Professional Documents
Culture Documents
Design of Linear Multivariable Feedback Control Systems The Wiener Hopf Approach Using Transforms and Spectral Factorization Joseph J. Bongiorno JR
Design of Linear Multivariable Feedback Control Systems The Wiener Hopf Approach Using Transforms and Spectral Factorization Joseph J. Bongiorno JR
https://textbookfull.com/product/multi-stage-and-multi-time-
scale-feedback-control-of-linear-systems-with-applications-to-
fuel-cells-verica-radisavljevic-gajic/
https://textbookfull.com/product/linear-canonical-transforms-
theory-and-applications-1st-edition-john-j-healy/
https://textbookfull.com/product/networked-control-systems-with-
intermittent-feedback-1st-edition-hirche/
https://textbookfull.com/product/classical-feedback-control-with-
nonlinear-multi-loop-systems-with-matlab-and-simulink-3rd-
edition-boris-j-lurie/
Linear Feedback Controls: The Essentials 2nd Edition
Mark A. Haidekker
https://textbookfull.com/product/linear-feedback-controls-the-
essentials-2nd-edition-mark-a-haidekker/
https://textbookfull.com/product/theory-of-stabilization-for-
linear-boundary-control-systems-1st-edition-nambu/
https://textbookfull.com/product/math-302-lecture-notes-linear-
algebra-and-multivariable-calculus-kenneth-l-kuttler/
https://textbookfull.com/product/the-great-facilitator-
reflections-on-the-contributions-of-joseph-f-hair-jr-to-
marketing-and-business-research-barry-j-babin/
https://textbookfull.com/product/factorization-of-boundary-value-
problems-using-the-invariant-embedding-method-jacques-henry-and-
angel-m-ramos-auth/
Joseph J. Bongiorno Jr.
Kiheon Park
Design of Linear
Multivariable
Feedback Control
Systems
The Wiener–Hopf Approach using
Transforms and Spectral Factorization
Design of Linear Multivariable Feedback Control
Systems
Joseph J. Bongiorno Jr. Kiheon Park
•
Design of Linear
Multivariable Feedback
Control Systems
The Wiener–Hopf Approach using
Transforms and Spectral Factorization
123
Joseph J. Bongiorno Jr. Kiheon Park
Department of Electrical and Computer College of Information
Engineering, NYU Tandon and Communication Engineering
School of Engineering Sungkyunkwan University
Polytechnic Institute, New York University Suwon-si, Korea (Republic of)
Brooklyn, NY, USA
Mathematics Subject Classification (2010): 93-01, 93C05, 93C15, 93C35, 93C55, 93C57, 93C62,
93D09, 93D15, 93E20, 47A68
This Springer imprint is published by the registered company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
This book is dedicated to Dante C. Youla our
mentor and colleague for his inspiration,
encouragement, and collaboration.
Preface
This book is based on some of the research done by the authors and their colleague,
Dante C. Youla, over a period that spanned almost 50 years beginning with Youla
(1961) and continuing most recently with Park and Bongiorno (2009). Specifically,
their work on the design of analog optimal time-invariant linear multivariable
feedback control systems using transforms and the Wiener–Hopf methodology is
presented within a unified framework for both the analog and the digital case. The
authors are especially grateful to Dante C. Youla for helping us to resolve a number
of mathematical issues connected with the material presented in this book.
The focus is on feedback control systems that can be modeled as an intercon-
nection of subsystems each of which is specified except for one that is chosen by
the designer and called the controller. Attention is restricted to those systems in
which all subsystems can be modeled with linear constant-coefficient differential
equations (analog systems) or to those in which all subsystems can be modeled with
linear constant-coefficient difference equations (digital systems). Exogenous inputs
that are deterministic and persistent (steps, ramps, etc.) and/or ones that are
stochastic are considered. In the book, a quadratic functional is used to measure
system performance. The design task is to choose and synthesize the controller
transfer matrix so that the desired overall system performance is achieved. The
relationship of the works of others to the work presented here is described
throughout the book in sections entitled “Historical Perspective and Discussion”.
A key first step in the design process is the parameterization of the class of
stabilizing controllers, which is the topic of Chap. 2. Larin, Naumenko, and Suntsev
(1971, 1972, 1973) were the first to solve this problem. However, this achievement
was not recognized at the time and solutions were also provided in Kučera (1974)
and Youla, Jabr, and Bongiorno (1976b) with the parameterization in the latter
attracting particular notoriety. In order to apply the Wiener–Hopf methodology
when tracking of persistent deterministic reference inputs and rejection of persistent
deterministic disturbance inputs is required, a refinement of the parameterization of
stabilizing controllers is needed. In particular, with these inputs the subset of sta-
bilizing controllers for which the error variables and the controller output variables
are stable must be employed. The required parameterization is presented in Chap. 3
vii
viii Preface
and used to obtain the optimal controller with respect to a quadratic performance
functional for the standard configuration. In addition, a methodology is described
that enables one to trade off optimality for improved stability margin and for
reduced sensitivity to plant model uncertainty. Results are also provided for
one-degree-of-freedom (1DOF) and three-degree-of-freedom (3DOF) systems.
Chapter 3 contains material that first appeared in Park and Bongiorno (2009).
Optimal design of decoupled systems is addressed in Chap. 4 which contains results
not previously published. The numerical calculation of the optimal controller
parameters and realization of the controller are treated in Chap. 5.
