 1.6, output feedback structures have been assumed . These structures
are convenient for the specification of closed-loop eigenvalue regions as they will be
discussed in Chapter 3. No knowledge of the spectrum of input signals is assumed .
Then, a good eigenvalue location guarantees that th e initial conditions resulting from
1.7 Controller Structures for Partially Known Inputs 25

previous input signals decay rapidly and well damped to th e equilibrium state. If,
however, some assumptions on the spectra of reference, disturbance and sensor noise
inputs can be made , then the control system may be tailored to specifications in different
frequency bands. It is then convenient to assume the controller structure in form of
transfer functions . In this section , we will first discuss the model reference approach for
the reference input and then feedback structures for disturbance and noise rejection .
Consider the control system of Figure 1.9; the principle is first explained for a
nominal plant model G(s) . K(s) is the feedback controller and F(s) is a prefilter .
Gd(s) is a desired transfer function for
GF
(1.7.1)
Gr y = 1 +GK
(we introduce the notation Gab(s) for the transfer function from signal a to signal b,
the complex variable s is omitted) .

r w u y

Figure 1.9. Modified twodegree of freedom structure with reference model Cd

The design goal Gr y = Gd can be satisfied only if Gd satisfies three feasibility con-
straints [185] .
I. Gr y must have at least the same relative degree as the plant G, otherwise F would
not be realizable.
ii. Gr y must have at least the same dead time e- T s as the plant G, otherwise F
would not be causal.
iii. Gr y must contain all zeros of G in the (closed) right half plane, because the non-
minimum phase zeros (including zeros on the imaginary axis) cannot be cancelled.

The design procedure for model reference following then is as follows [111], [128] :
1. Specify a reference transfer function G d that meets the feasibility constraints.
2. Choose the prefilter
(1.7.2)

3. The ideal open-loop transfer function

(1.7.3)

would be obtained for e = O. Then K has no influence on Gr y . K is designed to

keep e small; this design step is independent of F and Gd .
26 1 Parametric Plants and Controllers

d
w=o y

Figure 1.10. Plant disturbance d and sensor noise n

Now consider T == 0 and a plant disturbance d at the plant output and sensor noise n
as shown in Figure 1.10. Assume that d and n cannot be measured. There is, however, a
control system structure specifically suitable for disturbance rejection , see Figure 1.11.

d
w=o u y

~ 11+-_ _--<>-
Q --11 n
G-I Ik-....¢.ol~
~~--ll G-I(n+d) I I

Figure 1.11. Disturbance rejection feedback structure

The inverse G-I is in general not realizable, but in combination with a filter Q it
can be made realizable . Q must meet the same feasibility constraints as the reference
transfer function Gd in terms of relative degree, dead time and zeros in the closed right
half plane of G, otherwise the controller transfer function QG- I would not be realizable
and internally stable. Within these constraints, Q can be chosen by the designer for a
trade-off between the influences of d and n on y. In the nominal case, we have

y = (1 - Q)d - Qn. (1.7.4)

Both Q and 1 - Q should be small, but the sum of the two terms is 1. This conflict will
be dealt with in terms of sensitivity functions and complementary sensitivity functions
in Chapter 5. Typically, 1 - Q is made small at low frequencies and Q is made small
at high frequencies. In some applications, the frequency of d is known such that a dis-
turbance rejection system may be specifically tuned. An example is vibration isolation
of a helicopter from the rotor excitation , where the rotor frequency is also controlled .
1.7 Controller Structures for Partially Known Inputs 27

Another example is the periodic verti cal perturbation on a magnetically levitated train
from the track that hangs like a garland between its supports. The control system
structure of Figure 1.11 will be used in Section 6.7 for skidding avoidance by act ive car
steering, and in Section 7.3 for longitudinal flight control.
So far, we have assumed a nominal plant model G, such that G- I in the feed-
back path is known. Deviations of the real plant G from its nominal model Gn with
G- I replaced by G;;I in the feedback path may have different origin and appropriate
descri ptions.
For uncertainty of physical parameters (like mL , f , me in the crane example), the
perturbation structure must be derived by analyt ical modelling. In Chapter 6, a car
steering model with the uncertain parameters velocity, mass, moment of inertia, and
road-tire contact will be derived. Only the effects that have been modelled are consid-
ered and the order n of the state space model is fixed.
On the other hand, there are unmodelled uncertainties like neglected actuator and
sensor dynamics, neglected details of the plant like flexible structure modes , etc . Here,
a standard uncertainty structure with dynamic model uncertainty ~m may be chosen
to describe the deviation of G from the nominal G n in a vague, non-physical sense as
an aggregation of several model inaccuracies. Typical standard structures are :

Additive perturbation G= Gn+~m ' (1.7.5)

Multiplicative perturbation G= Gn(l + ~m) ' (1.7.6)
Gn
Feedback perturbation G= (1.7.7)
1 +Gn~m

For thecrane example, think of the neglect ed actuator dynamics as a multiplicative

perturbation. Elasticity of the rope may be approximated by an additive perturbation.
Typically, the low-frequency control system dynamics are dominated by uncertain phys-
ical parameters in a given model structure, and the high-frequency dynamics can only
be described by unmodelled dynamic uncertainties ~m, denoted non-parametric uncer-
tainty or unstructured uncertainty. A useful assumption that we will apply in Chapter 5
is a frequency response ~m Ow) with bounded magnitude and arbitrary phase.
A combination of the structures of Figure 1.9 and Figure 1.11 is shown in Figure 1.12.
28 1 Parametric Plants and Controllers

d
T I F I w u I C I y
I I - - I I

---1 Q I -
I
I C- 1 I
I n I
n

I
I K ~-
I Cd I
I I

Figure 1.12. Control system st ruc ture for separation of design tasks. F for inp ut T , Q for
a tradeoff be tween inputs d and n an d K for rob ust stabilizatio n, Cd for mod el reference
dyna mics
2 Hurwitz-stability Boundary Crossing
and Parameter Space Approach

2.1 Critical Stability Conditions

The basic idea of critical stability conditions is that a starting point in form of a stable
characteristic polynomial p(s, q), q = qo is given. Assume that the (real) coefficients
of p(s, q) are continuous in q. Then also the roots of p(s,q) are continuous in q, i.e.
they cannot jump from the left half plane to the right half plane without crossing the
imaginary axis. The stable neighborhood of qo is bounded by the values of q, where
for the first time one or more eigenvalues cross the imaginary axis under a continuous
variation of q starting at qo. Crossing of eigenvalues over the imaginary axis can occur
in one of three ways: at s = 0, at s = 00 and at s = ±jw.
An example of an infinite root boundary at s = 00 is the combination of the crane
with an inverted pendulum. Replace the rope by a rod and shift the load mass through
f = 0 to negative values of f. The eigenvalues at S3,4 = ±h/1 + mL/mc .Ji1i (see
(1.1.9)) go to ±joo and return as a real pair of eigenvalues from +00 and -00. The
infinite root boundary (IRB) is characterized by a degree drop at an(q) = 0, see (1.1.11).
For the real root boundary (RRB) at s = 0 and the complex root boundary (CRB)
at s = ±jw, Frazer and Duncan [89] have shown that the critical stability boundary
is det H n (q) = O. The result is presented as the following algebraic boundary crossing
theorem.

Theorem 2.1 (Frazer, Duncan)

The family of polynomials

P(s, Q) = {ao(q) + al(q)s + . . . + an(q)sn Iq E Q} (2.1.1)

with continuous real coefficient functions aj (q) is robustly Hurwitz-stable, if and

only if:
i. There exists a stable polynomial p(s, q) E P(s, Q).
ii. an(q) =!' 0 for all q E Q.
30 2 Boundary Crossing and Parameter Space Approach

iii. detHn(q) =f 0 for all q E Q.

o
Proof

Necessity of the three conditions follows from the Hurwitz- stability condition. To
prove sufficiency, it shall be shown that satisfaction of Conditions ii and iii implies
that P(s, Q) does not have roots on the imaginary axis. Consider a polynomial

(2.1.2)

with real coefficients. It may be written as

where

Peven(S2, q) = ao(q) + a2(q)s2 + a4(q)s4 + .

podd(S2,q) = al(q) +alq)s2+ as(q)s4 + .
The polynomial p(s, q) has a root at s = ±jw if and only if both the real and the
imaginary part of

(2.1.3)

vanish. Note that s = 0 is a root of SPodd(S2, q) resulting in the RRB at ao(q) = O.

For s =f 0, the common root condition for Peven(s2, q) and podd(S2, q) is written in
terms of a new complex variable v := S2

ao(q) + a2(q)v + a4(q)v 2 + = 0,

(2.1.4)
al(q) + a3(q)v + as(q)v 2 + = O.
These two polynomials share a root if and only if their resultant matrix has a zero
determinant (see Appendix A). If n is even, then the n x n resultant matrix is

an-l an-3 an-s al 0 0 0 0

0 an-l an-3 a3 al 0 0 0

0 0 an-l an-3 an-s a3 al 0

R(q) = (2.1.5)
an an-2 an-4 a2 ao 0 0 0
0 an an-2 a4 a2 ao 0 0

0 0 an an-2 an-4 ... a4 a2 ao

where the dependence of the coefficients on q has been omitted for brevity. By
interchanging rows, it is possible to transform R(q) into the Hurwitz matrix
2.1 Criti cal Stability Conditions 31

Hn(q). (The relat ion of R(q) and Hn(q) holds both for n even and for n odd. )
This relationship implies that

det R( q) = ± det H n (q). (2.1.6)

By Condition 3 of the theorem, det H n (q) =1= 0 so t he determinant of the resultant

matrix is also non-zero. Therefore,

pOw, q) =1= 0 for all w. (2.1.7)

This implies that no p(s) E P(s, Q) has a root on the jw-axis.

Theorem 2.2

Note that by det H n = ao det H n - 1 , the three conditions of Theorem 2.1 may be
replaced by the following four conditions:
i. There exists a stable polynomial p(s, q) E P(s , Q).
ii . an(q) =1= 0 for all q E Q.
iii. ao(q) =1= 0 for all q E Q.
iv. detH n _ 1 (q) =1= 0 for all q E Q.

Fictitious and Non-active Boundaries

A relationship between det H n-I and the roots of p(s) is given by Orlando 's formula
[164], [94) . Given an n-th order polynomial in factorized form
n
p(s)=anIT(s-si) , Si = O"i + jwi ' (2.1.8)
i=1

Orlando's formula relates detH n_ 1 to the polynomial's roots as follows:

IT
n
detH n- 1 = (_1)"(n+I )/2 a~-1 (s, + Sk) . (2.1.9)
i,k=1
i<k

If the polynomial has a pair of roots on the imaginary axis at 0"1 = 0"2 = 0, W I = -W2 ,
then the product term SI + S2 is zero and hence det H n-I also equals zero.
The condition detHn_1 = 0 is also satisfied for a real symmetric pair 0"1 = -0"2 ,
WI = W2 = 0 or for two complex and symmetric pairs SI/2 = 0"1 ± jW l , S3, 4 = - 0"1 ± jWI .
Thus, the condit ion det H n-I (q) = 0 not only generates the boundary in q-spa ce, where
32 2 Boundary Crossing and Parameter Space Approach

a root on the imaginary axis occurs, but also fictitious boundaries at complex frequencies
w . They cannot, however, intersect the stable neighborhood of qo, because one or more
eigenvalues have to cross the imaginary axis first before reaching a symmetric pattern
with respect to the imaginary axis with unstable eigenvalues.
The condition det H n (q) = 0 is also satisfied for a polynomial with unstable roots
plus a pair of roots on the imaginary axis. This situation gives rise to a non-active
boundary.

Example 2.3

e
For the crane example, let me = 5, mL = 5, = 2, k l = 1, k4 = 2. Find the set of
stabilizing k z, ks- The characteristic polynomial of (1.5.18) has the coefficients

ao = 10,
al lOk 2 ,
az 102 - k 3,
a3 2kz - 2,
a4 10.

The Hurwitz determinant ~3 = det H 3 is

~3
z
al aZ a3 - aoa 3 - a4al
z
20 [50k~ - 98k z - 2 - k 3kz(kz - 1)].

~3 vanishes for
k _ 50ki - 98k z - 2
3- kz(kz-l)
This boundary is plotted in Figure 2.1 It consists of three branches that divide the
plane into tree regions with ~3 > 0, and one complementary region with ~3 < O. For a
test point on the upper branch, the polynomial has a symmetric real pair of roots; this
is a fictitiuous boundary. For a test point on the left branch, the polynomial has an
imaginary pair of roots at ±jw, 0 < w < .j5 and two unstable roots; this is a non-active
boundary. Only the right branch is an active boundary corresponding to two stable
roots and one pair on the imaginary axis at ±jw, w > .j5. The real root boundary
(RRB) ao = k l9 = 10 > 0 and the infinite root boundary (IRB) a4 = eme = 10 > 0 do
not exist in this example .
The example also gives an explanation for the puzzle on the different Lienard-
Chipart conditions in Remark 1.14. Both LCI and LC3 require ao > 0 (satisfied
everywhere) and a3 > 0, i.e. kz > 1, which rules out the left and the upper region
with ~3 > O. The additional conditions from LCI (az > 0 =} k 3 < 102), and LC3
(al > 0 =} kz > 0) are non-active in this example.
The result of the boundary crossing theorem may also be illustrated by Figure 2.1.
If a stable starting point p(s, qo) is chosen, e.g. k z = 3, k 3 = 0, then the first crossing
of the imaginary axis under continuous variation of k z and k 3 occurs on the right lower
2.2 Th e Param et er Space Approach 33

bra nch of t he cur ve b. 3 = o. T he ot her th ree regions do not contain a st able start ing
point.
o

O<w <V5
V5<w<oo

Figure 2.1. The boundary t..3 = 0 for the crane example

2.2 The Parameter Space Approach

A drawback of the algebra ic boundar y crossing approach is that the condi t ion
b.n-1(q) = detHn-1 (q ) = 0 for large n leads to complicated symbolic expressions.
even for simple dependence of the coefficients a i on q. A second dr awback is that
the above condit ions also generate ficti ti ous boundari es. The reason is that the real
frequency w has been elimina ted from the two equations (see (2.1.3))

Re pOw, q ) ao(q) - a2(q)w2 + a4(q)w4 - = 0, (2.2.1)

Im pOw, q ) w [al (q) - a3(q)w2 + a5 (q)w4 - J=0

by t he resultant condit ion detR(q ) = ±detHn(q) . Therefore, complex values for w

ar e also included in t he boundary det H n (q) = O.
The basic idea of the parameter space approach is not to eliminate w ,
but to leave it as a parameter that is restricted to real non-negative values.
The stability boundaries in a two-d imensional slice through the q-space ar e calculated
by a sweep over w .
34 2 Boundary Crossing and Parameter Space Approach

Example 2.4

In the previous crane example, the two equations (2.2.1) are

Re p(jw, k2 , k3 ) 10 - (102 - k3)W2 + lOw4 = 0,

Im p(jw, k 2 , k 3 ) = 2]
W [10k 2 - (2k 2 - 2)w = O.

The equations are solved for

_w 2
=
5 -w2 '
-10 + 102w2 - lOw4
w2
For a sweep over real values of w, only the two lower branches of ~3(k2, k3) = 0 in
Figure 2.1 are generated in a parameterized form. The fictitiuous branch of Figure 2.1
is not generated by real w values. Figure 2.2 shows the two lower branches with the
parameter w. In the lower left region, there does not exist a stable polynomial, and
Condition 1 of Theorem 2.1 is violated .
k3
W = 1.58 100

w = 2.75

Figure 2.2. Parameterized complex root boundary for the example of Figure 2.1

o
The previous example was particularly simple, because only the CRB was active.
Therefore, we now look at an example where the RRB is also active.

Example 2.5

In the crane example, let me = 5, mL = 5, f = 2, k3 = 10, k4 = 1. Find the

stable region in the (k}, k2 )-plane. The characteristic polynomial of (1.5.14) has the
coefficients
2.2 Th e Parameter Space Approach 35

lOk2 ,
90 + 2kl ,
2k2 - 1,
10.

The equations for the real and imaginary parts read as

Re p(jw, kl , k2 ) = lOki - (90 + 2k1 ) W 2 + lOw 4 = 0,
(2.2.2)
1m p(jw, kl , k2 ) = w [lOk 2 - (2k2 - l)w 2 ] = O.
For w = 0, the equat ion for the imaginary part is always satisfied and the corresponding
RRB is kl = O. The CRB is described by

2(w2 - 5)'
Figure 2.3 shows the parametric stability boundary in the (k\, k2)-plane . For k1 = 0,

B k2 A

RRB p(s, qo)

w= 2.8 X
w 3.13
1 w 3.45
CRB
w=O
10 kl
w=0.28 w - > J5
D

Figure 2.3. At w = 0 the CRB branches off from the RRB

there is a real root at s = O. If also k2 = 0, t hen this becomes a double root at

s = O. This is the point, where the CRB branches off from the RRB. The CRB passes
through infinity at w = J5 and intersects the RRB a second time for w = 3 at kl = 0,
k 2 = 1.125.
Starting from a stable p(s, qo), e.g. at kl = 20, k2 = 2, a continuous variation of
kl and k2 brings either a real root to the active part of the RRB for k2 > 1.125 or a
complex pair to the active part of the CRB for kl > 0 (i.e. w > 3). The bound aries
decompose the (k\, k2) -plane into five regions, designated as A . . . E in Figure 2.3. A is
36 2 Boundary Crossing and Param eter Space Approach

the only stable region. B has an unstabl e real root , C has three roots in th e right half
plan e, D has four unstable eigenvalue s, an d E has two of th em. 0

The boundary crossing concept may now be formulat ed as:

Theorem 2.6 (boundary crossing)

The family of polynomials P(s, Q) is robustly st abl e, if and only if:

i. There exists a stable polynomial p(s, q) E P( s, Q).
ii. jw (j. Roots [P(s, Q) I for all w 2': O.
(Roots [P(s , Q)] denotes th e set of all roots of p(s, q) for all q E Q.)

The idea of a graphical representation of stability regions in a plane goes back to

Vishnegradsky [199] . In 1876, he studied th e stability of Watt 's flyball governor for
the steam engine. He analyzed th e effect of the two middl e coefficients of a third de-
gree characteristic polynomial. Neimark [156] generalized t his approach to n-th degree
polynomials depending linearly on two parameters and called it D-decomp osition, be-
cause the parameter plane is decomposed into regions with a fixed numb er of st able
roots. Mitrovic [149] extended th e paramet er spac e meth od to ot her eigenvalue regions
th an the left half plane, thereby relating cont roller paramet ers to ste p response speci-
fications. Several further results and references along t his line ar e reported in Siljak 's
book [200] . In 1980, the first au thor [2J used th e param et er space approach for th e
design of robust controllers, that meet eigenvalue specifications simultaneously for sev-
eral plant representatives (operating point s or nominal and faulted sensor situations).
The solution is the intersection of the admissible controller parameter sets for th e indi-
vidual representatives. A pragmatic design approach is, for example, to find a robu st
controller that simultaneously satisfies eigenvalue specifications for the four vert ices of
t he operating domain Q in Figure 1.2. In an ana lysis ste p, the eigenvalue boundaries
are then mapped onto the (ql' q2)-plane in ord er to check if the entire continuum of Q
meets the eigenvalue specifications. If necessary, the design st ep must be rep eated with
additional representatives of Q. The attitude may be describ ed as: be an optimist in
design and a pessimist in analysis!
Historically, paper-and-pencil graphic methods have a long tradition. The use of
computers shifted the attention towards numeri c optimization methods. This typicall y
required the formulation of all trade-offs in a scalar perform ance criterion. Engineers
continued, however, to visualize results in parameter spaces, because a figure tells more
th an an "optimal" point. Also, th e opt imizat ion at the low level of controller design
leaves no chance for a trade-off with requ irements from oth er fields dealing with dynam-
ics, e.g. structural dynamics, fluid dynamics, flight mechan ics, multi-body dynamics.
In a concurrent engineering design process, optimiz ation should be used on a high level
of syst em performance. Ideally, th e individual fields should provide sets of admissible
solut ions t hat satisfy the respective specificatio ns. Today, gra phic methods are easily
2.3 Pole Placement 37

computerized because graphical interfaces support the problem formulation and sym-
bolic computer programs (MATHEMATICA, MAPLE, etc.) take care of the generation of
mapping equations and producing the figures. And we can take advantage of the fact
that engineers have computers with rapidly growing 2D and 3D graphic capabilities on
their desk. Also, the use of colors opens new possibilities for color coding. Thus, the
design engineer can fully concentrate on the interpretation of figures and on decisions
for the next step in design or analysis. In this sense, the parameter space method is a
very useful tool for the control engineer, but not an automatic design procedure that
replaces her or him. Also, it is not meant to replace other useful tools.

2.3 Pole Placement

In Examples 2.4 and 2.5, the 2D-subspace for representing results was determined by
fixing all but two controller parameters. A very useful alternative is to fix all but two
eigenvalues in a design step . The design process then consists of sequential pole shift-
ing steps . This approach becomes feasible by combination of parameter space and pole
placement methods as will be explained in Sections 2.3 and 2.4. First, the pole place-
ment approach is formulated in a convenient way, where the closed-loop characteristic
polynomial is specified in factorized form.
In 1960, Kalman [122] introduced the definition of controllability and observability
and tests for these properties. The concept of placing the poles of a single-input system
in feedback canonical form appeared in [173] . The relation "controllability implies pole
assignability" for a single-input system was published in [45] . It was further formalized
by transformation to feedback canonical form in [1231. Another line of research places
the closed-loop poles without the need for a transformation to a canonical form. For the
single-input case, this was achieved in [44] . A particularly simple solution that directly
allows sequential pole shifting is Ackermann's formula [11 .
Consider a single-input plant
x=Ax+bu (2.3.1)
and assume controllability, i.e.
det [b Ab A 2b ... An-lb] f O. (2.3.2)
Pole assignment by state feedback
(2.3.3)
is the solution of
det(sI - A + be) = p{s) (2.3.4)
T
for the elements of k . By matching coefficients of the powers in s with
p{s) = ao + als + ... + an_ls n- l + s" , (2.3.5)
T
there are n equations in n unknowns of k . These equations have a unique solution, if
and only if the pair (A, b) is controllable.
38 2 Boundary Crossing and Parameter Space Approach

Theorem 2.7 (Ackermann)

The solution of det(sI - A + be) = p(s) with (A , b) controllable is

eTp(A),
(2.3.6)
[0 . .. 0 I] [b Ab A 2b . . . An-lb]-l .

The n-vector eT is the last row of the inverted controllability matrix and p(A) is
the matrix polynomial that is obtained by replacing the powers of s by the powers of
A, i.e. SO --+ A O = I, Sl --+ AI, etc.

Proof

Let F = A - bk T , expand F k into expressions of the form A k and Aibk T Fj ,

i + j = k - 1, and evaluate p(F) = aopO + alF l + ...+ an_lF n- l + F",

x ao

x1

p(F) = p(A) - [b Ab . . . An-lb] [~T ] .

(2.3.7)
The polynomial p(s) = det(sI - F) is the characteristic polynomial that shall
be given to F . Therefore, by the Cayley-Hamilton theorem [94], p(F) = O. Now
[~T ] = [b Ab . . . An-lb]-l p(A) . From the last row of this equation follows

e = eTp(A), (2.3.8)

where eT = [0 . . . 0 I] [b Ab . . . An-lb]-l is the last row of the inverted

controllability matrix.

o
2.3 Pole Placement 39

Note that Equation (2.3.6) may be written as

kT = eT(aoI + alA + . .. + an_1An-1 + An) = [a 1] E. (2.3.9)

]
The matrix

e~:
E=
[
eT~n
(2.3.10)

is called the pole placement matrix. The form (2.3.9) illustrates that it is not necessary
to evaluate p(A) by calculations with n 2-matrices. The calculation of E only requires
operations on n-vectors. For the numerical calculation of E , see [4], [13].
Note also that p(s) may be given in factorized form as

IT (S2 + b.s + Ci) for n even,

p(s) =
{ (s + d) g
i=l (n-1)/2
(S2 + bis + Ci) for n odd,
(2.3.11)

rather than in the multiplied form (2.3.6). The closed-loop system is then stable if and
only if all bi , Ci and d are positive. This is a characterization of all stabilizing state
feedback gains k in equation (2.3.6).
While the factorization of p(s) can be performed only numerically, the opposite step
of finding the polynomial from its roots by (2.3.11) can be performed symbolically and
is therefore particularly suited to parametric systems.

Remark 2.8
Equation (2.3.6) is related with (1.5.12), i.e.
aT = a~ + kTW (2.3.12)
for output feedback. In the special case of full state feedback, W is non-singular for a
controllable system, and .
e = (aT - a~) W- 1. (2.3.13)
Small gains can be expected for small changes in the coefficients of the characteristic
polynomial from 0. 0 to a. However, roots may be very sensitive to coefficient changes.
Therefore, we will look at shifting only one pair of eigenvalues in the next section rather
than changing coefficient vectors.
The connection between Equations (2.3.6) and (2.3.12) is established by substitution
of aT into (2.3.9).
e=[a~ +eW 1]E,
where by Cayley-Hamilton [a~ 1] E = 0, i.e,
1

w= e~: 1- (2.3.14)

reT ~n-1
40 2 Boundary Crossing and Parameter Space Approach

for a controllable system . For non-controllable systems, only the singular W exists ,
but not its inverse.
o

2.4 Sequential Pole Shifting

Consider a stepwise design procedure in which in each step n - 2 eigenvalues are fixed,
and only two eigenvalues are shifted by state feedback k T . In the first step, the char-
acteristic polynomial of A is factorized as
(2.4 .1)

The roots of do + diS + d2s2 are selected for shifting in the first design step, the
polynomial h(s) = h o + hiS + ... + h n_3s n- 3 + sn-2 shall be invariant under state
feedback, i.e.
det(sI - A + be) = h(s)(to + tiS + S2). (2.4.2)
The shift is denoted by

It may be visualized in the (I\;a, I\;b)-plane. How is a point in this plane related with the
full state feedback vector k T ? The answer follows immediately from (2.3.6) :
e = e Tp(A) = eTh(A)(toI + tlA + A 2).

Let er := eTh(A) = [h 1)
eT
eTA ]
(2.4.3)
[ eT In-2 '
e = entoI + t1A + A 2). (2.4.4)
No shift (to = do, t l = d 1 ) yields

0= er(doI + dlA + A 2). (2.4 .5)

The difference of Equations (2.4.4) and (2.4 .5) is

e = erl(to - do)I + (tl - d1)A)

eT
eTA ]
hn - 2 0 ]
hn - 3 hn - 2
[ eT In-l .
(2.4.6)
The (K: a , K:b)-plane is a linear subspace of the k-space. It is called the invariance plane.
It is spanned by the vectors eI and eIA and its coordinates are K:a = to-do , K:b = t 1 -d 1 ·
2.4 Sequential Pole Shifting 41

Remark 2.9
An alternative interpretation of the h-invariant subspace is th at the n - 2 roots of
h(s) are non-observable from the output k T x .
o
For the new system (F, b) = (A - be , b), again only two eigenvalues are shifted .
A new invariant polynomial
-
h(s) = -ho + -hIS + ...+ -hn_3sn- 3 + sn- 2
is chosen for n - 2 eigenvalues of F and

e~ 1
eTF
(2.4.7)
[
e~~-' .
No new calculation of the pole placement matrix is necessary, because its first n rows
are invariant under state feedback. The vectors eT and e~ are, by definition, the last
row of the respective inverse controllability matrices, i.e,
eT [bAb An-Ib] = [0 . .. 01], (2.4.8)
e~ [b (A - be)b (A - bkT)n-Ib] = [0 ... 01] . (2.4.9)
Substituting the first column of (2.4.9), i.e. e~b = 0 into the second one yields
e~(A - be)b = e~Ab = O.
The third column is

and so on until
e~ [bAb . . . An-Ib] = [0 ... 01] (2.4.10)
is obtained, and by comparison with (2.4.8),
e~ = eT for all e. (2.4.11)
The pole placement matrices E and E F of the two pairs (A, b) and (F , b) = (A-be , b)
are related by

(2.4.12)

Therefore, (2.4.7) may be written as

-: ~ e::~-' 1·
[AT 1] [
(2.4.13)

The above results for sequential pole shifting are summarized in the following:
42 2 Boundary Crossing and Parameter Space Approach

Recipe

1. Calculate

eT
eTA ]

[ eT In-I '
n
2. Factorize det(sI - A) = IT (s -
sil o
i=1
3.1. Select two eigenvalues, say SI and S2 , for shifting in the first design step .
•T n
4.1. Determine the coefficient vector h of the invariant polynomial h(s) = IT (s - s. )

er
i=3
·T-
and calculate = h E.
5.1. Choose a state feedback

(2.4.14)

in the (K a, Kb)-invarianc~plane that shifts S1, S2 to si, s;.

3.2. If necessary, repeat step 3.1; select, for example , S3 and S4'
2 n
4.2. Repeat step 4.1 with h(s) = fI (s - sj) fI (s - sil o
i=1 i=5
5.2. Repeat step 5.1 and shift S 3 , S4 to s3' S4 etc.

For shifting a single real eigenvalue, h( s) is of degree n - 1 and E is augmented by a

last row e T An .

Example 2.10

Consider the crane with parameters e= 12, mL = 1500, me = 1000. The state
space representation is given by (1.1.6)

0 1 0000]
~
05
X=
o 0
1o
o 0 1 ] x - [1/10 u,
[
o 0 -25/12 0 -1/-12000

and the open-loop characteristic polynomial is

The open-loop system has a double pole at the origin, d(s) = S2 , and a complex
conjugate pole pair on the jw-axis, h(s) = 25/12 + S2 . Choose an invariance plane in
2.4 Sequential Pole Shifting 43

which only the double pole at the origin is shifted to the roots of to + tIS + S2 . By
(2.4.14), the feedback vector is

T [ 2500 0 18000 0 ]
kl=[totd 0 2500 0 18000 '

The pole pair is shifted by to > 0, t l > 0, let, for example, to = 0.1, t l = 0.6, then

ki = [250 1500 1800 10800] .

In a second step, the roots of h(s) = 0.1 + 0.6s + S2 are kept invariant, and the
roots of d(s) = 25/12 + S2 are shifted to the roots of a new t(s) = to + tIS + S2,
i.e. Kc = to - do = to - 25/12 , Kd = t l - d l = tl . The resulting feedback vector is

T [120 720 -10 560 8 640 ]

k 2 = [Kc Kd] 0 120 -7200 -19560 .

The total state feedback gain is k = kl + k 2 = [kl k 2 k 3 k 4 ]T , with

kl 250 + 120Kc ,
k2 1500 + 720Kc + 120Kd,
k3 1800 7 10 560Kc - 7200Kd,
k4 10 800 + 8640Kc - 10 560Kd.
- - 25
The system is stable for to > 0, t l > 0, i.e. Kc > --, Kd > o.
A practically useful choice of Kc , Kd from the slible region is one for which we do
not have to feed back the rope angle rate, i.e,

k« = 10800 + 8640Kc - 10 560Kd = O.

The line k 4 = 0 is shown in the (Kc , Kd)-plane of Figure 2.4.

Remark 2.11

Figure 2.4 also contains the image of a more restrictive boundary than Hurwitz-
stability. Such "r-stability" will be treated in Chapter 3. The chosen boundary is the
left branch of the hyperbola w2 = 40- 2 - 0.25. The point K(2) is chosen such that not
only Hurwitz-stability is given, but also r-stability in the sense that all poles are on
the left side of the above hyperbola.
o
We choose Kc = 0.83, Kd = 1.7 on the line k4 = 0, then the second pair of roots moves
from the imaginary axis to the root of r(s) = Kc + 25/12 + KdS + S2 = 2.91 + 1.7s + S2,
and the total feedback vector is e
= [349 2300 - 19181 0].
o
44 2 Boundary Crossing and Parameter Space Approach

K.d 4 I
I
I
I
I
I
3 I
I
I
I
I
2 I
I
I
I
I

1 I
I
"
I
"
I
"
I
""
I
"
-2 -1 o 1 2 3

Figure 2.4. Invariance plane for the second design step. Feedback of the angular velocity X4
is avoided on the line k4 = 0

At this point, the reader may wonder why both pole shifts are not executed in one
shot. However, in the simultaneous stabilization of several operating points, it turns out
that sequential pole shifting is a very useful tool because it allows sequential graphical
design steps in 2D-subspaces. In fact, the idea of sequential pole shifting originated
from the application example of a track-guided bus [32], see Section 6.9.

2.5 Singular Frequencies

In the previous examples, a continuum of frequencies contributed to the CRB (see, for
example, Figure 2.2) There are , however, special cases, where a singular frequency gen-
erates an entire branch of the stability boundary. Consider a polynomial that depends
linearly on two uncertain real parameters ql and q2 (these may be plant or controller
parameters) . It may be written as

(2.5.1)

where PO(S),PI(S),P2(S) are known polynomials. For s = jw, their real and imaginary
parts are
Po(jw) Ro(w) + j/o(w) ,
PlOW) RI(w) + j/I(W), (2.5.2)
P2(jW) R2(w) + j/2(W),
and the real and imaginary parts of

(2.5.3)
2.5 Singular Frequencies 45

are zero for

Rp(W , qI,q2) Ho(w) + q1R1(w) + q2R2(W) = 0,

Ip(w, ql, q2) Io(w) + q1I1(w) + q2I2(w) = O.

In matrix notation,

(2.5.4)

A singularity occurs at a frequency w for which the determinant of the matrix in (2.5.4)
vanishes, i.e,
(2.5.5)

In this case, the two straight lines represented by (2.5.4) are parallels in the (qI, q2)-
plane, i.e. a solution exists only if the two lines become identical, i.e, if also

(2.5.6)

Real frequencies w = Wk that satisfy both Equations (2.5.5) and (2.5.6) are called
singular frequencies. At such a singular frequency, the solution in the (qI, q2)-plane is
not just an intersection point (like for non-singular frequencies) but an entire straight
line. We will exploit this fact for the design of robust PID-controllers in Section 2.6.
The real parts Ho, R 1, R 2 are polynomials in w2, hence R(w) = R(w 2). The imaginary
parts Io,!I, 12 have the form I(w) = wJ(w2). The case w = 0 yields the RRB. For the
CRB with w f. 0, divide (2.5.5) and (2.5.6) by w to get

R 1(w 2)12(w2) - R 2(w2)1 1(w2) = 0, (2.5.7)

R 1(w 2)lo(w2) - Ro(w2)1 1(w 2) = O. (2.5.8)

The real positive solutions wl of d(w 2 ) = 0 yield real frequencies Wi that are candidates
for singular frequencies. Substitute Wi in d) (w2 ) = 0, i.e.

(2.5.9)

If this condition is satisfied, then Wi is a singular frequency.

Remark 2.12

The equations d) = R 1lo - Rol) k 0 and d2 = R 2Io - R oI2 = 0 are linearly

dependent by construction and it does not matter which one is used. The only exception
is the trivial case d 1(W2) == 0, then d2(W2) = 0 must be used.
o
Figure 2.5. Three stability regions in the (ql , q2)-plane . a) q3 = 0.56, connected, b) q3 = 0.57,
singular case, c) q3 = 0.58, disjoint
2.6 Stability Regions for PID-controllers 47

Example 2.13

Consider the uncertain polynomial

Plot the stable region Qstable in the (ql, q2)-plane for fixed q3.
The RRB for w = a is

ao(q) = -0.3ql + 2q2 + (14 + 2q3) = O.

Equation (2.5.7) reads d(w2) = -1.6 + w2 = o. The candidate frequency w = Jf.6 is
substituted in (2.5.8):

d l (1.6) = -1.368 + 2.4q3 = O.

This equation yields q3 = 0.57 as the only possibility that w = Jf.6 is a singular
frequency.
The complex root boundary for all frequencies follows from (2.5.4), with the second
row divided by w :
2
ql ] = _ [ -0.3 2] -I [ (14 + 2q3) - IOw + w4 ]
[ q2 2 - 2w2 8 10 - 2w2 ,

19w + 23 + 4q3
2
2 ) _ -5 2w4 -
ql ( W , q3 - 5w2 _ 8 '
2 ) __ ~ IOw 6 - llOw 4 + (243 + 20q3)W 2 - 155 - 20q3
q2 ( W , q3 - 4 5w2 _ 8 .
Figure 2.5 shows the three cases with a) q3 = 0.56, b) q3 = 0.57, and c) q3 = 0.58.
The asymptote for w = Jf.6 is approached in opposite directions in the cases a) and
c). In case b) , it is a singular boundary. Obviously, the topology in the (qll q2)-plane
changes at q3 = 0.57 from a simply connected stability region to two disjoint stable
regions. 0

2.6 Hurwitz-stability Regions for PID-controllers

In the previous section, the occurrence of singular frequencies appeared as a very special
case. In this section, it will be shown, that singular frequencies occur in PID-control
systems and can be exploited for their design . Consider the PID-control system in
Figure 2.6. Gj(s), i = 1,2, ... , N is a finite family of plant models for N representative
operating conditions. The problem is to design a robust PID-controller

PID(s) = K 1 + Kps + KDS2 (2.6.1)

sO + TRS)
48 2 Boundary Crossing and Param eter Space Approach

r y
Gi(s)

Figure 2.6. PID-control system

that simultaneously stabilizes all N members of the plant family. A small time constant
TR is introduced for realizability of the PID-controller. TR = 0 may be assumed for the
design of the essential parameters K/, K p and K D. In order to allow for trade-offs with
other design specifications, e.g. in the frequency domain , it is of interest to calculate
the set of all simultaneously stabilizing PID-controllers. This set can be visualized in a
tomographic rendering, e.g. by representing the stable region in a (K p , KD)-plane for a
grid of Krvalues. In this plane, the stable region is non-convex and bounded by curves,
which may be calculated, for example, by the software tool PARADISE , see Appendix
B. In each plane , the intersection of the admissible sets for the N operating conditions
is determined in order to find the set of simultaneous stabilizers.
The above process is considerably simplified by a recent result of Bhattacharyya, Ho
and Datta [108], [711. They have shown that for fixed K» the Hurwitz-stable region in
the (K D , K/ )-plane consists of convex polygons . Thus, also the intersection for several
operating conditions consists of convex polygons , and a tomographic rendering for a grid
on K» yields a visualization of the set of all simultaneously stabilizing PID-controllers.
The above authors assume T R = 0 and use a modified Hermite-Biehler theorem for
the proof. An alt ernative proof was given by Munro and Soylemez [155] by calculating
the real axis intersections of the Nyquist plot . In this section , the param eter space
approach is applied to the problem .
The theory will be first developed for one operating condition

A(s) numG(s)=aO+als+ . .. +aksk, akfO

B(s) s(l + TRs) den G(s) = bo + bls + . .. + bmsm, bm f o.
The characteristic polynomial of the PID-control system in Figure 2.6 has th e structure

(2.6.2)

(The following derivation also holds for bo f 0.) The polynomial (2.6.2) has a root at
s = jw if and only if:

Re P(jw, K/ , K p , K D) = 0, and
(2.6.3)
Im P(jw, K/ , K p , K D) = O.

For the separation of real and imaginary parts, write

B(jw) = R B + JIB .
2.6 St ability Regions for PID-controllers 49

Then for fixed K p , Equation (2.6.3) may be written as

[ n; ]= [R A -RAW:] [K[ ] + [ -KpIAw + R s ] = [ 0 ] . (2.6.4)

t; I A -lAw KD KpRAw + Is 0

Note that the matrix multiplying [K[ KD]T is always singular, i.e. geometrically (2.6.4)
represents the intersection of two parallel lines. A solution exists, if and only if the
parallels are identical , i.e. for real frequencies Wk such that

g(W) det [R A -KpIAw + R s ]

IA KpRAw+Is
RA(KpRAw + Is) + IA{KpIAw - Rs)
wKp(R~ + I~) + RAIs - lARs = O. (2.6.5)

It is easily verified that g(w} is w times a polynomial 9 in w2 , i.e. g(w} = wg(w2 } and
Wo = 0 is always a solution of g(w} = O. The degree of the remaining polynomial g{w 2 }
results from the highest degree of the product terms in (2.6.5). Only its positive real
solutions w~ lead to real frequencies Wk . The solutions Wk (k = 0, 1,2 . . . K) are singular
frequencies.
The interpretation of (2.6.5) is that the root set of P(S ,qO,q2} (e.g. qo = K[ ,q2 =
K D ) can cross th e imaginary axis only through the "holes" at the singular frequencies
Wk ·

Example 2. 14

qo E [-2; 0.1],
q2 E [0.1;0.7].
Figure 2.7 shows the two-parametric root set generated by gridding qo and q2 . It can
cross the imaginary axis only through the hole WI = v'2. 0

Remark 2.15

This example also serves as a warning that gridding of w will miss the value WI = v'2.
o
Starting from a stable polynomial (2.6.2), there are three possibilities of how a root
can cross over the imaginary axis by variation of K[ and K D •
a} A real root crosses at s = O. For w = 0, we have I p = 0 and th e real root
boundary (RRB) is obtained from

Rp(w = O} = aoK[ + bo = 0
50 2 Boundary Crossing and Parameter Sp ace Approach

2r-------,--------,

O+------+-------l

-1

_2L---.,.----!-----:------i
-1 o 2

Figure 2.7. The root set can cross the imaginary axis only at the hole WI = /2

The RRB is

(2.6.6)

b) A real root crosses at s = 00 . This infinite root boundary (IRB) is characterized

by Pn = 0 in P(s) = Po + PIS + ... + Pnsn.
for m < k + 2
for m = k+2
for m > k+2
In the (K D ; KJ)-plane , the IRB is
0 for m < k + 2
KD = -bm/ak for m = k+ 2 (2.6.7)
{ no boundary for m > k+2

c) A conjugate pair of roots crosses at one of the non-zero singular frequencies Wk .

These complex root boundaries (CRB) are obtained by substitution of Wk into
the first row of (2.6.4) (assuming R A =I- 0) or second row (assuming fA =I- 0). The
first row becomes

(2.6.8)

In the (K D , K J )-plane, these are straight lines with positive slopes w~, one line
for each singular frequency.
In summary: in the (K D , K J )-plane for fixed K» , roots cross over the imaginary
axis :

i. On the line K J = -bo/ao with slope zero (RRB , Wo = 0).

ii. On the lines of Equation (2.6.8) with positive slope (CRB, W = Wk, k = 1,2 . . . K) .
2.6 Stability Regions for PID- controllers 51

iii. Possibly on th e boundary for K D of (2.6.7) with infinite slope (IRB , W = 00).

These straight lines form convex polygons (also infinity may be a vertex), where
neighboring polygons correspond to a different number of poles in t he left half plane.
There remains the t ask of identifying the stable polygons. One possibility is to check
one point in each polygon in order to find the stable ones.

Example 2.16

For K p = 1, the exampl e is identical to Example 2.14.

i. RRB for Wo = 0:

ii. CRB for WI = V2:

K[ = 2K o - 1, slopewi = 2.

iii. IRB for Woo = 00:

K D =0.

The three boundaries are shown in Figur e 2.8 together with a stable test point in th e
tri angle . 0

K[
K[ = 1 (RRB)
1

K o = 0 (IRB) K[ = 2Ko -1 (CRB)

1 Ko

Figure 2.8. A test point K[ = 0, K o = 0.25 is stable, i.e, the triangle is stable
52 2 Boundary Crossing and Parameter Space Approach

Another possibility to find th e stable polygons is to find the intersections of all lines,
bisect each boundary segment and divide the polynomial by the known elementary
factor of the line. If the remainder polynomial is stable, then the segment is an active
part of the boundary, i.e, it divides a stable polygon from one with one or more roots
in the right half plane . This second approach is more suited to an automated computer
test , in particular if several singular frequencies occur, like in the following exampl e.

Example 2.17

A(s) _S4 - 7s 3 - 2s + 1,
B(8) 8(8 + 1)(8 + 2)(8 + 3)(8 + 4)(8 2 + 8 + 1).

Find the stable region in the (K D , K [)-plane for K p = -4.4.

RRB : K[ = -bo/ao = O.
IRB does not exist , because Pn = b6 = 1.
CRB :

d(w2) wl O + (31 + K p)w8 + (-579 + 49Kp)w6

+ (730 - 30K p)w4 + (-257 + 4Kp)w 2 + 24 + K» = O.

One root with w2 < 0 does not lead to a singular frequency. The other Wk and corre-
sponding stability boundaries for K p = -4.4 are

WI = 0.316 , K[ = O.OlKD + 7.09,

W2 = 0.695 , K[ = 0.48KD + 15.5,
W3 = 0.730 K[ = 0.53KD + 14.1,
W4 = 4.13 ,K[ = 17.1KD - 67.0.
Figure 2.9 shows the five boundary lines. The reduction to the active segments by
bisection and division .is shown in Figur e 2.10.
The idea of a tomographi c represent ation of all stabilizing PID-controllers, by grid-
ding K p , is illustrated in Figur e 2.11. Th e figure shows that the two stable regions
of Figure 2.10 are simply connected in 3D space. The cross-section of Figure 2.10 is
enhanced. 0

In order to decide on the K p intervals for which stabilizing polygons exist in the
(K D , K[ )-plane, it is useful and simple to replace the factorization of (2.6.5), i.e.

(2.6.9)

for wK(Kp) by the evaluation of the inverse function Kp(WK) from

(2.6.10)
2.6 St ability Regions for PID-controllers 53

Figure 2.9. Five boundary lines in (KD, K] )-plane for Kp = -4.4

Example 2.18

For Example 2.18 Kp(w) is plotted in Figure 2.12. For fixed K» = -4.4, the WK
can be read off from the graph, e.g W I = 0.3156, W2 = 0.695, W3 = 0.730, W 4 = 4.13.
For W ---. 0, the CRB with K p = -24 approaches the real root boundary. At th e
minimum "-'2 = W 3 =0.712, K p = -4.51 , two singular frequencies merge into one and
two boundaries disappear. Figure 2.13 shows th e slice for K p = -4.51, which is closely
neighboring the slice in Figure 2.10 for K p = -4.4.

8 - r - - - -- - --------------,

-2 -60 -20 o

Figure 2.10. The active parts of the stability boundary form two stable convex polygons
54 2 Boundary Crossing and Paramet er Space Approach

Figure 2.11. The set of all stabilizing PID-controllers

From the maxima and minima of the graph, the following Kp-intervals are obtained:
Kp < -24 no real Wk
-24 < Kp < -4.51 (minimum) triangle
-4.51 < Kp < 3.99 (first maximum) two polygons
3.99 < Kp < 6.15 (second maximum) triangle
6.15 < Kp no real Wk
0

The design of simultaneous PID-stabilizers for an infinite plant family G(s, q) , q E Q

proceeds in three steps:
1. Choose representatives q(i) , i = 1,2 . . . N , typ ically the vertices of the Q-box. The
result is a finite plant fam ily.
Gj(s) = G(s, q(j»), i = 1,2 ... N. (2.6.11)

10..-----------------,

o
_S.!· -,i " .
Kp
-10

-IS

-20

_2S.l.-------~------!
2 W 3 4 S

Figure 2.12. K» as function of w

2.6 St ability Regions for PID-controllers 55

8,--------------------,

-2 -20 o

Figure 2.13. Slice for K» = -4.51.

For each i , the set of all stabilizing controllers is determined.

2. For the system in Figure 2.6 find a PID-controller that stabilizes the loop for all
representatives. For each gridded value of K» , the set of all such controllers is
the intersection of the convex polygons for the individual representatives, i.e. a
convex polygon. Choose a centrally located point for the controller.
3. Do a robustness analysis for the continuum of the parameters q E Q. This
pro cedure is easy for two uncertain parameters, as will be shown in the following
exampl e. Additional parameters must be gridded , see Figur e 2.11. If there is an
unstable subregion in Q, choose additional representatives that were unstable in
the first design iteration and go back to step 1.

~xa~ple 2.19 ((71J)

Consider the continuous plant family
K
G(s, K , T) = (1 + TS)8' K E [0.5; 2], T E [0.5; 1.5].

Choose the four vertices of the Q-box as representatives. Figure 2.14 shows the four
stabilizing triangles in a (KD , K[ )-plane for K p = 0.35. In their intersection, the
centrally located point K D = 0.50, K[ = 0.05 is chosen. For the robustness analysis,
the controller is now fixed and the stability region in the plane of plant parameters
T and K is determined. The mathematical details of mapping th e imaginary axis (or
other boundaries in the s-plane) into a parameter plane will be treated in Chapter 4,
in Example 4.19. Returning to the present example , the result is Figure 2.15. It shows
that not only the four vertices of the Q-box are stable, but the entire continuum of
parameter values in Q is stable. D
56 2 Boundary Crossing and Parameter Space Approach

2.5

1.5
Kj

0.5

-10 10 20

Figure 2.14. Simultaneous stabilizers in the plane K p = 0.35 in the intersection of four
triangles for the vertices of the Q-box

2.7 Hurwitz-stability Regions for a Compensator or

Plant Subpolynomial

A generalization of th e characteristic polynomi al (2.6.2) is

p(s , q) = A(s)Q(s) + B(s) , (2.7.1)

where Q(8) = qo + qls + .. . + qtst is a numerator or denomin at or polynomial of the

controller (for design) or of th e plant (for robustness analysis) . In thi s sect ion, th e PID
results of Section 2.6 will be generalized to a fixed even or odd part of Q(jw).
For s = jw , all polynomials are split into their real and imaginary parts, i.e. (omitting
the arguments )

(2.7.2)

The polynomial p(s , q) has a root at s = jw on the imaginary axis if and only if R p = 0
and I p = O.

(2.7.3)

With the detailed not at ion

RQ qo - q2w2 + q4 w4 =F ,
IQ qlW - q3w3 + qswS =F ,
2.7 St ability Regions for a Com pensator or Plant Subp olyn omial 57

6,---=----------------,

5 unstable

K3

stable

o 2 3 4 5
T

Figure 2.15. The simultaneously stabilizing PID controller not only stabilizes the representa-
tives (vertices of Q) but the continuum Q of operating conditions

Equation (2.7.3) becomes

(2.7.4)

Obviously, both matrices in Equation (2.7.4) have rank 1. (Th e trivi al case A(s) == 0 is
excluded here.) For fixed 1Q , Equation (2.7.4) represents two parall el hyperplanes in the
par ameter space with coordinates qo , qz , q4, . . . The two hyperplanes become identical
at real frequencies w , for which

RA RB - 1A1Q ]
det
[ 1A 1B + RA1Q

RA(IB + RA1Q) - 1A(RB - 1A1Q)

(R~ + 1~)1Q + RA1B - 1ARB = O. (2.7.5)

It is easily verified th at gR(W) is W t imes a polynomial in w2. Its degree results from
the highest degree in w 2 occuring among the products RA1B, RVQ , 1ARB and 11R Q.
w~ = 0 is always a solut ion. Further solutions are the positive real roots in w 2 that
result in a real frequency w . These real solutions Wk , (k = 0,1 ,2, ... , K) are the singular
frequencies.
58 2 Boundary Crossing and Parameter Space Approach

Remark 2.20

A dual result is obtained by fixing RQ . The singular frequencies are then the real
roots of

-IA(IB + IARQ) - RA(R B + RARQ)

-(I~ + R~)RQ - lAIR - RAR B = 0 , (2.7.6)

which is a polynomial in w 2 •
o

The stability boundaries are

a) RRB

(2.7.7)

b) CRB
For the singular frequencies Wk , the first row of (2.7.4) becomes

(2.7.8)

For fixed IQ, Equation (2.7.8) represents K hyperplanes in the parameter space
of RQ-coefficients.
c) IRB
In the RQ-space, the IRB is

0 for m < k + €
qf= -bm/ak form=k+€ (2.7.9)
{
no boundary for m > k + €

Based on the two dual results , it should be possible to develop a design procedure
for a compensator polynomial Q(8) by alternat ing design steps on th e even and odd
part of Q(8).
3 Eigenvalue Specifications

In Chapter 2, we were primarily concerned with Hurwitz-stabilization and Hurwitz-

stability analysis of feedback systems with uncertain parameters. In this third chapter,
the parameter space approach is extended to Gamma-stability, i.e. the left half s-plane
for Hurwitz-stability is replaced by other regions r that guarantee, for example, some
damping or negative real part for all eigenvalues. A linear system is called r -siable if all
roots of its characteristic polynomial are located in a specified region r in the s-plane.
This is an indirect approach to meet time-domain specifications. A direct approach
would require simulations of responses of several state variables and actuator signals
of the closed-loop system to typical reference inputs and disturbances. Such numerical
solutions with a more detailed model including plant and actuator non-linearities are
frequently necessary in a final analysis of the control system before its hardware and
software implementation. For systems with uncertain parameters, then gridding of
the operating domain (see Figure 1.2) is unavoidable. Such a detailed analysis would
not be practical in design iterations. Therefore, we work in the design process with
simple linearized plant models in order to understand and accommodate the influence
of uncertain parameters without brute-force gridding. In this context it is important
to know some elementary relations between eigenvalue locations and time responses ,
where the step responses are of primary interest.

3.1 Poles, Zeros and Step Responses

The performance of control systems is frequently evaluated by typical responses to

reference and disturbance inputs. Examples are:

• Transport of a load by a crane , where the initial and final values of the rope angle
and its rate and also of the crab velocity are zero.
• Transition of a vehicle from a straight track into a curve.
• Rapid increase of the crosswind acting on a car.
• Transition of an aircraft to a higher altitude.

Undesirable properties of a response and appropriate countermeasures are :

60 3 Eigenvalue Specifications

i. Th ere are oscillat ions t hat do not decay fast enough and cause excessive over-
shoot . Th e frequency w· = 21l" IT is determined from th e period T of the undesired
oscillation. In a distance w· from the origin of the s-plane, there is a complex
pair of eigenvalues with insufficient da mping. Improvement of this damp ing has
priority in the next design step. Note, however, that excessive overshoot may also
be caused by poles placed furth er away from th e origin t han a zero of the plant
transfer funct ion, see Figure 3.4.
ii. The response is "sluggish", it creeps towards the stat ionary value. A negative
real eigenvalue is too close to the origin and must be moved to t he left in the next
design step.
iii. Th ere is undesired high-frequency content in th e act uator signa l. Countermea-
sures are reduction of th e feedback bandwidth, increase of th e relative degree of
the controller, and shift ing far left eigenvalues closer to th e origin on a circle with
rad ius Wb such th at the absolute value of the frequency response drops rapidly for
frequencies W > Wb . In sampl ed-dat a systems, frequently th e anti-aliasing filter is
a cure.
For th e single-input single-output plant , some further asp ects of pole shifting must be
observed:
iv. If th e closed-loop eigenvalues are shifted too close to open-loop zeros, then large
loop gains may result as can be seen from th e root locus. High-loop gains are
und esirable for our exampl es of mechanical syst ems in view of th e limited actu at or
forces and amplification of sensor noise. If, however, open-loop poles are located
nearby zeros such th at an almost cancellat ion occurs, t hen they have only minor
effect on responses.
v. If th e open loop has a relative degree of two or more (t his is a common case), then
the center of all eigenvalues of plant and compensator cannot be shifted . In this
case, the coefficient am-I in a polynomial n:1 (s - s.) = ao +... + am_ISm- 1+ sm
is determined by t he plant and compensat or poles only and not by any zeros. Th e
coefficient represents th e center of the s, by am-I = - L::I Si ' If some eigenvalues
must be shifted to th e left , then necessarily other eigenvalues migrate to the right.
In thi s case, th e compensator poles must be chosen sufficiently far left, such th at
the plant poles can be shifted to th e left.
vi. Pole-zero cancellations out side th e desired closed-loop eigenvalue region r should
be avoided. From the frequency domain point of view, th ey may still look ac-
ceptable for a specific transfer function, but there may be a disturbance or init ial
condition that excites th e subsyste m with the undesired response. The control
syst em should be "internally r -stable" , i.e. th e tra nsfer functions from each pos-
sible input to each signal inside th e feedback loop should have all th eir poles in
th e region r.
In t he following paragraphs, we recap itul at e some simple relationships between ti me re-
sponses and eigenvalue locati ons for second and third order syste ms. Such relation ships
are useful if a "dominant behavior" similar to such simple syste ms is desired.
3.1 Poles, Zeros and St ep Responses 61

If all eigenvalues are locat ed to the left of a parallel to the imaginary axis of th e
s-plane at (7 = -a, then all solution terms decay at least like e- at . Figure 3.1 shows
two examples: Yl and Y2 have the same negative real part (7 = -a of the eigenvalues.

" 1 x jw
"""'"' 2 x
,,
q"'
(7

-a
x
x

Figure 3.1. Two solution terms with the same negative real part (7 = -a of the eigenvalues
Terms of the type Yl are, however, undesirable because more overshoot and larger
oscillations occur inside the envelope ±e- at than for Y2 with a lower frequency W2 '
The high-frequency eigenvalues should be located further to the left in the s-plane.
Therefore, a minimum value of the damping D is required. A complex conjugate pair
of eigenvalues a, ± jWi may be written as a second order factor of the dosed-loop
characteristic polynomial

The distance of the eigenvalues from the origin is the natural frequency Wo = (7[ + W[, J
and D = -(7i/wo is the damping. The inverse relation for the real part a, and the imag-
inary part Wi of the eigenvalue is (7i = -Dwo, Wi = woVI - D2. Figure 3.2 illustrates
these relations in the s-plane. A damping value D corresponds to an angle 0:" with
respect to the imaginary axis, where

D=sinO:". (3.1.1)

The corresponding solution term in the time-domain is

Yi(t) = A e- Dwot cos ( Vi - D2 wot + 'P) for IDI < 1. (3.1.2)

'P
Amplitude A and phase angle depend on input and initial conditions. Natural fre-
quency Wo and time t appear only as product wot, i.e. Wo may be considered as a scaling
factor for the time .
62 3 Eigenvalu e Specifications

-----'----+----(1

Figure 3.2. Natural frequency Wo and damping D of a complex conjugate pair of poles

Example 3.1

As an example, consider the step responses of the system

1
g(s) = ------,------,:-...,......-:, (3.1.3)
1 + 2Ds/wo + s2/ w5

for different damping values. For u(s) = l/s, and Y(8) = g(s)u(s) , we have
1 1 (s+Dwo) + Dwo
y (s) - -;-::;-~----::-:=--=---;­
- (S2/w5 + 2Ds/wo + l )s s (s + DwoF + (1 - D2)w'lJ"
The inverse Laplace transform is

y(t) = 1- e- Dwot [cos (Vl- D2wot) +h Sin (Vl- D2wot)] . (3.1.4)

The responses for D = 0.5, D = 1/ V2, and D = 0.9 are shown in Figure 3.3 for the
scaled time wot. The step response for D = 1/V2 : : : 0.7 is considered as particularly
favorable. Its maximum overshoot of 4.3 % occurs at wot = 4.4. The value D = 1/V2
is characterized by the fact that th e magnitude of the frequency response Ig(jw)1 has a
maximum only for smaller damping ; the resonance frequency is Wo = VI - 2D2. For
larger damping no resonance occurs. In Figure 3.2 we have Q: = 45° for the damping
value D = 1/V2. 0

Example 3.2

Next , we assume a zero at s = -bwo.

g(s) = (1 + s/( bwo)) . (3.1.5)

1 + 2Ds/wo + S2 /w~
3.1 Poles, Zeros a.nd Step Responses 63

Y D = 0.7

o
o 2 4 6 8

Figure 3.3. Step responses of the system (3.1.3)

The step response is

y(t) = 1- e- Dwot
[cos ("'1- D2wot) + ~sin ("'1- D2wot)] . (3.1.6)

Again, wot can be introduced as scaled time. For D = 1/ V2 and some values of b, the

b= 0.5
y
b = -0.5
1
b--s co ...... -
--- -----
/
/<b = -0.5
/ with prefilter
/
/1TV2
o
4 6 8

b=2 1 -0.5

Figure 3.4. Step response of the system (3.1.5) for different zero locations b

pole zero locations and the step responses are shown in Figure 3.4.
All curves are focused at wot = 1TV2 = 4.44, becau se here the sine t erm in (3.1.6)
vanishes and that is the only term which contains the varying b. The curve for b -+ 00
64 3 Eigenvalue Specifications

is identical to the center curve of Figure 3.3. It is the only one with relative degree
two such that the step response begins with slope zero by the initial value theorem
of Laplace transforms. A zero for b = 2 does not change the response significantly.
Even for b = 1 the response is acceptable, but if the zero is closer to the origin than
the poles, then a large overshoot occurs (40.7% for b = 0.5). It could be removed in
the step response by cancellation of the zero at s = -0.5. If we want to avoid this
cancellation, then the distance Wo of the eigenvalues from the origin must be reduced ,
i.e. for a fixed zero location at 8 = -bwo, b is increased and Wo is reduced .
For negative b, the system has non-minimum phase behavior ; the step response
starts in the negative direction. Cancellation is impossible here, because it would give
the system an unstable eigenvalue. The unfavorable undershoot can be reduced by
cancelling the mirror image of the zero by a compensator 1/(1 - 8/(bwo)). In the
example, it is 0.5/(8 + 0.5). The step response with this compensator is shown in dashed
lines in Figure 3.4; it is significantly slower. 0

Example 3.3

Next , the mutual influence of poles will be studied. Extend the system (3.1.5) by a
low-pass filter with a pole at 8 = -awo, i.e.
1
(3.1.7)
9(S) = (1 + 2Ds/wo + s2/ w;5)(1 + s/(awo)) ·
The step response is

y(t) = 1- 2 1D {e- awot + (3.1.8)

a - 2a + 1
~ a(1+aD-2D 2) ~
e-Dwot[a(a - 2D) cos( v l - D2wot ) + ~ sin( v 1 - D2wot )]} .

Now, a smaller damping may be chosen because the additional pole counteracts a
resonance in the frequency response. The chosen value is D = 0.5 such that the third
order Butterworth filter is contained as a special case with a = 1. For a --+ 00 , i.e.
without the additional pole the response is identical to the curve for D = 0.5 from
Figure 3.3. Figure 3.5 shows some typical step responses. The pole for a = 2 has little
influence. For a = 1, the overshoot is reduced from 15.5 % to 8.1 % and for a = 0.5
there is no overshoot . The solution gets slower as the real pole migrates to the right .
Finally, the real pole dominates and the complex poles show their influence only in the
beginning of the step response. The later response is "sluggish" . 0

Summary
The question of when the step response of a robustly stable system approaches its
steady state value depends only on the dominant poles with the smallest distance from
the origin 8 = O. Further left and more remote poles and zeros (more than twice the
3.1 Poles, Zeros and St ep Responses 65

a=2 1

Figure 3.5. Influenceof a pole at s = -a on the step response

distance) have an influence only on the initial part of the step response ; the correspon-
ding time response terms have died out before the entire solution approaches its steady
state value. They may have strong influence on the initial actuator signal magnitude.
Zeros in the left half plane have a similar effect as reduced damping. Their influence
can be reduced by cancellation in a prefilter or by a compens ator in the closed loop.
The unfavorable influence of right half plane zeros can be modified but not removed.
It is recommended to place several closed-loop poles at about the same distance
from the origin . If more poles on this radius contribute to the dominant behavior, then
a small lowest damping is acceptable.

Remark 3.4

The step response is taken as a reference in the preceeding discussion. It resulted

in th e "ideal" damping of D = 1/V2, see Figure 3.3. There are applications where
much smaller damping is acceptable, because typical input signals are milder than
a step. Take, for example, a car at maximum speed . It may have a damping of
only 0.3 or 0.4 in the yaw motion, but the driver will turn the steering wheel only
softly and not stepwise. Similarly, military specifications for aircraft require only 0.35
minimum damping in the short-period longitudinal mode. On the other hand, there
are applications where no overshoot at all is allowed (e.g. an aperiodic solution with
D > 1 or the response for a = 0.5 in Figure 3.5). An example is the increase of thrust
to 100 % for an aircraft engine. Here, the materials reach their temperature limits. and
66 3 Eigenvalue Sp ecifications

overshoot of th e temperature would reduce t he lifetim e of some parts of t he engine.

o

3.2 Root Sets, Gamma-stability

For cont rol systems with uncertain plant parameters, it is a brute-force analysis ap-
proach to grid the operating domain Q and to calculate and plot all closed-loop eigen-
values for the grid points. For stab ility, this root set must be locat ed in the left half
plane.

Example 3.5

Consider the cran e with fixed parameter me = 1000 and uncertain param eters
mL E [1000; 2000], e E [8; 16] . For the nominal case me = 1500, e = 12; the output
feedback cont roller

u = -K [500 2191 - 4299 0] x , K =1 (3.2.1)

shifts the eigenvalues to 51,2 = -0.59 ± 1.06j , 53 ,4 = -0.51 ± 0.16j.

elm]
16
xxxxxxxxx
xxxxxxxxx
xxxxxxxxx
12

8
~lga-Ia~~~

1000 1500 2000

Figure 3.6. An 11 x 9 grid over the operating domain Q

The operating domain Q of Figur e 3.6 is now gridded in order to find the root set
of Figur e 3.7 by numerical factorization. 0

The main difficulty in carry ing out a stability analysis by root set construction
is large comput at ion time . An algorit hm to find the roots of a single polynomial of
reasonably low order is not particularly time-consuming. Generally, repetition of the
algorithm up to a few t housand t imes for a one or two parameter root set is acceptably
brief. Now, consider t he general case of f. parameters. If an N point grid were used for
each parameter, th e root finding algorit hm would need to be repeated Nt ti mes. No
3.2 Root Sets , Gamma-stability 67

1.5 .:.
2 r-----~----,-----,

. ~.
0.5

,
8 0

-0.5

-1

-1.5 ..... . ..

-2
-3 -2 -1 0
o

Figure 3.7. Root set for the grid points of Figure 3.6

matter how fast the computer is, it is easy to select fairly small values of f and N that
would make the root computations last hours, weeks, even years. Motivated by these
limits, we will look at alternative, potentially more efficient stability analysis methods
in the parameter space. These methods are based on the boundary crossing concept
that we have introduced for Hurwitz-stability already in Chapter 2. For this purpose ,
r -stability is defined here.

Definition 3.6

A polynomial p(s) = (s - sd(s - S2) . .. (s - sn) is called r-stable if all s, E r . The

region r has a boundary ar that consists of one or more contours in the s-plane such
that the membership s, E r is well defined. 0

The boundary crossing theorem (Theorem 2.1) can be immediately generalized to

other pole regions. Only minor modifications have to be made .

Theorem 3.7 (boundary crossing theorem for pole regions r)

The family of polynomials P( s , Q) is robustly r-stable, if and only if:

i. There exists a r-stable polynomial p(s) E P(s, Q).
ii. 0"(0) + jw(o) f/. Roots[P(s, Q)] for all 0 E [0- ; 0+].

In Example 3.1, the controller was designed by pole placement for a nominal plant in
the center of the Q-box. Pole placement for single-input systems is primarily attractive
as a synthesis method, because it yields a unique solution for state feedback or for the
compensator. This supposed advantage in the synthesis procedure, however, turns into
68 3 Eigenvalue Sp ecificati ons

a disadvantage for uncertain parameters; a unique solut ion does not allow any flexibility:
if plant parameters change and the same pole locations are prescribed, th en also th e
controller must be changed. The preceding discussion on eigenvalue specification s has
shown that a precise assignment of all eigenvalues is more th an what actually follows
from the time-domain design specifications. For uncertain parameters, it is a reasonable
requirement that all eigenvalues are located in a specified region r in th e s-plane and
must remain in this region under all admissible parameter variations. We call thi s
approach pole region assignment.
Figure 3.8 shows a pole region r . Its bandwidth is bounded by a circular arc and
its damping by a hyperbola that guarantees damping according to th e asymptotes, and
its real part is bounded by the apex of the hyperbola.

bandwidth ~==~=-
, jw

s -plane

----'::'==3----------E'==3-=---'- a
o

Figure 3.8. Specification of a pole region I' th at guarantees damping, negative real part and
bandwidth limitation

Example 3.8

For the control system of Example 3.8, choose the left bran ch of the hyperbola
w2 = 40'2 - 0.25 as the r -stability boundary. It passes through the nominal poles at
81 ,2 = -0.59 ± 1.06j and is shown in Figur e 3.9.
It is now reasonable to modify the controller (3.2.1) such th at the root set becomes
r-stable without overdamping it . A trial- and-error approach would be time-consuming
here. We use this example to show a syst ematic approach by mapp ing the hyperbola:
i. to a (k2 , k3)-plane for simultaneous design of a fixed gain controller, and
ii. to an (mL, f)-plane for analysis ,
iii. to an (f, k4)-plane for gain-scheduling design in Section 3.5.
The mathematical details of mapping a boundary ar will be given in Chapter 4. The
present example should motivate the appro ach.
3.2 Root Sets , Gamm a-stability 69

2 r---~----r---r----,

0.5
:3 0

-0.5

-1

-1.5

-2
-3 -2 -1 0
(J

Figure 3.9. r-stability specified by a hyperbola

Consider the vertex mt. = 1000, £ = 8 of the operating domain Q and the controller

u = -[500 k2 k3 0] X . (3.2.2)

X 10'
1 r--......,..--~--,-~--~----,
of------,.-->.,,---c---.,----;----;----{
-1

-2
-3 .
~'"
-4 .

-5
-6
-7
-8 '--_~ _ _--'-_ _-'-_.l.._-'--_'__---'
o 1000 2000 3000 4000 5000
k
2

Figure 3.10. On the boundary lines there are closed-loop eigenvalues on the hyperbola

The boundary w2 = 40" 2 - 0.25 with parameter 0" is mapped into the (k2 , k3) -plane
in Figure 3.10. There is an RRB for 0" = -0.25 (dash ed line). The CRB branches
off the RRB at a double root at 0" = -0.25, goes through infinity for 0" = -0.548,
intersect s t he RRB and itself, and goes to infinity again . The boundaries decompose
t he (k 2 , k3)-plane int o six regions A to F . Only region D is r -stable and Figur e 3.10 is
70 3 Eigenvalue Specifications

x 10'
Or----~---~---~--_,

-1

-5
-6 .

-7 .

-8'----~---~---~---'
2500 3000 4000 4500

Figure 3.11. Active boundaries from Figure 3.10

f[m]
4 3
16

8 1
~ 2

1000 2000

Figure 3.12. The operating domain and four representatives

-2

-8

1500 2000 2500 3000 3500 4000 4500

~

Figure 3.13. r-stabilization of the four vertices of Q

3.2 Root Sets, Gamma-stability 71

reduced to the active boundaries in Figure 3.11. K~I) is the set of all controllers (3.2.2)
that f-stabilize the crane for the operating condition mL =1000, =8. e
For a simultaneous f-stabilization of the four vertices of Q as shown in Figure 3.12,
the procedure is repeated for the vertices 2, 3, and 4 and the active boundaries are
plotted together in Figure 3.13. Obviously, the four vertices are simultaneously f-
stabilized by all controllers in the intersection of the four regions K~I) to K~4) in Figure
3.13. Again, only the active boundaries of the intersection are relevant, they are shown
in Figure 3.14. For this set of controllers, it turns out that only the operating conditions
two and four contribute to the active boundary.
k3 E4

-1.80

-2.15

-2.50
,,
-2.85 ,,
,,
-3.20 ,,
,,
-3.55 ,,
,,
-3.90 , ~

2.5 2.6 2.7 2.8 2.9 3.0 3.1 3.2 3.3 3.4 3.5 E3
k2
Figure 3.14. Intersection of the four regions of Figure 3.5

Two candidate controllers

u = -[500 2639 - 15255 0] x (3.2.3)

and

u = -[500 2865 - 22800 0] x (3.2.4)

have been marked . The first one minimizes k~ + k~; it is fragile, however, in the sense
that all variations in k2or in ki lead to a loss of I'-stability, Therefore, the more central
second controller will be used and further analyzed. It has I'-stability margins in all
directions.
Note that the relation between mL and k3 is very simple, according to (1.5.18) both
parameters enter only into a2 = (mL + mc}g + kle - k3 • The I' stable variations in
the two parameters are related by Ci.k3 = lOCi.mL with Ci.mL = mt - mi. . Therefore,
Figure 3.14 allows a graphical optimization. The largest f-stable variation Ci.k3 = 9680,
corresponding to Ci.mL = 968 [kg] is obtained for k2 = 2900. The Ci.mL interval may
then be shifted by k3 . For example , for mL E [0; 1000],eE [8; 16], the output feedback
controller (k 1 = 500, k4 = 0) is
72 3 Eigenvalue Sp ecifications

2
1.5

0.5
a 0

-0.5

-1

-1.5

-2
-3 -2 -1 0
C1

Figure 3.15. Root set for the controller (3.2.4)

u = -[5002900 - 320000] x. (3.2.5)

So far, we can only be sure that the four vertices of the Q-box, Figure 3.12, are f-
stable. A robustness analysis must now show if this still holds for all operating points
in Q. The result of a gridding analysis is shown in Figure 3.15.

35 .------;-~--.----.---.....---.--__.
I
I
30 I
I
25 I
I
20 I
I I
- 15 I

500 1000 1500 2000 2500 3000

mLlkgj

Figure 3.16. Robustness analysis in the (mL,f)-plane

A more elegant and reliable robustness analysis is, however, to map the f-stability
boundary af into the (mL' f)-plane as shown in Figure 3.16. In this example, the
f-stable region Qr is convex. Therefore, the simultaneous f-stabilization of the four
vertices of Q also I'<stabilizes the entire continuum of operating conditions in Q. Note,
3.3 Phy sical Meaning of Closed-loop Poles 73

however , that stability regions in parameter planes are not convex in general such
that the robustness anal ysis is non-trivial. This is, in particular , true for non-linear
depend ence of the polynomial coefficients on the physical plant parameters, see Section
4.5. 0

3.3 Physical Meaning of Closed-loop Poles

For a controllable and observable nominal plant, the eigenvalues can be placed arbi-
trarily by feedback of measured states or reconstructed states from an observer. If this
process is executed in one shot (or two shots by the separation principle) , then the
idea of relating the physical meaning of open-loop and closed-loop poles is lost. We
may assign the same set of eigenvalues (with different time scaling) to a power plant ,
an aircraft or a distillation column. The pilot , however, likes to recognize the familiar
Dut ch roll or short-period longitudinal mode in the behavior of the controlled aircraft.
Otherwise, it does not "feel like an aircraft" . Therefore, it is of interest to relate th e
closed-loop poles one-to-on e or pair-to-pair to the open-loop poles.

Exampl e 3.9

In the crane controller (3.2.1), let the loop-gain K grow continuously from zero
(open loop) to one (the anal yzed control system). The result of this eigenvalue tracing
is illustrated by the root locus of Figure 3.17. The higher frequency closed-loop poles
can be associated uniquely with th e pendulum and the lower frequency eigenvalues with
th e crab. This one-to-one relation ship also holds for the neighborhood of the nominal
case as defined by the uncertain parameter intervals, see Figure 3.7.
When is this one-to-one relat ionship destroyed? The answer is: as soon as we cross
a branching point of the root locus. As an example , take th e controller

u = -K [500 0 15508 0] x (3.3.1)

for the crane of Example 3.1. For K = 1, it yields a double imaginary root at ±jO.803.
If K is increased beyond I , then it becomes impossible to distinguish between crab
and pendulum eigenvalues. It is possible, for example, to design a bad pole placement
controller that lets the crab oscillate faster than the pendulum and thereby wastes
actuator energy. 0

A branching point of a polynomial p(s) is characterized by th e property

p(s) o and
dp(s)
O. (3.3.2)
ds
74 3 Eigenvalue Specifications

Root Locus
1.5 r--------:::=-----,

.C
0.5
CIl
'x
III
g> 01-------'----;-----;
§
-0.5

C
-1

-1 0
Real axis

Figure 3.17. Eigenvalue tra cing for continuously growing loop gain

A desirable feature of cont rol systems is their transparency, that means we try to
avoid crossing branching points. If t he polynomial depends on two uncert ain parame-
ters , i.e, P(S,Qbq2) , then the two equat ions in (3.3.2) may be mapp ed into th e (QI , Q2)-
plane , and the design path should not cross t hese lines. The par ameter space approach
allows us to trace the effect of cont inuous varia tions of controller gains sta rting from
the open loop with controll er gain zero.

3.4 Remarks on Existence of Robust Controllers

For syst ems with nomin al parameter values, there are well-known necessary and suffi-
cient conditions for pole placement design. A sufficient condition is that the plant is
controllable and observabl e. Then, an observer with arbitrary stable poles can recon-
struct all state-variables. By the separat ion theorem, it may be combined with pole
placement by state feedback without changing the individually assigned observer and
feedback poles. In a more detailed analysis, the controllability and observability of
individual eigenvalues is considered. It may be checked by the Hautus test [104], [105] .

Theorem 3.10 (Hautus )

An eigenvalue s, of a system :i: = Ax + BtL is controllable, if and only if rank

[A - s.I B] = n. An eigenvalue s, of a system :i: = Ax, y = Cx is observable,
if an d on lvy Iif ran k [ A -C s;I ] = n.
I

o
3.4 Remarks on Existence of Robust Controllers 75

An eigenvalue can be shifted by feedback from y to u, if and only if it is controllable

and observable. Obviously, a necessary and sufficient condition for f-stabilization is
that all open-loop eigenvalues outside f must be controllable and observable.
A fundamental difference between the classical stabilization problem for a fixed plant
and the robust stabilization of a plant family G(s, Q) is illustrated in the following
example.

Example 3.11
Consider the finite plant family
g(1)(s) = _1_, g(2)(S) = 2 . (341)
s-1 s-1 ..
The unstable pole at s = 1 is controllable and observable, because there are no
cancellations and the numerator is non-zero . However, the two plants cannot be simul-
taneously stabilized by a common linear controller.
First, assume proportional feedback c(s) = k. Obviously, g(1)(s) can be stabilized
by k > 1 and g(2) (s) can be stabilized by k < -1, but there is no k that stabilizes both
plants simultaneously. If we assume a compensator c(s) = (co +CIS + ... + Cmsm)j(do+
dIS + ... + dmsm ) and write the plant family as g(s,q) = qj(s - 1),q E {I, -I}, then
the coefficients of the closed-loop characteristic polynomial are
ao coq - do,
al clq+do-dl,

am cmq + dm- I - dm,

am+1 dm •
Now, the necessary stability conditions aj > 0, i = 0,1 , ... , m , imply di < 0, i =
0,1, . .. , m, but dm < 0 contradicts the necessary stability condition am+! > O. Thus,
the two plants (3.4.1) cannot be simultaneously stabilized by any linear compensator.
An explanation is, that the two plants with gains 1 and -1 are on different sides of the
case num g = 0, where an unstable eigenvalue becomes uncontrollaDlejunobservable. It
is generally recommended to check first that in the operating domain Q there is no loss
of controllability or observability of f-unstable eigenvalues before a feedback structure
is chosen and a search for good parameter values in it is started.
o
The example gives rise to a fundamental question of robust control: given a plant
family G(s, Q), does there exist a fixed compensator c(s) that stabilizes the plant fam-
ily? What is the order of c(s) ?
Unfortunately, an answer to this question exists only for very simple systems, e.g.
for a first order uncertain system [175] . Thus, in a scientific sense, it is premature
to synthesize a robust compensator c(s) . On the other hand, there are many control
systems in operation that show remarkable robustness with respect to uncertainty in
the parameters q. There is a practical need to analyze and design such control systems
and to provide algorithms and software tools for this purpose. In summary: the design
of a robust controller is more of an art than a science.
76 3 Eigenvalue Sp ecifications

3.5 Further Potential of the Parameter Space

Approach

The concept of mapping I'<regions from the s-plane into regions in a parameter plane
may be extended to more informative displays, to the design of gain-scheduling con-
trollers, and for more than two parameters. The following examples indicate some of
the potential for further developments of the methods and software tools.

Grey-tone or Color Coding

In Figure 3.14, it was easy to choose the centrally located point (3.2.4) from the ad-
missible set of controllers K r . A more detailed discussion on the choice of a controller
from the admissible set is done in Section 7.2 for a flight control example . In Figures
7.19-7.24, fault tolerance is the criterion; in Figure 7.15, specifications are t ightened in
order to reduce the admissible solution set. Such stepwise tightening of specifications
is supported by grey-tone coding .
As an example, take the region of Figure 3.11. It is the image of the I'<stable region
to the left of the left branch of the hyperbola w 2 = 4a 2 - 0.25. This hyperbola is now
embedded into a family of hyperbolas and the spaces in between are coded by grey tones
from black (= forbidden) to white (good). Figure 3.18 shows the f-stability grades and
their images.

Figure 3.18. Grey-tone coding of r-stability grades and their images

Since we are interested in intersections, it is recommended to have the admissible

region white on white pap er printouts. It is then further reduced by other operating
conditions like in Figure 3.13 or additional criteria like in Figures 5.26 and 5.46.
On the computer screen, grey ton es may be replaced by colors (e.g. red = forbid-
den , yellow = possible, green = good) . For example, frequency domain performance
grades may be color coded and overlayed to P-stability regions in order to visualize
where eigenvalue and frequency domain criteria are in agreement and where they are
in conflict, see Chapter 5.
3.5 Further Potential of the Parameter Space Approach 77

Gain-scheduling Design
In the crane example, we have used the (k2 , k3)-plane (Figure 3.14) for the design of
a robust fixed-gain controller and the (mL, f)-plane (Figure 3.16) for the robustness
analysis. A third possibility arises if an uncertain parameter is measured and used for
scheduling a gain.

Example 3.12

In the crane example, let me = 1000, mL = 1500, k 1 = 500, k 2 = 2865, ka =

-22800 and find a gain-scheduling controller k 4 (f ) that complies with the f-stability
specifications postulated in Figure 3.9 even at very low rope lengths f . This is of
practical interest for a construction crane operating on the top floor of a building under
construction. The assumption here is that the rope length can be measured (e.g. from
the rotation of the rope drum) and used to schedule the feedback k 4 of the rope angle
rate. (The rope angle rate can be obtained by forming the derivative of the measured
rope angle by means of a suitable filter.) Figure 3.19 shows the mapping of f-stability
into the (f , k4)-plane for two different values of mL (mL = 1000 with grey line color and

5
4
3
2
..,. 1
-<
0
-1
-2
-3
-4
-1 0 0 10 20 30 40 50
( [m]

Figure 3.19. Gain scheduling of k4 (f) for the crane

mL = 2000 with black line color). The bold line in the f-stable region represents the
gain-scheduling k4 (f ) = -30 000 + 1778 f., which is adopted here. So the gain-scheduled
feedback system is Ivstable for a much larger interval of f , which is also recognizable
in the corresponding parameter space plot shown in Figure 3.20.
o
Gain-scheduling design in a (q, k)-plane shows that the interpolation between linear
controllers is quite arbitrary and may therefore be done as a linear interpolation (as in
the crane example) or by a simple polynomial. For the practical implementation, this
78 3 Eigenvalue Sp ecifications

50r---r---.-----.----.--~--..._-___.

1500 2000 2500 3000

me [kg]
Figure 3.20. An enlarged operating domain (in terms of f) becomes feasible by gain-scheduling
according to Figure 3.19

approach is more useful than other interpolations: e.g. based on poles/zeros or state
space models .

More than Two Parameters

So far, we have shown that the parameter space approach is a very intuitive tool : as long
as the user accepts the limitation of two parameters at a time. Design with more than
two controller parameters can be broken down to sequential steps with two parameters
in each step, as shown in Section 2.4 on sequential pole shifting.
Tomographic renderings for a third gridded parameter have been used for PID-
controller design in Figure 2.11. An early contribution to 3D renderings of stability
regions was given in [170]. In the meantime, much software has been developed to
visualize the 3D world on a 2D screen. Motion of objects or point of view and light
sources, stereoscopic vision, moving 2D slices under the control of a joystick and color
coding may also be exploited to visualize 3D f-stability regions and their intersections
for the purpose of design. The f-stable regions may thus be explored interactively.
This approach may be extended to more than three parameters. Then, however, the
user needs some help with navigation in higherdimensional spaces . Slices are best
suited for looking at intersections of sets as they occur in simultaneous design for
several representative operating conditions. The human user is very good and fast at
recognizing intersections of 2D-sets , even in cases where a lot of comput at ion would be
needed to calculate the intersection.
For robustness analysis for a given Q-box, a third and further parameters may
be gridded and all planes are projected onto one plan e. The projection of Q-box is
a rectangle and for robust f-stability it must be avoided by all projected f-stability
boundaries.
3.5 Further Potential of the Parameter Space Approach 79

Example 3.13

As an example, consider the robustness analysis for a crane . In Figure 3.16, the
r -stable region was shown for a crab mass me = 1000. It enters into the characteristic
polynomial (1.5.18) via a2 = (mL + me)g + klf - k3 (linearly) and via a4 = fme
(bilinearly). In Figure 3.21, me is gridded in the interval me E [800 ; 1200]. All curves
must avoid the operating domain Q for robust stability with respect to me.
o

,-m- ;
,-
30 .... ,...,. .:

25 -
,-
- ~ . . .:...
.
,-..

500 1000 1500 2000 2500 3000

"\.[1<91

Figure 3.21. f-stability boundaries by gridding me which enters together with e bilinearly
into the characteristic polynomial

Similarly, a fourth (or fifth, etc.) parameter may be gridded and all boundaries
are projected to a (qh q2)-plane where ql and qz are treated continuously. Also, any
other pair of two parameters may be treated continuously by choosing the appropriate
projection plane .
Alternative methods for dealing with a larger number of parameters in robustness
analysis of polynomials with specially simple coefficient functions a(q) will be treated
in Chapters 8 and 9.
4 Gamma-boundary Mapping into
Parameter Space

Chapter 3 has provided th e motivation for mapping boundaries or of r -stabilit y regions

into various planes of two real par ameters. In Figure 3.10, a hyperbola was mapp ed
into a plane of two gains for design. In Figure 3.16, the same hyperbola was mapp ed
into a plan e of two plant parameters for robustness analy sis, and in Figure 3.19 it was
mapped into a plane of one measurable plant parameter and one gain for design of a
gain-scheduling system .
Th e mathematical formalism for these three mappings is exact ly th e same, although
th e interpretation and use of t he resulting r -stable regions is quite different in t he t hree
cases. Thi s chapter deals with th e mathemati cal formalism and we call all (real) pa-
ram eters qi, independent of t heir meanin g as plant par ameters or contro ller par ameters.
For e param eters, i = 1, 2, .. . , e, t he par ameter space or q-space is e-dimensional. T he
decision of how we choose a two-dimensional linear or affine subspac e for grap hical rep-
resentation in a parameter plane may be taken lat er after we have derived the general
equat ions for mapping into the q-space.
An eigenvalue region r is charact erized by the equati on of the boundary or . In the
case of th e left half plane, the bound ary is described by

or := {s I s = jw , 0:::; w :::: 00 } . (4.0.1)

The restriction to non-n egative w is mad e becaus e polynomials with real coefficients
are assumed. Hence, for each complex eigenvalue, also a conjugate complex eigenvalue
exists . Therefore, in general th e r -root regions should be chosen to be symmetric with
respect to the real axis. Th e bound ary or of the desired root region r can be described
by
or := {s I s = 0-(0') + jw(O') , 0' E [0'- ; 0'+] }. (4.0.2)

The lower and upper bound s 0'- and 0'+ may also be + 00 or - 00. Th e scalar par ameter
0' is called th e generalized f requenc y. For generalizing th e results on Hurwitz-stability
to r -stability, s = jw has to be replaced by s = 0-(0') + jw(O') and s = 0 by s = a where
a is an int ersect ion point of o r with t he real axis.
Assume th at t he real functions 0-(0') and w(O') , or implicit descript ions by
f [o-(O' ),w(O')] = 0, are polynomials or simple trigonometri c functions. By choosing
polynomials of degree less or equal to two, root regions can be described , which are
4.1 Algebraic Problem Fotmuletion 81

bounded by conic sections, i.e. straight lines, circles, ellipses, hyperbolas and parabo-
las. For most applications, it suffices to compose the boundary or from segments of
conic sections.

4.1 Algebraic Problem Formulation

For Hurwitz-stability of a polynomial p(s, q) with real coefficients ai(q), continuous in

q, the three hypersurfaces

CRB: Re pOw, q) = 0 , Im pOw, q) = 0, (4.1.1)

RRB : ao(q) = 0, (4.1.2)
IRB : an(q) = 0, (4.1.3)

represent the boundary surfaces around a nominal point qO , which corresponds to a

stable polynomial. Eliminating w from the two equations (4.1.1) gives the last but one
Hurwitz determinant ~n-l, see Theorem 2.2.
The extension to other pole regions r is straightforward. Instead of s = jw in
(4.1.1), now s = da) + jw(a) is substituted. Compared to Hurwitz-stability, there are
only minor differences. The resulting equations are more complicated depending on the
complexity of the functions a = a(a), w = w(a) , which describe the boundary of r
Eliminating a by means of the resultant method (see Appendix A) gives the analog to
the last but one Hurwitz det erminant, that is,

Reso(Rep[a(a)+jw(a),qj, Imp[a(a)+jw(a),qJ) =0. (4.1.4)

The restriction to a real interval a E [a- ;a+] is lost by the elimination of a . Therefore,
fictitiuous boundaries are not only generated by complex values of a , but also by real
a-values outside the abov e interval. For the example of the hyperbola in Figure 3.8, the
symmetric branch in the right half plane would generate a fictitiuous boundary as well
as the extension of th e hyperbola beyond the bandwidth circle. Also, the bandwidth
circle would be mapped as an entire circle. Therefore, a piecewise definition of or
becomes impractical. We do not pursue the algebraic approach, because in the case of
r-stability there ar e even more good reasons to prefer the parameter space approach.

4.2 Parameter Space Mapping

A real root boundary (RRB) arises for each real axis intersection of or. Such intersec-
tion at s = ao is mapped into q-space by

p(aO' q) = o. (4.2.1)
82 4 Gamm a-boundary Mapping into Parameter Space

Similarly, an infinit e 'root bo und ary (IRB) may result from

lim p(a(O')
0 -00
+ jw(O'), q) = O. (4.2.2)

The mapping of the complex root boundary (CRB) into the set of all parameter vectors
q that yield a root pair at s = a(O') + jw(O') , i.e.

QCRB(O') := {q I p(a(O') + jw(O'), q) = 0, 0' E [0'- ; O'+]} (4.2.3)

is formulated in the following theorem [23] .

Theorem 4.1 (eRB Mapping, Ackermann, Kaesbauer)

Consider a polynomial family p(s, q) = [1 s ... sn] a(q) with real continuous co-
efficient function a(q) .
Now, q E QCRB(O') if and only if

(4.2.4)

for some 0' E [0'- ; 0'+], where

do(O') 1,
d1(O') 2a(O'), (4.2.5)
dH1(O') 2a(O')di(O') - [a 2(O') + w 2 (O' )]di-1(O') , i = 1,2, . . . , n - 1.

Proof

Consider the polynomial p(s, q) for fixed q = «: It has a complex conjugate pair
of roots on or
at a(O') ± jw(O') , if and only if

(4.2.6)

where
(4.2.7)
is an arbit rary polynomial of degree n - 2 with real coefficients. Equival ently
(omitting t he dependence of a and W on 0' for notational convenience) ,

1 - 2a a 2 +w2 0 0 0
0 1 - 2a 0
0 0 1 0 ro
a (q*) = (4.2.8)
a 2 +w2
-2a
0 1 rn - 2
4.2 Parameter Space Mapping 83

The matrix in (4.2.8) is triangular with identic al elements on the diagonals. There-
fore, its inverse b has the same structure. The entries dj of b are determined
from
1 -2a a 2 +w2 0 0

~l
do d 1
0 1 -2a
0 do
0 0 1 0
=1, (4.2.9)
a +w 2
2
d1
0 -2a
do 1
which implies

do = 1,
-2ado + d 1 0,
(a 2 + w2 )do - 2ad 1 + d 2 0,

(a 2 + w 2 )dn _ 2 - 2adn _ 1 + d n 0,

and solving for d; yields

do 1,
d1 2a ,
d2 Zad, - (a 2 + w 2 )do,

dn = 2adn _ 1 - (a 2 + w 2 )dn _ 2 .

By premultiplying (4.2.8) by b . it can be concluded that p(s(o),q*) = 0 if and

only if
o
o
a(q*) = (4.2.10)
. d1
o 0 do rn - 2
The last n -1 rows of (4.2.10) have an undetermined right hand side because the
remainder polynomial r( s,q*) is arbitrary. In other words, whatever the left hand
side requires, the equation is satisfied by the appropriat e choice of r j. There only
remain the first two rows. Hence, it can be concluded that q* E QCRB(O) if and
only if
do(o) d1(o) . . . dn(O)] a(q*) = [ 0 ] . (4.2.11)
[ o do(o) . .. dn-1(o) 0
The fixed value q* may be replaced by the general vector q to obtain (4.2.4).
Thus, the proof of the theorem is complete.
o
84 4 Gamma-boundary Mapping into Parameter Space

For short notation, Equation (4.2.4) is written as

D(o:) a(q) = o. (4.2.12)

This formulation clearly separates the r-boundary description by D(o:) from the q-
parameter-dependent coefficient vector a( q) . In the parameter space approach, the
scalar parameter 0: is gridded and the complexity of (4.2.12) depends on the coefficient
function a(q) . If it is linear, then the edge theorem (see Chapter 8) offers an interest-
ing alternative, also for more parameters. If the coefficient function is multilinear or
polynomial, however, then the parameter space approach is superior.
The polynomial p(s, q) has a complex conjugate pair of roots on ar, if and only if
there exists an 0: E [0:- ; 0:+1 such that (4.2.4) holds. In other words, in analogy to the
w-sweep over the imaginary axis in the robust stability test of Chapter 2, an o-sweep
along all branches or segments of ar must be made. For each 0:, the set QCRB(O:) can
be calculated from (4.2.4).

Example 4.2

Let ar be the imaginary axis, i.e. a = 0, choose the parametrization 0: = w2 , 0: E

[0; 00).

do 1,
di 0,
di + i -o:di - i,

[~ 0
1
-0:
0
0
-0:
0:
0
2
::: ] a(q) = [ ~]. (4.2.13)
0

Example 4.3

Let ar be parallel to the imaginary axis at a = ao.

do 1,
di 2ao,
di+i 2aodi - (a~ + w 2 )di _ i .
Again, 0: = w 2 is chosen.

[~ 2a 3a5 -
1 2a
0: 4ao(a5 - 0:)
3a5 - 0: : : : ] a(q) = [ ~]. (4.2.14)

o
4.2 Parameter Space Mapping 85

Example 4.4

Let or
be a line of constant damping D , see Figure 3.2, with w 2 = 0'2{1 - D2 )j D2 .
Choose the parametrization 0 = 0' , 0 E [0; -00).

do 1,
e, 20 ,
di+l 20d; - {02jD 2)d;_l '

(4.2.15)

Example 4.5

Let or be the left branch of the hyperbola

-
(~r (~r = 1, 0' E [-a ; -00) . (4.2.16)

The hyperbola is shown in Figure 4.1. This hyperbola ensures a minimum damping

jw
of
r
b

I
I
I
/
/
I
/
/
/
/
/
/

Figure 4.1. A hyperbola combines constraints on damping and real parts of the eigenvalues

of D = ~ , corresponding to the asympt ote s, and a maximal real part of the

2
a +b
2
eigenvalues of 0' = -a.
The parametrization is 0 = 0' , 0 E [-a ; - 00) .
From {4.2.16},
86 4 Gamma-boundary Mapping into Parameter Spa ce

do 1,
d, 20' ,
di + 1 = 20'd; - [ci{1 + (b/a) 2W] d i _ l ,

(4.2.17)

Example 4.6

Let ar be a circle with center s = 0 and radius R, i.e. 0'2 +w2 = R 2. Choose 0' = a ,
0' E [-R ; R].
1,
20',
20'di - R 2 d;_I,
2 2
.[ 1 20' 40'2 - R 40'{20'2 - R ) . •. ] () = [ 0 ]
(4.2.18)
o 1 20' 40'2 - R 2 . .. a q O'
RRB 's occur for 0' = - R and 0' = R . There is no IRB , because r is finite. o

Remark 4.7
Equation (4.2.18) may be further simplified in terms ofthe powers of 0', see Chapter
11. Since 0' is gridded anyway, the effect on computing time is not dramatic. 0

Example 4.8

Th e r-stability region of Figure 3.8 can be composed from the hyperbola equation
(4.2.17) and the circle equation (4.2.18).
a2(R2 + b2)
The circle and the hyperbola intersect for a real value 0'1 = - 2 b2
a +
Thus , (4.2.18) applies for 0' E [-R; ad and (4.2.17) for 0' E [0'1 ; -a]. There are
two RRB 's at 0' = ....;R and 0' = -a. There is no IRB because r is finite. 0

In the last example, the r-region was defined as an intersection of two regions.
Also, a union of regions can be useful. If the union consists of disjoint sets, then it
must also be specified how many roots are located in each disjoint set. A practical
use of a union of sets is the following. Assume you have designed a controller for a
nominal plant model, i.e. you know the nominal location of all closed-loop eigenvalues.
If the uncertain parameters vary, you want to keep the eigenvalues in some well-defined
neighborhood of their nominal locations.
4.2 Parameter Space Mapping 87

Example 4.9

Assume that the nominal polynomial closed-loop characteristic in factorized form is

p(8) = (4 + 28 + 82)(2 + 8)(5 + 8)(8 + 8)(72 + 128 + 82).

jw

f 1
f3

~
f2
-4 -2 a

~
Figure 4.2. Several components may constitute a r-stabi]ity region

The roots are shown in Figure 4.2, they represent a dominant Butterworth pole
configurat ion in a distance 2 from the origin and some eigenvalues that are located
further to the left. f-stability is now defined by the admissible migration of individual
eigenvalues or groups of eigenvalues . Take, for example, a circle of radius 0.3 around
the eigenvalues at 81,2 = -1 + j yI3. Then it is important to also keep the real eigenvalue
at 8 = -2 in the interval a E [-2.3 ; -1. 7], otherwise the low damping D = 0.5 of the
two eigenvalues 81,2 yields too much overshoot of the step response, see Figure 3.5. It is
not necessary to care about the other eigenvalues as long as they do not interfere with
the dominant behavior. Therefore, the requirement is that their real part must remain
less th an Ul = -4. The disjoint parts af 1 , af 2 , and af3 of th e boundary are shown in
Figure 4.2. Note that the real axis interval contains only one real root ; therefore, af 2
consists only of the real root boundaries at U2 = -2.3 and U3 = -1.7. The left region
f 1 has a real root boundary at Ul = -4 and its complex root boundary is 8 = -4 + jw,
w 2: O. The circle around the dominant complex poles is describ ed by

af 3 = {8 18= -1 + 0.3 coso + j(V3 + 0.3 sin 0 ) , 0 E [0; 271")}. (4.2.19)

o
88 4 Gamma-boundary Mapping into Paramet er Space

Example 4.10

For sampled-data systems (see Chapter 11), the constant damping lines a = cw,
(c = -DJv1- D2, D = damping) are mapped via z = esT = el7TeiwT = e=TeiwT into
the z-plane.
Choose a parametrization by polar coordinates 0 = wT, 0 E [-7T ; 0]. Then the
real and imaginary parts of z = 7 + j"7 are 7 = eCQ cos 0 and "7 = eCo sin o . They now
play the role of a and win (4.2.5).

