r

i'

Linear Contro l Theory

Linear Control Theory

The State Space Approach

Frederick Walker Fairman

Queen's University,
Kingston, Ontario, Canada

John Wiley & Sons

Chichester New York Weinheim Brisbane Singapore Toronto

Copyright

1998 John Wiley & Sons Ltd,

Baffins Lane, Chichester.
West Sussex P019 IUD, England
01243 779777
National
International ( -r 44) 1243 779777

e-mail (for orders and customer service enquiries): cs-books(c!;wiley.co.uk

Visit our Home Page on http:/ jwww.wiley.co.uk
or
http://www. wiley .com
All rights reserved. No part of this publication may be reproduced, stored
in a retrieval system, or transmitted, in any form or by any means electronic,
mechanical, photocopying , recording, scanning or otherwise, except under the terms
of the Copyright, Designs and Patents Act 1988 or under the terms of a licence issued
by the Copyright Licensing Agency, 90 Tottenham Court Road, London WIP 9HE, UK
without the permission in writing of the Publisher.
Other Wiley Editorial Offices

John Wiley & Sons, Inc., 605 Third Avenue,

New York, NY 10158-0012, USA
Wiley-VCH Verlag GmbH, Pappelallee 3,
D-69469 Weinheim, Germany
Jacaranda Wiley Ltd, 33 Park Road, Milton,
Queensland 4064, Australia
John Wiley & Sons (Asia) Pte Ltd, 2 Clementi Loop #02-01,
Jin Xing Distripark, Singapore 129809
John Wiley & Sons (Canada) Ltd, 22 Worcester Road,
Rexdale, Ontario M9W ILl, Canada

Library of Congress Cataloguing-in-Publication Data

Fairman, Frederick Walker.

Linear control theory : The state space approach ; Frederick
Walker Fairman.
p. em.
Includes bibliographic al references and index.
ISBN 0-471-97489-7 (cased: alk. paper)
I. Linear systems. 2. Control theory. I. Title.
QA402.3.F3 1998
97-41830
629.8'312-dc21
CIP
British Library Cataloguing in Publication Data

A catalogue record for this book is available from the British Library
ISBN 0 471 97489 7
Typeset in part from the author's disks in 10jl2pt Times by the Alden Group, Oxford.
Printed and bound from Postscript files in Great Britain by Bookcraft (Bath) Ltd.
at least
This book is printed on acid-free paper responsibly manufacture d from sustainable forestry, in which
two trees are planted for each one used for paper production.

To Nan cy for her unt irin g sup por t

crt,,

Contents
xiii

Preface

Introduction to State Space

1.1
1.2

1.3
1.4

1.5

1.6

1.7

1.8

1.9

1.10
1.11

Introdu ction
Review of Second Order System s
Pattern s of behavi or
1.2.1
1.2.2 The phase plane
Introdu ction to State Space Modeli ng
Solving the State Differe ntial Equatio n
The matrix expone ntial
1.4.1
1.4.2 Calcula ting the matrix expone ntial
1.4.3 Proper and strictly proper rationa l functio ns
Coordi nate Transfo rmation
Effect on the state model
1.5.1
1.5.2 Determ ination of eAt
Diagon alizing Coordi nate Transf ormatio n
Right-e igenve ctors
1.6.1
1.6.2 Eigenv alue-ei genvec tor proble m
1.6.3 Left-eig envect ors
1.6.4 Eigenv alue invaria nce
State Traject ories Revisit ed
Straigh t line state trajecto ries: diagon al A
1.7.1
1.7.2 Straigh t line state trajecto ries: real eigenv alues
1.7.3 Straigh t line trajecto ries: comple x eigenv alues
1.7.4 Null output zero-in put respon se
State Space Models for the Comple te Respon se
Second order proces s revisite d
1.8.1
1.8.2 Some essent ial feature s of state models
1.8.3 Zero-s tate respon se
Diagon al form State Model
Structu re
1.9.1
1.9.2 Proper ties
1.9.3 Obtain ing the diagon al form state model
Compu ter Calcula tion of the State and Output
Notes and Refere nces

State Feedback and Controllability

2.1
2.2
2.3

Introdu ction
State Feedba ck
Eigenv alue Assign ment
Eigenv alue assign ment via the contro ller form
2.3.1

2
5
7
9
9
10
12
12
13
14
15
16
17
19
20
21

22
23

24
25
26
26
28
29
32
32
33
35
37
39

41
41
42

44
45

viii

2.4

2.5

2.6
2.7

2.3.2
Realizing the controller form
2.3.3
Controller form state transformation
2.3.4
Condition for controller form equivalence
2.3.5
Ackermann"s formula
Controllability
2.4.1
Controllable subspace
2.4.2
Input synthesis for state annihilation
Controllable Decomposed Form
Input control of the controllable subspace
2.5.1
Relation to the transfer function
2.5.2
Eigenvalues and eigenvectors of A
2.5.3
Transformation to Controllable Decomposed Form
Notes and References

State Estimation and Observability

47
49
50
52
55
56
59
60
61
62
63
64
66

67

Introduction
Filtering for Stable Systems
Observers
Observer Design
Observer form
3.4.1
Transformation to observer form
3.4.2
Ackermann's formula
3.4.3
Observability
A state determination problem
3.5.1
3.5.2
Effect of observability on the output
Observable Decomposed Form
3.6.1
Output dependency on observable subspace
3.6.2
Observability matrix
3.6.3
Transfer function
3.6.4
Transformation to observable decomposed form
Minimal Order Observer
3.7.1
The approach
3.7.2
Determination of xR(t)
3.7.3
A fictitious output
3.7.4
Determination of the fictitious output
3.7.5
Assignment of observer eigenvalues
Notes and References

67
68
69
71
72
73
74
75
75

Model Approximation via Balanced Realization

91

4.1
4.2
4.3
4.4

91
91
94
96
96
98
99
101
104
107
108
109
111
111
112
114

3.1
3.2
3.3
3.4

3.5

3.6

3.7

3.8

Contents

4.5
4.6
4.7

4.8

4.9

Introduction
Controllable-Observable Decomposition
Introduction to the Observability Gramian
Fundamental Properties of W0
4.4.1
Hermitian matrices
4.4.2
Positive definite and non-negative matrices
4.4.3
Relating E0 to -\[W0 ]
Introduction to the Controllability Gram ian
Balanced Realization
The Lyapunov Equation
Relation to the Gramians
4.7.1
Observability, stability, and the observability Gramian
4.7.2
Controllability Gramian Revisited
The least energy input problem
4.8.1
Hankel operator
4.8.2
Notes and References

77
78
79
80
80
81
82
82
83
84
85
86
90

Content s

"-

115

Quadratic Control

5_1
5.2
5.3

5.4

5.5

5.6
5.7
5.8

6.3

6.4
6.5

6.6
6.7

Introd ucti on
Observe r Based Controlle rs
Quadrat ic State Feedbac k Control
Motivatin g the problem
5.3.1
5.3.2 Formula ting the problem
5.3.3 Develop ing a solution
Solving the OCARE
Stabilizin g solutions
5.4.1
5.4.2 The Hamilton ian matrix for the OCARE
5.4.3 Finding the stabilizin g solution
Quadrat ic State Estimatio n
Problem formulat ion
5.5.1
5.5.2 Problem solution
Solving the QFARE
Summar y
Notes and Referenc es

Introduc tion
LOG State Feedbac k Control Problem
Problem formulat ion
6.2.1
6.2.2 Develop ment of a solution
LOG State Estimatio n Problem
Problem formulat ion
6.3.1
6.3.2 Problem solution
LOG Measure d Output Feedbac k Problem
Stabilizin g Solution
The Hamilton ian matrix for the GCARE
6.5.1
6.5.2 Prohibiti on of imagina ry eigenval ues
6.5.3 lnvertab ility of T11 and T21
6.5.4 Conditio ns for solving the GFARE
Summar y
Notes and Referenc es

7.3

Introduc tion
Time Domain Spaces
7.2.1
Hilbert spaces for signals
7.2.2 The L 2 norm of the weightin g matrix
7.2.3 Anticaus al and antistabl e systems
Frequen cy Domain Hilbert Spaces
7.3.1 The Fourier transform
7.3.2 Converg ence of the Fourier integral
7.3.3 The Laplace transform
7.3.4 The Hardy spaces: 1- 2 and
7.3.5 Decomp osing 2 space
7.3.6 The H2 system norm
The H:x. Norm: SISO Systems
7.4.1
Transfer function characte rization of the Hx norm
7.4.2 Transfer function spaces
7.4.3 The small gain theorem
The Hoo Norm: MIMO Systems
7.5.1
Singular value decompo sition

Hi

7.4

7.5

147
149
149
150
153
154
155
157
158
158
159
162
165
166
166

167

Signal and System Spaces

7.1
7.2

115
116
119
120
121
122
127
127
130
133
137
137
140
143
145
145

147

LQG Control
6.1
6.2

ix

167
167
168
170
172
173
173
175
176
177
178
179
181
181
183
184
185
185

Contents

7.6
7.7

System Algebra
8.1

8.2

8.3

8.4

8.5

8.6

8.7
8.8

Introduction
Parallel connection
8.1.1
8.1.2 Series connection
System Inversion
Inverse system state model
8.2.1
SISO system zeros
8.2.2
MIMO system zeros
8.2.3
Zeros of invertible systems
8.2.4
Coprime Factorizatio n
Why coprime?
8.3.1
Coprime factorizatio n of MIMO systems
8.3.2
Relating coprime factorizatio ns
8.3.3
State Models for Coprime Factorizatio n
Right and left coprime factors
8.4.1
Solutions to the Bezout identities
8.4.2
Doubly-cop rime factorizatio n
8.4.3
Stabilizing Controllers
Relating W(s) to G(s), H(s)
8.5.1
A criterion for stabilizing controllers
8.5.2
Youla parametriza tion of stabilizing controllers
8.5.3
Lossless Systems and Related Ideas
All pass filters
8.6.1
Inner transfer functions and adjoint systems
8.6.2
Summary
Notes and References

H00 State Feedbac k and Estimation

9.1
9.2

9.3

9.4

9.5
9.6
9.7

10

7.5.2
Induced 2-norm for constant matrices
7.5.3
The L.x. Hx norm for transfer function matrices
Summary
Notes and References

Introduction
Hoo State Feedback Control Problem
9.2.1
Introduction of P_
9.2.2
Introduction of G~(s)
9.2.3
Introduction of J-inner coprime factorizatio n
9.2.4 Consequen ces of J-inner coprime factorizatio n
Hoc State Feedback Controller
Design equations forK
9.3.1
On the stability of A+ B 2 K 2
9.3.2
Determinat ion of 6.
9.3.3
Hx State Estimation Problem
Determinat ion of Te(s)
9.4.1
Duality
9.4.2
Design equations for L 2
9.4.3
Sufficient Conditions
Summary
Notes and References

H00 Output Feedbac k Control

10.1
10.2

Introduction
Developme nt

186
189
190
191

193
193
193
195
196
197
198
199
200
201
202
204
205
206
207
209
212
213
214
215
217
219
220
221
223
223

225
225
227
229
229
230
231
234
234
236
239
242
242
243
244
245
246
246

247
247
248

Cont ents

10.3

10.4

10.5
10.6

10.2.1 Refor mulat ion of P"

10.2.2 An Hx state estim~tor
ack
10.2.3 Introd ucing estim ated state feedb
Hx Outpu t Feed back Contr ollers
10.3.1 Centr al contr oller
10.3.2 Contr oller param etriza tion
10.3.3 Relat ion to Youla param etriza tion
Hx Sepa ration Princ iple
10.4.1 A relati on betwe en Hami ltonia ns
10.4.2 Relat ing stabil izing soluti ons
10.4.3 Deter minat ion of L0
Summ ary
Notes and Refer ences

Linear Alge bra

A.1
A.2
A.3
A.4
A.5

Multi ple Eigen value s and Contr ollabi lity

Block Uppe r Trian gular Matri ces
Singu lar Value Deco mpos ition (SVD)
Differ ent Form s for the SVD
Matri x Inver sion Lemm a (MIL)

Reduced Ord er Model Stability

Problems
C.1
C.2
C.3
C.4
C.5

Probl ems Relat ing to Chap ter 1

Probl ems Relat ing to Chap ter 2
Probl ems Relat ing to Chap ter 3
Probl ems Relat ing to Chap ter 4
Probl ems Relat ing to Chap ter 5

MATLAB Experiments
0.1

0.2

D.3

0.4

State Mode ls and State Resp onse

0.1.1 Contr oller form
0.1.2 Seco nd order linear beha vior
0.1.3 Seco nd order nonlin ear beha vior
0.1.4 Diago nal form
Feed back and Contr ollabi lity
0.2.1 Contr ollabl e state mode ls
0.2.2 Unco ntroll able state mode ls
Obse rver Base d Contr ol Syste ms
0.3.1 Obse rver based contr ollers
vior
0.3.2 Obse rver based contr ol syste m beha
State Mode l Redu ction
ble syste ms
0.4.1 Deco mpos ition of uncon trolla ble and/o r unob serva
ty
0.4.2 Weak contr ollabi lity and/o r obser vabili
rvabi lity
0.4.3 Energ y interp retati on of the contr ollabi lity and obse
Gram ians
0.4.4 Desig n of reduc ed order mode ls

xi
248
251
253
254
255
256
260
261
262
267
269
269
270

271
271
272
274
276
277

279

283
283
285
287
288
290

293
293
293
293
295
296
297
297
298
299
301
303
303
304
305
306
307

References

309

Index

313

Preface
g control
This book was written with the intent of prmidin g students and practicin
system
control
use
to
needed
theory
engineers with the basic backgro und in control
those
of
nt
treatme
detailed
a
with
begins
design software more productively. The book
the
of
er
remaind
the
in
needed
are
that
aspects of the state space analysis of linear systems
manner:
text. The book is organized in the following
n \ia
The first four chapters develop linear system theory including model reductio
balanced realization.
Chapter s S and 6 deal with classical optimal control theory.
Hx control
The final four chapter s are devoted to the development of suboptim al
theory.
needed
The mathem atical ideas required in the development are introduc ed as they are
beyond
venture
to
reader
the
e
motivat
to
using a "just-in -time" approac h. This is done
ic contror ',
.the usual topics appeari ng in introduc tory undergr aduate books on .. automat
leYel books
duate
postgra
to
d
restricte
been
to more advanced topics \Yhich have so far
titles.
their
in
control"
"robust
and
having the terms "mathem atical control theory"
the final
at
course
ester
two-sem
or
one
a
This book can be used as the text for either
lewl.
duate
postgra
g
beginnin
the
at
course
r
year undergr aduate level or as a one semeste
" or
S}Stems
and
''signals
either
in
course
Students are assumed to have taken a basic
state
the
of
ge
knowled
tory
introduc
an
.
"autom atic control ... Althoug h not assumed
would
space analysis of systems together with a good understa nding of linear algebra
book.
this
in
d
presente
ideas
benefit the reader's progress in acquiring the
t Yiew of
Ideas presented in this book which provide the reader with a slightly differen
are as
ks
textboo
other
reading
by
d
obtaine
control and system theory than would be
follows:
and 1or
The so-called PBH test \Yhich is usually presented as a test for controllability
ues in
eigenval
obserYability is used through out the present book to characterize
output
and;or
control problems involving eigenvalue assignment by state feedback
injection.
deYelop An easy to underst and matrix Yariational technique is used to simplify the
LQG
and
ic
ment of the design equation s for the time invarian t, steady-state. quadrat
controllers.
ment of the
The relatively simple idea of the L 2 gain is used as a basis for the develop
H 00 controller.

xiv

Preface

Concerning the style of the book, the beginning section, "Introduction", for each
chapter contains motivational material and an overview of the ideas to be introduced in
subsequent sections in that chapter. Each chapter finishes with a section called "Notes
and References", which indicates a selection of other sources for the material treated in
the chapter, as well as an indication of recent advances with references.
I would like to thank the following colleagues in the Department of Electrical and
Computer Engineering at Queen's University for proof-reading parts of the manuscript:
Norm Beaulieu, Steve Blostein, Mingyu Liu, Dan Secrieu and Chris Zarowski. Special
thanks go to my former research student Lacra Pavel for proof-reading and advice on
Chapters 6, 9 and 10 as well as to Jamie Mingo in the Department of Mathematics and
Statistics at Queen's University for his help with some of the ideas in Chapter 7. Thanks
go also to Patty Jordan for doing the figures. Finally, I wish to acknowledge the
contribution to this book made by my having supervised the research of former research
students, especially Manu Missaghie, Lacra Pavel and Johannes Sveinsson.
The author would appreciate receiving any corrections, comments, or suggestions for
future editions should readers wish to do so. This could be done either by post or e-mail:
< [email protected] >.

1
Introduction to State Spa ce

1.1

Introduction

A well known behavioral phenomen on of dynamic systems is the appearanc e of an output

in the absence of an input. This effect is explained once it is recognized that the internal
storage of energy in the system at the beginning of the response time will produce an
output. This kind of behavior is referred to as the system's zero-input response.
Alternative ly, the production of an output caused solely by an input when there is no
energy storage at the start of the response time is referred to as the zero-state response.
These two classes of response are responsibl e for all possible outputs and in the case of
linear systems we can always decompose any output into the sum of an output drawn
from each of these classes. In this chapter \Ye will use the example of a second order system
together with both the zero-input response and the zero-state response to introduce the
reader to the use of the state space in modeling the behaYior oflinear dynamic systems.

1.2

Review of Second Order Systems

A commonly encountere d physical process which we will use in the next two sections to
introduce the state modeling of linear dynamic systems is the electric circuit formed by
connecting an ideal constant resistor Rc, inductor Le, and capacitor Ce in series in a closed
loop as shown in Figure 1.1
Suppose the switch is closed at t = t, < 0 so that there is a current flow i(t). t ::.0: 0. and a
voltage across the capacitor y(t), t ::.0: 0. Then applying Kirchhoff' s voltage law yields

Rei(t)

+Led~~)+ y(t) =

where the current in the circuit depends on the capacitor voltage as

"( ) =

ce dy(t)
dt

Combining these equations gives a second order differential equation in the capacitor
voltage, y(t),
d l'(t)
d\(t)
(1.1)
--+a2 y(t) = 0
--+a
I dt
dt 2

Introduction to State Space

i(t)

L,

Switch

C,

Figure 1.1

y(t)

Electric circuit with charged capacitor. Switch closed prior to t = 0

where

and we refer to the capacitor voltage as the system's output.

1.2.1

Patterns of behavior

The differential equation (1.1) is said to gm;ern the evolution of the output, y( t), since it
acts as a constraint relating y( t), d~~t) , and ddt~r) to each other at each instant of time. We
will see now that once the initial conditions, i.e., the values of initial output, y(O), and
initial derivative of the output, y(O), are specified, the differential equation, (1.1 ),
completely determines the output, y(t), for all positive timet E (0, oo). We obtain y(t)
as follows.
Suppose we have y(t) such that (1.1) is satisfied. Then denoting the derivatives of y(t)
as
dy(t) = g(t)
dt

d2y( t) = h( t)
dt 2

we see that equation (1.1) becomes

( 1.2)

Now the only way this equation can hold for all t > 0 is for h(t) and g(t) to be scalar
multiples of y(t) where
g(t)

sy(t)

( 1.3)

Otherwise equation (1.2) can only be satisfied at specific instants of time. Therefore with
this assumption assumption (1.2) becomes
p(s)y(t) = 0

( 1.4)

Review of Second Order Systems

where p(s) is the second degree polynomial

pi,sj =

l- + a 1 1--:- a2

Finally, equation (1.4) holds for all time, when y( t) is not zero for all time, i.e., the trivial
solution. if and only if sis any one of the roots. {..\, : i = 1, 2} of p(s),

(1.5)
Returning to the requirement that y(t) and its derivatives must be constant scalar
multiples of each other, equation ( 1.3), the function that has this property is the
and has series expansion
exponential function. This important function is denoted as

est=~ (st)i
~

i=O

.,

( 1.6)

/.

where i!. (factorial i). is the product

i! = ( i) ( i - 1) (i - 2) ... ( 1)

i
i

=1

>0
=0

Notice that a bit of algebra shows us that the derivative of ~.Y. equation (! .6 ). has the
desired property of being an eigenfunction for differentiation,
dest

-=se
dt

sr

Now we see from the foregoing that est satisfies equation (1.1) when s = )q or .A 2 .
Therefore any linear combination of e)'~t and e>-,t satisfies equation (1.1) so tharthe output
y(t) is given in general as

(1.7)
where the k;s are constant scalars chosen so that y(t) satisfies the initial conditions. We
can be do this by solving the equations which result from setting the given values for the
initial conditions, y(O) and y(O). equal to their values determined from equation ( 1. 7), i.e.,
by solving
( 1.8)

for k1, k2. Notice that we can do this only if .A 1 t- A.c- In orderto proceed when A. 1 = .A2 we
replace equation (1.7) with
( 1.9)

Introduction to State Space

and determine the k;s from the appropriate equations to ensure that the initial conditions
are satisfied.
Returning to the behavior of the physical process that is under analysis, notice that
since R"' Le, and Ce are real, the a;s are real. As a consequence the roots >.; of p(s).
equation (1.5), are both real or are both complex. Moreover when these roots are complex
they are conjugates of each other, i.e., ).. 1 = >.2.
More generally, if all the coefficients of a polynomial of any degree are real, each
complex root must be matched by another root which is its complex conjugate. This
property is important in the context of the behavior oflinear physical processes since the
parameters of these processes, e.g., mass, heat conductivity, electric capacitance, are
always real so that the coefficients of p(s) are always real.
Now a plot of the output, y( t), versus time, t, reveals that there are two basic patterns
for the behavior of the output depending on whether the >..;s are real or are complex
conjugate pairs.
If the A;s are real, we see from equation (1.8) that the k;s are also real and the output
y(t): t E (0, oo) is given as equation (1.7) or (1.9). In this case we see that the output
voltage y(t) exhibits at most one maximum and decays without oscillation to the time axis
as t tends to infinity. Notice from equation (1.5) that the A;s are real provided the
parameters Re, L"' Ce have values such that C'i/ ~ a2 .
< a2 , then we see from (1.5) that
Alternatively, if the >..;s are complex, i.e., if
>.. 1 = X2 and from (1.8) that k 1 = k;. Thus k 1 e>'~ 1 and k 2e>- 21 are complex conjugates of
each other and their sum which gives y(t), equation (1.7), is real. Incorporating these
conjugate relations for the >..;sand the k;s in equation ( 1. 7) allows us to write the output as
a damped oscillation

C1/

(1.10)
where
k 1 = Re[kt]

+ jim[kt]

_ _1 (Im[ktl)
e -tan
Re[ki]
Thus we see from (1.1 0) that the output voltage across the capacitor, y( t), swings back
and forth from its initial value to ever smaller values of alternating polarity. This behavior
is analogous to the behavior of the position of a free swinging pendulum. The capacitor
voltage (pendulum position) eventually goes to zero because of the loss of heat energy
from the system resulting from the presence of Re (friction). In this analogy, voltage and
current in the electric circuit are analogous to position and velocity respectively in the
mechanical process. The inductance Le is analogous to mass since the inductance resists
changes in the current through itself whereas the inertial effect of mass causes the mass to
resist change in its velocity.
In addition, notice from equation ( 1.1 0) that the frequency of the oscillation, Im[).. J], as
well as the time constant associated with the decay in the amplitude of the oscillation,
(Re[>..J])- 1 , are each independent of the initial conditions and depend on the system
parameters, Re, Le, Ce only, i.e., on a1, az only.

Review of Second Order Systems

The previous discussion leads to the following characterization of the zero-input

response of dynamic processes whose behavior can be modeled by second order
differential equations with constant coefficients.
1
(i) The zero-input response, y(t): t > 0, depends on the set of signals {e'\; : i = 1, 2}
referred to as modes of the system where the constants >..il (system eigenvalues), are
roots of the polynomial p(s), (characteristic polynomial).
(ii) The steady state zero-input response is zero, i.e., limy(t) = 0, for any initial
conditions if and only if all the >..;s are negative or have negative real part, i.e.,
Re[>..;] < 0, i = 1, 2 . In this situation we say that the system is stable.
(iii) We have Re[>..;] < 0, i = 1, 2, if and only if a 1 > 0 and a 2 > 0. More generally, the
condition a; > 0, i = 1, 2, n for systems whose behavior is governed by differential equations in the order of n > 2, is necessary but not sufficient for the system to
be stable, i.e., is necessary but not sufficient for all >..;s to have negative real part
t~oo

1.2.2

The phase plane

We have just seen that, when there is no input, a second order system having specified a;s
has output,y(t), which is specified completely by the initialconditions,y(O) andy(O). This
important observation suggests that the same information concerning the behavior of the
system is contained in either (a) a plot of y(t} versus tor (b) a plot of y(t) versus y(t).
Thus if we make a plot of y(t) versus y(t), the point representing y(t),y(t) in the y(t)
versus y(t) plane traces out a curve or trajectory with increasing time.
The two dimensional space in which this trajectory exists is referred to as the state
space and the two-element vector consisting of y(t) and y(t) is referred to as the state,
denoted as x(t) where

x(t)

[~(t)]
y(t)

This approach to visualizing the behavior of a dynamic process was used by

mathematicians at the end of the last century to investigate the solutions for second
order nonlinear differential equations, i.e., equations of the form (1.1) but with the a;s
functions of y(t) and/or y(t). The term phase plane plot was used to refer to the state
trajectory in this case. Since, in general, the dimension of the state space equals the order
of the differential equation which governs the output behavior of the process, the state
space cannot be displayed for systems of order greater than two. Even so, the mathematical idea of the state space has become of great practical and theoretical importance in
the field of control engineering.
Referring to the previous section, we see that the state trajectory for a dynamic process
whose behavior can be modeled by a second order differential equation with constant
coefficients; can exhibit any one of the following four fundamental shapes.
(i) If the A;s are complex and Re[>..;] < 0 the system is stable and the state trajectory
spirals inwards towards the origin.
(ii) If the \s are complex andRe[>..;] > 0 the system is unstable and the state trajectory
spirals outwards away from the origin.

Introduction to State Space

y(t)

b
d

y(t)

y(t)

Figure 1.2

Plot of y(t) vs. t and y(t) vs. y(t) when A;s complex

(iii) If the A.;s are real and both A;s are negative the system is stable and the state
trajectory moves towards the origin in an arc.
(iv) If the A.;s are real and one or both are positive the system is unstable and the state
trajectory moves away from the origin in an arc.
Notice that state trajectories (ii) and (iv) do not occur in the present example of an
electric circuit. This results from the fact that the parameters R"' Le, Ce are positive. Thus
the coefficients, a; : i = 1, 2 of the characteristic polynomial, equation (1.4) are positive so
that the A;S are negative or have negative real parts. This implies that we are dealing with a
stable dynamic process, i.e., state trajectories tend to the origin for all initial states.
So far in this chapter we have used an electric circuit as an example of a system. We
used the character of the behavior of this system in response to initial conditions to

Introduction to State Space Modeling

introduce the concept of the state of a system. In the next section this concept is made
more specific by introducing the mathematical characterization of a system referred to as
a state model.

1.3

Introduction to State Space Modeling

We saw in the previous section that once a second order system is specified, i.e., once the
a;s are given numerical values, the zero-input response is determined completely from the
system's initial conditions, y(O),y(O). In addition, we noted that the second derivative of
the output is determined at each instant from y(t) and y(t) through the constraint (1.1).
These facts suggest that it should be possible to obtain the zero-input response by solving
two first order differential equations involving two signals, x 1 ( t), x 2 ( t), which are related
uniquely to y(t),y(t). One straightforward way of doing this is to identify y(t) with
x 1 (t)and y(t) with x 2 (t), i.e.,

y(t)

= XJ (t)

(1.11)

y(t) = x 2 (t)

( 1.12)

An immediate consequence of this identification is that at every instant the derivative of

x 2 (t) equals x 1 (t)

x2(t)

( 1.13)

= x1 (t)

Moreover, rewriting the second order differential equation, (1.1), as

:t (y(t)) =

~a 1 y(t) ~ a2y(t)

andusing equations (1.11-1.13) gives us the differential equation for x 1 (t) ail(1.14)

Thus we see from equation (1.13) and (1.14) that the derivative of each of the x;s is a
(linear) function of the x;s. This fact is expressed in matrix notation as
( 1.15)

x(t) = Ax(t)
where

A=

[ ~al
I

x(t)=

[xi(t)]
x2(t)

with the vector x(t) being referred to as the state, and the square matrix A being referred
to as the system matrix. In addition we see from equation (1.12) that

y(t) = Cx(t)

(1.16)

Introduction to State Space

where

[0

1]

t matrix.
with C being a row vector referred to as the outpu
ion (1.1) is equivalent to the vector
equat
ential
differ
In summ ary the second order
(I.l6) . These equat ions, (1.15, 1.16)
differential equat ion (1.15) and the outpu t equat ion
system in the absence of an input.
constitute a state mode l for the second order
d by a block diagr am involving the
Alternatively, the state model can be represente
integrators, and scalar multipliers
interc onnec tion of blocks which opera te as summers,
tor 1/sis used to indicate integration.
on the comp onent s of the state. The Laplace opera
to obtai n a state model for the
More generally, we can use the foregoing procedure
sses as follows.
zero-input response of higher order dynam ic proce
1h order process is gover ned by
n
an
of
nse
respo
Supp ose the zero- input
( 1.17)
where

fy comp onent s of the state with

Then we proce ed as in the second order case to identi
derivatives of the outpu t as

(t) = y(n~l)(t)
x2(t) = Y(n~2)(t)

X]

(1.18)

equat ion (1.15) having a system

Thus using (1.17, 1.18) we obtai n a vecto r differential

x.(t)

i,(t)

x,(t)

x,(t)

-a,

-a,

Figure 1.3

Block diagra m repres entati on of the state model

Solving the State Differential Equation

matrix A given as
-a I
1

A=

-a2

-a3
0

-an

0
(1.19)

and output equation, (1.16), having an output matrix C given as

c=

[0

... 0

1]

The pattern of zeros and ones exhibited in A, (1.19), is of particular importance here.
Notice that the coefficients of the characteristic polynomial

appear as the negative of the entries in the first row of A. Matrices exhibiting this pattern
are referred to as companion matrices. We will see shortly that given A in any form, the
characteristic polynomial is related to A as the matrix determinant

p(s)

det[s/- A]

This fact is readily seen to be true in the special case when A is in companion form.

1.4

Solving the State Differential Equation

Recall that the solution to a scalar differential equation, e.g., (1.1), involves the scalar
exponential function, e>.t. In this section we will show that the solution to the state
differential equation, (1.15), involves a square matrix, eAt, which is referred to as the
matrix exponential.

1.4.1

The matrix exponential

Suppose we are given the initial state x(O) and the system matrix A, either constant or time
varying. Then we obtain a solution to the state differential equation, (1.15), by finding
</>(t), the square matrix of scalar functions of time, such that

x(t) = <f>(t)x(O)

(1.20)

where </>(t) is referred to as the transition matrix.

Since the state at each instant of time must satisfy the state differential equation, (1.15),
the transition matrix is a matrix function of the system matrix A. In this book A is
constant. In this case the dependency of <f>(t) on A is captured by the notation

<f>(t)

=eAt

(1.21)

10

Introduction to State Space

where the square matrix eAt is referred to as the "matrix exponential of At" since it can be
expressed as an infinite series reminiscent of the infinite series for the exponential of a
scalar. (1.5), i.e.,
( 1.22)

In order to show that the transition matrix given by (1.22) solves the state differential
equation, (1.15), we differentiate the foregoing series expansion for the matrix exponential of At to obtain
2A 2 t 3A 3 t 2 At
deA 1
~=A+--+--Ae

2!

dt

3!

4A 4 t3

+--+
4!

AeAt

Then using this relation to differentiate the assumed solution

x(t) =eAr x(O)

( 1.23)

yields

x(t) = AeA 1x(O) = Ax(t)

and we see that (1.23) solves the state differential equation, (1.15).

1.4.2

Calculating the matrix exponential

There are many ways of determining eAt given A. Some of these approaches are suitable
for hand calculation and others are intended for use with a digital computer. An
approach of the first kind results from using Laplace transforms to solve the state
differential equation. We develop this approach as follows.
We begin by taking the Laplace transform of (1.15) to obtain
( 1.24)

sX(s) - x(O) = AX(s)

where A is 2 x 2 we have
x;

Xi(s) =

j xi(t)e-' dt
1

Then rewriting (1.24) as

(sf- A)X(s)

= x(O)

( 1.25)

we see that provided sis such that (sf- A) is invertible, we can solve (1.25) for X(s) as
X(s) = (sf- A)- 1x(O)

( 1.26)

Solving the State Differential Equation

11

Now (sl- A)- 1 can be expressed using Crammer's rule as

(

sl -A

adj[s/- A]
- det[s/ -A]

_1 _

where when A is an n x n matrix, the adjugate matrix, adj [s/ - A], is an n x n matrix of
polynomials of degree less than n and det[s/ - A] is a polynomial of degree n.
Finally, taking the inverse Laplace transform of (1.26) yields

x(t) = - 1 [(s/- A)- 1)x(0)

and we see, by comparing this equation with (1.23), that

Now in the case where A is the 2 x 2 matrix given by (1.15), we have

(sl _A)-I= adj[s/- A]
det[s/- A]

s+ai
-1

a2]-I

(1.27)

where

adj[s/- A]=

[s s -a2]
+a
1

Notice from the previous section that det[s/- A]= p(s), (1.4), is the characteristic
polynomial. In general any n by n system matrix A has a characteristic polynomial with
roots {>..i : i = 1, 2 n} which are referred to as the eigenvalues of A. The eigenvalues of
the system matrix A play an important role in determining a system's behaviq.r.
Returning to the problem of determining the transition matrix for A, (1.15), we apply
partial fraction expansion to the expression for (sl- A)- 1 , (1.27), assuming det[s/- A]
has distinct roots, i.e., >.. 1 -1 >..2 , to obtain
(1.28)
where

12

Introduction to State Space

Finally, taking the inverse Laplace transform of (1.28), we obtain the transition
matrix a~~
-)q)'l(e>" 1
-.\2e;.,,

e.>.'')]

+ AJe;.,,

(1.29)

We will show in the next section that there are other ways of modeling a dynamic
process in the state space. This non-uniqueness in the state model representation of a
given dynamic process results from being able to choose the coordinates for expressing
the state space. In the next section we will use this fact to simplify the determination of eAr
by working in co-ordinates where the state model has a diagonal A matrix.

1.4.3

Proper and strictly proper rational functions

Before continuing to the next section, notice that when A is ann x n matrix, adj [sf - A] is
an n x n matrix of polynomials having degree no larger than n - 1. Thus, since the
characteristic polynomial for A, det[sl- A], is of degree n, we see from (1.27) that
1 is ann x n matrix of strictly proper rational functions.
(sfIn general a rational function

Ar

n(s)
r(s) = d(s)
is said to be;
(i) strictly proper when the degree of its numerator polynomial is less than the degree of
its denominator polynomial, i.e.,
deg[n(s)] < deg[d(s)]
(ii) proper when the degree of its numerator polynomial equals the degree of its
denominator polynomial, i.e.,
deg[n(s)]

deg[d(s)]

In subsequent chapters we will see that this characterization of rational functions plays
an important role in control theory.

1.5

Coordinate Transformation

In Section 1.3 we saw that the zero-input response for a system could be obtained by
solving a state vector differential equation where the components of the state were
identified with the output and its derivatives. In this section we examine the effect of
changing this identification.

..,.,.

~.

Coordinate Transformation

1.5.1

13

Effect on the state model

Referring to the second order process used in the previous section, let x( t) denote the state
obtained by setting

= v[x1(t)]
[.Y(t)]
y(t)
x (t)

(1.30)

where Vis any invertible (nonsingular) 2 x 2 matrix of constants. In the previous section
V was the identity matrix.
Now we see from (1.11, 1.12, 1.30) that the state x(t) used in the previous section is
related in a one-to-one fashion to the state x(t) as

x(t)

(1.31)

Vx(t)

where we say that x(t) is the state in the old or original coordinates and x(t) is the state in
the new or transformed coordinates. Then the state model parameters in the old
coordinates, (A, C), are transformed by a change of coordinates to (A, C) in the new
coordinates as

- -

(A, C) ~ {A, C)

(1.32)

where

We can develop this relation as follows.

First using (1.31) in (1.15) we obtain

V x= AVx(t)
which, since V is invertible, can be multiplied throughout by

x (t) = Ax(t)
where

Again, using (1.31) in (1.16) we obtain

y(t) = Cx(t)
where

v-I

to give

14

Introduction to State Space

Notice that the transition matrix, eAt, which applies in the new coordinate s is related to
the transition matrix, eAt, in the original coordinate s as

(1.33)

1.5.2

Determination of eAt

The flexibility provided by being able to choose the coordinate s for the state model
representa tion of a dynamic process is often of considerab le use in the analysis and design
of control systems. We can demonstra te this fact by using a change of coordinate s to
calculate the transition matrix.
Suppose we are given a two dimension al system matrix A having a characteris tic
polynomia l, det[s/- A], with distinct roots (eigenvalues), i.e., )q -=f. >. 2 . Then we can
always find a coordinate transforma tion matrix V so that the system matrix A in the new
coordinate s is diagonal and

i (t) = Ax(t)

( 1.34)

where

with entries along the diagonal of A being the eigenvalues of A.