With regard to applications, three examples are provided. Example 3.23 con-
cerns the optimal H2 design of a 1DOF depth and pitch controller for a submerged
submarine traveling at 30 knots speed using the linearized model for submarine
motion given in Sect. 10.4 of Grimble (2006). Example 3.24 concerns the optimal
H2 design of a 2DOF controller for a Rosenbrock process: i.e., a multivariable
process whose transfer matrix exhibits a severe coupling phenomenon (Åström
et al. 2002). Specifically, the transfer matrix has a nonminimum phase zero at s ¼ 1
and this yields fundamental limitations on control system performance. The design
of an optimal decoupling 2DOF H2 controller is determined in Example 4.10 for
the same Rosebrock process so that the price paid in performance for imposing the
decoupling requirement could be assessed.
This book is intended for readers familiar with linear systems and matrix theory.
In particular, the reader should be familiar with the material in Appendix A and
enough matrix theory to follow the presentation in Appendix B. Many examples
with solutions are provided which contain important insights, extensions, and/or
elaborations on the topics treated in the book. Examples of particular importance
are identified with an asterisk.
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ............ 1
1.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ............ 1
1.2 System Models . . . . . . . . . . . . . . . . . . . . . . . . . ............ 2
1.3 Stability, Signal Models, Performance Measures,
and Saturation Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.4 Notation and Terminology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.5 Historical Perspective and Commentary . . . . . . . . . . . . . . . . . . . 13
1.6 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2 Stabilizing Controllers, Tracking, and Disturbance Rejection . . . . . . 23
2.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.2 Polynomial Matrix Descriptions . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.3 Stability Analysis of the Standard Configuration . . . . . . . . . . . . . 26
2.4 Stability Analysis of the 2DOF Standard Configuration . . . . . . . . 42
2.5 Stability Analysis of the 3DOF System . . . . . . . . . . . . . . . . . . . 48
2.6 Tracking and Disturbance Rejection in the 3DOF System . . . . . . 55
2.7 Parameterization and Realization of Proper Stabilizing
Controllers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ..... 75
2.8 Historical Perspective and Commentary . . . . . . . . . . . . . . ..... 84
2.9 Additional Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . ..... 85
3 H2 Design of Multivariable Control Systems . . . . . . . . . . . . . . . . . . 115
3.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
3.2 Preliminary Considerations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
3.3 Tracking and Disturbance Rejection with the Standard
Configuration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
3.4 Tracking and Disturbance Rejection in the 3DOF System
Revisited . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
3.5 H2 Design of Digital Multivariable Control Systems . . . . . . . . . 133
3.6 H2 Design of Analog Multivariable Control Systems . . . . . . . . . 139
ix
x Contents
1.1 Overview
The design of analog and digital feedback control systems is addressed in this book.
The focus is on control systems that can be modeled as an interconnection of
subsystems each of which is specified except for one, called the controller, which is
to be chosen by the designer. Associated with each subsystem is a set of signals that
are the elements of a designated vector called the input and which generate
responses that are the elements of a designated vector called the output. The input
and/or output vectors can contain one or more elements; hence, the subsystems are
multivariable ones that include single-input-output (single-variable) subsystems as
a special case. Also associated with each subsystem is a transfer matrix, which
connects the input vector to the contribution it makes in the output vector.
The symbol used to designate the transfer matrix of a subsystem is shown within
the block representing it in the system block diagram. The design task is to choose
and synthesize the controller transfer matrix so that the desired overall system
performance is achieved. In this book, quadratic measures of system performance
are used exclusively. The essential elements of complex function theory, trans-
forms, and stochastic processes needed are summarized in Appendix A. The
essential elements of matrix theory needed are reviewed in Appendix B.
When dealing with analog systems, attention is restricted to subsystems that can
be modeled by a system of ordinary linear differential equations with real constant
coefficients. As a consequence, all analog subsystem transfer matrices contain
elements that are ratios of polynomials in the Laplace variable s with real coeffi-
cients or, equivalently, real rational functions of s. The traditional transform used by
engineers for digital systems is the z-transform. It turns out for the problems of
interest here that it is more convenient to use the k-transform, where k ¼ 1=z. When
dealing with digital systems, attention is restricted to subsystems that can be
modeled by a system of ordinary linear differential equations with real constant
The block diagram of an analog control system that incorporates most of the
features one might ever encounter in practice is shown in Fig. 1.1. All symbols
designate Laplace transforms and the sizes of all column vectors representing the
inputs and outputs in the system are consistent with the sizes of the subsystem
transfer matrices. The subsystem that is to be controlled is referred to as the plant
and the vector d(s) represents the environmental disturbances that impact the plant.
The vector v(s) represents the variables that are available to control the plant.
Whenever feasible and economical, a sensor capable of measuring some or all of
the environmental disturbances impacting the plant is included in the feedforward
path as shown. The matrix L(s) is the sensor transfer matrix and the vector nl ðsÞ
accounts for sensor error. The feedforward path and the controller provide a
mechanism for modifying the plant input v(s) so as to suppress the impact of the
environmental disturbances d(s) on the plant output vector y(s). The same is true for
the feedback path that contains a sensor that measures the plant output y(s). The
matrix F(s) is the sensor transfer matrix and the vector nm ðsÞ accounts for sensor
error.