1,
27 = 2eco cos 0 , (4.2.20)
27 d, - (72 +w2 ) di - I
2eCQ coso di - e2 CQ di - I •

This is an example where the algebraic (Hurwitz-like) approach fails. The bound-
ary mapping, however, is not particularly difficult. The RRB's are p(l , q) = 0 and
p(_e C7r , q) = O. There is no IRE. 0

Example 4.11
In Example 4.9 there was a real axis segment a E [0'- ,0'+] with only one root , and
the two RRB's pia> , q) = 0 and pia" ; q) = O. If there are more roots in the segment,
then they can also leave it through a double root into imaginary direction (aperiodicity
constraint) . A polynomial p(s) has a double root at s = s", if and only if p( s*) = 0 and
p'(s*) = 0, where p'(s) = dp(s)Jds . Therefore, (4.1.1) has to be replaced by

p(o,q) = 0, p'(o, q) = O. (4.2.21)

The resultant with respect to p(s) and p'(s) is called the discriminant of p(s) . 0

Remark 4.12

A related boundary representation was derived by Siljak [200]. Instead of the start-
ing point (4.2.6), he uses the real and imaginary part of p[o'(o) + jw(o)] . The k-th
power of s is expressed by [s(o)jk = [0'(0) + jw(o)jk = Xk(o) + jYk(o). Then

p[o'(o) + jw(o)] Re p(XI(o) + j YI(o)) + j Im p(XI (0) + JYI (0))

n n
L akXk(o) + j L bkYk(O) = O. (4.2.22)
k=O k=O

The advantage of the form (4.2.4) is that only one recursion for dk is needed instead
of two recursions for X k and Yk • The two equations for real and imaginary parts are
4.2 Parameter Space Mapping 89

related to (4.2.4) by

Rep[a(n) + jw(n)] ]
[ Imp[a(n) + jw(n)]
= [1 0
-a(n)] [do(n) d1(n)
w(n) 0 do(n)
dn(n) ]
dn-1(n) a(q).
(4.2.23)
o

The toolbox PARADISE (see Appendix B) contains a r-editor [27] . The user com-
poses the r-region from standard elements by a graphical user interface or specifies it
numerically. The r-editor then generates D(n) for (4.2.12). The second ingredient a(q)
for (4.2.12) is provided by symbolic generation of the closed-loop characteristic poly-
nomial, i.e. a(q) , for example , from a SIMULINK model. This completes the mapping
into the t'-dimensional q-space.

Example 4.13

For the crane example , choose the hyperbola w 2 = 4a 2 - 0.25, i.e.

(0~5) 2 = 1, a = 0.25, b = 0.5. Equation (4.2.5) becomes

do(a) 1,
d1(a) 2a,
d2 (a ) 0.25 - a 2 ,
d3 (a ) a - 12a 3 ,
d4(a) 0.0625 + 0.5a2 - 19a 4 ,

and the CRB-mapping equation is

The coefficient vector is, by (1.5.8) ,

All plots for the crane in Chapter 3 are generated from D(a)a(q) = 0 by fixing five of
the seven parameters. The two free parameters are ka, k3 for design, mL , t' for robustness
analysis and t', k4 for gain-scheduling. In each case, the resulting equations are linear
in the parameters. They would be bilinear, however, for combining t' with k 1 , k 2 or me .
o
90 4 Gamma-boundary Mapping into Parameter Space

For the visualization of the result, the user specifies an interesting subspace of the
q-space. In the case of robustness analysis, the choice of two plant parameters qi, qj is a
natural one. Further plant parameters may be gridded and projected, as in Figure 3.21,
or they are assigned to a joystick such that the user can move them back and forth in
order to explore the higher dimensional stability region. For design, two gains ki , kj in
one design step are appropriate, but also a linear subspace at some odd angle may be
useful, like in the invariance planes of Section 2.4. Again, further gains or eigenvalues
may be assigned to a joystick.
In gain-scheduling design, it may be of interest to schedule two gains by the same
measured parameter. If the 3D-stability region (with the f-unstable region non-
transparent) can be rotated such that it is possible to peek through it, then this line of
sight describes a linear gain-scheduling control law.
In all cases, it is important that the symbolic calculations for D(o)a(q) are done
beforehand, such that the remaining numeric calculations can be executed online.
An interesting application is also a comparison of three controller candidates
k(1) ,k(2) and k(3) (one of them may be the open loop k(l) = 0) as they are obtained by
different design approaches. The three points define an affine subspace

(4.2.24)

for visualizing f-stable regions that help in the decision on a trade-off.

4.3 Generalized Singular Frequencies

In Section 2.5, the concept of singular frequencies Wk on the imaginary axis of the s-
plane was introduced. It is generalized now to singular frequencies Ok on the f-stability
boundary s = a(o) + jw(o), 0 E [0- ; 0+1.
Consider again the polynomial (2.5.1) with linear dependency on ql and qz, i.e.

(4.3.1)

Let s = a(o)+jw(o) and determine the two equations (4.2.4) for the chosen f-stability
boundary. Substitute the coefficient vector

(4.3.2)

to obtain two linear equations in ql and q2 of the form

[ ~~~~~~ ~~:~~~] [~~] + [ ~~~~~~ ] = [ ~ ] . (4.3.3)

Again, a generalized singular frequency 0 = Ok is defined by the condition

rank [Cll(O) C12(0) CIO(O)] = l. (4.3.4)

C21(0) C22(0) C20(O)
4.3 Generalized Singular Frequencies 91

Exampl e 4.14

Continue Example 4.13 with the fixed parameters me = 1000, mL = 1000, k j = 500,

I
k4 = O. The r-stability boundary is w 2 = 4a 2 -
0.25, a :::; -0.25. Equation (4.2.4) with
Q = a becomes

1 2a 0.25 - a
[0 1 2a
2 3
a - 12a 0.0625 + 0.5a - 19a
0.25 - a 2 a - 12a3
2 4
] [~~~~o +
k
500i - k
3
= [ 0 ] .
O '
2f
1000f

(20 + f)a - 12fa -0.25 + a

3 2
] [k 2 ]
[ 10 + 0.25f - fa 2 -2a k3

10000 + 187.5f - 20000a 2 - 19000fa4 ]

(4.3.5)
+ [ (40000 + 20 OOOf)a - 12000fa3

Is there a rope length f , for which a generalized singular frequency aj on ar occurs?

Candidates for aj are the roots of the determinant

d(a 2 ) [(20 + f)a - 12fa 3 ]( -2a) - (10 + 0.25f - fa 2 )( -0.25 + a 2 )

2.5 + 0.0625f + (-50 - 2.5f)a 2 + 25fa 4 = O.

2
The solutions of this quadratic equation in a 2 are a~ = f + 0.05 and a~ = 0.05. In the

range of a s -0.25, the value a2 = ±VO .05 is ruled out, and al = -J~ + 0.05 :::; -0.25
2
is possible for f :::; 160. The second condition for satisfying (4.3.4) is for a~ = f + 0.05

-0.25 + a~ 10 000 + 187.5t - 20 OOOa~ - 19000fa4

-2 40000 + 2000( - 12000ea~

= 1-0.2 + 2/ f 140e + 5200 - 116000/f 1

-2 1400e + 16000

= ~ (-0 .2f + 2)(1400f + 16000) + 2(140f 2 + 5200f - 116000)

= lOOO (f - 20) = O.
f
The answer to the above question is: yes, f = 20, the corresponding a -value is aj =
-VO.15, i.e. the roots can cross ar only at s = -VO.15 ±jvO.35.
The example illustrates that non-generic situations in a world of nominal parameters
(the probability of hitting exactly f = 20 is zero) become likely events in the world of
interval parameters.
92 4 Gamma-boundary Mapping into Parameter Space

For all £ =I- 20, Equation (4.3.5) is solved for

1000a[1O - 2£(0.25 - 5( 2)2J
(0.25 - 5( 2)[10 + £(0.25 - 5( 2 ) ] ,
5
2a£k2(a , £) + 1000 [0.25 _ 5a 2 + 20 + £(0.75 - 2
= ( ) ].

The RRB at a = -0.25 is the straight line

k3(£) = -(40 + 0.25£)k2 + (10000 + 562.5£).

The f-stability boundaries for a) £ = 19, b) £ = 20, and c) £ = 21 are plotted in Figure
4.3 . Like in Figure 2.5 two regions (A, F) that are disjoint for £ < 20 become connected
for £ > 20. In this case, the two regions are unstable. The stable region D does not
change much by variations of £ around 20.
o

4.4 Gamma-stable PID-control

In Section 2.6, it was shown that Hurwitz-stable regions in a (K D , K/ )-plane of a PID-

controller (i.e. for fixed K p) are convex polygons. Does this polygon property also hold
for other regions of f-stability? A complete characterization of all f-stability regions
with the above property will be given in this section. The polynomial

(4.4.1)

has a root at a + jw if and only if

Re Pio + jw , K/, K» , K D ) = 0, and
(4.4.2)
Im Pia + jw , K/ , K» , K D ) = O.
For the separation of real and imaginary parts, write

A(a + jw) = R A + jIA ,

Bio + jw) = R B + JIB'
Then

[~: ] (4.4 .3)

4.4 Gamma-stable PID-control 93

a)

b) 'r"''-'~""""""~~~,......,,~~-__-_---,

b)

c)

Figure 4.3. a) f = 19, b) f = 20, c) f = 21. In the neighborhood of the singular rope length
f = 20, the curves change direction of approach to the asymptote. For f = 20, (J = -VO.15,
a singular boundary in form of a tangent occurs instead of the asymptote
94 4 Gamma-boundary Mapping into Parameter Space

In [71J a coordinate transformation is introduced in order to treat f-stability for the half
plane to the left of a parallel to the imaginary axis at s = ao. This idea is generalized
here in order to answer the question : which other I'-stability regions yield the same nice
geometric properties as the left half plane? The answer follows from Equation (4.4.3)
by introducing a linear coordinate transformation

K J = turo + t12rl + t13 r2,

(4.4.4)

Choose Tl as the fixed parameter (other choices only lead to index permutations) . Then
Equation (4.4.3) for R p = 0, I p = 0 read

(4.4.5)

RB + rdt l2RA + t22(RAa - lAw) + t32[RA(a2 - w2) - 2IAaw]} ] [ 0]

+[ I B + rdtl2h + t22{IAa - 2 2
RAW) + t32[I A(a - W ) + 2R Aaw]}
=.
0
The 2 x 2-matrix multiplying [ro r2]T is analyzed for singularity.

(4.4.6)

J can be symbolically factorized into an A-dependent term, a factor w , and a factor

that only depends on the transformation but not 011 A.
..J (4.4.7)

a t 21t33 + t23t31 ,
b tu t33 + t 13t31l

J vanishes for the RRB at w = 0 and for

a(a 2 + w2 ) + 2ba + c = o. (4.4.8)

For a = 0, Equation (4.4.8) represents a parallel to the imaginary axis at

ao = -c/2b. (4.4.9)
The only other possibility is a i- 0, leading to the circle

(a + b/a? + w 2 = cia - (b/a)2, (4.4.10)

4.4 Gamma-stable PID-eontrol 95

with center m = -bla and radius r = Jela - (bla)2.

Such circles are useful for bandwidth limitation in th e s-plane and for stability and
performance analysis of discrete-time systems in the z-plane.
Note that the transformation matrices are not unique. Simple forms are

(4.4.11)

T
- [r2_~m2 ~1 -~m] for circle-stability. (4.4.12)

An alternative to the linear transformation (4.4.4) is a conformal mapping that maps

the r-region onto the left half plane.
For the shifted half plane, substitute

s = v + C10. (4.4.13)

The original polynomial is C1o-stable if and only if the transformed polynomial

P'(v) P(v + (10) = A(v + C1o)[K/ + Kp(v + (10) + Ko( v + (10) 2] + B(v + (10)
A(v + C1o)[(K/ + KpC10 + KoC1~) + (K p + 2Ko(10 )v + K ov 2]
+ B( v+C1o) (4.4.14)

is Hurwitz-stable. Introduce the transformed PID parameters

(4.4.15)

Then for fixed K~ , convex polygons in the (K~, KJ)-plane are obtained as described
in Section 2.6. Also, the intersection for several representative operating conditions is
done in this plane. By gridding K~, the tomographic rendering of all simultaneously
C1o-stabilizing PID-controllers is obtained. A (centrally located) controller K~, K~, K~
is chosen and transformed back to the PID-controller parameters

(4.4.16)

Also, a circle with real center and real axis intersections at C1L and C1R , Figure 4.4, can
be mapped into the left half v-plane by the bilinear transformation
VC1R - C1L S - C1L
s= V=--. (4.4.17)
v-I ' s - C1R
96 4 Gamma-boundary Mapping into Parameter Space

jw

(1

Figure 4.4. The circle is mapped to the left half plane

As shown by (4.4.8), these are the only two examples for which the convex stable
polygon property holds for a space of the same dimension as given by the degree of
Pis, q) . If we allow also a non-linear transformation
v + b1v + bo
2
s = ---,,----- (4.4.18)
v 2+alv+aO'
then a large class of eigenvalue regions can be treated, which includes, for example, all
conic section boundaries [187] . The price is, however, that the transformed polynomial
in v has twice th e degree of P(s, q).

4.5 Non-linear Coefficient Functions

A particular strength .of the parameter space approach in comparison to most other
methods of robust control is that the two parameters, for which f-stability regions are
plotted, may also enter in a multilinear or polynomial fashion into the coefficient vector
a(q) of the closed-loop characteristic polynomial. This will be illustrated first for a
bilinear example .

Example 4.15

Consider Hurwitz-stability of the uncertain polynomial

The equations of the real and imaginary part of p(jw, q) are

(2.04 + 6ql + 6q2 + 2qlq2) - (2 + ql + q2)W 2 0,

w [(2 + ql + q2) - 2
w ] O.
4.5 Non-linear Coefficient Functions 97

Th e RRB for w = 0 is the hyperbola

Substituting q2 = -qj - 2 + w 2 from the imaginary part into th e real part gives
2
2q; + (4 - 2w )qj + w4 - 6w 2 + 9.96 = O.

The solution to this quadratic equat ion is qj = -(1- 0.5w2) ± V-3.98 + 2w2 - 0.25w 4 .
The interval of real qj is bounded by the roots of

w4 - 8w2 + 15.92 = 0,

i.e. w 2 = (20 ± V2)/5 , and real qj occurs in the frequency interval w 2 E [(20-
V2)/5 ; (20 + V2)/5] ~ [3.71; 4.28], w E [1.93 ; 2.07]. At the w2-boundary, we have

qj = -(1 - 0.5w 2) = 1 ± V2/1O,

and q2 = -qj - 2 + w 2 = 1 ± V2/1O.
For w = 1.93, the point qj = 0.86, q2 = 0.86 is obtained and for w = 2.07 the result
is qj = 1.14 , q2 = 1.14. Correspondingly, the two solutions qj(w 2) and the associated
values q2 = -qj - 2 + w 2 are calculated for frequencies in th e admissible interval.
Figure 4.5 shows th e circular stability boundary. The center ql = 1, q2 = 1 yields th e
unstable polynomial 16.04 + 48 + 4s 2 + S3 , thus the system is unstable inside the circle.
D

w = 2.07

'P
/
CRB

w = 1.93 st able

unstable

Figure 4.5. The polynomial is unstable inside the circle

Remark 4.16

For t he calculation of t he CRB , it is tempting to eliminate w 2 from the equa tions

for the real and imaginary parts, yielding
98 4 Gamma-boundary Mapping into Parameter Space

Figure 4.6. Root set of Example 4.14. The edges of the Q-box are stable, but some interior
points are unstable

i.e. the equation of a circle with center ql = 1, q2 = 1 and radius 0.2. (This is, in fact ,
how the example was constructed.) But for higher degree polynomials, this algebraic
approach leads to very complicated expressions for det H n - 1 = 0, see Section 2.1.
o

Choose Q by ql E [0.6 ; 1.3], q2 E [0.7 ; 104], this operating domain encloses the
unstable circle. Plot the root set for a 15 x 15 grid. Figure 4.6 shows the part of the
root set near the crossover of the imaginary axis. There is also a third real root, which
is stable for all grid points . The solid lines are the images of the edges of Qj they are
stable .
In the frequency interval w E [1.93; 2.07], the root set crosses the imaginary axis.
Note that these boundary frequencies are independent of the chosen Q-box.
It will be shown in Section 8.4 that an unstable island as in Figure 4.5 cannot occur
if the coefficients of the characteristic polynomial depend linearly on the uncertain
parameters. Then, a f-instability inside Q will also show up on the edges of Q. It is
the bilinear term qlq2 in ao that destroys this nice property.

Singular Frequencies
In Section 2.5, we have introduced the notion of singular frequencies for polynomials
with linear parameter dependence. It is now generalized to the case of non-linear
parameter dependence . Consider a root that moves in the root set in Figure 4.6. When
it reaches the right hand boundary of this set, it must return. Finding such extremum
means that a derivative becomes zero. In the case of two independent variables, i.e.
Re p(jw, ql, q2) and 1m p(jw , q},q2) this derivative is the Jacobian matrix
4.5 Non-linear Coefficient Functions 99

(4.5.1)

and the extremum condition is

det J{w, q},q2) = o. (4.5.2)

Example 4.17

For Example 4.15, the Jacobian is

det J{w, q},q2) = 2w{q2 - ql) = O.

The case w = 0 is the RRB. The other factor gives ql = q2, which is w-independent
in this simple example . Therefore, the envelope of the root set of Figure 4.6 can be
calculated from the edges of Q and the segment of the line q2 = ql that is contained in
Q. For ql = q2 , the polynomial becomes

The boundary generating q's and their image in the s-plane are shown in Figure 4.7.
The dense gridding of Q is avoided in this approach.
q2
2.2 ,/
,/
,/
,/
2.1 ,/
1.4
S
2 ~/'--.detJ~ 0
1.9
0.7 ,/
,/
1.8 ,/
,/
,/
,/
1.7
-8 -6 -4
a
-2 0 2
0.6 1.3 ql
x 10-3

Figure 4.7. The image of points on J = 0 inside Q is the right hand boundary of the root set

o
100 4 Gamm a-boundary Mapping into Parameter Space

For £ uncertain paramet ers, t he Jacobian is a rectangular matrix

J(w ,q) = (4.5.3)

and the extremum condition is

rank J(w, q) < 2. (4.5.4)

The test of thi s rank condition involves £-1 determinant conditions on 2 x 2 submatrices
of J.
With these preparations we can now give a general definition of singular frequencies.

Definition 4.18

The non-negative freqency w. is a singular frequency of the uncertain polynomial

p(s , q) if there exists a qO E lie
such that the three conditions

0, (4.5.5)

rank J(W., qO) < 2,

are simultaneously satisfied. o

Note that w = 0 always satisfies the rank condition and makes 1m p = 0, thu s
the RRB ao(q) = 0 is a singular frequency. The notion of singular frequencies is not
standard in the literature; the concept has been used, however, by several authors, e.g.
[75], [200], [201] .
In the above example, the Jacobian is linear in ql and q2 and can be treated easily.
In general , J(w , q) is non-linear in q and also w-dependent. Then the calculation of
the additional rank condition in (4.5.4) and finding the simultaneous solutions of the
three equations in (4.5.4), for example, by the resultant method of Appendix A, is quite
involved. In this case, gridding of the Q-box may be more practical.
If the coefficients a( q) depend polynomially on the uncertain parameters, then
(4.2.4) may be written as

f = fO(q2) + !I (q2)ql + + !k(q2)q~ = 0, (4.5.6)

9 = 90(q2) + 91(q2)ql + + 9m(q2)q;" = O. (4.5.7)
fi and 9i are polynomials in q2 and continuous functions of Q . (4.5.6) and (4.5.7)
describe the intersection of two algebraic curves in the (ql , q2)-plane. These solution
pairs (ql, q2) can be calculated either with th e resultant method, see Appendix A, or
by construction of a Grabner basis.
4.5 Non-linear Coefficient Functions 101

0.05 +·0.35s + 0.50s 2 K

- r---- (1 + ST)8
S

Figure 4.8. Robustness analysis for uncertain K and T

Example 4.19

In Example 2.19 we have designed a PID-controller for a plant with non-linear pa-
rameter dependency. Then a robustness analysis for the loop of Figure 4.8 is necessary.
The characteristic polynomial is

p(s, K, T) = K(0 .05 + 0.35s + 0.50s 2 )

+ s(1 + STs + 2STs 2S2 + 56T3s3 + 70T4s4
+ 56T5s5 + 2Sy6 s 6 + ST7S7 + T 8s8).

Let s = jw , then the equations Rep(jw , K, T) = 0 and Imp(jw , K, T) = 0 have the

Jacobian

oRep ORe p]
J(w,K ,T)
e« or
8Imp
[ - 8Imp
- --
ex er
1/20 - 1/2w 2 -Sw 2 + 16Sw 4T2 - 280w6T4 + 56w8T6 ]

[
7/20w -56w3T + 2S0w 5T3 -16Sw7T5 + 8w9T7
The resulting boundary in the (T, K)-plane was shown already in Figure 2.15.
o

Example 4.20
8
Consider the polynomial p(s, qI, q2) = L:: aisi with
i=O

aO 453.106 q;,
al 528 . 106 q; + 3640 . 106
qI,
a2 5.72.10 q2q; + 113.10 q; + 4250.106 ql,
6 6

a3 6.93 . 106 q2q; + 911 . 106 ql + 4220 . 106 ,

102 4 Gamma-boundary Mapping into Parameter Space

a4 '1.45.106 q2qi + 16.8 .106 q2qI + 338.106 ,

as 15.6 . 10 q~qi + 840q2qi + 1.35 . 106 q2qI + 13.5 . 106 ,
3

a6 = 1.25 · 103 q~qi + 16.8q2qi + 53.9.10 3 q2qI + 270.10 3 ,

a7 = 50 q~qi + 1080q2qI,
ag q~qi.

This problem comes from the robustness analysis of a track-guided bus, where qi is the
velocity and q2 is the quotient of mass and friction coefficient. The technical background
will be described in Section 6.8.

m[103 kg] 70

60

50

40

30

20

10

10 20 30 40 50 60 70
v[m ·s-1J
Figure 4.9. The track-guided bus is robustly r-stable in the operating domain Q

The closed-loop poles should be located to the left of the left branch of the hyperbola

(4.5.8)

The real root boundary for a = -0.35 is described by

p( -0.35, qI, q2} =

-79 .66(qiq~ - 5339qiq2 - 3077qIQ2 - 3540 213qi + 9946 676ql + 22088 293} = O.
(4.5.9)

This curve can be plotted by gridding one of the parameters, say q2, and solving the
remaining quadratic equation in ql.
The complex root boundary or is parameterized by a(o) = 0, w2(0 ) = 250 2 -
1.752 , 0 E (-00; -0.35]. The coefficients d; in the boundary representation (4.2.4) are
4.5 Non-linear Coefficient Functions 103

given by

do(a) 1, (4.5.10)
d1(a) 2a, (4.5.11)
di+1(a) = 2adi(a) - (26a 2 - 3.0625) di- 1(a). (4.5.12)

Substituting the di and a(q) into (4.2.4) and collecting terms with the same power of
q1, the following form is obtained:

foo(a) + [ho(a) + f11(a)q2]q1 + [J20 + 121 (a)q2 + h2(a)qi)]q; 0, (4.5.13)

900(a) + [91O(a) + 911 (a)q2]q1 + [920 + 921 (a)q2 + 922(a)qi)]q; = O. (4.5.14)

The resultant has the form

(4.5.15)

If all its roots are complex for a given a = o", then there exists no real pair (qb q2)
for which the closed loop has an eigenvalue at u(a·) + jw(a·) . On the other hand, if
there are real solutions q~i)(a·), then the corresponding q~i)(a·) is given by the roots
of the greatest common divisor of (4.4.13) and (4.4.14). In this example, the q~i){a·)
corresponding to q~i)(a·) can be expressed immediately by the coefficients fij and 9ij,
omitting the argument a (see Appendix A):

q1 = _I f20 + h1q2 + h2q~ flO+ f11q1 I:I 120 + h1q2 + h2q~ foo I. (4.5.16)
920 + 921 q2 + 922q~ 910 + 911 q1 920 + 921 q2 + 922q~ 900

It turns out that the polynomial (4.5.15) has two real roots so that the complex root
boundary consists of two branches. In Figure 4.9 the root boundaries are plotted in
the interesting domain of the parameters velocity (v = qd and virtual mass (m = q2) .
The real root boundary is the dashed line, the complex root boundary is the solid line.
Assuming the operating domain is given by q1 E [3; 20] and q2 E [9.9; 32], then the
track-guided bus is robustly r-stable.
o
5 Frequency Domain Analysis and Design

In the previous chapters, all specifications were restricted to the eigenvalue domain .
With respect to the dosed-loop transfer functions, their denominator p(s, q, k) was
considered. Now, the effect of the numerators will also be taken into account . The aim
of this chapter is to combine a variety of frequency domain criteria with the parameter
space approach. The sequence of the chapters thus follows the historic evolution of
the parameter space technique. It was initially developed for mapping of f-stability
boundaries into parameter space. However, it is obvious that complete design and
analysis of control systems requires consideration of more criteria beyond mere eigen-
value specifications. This chapter distinguishes between two types of frequency domain
specifications, which will be made accessible for parameter space mapping:

• Frequency loci specifications

The most common locus used for linear control system analysis in frequency do-
main is the Nyquist plot . Moreover, there are non-linear stability criteria that
refer to frequency loci, like the dual locus method, which uses describing func-
tions as an approximation for non-linear elements, and the Popov criterion which
is used for proving absolute stability in the presence of sector non-linearities .
• Frequency magnitude criteria
Frequency domain magnitude specifications put bounds on the magnitude fre-
quency response of specific transfer functions or sensitivity functions. Thus , track-
ing quality, noise and disturbance rejection , and robustness against unstructured
uncertainty (e.g. unmodelled dynamics) can be addressed . Using these criteria
with the parameter space technique provides the possibility of combining the ad-
vantages of non-conservative mapping of inequalities into parameter space with
knowledge and experience in the field of frequency domain control design (e.g.
Hoo-loop shaping) and analysis (e.g. j1.-analysis).

5.1 Frequency Loci Specifications (Theta-stability)

In this section , a number of frequency loci type of specifications is introduced before

e-stability [631 ,[641 is defined as a generalization .
5.1 Frequency Loci Specifications (8-stability) 105

Nyquist Plot-related Specifications

The following example illustrates that it may not be sufficient to consider only eigen-
value specifications.

Example 5.1

Consider the uncertain single-input single-output (8180) system with open-loop

transfer function
L(s q) = y(s) = -ql(5q2 S- 8) (5.1.1)
, e(s) 5s2+5(q2+2)S+2 '
It depends on the uncertain parameter vector q = [q}, q2]T. Assuming unity feedback
e = w - y according to Figure 5.1, the closed-loop transfer function is

(5.1.2)

At first, an eigenvalue specification in terms of an admissible region I' is formulated : all

d
e y
L

Figure 5.1. 8180 system feedback structure

eigenvalues of T(s, q) are required to be in the region to the left of the hyperbola being
displayed in Figure 5.2. By mapping the f-stability boundaries into the (ql, q2)-plane,
the f-stability region Qr in Figure 5.3 is obtained.
In the sequel, the effect of variations of q2 is investigated , while ql = 1 is assumed.
Now the characteristic polynomial (denominator of T(s , q) in (5.1.2)) does not depend
on q2 anymore. The closed-loop poles are Sl,2 = -1 ± j, thus the system is f-stable
for all values of q2 . However, the Nyquist plot G(jw, q) = L(jw, q) according to (5.1.1)
still depends on Q2. The positive frequency branches (w :2: 0) of the Nyquist plots are
displayed in Figure 5.4 and Figure 5.5 for values Q2 = 1 and Q2 = 10, respectively.
Neither of the frequency responses encircle the critical point -1 , so the corresponding
closed-loop systems are both stable . However, if gain and phase margins of the Nyquist
plots are considered, the situation turns out to be completely different for the two
different Q2 parameter values.
Now, as an additional specification, it is postulated that the Nyquist plot must be
completely contained in the region e, which is displayed in Figure 5.4 and Figure 5.5.
In other words, the Nyquist plot must avoid the region e, which is the complement of
e
8. (The region continues towards -00 beyond the left edge of the diagrams .) This
106 5 Frequency Domain Analysis and Design

0.5
o
.§ o~---t---H So -2 ....--'

..,l.5

-. -6

" '!' .5
'-_ -"- u
-, -05 1.5 2 2.5 3 3.5
Re ql

Figure 5.2. Definition Figure 5.3. Mapped f-stable region Qr of T(s , q)

of a f-region in the in parameter space
eigenvalue plane

implies that the locus must not intersect the boundary 8e. At first , this specification
guarantees Hurwitz-stability for the closed loop, if the open loop is stable (i.e. q2 > -2)
according to the Nyquist-stability criterion. Moreover, it warrants a phase margin of
7r/ 4 and a gain margin of 2. The Nyquist plot for q2 = 1 satisfies this new specification.
However, for q2 = 10, it does not . 0

If the new criterion is to be included in robustness considerations, the question is: for

q2 = 10
O.5~ · ·' · · · · · · · · · · ·· · · · · · · ·:- · · · · · · · · · · · I 0.5

01-i-.....,.--'!ir----+--1

S
..... -0.5 .,:.
8 8

-11- ··:······ ·· ·· · · ····· ···· ···· " -11-- -:·· ··· ···· .., :- :: ,

-1 -0.5 o -1 -0.5 o
Re Re

Figure 5.4. Nyquist plot for ql = 1, q2 = 1 Figure 5.5. Nyquist plot for ql = 1, qz =
avoids the forbidden region e 10 intersects the forbidden region e
5.1 Frequency Loci Sp ecifications (8-stability) 107

which parameter values of q does the open loop comply with the postulated properties
regarding the shape of the Nyquist plot? This question will be answered in this section
for:

• Frequency loci of rational functions of the frequency w, where the coefficients of

the rational functions depend polynomially on the parameters q•.
• Almost arbitrary shapes of the region 8 in the complex plane in which the lo-
cus must be contained. This type of specification is denoted 8-stability and is
analogous to the notion of r -stability in the eigenvalue plane.

The suggested specifications will be mapped using the parameter space approach. This
provides the possibility of including various locus criteria for the synthesis and analysis
of parametric control systems. In particular, it creates the option of robustness analysis
of systems w.r.t . non-linear locus criteria like the dual locus method and the Popov
criterion. The foundations of this method have been developed and demonstrated by
the challenge of practical problems [62],[16] .
Over the next pages, the foundations of two non-linear frequency locus criteria are
recapitulated, which can be applied to non-linear systems. If the reader is familiar with
the dual locus method and the Popov criterion, then he or she may directly proceed to
the introduction of 8-stability.

Describing FUnctions and Dual Locus Method

Limit cycles are periodic oscillations performed by non-linear systems. Without external
input signals, the oscillations sustain with a certain frequency and amplitude. Since in
many cases they are not desired, criteria that enable non-linear stability analysis are
useful. One approximation method, that can be applied for analyzing the existence
and properties of limit cycles for a class of non-linear systems is the dual locus method
[95],[200] . This method requires several assumptions. The first assumption is that it
is feasible to represent the open-loop system as a series connection of a single non-

-- D
u y
G(s) f--

Figure 5.6. Single loop non-linear system

linearity n and the remaining linear part. In the sequel, the linear part is represented
by its frequency response G(jw). Therefore, the (control) system consists of a single
loop as depicted in Figure 5.6.
108 5 Frequency Domain Analysis and Design

The analysis of limit cycles starts with the assumption that the system is in the
state of a sustained oscillation. Another assumption is necessary for the application of
the dual locus method: the linear part COw) needs to have distinct low-pass properties
in the frequency range of the considered oscillation and at higher frequencies. As a
rule of thumb, a drop of -40 dB/decade is required . Then it is reasonable to assume
that the signal u, which is present at the input of the non-linearity, is sinusoidal. This
holds since it is equal (except for the sign) to the output signal of the linear part. Since
the higher harmonics are attenuated by the low-pass effect of C(jw), the output of the
non-linearity can be approximated by its first harmonic .
Thus, the consideration of the non-linearity can be restricted to its transmission
of sinusoidal input signals, and a linear approximation of the non-linear system can
be obtained. This approach is called harmonic linearization. A describing function is
defined as the frequency response from a sinusoidal input signal

u = Asinwt (5.1.3)

to the first harmonic Yg of the output signal Y, i.e.

= Yg(w)
N( w, A) u(w) . (5.1.4)

For static characteristics, the describing function depends only on the input amplitude
A, because input and first harmonic of the output are in phase and N(A) is real.
The describing function may also be applied to non-linear elements that produce a
frequency- and amplitude- dependent phase shift. Then the describing function N(w , A)
is complex. For some elementary non-linearities, the describing functions can be derived
analytically [200] by Fourier series expansion of the periodic signal y.
If the system shown in Figure 5.6 is in a sustained oscillation and the above-
mentioned assumptions hold, then the transmission properties of the non-linearity can
be approximated by its describing function . This leads to the condition
-1
N(w ,A) C(jw) =-1 or C(jw) = N(w, A)' (5.1.5)

which is denoted harmonic balance. If this equation holds for a pair (w, A), then the
system is capable of performing an oscillation with this frequency and amplitude. The
representation on the right hand side in (5.1.5) provides the foundation for the graphical
dual locus method: limit cycles are possible if there are intersection points between the
locus C(jw) of the linear part and the locus of the negative-inverse describing function
-1/N (w, A). From the parameterization of both loci at the intersection, the values of
wand A can be determined as properties of the corresponding limit cycle. Whether
the limit cycle is stable (i.e. persists after a perturbation) or unstable may also be
determined. In the latter case, the oscillation either decays or, after some time, turns
into another limit cycle with different frequency and/or amplitude or the amplitude
goes to infinity. For determination of the limit cycle's stability, the locus of COw) is
pursued in the direction of increasing frequency in the vicinity of the intersection point
with the locus of the negative-inverse describing function -1/ N(w , A) . If the left hand
side of -1/ N(w , A) corresponds to higher input amplitudes than on the right hand side
5.1 Frequency Loci Specifications (8-stability) 109

of the intersection point, then the limit cycle is stable. Otherwise, it is not stable. In
the example depicted in Figure 5.7, the lower of the two intersection points is stable
according to the criterion just mentioned. The upper intersection point corresponds to
an unstable limit cycle. Note that the stability of limit cycles has nothing to do with

O~-~"""""'2IIII'-~----l

-1

......s -2

-3
-1
N(w, A)
-4 ~-----'-----'--------'
-2 -1 o
Re

Figure 5.7. Determining the stability of limit cycles

the stability of the respective system . The goal of applying the dual locus method in the
context of this book is to provide a means for robustness analysis w.r.t . non-existence
of both stable and unstable limit cycles. The parameter space method can be applied
to determine stability regions where no intersections of the two loci exist. However, it
has to be kept in mind that an approximation is used and, consequently, the results
are not exact . In particular, the low-pass property has to be checked separately for
the stability regions obtained to investigate the reliability of the results. The notion of
8-stability here refers to ,the requirement that the locus G(jw) of the linear part does
not intersect the locus of the negative-inverse describing function .

Remark 5.2

Note that the existence of intersection points between the two loci does not neces-
sarily mean that the corresponding system will perform limit cycles. The system will
not perform limit cycles if its state does not leave the region of attraction of an asymp-
totically stable equilibrium state. The region of attraction cannot be determined by
the dual locus method. If this is the aim , then, e.g. Lyapunov theory can be applied .
We understand the specification established here in the way that no limit cycles may
occur (independent of the initial system state). 0
110 5 Frequency Domain Analysis and Design

For the analysis of limit cycles by means of the dual locus method, it is prerequisite
to obtain the describing function of the non-linearity involved. For that purpose, the
describing functions of two elementary non-linearities are now derived as examples. For
a detailed description of describing functions and their derivation, see [200].

Example 5.3 (describing junction of a saturation)

With a saturation (also denoted limiter) , the absolute value of its output signal is
limited to r.:

iflul~rs
(5.1.6)
y ee { :, signlu) if lui> r .
s

The symmetric characteristic curve is linear until the saturation value is reached . For
the sake of simplicity, unity slope about zero is assumed . The symbol for a saturation is

-'I~]I-
u y

Figure 5.8. Block diagram of a limiter

shown in Figure 5.8. For derivation of the describing function, the saturation is excited
with sinusoidal input signals (5.1.3) of various amplitudes A. Figure 5.9 shows the
input signals and the corresponding saturated outputs (dashed line style). If A ~ rs ,
then the output is equal to the input and the describing function value is equal to 1.
The more the input amplitude A exceeds the saturation value rs , the more of the input
signal gets truncated, which causes the amplitude of the output's first harmonic to
decrease. Thus, the gain of the describing function continuously decays with increasing
Alr, > 1. Since there is no phase shift between input and output, the describing
function of the saturation is real. It does not depend on the frequency. In [200], an
analytical expression is provided :

if A ~ r,

(5.1.7)
if A > r s .

Figure 5.10 shows the locus of Ns{A). It starts with A ~ r, at 1 and tends to zero
along the real axis for A/r. -+ 00 . The negative-inverse describing function -l/Ns{A)
starts at -1 and tends towards -00. 0
5.1 Frequency Loci Specifications (8-stability) 111

AtsJ AtsJ -1

-
0.5 Ns(A) <. N:,(A)
o 0_
01-·. ·••·• ·••:.·. · ... ._ .

AESJ AtsJ
-A -A : -
Nr 5 < 1 Nr 5 = 1.2
_S -0.5 . ... A
· . Ii- s . Alr~ ·
-1
0 ·
- -- · o~~~ ~_w~-
-1.5
__ _ _-'-_ _--...l
-A -A _21.-~- ~

Nr 5 =5 Nr 5 =20 -4 -2 o 2
Re
Figure 5.9. Sinusoidal input signals Figure 5.10. Loci of the saturation de-
(solid line style) with various amplitudes scribing function Ns(A) and its negative-
and respective saturation time responses inverse -l/Ns(A)
(dashed)

Example 5.4 (describing function of a rate limiter)

The effect of a rate limiter is to limit the absolute value of the derivative iI of a
signal y, i.e.
iI = R sign(u - y). (5.1.8)
This dynamic non-linearity is not representable by a static characteristic. Nevertheless,
a describing function can be derived . Figure 5.11 shows an ideal realization of the

~
u il 1 Y
0- -
s
- -

Figure 5.11. Block diagram of a rate limiter

rate limiter. It consists of a closed loop featuring an integrator and a symmetric two
point switch (output value - R or R respectively) . The derivative iI of the rate limiter
output signal y is present at the output of the two point switch and thus its absolute
value cannot exceed R. On the other hand, with a constant input u the two point
switch perpetually chatters between -Rand R and thus retains y at the input value u.
In Figure 5.12, some time responses of the rate limiter are shown. Various sinusoidal
input signals are applied with different frequency wand amplitude A. The shape of the
output signal in relation to the input signal only depends on the ratio wAI R. Therefore,
the describing function only depends on this composed parameter. The steady state
oscillation may be categorized into three ranges [80] depending on the value of wAI R:
112 5 Frequ ency Domain Analy sis and Design

J:~ J:SJ
o
I~i ' \N~ (w , Ar~
'~ '
-0.5 ' , i tfa, (~! ~) , ... .. .. . . .. - -

• wA/R
•i
w*NR < 1 w*AlR =1.2
-S -1

J=SJ ·•
-1 .5
!~ wA/R
'.' :
-
' , ,',

:
-A ' -_ _---='---J
-1 -0.5 o 0.5
w*NR = 1.862 w*NR = 20 Re

Figure 5.12. Sinusoidal input signals (solid Figure 5.13. Loci of the rate lim-
line style) at various values of wAfR iter describing function Na{w, A) and its
and respective rate limiter time responses negative-inverse -l/Na {w, A)
(dashed)

• Range I: if wAf R ~ 1, then the rate limiter does not change the signal (y = u) .
Consequently, in range I the describing function value is 1.
• Range III: if wAfR ~ J(7r /2)2 + 1 ;;:,; 1.862, then the output signal y is a "tri-
angular function" of time with maximum rate of change . Only at the reversal
points is the output equal to the input. In range III , an analytic expression for
the describing function can be derived [80] :

_ 4R -J. arccos ( 2wA

7rR )
Na(w,A) - - A e . (5.1.9)
7rW
• Range II: between range I and range III, there is a transition range where the
rate of change of the input signal is only partly greater than the maximum rate
R, which the rate limiter permits. The output signal y matches the input signal
only if litl < R, otherwise it follows the input -signal at maximum rate of change
R. In this transition range, sophisticated mathematical effort would be needed
to derive an analytical describing function . Alternatively, th e determination of
Na(w, A) may be done numerically for range II. The Fourier coefficient of the first
harmonic of the output signal is computed by numeric integration. For a grid
of values of wAf R in the interval [1; 1.862], the gain and phas e of the output
signal 's first harmonic are put in relation to the input signal. Note that this
procedure can be generically applied to other non-linearities, if no convenient
analytic representation of their describing functions is available .

Figure 5.13 shows both the locus and negative-inverse locus of Na(w, A) . In contrast
to the saturation, a phase shift occurs as soon as th e rate limiter gets activated. In
the limit case (wA/R -+ 00), the phase shift is -7r/2. For large values of wA/R,
the amplitude of the output signal drops towards zero. Therefore, the locus of th e
describing function Na(w, A) ends at the origin ofthe complex plane, with the imaginary
axis as an asymptote. Note that (5.1.9) is a parameterized representation of a circle
5.1 Frequency Loci Specifications (8-stability) 113

in the complex plane (radius 4lrr2 , center (0,4/rr 2 ) ) . In range III this is the locus of
the rate limiter describing function. If the locus of its negative-inverse -1/ Na(w,A) is
plotted, this circle becomes mapped to a straight line, which is parallel to the imaginary
axis. This straight line's real part is _1r 2 /8 ~ -1.234. The maximum imaginary
part of the range III section of -l/Na(w, A) is -1r/4 ~ -0.785, which corresponds to
wAf R = J(rr /2)2 + 1. There, the transition range (range II) begins, which ends at the
point -1. The latter represents range 1. Figure 5.22 shows how the negative-inverse
describing function locus of the rate limiter can be well approximated by three sections
formed by segments of diverse conic sections. This is a suitable representation for the
application of the parameter space method, which will be explained in Section 5.2. 0

With the dual locus method and the Popov criterion (which will be explained next)
non-linear systems can be analyzed for the existence of limit cycles or absolute stability,
respectively. Note that with the describing function approach, we assume that the
parametric uncertainty only affects the frequency locus representing the linear system
part, the non-linearity is assumed certain. With the Popov criterion, however, some
uncertainty of a considered non-linear characteristic is allowed in terms of the sector in
which it resides.

Popov Criterion
In this section, we consider the feedback loop of the system

u(s) -G(s), y(s)

(5.1.10)
y f(u),

which is shown in Figure 5.14. Let us first assume that the rational transfer function

-0 y
G(s) fo--

Figure 5.14. Non-linear function in the feedback loop

G(s) is known. However, the function f is considered uncertain; it is assumed to be

continuous and to belong to the class of functions

F k := U(u) I f(O) = 0, 0::; uf(u) s ku 2

} , 0< k < 00 . (5.1.11)
114 5 Frequency Domain Analysis and Design

That means it is only known that the graph of f is within a sector [0; k], limited by
the abscissa and by the line
f(u) = ku. (5.1.12)
We call this sector the k-sector, with notation [0; k]. In the k-sector, the graph of the
continuous function f can be arbitrary, see Figure 5.15. If for a given transfer function
G(s), the above non-linear closed-loop system is globally asymptotically stable, then it
is called absolutely stable in the sector [0; k] . A classical result for absolute stability is
the sufficient Popov criterion [168],[200],[198] :

Figure 5.15. Sector of non-linearity

Theorem 5.5 (Popov)

For f a continuous function in the class F k of (5.1.11) and G(s) a strictly proper
rational transfer function with all poles in the open left half complex plane, the
feedback loop of Figure 5.14 is absolutely stable, if there exists a ( E lR such that
the inequality
11k + Re {(I + jw()G(jw)} > 0, Vw ~ 0 (5.1.13)
holds.

o
The classical Popov criterion represents a strong robustness result , since f is as-
sumed to be an uncertain continuous function within a sector [0; k] .
If G(s) is unstable, then the Popov criterion obviously cannot hold for any k > 0
because the zero function (J(u) = 0, Vu E R) is in Fk. However, we can now ask for a
sector [k1 ; k21 for which absolute stability can be proven. For instance, in [200],[198] it
is shown that this more general problem can be transformed to the canonical problem
given above: if the open-loop transfer function G(s) of Figure 5.14 is unstable, but
can be stabilized by a constant negative feedback with gain p > 0, then replace in the
Popov criterion above the transfer function G(s) by O(s) = G(s)1 (1 + p G(s)) and the
non-linear function f(u) by j(u) = f(u) - p u, see Figure 5.16. It is obvious that the
two paths involving p cancel each other. If application of the Popov criterion to the
5.1 Frequency Loci Sp ecifications (8-stability) 115

j(u)
------------------------------------------- ----------------------,-IC(S)
1
p p 1
I
1
1
1
1
1
C(S)
1
II 1
II I
I_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_--=-_--=-.!I -------------------- -~

Figure 5.16. Transformation to the standard form

transformed system yields a sector [0; k] for t. then the original system is absolutely
stable in the sector [p; k + pl.
Values for k guaranteeing absolute stability can easily be determined graphically by
means of the Popov plot

Cp(jw) := Re C(jw) + jw 1m C(jw), w 2: 0 (5.1.14)

of the stable transfer function C(s). Consider any straight line y = (1/()(x + 11k) in
the (x, y)-plane with k > 0, intersecting the negative x-axis at -11k and with a slope
of 1/(. For all points on the right hand side of this line, we have 11k + x - ( y > O.
Let x = Re C(jw), y = wlm C(jw) , then

11k + Re C(jw) - (w 1m C(jw) > 0 (5.1.15)

for all w 2: 0, if the Popov plot Cp(jw) := Re C(jw) + jw 1m G(jw) lies on the right hand
side of the above line. But (5.1.15) is identical to the expression (5.1.13) , so we have
the following result : any straight line in the complex plane crossing the negative real
axis at the point z < 0, and with the property that the Popov plot lies totally on its
right hand side, yields k = -1/ z for a sector [0; k] of absolute stability. For such a k
and with 1/( as the slope of the corresponding Popov-line the Popov criterion (5.1.13)
is satisfied.
With the Popov plot , we can easily find the maximal value of k with absolute
stability. It follows from the tangent Popov-line as illustrated by Figure 5.17. The
sector with maximal value of k is called the Popov-sector. For instance , k = -1/ Z
determines the Popov-sector for a C( s) with the Popov plot of Figure 5.17 (with the
corresponding slope 1/( = tan(a) of the line). But note that the Popov criterion is a
sufficient condition . Therefore, we cannot exclude in general conservative results , i.e.
the possibility that for absolute stability a greater sector than the Popov-sector exists.
116 5 Frequency Domain An alysis and Design

1m

Re

Figure 5.17. Absolute stability and Popov-sector

Remark 5.6

A good hypothesis for testing is the so-called Aizerman Conjectur-e. It relates the
Popov-sector with the Nyquist-sector corresponding to the interval of stable linear gains .
The latter is obtained from the intersections of the Nyquist plot with the negative real
axis. In view of (5.1.14), these intersections are identical to those of the Popov plot.
A situation that frequently arises is shown in the Popov plot of Figure 5.18. Here,

Figure 5.18. An example for identical Nyquist- and Popov-sectors

the tangent to the Popov plot in its leftmost real axis intersection is a feasible line for
satisfaction of the Popov criterion, i.e. the Popov-sector and the Nyquist-sector have
the same upper bound which in this case is obviously both necessary and sufficient for
absolute stability.
Note that Aizerman's conjecture does not hold for the Popov plot shown in Fig-
ure 5.17. 0
5.1 Frequency Loci Specifications (8-stability) 117

Remark 5.7

A family of Popov plots gp(jw, q), q E Q results in a family of tangent Popov-lines .

The leftmost real axis intersection of all tangent Popov-lines yields -1/k+ (i.e. the

Re

Figure 5.19. Different Popov-lines apply to different members of a Popov plot family

maximal common Popov-sector [0, k+]), see Figure 5.19. Note that this e
is greater
or equal to the k resulting from a common Popov-line (dashed line in Figure 5.19) for
the family Cp(jw , q) , q E Q. 0

8-stability
Analogous to the definition of P-stability for eigenvalues is the notion of 8-stability
for frequency loci, which will now be defined. The following depiction is made with
a Nyquist plot (see Figure 5.20). However, all statements can analogously be applied
to other frequency loci (e.g. the Popov plot) . Only the locus branch with positive
frequency (w :::: 0) is considered .
With the example shown in Figure 5.4, the 8-definition specified that the region
e is to be avoided by the Nyquist plot . In Figure 5.20, another practically relevant
example for Nyquist diagrams is shown. Here, the admissible region 8 is the exterior
of a circle centered at the critical point. The non-zero radius of this circle guarantees
some Nyquist-stability margin .

Definition 5.8

Let 8 be an open region of the complex plane . The boundary of this region is

88 = {x + jy I Fae(x,y) = O} . (5.1.16)

The complement of 8 is denoted e. Since 8 is open , B8 c e. The frequency locus

COw, qo) is defined to be 8-stable, if

C(jw , qo) c 8 , vw > 0, (5.1.17)

118 5 Frequency Domain An alysis and Design

88

Figure 5.20. Example of a 8-stability definition in the Nyquist diagram

i.e. if the locus resides completely within 8. The frequency locus C(jw , qo) with w 2: 0
is defined as 8-limit-stable, if C(jw, qo) is 8-stable except for the points of contact with
the boundary 8e. 0

Robust e-stability
Analogous to robust f -stability is robust e-stability, which is defined: a system is
robustly e-stable in the operating domain Q, if for all operating points q E Q the
frequency locus C(jw , q) of the respective system is e-stable. For the sake of linguistic
simplicity, th e notion of e-stability is also applied to a parameter vector q" , for which
C(jw, q*) is 8-stable.
Assume qo is a 8-stable operating point, i.e. the locus C(jw , qo) avoids the forbidden
region 8 . Now we are interested in the set of all possible continuous variations q of qo
such that the locus C(jw, q) does not lose the property of e-stability. If the continuous
variation is started with a different e-stable operating point ql ' then possibly a different
set of 8-stable q-values is obtained. This set consequently does not cohere with the
first set. The union of all existing subsets of this kind is called the 8-stability region
Qs :

Definition 5.9

The e-stable region in the q- space is defined as

Qs = { q I C(j w, q ) c e, Vw E lRt }. (5.1.18)

o
5.1 Frequency Loci Specifications (8 -stability) 119

Figure 5.20 shows a nominal 8-stable locus COw , qo) (bold line style) and a fam-
ily of loci COw, q) (normal line styl e) with varying parameters q. For certain values
of q the locus COw, q) becomes 8-limit-stable (dashed) , where the locus is tangent
to the boundary 88 of the forbidden region 8 . The point of contact is indicated by
"+". The special situation where COw, q) becomes 8 -limit-stable will be used in Sec-
tion 5.2 to establish mathematical conditions for mapping 8-stability boundaries into
parameter space. This is analogous to f-stability boundary mapping. In Section 2.2
and Section 4.2, the condition that eigenvalues are exactly located on 8f provides the
equations to be solved for the determination of f-stability boundaries in parameter
space.

Representation of the 8-stability Boundary

Before mapping frequency loci margins, the specific 8-stability definition has to be made
by fixing the admissible region 8 and thereby also fixing its forbidden complement 8 .
In the sequel, it is assumed that 8 is defined by establishing adequate boundaries 88 of
8 in terms of mathematical curve equations Fae(x , y) = O. It is further assumed that
the boundary is put together from piecewise smooth boundary sections. (This way of
forming 88 is very similar to the way f-boundaries were composed in Section 3.2.) The
transition points where the boundary sections border on each other in a non-smooth
but continuous way are denoted jun ctions Zi .j ' For illustration, Figure 5.21 refers to
the example shown in Figure 5.4 and Figure 5.5. As will be discussed later, for the

0.2 .

01-.....- .....- ----4.----+----j

-
-0.2 .
S
-0.4 .

-0.6 '
e .
8

-0.8 .

_1l-.----i.----'------'-----'----'
-1 .5 -1 -0.5 o
Re

Figure 5.21. Complete definition 86 for the example

mathematical effort, it is very advantageous to compound the boundary from segments

of conic sections (i.e. straight lines, circles, ellipses, hyperbolas and parabolas) . In this
way, a good approximation of practically relevant 8-stability region boundaries can be
constructed. See Figure 5.21 as an example for a compound boundary.
The complete compound boundary 88 according to (5.1.16) is formed by the col-
120 5 Frequency Domain Analysis and Design

lectivity of its sections:

Fae! (z, y), if x + jy E Cae! (x, y)

Fae(x,y) = Fae2(x, y), if x + jy E Cae2(X, y) , (5.1.19)

where Cae; is a region in the complex plane in which the respective function Fae;
is defined. The upper horizontal boundary segment in Figure 5.21, for instance, is
characterized by Fae2 = Y = 0 and Cae2 = {x + jy I x :S -0.5}.

Example 5.10 (approximation of the negative-inverse locus of the rate limiter describing
function by segments of conic sections)

- - - - - - ~- - -
,I
1- _

-,

\}\ .~~
-0.1621

\ z23

212

1 - - - ' - dThelal

-1
Re

Figure 5.22. Approximation of rate limiter negative-inverse describing function by conic sec-
tion segments

Conic sections are second order curves, which can be represented in the (x, y)-plane
in the following form:
1:'
rae; ( (i) 2
x,y ) =anx + 2 a(i)12xy+a22y
(i) 2
+ 2 alOx+
(i) 2 a(i) (i)
2oy+aOO =
0 (5.1.20)
As illustrated by Figure 5.22, the locus of the negative-inverse describing function of
a rate limiter (see p. 112) can be piecewisely approximated by three segments of conic
sections:
i. Straight line:

= 8'
(1) (1) (I) (I) (I) / (I) 11"2
Fae1(x , y) an = a l 2 = a 22 = a 20 = 0, a lO = 1 2, a oo

Cae!(x,y) {x + jy I y :S -~} .
5.2 Mapping of Frequency Loci Margins into Parameter Space 121

ii. Parabola:

Fae2(x, y) a~;) = a~~ = 0, a~~ = -1/2,

(2) (2) a 11" (2) 11"2
a22 = a, a 20 = 4 ' aoo = 16 (a - 2), with a = 0.405424,

Cae2(X, y) {x + jy I - ~ < y $ -0.1621} .

iii. Hyperbola:

Faea(x, y) ag) = -14.7846, a~;) = -370.3541, a~;) = 381.350,

(3) (3)
a~~ = -371.403, a~ = 0, a~~ = all ; aoo = -7.39229,

Caea(x, y) = {x + jy I 0.1621 < y $ O} .

The dashed lines in Figure 5.22 show the idle continuation of the conic sections beyond
the segments, which are used for the locus approximation. 0

5.2 Mapping of Frequency Loci Margins into

Parameter Space

Representation of the Frequency Locus

We consider the family of uncertain frequency loci
m
. L: a;(q) (jw);
'
C( jW, ) = NG(jw, q) = ;=0 EQ (5.2.1)
q D (. ) n ' q
G jW, q L: bi(q) (jw);
;=0

with m $ n. This covers Nyquist plots of proper open-loop transfer functions and
Popov plots of strictly proper linear part transfer functions, respectively. Moreover,
it is assumed that the coefficients a;(q) , b;(q) are real, and depend polynomially on
the e elements of the parameter vector q . (This includes the simpler cases of interval
coefficients as well as affine and multilinear coefficient functions.) For the sake of
simplicity, the real and imaginary parts of the loci are represented according to

C(jW, q) = Rc(w, q) + jIG(w, q). (5.2.2)

The goal of applying the parameter space method is to display the regions in a plane
of parameters q, where C(jw , q) is 8-stable. In particular, this means determination
of the boundaries of the 8-stability regions. Analogous with r-stability, here for map-
ping into the (ql' Q2)-plane, the parameter space method complies with th e following
perception: if (after starting from a 8-stable point (Ql ,O , Q2,0)) a boundary curve in the
122 5 Frequency Domain Analysis and Design

parameter plane is crossed due to continuous parameter variations, then this means that
the system becomes 8-unstable. Thus , for mapping 8-stability margins into parameter
space, the values of q for which the frequency locus C(jw, q) changes from 8-stable to
8-unstable need to be determined.
As with r-stability in Chapter 2, critical 8 -stability conditions will now be es-
tablished. Therefore, it is assumed that the frequency locus is continuous in q, i.e.
continuous changes in the parameters result in continuous changes of the locus (except
for values of q, for which Dc(jw, q) = 0). This analogy with Frazer and Duncan's
boundary crossing theorem to frequency loci is not proven here since it follows anal-
ogously from the assumed coefficient structure (see (5.2.1)). However, the case where
C(s, q) has poles on the imaginary axis must be specially considered.

Definition of Critical 8-stability Conditions

Critical 8-stability condition here denotes a geometrically interpretable condition under
which a frequency locus C(jw, q) potentially changes from 8-stable to 8-unstable with
variation of q. Depending on both the specific' definition of 8 and the properties of
the considered locus, various critical conditions may come into play. Which of the
critical conditions he or she considers is left to the control engineer. Any previous
knowledge concerning properties of the system (e.g. the order , relative degree, etc.) may
be taken into account and additional critical conditions may be established according
to requirements. The following is a compilation of some critical 8-stability conditions
that are reasonable and practical:

a) The point condition

C(jw , q) runs through a junction or closing point (see below) of the boundary a8.
b) The tangent condition
C(jw, q) tangentially touches a smooth section of the boundary a8. (The dashed
locus in Figure 5.20 complies with this case. The point of contact is indicated by
"+". )
c) Frequency endpoint condition
C(jw, q) starts (i.e. w = 0) or ends (i.e. w -+ 00) on oe.
d) The infinite locus magnitude condition
C(s, q) has poles on the imaginary axis. At the corresponding frequencies, the
frequency locus C(jw, q) tends to infinitely high magnitudes. On the 8-stable
side of this critical condition, the frequency locus completely resides in 8. On the
e
other side, intersections with a8 occur. This case is only relevant if extends to
infinity, like in Figure 5.21.

These geometric conditions are easily translatable into mathematical equations, which
allow the mapping of 8-stability boundaries into parameter space. Before the mapping
is explained in detail for each critical condition, the treatment of compound and closing
boundaries is illustrated.
5.2 Mappin g of Frequency Loci Margins into Parameter Space 123

How to Treat Compound and Closing 8-boundaries

We assume that the boundary ae is given in the form as represented by (5.1.19) . A
reasonable procedure for mapping boundaries of this kind into parameter space takes
into account the following:

• For each boundary section ae j , an appropriate tangent condition is established

and mapped. A circumstance that has to be considered is illustrated by means
of the tangent condition for ae 2 in Figure 5.21: the mapping of the boundary
equation Fas2 = y = 0 also yields solutions that belong to x > -0.5. The function
Fae2{x, y) , however, is defined only for values of x + jy that are contained in the
definition region Cae2. Consequently, the relevance of each solution (q ,w,x,y)
resulting from the tangent condition needs to be checked. If x + jy E Cae2 =
{x + jy I x 5 -0.5}, then the corresponding solution is relevant, otherwise it is
denoted fictitious. For mapping, the fictitious solutions are skipped since they do
not contribute to the stability boundaries in parameter space.
• As with the tangent condition, the relevance of solutions corresponding to a fre-
quency endpoint condition has to be checked, and fictitious solutions must be
skipped when mapping .
• In general , at a junction , the slope of the boundary ae is not continuous. Thus ,
for each junction , a separate point condition needs to be pursued. The mapping
of the point condition Zi.j provides a continuous transition between the tangent
conditions, which belong to the neighboring boundary sections ae j and aej'
• With the dual locus method, e-definitions are reasonable, where e, i.e. the for-
bidden complement of e, is not a region but a non-closed curve (not enclosing any
area), see -1/Ns{A) in Figure 5.10 and -1/Na {w, A) in Figure 5.13 as examples.
In this case, the boundary of ae is identical to e.The boundary curve may close
amidst the complex plane like with the locus of the negative-inverse describing
function of a saturation . For such a closing point , a point condition also needs to
be established.

Mapping Critical 8-stability Conditions

a) Mapping the point condition
The goal of mapping the point condition is to show the set of parameter vectors q
in parameter space for which the frequency locus G(jw, q) contains the point z* =
x* + jy* E C:

Qz· = { q E lR 2 I G(jw, q) = z*, w E lRt } . (5.2.3)

The equation G(jw, q) = z", using (5.2.1), yields

p(jw, q, z") = NG(jw, q) - z* DG(jw , q) = O. (5.2.4)

124 5 Frequency Domain Analysis and Design

Remark 5.11

Note the case that COw, q) represents a SISO Nyquist plot and z' is chosen to be th e
critical point (z' = -1). Then, in the Laplace domain , p(s , q , z' ) is the charact eristic
polynomial of the closed loop. For all q , for which (5.2.4) is true, at least one root
of p(s , q , z') is located on the imaginary axis. The point condition with z' = -1 in
this case is equivalent to the condition that the closed loop has poles on the imaginary
axis. Thus, it may be used as an alternative to the method described in Section 2.1 for
mapping Hurwitz-stability boundaries. 0

The first step to accomplish the mapping is to establish the equations that represent
the point condition. At the frequency w where COw) runs through the point z'
x' + jy' , the real and imaginary parts of COw) and z' have to be identical:

Ra(w, q) = x' ,
(5.2.5)
Ia(w , q) y' .

Usually, the solution of these equations is mapped into a two-dimensional cross-section

of the parameter space , i.e. into a plane of two parameters qi and q2 . The remaining
parameters, therefore, are set constant, or gridded where appropriate. Including the
frequency, there are three unknowns (qI' q2, and w) and two real equations (5.2.5). The
subset Qz. of the parameter space, for which (5.2.3) is true, is represented by curves in
the qI,q2-plane. These curves are parameterized with the frequency w. Solving (5.2.5)
may be realized in the following way: (5.2.5) is transformed into

(5.2.6)

with two polynomials PI and P2. These two equations are of the form of (A.1). There-
fore, the mapping for rendering the graphical display of (5.2.6) can be performed as
generically described in Appendix A.

Example 5.12

For illustration, Example 5.1 is continued here: the point conditions for ZI,2 =
-1/2 and ZI,3 = ei-S/ 41r from Figure 5.21 are mapped. For the open-loop transfer
function (5.1.1), the real and imaginary part of the Nyquist plot can be determined
after replacing s by jw :

16qi - 5 qi (8 + 10q2 + 5 q22) w 2

(5.2.7)
Ra(W,qI ,q2) = 4+5(16+20q2+ 5 q22) w2+25w4 '

1 w _ -10 qi (8 + 5 q2) w + 25 qi q2 w3
(5.2.8)
a( , q1> q2 ) - 4 + 5 (16 + 20q2 + 5q22) w2 + 25w 4
5.2 Mapping of Frequency Loci Margins into Parameter Space 125

For the point condition at Z 1,2 = -1 /2 , i.e. x ' = -1/2, y' = 0, the polynomials PI and
P2 corresponding to (5.2.6) are

PI = -5 (-16-20q2-5q22+2ql (8+10q2+5q22))w2

+4 + 32ql + 25w 4, (5.2.9)

P2 = 5qlw (-16 -1Oq2 + 5q2 W2) .

Multiple solutions exist, which can be found by looking at the various combinations of
factors of these polynomials separately:

• For w = 0, the polynomial P2 is always zero and PI = 4 + 32 ql = 0 yields

ql = -1/8.
• For ql = 0, the polynomial P2 is always zero, but PI = 0 does not yield real
solutions for q2 .
• For the last factor combination, simplification of the equations is obtained through
substituting w 2 by o. The resultant (eliminating ql) is

(5.2.10)

The first factor solution is q2 = 16/ (5 (-2 + a)) . Back substituting and solving
PI for ql then yields ql = (50 - 2)/16. (This is a parameterized representation
of the hyperbola 2 ql q2 - q2 - 2 = 0 with ql ~ -1/8 for a ~ 0.) For plotting
this curve in the (qj, q2)-plane, a runs from zero to infinity. The second factor
solutions for qz are not relevant, since they are not real for a > O.

The boundaries are displayed in Figure 5.23.

4 , .• 4 ..., ... ; •.

2 2

o o -.
s -2 ....
e -2

..... ..... -4 ...,

-8 ..:... -8 " :'

1.5 2.5 3.5 1.5 2 2.5 3.5

ql ql

Figure 5.23. Point condition for ZI .2 = Figure 5.24. Point condition for ZI .3 =
-1/2 e)-5/4rr

The boundaries of the point condition for z' = &.5/4rr are shown in Figure 5.24. 0
126 5 Frequency Domain Analysis and Design

b) Mapp ing the tangent condition

The goal of mapping the tangent condition is to show, in parameter space, the set
of parameter vectors q for which the frequency locus COw, q) tangentially touches a
smooth section of the boundary ae. This situation is illustrated by the dashed locus in
Figure 5.20. The tangent condition is made up of two subconditions. The first is that
the curves COw, q) and ae share the common point of contact, which may be expressed
by substituting x = Rc(w , q) and y = Ic(w , q) from (5.2.2) into the definition of the
respective boundary section (5.1.16):

Faei(Rc(w, q), Ic(w, q)) = 0. (5.2.11)

Second, matching the tangent condition means equality of the slopes of both the bound-
ary ae and the locus COw, q) at the point of contact:
a
awFaei(Rc(w , q),Ic(w, q)) = o. (5.2.12)

With all boundary sections, the tangent condition for mapping into a qll q2-plane can
be transformed into a set of two polynomial equations that match the generic form of
(5.2.6):

(5.2.13)
a
P2(W, qll q2) = s» (w, ql, q2) = O.
Thus , the mapping can be performed as generically described in Appendix A.

Example 5.13

Again, Example 5.1 is used for illustrating the tangent condition: the slanted
straight line ael , as a part of the boundary ae shown in Figure 5.21, may be rep-
resented as
Fae1(x,y) 2x + (V2 - 2) y + 1 = 0,
(5.2.14)
Cae1(x,y) = {x+jYI -1/V2~x~ -1/2} .
With x = Rc(w, qll q2) and y = Ic(w, ql, q2) substituted from (5.2.7) and (5.2.8) in
(5.2.14), the equations (5.2.13) look like

4 + 32ql - 20 (1/V2 - 1) qd8 + 5q2) w

+ 5 (4 (1 - ql)(4 + 5 q2) + 5 (1 - 2 qd q2 2) w2
(5.2.15)
+50 (1/V2 - 1) ql q2w3 + 25w4

0,
5.2 Mapping of Frequency Lo ci Margins into Parameter Space 127

-2 (1/\1'2-1) qd8 + 5q2)

+ (4 (1 - qd (4 + 5q2) + (5 - lOqd q22) w
(5.2 .16)
+ 15ql q2 (1/\1'2 - 1) w2 + lOw3

O.
The resultant obtained by elimination of ql is

resq\ (Pll P2) = 8 (1/\1'2 -1) (8 + 5q2) + (8 (6 + 5q2) (12 + 5q2)) w

-10 (1/\1'2 - 1) (128 + 246q2 + 140 q22 + 25 q2 3 ) w2
+320w3
(5.2.17)
- 25 (1/\1'2 - 1) (48 + 46q2 + 20 q22 + 5q23 ) w4
- 50 (8 + 10 q2 + 5 ql) w5 + 125 (1/\1'2 - 1) q2 w6

O.

With these equations, the mapping into the ql, q2-plane can be done: the frequency w
is gridded. Then, for all real solutions of q2 obtained by (5.2.17) for one value of w, one
corresponding solution for ql can be determined from solving either (5.2.15) or (5.2 .16)
for ql. This approach provides the boundary curves in the qll q2-plane. Note that
real solutions do not necessarily exist for each frequency w . The particular frequencies
where the number of solutions changes may be determined from the discriminant of
(5.2.17) , see Appendix A. In this example , the positive real discriminant frequencies
are Wd E {O, 0.632456, 0.684375, 1.63691}.
The tangent condition for the entire straight line 2 x+ (\1'2 - 2) y+ 1 = 0 is shown in
Figure 5.25 with gray lines. However, not the entire straight line is relevant. Regarding
the additional constraint -1/\1'2 :s x :s -1/2 , the relevant tangent condition is plotted
in bold black line style. The distinction between relevant and fictitious parts of the
boundary can be done in the following way: assume the real triple (w' , qi, qi) is a
solution of (5.2.15) ,(5 .2.16) . If

Rc(w', qi, q;) + jIG(w',qi ,q;) E Cae\(x , y) = {x + jy I- 1/12:s x s -1/2}

(5.2.18)
with Rc from (5.2 .7) and I G from (5.2.8) holds, then (w' ,qi ,qi) contributes to the
mapping of the e-stability boundary, otherwise it is discarded since it is fictitious .
Note that in this specific example the relevance of the solution can be checked by
means of the simpler expression -1/\I'2:s Rc(w',qi ,qi}:S -1/2.
The tangent conditions for the two horizontal straight line segments 08 2 and ae 3
in Figure 5.21 are mapped analogously. 0
128 5 Frequency Dom ain Analysis and Design

,"

~
.

2 ...
. .. ..
0
..,.'"- 2 I
.
.
..
--+-
J'
o - --
---.-

•••• -2..--..........c..
-4

-6

-8 1.5 2.5
-0 .5 0 0.5 1.5 2 2.5 3.5
ql q,

Figure 5.25. Tangent condition Figure 5.26. Boundaries and stability re-
gion Qr n Qe for simultaneous p, and 8-
stability

c) Mapping of frequency endpoint conditions

Frequency endpoint conditions deal with the situation that C(jw, q) starts [i.e. w = 0)
or ends (i.e. w - t 00) on 8e. Let w· denote the considered endpoint frequency. The
mapping equations are established in a straightforward manner by postulating both
real part and imaginary part equality of the 8-stability boundary 88 and the frequency
locus C(jw, q) at the respective frequency w' :

Fa8i(Ra(w' , q), Ia(w', q)) = o. (5.2.19)

Taking the numerator of Fee, results in one single polynomial equation:

(5.2.20)

For display in a (qil q2)-plane, even explicit solutions ql = II (q2) or qz = h( ql) may
exist. Otherwise , gridding of one of the parameters may yield numeric solutions . As
with the tangent condition , the relevance of the solutions has to be checked and fictitious
solutions must be skipped during mapping.

Example 5.14

For Example 5.1, the only endpoint frequency is w = O. It applies along the bound-
ary section 88 2. For all ql :::; -1/8, the locus of C(jW,ql,q2) starts on 88 2 . At the
limit of Oa82 (i.e. at x = - 1/ 2), the locus of C(jw, -1/8, q2) starts at the junction
Zl,2 = -1/2. The solution ql = -1/8 has already been determined as a solution of the
corresponding point condition . 0

d) Mapping the infinite locus magnitude condition

If 8 extends to infinity in any direction of the complex plane, then this condition needs
to be considered . However, it is recommended that this condition is checked in any
5.2 Mapping of Frequency Loci Margins into Parameter Space 129

case for verification of the mapping results of the previous conditions . For instance, in
(5.2.11) or in (5.2.19) the expression % may occur, which is not taken into account
when only considering the respective numerator. The parameter values for which both
the corresponding crit ical 8-stability condition and the infinite locus condition apply
may subsequently be treated. The mathematical equations may be derived from the
Hurwitz-stability boundary for G(s , q).

Example 5.15

For Example 5.1, q2 = -2 is the boundary for the infinite locus magnitude condition,
which can be directly seen from the denominator of G(s , q) = L(8, q) in (5.1.1). 0

Multiple Criteria Mapping

An advantage of the par ameter space method is that parameter values for which multiple
criteria are satisfied may be determined from forming the intersection of the respective
stability regions. For Example 5.1, the f-stability boundaries (Figure 5.3) plus both
point conditions (Figure 5.23, Figure 5.24) and the tangent condition (Figure 5.25) are
displayed together in Figure 5.26. Here, all fictitious boundary parts were skipped . The
dashed line represents th e infinite locus magnitude condition. For all parameter values
q in the shaded stability region Qr n Qe, both f-stability (as defined in Figure 5.2)
and 8-stability (as defined in Figure 5.21) are satisfied.

Absolute Stability 8-margins for the Crane Example

Example 5.16

The Popov criterion is used to check robust absolute stability of the crane with
output feedback: k, = 500, k 2 = 2865, k 3 = -22800, k4 = 0, 9 = 10, me = 1000 (see
Example 3.8 with (3.2.4)) . The open-loop transfer function is
2+573£83
G(8,mL, £) = 1000+57308+ (4560+ 100£)8 •
(5.2.21)
2 82 (1000 + mL + 100£ 82 )

The absolute value of th e input to the crane (i.e. the propelling force of the crab)
is now considered physically limited to Iyl :::; 2000 [N] . A corresponding saturation
characteristic curve is shown in Figure 5.27. This non-linearity can be considered a
sector limited characteristic as introduced with Figure 5.15. The sector assigned to
the characteristic is [p, k + p] with p = 0.07 and k = 1. The lower sector needs to
be greater than zero since zero controller gain does not stabilize the undamped poles
of the crane . The fact that the saturation characteristic leaves the assigned sector for
lui = I[kI, k2 , k3 , k4 1.xl > 28571 can be neglected since no realistic state of the crane
130 5 Frequency Domain Analysis and Design

3000

2000
/
'000 ~.. .. .. ... ...........
<:
./
0
;;;-1000 ..... ...... .....
- ....... ..
-2000

-3QO()
: /
-1 -0.8 -08 -0 0 -02 0 0 .2 0 .0 0.8 0.8 1
U IN] .'0'

Figure 5.27. Saturation of the crane input (i.e. crab acceleration)

produces any bigger absolute value of the controller output than 28571. The system is
transformed to standard form as described in the context of Figure 5.16. The open-loop
linear part of the transformed system is
1000 + 5730 s + (4560 + 100 f) S2 + 573 f. S3
G(s,mL,f.) =
1000p + 5730 p s + a2 S2 + 573 p f. S3 + 200 f. S4 '
(5.2.22)
with a2 = 2000 + 4560 p + 100 pf. + 2 mt. .
The 8-stability region to be defined before mapping it into the parameter space is

30 5

25

_ 20 .,
.§. I .
- 15 I ; ....
I .
I
I
I
5 , . .. i , .

'--'
1000 1500 2000 2SOO 3000 - 10 o 10 20
"\ [kg) Re

Figure 5.28. 8 -stability region of the Figure 5.29. The Popov plot (dashed line)
crane for two Popov-line slopes (gray: 1.8, of the transformed system for tnt: = 1500,
black: 0.8) e= 12 liesentirely to the right hand side of
the two Popov-lines with slopes 0.8 (gray)
and 1.8 (black)

the region to the right hand side of the Popov-line , which intersects the real axis at
-11k = -1. The slope of the Popov-line may be chosen arbitrarily, however, it will
affect the result in terms of the obtained 8-stability region. Figure 5.28 shows the 8-
stability boundaries for two different slopes of the Popov-line. The center (mL = 1500
and f. = 12) of the operating domain Q turns out to be 8-stable in both cases. The
corresponding Popov plot resides entirely to the right hand side of both Popov-lines ,
see Figure 5.29. However, a little bit of the operating domain is cut off at different
5.3 Frequency Response Magnitude Specifications (B-stability) 131

,
,,
5

o 500 1000 1500 2000 2500 3000

mL [kg]

Figure 5.30. Stability region of the crane for simultaneous r- and 8-(absolute) stability
(compare to Fig. 3.16)

vertices by the stability boundaries. For the specific example, the results shown in
Figure 5.28 and Figure 5.29 suggest trying of an intermediate value for the Popov-line
slope. With a slope of 0.95, the e-stability-region shown in Figure 5.30 is obtained.
Also, the r-stability region from Fig. 3.16 is included . The entire operating domain is
contained in the intersection of the r- and the e-stability regions. 0

In the above example, it was possible to choose a common Popov-line in order

to prove robust absolute stability for the entire operating domain Q. Nevertheless,
Figure 5.28 also proves absolute stability by means of two Popov-lines with different
slopes. The system is absolutely stable if the Popov plot resides to the right hand side
of a Popov-line with arbitrary slope. Thus, the overall e-stability region is formed by
the union of all 8-stability regions that are obtainable at different Popov-line slopes. In
practice, the procedure for computing e-stability boundaries can be repeated with some
exemplary Popov-line slopes. The result will provide at least a good approximation of
the attainable absolute stability region in parameter space.

5.3 Frequency Response Magnitude Specifications

(Beta-stability)

The parameter space approach has been proven to be a very useful control design tool
in a number of applications involving parametric linear systems. This is due in part to
132 5 Frequency Domain A nalysis and Design

the non-conservative treatment of uncertain param eters, the simplicity of t he resulting

controller, and the transparency provided by the design method.
On the other hand , Hoc-based techniques do facilitate the design and analysis of
controllers for syst ems with both parametric and unstructured uncert ainty subject to
Frequency Response Magnitude (FRM) specifications. However , even for very simple
syst ems, these methods often result in complex controllers that provide conservat ive
results, which is due in par t to th e conservative treatment of parametric uncertainty.
In this section , typi cal frequency domain specifications used to ensure a number
of closed-loop characteristics (e.g. disturbance attenuation, noise rejection , reference
tracking, model matching, robust stability w.r.t . structured (par ametric) and unstruc-
tured uncertainty) are reviewed to motivate the use of frequency response magnitude
(FRM) specifications. Then, the notion of B-stability is formally defined to enable
the treatment of FRM specifications in the parameter space framework. B-stability is
analogous to f-stability and defines acceptable and unaccept able frequency response
magnitudes just as f-stability defined acceptable and unaccept able closed-loop pole lo-
cations. Using the concept .of B-stability, the parameter space approach can be used
to design controllers that provide desired closed-loop characteristi cs stated in t erms of
FRM specifications. Some elementary examples are provided in Section 5.3 and Sec-
tion 5.4 to demonstrate th e concept of using FRM specifications in parameter space.
In Sections 6.7, 6.8, and 7.3, case studies in vehicle dynamics and flight cont rol are
presented that demonstrat e t hese concepts in "real-world" cont rol design problems.

Remark 5.17

In t his sect ion, only SISO FRM specificat ions are considered . While any number
of these SISO specifications can be considered simultaneously, in some cases it is more
desirable to consider MIMO FRM specifications, as presented in Sect ion 5.5. 0

Remark 5.18

SISO FRM specifications are often given in terms of the oo-norm of transfer func-
tions, where the co-norm of th e transfer function G(s) is given by

IIG(s)lI oc = sup IG(jw)l · (5.3.1)

WEIR;;

The co-norm of G(s) is the peak value of its FRM plot (i.e. Bode magnitude plot) . It
is also the distance in the complex plane from the origin to the farthest point on t he
Nyquist plot of G(jw).
o

Before present ing the general discussion on frequency domain specifications , Exam-
ple 5.1 will be revisited to introduce t he concept of FRM specifications.
5.3 Frequency Response Magnitude Specifications (B-stability) 133

Example 5.1 9

Consider the parametric single-loop unity feedback system in Figure 5.1 with para-
metric open-loop transfer function £(s, q) defined by (5.1.1). Here, both the reference
input wand the disturbance d are considered. Thus, in addition to reference tracking
performance , disturbance attenuation qualities of the closed-loop system can also be
addressed. Therefore , we consider the two closed-loop transfer functions

y(s) ql(-5q2S+8)
T(s, q) = (5.3.2)
w(s) 5s 2 + (10 + 5q2 - 5qlq2)S + 2 + 8ql'
y(s) 5s 2 + (5q2 + lO)s + 2
si« q) (5.3.3)
d(s) = 5 S2 + (10 + 5q2 - 5qlq2)S + 2 + 8ql '

i.e. the tracking transfer function T and the disturbance transfer function 5 , also de-
noted complementary sensitivity function and sensitivity function , respectively. A de-
tailed description of the sensitivity functions will be given later in this section.
As in Example 5.1, the effect of variations of q2 is investigated while ql = 1 is
assumed. For this value of ql, the characteristic polynomial is robustly stable for all
values of q2, in fact it does not depend on q2. However, variations in q2 affect the zeros
of the closed-loop system and thus also the tracking and disturbance transfer functions
T and 5 . The Bode magnitude (i.e. FFtM) plots of 5 and T are shown in Figure 5.31
for values q2 = 1 and qz = 5 respectively.
For interpretation of the Bode magnitude plots in Figure 5.31, it is useful to consider
three frequency ranges:

i. Low-frequency range: the low-frequency range (w -+ 0) determines the steady

state behavior of the closed-loop system. Here, the steady state values of 5 and
T are given by

.-0
limT(s,q)= 8ql
2 + 8ql
and
.-0
lim 5(s, q) = 2 2
+. 8 ql
'

and do not depend on q2 and thus , the steady state behavior does not depend on
q2' However, this shows that the steady state error (limt_oo e(t)) will not be zero,
since T(O) # 1 and 5(0) # O.
ii, High-frequency range: the high-frequency range determines the immediate re-
sponsiveness of the closed-loop system, which may be characterized by the direct
transmission values

lim T(s, q) = 0 and lim 5(s, q) = 1 .

8-00 8-00

These values differ from 0 and 1, respectively, only for systems where the open-
loop transfer function L is of relative degree O.
iii. Intermediate frequency range: the intermediate frequency range determines the
dynamic behavior of the system. Although 5 and T are stable transfer functions,
their magnitude plots vary largely with q2 within this frequency range.
134 5 Frequency Domain Analysis and Design

q2 = 1 q2 =5

3 3

12 2
Ci)
1 1 ....

0 0
10. 1 10° 10
1
10- 1 10°' 10'

20 20

10 10
2l
2- 0 ······ ·· ··· ·· · . .. . .. ... .. .. ... .. .. .. ... . o .......... ...
3'
0 - 10 - 10
h
- 20 Br -20 Br
- 30 -30
10- 1 10° 10
1
10-' 10° 10'
w [rad/s] w [rad / s]

Figure 5.31. Bode magnitude plot of sensitivity function S (top) and complementary sensi-
tivity function T (bottom) for ql = 1, q2 = 1 (left) and qi = 1, qz = 5 (right). The desired
regions Bs and l3r are shaded in gray

In Example 5.1, f- and e-specifications were considered. Here, two new specifi-
cations (B-specifications) with regard to the FRM plots of S and T are introduced.
Namely, the FRM plots of S and T shall reside with in the desired regions Bs and BT
shown in Figure 5.31. Both specifications are satisfied for q2 = 1. However, for q2 = 5
the sensitivity specification is violated (i.e. the FRM of S does not reside in the desired
Bs region). Then, if the FRM specifications on Sand T are to be included in a param-
et er space robustness analysis, the task is to determine the parameter values of ql and
q2 for which both FRM specifications are satisfied. Th is task will be addressed fully in
Section 5.4.
o

Basics of Frequency Domain Feedback Control

Before showing how frequency response magnitude specifications can be incorporated
into the parameter space approach, some basics of frequency domain feedback control
are reviewed. For a more detailed descript ion of feedback control theory with frequency
response magnitude specifications , the reader is referred to [79] and [203],[137],[171],[38].
5.3 Frequency Response Magnitude Specifications (B-stabiUty) 135

Remark 5.20

Since we are working in the frequency domain, all variables (both signals and transfer
functions) are functions of the Laplace variable s. For the sake of simplicity, in the
sequel, s is being suppressed in most instances.
o
The following discussion considers the standard single-loop feedback system in Fig-
ure 5.32, which serves as a generic structure for many control systems. However, all
of the following can be easily extended to arbitrary feedback systems. The open-loop

Ym n

y controlled signal d1 (plant) input external signal

Ym measured controlled signal d2 (plant) output disturbance

u actuating signal y' plant output

e tracking error (e = w - y) u' controller output

em measured tracking error (em = w - Ym) u control signal

r reference input n sensor noise

w filtered reference input

Figure 5.32. Standard single-loop feedback system

transfer function for the standard loop is defined as

L=KG, (5.3.4)
and the characteristic polynomial of the closed-loop is

p = num (1 + KG) . (5.3.5)

The transfer function from the reference input r to the controlled signal Y, i.e.
H= FKG (5.3.6)
I+KG
denotes the tracking transfer function, often referred to as the model regulator transfer
function.
The two basic design steps are:
136 5 Frequency Domain Analysis and Design

i. Feedback step : design the controller K such that the input and output distur-
bances d 1 and d2 are attenuated, the sensor noise n is rejected , and the filtered
reference signal w is tracked.
ii. Feedforward step: design the prefilter F such that the tracking performance with
regard to the reference signal r matches a desired transfer function Gd ("model
matching", see also Figure 5.33).

The first design step (i.e. the feedback step) is the primary focus in the sequel.
Therefore, except when otherwise stated, F = 1 (w = r) is assumed. The input / output
relation between all external inputs w, db d2 , and n and all signals. of interest, i.e,
tracking error e, control signal u and controlled signal y is given by

w
e S T -Sa -S
n
u SK -SK S -SK (5.3.7)
d1
y T -T Sa S
'-v-'
Z
.
R
d2
'---v--'
w
where w is the vector of external input signals, and z is the vector of internal signals to
be used to evaluate internal stability and performance of the dosed-loop system. The
elements of the transfer matrix R are the four sensitivity functions S , T , Sa , and SK
defined as:

• .Sensitivity junction:
(5.3.8)

S describes the sensitivity of the dosed-loop system w.r.t . output disturbances

and changes in the filtered reference signal.
• Complementary sensitivity junction:

KG
T= = KG S = 1 - S. (5.3.9)
l+KG
T describes the influence of measurement noise on the controlled signal and the
tracking error, as well as the response of the controlled signal to changes in the
filtered reference signal.
• Input sensitivity junction:

K
SK = 1 + KG = K S. (5.3.10)

SK describes the influence of the filtered reference signal, measurement noise, and
the output disturbance on the control signal.
5.3 Frequency Response Magnitude Sp ecifications (B-stability) 137

• Output sensitivity function :

G
Sc = 1 + KG = G S. (5.3.11)

Se describes the sensitivity of the closed-loop system w.r.t. external input signals
d 1 , i.e. the effect of d 1 on y and e.
These four sensitivity functions (i.e. S, T, Se, SK) along with the characteristic
polynomial p and the model regulator transfer function H can be used to thoroughly
evaluate the stability and performance of the nominal closed-loop system in Figure 5.32
as well as that of the uncertain closed-loop system in Figure 5.39. In the discussion
to follow, investigating the stability and performance of closed-loop systems is divided
into the following four categories :
• Nominal stability: stability with no model uncertainty.
• Nominal performance: performance with no model uncertainty.
• Robust stability: stability with model uncertainty.
• Robust performance: stability and performance with model uncertainty.

Nominal Stability
First, the internal stability of the nominal closed-loop system must be verified. The
standard feedback loop in Figure 5.32 is internally stable, if all internal signals fade to
zero after the external signals w, d}, d2 and n have vanished. A necessary and sufficient
condition for internal stability of the standard feedback loop is [79]:

Theorem 5.21 (internal stability)

The standard feedback loop is internally stable, if and only if:
• All roots of p = num (1 + K G) are located in the left eigenvalue half plane
Re{s) <0.
• There is no pole-zero cancellation in Re (s) ;::: 0 while forming the product
KG.
o

Remark 5.22

The notion of internal stability can be extended to r-stability, which was defined in
Chapter 3. The standard feedback loop is internally r -stable, if and only if:
• All roots of p = num (1 + K G) are located in the desired r-region.
• There is no pole-zero cancellation outside r in the product KG.
o
138 5 Frequency Domain Analysis and Design

Nominal Performance
Next, the performance of the nominal closed-loop in Figure 5.32 will be investigated.
A number of closed-loop performance characteristics (e.g. tracking performance , dis-
turbance attenuation, noise rejection, required control input and model match ing) can
be evaluated by examining the magnitudes of the individual frequency responses (i.e.
FRMs) of the transfer matrix R (i.e. the four sensitivity functions) as described below.
Insight regarding the desired magnitudes of these sensitivity functions can be obtained
by observing the effect that the external inputs w have on the internal signals z : Of
particular interest is the effect of the external inputs on the tracking error,

e w-y (5.3.12)
S w + Tn - Sc dj - S d2 , (5.3.13)

as this provides information regarding the magnitudes required for three of the sensi-
tivity functions in order to achieve good performance. Since the FRMs of these transfer
function are frequency-dependent , further insight can be gained by considering the fre-
quency range of the external input signals. Typically, reference and disturbance inputs
are low-frequency signals , while sensor noise generally occurs at high frequency.

i. Tracking: in order to have good tracking perform ance in the frequency rang e
[0 ; wsl, changes in the filtered reference signal w can only have a small effect on
the tracking error e (5.3.12), which requires

IS{jw) 1«1 for O:S w :S Ws . (5.3.14)

ii. Disturbance attenuation: if inequality (5.3.14) is sat isfied, then output distur-
bances d2 are also attenuated (i.e. don 't have a significant effect on the tracking
error) in the frequency range [0; ws] . Input disturbances d, are at tenuated in the
frequency range [0; ws] if

ISc(jw) I «1 for O:S w :S Ws . (5.3.15)

iii. Noise rejection : the effect of noise on the tracking error can be rejected to a great
extent in the frequency range [WT; 00) if

IT(jw)1 «1 for WT:S w < 00 . (5.3.16)

iv. Required control input: since all actuators have performance limits , it is often
necessary to ensure that these limits are not exceeded. The control signal (i.e.
actuator command) is limited within the actuator bandwidth [0;Wactl if

ISK(jW)/ is small for O:S W < Wact . (5.3.17)

v. Model matching: the model matching performance also depends on the prefilter
F used to generate the filtered reference signal w (i.e. w = F r). If Cd is the
desired transfer function that the closed-loop syst em in Figure 5.33 is to track,
then the model matching error is
5.3 Frequency Response Magnitude Specifications (B-stability) 139

Figure 5.33. Model matching error

eu = (Cd - H)r = Er , (5.3.18)

where
(5.3.19)

The desired matching performance is achieved in the frequency range [0; w M] if

IE(jw)1 is small for wE [O jWM] ' (5.3.20)

Remark 5.23

In order for the model matching error eM defined by (5.3.18) to be small, IE(jw)1
must be small, which means that Cd must be a feasible closed-loop transfer function .
That is, by a proper choice of K and F it must be possible that the model regulator
transfer function H defined by (5.3.6) approximately matches Cd, see Section 1.7.
o

As briefly discussed above, a number of performance specifications can be stated

in terms of keeping the FRMs of various transfer functions small in certain frequency
ranges. While all of the specifications outlined may be of interest , for standard control
problems we are especially interested in the magnitude specifications on Sand T, and
specifically in ensuring that IS(jw)1 « 1 for small frequencies and IT(jw)1 « 1 for high
frequencies. More specific details regarding the desired magnitude or T will be given
later in th is section, but for now we will focus on S. Magnitude specifications on S , often
referred to as nominal performance criteria in SISO H oo control theory, can be used to
formulate three design goals concerning the steady state error , the stability margin, and
the transition frequency. Specifically, the maximum steady state error can be limited to
S(O) < 19 , the high-frequency disturbance amplification can be limited to IS(jw)! < hg ,
and the transition frequency Wt for which IS(jwt)I = 1 (i.e. the frequency at which
disturbances become amplified instead of attenuated) can be limited to Wt ~ Ws if

IS(jw)1 < 8Bs(w) , (5.3.21)

140 5 Frequency Domain Analysis and Design

where the magnitude specification, given by the frequency-dependent bound 8B s , can

either be represented piecewisely by straight line segments, e.g.

19 Vw E [O;wsJ
8B s (w) = (5.3.22)
{ kg Vw E (ws; 00)

or by the magnitude of a transfer function , e.g.

s+wsl
8B s (w) = IBOw)l, where B(s) = kg ; (5.3.23)
s+ws g
These typical magnitude bounds on S are demonstrated in the following example.

Remark 5.24

The disturbance amplification specification ISOw)1 < kg also determines a lower

bound for the Nyquist-stability margin Ps, which is defined as the minimum spacing
between the Nyquist locus and critical point (-1 ,0), i.e, by the relation

Ps inf
wE[O,co)
11 + LOw)1
(5.3.24)
= IIS(jw)lI~l .

Thus, if ISOw)1 < kg , the Nyquist-stability margin must satisfy

Ps> l/k g • (5.3.25)

Example 5.25 (nominal performance)

Let the open-loop transfer function be given by

w2
L(8) = k
S2
° 8 + Wo2
+ 2 D Wo ' (5.3.26)

where k = 1, D = 0.3, Wo = 61T. The two diagrams in Figure 5.34 show the Bode
magnitude plots of the corresponding sensitivity function S, together with the magni-
tude bounds 8B s (w) corresponding to (5.3.23) and (5.3.22) . The permissible steady
state error is 19 = 0.6 and the permissible high frequency disturbance amplificat ion is
kg = 1.9. The required minimum transition frequency is Ws = 21T. The magnitude plot
of S remains below 8B s (w) in both cases, i.e. the specifications are satisfied . 0
5.3 Frequency Response Magnitude Specifications (B-stability) 141

~ 0 .g 0
::l .;:;
.-= '2
~
- -5 ISOw)1
eo
~-5 ISOw)1
~

10' 10'
w [rad/s] w [rad/s]

Figure 5.34. Graphic interpretation of the requirements on the sensitivity function in the
Bode magnitude diagram

Modelling Uncertainty
Before addressing the issues of robust stability and robust performance, a brief dis-
cussion of uncertainty modelling will be given. Model uncertainty can be represented
in two distinctly different forms, parametric uncertainty and unstructured uncertainty.
Parametric uncertainty typically arises from a physical model that has uncertain or
changing parameters. On the other hand , unstructured uncertainty is typically used
to account for neglected or unmodelled dynamics (e.g. flexible body modes, actua-
tor dynamics, sensor dynamics) , and both magnitude and phase are considered to be
uncertain.
Parametric Uncertainty
In the case of parametric uncertainty, the plant is represented by a parametric trans-
fer function G(s, q) (5.2.1). When the family of transfer functions G(s, q) (5.2.1) is
substituted for G(s) in the standard feedback loop in Figure 5.32, the corresponding
sensitivity functions also become a family of transfer functions that depend on the plant
parameters q. Furthermore, if the controller K is parameterized, then the sensitivity
functions also depend on the controller parameters k .
Unstructured Uncertainty
There are a number of models used in robust control design to account for unstruc-
tured uncertainty, the most common of which is the multiplicative perturbation modeL
which can be used to capture a wide variety of model uncertainty including unmod-
elled dynamics as well as parameter variations. The multiplicative perturbation model
corresponding to Figure 5.35 is described by
O(s) = (1 + Wt,(s)b.(s))G(s) . (5.3.27)
b.(s) is the multiplicative perturbation and is assumed to satisfy 1Ib.(s)lIoo ::; 1. Th e
phase Lb.(jw) is considered arbitrary. Th e weighting function IWt,(jw) I scales b.(s) to
142 5 Frequency Domain Analysis and Design

I
I I
I ~

Figure 5.35. Plant with multiplicative perturbation

account for the frequency-dependent magnitude of the uncertainty, where it is assumed

that the model uncertainty is less than ±IW~(jw)l.
A typical weighting function magnitude plot is shown in Figure 5.36.

hg

'3
-;
0

19
WT
W [rad/s]

Figure 5.36. Typical weighting function magnitude plot for a multiplicative perturbation
model

As shown in Figure 5.36, IW~(jw)1 is typically small (i.e. IW~(jw)1 = 19) at low
frequencies where G(s) accurately represents the system , and large (i.e. IW~(jw)1 =
hg) at high frequencies w » WT where the influence of unmodelled dynamics may
be significant. The frequency WT indicates the transition frequency where the model
G(s) becomes unreliable . Typical weighting functions with which the above-mentioned
requirements may be realized, i.e. liIIlw_o IW~(jw)1 = 19 and limw _ oo IW~(jw)1 = hg ,
are:
(5.3.28)

or
Vw E [O;WTJ
(5.3.29)
Vw E (WT ; (0)
5.3 Frequency Response Magnitude Specifications (B-stability) 143

Remark 5.26
The transition frequency wT may be considered closely related to the desired band-
width WB of the closed-loop system. The latter is defined as the frequency for which
the magnitude of T(8)/T(0) falls below the -3 dB-line, i.e. for which
ITOWB)/T(O)\ = 1//2 . (5.3.30)
o

Im(G(jw))

Re(G(jw))

W A (jw· )G(jw·)

Figure 5.37. Frequency locus of the multiplicative perturbation model

Figure 5.37 shows the frequency locus of the multiplicative perturbation model.
For a fixed frequency co" , a disk centered at GOw*) with frequency weighted radius
IW~(jw·)GOw*)1 is plotted. The multiplicative model C(8) represents the family of
transfer functions for which the locus COw) remains within the disk for the respective
frequency w. The phase of COw*) is uncertain within the range rp- < <p < rp+.