Now when the system matrix is diagonal, the correspond ing transition matrix is also
diagonal. We can see this by noting that the state differential equation in these
coordinate s, (1.34), consists of two scalar first order differential equations which are
uncoupled from each other

x 1(t)

.A 1x 1(t)

x2(t)

>-2x2(t)

so that their solution can be immediate ly written as

x1 (t) = / 1 t x1(0)
x2 (t) = e>.'tx 2 (0)

( 1.35)

which in matrix form is

( 1.36)

Diagonalizing Coordinate Transformation

15

Thus we see that the transition matrix is indeed diagonal

Having determined the transition matrix for A, we can use (1.33) to determine the
transition matrix for A as

(1.37)
with V being the coordinate transformation matrix which makes A diagonal.
Now we will see in the next section that, in general, the coordinate transformation
matrix V needed to make A diagonal depends on the eigenvectors of A. However in the
special case when A is a 2 x 2 companion matrix, (1.15), with >.. 1 -=1- >..2, the required
coordinate transformation matrix is simply related to the eigenvalues of A as

(1.38)
We can see that this coordinate transformation gives rise to a diagonal system matrix by
/ .
using

(1.39)
to obtain

Then since J + a1s + a2 = (s- A.I)(s- >..2) we have a 1 = -(>.. 1 + >.. 2) and""a2 = >..,>..2.
Therefore the foregoing expression for A reduces to

(1.40)
Finally, the expression obtained for eAt using V, (1.38), in (1.37) equals (1.29) which
was obtained at the end of Section 1.4 through the use of Laplace transforms.
The foregoing approach to the determination of the transition matrix requires the
determination of a coordinate transformation matrix V which diagonalizes the system
matrix A. We will see in the next section that the columns of the coordinate transformation matrix required to do this are right-eigenvectors for A.

1.6

Diagonalizing Coordinate Transformation

As mentioned previously, the roots of the characteristic polynomial for a square matrix A
are called the eigenvalues of A. In this section we will see that corresponding to each of A's

Introd uctio n to State Spac e

16

igenvector. More over we will see that

eigenvalues there is at least one right and one left-e
transf orma tion V requi red to make
when the eigenvalues of A are distinct, the coord inate
A. [n addit ion we will see that
A diagonal has columns equal to the right-eigenvectors of
of A.
ectors
v-I has rows which are the transpose of the left-eigenv

1.6.1

Righ t-Eig env ecto rs

matri x in comp anion form, (1.15),

Cons ider the special case when A is a two-b y-two
g the chara cteris tic polyn omial as
having unequ al eigen value s{.\: 1,2}. Then writin
At= -a 1 Ai- a 2 we see that
i

= 1, 2

(1.41)

or

i = 1, 2

( 1.42)

where

i = 1, 2
the i1h eigenvalue, right-eigenvector
Notic e that (1.42) is a general expression relating
vi is said to be the right-eigenvector
pair (Ai, vi) for the any squar e matri x A, where
a major role in the state analysis of
corre spond ing to the eigenvalue Ai These pairs play
prese nt instan ce is a result of A being
systems. The partic ular dependence of vi on Ai in the
in comp anion form.
v- 1A V diagonal, we combine the
Conti nuing with the const ructio n of V to make
n
equat ions given by (1.42) fori= 1 and i = 2 to obtai

AV= VA

( 1.43)

where V, A are given as

A=[A01 A2OJ
A , V is invertible and we can preNow when A has distinct eigenvalues, i.e., A1 i= 2
multiply (1.43) by v- 1 to obtai n

matri x requi red to make A diagonal.

Thus we see that Vis the coord inate trans forma tion
necessarily in comp anion form,
x,
More generally, suppo se A is any n x n matri not
out that this condi tion of
turns
it
Now
j.
which has distin ct eigenvalues, ,\ i= A1 : i ic
of A to be indep enden t, i.e., vi and vi
distinct eigenvalues is sufficient for the eigenvectors
enden t colum ns and is theref ore
point in different directions. There fore V has indep

Diagona lizing Coordina te Transfor mation

17

invertible where
(1.44)
i

1, 2,- ,n

( 1.45)

and v- 1AVis diagonaL

In the special case when A is in compani on form, (1.19), its eigenvalues,
{>..- : i = 1, 2, , n} are related to its eigenvectors, {vi : i = 1, 2, , n} as
,n-1
v iT -:_ [ /\"
I

.A;

1)

In order to see that this result holds, set the last entry in vi equal to one. Then taking A
in compani on form, (1.19), solve the last scalar equation in (1.45) and use the result to
solve the second to last scalar equation in (1.45). We continue in this way solving
successive scalar equation s in (1.45), in reverse order, until we reach the first scalar
equation. At this stage we will have all the entries in vi. These entries satisfy the first scalar
equation in (1.45) since A; is a root of the characteristic polynomial whose coefficients
appear with negative signs along the first row of A.
Jn general, when A is not in any special form, there is no special relation between the
eigenvalues and the corresponding eigenvectors. Thus in order to determine the eigenvalue, right-eigenvector pairs when A is not in any special form we need to determine
(.A;, vi) : i = 1, 2, n so that the equations
i = 1,2, n

(1.46)

are satisfied.

1.6.2

Eigenvalue-Eigenvector problem

The problem of determining (.A;, vi) pairs which satisfy (1.46) is referred to as the
eigenvalue-eigenvector problem. There are well established methods for solving this
problem using a digital computer. In order to gain additional insight into the nature of
the eigenvalue-eigenvector problem we consider a theoretical approach to finding
eigenvalue-eigenvector pairs to satisfy (1.46).
To begin, suppose we rewrite (1.46) as
(AI- A)v = 0

(1.47)

where 0 denotes the null vector, i.e., a vector of zeros. Then in order for the solution v to
this equation to be non-null we must choose .A so that the matrix .AI - A is singular, i.e.,
does not have an inverse. Otherwise, if AI- A is invertible we can solve (1.47) as

~nd the only solution is the trivial solution v = 0. However when (AI - A) is not
Invertible, (1.47) can be satisfied by v f= 0.

Introduction to State Space

18

Now from Cramme r's rule for matrix inversion we have

(AI _ A )_ 1= adj[M- A]

det[M- A]
an
Therefor e )...] -A does not have an inverse when det[>..I -A] = 0, i.e., )... = >..;
eigenvalue of A.
Next recall that singular matrices have dependen t columns. Therefor e A;!- A has
dependen t columns so that we can find scalars {vk : k = 1, 2, , n} not all zero such that
n

L[(\1 - A)]ku~

= 0

( 1.48)

k=l

be
where [(>..J- A)]k: k = 1,2 ,n denote columns of >..J- A. Notice that (1.48) can
where
vi
=
v
and
>..;
=
)...
rewritten as (1.47) with

Since we can always multiply (1.48) through by a nonzero scalar a, the solution, vi, to
(1.48) or (1.47) is not unique since m/ is another solution. More generally, we say that the
eigenvectors of a given matrix are determin ed to within a scalar multiple, i.e., the
directions of the eigenvectors are determin ed but their lengths are arbitrary .
Assumin g that A has a complete set of (n independ ent) right-eigenvectors, we can
decompo se any initial state as
n

x(O) = l:a;vi = Va

(1.49)

i=l

where

with the a 1s being found as

a=

v- 1x(O)

This decompo sition of the state into a linear combina tion of right-eigenvectors of A
plays an importan t role in the analysis of system behavior. This is illustrate d in the next
of
section where we will use the eigenvectors of A to reveal certain fundame ntal properties
state trajectories.
Unfortun ately, when an n x n matrix A has multiple eigenvalues, i.e., when
det[>..I - A] does not have n distinct roots, the number of eigenvectors may or may not
be equal ton, i.e., A may or may not have a complete set of eigenvectors. When A does not
ed
have a complete set of eigenvectors, other related vectors, referred to as generaliz
a
provide
to
tors
eigenvectors, (Appendix), can be used together with the eigenvec
it
case
this
in
However
(1.49).
decompo sition of the state space in a manner similar to

Diagonalizing Coordinate Transformation

19

is impossible to find a nonsingular matrix V such that v- 1AVis diagonal. Then A is said
to be not diagonalizable.
In summary, the condition that A has distinct eigenvalues is sufficient but not
necessary for A to be diagonalizable. Most of the time, little additional insight into
control theory is gained by discussing the case where A does not have a complete set of
eigenvectors. Therefore, we are usually able to avoid this complication without loss of
understanding of the control theory.

1.6.3

Left-Eigenvectors

Suppose A is any n x n matrix having n distinct eigenvalues. Then we have seen that the
matrix V having columns which are right-eigenvectors of A is invertible and we can write
(1.50)
where Avi =>..;vi: i = 1, 2, nand
/

Now suppose we post-multiply both sides of (1.50) by

v- 1 to obtain

v- 1A = Av- 1
Then if we denote

(1.51)

v- 1 in terms of its rows as

and carry out the matrix multiplication indicated in (1.51) we obtain

A] [)..>..2w2T
w Tl

w:T A
ws
[

..

...

wnT A

)..n wnT

Therefore it follows, by equating corresponding rows on either side of this equation, that
i = 1,2, ,n

(1.52)

20

Introduction to State Space

which transposing through out gives

i = 1, 2, , n

( 1. 53)

of AT
Thus we see from equatio n (1.53) that the column vector w; is an eigenvector
left
the
on
with corresponding eigenvalue .A.;. However, since the row vector wiT appears
it
ish
distingu
to
side of A in equatio n (1.52), w' is referred to as a left-eigenvector of A
from the correspo nding right-eigenvector of A, namely, v;.
rs just
Notice that, since v-I V is an identity matrix, the left and right-eigenvecto
defined are related to each other as

=0 i=fj}
= 1 i =i

(1.54)

In addition notice that any nonzero scalar multiple of w;

satisfies (1.52), i.e.

Therefore zi is also a left-eigenvector of A and we see ftom equation (1.54) that

the left and right eigenvectors are related as
ziT v

=0
= 'li =I 0

=I j

=j

in general

nality
This basic characteristic ofleft and right-eigenvectors, referred to as the orthogo
r of
behavio
the
to
property, is responsible for a number of fundam ental facts relating
state models.

1.6.4

Eigen value invari ance

eigenvalues
Before we go to the next section, it is importa nt to note the basic fact that the
invertible
of A and of A are the same whenever A is related to A as A = v-I A V for any
v-I and
matrix V. We can see this by premultiplying both sides of equatio n (1.45) by
inserting vv-I between A and vi, viz.,

Then setting

ti

= v-I vi and taking v-I A v =A gives

Thus the
which implies that the eigenvalue, right-eigenvector pairs for A are (A.;, t;).

State Trajectories Revisited

21

eigenvalues of A equal the eigenvalues of A independent of the coordinate transformation

matrix V.
Alternativelv. another way we can see this fact is to carry out the following
manipulations
det[s/- A]= det[sl- V- 1AV] = det[V- 1 (s/- A)Vj

v-I det[s/- A1 det V

det

det[s/- A]

Thus A and A have the same characteristic polynomial, and since the roots of a
polynomial are uniquely dependent on the coefficients of the polynomial. A and A
have the same eigenvalues.
Finally, since the differential equations modeling the behavior of dynamical processes
must have real coefficients, we can always work in coordinates where the state model
parameters, (A, C), are real. As a consequence, if (.\,vi) is a complex eigenvalue-eigenvector pair for A, then (,>,j. l,tx) is also an eigenvalue-eigenvector pair for A.

1. 7

State Trajectories Revisited

We saw in Section 1.6.2 that assuming A has a complete set of eigenvectors. any initial
state can be written in terms of the eigenvectors of A, (1.49). In this section this fact is used
~to gain additional insight into the nature of a system's state trajectories and zero-input
response.
More specifically, under certain conditions on the matrix pair, (A, C), a system can
exhibit a null zero-input response, y(t) = 0 for all r > 0, for some non-null initial state,
x(O) -=1- .When this occurs we say that the state trajectory is orthogonal (perpendicular)
to cT, denoted cT __Lx(t), since the output depends on the state as y(t) = Cx(t). Two
vectors a, f3 are said to be orthogonal if
aT

f3 = 0

When the state space is n dimensional, a state trajectory which produces no output lies
in an n - 1 dimensional subspace of state space which is perpendicular to the vector cT.
For example, if the system is second order, this subspace is a straight line perpendicular to
cT: if the system is third order. this subspace is a plane perpendicular to cT Thus, in the
second order case, we obtain a null output if we can find an initial state such that it
produces a straight line trajectory

x(t)

1(t)v

satisfying
Cv

=0

where !(t) is a scalar function of time and vis a constant two-element vector. When n > 2,

Introduction to State Space

22

straight line
any trajecto ry orthogo nal to cT can be decomp osed into a sum of
re an
Therefo
cT.
to
nal
trajecto nes all lying in the n - 1 dimensi onal subspac e orthogo
property
the
of
nding
understa
an
understa nding of straight line trajectories is essential to
states.
posed by certain systems of having a null zero-inp ut response to certain initial

1.7.1

Straig ht line state trajectories: diago nal A

a straight line
Suppose A is a 2 x 2, real, diagona l matrix. Then the state trajecto ry is
can see this
We
axes.
ate
whenever the initial state lies only on one of the two coordin
immedia tely as follows.
Conside r the initial states

x1 (0) =

.X1 (0)]

1.5.2 that
where x1 (0) and, x2 (0) ~re any real nonzero scalars. Then recalling from Section
the state
that
(1.36)
from
see
we
l,
the transitio n matrix eAt is diagona l when A is diagona
correspo nding to each of these initial states is

x(t)=

[.xJ(OJe,x'r]

x(t) =

[.x2 (~e,\2 t J

for x(O)

= x1(0)

for x(O)

(1.55)
x 2 (0)

ates
The foregoin g suggests that the trajecto ry for any initial state in these coordin

x(O)

[~I (0)]
x2(0)

can be written as

( 1.56)
where

/ :k

1, 2

are columns from the 2 x 2 identity matrix, viz.,

ry
More generally, when A is a real, diagona l, n x n matrix, the state trajecto

(1.57)
results when the initial state

State Trajectories Revisited

23

has components which satisfy

.X;(O)

=0
#0

fori# k

fori= k

where ik is the k 1h column from the n x n identity matrix.

The foregoing result implies that the zero-input state response for any initial state can
be written as
n

x(t)

= .L:(xk(O)/'k 1) /
k=I

when
n

x(O)

L .xk(O)ik
k=I

In summary, the state trajectory is the vector sum of state trajectories along coordinate
axes where each coordinate axis trajectory depends on one of the system modes in the set
). I
.
of _system modes, {e k : k = 1, 2 , n}.
1 Now in order to generalize the foregoing result to the case where A is not diagonal, we
suppose, in the next section, that the foregoing diagonal case resulted from a coordinate
transformation from the original coordinates in which A is given.

1.7.2

Straight line state trajectories: real eigenvalues

Recalling that V, equation (1.44), is the coordinate transformation needed to diagonalizeA and taking x(O) as in (1.57) we have
(1.58)
where Avk = >.kvk. Then, using the series expansion for the matrix exponential, (1.22), we
see that when x(O) is given by equation (1.58) we have

(1.59)
Now with >.k real we see from equation (1.59) that the point representing the state
moves along the eigenvector, if, towards the origin when >.k < 0 and away from the origin
when >.k > 0. The case where >.k is complex is taken up in the next subsection.
More generally, assuming A has a complete set of eigenvectors, we can write any initial
state as
n

x(O)

= 2:: "(;Vi
i=I

(1.60)

Introduction to State Space

24

where 'Y

v-t x(O) and

Then the state trajectory can be written as

n

x(t) =

2,:,)'Y;e.\')v;

(1.61)

i=l

Now recall, from Section 1.2, that a system is stable if

lim x(t)

for all x(O)

( 1.62)

t--+'XJ

Therefore assuming A has real eigenvalues, we see from equation (1.59) that the system is
stable if and only if
i = 1, 2, , n

It should be emphasized that a system is stable if and only if equation (1.62) is satisfied.
Therefore if we need to restrict the initial condition in order to ensure that the zero-input
state trajectory goes to the origin, the system is not stable. For example, referring to the
expansion (1.60), any initial state, x(O), which is restricted so that its expansion satisfies

ri = o

when\ 2': 0

has a state trajectory which goes to the origin with time even though A has some nonnegative eigenvalues.

1.7.3

Straight line trajectories : complex eigenvalue s

Since the differential equations governing the input-output behavior of the physical
processes we are interested in controlling have real coefficients, we can always choose to
work in coordinates so that A, Care real matrices. Then since the eigenvalues, if complex,
occur in conjugate pairs we see that the corresponding eigenvectors are also conjugates of
each other, i.e., if(>., v) is a complex, eigenvalue-eigenvector pair, then(>.*, v*T) is also an
eigenvalue-eigenvector pair of A.
Now if the initial state is any scalar multiple of the real or imaginary part of v, the
resulting state trajectory lies in a two dimensional subspace of state space composed from
the real and imaginary parts of v. More specifically, suppose
( 1.63)

where 'Y is any real scalar and

v

= Re[v] + jlm[v]

25

State Trajectories Revisited

Then frow (1.59) we obtain

x(t)

O:re(t)Re[v]

+ O:;m(t)Im[v]

( 1.64)

where

a,e(t) =

1'eRe[.X]t cos(Im[A.]t)

O:;m(t) =

-1'eRe[.X]t

sin(Im[A.]t)

The foregoing generalizes the observation made at the beginning of the chapter that,
for second order systems, spiral shaped trajectories result when the roots of the
characteristic equation consist of a complex conjugate pair. In the present case where
the system order n > 2, the spiral shaped trajectory lies in a two-dimensional subspace of
the n-dimensional state space.
Notice that the initial stat~ was chosen to ensure that the state x( t) is real. However if
we choose the initial state as

/'

with the real scalars o:; satisfying

. then x(O) would be complex and the resulting trajectory would lie in a 2-dimensional
subspace of a complex state space.

1.1.4

Null output zero-input response

Having discussed straight line state trajectories, we return to the problem $.tated at the
beginning ofthis section concerning the possibility of having initial states which produce
null outputs.
Suppose the output matrix happens to satisfy
(1.65)
for some eigenvector, v\ of A. Then it follows from (1.59) that if x(O) = 1'kvk then

and we see the output is null for all time when the initial state lies along vk.
This effect of the existence of non-null initial states which do not affect the output, is
related to a property of the state model's A and C matrices which is referred to as the
system's observability (Chapter 3). More is said about this matter in Section 3.5.2.
In order to obtain some appreciation of the importance of this effect consider a state
model with A having all its eigenvalues, except A.k, in the open left-half plane. Then this
state model is internally unstable since x( 0) = vk produces a trajectory which moves away

Introduction to State Space

26

equa tion (1.65)

from the origin. How ever if C satisfies
outp ut a;Jd in this case we have

this traje ctor y has no effect on the

for all x(O)

lim y(t) = 0

t-v:yc

and the system is exte rnal ly stable.

beha vior
mod el is inter nally unst able its outp ut
This dem onst rate s that while the state
ical process,
ly impossible to exactly mod el a phys
is stable. How ever , since it is practical
r only and is
resp onse to initial state s exists on pape
the fore goin g stab ility of the outp ut in
on, we say
not robu stly outp ut stable. For this reas
referred to by saying that the system is
open left-half
A matr ix has all its eigenvalues in the
that a system is stab le if and only if its
plane.
which
uctiv e to cons ider cond ition s on C
Before we leave this section, it is instr
t and leftfied. Recall, from Section 1.6.3, that righ
guar ante e that equa tion (1.65) is satis
eigenvalues are orth ogon al,
eigenvectors corr espo ndin g to different

of eigenvectors so that
The refo re supp ose A has a com plete set
of the left-eigenvectors of A,

we can expa nd C in terms

( 1.66)

L...a ;wiT
= "'"'
I

when ak = 0
fied whe n Cis inde pend ent of wk, i.e.,
The n we see that equa tion (1.65) is satis
pter 3 to
acte riza tion of C will be used in Cha
in equa tion (1.66). This stru ctur al char
h relat e to their observability.
develop prop ertie s of state models whic

1.8

lete Response
State Space Models for the Comp

section
el a syst em's zero -inp ut response. In this
So far we have used the state space to mod
zeroem's
syst
a
of
g
conn ectio n with the mod elin
we take up the use of the state space in
state response.

1.8.1

Sec ond ord er pro ces s rev isit ed

age source
Sect ion 1.2, supp ose we conn ect a volt
Retu rnin g to the electric circuit used in
(cap acito r
t)
al equa tion gove rnin g the outp ut, y(
as show in Figu re 1.4. The n the differenti
voltage), beco mes
d2y(t) +a dy(t ) +a2y(t)
1 dt
dt2

and
where u(t) is the inpu t (sou rce voltage)

alla 2

= b2u(t)

are as in (1.1) with b2

( 1.67)

= a2 .

State Space Models for the Complete Response

27

i(t)

u(t)

Vollagc
Souroc

y(t)

Figure 1.4 Electric circuit with voltage input

Suppose, as in Section 1.3, that we choose the components of the state as

[ xx 1 (t)]
2 (t)

[y(t)]

(1.68)

y(t)

so that
(1.69)

"Then we see from (1.67) that

(1.70)

and from (1.68-1. 70) that the state differential equation and output equation are given as

x(t) = Ax(t) + Bu(t)

(1.71)

y(t)

(1.72)

Cx(t)

where

A= [

-a,
1

B=

[~]

C= [ 0

with B being referred to as the input matrix. This state model is complete in as much as it
can be used to obtain the output caused by any specified combination of initial state, x(O),
and input, u(t). We will see that the matrix product of the initial state and transition
matrix, which gives the zero-input response, is replaced in the calculation of the zero-state
response by an integral, referred to as the convolution integral, of the i:nput and
transition matrix. Before showing this, consider the following modification of the
foregoing state model.

Introduction to State Space

28

Suppose we rewrite (1.67) as

where

and proceed in the manner used to get equation (1.71, 1.72). This gives the alternative
state model

+ Bcu(t)

:<:(t) = Acxc(t)
y(t)

(1.73)

Ccxc(t)

where Ac = A and
Be= : 2 B =

[~]

Cc=b 2 C= [0

b2 ]

This state model, equation (1.73), is an example of a controller form state model.
Controller form state models are characterized by having Be as the first column of the
identity matrix and Ac in companion form, (1.19). The controller form state model is used
in the next chapter to provide insight into the dynamics of closed loop systems employing
state feedback.

1.8.2

Some essential features of state models

First notice that when we change coordinates by setting x(t) = Vx(t) for some constant
invertible matrix V, the resulting state model in the new coordinates is given by
.X (t)

Ax(t) + .Bu(t)

y(t) = Cx(t)
where the parameter s for the original and transformed state models are related as
(A, B, C)

where

v
f-+

- - (A, B, C)

State Space Models for the Complete Response

29

Second, notice that there are physical processes which transmit the effect of the input
directly to the output. In this situation the output equation for the state model has an
additional term, Du(t), i.e., the output equation is

y(t) = C:dt) + Du(t)

Notice that, unlike the other state model parameters (A, B, C), the D parameter is
unaltered by a state coordinate transformation.
Third. some dynamic processes have more than one scalar input and or more than one
scalar output. For instance a dynamic process may have m scalar inputs.
{ui(t): i E [I.m]} and p scalar outputs. {r,(t): i E [l,p1}. In this situation the system
input, u(t), and system output, y(t), are column vectors of size m andp respectively

z?(r)

[u 1 (t)

u2 (t)

rT (r) =

[Y! (t)

Y2(t)

lim ( t)

Further, in this case, the state model has an input matrix B with m columns and an output
ma!rix C with p rows. More specifically, the general form for state models used here is
given as

x(t)

Ax(t)

+ Bu(t) }

y(t)

Cx(t)

+ Du(t)

(1.74)

where x(t). u(t) and y(t) are column vectors of time functions of length n,m and p
respectively and the state model parameters, (A, B, C, D), are matrices of constants of size
n x n, n x m, p x n, and p x m respectively.
Finally, systems having p > 1 and m > 1 are referred to as multiple-input-multipleoutput (MIMO) systems and systems having p = 1 and m = 1 are referred to as singleinput-single-output (SISO) systems.

1.8.3

Zero .. state response

Recall from Section 1.3 that the zero-input response, Y:,(t), depends on the transition
matrix and initial state through multiplication as Yzi(t) = CeAr x(O). In this section we will
show that the zero-state response, Yzs(t), depends on the transition matrix and the input
through integration as

where the integral is known as the convolution integral.

We begin the development of the foregoing relation by assuming that the initial state
is null, x(O) = o, and that the state model parameters. (A, B, C, D) are known. Then
taking Laplace transforms throughout the state differential and output equations, ( 1. 74 ),

lntf'oduction to State Space

gives
sX(s)

AX(s)

+ BU(s)

( 1.7 5)

Yzs(s)

CX(s)

+ DU(s)

(1.76)

Next solving (1.75) for X(s) yields

X(s) = (sf- A)- 1 BU(s)
and substituting this expression for X(s) in (1.76) yields
Yzs(s) = G(s) U(s)

( 1. 77)

G(s) = C(sf- A)- 1B + D

(1.78)

where

Now, recalling from Section 1.4.3 that (sf - A) - 1 is a matrix of strictly proper rational
functions, we see that Gsp(s) is strictly proper where

so that

=0

lim Gsp(s)
S--->00

Thus G(s), (1.78) is given by

G(s)

= Gsp(s) + D

( 1.79)

and
lim G(s) = D
S--->00

Notice that two state models which are related by a coordinate transforma tion have
the same transfer function. We can see this by calculating the transfer function using the
state model in the transformed coordinates, (A, ii, C, D)
-

1-

G(s)=C( sl-A)- B+D

and substituting A=

v- 1AV, ii = v- 1B, C = CV, I>= D to obtain

G(s)

= CV[V- 1 (sf- A) vr 1 v- 1B + D
= C(sf- A)- 1B + D = G(s)

State Space Models for the Complete Response

31

Next recalling the following general property of Laplace transforms

J
00

Yzs(s) = F,(s)Fz(S)=}-Yzs(t) =

J,(t- T)fi(T)dT

we see that setting

F1 (s) = C(sl- A)- 1

so that

fi (t) = CeAt

f 2 (t) = Bu(t)

gives the inverse transform of (1. 77) as

Yzs(t) =

00

eA(t-r) Bu(T)dT

+ Du(t)

Notice that when Dis null and the input is an impulse, u(t)

(1.80)

= 8(t), (1.80) gives

( 1.81)

In addition recalling from section 1.4.2 that

we see that Yzs(t), (1.81), has Laplace transform equal to Gsp(s) as we expect since the
Laplace transform of the impulse response equals the transfer function. In addition we
see from (1.81) that the zero-input response equals the impulse response when the initial
state is x(O) = B.
Now we need to define eAt as a null matrix when t < 0 in (1.80, 1.81). This is done to
match the mathematics to the physical fact that the future input, u(T) : T > t, does not
affect the output at time t, i.e., y 2 s(t) is independent of future values of the input. This
property of the transition matrix, i.e., <f>(t) = 0 for all t < 0, is referred to as the causality
constraint and applies when we use the transition matrix in connection with the zero-state
response. Thus the causality constraint forces the integrand in (1.80) to be null forT> t
and enables (1.80) to be rewritten as

Notice that when the transition matrix is used in connection with the zero-input
response we can interpret( -t) fort> 0 as the matrix needed to determine the initial
state from the state at timet, i.e., x(O) = >( -t)x(t), which implies that( -t) is the inverse
of <f>(t).
At this point we can see, by recalling the principle of superposition, that when a system
is subjected to both a non-null initial state, x{O), and a non-null input, u(t), we can write

32

Introduction to State Space

the output as
y(t)

= Yz;(t)

+ Yz (t)

(1.82)

where
Yz;(t)

= CeA 1x(O)

Yzs(t)

=Clot

eA(t-T) Bu(T)dT

+ Du(t)

Notice that:
(i) y 2 ;(t), the zero-input response, is caused solely by to x(O)
(ii) Yzs(t), the zero-state response, is caused solely by to u(t).
In this section we have developed several essential relations involving the state model
of a dynamic process. These important relations provide the complete response of a
system to any specified input and any specified initial state.

1.9

Diagonal Form State Model

In subsequent chapters we will encounter a number of fundamental properties of state

models in connection with their use in the design of feedback control systems. A simple
way of beginning to appreciate these properties is to consider state models in coordinates
where the system matrix A is diagonal. This type of state model is referred to as a diagonal
or normal form state model. We encountered this model earlier in Section 1.5.2 in
connection with the determination of the matrix exponential, and in Section 1. 7.I in
connection with straight line state trajectories.

1.9.1

Structure

Suppose the state model for a given nth order SISO system has an A matrix which is
diagonal. Then the state vector differential equation decomposes into n scalar equations

x;( t)

.\x;( t)

+ b;u( t)

i = I, 2, , n

(1.83)

with the output being a scalar multiple of these components

11

y(t)

L c;x;(t) + Du(t)

( 1.84)

i=l

We can visualize these equations as a block diagram of the sort introduced in Section 1.3
to model the zero-input response.
Alternatively, when we view this state model in the frequency domain so as to obtain
the plant's transfer function, we see that since the state model's system matrix, A, is

Diagonal Form State Model

Figure 1.5

33

Block diagram representation for diagonal form state model

di~onal, we have (sf- A)- 1 diagonal,

0
0
(1.85)

This fact simplifies the dependency of the transfer function on the elements in the Band C
matrices since we have
G(s) = C(sl- A)- 1 B + D

Cb
=~
L-'-'-+D
i=1

s- .A;

(1.86)

where

1.9.2

Properties

Notice from (1.83) that the evolution of the components of the state, xk(t), are dccoupled
from each other, i.e., xk(t) is independent of {x;(t): i E [O,n], i # k}. This fact plays an
important role in the following discussion.

34

Introduction to State Space

We begin by noting that this decoupling of components of the state from each other
implies that any component of the state, say xk(t), depends on the input through the kth
component of the input matrix, bkl only. Therefore, if one of the components of B is zero,
the corresponding component of the state is not affected by the input. Alternatively,
notice from (1.86) that when bk = 0, the transfer function, G(s), does not have a pole at ).,k
and the zero-state response does not involve the mode e>.k 1.
These phenomena are related to a property of state models known as controllability.
Roughly speaking, a system is said to be controllable if its state can be forced, by the
input, to equal any point in state space at some finite time . Therefore if bk = 0, the
evolution of the kth component of the state, xk(t), is independent of the input, u(t), and
the state model is not controllable since there is no way of manipulating the input to make
xk(t1 ) equal some specified value after the passage of a finite amount of time, t1 . We will
see in the next chapter that controllability is a necessary and sufficient condition for being
able to arbitrarily assign the closed loop system matrix eigenvalues by using state
feedback. Unlike the diagonal form state model, the controllability of a state model in
other coordinates, depends on both the A and B matrices.
The foregoing observations were obtained by viewing the system from the input side.
We can obtain dual observations by viewing the system from the output side involving the
output instead of the input. We develop these observations as follows.
First, since equation (1.83) decouples the evolution the components of the state from
each other, we see that the output, y(t), (1.84), depends on the kth component of the state,
xk(t), through ck only. Therefore, if one of the components of Cis zero, the corresponding
component of the state does not affect the output. Alternatively, notice from (1.86) that
when ck = 0, the transfer function, G(s), does not have a pole at )..k and the zero-state
response does not involve the mode e>.kt.
The foregoing phenomena are related to the property of state models known as
observability. Recall that we encountered this property in the previous section in
connection with the system's zero-input response. In the present instance we see that
the observability of a system also relates to the system's zero-state response. Notice that
unlike the diagonal form state model, the observability of a state model in other
coordinates, depends on both the A and C matrices.
In summary, we see from equation (1.86) that if either bk or ck is zero for some integer
k E [0, n], then the system transfer function G(s) does not have a pole at Ak and the zerostate response does not involve the mode e>.kt. Moreover, under these conditions, if we
transformed coordinates and recomputed the transfer function from the parameters of
the state model in the new coordinates, i.e.,

we would find that the both the numerator polynomial, Cadj(s/- A)B, and denominator
polynomial, det(s/- A) (or det(s/- A) since the characteristic polynomial is invariant
to coordinate transformation), would have a zero at )..k Since s- Ak would then be a
common factor of Cadj(s/- A)B and det(s/- A), it would cancel from the ratio of these
polynomials and be absent from the transfer function. Thus the eigenvalue of A at )..k is
not a pole of G(s) and has no effect on the system's input-output.
Given a transfer function, any state model that is both controllable and observable and
has the same input-output behavior as the transfer function, is said to be a minimal

Diagonal Form State Model

35

realization of the transfer function. Notice that this terminology reflects the fact that lack
of either controllability or observability increases the dimension of the state model over
the minimal dimension needed to match the zero-state behavior of the state model with
that of the transfer function. As we will see, this fact play's an important role in control
theory.

1.9.3

Obtaining the diagonal form state model

The relation between the form for the transfer function, equation (1.86) and the diagonal
form state model suggests a way of determining a diagonal form state model for a given
SISO transfer function, G(s), having simple poles. This is done as follows.
We begin by using one cycle of division when the given transfer function is proper, to
separate out the strictly proper part of the transfer function. Thus suppose we are given a
transfer function which is proper
n

L
Gp(s)

f3n-iS 1

i=On-1
sn + L an-isi
i=O

Then dividing the denominator into the numerator enables Gp(s) to be written as the sum
of a strictly proper transfer function plus a constant, i.e.,

where

i = 1,2, n .,..

Then assuming the strictly proper part, Gsp(s), can be expanded in partial fractions as
n

Gsp(s) =

L s_
1

i=l

'

( 1.87)

/\l

we can obtain the diagonal form state model by factoring the k;s as
k;

.l

; ~:
'

'

= C;b;

and referring to equation (1.86).

Notice it is only possible to obtain a diagonal form state model if the transfer function
can be expanded as equation (1.87), i.e., if the system can be decomposed into first order
systems. When Gsp(s) has simple poles we can always expand Gsp(s), as equation (1.87).
Notice that if Gsp(s) has gjmple poles some of which are complex, the diagonal form

Introduction to State Space

36

state model will be complex. In some situations we may want the benefits of the diagonal
form state model, i.e., decoupled state differential equations, but require the state model
to be real. In this case a reasonable compromis e is to combine complex conjugate pairs of
terms in equation (1.87) so that the resulting realization has an A matrix which is block
diagonal with each pair of complex conjugate poles correspond ing to a 2 x 2 real block
on the diagonal of A. More specifically we expand Gsp (s) as

G,p(s)

G,pr(s)

+ G,p,(s)

(1.88)

where
n,.

Gspr(s)

L s _'A
i=l

A; real
1

with the nc complex pairs of poles of Gsp(s) being the poles of Gspc(s), i.e.,

Now we can obtain a state model from (1.88) as

A=

[A1

A3

= [

A2]
A4

cl c2]

where (A 1 ,B 1 , C 1) constitute a diagonal form state model for Gspr(s), i.e., A 1 is diagonal
with diagonal elements equal to the real poles and the elements of B 1 and C 1 satisfy
cub; 1 = k; fori= 1, 2, n1 . Both off diagonal blocks of A, namely A 2 and A 3 are null.
The remaining blocks constitute a state model for Gspc(s) as
A4I

A42

0
0

0
=

B22
B2 =

A4 =

c2

B21

[ c21 c22

A4n,

B2n,

C2n,]

where {(A 4 ;,B2;,C2 ;): i= 1,2,nc} are 2 dimensional real parameter state models
which we obtain from { Gci : i = 1, 2, nc} with

Computer Calculation of the State and Output

37

One such real parameter state model is given as

A4;

= [ -~il

C2;

= [hi]

-a,.2l
0 J

1.89)

bi2)

Finally notice that we encountered a restricted version of the state model (1.89) in
Section 1.8.1. In the next Chapter we will make extensive use of this form which is referred
to as a controller form. At that time we will provide a means for its determination from a
given transfer function.