Often it is not physically possible or economical to measure the disturbance
vector d(s). The task of suppressing the impact of disturbances on system perfor-
mance in those cases must be accomplished only with the aid of the feedback path.
Feedback also plays an essential role in reducing the sensitivity of system perfor-
mance to changes in the plant transfer matrix P(s) and is essential for the control of
unstable plants. The block with the transfer matrix Td ðsÞ does not represent a
1.2 System Models 3
nl (s ) Feedforward
+ d (s )
l( s )
L (s )
+
Path
z( s ) Plant Pd (s )
n(s )
+ +
r (s ) + u (s) v (s ) + y (s ) ε (s )
C (s ) P (s)
Controller
+
yd(s)
w( s ) Feedback
m ( s)
F (s )
+
n m (s ) +
Path
Desired
Td ( s )
Tracking
Fig. 1.1 Three-degree-of-freedom system (© 1990 Taylor and Francis Ltd. http://www.
tandfonline.com. Reprinted, with permission, from Park and Bongiorno (1990))
component of the physical system. It is simply the mechanism used in the block
diagram to define the system tracking error eðsÞ. Accordingly, the path in which it
lies is indicated with a dashed line. Obviously, yd ðsÞ represents the desired tracking
of the reference input r(s) that can be corrupted by noise n(s) when applied to the
controller.
There are three vector inputs to the controller and one can partition the controller
transfer matrix accordingly so that
and
The negative signs in (1.2) are introduced to reflect the fact that the feedforward
path and feedback path are intended to cancel the undesired components present in
the output y. Clearly, there are three submatrices associated with the controller that
the designer can choose. For this reason, the system shown in Fig. 1.1 is referred to
as a three-degree-of-freedom (3DOF) system. The second term on the right-hand
side of (1.1) and the second submatrix on the right-hand side of (1.2) disappear
when feedforward compensation is not included. The system is then called a two-
degree-of-freedom (2DOF) system. In some control applications, there are
4 1 Introduction
where I denotes the identity matrix. Now there is only one matrix to choose in the
design of the controller and the system is called a one-degree-of-freedom (1DOF)
system. The block diagram of a 3DOF digital control system is identical to the one
shown in Fig. 1.1 with the Laplace variable s replaced by the digital transform
variable k. Whenever it is not essential to distinguish between analog and digital
systems, the functional dependence on s or k is often omitted.
All systems of interest here can be divided into two parts: one part is just the
controller and the other part includes all the remaining subsystems. The block
diagram shown in Fig. 1.2 is a generic way to show all such systems and it is called
the standard configuration. The symbols G and C represent Laplace transform
transfer matrices when the system is analog and k-transform transfer matrices when
the system is digital. The block G is called the generalized plant and contains the
physical plant and all subsystems other than the controller. The column vector
e contains all the exogenous inputs to the system. Specifically, in the standard
configuration representation of the 3DOF system shown in Fig. 1.1 the column
vector e is an amalgamation of the column vectors r, n, nl , nm , and d. The mea-
surements column vector ym is an amalgamation of the column vectors u, w, and
z. For the column vector yc , one can choose an amalgamation of the controlled
variables e and v since one desires small tracking errors with minimal control effort.
One of the earliest papers in which the standard configuration is used to model
linear multivariable feedback control systems is Cheng and Pearson (1981), but a
similar configuration was also treated in Larin et al. (1971, 1972, 1973).
The standard configuration is a useful representation for establishing funda-
mental properties of feedback control systems irrespective of the subsystems and
their interconnection within the generalized plant. On the other hand, engineering
insights are generally obtained from the analysis of specific feedback system
implementations. A comprehensive approach covering both methodologies is
employed in the sequel.
Controller
1.2 System Models 5
v ¼ Cy y þ Cu u ð1:4Þ
and
C ¼ ½ Cy Cu : ð1:5Þ
A key element in the design of many feedback control systems is the require-
ment that the system tracks certain persistent reference input signals such as steps
and ramps. In these cases, the 2DOF standard configuration is a natural represen-
tation for the system since it clearly exposes the tracking requirement and is
especially useful when decoupling is needed. The system is decoupled when the
transfer matrix connecting r to yt is diagonal. Two of the earliest papers in which
the 2DOF standard configuration is used to model linear multivariable feedback
control systems are Pernebo (1981a, b).
Generalized
d ys
y ε
t −
G
v
+ yd
Plant Td
y
C
u +
r
Controller n +
X
1
1 ð 1 ð
J¼ ^y0c ðkÞ^yc ðkÞ ¼ yc ðkÞyc ðkÞdk=k ¼
H Tr½yc ðkÞyc ðkÞdk=k;
H
k¼0
2pj 2pj
jkj¼1 jkj¼1
ð1:6Þ
where ^y0c ðkÞ is the transpose of the inverse k-transform of yc ðkÞ and yc ðkÞ is the
transpose of yc ð1=kÞ. The notation Tr denotes taking the trace of a matrix which for
a square matrix is the sum of its diagonal elements. The quantity J can be used as a
measure of system performance. The validity of this lies in the fact that when any
element of ^yc ðkÞ is nonzero a positive contribution is made to J and the larger the
magnitude of this element, the larger the contribution. So the smaller J, the better
the performance one can expect since ideally ^yc ðkÞ ¼ 0 is desired. Contrary to
1.3 Stability, Signal Models, Performance Measures, and Saturation Constraints 7
Z1 Zj1 Zj1
1 1
J¼ ^y0c ðtÞ^yc ðtÞdt ¼ yc ðsÞyc ðsÞds ¼ Tr½yc ðsÞyc ðsÞds; ð1:7Þ
2pj 2pj
0 j1 j1
where ^y0c ðtÞ is the transpose of the inverse Laplace transform of yc ðsÞ and yc ðsÞ is
the transpose of yc ðsÞ. Again J can be used as a measure of system performance.
