Design of Linear Multivariable

Feedback Control Systems The Wiener

Hopf Approach using Transforms and
Spectral Factorization Joseph J.
Bongiorno Jr.
Visit to download the full and correct content document:
https://textbookfull.com/product/design-of-linear-multivariable-feedback-control-syste
ms-the-wiener-hopf-approach-using-transforms-and-spectral-factorization-joseph-j-bo
ngiorno-jr/
More products digital (pdf, epub, mobi) instant
download maybe you interests ...

Multi Stage and Multi Time Scale Feedback Control of

Linear Systems with Applications to Fuel Cells Verica
Radisavljevi■-Gaji■

https://textbookfull.com/product/multi-stage-and-multi-time-
scale-feedback-control-of-linear-systems-with-applications-to-
fuel-cells-verica-radisavljevic-gajic/

Linear Canonical Transforms: Theory and Applications

1st Edition John J. Healy

https://textbookfull.com/product/linear-canonical-transforms-
theory-and-applications-1st-edition-john-j-healy/

Networked control systems with intermittent feedback

1st Edition Hirche

https://textbookfull.com/product/networked-control-systems-with-
intermittent-feedback-1st-edition-hirche/

Classical Feedback Control with Nonlinear Multi-Loop

Systems: With MATLAB® and Simulink® 3rd Edition Boris
J. Lurie

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/

Theory of stabilization for linear boundary control

systems 1st Edition Nambu

https://textbookfull.com/product/theory-of-stabilization-for-
linear-boundary-control-systems-1st-edition-nambu/

Math 302 Lecture Notes linear algebra and multivariable

calculus Kenneth L. Kuttler

https://textbookfull.com/product/math-302-lecture-notes-linear-
algebra-and-multivariable-calculus-kenneth-l-kuttler/

The Great Facilitator: Reflections on the Contributions

of Joseph F. Hair, Jr. to Marketing and Business
Research Barry J. Babin

https://textbookfull.com/product/the-great-facilitator-
reflections-on-the-contributions-of-joseph-f-hair-jr-to-
marketing-and-business-research-barry-j-babin/

Factorization of Boundary Value Problems Using the

Invariant Embedding Method Jacques Henry And Angel M.
Ramos (Auth.)

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

ISBN 978-3-030-44355-9 ISBN 978-3-030-44356-6 (eBook)

https://doi.org/10.1007/978-3-030-44356-6
MATLAB is a registered trademark of The MathWorks, Inc. See https://www.mathworks.com/
trademarks for a list of additional trademarks.

Mathematics Subject Classification (2010): 93-01, 93C05, 93C15, 93C35, 93C55, 93C57, 93C62,
93D09, 93D15, 93E20, 47A68

© Springer Nature Switzerland AG 2020

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part
of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission
or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar
methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt from
the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authors or the editors give a warranty, express or implied, with respect to the material contained herein or
for any errors or omissions that may have been made. The publisher remains neutral with regard to
jurisdictional claims in published maps and institutional affiliations.

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.

Brooklyn, USA Joseph J. Bongiorno, Jr.