1.10

Computer Calculation of the State and Output

A fundamental problem encountered in using digital computers to analyses control

problems is created by the fact that the dynamic processes needing to be controlled
operate in continuous time whereas digital computers perform operations in discrete
time. In this section we indicate how this fundamental mismatch is overcome.
In order to be able to use a digital computer to calculate the system output y(t) from
(1.82) we need to employ a piecewise constant approximation of the input u( t) owr the
time interval of interest, [0, tN] This is done by partitioning the interval into N equal nonoverlaping contiguous subintervals each of time duration T. Then the piecewise constant
approximation, u( t), of the input is given by
u(t)

= u(t;.;)

(1.90)

where

tk = kT: k = 0, 1, 2, N.
The set {u(tk) : k = 0, 1, 2, N} is referred to as the discrete-time equivalent of the
continuous-tim e input, u(t), and the values, u(tk), of the discrete-time equivalent input
are referred to as the sampled values of the input. The continuous time "staircase like"
function u(t) generated by this approximation is referred to as the sample and hold
equivalent of the input. Figure 1.6 shows a plot of a typical input and its sample and hold
equivalent.
The output, y( t), at the sample times, tk : k = 0, I, 2, . , N, is computed by repeating
the calculation indicated in equation (1.82) for each subintenal with kT considered as the
start time for the computation of y([k + 1] T). This is carried out in a cyclical fashion with
the state that is computed at the end of a given subinterval being used as the initial state
for the computation of the state at the end of the next subinterval.
More specifically, the computations are started by specifying the initial state, x(O), and
the input. u( t), over the interval of interest, [0, t\;. The computer program then decides on
the sample spacing T. This is done so that the sample and hold version of the input is a
"close" approximation to the actual input in the time interval [0, tN] Then taking the

Introduction to State Space

38

u(t)

ii(t)

Figure 1.6

Typical input and its piecewise constant approximation

input as u(O) over [0, T] and using (1.82) we compute y(T) from
T

x(T)

=eAT x(O)

eA(T-T) Bu(O)dT

y(T)

Cx(T)

+ Du(T)

which can be rewritten as

x(T)

= Fx(O) + Gu(O)

y(T) = Cx(T)

( 1.91)

+ Du(T)

where

eA(T-T) EdT

Then to compute y(2T), we replace x(O) by x(T) and u(O) by u(T) in (1.91). This gives
x(2T)

= Fx(T) + Gu(T)

y(2T) = Cx(2T)

+ Du(2T)

This procedure continues with the output at the (k + l)th sample time being given as

+ I)T] = Fx(kT) + Gu(kT)

y[(k + I)T] = Cx[(k + l)T] + Du[(k + l)T]

x[(k

r-

Notes and References

-~-'

39

There are several ways in which the computation ofF and G can be carried out. A
simple way of doing this is to truncate the series expansion for the matrix exponential,
(1.22), with the upper limit, n0 , of the sum being chosen large enough to achieve some
specified accuracy. Thus F, Gin (1.91) can be expressed as

G=

(lo

no A;

L~da B+Eo
i=O

l.

where the error matrices Ep and E 0 are made negligible by choosing n0 "large enough" so
that we can take F and G as

Notice that the choice ofthe number of subintervals needed to cover the main interval,
i.e., [O,tN], also plays a role in the accuracy of the computation, with an increase in
accuracy being achieved at the expense of more computation time i.e., the accuracy goes
up the smaller T and the larger N.

1.11

Notes and References

There have been a number of excellent textbooks expounding on state space analysis of
systems. Some of these books are [5], [4], [23], [6], [9], [15].
Throughout the eighties the book by Kailath [23] was used widely as a text for courses on
state space analysis of linear systems. Although Brockett's book [5], is quite old and was
. written for a more mathematically sophisticated reader, the first half of the book does
give the reader a nice exposure to the more general nature of the subject. The treatment of
state modeling in the brief book by Blackman [4], shares the view with the present book
that understanding basic ideas in the state space can be facilitated by thinking geometrically.
Concerning books that provide insight into matters pertaining to linear algebra,
Strange's book [39] gives an excellent introduction. The book by Golub and Van Loan
[17] provides a more advanced treatment which is more orientated towards computer
computation. In addition the book by Brogan [6], plus the appendix to Kailath's book
also provide useful information on linear algebra.

2
State Feedba ck and
Controllability

2.1

Introduction

An important and well known consequence of feedback is behavioral modification.

F~edback is present when the input to a system depends on that system's output. When
feedback is used to modify a system's behavior the resulting closed loop system is referred
to as a feedback control system. For the purposes of discussion we can separate the
control system into a plant and a controller. The plant consists of the physical process to
be controlled plus any transducers, actuators or sensors, needed to interface the physical
process output to the controller and the output of the controller to the input of the
physical process. While controllers were implemented in the past by a physical process
such as an electric filter, their implementation now is often done by a digital computer.
Figure 2.1 shows the general setup for a control system.
Ut(t)

Yt(t)

PLANT

u2(t)

Y:(t)

CONTROLLER

Figure 2.1

Block diagram of the feedback control scheme

external input
controlled input

desired output
measured output

where u 1 (t), u2 (t),y 1(t),y2 (t) have dimensions mh mz,PI ,p2 respectively.

State Feedba ck and Controllability

42

The general form of the state model for the plant is

,"\:(t) = Ax(t) + B 1 (t)u 1 (t) + B2 u2 (t) }
YI(t) = C1x(t) + D 11 u1(t) + D 12 u 2 (t)
Y2(t) = C2x(t) + D21u 1(t) + D 22 u2 (t)

(2, 1)

the
This chapter focuses on the effect of using a feedback signal compos ed from
compon ents of the plant's state as the controlled input,
(2.2)
input. This
where K has m 2 rows, one row for each scalar signal making up the controlled
nts
ingredie
two
of
one
is
and
k
feedbac
state
as
to
referred
is
importa nt kind of feedback
er.
controll
the
of
needed to complete the design
it is
The need for something in addition to state feedback can be appreciated when
state
use
to
are
we
if
re
Therefo
n.
unknow
usually
is
recognized that the plant state
ng the
feedback as a basis for the design of a controller, we must have a means for estimati
is
which
r,
estimato
state
a
as
to
referred
model,
state
plant state. This is done by a
state
plant
the
on
tion
informa
The
model.
state
plant's
designed from knowledge of the
signals is
containe d in measurements of all or some of the plant's input and output
measurethese
to
equal
input
r's
estimato
the
supplied to the state estimato r by setting
There
plant.
the
of
state
the
mimics
state
its
that
so
ments. The state estimato r is designed
about
made
ions
assumpt
the
on
ng
dependi
r
are several different types of state estimato
exact,
e.g.,
ments,
measure
their
and
signals
the relation between the actual plant
chapters
ent
subsequ
In
n.
unknow
partly
or
corrupte d by random measurement noise,
ance of the
we will study both the design of these estimators and the effect on the perform
feedback.
state
with
tion
conjunc
in
use
their
by
feedback control system that is caused
so that we
known
is
model
state
plant's
the
of
In this chapter we assume that the state
in using
goal
Our
(2.2).
equation
by
given
can use it to generate the state feedback signal
input
zero
and
state
zero
the
which
in
system
state feedback is to produce a closed loop
which
to
extent
The
use.
plant's
the
to
al
behavior is changed in some way which is benefici
a system's
the plant's controlled input can affect the plant state, is referred to as
es
determin
model
state
plant
the
of
ability
controllability. We will see that the controll
plant's
the
from
differ
can
system
loop
the extent to which the state model of the closed
model
state model when we use state feedback. The importa nt property of state
chapter
this
out
through
guises
t
differen
controllability will be encountered in several
and the rest of the text.

2.2

State Feedback

state model
Applying state feedback, equation (2.2), to the plant, equation (2.1) gives the
for the feedback system as
x(t)
y 1 (t)
Y2(t)

=
=
=

(A+ B 2 K)x(t) + B 1 (t)uJ (t) }

(C 1 +D 12 K)x(t) +D 11 u 1(t)
(C2 + D 22 K)x(t) + D21u1 (t)

Figure 2.2 gives a block diagram interpre tation of state feedback.

(2.3)

State Feedback

43

YI(t)

u1(t)

PLANT
u2(t)

x(t)

L__

Figure 2.2

CONTROLLER

1.-----

Block diagram representation of state feedback.

Notice that the main effect of using state feedback is the transformation of the system
matrix from A, when the plant stands alone, to A + B 2 K, when the plant is embedded in a
state feedback loop. Thus we see that the main purpose of state feedback is the assignment
of eigenvalues to the closed loop system matrix A + B 2K to achieve a set of specified
performance criteria for the behavior of the closed loop system from the external input
u1 (t) to the desired output y 1 (t). This fact leads immediately to the question: can we
choose K so that the eigenvalues of A+ B2K are assigned to any specified set of values?
The answer to this question is obviously of great importance to the use of state feedback
in the design of feedback control systems. The answer to this question, which is given in
the following theorem, involves the use of the left-eigenvectors of A. Since this property
applies in general to any state model (A, B, C, D) we will drop the subscript on B, i.e., B2
will be replaced by B.
Theorem 2.1 Whenever A has a left-eigenvector w;, i.e., wiT A= >.;wiT, such that

the corresponding eigenvalue of A, ,X;, is invariant to state feedback, i.e., >.; E >.(A + BK)
for all K.
Proof Suppose the condition in the theorem is satisfied. Then multiplying the closed
loop state feedback system matrix, A+ BK, on the left by wiT gives

wiT(A +BK) =wiT A +wiTBK

However since wiT A= >.;wiT and wiT BK = 121, the previous equation becomes

and we see that >.1 is an eigenvalue of the system matrix for the state model of the feedback
control system, since>.; E >.(A+ BK) for all K.

It turns out that when there are no left-eigenvectors of A which satisfy

B = 0, the
eigenvalues of A are all assignable to any specified values by state feedback. This is shown
later in Theorem 2.4. Notice that when we have the condition
B = 0 so that the
corresponding eigenvalue >.1 cannot be assigned by state feedback, we refer to >.1 as an

wT

wT

State Feedb ack and Controllability

44

ion wf B 1- 0 we refer to
uncontrollable eigenvalue. Conversely, when we have the condit
state feedback, as a conthe corresponding eigenvalue A; which can be assigned by
trollable eigenvalue.
llable (uncontrollable)
Definition 2.1 An eigenvalue >..; of A is said to be a contro
A;~ >.[A+ BK]
eigenvalue of the pair (A, B) if we can (cannot) find K so that
invari ant to changing
Uncon trollab le (controllable) eigenvalues for a pair (A, B) are
coordinates. This can be seen as follows.
have
Suppose A; is an uncontrollable eigenvalue of (A, B). Then we
(2.4)
(2.5)

n wiT and A, in equation

Now insert rr-l between wiT and Bin equati on (2.4) and betwee
by T. Then we obtain
(2.5). In additio n, postm ultiply both sides of equati on (2.5)

where

of the state model in the

Thus we see that A; remains an uncontrollable eigenvalue
transfo rmed coordinates.
have no eigenvalues
Reflection on the foregoing reveals that plant state models which
ty that state feedback
of A which are both uncontrollable and unstable have the proper
having this proper ty are
can be used to make the closed loop system stable. State models
is not stabilizable when
referred to as stabilizable state models. Conversely, a state model
impor tant proper ty since
it has uncon trollab le eigenvalues which are unstable. This is an
condit ion for a control
the stability of a contro l system is almost always a necessary
system to be of arty use.
if all its uncontrollable
Definition 2.2 A plant state model is said to be stabilizable
eigenvalues are stable.
feedback control
Recall that the poles of the transfer function for an SISO output
the present case of state
system are constr ained to lie on branches of the root locus. In
matrix A + BK are
system
loop
closed
the
of
feedback, the locatio n of the eigenvalues
nal flexibility
additio
This
llable.
contro
are
they
that
,
uncon straine d provided, of course
one adjustable param eter
of state feedback over output feedback arises from there being
ck, whereas there is only
in K for each compo nent of the state when we use state feedba
ck.
one adjustable param eter when we use consta nt output feedba

2.3

Eigenvalue Assignment

has the same eigenvalues

Recall, Section 1.6.4, that a state model's system matrix
. In addition, notice that
independent of the coordi nates used to express the state model

Eigenvalue Assignment

45

if the state, x(t), in the original coordinates i& related to the state, x(t), in the new
coordinates as

x(t) = Tx(t)
then the state feedback matrix is transformed as

K = KT since

u(t) = Kx(t) = KTx(t)

This fact provides us with the flexibility of being able to carry out the determination of the
state feedback matrix in coordinates which are the most suitable for this computation.
We saw in the previous section, that we can choose K to achieve a specified set of closed
loop eigenvalues, {JL; : i = 1, 2, , n }, provided that each uncontrollable eigenvalue of
the pair (A, B) equals one of the desired closed loop eigenvalues. Assuming this condition
is satisfied, the state feedback matrix, K, which is needed to make the set of closed loop
eigenvalues equal the specified set of eigenvalues can be obtained by equating the
characteristic polynomial for A+ BK, denoted a(s), to the characteristic polynomial
1(s) having roots equal to the specified set of eigenvalues, {JL; : i = 1, 2, n}, as
/

a(s)

')'(s)

(2.6)

where

()= II(
n

S -

JL;

)=

+ 'YI S n-I + 'Y2S n-2 + + /n

i=l

Notice that the coefficients of a(s) are functions of the elements of K. Thus we see that
equation (2.6) gives rise to n equations in the unknown elements of K, i.e.,
a;=/;

i = 1,2, n

In general, these equations in the elements of K are coupled in ways which depend on
the A and B matrices. Therefore it becomes difficult to set up a computer program to form
and solve these equations for all possible (A, B) pairs. However if the plant's state model
is in controller form these equations are uncoupled and therefore easy to solve.

2.3.1

Eigenvalue assignment via the controller form

A state model for an nth order SISO system is said to be in controller form when A = Ac is
a companion matrix and B = Be is the first column of the n x n identity matrix. More

46

State Feedback and Controllability

specifically the controller form state model has parameters

~a I

Ac =

~a2

~an-I

~an

0
0

0
B

c~

with D and C being whatever is needed to model the plant's behavior in these coordinates.
Use of the controller form state model greatly simplifies the calculation of the state
feedback matrix Kc needed to make Ac + BcKc have a specified set of eigenvalues. This
results from the fact that in these coordinates the state feedback system matrix,
Ac + BcKn is also in companion form
~a 1

+ kc1

~a2

+ kc2
0

Ac +BcKc =

0
0

~a3

+ k3

~an-I+ kn-1

~an+

kn

with the kc 1s being the elements of the state feedback matrix

Recall from Chapter 1 that since Ac + BcKc is a companion matrix, the coefficients of
its characteristi c polynomial appear, with negative signs, along its first row. Therefore the
coefficients of the characteristi c polynomial for the closed loop system o:(s), equation
(2.6), are related to the elements of the state feedback matrix K as

i = 1, 2, n

(2.7)

Then if we want the closed loop system matrix, Ac + BcKn to have eigenvalues at
{p,1 : i = 1, 2, , n }, the o: 1s are obtained from

and the elements {kci : i

from equation (2. 7) as

1, 2, , n} of the feedback matrix K are easily determined

1, 2, n

(2.8)

Before showing how to transform coordinates so that a given state model is

transformed to controller form, we consider the problem of determining a controller
form state model from a given transfer function.

Eigenvalue Assignment

2.3.2

47

Realizing the co-ntroller form

Having seen that the assignment of eigenvalues is easy to do when the system state model
is in controller form, we turn now to the problem of getting a controller form state model
for a system specified by a transfer function or differential equation.
Theorem 2.2 Suppose we are given the a;, b; parameters in the differential equation or
transfer function model of some plant, i.e.,
n-1

n-1

Y(n)(t) + I:>n-iY(i)(t )

Lbn-iU(i)(t )
i=O

or

(2.9)

where

Y(s) = Gsp(s)U(s)

= 0 for all i = 0, 1,2, : ,n- 1

Then the controller form state model for this plant has parameters (Ac, Be, Cc) given by

with y<O(o) = u(i)(O)

-a I

Ac

Cc

=[

-a2

-a3

-an-I

-an

0
Bc-

bl

b2

b3

bn ]

(2.10)

Proof Factor the transfer function as

Gsp(s) = N(s)D(s)
where
n-1

N(s) = Lbn_;Si
i=O

D(s) =

--n--1--

sn +

L: an-is;

i=O

Then we see that

Y(s)

N(s)Z(s)

(2.11)

Z(s)

D(s)U(s)

(2.12)

JB

State Feedback and Controllability

where Z(s) is the Laplace transform of the signal z(t), i.e.

z(t) =

.c- 1 [Z(s)]

Since (2.12) gives the zero state response, z(t), of a system having transfer function
D(s) and having u(t) as input, we see that (2.12) implies that
n-1

z(n)(t)

+ L.>n-;Z(i)(t)

u(t)

(2.13)

i=O

and setting
i

1, 2, , n

(2.14)

in (2.13) gives

where the matrices Ae, Be are as stated in the theorem.

Finally, since the initial conditions are all assumed zero, we see from (2.11) that
n-1

y(t) = l::)n-iz(il (t)

i=O

and using the relation between the componen ts of the state and the derivatives of z(t),
(2.14), we see that the foregoing expression for the output can be written as

where the matrix Ce is as stated in the theorem.

Notice that since Gsp(s) is strictly proper, De = 0. Alternatively when the given transfer
function, Gp(s), is proper, we can proceed as in Section 1.9.3 to obtain its controller form
(Ac, Be, Cc, De) with Ae, Be, Cc as in Theorem 2.2 and D = {30 where
n

2:: f3n-i Si
GP (s) =

0- - -''-.=-'--

n-1

sn

+ "'a
0

n-1

si

i=O

with
i

= 1, 2, n

In summary, Theorem 2.2 shows that the controller form state model is easily obtained
when we are given either the transfer function or the differential equation governing the
process behavior.
In some cases we may be given a state model for an SISO plant which is in a form other
than the controller form. In this situation it may be possible to simplify the determina tion
of the state feedback matrix K by:

Eigenvalu e Assignme nt

49

(i) determinin g the coordinate transforma tion matrix, Tc, to transform the given plant
state model to controller form
(ii) choosing Kc using (2.8) to assign the eigenvalues to the controller form state model
(iii) using the coordinate transforma tion matrix Te to obtain the require state feedback
matrix asK= KeT-; 1 .
This approach, while not of practical significance when applied as just described, does
enable the developme nt of a practical computati onal algorithm for determinin g K. This
algorithm which is referred to as Ackerman n's formula will be developed in Section 2.3.4.
At this point we need to develop both a method for determining Te and conditions on
the given state model which enables its transforma tion to controller form.

2.3.3

Control ler form state transformation

Suppose we want to find a coordinate transforma tion matrix, Te, which transforms a
given state model, (A, B, C), not in controller form to a state model, (A, ii, C), which is in
controller form. Then assuming, for the time being, that such a transform ation exists, we
can solve this problem by proceeding as follows.
Since, in general A, and A= T- 1AT have the same characteris tic polynomia l, we can
use the coefficients of the characteris tic polynomia l for the given A matrix, det[s/ - A]. to
construct the controller form system matrix A = Ae from Theorem 2.2. In addition, since
Be is the first column of the n x n identity matrix we have ii = Be without doing any
calculation. Therefore neither Ae nor Be requires the determina tion of Tc. However Tc is
needed to determine C = CTc = C,. In what follows we use Ac- Be and their relation to
. Te to develop an algorithm for determinin g Tc, one column at a time.
To begin with, notice that the first column of Tc can be obtained immediate ly from
inspection of the relation Tcii = B when T e is written in terms of its columns.

where ti is the irh column of Tc.

Having determined the first column of Tc as B, we can see how to determine other
columns of Tc by inspection of the relation TeA= ATe with A set equal to Ac and Te
expressed in terms of its columns,

-an-1

-an

A[t 1

t"

t"

:Ju

;,(ale

l"eeaoacle and Controllabili ty

Thus equating like positioned columns on either side of this equation gives
-a 1 t 1 +t 2 =At 1
-a2 t 1 + t 3 = At 2
-ant 1 +

(2.15)

t~ = Atn-!

(2.16)
and we see that the remaining columns of Tc can be generated successively from (2.15) by
starting with t 1 = B and proceeding as
t 2 =a 1t 1 +At 1
t 3 = a2 t 1 + At 2

(2.17)
tn

= an_ 1t 1 + Atn-l

To recap, we can generate the columns of Tc by

(i) obtaining the coefficients of the characteristic polynomial for the given A
(ii) setting the first column of Tc equal to the given B.
(iii) using (2.17) to determine the remaining columns of Tc in succession.
Notice that so far we have no assurance that the matrix Tc which results from using
(2.17) will be invertible. However T c must be invertible if it is to be a coordinate
transformati on matrix. We show now that only those state models which have the
property of controllability will produce a nonsingular Tc matrix as a result of applying
the foregoing algorithm. Before doing this notice that we did not need (2.16) to get (2.17).
This extra equation will be used in Section 2.3 .5 to derive the Cayley-Ham ilton theorem.

2.3.4

Condition for controlle r form equivale nce

In order to determine conditions on the given state model which guarantees that it can be
transformed to a controller form, i.e., to insure that the matrix Tc which is generated by
(2.17) is invertible, we expand (2.17) by successive substitution to obtain

(2.18)

Eigenvalue Assignment

51

Then writing these equations in matrix form gives

(2.19)

where

fl= [B

AB

A 2B

...

An-Is'

R=

al

az

an-I

a!

an-2
an-3

Now since the product of two square matrices is invertible if and only if each matrix in
the product is im ertible. Tc is imertible only if both \2 and R are invertible. However
since R is an upper-triangula r matrix having nonzero elements along its diagonal, R is
invertible for any given state model. Therefore T, is invertible if and only iffl is invertible.
The matrix n, which depends on the interaction between the given state model's A and
B matrices, is referred to as the given state model's controllability matrix. For systems
having only one input the controllability matrix is square and the state model is said to be
controllable if its controllability matrix is invertible. However when the input consists of
m scalar inputs, n. has nm columns and only n rows. In this case the state model is said to
be controllable if its controllability matrix is full rank, i.e., all its rows are independent or
equivalently, n of its nm columns are independent.
More generally. the property of controllability of a given state model is preserved
under coordinate transformation. We can see this by noting that the controllability
matrix n of the state model in the transformed coordinates is related to the controllability
matrix n of the state model in the original coordinates as

n = [JJ

.4.e ...

= [T- 1B

;p-Iel

T- 1ATT- 1B

T- 1An-IB]

r- 1n

(2.20)

Therefore since D has fewer rows than columns we have

for some a -1 0 only iffl has dependent rows, since ,BT =

implies that

aT T- 1

= 0T only if a

= 0.

This

rank[n] = rank[T- 1D] = rank[D]

where the rank of a matrix equals the maximum number of its columns which are
independen1. However since, for any matrix, the number of independent columns equals
the number of independent rows. we see that, in general
rank[D] .::::; n

52

State Feedback and Controllability

since n is ann x nm matrix for any n dimension state model having m scalar inputs. This
fact will be used later in the fourth section of this chapter.
When the state model is single input and controllable, m = 1 and n, 0 in (2.20) are
square and invertible. Therefore we can obtain the coordinate transformation matrix, T,
from (2.20) as

nrz-l

(2.21)

Then comparing this expression with (2.19) we see that rz-l = rl;:- 1= R. Moreover the
controllability matrix for a state model in controller form is readily calculated as

A~- I Be]

nc = [Be AcBc
-a I

2
al-

a2

-a I

*
*
*

(2.22)

which, being upper-triangular with fixed nonzero entries along the diagonal, is invertible.
Thus all controller form state models are controllable.
To recap, a given single input state model can be transformed through an appropriate
coordinate transformation to a controller form state model if and only if the given system
is controllable. Moreover we have just seen that an uncontrollable state model cannot be
made controllable by a coordinate transformation and that any controller form state
model is controllable. In addition, we will see that there are several other aspects of system
controllability which play important roles in control theory. A more basic view of system
controllability is developed in the fourth section of this chapter.

2.3.5

Ackermann's formula

As mentioned at the end of Section 2.3.1, we could calculate K to assign the eigenvalues of
+ BK, for a given single input controllable system by transforming the given system to
controller form , finding Kc, and transforming Kc back into the original coordinates
through the relation K = KcT;:- 1 We will see now that we can use this idea to develop a
computational algorithm for obtaining K which is known as Ackermann's formula. In
order to do this we will need the following result.
Theorem 2.3 (Cayley-Hamilton) Any square matrix A satisfies its own characteristic
equation.
Before proving this theorem, the following explanation may be needed. Recall that the
characteristic equation for A,
A

(2.23)
is a scalar equation which is satisfied by the eigenvalues of A.

53

Eigenvalue Assignment

Therefore Theorem 2.3 tells us that if we replace the scalars in (2.23) by the matrix A,
we obtain the result

We can readily show the Cayley-Hamilton theorem as follows.

Proof Premultiplying (2.18) by A and using (2.16) yields
(An+ a 1 An-I

+ a2An- 2 + anl)B = 0

(2.24)

Now (2.24) holds for any B. Therefore letting B = I in the foregoing equation yields
An+ a 1An-I

+ a2An- 2 +ani= 0

Theorem 2.4 (Ackermann' Formula) Given a single input controllable state model
(A, B, C, D) and the desired characteristic polynomial, a(s), for the closed loop system

a(s)

det[si- (A+ BK)]

then the required feedback matrix K is given by

K

= -qT[a(A)]

where qT is the last row of the inverse of the controllability matrix, 0, for the given pair
(A, B).
Proof For convenience, and without loss of generality, suppose the given plant is of
dimension n = 3. Then the plant state model's system matrix A has characteristic
polynomial

a(s) = det[si- A] = s 3 + a 1s 2 + a2 s + a3
Let the desired characteristic polynomial, a(s), for the state feedback system matrix be
denoted as

Then using the coordinate transformation matrix Tc developed in Section 2.3.3 we

transform the given state model for the plant to controller form

so that the state feedback system matrix in these coordinates is given by

State Feedback and Controllability

54

where

Therefore the desired characteristic polynomial can be written by inspection from the
companion form for Ac + BcKc as

a(s) = s3 + (al- kc1)i

+ (a2- kcl)s + (a3- kc3)

= a(s)- (kc1i + kc2s + kc3)

(2.25)

Next recalling that A and T- 1AT have the same characteristic polynomial for any
invertible T, we see that a(s) is the characteristic polynomial for both A and A c. Therefore
from the Cayley-Hamilton theorem (Theorem 2.3) we have

a(A) = a(Ac) = 0

(2.26)

Therefore we see from (2.25, 2.26) that

(2.27)
where I is the 3 x 3 identity matrix

However we can readily verify that

Therefore we see that pre-multiplying (2.27) by the third row of the identity matrix gives

-3T a (A c ).

= -

k ellIT

k c2l-2T

k c3l3T

(2.28)

Then inserting the identity matrix T;; 1 Tc between i 3 T and a(Ac) on the left side of (2.28)
and post-multiplying throughout by r;; 1 yields
(2.29)
Moreover using the facts that

(2.30)

Controllability

and that KcT-;: 1

55

= K we see that (2.29) can be rewritten as

i 3TT;: 1a(A)

= -K

(2.31)

Finally recalling (2.19) we see that

"' R-n-

~ [~ ~

}-

.(2.32)

where

and* are elements of R- 1 which are not needed. Thus using (2.32) in (2.31) gives

and the theorem is proved.

Notice that since only the last row of n- 1 i.e., qT, is required when we use this result to
compute the feedback matrix, K, we need only compute q which satisfies

with in being the last column of the n x n identity matrix. This avoids having to do the
more intensive computation of n- 1 .

2.4

Controllability

So far we have encountered the effects of system controllability twice, once in connection
with the condition for eigenvalue assignment by state feedback, (Theorem 2.1 ), and again
in connection with the existence of a coordinate transformation matrix which transforms
a given state model to controller form, (Section 2.3.4). In this section we encounter system
controllability in the context of the basic problem of the input's ability to manipulate the
state of a given state model. The following definition of a state model's controllability is
made with this problem in mind.
Definition 2.3 A state model (A, B, C, D), or pair (A, B), is said to be controllable iffor
every possible initial state, x(O), we can find at least one input u(t), t E [0, t1 ] and some
finite final time t1 < oo so that x(t1 ) = !ll, i.e., so that the state is annihilated by the input in
a finite time.
. It is important to notice in the foregoing definition, that t1 is required to be finite. This
IS done to prevent all stable systems from being considered to be controllable. More
specifically, since stable systems have the property that
lim x(t) = !ll

1--+oo

for any x(O) when u(t) is null

56

State Feedback and Controllability

all stable systems would be controllable if the final time, tr, in Definition 2.3 were allowed
to be infinite.

In order to develop the implications of the foregoing definition we need to recall, from
the first chapter, that the state x( tr) which results from having an initial state, x(O), and an
input, u( t), is given by

J
It

x(tr)

= eA 11 x(O)

eA(tr-T) Bu(T)dT

(2.33)

Then if the state is annihilated at time t1 , i.e., if x(tr)

u(t), must be chosen so that

0, we see from (2.33) that the input,

J
It

eA 11 x(O)

=-

eA(It-Tl Bu(T)dT

(2.34)

However from the series expansion for

eAt,

(Section 1.4.1) we see that

and
so that we can simplify (2.34) as

J
It

x(O)

=-

e-AT Bu(T)dT

(2.35)

This equation is the basic constraint which must be satisfied by any input which drives
the system state from x(O) to the origin in state space in finite time, t1 < oo. Notice that
the following three questions are immediately evident:
1. For each initial state, x(O), is there at least one input, u(t), which satisfies (2.35)?
2. If it is not possible to satisfy (2.35) by some input for each initial state, how should the
initial state be restricted to enable (2.35) to be satisfied by some input u(t)?
3. Iffor a specific initial state, x(O), it is possible to satisfy (2.35), what is the specification
of the input u(t) which does so?
Notice that when the answer to the first question is in the affirmative the system is
controllable in the sense of Definition 2.3.

2.4.1

Controllable subspace

We will show now that a criterion for a given state model or pair (A, B) to be controllable
in the sense of Definition 2.3 is that the controllability matrix be full rank, i.e.,
rank[S1] = n, when the state model is n dimensional where

Controllability

57

Notice that when rank[fl] = n, n has n independent columns so that we can always find
a constant vector, 1, for any initial state, x(O), such that
x(O) = !:11

(2.36)

However, if rank[OJ < n, then n has fewer than n independent columns, and it is not
possible to find 1 to satisfy equation (2.36) for some x(O). However, those x(O) for which
we can find 1 to satisfy equation (2.36) are said to lie in the range or image ofn, denoted as
x(O) E range[fl]
We we will show, later in this section, that those initial states, x(O), satisfying
x(O) E range[fl], are those initial states for which we can find u(t) to satisfy equation
(2.35). This fact answers question 2 and leads to the following definition.
Definition 2.4 Initial states, x(O), for which we can (cannot) find u(t) to satisfy
equation (2.35) are said to lie in the controllable subspace Sc (uncontrollable subspace
Sc) of the state space.
Thus we see from the answer just given to question 2 that
Sc = range[fl]

An important property of the controllable subspace is given in the following theorem.

Theorem 2.5 If x(O) Erange[n] then Ax(O) Erange[fl].
Proof Suppose x(O) E range[fl]. Then we can find a constant vector, 1, to satisfy
(2.37)
where the I;S are constant vectors having the same length as u( t). However, we know from
the Cayley-Hamilton theorem, (Theorem 2.3), that
n

An

= -

L a;An-i
i=l

where the a;s are coefficients of the characteristic polynomial

Therefore multiplying (2.37) by A and using the Cayley-Hamilton theorem gives

Ax(O) = AB11 +A 2 B12 + An-lBin-l

(t

-a;An-iB )In

=fl')'

(2.38)

where

ln-2 - lna2

In- I

lnaJ]

State Feedback and Controllability

58

and we see from (2.38) that

Ax(O) E range[n]

Notice that this theorem implies that if x(O) E range[n], then Akx(O) E range[n] for all
integers k. This fact enables us to show that rank[D] = n is a necessary and sufficient
condition for S, to be the entire state space so that there is an input satisfying (2.35) for all
initial states. Thus rank[D] = n is a necessary and sufficient condition for a state model to
be controllable in the sense of Definition 2.3. This fact is shown now in the following
theorem.
Theorem 2.6 If a state model (A, B, C, D) or pair (A, B) is controllable in the sense of
Definition 2.3 then
rank[D]

where

Proof Suppose (A, B) is controllable in the sense of Definition 2.3. Then substituting
the series expansion for eAT in (2.35) yields

= Bro + AB11 +A 2B12 +

=

n,o +Ann/! + A2nn/2 + ...

(2.39)

where
I

kT

T
rkn

T
rkn+l

= 0, 1' 2 ...

with
i=0,1,2

However from Theorem 2.5 we have

k

0, 1, ...

(2.40)

Therefore the columns on the right side of (2.39) span the entire state space if and only if
rank[D] = n. Otherwise when rank[D] < n there are initial states x(O) :. range[D] so that
(2.39) is impossible to satisfy by any u( t) and the state model is not controllable in the
sense of Definition 2.3. This implies that rank[O] = n is necessary for the state model to be

controllable in the sense of Definition 2.3.

59

Controllability

2.4.2

Input synthesis for state annihilation

The following theorem shows that rank[n] = n is sufficient for a state model (A, B, CD)
or pair (A, B) to be controllable in the sense of Definition 2.3.
Theorem 2.7 Ifrank[n] = n then we can satisfy (2.35) for any x(O) by choosing u(t) as

where

Proof Suppose W is invertible. Then substituting the expression for u(t) given in the

theorem into the right side of (2.35) gives

tf e-ArBu(T)dT= ~ }otf e-Ars [-BTe-A T]r dTW- 1x(O)

- }o
/

ww- 1x(O)

x(O)

which is the left side of (2.35). Thus we have shown that the state model is controllable in
the sense of Definition 2.3 when W is invertible.
In order to show that W is invertible, suppose rank[n] = n and W is not invertible.
Then we can find a =f. 0 such that
(2.41)
Now since the integrand in the definition of W is quadratic, i.e.,

where

we can only satisfy (2.41) if

for all
'

E [0, tf]

(2.42)

However using the power series expansion of the matrix exponential we see that

60

State Feedback and Controllability

where
kT _

{3

l"

(-T )kn

(kn)!

( -T )kn+l

(kn

+ 1) !

(-T)(k+l)n-1

([k + l]n- 1)!

Thus (2.42) is satisfied only if

(2.43)

However since rank[r!] = nand n is n x nm, we see that n has independent rows and only
a= 0 satisfies (2.43) and (2.41). This proves that W is invertible when rank[r!] = n.
In the next section we will see that when rank[r!] < n we can still use the foregoing
approach to provide a method for determining an input which annihilates an initial state
x(O) in a finite time provided x(O) Erange[r!]. This is achieved by using a controllable
decomposed form for the system's state model so that the controllable part of the system
is immediately evident.

2.5

Controllable Decomposed Form

So far we have considered two special types (canonical forms) of state model, the
controller form and the normal or diagonal form. In this section we introduce another
form for the state model of an n dimensional system. The purpose of this form of state
model is to simplify the specification of the controllable and uncontrollable subspaces,
denoted as Sc and S 2 respectively, by aligning them with subsets of the coordinate axis.
More specifically, when a state model is in controllable decomposed form the state is
given by

where x;(t) is of length n; and

x(t) ESc

if x 2 (t)

x(t) ESc

if x 1 (t) =0

= 0

Then any state, x(t) can be decomposed as

(2.44)

where xc(t) ESc and xc(t) E S 2 with ac, ac being scalars. Notice that Scl_Sc
A state model whose state space is orthogonally decomposed in the foregoing manner
is referred to as being in controllable decomposed form. The state model for a system to
be in this form is defined as follows.
Definition 2.5 A state model having m scalar inputs is said to be in controllable

61

Controllable Decompose d Form

decomposed form when its A and B matrices have the following block matrix forms

B=[~]

A=[~

(2.45)

where (All B!) is a controllable pair; All A 2 , A 4 and B 1 have dimensions: n 1 x n 1 , n 1 x n2 ,

n 2 x n 2 , n 1 x m respectively with n = n 1 + n2 being the dimension of the state space.

2.5.1

Input control of the controllable subspace

We begin the discussion of the effect of an input on the state when the state model is in
controllable decomposed form by noticing that the product of upper-triang ular matrices
is upper-triang ular, and that the coefficients of the series expansion of eAt involve powers
of A. Therefore the transition matrix, eAt, for a state model in controllable decomposed
form, (2.45), is upper-triang ular,
At_

f[A~ * ~
i=O

A~

il

where the block marked * is not of importance here.

Next recall, from the previous section, that if u(t) drives the state from x(O) to the
origin in finite time, t1 , then (2.35) is satisfied. Therefore since the eAt is upper-triang ular
and B has its last n2 rows null when the state model is in controllable decomposed form,
we see that (2.35) can be written as

x'(O)l
[~(0)

=-

lot (oo2:.)-l)i [A;Bl. ~Ii) U(T)dT

1

i=O

(2.46)

l.

or

x 2 (0)

Thus we see that (2.46) can not be satisfied by any input u(t) ifx2 (0) f= 0. Alternatively ,
1
since (A 1 , B!) is a controllable pair, we can always find u(t) to force x (tr) = 0 for any
1
x (0) and ~(0).
More specifically, when the state model is in controllable decomposed form, (2.45), we
see that

so that

State Feedbac k and Controllability

62

Therefore, proceeding in the&ame manner as was done to obtain (2.35) we see that if
x 1 (tr) is null then we must choose u( t) to satisfy the following equation
(2.47)
where

Then using the result given in Theorem 2.7 we see that one input which satisfies (2.47) is

where

To recap, we have just shown that we can always find an input to annihilate the
projection of the initial state onto the controllable subspace. More specifically since any
initial state can be written as (2.44) with t = 0, we can find u(t) so that xc(tf) = 0 for any
tf

< oo.