The usefulness of (1.7) as a performance measure lies in the fact that when any
element of ^yc ðtÞ is nonzero over any finite time interval a positive contribution is
made to J and the larger the magnitude of this element, the larger the contribution.
So the smaller J, the better the performance one can expect since ideally ^yc ðtÞ ¼ 0 is
desired. Quadratic performance measures in the form of the last integral in (1.7) are
the primary ones used here when designing the controller for analog systems to take
into account transient performance in response to deterministic inputs. Since this
integral can also be used as a norm on the Hardy subspace of all real rational stable
strictly proper yc ðsÞ, these quadratic performance measures are also often called H2
performance measures in the literature.
When the deterministic inputs are shape-deterministic instead, then \yc yc [ is
used instead of yc yc in the integrands where \ [ denotes the ensemble average.
When only stationary zero-mean stochastic signals are applied to a stable standard
configuration, then for digital systems (see Appendix A)
1 ð
\^y0c ðkÞ^yc ðkÞ [ ¼ Tr[Uyc yc ðkÞdk=k
H ð1:8Þ
2pj
jkj¼1
can be used as a measure of system performance since it represents the mean square
value of the norm of the vector ^yc . The matrix Uyc yc ðkÞ is the two-sided k-transform
of the covariance matrix \^yc ðk þ lÞy^0 c ðlÞ[ and must be free of poles on the unit
circle for the integral to exist. The relationship corresponding to (1.8) for analog
systems is (see Appendix A)
8 1 Introduction
Zj1
1
\^y0c ðtÞ^yc ðtÞ [ ¼ Tr½Uyc yc ðsÞds: ð1:9Þ
2pj
j1
The matrix Uyc yc ðsÞ is the two-sided Laplace transform of the covariance matrix
\^yc ðt þ sÞ^y0c ðsÞ[ . For the integral to exist, s Uyc yc ðsÞ must be strictly proper and
Uyc yc ðsÞ must be free of poles on the imaginary axis.
It should be emphasized that (1.6) and (1.7) and their counterparts for
shape-deterministic inputs are typically used to measure transient performance
while (1.8) and (1.9) measure steady-state performance after all transients have
disappeared. When a system is subjected to both shape-deterministic and stochastic
inputs both measures can be combined into a single performance functional
1 ð
J¼ Tr½QðkÞUyc ðkÞdk=k;
H ð1:10Þ
2pj
jkj¼1
where
Zj1
1
J¼ Tr½QðsÞUyc ðsÞds; ð1:12Þ
2pj
j1
where
in the case of analog systems. The matrix Q is included to allow for frequency
weighting and is chosen so that Q ¼ Q and so that it is nonnegative definite on the
unit circle in the case of digital systems and is nonnegative definite on the imag-
inary axis in the case of analog systems. The scalar constants a1 ; a2 are nonnegative
and allow for weighting the relative importance of the contribution from deter-
ministic and stochastic components to the controlled vector yc . Traditionally, Uyc yc
is called a spectral density matrix. In the sequel, matrices that are of the form of the
generalized spectral density matrix Uyc are also referred to as spectral density
matrices and only a single subscript is used for simplicity to distinguish them from
the traditional ones.
The subset of all stabilizing digital controllers that yield finite values for (1.10)
and the subset of all stabilizing analog controllers that yield finite values for (1.12)
are parameterized in this book and the controllers that minimize these performance
1.3 Stability, Signal Models, Performance Measures, and Saturation Constraints 9
where ytd1 and ysd1 represent the transforms, respectively, of the desired responses
of the deterministic tracking signals yt1 and the deterministic signals likely to cause
saturation ys1 . When C is a stabilizing controller and et1 is stable, asymptotic
tracking is assured. When v1 denotes the deterministic component of the output
from C; when as is typical ys1 ¼ Gac v1 ; where Gac accounts for the actuators and
any pre-compensation of the plant; and when Gac is stable except possibly for
simple poles on the stability boundary; then ys1 is the pseudo-stable transform of a
bounded temporal function for all stable v1 . Thus, attention is restricted to deter-
ministic exogenous inputs e1 for which a stabilizing controller C exists such that et1
and v1 are stable. When ys1 is stable and ysd1 ¼ 0 or when ys1 is pseudo-stable and
ysd1 is such that any poles of ys1 on the stability boundary are not poles of es1 , then
es1 is also stable. In this case, the performance functional (1.6) with yc ¼ yc1 can be
used for the digital case. The performance functional (1.7) can be used for the
analog case provided yc1 is also strictly proper.