Remark 5.27
In addition to the multiplicative perturbation model, three other uncertainty model s
are commonly used in frequency domain feedback control theory to incorporate struc-
tured and unstructured uncertainty information into robust controller design, namely :
• The additive uncertainty model (left plot in Figure 5.38),
C= G + 6. W~ . (5.3.31)
• The two feedback perturbation models (middle and right plot in Figure 5.38),
- G
G = 1 + 6. W~' (5.3.32)
-=
G
G (5.3.33)
l+6.W~G
Examples for these unstructured uncertainty models can be found e.g. in [38] .
o
144 5 Frequency Domain Analysis and Design

~ y
~
~u ~
Figure 5.38. Common uncertainty models

Robust Stability
Theorem 5.21 describes a necessary and sufficient condition for internal stability of th e
control loop in Figure 5.32, assuming a nominal plant model G. However, if the nominal
model is replaced with a parametric model G(q) , unstructured uncertainty model G, or
a parametric model with unstructured uncertainty G(q) , appropriate robust stability
conditions must be considered.
Robust Stability for Parametric Uncertainty
Robust stability in the presence of parametric uncertainty is discussed in detail in
Chapter 2.
Robust Stability for Unstructured Uncertainty
We will now focus on the effect of unstructured uncertaint y on th e robu st stability
condition. Figure 5.39 depicts the standard feedback loop with a multiplicative pertur-
bation corresponding to (5.3.27). Ext ern al input signals are not considered. Assuming
th e non-perturbed closed-loop syst em (i.e. ll(s) = 0) is int ernally stable, a necessary
and sufficient condit ion for robu st stability of th e perturbed syste m is [79J:

Theorem 5.28 (robust stability)

The uncertain closed loop in Figure 5.39 is robu stly stable for all II such that
1I11(s)lIoo S; 1, if and only if

IIW~( s) T(s) 11 00 <: I. (5.3.34)

Figure 5.39. Application of the small-gain theorem to the sta ndard feedback loop with a
multiplicative perturbation
5.3 Frequency Response Magnitude Sp ecifications (B-stability) 145

For a simple interpretation of condition (5.3.34), consider the closed-loop system

shown on the left side of Figure 5.39, and note that the transfer function from d to v is
given by
v
d = -WL\T . (5.3.35)
Then, it is clear that the system in 2-block form on the right side of Figure 5.39 is
equivalent to the system On the left w.r.t. the signals d and v. The small-gain theorem
[203] can be easily applied to this 2-block structure to obtain (5.3.34) in Theorem 5.28.
Table 5.1 lists the robust stability conditions for the four uncertainty models corre-
sponding to Figure 5.35 and Figure 5.38 [79].

Table 5.1. Robust stability conditions for unstructured uncertainty models

a) multiplicative perturbation IIWL\(s)T(s)lIoo < 1

b) additive perturbation IIWL\(s)SK(s)lIoo < 1

c) feedback perturbation IIWL\(s)S(s)lIoo < 1

d) feedback perturbation IIWL\(s)Sc(s)lIoo < 1

Example 5.29 (robust stability)

Th e control loop in Example 5.25 will be revisited to investigate its robust stability
subject to a multiplicative perturbation model (i.e. unstructured uncertainty). The
open-loop transfer function is given by (5.3.26). For frequencies w « 1211' [rad/s] , the
(unstructured) model error is assumed to be smaller than 10% (i.e. IWL\Ow)1 = 0.1),
and for frequencies w » 1211' [rad/s] it is assumed that the model error does not exceed
200 % (i.e. IWL\(jw)1 = 2). The uncertainty is assumed to transition continuously from
its low- to high-frequency magnitude with a transition frequency of 1211' [rad/s] . A
weighting function that satisfies these specifications is
s + 211' ·6·0.1
( )
WL\s=2 . (5.3.36)
s + 211' ·6 ·2
Figure 5.40 shows the graphical interpretation of th e robust stability criterion in a
Bode- and Nyquist-diagram. The magnitude of the complementary sensitivity function
resides below the magnitude plot of the invert ed weighting function , i.e. IT(jw)1 <
IWL\(jw)I-1'Vw, as shown in the left plot of Figure 5.40. Hence, (5.3.34) is satisfied and
the closed-loop system is robustly stable w.r.t . the assumed multiplicative uncertainty
model. The right plot in Figure 5.40 shows an interpretation of the robust stability
condition IWL\Ow)T(jw)1 < 1 in the Nyquist-diagram. From (5.3.9) and (5.3.4), it
follows that
IWL\(jw) L(jw) I < 11 + LOw)l, 'Vw E [0; 00) . (5.3.37)
146 5 Frequency Domain Analysis and Design

Bode-diagram 1 yquist-diagra m
20.- __ 0.5 .....----------~-__.,
o
::::::-0.5
e 3
-0 ~ -1
IT Ow)!
'a-
;:l
'-"
20 .§ - 1.5
~
~ -2
-40
-2.5 ........._ _ ~ __ ~ ---.J

10
1
-2 -1 0 2
w [rad/s] Re (L(jw))

Figure 5.40. "Robust stability" in Bode- and Nyquist-diagram

This inequality means that the closed-loop system is robustly stable if at all frequencies
w the point (-1,0) lies outside the disk centered at L(jw) with radius IW~(jw) L(jw)l .
o

Robust Stability for Mixed Parametric and Unstructured Uncertainty

For a system with mixed parametric and multiplicative uncertainty (i.e. the closed-
loop system represented in Figure 5.39, with G replaced by G(q)), the robust stability
criterion is formulated by merging the robust Hurwitz-stability condition and the robust
stability condition (5.3.34) to obtain:

Theorem 5.30

The uncertain closed-loop in Figure 5.39, where G = G(s, q), is robustly stable
for all q E Q and all Ll such that IILl(s)lIoo ~ 1, if and only if

IIW~(s)T(s,q)lIoo < 1, T/q E Q . (5.3.38)

Robust Performance
In general , robust performance means that some performance specifications (e.g. eigen-
value specifications (Chapter 3), frequency loci specifications (Section 5.1), frequency
magnitude specifications (this section)) are satisfied robustly w.r.t. parametric and/or
unstructured uncertainty. Therefore, there are a large number of conceivable criteria
for robust performance.
5.3 Frequency Response Magnitude Specifications (B-stability) 147

In this section, we will consider the robust performance criterion expressed in terms
of FRM specifications , and commonly used in Hoc control theory [79] . Consider Fig-
ure 5.41 where the plant G is perturbed by a multiplicative perturbation according to
(5.3.27). In addition to ensuring robust stability w.r.t. the multiplicative uncertainty,

e y Yw
_ K G %

Figure 5.41. Weighted controller structure

we would like to achieve good disturbance attenuation properties for the perturbed
system , and therefore introduce a weighting function W s , whose inverse is an upper
bound on the perturbed sensitivity function S, i.e.
(5.3.39)

or equivalently,
IIWs(jw)S(jw)lIoc < 1 , (5.3.40)
where
- 1
S(s) = 1 + K(s) G(s)(l + WL\(s)6(s)) .
From (5.3.39) or (5.3.40), the following robust performance condition may be derived
[79] :
IWL\(jw)T(jw)1 + IWs(jw)S(jw)1 < 1, Vw E [0;00) . (5.3.41)
If G in Figure 5.41 also depends on uncertain plant parameters q E Q, then the
robust performance specification is extended to

IWL\(jw)T(jw, q)1 + IWs(jw)S(jw, q)1 < 1, Vw E [0; 00) , Vq E Q . (5.3.42)

Limiting Factors of Attainable Performance

In the preceding discussion, stability and performance specifications in terms of the
FRMs of various transfer functions were reviewed. However, some of the desired stabil-
ity and performance specifications are conflicting and require compromises. To provide
some insight regarding the compromises that must be made , some of the limiting factors
of the attainable closed-loop performance are reviewed here. These limitations are due
to the relat ion of the sensitivity funct ions and the pole-zero locations of the open-loop
transfer function as well as its relative degree.
An algebraic constraint for the performance of the closed-loop system is the relat ion
between the sensitivity and the complementary sensitivity function:

S+T=l , (5.3.43)
148 5 Frequency Domain Analysis and Design

which requires
IS(jw)1 + IT(jw)1 ~ 1. (5.3.44)
Therefore, it is not possible to have arbitrarily good tracking and noise rejection , (i.e.
cannot have both IS(jw)1 « 1 and IT(jw) I « 1) at a given frequency.
Furthermore, increasing demands on the magnitude of S (i.e. making IS(jw)1 « 1)
at low frequencies results in magnitude amplification at high frequencies. This effect
is termed the uiaterbed effect and implies a reduction of the Nyquist-stability margin
Ps corresponding to (5.3.24). For systems with relative degree of at least two, the
waterbed effect can be quantified by Bode's integral theorem, often referred to as t he
area formula, which is [90],[55]

J
00

log IS(jw)1 dw = 1rloge L ; Re(s;J . (5.3.45)

o
In (5.3.45), {Sj} is the set of unstable poles of the open-loop transfer function . The
integral of the logarithmic magnitude of the sensitivity function is a constant and does
not depend on the controller parameters. If all poles of L are stable, then the right
hand side of (5.3.45) equals zero [55] . In this case, the area below log IS(jw)1 = 0,
termed the negative area, equals the positive area, which is above log IS(jw)! = 0
as shown in Figure 5.42. This clearly shows that disturbance att.enuation in some

o l---~~========~

o
w [radj s]

Figure 5.42. Graphic interpretation of Bode's integral theorem

frequency range must be accompanied by disturbance amplification in another frequency

range . If L has unstable poles, the right hand side of (5.3.45) equals a positive number
that depends on the number and magnitude of the unstable poles. Then, the ratio of
disturbance attenuation and amplification (i.e. the ratio of negative and positive area)
decreases, which means that either less disturbance attenuation can be achieved or
greater disturbance amplification must occur at some frequencies.

Quantification of Frequency Response Magnitude Specifications

A number of FRM specifications were reviewed to show that they provide a convenient
way of rating the stability and performance of both nominal and uncertain systems , and
5.3 Frequency Response Magnitude Specifications (B-stability) 149

to motivate incorporating FRM specifications into the framework of parameter space.

All of t hese specifications result in ensuri ng t hat the FR M of a transfer funct ion (i.e.
sensitivity functions S, T , Sc and SK , model regulator and model matc hing trans fer
functions H and E) is less than a bound specification as summarized here:

IS(jw)1 < aBs(w)

IT(jw)1 < aBr(w)

ISK(jW)1 < aBSK(w)

Vw E [w- ;w+] . (5.3.46)
ISc (jw)1 < aBsa(w)

IH(jw)1 < aBH(w)

IE(jw)1 < aBE(w)

By choosing appropriate bounds, as shown in t he preceding review, t hese speci-
ficat ions can be used to guara ntee a number of performance crite ria (e.g. tr acking,
dist urbance attenuation, noise rejection, model matching) and to guarantee robust ness
for both parametric and unstruct ured uncertainty. Th e specifications can be formulated
for different frequency ranges [w-,w+1 as ap propriate to meet design requireme nts and
to make compromises between the limitations in attainable performance.

Beta-stability
The definition of B-stability is ana logous to r- and 8-stability considered in Chapter 4
and Section 5.1, respectively.
As discussed previously in this section, stability and performance specifications may
be clearly represented by magnitude bounds on the four sensitivity funct ions S, T , Sc
and SK and on the model regulator and model matching transfer functions H and E ,
or, more genera lly, by magnitude bounds on an arb itrary transfer function M . Now,
a t rans fer function is considered to be B-st able if its frequency response magnitude
(FRM) lies within a desired region in the FR M-plane referred to as B. The region B is
bounded by t he B-boundary
(5.3.47)
Here, for the sake of simplicity, the bound aB is assumed to be the FR M of B (s ),
where B (s ) is assumed to be a rat ional tra nsfer function in s with real coefficients.
However, aB could be a piecewise curve in t he FR M-plane composed of multiple, not
necessarily contin uous , segments . Assuming non-negative real frequencies w , the upper
and lower bounds w+ and w- may take on values between 0 and 00. Fig. 5.43 shows
some examp les of B-stability regions for various B-specifications.
The notion of B-stability can be app lied to both LTI systems and parametric LTI
systems. For th e LTI case, a nominal transfer function G(s) is called B-stable in the
interval [w - ;w+], if
(5.3.48 )
150 5 Frequency Domain Analysis and Design

. '" ... .-. . -

w [rad/s] w [rad / s]

, . .. ..
' oB? .. " .. .. .. oB., · .
~~ 1 : I o!3~ -,
~B l
C1S3
B r

(}lh
B
oe,
-.. , ..
8B6
w [rad/5) w [rad/s]

Figure 5.43. Represent ation of various B-stability regions in the FRM plane

with w - E lRt , w+ E lRt.

In the case of a parametric LTI system with varying or uncertain parameters q ,
where q is included in a predefined operating domain Q, the family of transfe r functions
9(jw, Q) defined by (5.2.1) is called B-stable if

(5.3.49)

Application of B-stability
The notion of B-stability can be applied to both robustness analysis and robust con-
troller design.

i. Analysis: for robustness analysis, the task is to determine the parameter set QB
for which robust B-stability is satisfied. This set is defined by:
5.4 Mappi ng of Magnitude Sp ecificati ons into Parameter Space 151

Definition 5.31

(5.3.50)
o

If the operating domain Q is included in Qs , i.e, Q ~ Qs , then the family of

transfer functions 90w, Q) is robustly B-st able.
ii , Design: for robust controller design, the task is to determine the set of controller
parameters K s such that for all controller parameters k E K s and for a predefined
operating domain Q the family of closed-loop transfer functions 90w, Q, K s ) is
B-stable. This set is defined by:

Definition 5.32

Ks = { k 1190w, Q, k)1c B, Vw E [w - ;w+] } (5.3.51)

o

Thereby, in a first design step, K s is determined assuming a multi-model ap-

proach, meaning that only a finite number of operating points (representat ives)
are considered , e.g. the vert ices of Q. Then, after having selected a controll er
k E K s in a subsequent analysis step, t he set Qs is determined. If Q ~ Qs , th en
robust B-stability is satisfied for all q E Q.

In the sequel, we do not disting uish between anal ysis and design since t he mathematics
are the same. Therefore, except when oth erwise stated, the parameter vector q appli es
for both controller and plant parameters.

5.4 Mapping of Frequency Response Magnitude

Specifications into Parameter Space

In this sect ion, the mathematical equations required to map the FRM specifications of
the previous section int o parameter space are developed . This development is motivated
by the mean value theorem [61] . Consider the magnitude of a continuous transfer
function GOw, ql) at a fixed frequency w = w' given by

g(qd = IGOw', qdl , (5.4.1)

where ql E R The mean value theorem of continuous functions states that the funct ion
g(ql)' being continuous in t he interval [a ;bj with g(a ) =I- g(b), takes on all values in-
between g(a) and g(b). Now, let g(a) < IBOw') 1< g(b), where IBOw')1represents an
152 5 Frequency Domain Analysis and Design

FRM boundary (e.g. separating a desired region from an undesired region) . Then , for
continuous variation of ql between a and b, the boundary IB(jw')1 is crossed.
From this consideration, the boundary crossing theorem for frequency response mag-
nitude specifications may be derived [159] :

Theorem 5.33
The set of transfer functions

Q(jw, QB) = {C(jw , q) I q E QB} (5.4.2)

is B-stable if and only if:

1. There exists a B-stable transfer function C(jw) E Q(jw, Ql3) .
2. IQ(jw, Ql3)1 i- aB for all w E [w-,w+].
o
Theorem 5.33 states that starting from a B-stable transfer function C(jw,qo), qo E
Ql3 , B-stability is preserved for continuous variation of qo until a boundary aB is crossed.
If the boundary aB consists of multiple segments aBi , then Condition 2. applies to all
boundary segments .
Now, consider the FRM of the transfer function C(jw, q), which shall remain in the
B-stability region bounded by aB = IB(jw)l. A generalized B-specification of this typ e
can be given by the inequality

IC(jw,q)! < IB(jw)/ , Vw E [w- ,w+j . (5.4.3)

Rema rk 5.34

The generalized B-specification given in (5.4.3) can be given in the form of an

equivalent oo-norm specification , given by

IIC(jw, q) B(jw)-liloo < 1. (5.4.4)

Remark 5.35

Instead of an upper bound, a lower bound may also be assumed as suggested in

Figure 5.43, i.e.
IC(jw,q)l > IB(jw)!. Vw E [w-,w+] . (5.4.5)
o
5.4 Mapping of Magnitude Specifications int o Parameter Space 153

Remark 5.36

For systems where the dynamics have a significant dependence on the uncertain
parameters, it might be useful to adapt the boundary 88 to the plant parameters, i.e.
88(w, q) = IB(jw, q)1 . (5.4.6)
o
For mapping the generalized 8-specific ation (5.4.3) into parameter space, the task
is to determine the parameter set QB such that Q(jw, QB) is 8-stable. To establish the
mapping equations for (5.4.3), it suffices to consider two mathematical conditions, the
point condition and the tangent condition :
• The point condition is applied when IG(jw , q)1 starts or ends on the boundary 88
or runs across a beginning or an end point of 88 or a boundary segment 88 i .
• The tangent condition allows for the mapping of touching points, i.e. the points
where IG(jw , q)1 becomes tangent to a smooth branch of the boundary 88.
The Point Condition
For the point condition w.r.t. an arbitrarily fixed frequency w = w' , the set of parame-
ters QB <; ]R2 is determined by finding the boundary 8QB such that
IG(jw', q)1 = IB(jw')I, Vq E 8QB , (5.4.7)
i.e. such that the FRM of G and B have a common point at W·. The boundary 8QB
divides the (q}, q2)-plane into a finite numb er of regions. All of these regions have to
be checked by applying the boundary crossing theorem for FRM specifications, Theo-
rem 5.33. Rearranging and squaring (5.4.7) yields

c '=R (G(jw ,q)) 2+ 1 (G(jw ,q)) 2_ 1=o (5.4.8)

.. . e B(jw) m B(jw) ,
where ~ = ~(w , q) . Let
PI(W , q)
~ (w, q ) = d(w , q) . (5.4.9)
Further, assuming d(w, q) =I 0 (this condition must be verified separ ately), the point
condit ion can be stated as follows: for a fixed w = w· find q, i.e. ql and q2 , such that
PI(W' ,q) := O. (5.4.10)
Equation (5.4.10) is a polynomial equation in two variables and may be interpreted as a
real curve in implicit form depending on two parameters ql and q2. The corresponding
curve may be plotted by gridding one parameter, e.g. q, = qi; the parameter value
for q2 then results from computation of the roots of PI(qi, q2) = O. The number of
real solutions may vary since conjugate complex solutions for q2 may exist. To t ake
this into account, it is useful to determine the values of ql for which the number of
solut ions changes, which can then be used for intelligent gridding. These values of ql
are determined by computation of the roots of the discriminant
P1(W'
·
d Iscr ((.
q2 PI w , q},q2
)) = resq2 ( PI(w• . qi , q2 ) , 8 8q2,ql ,q2)) =.
0 (5.4.11)
154 5 Frequency Domain Analysis and Design

The Tangent Condition

For mapping the marginal condition that occurs when ICOw, q)1 touches (i.e. is tangent
to) the boundary IB Ow)I for a fixed frequency w = u/' , (5.4.10) must hold as well as

oiCOw, q)1 = -'--'-....;..;.

olBOw)1
(5.4.12)
ow ow
It is easy to show that the solution of (5.4.12) is equivalent to the solution of

of,(w , q) = !.- (Pl (w, q)) = 0 . (5.4.13)

ow ow d(w,q)

Further, by applying the quotient rule it holds that

(5.4.14)

Again, assuming d -:F 0, then if (5.4.10) is satisfied (i.e. Pl = 0), it can be seen from
(5.4.14) that (5.4.13) is satisfied if Pl' = O. Letting

,Opl(W,q)
P2 (w,q ) = Pl = ow ' (5.4.15)

the condition for a touching point, i.e, the tangent condition can be stated as follows:
For a fixed w = co" , find q such that

Pl(W, q) = 0 ,
(5.4.16)
P2(W,q) = O.

Note that we do not want to map touching points only at discrete frequencies , but
rather the complete branch of the boundary for wE [w- ;w+] is mapped into parameter
space . This is done by algebraically eliminating either w, ql or q2. The mathematics
used to map the point and tangent conditions, i.e. Equations (5.4.10) and (5.4.16) are
described in Appendix A.
For the case of multiple B-specifications or additional specifications (e.g. r-
specifications) , the intersection of all of the individual regions in the parameter plane
that satisfy the individual specifications form the desired parameter region. Thus, mul-
tiple specifications can be addressed simultaneously (e.g. sensitivity, loci and eigenvalue
specifications) when designing controllers or analyzing uncertain plant parameters for
parametric LTI systems.

Examples
The mapping of frequency response magnitude specifications is demonstrated by means
of two examples; an uncertain mass-spring-damper-system and the example introduced
in Section 5.1 and Example 5.19.
5.4 Mapping of Magnitude Specifications into Param eter Space 155

Example 5.37 (Uncertain mass-spring-damper-system)

Consider the transfer function of a mass-spring-damper-system with transfer func-

tion 2
wo
G(s,wo,D)= 2 2D 2 ' (5.4.17)
s + wos +wo
with uncertain parameters D E jR+ and Wo E jR+. We are looking for parameter values
D and Wo for which the magnitude plot IG(jw, Wo, D)I resides between a lower and upper
bound IG-(jw)1 and IG+(jw)l , respectively, as shown in the left plot of Figure 5.44, i.e,

The selected B-boundaries are the magnitudes of the two transfer functions

G-(s)

i.e.

IG-(jw)l ,
IG+(jw)1,

with parameter values

9/10 , 7/10 , 10,

11/10 , 1/10 , 25 .

Since it holds that

lim G(wo , D,jw) = 1 , lim G-(jw) = 11:1 , lim G+(jw) = 11:2 ,

w- o w-o w-o
it also holds that the locus G(jw,wo,D) and the upper and lower bounds 8B 1 (w) and
8B 2 (w), respectively, do not depend on Wo and D. The three loci start on the real axis
at 1, 11:2 and 11:1 , respectively. For 11:2 i:- 1 and 11:1 i:- 1, no common starting point oflocus
and upper or lower bound, respectively, exists . Further, the three loci G(jw,wo, D),
G+ (jw) and G- (jw) end at the origin, since

lim G(jw,wo,D) = lim G+(jw) = lim G-(jw) =

w- oo w-oo w-oo
o. (5.4.18)

Thus, the locus has a common point with its upper and lower bound for w -+ 00 .
Applying the point condition (5.4.10), the two results for upper and lower bound are

Wo = WOI ~ and Wo = w02VK2 .

156 5 Frequency Domain Analysis and Design

20 r-- - - ----------,
1.5

(; 1--
c::>
"§
-
.., - 20 IG( D.D'.j:.J)\ Q

b"o 0.5
c::: - 40
~
- 60 L..- --J

0 1 2 0
10 10 10 10 15 20 25 30
W [rad/ J Wo

Figure 5.44. Upper and lower bound in parameter space

The two solutions, i.e, the two straight lines are plotted with dashed linestyle in the
right plot of Figure 5.44. In the next step, the tangent conditions for upper and lower
bounds are formulated . According to (5.4.16), the tangent condition for the lower bound
(i = 1) and uppe r bound (i = 2) are:

In this case, only even powers of w, Wo and D appear in PI and P2 . Therefore, we

may substitute w = w2 , Wo = w5 and b = D 2 • Note that W, Wo and jj still enter the
coefficients of PI and P2 in a polynomial way. The computation of the parameter space
boundaries is done according to Appendix A with the results shown in solid linestyle
in the right plot of Figure 5.44.
For the B-stable point (wa, DO) E QB marked with x in the right plot of Figure 5.44,
the corresponding magnitude plot is drawn in the left plot with dashed-dotted linestyle.
As expected, the magnitude plot resides between lower and upper bound.
o

Example 5.38

The closed-loop system described in Example 5.19 shall be analyzed with respect to
the loop-shaping FRM criteria "robust stability" and "nominal performance" . Consider
the single-loop unity feedback system in Figure 5.1 with open-loop transfer function
Li» , q) according to (5.1.1), sensitivity function S(s , q) according to (5.3.3) and com-
plementary sensitivity function T(s , q) according to (5.3.2). In addition to parametric
uncertainty captured in qi and qz , a mult iplicative perturbation model according to
(5.3.27) is assumed to account for unmodelled dynamics. The weighting function of th e
5.4 Mapping of Magnitude Sp ecificati ons int o Param eter Sp ace 157

multiplicative perturbation model is selected as

WA(S) = hg s + WTlg ,
s +wThg
where 19 = 0.2, hg = 5 and WT = 21r [rad / s], meaning that for low frequencies (w <
21r [rad/s]) the model error might be up to 20 % and for high frequencies (w > 21r [rad/s])
up to 500 %. Furthermore, to account for good disturbance attenuation properties: we
select an upper bound for the magnitude plot of the sensit ivity function S , i.e,

8B s(w) = IhgJw!W+WTh
+ WTlgl
g
where 19 = 0.5, h g = 2 and WT = 0.11r [rad/s]. This means that for low frequencies ,
disturbances shall be attenuated to at least 50 % and for frequencies W > O.h [rad/s]
disturbance amplification shall be limited to a maximum of 200 %. The corresponding
required Nyquist-stability margin is at least l/hg = 0.5.
To satisfy the above-mentioned specifications, three conditions have to be taken into
consideration:
i. Hurwitz-stability of the characteristic polynomial, i.e, the denominator polyno-
mial of (5.3.3) or (5.3.2) , respectively, must be proven . Therefore, in the first
step the Hurwitz boundaries are mapped into parameter space , i.e, in the plan e
(ql , q2)' The resul ting Hurwitz-stable region is
QH = {ql Re (Roots (num (1 + G(s ,q)))) < O} .
ii . The robust stability condition according to (5.3 .34) requires (besides Hurwitz-
stability) a magnitude specificat ion on T. Mapping this condition into parameter
space yields the region
QBT = {q IIWA(jw)T(jw, q)1 < 1, for wE [0; oo)}.
iii. The nominal performance specification requires (besides Hurwitz-stability) a mag-
nitude specification on S. The corresponding region in parameter space is defined
as
QBS = {q IIS(j w, q)1 < 8B s(w), for w E [0; oo)} .

The resulting Hurwitz-, Bs - and Br-stable regions in the parameter plane (ql , q2) are
shown in Figure 5.45. The parameter region for which robust stability w.r.t. parametric
and unstructured uncertainty can be guaranteed is the intersection of Hurwitz- and B-r-
stable region , i.e, QH n QBT ' Good disturbance attenuation properties are guaranteed
for all parameter values q E QHnQBs' Finally, the B-stable region might be intersected
with the r- and 6-stable regions to additionally account for good damping and phas e
and magnitude margin. The parameter values for which all r-, 6- and B-specifications
are satisfied simultaneously is defined by
QrnQenQB ,
where

o
158 5 Frequency Domain Analysis and Design

:-1 [;.
6 ~---.r-----.-------,

o N
6 ,----,-r-----r-- - - - - - - ,

o
-2 "" - 2
-4 -4

-8 _8 ' - ---L- -L-- - - - --' _8 LJJL...-- ' -- - - - - -'

o 2 3 o 3 o 2

Figure 5.45. Individual parameter regions

601,--- - -,-- - - - ---, 6,----,-......,,---,--.---.--- - --,

of s-2
o

~:l -4 -4

-6
-8 --JL--1. ---J _8 L....1IL -.- ' - --l
o 2 3

Figure 5.46. Nominal performance (left), robust stability (middle) and simultaneous f-, 6-
and B-stability (right) in parameter space

For more details about the mapping of frequency response magnitude specifications
into parameter space the reader is referred to [158] .

5.5 MIMO Systems

In this section, we describe methods to use the parameter space approach to analyze and
design control systems for multivariable or multi-input, multi-output (MIMO) systems.
While specifications on eigenvalues can be expressed by the characteristic polynomial
for both SISO and MIM0 systems (see Section 1.5), there are many ways to treat MIM0
systems in the frequency domain. Some of the specifications and design techniques for
MIMO systems are derived from their SISO counterparts, while the more potential ones
take the multivariable nature into account. There has been an enormous interest in the
5.5 MIMO Systems 159

design of multi variabl e control systems in the last decades, especially frequency domain
approaches [78],[86],[1371. We do not intend to give a comprehensive treatment of all
the aspects of multivariable feedback design and refer the reader to the cited literature.
Thus, the scope of this section is limited to the presentation of the basic concepts and
some examples. As the main result, this section presents the mapping equations for
widely used MIMO specifications, which enable the incorporation of these specifications
into the parameter space approach.
The model description is given as a transfer matrix representation

y = G(8 ,q)U, (5.5.1)

where u E IR m and y E IRP are vectors of signals , or as a parametric state space model

x] = [A(q) B(q)] [x] .

[y C(q) D(q) u

For ease of presentation, we consider only systems with the same number of inputs
and outputs, i.e. m = p. Nevertheless, all results are valid for non-square systems
with m f. p.

SISO Methods
For SISO control systems , classical gain and phase margins are good measures of robust-
ness. Furthermore, loop-shaping techniques provide a systematic way to attain good
robustness margins and desired closed-loop performance. The methods introduced in
Sections 5.1-5.4 facilitate such a design . However, the classical gain and phase margins
are not reliable measures of robustness for multivariable systems .
The simplest approach to multivariable design is to ignore its multivariable nature
and just look at one pair of input and output variables at a time . In that manner,
we can apply all SISO methods described in the previous sections of this chapter. A
classical design procedure using this idea for multi variable systems is the sequential
loop closing method, where a SISO controller is designed for a single loop. After this
design has been done successfully, that loop is closed and another SISO controller is
designed for a second pair of variables, and so on.

Example 5.39

Consider the following plant

(5.5.2)

In the first step, we design a constant gain controller ku (8) = k1. The transfer function
seen by this controller is 9U(8) = 2/((s + 1)(8 + 2)) . Setting k1 = 1 leads to a stable
160 5 Frequency Domain Analysis and Design

transfer function . After closing this loop, the transfer function seen by a controller from
output 2 to input 2 is

A stabilizing controller for this transfer function is kds) = 1. The resulting decentral-
ized controller is thus given by the identity matrix K(s) = I . We will come back to
this example with a better solution in Example 5.41 and Example 5.43. 0

This method has a number of weaknesses. During the controller design, the re-
sulting scalar transfer function for the i-th step might be non-minimal-phase , although
all members of G(s) are minimal-phase transfer functions. This might pose a severe
constraint for the control design, since non-minimal-phase transfer functions limit the
maximal usable gain.

MIMO Specifications
Apart from stability, the most important objective of a control system is to achieve
certain performance specifications. One way to describe these performance specifica-
tions is to use the size of certain signals of interest. For example, the performance of
a regulator could be measured by the size of the error between the reference and mea-
sured signals. The size of signals can be defined mathematically using norms. Common
norms are the Euclidean vector norms
n n
IIxliI := :L Ix;!,
i=l
IIxl12 := :L I iI2,
i=1
X Ilxll oo := max
19~n
IXil.

The performance of a control system with input and output signals measured by one
of the above norms (not necessarily the same) can be evaluated by the induced matrix
norms. The most prominent matrix norms used in control theory are the H 2- and
Hoo-norms , which will be considered in this section. A detailed treatment of mapping
MIMO specifications into parameter space is given in [151) .

The H2-norm is a widely used performance measure that allows to incorporate time-
domain specifications into control design. The H2-norm arises, for example, in the
following physically meaningful situation. Let the system input be zero-mean stationary
white noise of unit covariance. Then, at steady state, the variance of the output is given
by the square of the H2-norm .
The H2-norm is defined as

1 Joo ) 1/2
IIGII 2 := ( trace 211" -00 G(jw)G(jw)*dw ,
5.5 MIMO Systems 161

where G(s) is a stable transfer ma trix and' denotes the complex conju gate transpose of
a matrix. The above norm definition can be used for generic square integrable functions
on the imaginary axis and is then called L 2 -nor m . Thus, strictly speaking without the
stability condition, we are mapping the L 2 norm instead of the H 2-nor m .
The H 2-norm defined above is finite if and only if the transfer matrix G(s) is strictly
proper, i.e. the dire ct feedthrough matrix D = 0 for a state space realization of G{s) .
Hence, we assume D = 0 in the following presentation.
The H2 -nor m can be expressed as [58]

IIGII; = trace (c 1 eAtBBTeATtdtCT)

00

trace (Clv,;C T) , (5.5.3)

where
lv,; = 1 00
eAtBBTeATtdt

is the controllability Gramian of the realization. The controllability Gramian can be

obtained from the Lyapunov equation
Alv,; + lv,;A T + BBT = O. (5.5.4)
Equation (5.5.4) is affine in the entries of the unknown matrix lv,; = lv,;T. This system
of ~n(n + 1) linear equations, where n is the system order, can be solved for any para-
metric dependency of A(q) , C{q) . Thus, the parametric controllability Gramian lv,;{q)
can be obtained easily.
Substituting this parametric solution lv,;(q) into (5.5.3), we get the following poly-
nomial mapping equat ion
IIGII; = trace (C(q)lv,;{q)CT(q)) = 'l , (5.5.5)
where,2 specifies the desired performance level. Equation (5.5.5) is an implicit equation
in the uncertain parameters q . Since the desired performance level is not known a priori,
it is recommended to vary , or determine the set of parameters Q2 for which IIGII; = , 2
holds for multiple values y. Also, a gray-tone or color coding of the different sets Q2(,)
is useful.
Note : unlike other specifications described in this book, the H 2-nor m mapping
equation (5.5.5) does not contain an auxiliary parameter or frequency . This is in line
with the fact that the H 2 -norm can be expressed as an integral over all frequen cies. This
fact makes (5.5 .5) useful on its own, as opposed to being solely used for mapping H 2
specifications into a parameter plane, especially to analyze and design control systems
for more than two parameters.

Remark 5.40
The classical linear-quadratic-regulator (LQR) problem [122] , which aims to mini-
mize the objective function

J=-
2
11 0
00
(XTQX+uTRu)dt
162 5 Frequency Domain Analy sis and Design

for a state-feedba ck controller u -Kx, can be cast as a H 2-norm problem [57].

Namely, th e LQR problem considers a linear time-invariant system described by the
state equations

Ax+Bu+w , (5.5.6)

(5.5.7)

where u is the control input, W is unit intensity white noise, and z is the output signal
of interest. The LQR problem is then to design a state-feedback controller u = -Kx ,
which minimizes the H 2-norm between wand z . The performance index J is then
given as
J = IIGw_zll~ .
The parametric LQR control design described here applies to parametric SISO systems
as well. It allows us to explicitly incorporate control effort specifications into a robust
controller design, which is not possible with pure eigenvalues specifications. 0

Example 5.41

Revisit Example 5.39. We will design a decoupled constant-gain output-feedback

controller u = - K y with

(5.5.8)

which minimizes the following LQR-Iike performance index

J= - 11
2 0
00

(yTy+auTu}dt. (5.5.9)

This performance index treats both outputs equally, which is reasonable since the open-
loop plant has similar gains for these outputs. The parameter a provides an adjustable
design knob, which allows an intuitive trade-off between the integral error of the com-
manded output and the actuator effort. For this specific example, we assume a = 1.
The open-loop plant has pure real eigenvalues at {-I, -2}. Thus , we additionally re-
quire the rather stringent specificat ion that all closed-loop eigenvalues should have a
minimal damping of 1.0.
We will solve this problem by mapping the design requirements into t he kl , k2 con-
troller parameter plane . To this end, we formulate the LQR output problem (5.5.9) in
t he H2-nor m framework by employing Remark 5.40.
Using the fact that y = Cx , the LQR weight matrices in (5.5.7) for this problem
are given by
Q=cTc, R=l.
5.5 MIMO Syst ems 163

In order to apply the algebraic mapping equations (5.5.5), we need a state space
description of the system. A minimal realization of the system (5.5.2) is given by

-2 0 -2 -4

[:]~I-t--
o
[:] 1 -1
-1 -2 -2

0 0
(5.5.10)

-1 1/2 0 0

We incorporate the controller u = -Ky in parametric form into (5.5.6) and (5.5.7) to
get the state space system G(s) defined in the H 2-norm mapping equation (5.5.5). For
the particular problem considered in this example, these equations are given by

x = (A-BKC)a::+w, (5.5.11)

(5.5.12)

And the parametric transfer function G(s)w_z becomes

2(k) - 2k2 - 1) -2(k1 - k2) 1 0

2(k1 - k2) -2k1 + k2 - 1 0 1

-k) k1 0 0
G(s)w_z =
k2 -~k2 0 0

1 -1 0 0

-1 1/2 0 0

The controllability Gramian for this problem is obtained from (5.5.4) as

1
l-Vc(kI, k2) = 6(k 1 + l)(k2 + 1)2 •

4kr - 5k1k2 + 5/2k~ + 3k1 + 3/2 (k 1 - k2)(4k1 - 5k2 - 1) ]

[ (k) - k2)(4k) - 5k2 - l,) 4kr - 11k1k2 + lOk~ - 3k) + 9k2 + 3

Finally, the resulting performance index can be computed as

J = (4k1 + 5k2 + 9)( 2kr k2 + 2kr + klk~ + k~ + k1 + 2k2 + 3) . (5.5.13)

24(k1 + l)(k 2 + 1)2
164 5 Frequency Domain Analysis and Design

/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
:.'

- 3 2 3
k1

- 2

- 3

Figure 5.47. LQR boundaries (solid) and f-stability (dotted) for MIMO control example

For a given performance level J = J*, (5.5.13) provides an implicit mapping equa-
tion in the unknowns k1 and k 2 • The minimal achievable performance level J for
the decentralized output feedback (5.5.8) is bounded from below by the performance
level J l ull obtainable with a dense static-gain feedback controller. Using classical LQR
theory, J l ull is easily calculated as Jlu/l = 0.824. In general J > J l ml , therefore, we will
map J = 1 into the parameter plane. Figure 5.47 shows the resulting parameter set Q2
for J = 1. The region satisfying both LQR and f -stability requirements is shaded
in the figure. The figure actually shows that we can set k, = 0, while still obtaining
reasonable controllers. Note that without the damping specification on the closed-loop
eigenvalues, we would still need to map the Hurwitz-stability requirement to get the
correct LQR set . For this example, stability is assured if

k1 > -1 1\ k2 > -1 .
The actual boundary values k1 = k 2 = -1 appear in the denominator of (5.5.13) .
The minimal obtainable performance level J for the decentralized controller can
be computed using the algebraic equation (5.5.13). For k1 = 0.2074, k 2 = 0.9329, we
get J = 0.8421 , which is only slightly higher than Jlu/l . Compare this to J = 1.125 for
the controller designed in Example 5.39.
The numerical values of J change with the state space representation considered .
Thus, the actual values of J should be only considered to measure the relative per-
5.5 MIMO Systems 165

formance of a controller compared to a reference controller, e.g. the dense optimal

controller or zero-gain controller (open-loop system) . This becomes apparent when one
is computing J for the state space representations (5.5.10) and (5.5.23), which both
lead to the same input-output behavior but different quantitative values of J . 0

Hoo-norm
The Hoo-norm of a transfer matrix G(s) is defined as
IIG(s)lloo := sup a(G(jw)), (5.5.14)
w
where a is the largest singular value or maximal principal gain of an asymptotically
stable transfer matrix G(s). Note that (5.5.14) defines the Loo-norm if the stability
requirement is dropped.
For a scalar transfer function C(s), the infinity norm can be interpreted as the
maximal distance of the Nyquist plot of C(s) from the origin or as the peak value of
the Bode magnitude plot of IC(jw)l. In that sense, the frequency response magnitude
specifications in Section 5.3 could be recast as a scalar Hoo-norm problem .
For MIMO systems, the Hoo-norm describes the maximum amplitude of the steady
state response for all possible unit amplitude sinusoidal input signals. In the context
of stochastic input signals, the Hoo-norm can be interpreted as the square root of the
maximal energy amplification for all input signals with finite energy.
We use robust stabilization as a classical control problem that fits into the Hoo frame-
work to motivate the mapping of Hoo-norm specifications. Different from the traditional
literature about Hoo control theory [203],[86] we will treat structured (parametric) and
unstructured uncertainties.

Robust Stability
The well-knownsmall-gain theorem (see, e.g. [203]) states that a simple feedback system
is internally stable if the Hoo-norm of the resulting open-loop transfer matrix is smaller
than l.
As an example, consider a plant G(s) with additive, unknown perturbation A(s)
and a controller with transfer function C (s). The block diagram is shown in Figure 5.48.

Figure 5.48. Feedback system including plant with additive perturbation

The problem is, how large can IIAlloo be so that internal stability is preserved?
Using simple loop transformations, we can isolate the perturbation A, which can be
seen in Figure 5.49.
166 5 Frequency Domain Analysis and Design

Figure 5.49. Feedback system including plant with additive perturbation

The small-gain theorem now states that the feedback loop is internally stable if

(5.5.15)

Since the Hoc-norm is an induced norm, the following multiplicative property holds:

IIABlloc < IIAllocllBlloc. (5.5.16)

And we get the following sufficient condition for internal stability with respect to un-
structured uncertainty:
(5.5.17)

Mapping equations
We now consider the mapping equations for general Hoc norm specifications. Here, a
condition IIG(s, q)lIoc < , is mapped into the q-space to obtain the set of parame-
ters Qoc for which the condition holds.
In order to get the mapping equations for a parametric system with state space
realization G(s, q) C(q)(s[ - A(q»-l B(q) + D(q), we define the Hamiltonian
matrix

(5.5.18)

where R = DTD - ,2[ and S = DDT - ,2[.

The following well-known theorem can be used to compute the Hoc-norm.

Theorem 5.42 (Boyd, Balakrishnan, Kabamba [56})

Let A be stable and , > a(D).

Then IIGlloc ~, {::::::> H.., has pure imaginary eigenvalues (i.e. at least one).

o
5.5 MlMO Sys tems 167

Thi s theor em has been used to obtain a numerical bisection method to compute
the Hoo norm of a given transfer matrix [56]. We use the above theorem in its symbolic
form to derive mapping equations for the Hoo-norm. The boundary of a parameter
region with II G lloo ~ '"'I is given by parameters for which H.., defined in (5.5.18) has
pure imaginary eigenvalues. The eigenvalue equation

det[sI-H..,)=O

can be used to check if s is an eigenvalue of H..,. Using Theorem 5.42, we are interested
in pure imaginary values, and it follows that

(5.5.19)

is a necessary condition for parameters q for which IIGlloo = '"'I holds. Since H..,
is a Hamiltonian, pure imaginary eigenvalues will exist only as double eigenvalues.
Therefore, the following double root condition for polynomial equations applies

_ oel (S,'"'I ,q} ! -0

e2(w, '"'I, q) - 0 - . (5.5.20)
s s=j",

Equations (5.5.19) and (5.5.20) define two polynomial equations that can be used to
map a given Hoo-norm condition into parameter space. The following simplifications
can be made for (5.5.19) and (5.5.20) . Equation (5.5.20) contains the factor w, which
can be neglected since the equation is only valid for w > o. After dropping this factor,
both equations contain only terms with even powers of w. Thus, it is favorable to
substitute w 2 = n in order to reduce the degree of the equat ions.
Additionally, we have to consider both ends of the imaginary axis. Real eigenvalues
can merge at the origin and form a pair of pure imaginary eigenvalues:

detfjwI - H ..,)"'=0 = det H.., = O. (5.5.21)

The same applies to eigenvalues becoming imaginary at infinity:

detfjwI - H..,)",=oo = O. (5.5.22)

Condition (5.5 .22) is just the coefficient of the term with the highest degree in s of the
determinant det[sI - H..,).
Note that Equations (5.5.21) and (5.5.22) are not sufficient since they determine all
parameters for which the Hamiltonian (5.5.18) has a pair of eigenvalues at the origin and
infinity, respectively. This includes real pairs of eigenvalues that are just interchanging
without becoming imaginary. Thus, in order to get sufficiency, we have to check all
parameter s sat isfying (5.5.21) and (5.5.22) if there are only real eigenvalues.
Th e mapping equations (5.5.19}-(5.5.22) have a similar structure like the well-
known equations for f -stability, where (5.5.19) and (5.5.20) can be interpreted as the
complex root boundary condition , (5.5.21) as the real root and (5.5.22) as the infinite
root boundary condit ion (see Example 2.6).
Theo rem 5.42 has two prerequisites, namely the system should be stable and the
maximal singular value of t he direct feedthrough matrix D should be less than the
168 5 Frequency Domain Analysis and Design

desired performance level 'Y. The stability requirement can be mapped using the map-
ping equations for Hurwitz-stability described in Chapter 4. Using a frequency domain
formulation , it can be shown that the condition O"max(D) < 'Y is implicitly mapped
by (5.5.21) and (5.5.22).
Note that the mapping equations for the Hoo-norm are not just characterizing the
parameters for which IIGlloo = 'Y . Actually, all parameters for which G(s, q) has a
singular value such that 0"(G) = 'Y are determined. Thus , we might get boundaries in
the parameter space for which the i-th biggest singular value 0"; has the specified value 'Yo
This is similar to eigenvalue specifications, where we get boundaries for each eigenvalue
crossing of the r -boundary. Thus, similar to the r-case, we have to check the resulting
regions in the parameter space if the specifications are fulfilled or violated (possibly
multiple times) .

Example 5.43

We analyze the robust stability of the plant given in Example 5.39 for decentralized

:]
static-gain controllers

C(s) = kOl
[

with respect to unstructured additive uncertainty. Using (5.5.18), the Hamiltonian H"(
for this problem becomes

-2 - k, -2k2 5h 3h
-k1 -1- k2 3h 2h
H"(=
-8h 6h 2+k 1 k1

6h -5h 2k2 1 + k2

Figure 5.50 shows the gray-tone coded sets of parameters that correspond to different
tolerable uncertainty sizes. Dark areas correspond to poor robustness whereas areas
with lighter colors indicate good robustness. The initially designed controller C(s) = I
yields approximately the same robustness as the open-loop system. It can be seen from
Figure 5.50 that decreasing both gains, e.g. to C(s) = 0.5 I , will result in a system
that is more robust for the uncertainties considered.
o

Remark 5.44

The mapping equations establish conditions that allow the direct computation of
the Hoo-norm . Thus , instead of using the numerically attractive bisection algorithm
5.5 MIMO Systems 169

r-~--- ....
Robust stability for add.tive uncerta inty

12 1.4

Figure 5.50. Stability with respect to unstr uctured, additive uncertainty

widely used in control software , the Hoc-norm can be computed by solving the two
algebraic equations (5.5.19) and (5.5.20) in the two unknowns wand , . These positive
solutions provide candidate values, for the Hoc-norm. Additionally, th e solutions of
(5.5.21) and (5.5.22) are computed and the Hoc-norm of the system is given by the
maximal value over all candidate solutions. As a byproduct, we get th e frequency w ,
for which th e maxim al singular value occurs . 0

Example 5.45

We use the mapping equations to directly compute the Hoc-norm for the open-loop
transfer function G (s ) given in Example 5.39. A state space representation for G (s) is
given by
- 2 0 1 2

0 - 1 1 1
G (s) = (5.5.23)
- 2 2 0 0

2 - 1 0 0
The mapping equations (5.5.19) and (5.5.20) become

el , 4W4 + (5,4 -=- 14, 2)w 2 + 4,4 - 8,2 + 4,

e2 w(4,4 w2 + 10,4 - 28,2) .

This polynomial system of equations has only a single relevant solution w = 1" = .j2
with w" > O.
170 5 Frequency Domain An alysis and Design

The DC-gain condition (5.5.21) is

detH-y = -l- 2·l + 1,

which has the positive solution I = 1. Note that this can be easily observed by the fact
that G(jO) = I, and thus the singular values are given by 0"1 = 0"2 = 1.
There is no solution for (5.5.22) and we can conclude that the maximal singular
value of G(s) is given for w = 1 and IIG(s)lIoo = J2. 0
6 Case Studies in Car Steering

Many dangerous situations occur on the roads because a car driver does not react
fast enough at the beginning of skidding or rollover. Automatic feedback systems can
assist the driver to overcome such dangerous situations. A further step is automatic
driving by following a lane reference. In both cases, robustness of the control system
with respect to the uncertain road-tire contact is an important requirement. Further
uncertain parameters may be, for example, vehicle mass and velocity, or the load-
dependent height of the center of gravity (CG) above road level.
In the present chapter, a relatively simple car model is derived. It takes into account
only the most essential non-linearities of the lateral tire force characteristic and later
the rate limitation of the steering actuator.

6.1 Tires, Braking, and Steering

A car driver uses the pedals for commanding the longitudinal motion, and the steering
wheel for commanding the lateral motion and heading. What looks like two indepen-
dent actuators becomes coupled in extreme driving situations, because all commands
use the same limited resource, that is, the contact force between tire and road . Its sat-
uration limit is approximately described by the Kamm circle (Coulomb dry friction),
see Figure 6.1. In tire coordinates xt, yt, zt , the resultant of the tire forces Fxt and
Fyt cannot exceed the limit given by the normal force F zt and a maximal adhesion
coefficient J.L ~ 1.
(6.1.1)
Under braking or acceleration of a single tire, IFxtl is increased. If this happens close to
tire saturation, then the available lateral force IFlitl is decreased. For acceleration and
gas release, there is an obvious difference of the vehicle dynamics between front-wheel
and rear-wheel drive vehicles. The available maximal force J.L Fzt varies under:

i. changing road surface (e.g. ice patches on a dry road). This may happen rapidly
and unforeseeably.
ii . braking and acceleration. This changes Fzt , e.g. while braking, Fzt is increased at
the front wheels and decreased at the rear wheels.
172 6 Case Studies in Car St eering

Figure 6.1. Kamm circle.

For the design of vehicle dynamics control systems, the uncertainty i. means that ther e
is no time for identification and adaptation; robust control is needed. The uncertainty
ii. means that gain-scheduling , e.g. by the longitudinal acceleration is possible, but in
view of i. it is not worthwhile to model the effect very accurately.
For a stepwise change in u, the system inherits an initial condition from the dynami cs
before the step . The car steering has now to be stable for the new constant value of u.
Therefore a constant, but unknown value of p. is assumed in this chapter .

Remark 6.1

The reader with a control engineering background should be warned that the notions
of controllability and stability are used in the automotive literature in the following
sense: loss of controllability means front-wheel tire saturation, loss of stability means
rear-wheel tire saturation. These notions are not identical to the definitions in system
theory!
o
The saturation limits are not reached under moderate handling maneuvers on a dry
road . Then, the commanded forces in the (x, y)-plane are below saturation. In this
normal driving situation, the foot and hand commands of the driver are independently
executed . This is what the driver is used to. Safety-critical situations in handling occur
when the response of the car changes suddenly as the car gets onto a slippery road, e.g.
on an icy bridge or on wet leaves on the road.
6.1 Tires, Braking and Steering 173

Safety-critical situations also arise under disturbance torques like side wind, asym-
metric road surface or tire burst (fL-split braking), trailer pulling in gusty wind or bad
road surface. Also, an escape maneuver to avoid a collision with a surprisingly ap-
pearing human, animal or other car may drive the tires to the limits. In all surprising
situations, it takes the driver 0.5 to 1 second to react, and frequently it is an overre-
action that makes things worse. Roughly speaking , we may compare the driver in this
situation with a feedback system with time delay and high gain, which causes instabil-
ity. For small continuous disturbances there is also a comfort issue; the driver should
still be relaxed after pulling a trailer in gusty wind for some time.
There exist already driver assistant systems (ESP) that use braking (or increased
traction) at the individual wheels [195] , [153], [72] . These systems are cheap, because
they use the hardware of the existing ABS braking system with an additional yaw rate
sensor and do not require a new actuator. So, why is it of interest to look at the steering
alternative?
The first reason is that we need a torque to compensate yaw disturbance torques .
Torque is tire force times lever arm . For braking, the lever arm is half the trackwidth
t, for steering it is approximately half the wheelbase, which is bigger by a factor of
approximately two. In other words, only half the tire force is needed to generate a
required compensation torque by steering rather than by braking.
The second reason is that different friction coefficients fL on the left and right sides
(fL-split braking) may be the cause of the disturbance torque, which can be reduced
only by reduced braking, i.e. longer braking distance . In contrast, a steering torque can
compensate the braking torque and achieve a straight short braking path. Figure 6.2
shows the optimal combination of full braking and steering under the assumption, that
the CG is in the center of the wheelbase and the wheelbase is twice the trackwidth.
(The CG is projected to the road level, where the tire forces occur.) Also, equal vertical
forces Fzt for the front and rear wheels are assumed.

dry side

I
I
I
I
I
- - - - - - of - - +--------+-+--;...,.----------i
t:
I
I
I

slippery side
<:» "t

Figure 6.2. Compensation of j.l-split brake torque by steering torque

174 6 Case Studies in Car Steering

The disturbance torque

Mzbrake = (J.Lle/t - J.Lright)Fzt(1 + a)t/2 (6.1.2)

is compensated by the steering torque

Mz.teer = (J.Lle/t + J.Lright)Fzt bt (6.1.3)

with the saturation constraint (6.1.1)
va 2 + b2 ::; 1. (6.1.4)
The retarding force
Fz = (J.Lle/t + J.Lright)(1 + a) F zt (6.1.5)
is maximized under the constraint, that the torques compensate each other.
As a numerical example, let J.Lle/t = 1, J.Lright = 0, then a = 0.6, b = 0.8 makes the
most efficient use of the available tire force J.LFzt .
For other CG-Iocations and ratios trackwidth/wheelbase, it is easy to calculate the
Pythagoras ratio between braking and steering forces at the front wheels. It may also
be scheduled by the rear-front shift of F; under braking or acceleration. This braking-
steering combination fully exploits the physical limits given by J.LmazFz.
A third reason for driver assistance by steering (rather than by braking) is energy
conservation, reduced wear of tires and brakes and smooth operation around zero cor-
rection. Practically, braking systems cannot immediately compensate small errors , but
they have to intervene relatively late in detected emergency situations when the car
is close to skidding and the driver is unpleasantly surprised by the intervention. Only
the steering system is feasible for continuous operation, also for better comfort under
continuous disturbances.
The implementation of a steering control system is schematically shown in Fig-
ure 6.3. The main steering angle asis commanded by the driver from the steering
wheel. A small corrective steering angle is set by an actuator with input from a feed-
back controller. The main feedback signal is the yaw rate. (Yaw rate sensors are already
used in individual wheel braking systems and are therefore cheap.) The velocity signal
from the ABS system may be used for gain-scheduling the controller. The velocity is ap-
proximated here by the wheel rotation, neglecting the longitudinal slip. Under normal
driving conditions, this slip is smaller than 2 %. From the velocity and the handwheel
angle, a reference rref for the yaw rate r is generated in the feedforward path of the
controller.
The superposition OF = os+oc may be done mechanically, as indicated in Figure 6.4.
An alternative is a steer-by-wire system with an electrical addition os+oc and a common
electric actuator. The cost and safety issues related to such an actuator are presently
(2001) the bottleneck for the widespread use of active steering.
The tire forces depend on the steering angle OF and on state variables of the chassis.
Figure 6.5 shows the velocity vector VF under a sideslip angle {3F at the front axle with
respect to the longitudinal axis (x-axis) of the chassis. The lateral force FYt F in tire
coordinates is a function of the tire slip angle OF = OF - {3F. Its dominant component
in chassis coordinates is
(6.1.6)
6.1 Tires, Braking and Steering 175

Yaw rate
sensor

Figure 6.3. Feedback controlled additive steering angle 8c

The small retarding component FxF = - FytF sin 8F does not generate a yaw torque, if it
occurs symmetrically at the left and right wheel. Its longitudinal effect is compensated
by speed control (automatic or by the driver) . In a static tire description, the tire side
force FYtF is a function of the tire slip angle OF, i.e.

FYtF = f(OF) = f(8 F - 13F) ,

(6.1.7)
FyF = f(8 F - 13F) cos 8F.

The index F indicates the front wheels; it is replaced by R for the rear wheels.
The non-linear functions are determined experimentally for a specific tire and
smoothed by a mathematical approximation, see Figure 6.6. If the velocity vector
VF is aligned with the tire , then the lateral force is zero, f(O) = O. For 0A > 10°, th e
lateral force is close to saturation. Of course, a control system cannot overcome such

Figure 6.4. Mechanical implementation example for additive steering by linear actuator
176 6 Case Studies in Car Steering

J_-_-x
Y

Figure 6.5. Lateral forces FytF at the front wheel in tire coordinates and FyF in chassis
coordinates

physical limits. Therefore, it is very important to design steering controllers such that
only small tire sideslip ang les occur .
In a dynamic description , the lateral tire force does not respond immediately to a
step in the steering angle . The steady state deformation of th e rubber tire is reached,
however, after approximately one half wheel rotation. This phase delay, together with
other elasticities in the connection between the steering mot or and th e steering angl e
and delays in the steering motor itself, call for sufficient phas e margin in the control
design .

... ...
-; ~ . . .. . . . . . ... ...

::./
.....
2000

'"
OJ

//
U
....
.8
OJ
..,
.!::

j
:J - front
..... rear
10
sideslip angle [degree]
Figure 6.6. Lateral tire forces as a function of the tire sideslip angle
6.2 The Two Steering Tasks 177

6.2 The Two Steering Tasks

An important requirement in all man -machine systems is that the automatic control
system should not interfere with the human control task and vice versa. We will dis-
tinguish two steering tasks, one to be controlled by the driver , the other one under
automatic contro!' The driver should execute the primary steering task, that is, path
tracking. The vehicle is considered here as a point mass mDP and its velocity vector v
must be reoriented by a lateral acceleration ayDP in order to follow the planned path,
see Figure 6.7. The planned path exists only in the mind of the driver (e.g. change lane,
enter into a curve) , and v is the actual tangent to this path. The path is continuously
updated by the driver . The secondary steering task of disturbance compensation is

planned path

Figure 6.7. Primary steering task: path tracking

assigned to the automatic control system . More realistic than a point mass, the vehicle
must be modelled as a rigid body with a moment of inertia J, represented here by
a second mass mR , rigidly connected to mDP, see Figure 6.8. An essential variable
to describe the yaw motion around a vertical axis through the CG is the yaw rate
r . Safety-critical situations arise if a disturbance torque M z D acts on the vehicle, for
exampl e, by side wind, asymmetric road-tire contact (p,-split braking}, tire burst. An
automatic control system can do this disturbance compensation faster and more pre-
cisely than a driver , who is subject to some reaction time and whose late compensation
action is frequently an overreaction. The disturbance torques enter directly into the
rotational differential equations Jf = E torques. Their indirect influence on the lateral
acceleration ayDP via the vehicle dynamics should be compensated such that the driver
controls the undisturbed ayDP . In system theoretic terms, the task separation requires
to make the yaw rate r non-observable from ayDP . This unilateral decoupling must be
robust for all operating conditions.
Usual vehicle parameters are mass m, moment of inertia J and location of the CG.

1 1 m

[. , ] o
J
(6.2 .1)
178 6 Case Studies in Car Steering

I
I
I
I
I
I
I
I
I
I
I
I
.,10(
I
I

Figure 6.8. Secondarysteering task: disturbance compensation

The bar-bell model of Figure 6.8 is now attached to the vehicle geometry of Figure 6.9.
The mass mR is fixed at the rear axle. The location i DP of the decoupling point is
calculated from the requirement, that its lateral acceleration ayDP does not depend on
the rear axle lateral force F yR . (Note that all axle forces are the sum for the left and
right wheel).
The lateral acceleration at the decoupling point is

(6.2.2)

with

ayCG = (FI/R + FI/F + FI/D)/m

i = (FI/FiF-FI/RiR+MzD)/J
and by substitution

(6.2.3)

The factor of FYR is zero for

i DP = J/mi R (6.2.4)
and with this choice

(6.2.5)
6.2 The Two St eering Tasks 179

O-yCG

I
I
I
I
1«
I
I
1«
I

Figure 6.9. Definition of the decoupling point

where f = f R + f F is the wheelbase.

For simplification, the lateral disturbance force Fy D at the CG is set to zero. In
a driver support system, the effect of FIID is compensated by the driver within their
normal lane keeping task, while the effect of the disturbance torque M z D is an undesired
yaw rate that should be compensated by an automatic control system .
The desired robust decoupling effect is achieved by cancellation of the effect of r in
the argument of
FyF({JF,8F) = f(8 F - (3F) cos 8F. (6.2.6)
For this purpose, we need a dynamic model for the dependency /IF(r) .

Remark 6.2

The reader may wonder why we did not begin with the dynamic model for /IF(r) .
The reason is that there are many such models. The single-track model will be intro-
duced in the next section. A more detailed model uses 14 degrees of freedom (six for
the chassis, four for the wheel rotation, four for suspension hub) plus the tire model
and steering actuator dynamics. More detailed multibody models also consider several
masses of the suspension and steering system and car body elasticity. Also, aerodynamic
forces have an influence on the yaw motion. Such models are useful for simulations if all
180 6 Case Studies in Car Steering

the parameters are known. They are, however, not suited to controller design, because
a higher complexity of the relation t3F(r ) makes it less realistic to implement the exact
cancellation of r . If we have to reduce the controller to the essential and measurable
effects anyway, then it suffices to model only these effects.
o
We emphasize that the robust decoupling concept is not restricted to certain model
simplifications (single-track model, constant velocity, linear tire characteristics, tire
dynamics) . Therefore, it is the general basis for the separation of the driver task of
path tracking via ayDP from the automatic control of the yaw rate. It is structurally
robust with respect to the uncertain road-tire contact and braking and acceleration
effects.

6.3 The Non-linear Single-track Model and its

Robust Unilateral Decoupling

The essential features of car steering dynamics in a horizontal road plane with the
CG projected to this plane are described by the single-track model by Riekert and
Schunk [172] . It is well described in the automot ive literature, e.g. [150], [204] . We
present it here from a robust control point of view. The possibility of additional rear-
wheel steering (4WS = four-wheel steering) is considered , which makes it an interesting
MIMO example . Otherwise, a simple version is presented for clarity of the essential
effects.

Remark 6.3

In the English language th e term bicycle model is common . This term may be
misleading; the dynamics of balancing a bicycle are much more complicated. Therefore ,
we prefer the term single-track model. 0

The single-track model is obtained by lumping the two front wheels into one wheel
in the center line of the car, the same is done with the two rear wheels. Thereby, the
car model of Figure 6.9 is reduced to that of Figure 6.10, and the coupling with roll,
pitch , and heave motions is not modelled. The angles bF and bR are the front and rear
steering angles. The distance between the center of gravity (CG) and the front axle
(resp. rear axle) is f F (resp. f R ) and the sum f = f R + f F is the wheelbase.
The angle /3 between the vehicle center line and the velocity vector v at the CG is
called the vehicle sideslip angle. In the horizontal plane of Figure 6.10, an inertially
fixed coordinate system (xo, Yo) is shown together with a vehicle fixed coordinate system
(x, y) that is rotated by a yaw angle ib. In the dynamic equations, the yaw rate r := ~
6.3 The Non-lin ear Single-Track Model and its Robust Unilateral Decoupling 181

I Y
I I I
I I I
I
I
I
I
I
I
Yo

:.
I

I
€R .:.
I

I
€F .1
I

I
I
X

Xo

Figure 6.10. Single-track model for car steering

will appear as a state variable . The yaw angle 7/J itself will be included in the model in
Section 6.8 in the context of automatic car steering, where the position of the vehicle
relative to the lane is considered . For a driver assistant for yaw stabilization, it suffices
to model only the velocities as the path tracking task is left to the driver by visual
feedback. The velocity vector v has the absolute value v = Ivl and the components V x
an d "v in chassis coordinates.
Th e forces transmitted between the road surface and the car chassis via th e wheels
are represented in Figure 6.10 by th e side forces Fyt F and F yt R . The forces in the
longitudinal direction of the tires are assumed to be zero, i.e. the wheels are freely
spinnin g.

Figure 6.11 shows a block diagram of the model. Input to th e vehicle dynami cs are the
lateral forces at the front and rear axles:

Fyt F COSbF , (6.3.1)

FYt R COS b R ,

and a longitudinal force component

(6.3.2).

Note that th ese axle forces represent the sum of th e forces at the left and right tire.
Via the dynamics model, the forces control state variables (3, v , r . The equations of
motions for three degrees of freedom in th e horizontal plane are :
a) Lat eral motion

m v{ /3 + ~ ) cos (3 + m il sin (3 = FyF + Fy R . (6.3.3)

182 6 Case Studies in Car Ste ering

MZD

F OF FytF
! I FyF)
1 {J
Tire ~ cos OF
- t;

r
FYtR Dynamics ~
R OR FYR
-
model ~ cos OR IV ~
~

{JR

{JF Kinematic

Figure 6.11. Block diagram of car steering

b) Longitudinal motion
-mv(~ + "j;) sin {J + mi: cos {J = Fx · (6.3.4)
c) Yaw motion
J;j; = FyFf F - FyRt R + M z D . (6.3.5)
With r := '0, we obtain from (6.3.3) to (6.3.5)
mv(~ + r) - sin{J cos{J 0 F.T
mv cos {J sin {J 0 FYF + FYR (6.3.6)

o o 1

In the next step, the sideslip angles OF and OR at the front and rear tires are obtained by
a kinematic model from the steering angles OF , OR, and from the state variables {J, r, and
v . Figure 6.12 illustrates the vehicle motion around a momentary pole MP. The local
velocity vectors in front (VF) and rear (VR) and at th e CG (v) are oriented perpendicular
to the connecting line to the momentary pole. The front and rear chassis sideslip angles
are {JF and {JR' The velocity components in the direction of the longitudinal center line
of the vehicle must be equal, i.e.
(6.3.7)
The velocity components perpendicular to the center line depend on the yaw rate r , as
VF sin {JF = v sin {J + eFr,
(6.3.8)
6.3 Th e Non-linear Single-Track Model and its Robust Unilateral Decoupling 183

Figure 6.12. Kinematics variables

The velocity terms V F and VR are eliminated by division by the corresponding terms
from (6.3.7). Thus , the kinemati c model is

- Vsin 13 + eFr - 13 ~
tan 13F -
V cos
13 - tan + V cos 13 '
(6.3.9)
v sin 13 - eRr 13 eRr
tan 13R = = tan - - -
v cos 13 v cos 13 '
and the tir e sideslip angles (shown in Figure 6.5) are

(6.3.1O)

The feedback-structured model of Figure 6.11 is now completed by the non-linear tire
model:

FYt F (6.3.11)
Fyt R

Linearization for Constant Velocity and Small Angles

All four blocks in Figure 6.11 are non-linear. In normal driving situations (except
slow parking maneuvers) , the most important non-linearity is the uncertain tir e model.
Therefore , we linearize the three other non-linearities first, and then derive the feedback
184 6 Case Studies in Car Steering

control law for robust decoupling without further assumptions on the non-linear tire
dynamics. Later, also the static tire characteristics are linearized to allow for a linear
analysis of the uncontrolled and of the robustly decoupled cars.

i. The vehicle dynamics (6.3.6) are non-linear. They may be linearized by the as-
sumption i.J = O. This is justified, because the velocity v is changing more slowly
than the state variables rand fJ. The velocity v is now treated as an uncertain
constant parameter. Also, the force component Fx sin fJ is neglected . Then, the
linearized version of (6.3.6) is

(6.3.12)

ii. The chassis sideslip angles fJ, fJF and fJR are small, i.e. cosfJ = 1 in (6.3.12), and
(6.3.9) becomes

fJ + eFr/v, (6.3.13)
fJ - eRr/V.

iii. The steering angles OF and OR are small, i.e. cos OF = 1 in (6.1.7) and COSOR =1
for the rear wheels. Then,

FyF(OF), (6.3.14)
FyR(OR) ,

are the unknown tire characteristics. Equation (6.3.12) now reads

(6.3.15)

Under feedback control according to Figure 6.3, the steering angle OF is the sum of
the driver command Os and a corrective angle oc, generated by the feedback system .
The relationship between os,oc and OF , fJF , as shown in Figure 6.5, is illustrated by
Figure 6.13.

For feedback control , the angle

, := fJF - Oc (6.3.16)

is of particular interest. fJF is a plant state, and the feedback signal Oc is composed of
plant and /or controller states, thus I is continuous and may be used as a state variable
6.3 Th e Non-linear Single-Track Model and its Robust Unilateral Decoupling 185

Figure 6.13. Front wheel angles

of t he closed-loop system. Without control, 8c == 0" is t he (front axle) slip angle fiF.
With contr ol, "t is the slip angle correction by control , in short, t he corrective slip angle.
Similar t o the crane example in Chapter 1, we will derive a feedback structure for the
car. A first requirement is t hat it is robust with respect to the uncertain non-linear t ire
characterist ics FyF(QF) and FyR(QR). It will be shown that a robust decoupling of the
two steering tas ks is feasible. As described in Section 6.2, these are t he path tracking
t ask of the driver, and the aut omat ically controlled yaw stabilization and disturbance
compensat ion. This generic decoupling controller separates t he two tas ks, such t hat
th ey can be improved by independent feedback loop designs.

Robust Unilateral Decoupling Controller

By (6.2.5), with FYD = 0 and FyF(fiF, 8F ) = FyF(8s + 8c - fiF), the lat eral acceleration
ayDP is decoupled from th e yaw motion, if
(6.3.17)
does not depend on r , where r enters through the differenti al equat ion as fiF(r ). If we
assume th at fiF can be measured, then
(6.3.18)
is the ideal decoupling control law, yielding Q F = 8s , and , == O. The driver directly
commands the t ire slip angle Q F and ayDP is decoupled from r .
Since fiF is not available for feedback , we use a backstepping approach, i.e. th e
compensation of r is done at the input of an integrator with inpu t 6c . This additional
integrator must be stabilized by closing the loop , such t hat , remains small.
186 6 Case Studies in Car Steering

Theorem 6.4 (Robust unilateral decoupling, Ackermann)

The feedback control law
.
Oc = rref -
[f
r +
DP - fF ]
v T (6.3.19)
makes the yaw rate r non-observable from the lateral acceleration ayDP at the
decoupling point .
o

Remark 6.5

Practically, the control law (6.3.19) is turned on softly for v > Vi > 0, such that no
division by zero occurs. 0

Laplace transformation of (6.3.19) yields

oc(s) = rref(s) - r(s) _ f DP - f F r(s) . (6.3.20)
s v
The decoupling feedback controller is illustrated by Figure 6.14.

IMZD

G
Os OF

Oc
- 0;- r

1 f DP - f F V
-
S v

rref

Figure 6.14. Robust unilateral decoupling by feedback of r to Oc

The generation of a reference yaw rate Tre] will be discussed later. The steering com-
mand Os will also enter there. Such a prefilter will be used to shape the response of
ayDP and r to a steering wheel input. This feedforward path has no influence on the
robust decoupling concept in the closed loop. We first prove the robust decoupling
property and stability of the two resulting subsystems.
6.3 The Non-linear Single-Track Model and its Robust Unilateral Decoupling 187

Proof of Theorem 6.4

Substitute J = mfRf DP from (6.2.4) in (6.3.15)

(6.3.21)

where OF = 6s + 6e - (3 - fFTIv and OR = 6R - (3 + fRT [u , and close the loop by

the control law (6.3.19).
In a classical approach, we would solve (6.3.19) and (6.3.21) for ~, r and be. A
formal proof would then proceed with a transformation to the state vector

(3

T -0 1 o T (6.3.22)

It is then, however, not clear, why this state vector was chosen. In order to
introduce a general approach that is applicable to other plants as well, we use a
different line of proof. As was discussed for the crane example in (1.6.11), it is a
basic rule of modelling for robust control to keep the uncertainty in a single place
in the model, and not to "smear" it over several model coefficients.
In our case of th e car, the dominant uncertainty is in FyF(OF) and FyR(OR) .
Therefore , solve (6.3.21) first for these quantities:

mfR[v(~+T)+fDpr] = FyF(OF)f+MzD, (6.3.23)

m[vfF(~ + T) - fRfDpr] = FyR(OR)f - M zD. (6.3.24)

Substitute

T = Tr e] - be - (f DP - fF)i/v

from the control law (6.3.19) into (6.3.23), to obtain

mvfR[(~ - be) + fFrlv] = FyF(OF)f + M ZD - mVfRTreJ. (6.3.25)

The quantity in square brackets is "r, i.e,

(6.3.26).

as defined in (6.3.16). Then, with OF = 6s - "

"r = [FyF(6s -,)f + MzDllmvf. R - Tref - (6.3.27)

This first order differential equation for , does not depend on the state variable
T,it is unilaterally decoupled. The decoupling effect is robust with respect to
188 6 Case Studies in Car Steering

the uncertain mass m and velocity v and , most importantly, with respect to the
uncertain lateral forces F y F and FYR at the front and rear axles. The only quantity,
that must be known for the implementation of the control law (6.3.19) is £DP ' It
depends on the longitudinal mass distribution, which is constant during a ride of
the car .
The lateral acceleration ayDP at the decoupling point is, by (6.2.5),

(6.3.28)

and , by comparison with (6.3.27),

(6.3.29)

The decoupled lateral subsystem (6.3.27) is illustrated by Figure 6.15.

Figure 6.15. Lateral dynamics subsystem with output ayDP

Also, ayDP is robustly decoupled from the influence of r. This shows that the
yaw rate r is non-observable from the lateral acceleration ayDP at the decoupling
point.
o

The Decoupled Yaw Subsystem

For the rear-wheel subsystem also, by analogy with (6.3.16), the state variable

(6.3.30)

is chosen, then
6.3 The Non-linear Single- Track Mod el and its Robust Unilateral Decoupling 189

and with /J from (6.3.24),

(6.3.31)

The differential equation for r follows from the second row of (6.3.21)' as

(6.3.32)

Fy F is a coupling term from the lateral subsystem of Figure 6.15. From a practical
point of view, however, the quantity ayDP is better suited as a coupling term, because
it may be measured by an accelerometer, and also in view of the task separation as
illustrated by Figures 6.7 and 6.8. Therefore,

from (6.3.28) is substituted in (6.3.32), yielding

(6.3.33)

This i is now substituted in (6.3.31) and both equations are combined to the state
space model of the yaw subsystem

Figure 6.16 illustrates the yaw subsystem with the coupling input ayDP from the lateral
subsystem.

The signal flow diagram of Figure 6.17 illustrates the unilateraly decoupled system.
The lateral dynamics may be modified by accelerometer feedback of ayDP to front-
wheel steering 85 . If rear-wheel steering 8R is available, then feedback from r can be
used to improve the yaw dynamics. Also, a feedforward path form 85 to 8R can be
implemented, for example, to shape the response in the yaw rate r , or the sideslip angle
{JR , to a steering wheel input. The above additional control structures do not destroy
the unilateral robust decoupling property. Control system structures will be discussed
further after the linearization of the tire characteristics.
190 6 Case St udies in Car Steering

M ZD
.-- ' - -
f R (f DP - fF) 1 fF
8R f-:-:--
v ef DP f IPf DP
-.-- '-- .--

...-- ...--

~
/JR 1 f3R D:R. FYR - 1 r 1 r
- mf DP
-
s s
'--- '---

f DP + f R
V

Figure 6.16. Yaw dynamics subsystem

Rema rk 6.6

Note that the control law (6.3.19) was derived structura lly and its robustness does
not depend on any parameter values (except the decoupling point location f DP - f F ,
which depends only on the mass distribution and is const ant during a ride) . Compare
this with a classical approach, where the lateral accelerat ion ayCG at the CG would be
decoupled from r . Since thi s is not exactly possible, the influence of T on a yCG can be
only minimized by numerical optim izat ion for specific vehicle data. The transparency
of the structurally robust decoupling controller (6.3.19) would th en be completely lost.
o

Rem ark 6.7

1 ayDP
1
8s Lat eral dynamics Yaw Dynamics
~
State "I = f3F - 8c St at es f3R , T
8R

Figure 6.17. The yaw rate T is not observable from the lateral acceleration ayDP at the
decoupling point
6.3 The Non-linear Single-Track Model and its Robust Unilateral Decoupling 191

The assumption of small sideslip angles is not really needed for the derivation of a
robustly decoupling control law. In [9], [ll) this assumption is dropped . Then

where VyF and V x are the lateral and longitudinal velocity components of ii.
Instead of the control law (6.3.19), there results
2
-' [ ( f DP - f F) cos? f3F 'J
uc=rreJ- r+
Vx
r + f Fr V - ax . f3 f3
sin FCOS F , (6.3.35)
x

where ax is the longitudinal acceleration.

The implementation of (6.3.35) requires, however, an accelerometer for ax and a
reconstruction of f3F in an observer. It is, therefore, eventually simplified to the form
(6.3.19), which is used here from the beginning in order to simplify the proof. The
last non-linear term in the controller (6.3.35) shows that hard braking produces the
maximal value of the term €Fr 2 - ax. This situation should therefore be considered in
simulations. A simulation study for a limousine [178) indicated that the last term in
(6.3.35) has only minor influence on the closed-loop behavior.
o

Stability of the Decoupled Lateral and Yaw Subsystems

The robust decoupling control concept is practically useful, only if the resulting two
subsystems are stable, or can be stabilized separately without destroying the decoupling
effect.
For no external inputs, M zD = 0, Os = 0, rreJ = 0 and a symmetric tire character-
istic, the lateral subsystem is

The Lyapunov function V = ,2 yields

v = 2,7 = -2, F(r)fjmv€R. (6.3.36)

The condition ,F(r) > 0 is satisfied for tire characteristics, see Figure 6.6 Thus , V < 0
and the decoupled lateral subsystem is stable in forward driving (v> 0).
Alternatively, a Popov test may be applied. The Nyquist and Popov plots of the
linear part in Figure 6.15 have 90° phase angle, therefore the upper bound for the sector
of absolute stability is infinity.
For the yaw subsystem of Figure 6.16, the phase angle is below 180°, therefore
ideally the same conclusion applies. Here, however, we are much closer to the stability
boundary. Additional phase delay, e.g. by actuator dynamics or a delayed build-up
of the lateral forces FYF and FYR may cause instability. If the car is equipped with
rear-wheel steering, then a stabilization by feedback from r to OR is easy. In Section
6.6, alternatives to rear-wheel steering are discussed.
192 6 Case Studies in Car Steerin g

Experimental Verification
The effects of the control law (6.3.17) were experimentally tested on a BMW 735L The
data of the linearized model are given in Table 6.1. It turns out th at the decoupling
point is very close to the front axle, the difference EF - ED P is only 1 % of the wheelbase
E. Therefore, the simplified control law

(6.3.37)

was used. The test car is equipped with a steer-by-wire system such that the sum
Js+Jc = J F (see Figure 6.13) is formed electrically and the steer ing actuator moves the
entire steering angle JF . The generation of rre! from Js and v will be discussed later; the
emphasis of the tests was on the rejection of disturbance torques M z D ' Experimentally,
MZD was generated:
L In a side wind test by passing a series of lateral fans.
ii. In a /-L-split braking test, with the right tires on a slippery surface (water flooded
tiles) and the left tires on a high-grip surface .
The driver just had to keep the handwheel straight and L pass t he side wind machine,
or ii. apply full brake on the /-L-split test track, i.e. Js == 0 and Tre] == O.
Figures 6.18 and 6.19 show snapshots with 1 second sampling time . The left column
is for the conventional vehicle, the right column for our robust decoupling control law.
Both experiments show that the control system brings the car back to. the original
heading within one second.
The robustness of the control system was demonstrated in extensive simulations.
For more details, the reader is referred to [18) .
Alternative car steering concept s are usually more oriented towards parameter and
disturbance identification (e.g. for the road-tire friction coefficient, tire pressure, side
wind) , i.e. they look at the causes for M z D . It is, however, difficult to foresee all
possible causes of skidding, and to develop specialized sensors for each of them. The
robust control concept looks at the effect of disturbances M ZD of unknown origin. This
effect is seen in the error signal rre! - r .

6.4 The Linearized Single-track Model

In order to allow for a linear analysis of the car, the lateral tire forces are now linearized
about a zero tire sideslip angle as

FyF(Cl:F) = /-LCFCl:F , (6.4.1)

FyR(Cl:R) = /-LCRCl:R ·

The slope CF resp. CR of the tire characteristic is called corn ering stiffness. Note that
(6.4.1) applies to th e front or rear axle, i.e. to the sum of left and right wheel at the
respective axle. (Note that some authors use 2CF' or 2CR', where CF' or CR' refer to the
6.4 The Linearized Single-Track Mod el 193

Figure 6.18. Side wind test at v = 120 [krn/h] and wind velocity 80 [km/h]; left: conventional ,
right : decoupled
194 6 Case Studies in Car Steering

Figure 6.19. p-split-braking at 80 [km/h] ; left: conventional , right : decoupled

6.4 Th e Linearized Single-Track Model 195

individual tir e.) The friction coefficient j1 ~ 1 is assumed to be the same at the front
and rear wheels.
Typical values of the friction coefficient j1 are

j1 1 dry road

j1 = 0.5 wet road

j1 0.15 ice

Now, solve (6.3.15) for /J and i , and substitute the linearized tire characteristics Fy F
and Fy R :

/J ] [ J!:...-(CFCtF + CRCtR) - r ]

~ 5~CFtFOF ~M'D
(6.4.2)
[ i: - CRtRO R) + '

and substitute (6.3.10) and (6.3.13)

This linear state space model is written as

au -j1(CR + cF)/mv ,
a12 = -1 + j1(cRf R -,cFf F)/mv 2,
a21 j1(c RfR - cFfF)/J,

a22 = -j1(c RfR 2 + cFf F 2)/ Jv ,

bu j1CF/mv ,

b12 = j1CR/mv ,

b21 j1 cFfF/ J ,

~2 -j1 cRfR/J,

bD = I/J.
196 6 Case Studies in Car Steering

The lateral acceleration at the decoupling point is by (6.2.5) , (6.3.13) and (6.4.1) .

ayDP = p,CF(J:Felme R (6.4.5)

p,cFe M zD
= -(OF - (3 - eFTlv) + -.
meR meR

This classical form of the single-track model contains the four uncertain parameters
p" m, J, and v. These parameters may be reduced to two by the introduction of the
decoupling point location, see Figure 6.9 and Equation (6.2.4) :

(6.4.6)

It is invariant under changes of p, and v, and depends only on the mass distribution in the
vehicle. Vehicles are normally designed such that there are no significant differences
in eD P between the empty and the full car (See the data for the City Bus 0 305 in
Table 6.1). An extreme case is a heavy load in the trunk of a lightweight car, it
increases the ratio Jim and reduces eR . At least the parameter eD P does not vary
much during a ride, and eD P would have to be identified only once after starting the
car.
Substitute J = meReD P in Equation (6.4.3) or (6.4.4) , then the uncertain parameters
p, and m appear in their ratio ufm. A reduced p, on a slippery road has the same effect
as an increased mass m. Rescale the friction coefficient p, and the mass m by

, (6.4.7)
m

where iii. is fixed as iii. = m: , the mass of the unloaded vehicle (including only a light-
weight driver) . Then only one parameter

m m
ii=p,-, - ~ 1 (6.4.8)
m m

is needed, where ii E [p,-m- [m , 1] .

The four parameters p" m , J , and v have thus been replaced by two parameters

(6.4.9)

Then, (6.4.4) becomes

6.4 Th e Linearized Single-Track Model 197

all -ji(CR + cF)/mv ,

a12 -1 + ji,(cRfR - cFf F )/ mv 2,

a21 -ji,(cRf R - cFf F)/ ih f Rf DP,

a22 - ji(cRt k + cFf~)/mvfRfDP ,

bll ji,cF/mv,

b12 ji,cR/mv ,

b21 ji,cFfF/mfRfDP,

b22 -ji,cR/mf DP,

bD l/mf RfDP.

Without loss of generality, the notation may be simplified by omitting the tilde on p,
and m and observing th at the mass of the empty vehicle is substituted for m . The
actual mass only appe ars in bo -
A typic al operat ing domain for v and p, is shown in Figure 6.20.

v
v

Figure 6.20. Operating domain for car steering control system

At v = 0, the vehicle is obviously not controllable, therefore specifications for t he

controlled vehicle can be met only for v 2': u: > O. At the minimum velocity, th e
controller must guarantee good performance for the whole range from p,-(fullload , icy
198 6 Case Studies in Car Steering

road) to /l = 1 (empty vehicle, dry road). The maximum velocity v+ can be driven
only under reasonably good road conditions. The friction coefficient /l is a very critical
uncertain parameter, because it can vary suddenly when the car drives on an icy bridge
or on an icy road in the shade of a forest . There is no time to identify /l and a
steering control system must be robust with respect to p: Numerical values for the
fixed parameters are given in Table 6.1 for a city bus and for a high-speed limousine.
Note that the limousine has an almost ideal mass distribution , for which the front mass
would be at the front axle, i.e. f D P = f F . The difference is only 1 % of the wheelbase.
The city bus has a relatively short wheelbase for better maneuverability at low speeds.
Then, the decoupling point is in front of the front axle. It is remarkable, however, that
f D P does not vary between an empty and full vehicle.

Table 6.1. Numerical values for two vehicles

Parameter Daimler Benz BMW

City Bus 0 305 735 i

empty full

m 9950 [kg] 16000 [kg] 1864 [kg]

J 105700 [kgm2 ] 171300 [kgm2 ] 3654 [kgm'']

v+ 20 [m/s] 20 [m/s] 70 [m/s]

CR 470000 [N/rad] 470000 [N/rad] 213800 [N/rad]

CF 198000 [N/rad] 198000 [N/rad] 101600 [N/rad]

fR 1.93 [m] 1.93 [m] 1.32 [m]

fF 3.67 [m] 3.67 [m] 1.51 [m]

J
f DP = - - 5.50 [m] 5.55 [m] 1.48 [m]
mfR

f = fR+f F 5.60 [m] 5.60 [m] 2.83 [m]

f F - f DP
-0.33 -0.34 0.01
f
6.4 The Linearized Single- Track Model 199

Stationary Cornering
Consider the steady state (/3 = 0, i = 0) for the inputs t5F = t5Fstat , t5R = 0, M zD = O.
From (6.4.4), follows

(6.4.11)

-1
{3stat]
[ r stat
= [ a22 -a12 ] [ bb ll ] t5Fstat = [kk{3stat ] t5Fstad 6.4.12)
-a21 all 21 rstat

The evaluation yields

2
k _ J.I.CR£R£ - £Fmv
(6.4.13)
{3stat - J.I.CR£2[1 + (v/vcHF]'

v
(6.4.14)
k
rstat
= £[1 + (v/vcHF]'
with the characteristic velocity VCH .

(6.4.15)

For the empty bus, the characteristic velocity at J.I. = 1 is VCH = 40.3 [m/sec], for the
limousine VCH = 26.9 [m/sec].
The steady state gain krstat will be used later for generating a reference yaw rate rreJ .
For the lateral acceleration ayDP, the steady state gain is, by (6.4.5),

kayDPstat

(6.4.16)

Note that in stationary cornering the lateral acceleration is the same at all points,
otherwise i would not be zero.
Note also that in stationary cornering

V= Rr, (6.4.17)

where R is the curve radius as illustrated by Figure 6.21. The centripetal acceleration
is

(6.4 .18)
200 6 Cese Studies in Cer Steering

Center of curve

Figure 6.21. Relation between curve radius R, yaw rate T and velocity v

It is mathematically inconvenient that the curve radius goes through infinity for a
change from left to right turn. Therefore, introduce the curvature

e= I/R, (6.4.19)

with positive sign in a left curve. Its steady state is

1
estat = rstat/v = £[1 + (v/vcHF]Ostat. (6.4.20)

At low velocities, the driver commands the curvature estat = Ostat/ £; at high velocities
the lateral acceleration kayDPstat = OstatVCH 2/£ is commanded from the steering wheel.
At the characteristic velocity v = VCH , the driver commands essent ially the yaw rate
rstat = VCH /2£.
Th e steady state effect in curvature, yaw rate and lateral acceleration as a func-
tion of velocity is shown in Figure 6.22. In stationary cornering, the lateral force (or

Pstat aystat

.. . . . . . ... .. . . VbH/£
1
e
. ...... . ... .. . .. .. V~H/2f

o v

Figure 6.22. Steady state curvature, yaw rate and lateral acceleration as a function of velocity

acceleration) must balance the centrifugal force (or acceleration) .

Therefore,
_ 2
aystat - v estat, (6.4.21)
6.4 Th e Lin earized Single-Track Model 201

which agrees with (6.4.18).

For a given curve parameter , the required steering angle is

e[I + (VIVCH)2] lJstat

ell + (V IVCH)2]rstatlv (6.4.22)
e[1 + (Vlvc H)2 ]aystatlv2.
Note th at VC H is prop ortional to the uncertain friction coefficient u, see (6.4.15).

Rema rk 6.8

Th e first aut hor is frequently asked about his relationship with th e Ackermann angle
and Ackermann st eering. Accord ing to a footnote in [150], th e older Ackermann was
an English coach-builder, who filed in 1917 an English pat ent for Lankensperger, th e
inventor of th e knuckle steering. Nothing is known about a kinship.
For slow drivin g, e.g. in a parking maneuver, th e tire slip angles O F and OR are zero,
i.e, in Figure 6.12 th e momentary pole is locat ed on a line th rough the rear axle. Th e
steering angle 6F is th en calculated from t he tri angle of Figur e 6.23 as

sin6 F = eiR.
6F is called the Ackerm ann angle.

Th e Ack ermann steering is also designed for slow dri ving wit h zero sideslip angle. It
takes into account that in a curve t he inner front wheel is running on a smaller radius
t han the out er front wheel. Today, steering systems are built with almost par allel
deflection of t he left and right front wheel [1 50]. 0

Poles and Zeros

We first analyze th e open-loop system (6.4.4) for eigenvalues and transfer functions
from inputs bF,6R and MzD to r , and ayDP . The characterist ic polynomial of th e
syst em matrix in (6.4.4) is

(s - all)(s - a22) - al2a2l

(6.4.23)
ao + als + S2 ,
with t he parameter-dependent coefficients
202 6 Case Stu dies in Car St eering

MP

i + - - - f -----+i

Figure 6.23. Ackermann angle OF

The uncertain parameters ql = J.l and qz = l/v enter polynom ially int o t he coefficients
of the characteristic polynomi al

(6.4.24)

We continue here to use th e uncertain physical parameters J.l and v, rather than ql and
q2'
The syst em is stable for ao > 0, al > 0. Two cases must be distinguished:
a) Understeering: cni n - cFf F ~ 0, t he system is stable.
b) Overst eering: CR( R - cFf F < 0, the system is stable if th e cri tical speed ve« is not
exceeded , where
J.lCFc 2 Rf
v2 •- ----,,.--:-..,,..:-....:.:..._.,....-,- (6.4.25)
crit .- m(cFfF - cRf R) '

Typic al road vehicle data correspond to the understeering case, see Table 6.1.
The characteristic polynomial PA{S) may be written in terms of damping D and
natural frequency Wo as

(6.4.26)

where Wo and D depend on the physical parameters as follows:

2( ) _ J.l 2CFCRf 2 + mV 2J.l(cRfR - cFfF)

Wo u,v - 2~ o 2 ' (6.4.27)
m {.R{.DPV
6.4 The Linearized Single-Track Model 203

D(v) = + CF) + CReR 2 + CFeF 2

eReDP(CR
(6.4.28)
2JeReDP[cRCFe2 + mv 2(C ReR - CFeF)/J.L]

For large velocities v, the damping D goes to zero and the eigenvalues of understeering
vehicles approach the imaginary axis at

J.L(CRe R - CFe F)
Woo = lim Wo = (6.4.29)
v-oo meReDP

With decreasing velocity v, the damping D becomes larger than 1, i.e. the pair of
complex eigenvalues meets at the real axis and branches into a pair of real eigenvalues.
For v -> 0, two real eigenvalues go to minus infinity. Figure 6.24 illustrates the velocity
dependence of the yaw eigenvalues. Figure 6.24 is not a standard root locus because

jc.J

v--+oo

Figure 6.24. Velocity dependence of yaw eigenvalues

the parameter venters non-linearly. At some velocity vm , one of the real eigenvalues
has a maximum .

Transfer Functions
For the calculation of various transfer functions, first determine

(6.4 .30)

where
hll(s) S - a22 s + J.L(CRe R 2 + CF'-F 2)/mVeReDP,
h 12 = a12 -1 + J.L(CRe R - CFeF)/mv 2,
h21 a21 J.L(CRe R - CFeF)/mlR'-DP,
h22(S) s-all s + J.L(CR + cF)/mv.
204 St eering
6 Case Studies in Car Steering

Outputs of particular interest are the

th e yaw rate
rat e r and the
the lateral
lat eral acceleration
accelerat ion at the
th e
decoupling point.
point . By (6.4.5),
(6.4.31)
with Cl == -J-LcFf/mi R , d = 1jme
-j.LC F ej m1R, l/mf R.
R.
The
The six transfer functions
function s of the system 8F ,8tJRR and M
syst em (6.4.4) with inputs tJF' Mz D and
outputs rand ayDP are

r( s) ]
[ ayDP(s)

= PA~S) { [ ~
(6.4.32)

The
The transfer functions are written as

(6.4.33)

with the numerator terms

J-LCRi) J-tcFi F
nFr(S) ( S +--
mvi miRi DP'
F
_(S + /JCFf)
mvfR
/JeR
mf DP'
. J-L(CF+CR)) 1
( S+ ,
mv mfRi DP

(
S2 +S J-LCR(fR + f DP) +J-LCR)
-- - J-LcFi
-
mvf DP mi DP mi R'
2CFC
- ( s+ -v- - ) J-L
i DP - f F
Ri(iDP - iF)
m 2vi Ri DP ,

(
S
2
+s J-tCR(fR + fDP) +J-LCR)
- - --.1
mvf DP mi DP mi R
The numerator of the transfer function from front-wheel steering to the sideslip angle
{3 is

nF/3(s) = [ s+ - V (J-lCRi
- - - -iF)] -
J-lCF. (6.4.34)
i DP mv 2 f R mv

Controllability
For front-wheel steering of the car (6.4.4), the controllability matrix is

[b Ab]
[b = [b1lll allb
Ab] = bil
ali ll + aI2
a12
b21 ] ,
bl1
a21 b
b21 a2l ll + a22 b21
6.4 The Linearized Single-Track Model 205

and its determinant is

2 2
det[ b Ab] = 3 ~fc~f 2 [J.tCRUR(f DP - f F) + f}mv 2) . (6.4.35)
m v R DP

Controllability is lost for J.t = 0 and at a speed V nc, where

2 J.tCRUR(f F - f DP)
V ncF = n2 (6.4.36)
m<-F

If J > mfFf R, i.e. f F - f DP < 0, then VncF is imaginary and of no practical interest.
Otherwise non-controllability occurs at a real velocity. Substituting v 2 = V~cF into ao
of (6.4.23) and assuming f F - f DP > 0, ao > 0 is obtained. Since also al > 0, the
uncontrollable modes are stable. Thus, the steering dynamics are robustly stable, but
not robustly controllable.
For the bus of Table 6.1, we have f F - f D P = -1.73 [m], i.e, there is no real V nc '
For the limousine, f F - f D P = 0.03[m]. For J.t = 1, the controllability is lost at a
velocity VncF = 2.38 [m/sec] (8.6 [km/hj} . Typically, the controller is turned on above
this velocity. The non-controllable case is relevant for cars with a small moment of
inertia (e.g. central engine) or relatively long wheelbase (e.g. Mini type). If VncF is in
the operating domain of the controller, and a pole placement approach is used , then
the non-controllable stable eigenvalue must also be assigned to the closed loop .
For rear-wheel steering of the car (6.4.4), the controllability matrix is

det[b Ab] = 32/

m v
J.t2c
f2
R DP
2
[J.tcFf(f F - f D P) + f Rmv 2) . (6.4.37)

Controllability is lost at a velocity Vn cR, where

J.tcFf(f D P - f F)
V~cR = mf R
(6.4.38)

It has the opposite sign as Vn cF , thus here the case of large moment of inertia or relatively
short wheelbase is critical. For the data of the empty bus on a dry road VncR =10.28
[m/sec] (37 [km/hj} . Cars with four-wheel steering are controllable, because either Vnc F
or Vn cR is imaginary.

observabili ty
The determinant of the observability matrix for the output (3 is

det [
c
(/A] = det [1 0] all a12
c: a12 = -1 + J.t(cRf R - 2
c F f F )/ m v . (6.4.39)

In the understeering case ( cRf R - c F f F > 0) , the observability is lost for

(6.4.40)
206 6 Case Studies in Car Steering

The determinant of the observability matrix for the output T is

Here, observability is given independent of the velocity. Thus , from a robust observ-
ability point of view, the yaw rate T is more useful for feedback than the sideslip angle
{3. This fact coincides favorably with the fact that cheap yaw rate sensors are available
(see ESP application), whereas there is no cheap sensor for the sideslip angle {3. We
will assume robust controller structures with feedback of T. Note that the neutral steer
car with cRf R - cFf F = 0 is not observable from T. It is stable, however, because
ao > 0, a1 > 0 in (6.4.23).
By (6.4.31), the measurement of ayDP is related to both state variables {3 and T via
the uncertain friction coefficient u , such that no good reconstruction of {3 from ayDP
is possible. There are also practical disadvantages in the use of accelerometers. They
cannot be mounted on street level as assumed in the single-track model, therefore the
measurement includes components of the unmodelled roll acceleration. Also, it is gener-
ally not possible to mount the accelerometer above the decoupling point. Interpolation
between two accelerometer measurements circumvents this problem, but it increases
the cost .
From the above analysis of the choice of sensors, we conclude:

i. The yaw rate sensor will be used for feedback.

ii. The velocity is available from the ABS system and may be used for gain-
scheduling.
iii. The steering angle command 6s is measured at the handwheel (considering th e
gear ratio). It will be used together with the velocity to generate a yaw rate
reference TreJ.
iv. Also, the actual corrective steering angle 6c after the actuator (see Figure 6.3) is
a useful feedback signal , or alternatively the actual steering angle 6F = 6s + 6c .
It helps to achieve robustness with respect to unstructured actuator uncertainty.