2.5.2

Plelation to the transfer function

Concerning the input-ou tput behavior, suppose the state model has A,B as specified by
(2.45) and C partition ed as

with ci beingp X ni, i = 1' 2 where pis the number of scalar outputs. Then from the block
upper-triangular form for A, (2.45), we see that
(sf- AJ)- 1A 2 (sl- A4)- 1 ]
(sf- A 4 )- 1

Therefore the system transfer function is given as

G(s)

C(sl- A)- 1B

(I
_- C IS

C 1adj[s/- AI]B1
A )-IEI _
I
det[s/- AI]

(2.48)

Notice that the order of the transfer function is no more than n 1 This reduction in the
order of the transfer function results from the eigenvalues of A4 being uncontrollable. We

tion in the numb er of eigenvalues of A that

will see in the next chap ter that a furth er reduc
funct ion may occu r as a resul t of the
go over as poles of the corre spon ding trans fer
order to bette r appre ciate the fact that the
inter actio n of the state mode l's (A, C) pair. In
the following.
eigenvalues of A 4 are unco ntrol lable , consi der

2.5. 3

Eig env alue s and eige nve ctor s of A

ectors of A when the state mode l is given

Supp ose we attem pt to deter mine the left-eigenv
to solve
in contr ollab le deco mpos ed form. Then we need
WT A= AWT

struc ture of A, (2.45), we see that the

for the pair {.:\, w}. How ever from the block
,
foregoing equa tion expa nds to two equa tions
(2.49)
(2.50)
where

ied when we let w1 =

with w; being oflen gth n;. Then (2.49) is satisf

o and (2.50) becomes

(2.51)

2
1
.:\ is an eigenvalue of A 4 w is the
Ther efore for this choice of w , we see that
= [ 0 wf} is the corre spon ding leftcorre spon ding left-eigenvector of A 4 , and wT
s of A4 are eigenvalues of A. Notic e that
eigenvector of A. Thus we see that eigenvalue
in this case we have

H'T B

= [0 w

n [:

=0

s of A 4 are unco ntrol lable .

so that eigenvalues of A which are also eigenvalue
values of A 1 . We can see this by notin g
Now the rema ining eigenvalues of A are eigen
of A 1 and w1 is the corre spon ding leftfrom (2.49) that if w 1 i- 0 then .:\ is an eigenvalue
mpos ed form the pair (AJ>B!) is contr oleigenvector of A 1 . Since in contr ollab le deco
s of A I are contr ollab le. More over since
lable, eigenvalues of A whic h are also eigenvalue
s of any pair (A, B) are disjo int \Ve have
the contr ollab le and unco ntrol lable eigenvalue
1
is invertible. Ther efore when w is a leftthat if.:\ E .:\[Ad then .:\ :. .A(A 4 ] and .AI- A 4
2
red w 2 to make wT = [ w 1T w T J a lefteigenvector of A I, we can deter mine the requi
eigenvector of A from (2.50) as
w2T = w!T A2(A I-

A4)-1

A is imme diate ly evide nt when one

The foreg oing prop erty of the eigenvalues of
state feedback when the state mode l is in
attem pts to use state feedback. Thus if we use

64

State Feedback and Controllability

uncontrollable decomposed fDrm, the closed loop system matrix is

where the state feedback matrix is partitioned as

(2.52)

with K 1 , K 2 being m x n 1 and m x n2 . Now we just saw that block upper-triangular

matrices have the property that their eigenvalues equal the union of the eigenvalues of
their diagonal blocks. Thus the eigenvalues of the state feedback system matrix satisfies

and A+ BK has a subset of eigenvalues, .A[A 4 ], which is clearly unaffected by state

feedback. However since (A 1B 1 ) is a controllable pair, all the eigenvalues of A 1 + B 1K 1
can be assigned by K1 . Notice also, from the definition of stabilizability, Definition 2.2,
that a system is stabilizable if, when its state model in controllable decomposed form, the
A 4 partition of A is stable.
Finally notice that Ackermann's formula, which we developed to assign the eigenvalues of single input controllable state models, can also be used to assign the controllable
eigenvalues of uncontrollable systems. This is done by transforming coordinates to put
the given state model in controllable decomposed form. Ackermann's formula is then
applied to the controllable pair (A 1 , B 1) to determine the K 1 partition of K, (2. 52), with K2
being chosen arbitrarily. Finally, we can obtain the feedback matrix K 0 in the original
coordinates from K 0 = KT- 1 where Tis the coordinate transformation matrix needed to
transform the given state model to controllable decomposed form. In the next section we
consider how this coordinate transformation matrix can be determined.

2.6

Transformation to Controllable Decomposed Form

In this section we will indicate the role played by the controllability matrix in constructing
a coordinate transformation matrix to put any state model in controllable decomposed
form.
We begin by noting that when a state model is in controllable decomposed form with
the (A, B) pair being given by (2.45), the last n2 rows of the controllability matrix consist
of nothing but zeros, i.e.,

[i i]
where

(2.53)

Transformation to Controllable Decomposed Form

65

Notice that the columns of 0 2 depend on the columns of 0 1 , i.e., we can always find a
constant matrix 8 to satisfy
Therefore recalling that the subspace spanned by the independent columns of n is the
controllable subspace, Sc, we see that
range[O]

=range([~])= Sc

This fact together with computationally robust methods for finding a basis for the range
of a matrix, e.g., singular value decomposition, (Chapter 7), provides a means for
constructing a coordinate transformati on matrix T which takes any state model to
controllable decomposed form. The following theorem provides some insight into why a
coordinate transformati on based on the controllability subspace is able to achieve this
result.
Theorem 2.8 A coordinate transformati on matrix T transforms a given state model
(A, B, C, D) to controllable decomposed form, (Definition 2.5) if
range[TJ] = range[O] = Sc
where

n is the controllability matrix for the given state model and

with T 1 , T2 being n x n 1 , n x n2 and with T invertible where rank(O) = n1.

Proof Let QT denote the inverse of T, i.e.,

[~n [
T1

QTT=l

Tz] = [

~ ~]

(2.54)

Then the transformed A, B matrices are given by

Recall that for the transformed state model to be in controllable decomposed form,
(Definition 2.5), we need to show that A3 and B2 are each null.
To show that B2 = 0, notice from (2.54) that QJ T 1 = 0 so that the columns of T1 are
orthogonal to the columns of Q2 However range(TJ) = Sc so that we have
Qfx=!ll

when

ESc

(2.55)

State Feedback and Controllability

66

Now since B appears as the first m <:;olumns of rl,

range[B] c range[n]

S,

(2.56)

and we see from (2.55) that the columns of Bare orthogonal to the columns of Q 2 so that

B2
To show that A3
since

QJB = 0

= 0, recall from Theorem 2.5 that if x ESc than Ax E Sc- Therefore

range[TJ]

Sc

we have
range[ATt] C Sc

(2.57)

Therefore it follows from (2.57, 2.55) that the columns of AT1 are orthogonal to the
columns of Q2 so that

Notice, in the foregoing proof, that (2.55-2.57) imply that we can find constant
matrices 8 1 , 8 2 such that
B= T181
AT1 = T182

and since QJ T1 = 0 we achieve Q 2 B = 0 and Q 2 AT1 = 0 as required for T to transform

the given system to controllable decomposed form.

2. 7

Notes and References

The Ackermann formula was one of the first attempts to develop an algorithm for
assigning the eigenvalues to a single input state model by state feedback. More recent
work on this problem has concentrated on developing algorithms which are least sensitive
to errors caused by the need to round off numbers during the execution of the algorithm
using finite precision arithmetic on a digital computer. Further information on this
problem can be obtained by consulting [43].
The Cayley-Hamilt on theorem plays an important role in calculating matrix functions and can be used to provide an efficient method for calculating eAt once the
eigenvalues of A are computed. Modern computer oriented algorithms rely on methods
which truncate the infinite series expansion for eAt after having put A in a certain
computationall y beneficial form known as real Schur form, [17].
The approach to the problem of transforming a state model to controllable decomposed form which was discussed in Section 2.6 forms the basis for the command CTRBF
in MATLAB.

3
State Estimation and
Observabi lity

3.1

Introduction

In the previous chapter we saw that state feedback could be used to modify basic aspects
of a plant's behavior. However, since the plant state is usually not available, we need a
means for obtaining an ongoing estimate of the present value of the state of the plant's
state model. In this chapter we will see that this can be done by using measurements of the
plant's input and output.
The problem of determining the state of a state model for a plant from knowledge of
the plant's input and output is referred to in general as the state estimation problem.
There are two classical types of state estimation: deterministic state estimation and
stochastic state estimation. Deterministic state estimation, which is the subject of this
chapter, assumes that the system input and output are known or measured exactly. The
goal of deterministic state estimation is the determination of an estimate of the state
having error which tends to decrease with time following the initiation of the..estimation
procedure. However, in stochastic state estimation the input and output signals are not
assumed to be known exactly because of the presence of additi\"e stochastic measurement
noise having known statistics. In this situation the state estimation error is always
present. The goal of stochastic state e-stimation is the determination of an estimate of the
state so that, in the steady state, the average or expected value of the state estimation error
is null while the expected value of the squared error is as small as possible. Stochastic state
estimation is discussed in Chapter 6.
Both deterministic and stochastic state estimation are further subdivided according to
when the input and output signals are measured relative to when the state estimate is
needed. If the plant's input and output are measured over the time interval [0, T] and if we
need the state estimate at time te we have
1. a prediction problem if te > T
2. a filtering problem if le = T
3. a smoothing problem if le < T

The state estimation problem which is of concern in connection with the implementation of state feedback is a filtering problem since we need an ongoing estimate of the

68

State Estimation and Observability

plant's state at the present time. Prediction and smoothing problems, which are not
discussed it1 this chapter, arise, for example, in hitting a moving target with a projectile by
aiming at the target's predicted position and in estimating the true value of data obtained
from experiments done in the past when the measured data are corrupted by noise.
The computer implementation of a solution to the filtering problem, either deterministic or stochastic, takes the form of a state model which is referred to in general as a
filter. The filter's input consists of both the plant's input and output and the filter's output
is the state of the filter, i.e., the filter is designed so that its state is an estimate of the plant
model's state. In the deterministic case the filter is referred to as an observer. In the
stochastic case the filter is referred to as a Kalman filter. The Kalman filter is taken up in
Chapter 6.

3.2

Filtering for Stable Systems

Suppose we know the parameters (A, B, C, D) of a plant state model as well as the plant's
input and output, { u( t), y( t) : t E [0, tel}. Then, from Chapter 1, we can express the plant
state at any time tc in terms of the initial state and input as
(3.1)

However since the initial plant state is usually unknown we are only able to calculate the
zero state response. We use this fact to form an estimate of the plant state at time tc as
(3.2)
where from (3.1) we see that the plant state, x(te)) and plant state estimate, x(te)) differ by
the plant state estimation error, x(te),
(3.3)

Now if the plant's state model is stable, we see from (3.3) that the state estimation
error, x(te), approaches the null vector with increasing estimation time, te,
lim x(te) = 0

for all x(O)

x(O)

lc---+00

In this situation the state estimate, x(tc), given by (3.2), can be a reasonably good
approximation to the plant model's state, x( te), provided te is large enough to ensure that
the effect of the initial state estimation error x(O) is negligible. In this case the estimate is
said to be an asymptotic estimate of the state since the state estimate approaches the state
being estimated asymptotically with time.
Alternatively, we can view the state estimate obtained in the foregoing approach as the
output, Ye(t), from a system called a state estimator having state model (Ae, Be, Ce) More

Observers

69

specifically we have

x (t) = Aex(t)

Ye(t)

+ Beu(t)

(3.4)
(3.5)

Ce_x(t)

Then since the state of the plant model is governed by

x=

Ax(t)

+ Bu(t)

(3.6)

we see by subtracting equation (3.4) from equation (3.6) that the state estimation error is
independent of the plant input and is governed by the differential equation
x (t)

= Ax(t)

(3.7)

where
x(t)

= x(t) - x(t)

Therefore the state estimation error is independent of the plant input and is given by
(3.8)
Now if the initial state of the plant were known we could take x(O) = x(O) as the initial
state for the estimator, equation (3.4). Then .X(O) = 0 and we see from equation (3.8) that
we would have the desirable result of exact state estimation, x(t) = x(t), for all time.
Usually we don't know the initial plant state. In this case we could set x(O) = 0 in the state
estimator, (3.4), so that x(O) = x(O) and provided A is stable, the state estimate would
approach the actual plant state asymptotically for any initial plant state.
However, unstable plants are encountered quite frequently, and must be stabilized by
feedback, e.g., the feedback control of satellite rocket launchers. In these cases the
estimate obtained from the state estimator, (3.4), diverges from the plant st~te for any
x(O) -#0 since eAt becomes unbounded with time and the state estimation error given by
equation (3.8) grows indefinitely. Notice that it is impossible to know a physical
parameter such as the initial plant state exactly. Therefore in practice it is not possible
to set x(t) = x(O) so as to obtain x(t) = 0 from (3.8). Thus whenever the plant is unstable,
we are unable to use (3.4) as an asymptotic state estimator.
Not knowing the initial state and needing to estimate the state to implement stabilizing
state feedback for unstable plants, forces us to seek another approach to state estimation.

3.3

Observers

Notice that the foregoing simple approach to the plant state estimation problem ignored
the additional information on the plant state which is present in the plant output, i.e.,

y(t)

Cx(t)

+ Du(t)

(3.9)

In this section we show how to use this information to obtain an asymptotic state
estimator for unstable plants.

70

State Estimation and Observability

Suppose, for the moment, that the plant output, y(t), is unknown. Then, assuming we
have an estimate of the plant state, x(t), we can use it to obtain an estimate of the output,
ji(t), by substituting x(t) for the plant state in (3.9), to get

y(t) = Cx(t)

+ Du(t)

(3.10)

However, we assume in this chapter that the actual plant output, y(t), 1s known.
Therefore we can define a plant output estimation error as

(3.11)

y(t) = y(t) - Y(t)

Now since we can form.Y(t) from u(t),y(t) and x(t) all of which are known, we can use
y(t) to indicate the accuracy of the plant state estimate since y(t) t 0 is an indication that
the state estimate differs from the actual state. More importantly, we can use y(t) to
correct future state estimates by letting jl( t) affect x (t) by subtracting Ly( t) to the right
side of (3.4) to obtain

x= Ax(t)

+ Bu(t)- Ly(t)

which using (3.10, 3.11) can be rewritten as

x= Fx(t)

+ G [u(t)
y(t)

(3.12)

(3.13)

where
G= [B+LD

F=A+LC

-L]

When F is stable, the system represented by (3.13) is referred to as an observer. In order

to determine the effectiveness of an observer in obtaining an asymptotic estimate of the
state, consider the differential equation for the state estimation error, x(t) = x(t)- x(t).
We obtain this differential equation by subtracting (3.13) from (3.6) to obtain

.X (t)

Fx(t)

(3.14)

where
F =A +LC

Then the state estimation error is given by

Notice that unlike the simple approach to state estimation presented in the previous
section, we now have the possibility of getting an asymptotic estimate of the state, even if
the plant is unstable. We do this by choosing L to make F stable. Questions regarding the
possibility of doing this are answered by examining certain properties of the righteigenvectors of A.

Observer Design

71

Theorem 3.1 Whenever A has a right-eigenvector v; satisfying

the corresponding eigenvalue A; of A is an eigenvalue of A + LC for all L.

Proof Suppose the condition of the theorem is satisfied. Then multiplying A
the right by v; gives

+ LC on

However, since Cvi = 0 and Avi = A;Vi we see that the foregoing equation becomes

Thus A; is a fixed eigenvalue of the matrix A + LC for all L

We will see that if vi is an eigenvector of A such that Cvi =f. 0 then we can choose L so
that the corresponding eigenvalue, A;, of A can be assigned to any desired value as an
eigenvalue of A + LC. This leads to the following definition.
'Alefinition 3.1 An eigenvalue >..; of A is said to be an observable (unobservable)
eigenvalue for the pair (A, C) if we can (cannot) find L such that A; is not an eigenvalue of
A+LC.
Reflection on the foregoing reveals that we can design an observer for a given state
model, i.e., we can find L so that A + LC is stable, provided the plant state model has no
eigenv~lues which are both unstable and unobservable. Plant state models having this
property are referred to as being detectable.
Definition 3.2 A plant state model is said to be detectable if all its unobservable
eigenvalues are stable.
Thus we can only determine an observer for a given state model if that state model is
detectable. This is of obvious importance in the implementation of state feedback using

an observer for the stabilization of an unstable plant.

Recall, in the case of state feedback, that the controller form state model facilitates the
calculation of K to achieve a specified set of closed loop eigenvalues. In the present
situation there is an analogous form for the plant state model which facilitates the
calculation L to assign a specified set of eigenvalues to the observer. This form, which we
encounter in the next section, is referred to as an observer form.

3.4

Observer Design

Suppose we are given a state model (A, B, C, D) for the plant and we want to determine L
so that the observer eigenvalues, >..-[A + LC], equal a specified set {!1; : i = 1, 2 n}. Then
provided any unobservable eigenvalues of the pair (A, C) are in the specified set, we can
determine L so that the observer eigenvalues coincide with the specified set. This could be
done by equating coefficients of like powers on either side of the equation
a(s) = 7(s)

72

State Estimation and Observabi/ity

where
n(s) = det[sl- (A+ LC)]

( ) II(
11

I s =

S -

f-Li

= S

s"

+ o 1.1n-I + n 2 s"- 2 ++On

+ II S11-l + I2Sn-2 + + In

i=l

The resulting equations, li = ai: i = 1,2, n, in the elements of L are, in general,

coupled in a way that depends on the plant's A and C matrices. This makes it difficult to
set up a general procedure for determining L from these equations. However, when y(t) is
a scalar and the state model for the plant has all its eigenvalues observable, this difficulty
can be overcome by using coordinates which put the state model for the plant in observer
form.

3.4.1

Observer form

Recall that any state model (A, B, C, D), for a plant is related to the plant transfer
function G(s) as
G(s) = C(sl- A)- 1B + D
Then a system having transfer function GT (s) has state model (A, B, C, 15) where

State models (A, B, C, i5) and (A, B, C, D) are said to be duals of each other when their
parameters are related in this way. When G(s) is symmetric, the dual of any state model
for G(s) is also a state model for G(s).
In the SISO case G(s) is a scalar so that GT (s) = G(s). Therefore, in this case, we can
obtain a canonical state model referred to as an observer form, by forming the dual of the
controller form state model. Thus if the state model for G(s) is in controller form,
(A"' Be, Ce, De), the dual state model is in observer form, (Ao =A'{, B0 = C'{,
Co= B'{, D0 =De), or more specifically
0

-a!

-a2

-an-2

-an-I

-an

bl

b2

Eo

...

bn-2

bn-l

bn

OJ

Do

,j

Observer Design

73

An indication of the importance of this form is evident from the ease with which we can
assign observer eigenvalues through L. Thus when the plant is in observer form,
F 0 = A 0 + L 0 C 0 is in the same form as A 0

+ lol
-az + lo2

-al

0 0

/1

0 0

/2
L
o-

Fo-

+ lon-2
-an-I + lon-1

-an+ lon

-an-2

ln-2
ln-1

ln

Therefore if we want the eigenvalues ofF to satisfy a specified characteristi c polynomial,

'Y(s), we have
n

det[s/ - F] =

sn

')';Sn-i

i=1

and the required elements of L 0 , {10 ; : i = 1, 2, n} are easily obtained from

loi

=a,.-')';

i=l,2,n

In order to preserve the eigenvalues of A + LC under a coordinate transformati on we

see that if L assigns a desired set of eigenvalues to A + LC in the transformed coordinates
then we need to replace L by L = T L to assign the same set of eigenvalues to A + LC in
the original coordinates. This is seen by noting that

3.4.2

Transformation to observer form

In the previous chapter, (Section 2.3.3), we developed an algorithm for generating a

coordinate transformati on matrix T so that any single input controllable state model is
transformed to a controller form state model. This algorithm can be adapted to provide a
coordinate transformati on matrix so that a single output observable state model is
transformed to an observer form state model. This adaptation is made using the duality
between the controller and observer forms. Thus by replacing A, Bin the algorithm given
in Section 2.3.3 by AT, cT we obtain a matrix T which transforms the pair AT, cT to
controller form. Then r-T is the coordinate transformati on matrix needed to transform
the given state model to observer form. Just as the transformati on to controller form is
only possible if (A, B) is a controllable pair so here the pair (AT, cT) must be controllable
or equivalently the pair (A, C) must be observable, i.e.,
rank[U] = n

(3.15)

74

State Estimation and Observability

where

The matrix U is called the observability matrix. Notice, in general, that U is a pn x n

matrix where pis the number of elements in y(t). Also in general we have rank[U] :<:; n
since the rank of any matrix is never greater than its smallest dimension.
Notice that if a pair (A, C) is observable, i.e., rank[U] = n, we can assign all the
eigenvalues of A + LC by choosing L. Otherwise if we are not able to assign all the
eigenvalues, one or more of the right-eigenvectors of A satisfies Cv; = 0 (Theorem 3.1)
and we can show that
(3.16)

Therefore at least one of then columns ofU is dependent so that rank[ U] < nand the pair
(A, C) not observable.
Alternatively, if all the eigenvalues of A+ LC can be assigned, all the right-eigenvectors of A, v; satisfy Cv; =f. 0 and (3.16) is not satisfied by any right-eigenvector of A. Then
assuming A has a complete set of eigenvectors, any vector q having n elements can be
written as
n

= LO!;Vi
i=l

and

implying that the columns of U are independent and rank[U] = n. Thus all the
eigenvalues of A+ LC are assignable if and only if (A, C) is an observable pair.

3.4.3

Ackermann 's formula

Observability

3.5

75

Observabil ity

So far we have encountered two effects of system obsenabilit;. Both the capability of
assigning all the eigenvalues to A + LC and the capability of transforming coordinates so
that the state model is put in observer form requires that the given state model be
observable, i.e.,
rank:uJ = n
where U is pn x n observability matrix, (3.15), with p = l.
It is important to notice from the discussion so far, that we can still design an observer
to estimate the state of a given state model when the state model is unobservable, i.e.,
when rank[U] < n. provided the given state model is detectable, Definition 3.2. Thus
observability is sufficient but not necessary for the existence of an observer for a given
plant state model.
In order to gain further insight into the meaning of observability, we are going to
consider the problem of determining the state at some estimation time from the
derivatives of the input and output at the estimation time. Notice that in practice we
avoid using signal differentiation since noise acquired in measuring a signal can appear
gt;eatly enlarged in the derivatives of the measured signal. Therefore our intent in
discussing the problem of determining the state from the derimtives of the input and
output is to provide additional insight into the theoretical nature of observability. We will
show that we cannot solve this problem unless the state model is observable.

3.5.1

A state determinat ion problem

Suppose we are given the state model, (A, B, C, D), and the derivatives of the input and
output at some estimation time, te, i.e.,
{ y i) ( te ) 'u i) ( te ) .. l. --

0 ' 1' 2 '

...

'n - l }

Then from the output equation

(3.17)
we have p equations in then unknown components of x(te), where pis the number of
elements in y(t). Now if p = n, Cis square and if the rows of Care independent. or
equivalently if the elements in y( t) are not related by constants, i.e., are independent, then
we can solve (3 .17) for the state at time te as

without requiring derivatives of the input and output signals. However usually p <nand
we need to generate equations in addition to those obtained directly from the output
equation (3.17). We can do this by using the state differential equation
(3.18)
together with the output equation (3.17).

76

State Estimation an~ Observability

Suppose p = 1. Then we have one equation, (3.17), which we can use in the
determination of then components of the state x(tc) We can generate a second equation
for this purpose by taking the derivative of (3.17)
(3.19)
and using (3.18) to obtain
(3.20)
We can generate a third equation for determining x(te) by differentiating (3.20) and
substituting (3.18) for the derivative of the state. This gives

Continuing in this way we can generate additional equations in the state until we have
the required n equations. These equations can be collected and given as
(3.21)
where

y(te)

z = Y-

ru

Y=

CA

1\te)

U=
CAn-1

/n-1)(te)

f=

U=

CB

CAB

CB

CAn- 2 B

CAn- 3 B

CAn- 4 B

Now since we developed (3.21) from a consideration of the effect of the state and input
on the output, we have z Erange[U]so that we can always find x(te) to satisfy (3.21).
When (A, C) is an observable pair, all n columns ofU are independent, rank [U] = n, and
UTU is invertible SO that when p > 1 we have

and when p

1 we have

Observability

77

However when rank[ U] < n, U has dependent columns and we have

(IS=

(3.22)

Ux(te)

for some non-null x(te) which we denote by x 0 (te) Solutions to (3.22) when rank[U] < n
are said to lie in the null space of U denoted
when
Hence if x(te) is any solution to (3.21) then we have x(te) + x 0 (te) as another solution to
(3.21). Therefore we see that if (A, C) is not observable it is impossible to determine the
true state, x( le), from the derivatives of the system's input and output.
In summary we have shown that the observability of the state model is a necessary and
sufficient condition for being able to determine the actual state of the plant state model at
any time from derivatives of the input and output at that time.

3.5.2

Effect of observability on the output

Continuing the discussion begun in Section 1.7.4 we see that when (A, C) is an
unobservable pair with right-eigenvector satisfying Cvj =(IS then when x(O) = v; we
obtain the zero-input response as
y(t)

CeA 1x(O)

C [I+

;t

1
I
C[I +At+ 2 ! A 2 t 2 + 3! A 3 t3 + ]x(O)

,2 2 1 ,3 3
] ;
>-.-t C ;
+ 21! /\;
t + 3! /\; t + v = e ' v =

(IS

(3.23)

We can illustrate this important effect by considering the unobservable state model

C= [+1 +1]
Then since A is in companion form, (Section 1.6.1 ), we can express its right-eigenvectors
as
A1 =

-1,

A2 =

-2}

and any initial state given by

x(O) = av 1 =

[+a]
-a

gives rise -to a trajectory which lies in the null space of C for all time so that y(t) = 0 for
all t.

State Estimation and Observability

78

Notice also that the transfer function corresponding to the foregoing state model is given
by

G(s)

s+1

s 2 + 3s + 2

s+2

and the unobservable eigenvalue of A at -1 is not a pole of the transfer function. This
effect was encountered in Section 1.9.2 in connection with diagonal or normal form state
models.
We will see, in Chapter 4, that this property of having certain non-null initial states
which produce no output when a state model is unobsevable leads to a characterization of
observability in terms of the rank of a matrix called the observability Gramian. The
observability Gramian is involved in determining the energy in the output caused by a
given initial state .

. 3.6

Observable Decomposed Form

Recall, from the previous section, that we are unable to determine x 6 (t) Enull[U] from
knowledge of a state model's input and output derivatives at timet. We refer to null[ U] as
the state model's unobservable subspace, denoted S 6 , i.e., S 6 =null [U]. Moreover we
refer to range[ UT] as the state model's observable subspace, S 0 Now it turns out that any
solution to (3.21) can be written as
x(t)

= x (t) + x 6 (t)
0

where
(i) x 6 (t) Enull[U] = S6 and is arbitrary otherwise;
(ii) x 0 (t) Erange[UT] = S0 and depends uniquely on z.
Moreover the observable and unobservable subspaces are orthogonal, S 6 ..LS0 , i.e.,
(3.24)
for any x 0 (t) Erange [ ur] and any x 6 (t) Enull[ U]. This can be seen by using the fact that

if and only if we can find w to satisfy

so that we have

In this section we show that any unobservable state model can be transformed, using a
change of coordinates, so that in the transformed coordinates the observable and

Observable Decomposed Form

79

unobservable subspaces align with subsets of the coordinate axes. More specifically, when
a state model is in observable decomposed form the state is given by

x(t)

[:~~~n

(3.25)

where xi(t) is oflength n; and

x(t)

E Sa

if x 2 (t) =Ill

x(t)

S0

if x 1 (t) =Ill

Notice that (3.24) is satisfied when the state is given by (3.25). Now the structure
required for the state model to be in observable decomposed form is given in the following
definition. Notice that this structure is dual to the structure of the controllable decomposed form, Definition 2.5, in the sense described in Section 3.4.1.
Definition 3.3 A state model having p scalar outputs is said to be in observable
decomposed form when its A and C matrices are in the following block forms

where (AhC 1) is an observable pair and AhA 3 ,A4 , C 1 have dimensions

n1 x n 1, n2 x n 1, n 2 x n2 and p x n 1 respectively with n = n 1 + n 2 being the dimension
of the state space.

3.6.1

Output dependency on observable subspace

One way of seeing that a state model in observable decomposed form decomposes the
state space in the manner specified in equation (3.25) is to note that, in these coordinates,
the transition matrix is lower triangular,

so that the zero input response is given by

(3.26)
Thus the zero input response is independent of x 2 (0) and S 0 is as specified in
equation '(3.25). In addition since (All CJ) is observable, we could determine x 1 (t)
from derivatives of the output. Thus Sa is also as specified in equation (3.25).

80

State Estimation and Observabi lity

3.6.2 Observ ability matrix

Alternatively, we can also see that the decomposi tion of the state space given by equation
(3.25) is achieved when the state model is in observable decomposed form by inspection of
the observability matrix.

0
0

c1
CIA

U=

C An 1 -l
1 1
C1A~ 1

0
0

C1Ai- 1

~]

[u1
u2

where

rank[U] = rank[UJ] = n 1

and we have

x E null[U]

if and only if

= [ :2 ]

(3.27)

for any x 2 of length n - n 1 .

3.6.3

Transfe r function

Again, notice that the observable decompose d form gives the transfer function as
G(s)

= [C 1 0] [

(s/-AJ)- 1

*
with order equal to or less than n 1 depending on the controllability of the pair (A 1, B1).
Notice also that the eigenvalues of A 4 are not poles of the transfer function since they are
unobservable.
This latter observatio n is readily verified from the structure of (A+ LC) in these

Observable Decompose d Form

81

coordinates since

and since A

+ LC is block lower triangular we have

with >.[A4 ] being eigenvalues of A

(A, C).

3.6.4

+ LC for all L, i.e., being unobservabl e eigenvalues of

Transformation to observab le decomposed form

Now we saw in Chapter 2 that the controllable subspace is A-invariant, i.e., if x Erange[n]
then Ax Erange[n], (Theorem 2.5). The analogous result here is that the unobservabl e
subspace is A-invariant, i.e., if x Enull[U] then Ax Enull[U]. When the state model is in
ot>servable decomposed form we can see this fact directly from (3.27) and the block
structure of A. Alternatively, we can see that this result holds more generally in any
coordinates by noting that we can always express the last block row of UA using the
Cayley-Ham ilton theorem, Theorem 2.3, as

Therefore the last block row of UA depends only on block rows of U and thus if Ux = 0
thenUAx = !1).
Now the A-invariance of the unobservabl e subspace provides a means for determining
the coordinate transformati on matrix needed to put a given state model in observable
decomposed form.
Theorem 3.2 A coordinate transformati on matrix T transforms a given state model
(A, B, C, D) having observability matrix U with rank[U] = n 1 to observable decomposed
form if
?

range[T2 ] = null[U]
where

with T~o T2 being n x ni> n x n """- n 1 such that r- 1 exists.

Proof Let QT denote the inverse ofT, i.e.,
(3.28)

State Estimation and Observability

82

Then the transformed A and C matrices are given by

-

A= T

_1

AT=

[Qf AT1 Qf AT2]

QJ AT1 QJ AT2

C=CT=[CT1

CT2J= [cl

C2]

Now since C appears as the first prows ofU and T 2 is chosen so that UT2 = 0, we have
CT2 = 0.
Finally since range[T2 ] = S 8 and QT is the inverse ofT we see from (3.28) that

C2 =

when x E S 0

Q[x = 0

Then since AxE Sa if x E Sa, we have range[AT2 ] = S 0 and thus

3.7

Qf AT2 =

0.

Minimal Order Observer

Recall from Section 3.3 that use of the plant output enables the construction of an
asymptotic state estimator for unstable plants. This estimator, referred to as an observer,
has dimensions equal to the dimension of the plant state model. In this section we further
recruit the plant output into the task of estimating the plant state by using it to reduce the
dimension of the observer. More specifically, if the plant output y( t) consists of p
independent scalars, {yJt): i = 1,2, p}, then y(t) contains information about the
state in the form of a projection of the state onto a p dimensional subspace of state space.
This fact allows us to concentrate on the design of an estimator for the remaining part of
the state consisting of its projection on the remaining n - p dimensional subspace of the
state space. In this way we will see that we can obtain a plant state estimator having
dimension n- p which is referred to as a minimal or reduced order observer. We can
develop this reduced order observer as follows.

3.7.1

The approach

Suppose the model for the plant whose state is to be estimated is given as

x(t) = Ax(t)
y(t)

Cx(t)

+ Bu(t)

(3.29)
(3.30)

where

Notice that there is no loss of generality in assuming that the state model has D = 0,
since when D =1- 0 we can replace the left side of (3.30) by q( t) = y( t) - Du( t) and use q( t)
in place of y(t) everywhere in the following.
Now we assume throughout that the components {yi(t) : i = 1, 2, ,p} of the plant

Minimal Order Observer

83

output y( t) are independent , i.e., we assume that

p

l::O';Y;(t) = 0

for all t E [0, oc)

only if n;

for all i E [Lp]

i=i

This implies that C has independent rows. As mentioned earlier in this chapter, if p = n
then Cis invertible and we can solve (3.30) for x(t) as
x(t)

= c- 1y(t)

Since we usually have p < n, Cis not invertible. However, even in this case, we can still use
y(t) and (3.30) to obtain information on the state. More specifically, we can determine
xR(t),the projection of the state on range [cT], from
(3.31)
where c# = cT ( ccT)- 1is referred to as aright inverse ofC since cc# =I. Then we can
wrjte x(t) as
(3.32)
where
XN (t) E

null[CJ

Now since we can be obtainxR(t) from(3.31), we see from (3.32) that we can determine
the complete plant state provided we can develop a method for estimating xN (t). Before
doing this we show that xR (t) can be determined from (3.31 ).

3.7.2 Determination of xR(t)

Notice that the independenc e of the rows of Censures that CCT is invertible. We can see
this as follows. Suppose ccT is not invertible. Then we can find y # 0 such that

where w = cTy. However this is possible only if w = 0 and therefore cT has dependent
columns. Since this contradicts the assumption that C has independent rows we must
have CCT is invertible.
Now the expression for c#, (3:31), can be derived as follows. Since xR(t) Erange[CT] it
follows that
(3.33)
for an appropriate w(t). Then multiplying this equation through by C and using the

~-/--

84

State Estimation and Observabilit y

invertability of CCT enables

US

to determine w(t) as
(3.34)

However from (3.32) we see that

y(t) = Cx(t) = CxR(t)

and we can rewrite (3.34) as
(3.35)
Finally, premultiplyi ng this equation by CT and using (3.33) yields (3.31).

3.7.3

A fictitious output

Having determined the projection, xR(t), of x(t) on range [cT] along null[C], we need to
develop a method for estimating xN (t). This is done by introducing a fictitious output,
yF(t), and using it withy(t) to form a composite output, yc(t) as

y c (t)

y(t)
[ YF
(t) ]

c ] x(t)
[ CF

Ccx(t)

(3.36)

where C F is chosen so that Cc is invertible, i.e., C F has n - p independent rows each

independent of the rows of C. Then we can solve (3.36) for x(t) as
(3.37)
Before showing how to generate YF(t), it should be pointed out that we can ensure the
invertability of C c by choosing C F so that
range[CJ]

= null[C]

(3.38)

We can see this by using the general result for matrices which is developed at the
beginning of Section 3.6 in terms of the matrix U, i.e.,
range [ uT] l_null[ U]
Therefore we have
range [cJJ l_null[C Fl
and it follows from (3.38, 3.39) that
null[C] n null[CF] = {0}

(3.39)

Minimal Order Observer

85

so that
only if

X=0

which implies that Cc is invertible.