An example in which ys1 is pseudo-stable and yt1 ¼ Gp ys1 ¼ Gt e1 is now con-
sidered. When ytd1 ¼ e1 ¼ md =s and md is a constant vector; when the output from
a pre-compensator and actuator is ys1 ¼ Gac v1 ¼ G ~ ac ðsÞv1 =s, where G
~ ac is stable
and G ~ ac ð0Þ 6¼ 0; when Gp ðsÞ is analytic at s ¼ 0 and Gpo ¼ Gp ð0Þ; and when a
stabilizing controller is used so that the closed-loop transfer matrix Gt is stable and
Gto ¼ Gt ð0Þ ¼ I; then the error
10 1 Introduction
is stable since
Moreover, since ys1 is pseudo-stable and on the s ¼ jx axis possesses only a simple
pole at s ¼ 0, one can find a constant vector m such that
lim ^yt1 ðtÞ ¼ lim s yt1 ðsÞ ¼ lim sGp ys1 ¼ Gpo m ¼ Gto md ¼ md ð1:18Þ
t!1 s!0 s!0
that the vector m must satisfy Gpo m ¼ md . In this case, ysd 1 ¼ m=s can be chosen
and
Thus,
2
m ¼ m0 m ¼ m0 þ m0 ðmo þ mh Þ ¼ kmo k2 þ kmh k2 kmo k2 : ð1:22Þ
o h
So the solution m ¼ mo is the minimum norm solution and is in keeping with the
desire to avoid saturation. Moreover, mo as given by (1.20) is in terms of given
data: Gpo and md .
1.4 Notation and Terminology 11
The notation and terminology used throughout is collected here for easy reference.
All results are obtained working exclusively with rational transforms. The Laplace
transform in the complex variable s is used for analog systems and the k-transform
is used for digital systems. The k-transform is related to the z-transform through the
relationship k = 1/z. In the few instances that it is necessary to do so, a circumflex is
used to distinguish an inverse transform from its transform as in (1.6) and (1.7) for
example. Often function arguments are omitted for brevity when no confusion is
possible or when the discussion is applicable to both analog and digital systems.
Laplace transforms and k-transforms are distinguished from one another when
necessary by showing explicitly the dependence on s or k. The complex variable
p is used generically to represent either s or k when useful to do so. The conjugate
of p is denoted by p. The real part of p is denoted by Re p and Re p 0 is used to
denote the finite part of the closed right half p-plane. The notation j pj is used for the
magnitude of p and j pj [ 1 represents the finite part of the open region of the
complex p-plane outside the unit circle. A transform is called good in the analog
case when it is free of poles on s = jx, x finite. It is called good in the digital case
when it is free of poles on jkj ¼ 1. It is called stable in the analog case when all its
finite poles lie in Re s\0, the open left-half s-plane. It is called stable in the digital
case when all its finite poles lie in jkj [ 1, the region outside the unit circle.
A transform that is not stable is called unstable. In general, the finite poles of a
transform can be divided into those that lie in the stable region (Re s\0 or jkj [ 1)
and those that do not. The former ones are called the stable poles and the latter ones
are called the unstable poles. Transforms that have no poles in Re s 0 for the
analog case and no poles in jkj 1 for the digital case, except possibly for simple
poles on the stability boundaries (the finite s ¼ jx axis in the analog case or on
jkj ¼ 1 in the digital case), are called pseudo-stable.
A polynomial in s is called Hurwitz when all its zeros lie in Re s\0 and a
polynomial in k is called Schur when all its zeros lie in jkj [ 1. A polynomial in
p is called stable when its reciprocal is stable. Clearly, a polynomial is Hurwitz or
Schur iff it is stable. A monic polynomial is one whose leading coefficient is unity.
For any matrix G0 , G , det G, and Tr G are used for the transpose, conjugate
transpose, determinant, and trace of G, respectively. A diagonal matrix G of order
h is denoted by G = diag{g1, g2,…, gh}. The identity matrix is the special case of a
diagonal matrix with ones on the diagonal and is denoted by I. A positive definite
(nonnegative definite) matrix G is indicated by G > 0 (G > 0). A matrix of trans-
forms is called good when all its elements are good. It is called stable when all its
elements are stable. The matrix G ðsÞ is the conjugate transpose of GðsÞ. That is,
G ðsÞ ¼ G ðsÞ which for a real rational matrix reduces to G ðsÞ ¼ G0 ðsÞ. The
matrix G ðkÞ is the conjugate transpose of Gð1= kÞ. That is, G ðkÞ ¼ G ð1= kÞ
which for a real rational matrix reduces to G ðkÞ ¼ G0 ð1=kÞ. It is important to note
that G has a different interpretation for analog and digital systems, but it is always
clear from the context which interpretation needs to be made. A real rational matrix
12 1 Introduction
circumvent this issue by only incorporating the H2 -norm of the transfer matrix
between the reference input and the plant input into the system performance mea-
sure. As a consequence of these approaches, however, plant input saturation can
become a problem. In Park et al. (2002) on the other hand, the only persistent
deterministic exogenous inputs considered are ones with poles on the imaginary axis
that the plant is capable of tracking with a stable plant input. In this case, the H2 -
norm of the plant input is finite and can be included in the system performance
measure to avoid plant input saturation and, moreover, it is possible to make a
connection between state-space and Wiener–Hopf design formulas.
The parameterization of the subset of all stabilizing controllers that yield finite
H2 performance in Park and Bongiorno (1990) led to the consideration of a
trade-off between stability margin and nominal performance for 2DOF and 3DOF
analog multivariable systems which is described in Bongiorno et al. (1997).