6.5 Linear Analysis of Robust Decoupling

The second order plant (6.4.1O) is controlled by the first order controller (6.3.19). In
(6.3.27), linearize the front tire characteristic by FyF {6s - , ) = /-lcF{6s- r ) and introduce
the angle r (see Figure 6.13) as the third state variable.

e33rdec + h1 6s + h3 Mz D - Tr e] » (6.5.1)
-/-lcFf/mvf R,
/-lcFf/mvf R,
l/mvf R.
6.5 Linear Analysis of Robust Decoupling 207

Linearize the rear tire characteristic by FyR (OR - (3R) = j.tCR(OR - (3R) and substitute
from (6.5.1)

(6.5.2)

Then , (6.3.34) becomes

/J~dee
[ Tdee
] (6.5.3)

+ [ill
121

ell -J.tCR(lDP + lR)/mvlDP,

e21 j.tcR/mlDP,
e l3 - j.tcF(l DP -IF)/mvlDP ,
e23 -j.tcFlF/lRlDP,
ill j.tcF(lDP -IF)/mvlDP ,
hI = j.tcFlF/mlRlDP ,
122 -j.tCR/mi DP,
112 j.tcR(lDP + lR)/mvlDP ,
113 -l/mvlDP,
123 l/mlRlDP.
The linearized state space model of the robustly decoupled car is

[~=] ~ ['"
e21
-1
0 '"
e23 ] [ p-<
rdec ]

J[ M:
Idee 0 0 e33 Idee

!J2 il3 !J4 0

+ [ I"
121 122 123
~i
Os ]
D '
(6.5.4)
131 0 133
Tre!

r dec ]
[ ayDPdee =
[0 1
0 0
0 ] [{3Rdee]
Ve33 rdec
I dee
[0
+ VhI
0
0 ~] [t].
Tre]

The characteristic polynomial of the E matrix factorizes into

(6.5.5)
208 6 Case Studies in Car Steering

Plat S - e33 = S + -/-lC-Ft , (6.5.6)

mvER
Pyaw S2 - ell S + e21 (6.5.7)
2 j1CR(E DP + ER) JiCR
S + s +--
mvE DP mEDP
S2 + 2DdecwOdecS + W5dec ,
with

WOdec (6.5.8)

D dec (6.5.9)

Rem ark 6.9

An alternative approach for deriving th e closed-loop poles is to use the Laplace

domain description (6.3.20) for the controller <5c(s) = - Gdec(s )r (s )
Gdec(s ) = 1 + s(EDP - EF )/v =: ndec :
S S
Th en th e closed-loop characteristic polynomial is
p(s) = SPA(S) + nFr(s)ndec(S).
A symbolic factorization yields
p(s) = Plat(S)Pyaw(S) .
o

Mz D

rnDadecl I I nDrdec

<5s 1 ayDPdec 1 rdec

nSadec - nardec - -
Plat Pyaw

- InRrdec I
r!!!- <5 R

Figure 6.25. Structure of the transfer function model of the decoup led car
6.5 Linear Analysis of Robust Decoupling 209

In terms of transfer functions , several cancellations of Ptat (s) or Pyaw (s) occur. Th ey
are best illustrated by a linearized version of Figure 6.17 as shown in Figure 6.25. The
figure visualizes the main effects of robust decoupling:
i. Factorization of the characteristic polynomial into Ptat (s) and Pyaw (s).
ii. The lateral accelerat ion ayDPdec is controlled by the input as for the path tracking
task of Figure 6.9 via the lateral subsystem nSadec(S)/Ptat(s). The yaw rate Tdec
is not observable from ayDPdec'
iii. Rear-wheel steering OR controls only the yaw subsystem.
The numerators in Figure 6.25 are

nSadec(s) sJ.LcFf/mf R,
nDadec(S) s/mfR ,
nardec(s)
J.LCRf)
( s + -mvf- -f F
- (6.5.10)
F U DP'
nDrdec(S) s/ mU DP,
nRrdec(S) - sJ.LcR/mf DP.
For comparison with the uncontrolled car, i.e. the transfer functions (6.4.33), write

where
nsa(s)nar(S)
GSrdec(S)
Ptat(s)pyaw(S) ,
nRr(S)
pyaw(S) ,
nar(s )n Da(S) ()] 1
GDrdec(S) [ +nMr S - - ( - ) '
Ptat ()
S Pyaw S
S (J.L(CF + CR))
mfDpfRPtat(s)Pyaw(s) S + mv '
nsa(s)
GSadec(S)
Ptat(S) ,
0,
nDa(S)
GDadec(S)
Ptat(s) .

The s-factor in the numerators of GDadec(s) and GDrdec(S) indicates the disturbance
at tenuat ion at low frequencies and corresponding increased disturbance sensitivity at
high frequencies. The effect of the disturbance is quantified by the sensitivity function

S(s) = GDrdec(s) = GDadec(S) = SPA(S) (6.5.12)

GDr(s) GDa(s) Ptat(s)pyaw(s)
210 6 Case Studies in Car Steering

The acceleration subsystem (that is relevant for the driver task of path tracking) now
has only a first order transfer function
VS
GSadec(s) = mv R
f (6.5.13)
l+--s
J1-c Ff

Its eigenvalue at s = -J1-cFf/mvfR may be shifted by accelerometer feedback of ayDP

without affecting the yaw subsystem . This subsystem is, however, well-behaved and
there is no need to spend an additional sensor to the car steering system.
The poles of the yaw subsystem (6.5.7) may be expressed in terms of natural fre-
quency W o and damping D. Table 6.2 shows these values for the open-loop system and
for the decoupled system. The decoupled system is much more transparent in its pa-
rameter dependence such that we can continue with symbolic calculations . The natural
frequency of the decoupled system is velocity-independent . Solving D2 = D'i,c (which
is a quadratic equation in v 2 ) for v , a limit velocity V e is calculated , such that the
decoupled car has less damping than the conventional car for v > V e [10]. Figure 6.26

Table 6.2. Natural frequency and damping of the yaw motion

open loop decoupled

J1-~CFCR{,l + mV~J1-(cRfR - cFf F) J1-CR
w02
m2fRfDPv2 mf DP

D
fRfDP(CR

2VfRfDP[CRCFf2
+ CF) + cRf R 2 + cFf F 2
+ mv2 (cRfR - cF f F ) / J1-]
f R +fDP
2v
J J1-CR
mf DP

shows the velocity-dependent damping for the uncontrolled and decoupled car using
the data of a BMW 735i from Table 6.l.
In summary, the linear analysis of robust decoupling shows that the structurally
derived controller (6.3.19) without any tunable parameters (with the possible exception
of f DP , which is constant during a ride) yields the cascade structure of Figure 6.25,
where the lateral subsystem with output ayDPdec is relevant for the driver task of path
tracking, see Figure 6.7. In contrast to the oscillatory behavior of the uncontrolled car
(see nFa(S) in (6.4.33)), allDPdec is now generated by first order dynamics (see Gsadec(s)
in (6.5.11)). The path tracking task of the driver has been simplified.
The structure of Figure 6.25 indicates that the lateral and yaw subsystems may be
improved independently.
i. Accelerometer feedback for UyDP to {js changes only the lateral subsystem. Practi-
cally, there is no need for such a controller, because decoupling yields well-behaved
and sufficiently fast lateral dynamics.
ii. Yaw rate feedback to rear-wheel steering {jR changes only the yaw dynamics. As
Figure 6.26 shows, there is a need for improvement of the damping. It should
even in the worst case of J1- = J1-- provide a damping better than that of the
uncontrolled car. This can be achieved by the following result .
6.5 Linear An alysis of Robust Decoupling 211

1-F------...:~
...... --- D

L - . - - - - + - - - - - - - - - - - - - - - - + - _ v [m/s]
ve = 13.1 70
Figure 6.26. Decoupling causes reduced damping for v > Vi. Curves for p. = 1

Theorem 6.10 (Ackermann (7J, (31J)

The velocity scheduled control law

OR{V) = (e DPv+eR _ 2Ddes{V)/meDP)

P.CR
(rre! - r) (6.5.14)

assigns a desired yaw damping characteristic Ddes{v) to the decoupled car steering
J
system . It leaves the natural frequency WOdec = p.cR/meDP unchanged.
o

Remark 6.11

Again, the controller is turned on softly for v > VI > 0, such that no division by
zero occurs.
o

Proof of Theorem 6.10

By the control law (6.5.14), the transfer function from ORre! to r is

eDP + eR _ 2Ddes(V)/meDP) nRrdec{S)

rdamped(S) ( v P.CR
ORre!(S) Pyaw
v
ffii)
+ (e DP + eR - 2Ddec(v) -.!!.!:... nRrdec( s)
P.CR

DP R DP)
-SJ.LCR ( e + e - 2Ddes ()/me
V --
V J.LcR
212 6 Case Studies in Car St eering

The natural frequency is

WOdamped = WOdec = JJ.l~R

m{.DP
.

It does not depend on the choice of Ddes(v) and not on v . The damping is
Ddamped(V) = Ddes(v).
o
The controller (6.5.14) contains the uncertain parameter J.l/m. In order to achieve
the desired damping also in the worst case of slippery road (J.l = J.l-) and full load
(m = m+) , these parameter values must be used in the control law. On a dry road and
with lower mass, the damping then grows with J.l/m. J
The modified yaw dynamics have the characteristic polynomial

Pyawdes(S) = s 2 + ( 2D des(v) ~CR

- n - S + -J.lCR
m{.DP
n-'
m{.DP
(6.5.15)

and the sensitivity function of (6.5.12) is changed to

_ SPA(S)
S des()
S - (6.5.16)
Piat(S)Pyawdes(S) .
Peaks in ISOw)1 due to insufficient damping in Pyaw(s) are avoided by replacing it with
Pyawdes(S) in ISdesOw)l . The choice of Ddes(v) should take into account both the damping
(see Figure 6.26) and the effect on the sensitivity function (6.5.16).

Handling
So far, the robust decoupling concept has been analyzed with respect to the robust
stabilization and disturbance attenuation aspects. The handling transfer functions from
the handwheel angle to the lateral acceleration ayDP and to the yaw rate r can be shaped
by the second degree of freedom of prefilters providing the inputs bs , rrej , and bRrej to
the decoupled system . 'Here also, subjective performance ratings have to be considered
that go beyond th e scope of this book . It is, for example , possible to keep the sideslip
angle 13 small during all maneuvers by appropriate choice of the prefilter for bRrej [76] .
In the control systems structure of Figure 6.14 we require th at the immediate re-
action of ayDP and r to a steering step input is the same as in the uncontrolled car .
This requirement can be met in a robust way by assuring in the feedforward controller
structure, that the transfer function from the driver command b5 = bHandwheet!i (where
i is the steering gear ratio) to bF has high-frequency-gain 1, and the transfer function
to be has relative degree one, such that the step response at be begins with zero. For
a mechanical addition of angles as in Figure 6.4, bF = b5 + be is the only choice. The
prefilter for generating Tr e] from b5 allows more design flexibility.
A simple idea is to use a model following approach, where by (6.4.33) the open-loop
transfer function nFr(S)/PA(S) with one zero and two poles and desired parameters is
chosen as a prefilter . This would, however, reintroduce the oscillatory behavior of ayDP
that we have eliminated by robust decoupling . Two properties of the above filter should
be preserved :
6.5 Linear Analy sis of R obu st Decoupling 213

i. Th e relative degree one. This assures tha t th e ste p response in Tre/ - T begins
at zero with a non-zero slope. Th ereby, init ial peaks at the steering act uator are
avoided .
ii. Th e velocity dependence of t he gain according to (6.4.14).

A simple candidate t hat meets both requirement s is th e following prefilter.

krstat(v)
(6.5.17)
1+0.ls '

where
V 2 J.LCRCF£2
krstat(V) = £[1 + (v/vcH F] , VCH - m(CR£R - CF£F)"
(6.5.18)

Since J1./m is uncertain, only a nominal value for VCH can be assum ed for the prefilter .
Th e two signal paths from bs to bF are combined to

bFre/( s) [1 + s()
krstat(v) ]
s ( I + 0. ls ) SS,
(6.5.19)

n sadec(S)b () (6.5.20)
ayDP(s ) ( ) Fref S ,
Plat S
n ardec (S) ( )
Tdec(S) ayDP S . (6.5.21)
Pyawdes(S)

Th e handli ng ste p responses ayDP and Tdec may be furth er shaped by an additio nal
(velocity schedu led) prefilter with zero relative degree. Also, the desired yaw damping
characteristic Ddes(v ) enters into t he last transfer funct ion from ayDP to Tdec.

Fading Effect
Th e robust decoupling concept is important for the first 0.5 to 1 second after th e
occurrence of a disturbance torque M z D , when the driver has not yet reacted. It may
be modified such that th ereaft er the correct ive st eering angle bc (see Figure 6.3) returns
to zero. This is desirable in order to achieve the same st ati onary cornering behavior
that th e driver is used to, and also t o unload the actuat or for th e correct ive st eering
angle when it gets into saturation. Finally, also stability probl ems in th e yaw motion
are reduced if th e long-term behavior of the car is identical to th e convent ional car.
Since also th e immedi at e reaction to a ste p in the steering command bs is unchanged,
only th e transients are influenced by th e driver assist ant.
A disadvanta ge of th e t hree cascaded syste ms (6.5.19) to (6.5.21) is t hat th ey change
th e stea dy state response. Th e latter is determin ed by th e prefilter, because T re] - T
is integrat ed in t he decoupl ing feedback of Figure 6.14. Thi s long-term effect is not
what we want to achieve. Th e dri ver support syste m should give the idea l decoupled
behavior immediately after a ste p of bs or M z D ' In th e long run , however, t he driver
should take over t he cont rol of t he steady state, i.e. bc in Figur e 6.14 should return
to zero. Thi s effect is also necessary to unload t he act uator for bc , such t hat it is not
214 6 Case Studies in Car Steering

Step responses

0.4

0.3

0.2

0.1
OL.-.,,~-----'------'----'
0.5 2 4 6
tis
Figure 6.27. The initial response of the fading integrator is the same as that of the integrator

locked in saturation. For thi s purpose, th e li s-term in th e controller transfer function

is replaced by a fading in tegrator with transfer function
s
GF/ {s) = s 2 +as+ b' (6.5.22)

The immediate reaction corresponding to large s is lim s _ co GF/{ s ) = lis . For t --> 00 ,
the differentiating s in the numerator brings the response of be back to zero. For
a = 3 and b = 1 the ste p respon ses of the integrator and the fading int egrator are
shown in Figure 6.27. The fading integrator also eases the yaw damping problem of
robust unilateral decoupling, see Figure 6.26, because the system with fading integrator
behaves for large t like the uncontrolled vehicle. Also, the proportional feedback path
[(f D P - f F)lv]r may be faded by a filter s/{s + c). For a ralley driver with extremely
short reaction time, t he time scale of the fading effect may be changed by smaller values
of a and b .

Conclusions
In Section 6.5, an ideal car steering cont rol system has been presented that is based on

i. Yaw rate feedback to front-wheel steering for robust unilateral decoupling.

ii. Feedback of the yaw rate to rear-wheel steering for assignment of a desired
velocity-dependent yaw damping characteristic Ddes(s).
iii. Generat ion of a yaw reference r re f from the handwheel angle by a velocity-
schedul ed prefilter.
6.6 Skidding A voidance Based on Robust Decoupling 215

iv. The use of a fading integrator, in order to give the full driver-support only for
the first 0.5 seconds after a steering wheel or disturbance step, then the controller
smoothly fades out such that only the driver commands the steady state.
The control system requires only standard sensors for yaw rate, velocity and handwheel
angle . The control laws are derived structurally, thus only parameters of the single-
track model must be substituted for a specific vehicle. The decoupling controller under
i. requires no tuning at all, for the desired yaw damping controller under ii. , the designer
only has to specify the yaw damping characteristic Ddes(v) . The essential parameters
of the prefilter under iii. are structurally fixed, additional velocity-dependent dynamics
allow fine tuning of the handling transfer functions . The fading integrator under iv. has
two design parameters: parameter b essentially determines the timescale of the fading
effect, it depends on the human reaction time and not on the vehicle. Parameter a may
be tuned in view of closed-loop properties.
Like in the crane example of Chapter 1, the resulting controller structure is:
Lean only a few parameters of a low order controller must be adjusted.
Transparent : each controller par ameter or function has a clear physical meaning .
It influences only this property and leaves other input-output re-
lations unchanged .
Robust the control system is structurally robust with respect to uncertainty
in the road-tire friction , and the vehicle mass and velocity.
It is common practice to define progress in automobiles by comparison of a new model
with the predecessor model. If we compare a new model with the ideally controlled
vehicle, however, then we now have th e possibility to evaluate how close we can come
to the best .achievable car steering under technical and cost constraints.

6.6 Skidding Avoidance Based on Robust

Decoupling

The decoupling control concept of Section 6.5 was derived under idealized assumptions.
Practically, the following constrains must be considered :
i. No rear-wheel steering . This means that trade-offs in the front-wheel feedback
system must be made in order to achieve sufficient yaw damping at high velocities.

ii. Inclusion of the actuator dynamics . Its rate limitations may cause limit cycles that
must be understood and prevented . Its bandwidth is closely related to the cost ,.
i.e. only the absolutely necessary bandwidth should be specified for the actuator.

Yaw Damping by Double Cancellation

Most cars are not equipped with rear-wheel steering . Then, the yaw damping at high
velocities must be improved by yaw-rate feedback to front-wheel steering, rather than
216 6 Case Studies in Car Steering

to rear-wheel steering. If the decoupled system of Figure 6.25 is augmented by an

additional feedback path from Tde c to Os , then in general the decoupling effect and also
the factorization of the closed-loop characteristic polynomial , in a first order polynomial
Peat(s) and a second order polynomial Pyaw(s) , is destroyed. The latter factorization is,
however, very useful to calculate symbolically the yaw damping Ddec(v) as in Table 6.2.
This feature may be preserved by a cancellation compensation.

Theorem 6.12 (Ackermann)

A second feedback loop

(6.6.1)

with

(6.6.2)

around the robustly decoupled system of Figure 6.25 assigns the desired yaw
damping characteristic Ddes(v) to the system . Both the natural frequency WOdec =
vi/1cRjmf D P and the lateral dynamics pole Peat(s) = s+ /1c FfjmvfR are invariant
under the feedback (6.6.1).

Proof of Theorem 6.12

The proof proceeds in two steps:

1. It is shown that the control law

assigns the desired Ddes(v).

2. It is shown by substitution of WOdec and ddec that (6.6.3) is equal to (6.6.1),(6.6.2).

In the first step, the loop is closed with the plant

_ nSadec(s)nardec(S) J: ( ) (6.6.4)
rdec (S ) - Us S .
Peat (s)Pyaw(s)
A double cancellation of the two stable first order polynomials Peat (s) and nardec(s)
occurs, and the closed-loop characteristic polynomial is
6.6 Skidding A voidance Based on Robust Decoupling 217

WithPyaw(s) from (6.5.7) and nSadec from (6.5.10), the term in square brackets becomes

2mf p,cFf
+ 2Ddec(v) WOdecS + wOdec + (Ddes(v) - Ddec(v)) - - nR WOdec
2 2
S -n- s
P,CFf. mf.R
= S2 + 2Ddes(v) wOdecs + W5dec'
It has the desired damping characteristic Ddes(v) and the unchanged natural frequency
WOdec' Also, the factor Plat(S) of p(s) is unchanged.
In the second step , WOdec from (6.5.8) and Ddec(v) from (6.5.9) is substituted in the
control law (6.6.3) to obtain

os(s) = oSref{S) - (Dde.(V)

R
+ f ;v
f DP
J
:::;p) 2J:~:p ~: ~;;,
which agrees with (6.6.1).
o
The dynamic part of the controller (6.6.1) is

1 +TRs
1 +TFs
For the typical understeering car cRf R - cFf F > 0, i.e. T R > T F , the additional
controller has high-pass behavior (lead-lag filter) .
Practically, Ddes(v) = Ddec{v) may be chosen for small v, such that the full decou-
pling effect is preserved . For v > V2, then the additional yaw damping is softly turned
on. This destroys the non-observability of r from ayDP , but it yields good yaw damping
at high velocities, and the design remains transparent because the symbolic factoriza-
tion of the closed-loop characteristic polynomial is preserved . Note that the controller
(6.6.1) depends on the uncertain parameters p,lm and v . The velocity dependence may
be scheduled, for p,lm the worst case u: [m" is substituted. Ddes{v) should be chosen
such that in this worst case the damping gets never lower than for the conventional
uncontrolled car.

Analysis and Prevention of Active Steering Limit Cycles

In practical driving experiments, the effect of robust decoupling was demonstrated,
see Figs. 6.18 and 6.19. At that time , also the benefit of additional feedback of the
lateral acceleration ayDP at the decoupling point was experimentally investigated. At
low speed, stability problems in terms of oscillatory steering occurred and called for
further investigation. It turned out , that the oscillations were limit cycles, and due to
rate limitation of the steering actuator. The risk of unexpected occurrence of steering
oscillations obviously threatens driving safety severely. It constituted the need to have
a means to prove robust avoidance of limit cycles for any steering control at all op-
erating points of interest. Hence, the combination of the dual locus method with the
parameter space method was developed [16] . (Later this approach was generalized to
other frequency loci specifications as described in Sections 5.1 and 5.2.)
The scope of this paragraph is to
218 6 Case Studies in Car Steering

• explain the limit cycles which were experienced in driving tests,

• propose a remedy against limit cycles, for the robustly decoupled steering dynam-
ics,
• demonstrate the application of 6-stability mapping into parameter space for the
proof of robust limit cycle avoidance.

0
-2

-4 fact..... ideal
E
-6 - fact~2 HZ

-8
...... - l IN
•
- 10

Figure 6.28. Reconstruction of steering limit cycles at v = 5 mis, fJ = 1 with additional

feedback of lateral acceleration.

Figure 6.28 shows a reconstruction of the limit cycles at a specific operating point.
If (using additional feedback of the lateral acceleration ayDP at the decoupling point)
the gain exceeded a certain value, then limit cycles occurred in driving tests at that
operating point. The system used to generate Figure 6.28 consists of the linearized
single track model ((6.4.10) with BMW 735i data from Table 6.1) with robust decoupling
control (6.3.37) plus negative feedback of ayDP to the input 8s with a gain of 1.8. The
front steering angle 8F is set by a steering actuator 8F (s) = Ca(s) (8s(s) + 8c (s)). For
investigation, the control loop is cut open at the input of the steering actuator, and
zero steering command.de = 0 is assumed . The gray line in Figure 6.28 is the locus of
the transfer function of the open loop assuming an ideal steering actuator (Ca(s) = 1).
Since the steering actuator of the testing car was rate limited for hardware protection,
Figure 6.28 also shows the locus of the rate limiter negative-inverse describing function
as derived for Figure 5.13. No intersections between the two loci are present, hence no
limit cycles may occur. Now, a more realistic steering actuator model

(6.6.5)

with damping D a = 0.7 and bandwidth W a = 12.6[rad/s]~2 [Hz] is applied. The black
solid line in Figure 6.28 represents the corresponding Nyquist plot of the open loop
linear part. Here, an intersection point between this locus and the locus of the rate
limiter negative-inverse describing function appears. This indicates, that a limit cycle
may occur . Moreover, the Nyquist criterion indicates that the system is unstable in
the linear framework , i.e, even if no rate saturation is considered . This explains , why
limit cycles occurred in the driving experiments right after the controller had been
6.6 Skidding A voidance Based on Robust Decoupling 219

switched on. The reason for linear instability is the phase delay caused by the real
ste ering actuator in connection with the applied total loop gain . The effect of the rate
limiter caused the oscillatory instability to close in a steady limit cycle. The frequency
and amplitude of this limit cycle can be determined from the parameterization of the
two loci at the intersection point z = -1.23 - 6.27i. The frequency of the linear part
transfer function locus is wgz = 7.9 [rad/s] ;;1.26 [Hz] at the intersection point. If this
value is substituted in (5.1.9), iterative solution yields

R/A = 0.97[1/8]. (6.6.6)

The rate limitation of the real steering actuator was determined experimentally. It
depends on the actuator load but can be restricted to be in the interval R = 400...800°/s
(referring to the steering column angle). Using (6.6.6), the theoretic amplitude A of
the limit cycles corresponds to one or two revolutions of the steering wheel. Both the
frequency and the amplitude comply approximately with the experimental limit cycles.
Hence, the real limit cycles can be well explained by applying the presented system
modeling and analysis.
Further theoretic investigations resulted in the awareness, that even with the em-
ployment of a significantly higher bandwidth steering actuator (wa = 63[rad/s]) and
even with no lateral acceleration feedback the risk of limit cycles is present. Consider
the operating point with 70 tn]« vehicle speed on dry road . The respective loci are

-1

-3 .

_4'--~~-----'-_..L..--'

-2 -1 0 o
Re t[s]

Figure 6.29. Decoupled steering dynamics Figure 6.30. Rate limiter input u and out-
loci at v = 70 mis, J.L = 1 for rate limited put y plots of steady limit cycle simula-
steering actuator with W a = 27f . 10 [Hz} tions. The dashed line shows a sinusoidal
bandwidth. approximation of u.

displayed in Figure 6.29. There are two intersection points indicating possible limit
cycles. However, only the intersection point marked by a star represents a stable limit
cycle. A numeric simulation of the closed loop system was performed starting with an
initial state, which lies outside the region of attraction of the equilibrium state. In fact,
a steady limit cycle can be observed , which is depicted in Figure 6.30 in terms of the
actuator rate limiter input and output signals. The oscillation frequency (wgz ::::: 4.15
[rad/s] and amplitude are well in accordance with the par ameterization of the loci in
220 6 Case St udies in Car Steering

Figure 6.29 at t he intersect ion point indicated by the star marker. The input to the
rate limiter input signal can be well approximate d by a sinusoidal. Th is indicates that
the harm onic linearizat ion of t he rate limiter to its describing funct ion is legit imat e,

..
which is due to distin ct low-pass properties of the open cont rol loop. A rule of t humb

i;: m
m!mm
in 5~t:=IB

- 150 -

~m20:l l mPSIJ
-40&01 0.1 1 10 100 1000
co [rad/s]

Figure 6.31. Bode plot of the open loop linear part at v = 70 ui]«, J-L = 1.

requires an open loop relat ive degree of at least two. Moreover, it is important that
the associated gain reduction is already effective at the frequency of the limit cycle
oscillation . Figure 6.6 shows t he Bode plot. Th e vertical lines indicate the frequency
region above the limit cycle. At all frequencies great er than th e limit cycle frequency
there is a decline of at least [-40 dB/decade]. Hence the basic assumptions for applying
the dual locus meth od are met, see Section 5.1.

Nonlinear Control Enhancement for Limit Cycle A voidance with Ro-

bustly Decoupled Steering Dynamics
One fact , which essentially cont ributed to the occurr ence of limit cycles, is th e rate
limit ati on of the steering act uator. This physical limit has to be considered for control
design. The goal followed here is, in the first place, to reduce t he cont rolled system 's
susceptibility to limit cycles. Secondly, for t he sake of safety, robust avoidance of limit
cycles must be proven for the whole opera t ing domain of t he act ively steered car.
It tu rns out, that reducing th e suscept ibility to limit cycles is fairly simple in the
context of robust decoupling cont rol [16] as will be demonstrated in the sequel. Th ere-
fore, note that right before the cont rol signal enters th e act uator as its set point , t here
is an integrat ion of the signal Tr e] - T, see (6.3.37). The environment of t his part of the
control loop is illustrated in Figure 6.32. The out put JC r e! of the decoupling controller
(integrator) JC r e! = T r e! - T ente rs the rate limit er, which is a part of the actuator. At
the input of the int egrator the re is the derivative W.r.t. tim e of th e input to the rate
6.6 Skidding A voidance Based on Robust DecoupJing 221

limiter , i.e. th e demanded rate. This quantity is limited by the rate limiter for the sake
of hardware protection. Now a saturation of JC re! is introduced as a part of the control
with
r, ~ R (6.6.7)
as illustrated in Figure 6.32. Hence, by the effect of this modification the rate limiter
will never be activated. The demanded rate (i.e. the absolute value of the derivative of
OCre!) is a priori limited to T s. As a consequence, for any consideration and analysis , the
rate limiter can be neglected. However, the saturation needs to be taken into account
instead. There is a great benefit from this nonlinear control enhancement in terms of
stability. This is due to the fact , that in contrast to a rate limiter , the saturation does
not introduce any phase delay into the control loop, when the demanded rate exceeds
the allowed one. In fact , it turns out that the region in the ((v, f..L)) plane of operating
parameters, where limit cycles may occur with the present system , shrinks considerably,
since the saturation does not cause phase delay in the control loop [16] . Figure 6.33
illustrates this for the operating condition, which has already been considered in Fig-
ure 6.29. By means of the respective negative-inverse describing function loci the effect
on limit cycle existence can be directly compared for both nonlinearities. The real lo-
cus -1/ Ns(A) , which belongs to the saturation, is positioned much more advantageous
(i.e. farer away from the Nyquist plot) than the locus -I/Na(w , A) of the rate limiter.
With the saturation, there are no more loci intersections, hence no limit cycles exist .
Moreover, there is significant space between the loci, indicating a convenient safety
distance in terms of phase and gain margins. The introduction of the saturation to the
controller thus drastically reduces the risk of limit cycles. Nevertheless, the saturation
now takes over the task of rate limiting the actuator input. The risk reduction can
only be stated for the considered operating point. However, it is important for stability
robustness, that limit cycles may not occur at any operating point in the operating
domain (Figure 6.20).
The presented nonlinear control enhancement can be combined also with the fading
integrator, as illustrated in Figure 6.34. A controller-internal feedback is applied from
the integrator output OCre! to the saturation input via the filter

(6.6.8)

Tre! -T ~

Figure 6.32. By limitation of the control signal the rate limiter becomes disarmed.
222 6 Case Studies in Car Steering

-1/N (A)
s
-1

-2
s
......
-3

_~L....._---'- __ _ _.L-
"""""_~-'- '---'
o

Figure 6.33. Nyquist plot (solid line) for v = 70 ta]», J.L = 1 and loci of negative-inverse
describing functions of rate limiter (dash-dotted) and saturation (dashed).

JCrel

Figure 6.34. Nonlinear controller enhancement of the fading integrator .

such that the linearized transfer function from Tre] - T to 8Crel is just the one of the
fading integrator (6.5.22)
S
(6.6.9)

Still the absolute value of the derivative w.r.t. time of the input to the rate limiter
can not exceed T., and therefore the rate limiter never gets activated. For the appli-
cation of the dual locus method for analysis of limit cycle existence, the control loop
is cut open between the saturation and the integrator. The transfer function GI(s)
becomes part of the open loop linear part, and thereby affects the course of the Nyquist
plot. It contributes to reducing the risk of limit cycles, since the loop gain at low
frequencies is significantly lower. This reduces the input to the actuator at low frequen-
cies, which reduces the probability of the saturation in the controller to be activated.
Figure 6.35 shows the result of 8-stability margin mapping into the (v,l',)-plane for
active car steering with fading integrator, where e
the locus of the negative-inverse
6.7 Skidding A voidance Based on the Disturbance Observer 223

saturation describing function . For any operating point to the left hand side of the
boundaries, the open loop transfer function locus does not intersect -1/Ns(A) , i.e. this
is the 8-stable region. The boundaries were produced assuming two different actuator
bandwidths. With the gray line, the actuator bandwidth is 1 [Hz], and the boundary is
quite close to the operating domain . To ensure some safety margin, the actuator band-
width should therefore be greater. Assuming 2 [Hz] actuator bandwidth, the 8-stability
boundary recedes from the operating domain which enlarges the security distance. This
trend continues for further increase of the bandwidth. By this approach, an actuator
bandwidth can be specified such that limit cycles can be excluded.

Remark 6.13

If the results shown in Figure 6.35 are investigated in more detail, then it turns out
that no tangent condition, but only the point condition for the dosing point -1 (where
-1/Ns (A) starts), contributes to the boundaries being present in the depicted part of
the (v, j.t)-plane. These boundaries can be considered as Hurwitz-stability boundaries
of the corresponding linear systems if the saturation in the controller is neglected, since
the dosing point of -1/Ns(A) is the critical point -1 w.r.t . the Nyquist criterion. 0

6.7 Skidding Avoidance Based on the Disturbance

Observer

In this section, another robust steering controller is introduced for improving the yaw
dynamics of a passenger car. It is independent of the idea of robustly decoupling the

1 roi(21t)=1 Hz
::1. 0.8 roi(21t)=2Hz

0.1 '--'- -'-_ _---'

5 70
v [m/s]
Figure 6.35. Limit cycle a-stability boundaries in the (v,j.t)-plane.
224 6 Case Studies in Car Steering

lateral motion from the yaw dynami cs of the car. Here, rather the disturbance rejection
control architecture shown in Figure 1.11 is adapted to the vehicle yaw dynamics con-
trol problem and shown to robustly improve performance in terms of driving safety and
handling qualities. The relevant design specifications are formulated in terms of eigen-
values (f-stability) and in the frequency domain as bounds on weighted sensitivity and
complementary sensitivity functions (B-stability) . The parameter space method is used
to consider the specifications for controller design. A Popov criterion - based non-linear
stability analysis is also carried out to prove robust absolute stability (8-stability) in
the presence of actuator rate limitation. Simulations demonstrate the effectiveness of
the design.
The steering control system should be robust w.r.t. large variations in longitudinal
speed, payload and road adhesion and w.r.t. unstructured uncertainty (unmodelled
dynamics) as well. Moreover, its actions should not be uncomfortable for the driver
and passengers . A steering controller structure that effectively satisfies the requirements
outlined above is presented along with its associated design and analysis procedures.

Vehicle Model and Steering Actuator Model

For modelling the vehicle yaw dynamics , the linearized single track model from Sec-
tion 6.4 is employed as a parametric linear time-invariant plant (parametric LTI). The
transfer function from the front wheel steering angle OF to the yaw rate r can be com-
puted from (6.4.4):

(6.7.1)

with
bo J-t2 CF CR(lF + lR)V,
b1 J-tc FlFmv 2,
ao J-t2 CF CR(lF + lR)2 + J-t(cR lR - CF IF)m v 2,
al J-t(CF(J + l} m) + CR(J + l~ m))v,
a2 .l m u" .

The steady state gain of the nominal vehicle steering transfer function is

Kn(v) = 5-0
lim G(s)!
1'=l'n
(6.7.2)

at a nominal friction coefficient J-t = J-tn '

In this section, both the conventional car and the controlled car are equipped with a
steer-by-wire actuator, which sets the total front steering angle . Therefore, the steering
transfer function of the conventional car is
r(s)
Gconv(s) =8 = G(s) Ga(s), (6.7.3)
s(s)
where Os denotes the steering wheel angle and G a represents the actuator dynamics
(see (6.7.5) below).
6.7 Skidding A voidance Based on the Disturb ance Observer 225

Vehicle Model Data and Operating Domain

The nominal parameter values for th e linearized single track model assumed in thi s
section are IF = 1.25 [m), IR = 1.32 [m], m = 1296 [kg], J = 1750 [kg m"], CF
84243 [N/rad] and CR = 95707 [N/rad]. The nominal friction coefficient is J.Ln = 1,
which corresponds to a dry road.

1
0.8
:::l.
0.5

0.2

10 30 50
V (m/s)

Figure 6.36. Operating domain and considered operating points

The tire cornering stiff'nesses J.L CF and J.L CR can exhibit large variations due to vari-
at ions in the friction coefficient J.L between the road and th e t ires. The longitudinal
speed v is assumed to vary only slowly dur ing opera t ion in the range between zero and
50 [m/s]. Gain-scheduling will be necessary firstly to switch on the cont roller st art ing
from zero speed. Secondly, gain scheduling is required to meet all specifications also at
low speed. Th erefore, in thi s section only high-speed operation is investigated at two
exemplary speeds: v = 30 [m/ s] and v = 50 [m/s ]. Th e assumed operat ing domain
of the vehicle in terms of th e speed v and t he road adhesion coefficient J.L is displayed
in Figure 6.36. The maximum value of J.L is 1 (dry road) at all speeds while it is as-
sumed that th e minimum value increases linearly from 0.2 (icy road) at low speeds
up to 0.8 (wet road) at high speeds. (Additional uncert ainti es will be considered on
p. 230.) Cross markers in Figur e 6.36 indicat e specific operatin g points , which will be
particularly looked at later.

Steering Actuator Model

An electric motor und er position control is assumed here to be a steer-by-wire act uator,
which is used to set t he total front wheel steering angle. The act uat or model used in t his
section is shown in Figure 6.37. Notations and assumed numerical values of t he model
var iables are compiled in Table 6.3. The third order transfer function of t he linearized
act uator model according to Figure 6.37 while neglectin g the voltage limit at ion, is given
226 6 Case Studies in Car Steering

Figure 6.37. Actuator model

by

Ga(s) = c5F (s) (6.7.4)

c5Fre/( S)

With the data from Table 6.3, th e linearized steering actuator model transfer function
is
Gas( ) = c5 c5 F(S)
()
= 2 DW~ (
6.7.5
)
W + 2 aWa S + S
2'
Fre] S a
with Da = 0.7 and W a = 21r . 5 [rad/s]. Under the assumption of a 12 V power supply,
the voltage is limited to lui :::; iJ. = 12 V.

Problem Specifications
The aim in yaw dynamics cont roller design is t o make sure that stable and improved
yaw dynamics is achieved for all operating conditions and all possible values of th e
uncertain parameters despite the presence of add itional unstructured uncertainty rep-
resenting unmodelled dynamics , e.g. sensor dynamics, neglected flexible body modes,
etc. Here, improved yaw dynamics means good disturbance rejection properties where

Table 6.3. Notations and numerical values of actuator variables.

Ja 0.004053 [kgm:.!] motor inertia

k/ 0.01625 [Nm s/rad] motor damping coefficient
La o [H] inductivity (electric time constant neglected here)
Ra 5[n] electri cal resistance
k me = kem 0.9 [Nm/A] electromagnetic constant
ka 22.22 [V/rad] cont rol gain of actuator posit ion control
iJ. 12 [V] volt age limitation
6.7 Skidding Avoidance Based on the Disturbance Observer 227

the possible disturbances include the effect of wind forces and I-t-split braking. Good
steering tracking performance is required as well. A disturbance observer-based steering
controller is designed and shown to effectively achieve the desired aim in the following
sections.

The Disturbance Observer

The disturbance observer is a specific method of designing a disturbance rejection ar-
chitecture (see Figure 1.11) to achieve insensitivity to both modelling errors and dis-
turbances [163], [193]. It has been used successfully in a variety of motion control
applications, including high-speed direct drive positioning [124] and friction compen-
sation [102]. The disturbance observer can also be used to achieve dynamics similar
to those obtained by the use of inner feedback loops through control [100]. The appli-
cation of the disturbance observer control scheme to vehicle yaw dynamics control is
introduced in the following.
The vehicle model can also be expressed as

(6.7.6)

where G is the actual vehicle dynamics input-output relation between steering wheel
angle 8F and yaw rate r . The term ~m is a multiplicative uncertainty w.r.t . to all
adopted nominal model Gn, which may as well represent the desired dynamics . So
~m comprises any dynamic divergence (also structured and unstructured uncertainty)
between G; and the actual G but the effect of external disturbance d. The aim in
disturbance observer design is to approximately obtain
r
8 = GnGa (6.7.7)
5

(rather than (6.7.3)) as the input-output relation (steering transfer function) despite
the presence of ~m and external disturbance d. This aim is achieved by treating the
effect of ~m and d as an extended disturbance e in disturbance observer design and
solving for it as

r = Gn 8F + (Gn ~m 8F + d) = Gn 8F + e , (6.7.8)
e r - Gn8F , (6.7.9)

and using the new control signal 85 given by

(6.7.10)

to approximately cancel its effect in (6.7.8). Substituting (6.7.10) into (6.7.8) shows that
the desired steering transfer function (6.7.7) is achieved if a good actuator (Ga -> 1)
is used. The front wheel steering angle 8F is assumed to be the output of the steering
actuator Ga . With the aim of trying to limit the compensation to a preselected low
frequency range (in an effort not to overcompensate at high frequencies), the feedback
228 6 Case Studies in Car Steering

Disturbance observer I
r
I
I

Figure 6.38. Closed-loop structure with disturbance observer

signals in (6.7.10) are multiplied by the low-pass filter Q to obtain the implementation
equation
(6.7.11)
which can also be seen in the block diagram of Figure 6.38. Including G a in th e inner
feedback loop helps in reducing the effect of actuator saturation on disturbance observer
performance [34] . The relat ive degree of the unity d.c. gain low-pass filter Q is chosen
to be at least equal to the relative degree of G; for causality of Q/Gn . In the sequel ,
the structure of Q is assumed to be
1
Q(S)=TQS+ 1 (6.7.12)

The nominal steering t ransfer funct ion is chosen as a first order syst em, here given by

(6.7.13)

and K n (v) is the steady st ate gain of the nominal single-track model (i.e. on a dry road ,
J.1.n = 1) at th e actual longit udinal speed v from (6.7.2) .
The open-loop transfer function at signal T of th e disturbance observer compensated
yaw dynamics model is (for th e sake of brevity, th e Laplace operator s is partly omitted
in the sequel)
L := GaQ .£ (6.7.14)
1 - GaQG n
The steering transfer function, disturbance rejection (i.e. sensitivity funct ion S) and
sensor noise rejection [i.e, complementary sensitivity function T) transfer functions
given are then
T GnGaG
-= (6.7.15)
85 G n(l - G« Q) + G« G Q '
T 1 Gn(l- Ga Q)
S .- (6.7.16)
d l+L Gn(l- o; Q) + Ga G Q '
T L GaGQ
T .- -- (6.7.17)
n 1 +L Gn(l- GaQ) + GaGQ '
6.7 Skidding A voidance Based on the Disturbance Observer 229

from which it is obvious that for good performance, Q must be a unity gain low-pass
filter (e a is a unity gain low-pass filter as well). This choice will result in r / 8s -+ en,
r / d -+ 0 at low frequencies where Q -+ 1 and r / n -+ 0 at high frequencies where Q -+ 0
as is desired. Disturbance observer design is thus mainly shaping the filter Q to satisfy
the design objectives. The first limitation on the bandwidth of Q comes from the sensor
noise rejection at sensor noise frequencies. The second limitation is that the bandwidth
of Q should not be larger than the bandwidth of the actuator used as it makes no sense
to command what cannot be achieved. The other entity that significantly affects the
properties of the controlled system is the choice of the desired dynamics en' With the
assumed structures (6.7.12) and (6.7.13), tuning the implemented disturbance observer
controller (6.7.11) means tuning the parameters TQ and Tn . This is the scope of the
following section in order to achieve robust matching of the design objectives.

Parameter Space Design

The parameter space approach is applied for robust controller design and analysis .
Besides eigenvalue specifications (f- stability), Bode magnitude type specifications (B-
stability) on sensitivity functions are also incorporated into the controller design to
enhance vehicle stability as well as performance and handling qualities robustly w.r.t .
structured (v, /1-) and unstructured uncertainty (unmodelled dynamics) .
First, the problem set-up is stated. Then, the stability and performance specifi-
cations in eigenvalue and frequency domain for controller design are described. The
parameter space approach is used to map these specifications into the plane of the two
controller parameters TQ and Tn . Finally, the resulting controller is analyzed in terms
of sensitivity.

Problem Set-up
Consider Figure 6.38 and (6.7.12). Q is chosen as a unity d.c. gain-first order low-pass
filter with time constant TQ. The open-loop single-input-single-output transfer function
L, sensitivity function S and complementary sensitivity function T are given by (6.7.14),
(6.7.16) and (6.7.17), respectively. Furthermore, the characteristic polynomial of the
closed-loop system is defined as

p = num (1 + L) . (6.7.18)

In the following demonstration of the design procedure applying f- and B-stability, the
time constants Tn of the nominal model en in (6.7.13) and TQ of the filter Q in (6.7.12),
respectively, are considered as controller parameters. These time constants shall be
tuned such that the feedback provides f- and B-stability for the whole operating domain
given in Figure 6.36. In the first design step being described in this section, only the
four operating conditions marked by crosses in Figure 6.36 are considered. For each of
them, the p, and B-stability specifications given below are mapped into the controller
parameter plane (Tn, TQ) .
230 6 Case Studies in Car Steering

Control Design Specifications: r -stability

Figure 6.39 shows the f-region representing the eigenvalue specifications t hat are one
part of multiple specifications for the design of Tn and Tq for all (J-L, v) in the operating
domain. The boundary of of I' is composed of three geometric elements . The first

"
. . >~" ~' . .!. .
50
40 . ~" . . ...!. .
I
30 . . . . ./ ..... - T·
20 _.

10 . - ~ •. .

s
>-<
0
-10
-20 -" ;' .

-30 . -. .:',
. "' ,
-40 '''~" .

-50 ' ~" , rr

;1"1 " ..... :

-60 -40 -20 o

Re

Figure 6.39. r-specifications in eigenvalue domain

boundary is a shifted imaginary axis s = -2, which ensures that the system 's settling
time is limited . Two lines of constant damping guarantee a minimum damping of 0.5
of all poles, and a circle centered at the origin guarantees that the natural frequency of
any pole does not exceed 211" . lO[rad/sJ. The poles of the dosed-loop system shall be
located in the admissable eigenvalue region I' as shown in Figure 6.39, i.e.

roots{p(s, Tn, TQ, V, J-L)} C I' . (6.7.19)

Control Design Specifications: B-stability

Disturbance rejection

a) To ensure disturbance attenuation at low frequencies, IS(jw)1 « 1 must hold.

b) However, since L is a robustly stable transfer function with a relative degree of
three, Bode's integral theorem [55) applies, which means that attenuating distur-
bances at low frequencies unavoidably results in amplifying them at high frequen-
cies. Therefore, the amplification must also be restricted to an acceptable level,
i.e. /IS(jw)/Ioo < lips , where Ps is equivalent to the required Nyquist-stability
margin.
6.7 Skidding A voidance Based on the Disturbance Observer 231

These two specifications a) and b) are combined into one specification by selecting
an upper bound on the FRM of S (see Figure 6.44), i.e,

(6.7.20)

where
-I 8+0.7
Ws(s) =1.8 26 '
8+1 .
Robustness ui.r.i. unstructured uncertainty Two magnitude bounds on T, i.e. the mag-
nitudes of WT,I(st l and WT,2(8t l are selected to capture robustness w.r.t . unstruc-
tured uncertainty, i.e . using a multiplicative uncertainty model.
a) The disturbance observer shall guarantee robustness to 10 % magnitude uncer-
tainty at low frequencies, i.e. where the model of the vehicle and the actuator is
reasonably accurate, and 500 % uncertainty at high frequencies, i.e. where unmod-
elled dynamics come into play. Thereby, a transition frequency of 21r . 6 [roofs)
between low gain and high gain of the weighting function WTl ( 8) is selected based
on the knowledge of the vehicle model's accuracy and the performance specifica-
tion of the steering actuator.
b) A second bound on T , i.e. WT,2(8t l is used to cover the disturbance observer
stability specifications subject to model uncertainty in m and J [65] .
Hence, the following 13-specifications will be used as constraints for T :

(6.7.21)

where
5 s + 3.77
8 + 188.5 '

W ,2(8) = 0.12804(8 + 43.98)(8 + 0.4833)

T
(8 + 6.124)(8 + 2.882)
(see Figure 6.45) . Note that (6.7.21) represents the "robust stability" criterion for two
different multiplicative uncertainty models. At low frequencies, unstructured uncer-
tainty subject to uncertainty in m and J is relevant and for high frequencies uncertainty
subject to unmodelled dynamics is relevant.

Mapping of r - and B-stability Boundaries into Parameter Space

The mapping of the f-stability boundaries defined on p. 230 is based on the limit
condition that the characteristic polynomial p(8,Tn, Tq , V, J.L) has a root exactly on the
I'<region boundary of. Therefore, the condition of = 8(a) = u(a) + j w(a) , where a
is the parameterization of of, is substituted into the characteristic polynomial given
by (6.7.18) , which is separated into two equations for its real and imaginary parts,
respectively:
Re (p(a,Tn, TQ , V, J.L)) = 0,
(6.7.22)
Irn (p(a,Tn, TQ'V ,J.L)) = O.
232 6 Case Studies in Car Ste ering

The mapping of the P-stab lllty boundaries into th e parameter plane (Tn ' TQ) re-
quires algebraic solution of these two equations, see Appendix A. Figure 6.40 shows
the result for the mapping of the f-stability boundaries (as defined in Figure 6.39)
for the two operating conditions with v = 50 [m/s] and J.L E {0.8; 1} (see Fig-
ure 6.36). The (Tn' TQ)-region, which simultaneously I'-stabilizes the two operating
points, is denoted K r . In this plot and also in Figure 6.41 and Figure 6.42 th e
grey lines correspond to the operating condition with J.L = 0.8, and the black lines
to that with J.L = 1. For a better distinction between I'- and B-stability bound-
aries, dashed linestyle is used for f-stability in this plot and in the subsequent
plots. For the sake of consciseness, the f -stability boundaries for v = 30 [m/s]
and J.L E {0.5 ; 1} are not displayed in Figure 6.40, Figure 6.41 and Figure 6.42. To
establish the mapping equations of the B-stability boundaries defined by Equations
(6.7.20) and (6.7.21), it suffices to consider two mathematical condi tions, the point and
the tangent condition [159]. The point condition applies when IS(jw, Tn , TQ , V , J.L)! or
IT(jw, Tn, TQ , V, J.L)I , respectively, st arts (w = 0) or ends (w --> 00) on the boundary
8Bs = IWS(jW)-11 or 8BT,i = IWT,i(jW)-11, respectively, The tangent cond iti on al-
lows for the mapping of touching points, i.e, the points where IS(jw, Tn , TQ, V , J.L)I or
IT(jw, Tn , TQ , V , J.L)I , respectively, becomes tangent to 8Bs or 8BT ,i, respectively. Details
about the mapping of B-spe cifications are described in Section 5.3.

0.15L_--JI--f~-===J

~ 0.1

~
.... " 0.05

0.15 0.2 0.25 0.3 0.35

Tn [S]
Figure 6.40. Ivstable region for v = 50 [m/s] Figure 6.41. Nominal performance in pa-
and J.L E {O.S; I} rameter space for v = 50 [m/s.] and J.L E
{O.S;1}

To ensure that the magnitude of the sensitivity function IS(jw, Tn, TQ , V , J.L) I remains
below its upper bound, Condition (6.7.20) is mapped into the (Tn ,Tk) controller param-
eter plane. F igure 6.41 shows the resulting region K Bs for the two operating points at
V =50 [m/s] with J.L E {0.8; 1}. For any parameter combination (Tn ,TQ) taken from
this region, Condition (6.7.20) is satisfied for both operating points simultaneously.
For consideration of robust st ability, Condition (6.7.21) is mapped into the (Tn , To)
controller parameter plane (see Figure 6.42). The dotted lines correspond to t he
Hurwitz-stability condition, i.e. the nominal stability condition, for the two operating
points considered. This non-conservative mapping shows which set of (Tn , TQ) controller
parameters (i.e. the region K BT in Figure 6.42) will guarantee robust stability in the
6.7 Skidding A voidance Based on the Disturbance Observer 233

-0.1 L-_~ __ ~_~ __ ~_-J

OL-_-~--~--~-_-J

- 0. 1 o 0.1 0.2 0.3 0.4 0.12 0.14 0.18 0.2

Tn [Sl
Figure 6.42. Robust stability in parameter Figure 6.43. Parameter region which meets
space for v = 50 [m/s] and J1- E {0.8; I} all B- and f-specifications for v = 50 [m/s] ,
J1- E {0.8; I} and v = 30 [m/s], J1- E {0.5; I}

presence of actuator uncertainty, unmodelled dynamics and/or structured uncertainty

due to the disturbance observer specifications simultaneously for the two operating
points at v = 50 [m/s] with J1- E {0.8; 1}. Similiar results are achieved for the other
two operating points v = 30 [m/s] and J1- E {0.5; I} .
The final step in controller design is superimposing the previously determined sta-
bility regions in parameter space . The intersection of all of the individual regions in the
(Tn, TQ) plane that satisfy the formulated f- and B-specifications for all four operating
points form the desired parameter region K shown in Figure 6.43. For any choice of
(Tn' TQ) E K , all previously described requirements in terms of p, and B-stability are
met simultaneously for the four operating conditions.

Controller Selection
Note that satisfying all specifications postulated also for low speed (e.g. v = 10 [m/s])
requires gain-scheduling of Tn and TQ with speed since the vehicle yaw dynamics is
much faster at low speeds . This is not considered here for the sake of brevity. Instead,
from the region K in Figure 6.43 the parameters Tn = 0.165 [s] and TQ = 0.0318 [s]
are chosen as fixed controller parameters (marked with a cross) . Figure 6.39 shows all
eigenvalues of the closed-loop system for the four operating conditions. They are all
included in the desired r-region.

Sensitivity Analysis
Figures 6.44 and 6.45 show IS(jw)1 and IT(jw)1 for the four operating points. They
all remain below IWS(jW)-11, IWT,1(jW)-ll and IWT ,2(jwt 11, respectively, i.e. they are
entirely included in the B-stable regions.
 234 6 Case Studi es in Car St eering

2r------~-----_,

1.5 afl0

'3
... . ...
...... ::£ol------~~~~
.:.:::J 1 ..-..
Cf) . ~-1 0
0.5 h_20
-30'--:-------'-:--------'
10° 10' 10-' 10° 10'
Frequency [Hz] Frequency [Hz]

Figure 6.44. FRM-plot of the sensitivity Figure 6.45. FRM-plot of the complemen-
function tary sensitivity function

Simulation Results

Two versions of the vehicle are compared in this section by means of linear simula-
tion : the conventional car and the controlled car . For the sake of comparability, the
convent ional car is assumed to be a st eer-by-wire vehicle equipped with the same steer-
ing actuator as the controlled car (see (6.7.3}) . The steering transfer functio n of t he
controlled car is given by (6.7.15). Th e controller parameters are set according to the
design results: T« = 0.165 [s] and TQ = 0.0318 [s] . Two maneuvers are investiga ted:
A steering wheel step inpu t and a yaw disturbance torque step inpu t. The man euvers
are performed at th e four operating points marked in Figure 6.36. The results are
shown in Figures 6.46-6.49 in terms of th e yaw rate T , the steering wheel angle 6s and
the front wheel steering angle 6F . Th e simulations show that t he controller provides
excellent disturbance rejection and a good steering tracking at all investigated operat-
ing points. The yaw disturbances are robustly attenuated within 0.5 s. The control
action is finished before the driver is even capable of starting his count ersteering due
to his reaction time of about 1 s. Th e yaw rate responses to a steering wheel input
are well in accordance with th e desired steering transfer funct ion (6.7.7) not exhibit ing
any overshoot . Zero steady state error is inherent to the control structure as already
discussed by means of (6.7.15}-(6.7.17). Th e steady state gain of the controlled vehicle
steering transfer function is Kn(v}. This gain has been calculated to be identical to th e
conventional vehicle steady state gain for /In = 1. Therefore, the steady state yaw rate
of the conventional vehicle and the controlled vehicle after a steering wheel input are
identical for /l = 1 (Figure 6.46 and Figure 6.48) but differ otherwise.

Note that other choices for Q and Gn than given by (6.7.12) and (6.7.13) are feasible
and may be reasonable depending on th e specifications. For example , a second order
filter for G; to represent a cert ain desired single-t rack st eering transfer function can
be used. If this model is well in accordance with th e actual vehicle, th en th e steering
transfer funct ion is not significantly changed by control. However, good disturbance
rejection is retained. On th e other hand , a band-pass filter can be applied for Q instead
of a low-pass filter . Thi s removes control act ion at low frequencies ("fading controller")
and transfers th e task of compensat ing steady st at e errors back to t he driver [33].
6.7 Skidding A voidance Based on the Disturbance Observer 235

Steering wheel angle step response

0.12

rn
0.1

10.08
u,
~0.06 .... - l):contr .
en
<0
0.04 .. . - - l)FCOOY .

0.02
-0.5 L..-_--''---_--'-_ _--'-_ _-' 0
o 0.5 1 1.5 2 0 0.5 1 1.5 2
t[s) tis)

,
Yaw disturbance torque step response

,,.,
1.5 0.05
" ..... .
....... _------
v=SOniJs, Ii= 1
o ---_._ --,,-
6
~
iO 0.5 , .
, I
. ., .
-
--
5
l)Foontr
6Fconv
~ I . ..

-0.15 V

-0.5 -0.2
o 0.5 1 1.5 2 0 0.5 1 1.5 2
t[s) t is)

Figure 6.46. Linear simulation results at v = 50 [m/s], J.t = 1.

Sleering wheel angle slep response

1. 5.-----~--------___, O .2.-----~--------___,

Y=5OmIs. ,,=0 .8
~ ....._ ._ .~- _ ._-- 0.15 . . .... .. . . . . .. '-.:..:.
. ~~ .........,
g;
~
1IQu. 0.1
0>
<0

0.05

OL- --.J
-0 .5'---~--~--~---'
o 0.5 1 1.5 2 o 0.5 1.5 2
t[s)
Yaw disturbance torque step response
2 O.05 r - - - - - - , - - - - - - - , - - - ,
.. '!' .. V=5Oni.ts. ,,=0 .8
1.5 / : ~
I ..... - - :- - - -

I" I:.: =~II

1-0·05
<0"-

0.5 I' d 0> -0.1

,,""---------l
<0

o
-o.5'----~--~--~------'
o 0.5 1 1.5 2 0.5 1 1.5 2
I [5] 1[5]

Figure 6.47. Linear simulation results at v = 50 [m/s] , J.t = 0.8.

236 6 Case St udies in Car St eering

Ste e ring wheel a ngle step respon s e

2.5,---,..---,..---,..-----, 0.25
v-sornzs, fl=1
2 ,. ~.:..::..:. --...., 0.2

0;
~ 0.15
u.
eo
'" 0.1
""
0.05

-0.5 ' - - - ' - - - - ' - - - - ' - - - - - - ' 0

o 0.5 1.5 2 0 0.5 1 1.5 2
t[s]

2.5 0.05,----,-------,-------,---------,

2 .... ,.~..~.-._. _.~ .~.

v=30ni/s, fl=1
;';"":"; '';;'';.;,'~ .~ ..
o .__._.,_.._.._.-:-.._.._._.;_._.. _.-
I -0.05
1.5 . '" . Os
I ~ -0.1 - °Fcontr

0.5
·1·
I
. . . . .. -
1- - conventional
controlled
u.
": -0.15
",,(J)
- ;. °Fconv

-0.2
o "-~
.._.- - - - - - - . . . . ,
-0.5L---~--~--~---'
o Q5 1 1~ 2 Q5 1 1.5 2
t[s] t[s]

Figure 6.48. Linear simulation results at v = 30 [ru/s], j.t = 1.

Steering wheel angle step response

0.35

0.3

0.25
0;
'"u.
:s 0.2

":(J) 0.15
ec
0.1

0.05
_0.5L---~--~---~----.J 0
o 0.5 1 1.5 2 0
t[s]
Yaw disturbance torque step response
4 0.1,----------------,
,. r: ... ... v=30nits, fl=O,5 o '_"__'.-_' __".-"_'_'_"_"_'''_'''-
3 /".' .... ~'''''-'''''~'-'-'-'
I -0.1 Os
0;
2 I
I
~-0.2 - °Fcontr

I'
I

-
conventional
controlled
1- - I': ,,:u._
",,(J)
0 .3
- - °Fconv

of'.....~...,.... -I -0.4

-1 '--~~-~-----'------'
o 0.5 1 1.5 2 0.5 1 1.5 2
t[s] t[s]

Figure 6.49. Linear simulation results at v = 50 [m/s], j.t = 0.5.

6.7 Skidding A voidance Based on the Disturbance Observer 237

Non-linear Stability Analysis

Previous experimental experience shows that limit cycles with actively steered cars can
occur due to actuator non-linearities [16] . In this section, the rate limitation of the
steering actuator is considered in terms of the limitation of the electric voltage of the
motor. The Popov criterion [168] is employed to prove absolute stability for the con-
trolled car such that the existence of limit cycles can be excluded in the whole operating
domain of the car . The foundations of mapping the Popov criterion to parameter space
are described in Section 5.1. With the Popov absolute stability criterion, the Popov plot
of the system's linear part is considered, whereas for the static time-invariant non-linear
part only a sector needs to be known, which includes the non-linearity characteristics.
The transfer function of the system's linear part is the transfer function from signal
U to signal -Ul in Figure 6.37. Therefore, the actuator is considered to be connected
to the control loop as shown in Figure 6.38 with flF re! = fls + fie. The Popov plot is
derived from the locus of the linear part transfer function, whereupon its imaginary
part is multiplied with the respective frequency. The saturation in Figure 6.37 is a
characteristic that is included in a sector formed by the two straight lines U = 0 and
U = k . Ul with k = 1. According to the Popov criterion, absolute stability is proven if

0.8

- I :1.0.6
-20[
0.4
->OL--=--____:':---=----:":-~____:':,----:--'
-,
-3 - 2S
-. R'J 0.2

Figure 6.50. Popov-line (black) and Popov o"""''''--------~-_.....J

o 10 20 30 40 50
plots (grey) for six operating conditions v [m{s]
marked in Figure 6.51
Figure 6.51. Region Qe of absolute stability
using a Popov-line with slope M = 6.0

the linear part is Hurwitz and the Popov plot lies to the right hand side of the so-called
Popov-line . (This definition is adopted for 8-stability here.) The Popov-line intersects
the real axis at -1/ k = -1. The slope M of the Popov-line may be chosen arbitrarily.
We chose M = 6.0 (see Figure 6.50), which turns out to be favorable in terms of the re-
sulting 8-stable parameter region Qe. Thus, the whole operating domain of the vehicle
is included in Qe, i.e. Qe C Qas illustrated in Figure 6.51. Figure 6.50 also shows six
Popov plots corresponding to six operating points marked in Figure 6.51. Solid lines
are used for the Popov plots of the operating points indicated by cross markers . The
Popov plot belonging to the circular mark is plotted dashed , the one belonging to the
triangular marker is plotted with dotted linestyle. The latter two cases are exemplary
8-limit-stable operating points since the points are located on a 8-stability boundary in
the (v, J1)-plane. This illustrates the limit case of a Popov plot touching the Popov-line,
238 6 Case Studies in Car Steering

which was used for mapping stability boundaries into parameter space. By this ap-
proach , the robust non-existence of limit cycles in the operating domain is proven while
accounting for the presence of the saturation non-linearity. A similar proof , however.
applying the dual locus method has been shown in [16] for a different control struct ure.

Conclusions
A two degree of freedom steering controller based on the disturbance observer for vehi-
cle yaw dynamics improvement was introduced here. For robust controller design, fre-
quency magnitude specifications (B-stability) and eigenvalue specifications (f-stability)
were mapped into the parameter plane of two controller parameters. Simulation result s
based on the linear single-track model were used to demonstrate the achievement of
excellent disturbance rejection and steering command responses that match well th e
desired dynamics. In addition, considering the possibility of steering actuator rate lim-
itation due to electric voltage limitation, a Popov criterion - based non-linear analysis
was performed for proving robust absolute stability. Therefore , respective G-stability
boundaries were mapped into a plane of operating parameters.

6.8 Rollover Avoidance

Rollover accidents very often cause severe injuries or even death of th e involved road
users and are therefore in th e focus of public interest. By continuously improving active
and passive safety syst ems, the number of accidents with severe personal damage has
been reduced significantly during the last decade . However, these systems have not
been developed in the cont ext of rollover avoidance , with the result that the percentage
of rollover accidents has not decreased .
Common accident causes for passenger cars are skidding induced by braking or
accelerating on a unilateral icy road , sudden loss of tire pressure , or side-wind. For
vehicles with an elevated center of gravity, the same critical driving situations may
yield rollover even though all tires of the vehicle are far away from their saturation
limits .
Driving situations can directly induce vehicle rollover that are excessive speed when
entering a curve, severe lane change or obstacle avoidance maneuvers , or disturbance
impacts like side-wind gust s. Vehicles with an elevated CG are especially prone to
rollover. Moreover, many driver mistakes result from an overestimation of the vehicle's
roll stability, which varies due to large changes of the payload-dependent CG height.
Also, very often heavy truck drivers do not keep in mind the load, mass of th e load and
mass distribution. From common sense, it is clear that the ratio of the track width and
the height of the CG (the so-called track width ratio) is the most important vehicle
parameter affecting the rollover risk, and accident analysis results [35] confirm this fact .
In this section , a vehicle dynamics control concept for rollover avoidance of vehicles
with an elevated center of gravity is introduced. The two main goals of the cont rol
6.8 Rollover A voidan ce 239

concept are first to avoid rollover in emergency situations by combined steering and
braking control, and second to impr ove th e roll damping of the vehicle and thu s th e roll
stability, which reduces the roll overshoot in transient driving maneuvers . It is shown
that these design goals can be achieved for a wide range of varying or uncertain operating
conditions. Here, together with parametric (structured) uncertainty (vehicle speed and
height of CG), unstructured uncertainty is considered to account for the influence of
unmodelled dynamics , e.g. due to neglected flexible modes of the chassis or steering
and tir e elasticities. The desired stability and performance specifications, i.e. good
handling qualities and stability requirements, are formulated in terms of eigenvalues
(f-stability) , frequency loci (e-stability) and frequency magnitudes (B-stability) . For
controller design , an augmented single-track model, denoted single-track model with roll
augmentation, is introduced to describ e the interrelation of lateral and yaw dynamics on
th e one hand and the roll dynamics on th e other hand. In contrast to passenger vehicles,
this interrelation is particularly relevant for vehicles with an unfavorable relation of CG
height and track width in context of rollover. The active steering concept is based on
the assumption of an additive front-wheel steering actuator, i.e, the front-wheel steering
angle is composed of the steering angle commanded by the driver and an additional
steering angle set by the steering actuator. The effectiveness and robustness of th e
vehicle dynamics control structure, which is based on combined act ive steering and
braking , is shown by means of sensitivity analysis and simulation results.

Rollover Coefficient
Consider Figure 6.52. As a good indicator for the quantification of rollover risk, the
distribution of the vertical t ire loads is used here , resulting in the following definit ion
of a rollover coefficient:
(6.8.1)

Th e tir e loads left F ZL (right F z R ) front and rear are added. For straight driving on a
horizontal road , assuming symmetric load distribution, it holds th at F z R = Fz L , which
means that R = O. When Fz R = 0 (FzL = 0), th e right (left) wheels lift off the road
and the rollover coefficient takes on t he value R = -1 (R = 1). For IRI < 1, the vehicle
is termed rollover-stable. 'From th at , it can be seen that R is an intuitive measur e for
the risk of rollover.
If one wants to use the rollover coefficient as a control signal and/ or as a visual, an
acoust ic, or haptic warning signal for th e driver , it must be determined permanently
dur ing the ride. For the determination of R, different sensor signals and measuremen t
principle s may be applied.
First of all, R can be determined directly by measuring or observing the vert ical
tire loads . The spreading of air suspensions, especially in heavy t rucks, has increased
significantly during the last decade. This technology facilitates determination of th e
spring and damping forces based on the measurement of air spring pressure while also
spring and damper characterist ics are well known. Furthermore , vert ical acceleration
sensors mounted on the front and rear axles can be used for measuring a correction
term describing the influence of the unsprung mass, i.e, of tires and axles. Thu s, spring
240 6 Case St udies in Car Steering

·-- ~~1° ------

mIg
road

T
Figure 6.52. Variables for deter mination of a rollover coefficient

and damping forces plu s correcti on term ap prox imately equa ls the t ire vertical load.
However , for vehicl es with convent ional suspensions additional, effort (sensors) is
required to det ermine spring and damping forces. For t hese types of vehicles, an ap-
proximati on of the rollover coefficient may be used . Therefore, F igure 6.52 shall be
int erpret ed as a front view of a vehicle in steady state corne ring . The following as-
sumpt ions are made:
• Pi tch and plunge mod e are neglected.
• CG of chassis (unspru ng mass) is in t he road plane.
• The location of t he roll axis is assumed constant in a height h.n parallel to t he
ground .
• CG of roll bod y (sprung mass plu s payload ) is in a height h above t he roll axis.
• The payload is assumed symmet ric (no excentricity of the roll b ody CG) .
• The spring deflection of the tires is neglect ed.
The forces acting on the vehicle (which are relevant for det erminati on of t he rollover
condit ion) are t he grav itational forces m l 9 an d m 2 g , t he vert ical tire loads Fz L and
Fz R , and t he cent rifugal force of t he roll body m2 a y,2' The equilibrium condi tion for
t he vertical forces is

L F, = F zR + Fz L - (mj + m2)g == O. (6.8.2)

T he equilibrium condition of torques w.r.t. point 0 in Figur e 6.52 is

LM x = ~(FzL - Fz R ) + m 2g h sin ¢ + (h R + h cos ¢ )m2ay,2 == O. (6.8.3)

6.8 Rollover Avoidance 241

Substituting (6.8.2) and (6.8.3) into (6.8.1) yields an approximation of the rollover
coefficient :
R= m";;2 ((hR+hCOS4»a~ '2 +hsin4» , (6.8.4)

where m = ml + m2 is the total mass of the vehicle. To determine R corresponding

to (6.8.4) , the mass of the payload or mass of the roll body m2, respectively, and the
height h have to be estimated at the beginning of each ride and the roll angle has to be
measured . The latter might be done by measuring the spring deflection. Linearization
(4) small) yields

(6.8.5)

where
m2(h R + h)
aR = mgT/2 '
Further, assuming ml « m2 and the second term of (6.8.5) to be negligible compared
to the first one, yields
(6.8.6)

This approximation reflects the accident analysis results in [35), where it was stated
that the track width ratio (hR + h)/(T/2) is the most important vehicle parameter
affect ing rollover risk.
Determination of R corresponding to (6.8.4), (6.8.5) or (6.8.6) requires the knowl-
edge of the height h, which for a truck may vary largely from ride to ride due to varying
payloads. Online-estimation methods may be applied to estimate h at the beginning
of each ride as suggested in [96], [103], [1151 . After a reliable estimate of h is available,
R can be calculated subject to the lateral acceleration of the roll body at the CG a y,2
(and, if available, the measur ement of th e roll angle). The lat eral acceleration of the roll
body may be determined by interpolation between different acceleration signals (and
using h), measured, e.g. by lateral and vertical acceleration sensors.

General Considerations for Rollover A voidance

In the sequel, we will distinguish between direct and indirect rollover avoidance meaning
that the strategy for rollover risk reduction explicitly depends on R (direct rollover
avoidance) or it does not (indirect rollover avoidance) .

Direct Rollover A voidance

The vehicle becomes hard to control if front and rear wheels on one side lift off the road.
Therefore, the goal of the pursued direct rollover avoidance strategy, [158], [161]' [160),
is to avoid any wheel lift-off meaning to keep the vehicle in a rollover-stable mode, i.e.

IRI < 1 . (6.8.7)

242 6 Case Studies in Car Steering

Based on steady state considerations, the objective is to satisfy this requirement by

keeping the absolute value of the roll body 's lateral acceleration below:

T/2
lay,21< ay,RL = h+h 9 .
R
(6.8.8)

The steady state lateral acceleration limit due to rollover ay,RL may vary from ride to
ride, subject to largely varying payloads. Its value decreases with an increasing height
er cc.

Remark 6.14

Vehicles with moving or swashing load may roll over in transient driving situations
far before the steady state threshold ay,RL has been reached. For other vehicles, the
dynamical rollover limit, e.g. in transient lane change or obstacle avoidance maneuvers ,
may exceed the steady state threshold. 0

The lateral acceleration limit due to skidding (compare to (6.1.1)) is approximately

defined by
(6.8.9)

Hence, to avoid skidding and rollover at the same time, the lateral acceleration has
approximately to be limited by

layl < min(ay,sL, ay ,Rd . (6.8.10)

Also, from this formula it might be derived that for most of the passenger cars the skid-
ding limit is relevant . For these vehicles, rollover may only occur posterior to skidding.
In steady state cornering, the lateral acceleration mainly depends on curvature p and
speed v, i.e.
_ 2
ay,st - v (l . (6.8.11)

From this simple formula, it can be directly seen what has to be done in critical driving
situations (skidding or rollover). To reduce ay,st, either curvature or speed has to be
reduced. This can be done by braking to reduce speed and/or by active steering to
decrease curvature by means of driving into a wider curve (increasing radius of driven
curve). The active steering impact can be performed very fast because of the small
masses (steering wheel, column, steering rod, tires) thath have to be moved. Braking
requires deceleration of the total vehicle mass and is therefore much slower.
The strategy for direct rollover avoidance applied here is to combine active steering
and braking. Thereby, rollover avoidance is given priority over ideal lane keeping. The
active steering impact on the vehicle dynamics is prior to the effect of braking. By
active steering, curvature can be decreased to avoid rollover. However, at the same
time it has to be guaranteed that the vehicle remains on the road. Therefore , active
braking is applied to reduce speed, and which allows the return of steering command
to the driver after a little while.
6.8 Rollover A voidance 243

Indirect Rollover A voidance

Indirect rollover avoidance here means to decrease the rollover risk by improving driving
safety. Vehicle dynamics control systems using individual wheel braking (e.g. Electronic
Stability Program, ESP [195]) or active steering for stabilization of the yaw movement
as described in the previous sections have been firstly established for passenger cars
with a low CG. Some effort has been done to adapt individual wheel braking concepts
to vehicles with an elevated CG [106], [166] . These concepts can , in general, help to
avoid critical situations and thus indirectly help to reduce rollover risk. An effective
way to reduce the rollover risk in transient driving maneuvers such as lane change or
obstacle avoidance is to apply active and semi-active chassis control systems such as e.g.
ABC (active body control , [91]) or dynamic drive (ARC == active roll control) [126].
Besides these chassis control systems, active steering can also be used to shape the
vehicle dynamics such that the roll dynamics is damped and thus the transient roll
overshoot of the vehicle's body is reduced [158], [25], [26] .

Rollover A voidance Control Concept

The vehicle dynamics control approach for rollover avoidance described here aims at
two goals. One is to improve the damping of the roll dynamics by active steering . This
reduces the risk of rollover in dynamic steering maneuvers . The other task is to improve
safety in emergency situations when the vehicle is already very close to rollover. Here,
in addition to active steering, adapted braking is applied . The key idea of the rollover
avoidance control concept is that in critical situations, rollover avoidance is temporarily
given priority over ideal lane keeping.
The structure of the entire rollover avoidance control system is shown in Figure 6.53.
The system consists of the plant, which is the vehicle, sensors for measuring the vehi-
cle dynamics states, steering and braking actuators, and the vehicle dynamics control
system . The vehicle dynamics control system itself combines three feedback loops: con-
tinuous operation steering control , emergency steering control and emergency braking
control.
Subsequently, the design models for the vehicle dynamics and the steering actuator
are described and the design of the three feedback loops is shown.

Single-track Model with Roll Augmentation

The single-track model [172] has been proven to be an adequate model for investigation
of the vehicle dynamics of passenger cars with a low center of gravity. In this model, the
roll, pitch, and plunge mode of the vehicle are neglected. However, for vehicle dynamics
investigations of vehicles with an elevated center of gravity, it is necessary to at least
additionally consider the roll dynamics of the roll body relative to the chassis. For this
reason, a single-track model with roll augmentation is considered, which will serve as a
design model for the rollover avoidance control system in this section.
The vehicle is modelled as a multibody system consisting of two rigid bodies: 1)
The chassis with mass ml is composed of the front and rear axles, the four wheels
and the frame, and 2) the roll body (payload included) is the sprung mass m2' The
244 6 Case Studies in Car St eering

PD

vehicle dynami cs

Figure 6.53. Structure of rollover avoidan ce control concept

6.8 Rollover A voidanc e 245

position of th e vehicle's roll axis , i.e. the front and rear roll cent er, depend s on th e
susp ension kinematics. The model assumes a fixed roll axis parallel to the road plan e
in the longitudinal direction of the vehicle at a height h R above the street. Hence,
body 2 is linked to body 1 with a one degree of freedom joint. The roll movement of
the roll body is damped and sprung by suspensions and stabilizers with an effective
roll damping coefficient d,p and roll stiffness c¢'. The CG of the chassis, i.e. CG1 is
assumed to be in the road plane below CG2 since its contribution to the roll movement
is considered to be negligible. For the chassis , the same assumptions as for the single-
track model [1721 are used in order to represent the main features of vehicle steering
dyn amics in the horizontal plane . Linear spring, damping and tir e characteristics are
assumed. The latter assumption is permissible for vehicles and operating condit ions
where ay,RL « ay,SL '

Yo

x Xo

Figure 6.54. Model assumptions for the chassis

The mod el parameters used here are compiled in Tabl e 6.4. The data are taken
from [1331.
The multibody syste m describ es the vehicle's longitudinal, lat eral , yaw, and roll
dynamics. A similar description of a vehicle model can, e.g. be found in [179] . Applying
Jourdain 's prin ciple of virtual power , the non-linear equati ons of moti on are obt ained
according to a prop er choice of minimal velocities and minimal coordinates , respect ively..
e.g.
i [v x vy r ¢ ]T ,
Z [ J~ vxdt J~ vydt 'l/J 1> ( ,
where V x and v y are th e velocity components of body 1 in longitudinal and lat eral
dir ection, 'l/J is the yaw angle, r = "j; is the yaw rat e of body 1, 1> is the roll angle and ¢
246 6 Case Studies in Car Steering

Yl

Figure 6.55. Model assumptions for the roll body

Table 6.4. Numerical vehicle data

CF = 582 [kN/rad] front cornering stiffness
CR = 783 [kN/rad] rear cornering stiffness
c", = 457 [kN m/rad] roll stiffness of passive suspension
dJ, = 100 [kN/rad] roll damping of passive suspension
9 = 9.81 [m/s''] acceleration due to gravity
b« = 0.68[m] height of roll axis over ground
h = 1.15 [m] nominal height of CG 2 over roll axis
hx = 24201 [kgnr'] roll moment of inertia, sprung mass
J. = 34917 [kgm2 ] overall yaw moment of inertia
IF = 1.95 [m] distance front axle to CG 1
lR = 1.54 [m] dist ance rear axle to CG 1
m = 14300 [kg] overall vehicle mass
m2 = 12487 [kg] sprung mass
J.L=1 nominal road adhesion coefficient
T = 1.86 [m] track width
6.8 Rollover A voidance 247

is the roll rate of body 2 relative to body 1. Note that the triple U;
vxdt, vydt, 'l/J} J;
describes the vehicle location and orientation. It does not affect the vehicle dynamics.
The model inputs are the front wheel steering angle OF, the disturbance side force FyD,
the disturbance yaw torque M zD and the disturbance roll torque MxD, summarized to
the input vector
u = [OF FyD MZD MxD] . (6.8.12)
I
Note that any external disturbance like side-wind gusts or sudden unilateral loss of tire
pressure may be composed of these three disturbance inputs. The non-linear equations
of motion are given by
M(z) E + k(z , z) = Q(z, z, u) , (6.8.13)
where
m o hm2 sin(¢» 0 ]
M z = 0 m o -hm2 cos(¢»
(6.8.14)
( ) hm2 sin(¢» 0 M(3 ,3) 0 '
[ o -hm2cos(¢» o Jx ,2 + h2m2

M(3,3) Jz ,l + Jz ,2cos2 ¢> + (Jy,2 + h 2 m2) sin 2 ¢>,

-m r v y + 2 h m2 ¢ r cos(¢» ]
k(z,z) = mrvx + h m 2 (¢2+ r2) sin(¢» . (6.8.15)
[ r sin(¢»( -h m2 v y + 2 (J y ,2 - J z ,2 + h 2
m2) ¢> cos(¢»)
-r cos(¢» (h m2 V x + (J y ,2 - Jz ,2 + h m2) r sin( ¢»)
2

Fl,H + Fl,vcos(8F) - Fs,v sin(8F) ]

Q . _ Fs,vcos(8F) + Fl,vsin(8F) + Fs,H +FYD ()
(z ,z,u) - IF(Fs,vcos(8F)+Fl,vsin(OF))-lRFs,H+MzD . 6.8.16
[
hm2g sin(¢» - F", - FJ, + MXD
Further, the change in vehicle speed is assumed to be small compared to the lateral,
yaw and roll motion. Therefore, the speed is considered an uncertain or varying plant
parameter. Linearization w.r.t . straight driving at constant speed (vx = v = canst)
yields the linear second order equations of motion:
ME + (D + G)z + (K + N)z = Su, (6.8.17)
where

M
~~Iz st ' (6.8.18)

D+G=:DG
~: 1st' (6.8.19)

K+N= :KN =
~: 1st ' (6.8.20)

S
-;~L · (6.8.21)
248 6 Case Studies in Car Steering

From a mechanical point of view, (6.8.17) is a gyroscopic dissipative system , where

M is the mass matrix (symmetric, positive definite) ,
D the damping matrix (symmetric),
G the gyroscopic matrix (skew-symmetric),
K the conservative location matrix (symmetric) ,
N the circulatory matrix (skew-symmetric), and
S the operating matrix.
The matrices of the single-track model with roll augmentation become:

; J -h m
o
2
],
M =
[ -hm2 z
0 Jx ,2 + h 2m2
{CF+CR) I! (cFlF- CRlR)I!.+~mv
v 1 ( 2 v 2)
D (cFlF-cRlR);+"2mv cFlF + cRlR ;
[ o -~hm2v

0 0
K 0o
[ o 0

N 0,

S ~ [c;Vp ! H]
Equations of Motion in State Space Form

Now, let the state vector be defined as

X = [ 1> . ]7 .
"v r 1> (6.8.22)

Then, the state space equations resulting from (6.8.17) are

x=Ax+Bu , (6.8.23)

where

A (6.8.24)

B (6.8.25)

t (6.8.26)

and b2 = M- 1 S . This fourth order state space system describes the vehicles lateral,
yaw and roll modes.
6.8 R ollover A voidance 249

Equations of Motion in Transfer Form

From (6.8.17), the equations of motion in transfer form are derived:

...,
=:G N
(6.8.27)
where (J = vy/v is t he sideslip angle and

GN(3,3) = ct/> - m2 9 h + dJ,s + (Jx ,2 + h2 m2) 82 .

Th e transfer funct ion matrix from [ <5 F Fy D Mz D Mx D ( to [ (3 T 1; ( becomes

G = G,./ Gz , (6.8.28)

Th e inverse of G N can be determined symbolically by comput ing th e adjunct matrix.

Following Cramer 's formula, th e element [GNI] ik of th e inverse of GNI is given by

G - ]. _ adjidG N )
[ NI ik - det (G N ) 1
(6.8.29)

where the polynomial adj., (G N) is t he adjunct of t he element G N k . , which results from

G N by crossing out the k-th row and t he i-th column multiplied by (_ 1)iH.
Th e two tra nsfer functions required for controller design are t he tra nsfer funct ions
C:
from front-wheel steering angle <5 F to either roll rate ¢, i.e. F , and rollover coefficient
R corresponding to (6.8.5), i.e. C:F • Here, the lateral acceleration of t he roll body is
given by
ay,2 = v(/3 + T) - h ¢ . (6.8.30)

Operating Domain
Th e most essential uncertain or varying par ameters entering into the rollover coefficient
R are the vehicle speed 11 and t he height of CG h, which is uncert ain due to varying
payloads. For controller design, v and h are assumed uncertain in an operating domain
Q corresponding to Figure 6.56. Operating domain Q and set of vertices Qv are defined
as:

Q {q = [ v h f Iv E [v- ,v+], n « [h- ,h+]} , (6.8.31)

Qv {q = [ v h f Iv E {v- , v+} , h E {h - ,h+}} , (6.8.32)

r, r,
VI :
V3 :
qVI = [ »: ti:
qV3 = [ v+ h+ r V2 :
V4 :
qV2 = [ v+ h:
q v. = [ »: h+ (
where
»: = 20 [km/h] h: = 0.67 [m]
v+ = 100 [km/h] h+ = 1.55 [m]
250 6 Case Studies in Car Steering

1.53
V *-._'-'-'-'-'-'-'lV3
. ·· i Q .
S 1.15
·· ..
-e 0.77 ...)IE _ ._ .- _.-:-._._ .- - .
· .
V1 V2

Ol-..~---~----'-----'
o 20 60 100
v [km/h]

Figure 6.56. Operating domain Q

The robustness of the resulting controlled system w.r.t . the uncertain friction coefficient
f.L is not considered here, since a dry road is the most critical driving condition subject
to rollover. However, f.L may be considered for robustness analysis .
Further vehicle parameters may depend significantly on the payload that are the
mass of the roll body m2, the inertias Jx ,2, Jz ,2, the roll stiffness cq, and damping
coefficient d¢" the nominal cornering stiffnesses CF and CR or their ratio, and the locus
of the CG in longitudinal direction, i.e. IF and le - To account for these uncertainties,
a multiplicative perturbation model is introduced, assuming that for these parameters
only small deviations from their nominal values occur. The multiplicative perturbation
model may also be used to account for unmodelled sensor and/or actuator dynamics.
In this context, the multiplicative perturbation is applied to the open-loop transfer
function , i.e.
£(8, q) = £(8 , q)(l + We.(8}~(8}) . (6.8.33)
The weighting function' We. (8) can be determined by experiments or in simulations with
varying payloads.
In the sequel, L is assumed to be well known for low frequencies, i.e, the model
error does not exceed 10 %. For high frequencies, the effect of unmodelled dynamics
is assumed to be evident . The model error might be up to 200 % for frequencies w >
1011" [rad/s]. The weighting function We. is selected to

We.(8} = 2 8 + 0.1·2· 11" ·5 . (6.8.34)

8+2·2·11"·5

Stability
The characteristic polynomial of the uncontrolled vehicle is

(6.8.35)
6.8 Rollover A voidance 251

where

ao (c,p-m2g h) /-L (cFcR l 2/-L+ (c RlR -CF lF) mv 2) ,

al J1. (( cF cRdJ, l2J1.) + (CF (Jz + l}m) + CR (Jz + lh m ) ) (c,p - m2gh) v+

dJ, (cRlR - CFlF) mv2) ,

a2 (CF CR 12 (Jz ,2 + h2 m2) J1.2) - dJ, (CF (J z + l} m) + CR u. + lh m)) Jl v -
(Jzm (c4> - m2g h) - (cF lF - cR lR) (Jz,2m + h2ml m2) /-L) V 2,
a3 v (CFu. (Jz ,2+h2m2)+l} (Jz ,2m+ h2mlm2)) Jl+
CR u. (Jz ,2+h2m2) +lh (Jz,2m+h2m lm2)) /-L +JzdJ,mv) ,
a4 = Jz (Jz ,2m + h2m l m 2) v2 .

Note that t he parameters v and h enter in a polynomial way into t he characteristic

polynomial. Masses and inertia ente r mult ilinearly. The locat ion of t he eigenvalues of
the uncont rolled system for t he edges of t he operating doma in Q are shown in Fig-
ure 6.57, i.e. the root locus of p(s) for variation of vehicle speed v and height of CG h
along the edges of Q. For q E Q, all eigenvalues robustly lie in the left half plane. T he
eigenvalues that correspond to t he roll dynamics are shown in the magnified right plot
of Figure 6.57. T he roll damping increases with increasi ng speed but is nearly indepe n-
dent of the CG height . The natural roll frequency is nearly independent of speed but
decreases with increasing h . The eigenvalues t hat correspond to the lateral and yaw
mode are well damped for q E Q.
Table 6.5 shows damping coefficients DJ" o' natural frequencies Wo in rad/s and max-
imum real parts ao of the corresp onding roll dynamics eigenvalues of the four vertices
of the operating domain of the uncontrolled system.

4 -S
..• ~
_ .•...~, ... ,

2
..
, ...... 2 . . .
.. '. .
......S 0
_._._ ._--~.
S
... ....... ......
" . :'.:.\:
1
. '... . . .
-2
-,
' 'f;\ .
-4
-30 -20 -10 0 -2 -1 0
Re Re

Figure 6.57. Eigenvalues of the uncontrolled vehicle for the edges of the operating domain
(the right plot is a magnification of the left one)
252 6 Case Studies in Car St eering

Table 6.5. Roll dynamics of the conventional vehicle

v = 20lkm /hJ v = 100 [km/h]

D,p.o= 0.44 D,p,o= 0.52
h = 1.53 [m] V3 ; wo= 2.19 V4 ; wo = 2.11
<10 = -0.96 <10 = -1.10

D,p.o= 0.48 D,p.o= 0.55

h = 0.77 [m] VI ; wo= 3.43 V2 ; wo= 3.29
<10 = -1.61 <10 = - 1.82

Controllability
For front -wheel steering of the single-track model with roll augm entation, controllability
is lost if
(6.8.36)
where

Ab

Symboli c evaluation of (6.8.36) yields that the system is controllable for: CF =I 0, h =I 0,

m2 =I 0, J.L =I 0, J, =I 0, and Jx •2 m + h 2 ml m2 =I O. The latter condition results from
the requirement of positive definiteness of the mass matrix, namely

(6.8.37)

observabili ty

l~~, ]
Observability is lost if

det = °. (6.8.38)
cT A 3
Subs equently, th e four states are checked for controllability. Note that only the non-
trivial solutions (h =I 0, J.L =I 0, m2 =I 0, CF =I 0) are investigated:
6.8 Rollover A voidance 253

• y = (3, i.e. cT = [1 0 0 0]: (3 is observable if:

• CFcRl RlR - cFlF)jl- (CF + CR)2Jzv 2 =J O.
2(c
In the understeering case (cRlR - crir > 0), observability is lost at a velocity
2(c
2 = CFcRl RlR - cFlF)jl ...J. O.
V 0{3 )2 r (6.8.39)
CF + CR J zv 2
(
n

• Y = r , cT = [0 1 0 0] : r is observable if:
• (cRlR-cFlF)=JO.
• m2 9 h - c'" =J O.
Note that for neutral steering (cRlR = cFlF), the vehicle is not observable from
r and thus the vehicle cannot be stabilized by yaw rate feedback. Controllability
is also lost for m2 9 h - c'" = O. The roll body might be considered an inverted
pendulum mounted on the single-track model. For m2 9 h = c"', the equilibrium
becomes metastable. Note that in this case, the conservative location matrix K
in (6.8.17) becomes zero.
• y = 1>, i.e. cT = [0 0 1 0]: 1> is observable if:
• CFc Rl2(-(cFlF) + cRlR)jl- (CF + CR)2Jzv 2 =J O.
In the understeering case (cRlR - cFlF > 0), observability is lost at a velocity
2(c
2 = CFcRl RlR - cFlF)jl ...J. O.
CF + CR )2Jzv 2 r (6.8.40)
V ,.A, (
n"",

• y = 1>, i.e. cT = [0 0 0 1] : 1> is observable if:

• m2 9 h - c'" =J O.
• CFcRl2jl2 + (cRlR - crir )jlmv 2 =J O.
Observability is lost at the characteristic velocity
2 2 jlCRc Fl2
(6.8.41)
Vnexi> = VCH - m(cFlF - cRlR) .

Actuator Model
The actuator model used here has the same structure as the actuator used in the
previous section, which is given by Figure 6.37. However, in this section the actu ator
is considered to be an additional steering actuator. The data of the model used here
are compiled in Table 6.6. For the actuator transfer function Ga(s), it is assumed that
all poles have the same natural frequency W a , i.e.
w3
Ga(s) = (2 a 2)'{s+w ) .
d (6.8.42)
S +2 aWaS+Wa a
Superposition of (6.7.4) and (6.8.42) yields

and
254 6 Case Studies in Car Steering

Table 6.6. Notations and numerical values of actuator variables

Ja 0.00619 [kgm"] motor inertia
kf 0.0619 motor damping coefficient
La 0.0759 [H] inductivity
Ra 5 [0) resistance
kme = kem 0.9 electromechanical constant
ka 16.186 control gain of actuator position control

Substituting numerical parameter values from Table 6.6 into (6.7.4) results in a 3rd
order actuator transfer function with an actuator damping coefficient da = 1/V2 and
a natural frequency of 5 [Hz), i.e, W a = 101r[rad/s) . This correlates to an actuator
bandwidth of 3.7 [Hz], i.e, WB = 7.41r [rad/s], because of IGa(jWB)1 = IGa(O)I/ V2.

Roll Damping by Continuous Steering Control

In transient driving maneuvers such as lane change or obstacle avoidance (elk test)
vehicles with an elevated center of gravity tend to distinctive roll movements that may
lead to rollover. Especially steering or disturbance excitations in the range of the natural
roll frequency may yield to a swinging up of the roll body in such situations. Then ,
the driver even may feel urged to destabilizing steering commands (in flight dynamics a
comparable effect is called pilot induced oscillations [80)). In the sequel, it is shown that
besides active and semi-active chassis control systems , an active front-wheel steering
system can be used to improve roll damping, and it thus decreases the risk of rollover
in transient maneuvers .

Controller Structure
For controller design of the roll damping, a single-loop feedback structure correspond-
ing to Figure 6.58 is assumed . Proportional feedback of the roll rate ~ and the roll

Figure 6.58. Controller structure for roll damping by active steering

6.8 Rollover A voidance 255

acceleration ¢ is used , i.e.

KJ,(s , k) = kp + kd S , (6.8.43)
where k = [kp kdJT is the controller parameter vector for which the gains will be
determined subsequently. The exogenous inputs of the closed system are the roll rate
reference input ¢re/, the steering angle commanded by the driver Os, output disturbance
d and sensor noise n. The roll rate reference is assumed to be ¢re/ = O. Formally, Os
may be interpreted as external plant input signal (compare to dl in Figure 5.32). The
output disturbance d is considered representatively for the three disturbances FyD , M z D ,
and Mx D ' If good disturbance attenuation properties for d are achieved, then this result
can be directly transmitted to the three output disturbances and thus to any impact

ct
on the vehicle like side-wind .
In Figure 6.58, = CfF(s, q) denotes the transfer function

(6.8.44)

which depends on the plant parameters q = [v hjT.

The closed-loop performance of Figure 6.58 is rated by the transfer matrix of the
external inputs ¢ret> n, Os, and d to the internal closed-loop signals tracking error e,
additional steering angle 0c, front wheel steering angle OF , and controlled input ¢, i.e.
by

e]
Oc _ [S T
SK -SK -SG -S]
T -SK [¢re/]
n
[¢OF - SK -SK S -SK
T -T SG S
Os
d
.
(6.8.45)

Open-loop transfer function L and closed-loop characteristic polynomial p are given by

L(s,q,k) KJ,(s, k) Ca(s)CfF(s, q), (6.8.46)

p(s,q,k) num (1 + KJ,(s, k) Ca(s)CfF(s, q)) (6.8.47)

Sensitivity function S , complementary sensitivity T, input and output sensitivity func-

tion SK and SG of the closed-loop system are given corresponding to (5 .3.~), (5.3.9),
(5.3.10), and (5.3.11), i.e. by
1
S(s,q,k) , (6.8.48)
1 + KJ,(s, k) Ca(s)CfF(s, q)

KJ,(s , k) Ca(s)CfF(s, q)
T(s,q,k) , (6.8.49)
1 + KJ,(s, k) Ca(s)CfF(s, q)
KJ,(s,k)Ca(s)
SK(s,q,k) , (6.8.50)
1 + KJ,(s , k) Ca(s) CfF(s,q)

CfF(s, q)
SG(s, q, k) (6.8.51)
1 + KJ,(s, k) Ca(s)CfF(s, q)
256 6 Case Studies in Car Ste ering

Note that the output sensitivity function Se corresponds to the closed-loop steering
transfer function from 6s to the roll rate ~.
For controller design, a third order linear transfer function for Ga (s) corresponding
to (6.8.42) is assumed.
Design Specifications
The primary goal of the continuous operation steering control in this section is the
improvement of the roll dynamics. At the same time, it is useful to take some other
design requirements into account. The design goals considered are :
L The roll damping of the controlled vehicle shall be better than that of the uncon-
trolled one.
ii . The steering performance must not get worse (yaw damping, responsiveness) .
iii. Following a basic principle of robust control, the dynamics of the modes (modes
corresponding to roll, yaw and lateral dynamics, actuator modes) shall not change
fundamentally, i.e. a slow mode should stay slow, a fast mode should stay fast .
iv. Disturbances, e.g. induced by side-wind or J.L-split braking shall be attenuated in
the lower frequency range that is relevant for vehicle dynamics, i.e, in between
and 3 [Hz].
°
v. It follows from Bode 's intergral theorem (5.3.45) that if disturbances are attenu-
ated at low frequencies they will be amplified at high frequencies. In the frequency
range from 4-8 [Hz] oscillations of the vertical and roll dynamics feel uncomfort-
able to the driver . Therefore, it has to be ensured that disturbance amplification
is limited up to a tolerable value.
vi. Unmodelled dynamics should not have an effect on the system stability.
vii. Sensor noise should not affect stability and performance.
viii. From the outset, limit cycles induced by the actuator rate limitation should be
avoided .
ix. All above-mentioned specifications shall be met robustly w.r.t . varying or uncer-
tain operating conditions, i.e. speed v and uncertain height of CG h.
Specifications i and iii .are realized by the definition of f-stability regions as shown in
Figure 6.59. The of-parameters (10 and D4>,O are chosen separately for each operating
condition (the vertices in Figure 6.56 as given in Table 6.5). This guarantees that
damping and real parts in the closed loop cannot be worse than the respective open-
loop values for the vertex operating conditions. The eigenvalue locations of the yaw
and lateral mode , i.e, specification ii , are not considered separately. However, they have
to be taken into account for analysis.
In steady state, roll rate ~ and roll acceleration ¢ are zero and therefore it holds
that 8(0, q, k) = 0. Disturbances shall be attenuated in the frequency range [0,3] [Hz] .
Therefore, as in specification iv, an upper bound for the sensitivity function 8 is used,
i.e.
IS(jw, q, k)1 ::; 1, 'if wE [0 ,61T) . (6.8.52)
On the other hand, disturbances shall not be amplified more than 10 % for high fre-
quencies (for frequencies above the transition frequency) , i.e.

IS(jw, q, k)1 < 1.1, 'if wE (61T,00) , (6.8.53)

6.8 Rollover A voidan ce 257

Figure 6.59. Definition of q-dependent f-stability regions

which corresponds to a Nyquist-stability margin of l!s(q) > 1/1.1.

The sensors are assumed to work well in the frequency range [0,20] [Hz] . Thus,
noise rejection for frequencies w > 407r [rad/s] is guaranteed if

jT(jw, q, k)1 « 1 for w > 407r [rad/s] . (6.8.54)

To ensure robustness W.r.t. un modell ed dynamics, a multiplicative perturbation cor-

responding to (6.8.33) with weighting function (6.8.34) is taken into account. Robust
stability then requires

IWA(jw)T(jw, q, k)1 < 1, wE [0,00) , (6.8.55)

where WA is given by (6.8.34).

The specifications as shown in Figure 6.59, (6.8.52) and (6.8.55) will be used for
controller design . The corresponding f - and B-stability regions I', Bs and BT are
bounded by

CTo(q) + jw(q) ,
v'l -d (q)2 ]
for w(q) E [ 0, - ...0 CTo(q) ,
d~,o(q)

-y'w;'ax - w(qp + jw(q) ,

for w(q) E [O ,WmaxV'I---d-4>,-o(-q-p] ,
1 , V w E [0, 67r)
IW6 1(jW)I , V wE [0,00) .
258 6 Case Studies in Car Steering

Controller Design
For robust control design, the parameter space in combination with a multi-model
approach is applied . The design task is to det ermine the set of controller parameters
k p and k d for which all design specifications are robustly met . Thereby, the operating
domain is represented by its four vertices V; , i = 1, .. . , 4, i.e, by the parameter values
Vi and hi.
The I'- and B-stability regions will now be mapped into parameter space, i.e. into
the plane of controller parameters k p and k d • The desired regions in parameter space
are defined as
K ri = {k I Roots [p{ s, kp , kd , Vi , hi}] c I'( Vi , hi)} , {6.8.56}
K Bs .i {k IIS{jw, kp , i: Vi , hi}1 CBs} , {6.8.57}
K BT .; {k IITOw, k p , k d , Vi , hi}1 C BT } . {6.8.58}
Figure 6.60 shows the resulting f -stable controller parameter regions for the four vertices

..;:
0.01

0 1 2 0 0.04
kp kP
0.03

""
0.02

""
0.1
..;: ..;: "
0.01 "
" ""
.- ....., " "... .. ....... ... ..
"
0 1 2 0 0.04
kp kp

Figure 6.60. f-stable controller parameter regions for VI - V4

of the operating domain as given in Figure 6.56. For any controller parameters [kp kd]T E
K r .i , the eigenvalue design specifications for vertex V; are met . By intersection of all four
regions, the controller parameter values are found for which the eigenvalue specifications
are satisfied robustly for all vertices, i.e,
4

Kr=nKr.i'
i= 1
6.8 Rollover A voidance 259

as shown in Figure 6.61. Note that in Figure 6.60 different scalings for the plots were

0.01 0.015 0.02 0.025

kp

Figure 6.61. Region of robustly r-stabilizing controllers

used.
Figure 6.62 and Figure 6.63 depict the controller parameter regions for which the B-
stability specifications are met. Finally, in Figure 6.64 the controller regions are shown
that meet all f- and B-specifications simultaneously, i.e,

K, = KB ,i n K r ; for i = 1, .. . ,4 , (6.8.59)

where KB ,i = KST,i n KBs ,i' Note that there are no controller parameters for which all
specifications are met robustly, i.e,
4

nKi = {}.
i=l

Therefore, gain-scheduling of the controller parameters with speed v is required. Th e

procedure for controller selection is as follows. First, a controller in each region K,
is selected subject, e.g. to additional design specifications. Alternatively, the specifi-
cations already used may be tightened as illustrated in Figure 6.65. Here, the design
specification corresponding to damping is tightened exemplarily for one operating con-
dition . In the left plot , the grey-scale coded f-regions for improvement of roll damping
of 0% , 10%, 20% is shown, while in the right plot the shrinking I'-stable region is
illustrated. The resulting control law is represented by a look-up-table as shown in
Figure 6.66. Here, gain-scheduling with both speed and height of CG is assumed. Note
that if the height is uncertain, i.e. cannot be estimated at the start of each ride, gain
scheduling only with speed is sufficient (all specifications are still met) .