Suwon-si, Korea (Republic of) Kiheon Park
Contents

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

3.7 H2 Design of 1DOF Systems . . . . . . . . . . . . . . . . . . . . . . . . . . 151

3.8 H2 Design of 3DOF Systems . . . . . . . . . . . . . . . . . . . . . . . . . . 166
3.9 Trade-off of Optimal Performance for Reduced Sensitivity
to Plant Uncertainty . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
3.9.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
3.9.2 Sensitivity to Small Plant Parameter Changes . . . . . . . . . 186
3.9.3 Stability Margin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
3.9.4 The Analytical Solution for Sensitvity and Stability
Margin Combined . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
3.10 Historical Perspective and Commentary . . . . . . . . . . . . . . . . . . . 197
3.11 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
4 H2 Design of Multivariable Control Systems with Decoupling . . . . . 241
4.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241
4.2 Characterization of Realizable and Acceptable T . . . . . . . . . . . . . 242
4.3 Characterization of Realizable Diagonal T and Acceptable
Diagonal T . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 256
4.4 H2 Design of Decoupled Systems for Acceptable Inputs . . . . . . . 261
4.4.1 H2 Design for the Digital Case . . . . . . . . . . . . . . . . . . . 268
4.4.2 H2 Design for the Analog Case . . . . . . . . . . . . . . . . . . . 269
4.4.3 H2 Design of Decoupled Systems for Strictly
Acceptable Inputs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 272
4.5 H2 Design of 1DOF Decoupled Systems . . . . . . . . . . . . . . . . . . 276
4.6 H2 Design of 3DOF Decoupled Systems . . . . . . . . . . . . . . . . . . 286
4.7 Historical Perspective and Commentary . . . . . . . . . . . . . . . . . . . 302
4.8 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 304
5 Numerical Calculation of Wiener–Hopf Controllers . . . . . . . . . . . . . 319
5.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 319
5.2 Basic Machinery for Analog Systems . . . . . . . . . . . . . . . . . . . . 319
5.3 State-Space Representation of Wiener–Hopf Controllers:
The Analog Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 324
5.4 Basic Machinery for Digital Systems . . . . . . . . . . . . . . . . . . . . . 337
5.5 State-Space Representation of the Wiener–Hopf Controllers:
The Digital Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 341
5.6 Historical Perspective and Commentary . . . . . . . . . . . . . . . . . . . 353
5.7 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 354

Appendix A: Complex Functions, Transforms, Parseval’s Formula,

and Stochastic Processes. . . . . . . . . . . . . . . . . . . . . . . . . . . . 383
Appendix B: Coprime Polynomial Matrix Fraction Descriptions
for Rational Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 409
Contents xi

Appendix C: Spectral Factorization of Rational

Parahermitian-Positive Matrices . . . . . . . . . . . . . . . . . . . . . 435
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 443
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 449
Chapter 1
Introduction

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

© Springer Nature Switzerland AG 2020 1

J. J. Bongiorno Jr. and K. Park, Design of Linear Multivariable Feedback
Control Systems, https://doi.org/10.1007/978-3-030-44356-6_1
2 1 Introduction

coefficients. As a consequence, all digital subsystem transfer matrices contain

elements that are ratios of polynomials in the variable k with real coefficients or,
equivalently, real rational functions of k. So it should not be surprising that the
methodologies used for analog and digital systems parallel one another.
Sampled-data systems contain both analog and digital elements and represent a
greater modeling challenge. Typically in such systems, all subsystems except for
the controller are analog. The controller consists of a digital processor operating at a
fixed rate followed by a hold circuit. The input to the digital processor is a sequence
of sampled measurements and the digital processor output samples are the inputs to
the hold circuit. Usually, the output of the hold circuit is held constant during the
inter-sample interval at the value of the sample at the beginning of the interval.
When attention is restricted to the behavior of the system at the sampling instants,
the sampled-data system can be easily modeled as a digital system. When the
system behavior in the interval between the samples needs to be taken into account,
it is still possible with a more complex digital system model to account for the
performance of the sampled-data system (Rosenwasser and Lampe 2006). Hence,
all the results for digital systems presented in this book can be applied to
sampled-data systems.

1.2 System Models

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

vðsÞ ¼ Cw ðsÞwðsÞ  Cz ðsÞzðsÞ þ Cu ðsÞuðsÞ ð1:1Þ

and

CðsÞ ¼ ½Cw ðsÞ  Cz ðsÞ Cu ðsÞ: ð1:2Þ

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

restrictions on the implementation of the controller that correspond to Cw ¼ Cu ,

Cz ¼ 0. The controller is then of the form

C ðsÞ ¼ Cu ðsÞ½ I I ; ð1:3Þ

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.