Basically, an approximate measure of stability margin is embodied in an H2 per-
formance measure which is a functional on the free stable matrix parameter. Then
this stability margin performance measure is minimized subject to a constraint on
the allowed increase in nominal performance that is given by the H2 -norm of the
free stable matrix parameter. The extension of this methodology to the standard
configuration of Fig. 1.2 is treated in Sect. 3.9. The trade-off problem turned out to
be a special case of a more general one treated in Khargonekar and Rotea (1991)
and involved obtaining a Wiener–Hopf solution in terms of Kronecker products.
In the papers by Doyle et al. (1989), Khargonekar and Rotea (1991), and Hunt
et al. (1994), the standard configuration exogenous input is assumed to be a white
noise or simply an indeterminate to indicate an input channel. In this setting, the
dynamics of the persistent inputs are absorbed into the generalized plant block.
However, in these papers certain stabilizability assumptions are made which lead to
a distinct disadvantage: the results are not applicable to 2DOF and 3DOF systems
with unstable reference inputs. In Park and Bongiorno (1989), this problem is
avoided by keeping the model of the persistent inputs separate from the generalized
plant; however, persistent inputs with unstable poles only on the stability boundary
can be accommodated. In Corrêa and Da Silveira (1995) and Xie et al. (2000), a
generalized 2DOF configuration is considered and unstable persistent inputs are
treated. The approach begins with a parameterization of the subset of stabilizing
controllers for which the error transform is stable. However, the parameterization
does not ensure that actual plant inputs remain bounded with time. That is, the real
issue of plant saturation is not specifically respected.1
A methodology for overcoming these limitations is presented in Park and
Bongiorno (2009), which begins with a parameterization of the subset of stabilizing
controllers for which the error transform is stable and the actual plant inputs remain
bounded with time. The optimal H2 solution is then obtained from within the subset
of free parameters for which the cost functional is finite. A key element in this
1
© 2009 Taylor and Francis Ltd. http://www.tandfonline.com. This paragraph is reprinted, with
permission, from Park and Bongiorno (2009).
1.5 Historical Perspective and Commentary 17
approach which distinguishes it from the earlier works cited above is the explicit
recognition and imposition of the requirement that the controller output should be
stable for the reasons described in the last few paragraphs of Sect. 1.3. This is the
approach followed in Chap. 3 of this book. Additional remarks on Park and
Bongiorno (2009) are contained in the reply by Bongiorno and Park (2010) to
comments by Aliev and Larin (2010).
The Wiener–Hopf design methodology has also been successful in connection
with the design of optimal decoupled linear multivariable systems. This method-
ology is used in Lee and Bongiorno (1993a, b) for analog 2DOF and 3DOF systems
and in both Youla and Bongiorno (2000) and Bongiorno and Youla (2001) for
analog 1DOF systems. Optimal tracking under decoupling constraints has also been
considered by Corrêa et al. (2001). In Park et al. (2002), the results contained in Lee
and Bongiorno (1993a) are extended to the 2DOF standard configuration and
computational algorithms based on state-variable representations are also given. An
important step in the Wiener–Hopf solution of the optimal decoupling problem is
the recognition in Lee and Bongiorno (1993a) that the Schur matrix product plays a
key role. Another essential step in the solution of the optimization problem is the
parameterization of all closed-loop transfer matrices that can be realized with sta-
bilizing decoupling controllers. Relevant results in this regard for 1DOF systems
are contained in Vardulakis (1987), Lin and Hsieh (1993), Linnemann and Wang
(1993), and Gómez and Goodwin (2000) for unity-feedback systems. The extension
to nonunity-feedback systems is given in Youla and Bongiorno (2000). Relevant
results with regard to stabilizing decoupling controllers for 2DOF and 3DOF sys-
tems containing plants with square transfer matrices are given in Desoer and
Gündes (1986) and the extension to plants with rectangular transfer matrices is
given in Lee and Bongiorno (1993b).
In this book, the comprehensive analog H2 Wiener–Hopf frequency-domain
analytical design methodology developed by Youla, Bongiorno, and coworkers for
linear time-invariant multivariable feedback control systems is presented in a uni-
fied way for analog and digital systems. Moreover, the tracking of persistent
exogenous deterministic inputs is explicitly accounted for. The focus is on theo-
retical insights regarding engineering issues that are not easily obtained from a
state-variable approach.
1.6 Examples
Example 1.1 The standard configuration in Fig. 1.2 represents the mathematical
relationship
yc G11 G12 e e
¼ ¼G ;
ym G21 G22 v v
18 1 Introduction
where the submatrices of G are compatibly sized with respect to the vectors yc , ym ,
e, and v. (a) Find Gyc e , the transfer matrix between e and yc for the standard
configuration. (b) Find the expressions for the Gij that correspond to the 3DOF
system shown in Fig. 1.1 when y0c ¼ ½ e0 v0 , y0m ¼ ½ w0 z0 u0 , and
e0 ¼ ½ d 0 r 0 n0 n0m n0l . (c) Evaluate Gyc e for the 3DOF system. (d) Find
explicit formulas which give e and v in response to d, r, n, nm , and nl for the 3DOF
system using the results obtained in part (c).