Now assuming we chose C F to satisfy (3.38) we can readily show that
(3.40)
where

This enables (3.37) to be rewritten as

(3.41)
I

3.7.4

Determination of the fictitious output

Having obtained the state of the plant model in terms of the known plant output and the
fictitious plant output, (3.41 ), we need a means for generating the fictitious output. This is
done by using the plant state differential equation to set up a state model which has the
plant's input and output, (y(t), u(t)), as input and has state which can be used as an
estimate of the fictitious output. The derivation of this state model is given as follows.
We begin by multiplying the plant state differential equation (3.29) by CF to obtain
(3.42)

Then differentiating x(t), (3.41), and substituting the result in (3.42) gives

However from (3.38, 3.40) we have CFC#

written as

0 and CFC# =I. Therefore (3.43) can be

(3.44)
Now equation (3.44) suggests that an estimator for yF (t) has differential equation
(3.45)
where YF(t) will be an asymptotic estimate of y(t) provided CFAC# is stable. This
follows from the differential equation for the estimation error, yF (t), which we can

86

State Estimation and Observability

determine by subtracting equation (3.44) from equation (3.45) to obtain

where

Thus we have

so that
for all

yF (0)

(3.46)

if and only ifF R is stable.

3.1.5

Assignmen t of observer eigenvalue s

Now the eigenvalues ofF R = CFACt are fixed by the plant state model. Therefore when
some of these eigenvalues are unstable we cannot satisfy (3.46). Recall we encountered the
same sort of difficulty in Section 3.2 in connection with trying to estimate the state of a
plant state model without using the plant output. We can overcome the present problem
when the state model, whose state is to be estimated, is detectable by replacing CF
everywhere in the foregoing development by
(3.47)

where LR is used in the same way that Lis used in Section 3.3 to assign the eigenvalues of
an observer's system matrix, A+ LC. This replacement is effected by replacing yF (t) by
z(t) = Tx(t) in (3.36) to obtain
[ y(t)] = Crx(t)

(3.48)

z(t)

where
Q= [ I
LR

0]
I

Notice that since Cc is invertible and Q is invertible, independent of LR, Cr is invertible

for any LR. Therefore we can determine x(t) from (3.48) for any LR as
x(t)

My(t)

+ Nz(t)

(3.49)

Minimal Order Observer

87

where
(3.50 i

Now by forming C rCi 1 =I we can show that .'vi and X must satisfy
CN=0
TN=!

(3.51)

Therefore recalling from (3.38. 3.40) that

[cFc] Lc~

(3.52)

we see that the constraints (3.51) on Jf and X are satisfied when

S=Ct

In order to obtain a differential equation for z(t), we follow the approach used to get a
differential equation for yp(t). (3.44). Therefore multiplying equation (3.29) by T and
substituting for x(t) from (3.49) yields
:i(t) = TAXz(t)

where z(t)

+ TAMy(t) + TBu(t)

(3.53)

Tx(t) and
TAN= CpAC1

+ LRCAC1

TAM= CpAC#

+ LRCAC#- CpAC1LR- LRCAC1LR

This differential equation suggests that an estimator, having the same form as an
observer, (3.13), can be formulated for z(t) as

z (t) = Fri(t) + Gr [u(t)]

y( t)

(3.54)

where
Fr =TAN

Gr=[TB

TAM]

Then the estimation error for z(t) is given by

I(t)

(eFT 1)i(O)

where i(t) = z(t)- z(t) and we have an asymptotic estimate of z(t) if and only ifF r is
"
stable.
Thus provided F Tis stable, we can usez(t) generated by equation (3.54) in place ofz(t)

88

State Estimation and Observabilit y

in (3.49) to obtain an asymptotic estimate of x(t) as

x(t)

My(t)

+ Nz(r)

(3.55)

Notice that the dependency ofF T = TAN on LR, (3.53), implies that we are only able
to choose LR so that TAN is stable if the pair (CpAC'J;, CAC'J:) is detectable. We show
now that this detectability condition is satisfied when the plant state model or pair (A, C)
is detectable.
Theorem 3.3 The pair (CpAC#, CAC#) is detectable if and only if the pair (A, C) is
detectable.
Proof Suppose (>., v) is an eigenvalue, right-eigenvector pair for A. Let r be a
coordinate transformati on matrix given as

r = [ c# c#]

(3.56)

Then premultiplying the eigenvalue-e igenvector equation Av

rr- 1 between A and v gives

AV by r-l and inserting

Aq = >.q
where q =

(3.57)

r- 1v and

Then from (3.52) we see that Cis transformed as

c = cr = C[ c# c# ] =

[1

0]

Now suppose A is an unobservable eigenvalue of the pair (A, C). Then recalling that
unobservable eigenvalues are invariant to coordinate transformati on, we see that q,
(3.57), satisfies

Cq =

which from the form for C requires q to have its first p elements zero

where q2 has n- p components. Then using the form just obtained for
eigenvalue-e igenvector equation, (3.57) becomes

A, we see that the

Minimal Order Observer

89

implying that ,\is an unobservable eigenvalue of the pair (A 4 , A2 ) which equals the pair
(CFAC1, CAC1) specified in the statement of the theorem. This shows that the
observability or detectability. (if Re )] ~ 0 I. of (A. C) is necessary for the obsen ability
or detectability of the pair (C FAC1, CAC1). Sufficiency can be shown by reversing the

foregoing argument.
Notice that when the plant state model's (A, C) pair is observable the coordinate
transformation r, (3.56), just used in the proof of Theorem 3.3, can be used to determine
LR, (3.47), to achieve some specified set of eigenvalues for the minimal order observer's
system matrix, F T, (3.54). One way of doing this \Vouid be to use Ackermann's formula,
introduced in the previous chapter, on the pair (A 4 . A 2 ). (3.57).
The proof of Theorem 3.3 suggests that if we work in coordinates where the state
model's C matrix is

C= [I 0]
then a minimal order obsener to estimate the state in the transformed coordinates can be
determined by taking T, (3.47), as

where Q is any nonsingular matrix of appropriate size. This ensures that CT, (3.48), is
invertible independent of LR. Notice that Q plays the role of a coordinate transformation
matrix for the coordinates of the minimal order observer and therefore Q does not affect
the minimal order observer eigenvalues. Therefore referring to (3.54, 3.55) and taking
Q =I for simplicity. we obtain a minimal order obsener for the estimation of the
transformed state as

ij(r) = My(t)

+ ~vz(t)

z (t) = T ANz(t)

T AMy(t)

+ T Bu(t)

where
N=

TAN = A4

[~]

+ LRA2

TAJ1 = LRAI- A4LR- LRA2LR

+ A3

In summary, we have seen that a minimal order observer for ann-dimension al plant
state model is (n- p )-dimensional, where p is the number of independent scalar plant
outputs available for use in estimating the plant state. As in the case of the full order
observer introduced in the first few sections of this chapter, unobservable eigenvalues of
the plant state model's pair (A, C) are fixed eigenvalues for the minimal order clJserver's
system matrix FT, (3.54). Thus as in the case of the full order observer we are unable to
design an asymptotic observer when the plant state model is not detectable.

90

State Estimation and Observabilit y

3.8

Notes and Referenc es

The obse1 ver was originally proposed by David G. Luenberger in his Ph.D. thesis in 1964.
For this reason the observer is sometimes referred to as a Luenenberge r observer.
Previously, in 1960, the use of a state model having the same structure as an observer
was proposed by R. E. Kalman for the purposes of least squares signal estimation in the
presence of additive noise. We will encounter the Kalman filter in Chapter 6. For further
discussion of observers the reader is referred to Chapter 4 of [23]
Finally, the term " output injection" is used by some authors to refer to the system
signal manipulatio n we use to obtain an observer having system matrix A + LC. This
terminology should be compared with the use of the term "state feedback" in connection
with the formation of a system having system matrix A + BK.

4
Mod el Approximation via
Bal anc ed Rea liza tion

4.1

Introduction

In previous chapters it was noted that whenever a state model (A, B, Ci is either not
cotitrollable or not observable the transfer function resulting from the calculation
C(sl- A)- 1B has pole-zero cancellations. Thus system uncont;oll ability and/or unobservability gives rise to a transfer function with order less than the dimension of the state
space model. Therefore systems that are "almost uncontroll able" or "almost unobservable", should be able to be approxima ted by a transfer function model of order less than
the dimension of the state model. Reduced order model approxima tion can be of
importanc e in the implement ation of feedback control systems. For instance, model
order reduction may be beneficial when the large dimension of the full state model gives
rise to computati onal problems during the operation of the control system.
This chapter treats the problem of extracting a subsystem of a given full order system
which works on that part of the state space which is most involved in the input-outp ut
behavior of the original system. This is accomplished by introducin g t\m m;trices called
the observability Gramian and the controllab ility Gramian. It is shown that these
matrices can be used to quantify the energy transfer from the input to the state and
from the state to the output. In this way it is possible to identify which part of the state
space is controlled most strongly and which part is most strongly observed. Then in order
to construct a reduced order model by extracting a subsystem from the given system. it
will be shown that the strongly controlled subspace must coincide with the strongly
observed subspace. To achieve this coincidence we change coordinate s, if necessary, so
that the Gramians become equal and diagonal. The resulting state model is said to be a
balanced realization. The reduced order model is then obtained by extracting a subsystem
from the balanced realization.

4.2

Controllable-Observable Decomposition

The idea for model order reduction being pursued in this chapter grew out of the basic
fact that only the jointly controllable and observable part of a state model is needed to
model the input -output behavior.

Model Approximation via Balanced Realizati on

92

The controlla ble and observable part of the state space can be viewed as one of four
possible subs paces for the state space of any given state model. Morepecifically, consider
the four dimensional state model in diagonal form which is shown in Figure 4.1.
Inspection of the block diagram reveals that the four dimensional state model has a
is
transfer function which is only first order since only the first compone nt of the state
which
space
state
the
of
part
that
Thus
output.
the
affects
both affected by the input and
is
is both controlla ble and observable forms a one dimensional subspace Sea which
specified as
if x 1(t) = 0 fori= 2, 3,4
x(t) E Sea
where

On further inspection of Figure 4.1 we see that there are three other subspaces sci5J SCOJ
Sea specified as

Sea
x(t) E Sea
x(t) E Sea

x(t)

if x,(t) = 0 fori= 1,3,4

if x 1(t) = 0 fori= 1,2,4
if x 1(t) = 0 fori= 1, 2, 3

where Sea' Sea' Sea' Sea are referred to as follows.

Sea
Sea
Sco

Sea

is the controllable and observable subspace

is the controllable and not observable subspace
is the not controlla ble and observable subspace
is the not controlla ble and not observable subspace

More generally when we are given a higher dimension state model, (A, B, C), which is

~~--~

u(t)

~
-~
Figure 4.1

Example of controllab le-observa ble decompos ition

Controllable -Observable Decomposit ion

93

not in any special form, we can extract a subsystem from the given system which has its
state space coincident with the controllable and observable subspace. This can be done by
using two successive coordinate transformati ons as described now.
First we change coordinates so that the state model is in a controllable decomposed
form
. (4.1)
with (Ac1, Bc1) being a controllable pair. The transformati on matrix to do this was
discussed in Chapter 2.
Notice that, in the new coordinates, we have the full n dimensional state space split into
controllable and uncontrollable subspaces, Sc, Sc. It remains to split Sc into observable
and unobservable subspaces, i.e., to split Scinto Sco and Sco
Next we carry out a second coordinate transformati on which affects only the
controllable part of the controllable decomposed state model, (4.1). This is done using
a coordinate transformati on matrix, T, which is constructed as
I

with To being chosen so that the controllable state model (Ac1, Bc1, Cc1) is transformed to
observable decomposed form, i.e., so that

0]

(4.2)

with (Aol, Col) being an observable pair. The transformati on T0 to do this was discussed
in Chapter 3. Since coordinate transformati on preserves controllability, the.state model
(4.2) is controllable. Therefore the subsystem of the observable part of the ~tate model
(4.2), i.e., (A 01 , B01 , C01 ) is both controllable and observable. Thus the transfer function
for the entire system has order equal to the dimension of this subsystem and is given by

G(s)

,~;

= Col (sf -

Ao!)

-1-

Bol

Notice that the foregoing model reduction procedure applies to full order state models
which are either not controllable or not observable or both not controllable and not
observable. More important, the reduced order model we obtain has exactly the same
input-outpu t behavior as the full order model.
In this chapter we are interestedin obtaining a reduced order model when the full order
state model is completely controllable and observable. In this situation any reduced order
model will have input-outpu t behavior which differs from the input-outpu t behavior of
the full order state model. Our goal is to develop a method for achieving a reduced order
state model having input-outpu t behavior which approximates the input-outpu t
behavior of the full order state model. We will see that we can develop an approach
for solving this model approximati on problem by replacing the concepts of controll-

Model Approximation via Balanced Realization

94

ability and observability by concepts of energy transfer. We will see that the distribution
of energy among components of the state is related to properties of Gramian matrices
which replace the controllability and observability matrices in this analysis. Just as
coordinate transformations are important by extracting the controllable and observable
part of a state model, so too are coordinate transformations importanUo being able to
separate out that part of a controllable and observable state model which is most
important in the transfer of energy from the system's input to the system's output. The
state model in coordinates which enable the identification of this part of the system is
called a balanced realization.

4.3

Introduction to the Observability Gramian

In order to define the observability Gramian we need to consider the related idea of
output energy. Recall that if the state model (A, B, C) is known, then the zero input
response, y(t), caused by any specified initial state, x(O), is given as

y(t) = CeA 1x(O)

(4.3)

The output energy, E 0 , for each initial state, x(O), is then defined as the integral of a
nonnegative, real, scalar function v( t) of y( t), i.e., v(t) ;:::: 0 for all t E [0, oo)

Notice that since v(t) is non-negative real for all t, the output energy E 0 for any initial
state x(O) must satisfy
for all x(O)

E0

;::::

E0

= 0 only if v(t) = 0 for t;:::: 0

A simple way to choose v(t) so that E 0 has these properties is as follows.

if y(t) is a real scalar
if y( t) is a complex scalar

then choose
then choose

if y(t) is a real vector

if y(t) is a complex vector

then choose
then choose

v(t) = /(t)
v(t) =I y(t)l2
v(t) = YT (t)y(t)
v(t) = y*(t)y(t)

Notice that when y(t) is a scalar, y*(t) denotes the complex conjugate of y(t) and
IY( t) 12 = y* (t)y( t). Alternatively when y( t) is a complex vector, y* (t) denotes the transpose
of y(t) and the complex conjugation of each entry in the resulting row vector. Thus if y(t)
is a complex vector of length p then
p

y*(t)y(t)

LIYz(t)l 2
1=1

More important, notice that each of the foregoing choices is a special case of the last

Introduction to the Observabillty Gramian

95

choice. Therefore, to avoid special cases, we define the output energy E 0 , in general as

Eo

looc y* (t)y( t)dt

(4.4)

so that after substituting from (4.3) we obtain

Eo= x*(O)
r

(loao eA'rc*ceA dt)x(O)

1

(4.5)

(4.6)

x*(O) W 0 x(O)

where the constant n x n matrix W 0 is called the observability Gramian.

Notice that in order for the foregoing integral to converge, or equivalently for W 0 to
exist, the observable eigenvalues of (A, C) must be in the open left half plane. This can be
seen by referring to section 7 of Chapter 1. Alternatively, if (A, C) is observable W 0 exists
only if A is stable. Notice also that W 0 must be
1. an Hermitian matrix, i.e., W o = W ;,
2. a non-negative matrix, denoted W 0 ~ 0.
I

Requirement number one results from the integral defining W 0 , (4.5). Hermitian
matrices have entries in mirror image positions about the diagonal which are conjugates
of each other, i.e., (Wo)ij = (W 0 )j;. When W 0 is real we require W 0 = W~ and W 0 is
referred to as a symmetric matrix. Hermitian (symmetric) matrices have eigenvalues and
eigenvectors with some rather striking properties which we will develop in the next
section.
Requirement number two results from the output energy being nonnegative for all
initial states since a non-negative matrix, W 0 , has the property that

for all vectors x(O) of appropriate length. Now it turns out that we can always find a
matrix M such that any Hermitian, nonnegative matrix W 0 can be factored as
W 0 = M* M. This fact allows us to write
x(O) W 0 x(O)

x(O)* M* M x(O)
n

= z*z =

L)d2
i=l

where z = Mx(O). Since I: 7= 1 lz;l 2 ~ 0 for any z we have W 0 ~ 0. Moreover since

L %:1iz;l 2 = 0 if and only if z =,we have x(O)* W 0 x(O) = 0 for some x(O) -=f 0 if and
only if M and hence W 0 is singular, i.e., if and only if W 0 has one or more zero
eigenvalues. More is said about this in the next section.
Finally the observability of a state model is closely tied to the nonsingularity of its
observability Gramian. This can be seen by noting from (4.4) that E 0 = 0 if and only if
y(t) = 0 for all time. However we saw in Chapter 3 that the zero input response is null,
y(t) = 0, for some x(O) -=f 0 if and only if (A, C) is unobservable. Therefore it follows that

Model Approximation via Balanced Realization

96

(A, C) is unobservable if and only if W 0 is singular. Further insight into this fact is given
in subsequent sections where it is shown that W 0 can be obtained by solving a linear
matrix equation which is referred to as the Lyapunov equation and which involves A and
C in a simple fashion.

4.4

Fundamental Properties of W0

In the previous section we saw that the observability Gramian is a member of the class of
matrices which are characterized as being Hermitian and nonnegative. In this section we
give some of the essential features of these kinds of matrices.

4.4.1

Hermitian matrices

Recall that a matrix W 0 is said to be Hermitian if it equals its conjugate transpose,

W 0 = W~. This implies the ijth and jith entries are conjugates of each other. The
eigenvalues and eigenvectors of Hermitian matrices have some rather striking properties
which are given as follows.
Theorem 4.1 If W 0 is Hermitian then:
(i) W 0 has real eigenvalues
(ii) W 0 has right eigenvectors which satisfy v;*v 1 = 0, when er 0 ;

#- er J
0

Proof (i) Let (er 0 , v) be any eigenvalue-eige nvector pair for W 0 , i.e.,

(4.7)
Then premultiplying (4.7) by v* gives
(4.8)
and taking the conjugate transpose throughout this equation gives
v* W~v

= er~v*v

(4.9)

Then subtracting (4.9) from (4.8) we obtain

v*(W 0

W~)v

=(era- er~)v*v

(4.10)

Now since W 0 is Hermitian, the left side of (4.1 0) is zero. Therefore either er o = er~ or
v* v = 0. However v* v = 0 holds only if v = 0, and since vis an eigenvector of W 0 , we have
v*v #- 0. Therefore (4.10) can only be satisfied if er 0 = er~ which shows that era must be
~.

era!

Proof(ii) Let (er 0 ;,v;: i =

#- ero2 Then we have

1,2) be two eigenvalue-eige nvector pairs for W 0 with

(4.11)
(4.12)

97

Fundamen tal Properties of Wo

Next operate on these equations as follows:

(i) premultipl y (4.11) by 12*
(ii) postmultip ly the second form for (4.12) by v 1
(iii) subtract the result of (ii) from (i)
This yields
(4.13)

However, since H" 0 is Hermitian , the left side of (4.13) is zero. Moreover we are assuming

that 0'01 =fc 0'~ 2 . Therefore (4.13) is satisfied only if vtv 2 = 0.

different
to
ing
correspond
rs
eigemecto
,
symmetric
..
i.e
reaL
is
Notice that when W 0
. .. v iT v j = 0 : 1. 1_;_ ],. 0'0 ; 1_;_ O'oJ
Iues are orth ogona,I I.e
.
e1genva
Recall that if a matrix has distinct eigenvalues, it has a complete set of eigenvectors and
can be diagonaliz ed by a matrix haYing columns made up from its eigemecto rs. Thus
when Wa has distinct eigenvalues we can use its eigenvectors {v; : i = 1, 2, n} to form a
matrix V 0 \vhich diagonaliz es W 0

v- 1 w v
0

-2: 0

(4.14)

0-

where

:E 0

["~1

(l'o2

0
=

0
V 0 = [v 1

v2

diag[0'0 J, O'o2'

, O'on]

O'on
vn]

HoweYer if the eigenvectors are normalized ,

then we see from Theorem 4.1 that

( 4.15)

1
Therefore V~ 1 = V~ so that V~W 0 V = 2: 0 . Matrices satisfying the condition V~
are referred to as unitary matrices.
Alternatively, when W 0 is real then (4.15) becomes

V~

V~Va =I
and therefore V~ 1

V~ so that V~ W 0 V 0 = 2: 0 Matrices satisfying the condition

98

Model Approximation via Balanced Realization

v,;

are referred to as orthogonal matrices. It can be shown that any Hermitian

V;,- 1 =
matrix car1 be diagonalized by a unitary matrix even if the Hermitian matrix has multiple
eigenvalues.
To recap, we have encountered the following matrix types.
Definition 4.1
If M is complex then
M is Hermitian

if M

M*

M is unitary

if

M- 1 =

M is symmetric

if

M = M

M is orthogonal

if M- 1 = M

M*

If M is real then

4.4.2

Positive definite and non-negative matrices

Suppose that we change coordinates, viz., x(O) = Tz(O), with T nonsingular. Then for
each initial state x(O) in the old coordinates, there is a unique initial state z(O) in the new
coordinates and vice versa. Thus we see from the expression for the output energy, (4.6),
that the output energy can be rewritten in terms of the initial state in the new coordinates
as
(4.16)
where W0 = T*W 0 T.
Now if we choose T as the unitary matrix V 0 , (4.14), then

and (4.16) becomes

11

E0

= z*(O)~oz(O) =

(4.17)

CT0 ;IzJO)I 2

i=1

However since there is a unique z(O) for each x(O) we have that E 0 2 0 for all x(O) if
and only if E 0 2 Oforallz(O). Thusweseefrom(4.17)thatE 0 2 Oforallx(O) ifandonly
if all eigenvalues of W 0 are non-negative, i.e.,
croi

20

i= l,2,n

Otherwise, if W 0 were to have some negative eigenvalues, we could choose z(O) so that
z1(0) c/c 0 if cr01 < 0

and

z1(0) = 0 if cr 01 > 0

i = 1, 2, , n

which would give E 0 < 0, for the initial state x(O) = Tz(O) and the requirement for Eo to
be non-negative for all initial states would be violated.

Fundamental Properties of Wo

99

Finally, we can show by a similar argument that E 0 > 0 for all x(O) of 0 if and only if
all eigenvalues of W 0 are positive.
To recap, we have defined the matrix properties of positive definiteness and nonnegativeness as follows.
Definition 4.2 W 0 said to be:

(i)
(ii)

when x*W 0 x > 0

positive definite, denoted as W 0 > 0

non-negative, denoted as W 0 ~ 0

when x* W 0 x ~ 0

for all x
for all x

of 0

Moreover these properties impose certain restrictions on the eigenvalues of W 0

namely

4.4.3

>0

(i)

then

U0;

(ii)

then

U0 ; ~

for

U0;

for

U0; E

>.[W 0 ]
>.[W0 ]

Relating E0 to .A[W0 ]

In order to begin to appreciate the role played by the eigenvalues of W 0 in achieving a

lower order state model approximation we consider the problem of finding the initial
state, Xmax(O), which produces the most output energy, Eomax when the initial state is
constrained so that
X~ax(O)Xmax(O)

=1

(4.18)

When x(O) is constrained in this way we say x(O) is normalized to one. Notice that this
constraint on the initial state removes the possibility of obtaining an unbounded output
energy by choosing an initial state with an unbounded component.
We begin the development of a solution to this problem by noticing that if we use a
coordinate transformation matrix T which is unitary, the constraint (4.18) is preserved,
i.e., if T satisfies T* T = I then we have
?

x*(O)x(O)

z*(O)T*Tz(O)

z*(O)z(O)

(4.19)

Therefore if x*(O)x(O) = 1 so does z*(O)z(O).

Now we have seen that using the unitary matrix T = V 0 as a coordinate transformation matrix makes the observability Gramian diagonal in the new coordinates. Therefore
our problem can be solved by first solving the problem in coordinates in which the
observability Gramian is diagonal and then transforming the solution, Zmax(O) in these
coordinates back into the original coordinates.
Suppose we are in coordinates where the observability Gramian is a diagonal matrix,
E 0 Suppose also that we have reorder the components of the state in these coordinates so
that the largest entry on the diagonal ofE 0 is first, i.e., u01 ~ CT0 ; : i = 2, 3, , n. Then we
see from equation (4.17) that the output energy satisfies
n

Eo=

Luo;iz;(O)I
i=l

::=;

U0 1

L
i=l

z;(O)

(4.20)

100

Model Approximation via Balanced Realization

However since we are assuming that x(O) is normalized to one, we see from (4.19) that
z(O) is also normalized to one
n

z*(O)z(O)

z;(O)

12

= 1

(4.21)

i=l

Therefore under this constraint we see from (4.20) that the output energy, E 0 , is bounded
by the observability Gramian's largest eigenvalue, i.e.,
(4.22)
Notice that if we choose

then (4.21) is satisfied and the output energy becomes

n

Eo=

L D"o;lz;(O)I

2 = D"oJ

i=O

and we achieve the upper bound on E 0 , (4.22). Thus the normalized initial state in the
original coordinates which maximizes the output energy is obtained as
(4.23)

1
0

Therefore our problem is solved by setting the initial state equal to the eigenvector of
W 0 corresponding to the largest eigenvalue of W a when that eigenvector is normalized to
one. Moreover, it is not hard to see from the foregoing that, in general, we have

when
x(O)

= vk

The foregoing analysis indicates that the relative importance of the components of z(O)
in producing output energy depends on the size of an eigenvalues of W 0 relative other
eigenvalues of W 0 . We can use this observation to suggest how we are going to be able to
obtain a reduced order state model approximation. We do this as follows.
Suppose we choose the coordinates for the state space so that the eigenvalues of W 0 are
in descending size, i.e., a 01 ?: a 02 ?: ?: a 0 n with W 0 diagonal. Suppose we consider z(O)

_.{.,;:;.

".

Introduction to the Controllability Gramian

to have all its elements equal, z;(O)

101

zi(O). Then the output energy is given by

0

0
0:
0

1z(O) = I~
11

!Joi

!Jon

where

Moreover, we can always choose an integer n 1 E [0, n]large enough so that

(4.24)

where

n,

Eo]

= ~.

L
i=l

IJc;

Eo2 =

~.

!Joi

i=n 1+1

Therefore in this situation a good approximation to the output energy is

Since the last n- n 1 components of the initial state are less important to the zero input
response than the first n 1 components under the foregoing condition of an equal element
initial state, a reduced order state model in these coordinates could be formed as
(AhB 1 , C\) where A1 is the n 1 x n 1 partition in the top left corner of A and .8 1, C\ are
corresponding partitions of B, C.
However, our interest in this chapter is in obtaining a reduced dimension state model
whose input-output behaYior approximates the input-output beha\io; of the full
dimension model. The foregoing argument for neglecting the last n- n 1 components
of z(t) focuses only on the zero input response. Therefore what is missing from the
previous argument, if we are interested in modeling the input-output behavior, is the
effect the input has on the neglected components of z(t). The effect of the input on
different components of the state is related to properties of a matrix known as the
controllability Gramian.

4.5

Introduction to the Controllability Gramian

The controllability Gramian. which we introduce in this section, is im olved in characterizing the energy delivered to the state when the input is a unit impulse and the initial
state is null. Since the Laplace transform of the zero state response to this input is the
given system's transfer function the distribution of energy among the components of the
state in this situation is appropriate to our goal of obtaining a reduced order transfer
function approximation. We begin the dewlopment of the controllability Gramian as
follows.

102

Model Approxim ation via Balanced Realizatio n

Suppose we are given a state model, (A, B, C). Then the state in response to a unit impulse
at time zero when the initial state is null is given by

Then the energy, Ec, transferred from this input to the state is given by
(4.25)
Notice that the integrand obtained here is strikingly different from the one obtained in
the integral for the output energy, (4.5). We proceed now to express Ec (4.25) as an
integral which appears as the dual of the integral expression of E 0 More specifically, we
show now that Ec can be expressed as the trace of the integral obtained earlier for E 0 , (4.5)
with A replaced by A* and C replaced by B*. The trace of a matrix is a scalar which is
defined as follows.
Definition 4.3 The trace of any square matrix M of size n, is defined as the sum of its
diagonal entries,
n

= 2: m 11

trace[M]

i=l

where mu is the ijth entry in M.

In order to proceed we need the following result involving the trace of the product of
two, not necessarily square, matrices
Theorem 4.2 Given two matrices M, N of size p x n, and n x p then
trace [M N]
Proof Let M N

trace [N M]

= Q. Then
p

trace[MN] =

2: q

11

i=l

where q 11 is the result of multiplying the

ith

q,, =

row of M by the

ith

column of N,

2:munii
j=l

Therefore we can rewrite the trace of the matrix product as

p

trace[MN]

= 2: L

mun;;

i=l j=l

Next interchang ing the order of the summation s gives

n

trace[MN]

L L n mu
11

j=l i=l

.~

Introductio n to the Controllab ility Gramian

103

and we see that

where SjJ is the /h entry on the diagonal of S = N M, and the result follows
the
is
matrix
square
any
of
trace
the
that
is
theorem
this
of
ce
consequen
An immediate
1
sum of its eigenvalues. More specifically if JJ = 1"- \Y, then
n

trace[M] = trace[V- 1AV] = trace[Avv - 1] =

2:..:::\
i=l

where>..; E .\[Jf1
Returning to the problem of re-expressing the energy En (4.25), we use Theorem 4.2
with

to obtain

Ee

= trace[Wc]

(4.26)

\Vhere We. the controllab ility Gramian. is given by

(4.27)
Notice that We is Hermitian and that We is non-negative since the integrand in (4.27)
is quadratic. Notice also that if we replace A by A' and B by C* in the foregoing definition
of the controllab ility Gramian. We. (4.27), we obtain the obsenabil ity Gramian, W 0 ,
(4.6). Thus all the properties discussed in connection with W 0 hold for WeMore specifically. since We is Hermitian , Tre has real eigemalue s and can be
diagonalized by a unitary matrix, V 0 i.e.,
(4.28)
where V;:- 1 = v; and E, is real and diagonal. Notice that Vc has columns which are
normalized right eigenvectors of We and the diagonal entries in the diagonal matrix Ec
are the eigenvalues of rr,. In addition recalling the discussion following the proof of
Theorem 4.2 that the trace of a square matrix equals the sum of its eigenvalues, we have

Now we will see in the next section that we can always transform coordinate s so that
the controllability Gramian is diagonal. Assuming we have done this transfon11ation, we
permute componen ts of the state, if need be. so that the entries along the diagonal of Ee
appear in descending size, i.e., uc; 2': uc(i+I) : i = 1, 2, n- 1. Then we can always

104

Model Approximat ion via Balanced Realization

choose n 1 so that

where
with

Eel =

L"'

aci

i=l

and with

Ec2 =

ac;

i=n 1+1

This inequality implies that the first n 1 components of the state in these coordinates
receive most of the energy from the input. If in addition, in these coordinates, the
observability Gramian happened to be diagonal with entries on the diagonal being in
descending size, the first n 1 components of the state would also be relatively more
important than the last n- n 1 components in the transfer of energy to the output. We
could then obtain an approximati on to the full order transfer function for the system as

where the full state model in these special coordinates has its system parameters,
(A 6 , Bb, Cb), partitioned to reflect the partitioning of the state, i.e.,

In the next section we will see that it is always possible to transform the coordinates so
that any controllable and observable state model has controllability and observability
Gramians which are equal and diagonal, i.e., so that in the new coordinates

with L:b having entries in descending size along its diagonal. A state model having this
property is referred to as a balanced realization. A reduced order approximati on, Gapx(s),
to the full order system is obtained from the balanced realization in the manner just
described.
Finally, we are assuming the pair (A, B) is controllable so that every component of the
state receives some of the energy supplied by the input. Therefore, from the foregoing, we
have all eigenvalues of We bigger than zero or We> 0 when (A, B) controllable. In
addition, it turns out that
rank[Wc]

4.6

rank[D]

Balanced Realizati on

Suppose we are given the controllability and observability Gramians, W"' W 0 , for a
controllable and observable state model, {A, B, C, D}, of the system to be approximate d.
Then in this section we will show how to construct a coordinate transformati on matrix so

~;

Balanced Realization

105

that the state model in the new coordinates is a balanced realization. Just as in Section 4.2
we saw that a succession of two state coordinate transformations are required to achieve a
controllable-observable decomposed form, so too in this section we will see that we need a
succession of two transformations to achieve a balanced realization. The first coordinate
transformation gives a new state model with controllability Gramian equal to an identity
matrix. The second coordinate transformation keeps the controllability Gramian diagonal and diagonalizes the observability Gramian obtained from the first transformation
so that the resulting Gramians are equal and diagonal.
To develop these transformations we need the fact that, if the coordinates are changed
using coordinate transformation matrix T, then the Gramians (We, W0 )for the state
model in the new coordinates are related to the Gramians ( W c W 0 ) for the state model in
the original coordinates as
(4.29)
where x(t) = Ti(t). These relations can be obtained by replacing (A, B, C) in the
integrals for ~lie and W0 by (T- 1 AT, T- 1 B, CT). We use (4.29) now to develop a the
sequence of three coordinate transformations T; : i = 1, 2 and P so that T b = T 1 T 2 P
transforms the given controllable arid observable state model to a balanced realization,
where Pis the permutation matrix required to put the diagonal elements of the Gramians
in decreasing size. We will show now that these coordinate transformation matrices are
determined from the Gramians for the given state model. In the next section, we will show
that the Gramians for the given state model can be determined by solving two linear
matrix equations known as Lyapunov equations.
For the time being suppose we are given the controllability and observability
Gramians, We, W 0 for the state model we wish to approximate. Then we form the first
coordinate transformation as

r,..~

where v; WcVc = Ee with v; = V;:- 1and E~ > 0 is diagonal with mE~= Ec Then
using (4.29) we see that the Gramians which result from the first coordinate transformation are given as

Next, we take the second coordinate transformation as

Tz

- _!

VoEo4

where V(; W0 V0 = E0 and E~ > 0 with (E~) 4 = E 0 . Then using (4.29) we see that the
Gramians which result from the second transformation are given by

-I

Wo

TiWoTz

and we have the Gramians equal and diagonal.

=E~

106

Model Approxim ation via Balanced Realizatio n

'

It remains to permute the componen ts of the state so that the entries on the diagonal of

the Gramians are ordered in decreasing size. This correspond s to a third coordinate
transforma tion, this time using a permutatio n matrix Pas the coordinate transforma tion
matrix so that

where

and the ub;s are referred to as the Hankel singular values for the system. More is said
about this in Section 4.9
At this stage we can obtain the balanced realization, (Ab, B1, Ch), from
(T- 1AT, T- 1B, CT) with T = T 1 T 2 P. Then partitionin g the balanced realization as
(4.30)

with n 1 large enough so that

n1

i=1

i=n 1+1

I:: ubi I::

ubi

we obtain the reduced order state model as (Ab 1 ,Bh 1 ,Ch 1,D) and the reduced order
transfer function model as

Recall that we are assuming that the full order state model is stable, i.e., Ab is stable.
Therefore we get a bounded output for every bounded input to the full order, unreduced
system. Thus if Gapx (s) is to be a meaningfully reduced order approxima tion to the full
order transfer function, we expect Gapx(s) to be stable. Otherwise the behavior of the
reduced order system would be dramatical ly different from the behavior of the original
full order system. The following theorem specifies conditions which guarantee that the
reduced order model is stable.
Theorem 4.3 If in (4.30) we have ubn 1 'I uhn 1 _,_ 1 then Ab 1 is stable.
Proof Appendix
In the next section we show that the Gramians for a given state model can be
determined by solving linear matrix equations called Lyapunov equations. In this way
the calculation of the integrals in the definition of the Gramians is avoided. We will also
see that further relations between the Gramians for a state model and the observability,
controllability, and stability of that state model are uncovered by considering the
Gramians to be solutions to Lyapunov equations.

The Lyapunov Equation

4.7

107

The Lyapunov Equation

We begin this section by generalizing the output energ~ as follows. Recall that in the
absence of an input the observability Gramian gives the output energy caused by any
initial state as

Now suppose we replace this constant quadratic function of the initial state by e0 (tl, a
time varying quadratic function of the state given as
(4.31)

e0 (t) = x*(t)Px(t)

where Pis a positive definite Hermitian matrix and xit)

function is always positive
for all t

= eA 1x(O).

(0, oo) and all x(O) #

"Kotice that this

(4.32)

Next assume that e0 (t) has a derivatiYe \\hich is a negative quadratic function of the
st;rte
de 0 (t)
_ = -x*(t)Qx(t) < 0
dt

for all t

(0, oo) and all x(O)

(4.33)

where Q is a positive definite Hermitian matrix. Then we see from (4.32, 4.33) that e0 (t) is
a positive, decreasing function of time which is bounded below by zero. This implies that
(4.34)

However since Pis invertible we have e0 (t)

lim x(t)

= 0

= 0 only if x(t) = 0 and (4.34) implies that

for all x(O)

(4.35)

t-+00

Therefore equations (4.31). and (4.33) imply that the state model is internally stable, i.e.,
the state model's A matrix has all its eigenvalues in the open left half plane.
NmY if we differentiate (4.31) and use :\"(t) = Ax(t) we can express the time derivative
of e0 (t), as

e (tl
0

x*(r)(A*P~ PA)x(tl

These observations lead us to formulate the following theorem concerning the stability
of A.
Theorem 4.4 If for any Hermitian Q > 0 we obtain P > 0 as a solution to the linear
matrix equation
A*P+ PA = -Q

then A is stable.