Fig. 1.2 Standard Generalized

configuration (© 1989 IEEE. e yc
Reprinted with permission,
from Park and Bongiorno G
(1989))
v ym
Plant

Controller
1.2 System Models 5

An insightful alternative to the standard configuration can be generated once the

following observations are made. First, one can always view the exogenous input
vector e as an amalgamation of a reference input vector r, the noise vector n cor-
rupting r, and a disturbance vector d that accounts for all other exogenous system
inputs. Second, one can view yc as an amalgamation of an error vector e ¼ yd  yt
and a vector ys that contains all the system responses that need to be kept small. In
particular, ys should include the signals which need to be kept small so that the
linear system model remains valid and minimal control effort is required. The vector
yt represents the system outputs, which need to track yd ¼ Td r. Third, r and its
noise n involve only low power level signals that can only enter the system through
the controller. An alternative to the standard configuration in this case is the one
shown in Fig. 1.3. It is called the 2DOF standard configuration because the
controller transfer matrix can be partitioned into two blocks so that

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 +

Fig. 1.3 Two-degree-of-freedom standard configuration

6 1 Introduction

1.3 Stability, Signal Models, Performance Measures,

and Saturation Constraints

As a consequence of superposition in linear systems, once the controller is selected,

each of the output vectors yc ; ym ; and v in the standard configuration is given by the
sum of two vectors: one completely determined by the exogenous input e and one
completely determined by initial conditions in the system. So it is necessary for
satisfactory performance that any contributions from the initial conditions or
transients in response to exogenous inputs disappear in time. Then, the output
vectors are determined essentially by e. The system is called stable in this case and
system design can focus on the behavior of the system in response to e; this is the
approach adopted here. First, the class of all stabilizing controllers is identified.
Then the one within this class is sought for which performance requirements are
met. Attention is restricted to exogenous input signals that are deterministic with
real rational transforms, or are stochastic with real rational power spectral densities,
or are combinations of these. Often an input signal is viewed as the product of a real
random variable with a deterministic function. Such signals are called shape-de-
terministic signals and attention is restricted to those for which the deterministic
part has a real rational transform. Since only physical inputs which have real
rational transforms and since only physical subsystems with real rational transfer
matrices are considered, it is assumed throughout whenever not explicitly stated
otherwise that all polynomial matrices and all transfer matrices are real (i.e., all
coefficients are real).
It is useful in addition to the notion of system stability to define a stable
transform. A Laplace transform whose finite poles all have negative real parts or a
k-transform whose finite poles all have magnitudes greater than unity are called
stable transforms. When the standard configuration is stable, when there are no
stochastic inputs, when any persistent deterministic reference inputs are ones for
which the controlled variables have zero steady-state value, and when all other
deterministic inputs have stable transforms, then the output vector transform yc is
stable. It follows under these conditions that in the digital case (see Appendix A)

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

standard practice a circumflex over the inverse transform is used to distinguish it

from its transform rather than the other way around. This is done to keep the
notation simpler because transform quantities are primarily used here. Quadratic
performance measures in the form of the last integral in (1.6) are the ones used in
this work when designing a controller for digital systems that takes 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 yc ðkÞ, these
quadratic performance measures are also often called H2 performance measures in
the literature.
It is also typical in the analog case for yc to be strictly proper (i.e., yc is zero for
s ¼ 1). It then follows from Parseval’s formula under the same conditions as cited
above for the digital case that (see Appendix A)

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

Uyc ðkÞ ¼ a1 \yc ðkÞyc ðkÞ [ þ a2 Uyc yc ðkÞ ð1:11Þ

in the case of digital systems and

Zj1
1
J¼ Tr½QðsÞUyc ðsÞds; ð1:12Þ
2pj
j1

where

Uyc ðsÞ ¼ a1 \yc ðsÞyc ðsÞ [ þ a2 Uyc yc ðsÞ ð1:13Þ

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

measures are identified. In addition, a methodology is presented that enables one to