Solution (a) Clearly, v ¼ Cym and it follows that
Hence, from
yc ¼ G11 e þ G12 v ¼ G11 e þ G12 Cym ¼ ½G11 þ G12 CðI G22 CÞ1 G21 e
one gets
Td r Pd d Pv Pd Td 0 0 0 P
v v 0 0 0 0 0 I
yc e
w FPd d FPv nm FPd 0 0 I 0 FP
ym v
z Ld nl L 0 0 0 I 0
u rn 0 I I 0 0 0
(c) Using (1.2) and the Gij obtained in part (b) yields
Gyc e ¼ G11 þ G12 CðI G22 CÞ1 G21 ¼ G11 þ G12 ðI þ Cw FPÞ1 ½ Cw Cz Cu G21
P
¼ G11 þ ðI þ Cw FPÞ1 ½ ðCw FPd þ Cz LÞ Cu Cu Cw Cz :
I
Hence, with
Ged Ger Gen Gem Gel Gee
Gyc e ¼ ¼
Gvd Gvr Gvn Gvm Gvl Gve
one gets
and
where
and
Fig. 245
Now I will tell you what this difference is; and I want you to try and
understand it clearly, so that you will be able to explain it to others,
for I doubt if the grown-up people could give any better answers than
you. I think your fathers and mothers will be both surprised and
pleased when you show them some summer day how truly different
are these two berries.
You remember that in the strawberry we saw plainly that it was the
flat flower cushion which swelled into the ripe strawberry,—the
cushion which was quite hidden by the many pistils; and though
these pistils were scattered thickly all over the ripe, red fruit, these
little pistils with their seedboxes were too small and dry to add flavor
or richness to the berry.
But if we watch the growth of this blackberry, we see that things
are different.
Fig. 246
We see that the pistils of this fruit do not remain small and dry, as
with the strawberry. No, indeed! their little seedboxes grow bigger
and juicier every day, and they turn from green to red and from red to
black. They do not remain hard to the touch, but become so soft that
a slight pressure will bruise them and stain your fingers purple. And
we enjoy eating the full-grown blackberry (Fig. 249) because a
quantity of these juicy seedboxes are so packed upon the juicy
flower cushion that together they make a delicious mouthful (Figs.
247, 248).
The flower cushion of the blackberry is long and narrow, not broad
and flat like that of the strawberry.
So do not forget that in the strawberry we enjoy eating the ripened
flower cushion, while in the blackberry the juicy seedboxes give to
the fruit more of its size and flavor than does the flower cushion.
ANOTHER COUSIN
Fig. 250
Here we see a branch from the raspberry bush (Fig. 250). How is
the raspberry unlike both strawberry and blackberry? Let us place
side by side these three berries (Figs. 251, 252, 253).
Fig. 254
Fig. 255
I hope you will remember how these three berries differ one from
another.
Why the blossoms of these three plants grow into berries in three
different ways, we do not know; but our time has been well spent if
we remember that they do change in these three ways.
The more we see and question and learn, the more pleasure we
shall find in our own lives, and the better able we shall be to make
life pleasant for others.
PEA BLOSSOMS AND PEAS
T HE Pea family is a large one, and it is worth our while to find out
what plan it uses in flower building.
Let us look at a pea blossom and see of what parts it is made up.
“There is the green cup, or calyx,” you say.
Yes, that is plain enough. It is cut up into five little leaves.
“And there is a circle of flower leaves, which makes the corolla.”
Let us pull apart both calyx and corolla, and place the separate
leaves as in the picture (Fig. 256).
The five smaller leaves, the ones marked ca, are the green of the
calyx.
The five larger ones, marked co, belong to the corolla. These, you
notice, are not all alike. The upper one is much the largest.
Fig. 256
As I told you, the two lower leaves of the corolla are joined so as
to form a sort of pocket (Fig. 257). Now, surely, a pocket is meant to
hold something. So take a pin and slit open this pocket. As the two
sides spring apart, out flies some golden pollen, and we see that the
little pocket is far from empty. It holds ten stamens and one pistil.
If you look at these carefully (Fig. 256), you see that one stamen
stands alone, while the other nine have grown together, forming a
tube which is slit down one side. This tube clings to the lower part of
the pistil.
Now, if you pull this tube away, what do you see?
You see a little, green, oblong object, do you not (Fig. 258)?
And what is it? Do you not recognize it?
Fig. 258
Why, it is a baby pea pod. Within it lie the tiny green seeds (Fig.
259) which are only waiting for the fresh touch of life from a pollen
grain to grow bigger and bigger till they become the full-grown seeds
of the pea plant,—the peas that we find so good to eat when they
are cooked for dinner.
Fig. 259
So, after all, the building plan of the pea blossom is nothing but the
old-fashioned one which reads
1. Calyx.
2. Corolla.
3. Stamens.
4. Pistil.
Had I not told you to do so, I wonder if you would have been bright
enough to pull apart the little pocket and discover the stamens and
pistil.
What do you think about this?
THE CLOVER’S TRICK
H ERE you see the bees buzzing about the pretty pink clover
heads,—the sweet-smelling clover that grows so thickly in the
fields of early summer.
Can you tell me what plan the clover uses in flower building?
You will not find this easy to do. Indeed, it is hardly possible, for
the clover plays you a trick which you will not be able to discover
without help.
You believe, do you not, that you are looking at a single flower
when you look at a clover head?
Well, you are doing nothing of the sort. You are looking at a great
many little clover flowers which are so closely packed that they make
the pink, sweet-scented ball which we have been taught to call the
clover blossom.