(4.36)

108

Model Approximation via Balanced Realization

Proof Suppose (.A, v) is any eigenvalue, right-eigenvector pair for A. Then pre and post
multiplying (4.36) by v* and v respectively gives
v*(A*P+ PA)v

= -v*Qv

which recalling v* A* = .A*v*becomes

(.A*+ .A)v* Pv = -v*Qv

(4.37)

However since v* Pv > 0 and -v*Qv < 0 for all v # 0, we see that (4.37) is satisfied only
if (.A*+ .A)= 2Re[.A] < 0. This shows that all eigenvalues of A lie in the open left half
plane and therefore A is stable.

The linear matrix equation (4.36) is called a Lyapunov equation after the Russian
mathematician who originated this approach in the study of stability.

4.7.1

Relation to the Gramians

We show next that when we take Q = C*C the solution P to the Lyapunov equation,
(4.36), is the observability Gramian, i.e., W 0 satisfies the Lyapunov equation
(4.38)

Analogously, the controllability Gramian, We, is the solution the Lyapunov equation
(4.39)

The foregoing relations between the Gramians and the solutions to Lyapunov
equations are shown in the following theorem.
Theorem 4.5 IfF is a square stable matrix and if
(4.40)

then

FW+ WF* = -Q

(4.41)

Proof We begin the proof by defining the matrix function

Then we differentiate Z(t) to obtain

Next recalling the series expansion of er 1 , we see thatF* and eF' 1 commute. Therefore

The Lyapunov Equation

109

the foregoing equation can be written as

dZitl

---crt= FZI t) + Z( t)F

(4.4.:2)

Finally, since A is being assumed stable, we hcne Zix) = 0. Therefore integrating

(4.42) from 0 to oo gives

lax dZ(tl =lax [FZ(tl Z(t)Fjdt

Z(oo)- Z(OJ =

F[laoo

-Q=FW+

eF1 QeF' 1dt]

+[lax

eF1 QeF'rdt]F*

wr

No\Y there are two important observations we can make when \Ye compare the
Lyapunov equation (4.36) with the Lyapunov equations for the Gramians (4.38, 4.39).
First, notice that unlike Q on the right side of(4.36), C*C and BB* in (4.38, 4.39) are
only non-negative.
Second, (4.38, 4.39) are linear equations in the elements of the matrices W 0 , We.
Therefore we may be able to solve (4.38, 4.39) for matrices W 0 and rv, when A is not
stable. However since the observability and controllability Gramians (4.5, 4.27) are only
defined when A is stable, the matrices W 0 , We obtained by solving (4.38, 4.39) are system
Gramians only if A is stable.
The implications of these observations are taken up next.

4.7.2

Observabil ity, stability, and the observabil ity Gram ian

What follows are theorems which describe how the concepts named in the title of this
section are interrelated.
Theorem 4.6 If (A. C) is observable and if P > 0 satisfies the Lyapunov'equa tion
AXP+PA = -C*C

(4.43)

then A is stable.
Proof Suppose, contrary to the theorem, that A is not stable, i.e., Av = ,\v with
Re[.A] :::0: 0 for some .A E .A[A]. Then proceeding as in the proof of Theorem 4.4 we obtain
2Re[.A](v*Pv) = -v*C*Cv

(4.44)

Now since (A, C) is observable we have Cv = ~ i=- 0 and L'c~ Cc > 0. However we are
assuming that Re[.A]::;:: 0. Therefore (4.44) is satisfied under these conditions only if we
have~ Pv < 0 which contradicts the condition P > 0 specified in the Theorem. Therefore

our assumption that A is unstable is false.

is
C)
(A,
and
unstable
is
A
when
that
proof,
Notice, in connection with the foregoing
have
we
observable
(4.45)

110

Model Approximation via Balanced Realization

is unbounded and the observability Gramian, W 0 , is undefined. However this does not
mean that we are unable to find P to satisfy (4.43) in this case.
The foregoing theorem has the following corollary.
Theorem 4.7 If A is stable and Pis positive definite and satisfies (4.43) then (A, C) is
observable.
Proof Suppose, contrary to the theorem, that (A, C) is not observable. Then we must
have a(?., v) pair, Av = >.v and Cv = 0. Therefore proceeding as in the proof of Theorem
4.6 we obtain
2Re[>.](v' Pv) = -v*C*Cv = 0

(4.46)

Now in order to satisfy (4.46) we must have Re[>.] = 0 or v* Pv = 0 or both. The

condition Re[>.] = 0 is impossible since we are assuming that A is stable. The condition
v* Pv = 0 is also impossible since we are assuming that P > 0. Thus (4.46) can not be

satisfied and therefore (A, C) must be an observable pair.

To recap, the foregoing pair of theorems state that A is stable when (4.43) is satisfied by
P > 0 if and only if (A, C) is an observable pair. Since the observability Gramian W 0 ,
(4.45) is defined when A is stable and positive definite when (A, C) is an observable pair, if
the solution P to (4.43) satisfies P > 0 when (A, C) is not an observable pair Pis not the
observability Gramian. The following theorem characterizes the effect of having a P > 0
as a solution to the Lyapunov equation (4.43) when the pair (A, C) is not observable.
Theorem 4.8 If the pair (A, C) is not observable and P > 0 satisfies the Lyapunov
equation (4.43) then A has imaginary axis eigenvalues.
Proof Since the pair (A, C) is not observable, we can choose a (>., v) pair so that
Av = >.v and Cv = 0. Then we see from (4.46), that
2Re[>.](v* Pv) = 0

(4.47)

However since P > 0, (4.47) is satisfied only if Re[>.] = 0.

To recap, we have shown in the foregoing theorems that A is stable when the Lyapunov
equation
A*P+ PA

= -C*C

is satisfied by P > 0 if and only if the pair (A, C) is observable. Alternatively, if this
Lyapunov equation is satisfied by P > 0 when (A, C) is not observable, the unobservable
eigenvalues of A lie on the imaginary axis. Notice from the proof of Theorem 4.7 that
eigenvalues of A that are not on the imaginary axis are observable and lie in the left half
plane.
Analogously, A is stable when the Lyapunov equation
AP+PA*=-BB*

is satisfied by P > 0 if and only if the pair (A, B) is controllable. Alternatively, if this
Lyapunov equation is satisfied by P > 0 when (A, B) is not controllable, the uncontrollable eigenvalues lie on the imaginary axis and the controllable eigenvalues are all in the
left half plane.

Controllabil ity Gramian Revisited

4.8

111

Controllability Gramian Revisited

In this section we develop another use for the controllability Gramian which provides us
with a different way of viewing the process of model reduction and a new way of thinking
about a system's input-outpu t behavior.

4.8.1

The least energy input problem

An alternative use for the controllability Gramian, We, arises in connection with the
solution to the problem of finding the input u0 (t) : t E ( -oo, 0] which expends the least
energy in taking the null state at time t = -oo to some specified state at time t = 0.
In order to solve this problem we need to recall, from Chapter 1, that the basic
equation relating the initial state and input to the final state is

x( t1 )

eA 1f x( -oo) +

lf

-oo

eA(tr -T) Bu(T)dT

(4.48)

Then, in the present situation, we have x( -oo) = !Zl, and t1 = 0 so that (4.48) becomes
x(O) =

1:

e-AT Bu(T)dT

(4.49)

Now we want to choose u(t) to satisfy (4.49) so that the input energy, EcUl given by
(4.50)

achieves its minimum value, Ecua, determined from

Ecua

min Ecu
u(t)

It turns out that one input that solves this problem is given as
(4.51)

where We is the controllability Gramian, (4.27). Notice that this input satisfies (4.49) and
has energy determined as
Ecu

1:

u*(T)u(T)dT

(1:

x*(O)W~ 1

x*(O)W~ 1 x(O)

e-AT BB*e-A'TdT ) W~ 1 x(O)

Eeuo

Notice that u0 (t) exists provided We is invertible. This requirement that the (A, B) pair
be controllable is a necessary condition for being able to satisfy (4.49) for any x(O), (see
Chapter 2 Section 2.4).
Recall that the coordinate transformati on matrix, Ve, (4.28), that diagonalizes We is
unitary, V~ 1 = v;. Thus we see that this coordinate transformati on also diagonalizes

112

Model Approximation via Balanced Realization

} I

w;:- 1 since
Therefore we can use this fact to rewrite E,uo as

(4.52)
where

z(O)

v:x(O)

Then if the components of z(O) are equal and have been permuted so the entries along
the diagonal of ~c are ordered, abi 2': ab(i+ 1l : i = 1, 2, n - 1, we see from (4. 52) that
most of the energy in the input is needed to achieve components in the lower partition of
z(O). This implies that the lower partition of z(t) is harder to control than the upper
partition.
Now in the following we make use of the foregoing alternative interpretation of the
controllability Gramian to provide further insight into the use of a balanced realization to
achieve model reduction.
Recall that a balanced realization has both Gramians equal to the same diagonal
matrix with the entries encountered in descending the diagonal forming a nonincreasing
sequence. Then we saw, in Section 4.4.3, that most of the energy supplied to the output by
an initial state with equal component values comes from components in the upper
partition of the initial state. However, we have just seen that the components in the upper
partition of the state are more easily manipulated by the input than components in the
lower partition. Therefore, in balanced coordinates, the subspace of the state space
corresponding to the upper partition of the full state is more involved in the input-output
behavior than the subspace corresponding to the lower partition of the state. This
suggests that we can capture most of the full order system's input-output behavior by
using a reduced order state model formed by truncating the balanced realization in the
manner given at the end of Section 4.6.

4.8.2

Hankel operator

We can obtain further insight into a system's input-output behavior by using the present
interpretation of the controllability Gramian. This is done by noting that:
(i) the controllability Gramian is used to characterize the system's transformation of
u(t) : t E ( -oo, OJ into x(O);
(ii) the observability Gramian is used to characterize the system's transformation of
x(O) into y(t) : t E [0, oo) when u(t) = 0: t E [0, oo).
This suggests that, in addition to the concept of a system implied by its transfer
function, G(s), i.e., that the system transforms or maps system inputs, u(t) : t E [0, oo ), to
system outputs, y(t) : t E [0, oo ), we can view the system as a map from past inputs,
u(t) : t E ( -oo, OJ to future outputs, y( t) : t E [0, oo ). This map is referred to as the

n
'

Controllabil ity Gramian Revisited

113

system's Hankel operator and is denoted as r G When a system is stable its Hankel
operator has an operator norm which is referred to as the Hankel norm. Operators and
operator norms will be developed in more depth in Chapter 7. For the time being we want
to show that the Hankel norm of a system equals the largest entry on the diagonal of the
Gramians when the state model for the system is a balanced realization.
Suppose we are given a strictly proper transfer function, G(s), and minimal state
model, (A, B, C), for some stable system. Then the Hankel operator, r G' for this system
has output y(t) : t E [0, oo) produced by input u(t) : t E ( -oo, OJ when x( -oo) = 0 and

u(t) = 0: t E [O,oo).

Now we define the system's Hankel norm, denoted IIG(s)IIH as

IIG(s)IIH

(max

u(~12..

y*(t)y(t)dtl)~
[It'
tao u*(t)u(t)dt

(4.53)

IE(O,oo)
x(-oo)=l'l

where (-)! indicates the positive square root. There is a close connection between a
system's Hankel norm and that system's balanced realization as shown in the following
t~orem.

Theorem 4.9 The Hankel norm IIG(s)IIH.'for a stable system having transfer function
G(s) is given as
(4.54)
where ab 1 is the largest entry on the diagonal of the Gramians in balanced coordinates.
Proof Recalling (4.29) we see that the eigenvalues of the product of the controllability
and observability Gramians are invariant to a coordinate transformati on since we have

ww
c

r- 1 wc r-rw

r-'wc w r
0

(4.55)

Thus when Tis the coordinate transformati on which puts the system in balanced form
we have

where {ab; : i = 1, 2, n} are referred to as the Hankel singular values.

Next recall from Section 4.4.3 that when the state model is a balanced realization, the
~aximum output energy, Eomax> produced ?Y ~n initial state, z(O), constrained to satisfy
z (O)z(O) =a , when u(t) = 0: t E [0, oo ), 1s giVen by
Eomax

[Eo]=
= max
"(0)

[z",.x(o)] *:Eb[zmax(O)] = a 2ab 1

z' (O)z(O)=a2

where
[~ax(O)f =[a

OJ

114

Model Approxim ation via Balanced Realizati on

Alternatively, recalling the solution to the least energy input problem, we see from
produce
(4.52) that using the least energy input u0 (t) : t E ( -oo, 0] when z( -oo) = 0 to
zmax(O) gives the least energy as
)

E cmin

[Ecu ]
mm

[z max ( 0 )] * :Eh-I z max ( 0 )

ull) IE( -x.O]

Cl=(}hI

.:;(-'X)=O

:(Oj=:max (O)

Therefor e we see that the maximum value of the ratio of energy ~ut to energy in is given as
Eomax

,2

(Tbi =/\max

we wo ]

cmm

which from (4.53, 4.55) yields (4.54 ).

4.9

Notes and References

Hermitia n matrices play an importan t role in many branches of science and engineering,
e.g., quantum mechanics. In general a square matrix having repeated eigenvalues may not
be able to be diagonalized by a similarity transform ation, i.e., we may not be able to find
V such that v-I MV is diagonal. However a Hermitia n matrix can always be diagonalized
even if it has repeated eigenvalues, [39].
The term "Gramia n" arose originally in connectio n with the problem of determining if
the vectors in a set of vectors {v; : i = 1, 2, , n} were linearly independent. The
Gramian G was defined as the scalar whose value was determined as G = det[M],
where (M)u = v; *v1. The independence of the v;s required G to be nonzero, [5]. As
used in system theory, the term Gramian describes a matrix, not a scalar, which is formed
in manner reminiscent of the foregoing matrix G.
Balanced realization was first used in the area of digital filters to combat inaccuracies
(roundoff) resulting from computa tion errors brought about by the requirem ent that
calculations be done using finite precision arithmeti c [29]. The use of a balanced
realization to obtain a reduced order model approxim ation and the method given here
for calculating a balanced realization was originally presented in [27]. Theorem 4.8 and
the proof of Theorem 4.3 (Appendix) were originally given in [34]. The idea of a balanced
realization was extended to unstable linear systems, [22] and to a class of nonlinea r state
models, [38].
Finally, we need to take note of the following importan t aspect of model reduction. We
will see in the remainde r of this book that optimal controller design techniques, e.g.,
quadratic , Gaussian , and H x feedback control theories, make use of an observer as part
of the feedback controller. Since the observer order matches the plant order, there can be
problems in implementing controllers for use with high order plants. Unfortun ately, the
stability of the closed loop system may not be preserved when the controlle r designed
using the optimal control theory is replaced by a reduced order model. A good summary
of approach es to this problem up until 1989 is given in [2]. A more recent approach to
solving this problem, involving H X! robust control theory is given in [30] for linear plants
and in [32] for a class of nonlinea r plants.

5
Qua drat ic Control

5.1

Introduction

In the preliminar y part of this chapter. we introduce the basic approach to feedback
control known as observer based feedback control. In this scheme the plant state x(t) in
the state feedback control signal u(t1 = Kx(t) + v(r) (Chapter 2) is replaced by the
estimated plant state i(r) obtained from an observer (Chapter 3). As we will see in the
remainder of this book, this scheme has enabled the developme nt of several optimizati on
based controller design techniques . In addition to quadratic control treated in this
chapter, we will see that observer based feedback control is used in linear quadratic
Gaussian, LQG, control and in HX! control.
Following this, we take up the use of quadratic control to determine the K and L
matrices in an observer based feedback control setup. Quadratic control arises\\ hen there
is a need to maintain the control system's steady-stat e output constant in response to a
constant control input. In this situation the control scheme is said to operate as a
regulator. In practice, impulse like disturbanc e signals, i.e .. signals having large amplitude and short time duration. enter the control loop and cause the output to jiggle about
its desired level. A well known example of this is the regulation of the pointi~g direction of
an outdoor dish antenna. Unfortuna tely the dish acts like a sail and experience s torque
disturbanc es because of wind gusts which tend to change its pointing direction. Therefore
we need to choose K and L so as to alter the closed-loo p system dynamics. i.e .. assign the
eigenvalue s of the closed-loo p system matrix, in a way which reduces the effect on the
output of these impulse like disturbanc e signals. In quadratic control this choice of K and
Lis done so as to minimize the energy in the output signal caused by the disturbanc e.
We will see that the determina tion of the optimal K and L matrices needed to achieve
this minimum requires the solution to two nonlinear matrix equations known as algebraic
Riccati equations. Special solutions to these equations, referred to as stabilizing solutions
are required. We will develop conditions on the plant and optimizati on criteria which will
ensure the existence of such solutions. We will encounter the Riccati equation again in
connection with other optimal control schemes developed in this book. More specifically,
w_e will see that the algebraic Riccati equation is invohed in the solution of the
di~turbance attenuatio n problem when the disturbanc es are viewed as rand::>m signals
With known statistics (Chapter 6) and when the disturbanc e signals are restricted to the
so-called L 2 class of signals (Chapters 9 and 10).

116

5.2

Quadrati c Control

Obser ver Based Contro llers

Although errors in the estimated plant state in an observer based control system produce
effects at the control system's output, we will see that the dynamics of the observer do not
contribut e to the control system's zero-state response. More specifically, we will show
that the control system's transfer function equals the transfer function for the closed loop
feedback system when exact state feedback is used, i.e., when the estimated state equals
the plant state.
Suppose the plant state model paramete rs (A, B, C, D) are given. Then recalling the
development of the observer, (Chapter 3), we see that the-plant state, x(t), and the
observer state, x(t), are governed by the following vector differential equation s

x(t)

Ax(t)

+ Bu(t)

x (t) =(A+ LC)x(t)

(5.1)

+ (B + LD)u(t) - Ly(t)

(5.2)

Now recall from Chapter 2 that to apply state feedback we set u(t) = Kx(t) + v(t).
a
Lacking exact knowledge of the plant state we replace it by its estimate, .X( t), to get
control input

u(t) = Kx(t)

+ v(t)

(5.3)

Then substitut ing (5.3) in (5.1, 5.2) and the plant's output equation , we see that the
control system is a 2n dimensional system having state model given as

Xc(t) = AcXc(t) + Bcv(t)

(5.4)

+ Dcv(t)

(5.5)

y(t) = Ccxc(t)
where

Ac =

BK

[A

- LC A + LC + BK

Cc = [C

v(t) +

x(t)]
Xc(t) = [ x(t)

DK]

u(t)

y(t)

PLANT

STATE
FEEDBACK

x(t)

OBSERVER

1<---

l
Figure 5.1

Setup for observer based feedback control

Obsen/er saseci cOiitroriers,

111

sion n, the closed-loop

Notice that since the plant and observer are each of dimen
state mo~el is minimal,
plant
the
if
even
ver
system state model is of dimension 2n. Howe
, (5.4, 5.5), IS not controli.e. observable and controllable, the closed-loop state model
the resulting state model is in
lable. We can see this by changing coordinates so that
_
as follows.
controllable decomposed form, (Section 2.5). We do this
to
Be
matrix
input
ormed
Recall that controllable decomposed form requires the transf
a
,
halves
equal
oned into two
have a lower partit ion which is null. Since Be is partiti
given by
coord inate transf ormat ion matrix, T, which does this is

the contro l system state

Therefore, after using this coord inate transf ormat ion matrix
model parameters, (5.4, 5.5), become
Ac =

[Acl ~c2]
0

Ac4

Cc = [C+D K

x(t)
x(t)

(5.6)

-DK]

where
Aci =A+ BK

Acz = -BK

Ac4

= A+L C

x(t) = x(t) - x(t)

plant state estimation error,
Notice, from the null block structure of Ac and Be, that the
er function is given as
x(t), is uncontrollable and that the contro l system transf
Gc(s)

= Cc(s l- Ac)-I Be+ De

= [C+D K -DK ][(sl -Acd -I
0

(C + DK)[ sl- (A+ BK)r 1B + D

*_
(s/- Ac4 )-

I]

[B] +D

0
(5.7)

i.e., (A, B) controllable and

Therefore provid ed the plant state model is minimal,
(A + BK, C + DK) observable
(A, C) observable, we have (A + BK, B) controllable and
ing transfer function and the
forego
the
in
s
so that there are no pole-zero cancellation
e this contro l system transfer
contro l system and plant have the same order. Notic
we used exact state feedback.
function is the same as we obtain ed in Chapt er 2 when
outpu t or zero-state behavior
!here fore we can conclude that the contro l system's inputer. However, the contro l system's zero-input
IS unaffected by the presence of the observ
follows.
response is affected by the observer. We can show this as
the differential equations
that
see
we
(5.6),
l,
mode
state
's
system
From the contro l

118

Quadratic Control

governing the plant state and plant state estimation error are
~\'(t) = A,lx(t)

.X (t)

+ Ac2~'X-(t) + Bv(t)

(5.8)

= A, 4 ~x( t)

Then the state estimation error is independent of the input and depends only on its initial
value as

x(t)

e(A+LC)tx(O)

(5.9)

Therefore substituting this expression for .X(t) in (5.8) yields

x(t) =(A+ BK)x(t)- BKe(A+LC)rx(O)

+ Bv(t)

(5.10)

and we see that the evolution of the plant state depends on the state estimation error.
However, this effect which is caused by the presence of the observer in the feedback
loop, disappears with time. More specifically, recall from Chapter 3 that L must always be
chosen to make A+ LC is stable. Therefore we see from (5.9) that
lim.X(t) = 0

for any .X(O)

(-'t(X;-

Thus, for some time following the initiation of the feedback control, the plant state,
x(t), and plant output, y(t), are both affected by the plant state estimation error, .X(t).
Eventually the term A, 2 x(t) can be neglected in comparison to other terms on the right
side of (5.8) so that for large t the derivative of the plant state can be approximate d as

x(t) =(A+ BK)x(t)

+ Bv(t)

(5.11)

Notice that (5.11) is the differential equation for the plant state when exact state feedback
is used.
Finally, recalling that the eigenvalues of the system matrix A, are preserved under
coordinate transformati on, (Section 1.6.4), and that block upper-triang ular matrices
have eigenvalues which are the union of the eigenvalues of the diagonal blocks,
(Appendix), we see from (5.6) that

>.[A,] = >.[A,] = >.[A + BK] u >.[A + LC]

Thus half the closed loop system eigenvalues depend on K and are independent of L, and
the other half depend on L and are independent of K. This spliting of the closed loop
eigenvalues is responsible, in part, for being able to break the optimal control problems
treated in this book into two simpler problems:
1. the state feedback control problem
2. the state estimation problem
This fact, referred to as the separation principle, has been a major theme running through
the development of linear control theory since its discovery in the 1960s.

Quadra tic State Feedba ck Control

119

state
In summar y, the use of an observer-estimated plant state in place of the true plant
having
system
control
k
feedbac
a
to
leads
ration
in a state feedback control configu
as we would
transfer function (steady-state. input-ou tput behavior) which is the same
system,
control
the
of
start-up
the
at
initially
r.
Howeve
.
feedback
obtain using exact state
the
affect
will
state
plant
true
the
from
state
ted
r-estima
observe
any departu re of the
state
ted
control system output. This effect diminishes \Vith time as the observer-estima
approac hes the true plant state.
ues for
Recall, from Chapter s 2 and 3. that we needed the specification of eigenval
control
k
A + BK and A + LC in order to determine K for the design of the state feedbac
, we are only
and L for the design of the state estimato r, respectively. However, in practice
re, in this
given some desired goals for the behavio r of the control system. Therefo
control
the
situation, we need methods for turning specified goals for the behavio r of
and in
system into values for the K and L matrices. In the remaind er of this chapter,
for
goal
subsequ ent chapters , we will show in each case that by interpre ting the specified
develop
can
the control system's behavio r as a mathem atical optimiz ation problem , we
equation s for determi ning the required K and L.

5.3

Quadratic State Feedback Control

mandato ry
There are many possible control system design objectives in addition to the
ents arises
requirem
al
addition
these
of
One
requirem ent that a control system be stable.
control
the
entering
signals
nce
disturba
ration
when there are large-amplitude, short-du
the
design
to
need
we
n
situatio
this
In
t.
constan
loop and we want to maintai n the output
,
possible
as
much
as
,
diminish
to
and
system
controller both to stabilize the closed-loop
optimal
the
to
solution
The
output.
plant
the
on
the effect of the disturba nce signals
L both to
quadrat ic control problem attempt s to achieve this goal by determining K and
control
the
of
re
departu
the
in
energy
the
e
minimiz
stabilize the closed-loop system and to
\\e
section,
s
previou
the
in
ed
mention
As
value.
tate
system output from a desired steady-s
lems:
subp.rob
simpler
two
into
problem
this
can use the separati on principle to split
l. the quadrat ic state feedback control problem and
2. the quadrat ic state estimati on problem
with the solution to 1. determining K, and the solution to 2. determining L.
ic state
In this section we develop the design equation s forK which solves the quadrat
of the
sum
the
s,
equation
feedback control problem . When K satisfies these design
ed.
minimiz
is
signal,
nce
energies in the output and feedback signal caused by the disturba
ic
quadrat
the
as
known
One of these design equation s is a nonline ar matrix equation
to
needed
und
backgro
cal
theoreti
control algebraic Riccati equatio n (QCAR E). The
solve this equatio n is provide d in the next section, Section 5.4.
ic state
!n S~ction 5.5 we develop the design equation s for L to solve the quadrat
c
algebrai
an
of
solution
the
involves
this
es!Imatlon problem . Again we will see that
the
as
known
is
and
QCARE
the
to
related
Riccati equation . This equatio n is closely
to the
quadrat ic filtering algebraic Riccati equatio n (QF ARE). The required solution
the
solving
for
5.4
Section
in
ed
develop
res
procedu
using
QFARE can be obtaine d
~
QCARE .

Quadratic Control

120

5.3.1

Motivating the problem

Let the disturbance input to the plant be denoted as u 1 ( t); let the controlled input involved
in implementing the state feedback be denoted by u2(t). Then the plant state model is
given by the equations

x(t)

Ax(t)

+ B1 u1 (t) + B2u2(t)

(5.12)

y(t)

Cx(t)

+ D 1u 1 (t) + D 2 u2 (t)

(5.13)

and setting u2 (t) = Kx(t) + v(t), where v(t) is the command input, we obtain a state
feedback control system having state model given as

x(t)

Ax(t)

+ Bl

+ B2v(t)

(5.14)

y(t)

Cx(t)

+ D 1u 1 (t) + D 2v(t)

(5.15)

Uj

(t)

. where

disturbance input
controlled input

y(t)
v( t)

control system output

control system command input

However, if
(i) the disturbance input is null, u 1 (t) = 0
(ii) the control system command input is constant, v(t) = v5 ,
(iii) the feedback matrix K makes the control system stable so that all eigenvalues of A lie
in the open left half plane,
then the constant steady-state control system state, x" is obtained from the control
system state differential equation, (5.14) by setting x(t) = 0. Then using x, in (5.15) gives
the steady-state control system output y 5 as

u,(t)

PLANT
v(t)

y(t)

u,(t)

I+--CONTROLLER

Figure 5.2

If

Setup for disturbance input attenuation

Quadratic State Feedback Control

121

where
Xs =

limx(t)

t-x

Now suppose the control system is operating at this steady-state level at time t = t 1
when a short-duration disturbance signal, lasting T seconds, arrives at the disturbance
input. Then the disturbance input can be written as
u 1 (t)

for t E h

t1

+ Tj

otherwise

=o

Therefore, shifting the time origin to the instant immediately following the end of the
disturbance, i.e., to time t 1 + T. we are left \vith the state immediately following the end of
the disturbance being given by

However since A

+ BK is assumed stable we have

lim xd(t) = o
{-H:X!

and the state of the plant returns to its steady~state Yalue with time,
limx(t) = x,
{-H)O

Our goal in choosing K is to decrease the effect of the disturbance input on the control
system output during the return of the output to its steady-state value. Since superposition applies. we can assume, without loss of generality, that the control system
command input is null, v( t) = vs = 0. Then the desired steady-state output is null and any
deviation of the output away from being null is caused by the presence of the disturbance
input. Thus we want to choose K so that the action of the feedback signal attenuates these
deviations as much as possible.

5.3.2

Formulatin g the problem

Recall, from the previous chapter, that we used the observability Gramian and its relation
to the output energy to obtain a measure of the zero-input response. We use this idea here
to reformulate the problem just discussed as the problem of choosing K to minimize the
~ffect of a non-null initial plant state on the sum of the energy in the output and the energy
In the feedback signal, i.e.,

where
lQc =

laoo (y'(t)y(t)- pu2(t)u2(t))dt

la

00

[x*(t)C*Cx(t)

+ u;(t)[pJ]u 2 (t)]dt

(5.16)

122

Quadrati c Control

with p being a positive scalar and JQc being referred to as the cost or the performa nce
index for the quadratic control problem.
Notice that the reason for including the feedback signal energy in the performa nce
t)
index, i.e., for including u 2 ( t) in the integrand , is to limit the feedback signal u 2 ( t) = Kx(
given
is
this
of
on
explanati
detailed
more
A
to a level that can be physically implemented.
as follows.
When we choose K to minimize J QCo we discover that the smaller p the larger the
elements inK. If we were to let p be zero so that there would be no contribut ion to the
to
cost, lQc, from the feedback control signal, u2 (t), we would find that the K required
feedback
required
the
case
this
in
Thus
entries.
minimize lQc would have unbound ed
control signal u 2 ( t) would be an impulse. This physically unrealizable requirement
prohibits the use of p = 0. However, while increasing the size of p has the beneficial
on
effect of reducing the size of the required feedback signal, the effect of the disturban ce
between
ise
comprom
a
pis
of
size
the
for
choice
the output is increased. Thus a beneficial
limiting the size of the control signal, u2 ( t) and limiting the effect of the disturban ce on the
outpu~.

In order to simplify the development to follow, we will assume that there is no direct
feed through of the feedback controlle d input u 2 ( t) to the plant output, i.e.,
we assume D 2

=0

In addition, since it is always possible to choose coordina tes for the state space so that
the paramete rs of the state model are real, we will assume that these paramete rs are real.
Finally, for generality in what follows, we will replace C* C and pi in (5.16) by Qc and
Rc respectively so that the performa nce index becomes
( 5.17)
all
where Qc, Rc are real, symmetric matrices with Qc ?: 0 and Rc > 0 implying that for
x(t) and u(t) we have

x*(t)Qcx(t)?: 0
u).(t)Rcu2(t) > 0
These restrictions on Qc and Rc are needed to ensure that the contribut ions to lQc from
both u2 (t) and x(t) are never negative and so that R, is invertible.
In summary , we have formulate d the problem of designing a state feedback controlle r
to combat the effect of unknown short-du ration disturban ce inputs as an optimiza tion
problem. The solution to this optimiza tion problem requires that K be determin ed so that
is
a quadratic performa nce index in the disturbed state and the controlle d input
minimized.

5.3.3

Develo ping a solution

From the discussion earlier in this section, we suppose that the comman d input is null,
has
v( t) = 0, and that the time origin is taken at the instant a short-dur ation disturban ce

Quadra tic State Feedba ck Contro l

just ended. This leaves the plant state at some non-null value, x(O)

have the plant state given by

= xd

123

and thereafter we

( 5.18)

and the controlled input given by

(5.19)

Thus using (5.18, 5.19) in the quadratic control cost, (5.17), yields
( 5.20)

where
Qc = Qc + K*RcK

J. =A +B2K
-

I
integral, (5.20),
Now since we are assuming that K is stabilizing, i.e., A is stable, the
converges to a real symmetric matrix PQc,
( 5.21)

and the performance index, (5.20), can be written as

(5.22)

ov AUation
Recall from Chapte r 4 that PQC (5.21), is the solution to the Lyapun
(5.23)

to this Lyapunov
Therefore, since A, Qn (5.20), depend on K, so does the solution, PQc,
writing PQc and
by
fact
this
denote
We
equatio n and the performance index, JQc, (5.22).
JQc as PQc(K ) and JQc(K, xd) respectively.
From the foregoing we conclude that our goal is to find K such that
(i) K is stabilizing, i.e., A + B 2 K is stable, and
all initial states, xd .
(ii) the solution, PQc(K ), to (5.23) minimizes JQc(K, xd), (5.22), for
satisfies (i) and
.. In the following development we use K 0 to denote the value of K which
(u). Thus K 0 satisfies
(5.24)

for all disturbed initial states, xd

124

Quadratic Control

Now we are going to use a matrix variational approach in connection with the
Lyapunov equation, (5.23) to develop an equation for P QCo, where P QCo denotes the
matrix PQc which solves (5.23) when K = K 0 We begin by noting that any real and
stabilizing K can be written as
( 5.25)
where Eisa positive real scalar and b K is a real matrix having the same dimension as K
with E(bK) being referred to as a perturbation on K,n the desired value forK.
Since P QC depends on K through the Lyapunov equation, (5.23), the perturbation,
E(bK), on K 0 produces a series expansion for PQc about its optimal value PQco,
( 5.26)
where HOT (higher order terms) stands for terms in the expansion involving powers of E
which. are 2 or greater. The term (b1P QC) is a symmetric matrix which depends on (b K) in
a manner to be determined.
Next we substitute the expansion of PQc, (5.26), into the expression for the cost lQc,
(5.22). This gives an expansion for lQc about its minimum value lQco
(5.27)
where

and again HOT stands for terms in the expansion involving powers of E which are 2 or
greater.
Now since the terms indicated by HOT in both (5.26) and (5.27) depend on
{Ei: i=2,3,}weseethat
lim[HOT]
E-->0

( 5.28)

Therefore it follows that the scalar, (b 11Qc), in (5.27) and the matrix, (b 1PQc), in (5.26)
can be expressed as
(5.29)

( 5.30)
Notice that (5.29) is reminiscent of the derivative in calculus. This fact together with
ideas we encounter in using calculus to obtain minima of a scalar function of a scalar
variable enables us to visualize the present problem of minimizing a scalar function J QC of
a matrix K as follows.

Quadratic State Feedback Control

125

Suppose, for simplicity we assume u2 ( t) is a scalar. Then K is a row vector with n scalar
entries. Then we can think of JQc together with the components of K as forming ann+ 1
dimensional space. Imagine making a plot of the scalar JQc on the "vertical axis" vs. K
with K being represented by a point in a "horizontal" hyperplane, or plane when n = 2.
The resulting surface in this space would appear as a bowl with the bottom of the bowl as
the "point" {JQCo K 0 }, i.e., the optimal point. At this point (8tJQc), (5.29), is zero both
for all8 K and for all initial disturbed states xd. This is analogous to the derivative of a
scalar function of a scalar variable being zero at points where the function has a local
minimum.
Notice from the dependency of (8 1JQc) on (8 1PQc), (5.27), namely

that (81J QC) is zero for all xd if and only if (81PQC) is a null matrix. Therefore the optimal
value forK, K 0 , makes (8 1PQc) null for all (8K) which maintain K, (5.25), stabilizing.
Therefore in order to determine K 0 we need to determine an equation for the matrix
(8 1PQc) which involves (8K). We can do this by substituting the expansions forK and for
PQc (5.25, 5.26), in the Lyapunov equation, (5.23). The results of doing this are
detc;rmined as follows.
begin by denoting the result of using the expansion of K, (5.25), in A and Qc, (5.20),
as

We

A= A + EB2(8K)
0

Qc = Qco + E[(8K)* RcKo + K~Rc(8K)] +HOT

(5.31)

where

Then substituting these expressions for A, Qc as well as the expansion of P Q~' (5.26), in
the Lyapunov equation, (5.23), yields
A~PQCo + PQCoAo + Qco + E(M + M*) +HOT= 0

(5.32)

where

Notice that the first three term~ in (5.32) sum to a null matrix since PQco satisfies the
Lyapunov equation, (5.23), when K = K 0 Thus (5.32) becomes

E(M + M*) +HOT = 0

(5.33)

Then dividing (5.33) through byE, letting Ego to zero, and using the limiting property,

126

Quadratic Control

(5.28), yields
M

+ M* = 0

Finally, substituting for M from (5.32) yields the following Lyapunov equation relating
(8 1PQc) to (8K)
(5.34)

Recall, from discussion earlier in this subsection, that K = K 0 when

matrix for al18K which maintain K, (5.25) stabilizing. Therefore when
matrix, (5.34) becomes

(8 1PQc)
(8 1PQc)

is a null
is a null

( 5.35)

which is satisfied for all stabilizing perturbations, 8 K, only if M 1 is a null matrix.

Therefore we see from the definition of M 1, (5.32) that
(5.36)

where we have used the assumption Rc > 0 to insure that Rc is invertible.