trade off optimality for improved stability margin and reduced sensitivity of the
performance functional to plant model uncertainty. Moreover, specialized results
for the systems shown in Figs. 1.1 and 1.3 are presented. The design of optimal
1DOF and 2DOF decoupled systems is also treated.
A key element in the development of these results is the recognition that the
deterministic components of the inputs to the physical process under control must
be bounded. For example, the rudder angle of an aircraft or boat is limited. When
such a limit is reached, saturation is said to have occurred. Clearly, a necessary
condition for avoiding saturation is that the deterministic parts of the inputs to the
physical plant have pseudo-stable transforms. That is, transforms which 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). Otherwise, the physical plant
inputs would continue to grow until the saturation level is reached and the
assumption of linearity would no longer hold.
For the standard configuration shown in Fig. 1.2, one can write yc ¼ yc1 þ yc2
for the controlled variables, where yc1 represents the deterministic components and
yc2 represents the stochastic ones. It is particularly convenient to set
   
yt1  ytd1 et1
yc1 ¼ ¼ ; ð1:14Þ
ys1  ysd1 es1

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

et1 ¼ yt1  ytd1 ¼ ðGt  I Þmd =s ð1:15Þ

is stable since

lim ^et1 ðtÞ ¼ lim s et1 ðsÞ ¼ limðGt  I Þmd ¼ 0: ð1:16Þ

t!1 s!0 s!0

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

ys1 ¼ m=s þ ~ys1 ; ð1:17Þ

where ~ys1 is stable. In addition, it follows from

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

es1 ¼ ys1  ysd1 ¼ ~ys1 ð1:19Þ

is stable. A similar approach is used in Polyakov (2001) for single-variable

sampled-data control systems.
Typically, Gpo has row rank and there is at least one solution for m given by
 1
m ¼ mo ¼ G0po Gpo G0po md : ð1:20Þ

Clearly, all solutions are given by m ¼ mo þ mh ; where Gpo mh ¼ 0. Hence,

m0h G0po ¼ 0 and one gets
 1
m0h mo ¼ m0h G0po Gpo G0po md ¼ 0: ð1:21Þ

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

1.4 Notation and Terminology

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

G in s or k is called parahermitian when G ¼ G . Any parahermitian matrix GðsÞ

or GðkÞ which is nonnegative definite on s ¼ jx or jkj ¼ 1, respectively, is called
parahermitian-positive.
A matrix GðpÞ is called proper (strictly proper)) when it is finite (zero) at
p ¼ 1. Otherwise, it is called improper. A square proper matrix whose inverse is
also proper is called biproper. The notation GðpÞ  0ðpm Þ means that no entry in
GðpÞ grows faster than pm as p ! ∞. When GðpÞ is square and polynomial and the
coefficient matrix of the highest power of p is nonsingular, it is called regular.. The
rank of a matrix GðpÞ is its normal rank: i.e., the highest order of all non-identically
zero minors of GðpÞ. The McMillan degree dðG ; po Þ of p ¼ po (finite or infinite) as
a pole of the rational matrix GðpÞ is the largest multiplicity it possesses as a pole of
any minor of GðpÞ. The McMillan degree d (G) of GðpÞ is the sum of the McMillan
degrees of its distinct poles.
In the partial fraction expansion of G(s), the contributions made by all its finite
poles in Re s < 0, Re s > 0, and by its poles at s = ∞ are denoted by
fGðsÞg þ ; fGðsÞg ; and fGðsÞg1 , respectively. Clearly, {G(s)}+ is analytic in Re
s > 0, {G(s)}_ is analytic in Re s < 0, and both are strictly proper. The contribution
{G(s)}∞ is polynomial. Similarly, in the partial fraction expansion of G(k), the
contributions made by all its finite poles in jkj  1; jkj \1, and by its poles at
k ¼ 1 are denoted by {G(k)}+, {G(k)}–, and {G(k)}∞, respectively. Clearly, {G
(k)}+ is analytic in jkj\1; fGðkÞg is analytic in jkj  1, and both are strictly
proper. The contribution {G(k)}∞ is polynomial.
For any parahermitian-positive matrix GðsÞ, it is established in Youla (1961,
2015) that there are factors Xl ðs) and Xr ðsÞ analytic together with their respective
left and right inverses in Re s [ 0 such that GðsÞ ¼ Xl ðsÞXl ðsÞ ¼ Xr ðsÞXr ðsÞ.
Similarly, for any parahermitian matrix GðkÞ which is nonnegative definite on
jkj ¼ 1, one can show as in Appendix C that there are factors Xl ðkÞ and Xr ðkÞ
analytic together with their respective left and right inverses in jkj \1 such that
GðkÞ ¼ Xl ðkÞXl ðkÞ ¼ Xr ðkÞXr ðkÞ. The matrix Xl ðXr Þ is called a left (right)
Wiener–Hopf factor for G in both cases. Examples of parahermitian-positive
matrices are power spectral density matrices. Such matrices are the transforms of
ensemble averages and such averages are denoted by \  [ as in (1.8) and (1.9)
for example.
The Schur product G R of two same-size matrices G and R is the matrix whose
i-row, j-column entry is gijrij. The Kronecker product G ⊗ R is the matrix whose
ij-block is given by gijR. The vector vec G ¼ ½ g01 g02 . . . g0n 0 is formed by
stacking all the columns of the matrix G. Useful references for the properties of
Kronecker matrix products are Brewer (1978) and Horn and Johnson (1991).
A useful reference for the properties of Schur matrix products is Horn and Johnson
(1985)
Additional notation and terminology needed for Appendices B and C is sum-
marized in Sects. B.2 and C.2.
1.5 Historical Perspective and Commentary 13