It is incorrect to speak of so many flowers as one; and whenever
we say, “This is a clover blossom,” really we ought to say, “These are
clover blossoms.” We might just as well take a lock of hair—a lock
made up of ever so many hairs—and say, “This is a hair.” Now, you
all know it would not be correct to do this, and no more is it correct to
call a bunch of clover blossoms “a blossom.” But as most people do
not understand this, undoubtedly the mistake will continue to be
made.
Fig. 260 shows you one little flower taken out of the ball-like clover
head.
Can you think of any good reason why so many of these little
flowers should be crowded together in a head?
What would happen if each little blossom grew quite alone?
Fig. 260
Why, it would look so small that the bee could hardly see it. And
sweetly though the whole clover head smells, the fragrance of a
single flower would be so slight that it would hardly serve as an
invitation to step in for refreshments.
So it would seem that the clover plant does wisely in making one
good-sized bunch out of many tiny flowers, for in this way the bees
are persuaded to carry their pollen from one blossom to another.
The moral of the clover story is this: Be very careful before you
insist that you hold in your hand or see in the picture only one flower.
MORE TRICKS
Fig. 261
Here you would have been quite mistaken; for instead of one large
flower, the picture shows you a number of tiny blossoms, so closely
packed, and so surrounded by the four white leaves, that they look
like only one blossom.
Try to get a branch from the dogwood tree (only be sure to break it
off where it will not be missed), and pull apart what looks so much
like one large flower.
First pull off the four white leaves. Then you will have left a bunch
of tiny greenish blossoms. Look at one of these through a magnifying
glass. If eyes and glass are both good, you will see a very small
calyx, a corolla made up of four little flower leaves, four mites of
stamens, and a tiny pistil,—a perfect little flower where you never
would have guessed it.
But all by themselves they would never be noticed: so a number of
them club together, surrounding themselves with the showy leaves
which light up our spring woods.
In Fig. 262 you see the flower cluster of the hobblebush.
The hobblebush has still another way of attracting attention to its
blossoms. It surrounds a cluster of those flowers which have
stamens and pistils, and so are ready to do their proper work in the
world, with a few large blossoms which have neither stamens nor
pistils, but which are made up chiefly of a showy white corolla. These
striking blossoms serve to call attention to their smaller but more
useful sisters.
Fig. 262
T HERE is one plant (Fig. 264) which you city children ought to
know almost as well as the country children. In the back yards
and in the little squares of grass which front the street, it sends up its
shining stars; and as for the parks, they look as if some generous
fairy had scattered gold coins all over their green lawns.
Fig. 264
Now, what is this flower which is not too shy to bring its brightness
and beauty into the very heart of the crowded city?
It is the dandelion, of course. You all know, or ought to know, this
plucky little plant, which holds up its smiling face wherever it gets a
chance.
And now, I am sure, you will be surprised to learn that this
dandelion, which you have known and played with all your lives, is
among those mischievous flowers which are laughing at you in their
sleeves, and that regularly it has played you its “April fool;” for, like
the dogwood and the clover, this so-called dandelion is not a single
flower.
No, what you call a dandelion is a bunch made up of a great many
tiny blossoms.
If you pull to pieces a dandelion head, you will find a quantity of
little yellow straps. Each little strap is a perfect flower.
Now, if you had been asked for the building plan of the dandelion,
you would have looked for the calyx, and you would have thought
you had found it in the green cup which holds the yellow straps.
And when you were looking for the corolla, perhaps you would
have said, “Well, all these yellow things must be the flower leaves of
the corolla.”
But when you began your hunt for stamens and pistils, you would
have been badly puzzled; and no wonder, for these are hidden away
inside the yellow straps, the tiny flowers of the dandelion.
So remember that when you cannot find the stamens and pistils
within what you take to be the single flower, you will do well to stop
and ask yourself, “Can this be one of the plants which plays tricks,
and puts a lot of little flowers together in such a way as to make us
think that they are one big flower?”
THE LARGEST PLANT FAMILY IN THE WORLD
Fig. 265
Now, if I had asked you some time ago for the building plan of the
daisy, I think you would have told me that the arrangement of little
green leaves underneath the flower head made up the calyx, and
naturally you would have believed the white leaves above to have
formed the corolla; and the chances are that the yellow center would
have seemed to be a quantity of stamens. As for the seed holders,
you might have said, “Oh, well! I suppose they are hidden away
somewhere among all these stamens.”
It would not have been at all strange or stupid if you had answered
my question in this way.
I know of no plant which dresses up its flowers more cleverly, and
cheats the public more successfully, than this innocent-looking daisy;
for not only does it deceive boys and girls, but many of the grown-up
people who love flowers, and who think they know something about
them, never guess how they have been fooled the daisy. Indeed,
some of them will hardly believe you when you tell them that when
they pick what they call a daisy, they pick not one, but a great many
flowers; and they are still more surprised when they learn that not
only the yellow center of the daisy is composed of a quantity of little
tube-shaped blossoms, but that what they take to be a circle of
narrow white flower leaves is really a circle of flowers, each white
strap being a separate blossom.
Fig. 266
I dare not try to tell you how many separate blossoms you would
find if you picked to pieces a daisy and counted all its flowers,—all
the yellow ones in the center, and all the white outside ones,—but
you would find a surprisingly large number.