Notice, at this stage of the development, that the state feedback matrix K 0 is readily
determined from (5.36) once we have P QCo It turns out that P QCo is a solution to a
nonlinear matrix equation known as the quadratic control algebraic Riccati equation
(QCARE). This equation can be developed by setting K = K 0 , (5.36), in the Lyapunov
equation (5.23). After carrying this out, we get the QCARE as
A* P QCo

+ P QCoA -

P QCoRQcP QCo

+ Qc = 0

(5.37)

where

Having determined the QCARE it remains to solve it for PQca such that A+ B2Ka is
stable when PQco is used to determine K 0 from (5.36).
We can recap this section as follows. We began by introducing a matrix variational
approach and using it to determine an equation for the optimal feedback matrix, Ka,
(5.36). This equation involves P QCo the matrix which solves the Lyapunov equation,
(5.23), when K = K 0 . Then we saw that by eliminating K 0 from the Lyapunov equation
by using (5.36) we obtained a nonlinear matrix equation in PQCo referred to as the
QCARE, (5.37).
In the next section we will be concerned with conditions on the plant state model
parameters and the parameters of the performance index, Rc and Qc, which ensure the
existences of the so-called stabilizing solution, P QCo to the QCARE, i.e., the solution
which makes A- B 2 R; 1B* PQco stable.

127

Solving the QCARE

5.4

Solving the QCARE

We have just seen that the determin ation of the state feedback matrix K 0 , (5.36), which
g
minimizes the performa nce index JQc, (5.17), requires the determin ation of a stabilizin
establish
to
is
section
present
the
of
goal
main
The
(5.37).
,
solution to the QCARE
condition s on the plant state model paramete rs and weighting matrices Qc and Rc which
ensure the existence of a stabilizing solution to the QCARE. We will be aided in this task
by resorting to a 2n x 2n matrix called the quadratic control Hamilton ian matrix. Before
doing this we consider aspects of the closed loop stability which are brought out using the
fact the PQco is the solution to the Lyapuno v equation , (5.23), when K = K 0

5.4.1

Stabilizing solutions

Consider the following theorem and its proof.

Theorem 5.1 A0 is stable if
(i) (A, B2 ) is a stabilizable pair and
(ii) P QCo satisfies the QCARE with P QCo 2': 0 and
(iii) (A, Qc) is a detectabl e pair.
wh(;e A0 =A+ B2 K 0 with K 0 given by (5.36) and PQco satisfying (5.37).
Proof The necessity of (i) is obvious since we have, from Chapter 2, that a pair (A, B) is
we
stabilizable if we can find K such the A+ BK is stable. Therefor e assuming (i) holds,
stable.
be
to
A
for
sufficient
0
will show that (ii) and (iii) together are
Notice from (5.36, 5.37) that

Therefor e adding and subtracti ng PQcoRQcPQCo on the left side of the QCARE , (5.37),
gives the Lyapuno v equation , (5.23), with K = K 0 ,
(5.38)
where

We proceed now to use the ideas from Chapter 4 concernin g stability in connectio n with
the solution of Lyapuno v equations .
First we rewrite (5.38) by expandin g Qc0 , (5.31), using K 0 , (5.36). This enables (5.38) to
be rewritten as
(5.39)

~en pre and post-mul tiply this equation by

Yields

vi* and vi respectively where

A vi

= >..vi

(5.40)

128

Quadratic Control

Next recalling that R, > 0 and that Q, 2: 0 we have

(5.41)
and the right side of (5.40) is either negative or zero, i.e., either
(a)
or
(b)

Suppose (a) holds. Then we see from (5.40) that

(5.42)
and neither Re[A;] nor vi* PQcoVi can be zero. Therefore we see from condition (ii) in the
theorem that v'* PQcov' > 0. Thus (5.42) is satisfied only if Re[A;] < 0.
Alternatively, suppose (b) holds. Then we see from (5.41) that both
and
This implies that
and

(5.43)

and we see that

'
i
/\;V

A~ o V ;

(A

1
B 2 Rc B*P
2 QCo ) V ;

AV i

(5.44)

Thus (A;, vi) is an eigenvalue, right-eigenvector pair for A satisfying Qcv; = 0 implying
that\ is an unobservable eigenvalue for the pair (A, Q,). However (A, Q,) is detectable,
condition (iii). Therefore if (b) holds we have Re[\] < 0.
To recap, since the right side of(5.40) is either negative or zero and we have shown that
in either case Re[A] < 0 when the conditions in the theorem hold. Thus we have shown
that the conditions in the theorem are sufficient for A0 to be stable.

We will show in the next subsection that condition (iii) in the statement of the
foregoing theorem is not necessary for A0 to be stable. Instead, we will show that only
condition (i) in Theorem 5.1 and the condition that any imaginary axis eigenvalues of A
be observable eigenvalues for the pair (A, Qc) are needed to ensure the existence of a
stabilizing solution to the QCARE. Some appreciation of this weakening of condition (iii)
can be obtained by reconsidering the foregoing proof as follows.
First, notice that when case (b) in the proof holds, (5.40) implies that

Solving the QCARE

129

and either
(bl): Re[.\] = 0
or
(b2): PQcoV; = 0

However, recall from the proof that when case (b) holds the eigenvalue).., of A0 is also an
unobservable eigenYalue for the pair (A, Q,). Therefore if we impose the condition that
(A, Qc) has no unobservable imaginary axis eigenvalues, then case (bl) is impossible.
Second, though not apparent at this stage in the discussion. we will see in the next
subsection that if both condition (i) in Theorem 5.1 and the condition that (A, Qc) haYe
no unobservable imaginary axis eigenvalues are satisfied, then eigenvectors of A can only
satisfy the condition for case (b2) to apply if the corresponding eigenvalue of A is stable.
Finally, notice in the proof of Theorem 5.1, that case (b) is impossible when the pair
(A, Qcl is observable and thus PQcu > 0 \\hen (A. Qc) is obseryable. Again, though not
obvious now, we will show that we can replace this condition for P QCo > 0 in case (b) by
the condition that (A, Q,) have no stable unobservable eigenvalues. Therefore this
condition together with condition (i) in Theorem 5.1 and the condition that (A, Q,)
haye no imaginary axis eigenvalues are the necessary and sufficient conditions for
PQco > 0.

The uniqueness of the stabilizing solution is shown in the following theorem.

Theorem 5.2 The stabilizing solution to the QCARE is unique, when it exists.
Proof Suppose P QCI and P QC are two stabilizing solutions to the QCARE. This
enables us to write

+ Qc = 0

A*PQcl

+ PQClA-

A* PQc2

+ PQc2A- PQc2RQcPQc2 + Q, = 0

PQciRQcPQc!

where A; = A - RQcP QCi : i = 1, 2 are each stable.

Next taking the difference between these two equations yields
A*(b.PQc)

+ (b.PQc)A- PQclRQcPQcl + PQc2FQcPQc2 =

where (b.PQc) = PQCI- PQc 2.

Then adding and subtracting P QCl RQcPQc 2 enables this equation to be rewritten as
(5.45)
Now pro_cee~ing in the same manner as was done to prove Theorem 4.5, we can shO\v
that when A 1 , A 2 are stable the matrix equation
"4i(~PQc)

+ (!:::,.PQc)A2 =

has solution !:::..PQc which can be written as the integral

.\1

130

Quadratic Control

However since the right side in (5.45) is null, M is null and the integrand of the
foregoing integral is null for all time. Therefore we have !::,.PQc = 0 which implies that

PQci = PQc 2 and the stabilizing solution to the QCARE is unique.

Having examined certain properties of stabilizing solutions to the QCARE, we turn
next to questions concerning the existence and calculation of such solutions. An
important tool for answering these questions is the Hamiltonian matrix.

5.4.2

The Hamiltonian matrix for the QCARE

In this section we will show the relation between the solutions of the QCARE and
properties of a 2n x 2n matrix made up from the parameters of the QCARE, (5.37). This
matrix, called the quadratic control Hamiltonian matrix, is defined as
HQc =

[AQ,

( 5.46)

where RQc = B 2 K; 1B;.

Notice that the blocks of H QC depend on all the matrices in the QCARE, (5.37), except
its solutions P QC Recall that we are interested in the stabilizing solution denoted P QCo In
the following theorem, we give a property of H QC which is important in establishing
conditions for the existence of a stabilizing solution to the QCARE.
Theorem 5.3 It is not possible to find a stabilizing solution to the QCARE, (5.37), if
H Qc, (5.46), has imaginary axis eigenvalues.
Proof Define the 2n x 2n matrix T as
(5.47)

where P QC is any Hermitian matrix. Notice that Tis invertible for all Hermitian matrices,
P QC, having inverse given by

r-I =

[P~c ~]

Next let HQc be related to H QC as

-HQc =

T -1 HQcT =

[Az-

R-- QAc*

where
A= A- RQcPQc
Z =A* PQc

Notice that

+ PQcA- PQcRQcPQc + Q,

( 5.48)

Solving the QCARE

131

independent of P QC Therefore if we choose P QC so that Z is null, H QC is block uppertriangular and

(5.49)
>.[HQc] =>.[A] u >.[-A']
However since complex eigenvalues of A occur in conjugate pairs, we have
A[A*] = >.[A]. This together with the fact that >.[-A] = ->.[A] allow us to rewrite (5.49) as
>.[HQcl =>.[A]

u -A[A]

(5.50)

This shows that the eigenvalues of H QC are mirror images of each other across the

imaginary axis, i.e., if 'TJ E >.[H Qcl then-ryE A[HQcl

Now when we choose PQc so that Z is null, PQc is a solution to the QCARE, (5.37).
Moreover, recalling the expression for the feedback matrix K 0 , (5.36), we see that A,
(5.48) can be rewritten as

when PQc = PQco, the stabilizing solution to the QCARE. However, if HQc has an
imaginary axis eigenvalue, (5.50) shows that this eigenvalue must also be an eigenvalue
for Xfor z!l Hermitian matrices PQC which satisfy the QCARE. Therefore it is impossible

to choose P QC to satisfy the QCARE, i.e., to make Z null, so that A is stable.

To recap, the 2n eigenvalues of H QC split into two sets of n eigenvalues each such that
each eigenvalue in one set has a corresponding eigenvalue in the other set which is its
mirror image across the imaginary axis. Thus if PQC is a stabilizing solution to the
QCARE, the eigenvalues of A, (5.48), equal n of the eigenvalues of H QC which are in the
open left half plane. However, if H QC has imaginary axis eigenvalues, H QC does not have
n eigenvalues in the open left half plane. Therefore, in this situation, a stabilizing solution
to QCARE does not exist since the eigenvalues of A must include at least one imaginary
axis eigenvalue of HQc for all solutions, PQc, to the QCARE.
We need now to establish conditions which ensure that H QC does not have imaginary
?~
axis eigenvalues. This is done in the following theorem.
. Theorem 5.4 The quadratic control Hamiltonian matrix H QC (5.46), is devoid of
Imaginary axis eigenvalues if
(i)
(ii)

(A,,RQc) has no uncontrollable eigenvalues on the imaginary axis and

(A, Qc) has no unobservable eigenvalues on the imaginary axis

Proof Suppose HQc has an eigenvalue atjw. Then we have

(5.51)

HQcv=jwv

[ A .

Qc

R.Q: ] [ v 1 ] = jw [ v 1 ]
-A

Vz

Vz

or
(jwl - A )v 1 = RQcVz
(jwl + A*)vz

Qcvi

(5.52)
(5.53)

Quadratic Control

132

Next, we premultiply (5.52) by

v2 and (5.53) by vr

to obtain

v}(jwl- A)v 1 = v.2RQcv2

v[(jwl

+ A*)v2

v[Qcvl

(5.54)
(5.55)

Then taking the conjugate transpose on either side of (5.55) and recalling that Q,. 1s
Hermitian enables us to rewrite (5.55) as
( 5. 56)
Now we see, by comparing (5.54) with (5.56), that
(5.57)
How~ver Qc and RQc are each non-negative. Therefore the only way (5.57) can be

satisfied is for each side to be zero. Thus we have

(5.58)
(5.59)
and (5.52, 5.53) become
(5.60)
(5.61)
This shows that jw is an eigenvalue of A and that v 1 , v 2 are the corresponding right and
left-eigenvectors of A.
However if v 2 f 0, (5.58, 5.61) imply that (A, RQc) has an uncontrollable eigenvalue
on the imaginary axis and condition (i) is not satisfied. Therefore if (i) is satisfied, we must
have v 2 = 0.
Alternatively, if v 1 f 0, (5.59, 5.60) imply that (A, Qc) has an unobservable eigenvalue
on the imaginary axis and condition (ii) is not satisfied. Therefore if (ii) is satisfied, we
must have v 1 = 0.
From the foregoing we see that if conditions (i) and (ii) are satisfied, (5. 51) holds only if
v = 0. Thus, contrary to the assumption we made at the beginning of the proof, H QC has

no imaginary axis eigenvalues when the conditions of the theorem are satisfied.
1
Notice that since R, is non-singular and RQc = B 2R-; B2, we have
only if
Therefore condition (i) in the foregoing theorem is equivalent to
(i)

(A, B 2 ) has no uncontrollable eigenvalues on the imaginary axis.

We will show in the next subsection that, though sufficient for H QC to have no

Solving the QCARE

133

iJDaginary axis eigenvalues, the conditions in the foregoing theorem are not sufficient to
ensure that the QCARE has a stabilizing solution.

5.4.3

Finding the stabilizing solution

The proof of Theorem 5.3 suggests a way of determining the stabilizing solution to the
QCARE. More specifically, we saw t~at when the stabilizing solution to the QCARE
exists, we can find T, (5.47), such that HQc, (5.48), is block upper-triangular with the top
diagonal block being stable. This suggests that the stabilizing solution to the QCARE can
be determined from the eigenvectors of H QC corresponding to the stable eigenvalues.
However, it turns out that we can avoid having to compute eigenvectors. Instead we can
use Schur decomposition on H QC to determine the stabilizing solution to the QCARE. In
order to see how to do this we need the following basic result form linear algebra.
Theorem 5.5 If
(5.62)

MU=UW

where M and Ware square with W being invertible and U having independent columns,
then/
(i) each eigenvalue of W is an eigenvalue ofM, i.e., .A(W] c .A(M]
(ii) range[U] is an M-invariant subspace corresponding to A.[W], i.e., range(U] is
spanned by the eigenvectors of M corresponding to the eigenvalues of W.
~oof ~et (.A;, vi) be any eigenvalue-eigenvec tor pair for W. Then premultiplying
. Wv' = .A;v' by U and using (5.62) we obtain

Mw=A.;w

where w = Uvi and we see that A; is also an eigenvalue of M. This completes the proof of
~

Next assuming the eigenvectors of W are complete, we can express any vect~r q of the
same dimension as these eigenvectors as

n,
q= LO'.;Vi
i=i

so that post-multiplying (5.62) by q gives

Ms=r
where

n,
r= Up

p = L(A.;a;)vi

i=i
S=

Uq

We see that M maps range(U] into itself, i.e., range[U] isM-invariant.

134

Quadratic Control

Finally, notice that if V 1 is a matrix whose columns are the eigenvectors of M

corresponding to the eigenvalues of M which are also eigenvalues of W, then
MVI =VIAl

where A 1 is diagonal with the eigenvalues of W along its diagonal. Then post-multiplying
this equation by the non-singular matrix e such that U = V 1e and inserting ee-l
between V 1 and A 1 gives
MU=VW

where

Thus .A[W] =.-\[AI]. Moreover since

e is non-singular, we have

range[U] = range[V 1e] = range[VI]

This shows that range[U] is spanned by eigenvectors of M corresponding to the

eigenvalues of W.
In order to develop a relation between the quadratic control Hamiltonian matrix and
the QCARE which we can use as a basis for solving the QCARE for its stabilizing
solution, we proceed in the following manner.
Suppose we have the stabilizing solution, P QCO" Then the QCARE, (5.37), can be
written as

and we see that

(5.63)
where
HQc

[A

Qc

Now since A- RQcPQco is stable, we see from Theorem 5.5 that (5.63) implies
range [ _ :QCo ]
is the H Qc-invariant subspace corresponding to the stable eigenvalues of H QC
The standard way of computing this subspace, when it exists, is to use Schur
decomposition. This numerically reliable method computes an orthogonal (unitary
when H Qc is complex) matrix T such that the transformed quadratic control Hamiltonian matrix, HQc, (5.48), is block upper-triangular with first diagonal block, H11, being

Solving the QCARE

135

stable, i.e.,
( 5.64)

Notice from (5.48) that H 11 =A- RQcPQco = "4o. the stable system matrix for the state
feedback control system.
Now since Tis unitary, T* = T -t. we see that premultiplying (5.64) by T and equating
the first n columns on either side of the resulting equation gives
(5.65)
where Tis partitioned into n x n blocks, T;i, with T 1 denoting the first n columns ofT,
i.e.,

Then we see from Theorem 5.5 and (5.65) that

(i/ >.[Hn] c >.[H Qcl
(ii) range[Td is an HQcinvariant subspace corresponding to >.[H 11 ].
Therefore since H 11 is stable, the columns ofT 1 span the stable eigenspace of H QC i.e.,
the columns of T 1 can be expressed in terms of the eigenvectors of Hoc whose
corresponding eigenvalues lie in the open left half plane.
Recall that range[Td =range[T 18] for any non-singular matrix 8 of appropriate
dimension. Hence if T 11 is non-singular the columns of T 1T]} also span the stable
eigenspace of H QC Therefore assuming T 11 is invertible, we postmultiply (5.65) by T]}
and insert T]}T 11 between T 1 and H 11 on the right side of(5.65) to obtain
HQc[T21IT]/]

= [T21/T)/](TuHIIT)/)

Then comparing this equation with (5.63) we see that PQco is given by
PQco

-T21TJ11

(5.66)

and we have the stabilizing solution to the QCARE.

Notice that in order to use the foregoing approach to solYe the QCARE for its
stabilizing solution, not only must H QC have no imaginary axis eigenvalues, (Theorem
5.3), but in addition T 11 must be invertible. Notice also that when T 11 is invertible we
have P QCo > 0 only if T 21 is also invertible. The following theorem establishes conditions
?n the parameters of the plant ahd control cost which ensure that T 11 and T 21 are
Invertible.
Theorem 5.6 Suppose we can find a unitary matrix T so that H 11 is stable and

Quadratic Control

136

where

Then we have the following results:

(i) T 11 is invertible if (A, B 2 ) is stabilizable;
(ii) if T 11 is invertible, then T 21 is invertible if (A, Qc) has no stable unobservable
eigenvalues.
Proof (i) From (5.65) and the blocks in H QC and H

+ RQcT 21

11

we have

= T 11 Htt

( 5.67)

QcT11- A*T2l = T21H11

( 5.68)

AT 11

letting let (,\, v) be any eigenvalue-right-eigenvector pair for H 11 and postmultiplying (5.67, 5.68) by v gives
Th~n

AT 11 v + RQcT 21 v = ,\Tllv

( 5.69)

QcT11v- A*T2lv = ,\T21 v

( 5.70)

Now suppose TIIv = 0. Then T 11 is not invertible and (5.69, 5.70) become
RQcw = 0
A*w

-,\w

(5.71)
(5.72)

where w = T 21 v. Then if w = 0 we would have

T ]v=0
[T21
11

implying that T has dependent columns and is therefore not invertible. This is impossible
since we are assuming that Tis unitary. Therefore 1v -:10.
Now since H 11 is stable, (5.71, 5.72) imply that (A, RQc) is not stabilizable, and since
RQc = B 2R-; 1B2 with R-; 1 non-singular, we have (A, B 2) is not stabilizable. This contradicts the condition given in (i). Therefore no eigenvector v, of H 11 satisfies T 11 v = 0 when
(A, B 2 ) is stabilizable. Thus assuming the eigenvectors of H 11 are complete, the foregoing
condition shows that T 11 is invertible when (A, B 2 ) is stabilizable.
Proof (ii) Suppose TII is invertible and T 21 v = 0 for an eigenvector v of H11 Then
(5.69, 5.70) become
As= ,\s

(5.73)
(5.74)

where Tllv = s.
Therefore since H 11 is stable and,\ is an eigenvalue of H 11 , we have that (A, Qc) has a

Quadratic State Estimation

137

Stable unobserva ble eigenvalue. This contradict s the condition given in (ii). Therefore no
eigenvector v ~f H 11 satisfies T 21 v # 0 if (A! ~c) has no stable unobs~rvable e_i~envalues.
Finally assummg H 11 has a complete set of eigenvectors. the foregomg cond1t1on shows

that r 21 is non-singul ar when (A, Qc) has no stable unobservable eigenvalues.

Although more involved, it can be shown that the foregoing theorem holds even when
done using generalized
H 11 does not have a complete set of eigenvectors. This is
.
.
.
.
eigenvectors.
In summary, in thts sectwn, we ha\e considered the problem of solvmg the QCARE
for its stabilizing solution. The so-called Hamiltoni an matrix was introduced for this
purpose. By using this matrix it was shown that _the comp~ta:ionally ~seful technique of
Schur decompos ition could be employed to obtam the stabthzmg solutiOn to the QCARE
when it exists. In addition, the Hamiltoni an matrix enabled the developme nt of important
conditions on the plant and performan ce index parameter s which are necessary and
sufficient for the existence of a stabilizing solution to the QCARE.

Quadratic State Estimation

5.5

So far we have solved the quadratic control problem under the assumptio n that the state
of th7 plant is known. In this case the controlled input is
I

However, when the plant state is unknown, we must resort to the use of an observer in
an observer based feedback control scheme as discussed at the beginning of this chapter.
Doing this and using the optimal quadratic state feedback matrix K 0 , developed in
previous sections, gives the controlled input as
(5.75)

where

x(t) = x(t)- x(t)

Since u2s(t) = K 0 x(t) is the controlled input \vhich minimizes the quadratic state
fee~back control cost, JQc, (5.17), the departure of the controlled input, u2E(t), from its
optimal value, u25 (t), causes the cost JQc to increase. Our goal in this section is to design
an observer to minimize this increase in the quadratic control performan ce index. We
approach this goal by solving the optimizati on problem of choosing the observer's L
~atrix so as to minimize the energy in the state estimation error caused by both an
un~uls~ disturbanc e in the measurem ent of the plant output and an initial plant state
esttmatwn error.

5.5.1

Problem formula tion

Recall, from Section 3.3, that ideally the observer estimates the state of the plant hased on
the true plant output, y(t) = Cx(t), where for simplicity, we assume, as in the quadratic
state feedback control problem, that the plant is strictly proper. However, in practice,

138

Quadratic Control

only a measured plant output is available and this output differs from the true plant
output by an additive measurement noise signal, w(t). Thus ify 111 (t) denotes the measured
output we have
Y111 (t) = y(t)

+ 1r(t)

It should be emphasized that, unlike the quadratic state feedback control problem where
the disturbance is at the plant input, the disturbance here is at the input to the observer.

v(t)

PLANT

u,(t)

:E

y(t)

~
CONTROLLER

Figure 5.3

Setup for output measurement noise attenuation

output measurement noise

y(t)

control system output

controlled input

v( t)

control system command input

Now applyingy 111 (t) to the input of an observer, we obtain an estimate of the plant state
from

x= Ax(t)

.Y(r)

+ Bu(t) + L[.Y(t) - Ym(t)]

cx(r)

and we see that the plant state estimation error, x(t)

differential equation

.x (r) =

Ax(t)

+ LHi(t)

x(t)- x(t), is governed by the

A= A +LC

(5.76)

As in the state feedback control problem, we assume the unknown disturbance is a largeamplitude, short-duration signal which we model as an impulse, w(t) = 15(t- to)
Suppose the control system has been in operation with a constant command input and
no measurement noise, for a time which is great enough so that the plant state estimate
has converged to the steady-state value for the plant state, i.e.,

x(t)

x,

x(t)

= 0

Quadratic State Estimation

139

. Then shifting the time origin to t 0 , the time where the output measurement noise impulse
occurs, we have w(t) = 8(t), x(O) = 0 and we see from (5.76) that the state estimation
error becomes
(5.77)
Now we want to return the state estimation error to its null value as rapidly as possible
following the occurrence of the measurement noise impulse. Therefore (5. 77) implies that
L = 0 is the best choice for this purpose since then x(t) = 0 and the disturbance would
have no effect on the state estimate. However, this approach to setting up the quadratic
state estimation problem does not take into account effects on the state estimation error
which arise from the initial plant state being unknown. For instance, if the plant is
unstable, we have seen in Chapter 3 that choosing L null causes any estimation error
present at start-up to grow without bound.
To overcome this problem, we reformulate the foregoing optimization problem by
assuming that the initial state estimation error has an additional component not caused
by the output m~asurement noise, w(t), i.e., we take x(O) as

..x(o) =

.xd

Thet'we see from (5.76) that the plant state estimation error, x(t), caused by both
x(O) = xd and w(t) = 8(t) is given by
(5.78)

Now we want to choose L to decrease the effect on the plant state estimation error of
having both a non-null initial state estimation error, and an impulsive output measurement noise. One way of doing this is to solve the problem of choosing L so as to minimize
the energy in the state estimation error, i.e., find L such that

where
JQE

=loco x*(t)x(t)dt

However recalling the use we made of the trace of a matrix in connection with the
deve!opment of the controllability Gramian in Chapter 4, we see here that JQE can be
rewntten as

Which using (5.78) becomes

(5.79)

Quadratic Control

140

where
A= A +LC

Now in order to make the optimization problem more tractable we replace M by

(2e,

where Re, Qe are real symmetric matrices with Qe 2' O,Be > 0. Then our problem is to
choose L to minimize J QE such that A is stable, where
( 5.80)

with
A =A+LC

The development of equations for doing this proceeds along lines similar to those used to
develop the equations for K 0 in Section 5.3.

5.5.2

Problem solution

We begin the development by noting that the required stability of the observer ensures
that the integral, (5.80), converges to a real symmetric matrix P QE so that
(5.81)

where
PQE =

(X)

At-

e Qee

A't

dt

( 5.82)

Recall, from Chapter 4, that PQE' (5.82), solves the Lyapunov equation
(5.83)

Notice that since A and Qe, (5.80), depend on L so does PQE and JQE We use the
notation PQE(L) and JQE(L) to denote this fact. Therefore our goal is to find L so that
(i) A(L) is stable, and
(ii) the solution, PQE(L), to (5.83) minimizes JQE(L), (5.81).
We denote the value of L that satisfies (i) and (ii) as L 0 . Thus L 0 , the optimal value of L,
satisfies

Quadratic State Estimation

141

Now we employ the matrix variational technique used in Section 5.3 to develop L 0
Thus suppose Lis perturbed away from L 0 as
L

L0

+ E(fJL)

(5.84)

where Eisa small positive real scalar and (8 L) is any matrix having the same dimensions
as L such that A is stable.
Then we can write the expansion of PQE about PQEo as
(5.85)
where HOT (higher order terms) stands for terms in the series involving powers of Ewhich
are 2 or greater. Next we obtain the expansion of J QE about its minimum\ alue, J QEo, by
substituting (5.85) in (5.81). Doing this and using the fact that the trace of sum of matrices
is the sum of the trace of each matrix gives
JQE

= trace[PQ] = trace[PQEo] + Etrace[(8 1PQ)] +HOT

(5.86)

= JQEO + E(SIJQE) +HOT

Notice that

( 5.87)
Now comparing the present development of equations for L 0 with the development
used to obtain equations for K 0 , Section 5.3. we see that the optimal value for L, L 0 .
makes (S 1JQE) = 0 for all8L. Therefore we see from (5.87) that an equation for (8 1P QE) is
needed.
W~ begin the determination of an equation for (S 1PQE) by using (5.84) to express "4
and Qe , (5.80), as

A= A + E(8L)C
0

Q" = Qo -t- E[L

Rc(8L)*-'- (8L)ReLo]-r- HOT

where

Then substituting these expressions for

Lyapunov equation, (5.83), yields

A, Q and

the expansion for

PQE

(5.85), in the

(5.88)

142

Quadratic Control

Now the first three terms on the left side of (5.88) sum to a null matrix since P QEo
the Lyapunov equation, (5.83), when L = L 0 Thus using this fact and recalling
from Section 5.3 that
satisfie~

l.im
,~o

[HOT] = 0
E

we see that dividing (5.88) byE and letting Ego to zero gives
(5.89)
Notice that (5.89) is a Lyapunov equation in (8 1PQt} Therefore, since A0 is required
to be stable we must have (81 PQE) ;::: 0. Thus no eigenvalue, <J;, of 81P QE is negative, i.e.,
{ <J; ;::: 0 : i = 1, 2 , n}. Moreover, referring to the development in Section 5.3, we see
that J QE is minimized by choosing L so that 81J QE = 0 for all 8 L. Thus after rewriting
(5.87) as
n

81JQE

= trace[(8IPQ)] =

L<J;

i=1

we see that 81J QE = 0 for all 8 L only if all the O";S are zero for all 8 L which implies that
(8 1PQE) is null for all8L.
Now referring to (5.89) we see that if (8 1PQ) is null then

(8L)Mt + M1 (8L)* = 0
which holds for all 8 L only if M 1 is null. Therefore L 0 must satisfy

which since Re

> 0 can be solved for

L 0 as

(5.90)
Notice that in the same way that K 0 depends on P QCo the stabilizing solution to the
QCARE, so L 0 depends on P QEo the stabilizing solution to an algebraic Riccati equation
commonly referred to as the quadratic filtering algebraic Riccati equation (QF ARE).
This equation is obtained by substituting L 0 , (5.90) in the Lyapunov equation, (5.83).
Thus after some algebra we obtain the QF ARE as
(5.91)
where RQE = C* R;: 1 C. This completes the development of the equations for the optimal
value of L.
In summary, the same variational approach which was used in Section 5.3 to determine
equations for the optimal state feedback matrix K 0 was used to determine equations for
the optimal observer matrix L 0 In the next section we will find PQEo so that the QFARE,

Solving the QFARE

143

I), is satisfied and A - PQEoRQE is stable by resorting to ~ Ham_il_t~nian m~trix

approach similar to the approach used to solve the QCARE for Its stabthzmg solutiOn.

5.6

Solving the QFARE

In this section we exploit the dual nature of the quadratic state feedback and state
estimation problems. We begin by recalling the QCARE, (5.37) and QFARE, (5.91)

QCARE:
QFARE:
where

RQE- C*R-e 1 C
Then comparing these equations we obtain th correspondences given in Table 5.1.
Recall that if (A*, C*) is stabilizable then (A, C) is detectable. Therefore the foregoing
corf!Spondences together with Theorem 5.1 in Section 5.4.1, imply the following
sufficient conditions for the existence of a stabilizing solution to the QFARE.
Theorem 5.7 A0 is stable if
(i) (A, C) is a detectable pair and
(ii) P QEo satisfies the QFARE with
(iii) (A, Qe) is a stabilizable pair.

P QEo :::::

0 and

where A0 =A+ L 0 C with L 0 given by (5.90) and PQEo satisfying (5.91).

Next recalling the quadratic control Hamiltonian matrix, H QC (5.46), we see from the
foregoing correspondences that the quadratic estimation Hamiltonian matrix, H QE
needed to solve the QFARE for its stabilizing solution should be defined as

_[A*

HQE-

Qe

RQE]

(5.92)

-A

where RQE = C* R;;'c

Then substituting PQE in place of PQc in the similarity transformation matrix T, (5.47)

Table 5.1
Estimation
A
C*

QCARE-QFARE Symbol Correspondences

Control
A*

Re
Qe

B2
Rc
Qc

pQEo

PQco

144

Quadratic Control

gives the transformed quadratic estimation Hamiltonian matrix as

T

-1

H QE T = H QE =

[A-*y

where
A=A-PQERQE

Y = PQEA*

+ APQE- PQERQEpQE + Qe

Notice that H QE and fi QE have the same eigenvalues for all T. Therefore we can see, by
choosing T so that Y is null, that H QE has eigenvalues which are mirror images of each
other across the imagin~ry axis. This means that any imaginary axis eigenvalues of H QE
are also eigenvalues of A for all P QE that make Y null, i.e., that satisfy the QF ARE. Thus
when H QE has imaginary axis eigenvalues the QF ARE does not have a stabilizing
solution, cf., Theorem 5). Conditions which ensure that H QE has no imaginary axis
eigenvalues are determined by analogy with Theorem 5.4.
Theorem 5.8 The quadratic estimation Hamiltonian matrix, H QE (5.92), is devoid of
imaginary axis eigenvalues if
(i) (A, C) has no unobservable eigenvalues on the imaginary axis and
(ii) (A, Qe) has no uncontrollable eigenvalues on the imaginary axis
Assuming a stabilizing solution to the QF ARE exists, we can use the Schur decomposition of H QE to determine this solution as was done in Section 5.4.3 to determine the
stabilizing solution to the QCARE. Therefore referring to Section 5.4.3, we use the first n
columns ofT, the unitary or orthogonal matrix which transforms H QE to a block upper
triangular matrix with the first diagonal block stable. Then the stabilizing solution to the
QF ARE is given as
PQEo

= -T21T]}

where

Conditions required for the existence of the stabilizing solution to the QF ARE can be
obtained by analogy with Theorem 5.6.
Theorem 5.9 Suppose we can find a unitary matrix T so that H 11 is stable and

where
HQE= [

A*

Qe

Summary

145

we have the following results:

(i) Tn is invertible if (A, C) is detectable;
(ii) if TIl is invertible, then T 21 is invertible if (A, Qe) has no stable uncontrollable
eigenvalues.
Finally we can show that the stabilizing solution to the QF ARE is unique, when it
exists, by proceeding in the same fashion as was used to prove Theorem 5.2. This
completes the discussion of the stabilizing solution to the Q FARE.

5.7

Summary

We began this chapter with a description of observer based state feedback and its
properties. This was followed by the derivation of desig_n equations f~r both an optimal
quadratic state feedback controller and for an optimal quadratic state estimator
(observer). In each case the design equations were nonlinear matrix equations referred
to as algebraic Riccati equations, i.e., the QCARE and QF ARE. The requirement for the
state feedback system and observer to be stable restricted the use of the solutions of each
of these equations to their so-called stabilizing solution. Conditions for the existence of
these stabilizing solutions were developed, in each case, in terms of the parameters of the
planyand performance index. This was done through the use of a related four-block
matiix known as a Hamiltonian matrix. We saw, in Theorems 5.4 and 5.6, that the
QCARE has a stabilizing solution if and only if (A, B2 ) is stabilizable and (A, Qc) has no
unobservable imaginary axis eigenvalues. Alternatively we saw in Theorems 5.8 and 5.9
that the QFARE has a stabilizing solution if and only if (A, C) is detectable and (A, Qe)
has no uncontrollable imaginary axis eigenvalues.

5.8

Notes and References

The algebraic Riccati equation has played a role of great importance in the evolution of
control theory. An extensive view of this subject is given in [3]. The use of the Schur
decomposition of the Hamiltonian matrix to solve the quadratic state feedback control
problem, is implemented in the MATLAB Control System Toolbox under the '"dommand
lqr. More detailed information on the Schur decomposition is given in [17].
The matrix variational technique used here to develop the design equations for the
optimal K and L matrices was used previously to reduce the effect of inaccurate modeling
ofthe plant on the performance of a quadratic control system [28]. For more recent uses
ofthis approach see [13] and the references therein.
Since the stabilizing solutions, P, to either of the algebraic Riccati equations
~countered in this chapter are uniquely determined from the corresponding Hamiltontan matrix, H, we have H f---+ Pis a function. This fact has lead to the use of H Edom(Ric)
to d~ote those H that admit a stabilizing P and P =Ric(H) to denote the stabilizing
. ~lution P determined fr~m H_. Thi~ notation has become standard in the control
terature. For a more detailed discussiOn see Chapter 13 of [47].
When either the plant is time-varying, or when the control cost is evaluated over a finite
of time, the optimal controller is time-varying and is determined by solving two
matrix differential equations. The development of these conditions is beyond the
ofthe present text. The reader is referred to [I] and [1 O]for more information on this

6
LQ G Control

6.1

Introduction

metho ds for choosing the

This chapte r continues the development, begun in Chapt er 5, of
As in Chapt er 5, our goal
K and L matrices in an observer based feedback contro l system.
on the contro l system's
here is to comba t the effect that unkno wn disturbances have
t loss of generality, that
output. Thus, as discussed in Section 5.3.1, we can assume, withou
l system output is caused
the contro l system's comm and input is null so that the contro
durati on charac ter of the
solely by the disturbance input. The assumed sporadic, short
characterized as being
disturbances used in Chapt er 5, is replaced here by disturbances
a linear system produce an
persistent random signals which when applied as an input to
average power if
output with bound ed average power, where y(t) has bound ed
lim
T->oo

2_ fT y*(t)y (t)dt < oo

T }o

energy unbou nded.