1.5 Historical Perspective and Commentary

The methodologies employed to design linear time-invariant feedback control

systems are generally either time-domain methodologies or frequency-domain
methodologies. Time-domain methodologies are based on a state-variable repre-
sentation for each given subsystem. Frequency-domain methodologies are based on
a transfer matrix representation for each subsystem. The design methodologies are
also distinguished by the performance measure employed. When tracking accuracy
and control effort with the nominal plant model is key, then a quadratic or H2
performance measure is often used. When robustness or insensitivity to uncertainty
or change in the plant model or exogenous inputs is key, then an H1 performance
measure is often used. The H1 -norm of a real rational stable transfer matrix is the
least upper bound for all the singular values of the matrix on the imaginary axis for
analog systems or on the unit circle for digital systems.
Examples of books focusing on the time-domain approach with an H2 perfor-
mance measure are Kwakernaak and Sivan (1972), Anderson and Moore (1989),
and Saberi et al. (1995). In Chen (2000), the time-domain approach with an H1
performance measure is considered and the results are used in practical applica-
tions: specifically, the control of the voice-coil-motor actuator of a hard disk drive,
the control of a piezoelectric actuator, and the control of a gyro-stabilized mirror
targeting system. Examples of books focusing on the time-domain approach and
covering designs for both H2 and H1 performance measures are Chen and Francis
(1995) and Zhou et al. (1996). Books focusing on the frequency-domain approach
are Kučera (1979), Vidyasagar (1985), Grimble (2006), and Rosenwasser and
Lampe (2006). In Kučera (1979), H2 performance measures are used exclusively;
in Vidyasagar (1985), the emphasis is on H1 performance measures; and in
Grimble (2006), both H2 and H1 performance measures are used. Grimble (2006)
covers robustness in depth and contains applications of the theory that include
designs of aircraft, ship, submarine, and industrial control systems. The H2 per-
formance measure is used in Rosenwasser and Lampe (2006) and as in Chen and
Francis (1995) the focus is on sampled-data systems. In Aliev and Larin (1998), H2
designs for both time-domain and frequency-domain methodologies are covered
with an emphasis on computational algorithms. H2 designs for both time-domain
and frequency-domain methodologies are also covered in Kučera (1991), with an
emphasis on the relationships between the two methodologies.
In addition to these analytically oriented design methodologies, a number of
graphically oriented frequency-domain design methodologies for multivariable
systems along classical lines have been proposed. The inverse Nyquist array of
Rosenbrock, the characteristic locus method of MacFarlane, and the quantitative
feedback theory of Horowitz are discussed in Maciejowski (1989) along with
designs based on H2 and H1 performance measures. Another alternative to ana-
lytically oriented design methodologies is the use of convex programming along the
lines taken in Boyd and Barratt (1991) or indirectly through linear matrix
inequalities as described in Boyd et al. (1994) and in Skelton and Iwasaki (1995).
14 1 Introduction