Eigenvalue Analysis
In Figure 6.67 damping coefficientsand maximum real parts of the roll dynamics eigen-
260 6 Case Studies in Car St eering

02 0.2

J(B S .4 J(B S ,3
,..-.

O'
~
0
-e
.:< ..;: 0

-02 -0.2

-0.4 - 0.4
-2 0 2 -2 0 2 4
kp kP

0.2 0.2

J(B S ,I J(B S ,2
0
."
.:<

- 0.2

-0.4
-2 0 -2 0
kp k:p

Figure 6.62. Controller parameter regions satisfying the spec ification s on S for VI - V 4

values are compared to those of the conventional vehicle by means of plotting the damp-
ing coefficients and the maximum real parts, respectively, versus velocity and height of
the CG 2 . The black surface plot belongs to the controlled system while t he grey one
belongs to th e conventional system. From this it can be seen th at the controller is not
only robustly f-stable with respec t to the representatives considered but also to th e
entire operating domain. Significant improvement is achieved w.r.t. both damping of
th e roll mod e and settling time of the roll mode.

Sensitivity Analysis
Figure 6.68 depicts t he Bode magnitudes of the sensitivity funct ion for VI - V4 (left)
and for a fine grid of t he operating domain (right) . The transmission frequen cy is bigger
than 3 [Hz] for all operating points considered . Since the set of suitable driver steering
excitations and disturbance impacts is assumed to lie in between 0 [Hz] and about
2.5 [Hz], this means that transient steering reduction can be assured for the entire
operating domain. Furthermore, transient steering redu ction is particularly distinctive
for the input range close to the natural frequency of the roll dynamics of the convent ional
system (marked as crosses in th e left plot of Figure 6.68). The amplificat ion becomes
at a maximum 108 %, meaning that driving comfort is only slight ly affected in th e
frequency range 4-8 [Hz]. The stability margin {!S = 1/1.08 is st ill sufficient.
In Figure 6.69 the magnitude of the output sensitivity function Be is compared
to the magnitude of the plant G, which illustrates the bandpass characteristic of the
roll damping cont rol law. Only in an intermediate frequency range of approximately
6.8 Rollover Avoidance 261

02 02

0.1 0.1

0 O·
""
-=< -=<""
- 0.1 -0.1

- 0.2 - 0.2

- 1 0 1 3 -1 0 1 3
kp kP

0.2 0.2

0.1 0.1

0 0
.:] ."
-=<
- 0.1 -0.1

- 0.2 - 0.2

- 1 0 1 3 -1 0 1 3
kp kp

Figure 6.63. Controller parameter regions satisfying the specifications on T for Vl - V4

0.2 - 2 [Hz] is the roll damping active. In steady state maneuvers (¢ ::::: 0) as well as
for high-frequency disturbances, the behavior of the controlled and uncontrolled vehicle
are nearly identical.
Absolute Stability in the Presence of Steering Actuator Rate Limitation
As in Section 6.7, limit cycles might be induced by the rate limitation of the front
wheel steering actuator, i.e. by the voltage limitation as shown in Figure 6.37. The
Popov-criterion is applied to prove for absolute stability. The transfer function from u
to Ul of the closed-loop system is given by

G(s q) = _ Ul = kemka(l + KJ,(k(q))Gtv(s ,q))

(6.8.60)
, U kemkme + (Jas + kf)(Las + Ra) .
In Figure 6.70 the Popov-line and the Popov-loci for the four vertices of the operating
domain Vl -\t4 are plotted. Since all loci reside to the right of the Popov-line, robust
absolute stability of the roll damping can be guaranteed.

Rollover A voidance by Coordinated Steering and Braking Control

With the feedback of ¢ and ~ to the front-wheel steering angle, the roll damping of
the vehicle is improved considerably. In fact, the steering transfer function has been
shaped such that the roll mode is excited less in the frequency range of the roll resonance
262 6 Case Studies in Car Steering

0.03p:====~==~;;;;;;;;:]

O.lr-'t-_~
0.02 ,,I
K~~
....
-se

\'~~~::7" ~
o 2 o 0.04
kp

0.02r-- - -L

o 0.04
kp

Figure 6.64. Simultaneous f- and B-stabilizing controller parameter regions for VI - V4

0.03 r-~-~-~-~-~---:_-""""'---,

jw

0.02

.;:
a 0.01

o
... o 0.02 0.04 0.06
kp

_____I d¢ ;:: d¢.o d¢ > 1.1 d¢.o _ d¢ > 1.2 d¢,o

Figure 6.65. Contraction of f-stable region

6.8 Rollover A voidance 263

0.4

~'77
60
1.15

100 1.53 h [m] 100 1.53 h [m]

v [km/h] v [km/h]

Figure 6.66. "Look-up-table" for gain-scheduling with speed and height of CG

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

....... :

.......... : ,'

0.77
20
60 1.15 1.15
v [km/h] 100 1.53 h [m] 1.53 h [m]

Figure 6.67. Characteristic roll dynamic values

1.2 1.2,....~--------~

••
~i~'h~J
1 r:: 1
-, -;-
~

.~ ....,
.:3
"""""
;~

.: ~\ ~::..::.~
~.8 ",: ~.8f".,,····
.~
.:
0.6 •
0.1 1 10 0.1 1 10
Frequency [Hz] Frequency [Hz]

Figure 6.68. Bode magnitude plots of the sensitivity function S

264 6 Case St udies in Car Steering

m 'l
,........... .. ... . . . . . .. . .. ..

~I
. ,
: : : :: : . : :: : : : : : : : : : : ::
2D ...
: '''.... ' ' ' ' ' ; .;.. .
OOI5= ~
'"D . : : : :: : :: . : : :: : : : : ::::
~....
. ;.. ' ...
..
...
...
..
.. .;.. "' "~''''..
. . ..
' ' ' ' '' .'...
.

~O; >.,c "" ,ii : !T!!1

.. ....
.. .. . ..
1D Y:·' : · ·~< ·'· ~ :· ; ; ; · · · ·~ ·· :··~ ·« i
: : : : :.::: :.. ::: ::::..

0 15EI
= ';" ~
; '.: :
. ...
D · ···· . D .....
: : ::: : : : : : :::. . .
D.1 1 1D D.1 1 1D
,.........,
'"D •
---..
: : :,:
.
: : : : : :: ::
.
.
:
.
.
: : : : :: ,
.
2D r- -
.. ....
IG(lw,v,hl1
ISa jw,v,h I

~O: nm T :~ " mi 1D

D
.... .
/ . . ::N
D.1 1 1D 0.1 1 1D
Frequ ency [Hz] Frequency [Hz]

Figure 6.69. Bode magnitude plots of the output sensitivity function Be and the plant G for
V1 - V4

0 ········
........
]:
'6-200
'--'
S
......
::3 -400

-5 o 5 10 15 20
Re (G(jw))

Figure 6.70. Popov-line and Popov-loci for V1-V4 considering voltage limitation of the steering
act uator
6.8 Rollover A voidance 265

frequency. Thus , the risk of causing a rollover by steering excitation has been reduced.
However, even the controlled vehicle can roll over if the steering input is large enough.
Also, in quasi-steady state driving maneuvers the roll dynamics are not excited and
thus the automatic steering system is deactivated.
Therefore, the rollover coefficient R is used for feedback. The non-linear control
introduced subsequently can be interpreted as an "intelligent" steering angle limitation
such that rollover on a plane road can be completely avoided. The key idea is that
rollover avoidance is given priority over lane keeping because a tipped vehicle is no
longer steerable. To drive the narrowest curve that is physically possible, maximum
lateral acceleration must be applied. The lateral acceleration is limited, however, by
the boundary where rollover occurs. This boundary is reached if the vehicle is steered
such that the inner wheels are just about to lift off the road, corresponding to IRI = 1.
The optimal strategy to keep the narrowest curve possible while avoiding rollover would
be to keep IRI = 1. With some safety margin , this idea is implemented in a non-linear
steering control law. Therefore, if the magnitude of R exceeds R, then the overstepping
difference OR = kn : sign(R) . (IRI- R) is fed back to the front-wheel steering angle OF
such that the curvature of the course is slightly reduced and rollover is avoided, i.e. the
emergency steering control feedback is described by the relation

OR = { k R · sign(R) . (IRI - R) V IRI > ~

(6.8.61)
o V IRI S; R.
In order to implement the prescribed effect, a dead zone element is introduced into
the emergency steering feedback loop. The black line in Figure 6.72 shows the char-

G(s, q)

f(e)
Figure 6.72. Popov-sector and the non-
Figure 6.71. Closed loop with linear and linear emergency steering control charac-
non-linear part teristics

acteristics of the dead zone with an absolute value threshold of R and a slope of k R .
This corresponds to the dead zone element in the emergency steering feedback loop
in Figure 6.53. However, this non-linear element in the loop induces the risk of limit
cycles. Therefore, a stability analysis is performed using Popov's sufficient criterion
on absolute stability as described in Section 5.1. Subsequently, mere feedback of R to
the front-wheel steering angle is considered. The control loop is subdivided into the
non-linear characteristics and the transfer function

(6.8.62)
266 6 Case Studies in Car Steering

as shown in Figure 6.71. The critical Popov-line or Popov-sector, respectively, can be

computed symbolically by evaluation of a tangent or double tangent condition, respec-
tively, or numerically by solving the optimization problem

k: = inf sup -Re ((1 + jwO)G(jw)) , (6.8.63)

6EIRwEIR;!"

which directly results from the Popov condition (5.1.13). In Figure 6.73 a comparison

28- -;n-- _ _-
60
v [km /h
100 1 53
.
1.15
h [m]
0.71 28- - -;;;:--_ _
60
-=--
100
1.53
1.15
h [m]
0 .77

v [km/h ]
Figure 6.73. Upper bound of Nyquist- (left) and Popov-sectorfright) for an equidistant grid
of the operating domain

of the upper bounds for Nyquist- (left) and Popov-sector (right) is shown. The edge
in the Popov-surface characterizes the operating conditions for which necessary and
sufficient condition for absolute stability start to differ. The upper bound kmax of the
maximum Popov-sector [0, kmax ) is computed by

(6.8.64)

The result is kmax = k+( qV3) = 0.3091 (for comparison, the numerical value of the
upper bound of the Nyquist-sector is inf qv3 k't( qV3) = 0.8031). e(qv3 ) is the value for
the critical operating point V3 with v = 100 [km/h] and h = 1.53 [m] . The slope of
the corresponding Popov-line becomes 1/0+( qv3) = 5.984. For the four vertices of the
operating doma in, i.e. q E Qv, the control loop is absolutely stable if kn E [0, k( qV3))
as shown in Figure 6.74. To ensure robust absolute stability for the continuum of the
entire operating domain , i.e. q E Q, the robust absolute stability condition is mapped
into parameter space as suggested in Section 5.1. The region

Qep.6 = {q 11+ Re((1 + jwO)G(jw, q)) > 0, V wE IRt, 0E IRo} (6.8.65)

shown in Figure 6.75 signifies the region in the plane of plant parameters q, for which
the Popov-locus for all q E Qep,6 resides to the right of the Popov-line with slope 1/0.
The control loop in Figure 6.71 is robustly absolute stable since the operating domain
is entirely included in this region, i.e, Q E Qep,6 '
6.8 Rollover A voidance 267

2Or-----.-----,.-----..,.,....-----,

....---
o
....---
is- -20
j
.:,:) -40

-
~
S
:3 -80

-100'--_ _--'- -'--_ _--:::=:..... ---'

-10 o 5

Figure 6.74. Popov-line and Popov-locus for the four vertices

o 50 100
v [km/h]

Figure 6.75. Popov region in parameter space

268 6 Case Studies in Car St eering

Combined Continuous Operation and Emergency Steering Control

For absolute stability of the closed loop with combined continuous operation and emer-
gency steering control, the linear part of the system becomes
.__ R _ Ga(s)GfF(s, q)
G(s,q.-
) - . , (6.8.66)
OR 1 + Ga(s)K~(k(q))GtF(s, q)
where controller parameters for K~(k(q)) are selected corresponding to the look-up-
table in Figure 6.66. Figure 6.76 depicts the course of the upper bound of the corre-
sponding Popov-sector for a fine grid of the operating domain Q. Thereby, the left plot
in Figure 6.76 corresponds to J-L = 1, the right one to J-L = 0.2. For all kR E [0, km ax ) with

2 6

:2
~
t 2

28- - --;;;;-- - 0.77 28 o.n

60
v [km/h]
100 60
v [km/h
100 1.53

Figure 6.76. Upper bound of Popov sector for J.L = 1 (left) and J.L = 0.2 (right) for combined
emergency steering and continuous operation steering control

km ax = infv,h e(v, h) = 0.394 (J-L = 1) or k+(v, h) = 0.372 (J-L = 0.2), the controlled
system corresponding to Figure 6.53 (neglecting braking control) is robustly absolute
stable.
Emergency Braking Control
Applying braking control requires the application of a non-linear dynamic model of
the vehicle with longitudinal velocity v as an additional state variable and the braking
pressure p or a corresponding longitudinal force Ix = m ax as an additional input
(see Figure 6.53). The force Ix is assumed to act on CG j in the vehicle's longitudinal
direction. The time delay effect of the brakes is modelled by a first order lag with a time
constant of 0.1 s. The intention of emergency braking is to make the deviation from the
desired course being induced by emergency steering control as small as possible. This
task is realized by decelerating the vehicle as soon as the rollover coefficient becomes
critical. Reducing speed means also to reduce lateral acceleration . Energy is taken out
of the system . Therefore, braking contributes essentially to rollover risk reduction. Two
non-linear effects must be taken into consideration for minimization of the deviation
from the desired path:
a) By the inertia of the vehicle mass, the loads on the front tires are increased while
the loads on the rear tires are decreased by the same amount . The front and rear
sideslip angles increase, the rear more than the front one. By magnification of the
rear sideslip angle, a yaw movement to the inside to the curve is induced [150J.
6.8 Rollover A voidance 269

b) Assuming understeering beh aviour steady st at e decelerating in cornering means

th at th e front wheel steering angle has to be taken back to remain on the desired
path, which also induces a yaw movement to the inside of th e curve.

Figure 6.77 depicts the structure of the emergency braking control system. Note th at PD

PD

T
P

I
braking act uator I ! vehicle dyn amics sta tes
t
I< >. - Pre! braking pressure-
'< )f determination

I~. It
-R Rr:1 R

Figure 6.77. Structure of the emergency braking control system

Figure 6.53 describes th e braking pressure set by the driver corresponding to th e brake
pedal force. Consider the left plot of Figure 6.78. Assuming an ideal braking act uator

-1 -R 1 R -1 -R-R R R R

Figure 6.78. Static and dynamic braking control characteristics

and no braking of th e driver (PD = 0), the automatic braking act ion is described by
the following relation:

P-
0
•
v IRI ~ R (6.8.67)
{ Pre! V IRI >R .
Altern ativel y, in a refined realization shown in the right plot of Figur e 6.78 a dynamic
characterist ic is applied to dist inguish between decreasing (Rsign(R) > 0) and in-
creasing (Rsign(R) < 0) rollover st ability. Assuming decreasing rollover stability,
270 6 Case Studies in Car Steering

braking shall be implemented as fast as possible, while in increasing rollover stability

the breaking force shall be withdrawn. Such a dynamical relation is, e.g. given by

:'1
'if IRI::; R

P= { 'if IRI > ill', Rsign(R} > 0 (6.8.68)

IRI- R
R-R Pre! 'if IRI > R 1\ Rsign(R} <0
where R denotes the (dynamic) maximum absolute rollover coefficient that is stored in
a memory while in increasing rollover stability state.
Here, the braking pressure reference value Pre! is assumed fixed, i.e, its value corre-
sponds to a predetermined deceleration.

Remark 6.15

However, Pre! also might be computed subject to other vehicle dynamics states. For
example , if the run of the road and also the road conditions are measured and available
to the driver assistance system (e.g, assuming a telematic system with electronic map
or an automated highway system), then automatic steering and braking impact can be
adapted such that rollover avoidance and lane keeping can be achieved simultaneously
without any driver impact.
o

Simulation Results
The functioning of the emergency control system are now explained by means of sim-
ulation results. The simulations were performed using the non-linear dynamic vehicle
model , assuming a dry road (J,L = I) and an unfavourably large height , h = 1.53 m.
Figure 6.79 shows the responses of the conventional (dashed-dotted line) and two con-
trolled vehicles (solid lines) when a ramp-like input signal is applied to the steering
wheel angle as. Both braking control approaches are investigated. The black solid line
corresponds to braking action due to (6.8.67), the grey line is according to (6.8.68).
This maneuver is similar to driving through a highway exit with increasing curvature
(clotoidal transition). After about 2 s, the rollover coefficient of conventional vehicle
aproaches 1 and the wheels on one side lift off the road . For the sake of comparability,
the simulation is continued until the end of the maneuver. Note that the simulation
model is no longer valid if IRI > 1.
Emergency steering and braking control is switched on after about 1.8 s when the
rollover coefficient R implies that the vehicle is close to rollover, i.e. IRI > 0.9. Due
to th e fast and precise steering intervention, the rollover is avoided. Speed is reduced
by braking control and therefore also the lateral acceleration is reduced such that after
about 1 s the steering command can be returned to the driver. However, only little
track error occurs in the vehicle's position plot (x, y) in Figure 6.79 because the vehicle
is simultaneously decelerated by the emergency braking system.
Comparably advantageous results were obtained when other maneuvers , e.g. lane
change maneuvers, and variations of v and h were investigated in further simulations.
6.8 Rollover A voidance 271

front wheel steering angle vehicle speed roll angle

60 . ' _ . _ . _ 0- _0_ .
15 /:

L ~ 40 10
S L
Ct.
'<:>
~
~ 20 -s.
;::> 5

0 . .. . .. .. . . . ... .. . 0 .'
0 2 4 0 2 4 2 4
rollover coefficient yaw rate roll rate
.' -: - 0_ . _ . _ .~ .- ._ ._ .

en
........ en
........
ex:: L L
0.5 l- .-s.

0
0 2 4 2 4 0 2 4
lateral acceleration sideslip angle t [s]
' ~' -.
40
4
N
'"
........ 820
£ 2
<:!:l..
r! -1 "" 0
2 4 0 2 4 0 20 40 60
t [s] t [s] x [m]

Figure 6.79. Simulation results for a driver steering input ramp

Concluding Remarks
The rollover control concept introduced in this section basically depends on robust
steering control. It is especially suitable for vehicles with an unfavorable relation of
height of CG and track width . By continuous steering control and using feedback of
the roll rate, the roll damping is improved significantly. Thus, especially in transient
driving maneuvers such as double lane change or obstacle avoidance the risk of rollover
can be reduced significantly. If the vehicle is already close to rollover, rollover is avoided
by additional feedback of a rollover coefficient, i.e. by suited driving state limitation.
Especially in quasi-steady state driving maneuvers, e.g. when entering a curve with
excessive speed, rollover can be prevented. By means of principal considerations it was
shown that the deviation from the desired course can be minimized by the automatic
braking impact . By combination of the roll damping and emergency actions clear
synergetic effects can be exploited.
By systematic use of the parameter space method for design and analysis of th e
rollover avoidance control system, a multitude of design specifications, covering eigen-
272 6 Case Studies in Car St eering

value (f-st ability) , frequency loci (8 -stability), and frequency magnitude (B-stability)
specificat ions were incorporated. While all design specifications could be treated ex-
actly for parametric uncert ainty (height h and speed v), th e effect of unstructured
un certainty (unmodelled dynamics) was also considered. On th e basis of the simple
controller structure, the transparency of the design pro cedure, and the results as well
as the multitude of performance and st ability criteria used, t he rollover avoidance con-
trol design procedure may be adapted easily to different types of vehicles.
For more details the reader is referred to [158).

6.9 Patents

The authors hold several patents on car steering methods and their implementation.

i. The basic idea of robust unilat eral decoupling for a car with linearized t ire and
ideal mass distribution is patented by [5). It involves th e feedback of the yaw rate
r to the front steering angle bF .
ii. For cars with rear-wheel steering bR , th e yaw damping can be impr oved by feed-
back of r do bR , [7] .
iii. A lateral acceleration a yp can be generated more quickly by feedback of a yp to bF ,
[8) .
iv. Modelling restriction s of i -iii have been relaxed in [9]. Th e method is generalized
for a non-linear uncertain t ire model with delayed generation of lateral force, non-
ideal mass distribution , acceleration and braking effects , feedback of longitudinal
velocity and longitudinal and lat eral accelerations. A desired function of th e yaw
damping on the velocity is implement ed and the fadin g integrator replaces the
ideal integrator.
v. As an alte rnat ive to robu st unilateral decoupling, a disturbance observer feedback
structure is assumed in [101) .
vi. The concept of assigning a desired dependence of yaw damping on th e velocity by
a double cancellation is developed in [12) for a car with front-wheel steering only.
vii. While items i to vi deal with skidding avoidance , [160] uses feedback steering for
rollover avoidance .
viii. A hardware implementation of mechanic al addition of st eering angles is describ ed
in [17].
ix. The electrical addition of steering angles requires a steer-by-wire system. Thi s
should give the same feeling for t he driver as a convent ial steering syst em [162].

6.10 Automatic Car Steering

In t he previous sections, the pat h tracking task was left with t he dr iver. Consider
now automatic following of a lane reference. Such a reference may be provided by
6.10 Automatic Car Steering 273

the electromagnetic field of a guiding wire in the center of the lane [69], [190] , or
equivalently by passive magnets on this lane center [196] or by processing the images
from a car-mounted video camera [74], [143],[77]. In this section, we assume that the
lateral deviation of the vehicle from the lane reference is available as a controller input .
In order to study automation of car steering, the steering model must be extended .
The model must include not only velocities, but also the vehicle heading and the lateral
position of the sensor with respect to the reference path. For simplicity, this extended
model will only be derived using a linear model that is valid for small deviations from
a stationary circular path. It is assumed that the reference path consists of circular
arcs. Figure 6.80 shows the transition from an arc with radius R 1 and center M 1 to an
arc with radius R2 and center M 2 . At the transition point, the tangent to the path is
continuous . There is, however, a step change in the reference input from Rre! = R 1 to
Rre! = R 2 • For straight path segments, the radius is Rre! = 00. It is more convenient to
introduce the curvature ere! := 1/Rre! as the input that generates the reference path.
The curvature is defined positive for left cornering, and negative for right cornering.
M1

Figure 6.80. The reference path is comprised of circular arcs

The vehicle motion in circular cornering will now be modelled for small deviations
from a stationary circular path, see Figure 6.81. The radial line from the center M
passing through the center of gravity CG of the vehicle intersects a point zre! on the
desired path. The distance from the reference point zre! to the CG is the deviation
YCG . Figure 6.81 shows an inertially fixed coordinate system xo, Yo and a car body fixed
coordinate system x, Y, which is rotated by the yaw angle 'l/J. The tangent to the path
at zre! - denoted by Vt - is rotated by a reference yaw angle 'l/Jt with respect to xo.
A model for the rate of change of YCG will now be developed. The component of
the car velocity v that is perpendicular to Vt is equal to the rate of change of YCG .
This perpendicular component is given by vsin(,B + ti.'l/J) , where,B is the car sideslip
angle and ti.7/-' := 'l/J - 'l/Jt is the angle between the tangent to the path at zre! and the
centerline of the car, see Figure 6.81.
With the linearization sin(,B + ti.'l/J) ~ ,B + ti.'l/J, the deviation YCG changes according
to

YCG = v(,B + ti.'l/J) . (6.10.1)

274 6 Case Studies in Car Steering

Guiding

Figure 6.81. Vehicle heading and measured displacement y from the guiding wire in stationary
circular cornering

Actually, the sensor S is not mounted in the CG but in a distance is in front of the
CG with es « Rre/ ' The measured displacement y from the guiding wire now changes
both with iJCG and under the influence of the yaw rate r = tiJ. Taking this into account,
the rate of change of the measured displacement is
iJ = v{{3 + !:i:lf;) + esr. (6.1O.2)
Determination of iJ requires knowledge of three variables {3, r , and ti.tf;. The variables
{3, and r are given by the basic car model (6.4.4). (Assume only front-wheel steering,
i.e. 6R == 0.) The angle ti.tf; will be obtained by integrating its derivative
ti.tiJ tf; - tf;t

The term r.t is the yaw rate of the path tangent , i.e. r.t = vi Rre/ = v(Jre/ in stationary
circular cornering. Hence,
ti.tiJ = r - v(Jre/ ' {6.1O.3}
Combining {6.4.4} , {6.1O.2}, and {6.1O.3} , the extended state space model is obtained
as
6.10 Automatic Car St eering 275

l/~mV l/me0ReDP ] [FYD] .

+ 0 M (6.10.4)
[ zD
o
In addition to the eigenvalues of (6.4.4), there is a double eigenvalue at S2 = O. The
reference curvature {Jre! appears as an additional input to the system . The transition
to a new curvat ure corresponds to a step input in {Jre! ' Fig. 6.82 illustrates (6.10.4) in
the form of a block diagram.
{Jre!
v

(J

l..-.----.l es 1-_ _...1

Figure 6.82. Block diagram of auto matic steering with reference cur vature (Jre!

By aggregating the single-track model, th e block diagram Figur e 6.83 is obtained.

The transfer function from the steering angle OF to the measured displacement from
the guiding wire y satisfies
n (s)
y(s) = ( 2) 20F(S ), (6.10.5)
ao+a\s+ s s
with ao and a\ given in (6.4.23). Th e numerator n(s) can be reduced to nF{3 and
nFr according to Fig. 6.82.

(6.10.6)
276 6 Case Studies in Car Steering

n(s) = S2{ (; + £:)nFr(s) + ~nF/J(s)} ,

and with nFr(S) from (6.4.33) and nF/J from: (6.4.34)

n(s) b2s 2 + bls + bo, (6.10.7)

j1,C
F S U
-'---- +v,
m£R£DP
j1,CF [v + j1,CR£ (US + 1)]
m£DP mV£R£F £R£F '
j1,2 CRCF e
bo
m 2£ F f R f D p ·

erel

fJ
Single-track
model r

Figure 6.83. Aggregated block diagram with external reference input

Controllability

The pair (A ,b) in (6.10.4) is written in 2 x 2 block form as

A = [All
A 21 A 22
0] ,b = [ b0
l
] , (6.10.8)

where

Au

o
[ V
00] , A~2 o.
6.10 Autom ati c Car Steering 277

Then

Ab [ Aoll ]b I,

The first two columns of the controllability matrix become linearly dependent, i.e.

(6.10.9)

und er the conditions already an alyzed in (6.4.35) for th e single-track model. Then,
also Ailb l and A~lbl are linearly dependent on b l. Further condit ions of non-
controllability can only arise if

d = det[(A 2IA ll + A 22A2db l (A 2IAil + A22A 21Al1)bd = 0

i.e, for th e same condition as ab ove and for

(6.10 .10)

which is not satisfi ed for J1- =I o.

Observability

For th e output y , the rows of th e observability matrix are

cT [0 cf], cr = [0 1] ,

cT A
3
cr [ A 21Ail + A 22A21Al1 0] .
The structure of th e observability matrix is
278 6 Case St udies in Car St eering

The last two columns relat ed to t he position state variables t::, Wand yare linearly
independent for v =I- O. Th e first two columns complement t he rank of t he observability
matrix t o 4 if

which is true for Jl =I- O.

The observability of (J and r can be improved considerably by yaw rate feedback. It
yields velocity-independent observability as shown in (6.4.41). The resulting feedback
structure is illustrated by Figure 6.84.

{lrej----......,

y
y
Actuator

'------I Yaw Rat e Feedback 1+-- ---'

Figure 6.84. Feedback structure for automatic car steering

The measured output y is the deviation between th e reference position Yr e! and the
absolute position Yabs ' Therefore,
only a single degree of freedom controller is feasible for feedback of y . A further
feedback path is obtained by a second displacement sensor in a distance f S R behind
the CG (if the second sensor is mount ed in front of the CG then f S R is negative). The
block diagr am Figure 6.83 is augmented and modified to Figure 6.85. The front sensor
quantities are now characterized by the index F .

Finally, OF may also be measured and fed back to the actuator input. This is helpful
for actuator uncertainties or if the actuator poles have been shifted to an unfavorable
locat ion by t he yaw damper and t he path tracking cont rollers.
6.10 Automatic Car St eering 279

{J
t-r--+<:>------ YF
Single-track
model

L..--+<>- YR

Figure 6.85. Two lateral displacement sensors for YF and YR

Remark 6.16

In the early days of automatic steering around 1980, rate sensors were still expensive
and gyros had not yet been considered . The front displacement sensor for YF is always
used. Typical sensor concepts for additional sensors are:
1. MAN [190] used YR for estimation of the road curvature, a sensor for bF and gain
scheduling by v .
2. Daimler-Benz started with front and tail displacem ent sensors and a sensor for
bF' Later, the system was reduced to just the front displacement sensor [69].
3. The sensor concept of Figure 6.84 with front displacement and yaw rate feedback
was proposed in [28].
4. In the PATH project [99], front and tail displacement sensors were used and
curvature preview is encoded in the road using binary polarity coding of the
magnetic lane reference markers.
o

Path Tracking Controller Design for the Bus

The aim of path tracking is to keep Y small at all times under the influence of
changing reference curvature (!rej , disturbance torques M z D , and lateral disturbance
forces FyD . The control system must be robust with respect to uncertain road-tire
contact, mass and velocity.
Like in Fig. 1.5 for th e crane , the controller in Figure 6.84 may still be non-linear,
adaptive, etc . In the following discussion , however, we design linear controllers. Th e
next question is then, what dynamic order of the controllers is required? The actuator
is modelled simply as a first order low-pass 1/(1 + Ts) for actuation at the steering
column or 1/s for force addition at the wheels by an hydraulic actuator without position
feedback [69]. For the latter case, the plant plus actuator transfer function is
280 6 Case Studies in Car St eering

= b(s2 + bjs + bo) ()

G(s) 2 ) 3' 6.10.11
(s +ajs+ao s
Figure 6.86 shows the root locus for proportional feedback of YF for the data of th e
bus in Table 6.1 at v = v+ and J.l = u> . The sensor position is f s = 6.12 [m] .
jw 2.0
,,
1.5

1.0 0.03 ' \

f07
0.5
,
\
& ~\
<;;:)":> <;;:). <;;:)';3 I
0.0
I
I

X
-0.5 I

-1.0

-1.5
I
I

-1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 2.0

a
Figure 6.86. Root locus of an automatic car st eering syst em with proportional feedb ack of
the front displ acement from the guideline

The transfer function (6.10.11) has a tripl e pole at s = O. Thus , th ere are three
root locus branches at zero with breakaway angles ±60° and 180 and th e closed loop is 0

unstable for small loop gains, see Figur e 6.86. The loop is also unst able for large loop
gains , because the asymptotes of the root locus have angles ±60° and 180 In order to 0
•

stabilize th e system at intermediat e gains, it is necessary to place two fine-tun ed zeros in

the left half plane that at tract the two right half plane branches of the root locus or allow
the two plant zeros to attract these branches. The first alt erna tive is favorable in view
of r-stabilization where ar is the dotted hyperbola in Figure 6.86. Thus, the controller
ord er must be at least two. In the road tests , it turned out that high-frequency noise
enters with YF originating from inaccuracies in the guiding wire reference. Therefore a
path tracking controller with relati ve degree one and a Butterworth pole configura t ion
was assumed , i.e,
2
1:
UF S
( )
= k j + k2 s + k3 s ( )
Ys . (6.10.12)
(1 + s/wo + S2 /w~)(l + s/wo)
For frequencies w < wo, the gain k j may be viewed as proportional gain, k 2 as derivative
gain , and k 3 as second derivative gain . During th e design, primarily k j , k2 , and k3 were
varied , i.e. a pair of zeros and the loop gain.
This controller structure is:
6.10 Automatic Car Steering 281

i. Lean: only four essential parameters k I , k2 , k3 and Wo must be tuned.

ii. Transparent: the parameters have a well-defined influence on the car steering
system. The controller is of the minimal required order.

Robustness with respect to uncertain operating conditions will be achieved by pa-

rameter space design.
The specification for the closed loop are given in the time-domain:

i. Less than 2 [em) track deviation on a straight lane without wind , less than 15
[em) under sidewind of 20 [m/sec].
ii. Less than 15 [em] peak deviation after entering a curve.
iii. Observing the 161 :5 23 [degree/s] steering rate constraint also in the worst case of
entering a narrow parking bay.
iv. Smooth transition from manual to automatic control.
v. Keep the lateral acceleration below 2 [m/sec 2 ] for passenger comfort .

A first design iteration showed that these specifications can be met with all poles in
a r-stable region to the left of the hyperbola

( ~) 2 _ (~)2 = 1 , a < -0.35.

0.35 1.75

Remark 6.17

A design based on three sensors for YF, YR and 8/ and an actuator transfer function
4.7/(8 + 4.7) was performed in [32]. The approach was to intersect r-stable regions
for the four vertex operating conditions. This was done in an invariance plane that
permitted only the shifting of the double eigenvalue at zero. This invariance plane
was chosen for an average velocity v = 10 [m/s] and the worst case {L = /l-m- [m" =
0.5·9.95/16 = 0.31. The unmeasured state variables flF and fiR were approximated by
low-pass filtered differentiation of YF and YR. For details of this design the reader is
referred to www.op.dlr/FF-DR-RR/paradise.
o

Design with Front Displacement Feedback

The design with only one sensor for YF, an actuator transfer function 1/8, and the
controller (6.10 .12) was performed in [154]. The approach was to begin with a high
controller bandwidth Wo = 100 and reduce it stepwise. This requires an adjustment
of k I in order to obtain a reasonably large simultaneously r-stabilizing region in the
(k2 , k3)-plane. This process led to the values Wo = 25, k I = 0.6. The four r-stable
regions in the (k2 , k3)-plane for the vertices of the Q-box are shown in Figures 6.87-
6.90. The set of simultaneous r-stabilizers in the intersection of the four operating
points is shown in Figure 6.91.
 282 6 Case Studies in Car St eering

k3 2.0 k3 2.0
1.8 1.8
1.6 1.6
1.4 1.4
1.2 1.2
1.0 1.0
0.8 0.8
I
0.6 0.6 I
I

I
0.4 0.4 I
I
0.2 0.2 I
I
I
0.0 0.0
I
-0.2 I
I -0.2 I
I
I I

0.5 1.0 1.5 2.0 0.5 1.0 1.5 2.0

k2 k2
Figur e 6.87. Set of r-st abilizing con- Figure 6.88. Set of r -stabilizing con-
trollers for m = 9950 [kg] , v = 3 [m . s-l], trollers for m = 9950 [kg], v = 20 [m . s- lj,
Wo = 25, k[ = 0.6 Wo = 25, k1 = 0.6

k3 2.0 k3 2.0
I
I
1.8 1.8 I
I
1.6 1.6 I
I
I
1.4 1.4 I
I
1.2 1.2 I
I
1.0 1.0 I
I

I
0.8 0.8 I
I
0.6 0.6
0.4 0.4
0.2 0.2
0.0 0.0
I I
-0.2 I -0.2 I
I I
I
0.5 1.0 1.5 2.0 0.5 1.0 1.5 2.0
k2 k2
Figure 6.89. Set of r-stabilizing con- Figure 6.90. Set of r-stabilizing con-
trollers for m = 32000 [kg], v trollers for m = 32000[kg] , v = 3 [m· s-lj ,
20 [m . s-l], Wo = 25, k 1 = 0.6 Wo = 25, k 1 = 0.6
6.10 A utomatic Car Steering 283

From the set" we pick the point k2 = 0.7 and k3 = 0.15. The resulting controller is
then
0.6 + 0.78 + 0.158
2
G (8) = 253 (6.10.13)
C (8 + 25)(8 2 + 258 + 625)"
This controller I'<stabilizes the plant at least for the extremal operating conditions.
In a stability analysis, the I'-stability of this controller is verified for the whole range
of operating conditions [231 . The stability boundaries in the (v,m = m/J.L)-plane for
this controller were already shown in Fig. 4.5. The vertices are stable and none of
the boundaries intersects the Q-box . A real root boundary passes by very close to the
Q-box. This happens, however, for low velocities and is therefore not critical because a
sluggish time response resulting from a real pole close to the origin means only a short
driving distance at low speeds.

Remark 6.18

The result of Fig. 4.6 is amazing , because I'-stability of the four vertices guarantees
f-stability for the entire Q-box. In Section 9.1, an example is constructed with one co-
efficient containing a bilinear term qlq2' This non-linearity destroys the nice properties
of linear parameter dependence (see Chapter 8) it does not suffice to check vertices or
edges of the Q-box . The track-guided bus example with ql = V, q2 = m] J.L has the terms
qll qr, ql q2, QrQ2 , QrQ~ in the coefficients of the characteristic polynomial. It looks much
more nasty than the above bilinear example, but Fig. 4.6 shows that it is well behaved.
o

8K(2)
,,
0.20 r

0.15
,lKC
8K(3)
r
X
,,
~
,
'8K(l)
r

0.6 0.7 0.8

Figure 6.91. Simultaneously stabilizing set for the extremal plants for Wo = 25, k1 = 0.6
The controller has now to be tested in simulations where the design specifications
are checked. A simulation of the transition from manual to automatic steering is shown
 284 6 Case Studies in Car St eering

y[m] 0.16
0.14
s, [rad] E-2
1.5
0.12
1.0
0.10 0.5
0.08 0.0
0.06 -0.5
0.04 -1.0
0.02 -1.5
0.00 -2.0

20 40 60 80 100 20 40 60 80 100
Track [m] Track [m]
Figure 6.92. Simulation of a transition from manual to automatic steering for v = 20 [m . s-1]
and in = 32000 [kg].

in Figure 6.92. This case is especially critical because here the largest deviations from
the guideline may occur. Additionally, this is the only operating case where a step
input appears. The controller is not started from a zero initial condition, instead, the
controller is connected to the plant before switching. Then, only the controller output is
switched to the actuator. Thus , no differentiation of a step input occurs. The simulation
was made for an initial deviation of 15 [cm]. At t = 0, the driver switches to automatic
steering. Also, in other cases like crosswind, entering a curve, etc. the controller shows
satisfactory behavior.

Design with Feedback of Front Displacement and Yaw Rate

In a later design study [28], it was shown that the performance of the automatic
steering system of the previous section can be considerably improved by additional yaw
rate feedback. The feedback struct ure is shown in Figure 6.84. The design idea is to
exploit the simplicity of the robust decoupling concept as illustrated by Figure 6.8.
Specify accurate and robust path tracking not for the sensor at the front bumper (f s =
6.12 [m] for the bus) but for the decoupling point (fop = 5.5 [m]), for example , by
placing the magnetometer in this position (f s = 5.5 [m]).
The lateral acceleration ayDP is integrated twice to the lateral position y( s)
ayDP(s)/s2, such that the transfer function from 8F(~8s) to y is, by (6.5.13):

(6.10.14)

and for the input uref(s),

(6.10.15)

This "robustified" system is much easier to control by the path tracking controller
than the bus without yaw rate feedback, i.e. with the transfer function (6.10.11).
6.10 Automatic Car Steering 285

For alternative parameter space designs of automatic steering systems the reader is
referred to [19], [99].
7 Case Studies in Flight Control

Models of aircraft dynamics and their linearization can be found in standard books on
aircraft dynamics and control, e.g. [81], [142) , [189) , [60) . When treated as a rigid body,
an aircraft has six degrees of freedom of motion: three coordinates of the center of
gravity position and three rotation coordinates for roll, yaw, and pitch. In stationary
flight, two subsystems may be considered as decoupled:

L The longitudinal motion in pitch, altitude, and longitudinal position.

ii. The lateral motion in roll, yaw, and sideward position.

The first case study deals with a severe perturbation in the yaw motion by an engine
fault (called EO=engine out). The normal flight control system cannot compensate this
effect, thus the pilot has to take over. That may take two seconds and significant effects
in the roll and sideward position occur during this reaction time. The strong pilot input
for reduction of the yaw rate then induces high peak loads (shear forces) at the root of
the vertical tail. These shear forces constitute design loads and essentially determine the
static design of the vertical tail and thereby its structural weight. An automatic control
system can alleviate this load by the same robust yaw-lateral decoupling concept that
was used in Chapter 6 for car steering . By a fast reaction , the disturbance compensation
occurs essentially in the yaw degree of freedom as a full simulation with a flight test-
validated high-precision simulation program proves.
The other two case studies deal with the longitudinal dynamics. They are char-
acterized by a slow oscillation of the CG around its trajectory (phygoid) and a much
faster pitch motion of the fuselage around its CG. The latter is called the short period
longitudinal mode. Both modes are weakly coupled. We consider here only the second
mode. It can be measured by a gyro and an accelerometer and it can be controlled
by elevator and horizontal canard rudders . The use of canards is efficient from a flight
mechanics point of view. However, they cause instability of the uncontrolled aircraft in
some flight conditions. Thereby, the robust pitch stabilization is a demanding problem.
The second section on robust and fault-tolerant Gamma-stabilization of the short
period longitudinal mode gives a practical motivation for the simultaneous Gamma-
stabilization of four representative flight conditions (three of them unstable) by a fixed-
gain robust controller. Paralleled sensors are used in order to maintain robust stabi-
lization immediately after a sensor fault .
The third case study investigates the task of designing a large envelope pitch-
rate (i.e. short period longitudinal) controller for a high-performance aircraft, and
7.1 Aircraft Load Alleviation by Robust Yaw-lateral Decoupling 287

demonstrates the use of I'- and B-stability specifications to design a controller that
is robust to both parametric and unstructured uncertainty. This is a challenging prob-
lem, as the controlled pitch-rate response must satisfy strict performance requirements
despite the fact that the aircraft dynamics vary significantly throughout the flight enve-
lope. The design uses an analytical model of the short period equations of motion that
describes the dependence of the aircraft 's pitch dynamics on altitude and Mach number
over a large flight envelope. In addition to the parametric " uncertainty " introduced
by this model, the design also incorporates unstructured uncertainty to account for
unmodelled dynamics (e.g. flexible modes of the aircraft) . f- and B-stability specifica-
tions are mapped into parameter space to obtain a robust controller that guarantees the
required performance throughout the design envelope despite the model uncertainty.

7.1 Aircraft Load Alleviation in case of an Engine

Out by Robust Yaw-lateral Decoupling

Engine out, abbreviated EO, is a design-critical operating condition for large trans-
port aircraft from the viewpoint of flight safety, handling qualities, robust design of the
electronic flight control system, comfort, structural dynamics (structural loads), and
static aircraft structure design. As the EO constitutes a yaw disturbance torque , see
Figure 7.1 for the stable lateral aircraft motion, the EO by itself is not harmful to the
aircraft. Nevertheless, large magnitudes of the yaw rate, the sideslip angle and the roll
rate occur. Several authors [85], [110], [141], [46], [129] , [127] have revealed the pot en-

Disturbance
YawMomentM

Figure 7.1. Disturbance torque from an engine out results in a shear force at the vertical tail

tial of modern control systems and electrohydraulic actuation technology to increase

the already high safety standards and to fulfill stronger requirements than those which
the above-mentioned disciplines impose today. Favre [85] discusses the integration of
the EO problem into the design of the standard lateral command and stability aug-
mentation system (CSAS). The CSAS is based on the measurement or the observation
of the yaw and the roll rate, the bank angle [140] and , optionally, the sideslip angle.
288 7 Case Studies in Flight Control

The aircraft is safely stabilized after the EO. The corresponding equilibrium condition
is characterized by significant magnitudes of the yaw rate, the sideslip, and the bank
angle, which clearly indicate the EO condition and cause the pilot to interact early, but
without overreaction in order to stabilize the aircraft with zero yaw rate and low bank
angle and to recover the original flight path. Hopkins [lID] and McLean [141] use stan-
dard controller design methodologies to design a command and stability augmentation
system, which is only activated after the detection of the EO. The system is therefore
termed the EO controller. In contrast to Favre [85], no significant magnitudes of the
yaw rate, the sideslip, or the bank angle occur . McLean [141] focuses on a general avi-
ation aircraft. As Hopkins [110] describes a new system of the Boeing B777, only a few
details with regard to the design and the architecture of the system are given. Benani,
Magni and Terlouw [46] and Lambregts [129] integrate the EO control problem into the
autopilot design. As the autopilot is based on more measured or estimated quantities
than the CSAS (e.g. quantities characterizing the flight path), good performance can
be achieved more easily than in the case of the command and stability augmentation
system . The above authors already cover the requirements of the initially mentioned
disciplines handling qualities, safety and comfort . Benani, Magni and Terlouw [46] and
Lambregts [129], moreover, tackle the robustness requirements.
Here, we investigate whether the aspects of structural dynamics (load alleviation
at the vertical tailplane) and thereby of the static aircraft structure design can be
integrated into the design of an EO controller. During an EO and the corresponding
corrective pilot action, large loads, i.e. shear forces, bending and torsion moments occur
at the vertical tailplane. In fact, the EO case is a sizing case for the structural dynamic
and the static design of the vertical tailplane. One third of the structural dynamic
simulation time is spent on this case because the changing mass and mass distribution
and the uncertain aerodynamic properties, in particular due to variations of the velocity,
the altitude, and the flap-jslat-setting, have to be covered. A considerable saving in
structural weight of the vertical tail plane is possible if the maximal loads are reduced .
The main contribution to the loads at the vertical tailplane are the aerodynamic forces
due to the sideslip angle, the rudder deflection and the inertia term [135].
Consequently, for the structural loads computation and the resulting design of
the vertical tail plane , a worst case rudder deflection commanded by the pilot and
a worst case sideslip angle have to be considered . According to the Joint-Aviation-
Requirements JAR 25.367 [114] , [135] the pilot begins to react at the first maximum
of the yaw rate, but not earlier than 2 seconds after the EO. In consequence of the
2 seconds delay, a large sideslip angle has built up. The pilot reaction is then to re-
duce the yaw rate by the maximal rudder deflection rate until the derivative /3 of the
sideslip angle passes zero, i.e, they pushes the pedal into the stops as quickly as they
can. Consequently, the rudder can even reach the rudder travel limitation. Afterwards,
the pilot reduces the rudder deflection angle to a value that is approximately required
to maintain /3 = O. Thereby, the essential characteristics of a human controller in a
disturbance rejection problem are covered:

1. The human controller is possibly too slow due to their reaction time .
2. In a delayed corrective action, they may even overreact and make a critical situ-
ation worse.
7.1 Aircraft Load Alleviation by Robust Yaw-lateral Decoupling 289

The same aspects can be observed in a yaw moment disturbance rejection in the case
of a car. An example is the skidding of a car on an icy road. The driver reacts when
already a considerable yaw rate has been built up and an overreaction may drive the tire
force into saturation. Such a safety-critical situation can be avoided by an automatic
yaw control system, with feedback of the yaw rate to a corrective steering angle that is
added to the driver-commanded steering angle. This control system will compensate the
disturbance torque faster and more accurately than the human driver, see Chapter 6.
This control concept is transferred to the EO problem of the aircraft, constituting
of a feedback of the yaw rate to the rudder: due to the early feedback of the yaw rate
measurement after the EO, such a system can react faster and more accurately to the
disturbance torque than the pilot. The yaw rate after the EO is kept small, there is no
reason for a corrective reaction of the pilot, and there is very little energy transfer to the
roll motion . Moreover, the shear force reduction at the vertical tailplane is explicitly
considered in the controller design.
The essential aspect of the control concept is that the effect of a yaw disturbance
torque is compensated so fast in the planar lateral motion that in the closed-loop
system the coupling into other degrees of freedom is small, in particular, the yaw rate
(which is an unavoidable initial consequence of the EO yaw disturbance torque) should
not influence the lateral acceleration. The planar lateral motion consists of the lateral
translational motion and the rotational (yaw) motion, which are bidirectionally coupled.
Within this planar lateral motion , the controller achieves a separation into the following
two subsystems:

• A first order system for generating the lateral acceleration at a decoupling point
which is close to the vertical tailplane.
• A second order subsystem involving the yaw rate.

In the controlled system, there is only a coupling from the lateral acceleration to the
yaw rate, but not vice versa. The system is unilaterally decoupled in the sense that
the yaw rate is made non-observable from the lateral acceleration at the decoupling
point, see Section 6.3. Hence, the yaw rate in consequence of the EO will not influence
the lateral acceleration . Exploiting the design freedoms in the controller allows us to
achieve a yaw rate that has approximately no influence on the shear force.
An important requirement for both the car and the airplane application is the ro-
bustness of the unilateral decoupling controller . For the car, the control concept allows
us to achieve robustness with respect to the uncertain tire side force characteristics,
the mass and the velocity. For the aircraft, robustness has to be achieved with respect
to the uncertain aerodynamics, in particular due to the variation of the velocity, the
altitude and the flap-/slat-settings, and with respect to the changing mass.
In contrast to the car, the aircraft is already equipped with a lateral control system .
As this system is designed to use only moderate control surface deflection angles [85], the
drastic EO controller is switched on with full authority after the detection of the EO. Si-
multaneously, the standard controller is immediately switched off. For simple detection
mechanisms, detection times up to 500 [ms] have to be assumed . More sophisticated
detection mechanisms yield a detection time of 20 [ms] .
290 7 Case Studies in Flight Control

Robust Unilateral Decoupling

For cars, it is usual to describe the steering motion in a plane on road surface level. In
the case of an aircraft, the planar motion is strongly coupled with the roll motion due
to the vertical tail (fin) and the different lift of the left and right wing due to different
velocities under a yaw motion. This coupling is important for stability and manoeuvra-
bility [82], [140] . The disturbance torque of an EO fault and the compensating torque
from the rudder deflection occur essentially in the body-fixed (x, y)-plane [135) . A de-
sign goal is to prevent the energy resulting from the unbalanced engine after the EO
from being transferred from the yaw motion into the roll motion. This should not affect
the yaw motion , which has been commanded by the pilot .
Only the disturbance-induced yaw motion and the corresponding disturbance-
induced energy flow should be kept low. To achieve this , the coupled yaw and lateral
translational motion resulting from EO has to be kept small. In order to distinguish
between a yaw rate that has been commanded by the pilot, and a yaw rate in conse-
quence of an engine out, the EO controller has to be combined with a prefilter generating
the reference yaw rate rre/ from the pilot command . The disturbance is already com-
pensated within the planar lateral motion after having induced only small deviations
from the reference yaw rate. If this goal is to be achieved, then this is an a posteriori
justification 'for using a planar model to derive the controller structure.
If the disturbance only induces small amplitudes in the closed-loop system , then
even non-linear effects [82], [140] can be neglected in the controller design. Such a
linear planar model allows us to design a simple controller for the aircraft in detailed
analogy to the car. Due to the superposition principle of linear system theory, the
prefilter design can be straightforwardly done after the controller design, so that only
the case Tre] == 0 is considered here for brevity. The prefilter is analyzed in non-linear
simulations in [127]. The planar model is illustrated by Figure 7.2. In the case of the
car, the starting point of the controller design are the lateral forces between the tires
and the road , see Chapter 6. They are concentrated at the front and the rear axes, Fy,F
and Fy,R in Figure 7.2. The corresponding lever arms are IF and IR.
In the case of the aircraft, such a pair of forces in the (x, y)-plane can be introduced
by discretizing the distributed aerodynamic forces and summing them up for wing,
fuselage and pods: Fy,w/p and for the tail : FY,tail, see Figure 7.2. Fy,w/p depends on the
state x ofthe aircraft and has a lever arm lw/p with respect to the center of gravity (CG).
The state x = [,8 rjT consists of the yaw rate r and the sideslip angle,8 . Within linear
aerodynamic theory, Fy,w/p(x) can be expressed by aerodynamic derivatives

Fy,w/p(X) = Yfj ,w/p,8 + (bjVTAS)Yr,w/p r, (7.1.1)

where b is a reference length like the half span length and V TAS is true airspeed [82] . The
force derivatives Yfj ,w/p , Y,.,w/p result from windtunnel tests , which are validated by flight
tests. They depend on varying parameters: velocity, altitude and wing configuration ,
i.e. flap-jslat-setting [140], [82). In that sense, Fy,w/p is an uncertain function . The
force FY,tail(X , bR) depends on both the state x and the input of the system , i.e. the
rudder deflection angle OR, see Figure 7.2. In analogy to (7.1.1), FY,tail is given by

(7.1.2)
7.1 Aircraft Load Alleviation by Robust Yaw-lateral Decoupling 291

Figure 7.2. Planar model for the compensation of a yaw disturbance torque MzD (engineout
torque/jl-split braking torque) by a rudder deflection angle OR / front tire angle 0

The lever arm of FY,tail is ltail' The derivatives YP,tail, Y,.,tail and Y 6n depend on th e
same varying parameters as Yp,w/p, Yr,w/p ' In that sense, FY ,tail is an uncertain func-
tion. Ideally, the desired balance of torques would be achieved by choosing the rudder
deflection angle JR such that

(7.1.3)

The forces and the disturbance torques are unknown, however, such that a robust
implementation of the control laws resulting from (7.1.3) is impossible .
An alternate approach was very successful in the automotive application and is
applied to the aircraft here. The idea is to split the vehicle dynamics robustly into
two subsystems. A first order subsystem with output lateral acceleration should not be
influenced by the yaw rate. A simple choice would be the decoupling of the lateral accel-
eration at the CG. Then, however, the yaw rate r as part of the state x enters into aY,CG
via both unknown forces Fy,w/p(x) and FY,tail (x ,JR ) , and again the ideal robustness
cannot be achieved. Therefore , we first define a decoupling point DP in a distance lDP
from the CG so that Fy,w/p(x) does not enter into the lateral acceleration aY,DP at the
292 7 Case Studies in Flight Control

decoupling point. The lat t er is

aY,DP = aY,CG - lDPT . (7.1.4)

Here, aY,CG and T are expressed in terms of the pair of forces:

Fy,w/p + FY,tail (7.1.5)

aY,CG
m
T
Fy,w/p lw/p - FY,tail ltail + M ZD (7.1.6)
Iz
where m is the mass and I z is the moment of inertia of the aircraft. This yields
Fy,w/p + FY,tail
aY,DP = m
Fy,w/p lw/p - FY,tail ltail + M ZD
- l DP
t,
= F
Y,w/p
(~ _ lDP lw/ p)
I (7.1.7)
m z

+
FY,tail ( -1 + -1--
lDP ltail)
- -I
lDP M
z D·
m z z

The location EDP of the decoupling point is calculated from the condition that th e factor
of Fy,w/p is zero, i.e.
Iz
lDP=-- (7.1.8)
mlw / p
and for this choice of lDP , we have

s )lw/p + ltail 1
aY,DP = FY,tail ( X ,UR l - - - l - Mz D . (7.1.9)
m w/p m wf p
Note that this derivation of the decoupling point is done in complete physical analogy to
the derivation of the decoupling point in the case of the car in (6.2.4). Equation (7.1.8)
corresponds to a representation of the mass and the moment of inertia by two masses
in the positions lw/p and lDP from the eG. The position lDP of the decoupling point is
close to the vertical tail and varies only slightly with the distribution of the mass, i.e,
with the payload and with the fuel consumption. For controller design, it is therefore
assumed to be constant. The induced error has to be analyzed in a detailed assessment
of the closed-loop system by a non-linear high precision simulation.
According to (7.1.9) and (7.1.2), the lateral acceleration aY,DP at th e decoupling
point depends on the yaw rate T only via th e uncertain function F Y ,tail (J3 , T, OR) =
Yp ,tail!3 + (b/ VT AS)Y,.,taiIT+ Y6 R OR . The crucial step for robustly eliminat ing t he influence
of T on aY,DP by a robust unilateral decoupling controller is to choose the rudder deflec-
tion angle OR such that it cancels T in the argument of the function F Y ,tai/(!3, T, OR) , see
Figure 7.3 and 7.4. More precisely, aY,DP shall be robustly decoupled from the yaw rate
by a controller, i.e. the yaw rate should be not observable from aY,DP for all variations
in mass and aerodynamics due to variations in velocity, altitude and flap-yslat-setting.
7.1 Aircraft Load Alleviation by Robust Yaw-lateral Decoupling 293

dynamicsof
ateral acceleratio
yaw atdecoupling
dynamics r point

Figure 7.3. Bidirectional coupling between the lateral acceleration aY,DP at the decoupling
point DP and the yaw rate r in the planar lateral aircraft motion in the case of an engine out
disturbance torque MzD

These variations are equivalent to variations in lw/p, ltail ' m, FY,tail , and Fy,w/p, [82]
and [140] . Physically speaking, the yaw rate r grows rapidly after the EO and induces
a large acceleration aY,DP and large shear forces at the vertical tail. It is this dominating
effect which the robust unilateral decoupling EO controller (in the following abbreviated
by r. u. d. EO controller) directly compensates. Detailed numerical simulations in the
next section reveal that the remaining immediate effect of the EO on aY,DP in terms
of (11m lw/p) N ZD has an acceptable level from the viewpoint of structural dynamics
and handling qualities.

dynamics of MaO
dynarne is of

yaw
latcn.lacceleration
al decoupliDg
ay,DP
L modified
lateral acceleration
at decoupling
f-~
. . .L _... poinl yaw dynamic point
dynamics c--

'-------' r \ This effect is cancelled

by the EO Controller
equivalem
blockdiagrams

Figure 7.4. Robust unilateral decoupling of the lateral acceleration aY,DP from the yaw rate r
by a dynamic feedback of the yaw rate to the rudder in case of an engine out. Thereby, the
shear force is reduced

So far, only the requirement of safe flight path stabilization is directly covered by
this physical controller design approach. Next , the requirement for low shear forces Qy,vt
at the vertical tail (load alleviation) has to be incorporated into the controller design
approach. Qy,vt is approximately given by Lomax [135) as

(7.1.10)

where mtail is the mass of the vertical tail. As the decoupling point is close to the vertical
tail for all operating conditions of the aircraft, the lateral acceleration at the decoupling
point aY,DP approximates the lateral acceleration aY,tail at the vertical tail. The yaw rate
294 7 Case Studies in Flight Control

is made robustly non-observable from FY,tail and by (7.1.10) also from Qy,vt . Hence, the
controller design approach covers the requirement to alleviate the loads at the vertical
tail.
Next, the r. u. d. EO controller is symbolically derived . The rudder angle is com-
posed of a reference part OR,pilot and a feedback part: OR,controi resulting from the
r. u. d. EO controller:

(7.1.11)

In the closed-loop system, the differential equation of FY,toil should not depend on the
yaw rate r, i.e. it should have the general form

(7.1.12)

Next, it is shown that already a first order controller structure is sufficient for fulfilling
the condition (7.1.12):

JR,control k",r + kr1' + Kpre/ilterOR,pilot , (7.1.13)

OR = 0R,control + 0R,pilot ,

with free controller parameters k", and kr • To show this, the differential equations for /3
and r are expressed in terms of Fy,w/p and FY,toil:

({Fy,w/p + FY,taid!mVT As ) - r ,
(7.1.14)
(lw/pFy,w/p -ltailFY,taii + NzD)!Iz'

Now, differentiating (7.1.2) and then substituting /3, l' and JR by the expressions in
Equations (7.1.11), (7.1.13) and (7.1.14) yields:

(7.1.15)

Then, requiring the explicit terms in r, i.e. {Y6 R k", - YiJ,toidr, to be cancelled, allows us
to derive the controller parameter k", symbolically:

(7.1.16)

However, r-dependent terms also implicitly enter in (7.1.15) via the terms in Fy,w/p'
By choosing

(7.1.17)
7.1 Aircraft Load Alleviation by Robust Yaw-lateral Decoupling 295

all terms in F y .w / p cancel. Herein, (7.1.8) has been used. With this choice of controller
parameters, the differential equation for Fy,tail is
.
F.y,tail =
(_b_y'
IT
.
T,tad
+ vLOR k)r -ltailFy,tail
I
+ M zD
I'TAS z
+ y; Fy,tail
(3,tail-1-rr:
ml'TAS

+ Y6R(Kpre/ilter6R,pilot + 8R,pilot ) . (7.1.18)

This equation shows that the yaw rate r of the controlled system has no influence
on FY,taii and by Equation (7.1.9) also no influence on aY,DP. Hence, the choice of
the controller structure is adequate to derive a robust gain-scheduling controller, which
makes FY,tail independent of the yaw rate for all variations of the mass and the aero-
dynamics . Since the variations of the aerodynamic derivatives of the tail Y{3,tail , Y,. ,tail
and Y6 R can be expressed in terms of the variations of the velocity, the altitude and the
flap-jslat-setting [82], [140], and as lDP is assumed to be constant, a gain-scheduling is
only required in the velocity, the altitude and the flap-jslat-setting, which can be easily
measured .
It is an essential robustness advantage of the controller that it does not depend on the
aerodynamic derivatives Y{3,w/p and Y,. ,w/p of Equation (7.1.1), because these derivatives
cannot be exactly expressed in terms of the velocity, the altitude and the flap-jslat-
setting, [82], [140]. Hence, these derivatives would have induced an uncertainty in the
gain-scheduling. As the rudder deflection angle 6R is the input, the controller consists
of a proportional and an integral feedback of the yaw rate. It has to be emphasized
that the controller is not a standard PI controller, which primarily ensures steady state
accuracy.
The controller , (7.1.13), (7.1.16), (7.1.17), essentially achieves the robust decoupling
of FY,tail from the yaw rate. By doing a state transformation in Equation (7.1.14)
from [,8 rjT to [aY,DP r]T and then computing the closed-loop system makes the
unilateral decoupling more evident. The yaw and the controller dynamics

r
[ 8R,control ]=[a
22 a23][
a32 a33
..
6R,control

+ [~:~] [ aY,DP ]
(7.1.19)
+ [~:~] [ 6 R,pilot ]

+ [::~] ( M zD ]

still depend on aY,DP , but the dynamics of the lateral accelerat ion at the DP
( aY,DP ] [ all ] [ ay,Dp ]

+ [bll b12 ] [ ~R,pilot ] (7.1.20)

6 R,pdot

+ [ell e12] [ Z:~ ]

296 7 Case Studies in Flight Control

are robustly decoupled from the yaw and the controller dynamics, Figure 7.4. The
decoupled first order dynamics have additional advantage that any maneuver that the
pilot performs in an EO will not excite any complex dynamics but only moderate first
order transients in aY,DP will occur.
To provide the coefficients of the robust unilateral decoupled state space model,
Equation (7.1.19), (7.1.20), it is useful to express the sum of the planar linear lateral
aerodynamic forces and corresponding moments by the standard aerodynamic force and
moment derivatives Yp, Yr , Y.s R, Np, N r, NOR [140], [82] :

Fy,w/p + FY.taii = ,(Ypw/p + YP,tail), /3 (7.1.21)

.
=:Y p
+ b/VTAS) ,(Y,.,w/p +
v
Y,.,tail))r + YoRO R
'

and, equivalently,

lw/pFy,w/p -ltaiIFY,tail = Np/3 + (b/VTAS)NrT

+ NoR' (7.1.22)

The differential equations of the planar lateral motion, (7.1.14), can then be rewritten
as

(7.1.23)

(7.1.24)

The coefficients of the state space model, (7.1.19), can be directly derived from (7.1.4),
(7.1.5), (7.1.21), (7.1.14), (7.1.23), and (7.1.13). They read :

K1 --
a22 V / (YpNr - Y,.Np) , where:
m TAS z
K1 VTAS/(Yp/m - NpID P / /z) ,
K1
a23 b21 = V / (YpNOR - Y.sRNp) ,
m TAS z
a21 KINp/VTAS/z,
e21 K1Yp/(m VTAS/z),
a32 k", + kra22 ,
a33 k ra23 ,
b31 Kpre/ilter + krb21
e31 k r e21'
7.1 Aircraft Load Alleviation by R obust Yaw-lat eral Decoupling 297

Th e coefficients of t he state space model, (7.1.20), can be directly derived from (7.1.9)
and (7.1.18). Th ey read:

all -K2ltail/lz + YC,tail/(mVTAs) , where:

K2 Y.sRkr + (b/VTAS)Y,.,tail,
bll YORKpre/ilter (lw/p + ltail)/(mlw/p) ,
b12 YOR(lw/p + ltait}/(mlw/p) ,
e ll K 2/(mlz ) + YC,tail/(m2VTAslw/p) ,
e 12 = -l/(mlw / p) .

Application to Aircraft Load Reduction

To analyze the application of th e robust unilateral decoupling EO controller (r. u. d. EO
controller) to an aircraft encount ering an EO , it has to be compared with a standard
lateral controller of an electronic flight control system (EFCS) in the following abbre-
viated by th e normal law. This normal law consists of a feedback of th e yaw rat e T ,
the sideslip angle (3, the roll rate p , and th e bank angle if! to th e rudd er and the aileron
deflection angles 6R and 6A . It has to ensure a sufficient stability margin and to fulfill
standar d handling quality and performance criteria [82], [140]. For an adequate com-
parison, th e r. u. d. EO controller has to be consist ently integrated into such a normal
law [127] .
The EO is simulated as a worst case, i.e. t he thrust is reduced in the form of a step
from maximum thrust to zero. This corr esponds to a bur st or th e initial dynamics in the
case of a flame out or a mechanical fault and yields a quicker loss in thrust t han in the
case of a very simple first order modelling of the engines [46]. Th e r. u. d. EO controller
is switched on after a detection t ime of 20 [ms] [60] . Detect ion t imes of 0.5 [s], resulting
from less sophisticat ed detection mechanisms, are considered later. For structural dy-
namic analysis , a second comparison is necessary, i.e. a corrective rudd er deflection of
th e pilot according to JAR 25.367 [114], [135] has to be sup erimp osed over the rudder
deflection from the normal law. This corrective pilot action was defined in the intr o-
duction and is illustrated by th e corresponding rudder deflection angle in Figure 7.5
and the yaw rate T in Figure 7.6. The large yaw rate r , which has been built up within
the first two seconds after th e EO , explains the large corrective rudder deflection of th e
pilot . This case is abbreviated by th e normal law + pilot. Simulations are performed
on a non-linear high-pre cision model with the following characteristics [82], [140] : i)
complete non-linear differential equation for the rigid body motion, ii) detailed non-
linear aerodynamic model, iii) st ruct ural flexibility, approximat ed by flexible factors
(residu al stiffness method [140]) , iv) non-linear actuator and electronic flight contro l
syst em (normal law) modelling.
Flight condit ions characterized by a low altitude and a low velocity are th e most
critical operat ing conditions in the case of an EO. These are take-off condit ions, with
maximum thrust and therefor e maximal disturbing yaw moment , requiring th e largest
rudder deflection angles for stabilization, which in t urn cause large loads at the ver-
tical t ail. To show robus tn ess, eleven critical operating condit ions (i.e. different mass
298 7 Case Studies in Flight Control

s r - - - - - - - - - , ;g 0 40
e......
.....
.s -20
;g 30
-0;
e......
2-40
.<: .!:
~-60
E

... .. ....
~ -80
_-~ _._.~._
°100
-1 ' - - - - - - - - - - ' :2"-
o 5 10 15 0 5 10 15 5 10 15
tin [5] tin [5] tin [5]

6 10
............
4
;g
e......
'§
!i-40 .
.JO
-- --
eo
c.oa:- eo -Go
-2

-80 L----"'----'_ _~_ _...J -4

o 5 10 15 0 5 10 15 5 10 15
tin [5] tin [5] tin [5]

Figure 7.5. Time histories of lateral load factor ny at ca, engine out moment MzD , sideslip
angle f3, rudder deflection 8R, bank angle <p and aileron deflection 8A for a critical flight
condition in the case of EO on left side (c-. normal law, *: normal law + pilot, 0 : r. u. d. EO
controller) . The yaw rate is given in Figure 7.6

normal law + pilot panic reaction

10 15
tln{s]

robust unilateral decoupling controller

0

5
~
.5
°l~
5

-1 0
10 15
tin [5]

Figure 7.6. Time histories of the yaw rate T for the comparison of the r. u. d. EO controller
switched on after a detection time of 20 [ms] and the normal law + pilot for eleven flight
conditions at M = 0.2 and 0.35 at "zero" altitude
7.1 Aircraft Load Alleviation by Robust Yaw-lateral Decoupling 299

and mass distribution) at M = 0.2 and 0.35 and zero altitude are considered. These
operating conditions yield the highest loads at the vertical tail in both cases with and
without the r. u. d. EO controller.

Structural Dynamic and Robustness Analysis

A first indicator for an adequate dynamic performance is the lateral acceleration n y
at the CG in units of the constant of gravity g, usually referred to as the lateral
load factor. The first transient peak is reduced to one third compared to the normal
law + pilot and to one half compared to the normal law, Figure 7.5, which shows that an
automatic control system can fully prevent critical scenarios within man's reaction time
(in Figure 7.5, 7.6 and 7.7 the reference or scaling values, ny,rel' etc. are maximal values
that are derived or roughly estimated from maneuvers of civil and military aircraft
covering the whole flight envelopes. As here only a maneuver of a civil aircraft is
considered, a ratio of 100 % does not occur).
In other words, what has been missed in the first two seconds, i.e. the two-third
extra amplitude in n y can be compensated for later on at only the cost of substantially
higher loads and unwanted coupling into the roll motion . In stationary values, there is
a reduction of more than one third (normal law) and of one half (normal law + pilot),
which again shows the advantages of an early drastic compensation.
In structural dynamics and in statics, the shear force at the vertical tail Qy,vt and
the corresponding bending Mx ,vt and torsion moments Mz ,vt at the vertical tail are
analyzed in phase diagrams, so-called phase loads diagrams, see Figure 7.7. Critical
operating conditions are represented by regions in phase loads diagrams. Here, a local
structural dynamic coordinate system has been chosen: the x- and z-axis are inverted
and rotated around the y-axis (by the swept-back angle of the vertical tail) so that
the z-axis is parallel to the elastic axis of the vertical tail. Figure 7.7 compares the
r. u. d. EO controller (center, right) with the normal law + pilot (left). Two detection
times for the EO are considered: a time of 20 [ms] (Figure 7.7, center) and a time
of 0.5 [s] (right). Figure 7.7 indicates that the regions of simultaneously large shear
force and torsion moment are avoided in a nearly ideal way in consequence of the
r. u. d. EO controller . The extreme values in shear force and torsion moment are nearly
reduced to one half by the r. u. d. EO controller . In more detail, the r. u. d. EO
controller reduces peaks in the transients of the shear force by more than one half
(normal law + pilot) , whereas the stationary values are reduced by 50 % compared to
the normal law + pilot. Concerning the torsion moment Mz ,vt , the r. u. d. EO controller
reduces the transient peak (normal law + pilot) by 55 % and brings about a much better
performance in transients than the normal law + pilot . This is a result of the achieved
reduction in sideslip angle, Figure 7.5, showing the interplay between flight mechanics
and structural dynamics. The achieved robustness is remarkable, i.e. the area covered
by all the various phase curves corresponding to different flight conditions is nearly
condensed to one curve. For static analysis, this means significantly reduced work in
assessment. Even more important is that in consequence of the achieved load reduction
for future large aircraft, the vertical tail can become lighter in combination with the
r. u. d. EO controller than for today's aircraft.
300 7 Case Studies in Flight Control

normal law+ pilot r. u. d. EOcontroller r. u. d. EOcontroller

z
detection time: 20ms detection time: 0.5s
~ ~
.£: 0 .£: 0 .£: 0
-x
-~20h~~~"; 20
E
~
'Lo ~
~40
';;;-60 critical'region' ';;;-60 critrcal region' .
o ~N -100 -50 o ~N -100 -50 o
Qy,v/IQy,vt,maxl in [%] Qy,v/IQy,vt,maxl in [%]

Figure 7.7. Phase loads diagram: shear force Qy,vt against torsion moment Mz,vt at the
vertical tail, comparison of the normal law + pilot and r. u. d. EO controller for eleven flight
conditions at M = 0.2 and 0.35 at "zero" altitude. In the case of the r. u. d. EO controller,
a detection time of 20 [ms] (center) and of 0.5 Is] (right) are considered

Flight Mechanical and Robustness Analysis

In contrast to the normal law and the normal law + pilot, the r. u. d. EO controller
avoids large magnitudes in transient and final states, in particular in the yaw rate r,
see Figure 7.5 and 7.6. Thereby, large aircraft motions in consequence of the EO are
avoided so that any pilot interference becomes unnecessary. The reference or scaling
values ny,ref' f3re/l 8R,re/l 8A ,re/l ¢>re/l and Tre] are maximal values that are derived
from maneuvers of civil and military aircraft covering the whole flight envelopes. The
r. u. d. EO controller achieves such a performance by an immediate and sufficiently
strong rudder command, nearly reaching the final rudder deflection angle for yaw mo-
ment compensation already after one second. Therefore, the EO is fully compensated,
i.e. the aircraft is in a nearly steady state after a second, which is half the pilot's re-
action time according to JAR 25.367, see in particular the yaw rate r in Figure 7.6.
Additionally, the commanded rudder rate is lower than that in the case of the corrective
pilot action . The later the yaw moment compensation is achieved, the more energy is
transferred into the roll motion and the more aileron activity is necessary to compensate
and to reduce this effect. Therefore, the reduced rudder deflection angle in the case of
the normal law has to be compensated by large aileron and spoiler angles, Figure 7.5.
Moreover, this means being far away from an optimal flight condition in terms of an
optimal lift, flight safety, passenger comfort, heading, and drag . Concerning the flight
mechanical quantities in more detail, the following improvements can be observed:
i. yaw rate and bank angle
The yaw rate r, which directly indicates the achieved yaw moment compensation,
is reduced to less than one third in transient peaking compared to the normal law
and the normal law + pilot, Figure 7.6. After one second, the yaw acceleration
already changes its sign, so that there will be no reason for any overreaction. The
yaw rate goes to zero sufficiently fast. Therefore, only little deviations will occur
in heading in contrast to the normal law and the normal law + pilot . The transient
peak in bank angle ¢>, Figure 7.5, is reduced to one third (normal law + pilot)
and to one eighth (normal law), which means significantly improved passenger
comfort.
7.1 Aircraft Load Alleviation by Robust Yaw-lateral Decoupling 301

ii. aileron angle

The peak in aileron angle OA is reduced by 60 % compared to the normal law +
pilot and by 80 % compared to the normal law, Figure 7.5.
iii. sideslip angle
The sideslip angle {3, see Figure 7.5, is reduced in transient peaking by more than
80 % and in its final value by 40 % (normal law + pilot) and 80 % (normal law),
which guarantees, combined with the achieved yawing performance, a significantly
lower lateral deviation from the flight path. Taking low roll controls into account,
the drag will be significantly reduced.

As two inputs (rudder and aileron deflection angle) are available to stabilize the aircraft
with zero roll and yaw rate , there are arbitrarily many combinations of steady states
and stationary inputs to achieve the stabilization. However, a sufficiently large rudder
deflection angle OR and a finite sideslip angle {3 ensure that the magnitudes of all states
and inputs remain in acceptable boundaries for the stabilization. Due to the different
stabilization strategies in the case of the r. u. d. EO controller, in the case of the normal
law and in the case of the normal law + pilot, different stationary values of the states
occur. The moderate use of the rudder in the case of the normal law requires large
stationary sideslip and aileron deflection angles. To avoid this , the pilot has to bring
down the sideslip by a large rudder deflection angle. The r. u. d. EO controller avoids
this large rudder deflection angle, the large sideslip and the aileron deflection angle by
an early rudder deflection.
These benefits of the r. u. d. EO controller and the resulting reduction of the loads
at the vertical tail become apparent in Figure 7.8. For an outer left EO at t = 0 [sJ, Fig-
ure 7.8 compares the r. u. d. EO controller (upper aircraft with the light grey vertical
tail, the corresponding shear force QlI ,vt is illustrated by the right, light grey column at
the right bottom of each subplot in the six subplots) with the normal law (lower aircraft
and left column in each subplot). Six time steps are considered. At 0.5 [s], the early
rudder deflection angle of the r. u. d. EO controller is illustrated and the following sub-
plots show the resulting early stabilization of the flight path and the loads in contrast
to the conventional aircraft. At 2.5 [sJ the corrective pilot action (JAR 25.367) becomes
evident. It is clearly justified by the yaw rate and the sideslip angle, and the resulting
heading angle and the flight path deviation that have been built up. The benefits of the
r. u. d. EO controller with regard to the shear force, the energy transfer into the roll
motion, the bank angle and the heading angle become obvious at 2.5 [s], 3 [sJ and 3.5 [s] .
Figure 7.6 and 7.7 show that the r. u. d. EO controller has better robustness prop-
erties than the normal law. The robustness of the normal law has been achieved by
applying an iterative optimization over the whole flight envelope [46] . It therefore serves
as a reference to evaluate the robustness of the r. u. d. EO controller. The amplitude
variations of the yaw rate time histories in the first peak, which are caused by parameter
variations , are reduced by more than one half by the r. u. d. EO controller compared
to the normal law. It should be emphasized that this performance was achieved with a
very simple controller structure.
302 7 Case Studies in Fligbt Control

3.5

-.:::-""""- '.
--
~ ....-

Figure 7.8. Aircraft and shear force dynamics at the vertical tail in an outer left EO for six
time steps: r. u. d. EO controller (upper aircraft with the light grey vertical tail, correspond-
ing shear force: right, light grey column at the right bottom of each subplot) compared to the
normal law + pilot (lower aircraft and left column)
7.2 Robust and Fault-tolerant Gamma-stabilization of an F4-E 303

The assumption of a linear planar model for controller design is justified by the
small amplitudes of all relevant lateral quantities (Figure 7.5 and 7.6), which occur in
the EO simulation based on the high-precision closed-loop system.

Conclusion
A new type of pilot assistance control system for the yaw disturbance moment compen-
sation has been transferred from cars (jL-split braking) to aircraft (engine out situation) .
This transfer includes the analytical controller derivation. The new pilot assistance sys-
tem is characterized by a very early interference within man's reaction time: obviously,
an automatic system can master critical situations in a much better way by avoiding
any time wasting in consequence of the pilot reaction time and by avoiding any overre-
action . Instead of first letting the aircraft drift into an extreme flight situation, which
then is hard to master, the early yaw moment compensation keeps the aircraft in a
moderate flight condition. Consequently, the pilot has time for other safety tasks like
the engine management (e.g. in case of a flame out) and is not subject to a panic or
an overreaction. Thereby, a task separation is achieved, which enables the pilot to
concentrate on planning and performing long-term tasks. Here, man is superior to any
automatic system. The short-term task, in particular within the reaction time where
an automatic system is superior to man, is performed by the automatic system. In sim-
ulations on a non-linear high-precision model, the control system has turned out to be
very robust to any parameter variations (mass, mass distribution, aerodynamics) and
to non-linearities. The above features have been achieved with a simple controller of
dynamic order 1 for all flight conditions. The design procedure is simple and physically
transparent and yields symbolic expressions for the controller parameters.

7.2 Robust and Fault-tolerant

Gamma-stabilization of an F4- E

Consider the short period longitudinal mode of the fighter aircraft F4-E in Figure 7.9.
The F4-E is a modified military aircraft. In particular, the maneuverability was in-
creased by additional horizontal canards . This results in a loss of longitudinal aerody-
namic stability, however. The short period mode is unstable in subsonic flight and only
weakly damped in supersonic flight. The equations of motion are linearized for small
deviations from a stationary controlled flight (i.e. constant altitude and velocity, small
angle of attack a). In flight mechanics, it is usual to take a and the pitch. rate q as
state variables. Here, we transform the equations to sensor coordinates, i.e. the normal
acceleration Xl = n z and the pitch rate X2 = q are introduced as states. This simplifies
the design for robustness with respect to accelerometer fault [3], [88]. The actuator for
the elevator is modelled as a low-pass filter with the transfer function 14/(8 + 14). Its
state variable is 5e • For the state vector, we take
304 7 Case Studies in Flight Control

Canard '"

Figure 7.9. Fighter aircraft F4-E

(7.2.1)
and the linearized state equations

X Ax+bu

have the following structure:

A ~ [~:: ~:: ~~4] ' ~ ~ b [ ] .

(7.2.2)

Remark 7.1

The model (7.2.2) is based on some simplifying assumptions:

a) Elevator (8.) and the canard rudder (8c ) are not used independently of each other
for small deviations from stationary flight. The two commanded input variables
are coupled by
8.com = u,
(7.2.3)
8ccom = -0.7u.
The factor -0.7 was chosen for minimum drag. Therefore, the system (7.2.2) has
only one input u.
7.2 Robust and Fault-tolerant Gamma-stabilization of an F4-E 305

b) Structural vibrations are not modelled. The bandwidth for the rigid body control
is limited to be below th e first structural mode frequency of 85 [rad / s] to avoid
excitation of structural modes.
o

The possible flight conditions of this aircraft are represented in the Mach-number
alt itude diagram, Figure 7.10.
80

60
.> -
'\
;:: f-"
o
o
....
o
40
-:
/
QJ

~
-0
0 0

-:
.....;:::l
3 4
~ /

I
20

o
o 0.4
01
0.8
02 ~
1.2 1.6 2.0 2.4
Mach number

Figure 7.10. Envelope of the possible flight conditions and four representative cases

Numerical values for four representative flight conditions (Fe) as indicat ed in Fig-
ure 7.10 have been taken from [47] and were transformed to the state equation (7.2.2) .
Th ey are listed together with th e respective eigenvalues 8 1 , 8 2 in Tabl e 7.1. The third
eigenvalue of (7.2.2) is fixed at 83 = -14.

Robustness with Respect to Flight Conditions

Based on pilot ratings, admissible intervals for damping and frequency of the short
period mode have been determined and must be verified in the qualification tests for
the certification of a new aircraft [146]. For the characteristic polynomial p(s) = ao +
alS + S2 = w~ + 2Dwos + S2 , the following bounds are prescribed for damping D and
natural frequency Wo :
(7.2.4)
The bounds W a and Wb depend on th e flight condition and are given in Table 7.2.
Comparing the eigenvalue distances from the origin for th e open-loop poles in Ta-
ble 7.1 and the closed-loop pole specifications in Table 7.2, it is obvious that a basic
rule of robust control is observed: keep a fast syst em fast and keep a slow system slow.
Th e admissible region is defined in the plane of coefficients ao, a1 by (7.2.4). It is
bounded by the lines ao = w~ and ao = w~ as well as the two parabolas ao = w~ ,
306 7 Case Studies in Flight Control

Table 7.1. Model data for an F4-E aircraft for four typical flight conditions. The eigenvalues
81and 82 result from (8 - aU)(8 - a22) - a12a21 = a

FC 1 FC 2 FC 3 FC4
Mach 0.5 0.85 0.9 1.5
Altitude [ft] 5000 5000 35000 35000
au -0.9896 -1.702 -0.667 -0.5162
a12 17.41 50.72 18.11 26.96
a13 96.15 263.5 84.34 178.9
a21 0.2648 0.2201 0.08201 -0.6896
a22 -0.8512 -1.418 -0.6587 -1.225
a23 -11.39 -31.99 -10.81 -30.38
bl -97.78 -272.2 -85.09 -175.6
81 -3.07 -4.90 -1.87
-0.87 ±j4.3
82 1.23 1.78 0.56

Table 7.2. Military specifications for the admissible natural frequency Wo of the short period
mode of an aircraft

Flight Velocity Altitude Wa Wb

Condition [Mach] [Feet] [Radian / Second]
1 0.50 5000 2.02 7.23
2 0.85 5000 3.50 12.60
3 0.90 35 000 2.19 7.86
4 1.50 35000 3.29 11.80

al = 2DminWo = 0.7wo and ao = w5, al = 2Dmaxwo = 2.6wo. The region for flight
condition 2 is represented in Figure 7.11. If D is increased beyond 1, then the complex
pole pair in the s-plane unites at a branching point for D = 1 and then separates into
two real poles of which one moves to the left and the other to the right as D increases.
The pole moving to the right often leads to an undesired decrease in the bandwidth
of the closed loop; this corresponds in the time-domain to a sluggish response. These
less desirable domains lie above the parabola for D = 1 in Figure 7.11. Therefore, the
admissible region in coefficient space will be contracted by replacing the upper bound
for al by the straight lines a and b (the tangents of the curve D = 1 at w~ = 3.52
and w~ = 12.62 ) . The coefficient region between a, b and the curve D = 1 describes
exactly the polynomials with a pair of real roots in the interval 8 E [-W a ; -Wb] . The
reduced coefficient region corresponds to the "pineapple segment", which is displayed
in Figure 7.12 with solid lines. The expanded region shown in dashed lines is chosen
for further eigenvalues that do not belong to the short period mode. The values W a ,
Wb, and the open-loop eigenvalues 1, 2, and 3 are displayed in Figure 7.12 for flight
condition 2.
7.2 Robust and Fault-tolerant Gamma-stabilization of an F4-E 307

Figure 7.11. Requiredstability regionfor flight condition2 in coefficient plane and replacement
of the boundary D = 1.3 by the two boundaries a and b

jCJ

.... ""
.:
;' \
\
I D= 0.35--\
/ 10
I
: 3
I 5
\
\
\
\ I
"
CJd' /
/
'-..1

Figure 7.12. r-stability region in the s-plane

The controller structure that is easiest to implement is

(7.2.5)

where Xl is measured by an accelerometer and X2 with a gyro. The third state ae is the
deviation of the elevator from its trim position. The latter is not known with sufficient
a
accuracy, therefore e is not used for feedback. The viewpoint of each of the two sensors
would suffice. But it is not clear at this point which sensor is better suited for robust
control. During the design, both alternatives will be considered and it will be shown
why the gyro is the better choice. This example illustrates that the controller structure
may not always be fixed before the controller design, it may also be a result of the
design process.
The resulting two-dimensional cutting plane through the three-dimensional region
of r-stability is shown in Figure 7.13 for Wd = 70 [rad/s] and flight condition 2.
For k = 0, the eigenvalues are those of the open-loop, see 1, 2, and 3, in the s-plane
figure. If the boundary (1a is crossed by a variation of k starting with k = 0, then
eigenvalue 2 moves across the point (1a to the right.
308 7 Case Studies in Flight Control

1
I
I
I
I
o
D

-1

-2

-3

-0.3 -0.2 -0.1 0.0

Figure 7.13. Two-dimensional cross-section of the r-stability region in k-space

The eigenvalue pair (1, 2), which describes the short period longitudinal mode, now
moves as a complex conjugate pole pair into the desired region if the boundaries Wa or
D are crossed. The desired region is thus ABCDEA. On AB, the eigenvalues 1 and 2
lie on the circle W a , on BC they lie on D, and on CD eigenvalue 3 lies at (1d. On DE,
one of the eigenvalues (1, 2) lies at (1b, on EA eigenvalue 3 lies at (1b.

Remark 7.2

A one-to-one correspondence of the eigenvalues of the closed-loop to those of the

open-loop is preserved on the described path in the (kq,kn.)-plane. Such a relation can
be lost, for example, if eigenvalues 2 and 3 combine to form a complex pair , then a
branching point is crossed. This would be the case, for example, if after leaving k = 0,
first (1b and subsequently Wb were crossed. Pole 3 then moves to the right first and the
pair (2, 3) moves over the boundary Wb in the s-plane, 0

Correspondingly, the r -stability regions are determined for the other three flight
conditions. All four regions intersect in the region represented in Figure 7.14. Thus,
the assumed controller structure leads to an admissible solution set .
Figure 7.14 provides the design engineer with essential information as to which de-
mands are critical in which flight cases. The flight condition numbers are used as indices
at the boundary line names. The stim ulation of structural oscillations is especially crit-
ical near the boundary (12 = -70, i.e, for flight condition 2 (high velocity, low-level).
Insufficient damping is critical near the boundary D 4 for flight condition 4 (fast flight
at high altitude) . The other two boundaries originate from flight condition 1 (landing
approach). Near the boundary (11 = -2.02, a real pole moves toward the origin. A real
pole would cross over to the left at the boundary (11 = -7.23. Flight condition 3 is not
critical in this controller structure.
7.2 Robust and Fault-tolerant Gamma-stabilization of an F4-E 309

k q 0.00

-0.25

-0.50

-0.75

-1.00

-1.25

-1.50

Figure 7.14. Intersection of the regions of r-stability for all four flight conditions

Information on the possible demands that can be met by a control system and the
possibilities for compromise is usually more helpful for the control engineer than an
optimization process, which requires that the trade-offs for all conceivable conflicts are
decided beforehand by the choice of weightings in a performance criterion.
Figure 7.14 suggests different possibilities for coming closer to the selection of a
point from the solution set. One possibility is to reduce the eigenvalue region r of
Figure 7.12. Figure 7.15 shows the tightened region obtained by a reduction of Wd
from 70 to 50 [rad/s], an increase of the minimum damping from 0.35 to 0.5 and an
increase of all minimum natural frequencies W a by 50 %. The boundary <71 = -7.23
serves only to keep the eigenvalues in separate regions; it remains unchanged. There
still exists an intersection for the four flight conditions and the designer may decide
which of these specifications they want to tighten even more and thereby narrow the
admissible region further.
However, other requirements can also be incorporated into the solution choice which
have not been taken into consideration so far. As an example, suppose the limitation
of the elevator deflection X3 and its derivative 3;3 = -14x3 + 14k T X is essential. Then,
a solution with a smaller loop-gain is preferred. In order to illustrate this effect, the c*-
response to a unit step input given by the pilot has been calculated, as well as the
required elevator deflection X3 for the points 91' 92' and 93 in Figure 7.15, see Fig-
ure 7.17. The quantity
c" = (n z + 12.43q)/coo
is a common variable in flight mechanics for the evaluation of the step response . The
stationary value Coo is used for normalization. The c*-step response should lie in the
envelope shown in Figure 7.16. If the eigenvalues are suitably located in the eigenvalue
region I', then the step responses are slightly faster than required. By a small delay in
a prefilter they can be modified such that they fit into the envelope in Figure 7.16. For
this design study, the prefilter 6/(8 + 6) was given.
The elevator deflection is decreased significantly for a small loop-gain 91 and the
310 7 Case Studies in Flight Control

k q 0.00

-0.25 gl "of
,,'1( I

-0.50 D 4 = 0.3Ji,,"""""" ,/\

.:>" .,

"",,"1)4 = O. ,. (Tl = -2.0

-0.75 K = -3.03
,,
r »>
r
(Tl
kO
-1.00 (Tl = -7.23
" (T2 = -50
-1.25
(T2
,
= -70 , g2

-1.50
,,
, »"
,,"
,," "

g3~""

-0.25 -0.20 -0.15 -0.10 -0.05

Figure 7.15. Tightenedspecifications reduce the admissible solution set

step response is correspondingly slow. A high loop-gain g3 is unsatisfactory due to too

much overshoot of the step response, greater amplitude, and extension of the bandwidth
to the vicinity of the structural-oscillation frequency. The medium loop-gain g2 is the
best in this comparison, it produces a solution that is closer to g3 than to gl and a
point kO slightly closer to gl is finally selected. The chosen controller is

u = _koT X = [-0.115 -0.8 0] x . (7.2.6)

The point kO is indicated in Figure 7.15. Calculating and testing the eigenvalues shows
that they lie in their respective prescribed regions for all flight conditions.

Fault Tolerance
Usually, control systems are designed under the assumption that sensors do not fail.
Redundancy management has then to provide the required measurements with only
very short interruptions due to faults of individual sensors. If the plant is an unstable
aircraft, this means that fault detection is vital for stabilization. Detection must operate
quickly and this is in conflict with the requirement of low probability of false alarms .
An alternative is the use of a hierarchical concept. Its basic level is a fixed-gain
control system, which is designed such that pole region requirements are robust with
respect to component faults and uncertain parameters [3]. All the more sophisticated
tasks, like fault detection and redundancy management , plant parameter identification
and controller parameter adaptation, or gain-scheduling, are assigned to higher levels
if they are required for best performance. The higher levels process more information
and are operating in a slower timescale than the basic level. Since the higher levels are
not vital for stabilization, they can make their decisions without panic.
7.2 Robust and Fault-tolerant Gamma-stabilization of an F4-E 311

c* 2.0
... -,

1.5 I
"" ""
""
I "
I "
I ' "
I '-,_
I --- _
I ----
1.0 I
I
-------
-- --- ---- -----
I
, I
... ...
I
0.5 "
I"
I
I

""
I

... " "

0.5 1.0 1.5 2.0
t [s]
Figure 7.16. Admissible envelope for the c*-response
c* 1.25 r----.--~--~-...., oe [rad] 0.15

1.00 0.10

0.05
0.75
0.00
0.50
-0.05
0.25 -0.10

0.5 1.0
2.0 0.5 1.51.0 1.5 2.0
[s] t t [s]
Figure 7.17. c* response and elevator deflection oe for low (91)' uedium (92) and high (93)
loop-gain. The elevator deflection is normalized to the stationary value of -oe(t, 91)

Assume that a sensor fault has two effects as illustrated by Figure 7.18.

i. The multiplicative effect reduces the gain V from its nominal value 1 to zero or
some value in between.
ii. The additive effect introduces a bias or noise d at the output.

As far as the eigenvalue location is concerned, only the multiplicative effect is important.
The additive effect may require that the failed sensor signal is switched off. This decision
of a fault detecting system may be slow, e.g. if the plant operates in a steady state and
one of the sensors sticks at one value.
If the measured variables are used as state variables ("sensor coordinates") , then
the sensor fault is equivalent to reducing the corresponding feedback gain. For V =
0, a solution for which r-stability is robust against fault of the sensor for the state
variable Xi is characterized in gain space by the fact that the projection of k on the
subspace k, = 0 is contained in K r . A two-dimensional cross-section through gain space
312 7 Case Studies in Flight Control

------t{2]t-----
a) Nominal sensor b) Failed sensor, 0 s V ~ 1

Figure 7.18. Model of sensor faults

is shown in Figure 7.19. Assume that the admissible region K r is the triangle ABC.
Robustness against fault of sensor i is achieved if the projection of the appropriate gain
is on GE. This property holds for DEFG . Similarly, HCJKL is robust against fault of
sensor j and KLMN is robust against fault of either i or j .

C
Figure 7.19. In KLMN, I'-stability is robust against faults of the type ki = 0 or k j = 0

If no such intersection of the r -stability region with the axes exists, then the designer
may use redundant sensors in parallel. The simplest possibility is to use two parallel
sensors. Their outputs are multiplied by a factor 1/2 and added to produce X i in the
unfailed case. If one of the two sensors fails, then k i is reduced by 50 %. In order to
achieve robustness, ki has to be selected such that it has a 50 % gain reduction margin
in k; for r-stability. If K r is the triangle ABC in Figure 7.20, then DEF is the region
for which r-stability is achieved after a 50 % reduction of k i . Thus , the triangle EGH
contains the admissible points which are f-stable and also rema in I'-stable after 50 %
gain reduction.
If an inters ection still does not exist for two parallel sensors, then three sensors in
parallel may be considered. An advantage of this choice is th at it can be combined
with a fault detection procedure on the next higher hierar chical level. Its structure is
7.2 Robust and Fault-tolerant Gamma-stabilization of an F4-E 313

c F
Figure 7.20. In EGH, r-stability is preserved under 50% gain reduction in ki

shown in Figure 7.21.

,--- Fault
.--- Detection

XiI
1
o} ,---
u Xi2 Xi
Plant 02
Xi3 03 -

Figure 7.21. A triplex system

The fault detection system forms the three decision functions :

d}(t) = [Xil(t) - Xi2(t)][Xil(t) - Xi3(t)],

d2(t) = [Xi2(t) - Xi3 (t)][Xi2 (t) - Xil(t)], (7.2.7)
d3(t) = [Xi3(t) - Xil(t)][Xi3(t) - Xi2(t)] .

The dk's are nominally zero; Idkl ~ e indicates a fault of sensor k , In order to avoid
false alarms from short impulses, dk(t) is low-pass filtered first and then compared to a
threshold value .
A(t) = a!r(t) + d}(t),
i2(t) = ah(t) + d2(t), (7.2.8)
i3(t) = a!J(t) + d3(t).
The decision logic is then

Nominal state:
/!r(t)1 < e, Ih(t)/ < e, 1!J(t)1 < e, (7.2.9)
o} = 02 = 03 = 1/3.
314 7 Case Studies in Flight Control

Fault of sensor k :
Ifk{t)! > e, Ih{t)1 < c for j =I- k,
(7.2.1O)
ak = 0, ai = 1/2 for j =I- k.
No changes are made in the a's after a second fault . The parameters a and c are chosen
in view of safety against false alarms, i.e. both are not too small. This means that the
decision may take some time . Between fault and decision times, the gain is reduced
to 2/3 . If a second fault occurs after the first decision, then the gain is reduced to 1/2 .
In the unlikely case that a second fault occurs before the first one is detected, the gain
is only 1/3. Thus, the basic robust control system should be designed for this gain
reduction margin of 50 % or 67 % for r-stability.