In this book attention is focused on an analytical frequency-domain design

methodology using an H2 performance measure.
The earliest efforts to design controllers for feedback systems by minimizing a
quadratic or H2 performance measure are described in Newton et al. (1957). These
efforts employ the calculus of variations to generate the integral equation, known as
the Wiener–Hopf equation, which the impulse response function of the optimal
controller satisfies. Equations of this kind had been treated in Wiener (1949).
Transforms, partial fraction expansions, and spectral factorizations are then used to
solve the Wiener–Hopf equation and an explicit formula for the controller transfer
function is obtained for analog 1DOF single-input-output systems containing stable
plants. Since the design methodology presented in this book also makes use of
transforms, partial fraction expansions, and spectral factorizations, it is referred to
as the Wiener–Hopf design methodology in recognition of the insightful work of
these early researchers. In Weston and Bongiorno (1972), the Wiener–Hopf design
methodology is extended to analog 1DOF multivariable systems with stable plants
by utilizing the variational technique in Bongiorno (1969) and the matrix spectral
factorization technique in Youla (1961).
The Wiener–Hopf methodology is applied in Chang (1961) to analog 2DOF
single-input-output systems containing stable plants. The restriction to stable plants
is needed in the cases just cited because the methodologies involve first finding an
optimal stable open-loop controller and then realizing it within a single-loop
feedback structure. This methodology leads to pole-zero cancelations in Re s  0
when the plant is unstable; hence, the system is unstable. In Newton et al.
(1957) and Chang (1961), it is suggested that unstable plants can be handled by first
stabilizing the plant with minor-loop feedback. Although this approach allows one
to make use of the methodology available for stable plants, it leaves open the
question of whether or not the performance one can realize is different among the
many possible choices one can make for the minor-loop transfer function. This is
especially a problem in the multivariable case and/or when the plant cannot be
stabilized with stable minor-loop feedback. That there exist unstable plants that
cannot be stabilized with a proper stable minor-loop feedback transfer matrix is
made clear in Youla et al. (1974). Also left open is whether the open-loop design
can actually be realized with a stable closed-loop feedback system. Indeed, it is
pointed out in Aliev and Larin (1998, pp. 66–67) that all stabilization by minor-loop
feedback is not equivalent to all stabilizing controllers for the original plant.
Obviously, an unequivocal approach is needed for unstable plants.
The first Wiener–Hopf solution for unstable multivariable analog plants is
contained in Larin et al. (1971, 1972, 1973). However, this notable achievement for
its time went unrecognized by other researchers and similar results were published
in Youla et al. (1976b) for analog systems and in Kučera (1979) for digital systems.
In these works, all stabilizing controllers are effectively parameterized in terms of a
free stable matrix and the optimal H2 choice for the free matrix parameter is
determined. In Kučera (1979), it was shown that all stabilizing controllers can be
generated from the parameterized solution set of a Bezout equation. (There is a
difference of opinion on whether the Bezout equation should actually be called a
1.5 Historical Perspective and Commentary 15