XiI
1/3 -
u Xi2 Xi
Plant 1/3
Xi3
1/3 -

ki
X;
ki : Filter I

Figure 7.22. Substituting the measurement of Xi by a filter producing Xi

Here, a filter produces a substitute feedback variable Xi for the true state-variable Xi '
The filter transfer function is chosen such that the transfer function from u to Xi at least
crudely approximates the transfer function from u to Xi within the desired closed-loop
bandwidth and over the range of plant parameter variations. The controller structure of
Figure 7.22 is especially useful if the transfer function from u to Xi is minimum-phase,
because then cancellations or near-cancellations by filter poles can be made. Instability
of the transfer function is no disadvantage if the same instability also occurs, as is usual,
in the transfer function from u to Xi '
In the filter structure, a fault effects both feedback channels simultaneously, i.e.
both gains k i and ki are reduced. For example, let a fault of one sensor occur such
that the sensor and feedback gain is reduced by a factor 2/3 . If ABC in Figure 7.23 is
the r-stability region, then DEF is the region for which r-stability is maintained after
one-third gain reduction in ki and ki' Thus , AGH is the region in which r-stability is
achieved both nominally and after the sensor fault .
The configuration of Figure 7.22 with one measured state variable immediately gives
the same gain reduction margin in both channels in the case of fault of one of several
parallel sensors.
The above ideas on fault tolerance are now applied to the F4-E example and robust-
ness with respect to gyro or accelerometer faults. Figure 7.14 shows that the r-stability
region does not intersect the axes kn • and kq • Therefore, one of the two feedbacks alone
is not sufficient.
A first attempt would be to use two parallel gyros and two parallel accelerometers.
From Figure 7.14, it is seen that there are no points in the r-stability region that
7.2 Robust and Fault-tolerant Gamma-stabilization of an F4-E 315

F
Figure 7.23. In AGH, r-stability is preserved under 1/3 gain reduction in both channels

admit 50 % gain reduction of both kn • or kq • Interestingly, the shape of K r in Figure 7.14

with its lengthy extension away from the origin is such that the flying qualities are
improved if a second fault in the other channel occurs, i.e. for a simultaneous 50 %
reduction of gyro and accelerometer gain. This indicates that a controller structure as
in Figure 7.22 with only gyros or with only accelerometers is advantageous . Here, any
gain reduction occurs in both feedback gains simultaneously.
Of course, the unmeasured variable cannot be really reconstructed for all four flight
conditions simultaneously by a constant filter. Also, the system order is increased by
the filter such that additional eigenvalues occur. Therefore, the new stability region
in the plane of the two feedback gains must be established. A similar form of the
stability region can be excepted, at least if the filter is chosen such that it yields an
approximation for the unmeasured variable.
In deciding whether to use only gyros or only accelerometers, we consider the zeros
of the two transfer functions . In the four flight conditions, the zeros in the accelerometer
transfer function vary between -0.4 ±j5.7 and -0.9 ±j9.1. They are minimum-phase
but outside the f -stability region. Also, they may vary widely such that an approximate
cancellation by the polynomial of averaged zeros 82 + 1.1728 + 49.9 is not advisable.
This would generate closed-loop poles that are not f -stable , possibly not even stable .
The gyro transfer function has one real zero varying between -0.64 and -1.57. Here,
an approximate cancellation by the polynomial of averaged zeros of the four operating
conditions (8 + 0.98) is no problem. By this near-cancellation, this filter pole will be
only weakly controllable from u and will not be shifted much by closing the loop. Also,
it is weakly observable from the pitch rate and can be exempted from the pole region
requirement .
Both transfer functions from u to n z = Xl and q = X2 have the same poles; one pole
is unstable in subsonic flight. The filter must replace the numerator of the gyro transfer
function by the numerator of the accelerometer transfer function . Taking the averaged
316 7 Case Studies in Flight Control

zeros of the two transfer functions, the filter is

2
f(8) = a.8 + 1.1728 + 49.9 . ~ . ( )
7.2.11
8 + 0.98 8 + 10
The term 10/(8 + 10) is introduced to make the filter realizable. The gain ratio a
varies between 0.527 and 0.577 in the four flight conditions and is chosen at the average
value a = 0.543.
By introducing the filter , the system order is raised to n = 5. The chosen controller
structure of Figure 7.22 defines a plane of the free controller parameters kn • and kq •
In this plane, the intersection of r -stability regions for the four flight conditions must
be analyzed. It is represented in Figure 7.24. The allowable region is clearly different
kq 0.00

-0.25

-0.50

-0:75

-1.00

-0.1

Figure 7.24. Intersection of the f-stability regions for the four flight conditions for the con-
troller structure using only gyros and a feedback filter

from the region in Figure 7.14. Also, flight condition 3 instead of flight condition 1
now contributes to the boundary. Nevertheless, it confirms our expectation that a large
simultaneous gain-reduction margin is now achievable , in the extreme case 80 %. If one
chooses the indicated point k = [-0.09 - 0.8 0lT, then the points 2k/3, k/2 , and
k/3 lie well in the admissible region. Therefore, the system can handle the fault of
two of the three gyros, even if the second fault occurs before the first one is detected.
For the c*-step responses in Figure 7.25, it is also worth remarking how little the step
response is changed in all four flight conditions if k (steeply climbing curve) is reduced
to k/2 . If the two feedback paths via kq and kn • are combined , then the controller
2
-us(8) = -0.8 _ 0.09 . 0.543 . 8 + 1.1728 + 49.9 . ~
X2S(8) 8 + 0.98 8 + 10
1.2982 + 9.368 + 32.23
= (7.2.12)
82 + 10.988 + 9.8
is obtained. The minus sign is explained by the common flight mechanics definitions of
the sign of the elevator deflection oe and the pitch angle q, which causes a minus sign
of the plant transfer function .
7.2 Robust and Fault-tolerant Gamma-stabilization of an F4-E 317

c* 1.5 c* 1.5

1.0 1.0

0.5 0.5
FC 1

0.5 1.0 1.5 2.0 0.5 1.0 1.5 2.0

t [s] t [s]
c* 1.5 c* 1.5

1.0 1.0

0.5 0.5
FC 3 FC 4

0.5 1.0 1.5 2.0 0.5 1.0 1.5 2.0

t~ t~
Figure 7.25. c*-step response in four flight conditions. Each case shows nominal and 50 %
reduced loop-gains. The nominal gain results in the steeper initial rise

The procedure for the design of robust stabilization for the F4-E is certainly not
a general design recipe, but similar results can be expected for other aircrafts. In the
robust design of a back-up controller for the Swedish fighter JAS 39, it was assumed
from the beginning that only the gyro is used for feedback. The controller in this case
was assumed as a proportional channel plus two channels with first order delay. A
robust level 1 controller for 10 flight conditions was designed using both the parameter
space technique and a vector optimization method [98].

This extensive F4-E example shows how the design tool of two-dimensional cross-
sections through regions of r -stability is applied and combined with other design con-
siderations. The controller structure was developed in steps . The design would be less
transparent if the second order controller of (7.2.12) with five free controller parameters
had been assumed from the beginning.
318 7 Case Studies in Flight Control

7.3 Large Envelope Flight Control of a High Per-

formance Aircraft

The design of flight control systems for fighter aircraft is an extremely demanding task .
These aircraft must meet stringent performance specifications [1471 throughout a large
envelope of operating conditions despite large variations in system dynamics (including
variations in open-loop stability), which result from varying operating conditions and
the presence of significant nonlinearities. These large variations in the dynamics of the
aircraft often make it impossible to achieve the required level of performance throughout
the flight envelope with a fixed controller. Practically, this results in the design of several
controllers at trim points throughout the flight envelope, which are then scheduled with
the operating condition . While this gain-scheduling procedure has been successfully
implemented on a number of fighter aircraft , it is extremely tedious and time consuming
to design. The complex task of traditional gain-scheduling has resulted in a significant
amount of research in recent years to develop methods of designing controllers that
are "automatically" gain-scheduled. While significant progress has been made in this
effort, e.g. [521, these methods are still quite involved and result in complex controllers
that are often very conservative.
In this section, the parameter space methods presented in the previous chapters are
used to provide a much simpler approach for designing large envelope flight controllers
that provide the desired performance throughout the design envelope without gain-
scheduling. The design procedure presented here [53] uses a variation of the control
architecture shown in Figure 1.12 along with r- and B-stability specifications to form
a unified approach for designing robust flight controllers for high performance aircraft .
The procedure enables the designer to explicitly define the desired closed-loop dynamics
and ensures that the closed-loop satisfies r- and B-stability (i.e. H oo ) specifications. The
result is a straightforward procedure that allows the design of a robust flight controller
that "forces" the closed-loop dynamics to behave like the specified "desired dynamics"
despite disturbances, modeling uncertainties, and variations in aircraft dynamics due
to changing flight conditions.
The procedure is presented here by designing a pitch-rate controller for the F-16
Variable Stability In-Flight Simulator Test Aircraft (VISTA), and clearly demonstrates
the "simplistic" power of the parameter space design method . The resulting controller
provides the desired performance throughout the large design flight envelope, which is
demonstrated using both linear simulations and a high fidelity, non-linear simulation.

Pitch-Rate Control of the F-16 VISTA

VISTA is a modified F-16 with the capability of simulating advanced aircraft configura-
tions and testing advanced flight control concepts. A detailed description of the VISTA
aircraft is given in [148]. In this section, a linear parameter-varying (LPV) model of the
F-16 VISTA's short-period dynamics and design specifications for an F-16 pitch-rate
controller are reviewed.
7.3 Large Envelope Flight Control Satisfying H-Infinity Specifications 319

Figure 7.26. F-16 Variable Stability In-Flight Simulator Test Aircraft (VISTA).

Remark 7.3

Linear parameter-varying (LPV) models and controllers are linear systems whose
coefficients have a parametric dependence. Note that the assumptions made about the
parameters appearing in an LPV system are problem specific; the parameters may be
constant or time-varying, they may be unknown or measured, or any combination of
these. For example, an LPV controller is a linear controller whose coefficients depend
on some measured parameters. 0

LPV Model of the F-16 VISTA Longitudinal Dynamics

The actual longitudinal dynamics of the VISTA aircraft , which vary significantly with
flight condition (i.e. altitude h and Mach number M), are denoted by 9(h, M). For the
pitch-rate control design, 9(h, M) is modeled using the standard short period equations
of motion (7.3.1) and a first order approximation of the actuator dynamics (7.3.2).

d [a]
dt q -
_ [Za
M a
1] [ aq ] + [Z6.]
Mq Mo.
0 ed
(7.3.1)

where, a is the angle-of-attack, q is the pitch-rate, and Oed is the elevator deflection.

-20.2
(7.3.2)
s + 20.2
where, oec is the elevator command.
320 7 Case Studies in Flight Control

Remark 7.4

Note that the negative sign in (7.3.2) is simply a result of the "command" sign con-
vention used. In standard aircraft modeling (e.g. (7.3.1)), a positive elevator deflection
produces a negative change in pitch-rate. The command sign convention used here
(7.3.2) was chosen so that a positive elevator command produces a positive change in
pitch-rate, which requires a negative elevator deflection. 0

At trimmed level flight, the dimensional coefficients Za, Ma, Mq, Zo. and Mo. depend
mainly on the altitude h and Mach number M . An LPV version of (7.3.1) for VISTA
was developed by trimming and linearizing the United States Air Force Research Lab-
oratory's high fidelity nonlinear simulation model of the F-16 VISTA at the operating
conditions

M = [0.35 0.45 0.55 0.65 0.75 0.85]

h = [1000 5000 15000 25000][ft]

and fitting the corresponding data for the dimensional coefficients with polynomial func-
tions of hand M . The resulting expressions for the coefficients are given by (7.3.3).
The dependence of these parameterized coefficients on flight condition is shown in Fig-
ure 7.27.

Za(h, M) = 0.22 - 4.1 . 1O-7h - 2.6M + 5.15 . 10-5 Mh

Ma(h, M) = 17.1 - 8.07· 1O- 4h - 68.4M
+3.31 .10- 3 Mh + 56.2M 2 - 2.92 . 10- 3 M 2h
Mq(h,M) -0.228 + 7.06 . 1O- 6h - 2.12M + 4.86 . 10- 5 Mh (7.3.3)
Zo.(h,M) = -1.38 .10- 3 + 8.75 · 1O- 8h - 0.34M + 7.98.10- 6 Mh
Mo.(h,M) -8.16 + 1.73 · 1O- 4h + 40.6M
-8.96 . 10-4 Mh - 99.3M2 + 2.42 . 10-3 M 2 h

The LPV model of the short period equations of motion (7.3.4) obtained when these
polynomial expressions (7.3.3) are used with (7.3.1) is given by

d[O']_[Za(h,M)
dt q -
1
Ma(h, M) Mq(h, M)
][0'] + [Zo.(h,M)]
q Mo.(h, M) 8 ed (7.3.4)

and denoted by GLPV(h, M) . Then, the model used to represent Q(h, M) for the pitch-
rate control design is given by

G(h, M) = GLPv(h, M) · Go (7.3.5)

and accurately represents the F-16 VISTA's short period dynamics throughout the flight
envelope h E [5000; 25000] [ft] and M E [0.4; 0.8]. Of course, even though G(h, M)
accurately represents Q(h, M) , there are obviously modeling errors , which results in
7.3 Large Envelope Fligbt Control Satisfying H-Infinity Specifications 321

x 10' l,.(h.M)
2.sF"7'""""J'-70

Figure 7.27. Dependence of the parameterized dimensional coefficients {7.3.3} on flight con-
dition .

model uncertainty denoted by D. m , where D.m represents the difference (in a multiplica-
tive sense) between the actual aircraft dynamics and the model (7.3.5). That is, we
assume

Q(h, M) = G(h, M)(l + D. m ) (7.3.6)

The unstructured (i.e. non-parametric) uncertainty D. m is often represented (as de-

scribed in Section 5.3) by

D. m= W a · D. (7.3.7)

where 1ID.lloo < 1 and Wa defines the frequency content of the uncertainty. As shown
in Figure 5.36, W a is typically small at low frequencies where the model represents the
actual system fairly well, and larger at high frequencies where unmodeled dynamics are
prevalent. Then, substituting (7.3.7) in (7.3.6) gives,

Q(h, M) = G(h, M)(l + Wa · D.) (7.3.8)

as depicted in Figure 7.28.

F-16 Pitch-Rate Control Design Specifications

The objective is to design a controller for the F-16 VISTA that provides robust 'Levell'
pitch-rate command tracking throughout the design envelope h E [5000; 25000) [ft)
and M E [0.4; 0.8) (i.e, the flight envelope for which (7.3.5) is valid). This design
322 7 Case Studies in Flight Control

Figure 7.28. Block diagram representation of (7.3.8).

=-oo
o
z,
Gl
"C
::J
; 2n1-- +--I
;(

0.2 0.4 0.6 0.8 1.2 1.4

Mach

Figure 7.29. VISTA flight envelope.

envelope includes a large part of the VISTA aircraft's test flight envelope [148], as shown
in Figure 7.29. 'Level l' (i.e. acceptable) time domain handling quality specifications
for the pitch-rate response [147] are shown in Figure 7.30 and Table 7.3.
A reference model that satisfies the 'Levell' pitch-rate 'Handling Quality Specifi-
cations' over the entire design envelope is given by

42
Ge = -2: - - - - - - - . , .2 (7.3.9)
8 + 2 . 0.5 . 4 . 8 + 4

In the sequel, the reference model Gd will be referred to as the "desired dynamics".
Then , the design objective can be satisfied by designing a controller that "forces" the
closed-loop system to behave like the desired dynamics Gd (7.3.9) throughout the design
envelope.
7.3 Large Envelope Flight Control Satisfying H-Infinity Specifications 323

Figure 7.30. Pitch-rate handling qualities specifications.

Table 7.3. P'Itch-Rate Handlinil: Qualiity Speeificati.ons.

Parameter Level I Level II
t l max 0.12 [s] 0.17 [s]
6.Q2/6.QI max 0.30 0.60
6.t max 500/ VT [s] 1600/ VT [s]
6.t min 9/ VT [s] 3.2/ VT [s]
VT represents the true velocity [ft/s]

Large Envelope Flight Control Design

Control Architecture and Design Objectives
The control architecture used in this example was first applied to large envelope flight
control in [53] and is based on the control structure shown in Figure 1.12, which pro-
vides insensitivity to disturbances and model uncertainty. Furthermore, as implemented
here, this architecture enables the desired dynamics of the closed-loop to be explicitly
defined. The control architecture used here for the VISTA pitch-rate controller is shown
in Figure 7.31, where 9(h, M) denotes the open-loop aircraft dynamics from elevator
command to pitch-rate response and Gd denotes the desired pitch dynamics of the air-
craft. The two filters Gd and Q form the pitch-rate controller when utilized as shown
in Figure 7.31. In general, designing the two degree of freedom controller shown in
Figure 7.31 would involve the design of Doth the filter Q and the desired dynamics Cd'
However, as in this example, Cd is usually selected based on some knowledge of the
desired closed-loop performance and independent of Q. Then, the objective is to design
the filter Q so that a given 'pitch-rate command Oc provides the desired response despite
changing flight condition (h, M), disturbances d, sensor noise n, and model uncertainty.
324 7 Case Studies in Flight Control

I
I
Pitch-Rate Controller I
I

8C I
I 8ec :

L...-_---' I
----------------------

Figure 7.31. Pitch-rate controllerarchitecture.

That is, design a Q so that the closed-loop represented by Figure 7.31 has

q = Gd8c (7.3.10)

as its input - output relation (within the frequency range of interest) throughout the
design envelope.

Remark 7.5

In contrast to the control architecture shown in Figure 1.12, the controller shown in
Figure 7.31 does not contain a nominal plant model, but rather the desired closed-loop
dynamics represented by Gs . 0

Remark 7.6

The relative degree of the filter Q is chosen to be at least equal to the relative degree
of Gd for causality of Q/G d • 0

Remark 7.7

It may not be immediately obvious that there is a Q such that the controlled aircraft
provides the desired response (7.3.10) . However, the following discussion will provide
some insight regarding the form of Q required to do just that and the design to follow
will in fact find such a Q. 0

Investigating the input-output transfer functions associated with the proposed ar-
chitecture provides insight regarding the required form of Q and the expected character-
istics of the resulting closed-loop. The loop-gain for the closed-loop aircraft represented
by Figure 7.31 is

L(h M) = ~. 9(h ,M) (7.3.11)

, 1- Q Gd
7.3 Large Envelope Flight Control Satisfying H-Infinity Specifications 325

As described in Section 5.3, the model regulation H, disturbance rejection (sensitiv-

ity) S, sensor noise rejection (complementary sensitivity) T, and performance (error)
E, transfer functions of the controlled system are given by
q Ge : Q(h,M)
H(h,M} .- = (7.3 .12)
Oc Ge : (1 - Q) + Q(h, M) . Q
q 1 G« : (1- Q)
S(h, M) .- = (7 .3.13)
d 1 +L(h,M) G d • (1- Q) + Q(h. M)· Q
-q L(h,M) Q(h,M} . Q
T(h,M) .- -=
n 1 + L(h, M) - Gs : (1 Q) +Q(h,M) . Q
(7.3 .14)

E(h,M) .- Gd - H(h,M}. (7.3.15)

It is obvious from (7.3 .12) - (7.3.14) that the design objective can be satisfied by the
selection of an appropriate low pass Q filter with unity gain, which results in q/oc ~ Gd
and qf d ~ 0 at low frequencies where Q ~ 1 and q/n ~ 0 at high frequencies
where Q ~ O. In order to explicitly show the characteristics of this control architecture,
Q is set to 1 in (7 .3.12) and (7.3 .13) and Q is set to 0 in (7.3.14) to obtain

HQ=l(h, M)
= .!L
Oc
= GdQ(h,
• Q(h,M) = G
M) d (7.3.16)

q 0
SQ';l(h, M} = d = Q(h, M} = 0 (7.3.17)

TQ=o(h,M} = -q = ~= 0 (7.3 .18)

n Gd
which clearly shows that good model regulation and disturbance rejection are achieved
within the bandwidth of Q (despite variations in Q(h, M)}. Of course, it is not possible
to achieve model regulation or disturbance rejection at frequencies above the actuator's
bandwidth, so nothing is gained by having the bandwidth of the Q filter significantly
larger than that of the actuator. Furthermore, as shown by (7.3 .18), limiting the
bandwidth of Q is required to reject high frequency sensor noise.
Parameter Space Control Design
In this section, the parameter space techniques of previous chapters are used to design
the filter Q to obtain a pitch-rate controller that robustly satisfies the design objective
(i.e. provides a pitch-rate response as defined by (7.3.1O) as closely as possible despite
varying dynamics and uncertainty) . This is accomplished by mapping r- and B-stability
specifications into the free parameter space of Q. The desired dynamics given by (7.3.9)
mandates a Q filter with a relative degree of at least two for implementation purposes
(i.e, causality of Q/Gd ) . Opting for a simple controller, the Q filter structure is chosen
as
2
W
Q= Q 2 (7.3.19)
S2 + 2DQwQS+ wQ
giving free parameters wQ and DQ. Then, parameter space methods are used to map the
r-stability specification (7.3.20) and the B-stability specifications (7.3.21) and (7.3.22)
326 7 Case Studies in Flight Control

into the (wQ, D Q) parameter plane resulting in the feasible region of the (wQ, D Q) space
(i.e. the set of (wQ, DQ) that satisfy the mapping equations).

Roots (Pce(h, M, WQ , DQ)) (Hurwitz Stability) (7.3.20)

I/WA . T (h, M ,wQ, D Q)l/ oo (Robust Stability) (7.3.21)
I/Wp ' E(h, M,wQ, D Q)I/ oo (Nominal Performance) (7.3.22)

Equation (7.3.20) ensures that all of the roots of the closed-loop characteristic equation
given by

pce(h, M,wQ,DQ) := num (1 + L(h, M,wQ,DQ)) (7.3.23)

are in the negative complex plane (i.e. ensures stability of the nominal closed-loop
throughout the design envelope).

Remark 7.8

T,E and Lin (7.3.21), (7.3.22), and (7.3.23) are the same as T, E and L defined in
the previous section, except that the actual VISTA longitudinal dynamics represented
by g(h, M) are replaced by the model of these dynamics G(h, M) as defined in (7.3.5).
o
The weights WA and Wp in (7.3.21) and (7.3.22) are given by [53]
W 28 + 1256·0.2 (7.3.24)
s + 1256 ·2
A

Wp = 0.4 s + 10 ·10 (7.3.25)

8 + 10 ·0.4

and shown in Figure 7.32.

As discussed previously, the uncertainty weight W A is used to account for model
uncertainties 6.m (7.3.6) in G(h, M) (7.3.5). Examining Figure 7.32 shows that WA
represents 20 percent model uncertainty at low frequencies and that the uncertainty
grows rapidly at higher frequencies, where unmodeled dynamics are significant. Then ,
satisfying equations (7.3.20) and (7.3.21) ensures robust stability throughout the design
envelope despite the uncertainties represented by W A .
The performance weight Wp penalizes the error between the desired and actual
pitch-rate response and is large in the frequency range where good model matching is
desired. Figure 7.32 shows that the error in the controlled aircraft's pitch-rate response
must be less than 10 percent within the frequency range of interest in order for the
performance specification to be satisfied. Thus, satisfying (7.3.22) ensures nominal
performance throughout the design envelope; that is, it ensures that the actual response
is within 10 percent of the desired response for the nominal model G(h , M) (7.3.5) (i.e.
(7.3.6) with 6.m = 0).
Further insight regarding the interpretation of the weights (7.3.24) and (7.3.25) can
be obtained by reviewing the descriptions of (5.3.23) and (5.3.28) and in [109], where
7.3 Large Envelope Flight Control Satisfying H-Infinity Specifications 327

Uncertainty Weight (7.3.24) Performance Weight (7.3.25)

£3(1.
6
5
-4
0.8
3
0.6
0.4

0.2t==;===:::;::::===~_----,-;-~
10· 10' 10' 10· lrf 10'
OJ fradls] OJ fradls]

Figure 7.32. Design weights.

a detailed explanation of weights appropriate for flight control is given. Additionally,

further interpretation of (7.3.21) and (7.3.22) can be seen by examining the weighted
design model in Figure 7.33, where

(7.3.26)

(7.3.27)

,, Pit ch-Rate Cont roller :

Jc ', Dec:

,
------- -- - - - - - - - - - - - - -

' - - - - -- - - - - - - .-1 C d 1 - - - - - -- - - -------'

Figure 7.33. Weighted design model.

328 7 Case Studies in Flight Control

Hurwitz St abilit y (7.3.20) Robu st Stability (7.3.21)

3 3

17
2

0'
Q
o
1 \/ -:
..
'-

~~ o 0 10 20 30 40 50 o 10 20 30 40 50
WQ WQ

Nominal Performance (7.3.22) Feasible Region for (7.3.20)-(7. 3.22)

~ I
3~~~~-~~~~---,

Q
0' 1 j
.
O

-1 .

~~O 0 10 20 30 40 50 ~~O 0 10 20 30 40 50
WQ WQ

Figure 7.34. Feasible (wQ,DQ) region for h = 5000[ft], M = 0.8.

Remark 7.9

The weighted" design model in Figure 7.33 along with the specifications given by
(7.3.20) - (7.3.22) are in a form typically used for "classical" Hoo control design . In
fact, a very similar problem formulation was used in [188] and [52] to design an LPV
controller for VISTA using J,L-synthesis (i.e. H oo design method) . Thus, this example
clearly demonstrates that the parameter space techniques presented in Chapter 5 can
be used to design controllers that satisfy Hoo design specifications. 0

After (7.3.20) - (7.3.22) are mapped to (wQ, D Q ) , the feasible region that simultane-
ously satisfies all three requirements is formed by the intersection of the feasible region
obtained from each individual requirement. Note that since (7.3.20) - (7.3.22) depend
on flight condition (h, M) , the set of feasible (wQ, DQ) satisfying these equations also
depends on (h, M) . Figure 7.34 shows the results of mapping (7.3.20) - (7.3.22) into
(wQ, D Q) for h=5000 [ft] and M=0 .8, where the X indicates the feasible region.
Since the objective is to design a controller that satisfies these specifications over
the entire design envelope, (7.3.20) - (7.3.22) are mapped for as many additional oper-
ating conditions deemed necessary to ensure that the specifications are met throughout
the design envelope. The region in the (wQ, D Q) plane that satisfies (7.3.20) - (7.3.22)
at the four "corners" of the design envelope is shown in Figure 7.35. Note that for
this design, the feasible region obtained by mapping the corners of the design enve-
7.3 Large Envelope Flight Control Satisfying H-Infinity Specifications 329

lope is not reduced by mapping additional flight conditions. The design is completed
by picking one point from this feasible region. This can be done arbitrarily; however,
if after mapping all design requirements, the feasible region is still large, additional
demands can be placed on the controlled system. For example, the feasible solution
that minimizes the structured singular value could be selected, or the original design
specifications could be "tightened", which was done here. In this design, after map-
ping (7.3.20) - (7.3.22) using the "initial" weights, the feasible region was rather large,
so the uncertainty weight W~ (7.3.24) was adjusted to account for more uncertainty,
which increases the guaranteed robustness of the resulting closed-loop system. After
changing the uncertainty weight to that given in (7.3.24), the final design, which is
marked with an X in Figures 7.34 and 7.35 was chosen (rather arbitrarily) as D Q= 0.7
and wQ=40 [rad/s] .

Feasible Region Feasible Region (zoomed)

3.---- - - - - - - - ---,

40 20 30 40 50
WQ

Figure 7.35. Feasible region obtained by mapping (7.3.20) - (7.3.22) into (wQ,DQ) at the four
corners of the design envelope .

Controller Evaluation
Verifying Mapping Equations
Selecting D Q and wQ from the feasible region, as shown in Figure 7.35, ensures that
the mapping equations (7.3.20) - (7.3.22) are satisfied, which can be easily verified.
Verifying that the r-stability specification (7.3.20) is satisfied simply involves checking
that the eigenvalues of the closed-loopsystem are negative at all flight conditions, which
they are. Verifying that the E-stability (i.e. H oo ) specifications (7.3.21) and (7.3.22)
are satisfied can be easily checked by examining the frequency response magnitude (i.e.
Bode magnitude) plots of W~ · T(h, M,wQ, D Q) and W p ' E{h ,M,wQ,DQ) to ensure
that they are less than one at all flight conditions and frequencies. Figure 7.36 shows
these frequency response magnitude plots at the four corners of the design envelope.

Linear Simulations
Linear simulations were performed throughout the design envelope using the controller
designed in the previous section. The results corresponding to the four corners of the
330 7 Case Studies in Flight Control

Robust Stability (7.3.21) Nominal Performance (7.3.22)

-;;0 0.8 ;;0 0.8

~
8" 0.6
8"
~ 0.6
s= s=
i=" 0.4
W 0.4
~<l
0.2
) ~
a.
0.2

~0-3 10·
~~
~
10' 10'
00"
1 10· 10' 10'
ll) [rad/s) ll) [rad/5)

Figure 7.36. Mapping equations {7.3.21} and {7.3.22} verified at the four corners of the design
envelope.

design envelope are shown in Figure 7.37, which also shows the pitch-rate command
(the step) and the response of the desired dynamics Cd (7.3.9). It is clear from this

1.5r----~--~--~--_,

2 4 6 8
timelsl

Figure 7.37. Linear simulation results.

figure that the pitch-rate response demonstrates predicted 'Levell' handling qualities
at all four corners of the design envelope. At three of these flight conditions, the
pitch-rate response is essentially the same as the desired response. While the response
at h = 25000[ft] and M = 0.4 does have a little more overshoot than the desired
response, it is certainly within the 'Level l' specifications.
7.3 Large Envelope Flight Control Satisfying H-Infinity Specifications 331

Non-linear Simulations
High-fidelity, nonlinear simulations were also performed, since achieving the desired
performance in linear simulations is not sufficientwhen assessing the controlled dynamic
behavior of the highly nonlinear F-16 VISTA aircraft . The nonlinear simulations also
demonstrated a predicted 'Level l' pitch-rate response throughout the design envelope.
Figure 7.38 shows the simulation results when the maneuver starts at the center of the
design envelope (i.e. h=15000 [ft], M=O.6). Again, the pitch-rate plot includes the step
command and the response of the desired dynamics.

Remark 7.10

In order to better assess the merits of the design procedure presented, the VISTA
pitch-rate controller designed here (i.e. parameter space controller) was compared with
previous VISTA pitch-rate control designs [52] and [188], which also provided 'Level l'
responses throughout the same design envelope. The pitch-rate controllers presented
in [52] and [188] were designed using H oo and It-synthesis methods and are both LPV
controllers (i.e. gain-scheduled). This comparison highlighted a number of noteworthy
points.
First, comparing the weighted design models Figure 7.33 used for the different de-
signs showed that the same VISTA model C(h, M), desired dynamics Cd, and perfor-
mance weight W p were used for all three designs. However, the parameter space design
used a considerably more demanding uncertainty weight Wa (i.e. accounted for more
uncertainty) than the LPV designs.
Second, a detailed robustness analysis (i.e. It-analysis) revealed that the parameter
space controller guarantees a much higher level of robust performance (i.e. guaranteed
level of performance despite uncertainty) than the LPV controllers. This difference in
guaranteed robust performance is actually quite remarkable, since the LPV controllers
were designed using methods (It/ H oo ) intended to "maximize" this robustness, while
robust performance was not even an explicit design specification for the parameter space
controller. Recall, the design specifications for the parameter space controller were
robust stability (7.3.21) (i.e. stability despite uncertainty) and nominal performance
(7.3.22) (i.e. performance without uncertainty) .
Third, comparing the design methods revealed that the parameter space proce-
dure presented provides a significantly more straightforward and transparent design
approach. With the parameter space approach, the design specifications are mapped
into the plane of two design parameters (e.g. D Q and wQ) and a feasible controller is
selected. If no feasible controller exists, the conflicting specifications are immediately
obvious, as the feasible regions of the individual specifications do not have an inter-
section. Additionally, if many feasible controllers exist (i.e. the intersection of feasible
regions is large), then it is immediately obvious which specifications can be strength-
ened. On the other hand, the LPV procedures involve several design iterations, and
it is not obvious whether or not the specifications are feasible until they are satisfied.
Furthermore, if after several iterations the design specifications are not satisfied, it is
not clear whether it is because they are infeasible, because additional iterations are
necessary, or because poor design choices were made in previous iterations.
332 7 Case Studies in Flight Control

12 14
(\
I -
10 12
8
0;10
~6 CD
"C
CD ";;8
~4 s:
sr
«6
Co

2
o- \ ~
4

2 4
V 6 8 2Q 2 4 6 8
time lsI time Is]
3 640
0;2
CD 620
"C
';1
0 ~600
13CD 0 ~
~
...0
0-1 ·i 58O

<;;-2 ~560
>
CD
ilL3 540

-4 520
0 2 4 6 8 0 2 4 6 8
time Is] time Is]
4
X1O 4
X1O
1.62 1.62
1.6 1.6
1.58 1.58
~ ~
CD 1.56 CD 1.56
"C "C
::;] ::J
E1.54 E 1.54
< -c
1.52 1.52
1.5 1.5
1.48 1.4eoo
0 2 4 6 8 550 600 650
time lsl Velocity Ift/s]

Figure 7.38. Nonlinear simulat ion results.

7.3 Large Envelope Flight Control Satisfying H-Infinity Specifications 333

Finally, comparing the controller architectures showed that the parameter space
controller is easier to implement and maintain. It only requires the feedback of the
pitch-rate (i.e. the controlled variable) , while both LPV controllers depend on the
flight condition (i.e. gain-scheduled) and require the feedback of both the pitch-rate
and the angle-of-attack. Furthermore, the parameter space controller consists of just
two 2nd order filters, while the LPV controllers were 7th and 10th order systems. 0

Concluding Remarks
This case study demonstrated the "simplistic power" and transparency of the param-
eter space design method by using r- and B-stability specifications to design a robust
controller for the extremely challenging "real-world" problem of large envelope flight
control of high-performance aircraft. The pitch-rate controller designed here for the
VISTA F-16 aircraft satisfies (Hoo) robustness and performance specifications through-
out the large design envelope without gain-scheduling, and high fidelity nonlinear sim-
ulations showed that it provides a robust 'Level l ' pitch-rate response throughout the
design envelope. Furthermore, this controller is far simpler and provides better ro-
bust performance than previous LPV controllers designed for VISTA to meet the same
objective (Le. robust 'Level l ' pitch-rate response throughout the design envelope)
using H oo and It-synthesis methods . This rather remarkable achievement is the result
of both the parameter space method's explicit , non-conservative treatment of physi-
cal parameters (e.g. flight condition parameters hand M) and the properties of the
two-degree of freedom control architecture used. This control architecture is extremely
attractive for a number of reasons, most notably, its insensitivity to disturbances and
model uncertainty and its unambiguous structure, which enables the explicit definition
of the desired closed-loop dynamics. Together, this control architecture and parameter
space techniques were shown to give a straightforward and transparent procedure for
designing robust, large envelope controllers for high-performance aircraft that provide
explicitly defined desired closed-loop dynamics without gain-scheduling.
8 Robustness Analysis by Value Sets

In this chapter, we return to Hurwitz-stability analysis of a polynomial family

p(s , q) = [1 s . . . s"] a(q), q E Q, (8.0.1)
where Q is a hyperrectangle ("box") .
The obvious limitation of the parameter space approach is that it can practically
handle only a small number of uncertain parameters in q . This limitation is avoided
by the value set approach in this chapter. On the other hand , this approach has other
restrictions. It is primarily suited for robustness analysis, not for design; the operating
domain q E Q must be fixed; and the uncertain parameters can enter a(q) only in
restricted structures.
The basic idea of the value set approach may be explained by the imaginary axis
crossing conditions (2.2.1):
RepUw, q) = ao(q) - a2(q)w 2 + a4(q )w 4 - .. . = 0,
(8.0.2)
ImpUw, q ) = al(q)w - a3(q)w3 + as (q )wS - . • . = o.
This condition may be tested graphically for a fixed q = q* from a plot of pUw , q*) =
RepUw , q*) + j ImpUw , q*) , w 2:: o. Equation (8.0.2) is satisfied for those special values
q*, for which this plot passes through the origin. For uncertain q , the plot blows up
to a family of plots (for example, generated by gridding the Q-box). The polynomial
family is stable if and only if i) it contains a stable polynomial and ii) for no q E Q the
plot passes through the origin, i.e. the origin is excluded from the family of plots. This
test is known as the zero exclusion condition.

8.1 Mikhailov Plot

Consider first a fixed q . The plot pUw ), w 2:: 0 in the complex p-plane is referred to
as a Mikhailov plot. Stability or instability of the polynomial can be determined by
inspection of this plot.
Stability conditions stated in terms of frequency plots have a long history. In their
most general form, these conditions were given by Cauchy's principle of the argument
around 1829. For the specific case of polynomials, conditions of this form also fol-
low from the Hermite-Biehler theorem [107], [50], see also [94]. The graphical use of
8.1 Mikhailov Plot 335

the argument principle was introduced to the engineering community by Nyquist [157]
in 1932. In 1938, Mikhailov [145] gave simpler graphical conditions for analyzing sta-
bility of known polynomials. In 1944 and 1947, Leonhard [131] and Cremer [68], re-
spectively, gave similar conditions. For this reason, the Mikhailov plot is sometimes
called the Cremer-Leonhard plot. Extensions of this method from known polynomials
to families of polynomials are also available. In the 1950s, a frequency plot method
for families of polynomials was originated by Curtis. This method was made known to
the control community by Zadeh and Desoer [202]. This approach has been labelled by
Barmish [40] as the value set approach using the zero exclusion theorem (Theorem 8.6) .
It is not fully clear which author should get the credit for this theorem. An assertion
close to Theorem 8.6 is given in a book by Zadeh and Desoer [202] . Also, Frazer and
Duncan [89] were not far from this result. A historical review of the zero exclusion
theorem was given in a survey by Barmish [39] .
This section will review both the Mikhailov stability conditions and the principles
of the value set approach.

Theorem 8.1 (Mikhailov , Leonhard and Cremer)

The polynomial
p(s) = ao + als + ...+ ansn , an> 0, (8.1.1)
is stable if and only if the frequency plot pOw), 0 ::; w < 00 satisfies the following
two conditions:
L pUO) = ao > 0, i.e, the plot starts on the positive real axis.
ii. As w increases, the plot of pOw) encircles the origin in a counterclockwise
direction and its phase goes to n~ for w -+ 00 .
o

Proof

To prove necessity, we first assume that the polynomial is stable and show that th e
Mikhailov conditions must be satisfied. The first condition, ao > 0, is satisfied for
a stable polynomial. The second condition follows from the argument principle.
The stable polynomial p(s) can be factorized :

IT Ijw - sile jarguw-.;)

n
IT Ow - s;)
n

pOw) = an an

fI
i= l i=l

= (an Ijw - Sd) e Ei=l arguw-.;),

j

.=1
where all Si have a negative real part. If a root is real, say S 1, then a :=
arg Ow - sd increases from zero to ~ as w goes from zero to 00 (see Figure 8.1) .
If a root is complex, say S 2, then S3 = 82 is also a root of p(s). The argument
of Ow - S2) changes' from -/3 to ~, the argument of Ow - S3) from /3 to ~ as w
goes from zero to 00 . So the total argument variation of a conjugate complex root
336 8 Robustness Analysis by Value Sets

jw

jw

/
/
/
/
/
/
/
/

_
/
/

s ,~~:y

Figure 8.1. The variationof arg (jw - s;)

pair is 2~ . Adding the contribution of each root , the total argument variation of
all roots is n~ . Thus, p(jw) must encircle the origin corresponding to the second
condition .
To prove sufficiency, assume that the Mikhailov conditions are satisfied. It
must be shown that the polynomial is stable . We will show the equivalent state-
ment: whenever a polynomial is unstable , it will violate the Mikhailov conditions.
Assume that the polynomial has m roots in the right half plane and n - m roots
in the left half plane. By the argument principle, each root in the right half plane
contributes a change of the argument of -1r /2 as w goes from zero to infinity.
Thus , the total phase change is (n - 2m)1r /2 and the second Mikhailov condition
is violated. In the special case of a root on the positive imaginary axis, it has to
be circumvented by a small semicircle into the left half plane in order to count
the root correctly as unstable . Thus , the phase jumps by -1r as w crosses the
imaginary root (Figure 8.2).
o
Note that the above proof also shows that the phase of the Mikhailov plot for a
stable polynomial is monotonically increasing, i.e,

8 arg~(jw)} > 0 for all w 2: O. (8.1.2)

For every polynomial, the total variation of the phase of p(jw) of (n - 2m)1r /2 determines
the number n - m of left half plane roots and the number m of right half plane roots.
8.1 Mikhailov Plot 337

jw

Figure 8.2. The variation of arg Uw - 8i) for imaginary roots

Note also that for one or more unstable roots, the phase of the Mikhailov plot
cannot exceed (n - 1)~, i.e. once the Mikhailov plot has reached the n-th quadrant as
w increases, stability is guaranteed. This condition gives an upper bound w+ for the
frequencies w that must be tested, where

argp(jw+) = (n - 1)~. (8.1.3)

The minimal distance between the Mikhailov plot of a stable polynomial and the origin
is not an absolute measure for the stability margin because p(s) may be multiplied with
an arbitrary factor without changing its roots. This distance is, however, a relative
measure that indicates frequencies that are critical for stability. If the Mikhailov plot
passes through the origin, then the polynomial has a root on the imaginary axis at
the corresponding frequency. Given the frequency plot of a polynomial, the Mikhailov
conditions indicate at a glance whether or not the polynomial is stable .
The Mikhailov condition is often interpreted in another form. The phase condition
is equivalent to requiring the frequency plot to intersect the axes in the following order:
positive real (for w = 0), positive imaginary, negative real, negative imaginary, positive
real, ... , until n such intersections have been made. This formulation is equivalent to
the Hermite-Biehler theorem [107], [50] that requires an "interlacing property" of the
roots of Re p(jw) = 0 and 1m p(jw) = O.

Example 8.2

Recall the characteristic polynomial of the crane from (1.5.18) with me = 1000,
e= 12, mL = 1500, k1 = 500, k 2 = 2850 and k 4 = O.

p(s, k 3 ) = 1200084 + 3420083 + (31000 - k3 ) S 2 + 28 500s + 5000.

For k 3 = 0, the plot satisfies the Mikhailov stability condition, for k 3 = 16000 it does
not, see Figure 8.3. Between these two values there must be a stability boundary, for
which the Mikhailov plot passes through the origin. 0
338 8 Robustness Analysis by Value Sets

4
X 10
2.--------.-----,----....-----,

O!-----,..c......;..----t\-..l..--..;..-----j

~-2 ... . .
S k 16000
--4 .

-6

-8 L...-_ _---'-_ _--->---'- -'---_ _- - '

-2 -1 o
Rep

Figure 8.3. Frequency plot of two polynomials

8.2 Value Sets and Zero Exclusion

An obvious way to use the Mikhailov conditions for a family of polynomials is to com-
pute the frequency plot of each polynomial in the family and to check the Mikhailov
conditions one polynomial at a time . As was discussed in the root set section, the family
normally has an infinite number of members, so this approach is not really possible.
In practical terms, however, computing the frequency plots for a sufficiently dense grid
of the parameter set is adequate. The frequency plots are placed together on a single
graph. This procedure is illustrated by two examples .

Example 8.3

Again, recall the characteristic polynomial of the crane from (1.5.18) with me =
1000, £ = 12, mL = 1500, k1 = 500 and k4 = O.
p(s, k 2 , k 3 } = 12 000S4 + 12k 2s3 + (31000 - k 3 }S 2 + lOk 2 s + 5000.

For the range of gains

1000 ~ k2 ~ 11 000, (8.2.1)
12 000 ~ k3 ~ 17 000, (8.2.2)
a 6 x 6 grid point grid was used to compute the frequency plots; they are displayed
in Figure 8.4. The set contains plots that pass through zero. The family of control
systems is not robustly stable.
This example was repeated using another range for k 3 :
o s k3 s 5000. (8.2.3)
8.2 Value Sets and Zero Exclusion 339

4
X 10
4,----....---..,.---.....-==---.----,
3 -

-
~
S 0 f----;----;---i~

-1

-2
-3 '

-4
-1 .5 -1 -0.5 0 0.5 1
Rep x 10 4

Figure 8.4. Frequency plots of an unstable family of polynomials

The corresponding collection of frequency plots were computed and displayed in Fig-
ure 8.5. The family defined by this range of gains is known to contain at least one
stable polynomial for k 3 = 0, see Figure 8.3. Figure 8.5 shows that the frequency plots
do not intersect the origin and hence none of the complex numbers pOw, k2 , k3 ) equal
zero. This implies there are no possible roots on the jw-axis, and hence by the boundary
crossing theorem, the polynomial is robustly stable for the given range of gains. The
line of reasoning used in these two examples is true in general and will be stated in
theorem form. 0

Theorem 8.4

The polynomial family P{s, Q) is robustly stable, if and only if:

i. There exists a stable polynomial p{s, q) E P{s, Q).
ii. There does not exist a q E Q such that the frequency plot pOw, q), w 2: 0,
intersects the origin.

o
The gridding approach suggested in this section has one of the same drawbacks as
computing the multiparameter root set. It is quite easy to select a grid that would
cause the computations to take an excessively long time. For this reason, it is generally
better to use a slightly different approach in computing the set of frequency plots.
Rather than computing the frequency plot pOw, q), w 2: 0, for each q on a grid
of Q, it is advisable to compute the value set

POw,Q) = {p(jw,q)1 q E Q}
340 8 Robustness Analysis by Value Sets

-
~
s 0
-1

-2

-3
-4 u....JIL>...LI..loL.I..........->-->O--"'-_ _---'- '--_--'
-1.5 -1 -0.5 0 0.5 1
Rep 4
x 10

Figure 8.5. Frequency plots of a stable family of polynomials

for each w on a grid of frequencies from a to +00. This slight modification makes it
possible to reap significant computational savings by exploiting the structure of the
uncertain polynomial in terms of the uncertain parameters qi in q .

Example 8.5

. Consider the uncertain polynomial of Example 8.3 with k 2 E [1000 ; 11 000] and k 3 E
[0 ; 5000]. Since the gains enter the coefficients in an affine manner , value sets of this
polynomial are four-sided convex polygons. In addition, the vertices of these polygons
are determined by the extreme values of the gains. These are the four complex numbers
p(jw,k; , k;), p(jw, k; , kt) , p(jw,kt, k;) , and p(jw, kt, kt) . From just these four gain
combinations rather than from a large grid of gains, it is possible to completely deter-
mine the value set at each frequency. For a grid on w in the interval [0.5 ; 1.5], the value
set is the family of rectangles in Figure 8.6. Obviously, it is much simpler to construct
the Mikhailov set by a one-dimensional grid on w rather than by a multi-dimensional
grid of q. 0

The result of the above discussion is summarized in the following zero exclusion
theorem .

Theorem 8.6 (zero exclusion)

Given a polynomial family P(s , Q) as in (8.0.1), the set P(s, Q) is robustly stable ,
if and only if:
i. There exists a stable polynomial p(s, q) E P(s , Q).
ii. a f/. P(jw, Q) for all w ~ O.
8.3 Interval Polynomials, Kbaritonov Tbeorem 341

r--

ill
3

~
·· ..
2 .... ·: . . . .... ....
· .
· .
· .
... . ·: .. . . .. .. .:
1 · .
~ · . '--

~
S 0
-1

-2
-3
-4
-1 .5 -1 -0.5 0 0.5 1
Rep x 10
4

Figure 8.6. Value set for some discrete frequencies

8.3 Interval Polynomials, Kharitonov Theorem

In some simple structures of a( q), it is not necessary to plot value sets . Simple stability
tests can be derived, however, using the concept of zero exclusion from the value set .
The simplest structure is the interval polynom.

Definition 8.7

A polynomial
p(s, a) = ao + alS + ... + ansn (8.3.1)

with the uncertain coefficient vector

(8.3.2)

is called an interval polynomial, if a ranges over the box

A := {a I ai E [a; ; at], i = 0,1, . . . , n} . (8.3.3)

o
342 8 Robustness Analysis by Value Sets

For an interval polynomial, the uncertain parameters are identified with the polyno-
mial coefficients aj, that means aj is not considered as a function of other parameters.
Then each uncertain coefficient in (8.3.1) is independent of all other coefficients. An
interval polynomial generates the polynomial family

P{s, A) = {p(s, a) Ia E A }. (8.3.4)

In 1953, Faedo [83] formulated the problem of a necessary and sufficient condition for
robust stability of an interval polynomial. Faedo obtained only a sufficient condition.
In 1979, Kharitonov published a surprisingly simple necessary and sufficient stability
condition for interval polynomials [125]. Kharitonov's result is: only four polynomials
of the continuum A have to be checked for robust stability.

Theorem 8.8 (Kharitonov)

The polynomial family

(8.3.5)

is stable if and only if the following four polynomials are stable :

r: (s) = at + at s + a2"S2 + at S3 + at S4 + ,
p++(s) = at + ars + a2" s 2 + ais3 + ats 4 + , (8.3.6)
p-+(s) = ail + ars + ats 2 + ais3 + a4s 4 + ,
p--(s) = ail + als + ats 2 + ats 3 + a4s4 + .
o
The polynomials (8.3.6) are called Kharitonov polynomials. The superscripts indi-
cate the upper and lower bounds of the coefficients ao and al '

Remark 8.9

A simple rule for constructing the Kharitonov polynomials is the "Kharitonov

melody" ... two high, two low, two high, . . . with four possibilities for the two initial
values. The upper and lower bounds occur in the four polynomials as

+--++--+
++--++--
-++--++-
--++--++
o
8.3 Interval Polynomials, Kharitonov Theorem 343

The original proof of Theorem 8.8 by Kharitonov is complicated. Meanwhile, a

much simpler proof was found [36], which is based on the zero exclusion from the value
set of pOw, a) . First, we prove the following lemma , [70J:

Lemma.
For each fixed w = w· ~ 0, the value set pOw· , A) = {pOw· , a) I a E
A} is a rectangle with edges parallel to the coordinate axes and with
vertices determined by the values of the four Kharitonov polynomials
p+-(jw·), P++Ow·), P-+Uw·), p--Uw·), see Figure 8.7.
o

Imp
p-+Uw·).---+-------'v++Uw·)

PUw·,A)

Rep

Figure 8.7. Value set of an interval polynomial for fixed w = w·

Proof
For all w ~ 0 and all a E jRn+l , the real and imaginary parts of pUw, a) =
RepUw, a) + j ImpOw , a) have the lower and upper bounds
ail - atw 2 + a;w 4 - .• • S; RepUw, a) S; at - a;w 2 + atw 4 - •• • , (8.3.7)
w(a1 - atw 2 + a sw4 - • • •) S; ImpUw, a) S; w(at - aaw2 + atw 4 - • • •). (8.3.8)
Since Re pUw, a) is a function of the even indexed parameters ai and 1m pUw, a) is
a function of the odd indexed parameters ai, all bounds are mutually independent.
Since the two functions are continuous, the set PUw·, A) must be a rectangle. Its
edges are parallel to the coordinate axes.
The four Kharitonov polynomials defined in Theorem 8.8, can be represented
as follows:
sw
p+-(jw) = at - a;w 2 + atw 4 - . , . + jw(al- atw 2 + a 4 - •• •),
p++(jw) = at - a;w 2 + atw 4 - . , . + jw(at - aaw2 + atw 4 - • •• ) ,
(8.3.9)
p-+(jw) = ail - atw 2 + a;w 4 - . • • + jw(at - aaw2 + atw 4 - • •• ) ,
p--(jw) = ail - atw 2 + a;w 4 - • . • + jw(a 1 - atw 2 + asw 4 - • .• ) .
Obviously, for each w ~ 0 the vertices of the rectangle PUw, A) are the values of
the four Kharitonov polynomials .
o
344 8 Robustness Analysis by Value Sets

Using the lemma , we now prove Kharitonov's theorem .

Proof
The necessary part of the theorem is trivial. To prove sufficiency, assume that the
four Kharitonov polynomials are stable. Then , the Mikhailov plot (see Section 8.1)
of each Kharitonov polynomial starts on the positive real axis and circles the origin
in counterclockwise direction until its phase is mr/2. Since the edges of the value
set are parallel to the coordinate axes and their end points satisfy the Mikhailov
stability condition, the origin cannot enter into the value set through an edge of
the value set. Hence, all Mikhailov curves for pOw, a) with a E A are stable and
P( s, A) is stable.
o
For low order polynomials, even less than the four Kharitonov polynomials suffice
for a stability test .

Theorem 8.10 (Anderson , Jury, Mansour)

An interval polynomial family

p(s , a) = ao + alS + .. . + ansn, ail > 0 (8.3.10)

is robustly Hurwitz-stable, if and only if the following polynomials are stable:

n=3 : p+-(s),
n = 4: p+-(s) ,p++(s),
n= 5: p+-(s), p++(s) ,p-+(s) ,
n 2: 6 : p+-(s), p++(s), p-+(s), p-- (s) .
o
For n = 2 and n = 1, the conditions ai > 0 are necessary and sufficient.

Proof
Let n = 3 and assume that p+-Ow) is stable. Then, its Mikhailov plot traverses
through the quadrants I, II, and III, as illustrated by Figure 8.8, and with ail >
0, the same is true for all p(jw, a) E POw, A). Let n = 4 and assume that
both P+-Ow) and p++Ow) are stable as illustrated by the Mikhailov plots in
Figure 8.9. Then, with ail > 0, the same is true for the entire rectangle P.
Finally, for n = 5, p-+Ow) must also be stable to guarantee that all pEP are
stable, see Figure 8.10. For n > 4, the condition ail > 0 needs not be tested
separately, it is implied by stability for p-+ (s).
o
8.3 Interval Polynomials, Kharitonov Theorem 345

Imp

Rep

Figure 8.8. For polynomials of degree three, only p+-(s) has to be tested

Imp

Rep

Figure 8.9. For polynomials of degree four, only p+-(s) and p++(s) have to be tested

For testing the stability of a polynomial, plotting its Mikhailov curve is not necessary.
Using the Hurwitz test or factorizing the polynomials are other possibilities.
Exampl e 8.11

Recall the characteristic polynomial of the crane with state feedback, Equa-
tion (1.5.18), with the fixed physical parameters 9 = 10, me = 1000, l = 10, mL = 1000,
with fixed feedback gains k1 = 600, k2 = 2000, and nominal values k3 = -10 000
and k4 = O. The controller parameters may vary in the intervals k3 E [-20000; 0]
and k 4 E [-10 000; 10000]. We check the stability of the polynomial family. With the
given parameter values, the characteristic polynomial is

This, obviously, is an interval polynomial. Because of the degree four, only p+-(s)
and p++(s) have to be tested .

p+-(s) = 2000(3 + lOs + 13s2 + 15s3 + 5s 4 ) , (8.3.12)

P++(s) = 2000(3 + lOs + 13s2 + 5s 3 + 5s 4 ) . (8.3.13)

346 8 Robustness Analysis by Value Sets

Imp

Figure 8.10. For polynomials of degree five, p+-(s),p++(s) and p-+(s) have to be tested

Both polynomials are stable as can, for instance, be shown by a Hurwitz test. Therefore,
the entire polynomial family is stable. 0

Example 8.12

The characteristic polynomial of the crane with the state feedback is, by (1.5.18),

Suppose all parameters except the load mass mL are fixed, then this is an interval
polynomial with a~ = (mE+mc)g+klf - k3 ) . Only a;, i.e. only mI; enters into p+-(s)
and p++(s). In other words:
for the stabilization of cranes by state feedback, it suffices to stabilize the minimum
load mI;.
o

Note that the Kharitonov theorem applies only to Hurwitz-stability, for r-stability
the value set is no longer a rectangle parallel to the axes. Figure 3.16 illustrates that it
does not suffice to r-stabilize the crane only for the minimum load mI;.
Also note that the Kharitonov theorem is necessary and sufficient only if the coeffi-
cients are truly independent. If they are dependent and overbounded by the maximal
and minimal individual coefficient bounds, then the criterion is only sufficient and may
give very conservative results as in the following examples.

Example 8.13

Analyze robust stability of the polynomial

p(s, ql, q2) = (1 + 2ql + q2) + (1 + ql + 2112)S + 2s2 + S3 ,
with ql E [0; 11, 112 E [0; 11·
8.3 Interval Polynomials, Kharitonov Theorem 347

Overbounding yields ail = 1, at = 4, a 1 = 1, at = 4.

The Kharitonov polynomial

p+- = at + a 18 + ai s + at 8 3
4 + 8 + 28 + 8
2 3
=

is unstable, thus the overbounding polynomial familiy is not robustly stable. However,
the coefficients ao and al do not vary independently. Their domain is a rhombus with
the following four vertices:
1. ao(ql,qi) = 1, al(ql ,qi) = 1,
2. ao(ql,qi) = 2, al(ql,qi) = 3,
3. ao(qt,qi) = 4, al(qt,qi) = 4,
4. ao(qt,qi) = 3, al(qt,qi) = 2.
It is shown in Figure 8.11 together with the stability boundary of ao + al 8 + 28 2 + 8 3,
i.e, 2al > ao . It is seen that the given polynomial family is robustly stable.

4 •............................ ......... ...... 3

3 stable

unstable

1 2 3 4

Figure 8.11. The rhombus (image of Q) is stable, the overbounding rectangle with the ver-
tex p+- is not robustly stable

o
348 8 Robustness Analysis by Value Sets

Example 8.14
Continue Example 8.12 with the fixed parameters mt. = 1, mc = 1000, k1 > 0,
k 2 > 0, k3 = 0, k 4 = O. Then, we know by (1.6.9) that the system is stable for all e> O.
We now try to show stability for e E [9.99; 10.01), by overbounding. Consider the
Kharitonov polynomial
pt+(s) = a;j+ats+a2"s2+ais3+ats4
= k 19 + k29S + (10019 + 9.99k1)S2 + 9.99k2s3 + 10 010s4.
Its coefficients are positive and the critical stability test is ~t+ > O.
~t+ = at a2" ai - a;j(ai)2 - (at? at
= k~ l [9.99 mL - 0.02 mc]
= -1O .01k~ l.
~r is negative, pt+(s) is unstable , the sufficient criterion does not even give a useful
e
answer for a 0.1 % rope length variation around = 10. The reason is that the polyno-
mial pt+ (s) does not describe a controlled crane, because a2" and ai contain t: = 9.99,
whereas at contains e+ = 10.01.
This example may serve as a warning that overbounding without a measure of
conservativeness may lead to a useless result . At least the correct solution should be
estimated in a small gap between overbounding (= sufficient stability criterion) and
underbounding (= necessary stability criterion). In this book, we do not deal with such
estimates, but only with exact stability boundaries . Therefore, we recommend the use
of the Kharitonov theorem only for interval polynomials like in Example 8.11 and 8.12.
o

Remark 8.15
In the first edition of this book, some results on interval plants have been described.
An interval plant has independent coefficient intervals for the numerator and denomi-
nator of its transfer function. These results are not repeated in this second edition for
three reasons:
i. Interval plants have little practical importance. Even in the very simple crane
example (1.3.3) the rope length enters both into numerator and denominator.
ii. It is easy to rewrite the problem in the characteristic polynomial format.
iii. If the open-loop transfer function includes a dynamic compensator, then an in-
terval property of the plant is, in general, destroyed. As an example, consider the
denominator of (1.6.10). The parameter mL now occurs in the coefficients of S2
and s3 by the low-pass filter term (1 + Ts) . A result worth mentioning is that
a first order controller stabilizes an interval system, if and only if it stabilizes 16
extreme points resulting from the Kharitonov conditions applied to the numerator
and denominator polynomials of the interval plant transfer function [41].
o
8.4 Affine Coefficients: Edge Theorem 349

Remark 8.16

The proof of Kharitonov's theorem is easy after the observation that the value
set pOw, A) of an interval polynomial for fixed w is a rectangle with edges parallel to the
real and imaginary axes. A natural question is: are there other classes of polynomials
which have such a rectangular value set? Indeed, it is easy to show that polynomials
with even-odd decoupling
p(s, q, r) = Pellen(s2, ao(q), a2(q), .. . ) + Spodd(S2, al (r), a3(r), . .. )
have the above property.
The uncertain parameters q enter only in the even order coefficients, the uncertain
parameters r enter only in the odd ordered coefficients. Then , as in (8.3.7) and (8.3.8),
the bounds of the real and imaginary parts are mutually independent and the value set
is a rectangle with edges parallel to the real and imaginary axis.

P;lIen (w2, ao(q), a2(q), ) S; Re pOw, q) S; P:lIen(w 2, ao(q), a2(q), ),

W P;;d.,(w 2, al (r) , a3(r) , ) S; 1m pOw , q) S; Wp~(W2 , al (r), a3(r), ).
The determination of P;"en and P~ is easy if the uncertain parameters q and renter
affinely into the coefficients ai' Then, a feasible testing set consists of all combinations
of extremal values of q and r, see Panier et al, [1671 . 0

8.4 Affine Coefficients: Edge Theorem

When Kharitonov discovered his amazingly simple result, it took some time before it
got published and then known in the international literature. The author remembers
that he first heard about it from A. Olbrot in 1981 at a meeting in Bielefeld, and could
not believe it. It became much better understood by the value set interpretation given
in the preceeding section. The Kharitonov theorem and the value set approach triggered
a research direction in the 1980s, see [41] for a historical account. Practially the most
important result of this period is the edge theorem by Bartlett, Hollot and Huang [431 .
Simply speaking, it states that for the stability of a polynomial with linear parameter
dependency, it suffices to check the edges of the Q-box.
Consider the polynomial family
t
P(s, Q) = {p(s , ql, q2 , .. " qt) = Po(s) +L qiPi(S) I qi E [q;;qtJ, i = 1,2, .. . , e}.
i=1
(8.4.1)
The coefficients depend linearly on the uncertain parameter vector q = [ql q2 . .. qtlT .
More precisely, it is an affine dependency because of the additional constant terms
in Po(s), i.e. each coefficient ai has the form
ai = a? + iIiql + hiq2 + ...+ fliqt, i = 0, 1, . . . , n. (8.4.2)
350 8 Robustness Analysis by Value Sets

The parameters qi vary in an i-dimensional box Q:

q, E [qi- .,qi+j, i = 1,2, . . . ,i. (8.4.3)

The polynomial family P(s , Q) = { p(s, q) I q E Q } represents a parallelepiped

(the generalization of the parallelogram to higher dimensional spaces) in the coefficient
space because of the affine mapping (8.4.2). Its vertices and edges are generated by
vertices and edges of the Q-box. If all parameter values qi take their minimum or
maximum value, then the corresponding polynomial is called a vertex polynomial. If
exactly one of these parameters varies between its minimum and maximum value while
the remaining i - I parameters stick either to their minimal or maximal values, then
this polynomial family is called an edge polynomial family.
For notational convenience, the original box is scaled to the unit box with center
zero and sidelength 2 by the transformation
- 2qi - qt - qi
qi = + _ i = 1,2, . .. ,i . (8.4.4)
qi - qi
It maps qi = qi to qi = -1 and qi = qt to qi = + 1. The tilde is now omitted and we
have without loss of generality qi E [-1 ; +1].
Imp

Rep
P2

Figure 8.12. The construction of the value set with three parameters and fixed w

First, the value set POw, Q) with affine coefficient functions will be constructed. In
the case of two parameters, the value set POw,Q) for fixed w is a parallelogram in the
complex plane.
8.4 Affine Coefficients: Edge Theorem 351

-++f---+---+-f--l

Figure 8.13. The edges a - f generate the boundary of the value set of Figure 5.7

Consider now a third parameter, i.e. we have the affine polynomial family

{p(s, qb qz, q3) = Po(s) + qIPI(S) + q2P2(S) + q3P3(S) I qi E [-1 ; +11 , i = 1,2,3 }.
(8.4.5)
For fixed frequency s = jw·, we calculate the four complex numbers Po(jw·), PI(jW·) ,
P2(jw·), P3(jW·) and draw four vectors at the origin that represent these numbers (see
Figure 8.12).
Starting from Po, add the vectors qiPI and qlPI and obtain the points +00
and -00 respectively. The segment [+00; -001 is the value set for the polynomial fam-
ily pes, qlt 0, 0). In +00, we add q;P2 (arriving at + + 0) resp. qiP2 (arriving at + - 0)
and repeat this procedure at -00. We obtain the parallelogram ++0, +-0, --0, -+0,
which corresponds to the polynomial family pes, qt, q2 , 0). The last step is adding qtp3
resp. q:;P3 at all four vertices to get a hexagon +++, +-+ , - -+, - - -, - + -, ++- ,
which is the value set of the polynomial family (8.4.5) for the fixed frequency W·.
Denote the six edges of the value set a to f, see Figure 8.12. They are the image of
six pairwise adjacent edges of the Q-box as indicated in Figure 8.13. The other six edges
and also two vertices are mapped into the interior of the hexagon. Observe that vertices
and edges that contribute to the boundary of the hexagon are not the same with varying
frequency. Changing the frequency leads to different values of Pi(jW), i = 0,1,2,3,
where Po(jw) determines the midpoint of the hexagon and the vectors PI(jW),P2(jW)
and P3 (jw) determine the size and orientation. The transition frequencies, where the
image of an edge of the Q-box changes from the boundary into the interior and vice
versa, may be calculated by comparing the phases of Pt(jw),P2(jW) and P3(jW) . If two
of them are equal, then a transition occurs. At the transition frequency, the value set
P(jw, Q) degenerates from a hexagon to a parallelogram.
Adding a further term q4P4(S) to (8.4.5) requires adding qtp4 resp. qip4, which
yields an octagon. Obviously, the value set of an affine family of polynomials for fixed
frequency is always a convex polygon with, at most, 2£ vertices (this number may
352 8 Robustness Analysis by Value Sets

decrease for special frequencies) . Two opposite vertices are parallel. This special kind
of polygon is called a parpolygon.
With these preparations, we are now ready to formulate and prove the edge theorem.

Theorem 8.17 (edge theorem of Bartlett, Hollot and Huang)

The polynomial family P(s,Q) = {p(s,q) = E7=oai(q)si I q E Q} with affine
coefficient functions a;( q) and Q = { q I qi E [qi ; qtJ, i = 1,2 , . . . , e}is stable if
and only if the edges of Q are stable .
o
Proof
The necessity is obvious. We use the zero exclusion principle for the proof of
sufficiency. With varying frequency, the shape of the parpolygon is changing and
is moving around the origin. If the polynomial family is unstable , then there is
a frequency where the boundary of the parpolygon, i.e. an edge goes through the
origin of the p-plane. Because this edge is the image of an edge of the Q-box, the
instability can be recognized by testing all edges of the Q-box .
o

Remark 8.18

.It is not necessary to test all edges as illustrated by Figure 8.13, however, the amount
of calculation to determine the critical edges for each frequency interval between the
transition frequencies is about the same as to test all edges. A procedure that avoids
the mapping of superfluous edges was given by Fu [92J. It reduces the problem of the
computation of all 2i extremal points to the computation of f points followed by a
sorting problem.
o

Remark 8.19

The edge theorem is valid for a more general class of polynomial families than those
bounded by a Q-box . It was shown in [43J that a "polytope" of polynomials is robustly
stable if and only if all its edges are stable . A polytope of polynomials is the higher
dimensional generalization of a convex polygon in two dimensions. It is the convex hull
of a finite number of points. A polytope of polynomials can be written in the form

Pi = conv {PI(s), P2(S) , . . . ,Pi(S) }

i e
= {p(s) = L AiPi(S) L Ai = 1, Ai 2 0, i = 1,2 , .. . .z}.
i=1 i=1
8.4 Affine Coefficients: Edge Theorem 353

Each edge corresponds to polynomials p(s, q) where exactly one parameter qi varies
between its given limits, while the remaining ( - 1 parameters take their minimum
or maximum value. The Ai are called barycentric coordinates. The name barycen-
ter (= center of gravity) stems from the fact that every point inside the polytope can be
made the center of gravity by distributing a unit mass over the vertices of the polytope .
A barycenter is outside the polytope if and only if at least one of the vertex masses is
negative. 0

The edge theorem reduces the robustness analysis of an affine polynomial family
to a finite number of one-dimensional tests. This is a drastic reduction of the testing
set compared with the complete set . An (-dimensional box has 2l vertices. From each
vertex there start (edges. The number (2 l would count each edge two times, thus, the
total number of edges is (2 l - i . For some values of (, the number of vertices and edges
is given in Table 8.1. Obviously, for a large number of parameters the number of edges
to be tested becomes prohibitively large.
Table 8.1: Number of vertices and edges of an i-dimensional box

( 2l vertices (2l - i edges

1 2 1
2 4 4
3 8 12
4 16 32
5 32 80
10 1024 5120
20 1048576 10485760

Consider an edge with end points Pb and Pc, i.e. the polynomial family
P(s , Q) = { (1 - ql)Pb(S) + qiPc(S) I qi E [0; I]}. (8.4.6)
A stability test for families of polynomials generated by a linear dependence on a single
uncertain parameter was given by Bialas [49] .

Theorem 8.20 (Bialas)

Let H~ and H~ be the Hurwitz matrices of
pb(S) = bo + bis + b2s2 + + bnsn , bn > 0,
Pc(s) = Co + CiS + C2s2 + + ensn , en> 0,
respectively. The polynomial family
P(s, Q) = { (1 - q)Pb(S) + qpc(s) I q E [0; 1] }
is stable, if and only if:
1. Pb( s) is stable.
2. The matrix (H~)-iH~ has no non-positive real eigenvalues.
o
354 8 Robustness Analysis by Value Sets

Proof
Polynomial Pb(S) is in P(s , Q), so obviously, Condition 1 of the present theorem is
necessary. Furthermore, its satisfaction implies the existence of a stable starting
point q = O. Now, assuming that Condition 1 is indeed satisfied, the second
condition is investigated. For an arbitrary polynomial
p(s,q) = (1- q)Pb(S) + qpc(s) ,
in P(s, Q) it can be shown by straightforward algebraic manipulations that the
polynomial's Hurwitz matrix satisfies the following relationship :
Hn(q) = (1- q)H~ + qH~ .
Condition 1 and Theorem 2.1 imply that for q = 0
detHn(O) = detH~ -# O.
This implies that H~ is invertible, so for arbitrary q E (0, 1], it can be written
that

detHn(q) det (-qH~ [C -=-/) I - (H~)-lH~])

= (-qtdetH~det (C-=-/) I - (H~tlH~) .
This implies that, for any q E (0, I], detHn(q) = 0 if and only if

det (C-=-/) I - (H~)-lH~) = O. (8.4.7)

Recall that by definition, sEC is an eigenvalue of a matrix A if and only if

det(sI - A) = O. This makes it clear that (8.4.7) is satisfied if and only if there is
a q E (0 ; I] such that -(1- q)/q is an eigenvalue of (H~)-l H~ . Since -(1- q)/q
is real and ranges from -00 up to 0 as q ranges from 0 to 1, it follows that
(8.4.7) is satisfied if and only if (H~)-l H~ has a non-positive real eigenvalue.
This argument has shown the equivalence of Condition 2 of this theorem with
Condition 2 of Theorem 2.1 (when Condition 1 is assumed to be satisfied). This
equivalence completes the proof.
o
The test of the second condition in Theorem 8.20 can be simplified. By 1.6.4, we
have

where
ao(q) = (1 - q)bo + qeo,
H n- 1 = (1- q)H~_l + qH~_l '
By Condition 1 Pb(S) is stable, thus bo > 0 and eo = Pc(O) > 0 must be required for
ao(q) > O. For the second factor, the proof of Theorem 8.20 may be written with H';
replaced by H n - 1 [14]. Thus, we have
8.4 Affine Coefficients: Edge Theorem 355

Theorem 8.21

pes, Q) = {(I - q)pb(S) + qpc(s) I q E [0; I]} is stable, if and only if:
1. Pb(s) is stable.
2. Pc(O) > O.
3. The matrix (H~_ltlH~_l has no non-positive real eigenvalues.
o

The Bialas test can be executed by a standard eigenvalue program for the calculation
of all eigenvalues of the Bialas matrices (H~_lrlH~_l'

Remark 8.22

If the two vertices Pb and Pc are exchanged, then the eigenvalues are reciprocal. This
fact follows from the relation

(8.4.8)

for square, non-singular matrices A and B. This means that the eigenvalues of
(H~ _l)-l H~_l and (H~_l)-l H~_l have the same angle with respect to the axes
and reciprocal absolute values. This has no influence on the existence of negative real
eigenvalues. 0

Example 8.23

Return to the crane of Example 8.14 with a larger rope length variation i E [5; 15]
and k1 = k 2 = 1, k 3 = k 4 = O. Note that the 0.1 % variation around 10 is embedded in
this interval . The vertex polynomials are

Pa(S) = pes,r) = 10 + lOs + 10 015s2 + 5s3 + 50oos\

Pb(s) = p(s, e+) = 10 + lOs + 10 025s2 + 15s3 + 15 000S4 .

The real root and infinite root conditions are satisfied, the complex root condition
follows from the Hurwitz matrices

5 10 0] , 5 10 0] .
H~ =
[ 5000 10015 10
o 5 10
H'3 =
[ 15000 10025 10
o 5 10

The MATLAB command eig(inv(Hb) *'Hc) yields the eigenvalues Al,2 = 1, A3 =

3, eig(inv(Hc) * Hb) yields Al ,2 = 1, A3 = 1/3. The interval [5;15] of the rope length
is stable. (As we know already, all Z » 0 are stable with any k1 > 0, k 2 > 0).
o
356 8 Robustness Analysis by Value Sets

8.5 Edge Theorem for Gamma-stability

The generalization of the zero exclusion theorem to I'-stability is obvious:

Theorem 8.24
The polynomial family P(s, Q) is robustly f-stable, if and only if:
1. There exists a f-stable polynomial p(s, q) E P(s, Q).
rt
2. 0 P(a(a) + jw(a), Q) for all a E [a- ; a+].
If of is composed of several boundary segments then all of them must be included
in checking the second condition.
o
The proof of the edge theorem in Section 8.4 was based on the fact that the value
set is a parpolygon whose edges originate from edges of the Q-box, and the origin can
enter the value set only through one of the edges of the parpolygon . In (8.4.5), s = jw'
was substituted. All arguments remain valid if s = a(a*) + jw(a*) is substituted. Thus,
the following theorem holds:

Theorem 8.25 (Bartlett, Hollot, Huang)

The polynomial family
n

P(s, Q) = {p(s , q) = I.:ai(q)si I q E Q }

i=O

with affine coefficient functions ai(q) and Q = {q I qi E [qi ; qtJ, i = 1,2 , . . . , e}

is f-stable if and only if the edges of Q are f-stable .
o
It remains to provide a f -stability test for an edge with endpoints qb and qc corre-
sponding to the polynomials pb(S) = p(s, qb) and Pc(s) = p(s, qc). Since a(q) is affine,
the set of polynomials corresponding to the edge of the Q-box is

P(s, Q) = { (1- q)Pb(S) + qpc(s) I q E [0; 1] }.

A generalization of the Bialas test (Theorem 8.20 or 8.21) requires an algebraic
formulation of the stability conditions. Such a generalization will be given for the unit
circle, i.e, Schur stability, in Chapter 11. For other f-stability regions, the algebraic
formulation of stability criteria is more complicated, see (4.4.18). But in principle, a
finite test may be derived.
A practical approach is a root locus plot for
8.5 Edge Theorem for Gamma-stability 357

s 0 + + ... -++ -- ---+ --, -+ - * + -++++++++~:E

++:+:+::
C
++ ++¥ +
-1

++
.t
-2 ...... . . : .... . .. .
. - - . . . . ..-...-.

_3 e::::.- -'- L- ---.J

-1 .5 -1 - 0.5

Figure 8.14. The four edges of Q are r-stable

which may be written as

1 + q Pc(s) - Pb(S) = O.
Pb(S)
The root locus for q E [0; 1] must have r -stable end points and must not intersect any
branch of ar. This is a necessary and sufficient condition for r-stability of the edge
connecting the vertices qb and qc.

Example 8.26

Consider the crane with parameters me = 1000, mL E [1000; 2000], f E [8; 16]
and output feedback u = - [500 2850 - 22800 O]x. It has been designed in
Chapter 3 for r-stability for the four vertices of Q, where ar is the hyperbola w 2 =
4er 2 - 0.25, a S; -0.25. From the parameter space analysis in Figure 3.16 we know
that the entire Q-box is r-stable. This result shall now be verified by checking the four
edges of Q. Figure 8.14 shows the four root loci for the four edges. They stay on the
stable side of ar. Then, by the edge theorem the entire Q is r-stable. 0

The comparison of the parameter-plane in Figure 3.16 and the gridded edge mapping
of Figure 8.14 allows some statements about the relative merits of the two approaches :
1. The parameter plane figure immediately shows the margins in terms of the uncer-
tain parameters ql = mt. and q2 = f. Since the entire r-stable region is shown, it
is easy to t est other operating domains Q for r-stability or to find the "largest"
one.
2. The gridded edge mapping figure immediately shows the margins in terms of the
definition of r -stability. Since all eigenvalues for q E Q are enclosed by the string
of eigenvalues along the four edges, it is easy to test other r -stability definitions
or to find the "t ightest" one.
358 8 Robustness Analysis by Value Sets

3. A third parameter qa entering linearly may be gridded in both approaches. In the

parameter space the discrete family of boundaries for the grid on qa is projected
on the (q1' q2)-plane, where the projection of the Q-box is a rectangle. This idea
may be extended to more uncertain parameters. In the gridded edge mapping
approach, a third parameter requires root loci for the 12 edges of the Q-box. Also,
this idea may be extended to more uncertain parameters with the obvious practical
limit given by the exponentially increasing number of edges, see Table 8.1.
4. For a third parameter qa = me (the crab mass) entering bilinearly, the parameter
space projection approach was applied in Figure 3.21. The edge result does not
apply to bilinear parameters, see Figure 4.6. Here also the interior of the Q-box
must be tested for I'-stability.
5. Two parameters entering polynomially into the coefficients of the characteristic
polynomial, have been analyzed in Figure 4.9 and the projection method for
further gridded parameters may be applied also.

The advantages of the edge result for robustness analysis are obvious for Hurwitz-
stability by the Bialas edge test and a larger number e of uncertain plant parameters
entering linearly. The test requires one eigenvalue calculation per edge. In other situa-
tions, the parameter space approach is advantageous for robustness analysis. Also, the
parameter space approach is superior for systematic design steps by the intersection of
admissible sets of controller parameters.

Remark 8.27

The derivation of the zero exclusion result from Mikhailov's theorem suggests the
construction of value sets in a complex plane

Rep[a(a) + jw(a)] + j Imp[a(a) + jw(a)] .

If the Ivstability boundary is represented by the recursion formula (4.2.6), then its
relationship with the real and imaginary parts of (4.2.21) may be used:

Rep[a(a) + jw(a)]] =
[ Imp[a(a) + jw(a)]
[10
-a(a)] [do(a) d1(a)
w(a) 0 do(a)
dn(a) ]
dn- 1(a) a.

The RRB, w(a) = 0, is treated separately, then for the eRB, w(a) =1= 0, the following
two equations must be analyzed:

ft(a)] = [do(a) d1(a)

[ h(a) 0 do(a)
dn(a) ] a
dn- 1(a)
= [10
~]
_1_
[
Rep[a(a) + jw(a)] ] .
Imp[a(a) + jw(a)]
w(cr)

The zero exclusion condition for p[a(a + jw(a)] directly translates to zero exclusion
conditions in the (ft , h)-plane.
o
8.5 Edge Theorem for Gamma-stability 359

The difficulty of a robustness test increases enormously from the interval or affine
case to multilinear or polynomial coefficient functions. Therefore, it is important to
bring a given polynomial first to its simplest form. We have done this already by
converting the rational coefficient function in p(s) of (1.5.17) to the coefficient function
in p(s) = tmcp(s) of (1.5.18).
Sometimes, simple polynomials look complicated but can be factorized into
p(s, q) = f(q)p(s , q) , f(q) f; 0 for all q E Q. (8.5.1)
Then, p(s, q) and p(s, q) have the same roots but p(s, q) may be simpler.

Example 8.28
p(s, q) = ql +q~S+qlq2S2+qlS3 can be converted into an interval polynomial because
p(s, q) = p(s, q) = (1 + qlS + q2 s2 + S3).
ql
o

Example 8.29

p(s, q) = (1 + ql)
q2
+ 1+ ql + (5 + 2+ ql) S2 + S3
lJ2
S
q2
can be converted into an affine polynomial
p(s , q) = q2P(S, q) = (ql + q2) + (1 + qds + (2 + ql + 5q2)s2 + q2s3.
o

Example 8.30
Consider the polynomial
p(s, q) = (1 + qi) + (q2eq~ + q~)s + s2, qi E [1; 2], i = 1,2 ,3 (8.5.2)
with three uncertain parameters
(8.5.3)
At first glance, this polynomial looks complicated for a robustness analysis. But note
that each independent variable qi appears only in a single coefficient function:
~(q) = ~(ql) = 1 + q~,
al(q) = al(q2,q3) = q2eq~+q~, (8.5.4)
a2(q) = a2 = 1.
The coefficient functions are independent of each other and generate a rectangle in
the (ql' q2)-plane. Therefore, the Kharitonov theorem may be applied.
o
360 8 Robustness Analysis by Value Sets

8.6 Singularity of Value Sets

In Section 2.7, we have investigated Hurwitz-stability of a polynomial family

P(s, q) = A(s)Q(s) + B(s), (8.6.1)

with given A(s) , B(s) and the uncertain polynomial

Q(s) = qo + qis + ...+ qm sm. (8.6.2)

The occurrence of singular frequencies for fixed K p will now be explained in terms of
value sets.

I*Q

Figure 8.15 . Value set of Q(jw)

For s = jw, QOw) is split into its real and imaginary parts

Q(jw) Ro + jIQ , (8.6.3)

Ro qo - q2w2 + q4 w4 - + ... ,
IQ W(ql - q3w2 + qsw4 - + ... ),

with the extremal values

R- = qo - qtw2 + qiw4 - + , (8.6.4)
R~ = qt - qiw2 + qtw4 - + .
Suppose I Q = I Qis fixed, in the PID case this corresponds to K p fixed. The correspond-
ing value set is constructed now for fixed w. Figure 8.15 shows the value set of Qdw),
it is a straight line segment parallel to the real axis. Multiplication with the complex
number A(jw) = IAlejo: rotates the line segment by an angle Q' and stretches or shrinks
it by a factor IAI. Finally, it is added to BOw) = IBlej13 • For the frequency w illustrated
in Figure 8.16, the line segment may be extended arbitrarily far without ever hitting the
origin. A stability boundary occurs only at the singular frequencies, when the straight
line through the segment passes through the origin. Then, 1m POw) = eRe POw) .
This is the singular case of the two equations Re POw) = 0 and 1m P(jw) = 0 that was
analyzed in Sections 2.6 and 2.7.
8.6 Singularity of Value Sets 361

........
....
.... )
I
I
ImP I
I

I
I

o &P
Figure 8.16. Value set of POw, q) = AOw) QOw) + BOw)
9 Value Sets for Non-linear Coefficient
Functions

Results for robustness analysis of polynomials with interval or affine coefficient func-
tions have been shown in Chapter 8. Simple testing sets (vertices, edges) have been
found whose stability guarantees stability of the entire Q-box . It is tempting to
look for further results (faces?, diagonals?) for multilinear coefficient functions (terms
like qlq2 , q2q3, qlq2q3) and polynomial coefficient functions (terms like q?q2 and no possi-
bility to introduce q3 = q? as an independent parameter). Such hopes for general classes
like multilinear or polynomial coefficient functions have been buried by the following
example. It shows that testing sets in general cannot be derived from the Q-box alone,
they also depend on the specific polynomial and its Jacobian. Special uncertainty struc-
tures are exploited for robust stability analysis, e.g. tree-structures in the polynomial,
e.g. p(s, ql , q2 , q3) = Pl(S, ql)P2(S, q2) + P3(S , q3) .

9.1 A Warning Example

Example 9.1 (see [21J)

Let

Note that this is a generalization of Example 4.17, there r = 0.2 was chosen. Let ql > 0,
q2 > 0, then ao > 0, al > 0, and a2 > 0, and the stability condition is

The system is stable outside a circle with center ql = 1, qz = 1 and radius r , see
Figure 4.5. Now, let r go to zero. There remains an isolated unstable point (IUP)
at ql = 1, q2 = 1. We can choose an arbitrary Q-box that contains the IUP and

°
obviously stability is not related to other points in Q than just the IUP. In this special
case, r = instead of r = 0.2 of Figure 4.6, the root set just touches the imaginary axis
9.1 A Warning Example 363

in a single point w = 2. The radius r does not enter into the Jacobian of Example 4.19,
thus it is unchanged (ql = q2) as shown in the right part of Figure 4.7. For the
construction of the value set, the segment inside Q of the Jacobi line (det J = 0) line
must be considered. The line obviously only depends on the polynomial; only the
endpoints of the line segment depend on the box Q. Figure 9.1 shows the value set at
the frequency w = 2 for th is example. At this frequency, the image of the Jacobi line
passes through the origin. At all other frequencies, zero is excluded from the value set.

Imp(jw)

Rep(jw)

Figure 9.1. Value set of the warning example at w = 2 (not to scale)

Remark 9.2

In order to rule out further conjectures, e. g. about faces of Q, the example was
generalized to a parameter space of arbitrary dimension f [6]. Let
(9.1.2)
with the bilinear coefficient functions
i i-I t
2+2(f+1)Lq;+2L
ao(q) = f(f-1)+r L q;qj,
;=1 ;=1 j=i+l

i
al(q) f+Lq; ,
;=1
364 9 Value Sets for Non-linear Coefficient Functions

Let q E Q+ = {q I qi > 0, i = 1,2, . . . , e}, then all coefficients ai(q) are positive and
the only remaining Hurwitz-stability conditions is a1(q)a2(q) - ao(q) > O. It is easily
verified that
l
a1(q)a2(q) - ao(q) = L (qi - 1)2 - r 2.
i=1
The resulting stability condition
l
L (qi - 1)2 > r 2 (9.1.3)
i=1

is satisfied outside a ball of radius r and center qO = [1 1 . .. If

Next, we let r go to zero. Thus, the unstable ball shrinks to an isolated unstable
point qO. With the exception of this point , the polynomial is stable for all q E Q+. This
example has been used as a benchmark test in many publications. It also illustrates
the risk in gridding methods without graphical warning on closeness to instability. D

The values set of Figure 9.1 may be generated by a chord approximation. For r = 0
and s = j2, the value of the polynomial is

For fixed q2, the expression is linear in qI, therefore the dotted lines in the Q-box of
Figure 9.2 map into straight lines segments in the value set . For example, the center

Imp(jw)

C'
D....- ---,C
B'

Rep(jw)

A B

Figure 9.2. Chord approximation of det J = 0 (not to scale).

point of A'D' is connected by a line segment with the center point of B'C'. The chord
approximation may be refined by a further bisection
In general, we can draw the following conclusions for bilinear coefficient functions :
9.1 A Warning Example 365

a) If the image A' B'G'D' of the rectangle ABGD is a convex quadrangle, then this
is the value set .
b) If A'B' G'D' is not convex, then the value set is contained in the convex hull, in
Figure 9.2 this is the triangle B'G'D'. This result will be generalized in Section 9.2.

In the previous example with only two uncertain parameters, there is no advantage
of the value set approach over the parameter space approach . For more parameters,
however, the value set and its union for all frequencies can be easily inspected for
stability. The complexity of the problem enters into the boundary of the value set in
the complex pOw)-plane, as the following example illustrates.

Example 9.3

Construct the value set of the polynomial

3
p(8, q) = II(8 + qi) (9.1.4)
i=1

for 8 = O.5j and qi E [-J3 ; J3]. The value sets of the three subpolynomials 8 + qi are
straight lines parallel to the real axis in the complex plane, see Figure 9.3. Thus , for

r r4
m m
O.5j I~j O.5j
:
-J3
Ik
+J3
x :
-J3
4
+J3
e
x :
-J3
e
+J3
Figure 9.3. Value sets of the subpolynomials

generating the set P, three line segments have to be multiplied with each other.
First, the four extremal points are determined:

A = (J3 + O.5j)3 £J3 + ¥j ~ 3.9 + 4.4j,

B = (J3 + O.5j)2(-J3 + O.5j) _llJ3 _llj ~ -5.6 - 1.6j,
4 8

G (J3 + O.5j)(-J3 + O.5j)2 ¥J3 - ¥j ~ 5.6 - 1.6j,

D = (-J3 + O.5j? -~J3
4
+ ~j
8
~ -3.9 + 4.4j.

They have all the same absolute value IAI = ¥J3 ~ 5.9 and the phase angles are

'PA 3 · tan"! O.5/V3 3.16.1° = 48.3°,

'PB 2 .16.1° + 163.9° 196.1°,
'Pc ·16.1° + 2 .163.9° 343.9°,
'PD 3 .163.9° 491.7°.
366 9 Value Sets for Non-linear Coefficient Functions

On the edge AB, two factors are fixed at J3 + O.5j, the remaining one moves
from J3 + O.5j to -J3 + O.5j. On the edge BC, one factor is fixed at J3 + O.5j and
another one at -J3 + O.5j. By continued bisection of AB and BC and connecting the
corresponding points for fixed aI, the image of the polynomial (J3+0.5j)(a1 +O.5j)(a2+
O.5j) for 0'2 E [-J3 ; J3) is generated. These images are straight lines because the
polynomial is linear in 0'2. By continuing this construction for the edges BC and DC,
Figure 9.4 is generated. Alternatively, the Jacobian condition

1m

, -,
/
/ -,
/
/
-,
D' " ' , -,A
,,
/
/ -,
/

/
/
,
/ \
/ \
I \
I \
I \
I \
I \
I \

,,
I \
I
I Re
I ,
1 I

B \l,L.:::::=2:~~3~~E~~~~::S=:::::~,/ c
\ I
\ I
\ I
\ I
\ I
\
,, /
I

, ,, /
/
/

, /
/

"'" ////

"
- --
3
Figure 9.4. Value set of the polynomial IT (8 + qi) for 8 = O.5j and qi E [-J3 ; J31
i=l

can be used. We have rank J < 2 for q1 = q2 = q3, i.e, a part of the boundary is
generated by (a + O.5j)3, a E [-J3 ; J3). Also, this Jacobian is plotted in Figure 9.4.
In Figure 9.5, the value set is plotted without its internal structure. The value set is
contained in the convex hull, i.e, the rectangle ADBC. Zero is not contained in the
value set. 0
9.2 Desoer Mapping Theorem 367

1m

Re

B c

Figure 9.5. Product of three straight line segments

9.2 Desoer Mapping Theorem

For multilinear polynomial families, the mapping theorem of Desoer [202] yields a useful
sufficient stability condition that can restrict frequency gridding to a subset of the non-
negative frequencies. The robust stability test is by zero exclusion from the value set,
see Theorem 8.6. The mapping theorem gives a simple description of the convex hull
of the value set.

Theorem 9.4 (mapping theorem of Desoer)

The convex hull of the value set POw', Q) of a polynomial with multilinear coef-
ficient functions is the convex hull of the images of the vertices of Q.
o

Proof

Without loss of generality, let qi E [0 ; 1]' i = 1,2 , . . . , f .We will demonstrate the
proof of the mapping theorem for three parameters ql, qz, q3 . The generalization
to f parameters is obvious (induction over f) . In four steps , we show that the
image of an interior point p of the Q-box is contained in the convex hull of th e
images of the eight vertices. Let conv{ PllP2,. . . ,Pm} denote the convex hull of m
points that is the set
368 9 Value Sets for Non-linear Coefficient Functions

q++-

q++

q+++

q q+
q I
Jir----
/q---
/
/

/ q
/
/
/
q--+.*--------
q+-+

Figure 9.6. Notations used in the proof of Theorem 9.4

m m

{plp=LAiP(i), LAi=l , Ai~O, i=1 ,2 , ... ,m} . (9.2.1)

i=1 i=1

The polynomials p(i) = p(q(i») are generated by the corresponding vectors q(i) as
indicated in Figure 9.6. For fixed q2 = q2 and q3 = qi , and ql E [0; 1], we have

(9.2.2)

where p: = p(jw',0 ,q2,qi) , p" = p(jw', 1,q2,qi). Using the notation of (9.2.1),
(9.2.2) may be written as
(9.2.3)
Analogously, we have
p- c conv {p-- ,p-+} , p+ C conv {p+- ,pH }. (9.2.4)

The third step is combining (9.2.3) and (9.2.4):

(9.2.5)

and the last step gives

(9.2.6)

o
9.2 Desoer Mapping Theorem 369

Example 9.5

The crane as given in (1.5.15) with e E [8; 16] [m], me E [100; 2000) [kg], mL =
2000 [kg] and 9 = 10 [ms- 2] with the control law

u = - [500 100 -100 0] x

has to be checked for Hurwitz-stability. Its characteristic polynomial is

p(s , m e , e) = 5000 + 1000s + (20 100 + 500e + lOme)s2 + 100es3 + emes4.

We can verify stability easily for a nominal point in the given domain of interest,
for instance e = 8 [m] and m e = 1000 [kg] . Then, the frequency is gridded and for
each grid point s = jw' , the convex hull of the value set is constructed. The result is
shown in Figure 9.7. For higher frequencies (which are not shown for reasons of scale of
the figure), th e quadrangles move away from the origin and into the fourth quadrant.
Th e frequency grid for this fourth degree polynomial can be terminated as soon as the
entire convex hull lies in the fourth quadrant. Obviously, the origin is excluded from

500 1m p

-=.Re p
5000

Figure 9.7. Convex hulls for th e given example with w' = k · 0.05, k = 0, 1, .. . , 20

t he convex hull of th e value sets at all frequencies. Thu s, th e origin is also excluded
from the value sets at all frequencies and the system is stable. 0

If th e origin is included in th e convex hull of the value sets for a frequency band
wEn, then th e mapping theorem does not give a conclusive answer about the stability
of t he polynomial family. A necessary and sufficient stability condition is then obtain ed
by the actu al construction of the value set for a grid on wEn and checking for
zero exclusion from the value set. For frequencies w (j. n, the value set need not be
constructe d in view of the sufficient condition obt ained from the convex hull of the
value set .

Remark 9.6

An interesting use of th e mapping th eorem is in domain splitting algorithms

[73],[180],[181]. The Q-box is split into two parts. The value set is contained in the
370 9 Value Sets for Non-linear Coefficient Functions

union of the convex 'hulls of the two parts. further bisection improves this outer ap-
proximation of the value set . In addition to the outer bound of the value set (sufficient
condition) the algorithms also generate an inner bound (necessary condition) from the
fact that images of the additional edges must belong to the value set. The algorithm
stops when either a sufficient condition for zero exclusion is satisfied or a necessary
condition for zero exclusion is violated. 0

For the case of multilinear coefficient functions, the sufficient condition of the convex
hull and construction of the actual value set by families of straight lines inside this hull
have been shown in this section . Note that the convex hull condition does not hold for
polynomial coefficient functions.

Example 9.7

Consider the polynomial family

P(S,ql)=S+q?-6ql+8, qlE[1 ;5).

Both p(s, 1) = s + 3 and p(s, 5) = s + 3 are stable, but p(s, 3) = s - 1 is unstable. The
value set p(jw, ql) E Uw - 1 i jw + 3) can be calculated from the extremal eigenvalue
locations. For w = 0, the value set contains the origin. The convex hull of the value set
is the set itself, it is not generated by the extremal values of the parameter ql.
This example suggests that further results may be obtained if the monotonous parts
of the non-linearity are considered separately, i. e. the intervals qi E [1 ; 3) and ql E
[3 ; 51. 0

For polynomial coefficient functions there remains the brute force gridding of Q for
generating the value set. This is computationally prohibitive even for a medium number
of parameters. For two parameters the parameter space method is clearly superior, see
Example 4.17.

9.3 Tree-structured Value Set Construction

In the previous chapter, the stability test by zero exclusion from the value set at all
frequencies was used as a concept for proofs of Kharitonov's theorem and the edge
theorem. For non-linear parameter dependence, there are no such simple results; it
may be possible, however, to construct the value set and to use it for the stability test
by zero exclusion. We will see that the construction of value sets can be performed
extremely fast if the system has a so-called tree-structure.
In Theorem 8.6, it was shown that an uncertain system with characteristic polyno-
mial p(s,q) is robustly stable if and only if:
• There exists a qo E Q for which p(s, qo) is stable.
• The value sets P(jw, Q) do not contain'the origin for all frequencies w E [0 ; 00).
Tree-structured Value Set Construction 371

One advantage of this approach is that even a high dimensional operating domain Q is
always mapped into the two-dimensional complex plane C. Therefore, the construction
of value sets is well suited to a graphical display. The sets can be visualized for various
frequencies on a computer display and the user can invest igate the stability of the
uncertain polynomial by visual inspection of the sets. If the construction of the value
sets is fast enough, then it is even possible to create a computer animation where the
value sets are displayed on the screen with increasing frequency.
If we are interested in generating the exact value set for non-linear parameter de-
pendency, dense gridding of the uncertainty domain will often be the only possibility.
In certain cases, however, the construction of the value sets can be simplified drastically
if the characteristic equation has special structural properties.
Models of mechanical and electrical systems frequently have a tree-structure, for
example a chain of mass-spring-damper (MSD) elements or a corresponding electric
RLC network or a robot in free motion(Le. without hand contact with the environment) .
Such physical tree-structures are preserved under state and output feedback. It is
therefore worthwhile, to exploit tree-structures in the construction of value sets.

Example 9.8

Consider the crane with outputs crab position Xl and rope angle X3. The respective
transfer function are, by (1.3.5) and (1.3.6),
Xl(S) s2f+g
u(s) s2[mefs2 + (mL + me)g]'
X3(~ -1
u(s) mefs2+(mL+me)g '
These transfer functions, and with feedback the resulting closed-loop characteristic
polynomials, are not treated as funct ions of s as in numerical analysis for given pa-
rameter values. In the construction of value sets for fixed w (or a on ar), s is just a
fixed complex number and we are only interested in how the value sets depend on the
uncertain parameters. Assume that f , mt. and me are uncertain. Now, consider the
(k
feedback system of Figure 1.8 with w = 0, i.e, u = - l + 1 ~2TJXl - k3X3, T fixed.
Its characteristic polynomial is (see also (1.6.10))

p(s, f, mL, me, kl, k2, k3 ) = (s2f + g)[(s2me + kl)(1 + Ts) + k2s] + s2(mL9 - k3 )(1 + Ts) .
(9.3.1)
All six parameters appear only once in (9.3.1). For construction of the value set
pOw, i, mL, me, kl, ka, k3 ) with given lower and upper bounds for the parameters, the
value sets of (s2i + g), (s2 me + kd, (mL9 - k3 ) and k2 are constructed independently
and combined to the value set p(s, f, mL, me, kl , k2, k3 ) , where s is a fixed complex
number on the boundary ar. Figure 9.8 illustrates the sequential operations on value
sets of the subpolynomials until p(s, q) is reached. The construction is started with
the term kl + mes2. Its value set for fixed s = jw is kl - mew2, i.e. it is the
real interval [k1 - m"t:w2 ; kt - mcw2] . The next step is to generate the set for
(1 + jwT)(k l - mew2) + kzjw. The multiplication by the complex number (1 + jwT)
372 9 Value Sets for Non-linear Coefficient Functions

p(s, q)
.---------<+t---------,

,-------{x:}--------,

(1 + Ts)(k l + mcs2) + k2s

,----{+)------,

k2 s
Figure 9.8. Tree-structure of the characteristic polynomial (9.3.1)

means a rotation and scaling corresponding to the phase and magnitude of (1 + jwT) .
The line segment PIP2 in Figure 9.9 is the result: The set of k~w has to be added. The
result is the parallelogram PIP2PaP4 in Figure 9.9. Its vertices are

PI = (1 + jwT)(k l - m6 w2) + kijw, P2 = (1 + jwT)(kt - mcw2) + kijw ,

Pa = (1 + jwT)(kt - mcw2) + ktjw , P4 = (1 + jwT)(k - m6 w2) + ktjw .
1
Similarly, the set of s2f + 9 and (1 + jwT)s2(mL9 - k a) can be constructed easily. The

1m Pa

PI Re

Figure 9.9. Value set of (1 + jwT)(k l + mcs2) + k2S for a fixed frequency s = jw
last construction step is then the multiplication of the sets s2e + 9 and (1 + jwT) (kl +
mcs2) + k2s. The tools for the execution of this set multiplication will be introduced
in the following sections. The main point is that the value set can be constructed
sequentially. Much more effort would have been spent if the uncertain parameters in a
six-dimensional Q-box had been gridded. 0

'free-structures can be found in several control system configurations. The question

is how to recognize these structures. The conventional way of modelling hides tree-
structures by smearing uncertain parameters over the coefficients of the characteristic
Tree-structured Value Set Construction 373

polynomial. A first advice is not to unnecessarily manipulate the system equations. For
example, in mechanical systems the differential equations are represented in the form

lv.1(q)X+IJ(q)X+J((q)x=u. (9.3.2)
In the usual way to arrive at a state space representation, this system of equations is
premultiplied by the inverse of the mass matrix lv.1(q) . This distributes the elements in
this matrix allover the system of equations . A better approach is to directly determine
the Laplace transform of (9.3.2). The characteristic polynomial is then

pes, q) = det [lv.1(q)s2 + IJ(q)s + J((q)] (9.3.3)

and the value set for s = jw is
p(jw, q) = det [J((q) - w2lv.1(q) + jwIJ(q)] . (9.3.4)
This determinant can be evaluated step by step, and possible tree-structures can be
recognized more easily.
Even if the system does have a mechanical loop, i.e. no tree structure, it is possible
to gain large computational advantages if only one of the connections is gridded rather
than all parameters.

Example 9.9

In Figure 9.10, a schematic representation of a mechanical system taken from [29]

is given. All elements are assumed to be uncertain. Its differential equations are

~/ 1 - - - - - - - i

X2
Figure 9.10. Schematic representation of a mechanical system

ml xl + dlxl + CIXI + CI2(XI - X2) 0,

m2x2 + d2X2 + C2X2 + C12(X2 - xd u,
and their Laplace transform is
ml s2 + dIs + CI + Cl 2 - C12 ] [ XI(S) ]
[ - C12 m2s2 + d2s + C2 + CI2 X2(S)
374 9 Value Sets for Non-linear Coefficient Functions

The characteristic polynomial is

(9.3.5)

with
2 .
Pi(S, mi, di , Gi, CI2 ) = m iS + diS + Gi + C12, i = 1,2 . (9.3.6)

The system does not have a tree-structure because the uncertain parameter CI2 ap-
pears both in PI ( S , q) and P2 (S, q). If, however, this single parameter is assumed to be
constant, then the system has a tree-structure in the remaining six parameters. The
parameter C12 has to be gridded and for each grid point the stability analysis can be
done very fast by exploiting the tree-structure. This is much better than gridding all
seven uncertain parameters. For a given grid point C12 = ch and s = jw', the sets of
Pi(S, mi, di , Gi, CI2), i = 1,2 are rectangles, see Figure 9.11. These two rectangles have

1m 1m

Re x Re

Figure 9.11. Computation of the value set by multiplication of two rectangles in the complex
plane

to be multiplied and c'f2 is subtracted. The value set of the characteristic polynomial
(9.3.4) for mi E [1 ; 3], dl E [0.5 ; 2], CI E [1 ; 2], m2 E [2 ; 51 , d2 E [0.5 ; 2],
C2 E [2 ; 4], Ci2 = 1, and w' = 1 is to be constructed.
For the multiplication of two rectangles in a complex plane, note that the product
of a test point on the boundary of the first rectangle (i.e. a complex number) with the
second rectangle is a rotated and scaled rectangle. As the testpoint wanders around the
first rectangle, many such rectangles are generated and the product set is the union of
rectangles as shown in Figure 9.12. The actual value set is bounded by the contour of
Figure 9.12, it is shown in Figure 9.13. This contour may be generated by a standard
contour algorithm [169] . This reduction step is particularly important if further value
sets of other subpolynomials are added or multiplied. Practically, between any two
additions or multiplications , a reduction to the contour is performed.
For a stability analysis, the value sets for all frequencies 0 ~ w < W max must be
constructed and checked for zero exclusion. For the example, W max is the frequency
where the set is for the first time entirely contained in the fourth quadrant. Figure 9.14
shows the union of value sets POD, Q). Zero is not included in the union, thus the
system is robustly stable . 0
Tree-structured Value Set Construction 375

1m

Figure 9.12. Value set for the mechanical system of Figure 9.10 at w' = 1

Remark 9.10

A stability test of the above mechanical system is trivial. It is a passive system

and therefore stable, independent of the values for masses, springs, and dampers. If,
however, one would like to check if the system guarantees a certain stability margin, for
example , a negative real part 0" < 0"0 for all q E Q, then the stability test is no longer
simple, but the same tree-structure can be exploited.
Note also that the origin is not contained in the value set of Figure 9.13, but it is
contained in its convex hull. Therefore the sufficient condition of Theorem 9.4 does not
give a conclusive answer. 0

Example 9.11 (Example 9.9 continued)

An alternative method is the graphical visualization of sets for increasing frequency.

The engineer can watch a "movie" of the animated set on the computer display and
thereby perform the zero exclusion test by visual inspection of the set . Some scenes of
the animated sequence of the value set of the mechanical system given in Example 9.9
are shown in Figure 9.15. The animated value set not only gives a yes/no answer about
robust stability, but it also shows critical frequency ranges, for which the value set
comes close to the origin , and where better compensation is required . In Figure 9.15,
we recognize that for this example the frequency range w E [0.75 ; 1.5] seems to
be especially critical. In case of a controlled system , controller par ameters can be
modified during the animation with appropriate input devices (e.g. dial box, joystick ,
steering ball) and the engineer can immediately recognize the consequences of this
modification. With this interactive search in the space of controller parameters, a
robustly r-stabilizing controller can be quickly found [24] .
o
376 9 Value Sets for Non-linear Coefficient Functions

1m

Re

Figure 9.13. Contour of the value set of Figure 9.12

1m 15
lO

5
0
-5
-lO
1'un,Q)
-15
-20

-15 -10 -5 o 5 lO 15 20
Re

Figure 9.14. Union of the value sets of the mechanical system

Tree-structured Value Set Constru ction 377

w = 0.125 w = 0.250 w = 0.375

.....--. c::J c::J

w = 0.500 w = 0.625 w = 0.750

b <::J [ b

.:
w = 0.875 w = 1.000 w = 1.125
( ) r<.D )
'-"

<. )
w = 1.250 w = 1.375 w = 1.500

r-.
D ..
/':~ r-:

~
w = 1.625 w = 1.750
'------
/
w = 1.875
r---

r
- r-.,
I »>
r--- r----
I
Figure 9.15. Scenes from a value set animation
378 9 Value Sets for Non-linear Coefficient Functions

9.4 Computer-aided Execution of Value Set

Operations

For the set operations, we assume all value sets to be closed and bounded . The complex
set operations we are interested in are addition and multiplication. They are defined as

e = A + 8 = {a + b I a E A, b E 8} , (9.4.1)

e = A · 8 = {a · b I a E A, s e 8}. (9.4.2)
Every point of set A is added to (or multiplied by) each point of set 8 .
For an interior point point a E A, there exists an open neighborhood N(a) in the
set A. If this open set N(a) is multiplied by any point b E 8, this will again result
in an open set. Therefore, boundary points in the set e = A . 8, for which no open
neighborhood in the set e exists, can only stem from points a E A and b E 8, which
also do not have an open neighborhood in the sets A and 8. These are boundary points,
a E 8A and b E 88, thus,
8C c 8A ·8B. (9.4.3)
Similarly, if e = A + 8, then
ec c 8A+88. (9.4.4)
e
The main use for constructing value sets is to check the resulting set for·exclusion of
the origin. For this purpose, it suffices to construct the boundary of the set , for which
(9.4.3) and (9.4.4) are exploited.
A first problem is a suitable way of representing the set boundaries in the computer .
Here, we could think of an analytic computation of the boundaries . Then, the bound-
ary is represented by piecewise continuous and differentiable curves. If two sets are
multiplied, not only will the order of these curves be increased but also the number of
curve segments forming the boundary. For further operations, it will become more and
more complicated to determine the correct representation of the resulting boundaries .
The complexity of the problem enters into the fine detail of the contour of the sets.
For our purpose, it is more practical to approximate the boundaries by chords.
Before an opesation A+8 or A8, both boundaries of A and 8 have to be approximated
by polygons. Both these polygons consist of line segments like the ones shown in
Figure 9.16. For addition and multiplication, each line segment representing a part
of the boundary of A is added to (or multiplied by) each line segment representing
the boundary of 8 . The result is the union of the sets obtained by these elementary
operations, which are shown next.
Addition
The result of the addition of two line segments is a parallelogram , like that shown in
Figure 9.17. The vertices are C;j = G.i + bj , i,j = 1,2 .
Multiplication
Two line segments have to be multiplied with each other. The line segments are de-
scribed by al + Q(a2 - ad, Q E [0; 1] and bl + (3(~ - bl ) , (3 E [0; 1]. The product is
9.4 Computer-aided Execution of Value Set Operations 379

b'\
1m az 1m

/
al Re
x
or
+
bz Re

Figure 9.16. Elementary set operations

Re
Figure 9.17. Sum of two line segments

then
C={C(O,,B)/OE[O ; 1]' ,BE[O ; I]}, (9.4.5)
with
(9.4.6)
We can consider this as an uncertain multilinear term in 0 and ,B. Thus , its value set
can be constructed easily using the Jacobian (see Chapter 4). Besides the edges of the
uncertainty domain, points at which th e Jacobian determinant

oRe c(0, ,8) oRe c(o, ,B)

00 0,B
J(O,,8) = (9.4.7)
81m c(0 , ,B) 81m c(o , ,B)
00 0,B
vanishes, also cont ribute to t he boundary of t he resulting set . Th e resulting curve is a
parabola, which again is approximated by a polygon. If the Jacobian does not vanish
for 0 E [0 ; 1) and,B E [0 ; 11, then the resulting value set is a convex quadrangle. The
vertices of this convex set are determined by th e four points Cij = a i' bj , i, j = 1,2. Two
possible forms of the result are shown in Figure 9.18. It is also possible to determine
the product of two line segments without computing th e Jacobian determinant. First,
380 9 Value Sets for Non-linear Coefficient Functions

a) 1m b)
C22 Im

Figure 9.18. Products of two line segments: a) the Jacobian determinant vanishes for Q E'
[0; 1) and f3 E [0; 1], b) the Jacobian determinant does not vanish, i.e. the set is convex

the four points Cij are computed and connected in the sequence CU-CI2-C22-C2I-CU ' If the
resulting set is convex, then the final result is already obtained. If the resulting set is
non-convex, then one of the line segments has to be gridded and the other line segment
is multiplied with these grid points. The result is shown in Figure 9.19. The parabola
generated by the Jacobian is approximated by a chain of chords, which are well suited
to next-set operation in the tree-structure. We have seen that tree-structures of the

1m

Figure 9.19. Construction of the product of two line segments by approximation

characteristic polynomial arise naturally from modelling of physical systems, which have
a tree-structure with respect to the uncertain parameters. If necessary, in a mechanical
or electrical system a loop may be cut by gridding the parameters of a connection that
does not satisfy the structural assumption . In the example of Figure 9.10 this was the
coupling parameter CI2.
On the other hand , a parametric polynomial may be given for which the existence
of a tree-structure is not obvious. In this case we try to decompose it , i.e. to find its
tree-structured decomposition (TSD).
First, a partition of the uncertainties qi , qz, .. . , qt in q is defined. q is decomposed
into subvectors q(l) and q(2) with no common elements, i.e. each uncertainty qi belongs
9.4 Computer-aided Execution of Value Set Operations 381

either to q(l) or to q(2) .

We call a polynomial p(s , q) sum decomposable, if there exists a partition q(l) , q(2)
such that
p(s, q) = PI (s, q(l») + P2(S, q(2»). (9.4 .8)
A polynomial p(s, q) is product decomposable, if there exists a partition q(l), q(2) such
that
p(s , q) = Pl(S, q(l)) . P2(S, q(2») + Po(s). (9.4 .9)
The polynomials Pl(S, q(l») and P2(S, q(2») are called subpolynomials of p(s, q). Further
subpolynomials may arise if PI(S, q(1») or P2(s, q(2») are sum or product decomposable.
A subpolynomial is called undecomposable if neither a sum nor a product decomposition
is possible.

Example 9.12

p(s, q) = (qIq2 + qa + q5) + (ql - 4q2 + qa + 3)s (9.4.10)

In the first step, a sum decomposition can be performed, that separates the parameters
ql and q2 from qa . It is

with

Pl(S,Ql ,q2) qIq2 + (ql - 4q2)S ,

P2(S ,qa) qa + q5 + qa s + 3s.
Note that the sum decomposition is not unique with respect to the constant term 3s.
Any polynomial 13(s) could be added to PI(S,qI, q2) and subtracted from P2(S, qa). The
polynomial PI (s, ql, q2) is further product decomposable by adding the term (-4s 2) and
incl uding (+4s 2) to Po( s), such that

with

Pll(S, ql)
pdS,q2)
Po(s)

This example demonstrates that the degree of the subpolynomials in s may be higher
than the degree of the polynomial itself. For the construction of value sets, larger
powers of s are not difficult to handle, since they are merely complex numbers for a
fixed frequency s = jw·. Also , a non-linear dependency on a single parameter qa in the
subpolynomial P2(S , qa) is no difficulty, see Example 9.7. 0
382 9 Value Sets for Non-linear Coefficient Functions

Remark 9.13

In more difficult cases, the tree-structure is not obvious, e.g. if we deal with an
expanded characteristic polynomial with a high number of uncertain parameters, or
the system is too complicated to recognize a tree-structure while already modelling.
These cases require a suitable tool that brings to light the structures. Algorithms for
the tree-structured decompositions of a parametric polynomial are given in the first
edition of this book of 1993, see also [182] . The basic idea for the decomposition (9.4.9)
is to factorize symbolically the derivative of p(s, q) with respect to q-elements, thereby
Po(s) does not enter into the factorization. 0

9.5 Tree-structured Transfer Functions

In the preceeding sections, we have analyzed control systems by their characteristic

polynomials. We now extend the analysis to frequency domain specifications. Also, in
rational transfer functions, we can exploit specific structures in order to construct, for
example, Nyquist and Popov value sets. The only additional operation on complex sets
is their inversion.
Consider a transfer function
n(s, n)
G(s, n, d) = d(s , d) (9.5.1)

with uncertain parameters n E Qn in the numerator and a e Qd in the denominator and

assume that nand d do not contain common uncertain elements. The two polynomial
value sets N(jw , Qn) and V(jw, Qd) can be constructed independently. Then the set
V(jw, Qd) must be inverted and multiplied by N(jw , Qn). An example is the continued
fraction form (1.6.11) of the crane transfer function.
Value set multiplication was already defined in Section 9.4. Thus, we have to show
the inversion of a complex set. The inversion of a complex set Ace , 0 f/. A is defined
as
A-I = {l/a l a E A}. (9.5.2)
Each point of the set A has to be inverted. If zero is excluded from A, then the inversion
w = liz is a bijective and continuous function. An interior point of A is mapped into
an interior point of A-I. A boundary point of A is mapped into a boundary point of
A-I. Thus , only the boundary of A has to be inverted to get the boundary of A-I :

(0 f!. A) . (9.5.3)

In a polar coordinate representation of the inversion map applied to a complex point

z = Izl&"', we obtain the inverse
(9.5.4)
9.5 Tree-structured Transfer Functions 383

It can easily be seen that th e interior (exterior) of the unit circle in the z-plane is
mapped onto the exterior (interior) of the unit circle in the w-plane. The unit circle
itself is mapped onto the unit circle in the w-plane . Written in real and imaginary parts
of z = x + jy and w = U + jv, the mapping equation is
. 1 x .-y
w=u+jv= - -.- = -2- -2+ j - --, (9.5.5)
x + jY x + y x 2 + y2
and thus
x -y
U=--- , V=--- (9.5.6)
x2 +y2 x2 + y2'
and conversely
U -v
x = u 2 + v 2' Y = u 2 + v 2' (9.5.7)
For a computer-aided value set construction, set boundaries are approximated by a chain
of chords. Therefore, we are mainly interested in how a line segment is transformed by
the inversion map.
We consider all points z = x + jy on the straight line determined by the equation

2ax + 2by = 1. (9.5.8)

Substituting x and y from (9.5.7) yields

U -v
2a-2--2 + 2b-2--2 = 1,
U +v U +v
and after some elementary transformations,

(9.5.9)

This is the equation of a circle with center (a, -b) and radius J a 2 + b2 . The circle
always passes through the origin (0 , 0) of the coordinate system. The point w = 0 is
the image of the point at z = 00 on the line. The resulting circle may also be determined
easily with two additional points on the line to be mapped. Since we want to map a
line segment with endpoints Zl and Z2, these points are the natural choice. The three
points determine the circle, see Figure 9.20. The image of the line segment ZlZ2 is the
arc not containing the origin.

Example 9.14

Consider the mechanical system given in Figure 9.10. In Example 9.9, the value
set of the characteristic polynomial of the system was constructed. In the example, it
was mentioned that the system is passive, which implies Hurwitz-stability for arbitrary
positive parameter values. However, if for instance the damping of the system is not
sufficiently high, feedback control has to be introduced. For this example we consider
the transfer function from input U to the position Xl of the mass ml
384 9 Value Sets for Non-linear Coefficient Functions

.... ....
.... Im z Im w
............. Zl

," ' ·~Izi = 1

I
.1..
I ZI
I Re z
.... ....

ZO

Figure 9.20. Inversion of a line segment

with P1(S, q(l), C12) and P2(S , q(2) , C12) given in (9.3.5). The parameter intervals were
given as follows: m1 E [1 ; 3], d 1 E [0.5 ; 2], C1 E [1 ; 2], m2 E [2 ; 5], d2 E [0.5 ; 2],
C2 E [2 ; 4]. In [29], the controller

nc(s) = 471250 0.5 + 1.9s + 1.7s + S3

2

dc(s) 19000 + 1450s + 62s2 + S3

was proposed for position control of the mass m1. The open-loop transfer function has
a tree-structure if the parameter C12 is fixed, C12 = ci2 . (If C12 is uncertain then this one
parameter must be gridded.) The characteristic equation is
(1) (2) _ nc(s) Ci2 _
1 + Go(s, q , q ) - 1 + d() (1) (2) • - O. (9.5.1O)
c s pp(s,q , q , C12 )
In the following, the value set of the open-loop frequency response Go(jw , q(l), q(2)} is
constructed and checked for exclusion of the critical point -1. (Note that the existence
of a stable Nyquist plot in the Nyquist set is guaranteed by stable pole assignment for
the center of the Q-box.)
For a fixed frequency w = w', the value set of the denominator polynomial
pp(jw', q(1), q(2) , ci2) is constructed as demonstrated in Example 9.9. The resulting
set is inverted and multiplied by ci2. The value set for w' = 1 and ci2 = 1 is displayed
in Figure 9.21. To obtain the value set of the open-loop transfer function, this set has to
be multiplied by the transfer function of the controller. For a fixed frequency, this is a
complex number. In Figure 9.22, the union of the Nyquist sets of the open-loop transfer
function for fixed C12 and a grid on w is displayed . Figure 9.23 shows a detailed view
of the neighborhood of the critical point -1. The critical point -1 is excluded from
the union of the sets . Hence, the closed-loop system is robustly stable. The minimal
distance of the Nyquist sets from the critical point is a measure for the stability margin
of the closed-loop system, see Chapter 5. In the example, the radius of this circle is
approximately 0.24.
o

With the use of transfer functions for value set operations, tree-structures can be
exploited in complex control system structures. For more complicated networks, Ma-
son's formula [202] can be used to compute the characteristic equation. Similar as in
the polynomial context, a basic rule for modelling of structured systems is:
9.5 Tree-structured Trensier Functions 385

1m 1.0

0.5

0.0

-0.5

-1.0 -0.5 0.0 0.5

Re
Figure 9.21. Plant transfer function set for w = 1, C12 = 1

never execute additions or multiplications of transfer functions with disjoint uncer-

tainties.
The tree-structured decomposition of polynomials and transfer functions is a pow-
erful tool for robustness analysis. It can handle an extremely large number of uncertain
parameters. In [30], for example, a chain of 17 mass-spring-damper systems with 51 un-
certain parameters is analyzed. In [183], a control system for a finite-element model of a
plate with 60 uncertain parameters is analyzed. The concept of tree-structures is also
fuseful or feedback connected subsystems with independent uncertainties as illustrated
by the following examples.

Example 9.15

In this example, we assume that the transfer functions have disjoint parameter sets
in the numerator and denominator polynomials and that different transfer functions
have no uncertain parameters in common.
In the feedback loop of Figure 9.24 the closed-loop characteristic polynomial

p(s,n,d) = nl(s,nl)n2(s ,n2)ne(s,ne)

+ d1(s, d 1)d2(s, ~)de(s, de) (9.5.11)

with n := [n[ nf nIV, d := [di dI d~]T has a tree-structure. However, the character-
istic polynomial of the feedback loop in Figure 9.25 does not have a tree-structure:

p(s,n,d) = [nl(s,ndd2(s ,~) +n2(s,n2)d1(s,d1)]ne(s,ne)

+ d1(s,d 1)d2(s, d 2)de(s, de) . (9.5.12)

The parameters d, and ~ appear in two terms of p. However, in the characteristic

386 9 Value Sets for Non-linear Coefficient Functions

Irn 50

25

-25

o 25 50 75
Re

Figure 9.22. Union of Nyquist sets for a grid on w

equation of this feedback loop

r(s, n , d) = 1 + g(s, n , d)
1+ [nl s,nl
dl s, d l
(9.5.13)

where g(s, n, d) is the open-loop transfer function, each uncertainty vector appears just
once, either in a numerator or in a denominator polynomial. Thus , the value set of
the open-loop transfer function g(s, n, d) can be constructed sequentially. Also, the
closed-loop transfer function
1
gc(s,n,d) .-
(9.5.14)

can be constructed sequentially from the value sets of the three subsystems. 0

Example 9.16

The transfer function of the crane with full state feedback from input u to the
position Xl is

Xl(S) g+fs 2
u(s) = gk 1 + gk 2s + (kIf - k3 + g(mc + mL))s2 + (k2f - k4 ) S 3 + fmcs 4 '
9.5 Tree-structured Transfer Functions 387

1m 1.0

0.5

0.0

-0.5

-1.0 -0.5 0.0 0.5

Re

Figure 9.23. Detailed view of the union of Nyquist sets

Figure 9.24. Serial connect ion with feedback

The numerator is unchanged from the open-loop transfer function (1.3.3). The denom-
inator has been changed by state feedback to the form (1.5.18). The numerator and
denominator have the uncertain parameter rope length e in common. Dividing both
the numerator and the denominator polynomial by the numerator, we get a continued
fraction with each uncertain parameter appearing just once:
Xl (s) 1
= ----------:02---=(- - -,-)--:<"2.
U( s ) k k gmLs - k3 + k 4s
g+ eS 2
2 S
1 + 2S + maS +
Thus, the value set of the closed-loop transfer function can be constructed sequentially.
o
A standard form of the characteristic equation is 1 + Go(s, q) = 0, where Go(s, q)
is the open-loop transfer function. Remember that Go(s, q) is not necessarily given in
388 9 Value Sets for Non-linear Coefficient Functions

nl(s,nd
dl(s,dl)

Figure 9.25. Parallel connected transfer functions with feedback

form of a ratio of two polynomials. The two examples have led to a continued fraction
form. Any other nested rational form would be feasible too. The important point for
the construction of the value set of G(s, q) is only that it is in a form such that each
uncertain parameter appears only once.
If the transfer function of a plant has a tree-structure, then it is also easy to construct
the family of Popov plots

P(jw , Q) = {pOw,q) = Re GoOw, q) + jw Im GoOw, q) I q E Q}.

For each frequency grid w = o", the imaginary part of the set Go(jw,q) must be
multiplied by io" , Thus , also robust stability of a system with a sector non-linearity
and a large number of uncertain parameters in the linear part can be analyzed, provided
that it has a tree-structure.

Example 9.17

Consider the crane (1.1.6), uncertain in the rope length f E [8; 16] [m], the load
mass mL E [50 ; 2000] [kg], and the crab mass me E [800 ; 1200] [kg]. Using full state
feedback u = k T X = ki Xl + k 2 X2 + k 3 X3 + k 4 X4, it can be shown that the closed-loop
system is robustly stable for a controller coefficient vector k with values arbitrarily
chosen in the intervals kl E [500 ; 700], k2 E [3000 ; 4000], k3 E [-30000; -25000],
k 4 E [-2800 ; -2400] . Now, we assume that the unmodelled actuator has a sector
non-linearity. We want to determine the greatest sector for that non-linear function,
for which the closed-loop system is absolutely stable. The linear part of the system is
described by the transfer function

G ( ) _ (kl + k2s)(f s2 + g) -
(k 3 + k 4s)s2
(9.5.15)
o s,q- s2[mc( s2f+g)+mLg] ,
9.6 The Stability Profile 389

where q = [fmL me k 1k 2 k 3 k4 ]T. Since Go is unstable , the Popov theorem cannot

directly be applied. Thus, the system first has to be transformed to the standard form
in Figure 5.16.
From the above, we know that the dosed-loop linear part of the system is robustly
stable for p = 1. That value is used for constructing the Popov value set of Go =
Go/(1 + Go) = 11(1 + I/G o). In the transfer function (9.5.15), all uncertain parameters
but the rope length appear just once, i.e. the transfer function has a tree-structure if
the rope length f is gridded. For the following, f is assumed fixed with f = 12 [m].
The Popov value set can be constructed sequentially, the union of the sets is shown in
Figure 9.26. From that plot, we get the Popov-sector [0; 00) for the non-linear function
Im 0.5

-1.0

-2.5

-4.0

0.00 0.25 0.50 0.75 1.00 1.25

Re

Figure 9.26. Union of Popov value sets

j(.) = (f - p)(.). Thus, for the transfer function Go, that means for the crane with the
above state feedback, absolute stability is guaranteed in a k-sector [1 ; 00) . 0

9.6 The Stability Profile

In Section 3.2, we have discussed root sets in the s-plane, see, for example, Figure 3.7.
For a large number of uncertain parameters, the computation times for root sets are
390 9 Value Sets for Non-linear Coefficient Functions

prohibitive. If the parameters appear in a tree-structure, however, then it becomes

feasible to compute the right hand boundary of the root set . This boundary is called
the stability profile. The stability profile contains important information, e.g. on worst
case damping or the real part of all eigenvalues for all operating conditions.
First, the Hurwitz-stability analysis is generalized to r-stability with respect to the
boundary
ar = {lTo + jw Iw ~ O}. (9.6.1)

All eigenvalues should have a real part smaller than lTo (lTo-stability). The polynomial
family
P(s, Q) = {p(s, q) I, q E Q}
is lTo-stable, if and only if:
1. There is a lTo-stable polynomial in the family.
2. The value set P(lTo + jw, Q) excludes the origin at all frequencies w ~ o.
In order to construct the values, we only have to substitute s = lTo + jw instead of
s = jw. A tree-structure of the system is not destroyed by this substitution.

Example 9.18

Consider the value set of Figure 9.13 for the mass-spring-damper system of Fig-
ure 9.10. For the fixed frequency w· = 1, the real part lTo is shifted to the left in small
steps until at lTo = -0.062 the contour passes through the origin, see Figure 9.27. The

Figure 9.27. The real part is shifted to the left until zero is no longer excluded for 0"0 = -0.062
9.7. Synopsis of Parametric Robustness Analysis 391

same procedure is repeated for a grid on w. The resulting stability profile is shown in
Figure 9.28. The two squares indicate the eigenvalues for the center of the Q-box. 0

1m

2
---------------

o
1
o

----------------------
Re
-0.1

Figure 9.28. Stability profile for the mass-spring-damper system of Figure 9.10

Rather than shifting a in small steps, it is more efficient to use a bisection as

schematically illustrated by Figure 9.29. For the DLR plate experiment with 60 uncer-
tain parameters in the finite-element model, the stability profile is shown in Figure 9.30
and Figure 9.31. For the details the reader is referred to [1841.

9.7 Synopsis of Parametric Robustness Analysis

The different approaches for parametric robustness analysis presented in this book are
summarized in the following synopsis.
Consider the polynomial family

P(s , Q) = {p(s, q)lq E Q}, (9.7.1)

where q = [qI q2 . . . qtl can take on any value in the operating domain

Q = {qlq; E [qi; qtJ, i = 1,2, . . . ,f}. (9.7.2)

Is the polynomial family robustly r-stable (defined by a given eigenvalue region r with
boundary ar = q(a) + jw(a) , a E [a- ; a+]) ?
392 9 Value Sets Ear Non-linear Coefficient Functions

·: /:
~
1m

'----------~- - -~-~~~~
~
~
~ "":~'-" .

.•---
•
••• Re

Figure 9.29. The real value 0"0 is founded by repeated bisection between a f-stable value 0"1
and a f-unstable value 0"2. Then O"max = max O"o(w)
w

1. Grid Q, factorize p(s,q) for grid points, and plot the root set in s-plane. For p,
stability it must be contained in f . For calculation of the right hand boundary
of the root set (stability profile) see 2.2.2.4.
2. Check an arbitrary q E Q for f -stability. Its stable neighborhood Qr is
bounded by

Rep [O"(a) + jw{a), q] o ,and

Imp [O"{a) + jw{a), q] = O. (9.7.3)

2.1 Eliminate a from (9.7.3). For f= left half plane this leads to the last Hurwitz
determinant

detHn{q) = O. (9.7.4)
2.2 Grid a E [a- ;a+].
Check for singular frequencies; in their vicinity use fine grid , otherwise coa.rse
grid.
2.2.1 For two parameters ql, q2 entering polynomially, multilinearly or linearly into
characteristic polynomial coefficients, solve for ql (a) , q2{ a) or It (a, ql, q2) = 0
and plot the f-stability boundary in the (ql' q2)-plane. Q is robustly stable if
Q c Qr. Grid further parameters qa, q4 . . . and project boundaries.
9.7. Synopsis of Parametric Robustness Analysis 393

1m

150

100

50

Re

-2.0 -0.5

Figure 9.30. Stability profile for a finite-element model with 60 uncertain parameters

2.2.2 Use zero exclusion from the value set with the following cases:
2.2.2.1 For interval polynomials and r= left half plane, check four Kharitonov poly-
nomials (8.3.6). '
2.2.2.2 For affine coefficient functions, check all edges of the Q-box for r-stability.
2.2.2.3 For multilinear coefficient functions, the Desoer mapping theorem is a useful
sufficient condition .
2.2.2.4 For non-linear coefficient functions, check for a tree-structure that simplifies
the construction of value sets and stability profile.
394 9 Value Sets for Non-linear Coefficient Functions

1m

10

I 5

---~
-~
_• ..a:-- Re

-2.0 -0.5
Figure 9.31. Clipping of Figure 9.30
10 The Stability Radius

In Chapters 8 and 9, the stability of polynomial families

P(s ,Q) = {p(s,q) I q E Q}, (10.0.1)

with p(s,q) = ao(q) + al(q)s + .. . +an(q)sn and q, E [qj;qtJ, i = 1,2 , .. . ,£ was

investigated. The primary interest was necessary and sufficient conditions for stability.
These conditions give primarily a yes or no answer to the robust stability problem. If the
answer is yes, then there is some kind of stability margin of the uncertain parameters,
the domain Q of the parameters can be enlarged without loosing stability. The next
question is: what is the smallest perturbation in the parameters q that destabilizes the
system?
Assume that p(s, q) is stable for the center point qO of the Q-box. Then , the
polynominal family is stable in a sufficiently small neighborhood of qO because of the
continuity of the roots . By shifting the origin of the parameter space, it is always
possible to have qO = O. The box centered at qO = 0 may now be blown up by a
dilation factor 'Y, i.e. the polynomial family

P(s, 'YQ) = {p(s , q) I q E 'YQ} (10.0.2)

is considered. By increasing 'Y , a value p must be reached where a member of the

polynomial family becomes unstable , i.e. the box 'YQ hits a stability boundary; this
value p is called the stability radius of the polynomial family, it is a measure for the
smallest destabilizing perturbation. For all q with IIqllao = max;=1.2,....llq,1 < p, the
polynomial family is stable. Note that the value p may be 00 , for example, consider
p(s,q)=s+l+q2.
The sidelength of the Q-box provides a scaling for the individual uncertain param-
eters q,. After that scaling, a unit box Qao = {qlq, E [-1; I]} is enlarfed by the
dilation factor 'Y . For computational simplicity, a unit ball Q2 = {q I 2::'=1 ql = I}
may be dilated . The stability radius P2 is then defined as the smallest 'Y for which 'YQ2
touches a stability boundary. The difficulty of finding the stability radius p depends
again on the type of coefficients functions a,(q), i = 0,1, ... , n. For affine dependency,
two methods are demonstrated. The case of polynomial dependence (which includes
multilinear dependency) will be treated in the third section. There, the problem is
solved in theory but in applications only a small number of parameters can be handled.
396 10 The Stability Radius

10.1 Tsypkin-Polyak Loci

In Chapter 8 in the interval and affine cases, algebraic tests have been used to check
the stability of testing sets. An alternative approach is the use of graphical frequency
domain stability criteria. This approach is particularly useful if the largest stable Q-box
around its stable center with coordinates
q+ + q-:-
°
qi = ' 2 ' i=I,2, .. . ,f (10.1.1)

has to be determined. A given Q-box with qi E [qi- ; qtJ may then be described by

°
Iqi - qi I ~ ai =
qt - qi
2 ' i = 1,2, . . . ,f. (10.1.2)

A variable size of the uncertainty box can now be introduced by a common real dilation
factor "( ~ 0 for all uncertainties, i.e.

Iqi - q?1 ~ "(ai, i = 1,2, . . . , f . (10.1.3)

The case "( = 1 coincides with the given box in (10.1.2). For "( < 1, the box size is
reduced and by increasing "( beyond 1 the uncertainty box is blown up. The "(-value,
for which the box around a stable center point (10.1.1) first hits the stability boundary,
is called the stability radius.

Case 1: Interval Coefficients

Consider the polynomial family

(10.1.4)

with the coefficients subject to the following constraints:

lai - a?1 ~ "(ai, i = 0, 1, . .. ,n, ag > o. (10.1.5)

This means that for each coefficient ai, there is a nominal value a? and a scaling factor
ai ~ 0 for the coefficient perturbation.
If pO(s) = ag + a?s + ...+ a~sn, ag > 0, is the nominal polynomial , then

p°(jw) = Re p°(jw) + jim p°(jw) = : U(w) + jwV(w), 0~w< 00 (10.1.6)

represents the usual Mikhailov curve. For testing the polynomial family, introduce the
two scaling polynomials

(10.1.7)

(10.1.8)
10.1 Tsypkin-Polyak Loci 397

and construct the frequency plot

z(w) = x(w) + jy(w), (10.1.9)

with
U(w) (10.1.10)
x(w) := S(w) ,

V(w) (10.1.11)
y(w) := T(w) .

Observe that the coefficients of the numerator polynomials only depend on the coeffi-
cients of the nominal polynomial, whereas the denominator coefficients depend on the
scaling factors O'i . If the O'i are positive, then the denominator polynomials are positive
so that z(w) is finite for all finite w > O. The boundary points z(O) and z(oo) are
o
x(O) = ao , (10.1.12)
0'0

o
y(O) = ~, (10.1.13)
0'1

r
and
.::n.
O'n'
n even
x(oo) (10.1.14)
a~_1
n odd

r
O'n-l'

.::n.
O'n'
n odd
(00) (10.1.15)
a~_l
O'n-l'
n even

Theorem 10.1 (Tsypkin and Polyak theorem for interval polynomials)

For the stability of the polynomial family (10.1.4) and (10.1.5), it is necessary
and sufficient that the following conditions hold for the plot z(w):

1. It goes through n quadrants in counterclockwise direction for wE [0; 00) .

2. It does not intersect the square centered at the origin with sidelength 2,.
3. Its boundary points z(O), z(oo) have coordinates with absolute values larger
than , .

o
398 10 The Stability Radius

I Y I
I I
I I
II I I I
I I
I I
------ ------
"I

"I x

------ ------
I I
I I
III I I IV
I I
I I
I I

Figure 10.1. The feasible domain for the Tsypkin-Polyak locus

Condition 2 requires that z(w) is outside the square. Condition 3 means that Ix(O)1 >
"I, ly(O)1 > "I, Ix(oo)1 > "I, ly(oo)1 > "I. Thus , z(O) and z(oo) must lie in the
regions I, II, III or IV of Figure 10.1.
This theorem is a frequency domain version of Kharitonov's Theorem . The advan-
tage compared to plotting four Mikhailov curves is that only one plot (even bounded if
Qo =1= 0, QI =1= 0, Qn-I =1= 0, Q n =1= 0) is needed and the stability radius can be determined
easily. When "I is enlarged until one of the conditions of the theorem is violated, then
the critical frequency is also known, i.e. the frequency where the root set crosses the
imaginary axis.

Remark 10.2
Theorem 10.1 is a simplified version of the original theorem in [191] . Tsypkin and
Polyak additionally allow other types of perturbations. For the long proof the reader
is referred to the original paper . A simpler proof was given by Mansour [138] . 0

Example 10.3

Recall the characteristic polynomial (1.5.18) of the crane with k l = 600, k 2 = 2000,
k 3 = -10000, k 4 = 0, me = 1000 [kg], e = 10 [m] . We want to check the stability of
the polynomial family for mL E [50 ; 2395] [kg] . This family
pi», mt} = 0.6 + 28 + (2.6 + 0.00Imt}82 + 283 + 84
has first to be scaled. With kg = -10000, k~ = 0, "I = 1 and
k3 = - 10000 + 10000ql , (10.1.16)
10.1 Tsypkin-Polyak Loci 399

(10.1.17)

the Tsypkin-Polyak locus z(w) of the transformed polynomial family

(10.1.18)

must avoid the unit square. The polynomials (10.1.6) (10.1.8) are U(w) = 3 - 18w2 +
5w4 , V(w) = 10 - lOw 2 and S(w) = T(w) = 5w2 • The frequency plot

3 - 18w2 + 5w4 • 10 - lOw 2

z(w) = 5w2 +J 5w2 (10.1.19)

must go through four quadrants and avoid the unit square. The boundary conditions
are satisfied because z(O) = 00 + joo and z(oo) = 00 - 2j. The plot starts in the
first quadrant and goes through four quadrants. Thus , the nominal polynomial for
ql = q2 = 0 is stable, see Figure 10.2.
To determine the stability radius, the box is enlarged until it hits the Tsypkin-
Polyak locus. From Figure 10.2, it follows that the box touches the Tsypkin-Polyak
plot for x(w) = y(w), i.e, 5w4 - 18w2 + 3 = -lOw 2 + 10 or 5w4 - 8w2 - 7 = O. The
only real positive root W m = 1.49 leads to Zm = -1.1 - 1.1j and to the stability radius
P = lmax = 1.1.
o

!
1m z
3

I 2

--
0
-4 -3 -1 0 1 2 3 4
~ -1
I--
-2

-3

-4

Re z

Figure 10.2. The Tsypkin-Polyak locus

400 10 The St ability Radius

Case 2: Affine Coefficients

The next step is now the extension of the frequency domain approach to affine depen-
dency of the polynomial coefficients. The polynomial family P(s , Q) can be rewritten
in the form
p(s, q) Po(s) + E~=1 qiPi(S) ,
Po(s) = ag+a~s+ +a~sn, a~ > 0 , (10.1.20)
Pi(S) ah + a~s + + a~sn , i = 1,2 , . .. , e.

q may vary in an e-dimensional box, i.e. qi E [q;; qtJ, i = 1,2, ... , e. Shifting the
origin and scaling the qi-axes in the parameter space, the box Q is transformed to an
e-dimensional cube with sidelength 2 and center at q = [0 . . . O]T , see (8.4.4). Let W s
be a common real zero of the rational functions

for k = 1,2, . .. ,e.

Now, a real-valued function T(W) is defined by
11m (Po (jW)/Pk(jW)) I
T(W) = max e 0<W < 00 , W =I- w s , (10.1.21)
L
l::;k::;t
11m (Pi(jW)/Pk(jW)) I
i=1

IPo(jw)I (10.1.22)
T(W) e
W =Ws ,

L Ipi(jW) I
i=1

T(O) ~
t
T(oo) = tla~ 1 (10.1.23)
Lla~1 Lla~1
i=1 i=1
For W = 0, W = 00, and W = ws , the function is, in general, discontinuous. The first
two cases correspond again to roots at s = 0 and s = 00 . The separate definition of
T(W) for Ws is necessary because T(W) is discontinuous at the singular frequencies w s , a
formal evaluation of T(W s ) with (10.1.21) would lead to 0/0.

Theorem 10.4 (Tsypkin and Polyak theorem for affine coefficients)

The polynomial family P(s , Q) is stable , if and only if:

1. Po( s) is stable.
2. T(W) > 1, 0 ~ W~ 00 .

o
10.1 Tsypkin-Polyak Loci 401

A slightly different form of this theorem, which resembles the Mikhailov plot, is

Theorem 10.5 (Tsypkin and Polyak theorem for affine coefficients)

The polynomial family P(s , Q) is stable, if and only if,

1. Po(jw) i' O.
2. z(w) = ~~ 1:) r(w) goes for 0 :5 w < 00 through n quadrants and does not
intersect the unit circle.

The difference between Theorem 10.4 and Theorem 10.5 is only the graphical rep-
resentation. For a stability test by Theorem 10.4, the function r(w) is plotted. Its plot
must be entirely above the line r = 1. In Theorem 10.5, a polar plot is generated,
where r(w) is the distance from the origin and the phase angle is that of the nominal
. I. po(jw)
po Iynomia i.e. Ipo(jw)I '

Remark 10.6

The function r(w) can also be defined using trigonometric functions. However, for
computational purposes, it is more convenient to have a non-trigonometric version. For
the proof, the reader again is referred to the original paper [192) . Using Theorem 10.4
or Theorem 10.5 beyond a yes or no answer, critical frequencies are recognized. 0

What is the amount of calculation for this test? First, the frequency w has to be
gridded and the polynomials Po(s), Pi(S), i = 1,2, .. . , ehave to be evaluated at s = jw.
Then, e functions have to be calculated and their maximum has to be determined.
The singular frequencies w. can be found by transforming 1m (Pk(jW)jPO(jw)) . With
Pk = Rk + jh and Po = Ro + jlo, it is
.
1m (Pk(jW)jPo(jw))
. Rk + jh
= 1m Ro
+j
'!<
0
= m+ 1
Roh - RkIo
2
0
= O. (10.1.24)

So w. must be a common (real) root of the polynomials

Roll - Rllo 0,
RoI2 - R2Io 0,
(10.1.25)

Role - Relo = O.
402 10 The Stability Radius

Example 10.7

Recall the characteristic polynomial family (1.5.18) of the crane with kI = 300,
e= 10/3 [m]. The transformation
k 4 = 0, me = 100 [kg], mL = 1000 [kg],

k2 = 535 + 5qI l k3 = 7662.5 + 87.5q2

allows qI and q2 to vary in the square of sidelength 2. The polynomial family is

(10.1.26)

with

Po(s) = 129 + 1668 + 23782 + 10883 + 8084 ,

PI(S) = -16+24s-1282+4s3 ,
P2(8) = -21 + 428 - 2182 •

At first, a check is made for the existence of isolated frequencies. The polynomials
(10.1.25) are

l1{) = 129 - 237w2 + 8Ow4 ,

10 166w- 108w3 ,
RI = -16 + 12w2 ,
II 24w - 4w3 I

R2 -21 + 21w2 ,
12 42w I

and

l1{)II - RIIo = -4w(w 2 - 2)(8Ow 4 - 881w2 + 719) ,

l1{)I2 - R 2Io = 84w(w2 - 2)(67w2 - 53) .

Common real roots are w. = 0 (which is always the case) and w. = J2.

The test could be finished because this value is smaller than 1 and the polynomial
family must be unstable. For completeness, the other values of r(w) are calculated.
They are

r(O)
r(oo) 00 ,
10.1 Tsypkin-Polyak Loci 403

and otherwise
IIm(Po(jw) /Pl (jw))I
IIm(p2(jw)/Pl(jW))I '

IIm(po(jw)/P2(jw) )1
IIm(Pl (jw)/ P2(jw))I '

r(w) =
From Figure 10.3 (the ordinate is logarithmically scaled) , it can be seen that r(w) > 1
for all w =/: V2. The minimum of the function r(w) is 25/29 for the frequency w =
V2. The modulus of the components ql and q2 must be smaller than 25/29. The
corresponding intervals for the original feedback gains are then k1 E [530.7 ; 539.3] and
k 2 E [7587 ; 7738]. The parameter combination ql = q2 = 25/29 produces a polynomial
with a root pair at s = jV2.

r(w)
102 -r--------------------.,

100 ---------0------------------------

o 1V2 2 3 4 5

Figure 10.3. The distance function r(w) for testing the stability. r(w) is> 1 at the continuous
part, but < 1 at the isolated frequency W s = y'2, so the polynomial family is unstable

An alternative method is using Theorem 10.5. The magnitude of the complex-valued

function z(w) = I:(~:~((w) is evaluated for all w and the infimum of this function
determines the stabMty radius. This is equivalent to enlarging the unit circle in the
complex plane until there is an intersection with the frequency plot z(w). In Figure lOA,
the alternative Tsypkin-Polyak locus is plotted for determining the stability radius. The
function z(w) is discontinuous at the singular frequencies Wsl = 0 and Ws2 = V2 (marked
with 0). The continuous part avoids the unit circle but z( V2) is inside the un it circle .
Thus, the polynomial family is unstable and to get stability the operating domain of ql
and Q2 must be reduced by a factor < z( V2) = 25/29. 0
404 10 The Stability Radius

Im z(w)

Re z(w)

Figure lOA. The alternative Tsypkin-Polyak locus for testing the stability radius, it is dis-
continuous and does not avoid the unit circle at the isolated frequency W s2 .= y'2, so the
polynomial family is unstable

10.2 Affine Dependence: The Largest Hypersphere

in Parameter Space

The evaluation of T(W) and the determination of the largest stable box is a rather
expensive task. A more effective method is now presented for a closely related problem,
namely calculating the largest stable hypersphere in the space of parameters q. This
problem was first solved by Soh et al. [186] . A simpler solution was given by the authors
[119], it is presented here.
Consider an uncertain polynomial p(s, q) = [1 s . . . sn] a(q) with affine dependency
of the coefficient vector a on q, i.e.

a = aD+Fq, (10.2.1)

with aD E jRn+I and F E Each element qi of q is scaled by a "high value"

jR(n+I)xf .
qt, and this scaling factor is included in the i-th column of F . Then, a hypersphere
L1=1 ql = const . in q-space is a reasonable measure for the neighborhood of q = O.
The parameter vector q = 0 is the nominal point , i.e. the coefficient vector of the
unperturbed polynomial is aD. Assume that p(s, 0) is stable. For e = n+ 1 and F = I ,
the polynomials are interval polynomials. The monic case (an = 1) is also included if
the last element of aD is 1 and the elements of the last row of F are zero.
10.2 Affine Dependence: Th e Largest Hypersphere in Parameter Space 405

A geometric approach is used to dete rmine t he largest hypersphere. Th e knowledge

of some fundamental definitions and formulas of higher dimensional geometry is needed.
In ]R3, a plan e is defined by

E = el ql + e2q2 + e3q3 + eo = O. (10.2.2)

Th e generalization to ]Rt is th e (e- I)-dimensional hyperplane

E = el ql + e2q2 + ...+ etqt + eo = 0 . (10.2.3)

The normal vector e(1 0) of a hyperpl ane is given by

(10.2.4)

Two hyperplanes are orthogonal if th eir normal vectors el , e2 are orthogonal, that is,
e l T e2 = O. The squared dist ance d? of th e origin from th e hyperplane is

(10.2.5)

Star ting from the st able nominal point and varying q, the boundary crossing theorem
states t hat t here are three possibilities for t he polynomial to become unst able:
i. A real zero goes t hrough t he origin (8 = 0).
ii. A real zero goes through infinity (8 = 00) .
iii. A pair of conjugate zeros crosses t he imaginary axis (8 = ±jw).
In th e parameter space, each of the t hree cases corresponds to a hypersurface. Parts
of these hypersurfaces are th e stability boundaries. Th e minimal distance of th e origin
from these hypersurfaces determines the largest hypersphere.
Th e first hypersurface is given by p(O, q) = [1 0 . . . O]a = [1 0 ... 0] lao+ Fq] = O.
Only t he first row of (10.2.1) is important . Because of t he affine dependency, th e
hypersurface is a hyperplane wit h th e equation

Eo = ag + fll ql + !I2q2 + ...+ fllqt = 0 , (10.2.6)

and th e squared distance to th e origin is with (10.2.5)

2 _ (ag/ (10.2.7)
f2
TO -
11 + f 122 + ... + f It2 .
ag
If I 0 and fll = !I2 = ... = !It = 0, then case a) is not possible and th e hyperplane
does not exist. No combinat ion of t he parameters ql, q2, . . . , qt will produce a zero at
8 = O. A reasonable choice for the distance is TO = 00 .
T he same conclusion leads to the hyperplane for case b). It is p(oo , q) = 0 (t he
leading coefficient must vanish) and

E oo = a~ + ! n+t ,1ql + !n+t,2q2 + .. .+ fn+l,tqt = 0, (10.2.8)

406 10 The Stability Radius

with the squared distance

(a~)2
r;, = f2 n+1 ,1 + f2n+1,2 + .. . + e.;.
n+1,t
(10.2.9)

For a constant leading coefficient an, again roo = 00.

For the more complicated case c), the polynomial p(s, q) = [1 s ... sn] a =
[1 s .,. Sn] [aD + Fq] must have a root on the imaginary axis. For s = jw with w =I- 0,
the complex equation is equivalent to the following two real equations:

E I = Re p(jw, q) = [1 0 -w 2 0 w4 . .. ) [aD + Fq] = 0 (10.2.10)

and

E2 = 6Im pOw,q) = [0 1 0 -w 2 0 w4 ... ) [aD + Fq] = 0 . (10.2.11)

The complex root boundary is generated by the intersection of the two hyperplanes
E I = 0 and E 2 = 0 that vary with w. The set of intersection points for fixed w is an
(f - 2)-dimensional hyperplane. The distance of the origin to this (f - 2)-dimensional
hyperplane is a function of w, call it rc(w) , and the distance from the complex root
boundary is the minimum of this function rc(w) .
The calculation of rc(w) would be easy ifthe two hyperplanes (10.2.10) and (10.2.11)
were orthogonal. This is, in general, not true, however it is possible to replace E2 = 0 by
a third hyperplane E a = 0, which is orthogonal to E 1 = 0 such that E 1 n E 2 = E 1 n Ea.
The important point to note is that the equation

(10.2.12)

is satisfied for all points lying on both E 1 = 0 and E2 = O. (10.2.12) represents another
(f - 2)-dimensional hyperplane, which contains the intersection points of E I = 0 and
E2 = 0, whatever the value of >.. As >. is varied, a set of (f - 2)-dimensional hyperplanes
is formed, two of which are E I = 0 (when>. = 0) and E 2 = 0 (when>. = 1). E 1 n E a
will produce the same set as E I n E 2 if>' =I- 0 (see Figure 10.5). The value of>. can now
be chosen in such a way that E 1 = 0 and Ea = 0 are orthogonal. The normal vector €a
of E a = 0 is
(10.2.13)
and €I and €a are orthogonal if

(10.2.14)

or T
\ _ €l €I
/\ - T T (10.2.15)
€I €I - €I €2

Note that>. is a function of w. This orthogonalization procedure is not possible if

(for singular frequencies w = ws ) the hyperplanes E 1 = 0 and E 2 = 0 are parallel.
This means a) that for these values, both hyperplanes have no common points and
rc(w s ) = 00, or b) E I == E2 and it is obvious how to determine the distance to the origin.
10.2 Affine Dependence: The Largest Hypersphere in Parameter Space 407

-- -- --- --
,---
,, 2
,E
,,
,

--- ---
Figure 10.5. For each frequency w the complex root boundary is generated by the intersection
of the two hyperplanes E 1 and E2, which is equal to the intersection of E 1 and E3

Case b) produces a discontinuity in the distance function . It is another appearance of

the isolated frequencies that were discussed in Section 4.4.
If the polynomial coefficients vary in the coefficient space between given bounds
(interval polynomial) rather than in the parameter space, then the hyperplanes E 1 = 0
and E 2 = 0 are orthogonal and there cannot be any singular frequencies (F = I and
the normal vectors are orthogonal, i.e, d;d 2 = 0).
The reason for orthogonalizing is that afterwards the distance function rc(w) can
be determined. (10.2.10) and (10.2.11) can be written as

D(ao+Fq)=Dao+Gq= [~], (10.2.16)

where
4
w ]
(10.2.17)
o .
G : = DF = [ :r ]. (10.2.18)

Thus, in the parameter space the equations of the two hyperplanes are

(10.2.19)

E2 = dT2 a °+ eT_O
2q - , (10.2.20)
408 10 The Stability Radius

where el and e2 are the normal vectors of E 1 = 0 and E2 = O. For each non-singular
frequency w, it is possible to construct the hyperplane E 3 = 0 of (10.2.12) with A of
(10.2.15) such that the two hyperplanes E 1 = 0 and E3 = 0 are orthogonal : efe3 = O.
This situation is illustrated by Figure 10.5. The nominal point is q = O. Seen from
there , the closest point on the complex root boundary is q c = q cl + qc3 ' The unit vector
parallel to el is e 1 / J efel and by (10.2.5) di a O/ J efel is the distance between the
origin and E 1 = 0, thus ,
dTa O
qcl = --T-el .
1
e 1 el
Analog reasoning for qc3 leads to

(10.2.21)

The distance rc(w) is

(diaO)2 + (dfaO)2
(10.2.22)
efel efe3
The squared distance rb(w) is a rational function of w and the necessary condition
for the minimum at w* is that the derivative vanishes. However, keep in mind that
(10.2.22) is only valid for non-singular frequencies. At a singular frequency ws , the
normal vectors are parallel. It follows that either the hyperplanes E 1 = 0 and E 2 = 0
are parallel and not identical then no (£- 2)-dimensional hyperplane is generated and
rc(ws ) = 00 can be set, or the hyperplanes are identical, i.e. E 1 = E 2 = 0 and the
formula used for Eo = 0 can be applied.
After the distances ro, roo, rc(w*) of the origin of the three boundaries are deter-
mined, the stability radius p is found as

p = min{ro, rex" rc(w*)} . (10.2.23)

In comparison to the method described in Section 10.1 for the affine case where £
functions have to be evaluated , here only one function has to be evaluated.

Example 10.8

Recall Example 10.7 The matrices a O and Fare

129 - 16
166 24 -21
42 ]
237 -21 . (10.2.24)
F~ -1~
108 [ o
80 o
The real root boundary for s = 0 is the straight line

Eo = -16ql - 21q2 + 129 = 0 , (10.2.25)

10.3 Polynomial Dependence 409

which gives ro = 129/V162 + 212 ~ 4.89. The real root boundary for s = 00 does not
exist, so roo = 00. The orthogonalization procedure leads for w =f V2 to
2() 1O(640w8 - 14096w 6 + 287085w 4 - 439886w 2 + 175573)
r w--'------------....,,---------'- (10.2.26)
c - 441(w2 + 1)2

l
and
67w2 - 53 ]
w2 + 1
(10.2.27)
qc = _ 80w4 _ 881w2 + 719 .
21(w + 1)
2

There is a singular frequency at w = V2 and E} (V2) = E 2(V2) = 8q} + 21q2 - 25 = O.

Thus ,
25
rc( V2) =
v8 2 + 212
~ 1.11 . (10.2.28)

A plot of rc(w) in Figure 10.6 (with logarithmic scaling on the ordinate) shows that this
is the minimum . Thus, the stability radius is P2 = rc( V2). Note that in Example 10.7,

1 2 3

Figure 10.6. The distance function rc(w) with two discontinuities at w = 0 and w = V2
a smaller stability radius of p = 25/29 was obtained for the maximum size of the box
uncertainty. The hypersphere, i.e. circle for e= 2 of Example 10.9 is contained in the
box and can be dilated further before it hits the stability boundary. 0

10.3 Polynomial Dependence

In this section, polynomials are considered whose coefficients depend polynomially on

the elements of an uncertain parameter vector . The size of perturbation is characterized
410 10 The Stability Radius

by the weighted norm of the parameter vector. The smallest destabilizing perturbation
defines the stability radius of the set of uncertain polynomials.
It will be shown that determining this radius is equivalent to solving a finite set
of systems of algebraic equations and then selecting the real solution with the small-
est norm. The number of systems of equations depends crucially on the dimension
( of the parameter vector, whereas the complexity of systems of equations increases
mainly with the kind of polynomial dependence a(q) and the degree n of the polyno-
mial. This method also yields the smallest destabilizing parameter combination and
the corresponding critical frequency.
Vicino et al. [197J transform the problem into an optimization problem and present
a numerical algorithm to find the solution. It will be shown that the problem can also
be solved in an analytical-numerical way, that is, systems of algebraic equations [118J
have to be solved. The derivation and solution of these systems of equations will be
demonstrated in this section.
Given an uncertain polynomial p(s, q) with q E ]Rt and real polynomials a;(q). The
nominal polynomial p(s, 0) is stable. Find the maximal p such that p(s, q) is stable for
all IIqllp < p. p is called the stability radius and the index p characterizes the type of
norm.
Concerning the choice of the norm, i.e, p, there are three important possibilities. For
p = 00 the set of admissible q describes an (-dimensional hypercube. Dual to this norm
is p = 1, which corresponds to a diamond. p = 2 yields an (-dimensional hypersphere
in Q-space. From the practical point of view, the case p = 00 is the most important
one, because there the bounds for the uncertain parameters are independent. The case
p = 00 will be handled in detail. For the other cases, it is not difficult to derive the
corresponding results.
It was shown in Section 10.2 that the stable set of q is bounded by three hyper-
surfaces, namely ao(q) = 0, an(q) = 0 and ~n-l(q) = 0, see Figure 10.7. The last
equation is the last but one Hurwitz determinant, which is the critical one. It results
from the elimination of w from the two equations Re pOw) = 0 and 1m pOw) = 0 as
was shown in the proof of Theorem 2.1. Here, linear or non-linear dependence make
a big difference. For fixed wand linear dependence Re p = 0, 1m p = 0 represent a
linear manifold; this means, for example, that for ( = 3, ~2(q) = 0 is generated by
the continuous movement of a straight line. In the case of non-linear dependency, a set
of curves in ]R3 generate ~2 = O. Computing ~n-l(q) must , in general, be done by a
computer algebra program .
All three equations of the hypersurfaces will be treated in the same way. The only
difference is that the third one will be the complicated one with respect to the number
of terms and degree of the terms . In the sequel, the notation F(q) = 0 is used for each
of these equations .
Consider first the case of two parameters for which the basic idea is demonstrated.
The polynomial family with IIqlloo ::; , (, sufficiently small) is stable and can be
described by a square of sidelength 2,. Enlarge this square continuously until there is
an intersection point with the curve F( ql , q2) = 0, as demonstrated in Figure 10.7. This
point of first contact with a stability boundary may lie on a vertex or on an edge of the
square (see Figure 10.8). The first situation is characterized by the fact that ql = ~
10.3 Polynomial Dependence 411

q2

Figure 10.7. The stable region is bounded by at most three hypersurfaces

(marked a) or ql = -q2 (marked b), which results in the two polynomial equations

F(ql ,qd = 0 ,
(10.3.1)
F(ql , -ql) = O.
The second case is an intersection point on an edge. This means that F( ql, q2) = 0
has a horizontal (marked c) or a vertical tangent (marked d). This necessary condition
leads to the two systems of equations in two unknowns

OF(ql, q2) = 0
Oql '
(10.3.2)
of(q), q2) = 0
Oq2 .
It may be possible that in the intersection point, the curve F(q), q2) = 0 cannot be
differentiated, i.e, the curve has, for example, a cusp or a singular frequency. For
these points, both partial derivatives vanish and these solutions are already obtained
by (10.3.2).
Finding the real roots of the two polynomials and the real solution vectors of the two
systems of equations gives a set of points (ql, q2) that are candidates for the first contact.
For the first case (vertex contact) 1I (±q, ±q)lIoo = IIqlloo = Iql and for the second case
(edge contact) lI (ql, q2)lIoo = max{lqI/ , Iq21} . The solution vector with the smallest norm
q" = [qj ; q2V yields the stability radius. This critical parameter combination q" also
determines the critical frequency. The polynomial p(s,qj ,q2) = 0 has a root at s = 0
412 10 The Stability Radius

,/
,/
,/
,/
,/

Figure 10.8. The case p = 00

or at s = 00, or a root pair at s = ±jw. All other roots are not in the open right half
plane.
Consider the case of three parameters. A surface F(ql , in, q3) = 0 replaces the curve
F(ql , q2) = O. This surface bounds the stable polynomials and now a cube instead of a
square is enlarged. The intersection points lie either on a vertex, on an edge, or on an
face of the cube. These subsets of the cube are called subpolytopes. Corresponding to
the eight vertices of the cube, the four polynomials

F(q, +q, +q) = 0 ,

F(q, +q, -q) = 0 ,
F(q, -q, +q) = 0 ,
F(q, -q, -q) = 0

describe this situation. In case of the twelve edges, two intersection points must coin-
cide, i.e. the partial derivatives must vanish and the six systems of equations
10.3 Polynomial Dependence 413

F(ql' +q, +q) = 0,

F(qll +q, -q) = 0,

F(+q,q2,+q) = 0, oFI
Oq2 q\=+q,q3=+q -
-0
,

F(-q,q2,+q) = 0, OFI
Oq2 q\=-qm=+q
=0
,

F(+q,+q,q3) = 0,

F(+q, -q,q3) = 0, oFI

Oq3 q\=+qm=-q -
-0

in two unknowns have to be solved.

If the intersection is on one of the six surfaces, then the normal vector of the surface
F(qll q2, q3) = 0 is parallel to one of the coordinate axes, which means that two of the
three partial derivatives vanish simultaneously and the three systems

of of =0 ,
F(ql' q2 , q3) = 0 , Oql = 0 , Oq2
of of =0
F(qll q2, q3) = 0, Oql = 0 , Oq3 '

F(ql ' q2, q3) = 0, of =0 of =0,

Oq2 ' Oq3

in three unknowns characterize this situation.

For i > 3 parameters, the number of polynomials and systems of equations obviously
grows with the number of subpolytopes of an i-dimensional cube. The derivation of
the polynomials and systems of equations for arbitrary i is straightforward. But, now
again occurs what is called the combinatorial explosion. The number of equations grows
exponentially with the number of parameters.
Using other norms (p = 2 or p = 1) gives similar equations. In any case, equations
that are linear combinations of F and their partial derivatives have to be solved. For
p = 2, the number of systems is smaller , but the equations are more complicated. In
that case at the intersection point, the vector q must be parallel to the gradient of F
as indicated in Figure 10.9.
414 10 The Stability Radius

,- ;'

I
I
--l_---
__~~-I__---l_- qi
\
\
' ...

Figure 10.9. The case p =2

Only one system of equations,

8F
F(q)=O , 8q = >..q ,

is necessary for describing this situation.

A restriction for practical applications is the problem to find all solutions of the
system of equations. The theorem of Bezout (see Appendix A) says that, for example,
a system of three equations with the degrees mi, i = 1,2,3 has up to m1 . m2 . m3
solution vectors. So the resulting polynomial in one variable after eliminating the other
two variables may have the degree m1 . m 2 . m3. Therefore, this method is recommended
for up to only three uncertain parameters. The restriction to few parameters seems
to be severe, but a quotation of Lazard [130) in 1991 on software for the solution of
algebraic equations gives hope for the future . "Five years ago, problems with four
or five unknowns were outside of the capabilities of most available softwares. Recent
progresses made or will make accessible problems with six or seven unknowns."

Example 10.9

Consider the track-guided bus of Section 6.8. The transfer function of the uncon-
trolled bus depends on the virtual mass iii and on the velocity v. For the data given in
Table 6.1, it is

_ 6.079 · 105 mv 2s2 + 3.886 . 1011 vs + 4.803. 1010 v2

9(S, m, v) = s3(m2v2s2 + 9.818 . 105 mvs + 1.663 . 104 mv 2 + 2.690 . 1011) .
10.3 Polynomial Dependence 415

In [154), a controller was designed that led to the closed-loop polynomial

s
p(s, m, v) = L aisi ,
i=O

with

ao 453 .106 v 2 ,
al = 528 · 10 v + 3640 . 10 v ,
6 2 6

a2 5.72.10 mv + 113.106 v 2 + 4250 .106 v,

6 2

a3 6.93· 106 mv 2 + 911 . 106 v + 4220 . 106 ,

a4 1.45· 106 mv 2 + 16.8 . 106 mv + 338 . 106 ,
as = 15.6 · 10 m V + 840mv + 1.35 . 10 iii» + 13.5 . 10 ,
3 2 2 2 6 6

a6 = 1.25· 10 m V + 16.8m + 53.9 . 103 mv + 270 . 103 ,

3 2 2 2v2

a7 50 m 2V 2 + 1080mv ,
as m2v2 •
Let the nominal point be v = 20 [ms- I ] and m = 20 [103 kg] . For determining the

m[103 kg] 70

60
~
50

40

30

20 X

10 /w = 16.3
r:
10 20 30 40 50 60 70
v[m·s-IJ
Figure 10.10. The maximal square around the nominal point (20,20)

stability radius, the parameter space method can be applied, i.e, plotting the stability
boundaries in the (v, m)-plane, see Figure 10.10. But, the analytical method of this
section is also applicable. The distances to the real root boundaries m = 0 and v = 0
are trivial, but for the distance to the complex root boundary the Hurwitz determinant,
which is of order seven, has to be calculated. This must be done by a computer algebra
416 10 The Stability Radius

program . Solving (10.3.1) and (10.3.2) with ql = v - 20 and q2 = in - 20 yields an

intersection point of the growing box with the complex root boundary at ql = 14.8,
q2 = -ql for w = 16.3, see Figure 10.10.
o
11 Robustness of Sampled-data Control
Systems

Controllers are usually implemented in a digital computer. Figure 11.1 shows a single-
loop sampled-data control system with cz(z ) representing the a-t ransfer function of th e
digit al cont roller, and (1 - e-·T )/ 8 as th e transfer function of th e hold element , T is
th e sampling interval. Th e problems and solution appr oaches tr eat ed in Chapters 1

F igu re 11.1. Sampled-dat a control sys te m wit h uncertain ph ysical parameters q in the
continuous-time plant

through 10 have their counterparts for sampled-data cont rol systems and Sections 11.1
t o 11.9 are organized accordingly. Before we go into det ails, we should however discuss
an obvious question:
Assume th at a robust continu ous controller was designed for the plant 9. (8, q). Is it
possible to find an approximately equivalent discrete-time controller cz(z) that preserves
robust stability?
Discretization of a continuous-time controller is a topic of research interest. For a
nominal plant 9.(8), a controller discretization procedure was shown in [139] that pre-
serves th e closed-loop stability. The cont roller, however, depends on th e nominal plant
model and th erefore does not meet our requirement of a fixed-gain robust cont roller for
an uncertain plant.
Since robust stability is not guaranteed after controller discretization, there remain
two alternatives:

a) Try a controller discretization procedure that does not involve the plant transfer
function , e.g. the Tustin approximation with

2(z - 1)
(11.0.1)
8 ~ T (z + 1) ,
418 11 Robustness of Sampled-data Control Systems

and do a robust stability analysis for the resulting discrete-time system. For "suf-
ficiently small" sampling intervals T, the approximation is good. Essentially, it re-
places each continuous integrator II s in the controller by a discrete approximation
using the trapezoidal rule for integration with z-transfer function T(z+ 1)/2(z-I).
b) Discretize the plant model and try any design tool to simultaneously stabilize the
plant for some representative operating conditions. Then, do a robust stability
analysis for the continuum of parameter values. If we adopt this multi-model
approach, then design of the discrete-time controller is not more complicated than
the continuous-time design, because numerator and denominator coefficients of
cAz) in Figure 11.1 enter linearly into the closed-loop characteristic polynomial
like in the continuous-time case.
For both approaches, a robust stability analysis of the resulting closed-loop discrete-
time system is required, in case a) because stability cannot be guaranteed after approx-
imate controller discretization , and in case b) because only representative operating
conditions were considered in the design.

11.1 Plant and Controller Discretization

In Chapters 6 and 7, we have seen examples of parametric state-space and transfer-

function models of continuous-time plants. Now, the plant inputs u(t) are generated
by a sampler and hold, i.e.
u(t) = u(iT) , tE[iT;iT+T) , i=O,I,2, .. . (11.1.1)

and the plant model may be discretized to obtain a description of states and signals at
the sampling instants t = iT. The discretization process is derived in many textbooks
on sampled-data control systems, e.g. [37] [87] [4] and the results are only summarized
here.
A state space model
x = A(q)x + B(q)u , (11.1.2)
y = C(q)x,
with input (11.1.1), yields the discretized system
x(iT + T) = Ad(q)x(iT) + Bd(q)u(iT) ,
(11.1.3)
y(iT) = C(q)x(iT) .

Let

J
T

R(q) .- eA(q)vdv , (11.1.4)

o
then
eA(q)T = 1+ R(q)A(q) ,
(11.1.5)
R(q)B(q) .
11.1 Plant and Controller Discretization 419

Each eigenvalue of A at Si is mapped to an eigenvalue of Ad at z, = es,T with the same

multiplicity.
The z-transfer function matrix of the system (11.1.3) with z = esT is

(11.1.6)

and , in the single-input, single-output case,

(11.1.7)

An alternative approach to plant discretization starts from the plant transfer function
9s(S, q) . (9s(S , q) may also be an element of a transfer function matrix.) First , the
a-transfer function for sampler, hold, and plant is determined as

h(
% z,q
)=Z-I Z { 9s (S, q) } , (11.1.8)
Z S

or, equivalently, by the Poisson sum [136] [134]

( sT ,q) = (1 _ e-sT) ~
h%e 9s(S + jm2rr/T, q) (11.1.9)
L.J T · 2
S + Jm rr
.
m=-oo

Example 11.1

1
9.(S, q) = s(1 + s/q) , (11.1.10)
(1 - qTe- qT - e- qT) - (1 - qT - e-qT)z
h%(z, q) = (T) (11.1.11)
q(z - 1) z - e- q
The poles of 9.(S, q) at SI = 0 and S2 = -q are transformed to poles of h%(z, q) at
ZI = e
S1T = 1 and Z2 = eS2T = e- qT. The zero of h%(z, q), however, does not even have
a continuous-time counterpart. In the numerator, it is not possible to introduce only
one new uncertain variable e- qT instead of q, because qT also enters in form of a sum
and a product. Therefore, we have to deal with an exponential parameter dependence
that also enters into the closed-loop characteristic polynomial.
The Poisson form of the z-transfer function is
1- -sT 00 1
h%(esT,q) = ; L
m=-oo
(s + jm2rr/T)2[1 + (s + jm2rr/T)/q] · (11.1.12)

For numerically given q, the form (11.1.11) of the z-transfer function is more conve-
nient. For uncertain q, however, (11.1.12) is a useful form because h%(esT,q) may be
approximated by truncation of the series, see Section 11.8. This approximation is
rational in q. 0
420 11 Robustness of Sampled-data Control Systems

11.2 Discrete-time Controllers

Controllability and observability of a controllable single-input single-output plant get

lost in the discretization process, whenever two distinct eigenvalues 81 f. 82 of the
continuous-time plant map into identical eigenvalues eS t T = eS2T of the discrete-time
plant [122] .

Example 11.2 (Crane)

By (1.1.7), the crane has eigenvalues on the imaginary axis at S3,4

±jJg(mL + mc)/emc. Controllability gets lost for eS3T = eS4T , i.e. n =
k7rJemC/g(mL + me) , k = 1,2 , . ... For T = T l , two distinct eigenvalues of the
continuous system coincide after transformation at z = -1 in the discrete-time model.
A good choice for T is T :::; Ttl4.
In the parameter-dependent case, the sampling interval must be chosen for the worst
case T, = T l- , i.e,
1r e-
T<- (11.2.1)
- 4 g(l+mt/m d '
If we let the minimum rope length t: go to zero then also T goes to zero. Therefore, it
is necessary to specify positive values for e: and me
in order to obtain a non-zero T .
o

Remark 11.3

From a practical point of view, one may want to fix the sampling interval T at an
early stage of analysis and design. In the case of uncertain parameters, the usual rules
for choosing T apply to the operating condition q* with the fastest plant dynamics
or largest bandwidth. Alternatively, T may be treated as an additional undetermined
parameter in robustness analysis. T is not an additional parameter if the system is
scalable [20] . Roughly speaking, a system with uncertain parameters q is scalable if
there is a rescaling of q that changes only the timescale of all solutions to the differential
equations describing the plant . It is then possible to introduce new parameters in a
vector r = r(q, T), where r has the same dimension as q. The number of uncertain
parameters is not increased by allowing an uncertain T . Conceptually, scalability is
important if we want to compare continuous and sampled systems by their respective
stability regions in the same scaled parameter space, where for T -+ 0 identical bounds
are obtained . 0

The general discussion on controller structures remains valid for discrete-time con-
trollers. There are, however, some additional considerations: discrete-time compen-
sators for single-loop systems are assumed with the same numerator and denominator
degree. A non-zero relative degree would introduce undesirable time delays into the
11.3 Eigenvalu e Specifications 421

loop. In continuous-time systems, frequently a strictly proper controller transfer func-

tion is assumed in order to reduce high-frequency disturbances (e.g. measurement noise,
high-frequency model uncertainty). In a sampled-data system, the same effect can be
achieved by an analog anti-aliasing filter before the sampler. This is a low-pass filter
with a bandwidth WB of, at most, half the sampling frequency, i.e. WB ~ 211" IT.
The problem classes of parametric polynomials (interval, affine, multilinear, polyno-
mial) must be augmented by polynomials with an even more complicated exponential
parameter dependence. Such polynomials must be tested for Schur-stability, i.e. all
their roots must be located inside the unit circle in the z-plane.

Example 11.4

Consider the sampled-data feedback loop of Figure 11.1 with 9s(S, q) as in (11.1.10)
and a proportional controller cz(z) = k. The closed-loop characteristic polynomial is
p(z, q, k) ao + alZ + a2z2 ,
ao qe- qT + k(1 - qTe- qT - e- qT ) ,
(11.2.2)
al = -q(1 + e- qT ) - k(1 - qT - e- qT ) ,
a2 q.
The coefficient functions are exponential for the uncertain plant parameter q, they are
affine for the controller parameter k. 0

11.3 Eigenvalue Specifications

The open-loop eigenvalue locations of sampled-data systems depend on the sampling

interval. An extreme case is obtained with T -+ 0 and all eigenvalues z, = eS i T approach-
ing Z = 1. Practically, T is chosen such that the eigenvalues are in some neighborhood
of z = 1. In [4], a rule of thumb is given: T should be chosen such that all open-loop
eigenvalues are located in a unit circle centered at z = .;2.
For closed-loop stability, these eigenvalues must be shifted inside the unit circle.
The eigenvalue locations s, recommended in Chapter 3 may be directly mapped into
the a-plane via z, = eS i T • The type of discrete-time solutions is quite similar to the
corresponding continuous-time solutions as long as Re z, > O. The mapping of lines of
constant damping and constant natural frequency from the s-plane to the z-plane is
illustrated by Figure 11.2.
A special feature of sampled-data systems is the existence of deadbeat solutions.
They are achieved with all eigenvalues located at Z = O. Practically, this requires
relatively long sampling intervals T in order to avoid excessive magnitudes of the plant
inputs u. Deadbeat control is an ideal case achievable with finite controller gains. We
can come close, the ideal case by shifting the eigenvalues close to the origin.
Another special case in sampled-data systems occurs for an eigenvalue at z = -l.
The corresponding solution term at the plant input is shown in Figure 11.3. Consider
422 11 Robustness of Sampled-data Control Systems

s-plane jw
0.35
~-.. j1r/2T

jIm Z (J

j r--_wnT
_ = 75°

z-plane

0.35

0.5

D= 1/.../2

o 0.46 1 Re z

Figure 11.2. Mapping of constant damping (D) lines and constant natural frequency (wn )
lines from the s-plane to the plane z = esT

the situation for a sampling interval T that is long compared to the settling time of the
plant step response. Then, the response has approximately reached its stationary value
before the next sampling occurs, and the plant response to the input signal shown in
Figure 11.3 consists of consecutive full length step responses. This type of solution and
its neighborhood is obviously very undesirable. It is an indication either of a too-long
sampling interval compared to the plant dynamics, or of a bad design that has placed
closed-loop eigenvalues too close to z = -1. Closed-loop poles on the negative real axis
at z, E (-1 ; 0) give rise to decaying solution terms with alternating sign of u(iT). Such
excessive actuator activity is inefficient and should be avoided.
In later sections, we will formally treat the real root stability boundaries at z = 1
11.3 Eigenvalue Specifications 423

0 T 2T 3T 4T t

Figure 11.3. Solution term corresponding to an eigenvalue at z = -1

and z = -1. We should keep in mind that a good practical solution cannot be close
to the z = -1 boundary. Closeness to the z = 1 boundary may be tolerable if the
sampling interval T is short. For example, an eigenvalue at z = 0.98 yields a solution
term 0.98i that decays only by 2% from sampling instant to sampling instant. For small
T, this may be a rapid exponential decay in real time t.
We will now move towards a definition of a useful r-stability region in the z-plane
that takes into account the above discussion and is also computationally tractable.
Consider the r -stability regions shown in Figure 11.4.

0.35 o
jw j77
D=O

s-plane

z-plane

Figure 11.4. Mapping of r -stability regions from s-plane to z-plane

Constant damping lines in the s-plane map into logarithmic spirals in the z-plane,
constant real part lines in the s-plane map into circles centered at z = O. Three
424 11 Robustness of Sampled-data Control Systems

examples with increasing damping (0.35, 0.5, 0.707) and decreasing real part boundary
are mapped via z = esT into the corresponding boundaries in the z-plane.
In the s-plane, we have smoothed the piecewise defined boundary by a hyperbola, see
Figure 3.8, because a hyperbola is reasonably well tractable as a f -stability boundary.
In the z-plane, we may smooth the image by an even simpler boundary that is a circle.
Let z = T + j11 and consider a circle

(T - TO? + 112 = r
2
(11.3.1)
The circle has center TO and radius r . Now let

r for 0 :::; r :::; 0.5,

TO = 1- r for 0.5 :::; r :::; 1, (11.3.2)
{
o for 1 :::; r .

Figure 11.5 shows some circles for r = 1, 0.8, 0.6, 0.5, 0.44, 0.33. Given that we are

1
T

z-plane

Figure 11.5. Approximation of r-stability regions of Figure 10.4 by circles

talking about a rule of thumb, the Ivstebility regions are reasonably well approximated
with the following correspondence:
D r TO
0 1 0
0.35 0.5 0.5
0.5 0.44 0.44
0.707 0.33 0.33
11.4 Classical Stability Tests 425

Conceptually, a design process may be viewed as follows. By choice of the sampling

interval T , the open-loop eigenvalues are located in a unit circle centered at v'2. Thus,
a circle with radius r = 1 + v'2 ~ 2.4 includes these eigenvalues for all operating
conditions q E Q. The open loop is (r = 2.4)-stable. In the design process, we move
away from the open loop k = 0 in k-space. The design direction should be such that r
is reduced. For r = 1, Schur-stability is achieved, for r = 0.6 a solution term (-0 .2)i is
still admitted but no alternating terms that decay slower. For r = 0.5, all alternating
terms are ruled out but the deadbeat solution is still admitted. A further improvement
in the degree of stability is now achieved by pushing eigenvalues in the right half z-plane
closer to the origin . The ideal deadbeat solution cannot be achieved simultaneously for
a plant family, but it is a well-posed question to ask for the smallest radius r min that
can be achieved simultaneously by a controller of fixed structure.

Remark 11.5

If a strict contraction is desired, then the relationship between 70 and r may be

changed to
70(1 - 70) = exr(l - r) , (11.3.3)
with ex < 1 instead of ex = 1 in (11.3.2). o

11.4 Classical Stability Tests

The approaches for testing Hurwitz-stability of a polynomial family can easily be mod-
ified for testing Schur-stability of polynomial families:
n
P(z, q) = {p(z, q) = L ai(q)zi I q E Q j . (11.4.1)
i=O

The root set calculation by gridding all parameter intervals is the same as for cont inuous
time, only the interpretation is different because we are now interested in the root
location relative to the unit circle. If we rely on a gridding approach for robustness
analysis, then also the discretization (11.1.3) may be performed numerically on a q-
grid .

Boundary Crossing
The boundary crossing idea for discrete-time systems was formulated by Jury and
Pavlidis [117]. There are three boundaries:

i) A real root boundary at z = 1, i.e.

p(l , q) = o. (11.4.2)
426 11 Robustness of Sampled-data Control Systems

ii) A real root boundary at Z = -1 , i.e.

p(-I ,q)=O . (11.4.3)

iii) A complex root boundary for a root on the upper half of the unit circle, i.e.
p(eiwT , q) = 0 , wT E [0 ; 1l"1. (llAA)
The polynomial with real coefficients then also has a symmetric root on the lower
half of the unit circle.
Algebraic Problem Formulation
The algebraic problem formulation corresponding to Theorem 2.1 (Frazer, Duncan) is:

Theorem 11.6 (Jury, Pavlidis)

The polynomial family P(z , Q) = {p(z, q) Iq E Q} with p(z,q) continuous is
Schur-stable, if and only if:
i. There exists a qO E Q such that the polynomial Po(z) = p(z,qo) is Schur-
stable.
ii. p(l , q) =I 0, for all q E Q.
iii. p( -1 , q) =I 0, for all q E Q.
iv. det S(q) =I 0, for all q E Q, where S(q) = X(q) - Y(q) and (omitting
the dependency on q)

an an-I an-2 a2 0 0 0 no
0 an an_1 a3
X= 0 0 an a4 Y= 0 0 ao an-4
0 no al a n-3
0 0 0 an no al a2 an-2
(11.4.5)
0

Proof
In [117], the following counterpart to Orlando's formula was shown ,
n-I n

det S = a~-1 II II (1 - ZiZk) , (11.4.6)

;=1 k=i+l

where Zi , Zk are roots of p(z). Consider now a continuous variation of q starting

with qO , i.e. a stable polynomial. If a conjugate pair of roots Zi, Zk crosses the
unit circle, then 1 - ZiZk = 0 and det S = O. Th e only other possibilities for the
polynomial to become unstable are that a real root crosses the unit circle at Z = 1
(Condition 2) or Z = -1 (Condition 3). Then, Theorem 11.6 follows from the
boundary crossing Theorem 4.1.
o
11.4 Classical Stability Tests 427

Remark 11.7

Complete tests for a polynomial to have all its roots inside the unit circle have been
given by Schur [176] , [177] and Cohn [67] . A simplified formulation for polynomials with
real coefficients was given by Jury 1116]. Since we always assume that we know a stable
polynomial to start with and use the boundary crossing theorem, we do not need the
complete set of algebraic inequalities . Here we only give the inequalities corresponding
to Conditions 2, 3 and 4 of Theorem 11.6.
p(I) > 0,
(_l)n p( -1) > 0, (11.4.7)
detS > 0.
o

Example 11.8

Consider the characteristic polynomial (11.2.2). Let T = 1, Q = {q, k Iq E

[0.1 ; 2] , k E [0.1; 2]}. Is this polynomial family Schur-stahle?

p(z, q, k) ao + alz + a2z2 ,

ao qe- q + k(I - qe:" - e- q ) ,
al -q(I + e- q ) - k(I - q - e- q ) ,
a2 = q .
i. Let q = 1, k = 1, then Pz(z, 1, 1) = (1 - e- l) - z + z2 has roots inside the unit
circle. The three stability boundaries are given by:
ii. p(I , q, k) = kq(I - e- q ) =f o.
iii. p( -1, q, k) = 2q(I + e- q ) + k[2(1 - e- q ) - q(I + e- q ) ] =f 0.
iv. X = a2, Y = ao, S = a2 - ao = q(l - e- q ) - k(I - qe- q - e- q ) =f 0.
°
Condition 2 yields the boundaries q = and k = 0, which are outside Q. Conditions 3
and 4 may be checked graphically by plotting the stability boundaries in the (q, k)-plane .
From Condition 3 follows the stability boundary at z = -1, with

k 1 = 2q(I + e- q )
- q(I + e- q ) - 2(1 - e- q )

From Condition 4 follows the complex root boundary at z = eJ wT , with

q(I - e- q )
kc = 1- qe- q - e- q
, lim k;
q-O
=2.
The stability boundaries k _ l and k; are plotted in Figure 11.6 together with the bound-
aries k = 0 and q = 0 resulting from Condition 2. Figure 11.6 not only provides a test
for the given Q-box, but the entire stable region is exhibited. For comparison, consider

qk + qs + 8 2, it is stable for q > °,

the continuous-time system (11.1.10) with feedback k. Its characteristic polynomial is
k > 0, i.e. in the entire first quadrant in Figure 11.6.
428 11 Robustness of Sampled-data Control Systems

5
k
4

D
2 QSchur

0 q
0 1 2 3 4 5

Figure 11.6. The stability boundaries do not intersect Q, and Q contains a stable point at
q = 1, k = 1. Therefore, the polynomial family is robustly Schur-stable

The situation that sampling reduces stability regions is typical, however, it is possi-
ble to construct (somewhat exotic) examples for which sampling stabilizes [201. Note
that in the example the loop-gain k enters affinely into the characteristic polynomial.
Therefore, it was easy to solve explicitly for the stability boundary k(q). In general, it
is not possible to test Conditions 3 and 4 of Theorem 11.6 without gridding q . 0

An alternative algebraic formulation of stability criteria uses a bilinear transforma-

tion of the polynomial p(z) by
z-1 l+w
w ·_-- z- -l-w
-- · (11.4.8)
.- z+ 1 '
It maps the unit disk of the z-plane into the left half w-plane. Substituting z
(1 + w)j(1 - w) into p(z) of (11.4.1), a new polynomial Pw(w) is obtained as

Pw(w) = (I_W t p(l+W)

l-w

L ai(1 + w)i(l- wt-i

i=O

(11.4.9)

p(z) is a Schur polynomial if and only if Pw(w) is a Hurwitz polynomial. Thus , results
of Chapter 2 can be adapted to the discrete-time case. Of course, an exponential
11.4 Classical Stability Tests 429

dependen cy of ai(q) is inherit ed by bi(q) and the robustness analysis in terms of physical
plant parameters q remains a difficult problem.

Remark 11.9

In addition to , the two cases of ai-intervals of p(z) and bi-intervals of Pw(w) there is
a further alternative by the 8-transform [144], where
z-l
8=---y;- .

The transformed polynomial is

n n
p<5(8) = p(8T + 1) = Lai(8T + l) i = LCi~.
i=O i=O

The polynomial p(z) is Schur-stable if and only if p<5(8) has all its roots inside a circle
centered at 8 = -liT and passing through 8 = O. 0

Example 11.10

Consider again Example 11.8. The transformed polynomial is

bo + bl W + b2w 2 ,
ao + al + a2 = kq(l - e- q ) ,
2(a2 + ao) = 2[q(1 - e- q ) - k(l - qe- q - e- q ) ] ,
ao - al + a2 = 2q(1 + e- q ) + k[2(1 - e- q ) - q(l + e- q ) . ]
The boundaries for Hurwitz-stability, i.e. bo = 0, bl = 0, bz = 0 are identical to the
Schur-stability bound aries of Example 11.8 with the result plotted in Figure 11.6.
o

The correspondence between the critical stability conditions of p(z, q) and Pw(w, q)
is the following:
p(l, q) o ¢:=} bo(q) 0,
p( -1, q) o ¢:=} bn(q) 0, (11.4.10)
det S(q) o ¢:=} det Hn -l(q) O.
The bilinear transformation also provides a set of relatively simple necessary stability
conditions :
bi(q) > 0 , i = O, l, . .. , n. (11.4.11)
The necessary condition (11.4.11), interpreted in terms of the original coefficients ai, is
closely related to the following result on th e convex hull of th e stable region in coefficient
space [84] .
430 11 Robustness of Sampled-data Control Systems

Theorem 11.11 (Fam, Meditch)

Consider the stability region of

(11.4.12)

in the space of coefficients ao, aI, . . . , an-I ' Its convex hull is a polytope with n+ 1
vertices corresponding to the polynomials

Pi(Z) = (z + l) i(z - l)"-i, i = 0,1, ... ,n (11.4.13)

It was shown in [4] that the above convex hull is formed by hyperplanes representing
the necessary stability conditions bi(ao , aj, . .. , a n- I) > 0, i = 0, 1, . . . , n .

Example 11.12

Necessary conditions for Schur-stability of a given polynomial are

The corresponding vertex polynomials are

Po (1 - W)4 = 1 - 4w + 6w2 - 4w3 + w 4 ,

PI = (1-w) 3(1+w)=1-2w+2w 3-w4 ,

P2 (1-w)2(1+w)2=1-2w 2+w4,
P3 + w? = 1 + 2w - 2w3 - w4 ,
(1 - w)(l

P4 = (1 + W)4 = 1 + 4w + 6w2 + 4w3 + w4 .

The transformed polynomial is

4 4

Pw(w) = LaiPi(W) = Lbiwi.

i=O i=O

Necessary conditions for Schur-stability of p(z) are

bo ao + al + a2 + a3 + 1 > 0,
bI 2(-2ao - al - a3 + 2) > 0,
~ 2(3ao - a2 + 3) > 0,
b3 = 2(-2ao + al - a3 + 2) > 0,
b4 = ao - al + a2 - a3 + 1 > O.
11.4 Classical Stability Tests 431

These five linear inequalities correspond to five three-dimensional hyperplanes in the

four-dimensional space with coordinates 0.0, al, a2 , a3' The hyperplanes form the convex
hull of the stability region.
The five inequalities can be combined as

lal + a31- (1 + 0.0) < a2 < 3(1 + 0.0) ,

lal - a31 < 2(1 - 0.0) .

o
Necessary stability conditions have only limited value in robustness analysis, be-
cause we have to be pessimistic in analysis. For design, however, (q replaced by the
undetermined controller coefficients k), necessary stability conditions are useful because
they yield a finite region in k-space where possible solutions must be located .
Frequency Domain Problem Formulation
In frequency domain problem formulations, we have to substitute s = jw, i.e, z = eiwT .
Then , an initially stable polynomial p(z , q) will become unstable by variation of q
whenever we have simultaneously
Re p(eiwT , q) = 0,
(11.4.14)
1m p(eiwT , q) = 0
for some wT E [0; Jr] . For wT = 0 or wT = n , the second condition of (11.4.14) is always
satisfied and there remains only the first condition p(1, q) = 0 or p( -1 , q) = O.
Parameter Space Approach
In the parameter space approach for two uncertain parameters q = [ql q2] , we try to
eliminate ql or q2 from the two equations (11.4.14). This was possible up to polynomial
coefficient functions as shown in Chapter 4. This approach is no more feasible if ql and
qz enter exponentially.
The parameter space approach will be used as a design tool in k-space. For single-
input plants , the controller can always be assumed in a form such that the undetermined
controller parameters in k enter affinely into the coefficients of p(z, k). In the multi-
input case, the controller .structure can be assumed such that the coefficient functions
are multilinear. For such design considerations, it is useful to have some idea of the
shape of stability regions in the space of coefficients ai ' We will treat the examples of
polynomials with degree n = 2 and n = 3.

Example 11.13

p(z) = 0.0 + alZ + z2

The polynomial is .Schur-stable for 0.0 = 0, al = O. Find the stable neighborhood of
this point in the coefficient plane.
By (11.4.2) and (11.4.3), the real root boundaries result from
p(l) = ao+al+l 0,
p(-l) = o.o-al+l = O.
432 11 Robustness of Sampled-data Control Systems

These two straight lines are plotted in the (at, a2)-plane of Figure 11.7. By (110404) ,
the complex root boundary is

It is more convenient, however, to parameterize the unit circle in the plane z = 7+j7J by
the real part 7 instead of the phase angle wT. On the unit circle, 7 and 7J are connected
by 7 2 + 7J2 = 1, 7 E [-1; 1], and a polynomial with a pair of conjugate roots on the
unit circle has the form

(z - 7 - j7J)(z - 7 + j7J) = z2 - 2TZ + 1, 7 E [-1; 1] .

By comparison with the second order polynomial, the complex root boundary is ob-
tained as
ao = 1, al = -27, 7 E [-1; 1].

This line segment c is also shown in Figure 11.7. Point 0 in Figure 11.7 is common to

P (-1)=0

D
o
P (1)=0

Figure 11.7. Partition of the polynomial coefficient plane for n = 2

p(&WT) = 0 and p(1)=0. A double root arises here for z = 1, i.e.

Po(z) = (z-I)2 = z2-2z+1 ,
and correspondingly,

(z-I)(z+l) Z2 -1 ,
(z + 1)2 z2 + 2z + 1 .

The two straight lines p(l) =0 and p( -1) =0, and the complex root boundary p(&wT) =0
partition the coefficient plane into the following regions (EV = eigenvalue, DC = unit
circle):

A: Both EV's in the DC.

B: One EV in the DC, one to the left.
C: One EV to the left, one to the right of the DC.
11.4 Classical Stability Tests 433

D: One EV in the DC, one to the right.

E: Both EV's outside the DC, either as a complex pair or both to the left, or both to
the right . (Note that these cases can be transferred into each other by continuous
motion of the eigenvalues across the branching points of a root locus without
crossing the unit circle.)

We are mainly interested in the stability region A, which is completely determined by

the vertices of the triangle 012. 0

Example 11.14

The polynomial

has real root boundaries

p(l) = ao + at + a2 + 1 = 0 ,
p(-1) ao - at + a2 - 1 = 0 .

In the spa ce of coefficients ao, all a2 the linear equation p(l) = 0 describes a plane;
in Figure 11.8 it is the plane containing the points 0, 1 and 2. The plane p(-1) = 0
contains the points 1, 2 and 3.

,--------------- a,
I,
I, I \
I , I \
I ,
I , I " \
I ' I \
I ' I \
I - r----~
I I I
I I I
I I
I I
I I
I I
I I
I I
I I
-""' \ I
\ I
\ I
, I

" , II
\I
------_.:~
1

Figure 11.8. Stability region in the polynomial coefficient space for n = 3

The polynomial p(z) = (z2 - 2rz + l)(z + r) = z 3 + (r - 2r)z2 + (1- 2rr)z + r has a
conjugate pair of roots on the unit circle for an arbitrary real rand r E [-1 ; 11. Thus ,
434 11 Robustness of Sampled-data Control Systems

the complex root boundary is

ao = r,
al = 1 - 2r7 , 7 E [-1; 1],
a2 = r - 27 .
The boundary is bilinear in r and 7, i.e, for constant r it is a family of straight line
segments beginning with 7 = -Ion edge 23 and ending with 7 = 1 on edge 01. For
constant 7 E [-1; IJ it is a family of straight lines through the edges 02 and 13. The
relevant part of the stability boundary is the line segment for r E [-1 ; IJ.
The corners of the stability region correspond to the four polynomials with zeros in
{-I, I}, i.e.
Po(z) (z-I)3 -1+3z-3z2+z3,
Pl(Z) (z+I)(z-I)2 l-z-z 2+z3 ,
P2(z) = (z+I?(z-l) -1-z+ z 2+z3,
P3(z) = (z+I)3 = 1+3z+3z2+z 3.
If the coefficients move along an edge from the corner i to the corner i + 1, then a real
root migrates from z = 1 to z = -1 and the other roots remain at z = -lor z = 1. If
the coefficients move along an edge from corner i to the corner i + 2, then a conjugate
pair moves along the unit circle from z = 1 to z = -1 and the remaining real eigenvalue
is located at z = 1 or z = -1.
The complex boundary is the saddle surface visualized in Figure 11.8. It is a ruled
surface that can easily be plotted by subdividing edge 02 and edge 13 into the same
number of equal segments and connecting the corresponding points. The same proce-
dure applies to edges 01 and 23. The stable region is the one that contains the origin ,
in Figure 11.8, below the saddle surface. Theorem 11.11 says for this example that the
tetrahedron 0123 is the convex hull of the stability region.
We will refer to Figure 11.8 several times in order to visualize various concepts for
the case n = 3. 0

If the polynomial coefficients are functions of uncertain parameters q , then the

boundary representation Theorem 4.1 provides a systematic way of generating the equa-
tions for mapping stability boundaries into the space of parameters q.
Let us apply Theorem 4.1 to the unit circle

do d 1 d2 ••• d; ] a(q) = [ 0 ] . (11.4.15)

[ o do d1 •.• dn - 1 0
This general expression may be simplified [23J for the unit circle parameterized by the
real part of z = 7 + j7], i.e. 7 E [-1; IJ, 7 2 + 7]2 = 1. Then
do 1,
d1 27 ,
d;+l 27 d; - d;-l ' i = 1,2, ... , n - 1.
Premultiplying (11.4.15) by
(11.4.16)
11.4 Classical Stability Tests 435

yields
0 do
d 1 d2 ••• dn-1] a(q) b [ 0 ] . (11.4.17)
[ -do
0 do d 1 ... dn - 2 0
The term d« with the highest power of 7 has been removed. This reduction procedure
may be continued n/2 times for n even and (n + 1)/2 times for n odd. The resulting
equations are for n even:
-d~_2 -do 0 do d1
d!! ] [ 0] (11.4.18)
[ -d!!-l -d1 -do 0 do d~~l a(q) = 0 '
2

and for n odd:

-d!!±!_2 . .. -do do d1
0 . .. d!!±!_l] [ 0] .
2 2 a(q) = (11.4.19)
[ -d!!±!_l . . . -d1
2
-do 0 do d!!±!_2
2
0
In (11.4.18) and (11.4.19) , the highest degree polynomials in 7 have been reduced to
about half the degree arising in (11.4.15).