Diophantine equation or an Aryabhatta equation as noted in the preface to

Vidyasagar (1985).) Additional commentary on the history and attributes of sta-
bilizing controller parameterizations is provided in Sect. 2.8. Also relevant are the
comments in Aliev and Larin (2007) and the reply by Kučera (2007).
The frequency-domain methodology in Kučera (1979) is applicable to 1DOF
multivariable digital plants and is called the polynomial equation approach. In
contrast with the Wiener–Hopf methodology in which a partial fraction expansion
is needed, the polynomial equation approach involves finding solutions to a pair of
linear matrix polynomial equations. Only open-loop optimal designs with zero
initial conditions are considered in connection with the tracking of deterministic
inputs. The stable closed-loop realization of these designs is not addressed for
arbitrary initial conditions and with output feedback instead of state-variable
feedback. Hence, the results for open-loop designs are of limited value for unstable
plants. Optimal H2 closed-loop designs are given for the regulator problem with
stochastic inputs; however, some important restrictions regarding signal and/or
system poles on the unit circle are not adequately generalized although specific
examples in which such poles can be handled successfully are provided.
The next major step forward in connection with the Wiener–Hopf design method
occurs in Youla and Bongiorno (1985) where analog 2DOF multivariable systems
are treated. Not only is the optimal H2 design determined in this paper, but all
designs which yield a finite value for the H2 performance measure are parame-
terized in terms of two free stable matrices. This important parameterization is
extended to analog 3DOF systems in Park and Bongiorno (1990) and to the analog
standard configuration in Park and Bongiorno (1989). Digital 2DOF and 3DOF
systems are treated in Grimble (1988) using the polynomial equation approach.
However, in contrast with the treatments cited above for the analog case,
shape-deterministic inputs and plants with poles on the stability boundary are not
treated directly. Instead, a limiting procedure is suggested. Moreover, only the
optimal solution is given and not all solutions for which the H2 performance
measure is finite. The standard configuration is also treated using a state-variable
approach in Doyle et al. (1989) for analog systems and using the polynomial
equation approach in Hunt et al. (1994) for digital systems. However, the first of
these two papers does not deal directly with optimal designs for persistent inputs
that have poles on the imaginary axis and the second paper does not deal fully with
the issue of stability. Also, the second paper does not parameterize all designs for
which the H2 performance measure is finite. It does provide a comprehensive
chronological account of the publications related to the polynomial equation
approach.
The analog 2DOF standard configuration is treated in Corrêa and Da Silveira
(1995), Xie et al. (2000), and Park et al. (2002). The first two papers allow persistent
deterministic exogenous inputs with poles in Re s  0 where the plant has none and
insists on asymptotic tracking with zero steady-state error. In this case, the plant
input must have poles in Re s  0 and the H2 -norm of the plant input is not finite.
Corrêa and Da Silveira (1995) circumvent this issue by only incorporating the stable
part of the plant input into the system performance measure. Xie et al. (2000)
16 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

ym ¼ G21 e þ G22 v ¼ G21 e þ G22 Cym , ym ¼ ðI  G22 CÞ1 G21 e:

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

Gyc e ¼ G11 þ G12 CðI  G22 CÞ1 G21 :

(b) It is evident from Fig. 1.1 that

 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 rn 0 I I 0 0 0

Thus, for the 3DOF system

 Pd Td 0 0 0 P
0 0 0 0 0 I
G11 G12
 FPd 0 0 I 0 FP
G21 G22
L 0 0 0 I 0
0 I I 0 0 0
1.6 Examples 19

(c) Using (1.2) and the Gij obtained in part (b) yields

CðI  G22 CÞ1 ¼ ðI  CG22 Þ1 C ¼ ðI þ Cw FPÞ1 ½ Cw Cz Cu :

It then follows from part (a) that

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

Ged ¼ Pd þ PðI þ Cw FPÞ1 ðCw FPd þ Cz LÞ;

Ger ¼ Td  PðI þ Cw FPÞ1 Cu ;
½ Gen Gem Gel  ¼ PðI þ Cw FPÞ1 ½ Cu Cw Cz ;

and

Gve ¼ ðI þ Cw FPÞ1 ½ ðCw FPd þ Cz LÞ Cu Cu Cw Cz :

(d) From the fact that yc ¼ ½ e0 v0 0 ¼ Gyc e e, it follows that

e ¼ ½I 0 yc ¼ ½ I 0 Gyc e e ¼ Gee e ¼ ed þ er þ en þ enm þ el ;

where

ed ¼ Ged d ¼ ½PðI þ Cw FPÞ1 ðCw FPd þ Cz LÞ  Pd d;

er ¼ Ger r ¼ ½Td  PðI þ Cw FPÞ1 Cu r;

and

en þ em þ el ¼ ½ Gen Gem Gel ½ n0 n0m n0l 0

¼ PðI þ Cw FPÞ1 ðCu n  Cw nm  Cz nl Þ:
Another random document with
no related content on Scribd:
 A COUSIN OF THE STRAWBERRY