Example 11.15

Let
p(z, ql, q2) = (-0.825 + 0.225ql + 0.lq2) + (0.895 + 0.025ql + 0.09q2)Z
+( -2.475 + 0.675ql + 0.3q2)z2 + z3 .
The real root boundaries p(1, ql, q2) = 0 and p(-1, ql, q2) = 0 yield two straight lines
in the (ql, q2)-plane of Figure 11.9. The complex root boundary is described by

[=~ _~ ~ ~] a(q) [~]

[-=-2~ ~1 ~ 2;]a(q) = [~] .

For the example

ao(q) ] [ -0.825 ] [ 0.225

a( ) = al(q) = 0.895 + 0.025
q a2(q) -2.475 0.675
[
1 1 0
the resulting equations

-1.65 + 27] [0.45 0. 2 ] [ ql] [ 0]

[ 0.105 + 1.657 + -0.025 - 0.457 -0.09 - 0.27 q2 = 0
can be solved for
ql ] _ [ 3.592 ] [-5.070 ] [ -11.268] 2 [ . 1
[ q2 - 0.169 + 1.408 7+ 25.352 7, 7 E -1 ,1 .
The complex boundary is also plotted in Figure 11.9. This figure is a cross-section
through an affine image of the stability region of Figure 11.8. 0
436 11 Robustness of Sampled-data Control Systems

Figure 11.9. The vertices of the operating domain Q are stable, but Q is not robustly stable

For n 2: 4, we cannot visualize the stability region but some conclusions are obvious:

i. The stability region is finite and simply connected. It is contractable to the origin
of the coefficient space corresponding to the polynomial p(z) = z" .
n. The stability region is bounded by two hyperplanes corresponding to p(l) = 0
and p( -1) = 0 and by a complex root boundary surface .
iii. The stability region for n 2: 3 is non-convex.
iv. The convex hull of the stability region is a polytope whose vertices correspond to
the n + 1 polynomials

Pi(Z) = (z + l)i(z - lr- i , i = 0, 1, ... , n;

see Theorem 11.11.

v. The edges with a vertex number difference of 1 or 2 are part of the stability
boundary. Edges with a vertex number difference m > 2 are outside of the
stability region. A motion along these edges corresponds to a motion of a complex
conjugate root pair along the root locus of (z-1)ffi+K(z+ l )?' = 0, which consists
of circles passing through z = 1 and z = -1 whose tangents intersect at 1800 1m .
11.5 Edge Test 437

Remark 11.16

A sufficient stability condition is due to Cohn [67] . The polynomial

is stable if
n-l

Lla;1 < 1. (11.4.20)

;= 0
For n = 3, this is the octahedron shown in Figure 11.10. The point A with coordinates

Figure 11.10. Sufficient stability condition laol + lall + la21 < 1 for n = 3
[0 1 0] is the saddle point, see Figure 11.8. The point B with coordinates [0 -1 0] is the
center of edge 12 of Figure 11.8. The illustration for n = 3 shows the conservativeness
of the sufficient condition. 0

11.5 Edge Test

The proof of the edge theorem [43] applies to the unit circle as well. Thus, the crucial
condition is that the polynomial coefficients depend affinely on the uncertain parameters
q. For this case, it suffices to check the edges of the Q-box. If this testing set is Schur-
stable, then the entire Q-box is Schur-stable.
The following Schur-stability test for a single edge was given in [14].
438 11 Robustness of Sampled-data Control Systems

Theorem 11.17 (Ackermann, Barmish)

Let Sb and Se be the stability testing matrices (11.4.5) of two polynomials

Pb(Z) = bo + biz + + bnzn ,

Pe(z) = CO + CIZ + + c"zn,
respectively. The convex combination of the two polynomials

p(z, q) = (1 - q)Pb(Z) + qPe(z) , q E [0; 1] , (11.5.1)

is Schur-stahle, if and only if:

i. Pb(Z) is Schur-stahle.
ii. sign Pb(l) = sign Pe(1) .
iii. sign Pb(-I) = sign Pe(-I) .
iv. The matrix S"b l Se has no non-positive real eigenvalues.

o
The proof follows closely the proof of Theorem 8.20.

Example 11.18

Consider Example 11.15 with the uncertainty rectangle Q = {q I q1 E [1; 5], q2 E

[1; 2]}. Figure 11.9 shows the rectangle (and the stability boundaries) .
The four extreme polynomials are
P1(Z) = p(z ,5 ,2) , P3(z) = p(z ,5 ,1) ,
P2(z) = p(z, 1,2) , P4(Z) = p(z , 1, 1) ,
and it is easily verified that their zeros are inside the unit circle. The four matrices
Si = S(Pi) are

S [1 0.8]
3 = -0.4 -0.11 '

1 -0.8]
S2 = [ 0.4 -0.1 ' S4 = [ 1 -1]
0.5 -0.Ql '

Only two of the Si must be inverted to obtain SiSj1 for the four edges.

i. Edge P1P2
S s:' 1 [-0.5 1.8]
I 2 = 0.22 0.13 -0.6
with eigenvalues Al = -0.289 and A2 = -4.711.
11.6 Construction of Value Sets 439

ii. Edge P4P2

8 8-) 1 [ 0.3 -0.2 ]
4 2 = 0.22 -0.046 -0.399

with eigenvalues A) = 2.079 and A2 = 1.098.

iii. Edge P)Pa

8 8-) 1 [ 0.29 0.2]

) a = 0.21 -0.025 0.2

with eigenvalues Al,2 = 1.167 ± j 0.260.

iv. Edge P4Pa

-1 1 [-0.51 -18]
8 48a = 0.21 -0.059 -O .:U

with eigenvalues A) = -0.621 and A2 = -3.760.

The edges P4P2 and PIPa do not yield negative real eigenvalues, i.e. they are Schur-
stable. Edges PIP2 and P4Pa, however, have negative real eigenvalues. Thus, the Q-box
is not robustly Schur-stable. 0

Remark 11.19

For small sampling intervals T, the control system of Figure 11.1 with cz(z) = k
behaves very similarly to its continuous-time counterpart without sample and hold ele-
ment. Therefore, an interesting question is: if the continuous system has a characteristic
polynomial with affine coefficient functions such that an edge result holds, then it can
be conjectured that an edge result also holds for the sampled-data system with suffi-
ciently small T . Indeed, this conjecture was proven in [112] . It is still an open question
whether robust sampled-data stability for affine continuous plants can be deduced from
edges. 0

11.6 Construction of Value Sets

Presently, there are no results available on tree-structures in characteristic polynomials

of sampled-data systems. In [112], an example with affinecontinuous plant is considered.
The value set p( ei wT , q) of the sampled system is bounded not only by images of the
edges of Q, but also by images of interior points of Q. Thus, the construction of value
sets for sampled-data systems is a difficult problem.
440 11 Robustness of Sampled-data Control Systems

11. 7 Real Radius of Stability

The problem of finding the real radius of stability of a sampled-data control system in
the space of plant parameters is extremely difficult. For controller parameters entering
affinely into the coefficients, the results of Chapters 10 can be used directly. Chapter 4
shows the modification for general r -stability regions, which can be applied for the unit
circle by using the simplified boundary representations in (11.4.18) or (11.4.19).

11.8 Single-loop Feedback Structures

In Chapter 5, we have discussed the use of a Nyquist value set for continuous-time
plants. The zero exclusion from the value set of the characteristic polynomial is thereby
replaced by the exclusion of the critical point -1 from the Nyquist value set . We will
now translate this result to the discrete-time case.
Consider the single-loop sampled-data control system of Figure 11.1 with the open-
loop frequency response

(11.8.1)

The discrete-time frequency response has the following properties:

i. It is periodic in wT with period 271" .

ii. It is symmetric with respect to the real axis, i.e.

Re ho(&wT , q) = Reho(e- jwT , q),

(11.8.2)
1m ho(eiwT , q) = -1m ho(e- jwT , q) .

iii. For wT = 71",

hO(ei2" ,q) = ho(-I ,q) (11.8.3)

is real, i.e.
ho(- I ,q) =-1 (11.8.4)
yields the real root boundary for z = -1 .
iv.
hz(l, q) = g.(O, q) (11.8.5)
(see [20]) . Thus,
11.0(1, q) = cz(l)hz(l , q) = cz(l)g.(O, q). (11.8.6)

From 1. and 2., it follows that it suffices to sweep wT from zero to 71" .
The four conditions of Theorem 11.6 have their counterparts in terms of the open-
loop frequency response as summarized in:
11.8 Single-loop Feedback Structures 441

Theorem 11.20

A single-loop sampled-data feedback system with open-loop discrete frequency

response ho(eiwT , q) is Schur-stable for all q E Q, if and only if:
i. There exists a qO E Q such that the loop is Schur-stahle.
ii. 11.0(1, q) i- -1 for all q E Q.
iii.ho(-I,q) i- -1 for all q E Q .
iv. 11.0 (ei wT , q) i- -1 for all q E Q. and all w E [0; 7T]

Condition 1 is easily tested for an arbitrarily picked qO E Q. Condition 2 is relatively

simple; in view of (11.8.5), q does not enter exponentially. Conditions 3 and 4, however,
contain exponential terms in q and are therefore difficult to test unless we are willing
to grid Q.
An alternate formulation of Conditions 3 and 4 is given by the Poisson sum form
of the z-transfer function, see (11.1.9). For z = -1, i.e. wT = 7T, this sum may be
simplified as follows:

9sU(1 + 2m)7TIT,q]
L:
oo
hz(-I,q) = 2 _
()
j 1 + 2m 7T
m--oo
2 ~ 9sU(1 + 2m)7T IT, q] 2 ~ 9s[-j(1 + 2i)7T IT, q]
~ j(1 + 2m)7T + f.;o' -j(1 + 2i)7T
(11.8.7)
2 ~ (9sU(1 + 2m)7TIT, q] 9s[-j(1 + 2m)7TIT, q])
~ j(1 + 2m)7T + -j(1 + 2m)7T
= 4fm=O
Re 9sU(1+2m)7TIT,q].
j(1 + 2m)7T

Thus, Theorem 11.20 may be reformulated in terms of the continuous-time transfer

function 9s(8, q).

Theorem 11.21

The single-loop sampled-data feedback system with continuous plant 9s(8, q),
sampler, hold and discrete compensator cAz) (see Figure 11.1) is stable for all
q E Q, if and only if:
i. There exists a qO E Q such that the loop is Schur-stable.
ii . cz(1)9s(0, q) i- -1 for all q E Q .
iii.
c, ( - 1) 4 L...
~ Re 9sU(1+2m)7TIT,q]-i._1J:
.( 2) r:
II Q
lor a q E . (11.8.8)
m=O Jl+m7T
442 11 Robustness of Sampled-data Control Systems

iv. For all q E Q and all w E [0; 7l"],

c%(eiwT)(I_ eiwT) ~ 9.(8 + jm.27l"/T, q) # -l. (11.8.9)

L.J
m=-oo
sT + Jm27l"

o
The advantage ofthe formulation of Theorem 11.21 is that for a proper plant 9As, q),
the infinite series can be calculated with any desired accuracy by a finite number of
terms and q enters into each term only with the same complexity as in the continuous-
time case.
Many plants 9.(S) have a low-pass characteristic with relative degree larger than
one. The convergence of the series (11.8.8) and (11.8.9) is enhanced by the l/s term
from the hold element. In this case, the sum converges rapidly and the dominant effect
of sampling is described by the term m = 0, i.e. for the
real root boundary at z = -1,

(11.8.10)

complex root boundary at z = &wT,

. 1- e- jwT
h%(dWT, q) ~ . T 9.(jW, q) . (11.8.11)
JW
Note that the factor (1 - e-jwT)fjwT is the frequency response of the hold element
divided by T . It may also be written as
1 - e- jwT eiwT/2 _ e- jwT/2 -jwT/2
jwT = jwT . e
(11.8.12)
sinwT 2. -jwT/2
= wT 2 e
For the calculation of the complex root boundary, the essential effect comes from the
term e- jwT/2, i.e, a phase lag equal to that of a dead time of half a sampling interval.
The gain factor is one for wT = 0 and decreases with increasing wT .
The approximation (11.8.11) is very good for small sampling intervals T, i.e. large
frequency intervals 27l" IT at which values of the frequency response 9.(jw, q)fjw are
taken in the sum (11.8.9). With increasing T, the quality of approximation deterio-
rates . Then, however, at some sampling interval the type of oscillation illustrated by
Figure 11.3 takes over, which is indicated by Condition (11.8.10). For such sampling
intervals, the approximation is good, again because it includes a second term of (11.8.9).

Example 11.22

Consider Example 11.1 with T = 1 again. The approximate boundary equations

are for z = -1 :
k
4Re 2 =-1,
(j7l") (1 + j7l"/q)
11.8 Single-loop Feedback Structures 443

and for z = ei wT from (11.1.12):

-(1 - e-jW)k
(w + m21T)2[1 + j(w + m21T)/q] = -l.
The approximate boundaries (dotted) are shown in Figure 11.11; the exact boundaries
(solid lines) are from Figure 11.6. The approximation is good for small q. Note that
a variable T would appear in a term qT, thus, small q has an effect like small T . For
larger q, the approximation of the complex root boundary k c for z = ei wT gets worse.
A good approximation is then achieved for larger q, where the stability boundary k_ 1
for z = -1 takes over. 0

5
k
4

1 Q

0 q
0 1 2 3 4 5 6 7 8 9 10

Figure 11.11. Approximation of Figure 10.6 by truncated Poisson series

Example 11.23

For the automatic bus steering Example 4.20 and (6.10.13), the compensator
2
C(8) = 253 0.15s + 0.78 + 0.6
(82 + 258 + 252)(s + 25)
was discretized by the TUstin approximation (11.0.1) with a sampling interval
T = 10 [ms], yielding

C (z) = 9.3464 (z - 0.988986)(z - 0.965028)(z + 1) .

z (z - 0.7788)(Z2 - 1.7238z + 0.7788)
The resulting stability boundaries are shown in Figure 11.12. The solid line is the f -
444 11 Robustness of Sampled-data Control Systems

5 W U W ~ W ~ ~ ~ W
v[m.s-1J
Figure 11.12. Continuous time (solid) and approximated discrete time (dashed) r-stability
boundaries in the (v, m)-plane

stability boundary af = {O" + jw I w2 = 250"2 - 49/16, 0" ~ -O .25} for the continuous
system. The dashed line is the approximated f-stability boundary for the system with
discretized compensator cz(z). The accuracy of the approximation was tested by finding
the roots of some polynomials on both sides nearby the approximate boundary. The
agreement was excellent such that the dashed line is also the exact I'-stability boundary
for the discrete-time system. This effect is not surprising because the continuous-time
plant has a relative degree three, i.e, the Poisson series converges like 1/w4 and the
sampling interval of T = 10 [ms] used for the controller implementation is small for this
plant. In this example, the I'-stable region is not much reduced by the compensator
discretization . 0

11.9 Circle Stability

In Section 11.3, we have discussed eigenvalue specifications for sampled-data control

systems. A conclusion was that circles centered on the real axis are useful regions of
f-stability for sampled-data systems. In this section, we show how the results obtained
for Schur-stability can be generalized to Ivstability for a circle with real center TO and
radius r, see Figure 11.13.

Definition 11.24

We call a polynomial circle stable if all its roots are located in a circle with given
real center and radius. 0
11.9 Circle Stability 445

T T

Figure 11.13. Circle stability

Circle stability is specified by two parameters, e.g. center TO and radius r, or alter-
natively,
T- .- r, TO -
(11.9.1)
T+ .- TO + r .

The circle of Figure 11.13 can be mapped onto the unit circle by an affine transformation

z := (z - To)/r, z = rz + TO . (11.9.2)

Thus, a polynomial p(z) = £to + alZ + ... + anzn is circle stable if and only if the
polynomial
p(z) = p(rz + TO)
n

L ai(rz + To) i (11.9.3)

i=O

L
n

= O,iZi
i=O

is Schur-stable.
The affine transformation (11.9.2) can be directly combined with the bilinear trans-
formation (11.4.7)

z -1 z- TO - r
= - - = ----'-- z=r--+To
l+w
(11.9.4)
W
z + 1 z - TO + r ' 1-w

or, written in terms of the parameters T- , T+,

z- T+
W=--
z- T-
, (11.9.5)
446 11 Robustness of Sampled-data Control Systems

A polynomial p(z) = ao + alz + ...+ anzn is circle stable if and only if the polynomial

L ai(T+ - T-w)i(1 - wt-i

n

(11.9.6)
i=O

is Hurwitz-stable.
Again, the necessary stability conditions b, > 0, i = 0,1 , . . . , n describe the convex
hull of the stability region in the space of coefficients ai.

Example 11.25

The zeros of the polynomial p(z) = ao + alZ + a2z2 + Z3 should lie inside the circle
with center TO = 0.4 and radius T = 0.4, i.e. T- = 0, 7'+ = 0.8.
3 3
Pw(w) = L ai 0.8i (l - W?-i = L b, Wi ,
i=O i=O
bo p (0.8) = ao + 0.8al + 0.64a2 + 0.512 , (11.9.7)
bl -(3ao + 1.6al + 0.64a2) ,
b2 3ao + a.8al ,
b3 -ao ·
The convex hull of the stability region is described by the inequalities b, > 0; the
critical conditions contributing to the boundary of the stable region are bo > 0, b3 > 0
and bibz - bob3 > O. 0

The modified form of Theorem 11.11 can be shown by barycentric coordinates [4] .
The result is:

Theorem 11.26

Consider the circle stability region of

(11.9.8)

in the space of coefficients ao, all ... ,an' Its convex hull is a polytope with n + 1
vertices corresponding to the polynomials

(11.9.9)

o
11.9 Circle Stability 447

A boundary representation (9.2.15) for frequency domain approaches is obtained by

substitution of

(11.9.10)

into (9.2.16):
do 1,
d1 27, (11.9.11)
di+l 27 d, - (r2 + 2707 - 7J) di - 1,

and (9.2.15) yields the boundary description.

Example 11.27

Consider again Example 11.25 with 70 = 0.4, r = 0.4:

do = 1,
d1 = 27 ,
d2 47 2 - 0.87 ,
d3 47 3 - 3.27 2 .

The coefficients ai of a third degree polynomial with roots on the circle of radius r = 0.4,
centered at 70 = 0.4, satisfy

1 27 47 2 - 0.87 47 3 - 3.27 2
[ o 1 27 472 - 0.87
] [:~
a2
] = [ 0 ]
0 ' 7 E [0; 0.8] .
1
The cubic term in 7 may be eliminated by a left multiplication of the above equations
by

yielding

o
A Polynomials and Polynomial Equations

This appendix summarizes some results from classical algebra concerning univariate or
multivariate polynomials. Only those properties of polynomials are mentioned that are
important for understanding and applying the methods of robust control in this book.
More details can be found in standard textbooks like [51] and [93] .

Introduction
The parameter space method [15] has as its basis the closed-loop characteristic poly-
nomial p(s, q, k), where q are the uncertain plant parameters, and k are the controller
parameters. The aim of the control design is to find the controller parameters k such
that the roots of the uncertain polynomial (the parameters q vary in a given operating
domain) are located in a prescribed region I' of the s-plane. Fixing all but two param-
eters, the feasible region in the parameter plane is represented. The parameter space
method can also be used in an analysis step . In this case, a test can be done in the
plant parameter plane to determine if the entire operating domain is I'-stable.
In both analysis and design, the basic equation is p(s, x , y) = 0, where x, y are
parameters and s varies on the boundary of of the region f . Substituting s = u(Q) +
j v( Q) E of in the uncertain polynomial and separating into real and imaginary parts
leads to the system of polynomial equations

Rep(x,y,Q) = 0, Imp(x,y,Q) = 0, (A.I)

which has to be solved for varying Q . The solutions represent parameter combinations
that lead to a polynomial with a root or a root pair on the boundary of. So, after we
have mapped the boundary of into the parameter plane, the next step is to find the
active boundaries that are the points (or boundaries) that contribute to the boundary
of the I'-stable area.
The complexity of the system of equations (A.I) depends on how the parameters
appear in the characteristic polynomial. We distinguish three types of dependence with
increasing complexity: affine, bilinear and polynomial dependence.
In the following, we discuss methods that can be used to solve the system (A.I)
and related problems. To aid in this discussion, we first introduce some definitions and
theorems from the theory of algebraic curves. We then present procedures for solving
various systems like (A.I) 'and discuss the applications in which the procedures are
useful. We demonstrate the methods presented using simple examples.
 449

Algebraic Curves
If f(x, y) is a bivariate polynomial in x, y:

f(x ,y) = L I aijXiyi (A.2)

i,j ?O
(the prime denotes that the sum is finite) with real coefficients aij, then the set {(x , y) E
]R2 : f(x, y) = O} is called an algebraic curve. The degree of f is the highest sum of the
exponents, and is equal to the number of common points with a straight line. We call
f(x,y) = 0 (A .3)
the implicit representation of the curve. Curves may have also an explicit representation
x = f(o:), y = g(o:), (AA)
where x(o:) and y(o:) are rational functions of the real parameter 0: . Equivalent is the
representation
nl(O:) n2(0:)
x = d1(0:) , y = d ,
2(0:)
(A.5)
with ni(O:), di(o:), i = 1,2 , polynomials in 0:. These curves are called rational curves.
Nyquist or Popov plots are trivial examples of rational curves. The root loci have, in
general, no parametric representation.

Elimination Theory. The Resultant

In this section, we follow the representation used in [121). More details of the algebraic
theory can be found in classical textbooks like [42) , [54) or [194). For the theory of
curves, see [59).
Consider the two polynomials
f(y) = anyn + an_lyn-l + an_2yn-2 + ... + ao }
(A.6)
g(y) = bmym + bm_1ym-l + bm_2ym-2 + . .. + ho, '
(an # 0, bm # 0), whose coefficients ai, b, are real numbers.
The theory of the resultant originates in the following:

Problem

Give a necessary and sufficient condition on the coefficients of (A.6), such that
f(y) = 0 and g(y) = 0 have a common root. 0

Theorem 1.1
f and g have a common root, if and only if the determinant of the matrix R, in
(A.8) is zero.
o
450 A Polynomials and Polynomial Equations

This determinant is called the resultant: of f and 9 with respect to y, i.e.

Res (f,g ,y) = det R . (A.7)
Assume now that f and 9 have a common root , then the degree of the greatest common
divisor of f and 9 (gcd(f, g)) is at least 1. For finding this greatest common divisor, an
additional definition is given.

Definition 1.2
The matrix R I of dimension m + n - 2, obtained by deleting the first and the last
rows and t he first and t he last columns in the Matrix R , is called the first inner of
R . Continuing on the deletion process, we obtai n the inners R 2 , R 3 , . . . of dimension
m + n- 4, m+n- 6, . . . respectively. The determinants of inners are called subresultants.
o

Example 1.3
For n = 4, m = 3 we have the matrix R and 3 inners:
a4 a3 a2 al an 0 0

0 a4 a3 a2 al ao 0
0 0 a4 a3 a2 al ao
R = 0 0 0 b3 bz bl bo (A.S)
0 0 b3 ~ bl bo 0
0 bs ~ bl bo 0 0

b3 ~ bl bo 0 0 0

a4 a3 a2 al ao
0 a4 a3 a2 al
RI = 0 0 b3 b2 bl

0 b3 b2 bl bo

bs b2 bl bo 0
a3
R2 = [~
0 b3 b2~]
b3 b2 bl
lIn the literature the matrix is often defined in a slightly different mann er, th e rows of th e b, are
exchanged , so only the sign of the resultant may change depending on th e numb er of exchanges.
 451

Theorem 1.4

If det R = det R, = det R k - l = 0 and det R k i- 0, then the degree

of gcd(f, g) is k. In this case, gcd(f, g) equals the determinant of the matrix
obtained from R k by replacing the last column in it by the column

[y m-k-lf( y ) , y m-k- 2f() y , . .. , (A.9)

f(y) , g(y) , yg(y) , . . . , yn-k-lg(y)]'.

an an-l an- 2 .. . al ao 0 0 0 0 0
(m rows)
0 an an- l al ao 0 0 0 0

0 0 0 al ao
R (A.lO)
0 0 0 bl bo
(n rows)
0 bm bm- l b, bo 0 0 0 0
bm bm- 1 bm - 2 .. . bl bo 0 0 0 0 0

Example 1.5

Suppose in Example 1.3 th at k = deg(gcd(f, g)) = 1, then

a4 a3 a2 al y f(y )
0 a4 a3 a2 f(y )
gcd(f,g) = det 0 0 b3 b2 g(y)

0 b3 b2 bl yg( y)

b3 ~ b1 bo y2g (y )

Using the fact th at gcd(f, g) is linear in y, gcd(f, g) = Ay+B, all quadrat ic and higher
452 A Polynomials and Polynomial Equations

terms in Y must vanish so we can simplify the determinant:

a4 a3 a2 al aoy
0 a4 a3 a2 alY + an
gcd(J,g) = det 0 0 b3 ~ bly + bo
0 b3 b2 bl boY
~ bz bl bo 0
Thus, gcd(J, g) = Ay + B, where A = det R I and
a4 a3 a2 al 0

0 a4 a3 a2 an
B=det 0 0 b3 ~ bo
0 b3 ~ bl 0

b3 ~ bl bo 0
For k = 1 in Theorem 1.4 we have 0

Corollary

If f and 9 have a single common root Ya (det R = 0, det R I i= 0), then

Ya = - det Rd det R I , (A.H)

where R1 equals the matrix obtained from R 1 by replacing the last column in it by the
column ~, an, bo, ~jT . 0
m-2 n-2

Application

Let p(s) be any real polynomial. If we set s = jw, we can write

= aO - a2W + a4 W + ...
. ) 2 4
P( JW
. (at
+ JW - a3w 2 + asw4. .)
.
= Peven + jw Podd·
Setting f2 = w2 , we can rewrite these polynomials as

Peven(f2) = ao - a2f2 + a4f22 . .. ,

podd(f2) = at - a3f2 + a s f2 2 • • • •

The resultant of the polynomials Podd(f2) and Peven(f2) with respect to f2 is the deter-
minant of the well-known Hurwitz matrix. Apart from the sign of some entries, R is
the Hurwitz matrix (1.6.2).
 453

Example 1.6
For n = 4, we have f = ao - a2n + a4n2 and 9 = a l - a3n , the matrix R is

R= [ a~
- a3
=:: :~] .
al 0
o

The Discriminant
A special case of the resultant is the discriminant Dt- The second polynomial 9 is the
derivative of the first polynomial I, i.e,

D/ := 2- ResU , 1', Y). (A.12)

ao
Let Yi, i = 1 . .. n , the roots of f(y) = O. Another definition of the discriminant is

IIII(Yi - Yk)2
n n
D/ = (-1) n(n2- 1) a~n-2 (i =1= k). (A.13)
i=1 k=1

Theorem 1.7
The equation D] = 0 is a necessary and sufficient condition that f(y) has a root
of multiplicity of at least 2.
o
Application

The breakaway points (real) and saddle points (complex) of the root locus can be
determined by computing the discriminant and finding the zeros.

Example 1.8
Let p(s) = (s + 2)( s + 1)( s - 1) + k = S3 + 2s2 - S - 2 + k , then p'(s) = 3s2 + 4s - 1

12-1 -2 + 0]
and
k
o 1 2 - 1 -2 + k
D; = det 0 0 3 4 -1
o 3 4 -1 0
3 4 -1 0 0
= 27k 2 - 40k - 36 = 0 .
The two solutions are kl ,2 = (20 ± 14/7)/27 and lead to a double root at Sl,2 =
(-2 ± /7)/3. 0
454 A Polynomials and Polynomial Equations

Intersection Points of Algebraic Curves

Consider now a system of two polynomial equations

f(x, y) = 0, g(x, y) = 0 (A.14)

with real coefficients. We assume that f and 9 have no common factor, and that both
polynomials are irreducible over the rational numbers. Each equation can be seen as a
curve. So the points (x , y) for which f(x, y) = 0 and g(x, y) = 0 are the intersection
points of the two curves.
We can see f(x , y) and g(x, y) as polynomials in one of the variables, say y, where
the coefficients ai, bi of (A.14) are now polynomials in x .
Using Theorem 1.1, the resultant is no longer a number but a polynomial in x. So
we have eliminated the variable y. The first component x· of the intersection point
(x', y') is then a root of the resultant. The other component is determined by the
gcd(f, g) = 0, where x· is substituted for x :

gcd (f(x',y),g( x',y)) =0 (A.15)

Theorem of Bezout

Let nf = deg(f), n g = deg(g). Then, the number of intersection points (real and/or
complex) of f(x, y) = 0 and g(x, y) = 0 is nfng . 0

Remark 1.9

If the degree of the resultant Res(f, g, y) = h(x) is not nfn g , then

i. Some roots are at infinity, and/or

ii. More than one component y' belongs to one component x' , and/or
iii. The multiplicity of one or more intersection points is greater than 1.
o
Application

Given an explicit representation of a curve

we can eliminate the parameter 0:' using the resultant method :

The inverse process, that is to find a parametric representation given an explicit repre-
sentation, is only possible for rational curves.
 455

The Intersection of Two Polynomial Families

Consider now a system of two parametric polynomial families
f(x , y, a) = 0, g(x, y, a) = 0, a E [a- , a+J (A.16)
For fixed a, we have the situation of the preceding section. For varying a , the solution
pairs (z", yO) generate a new curve consisting of one or more (real or complex) branches.
The parameter-free equation F(x, y) = 0 of this curve can be generated by elimination
of a :
F(x,y) = Res(f,g,a) = O. (A.17)
But the restriction to an interval of the parameter a gets lost, so F(x, y) = 0 has "more"
points than the parametric representation. Therefore, we eliminate y rather than a.
The procedure for solving (A.16) is the following:
First, compute the resultant of f and 9 with respect to y:

h(x, a) = Res (J(x, y, a) , g(x, y, a) , y) . (A.18)

Then, grid a . For a = a*, determine the roots x* of h(x, o") = o. For the corresponding
value(s) z", again the greatest common divisor has to be determined :

gcd (f(x* ,y,a*) ,g(x*,y,a*)). (A.19)

Varying a, the number of real roots can change (i.e. branches of the curve can terminate
or arise). This happens when two roots coincide on the real axis and go into the complex
plane (or vice versa).
The condition for double roots in this case is:
ah(x, a)
Dh(x ,a) = Res(h(x , a), ax ,x) = k(a) = O. (A.20)

The real zeros of k(a) = 0 divide the range of a into intervals where the number of real
roots x* are constant. These numbers are only different in the neighborhood of a zero
of k(a) = 0, if the sign of the discriminant changes.

Application

Solving (A.16) is the main task of the parameter space method . x , yare the un-
certain plant parameters and/or controller parameters. Varying a corresponds to the
movement on the boundary of the desired pole regions. If the dependence on x, y is
affine, the equations (A.16) can be solved easily (the equations are linear in x, y) and we
arrive at the explicit representation (A.4). Also, the solution of the bilinear case (the
equations (A.16) are bilinear in x , y) can be given in a closed form. According to the
theorem of Bezout, we should have four intersections points. Two points are at infinity,
since both curves are, in general, hyperbolas with asymptotes parallel to the coordinate
axes. So, the final equation (the univariate polynomial) has degree two. An algebraic
argument for this fact is the following: we can linearly combine both equations such
that the resulting equation is linear in x and y. Solving for x and substituting in one of
the basic equation s gives a quadratic equation in y. Examples can be found in [159J .
456 A Polynomials and Polynomial Equations

The Envelope
Given a one-parameter family of algebraic curves,

f(x, y, 0:) = O. (A.21)

An envelope is a curve that touches each member of the one-parameter family and is
touched at each of its points by some curve of the family. If the envelope exists (for
example, concentric circles have no envelope), then the defining equations are

f(x, y, 0:) = 0,
of
00: (x , y, 0:) = O. (A.22)

Again, the parameter 0: has to be eliminated, this leads to the implicit equation :

F(x,y) = D j = O. (A.23)

The notion of the envelope has still another meaning. It is also used for the extremal
curves or hull of a family of curves. Its equation can not be generated by means of the
discriminant.

Example 1.10

Let f(x, y, 0:) = (x - 0:)2 + y2 - 1 = 0, a family of circles with center on the x-axis
and radius 1. The derivative with respect to 0: is f' = 2(x - 0:)(-1) . Substituting the
solution 0: = x in f(x, y, 0:) = 0 gives the envelope y = ±1. 0

Application

Algebraic constraints or bounds on transfer functions (sensitivity function, comple-

mentary sensitivity function) are often formulated as inequalities f(x, y , w) > 0 for each
frequency [159] . Instead of gridding wand plotting f(x, y, w) = 0, the envelope (if it
exists) could be computed .

Example 1.11

The data for an uncertain transfer function used in this example are from [48] . For
each frequency w, we have a requirement f(x,y ,w) > O. The feasible region (necessary
but not sufficient) (Figure A.l) for x and y lies outside the ellipsoidal areas (dotted
lines). Instead of plotting all these areas, we compute the equation of the envelope
D] = F(x, y) = O. Because we have only the implicit representation, we must grid x
and/or y to get a pointwise plot (diamonds) of the envelope. 0
 457

.""",-- - - - - - - - - - - - - - ,

.'
"
O.CUI1

..,
.,,,,
.' ... ';
. ~ ...
.. :

Figure A.I. Requirements for different frequencies and the corresponding envelope

Singular Points

Definition 1.12

A point (x , y) on a curve f(x , y) = 0 is called singular if both partial derivatives

vanish. In this case, we have three polynomial equations

f(x, y) = 0, fx(x , y) = 0, fy(x, y) = o. (A.24)

If higher partial derivatives also vanish, we speak of a higher singularity.
There are three types of (normal) singularites depending on the sign of the expression

(A.25)

(also a discriminant) . If D > 0, we have a double point (node); the curve intersects
itself. If D = 0, we have a cusp and for D < 0 we have an isolated point . In its vicinity
there are no further (real) points.
The resultant method is also useful for finding the singularities.
If the curve is defined in parametric form:

x = f(a), y = g(a) (A.26)

(J, g rational functions), then the intersection points (double points) can be found by
solving the following system:
(A.27)

which has the trivial solution al = a2. So, instead we should solve the reduced system

(A.28)
458 A Polynomials and Polynomial Equations

Num (g(n d - g(n 2 ) ) = O. (A.29)

nl - n2
o
Application

Determining the f-stable region (not only the complex root boundary) requires the
knowledge of the intersection points of a curve with itself (see [159]).

Application

An algebraic curve is rational if

(n - l)(n - 2) = L rj(rj - 1) (A.30)

(r, are the multiplicities of the singular points) . Only rational curves have a para-
metric representation. Finding it, starting from the implicit representation, needs the
knowledge of all singular points and additionally n - 3 simple points.

Special System of Equations

Consider the following system of equations:

N(x,y ,n) _ 0 ~ N(x,y,n) = 0

(A.31)
D(x,y,n) - , an D(x, y, n)
(N, D polynomials). The first equation is called the point condition, the second tangent
condition.

Lemma

The system
N(x, y, o) = 0,
a
an N(x, y, n) = 0 (A.32)

is equivalent to (A.31).

Proof
DN'-D'N N'
From N/D = 0 follows N = O. Therefore, (N/D)' = = - = 0 is
D2 D
equivalent to N' = O.

Application

The conditions that the Nyquist curve (or any curve) touches another curve (for
example, a circle around the critical point -1) leads to a system like (A.32). Further
 459

examples can be found in [63J .

Application

Computing the stability radius [15] also leads to a system of type (A .32) .

Double Tangent and Popov-line

Find the double tangent of a curve x = f(a), y = g(a) .
The tangent t at a point a = a" is

(A.33)

These equations are substituted in the parameter-free representation

(A.34)

Since the tangent t touches F , A = 0 is a double root . To touch the curve once more,
we have the condition that F(a", A)IA2 has a double root . So again, the discriminant
with respect to A yields the corresponding polynomial whose roots are the values for a,
which give the touching points.

Application

The Popov-line is either the tangent at an intersection point of the Popov plot with
the real axis or a double tangent to the Popov plot.

Example 1.13

Let g(8) = (0.018 + 1)/(8 5 + 584 + 1083 + 4182 + 118 + 1). We want to determine
the double tangent which is also the Popov-line. Since the real part and the imaginary
part of the Popov plot are always quadratic functions of w, we can replace w2 by
a . The degree of the parametric representation, as well as the degree of the implicit
representation in a, is five, the degree of the denominator polynomial of g(8). The
parameter values for the touching points are al = 0.0766 and a2 = 8.23, resulting from
the discriminant polynomial which has degree 118. The Popov-line intersects the real
axis at -11k = -0.136 and leads to the sector with k = 7.36 (Figure A.2) .
o
460 A Polynomials and Polynomial Equations

0.3

0.25
1
0.2

/
I
0.15
•

I
O.1

O.ol5

J .

~.05
V ~

~~ -----
~.1
.f-. .

~.2 0.2 0.0 0.8 0.8

Figure A.2. Popov plot and Popov-line

B PARADISE: Parametric Robustness
Analysis and Design Interactive Software
Environment

This appendix gives a brief introduction to the PARADISE toolbox .

It is not our intention to describe a specific set of features or a certain ver-
sion of this toolbox . Moreover, the concept and basic usage of this toolbox is ex-
plained such that the reader is able to repeat the case studies in Chapters 6 and 7
and to try his or her own design. The latest version of PARADISE is available at
http://www.robotic.dlr.de/control/paradise. More details and examples can be found
at this site and in the comprehensive toolbox manual [1521.

Concept
PARADISE is a MATLAB-based toolbox that implements parametric robust control
methods. Th e aim was to design a user-friendly toolbox with graphical user inter-
faces (GUI) , which hides all calculations from the user as much as possible. Rather
than doing tedious calculations, the user should concentrate on analyzing the graphical
results. PARADISE uses both symbolical and numerical code to get the benefits from
both worlds.

Architecture
The architecture of the CACSD (Computer-aided Control System Design) toolbox PAR-
ADISE is driven by the following objectives: user friendliness, code maintenance, code
efficiency and reusability of code. For this reason the toolbox was split into several
modules. The modularity allows the realization of these objectives.
A basic partition of any interactive CACSD toolbox is the subdivision into interac-
tive and computational modules. Figur e B.1 shows the coarse structure of PARADISE.
The individual modules will be discussed in subsequent sections .

Intemctive Modules

In order to achieve maximal user friendliness, the user can choose between numer-
462 B PARADISE

Spezlflcatlons:
r-Edltor

Figure B.1. Modular design of the toolbox

ous possibilites to interact with the toolbox . Apart from the GUls, the toolbox can
be controlled from the command-line or batch-file driven. Data can be imported using
files, e.g. SIMULINK models, or by direct exchange of MATLAB workspace variables .
The main interactive modules will be described in more detail in section "Working with
the toolbox" .

Calculation Modules

The model transformation module has two purposes ; it analyses a given SIMULINK
model to determine the inputs and outputs of the model and generates a symbolic model
of the system . Additionally, this module is used to find the characteristic polynomial
for eigenvalue specifications and to transform systems from state space into transfer
function format .
The general mapping equations for specifications presented in this book are given
by two polynomial equations with three real parameters ql, qz, 0::

PI(ql,q2,0:) = 0,
(B.l)
P2(ql,q2 , 0:) = 0,

where 0: is an artificial parameter stemming from the specification with 0: E [0:-; 0:+] .
Note: parameters ql, q2 might be uncertain plant or controller parameters. The map-
ping equations (B.l) are generated by the equation generator module using the system
description and the user-defined specifications .
The generic problem is now to find all curves ql(0:),q2(0:) that satisfy (B.l) . These
curves form the boundaries in a parameter plane . The curves are determined by the
mapping solver module .
In order to get short execution times, the mapping solver module first classifies the
polynomials PI, P2 with respect to the parameter dependence for ql, q2 in affine, bilinear
 463

and polynomial. For each class, there exists a special code, which is run automatically
such that the user needs not to choose the appropriate method. Furthermore, the
modular design of the toolbox allows incorporation of code for special cases easily, e.g.
PIn-controllers [22] .
Generating the mapping equations (B.1) and finding all solution branches hereof
requires numerical and symbolical methods. The Extended Symbolic Math Toolbox
adds the symbolic mathematics of the MAPLE symbolic engine to the powerful numeric
capabilites of MATLAB. All symbolic computations in PARADISE are done using this
toolbox.

Working with the Toolbox

Input

The first step to solve a parametric robust control problem involves the definition of
the plant with the used controller structure, determination of the operating range and
specification of the control requirements.
A parametric plant model can be imported into PARADISE using a parametric state
space description provided as a text-file. Apart from the user-generated symbolic plant
description, PARADISE can work with parametric SIMULINK models. These are ordi-
nary SIMULINK models, where the model parameters are entered as strings instead of
numerical values. PARADISE analyzes the given block interconnection and generates a
parametric state space representation. PARADISE can also work with the multi-model
formulation described in Section 2.6, or if no continuous parameter dependence is avail-
able like in the aircraft example of Section 7.2.
After the model description has been imported, all parameters existing in the model
are determined. The user can now classify these parameters into controller, variable
or constant parameters. Subsequently, the numerical values for controller and constant
parameters and the numerical ranges for variable parameters have to be entered. A
special parameter specification GUI therefore is available.
The control-related specifications can be entered using special GUIs. As an example,
we consider the f-editor, which is used to define f-specifications. The f-region is formed
by choosing different basic elements that form the boundary of the f-region. There is
a library of basic elements, e.g. real part limitation or hyperbola, all of which are
conic sections. The geometrical parameters of all basic elements can be entered either
graphically using the GUI or by entering the numerical values. All basic elements can
be logically combined through logical union or intersection such that almost any region
can be described using basic elements. Figure B.2 shows the f-editor.

Design and Analysis

The main purpose of PARADISE is to map specifications into a parameter plane.

Through mapping given specifications into a parameter plane, the user can easily do a
robustness analysis. The specifications can be mapped into planes of given parameters
or the user can define his own parameter planes. Actually, there is special support for
464 B PARADISE

- - --- -
61'<'""""- - - - - - - ....,......., F
- 0..... reqi on ' 1
S
.:t'
e t rucUon : def ault
t ' 1: Ileal. pu t b a l.t.. tion
.ipa· · 1
trwu t4"4 : 10
· t lt·..nt n : DupinlJ
. ....1n9 • a7
Im9["4: 10
· El..ent , l : Circl•
• i 9U . O

Innrud.: 1 0

SII

Figure B.2. f-Editor

invariance planes, which allows successive pole-shifting, see Section 2.4.

Another possiblity to verify a control design is to display the resulting dosed-loop
poles directly in the I'-editor.
After specifying the plant model and the parameters, a robust design of a control
system is done iteratively by performing the following steps:

i. Define the specifications.

ii. Choose an appropriate parameter plane .
iii. Map the current specifications into parameter plane .
iv. Check if specifications are fulfilled or tune parameters such that specifications are
achieved/improved. (Maybe go back to step 1. and relax specifications) .
v , Go back to step 2. until specifications are robust fulfilled.

Output

All results of the analysis and design steps are displayed graphically to facilitate the
interpretation. There are interfaces to save and export the data in appropriate formats ,
e.g. MATLAB variables.
Bibliography

[I] J. Ackermann, "Der Entwurf linearer Regelungssysteme im Zustandsraum (State

space design of linear control systems)," Regelungstechnik, vol. 20, pp. 297-300,
1972.
[2] J . Ackermann, "Paramet er space design of robust control systems," IEEE Trans.
on Automatic Control, vol. 25, pp. 1058-1072, 1980.
[3] J. Ackermann, "Robustness against sensor failures," Automatica, vol. 20, pp . 211-
215, 1984.
[4] J . Ackermann, Sampled-data control systems: analysis and synthesis, robust sys-
tem design. Berlin: Springer , 1985.
[5] J . Ackermann, "Verfahren zum Lenken von Strai3enfahrzeugen mit Vorder- und
Hinterradlenkung I (Method of steering road vehicles having both front-wheel
steering and rear-wheel steering I)." German Patent P 4028320 , European Patent
0474130, US Patent 5375057, Japanese Patent 3158347, Priority Sept. 6, 1990.
[6] J . Ackermann, "Does it suffice to check a subset of multilinear parameters in
robustness analysis?, " IEEE Trans. on Automatic Control, vol. 37, pp . 487-488 ,
1992.
[7] J . Ackermann, "Verfahren zum Lenken von Strai3enfahrzeugen mit Vorder- und
Hinterradlenkung II (Method of steering road vehicles having both front-wheel
steering and rear-wheel steering II)." German Patent P 4206654, European Patent
0559114, US Patent 5515275, Japanese Patent 3210471, Priority March 3, 1992.
[8] J . Ackermann, "Verfahren zum Lenken von StraBenfahrzeugen mit Vorder-
radlenkung (Method of correcting steering of a road driven vehicle)." German
Patent 4307420, European Patent 0614795, US Patent 54 28 536, Japanese Patent
Application 6-59841, Priority March 9, 1993.
[9] J . Ackermann, "Verfahren zur Lenkunterstiitzung eines Fahrers eines StraBen-
fahrzeugs (Method of correcting steering of a road driven vehicle)." German
Patent P 19650691, European Patent 97121452.3-2206, US Patent 5925083, Pri-
ority Dec. 12, 1996.
[10] J . Ackermann, "Yaw disturbance attenuation by robust decoupling of car steer-
ing," in Proc. 13th IFAC World Congress, vol. Q, (San Francisco) , pp. 1-6, 1996.
[11] J . Ackermann, "Robust control prevents car skidding ," IEEE Control Systems
Magazine, pp . 23-31, June 1997. Bode Prize Lecture 1996.
466 Bibliography

[12J J. Ackermann, "Verfahren zur Vorgabe einer gewiinschten geschwindigkeits -

abhangigen Gierdampfung bei der Fahrzeuglenkung (Method of assignment of
a desired velocity-dependent yaw damping for vehicle steering) ." German Patent
Application 10206730.9, Priority Feb. 18, 2002.
[13J J . Ackermann, Pole placement control. UNESCO, 2002 (to appear) . Encyclopedia
of Life Support Systems .
[14J J . Ackermann and B. Barmish, "Robust Schur stability of a polytope of polyno-
mials," IEEE Trans. on Automatic Control, vol. 33, pp . 984-986, 1988.
[15J J. Ackermann, A. Bartlett, D. Kaesbauer, W. Sienel, and R. Steinhauser, Robust
control: Systems with uncertain physical parameters . London: Springer, 1993.
[16J J. Ackermann and T . Biinte, "Robust prevention of limit cycles for robustly de-
coupled car steering dynamics," Kybernetika, vol. 35, no. 1, pp . 105-116, 1999.
[17J J . Ackermann, T . Biinte, J . Dietrich, B. Gombert, and B. Willberg, "Aktuator
zum Korrigieren eines iiber das Lenkrad eines Fahrzeugs an die Rader einer ge-
lenkten Achse eingegebenen Lenkwinkels (Actuator for generating an additional
steering angle for road vehicles)." German Patent 19750585, European Patent
0916568, US Patent 6343671, Priority Nov. 17, 1997.
[18J J . Ackermann, T. Biinte, W. Sienel, H. Jeebe, and K. Naab , "Driving safety by
robust steering control," in Proc. Int . Symposium on Advanced Vehicle Control,
(Aachen , Germany) , June 1996.
[19J J. Ackermann, J . Guldner, W. Sienel, R. Steinhauser, and V. I. Utkin, "Linear
and nonlinear control design for automatic steering," IEEE Trans. on Control
Systems Technology, Special Issue on Automotive Control Systems, vol. 3, no. 1,
pp . 132-143, 1995.
[20] J . Ackermann and H. Hu, "Robust ness of sampled-data control systems with
uncertain physical plant parameters," Automatica, vol. 27, pp . 705-710, 1991.
[21J J . Ackermann, H. Hu, and D. Kaesbauer, "Robustness analysis : a case study,"
IEEE Trans. on Automatic Control, vol. 35, pp . 352-356, 1990.
[22J J. Ackermann and D. Kaesbauer, "Design of robust PID controllers," in Proc.
European Control Conference, (Porto), 2001.
[23J J . Ackermann, D. Kaesbauer, and R. Miinch, "Robust f-stability analysis in a
plant parameter space," Automatica, vol. 27, pp . 75-85, 1991.
[24J J. Ackermann, D. Kaesbauer, and W. Sienel, "Design by search ," in Proc.
First IFAC Sympos ium on Design Methods of Control Systems, vol. 1, (Ziirich),
pp . 181-186, 1991.
[25] J . Ackermann and D. Odenthal, "Robust steering control for active rollover avoid-
ance of vehicles with elevated center of gravity," in Proc. International Conference
on Advances in Vehicle Control and Safety, (Amiens, France) , pp . 118-123, July
1998.
[26J J . Ackermann and D. Odenthal, "Damping of vehicle roll dynamics by gain sched-
uled active steering," in Proc. European Control Conference, (Karlsruhe, Ger-
many), 1999.
[27J J. Ackermann and W. Sienel, "Computer-aided robustness analysis for character-
istic polynomials with polynomial parameter dependence," in Proc. IEEE Conf.
Decision and Control, (Tampa) , pp. 1934-1935, 1989.
 467

[28] J . Ackermann and W. Sienel, "Robust control for automat ic steering," in Proc.
American Control Conference, (San Diego), pp . 795-800, 1990.
[29] J . Ackermann and W . Sienel, "What is a 'large' number of parameters in robust
systems," in Proc. IEEE Con! Decision and Control, (Honolulu), pp. 3496-3497,
1990.
[30] J . Ackermann and W. Sienel, "On the computation of value sets for robust stabil-
ityanalysis," in Proc. First European Control Conference, (Grenoble), pp. 2318-
2327 , 1991.
[31] J . Ackermann and W. Sienel, "Robust yaw damping of cars with front and rear
wheel steering," IEEE Trans. on Control Systems Technology, vol. 1, no. 1, pp. 15-
20,1993.
[32] J . Ackermann and S. Tiirk, "A common controller for a family of plant models,"
in Proc. 21st IEEE Con! Decision and Control, (Orlando), pp . 240-244, 1982.
[33] B. Aksun Giivenc, T. Biinte, D. Odenthal, and L. Giivenc, "Robust two degree
of freedom vehicle steering controller design," in Proc. American Control Confer-
ence, (Arlington, VA, USA), June 200l.
[34] B. Aksun Giivenc and L. Giivenc, "Robust two degree-of-freedom add-on con-
troller design for automatic steering," IEEE Trans. on Control Systems Technol-
ogy, vol. 10, no. 1, pp. 137-148,2002.
[35] R. W. Allen, H. T. Szostak, D. H. Klyde, T . J. Rosenthal, and K. J . Owens,
"Vehicle dynamic stability and rollover," tech . rep., Systems Technology, Inc.,
Hawthorne, CA, 1992. U.S.-D.O.T ., NHTSA .
[36] J . Anagnost, C. Desoer , and R. Minichelli, "Generalized Nyquist test for robust
stability: Frequency domain generalizations of Kharitonov's theorem," in Robust-
ness in Identification and Control (M. Milanese, R. Tempo, and A. Vicino, eds.),
pp . 79-96, New York: Plenum Press , 1989.
[37] K. Astrom and B. Wittenmark, Computer controlled systems. Englewood Cliffs,
N.J. : Prentice-Hall, 1984.
[38] G. Balas, J . Doyle, K. Glover, A. Packard, and R. Smith, The f.L analysis and
synthesis toolbox. The MathWorks, Inc., 1998.
[39] B. Barmish, "New tools for robustness analysis," in Proc. IEEE Con! Decision
and Control, (Austin), pp. 1-6, 1988.
[40] B. Barmish, "A generalization of Kharitonov's four-polynomial concept for ro-
bust stability problems with linearly dependent coefficient perturbations," IEEE
Trans. on Automatic Control, vol. 34, no. 2, pp . 157-165, 1989.
[41] B. Barmish, New tools for robustness of linear systems. New York: MacMillan,
1994.
[42] S. Barnett, Polynomials and Linear Control Systems. New York: M. Dekker,
1983.
[43] A. Bartlett, C. Hollot, and Huang-Lin, "Root locations of an entire polytope of
polynomials: it suffices to check the edges," Mathematics of Control, Signals and
Systems, vol. 1, pp. 61-71, 1988.
[44] R. Bass and I. Gura, "High order system design via state-space considerations,"
in Joint Automatic Control Conference, pp. 311-318, 1965. Preprints.
468 Bibliography

[45] R. Bass and P. Mendelson , "Aspects of general control theory," Final report 2754,
AFOSR, 1962.
[46] S. Bennani, C. Magni, and J . Terlouw, Robust Flight Control. London: Springer,
1997.
[47] R . Berger, J . Hess, and D. Anderson, "Compatibility of maneuver load control
and relaxed static stability applied to military aircraft," AFFDL-TR-73-33 , Air
Force Flight Dynamics Laboratory, 1973.
[48] V. Besson and A. T . Shenton, "An interactive parameter space method for robust
performance in mixed sensitivity problems," IEEE Trans. on Automatic Control,
vol. 44, pp. 1272-1276, June 1999.
[49] S. Bialas, "A necessary and sufficient condition for the stability of convex com-
binations of stable polynomials or matrices," Bulletin of the Polish Academy of
Sciences, vol. 33, pp. 473-480, 1985.
[50] M. Biehler, "Sur une classe d'equations algebriques dont toutes les racines sont
reelles," J. Reine Angewandte Mathematik, vol. 87, pp . 350-352, 1879.
[51] G. Birkhoff and S. M. Lane, A survey of modern algebra. New York: The MacMil-
lan Company, 1965.
[52] P. Blue and S. Banda, "D-K iteration with optimal scales for systems with time-
varying and time-invariant uncertainties," in Proc. American Control Conference,
1997.
[53] P. Blue, L. Oiivenc, and D. Odenthal, "Large envelope flight control satisfying
H oo robustness and performance specifications," in Proc. American Control Con-
ference , (Arlington, VA, USA) , June 2001.
[54] M. Bocher, Introduction to Higher Algebra. New York: Dover , 1964.
[55] H. Bode, Network analysis and feedback amplifier design. New York: D. van
Nostrand Company, Inc., 1945.
[56] S. Boyd , V. Balakrishnan, and P. Kabamba, "A bisection method for computing
the H oo norm of a transfer matrix and related problems," Mathematics of Control,
Signals and Systems, vol. 2, pp . 207-219, 1989.
[57] S. Boyd and C. Barratt, Linear Controller Design: Limits of Performance.
Prentice-Hall, 1991.
[58] S. Boyd, 1. E. Ghaoui, E. Feron, and V. Balakrishnan, Linear Matrix Inequalities
in System and Control Theory. SIAM studies in applied mathematics, 1994.
[59] E . Brieskorn and H. Knorrer, Plane Algebraic Curves. Basel, Boston, Stuttgart:
Birkhauser Verlag, 1986.
[60] R. Brockhaus, Flugregelung. Berlin: Springer, 1994.
[61) I. N. Bronstein and K. A. Semendjajew, Taschenbuch der Mathematik. Thun und
Frankfurt/Main: Harri Deutsch Verlag, 1987.
[62) T . Bunte, Beitriiqe zur robusieti Lenkregelung von Person enkraftwagen (Contri-
butions to robust car steering control). Ph.D. thesis, RWTH Aachen, 1998. VDI
Fortschritt-Bericht, Reihe 12, Nr. 366, VDI-Verlag, Diisseldorf, 1998.
[63] T. Bunte, "Mapping of Nyquist/Popov Theta-stability margins into parameter
space," in Proc. 9rd IFAC Sympos ium on Robust Control Design, (Prague, Czech
Republic) , 2000.
 469

[64J T. Bunte, "Die Anwendung des Parameterraumverfahrens auf Ortskurvenkrite-

rien (Application of the parameter space method to frequency loci criteria) ,"
Automatisierungstechnik, vol. 49, pp. 547-556, December 200l.
[65J T . Bunte, D. Odenthal, B. Aksun Giivenc, and L. Giivenc, "Robust vehicle steer-
ing control design based on the disturbance observer," in Proc. 3rd IFAC Work-
shop on Advances in Automotive Control, (Karlsruhe) , pp . 107-116,200l.
[66J F. Cellier, Continuous system modelling. New York: Springer, 1991.
[67J A. Cohn, "Uber die Anzahl der Wurzeln einer algebraischen Gleichung in einem
Kreise," Mathematische Zeitschrijt, vol. 14, pp. 110-148, 1922.
[68J L. Cremer, "Ein neues Verfahren zur Beurteilung der Stabilitiit linearer
Regelungssysteme," ZAMM, p. 161, 1947.
[69J W. Darenberg, "Automatische Spurfiihrung von Kraftfahrzeugen," Automobil-
Industrie, pp. 155-159, 1987.
[70J S. Dasgupta, "Perspectives on Kharitonov's theorem: a view from the imaginary
axis," tech . rep ., Dept. Elect . Compo Eng., University of Iowa, Iowa City, 1987.
[71J A. Datta, M. Ho, and S. Bhattacharyya, Structure and synthesis of PID con-
trollers. London: Springer, 2000.
[72) M. Debes , E. Herb , R. Muller, G. Sokoll, and A. Straub, "Dynam ische Sta-
bilitiits Control DSC der Baureihe 7 von BMW - Teil 1/2," Automobiltechnische
Zeitschrijt, vol. 99, no. 3/4, pp . 134-140,208-213, 1997.
[73J R. deGaston and M. Safonov, "Exact calculation of the multiloop stability mar-
gin," IEEE Trans. on Automatic Control, vol. 33, pp . 156-171,1988.
[74J E. Dickmanns, "Entwicklungsschrit te zur Erhohung von Sicherheit und Komfort
durch sehende Autos," Automatisierungstechnik, vol. 44, pp . 243-251, 1996.
[75J T . Djaferis, "Stabilization of systems with real parameter variations," in Proc.
IEEE Conf. Decision and Control, (Tampa), pp. 1870-1871 , 1989.
[76] E. Donges, R. Aufhammer, P. Fehrer, and T . SeidenfuB, "Funktion und Sicher-
heitskonzept der Aktiven Hinterachskinematik von BMW," Automobiltechnische
Zeitschrijt, vol. 10, pp. 580-587, 1990.
[77J E. Donges and K. Naab, "Regelsysteme zur Fahrzeugfiihrung und -stabilisierung
in der Automobiltechnik," Automatisierungstechnik, vol. 44, pp . 226-236, Mai
1996.
[78J J . Doyle and G. Stein, "Multivariable feedback design: concepts for a classical
modern synthesis," IEEE Trans. on Automatic Control, vol. 26, no. 1, pp . 4-16,
1981.
[79) J. C. Doyle, B. A. Francis, and A. R. Tannenbaum, Feedback control theory. New
York: MacMillan Publishing Company, 1992.
[80) H. Duda, Fliegbarkeitskriterien bei begrenzter Stellgeschwindigke it. Ph .D. the-
sis, Technische Universitiit Braunschweig , April 1997. Forschungsbericht 97-15,
Deutsche Forschungsanstalt fiir Luft- und Raumfahrt e.V., Koln.
[81] B. Etkin, Dynamics of atmospheric flight. New York: Wiley, 1972.
[82] B. Etkin and L. Reid , Dynamics of Flight. New York: John Wiley and Sons, Inc.,
1995.
[83] S. Faedo , "Un nuovo problema di stabilita per le equazioni algebriche a coefficienti
reali," Ann. Scuola Norm. Super Pisa, Sci . Fis. Mat., vol. 7, pp. 53-63, 1953.
470 Bibliography

[84] A. Fam and J . Meditch, "A canonical parameter space for linear system design ,"
IEEE Trans. on Automatic Control, vol. 23, pp . 454-458, 1978.
[85] C. Favre, "Modern flight control system a pilot partner towards better safety," in
Proc. of the 2nd International Symposium on Aerospace, Science and Technology,
ISASTI96, (Jakarta), pp. 472-481, 1996.
[86] B. Francis, A course in H oo control theory. New-York: Springer, 1987.
[87] G. Franklin and J. Powell, Digital control. Reading, Massachusetts: Addison-
Wesley, 1980.
[88] S. Franklin and J. Ackermann, "Robust flight control: a design example," AIAA
J. Guidance and Control, vol. 4, pp . 597-605, 1981.
[89] R. Frazer and W. Duncan, "On the criteria for the stability of small motions," in
Proc. Royal Society A, vol. 124, pp. 642-654, 1929.
[90] J. S. Freudenberg and D. P. Looze, "Right half plane poles and zeros and design
tradeoffs in feedback systems," IEEE Trans. on Automatic Control, vol. 30, no. 6,
pp . 555-565, 1985.
[91] F. Friihauf and R. Rutz, "Innovisia - eine aktive Federung fur den Reisebus, "
Automatisierungstechnik, vol. 46, pp. 120-.127, 1998.
[92] M. Fu, "Polyt opes of polynomials with zeros in a prescribed region : New criteria
and algorithms," Systems and Control Letters, vol. 15, pp . 125-145, 1990.
[93] W . Fulton, Algebraic Curves. Reading, Massachusetts: Addison-Wesley, 1989.
[94] F. Gantmacher, The theory of matrices. New York: Chelsea, 1959. .
[95] A. Gelb and W. Vander Velde, Multiple-Input Describing Functions and Nonlinear
System Design. New York: MacGraw-Hill, 1968.
[96] S. Germann and R. Isermann, "Determination of the centre of gravity height of a
vehicle with parameter estimation," in IFAC Symposium on System Identification,
(Copenhagen), 1994.
[97] E. Gilbert, "Controllability and observability in multivariable control systems,"
SIAM Journal of Control and Optimization, pp . 128-151, 1963.
[98] G. Griibel, D. Joos, D. Kaesbauer, and R. Hillgren, "Robust back-up stabilization
for artificial-stability aircraft," in Proc. 14th ICAS Congress, (Toulouse), 1984.
[99] J. Guldner, W . Sienel, H. Tan, J. Ackermann, S. Patwardhan, and T . Bunte,
"Robust automatic steering control for look-down reference systems with front
and rear sensors," IEEE Trans. on Control Systems Technology, vol. 7, no. 1,
pp . 2-11, 1999.
[100] 1. Ciivenc, "Closed loop pneumatic position control using discrete time model
regulation," in Proc. American Control Conference, pp. 4273-4277, 1999.
[101] L. Giivenc, T . Bunte, B. Aksun Guvenc, and D. Odenthal, "Regelung des dy-
narnischen Verhaltens eines Fahrzeugs urn eine definierte Achse (Control of the
dynamic behavior of a vehicle around a defined axis) ." German Patent Appli-
cation 10061966.5, European Patent Application 01129567.2, Priority Dec. 13,
2000.
[102] L. Giivenc and K. Srinivasan, "Friction compensation and evaluation for a force
control application," J. of Mechanical Systems and Signal Processing, vol. 8, no. 6,
pp . 623-638, 1994.
 471

[103] C. Halfmann, H. Holzmann, R. Schwarz, and M. Wiirtenberger, "Identifikation

der Wankdynamik eines Kraftfahrzeugs mit Parameterschatzverfahren und Neu-
ronalen Netzen," Automatisierungstechnik, vol. 46, no. 9, pp. 420-425, 1998.
[104] M. Hautus, "Controllability and observability condition of linear autonomous
systems," Proc. Kon . Ned. Akad. Wetensch., vol. 72, no. ser. A, pp. 443-448,
1969.
[105] M. Hautus, "Controllability and stabilizability of sampled systems," IEEE Trans.
on Automatic Control, no. 17, pp. 528-531, 1972.
[106] F. Hecker, S. Hummel, O. Jundt, K.-D. Leimbach, 1. Faye, and H. Schramm,
"Vehicle dynamics control for commercial vehicles," SAE, no. 973284, 1997.
[107] C. Hermite, "Sur le nombre des racines d'une equation algebrique comprise entre
des limites donnees," J. Reine Angewandte Mathematik, vol. 52, pp. 39-51, 1852.
English translation Int. Journal of Control, 1977.
[108] M. Ho, A. Datta, and S. Bhattacharyya, "Design of P, PI and PID controller for
interval plants," in Proc. American Control Conference, (Philadelphia), pp. 2496-
2501, 1998.
[109] Honeywell Technology Center, "Application of multivariable control theory to
aircraft control laws, Final Report : Multivariable Control Design Guidelines,"
May 1996. WL-TR-96-3099.
[110] H. Hopkins, "Delivered with feeling," Flight International 20, pp. 31-34, 1991.
[Ill] 1. Horowitz, Synthesis of feedback systems. New York: Academic Press, 1963.
[112] H. Hu, C. Hollot, R. Tempo, and J. Ackermann, "Absence of extreme-point results
in sampled-data control systems," in Proc. First European Control Conference,
(Grenoble), pp. 1356-1359, 1991.
[113] A. Hurwitz, "Uber die Bedingungen, unter welchen eine Gleichung nur Wurzeln
mit negativen reellen Teilen besitzt," Mathematische Annalen, vol. 46, pp. 273-
284,1895.
[114] Joint Aviation Authority, "Joint aviation requirements JAR-25 large aeroplanes,"
1994. pp. 1-C-8-1O.
[115] C. Jung and W. Hirschberg, EUROMOTOR Seminar: Understanding human
monitoring and assesment, TNO Delft, November 20-21, 1997, ch. Payload mon-
itoring as one basis for commercial vehicles dynamics. Amsterdam: Swets &
Zeitlinger, 1998.
[116] E. Jury, Theory and application of the z-transform method. New York: John
Wiley, '1964.
[117] E. Jury and T . Pavlidis, "Stability and aperiodicity constraints for systems de-
sign," IEEE Trans. on Circuit Theory, vol. 10, pp. 137-141, 1963.
[118] D. Kaesbauer , "On robust stability of polynomials with polynomial parameter
dependency: two/three parameter case," Automatica, vol. 29, no. 1, pp. 215-217,
1993.
[119] D. Kaesbauer and J . Ackermann, "The distance from stability or gamma-stability
boundaries ," in Proc. 11th IFAC Congress, (Tallinn), pp. 130-134, 1990.
[120] T . Kailath, Linear systems. Englewood Cliffs, N.J.: Prentice-Hall, 1980.
472 Bi bliography

[121] E. Kalinina and A. Uteshev, "Determination of the number of roots of a poly-

nomial lying in a given algebraic domain," accepted for Linear Algebra and Its
Applications, 2000.
[122] R. Kalman, "On the general theory of control systems," in First IFAC Congress,
(Moskow), pp. 481-492, 1960.
[123] R. Kalman, "Mathematical description of linear dynamical systems," SIAM Jour-
nal of Control and Optimization, pp. 152-192, 1963.
[124] C. Kempf and S. Kobayashi, "Disturbance observer and feedforward design for
a high-speed direct-drive positioning table," IEEE Trans. on Control Systems
Technology, vol. 7, no. 5, pp. 513-526, 1999.
[125] V. Kharitonov, "Asymptotic stability of an equilibrium position of a family of
systems of linear differential equations," Differential Equations, vol. 14, no. 3,
pp. 26-35, 1979.
[126] D. Konik, R. Bartz, F. Barnthol, H. Bruns, and M. Wimmer, "Dynamic Drive-
the new active roll stabilization system from BMW Group - system description
and functional improvements," in Proceedings of AVEC 2000, 5th Int 'l Symposium
on Advanced Vehicle Control, (Ann Arbor! Michigan), June 2000.
[127] M. Kordt and J. Ackermann, "Robust design of a structural dynamic engine out
controller for a large transport aircraft ," AIAA Journal of Guidance, Control and
Dynamics 24 (2), pp. 305-314, 200l.
[128] G. Kreisselmeier, "Struktur mit zwei Freiheitsgraden (two-degree of freedom con-
troller structure)," Automatisierungstechnik, vol. 47, pp. 266-269, 1999.
[129] A. Lambregts, "Integrated system design for flight and propulsion control using
total energy principles," in AIAA Aircraft design, systems and technology meeting,
(Texas), 1983. paper no. AIAA-83-256l.
[130] D. Lazard, "System of algebraic equations (algorithms and complexity)," in Sym-
posium on symbolic and algebraic computation, (Bonn), 1991.
[131] A. Leonhard, "Neues Verfahren zur Stabilitatsuntersuchung," Archiv fur Elek-
trotechnik, vol. 38, pp. 17-28, 1944.
[132] A. Lienard and A. Chipart, "On the sign of the real part of the roots of an
algebraic equation," J. Math. Pures et Appl., vol. 10, pp. 291-346, 1914.
[133] R. C. Lin, D. Cebon, and D. J. Cole, "Optimal roll control of a single-unit lorry,"
in Proc. IMechE, vol. 210, Part D, pp. 45-55, 1996.
[134] W. Linvill, "Sampled-data control systems studied through comparison of sam-
pling with amplitude modulation," AlEE Transactions, vol. 70, pp. 1779-1788,
1951.
[135] T . D. Lomax, "Structural loads analysis for commercial transport aircraft: Theory
and practice," AIAA, Education Series , pp. 42-45,143-160, 1995.
[136] L. MacColl, Fundamental theory of servomechanisms, pp. 88-101 . NewYork: Van
Nostrand, 1945.
[137] J . M. Maciejowski, Multivariable Feedback Design. Electronic Systems Engineer-
ing Series, Addison-Wesley, 1989.
[138] M. Mansour, "On Robust Stability of Linear Systems," in Systems and Control
Letters, (North Holland), 1993.
 473

[139J A. Markazie and N. Hori, "A new method with guaranteed stability for discretiza-
tion of continuous-time control systems," in Proc. American Control Conference,
(Chicago) , pp. 1397-1402, 1992.
[140] D. McLean, Automatic Flight Control System. London: Prentice Hall Interna-
tional Ltd ., 1990.
[141] D. McLean, "An automatic engine-out recovery system," in Proc. of Intern. Avi-
ation Safety Conf, IASC 97, (Amsterdam), pp. 763-776, 1997.
[142] D. McRuer, I. Ashkenas, and D. Graham, Aircmjt dynamics and automatic con-
trol. Princeton: Princeton University Press, 1973.
[143] S. Mehring, U. Franke, and A. Suissa, "Optische Spurhaltung - Eine Un-
terstiitzung des Fahrers bei der Lenkaufgabe," Automatisierungstechnik, vol. 44,
pp. 238-242, 1996.
[144] R. Middleton and G. Goodwin, Digital control and estimation. Englewood Cliffs,
N.J.: Prentice-Hall, 1990.
[145] A. Mikhailov, "Method of harmonic analysis in control theory," Avtomatika i
Telemekhanika, vol. 3, pp. 27-81, 1938.
[146] "Flying qualities of piloted airplanes." MIL-F-8785 B (ASG), August 1969.
[147] "Flying qualities of piloted airplanes." MIL-STD-1797A, 1990.
[148] J. Minor, A. Thurling, and E. Ohmit , "VISTA - a 21st centuary UAV testbed,"
in Proc. 32nd Society of Flight Test Engineers Symposium, 2001.
[149] D. Mitrovic, "Graphical analysis and synthesis of feedback control systems,"
Trans. AlEE, vol. 77, no. 2, pp. 476-503, 1958.
[150] M. Mitschke, Dynamik der Kmjtfahrzeuge, vol. C. Berlin: Springer, 1990.
[151] M. Muhler, "Mapping MIMO control system specifications into parameter space,"
submitted to IEEE Conference on Decision and Control, 2002.
[152] M. Muhler, D. Odenthal, and W. Sienel, PARADISE User's Manual. Deutsches
Zentrum fiir Luft und Raumfahrt e. V., DLR Oberpfaffenhofen, 2001.
www.robotic.dlr.de/control/paradise.
[153] A. Muller, W. Achenbach, E. Schindler, T . Wohland, and F.-W . Mohn, "Das
neue Fahrsicherheitssystem Electronic Stability Program von Mercedes Benz,"
Automobiltechnische Zeitschrijt 96, vol. 11, pp . 656-670 , 1994.
[154] R. Munch, "Reglerauslegung und Robustheitsanalyse fiir einen spurgefiihrten
Bus," Tech. Rep. IE 515-86-14, DLR, Oberpfaffenhofen, Germany, 1986.
[155] N. Munro and M. T. Solyemez, "Fast calculation of stabilizing PID controllers for
uncertain parameter systems," in Proc. 3rd IFAC Symposium on Robust Control
Design, (Prague, Czech Republic), 2000.
[156] Y. Neimark, "On th e root distribution of polynomials," Dokl. Akad. Nauk.,
vol. 58, pp. 357-360, 1947.
[157] H. Nyquist , "Regeneration theory," Bell Systems Technical Journal , p. 126, 1932.
[158] D. Odenthal, Ein robustes Fahrdynamik-Regelungskonzept fur die Kippvermei-
dung von K mjtfahrzeugen (A robust vehicle dynami cs control concept for rollover
avoidance). Ph .D. thesis, TU Miinchen, 2002. VDI Fortschritt-Bericht , Reihe 12,
VDI-Verlag, Dusseldorf, 2002.
474 Bibliography

[159] D. Odenthal and P. Blue, "Mapping of frequency response magnitude specifica-

tions into parameter space ," in Proc. 3rd IFAC Symposium on Robust Control
Design, (Prague, Czech Republic), 2000.
[160] D. Odenthal and T . Bunte, "Verfahren zUltReduktion der Kippgefahr von Strafen-
fahrzeugen (Method of rollover avoidance for road vehicles)." German Patent P
19918597 C2, European Patent Application, Priority April 23, 1999.
[161] D. Odenthal, T . Bunte, and J. Ackermann, "Nonlinear steering and braking con-
trol for vehicle rollover avoidance," in Proc. European Control Conference, (Karls-
ruhe, Germany), 1999.
[162] D. Odenthal, T. Bunte, and R. Koeppe, "Regelung von Systemen mit
kinasthetischer Kopplung (Control of systems with kinesthetic coupling)." Ger-
man Patent Application 10216247.6, Priority April 12, 2002.
[163] K. Ohnishi, "A new servo method in mechatronics," Trans. Japanese Soc. Electr.
Eng., vol. 107-D, pp . 83-86, 1987.
[164] L. Orlando, "SuI problema di Hurwitz relativo aIle parti reali delle radici di
un'equazione algebraica," Mathematische Annalen, vol. 71, pp. 233-245, 1911.
[165] M. Otter, "Objektorientierte Modellierung physikalischer Systeme," Automa-
tisierungstechnik, vol. 47, 17 parts from 1/1999 to 12/2000. Series "Theorie fur
den Anwender".
[166] L. Palkovics, A. Semsey, and E. Gerum, "Roll-over prevention systems for com-
mercial vehicles - additional sensorless function of the electronic brake system, "
Vehicle System Dynamics, vol. 32, pp. 285-297, 1999.
[167] E. Panier, M. Fan , and A. Tits, "On the robust stability of polynomials with
no cross-coupling between the perturbations in the coefficients of even and odd
powers," Systems and Control Letters, vol. 12, pp. 291-299, 1989.
[168] V. Popov, "Absolute stability of nonlinear systems of automatic control," Au-
tom .& Rem. Control, vol. 22, pp. 857-875, 1962.
[169] F . Preparata and M. Shamos , Computational geometry. New York: Springer,
1985.
[170] P. Putz and M. Wozny, "A new computer graphics approach to parameter space
design of control systems," IEEE Trans. on Automatic Control, vol. 32, pp. 294-
302,1987.
[171] J. Raisch, Mehrgropenregelung im Frequenzbereich. Miinchen: Oldenbourg Verlag,
1994.
[172] P. Riekert and T . Schunck, "Zur Fahrmechanik des gummibereiften Kraft-
fahrzeugs," Ingenieur Archiv, vol. 11, pp. 210-224, 1940.
[173] J. Rissanen, "Control system synthesis by analogue computer based on the gen-
eralized linear feedback concept," in Proc. Int. Seminar on Analog Computation,
(Brussels , Belgium), Nov. 1960.
[174] E. Routh, A treatise on the stability of a given state of motion. London: MacMil-
lan, 1877.
[175] R. Saeks and J. Murray, "Fractional representation, algebraic geometry, and the
simultaneous stabilization problem," IEEE Trans. on Automatic Control, vol. 27,
pp . 895-903, 1982.
 475

[176] I. Schur, "Uber Potenzreihen, die im Inneren des Einheitskreises beschrankt sind,"
Journal fUr Mathematik, vol. 147, pp. 205-232, 1917.
[177] I. Schur, "Uber Potenzreihen, die im Inneren des Einheitskreises beschrankt sind,"
Journal fur Mathematik, vol. 148, pp. 122-145, 1918.
[178] I. Schwamm, "Analytische und simulative Untersuchungen zur Verbesserung der
robusten Lenkregelung," Tech. Rep. IE 515-97-14, DLR, Oberpfaffenhofen, Ger-
many, 1997. Betreuer J. Ackermann und T . Bunte.
[179] L. Segel, "Theoretical prediction and experimental substantiation of the response
of the automobile to steering control," in IMechE, pp. 310-330, 1956-1957.
[180] A. Sideris and R. deGaston, "Multivariable stability margin calculation with
uncertain correlated parameters," in Proc. IEEE Conf Decision and Control,
(Athens), pp. 766-771, 1986.
[181] A. Sideris and R. S. Pefia, "Fast computation of the multivariable stability margin
for real interrelated uncertain parameters," IEEE Trans. on Automatic Control,
vol. 34, pp. 1272-1276, 1989.
[182] W. Sienel, "Algorithms for tree structured decomposition," in Proc. IEEE Conf
Decision and Control, (TUcson), pp. 739-740, 1992.
[183] W. Sienel, Analyse und Entwurf von robusten Regelungssystemen durch Kon-
struktion von komplexen Wertemengen (Analysis and design of robust control
systems by construction of complex value sets) . Ph.D. thesis, Technische Uni-
versitat Miinchen, 1994. VDI Fortschritt-Bericht, Reihe 8, Nr. 476, VDI-Verlag,
Dusseldorf, 1995.
[184] W. Sienel, "Robust decoupling for active car steering holds for arbitrary dy-
namic tire characteristics," in Proc. Third European Control Conference, (Rome),
pp. 744-748, 1995.
[185] O. Smith, Feedback Control Systems. New York: McGraw-Hill, 1958.
[186] C. Soh, C. Berger, and K. Dabke, "On the stability properties of polynomials with
perturbed coefficients," IEEE Trans. on Automatic Control, vol. 30, pp. 1033-
1036, 1985.
[187] K. Sondergeld , "A generalization of the Routh-Hurwitz stability criteria and an
application to a problem in robust controller design," IEEE Trans. on Automatic
Control, vol. 28, pp. 965-970 , 1983.
[188] M. Spillman, P. Blue, L. Lee, and S. Banda, "A robust gain scheduling example
using linear parameter varying feedback," in Proc. 13th IFAC World Congress,
(San Francisco), 1996.
[189] B. Stevens and F. Lewis, Aircraft control and simulation. New York: Wiley, 1992.
[190] M. Stieber, Das Konzept des Tandemreglers zur automatischen Spurfuhrung nicht
schienengebundener Fahrzeuge. MAN, Internal Report B 094002-EDS-005, 1980.
[191] Y. Tsypkin and B. Polyak, "Frequency domain criteria for IP-robust stability of
continuous linear systems ," IEEE Trans. on Automatic Control, vol. 36, no. 12,
pp. 1464-1469, 1991.
[192] Y. Tsypkin and B. Polyak, "Frequency domain criteria for robust stability of poly-
tope of polynomials ," in Control of uncertain dynam ic systems (S. Bhattacharyya
and L. Keel, eds.), pp. 491-499, CRC Press, 1991.
476 Bibliography

[193] T . Umeno and Y. Hori, "Robust speed control of DC servomotors using modern
two degrees-of-freedom controller design," IEEE Trans. Ind . Electron ., vol. 38,
no. 5, pp. 363-368, 1991.
[194] J. Uspensky, Theory of Equations. New York: McGraw-Hill, 1948.
[195] A. v. Zanten, R. Erhardt, and G. Pfaff, "FDR - die Fahrdynamikregelung von
Bosch," Automobiltechnische ZeitschTift, vol. 96, pp. 674-689, 1994.
[196] P. Varaiya, "Smart cars on smart roads : Problems of control," IEEE Trans. on
Automatic Control, vol. 38, pp. 195-207, 1993.
[197] A. Vicino, A. Tesi, and M. Milanese, "An algorithm for nonconservative stability
bounds computation for systems with nonlinearly correlated parametric uncer-
tainties," IEEE Trans. on Automatic Control, vol. 35, pp. 835-841, 1990.
[198] M. Vidyasagar, Nonlinear systems analysis. Englewood Cliffs, N.J .: Prentice-Hall,
1978.
[199] I. Vishnegradsky, "Sur la theorie generale des regulateurs," Compt. Rend. Acad.
Sci., vol. 83, pp. 318-321, 1876.
[200] D. Siljak, Nonlinear systems: the parameter analysis and design. New York:
Wiley, 1969.
[201] E . Walach and E. Zeheb, "Generalized zero sets of multiparameter polynomials
and feedback stabilization," IEEE Trans. on Circuits and Systems, vol. 29, pp. 15-
23, 1982.
[202] L. Zadeh and C. Desoer, Linear system theory: the state space approach. New
York: MacGraw-Hill,1963.
[2031 K. Zhou, J . Doyle, and K. Glover, Robust and optimal control. Upper Saddle
River, NJ : Prentice-Hall, 1996.
[204] A. Zomotor, Fahrwerktechnik: Fahrverhalten . Wiirzburg: Vogel-Verlag, 1987.
Index

13-specifications, 231 automatic car steering, 180, 181, 272

13-stability, 131, 149, 152, 230
a-transform, 429 bandwidth, 60, 68, 143, 314, 325
f-editor, 463 Barmish, BoR., 335, 438
Ivstability, 43, 59, 66, 67, 230, 256, 356 barycentric coordinates, 353
a-stability, 104, 107, 117 Bhattacharyya, So, 48
a-stability margin, 222 Bialas, So, 353
jL-split braking test, 173, 192 bicycle model, 180
jL-synthesis, 328, 331 bilinear coefficient, 98, 358, 363, 434,
448
absolute stability, 114, 261 bilinear transformation, 95, 428, 445
accelerometer, 13, 189, 206, 286 bisection, 364, 370, 391
accelerometer feedback, 210 Bode magnitude plot , 133, 165
Ackermann angle, 201 Bode's integral theorem, 148, 230
Ackermann steering, 201 boundary crossing, 67, 425
Ackermann 's formula, 37 boundary crossing theorem, 29, 34, 36,
Ackermann, r., 38, 82, 186, 211, 216, 67, 339, 405, 426
438 boundary representation theorem , 434
active boundary, 32, 52, 71, 448 Boyd, So, 166
actuator, 16, 27, 60, 173, 223, 279, 303, braking, 171, 173, 261, 268
422 branching point , 73
actuator constraints, 22, 141, 215 Butterworth pole configuration, 87, 280
actuator model, 253
actuator uncertainties, 278 calculation modules, 462
additive perturbation, 27, 143 cancellation, 9, 60, 64, 292
admissible solution set, 36, 308 cancellation in the argument of the func-
affine coefficients, 349, 400, 404, 428, tion,292
437, 448 canonical form, 10
affine polynomial family, 351 cascade structure, 210
affine transformation, 445 Cauchy's principle of the argument, 334
aircraft dynamics, 286 Cayley-Hamilton theorem, 38, 39
algebraic curves, 448, 449 center of eigenvalues, 60
algebraic problem formulation, 98, 426 characteristic polynomial family, 7, 14,
analytical modelling, 2 201, 207, 374
Anderson , a.n.o., 344 characteristic velocity, 199
ant i-aliasing filter, 60, 421 Chipart, Ao, 21
approximation , 419 chord approximation, 364, 380
478 Bibliography

circle, 94, 424 Curtis, 335

circle stability, 444 curvature, 200, 273
coefficient function, 84, 431, 437
Cohn, A., 427 damping, 60, 61, 65, 68, 88, 202, 212,
color coding, 37, 76 305,421
compensator, 56, 65 dead time, 25, 442
complementary sensitivity function, 26, deadbeat solution, 421, 425
136, 147, 224 decoupled lateral subsystem, 188
complex root boundary (CRB), 29, 34, decoupled yaw subsystem, 188
50, 69, 82, 426 decoupling point, 178, 196, 291
complexity, 365, 378, 410, 442 describing function , 107, 108, 218
computer algebra, 416 design, 75, 358, 431, 448
computer animation, 371 design process, 59
concurrent engineering, 2, 36 Desoer mapping theorem, 367, 393
conformal mapping, 95 Desoer, C., 335
digital controller, 417
conic section boundaries, 81, 96, 119
construction of value sets, 370 dilation factor, 395
direct rollover avoidance, 241
contact force, 171
direct transmission, 133 -
continued fractions, 24
discontinuity in the distance function,
contour algorithm, 374
407
contractable, 436
discrete-time frequency response, 440
control architecture, 323
discriminant, 88, 453
controllability, 10, 12, 37, 74, 172, 197,
disjoint solution sets, 86
204, 252, 276, 420
displacement, 275
controllability Gramian, 161
distance, 405, 407
controller discretization, 417
disturbance attenuation, 26, 138, 147,
controller structure, 23, 215, 290, 294,
177,209,228,230
307
disturbance observer, 223, 227, 323
convex, 73
disturbance torque, 173, 177, 247, 289
convex hull, 352, 436, 446
DLR plate experiment, 391
convex hull of the stable region, 429, 431 domain splitting algorithms, 369
convex hull of the value set, 367 dominant poles, 60, 64, 87
convex polygons, 48, 92 double cancellation, 215
coordinate transformation, 94 double root, 69, 73, 88, 167, 432
cornering stiffness, 192 double tangent Popov-line, 459
corrective angle, 184 driver, 180
corrective rudder deflection, 297 driver assistant systems, 173, 181
corrective slip angle, 185 driver reaction time, 173
corrective steering angle, 206 driving tests, 218
crane, 2, 9 dual locus method, 107, 220
Cremer, L., 335 Duncan, W., 29, 335
critical a-stability condition, 122 dynamic model uncertainty, 27
critical frequency, 410 dynamic order of the controllers, 279
critical stability conditions, 29
crossing the imaginary axis, 29, 32 edge polynomial family, 350
 479

edge theorem, 349, 352, 393, 437 gain-reduction margin, 316

eigenvalue, 18, 41, 74, 203, 305, 358, 419 gain-scheduled controller, 76, 77, 225,
eigenvalue regions, 24 318,331
eigenvalue specifications, 68, 105 generalized frequency, 80
eigenvalue tracing, 73 generalized singular frequencies, 90
electric RLC network, 371 generation of mapping equations, 37
engine out, 286 geometric approach, 405
envelope, 456 Gilbert, Eo , 10
equation generator, 462 graphical frequency domain criteria, 396
even-odd decoupling, 349 graphical user interfaces, 37, 461
existence of robust controllers, 74 graphical warning on closeness to insta-
exponential parameter dependence, 419 bility, 364
exponential parameter dependency, 421, gray-tone coding, 168
429 gridding, 66, 72, 78, 100, 338, 358, 364,
extended state space model, 274 371, 374, 392, 428
gyro, 286
fading integrator, 213, 221 gyro or accelerometer faults, 314
Faedo, So, 342
Fam , s; 430 H 2-norm, 160
family of Popov plots, 388 Hoc design specifications, 318, 328
fault detection, 312 Hoc-norm, 132, 147, 165, 331
fault tolerance, 310 Hamiltonian matrix, 166
feasible transfer function, 25, 139 handling, 212, 288, 293, 322
feedback of the yaw rate, 289 harmonic linearization, 108
feedback perturbation, 27, 143 Hautus, Mo, 74
feedback structure, 12, 75, 185 Herm ite , Co , 19
fictitious boundary, 32, 33, 123 Hermite-Biehler theorem, 48, 334
finite plant family, 54 hold element, 417
finite test, 356 Hollot, C., 349
fixed eigenvalues, 37, 40 Huang, L., 349
flight conditions, 305 Hurwitz determinant, 19, 326, 392
four wheel steering, 180 Hurwitz, A., 19
Fourier series, 108 hyperbola, 68, 76, 280, 357
fragility, 71 hyperplanes, 59, 405
Frazer, n., 29, 335
frequency band, 369 ideal mass distribution, 198
frequency domain criteria, 76, 104 immediate reaction, 212
frequency endpoint condition, 122, 128 indirect rollover avoidance, 241
frequency loci specifications, 104, 121 infinite locus magnitude condition, 122,
frequency range, 375 128
frequency response, 27 infinite root boundary (IRB) , 29, 50
frequency response magnitude, 104, 132 inner bound, 370
friction coefficient , 195, 198 input sensitivity function , 136
Fu , Mo, 352 input-output modelling, 2
interactive modules, 461
gain , 159 interlacing property, 337
480 Bibliography

internal r -stability, 60 Mansour, M., 344, 398

internal stability, 26, 137 mapping solver, 462
intersection of sets, 76, 86, 95, 129, 154, Mason's formula, 384
258,463 mass-spring-damper elements, 371
intersection points of algebraic curves, MATLAB, 462, 463
454 measurement noise, 22
interval plants, 348 Meditch, J ., 430
interval polynomials, 341, 396 Mikhailov plot, 334, 344
invariance plane, 40, 464 Mikhailov set, 340
inverse Laplace transform, 62 Mikhailov, A., 335, 396
inversion of a complex set, 382 MIMO systems, 16, 158, 431
isolated unstable point, 362, 364 Mitrovic, D., 36
model reference, 25, 138
Jacobian, 98, 362, 363, 366, 379 modular toolbox, 461
Joint-Aviation-Requirements, 288 momentary pole, 182
junctions, 119 monic polynomial, 6, 16, 18
Jury, E.!., 344, 425, 427 multi-model approach, 151, 418
multilinear coefficient functions, 15, 16,
Kaesbauer, D., 82 336, 359, 362, 431
Kalman, R., 10, 37 multiplication of value sets, 378
Kamm circle, 171 multiplicative perturbation, 27, 141, 147
Kharitonov's theorem, 341, 393, 398 multiplicity, 419
Kharitonov, V., 342 multivariate polynomials, 448
Munro, N., 48
lane keeping, 243
lane reference, 272 naturalfrequenc~ 61,202,212,305,421
Laplace transform, 208, 373 necessary condition, 370
largest stable box, 404 Neimark, Y., 36
largest stable hypersphere, 404 noise rejection, 138, 148, 228
lateral acceleration, 200, 291 nominal performance, 326
lateral subsystem, 209 nominal plant model, 27
lateral tire forces, 174, 192 nominal polynomial, 87
lean controller structure, 24, 215, 281 non-active boundary, 32
Leonhard, A., 335 non-linear, 14, 297
Leverrier-Faddejew algorithm, 7, 15, 16 non-linear coefficient functions, 96, 98,
Lienard, A., 21 362, 370, 371, 393
Lienard-Chipart conditions, 32 non-linear simulation, 290, 318, 331
limit cycle avoidance, 220 non-linear transformation, 96
limit cycles, 107,215,217,237 non-minimum phase zeros, 25, 64
linear-quadratic-regulator, 161 non-observable, 9, 13, 41, 177, 186, 289
linearization, 3, 241, 247, 273 non-parametric uncertainty, 27
load alleviation, 287, 297 normal vector, 405
logarithmic spiral, 423 norms, 160
longitudinal dynamics, 319 number of edges, 353, 358
low-pass filtered differentiation, 22 numeric optimization methods, 36
low-pass properties, 108, 220 Nyquist plot, 48, 105, 117, 121
 481

Nyquist value set, 440 plant disturbance, 6, 26

Nyquist, H., 140, 145, 165, 335, 382 plant family, 75
Nyquist-stability margin, 140 point condition, 122, 123, 153
Poisson sum form, 441
observability, 10, 12, 13, 74, 205, 252, pole placement, 37, 67, 74
277, 420 pole placement matrix, 39, 41
observer, 191 pole region assignment, 68
octahedron, 437 Polyak, B., 397, 400
Olbrot, A., 349 polynomial coefficient, 359
one-to-one correspondence of the eigen- polynomial dependence, 100, 121, 358,
values, 73, 308 362
operating domain, 6, 12, 249 polynomial equations, 448
order of the controller, 14 polynomial family, 27, 334, 349, 395
Orlando's formula, 31, 426 polytope, 352, 430
orthogonal, 405 Popov criterion, 113, 191, 224, 237
orthogonalization procedure, 406 Popov plot , 115
oscillations, 61 Popov value sets, 382
output feedback, 14, 66 Popov-line, 115, 130, 237
output sensitivity function, 137 Popov-sector, 115
overbounding, 348 prefilter , 65, 186, 212, 290, 309
overshoot, 60, 61, 66 product decomposable , 381
oversteering, 202 projected r-stability boundaries, 78

Panier, E., 349 rank , 57

paper-and-pencil graphic methods, 36 rate limited, 111, 218
PARADISE, 461 rate sensor, 13, 279
parallel accelerometers, 314 reaction time, 286, 288, 299
parallel gyros, 314 real radius of stability, 440
parallelepiped, 350 real root boundary (RRB), 69
parameter space approach, 33, 36, 37, real root boundary (RRB), 29, 34, 49,
48, 67, 76, 78, 107, 266 100, 422, 425
parameter space control design, 229, 325 realizability, 26, 48
parametric state space representation, rear-wheel steering, 210
463 reduced number of uncertain parame-
parpolygon, 352, 356 ters, 196
partition of the uncertainties, 380 reduced order observer, 14
path tracking, 210, 279 reduction procedure, 435
Pavlidis, T ., 425 reference path, 273
performance, 318 regulator transfer function , 135
performance weight, 326 relative degree, 8, 25, 60, 133, 147, 213,
phase angle, 27, 61, 191, 336 324, 420, 442
phase margin, 159, 176, 177, 219, 442 representative operating conditions, 36,
PID-control, 47, 78, 92, 101 54, 55, 418
pitch, 286 required control input, 138
pitch-rate, 319 resultant method, 30, 81, 100, 103
planar model , 290 road curvature, 279
482 Bibliography

robot in free motion, 371 Siljak, D., 36, 88

robust controllability, 12 simply connected, 436
robust controller design, 150 simulation, 234, 270, 283, 286, 297
robust limit cycle avoidance, 218 SIMULINK, 462
robust observability, 12 simultaneous stabilizers, 48
robust performance, 146 simultaneously r -stabilized, 71
robust pitch stabilization, 286 simultaneously stabilized, 75
robust stability, 144, 326 single-track model, 180, 192, 243
robust stability analysis, 418, 431, 448, singular frequency, 44, 49, 98, 100, 360,
463 401
robust unilateral decoupling, 180, 185, singular points, 457
286, 293 singular value, 165, 168
robustness analysis, 55, 72, 150, 334 singularity of value sets, 360
roll damping, 254, 256 skidding avoidance, 174, 215
roll motion, 286, 289, 290 slices, 78
rollover avoidance, 238, 261 sluggish responses, 60
rollover coefficient, 239 small-gain theorem, 145, 165
root finding algorithm, 66 smallest destabilizing perturbation, 395,
root locus, 9, 60, 73, 203, 280, 356 410
root set , 7, 49, 66, 392, 425 specifications, 68, 132
Routh, E ., 19 stability, 172, 191, 202, 250
ruled surface, 434 stability analysis, 22
sampled-data control system, 417 stability profile, 389, 392, 393
sampling interval, 417, 420, 421 stability radius, 395, 408
saturation, 110, 171, 221 stability regions, 428
scaling, 61, 404, 420 stabilizing feedback, 13
Schur , I., 427 stable neighborhood, 29
Schur-stability, 421, 425, 426, 437 stable test point, 51
sensitivity analysis, 233, 260 state space model, 1, 4, 14, 195, 207,
sensitivity function, 26, 136, 147, 255 248, 296, 304
sensor, 13, 16, 27, 307 stationary cornering, 199
sensor coordinates, 303, 311 stationary flight, 286
sensor dynamics, 141 steady state behavior, 64, 133, 139, 199,
sensor fault , 286, 311 224
sensor noise, 25, 26, 325 steer-by-wire , 174, 192, 224
sensors, 312 steering, 171, 256, 261
separation, 14, 180 steering actuator model, 224
sequential pole shifting, 37, 40, 78 step response, 59, 65
set operations, 378 stepwise change, 172
shear forces, 293 stepwise design procedure, 16, 40
short-period longitudinal mode, 65, 73, strict contraction, 425
286, 303, 318, 319 strictly proper controller transfer func-
side forces , 181 tion, 421
side wind test, 192 structural dynamics, 288, 293, 305
sideslip angle, 174, 273, 290, 301 structural robustness, 190, 210
 483

structured singular value, 329 unilateral decoupling, 177, 289

subpolynomial, 24, 56, 381 union of sets, 86, 463
subpolytopes, 412 unit circle, 421, 434
subsystem, 60, 191 unmodelled uncertainties, 27
sufficient condition, 370 unstable island, 98
sum decomposable, 381 unstructured uncertainty, 27, 141, 321
symbolic computer programs, 37
system performance, 36 value set, 334, 335, 339, 363, 393
variance, 160
tangent condition, 122, 126, 153, 411, vehicle sideslip angle, 180
458 vertex polynomial, 350, 430
theorem of Bezout, 414, 454 Vicino, A., 410
time scaling, 61 Vishnegradsky, I., 36
time-domain specifications, 59, 68
weighted sensitivity, 224
tire side force, 175
weighting function, 141
tire slip angle, 174
wheelbase, 173, 180
tires, 171
tomographic rendering, 48, 52, 78 yaw angle , 180, 273, 286
trace of a matrix, 8 yaw damping at high velocities, 215
track-guided bus, 44, 102, 414 yaw disturbance torque, 287
tracking, 138, 148 yaw motion, 65
trackwidth, 174 yaw rate, 174, 180, 200, 288, 290, 291
transfer function , 25, 203, 224, 314, 384 yaw rate feedback, 278, 284
transfer function matrix, 1, 8, 249 yaw rate reference, 206
transformed PID parameters, 95 yaw rate sensor, 206
transition frequency, 143, 351 yaw subsystem, 209
transparency of controller structure, 14,
24, 74, 210, 215, 281 z-transfer function , 417
transparent design , 331 Zadeh , L., 335
trapezoidal rule for integration, 418 zero exclusion, 334, 335, 340, 352, 370,
tree-structure, 362, 370, 393, 439 393
tree-structured decomposition, 380, 382
tree-structured value set, 370
Tsypkin, Y. , 397, 400
Tsypkin-Polyak loci, 396
tuning parameters, 14
Tustin approximation, 417
two-mass systems, 24

uncertain aerodynamics, 289

uncertainty, 3, 146, 321
uncertainty structure, 27
uncertainty weight, 326
undecomposable, 381
underbounding, 348
understeering, 202