T HE strawberry is a member of the great Rose family. Among its

many cousins we find the blackberry and the raspberry.
The blackberry blossom (Fig. 245) also has five white leaves, and
a center made up of pistils and stamens.
When its white flower leaves fall, and its empty dust boxes wither,
we see the blackberry begin to take the place of the blossom, just as
we saw the strawberry take the place of the strawberry blossom (Fig.
246).
But now we are about to discover the way in which the blackberry
differs from the strawberry.
 Cannot some boy or girl tell me in what way they are different?
“One is black, and the other is red.”
But that is not the answer I want. Perhaps it is hardly likely that
any child could guess what I have in mind. Still a little exercise in
guessing is as good for your brains as gymnastics are good for your
bodies.

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).

Fig. 247 Fig. 248 Fig. 249

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. 251 Fig. 252 Fig. 253

Once more we observe that the strawberry is the flat flower

cushion grown big and juicy.
Again we see that the seedboxes of the blackberry packed upon
the swollen flower cushion make up much of the fruit.
But in the raspberry we find that the red, ripe seedboxes alone
make the berry which is so good to eat.
 When we pick this raspberry, we find that the flower cushion
remains upon the plant, instead of coming off in our fingers and
helping to make a luscious morsel, as with the other two fruits (Figs.
254, 255).

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

The two side ones are alike.

In the real flower the two lower ones are joined so as to form a
little pocket.
And what else do you find?
Now, if you do not pull apart the pea blossom, you find nothing
else. But you know that the seed-holding fruit is the object of the
flower’s life, and that so this flower is pretty sure to have somewhere
either a pistil with its seedbox, or stamens with their dust boxes, or
both; for without the seeds of the seedbox, and the pollen of the dust
boxes, no fruit can result.
So, knowing that the pea blossom cannot give birth to a pea pod
without stamens and pistil, let us have a search for these.
 Fig. 257

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

C AN you think of any other flowers that deceive us as the clover

does?
Early in May we see in the woods a tree that is very beautiful. It is
covered with what seem to be white blossoms. This tree is the
flowering dogwood, and it tricks us somewhat in the same way as
does the clover; for in this picture (Fig. 261) you see what nearly
every one believes to be a single flower of the dogwood. And if some
time ago you had been asked to give the building plan of the
dogwood flower, you would have been pretty sure to say that the four
large white leaves formed its corolla.

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

Sometimes a whole plant family will play this trick of putting a

quantity of flowers in one bunch or cluster.
 Fig. 263

The wild carrot (Fig. 263) is one of our commonest wayside

weeds, a torment to the farmer, but a beautiful plant nevertheless.
Each one of its lace-like flower clusters is made up of many flowers,
—flowers which are too small to live alone, and so have decided to
keep house together.
You will notice that here, as with the hobblebush, the outer flowers
are large and more showy than the inner ones. They seem to feel
that with them rests the reputation of the family; that they must make
the most of themselves, and do all in their power to attract the bees
and butterflies.
The wild carrot belongs to the Parsley family. All the members of
this family collect a great many little flowers into one fairly large
cluster.
 AN OLD FRIEND

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

T HE dandelion belongs to the largest plant family in the world. All

the members of this family have the dandelion trick of bunching
together a quantity of little flowers. From this habit the family takes
its name. It is called the “Composite” family, because with it, that
which looks like one flower is composed of many flowers.
To this great family belong some of the flowers which you know
best; and if you are not to be fooled again and again, you must learn
to tell by its blossoms whether a plant is a member of the Composite
family. This will not be difficult if you will be patient, and pull to pieces
a few of the flower heads which I am going to describe, and examine
carefully the building plan used by the separate flowers.
Fig. 265 shows you the field daisy. This pretty flower is an old
friend; and many of you know that its beauty is no comfort to the
farmer, who finds it a sign of poor soil and a nuisance, and does his
best to get rid of it.
As you know, the central part of the daisy is bright yellow, and the
narrow leaves which stand out in a circle around its yellow center are
pure white.

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.