Notice that the persistence of the disturbances makes the output
the output energy
use
longer
Therefore unlike the quadra tic contro l problem, we can no
to attenu ate
ability
's
system
l
caused by the disturb ance input as a measure of the contro
mance
perfor
a
as
used
is
disturbances. Instea d the steady-state average power output
measure.
zero mean Gauss ian
We will assume throug hout that the disturbances w;( t) are
random vectors with covariance given as
(6.1)

148

LQG Control

and E[] is the expectation operator from probability theory. Moreover we will assume
that the output caused by these random disturbances is ergodic so that the steady-state
average power at the output is given as
lim _!._ {T y*(t)y(t)dt
T }o

T -we

lim E[y*(t)y(t)]

(6.2)

t-->cc

Now we will begin by assuming that the plant to be controlled has state model specified
as

x(t)

y 1(t)

Y2(t)

+ wo(t) + B2u2(t)
I
C 1x(t) + w1 (t) + D12u2(t)
C2x(t) + w2(t) + D22u2(t)

Ax(t)

(6.3)
(6.4)

(6.5)

where
disturbances

controlled output

controlled input

measured output

and u2 ( t), y 1 (t), y 2 (t) have dimensions m 2 ,p 1 ,p2 respectively.

Our goal here is to choose K, Lin an observer based feedback control system so that
J cc is minimized where

fcc= {-'>CC
lim E[y7(t)yJ (t)]

(6.6)

Since the disturbance signals are assumed to be zero mean Gaussian random vectors,
this type of optimal control is known by the acronym LQG (Linear Quadratic Gaussian).
More recently this optimal control problem was interpreted as a minimum norm problem
and the term H 2 optimal control was also used. More will be said about this in Chapter 8
following the development of ideas involving signal and system spaces.
w,(t)

~ y,~)
-~~~------------~~+~
wo(t)

PLANT
u,(t)

CONTROLLER

Figure 6.1

LQG control configuration.

LQG State Feedback Control Problem

149

As in the quadratic control problem, the separation principle allows us to obtain the
solution to the LQG control problem by combining the solutions to two subproblems:
(i) the LQG state feedback control problem
(ii) the LQG state estimation problem.
The LQG control problem is also referred to as
(iii) the LQG measured output feedback control problem.
In what follows we assume that all matrices are real so that we can write ()Tin place
of()*

6.2

LQG State Feedback Control Problem

In this section we dewlop the design equations for the optimal state feedback control
matrix K which minimizes lee, (6.6). We will do this using the \ariational approach
introduced in the previous chapter. We will see that the determination of the optimal K
requires the solution of an algebraic Riccati equation.

6.2.1

Problem formulation

Suppose the plant state, x(t). is known. Then we want to determine the state feedback
matrix, K, involved in generating the controlled input, u2 (t) = Kx(t), so that1cc, (6.6), is
minimized. Notice from (6.3. 6.4) that under state feedback the system relating the
disturbances to the controlled output is given as
x(t) = Ax(r)

+ 1r0 (t)

(6.7)

Yt (t) = Cx(t)

+ w 1(t)

(6.8)

where

and w0 (t) and 1r 1 (t) are unknown zero mean Gaussian random wctors having covariances given by (6.1).
Now in order to proceed \Ye must have H 1 (t) = o. This requirement can be seen by
noting from (6.1) that when w1 (t) cf0 the performance index, fcc, (6.6), is unbounded for
all K because of the impulsive nature of the covariance of w1 (t). Therefore in order for the
LQG state feedback control problem to be properly posed we must have w1 ( t) = o.
Therefore under this assumption the statement of the LQG state feedback control
problem becomes:
minJcc
K

150

LQG Control

given that
lim E[yf (t)yl (t)]
lee= 1---+x
x(t)
YI (t)

Ax(t)

+ w0 (t)

= Cx(t)

(6.9)
(6.10)
(6.11)

with A stable where (A, C) are given in (6.7, 6.8) and w0 (t) is characterized from (6.1) as

The variational method used in the previous chapter is used now to develop equations
forK to solve this problem.

6.2.2

Development of a solution

Since A is stable and lee, (6.9), involves the steady-state behavior, the performance is
unaffected by the initial state of the plant. Therefore only the zero state controlled output
response is needed. Thus referring to (6.1 0, 6.11) we see that this response is given by
(6.12)
Then using (6.12) we see that the expectation needed in the cost, (6.9), can be written as

Now we can express the right side of this equation by employing the trace relation used
in the development of the controllability Gramian in Chapter 4. This relation is restated
here as
(6.14)
for any compatible vectors vi and matrix 8.
Therefore identifying w0 ( T 1 ) with v 1 and w0 ( T2 ) with v 2 enables (6.13) to be rewritten as

Next evaluating the expectation using (6.1) reduces (6.15) to

LQG State Feedback Control Problem

151

Finally, use of the sifting property of the impulse gives

E[yf (t)y 1 (t)] =trace

and substituting T = t index, JGe (6.9) as

T2

[1 W
1

and allowing t

00 e"F(t-T2 )CT CeA(t-T2 )dT2

--->

(6.17)

oo enables us to express the performance

J 6 e = trace[WooPGcl

(6.18)

where

Now recalling Theorem 4.5 in Chapter 4, we see that P 6 e satisfies the Lyapunov
equation
(6.19)
and isfthe o.>servability Gramian for the pair (A" C). Thus since (A, C) depends on K so
does PGe
Now the equations which determine K so that J 6 e, (6.9), is minimized can be
developed by following the variational approach used in the previous chapter. Thus if
K 0 denotes the optimal value forK we have

and expanding he and PGe about K 0 yields

he= heo + t:(l5,he) +HOT

= trace[WooPGeol

+ t:trace[W00 (8 1P 6 e)] +HOT

(6.20)

where

~d

t:

is a small positive real scalar with l5 K being any real matrix having the same

~ension as K such that A is stable. In addition, recall that HOT denotes higher order
'Fms in E and therefore satisfies

limHOT =0
<-->0

(6.21)

we see from (6.20, 6.21) that

(6.22)

152

LQG Control

Recall from the previous chapter that lee is minimum when (o 1Jcc) = 0 for all bK.
Thus we see from (6.22) that we need a relation between (61 Pcc) and 0 K. This relation is
developed by substituting the expansions for K and Pc;c as given in (6.20) into the
Lyapunov equation (6.19). This gives
(6.23)
where

A+B2Ko

Co = C 1 + D12Ko

Notice that the first three terms on the left side of (6.23) sum to a null matrix since P GCo
satisfies (6.19) when K = K 0 . Therefore dividing (6.23) through byE, letting Ego to zero,
and using (6.21) yields
(6.24)
where

and we see that (o 1PGc) is governed by a Lyapunov equation.

Now in order for (o 11Qc), (6.22), to be zero for all OK, we must have W 00 (o 1PGc) null
for all OK. However W 00 is independent of OK. Therefore we can have W 00 (o 1PGc) null
for all oK only if (8 1PGc) is null for all8K.
Thus we see from (6.24) that if (8 1PGc) = 0 for all8K then we must have M 1 = 0.
Therefore assuming that the columns of D 12 are independent so that Df2D 12 invertible, we
can solve M 1 = 0 for K 0 as
(6.25)
where

As in the case of the optimal quadratic control problem, the matrix PGco is obtained as
the solution to an algebraic Riccati equation. We develop this equation by substituting,
Ko, (6.25), in the Lyapunov equation, (6.19). Thus after substituting for A, C from (6.7,
6.8) in (6.19) we obtain

AT PGco

+ PGcaA + K"[; Qo + Q(; K + Cf C1 + K"[; Df2D12Ko = 0

0

which after substituting for K 0 from (6.25) becomes

AT PGco

+ PGer,A- 2QT; (Di;D\2)-l Qo + cf cl + QT; (Df2Dl2)- 1Qo = 0

(6.26)

LQG State Estimation Problem

153

Finally substituting for Q0 , (6.25), gives

[AT~ C[ Dn(D[zDtz)- 1 BfjPGco

~PGcoBz(Df2Dl2)- 1 BzPGco

+ PGco[A ~ B2(D[zDn)-I D[zCJ]

+ CfCt

~ CfDJ2(DfzDn)- 1DfzCt

=0

which we can write more compactly as

(6.27)
where
At= A~ Bz(D'fzDn)- 1DTzC 1
Rt = Bz(DfzDI2f 1B[
QI = C'f[I ~ DI2(D'fzDu)- 1D'fzJCt

Notice that

Now the algebraic Riccati equation (6.27) is referred to as the Gaussian control
algebraic Riccati equation, (GCARE). Thus we obtain the optimal K, K 0 , by solving the
GCARE for its stabilizing solution, i.e., the solution P GCo which makes A + B 2 K 0 stable.
As in the previous chapter, we will show that the GCARE can be solved for its stabilizing
solution by using an appropriate Hamiltonian matrix. This matter is considered further
after taking up the state estimation problem.
Before leaving this section, notice that the requirement that D[2D 12 be invertible is
similar to the requirement we encountered in Section 5.3.2 that Rc be invertible. Both Rc
and D[2D 12 play the role of introducing a penalty on the use of the controlled input in
"~
minimizing the control costs JQc and lac respectively.

6.3

LQG State Estimation Problem

In this problem we assume that the plant state x( t) is unknown and that we want to design
an observer to estimate it from the controlled input, u2 (t) and measured output, Yz(t),
(6.3, 6.5). Unlike the problem just treated, the controlled output y 1 (t), (6.4) is not used.
Our goal now is to choose the observer matrix L so that JGE> the steady-state average
power in the state estimation error, i(t), is minimized where
JGE

=lim E[.XT(t)i(t)] =lim trace[E[i(t)ir(t)]]

(----700

(6.28)

t-oo

Notice that this performance index involves the steady-state, state estimation error
covariance matrix P GE where

154

LQG Control

n of the
In order to proceed we need to review the role played by L in the evolutio
estimati on error. This is done as follows.

6.3.1

Probl em formu lation

having the
Recall, from the beginnin g of this chapter, that the state model for the plant
as
given
is
output
only
its
measure d output, Y2(t), as

+ wo(t) + B2u2(t)

x(t)

Ax(t)

Y2(t)

C2x(t)

( 6.29)

+ w2(t) + D22u2(t)

(6.30)

vectors
where the unknow n disturba nces w0 (t) and w2 ( t) are zero mean Gaussia n random
that an
see
we
3,
Chapter
to
referring
Then
(6.1).
with covarian ce matrices as specified by
as
observer for this state model can be cast

+ B 2u2(t) + L[y2(t) - Y2(t)]

.X (t)

Ax(t)

.Y2(t)

C 2x(t)

+ D22u2(t)

(6.31)
(6.32)

where we have replaced the unknow n disturba nces by their expected values.
on
Again as in Chapter 3, we can obtain a differential equatio n for the state estimati
gives
This
6.32).
(6.30,
using
and
(6.29)
error by subtract ing (6.31) from

(t)

Ax(t)

(6.33)

+ Bw(t)

where

w(t) =

[wo(t)l

x(t) = x(t)- x(t)

w2(t)

A= A+ LC2

B=

[I

L]

nce vector,
Notice from (6.1) that the compos ite zero mean Gaussia n random disturba
w(t), has covarian ce
(6.34)

where

that A is
Recall from Chapter 3, that in the absence of disturbances, we chose L so
hes the
approac
which
estimate
state
a
s
generate
r
observe
stable. This ensures that the
l
beneficia
this
nately
Unfortu
error.
on
estimati
state
plant
initial
plant state for any
t
persisten
of
asympto tic behavio r of the state estimato r is not possible in the presence

--..... ~'!".'"""''-<'1"1?:'11'.

toG st~,;; E:fdimiliion-:liriiiii"~'rir

155

.disturbances like those being considered here. However the expected value of the state
estimation error will be null in the steady-state when A is stable and the expected values of
the disturbances are null. Estimators which have this desirable property are said to be
unbiased. Therefore since we are assuming zero mean disturbances, we obtain unbiased
estimates of the plant state provid~d our choice of L to minimize h, (6.28), is
constrained by the requirement that A be stable. This plus the fact that the performance
index, hE, (6.28), concerns the steady-state behavior of the estimator, implies that we can
ignore the state estimation error response due to any initial state estimation error.
Therefore using (6.33) we obtain
(6.35)

Now this relation can be used to develop an expression for the state estimation error
covariance matrix PGE We do this by following the steps we used to develop PGc in the
previous section. Doing this yields

(6.36)

Finally, we can obtain the steady-state error covariance matrix,

= t- T 2 and taking the limit as t tends to infinity. This yields

PGE

by setting

(6.37)

Notice from Chapter 4, (Theorem 4.5), that PGE (6.37), is also the controllability
Gramian for (A, B) and satisfies a Lyapuno v equation
(6.38)

where P GE depends on the observer matrix L since both A and B depend on L. Thus we see
from (6.28) that we need to adjust L so that A is stable and so that the solution to (6.38),
PGE, has minimum trace. A solution to this problem is obtained now by using the duality
lletween the present problem and the LQG state feedback control problem which we
~lved in the previous section.

Problem solution

begin by using the dependency of and C, on K, (6.7, 6.8) to expand the Lyapunov
(6.19) and by using the dependency of A and Bon L, (6.33) to expand the

156

LQG Control

Lyapunov equation (6.38). This gives

AT Pee+ PeeA

+ KT sf Pee+ PecB2K

+Cf C1 + Cf D12K + KT Df2C1 + KT D[2D 12 K

APeE

+ PeEAr + LC2PeE + PeEcf Lr

+Woo+ LW2o

+ Wo2LT + LW 22 LT = 0

(6.39)

(6.40)

In order to compare the symbols in these Lyapunov equations we make use of the fact
that covariance matrices are symmetric. Therefore we can factorize W, (6.34), as
(6.41)

so that

Notice that these factorizations allow us to relate w0 (t) and w2 ( t) to a zero mean Gaussian
random vector u 1 (t) as
(6.42)
where

Now we can use the foregoing factorizations, (6.41), to rewrite the Lyapunov equation
(6.40) as

+ LC2PeE + PeEcT Lr
+B1Bf + LD21Bf + B1Df1LT + LD21Df1LT = 0

APeE + PeEAr

(6.43)

Then comparing terms in (6.43) with terms in (6.39) we obtain the symbol correspondences for the LQG state feedback control problem and for the LQG state estimation
problem given in Table 6.1.
Thus we see that the formulas for the optimal Land PeE, i.e., L 0 and PeEo' can be
obtained by substituting from these correspondences in the formulas for K 0 , Peeo given
in (6.25, 6.27)). Doing this yields
(6.44)
(6.45)

Table 6.1

GCARE -GFARE Symbol Correspo ndences

Control

Estimatio n

------

PGc

PGE

AT
KT

A
L

sr

Cz

cf

Bl

Dfz

D21

where

A2 =A- BtDf1 (DctDfJ )- 1 C 2

R2

Cf(D21 Dft)- 1 C2

Q2 = B 1 ~I- DJ1(D 21 Df~)- 1 D 21 ]Bf

Notice that

A2- pGEoR2 =A+ LoC2

filtering
Now the algebraic Riccati equation (6.45) is referred to as the Gaussia n
(6.44)
in
PGEo
ting
substitu
by
d
algebraic Riccati equatio n, (GFARE ). Thus L 0 is obtaine
makes
L
resulting
the
0
since
where PGEo is called the stabilizing solution to the GFARE
stabilizing
A+ L 0 C 2 stable. As in the previous chapter, the GFARE can be solved for its
6.5.
Section
in
this
show
will
We
solution by using an appropr iate Hamilto nian matrix.
on
estimati
state
LQG
the
to
solution
Before going on, notice that the foregoing
dent
indepen
have
must
D
that
21
problem requires that D 21 D[1 be invertible. This implies
(t) are
rows. Thus when this conditio n is satisfied the elements in the random vectorw 2
s in
element
the
of
subset
a
that
sense
deterministically indepen dent of each other in the
n
conditio
this
When
exactly.
(r)
w
in
2
w2(t) can not be used to determine another element
.
singular
being
as
to
is not satisfied, the estimati on problem is referred

6.4

LQG Meas ured Output Feedback Problem

estimati on
Having solved the LQG state feedback control problem and the LQG state
d output
problem , the separati on principle allows us to determine the LQG measure
feedback controll er by combini ng these solutions as follows.
in the
First, replacing the plant state by the observe r's estimate of the plant state
expression for the controll ed input gives

r. Then
\Vhere Ka is obtaine d from (6.25, 6.27) and i(r) is the state of the observe
(6.31,
n,
equatio
tial
differen
state
r's
observe
the
in
t)
(
u
for
on
2
substitu ting this expressi

158

LQG Control

6.32), yields the state model for the controller as

.X (t)

u2(t)

= K 0 .X(t)

A,cx(t)

L 0 Y2(t)

( 6.46)
(6.47)

where
Ace =A+ Lu( C2

+ DnKa) + B2Ko

and L 0 is obtained from (6.44, 6.45).

This completes the formulation of the controller which solves the LQG control
problem. It remains now to develop conditions on the plant state model paJameters
which ensure the existence of stabilizing solutions to the GCARE and the GFARE.

6.5

Stabilizing Solution

Since stabilizing solutions to both the GCARE and GF ARE are needed to implement an
LQG controller, sufficient conditions for the existence of these solutions, expressed as
conditions on the plant parameters, are of considerable importance to designers of these
controllers. As in the quadratic control problem, the stabilizing solutions to the GCARE,
(6.27), and to the GFARE, (6.45), can be computed, when they exist, by using
appropriate Hamiltonian matrices. We will use this fact, in the same manner as was
done in the previous chapter, to develop conditions on the parameters of the plant which
ensure the existence of stabilizing solutions for the GCARE and GF ARE. The duality of
the GCARE to the GF ARE allows us to obtain these results for both by treating the
GCARE in depth and extending the results obtained to the GF ARE through the use of
duality.

6.5.1

The Hamiltonian matrix for the GCARE

Referring to Section 5.4 we see that the GCARE has a stabilizing solution if and only if
the associated Hamiltonian matrix Hac has
(i) no eigenvalues on the imaginary axis,
(ii) T 11 invertible where

T,=[~~:J
is any 2n x n matrix whose columns span the
Hac corresponding to the stable eigenvalues of H GC

invariant subspace of

where Hac is composed from the GCARE, (6.27), in the same manner as H QC is
composed from the QCARE in Section 5.4.2. Thus we see that
(6.48)

Stabilizing Solution

159

with
AI =A~ B:(D[2DI2)- 1 Df~CI

R1

= B2(Df2DI2)-I BJ

QI

= Cf[/ ~ DciDf2Dic)- 1Df:JCI

Also recall. from the previous chapter, that condition (i) is necessary and sufficient for
the existence of a nonsingular 2n x 2n matrix T, (which can be orthogonal), so that H cc
can be transformed to block upper triangular form such that then x n matrix H II is stable
where
(6.491

Then condition (ii) is required to construct the stabilizing solution to the GCARE from

and we see that Pcco > 0 requires that T 21 be invertible.

In what follows we obtain necessary and sufficient conditions on the plant parameters
so that (i) and (ii) hold. We do this by modifying theorems obtained in the previous
chapter taking into account the differences in the dependency of the Hamiltonian
matrices H Qc and H cc on the plant parameters. Conditions on the plant parameters
which ensure that P ceo > 0 will also be dewloped.

6.5.2

Prohibition of imaginary eigenvalue s

Recall that in the quadratic state feedback control problem we obtained conditions on the
plant and performance index which ensured that H QC has no imaginary axis eigenvalues.
These conditions were given in Theorem 5.4. Therefore using Theorem 5.4 and comparing H QC with H cc, we see that H cc, (6.48), has no imaginary axis eigenvalues if:
(a) the pair (A 1 RI) has no uncontrollable imaginary axis eigemalues and
(b) the pair (A 1 , QJ) has no unobservable imaginary axis eigenvalues.
Notice that these conditions are not immediately applicable to the given data of a plant
state model and performance index since they do not explicitly state how these
parameters are involved in contributing to the presence of imaginary axis poles for
H cc This deficiency is overcome in the following theorem.
Theorem 6.1 H cc, ( 6.48), has no imaginary axis eigenvalues if
(i) (A.B 2 ) has no uncontrollable imaginary axis eigemalues,
(ii)
__,;E(-:x:.oo)

where B 2 is n x m2

160

LQG Control

Proof (i) Referring to the proof of Theorem 5.4, we see that smce Df2 D 12 1s
nonsingubr we have

if and only if
and therefore A is an uncontrollable eigenvalue of (A 1 ,R 1 ) if and only if A is an
uncontrollable eigenvalue of (A, B2 ). Thus condition (i) here is equivalent to condition
(i) in Theorem 5.4.
Proof (ii) We begin by considering the pair (A 1 , Q1 ) where
A, =A - B2(Df2Dt2)- 1Df2C1

Thus if jw is an unobservable eigenvalue of (A 1 , Q1 ) we have

which can be rewritten as

(A- jwl)v- B2(Df2D!2)- 1Df2C,v = 0

(I- D!2(Df2D12)- 1Df2)C,v = 0

Now these equations can be written as
(6.50)

where
8= [

A- jwl

B2

C,

D12

~]

Since .6. is invertible, 8.6. has independent columns only if 8 has independent columns.
However (6.50) implies that 8.6. has dependent columns. Therefore 8 must have
dependent columns and
rank[8] < n + m 2
This shows that condition (ii) is sufficient for (A 1 , Q1 ) to have no imaginary axis
eigenvalues. Now we want to use this fact to show that condition (ii) is sufficient for
(A 1, Q 1 ) to have no imaginary axis eigenvalues.
In order to proceed we need to be able to factor D, as
De=

') T

D~

D21

( 6.51)

Stabilizing Solution

161

where here

However, we can only factorize De in the foregoing manner if De is symmetric and nonnegative. The symmetry of De is obvious. We can show De~ 0 as follows.
Since D 12 has independent columns, we can solve the matrix equation
D 12 u

= y,

y,. E range[Dd

as

Therefore it follows that

and
(6.52)

However since any p 1 dimensional real vectors y can be written uniquely as

Y

= Yr + Y1.

(6.53)

where
y, E range[Dn]

YJ. E null [Dfz]

we see from (6.52) that

(6.54)

and D c is said to be a projector.

Finally, since in general y"[ y .1.

0, we see from (6.54, 6.53) that

(6.55)

for all real p 1 dimensional vectors, y. Therefore since the right side of the foregoing
equation can not be negative we have shown that De ~ 0 and therefore we can factorize
De as (6.51) and we can proceed to show that when condition (ii) is satisfied (A 1, QJ) has
no imaginary axis eigenvalues.
Suppose condition (ii) is satisfied. Then we showed earlier in the proof that any
eigenvector-imaginary axis eigenvalue pair (v,jw) of A 1 satisfies
(6.56)

where

162

LQG Control

Thus we must have

so that

or

This shows that

for any eigenvector v of A 1 associated with an imaginary axis eigenvalue which implies
that the pair (A 1 , Q 1) has no unobservable imaginary axis eigenvalues.

The foregoing proof reveals an important property of Dc which we will make use of in
Chapter 9 and 10 in connection with the Hx feedback control problem. Not only is D,
square and non-negative but in addition is a contraction since it has the property that
UT

u:::; YTY

for any p 1 dimensional vectors u, y which are related through De as u =DeY This can be
seen by noting from (6.53-6.55) that
T

Y Y

Yr Yr

T
+ Y1Y
.L

A matrix or operator having this property is referred to as a contraction.

In addition, notice that matrices which are contractions have eigenvalues which are
bounded above by one, i.e.,

Amax[Dc] :S: 1
This property can be seen by refering back to Chapter 4 and the discussion of the
significance of the largest eigenvalue of a non-negative matrix (Section 4.6.3).
Finally, notice that this constraint on the eigenvalues of De is also a consequence of De
being a projector since De projects from a higher to a lower dimenional space.

6.5.3

lnvertability ofT 11 and T21

Having established conditions on the plant state model which ensure that H cc has no
eigenvalues on the imaginary axis, it remains to determine conditions on the plant which
ensure that T 11 , (6.49), is non-singular. We can do this by appropriately modifying
Theorem 5.6 as follows.
Theorem 6.2 If H cc, (6.48), has no imaginary axis eigenvalues so that we can find T,
(6.49), so that H 11 is stable where
(6.57)

Stabilizing Solution

163

then we have
,
(i) T 11 invertible if i A, B2 ' is stabilizable:
DzC
.
(A
if
invertible
is
T
1 : has no stable unobservable
then
1
21
(ii) if T 11 is invertible,
eigenvalues where

A, =A- B2(Df2DJ2)- 1Df2C1

D, = I - DdDf2D12)- 1Df2

(nz)

T Dz

Proof (i) We begin by recognizing that (6.57) implies that

which we can expand as

(6.58)
(6.59)
Now suppose H 11 has an eigenvalue-eige nvector pair(.>., v) such that

Then post-multiplyin g (6.58. 6.59) by v we see that

(6.60)

-Afw = >.w

(6.61)

where w = T 21 v.
NO\Y ifw =owe would haYe

implying that T 1 and hence T has dependent columns. This is impossible since T is
invertible. Therefore w -=f. 0.
Finally since H 11 is stable we have Re[)..] < 0 and therefore (6.60, 6.61) imply that
(A 1 , R 1 ) is not stabilizable. This is equivalent to (A, B 2 ) not being stabilizable since
R1 = B2 (Df2 D 12 )- 1Bf, Therefore no eigenvector, v, of H 11 satisfies T 11 v = 0 when
(A, B 2 ) is stabilizable. Thus assuming the eigenvectors of H 11 are complete we haw
shown that T 11 must be invertible when (A, B 2 ) is stabilizable. If H 11 lacks a complete set
of eigenvectors. the proof is completed by resorting to generalized eigenvectors.
Proof(ii) Suppose T 21 v = o with s = T11 u 'I 0 for some eigenvector,'' of H 11 . Then

164

LQG Control

post-multiplying (6.58, 6.59) by v we obtain

A]S

(6.62)

=AS

(6.63)
where
I

M=D'cC1
However since, in general, we have null[Mr]l_ range [M], we see that (6.63) is satisfied
only if s Enull[M]. Therefore if (6.63) is satisfied then
(6.64)
I

Moreover since H 11 is stable, (6.62, 6.64) imply that (A 1, D;C 1) has a stable
eigenvalue. Therefore no eigenvector, v, of H 11 satisfies T 21 v = 0 when
(A 1 ,D~CJ) has no stable unobservable eigenvalues. Thus assuming ~he eigenvectors of
H 11 are complete we have shown that T 21 is invertible when (A 1, D;.c J) has no stable
unobservable eigenvalue. Again as in the proof of (i), if H 11 lacks a complete set of

eigenvectors, the proof is completed by resorting to generalize~ eigenvectors.

Notice that when (A, CJ) is not observable neither is (A 1, D~C 1 ). Therefore when the
GCARE has a stabilizing solution, a necessary condition for T 21 to be non-singular so
that Pcco > 0 is that (A, C 1) have no stable unobservable eigenvalues. However this
condition is not sufficient for P ceo > 0. This can be seen from the following considerations.
First, reviewing (6.51-6.56) we see that
unobs~rvable

where null [Di J is of dimension m 2 and it may turn out that C 1v # 0 but Dt.c 1v = 0.
Df2 D 12 = RD is symmetric and invertible it can be factored and its
Second, since
I
square root Rb is invertible. Hence we can write (6.51) as

(D;.1)T D'c1=I- D12 (Rb1)-1( Rb1)-T Df2

which implies that

[ (Rb)d

-T

D"(;

has orthonormal columns. Therefore it is impossible to have a non-null vector q such that
I

D'cq =

and

(Rb1)

-T

D'{;

Stabilizing Solution

Thus if vis an eigenvector of A which satisfies C 1 v fc 0 but

(Rb)-T D[2 C 1v # 0 so that

D~C 1 v = 0,

165

then we have

and vis an eigenvector of A but not of A 1 Notice in this case that unlike the eigenvectors
of A, the eigenvectors of A 1 depend on B2 implying that the positive definiteness of the
stabilizing solution depends on B2 .
In summary, the GCARE has a stabilizing solution, Poco: if
(lc) (A, B 2 ) is stabilizable;
(2c)

rank [[

A- jwl
wE

cl

(-oo,oo)

Finally if a stabilizing solution, Poco, exists then

!

(3c) Poco> 0 if (A 1 ,D~C 1 ) has no stable unobservable eigenvalues.

6.5.4

Conditions for solving the GFARE

In order to obtain conditions on the plant parameters which ensure that the GF ARE has
a stabilizing solution, we can use the correspondence between matrices in the GCARE
and the GF ARE given in Table 6.1. Therefore substituting these correspondences in (lc, 3c) we see that the GFARE has a stabilizing solution, PoEo if
(le) (AT, Cf) is stabilizable;
(2e)

rank [[

AT-jwi
T

Bt

cf]] =
T

D21

n + P2

wE

(-oo,oo)

.,~

Moreover, if a stabilizing solution, P GEo, exists then

1

(3e) POEo

> 0 if (AJ, D~Bi) has no stable unobservable eigenvalues, where

However these conditions can be restated in a more convenient form by using the facts
that the stabilizability of any pair (A, B) is equivalent to the detectability of the pair
(AT, BT), and that the rank ofamatrixequals the rank of its transpose. Thus the GFARE
has a stabilizing solution, PoEo if
(le) (A, C2 ) is detectable;
(2e)

rank [[

A- jw/

Cz

B1
D21

]]

n +p 2

wE

(-oo,oo)

166

LQG Control

Moreover, if a stabilizing solution,

PGo,

exists then

(3e) Pc;eo > 0 if ( A2, B 1Df) has no stable uncontrollable eigenvalues.

6.6

Summary

In this chapter we have given a derivation of the design equations for the LQG optimal
feedback controller for a linear time-invariant continuous-time plant. The performance
criterion or cost function which is minimized by the optimal LQG controller is the steadystate expected or average value of a quadratic form in the output vector to be controlled
(the controlled output) when the disturbance input is a zero mean Gaussian random
vector. Necessary and sufficient conditions were given for being able to design the LQG
controller.

6. 7

Notes and References

Again, as in the quadratic control problem treated in the previous chapter, we have seen
that the algebraic Riccati equation plays a central role in the design of optimal
controllers. For other instances where the algebraic Riccati equation arises in control
theory see Chapter 13 of [47]. The LQG optimal control problem was recently referred to
as the H 2 optimal control problem. The reason for this will be given in Chapter 7.
Assuming the appropriate existence conditions are satisfied, the LQG controller can be
calculated using the command h2lqg in the MATLAB Robust Control Toolbox.
The observer obtained by solving the LQG state estimation problem was originally
given by Kalman as an alternative to the Wiener filter for extracting a desired signal from
an additive combination of the desired signal and noise. This aspect of the LQG problem
has had a great impact on a wide variety of industrial problems. There are many books
which deal with this subject. For example, informative treatments are given in [7, 10, 25].

7
Signal and System Spaces

7.1

Introduction

In the previous two chapters we were concerned with the problem of designing controllers
to attenuate the effects of disturbance inputs on the output of a feedback control system.
We assumed that the disturbance input was either an impulse (Chapter 5) or a random
signa1with an impulsive covariance (Chapter.6). In the next two chapters we develop
ideas needed to solve the disturbance attenuation problem for a broader class of
disturbance input. Signals in this class have finite energy and are denoted by 2 [0, oo)
where is used in recognition of the mathematician H. L. Lebesgue, pronounced
"Lebeg", the subscript 2 is used in recognition of the quadratic nature of energy, and
the bracketed quantities indicate the time interval over which the signals are not always
zero. Therefore f( t) E 2 [0, oo) if
(i) f(t) has bounded L 2 norm (finite energy),

[lao J*(t)J(t)dt]!= IIJ(t)ll2<

00

(7.1)

where [] 2 denotes the positive square root andf*(t) denotes the conjugate transpose of

f(t).
(ii) f(t) is null for negative time

f(t) = 0

for all

t E ( -oo, 0)

(7.2)

In this chapter we will use the foregoing idea of the L2 norm for signals to develop
system norms defined in both the time and frequency domains. This is made possible by
Parseval's theorem which relates the L2 norm in the time domain to the L2 norm in the
frequency domain.

7.2

Time Domain Spaces

In classical control theory there was no apparent need to consider signals defined for
negative time since control problems usually have a well defined start time. However the

168

Signal and System Spaces

need for more sophisticated mathematics to deal with new approaches to control
problems requires that we enlarge the domain of definition of signals to include the
negative time axis and to consider the operation of systems in both positive and negative
time.

7.2.1

Hilbert Spaces for Signals

Consider a signal vector ,f( t), which unlike signals in .C 2 [0, oo), is not zero for all negative
time but still has finite energy over the time interval ( -CXJ, CXJ), i.e.,

(7.3)
Any f( t) which satisfies (7 .3) belongs to the normed space .C 2 ( -oo, oo).
Alternatively, another related normed space, denoted .C 2 ( -oo, 0], consists of finite
energy signals which are null for all positive time. Notice that .C2 [0, oo) and .C 2 ( -oo, OJ are
each subspaces of .C 2 ( -oo, oo ), i.e.,

Now from the foregoing we see that if[(t) E .C 2 ( -oo, oo) it is always possible to write

f(t) =f+(t)

+ f_(t)

(7.4)

wheref+(t) E .C2 [0, oo) andf_(t) E .C 2 ( -oo, OJ with

f+(t) = f(t)

>0

j~(O)

= ""f(O)

f_(t) =f(t)

<0

f_(O)

(1- "")f(O)

where"" E [0, lJ.

Notice that once"" is specified,f+(t),f_(t) are uniquely dependent onf(t). Moreover
since f+ (t) and j"_ (t) are nonzero on disjoint intervals, we see that

l:f~(t)f_(t)dt =

(7.5)

Now (7.4) together with (7.5) enables the L 2 norm of any f(t)
decomposed in the following fashion

[LCXJ [f~(t)f+(t)Jdt +

[Ill~ (t) II~ + III- (t) II~]~

1:

.C 2 ( -oo, CXJ) to be

[f*_(t)f_(t)Jdt],

which is reminiscent of the Pythagorean theorem for right-angled triangles.

(7.6)

Time Domain Spaces

169

In the foregoing signal analysis we encountered an integral involving two signals in

2(-oo, oo ). In general the quantity

1h

a*(t)p(t)dt

(7.7)

is referred to as the inner product of a(t) and p(t) for any equal length vectors
a(t),p(t) E 2 (a,b) and is denoted as< a(t),p(t) >,i.e.,

< a(t), p(t) > =

1h

a*(t)p(t)dt

(7.8)

Notice that, unlike the norm which involves only one signal and is always a real nonnegative scalar, the inner product involves two signals and is not restricted to be positive
or real. Notice that the norm and the inner product are related as
I

lla(t)ll2 =[<a, a >J2

(7.9)

where again [J 2 denotes the positive square root.

J\ signal space with an inner product is referred to as inner product space. Under
additional technical constraints (completion), an inner product space is referred to as a
Hilbert space. Hilbert space has found widespread use in applied mathematics, physics
and engineering. In addition to being normed spaces, 2[0, oo ), 2( -oo, oo ), and
2( -oo, OJ are each Hilbert spaces with inner product defined by (7.7).
An important consequence of inner products is the property of orthogonality. Two
, signals are said to be orthogonal or form an orthogonal pair if they have an inner product
which is zero. Thus we see from (7.5) thatf+(t),J_(t) E 2(-oo, oo) are orthogonal.
The classical example of signals which are orthogonal arises in connection with the
Fourier series decomposition of periodic signals. This decomposition exploits the
periodicity and orthogonality of the following signal pairs: {cos(wt), sin(wt)};
{cos(nwt), cos(mwt)}; {sin(nwt), sin(mwt)}; where n, m are unequal integefs and the
interval of integration used in the inner product is the period~ of the periodic signal being
decomposed.
Two spaces S 1 , S 2 are said to be orthogonal, denoted S 1 _L S 2 if

< a(t),p(t) >= 0

for all a(t) E S 1 and all p(t) E S2

Thus we see that 2( -oo, OJ _L 2[0, oo ). More important, 2(-oo, OJ n 2[0, oo) = 0 so
that 2(-oo, OJ and 2[0, oo) are orthogonal complements of each other, denoted by
2(-oo,OJ = t[O,oo)
and the decomposition of any signalf(t) E 2( -oo, oo) given by (7.4) is captured by the
following relation between the involved Hilbert spaces
(7.10)

where EB is referred to as the direct sum of the spaces involved.

170

7.2.2

Signal and System Spaces

The L 2 Norm of the Weighting Matrix

Recall, from Chapter 1, that the zero state response of a single-input, single-output
system can be calculated from its state model as

(7 .11)
with impulse response denoted y 1 ( t) being given as

(7 .12)
In addition recall from Section 1.8.3 that the causality constraint requires

YJ(t)

for all

<0

Therefore the square of the L 2 norm of YJ(t) for stable systems can be determined by
recalling the observability and controllability Gramians, W 0 , We from Chapter 4. Thus
we have

(7.13)

Alternatively, since YJ(t) is a scalar, we can determine the square of the L2 norm ofy1 (t) as

(7.14)

Thus we see from (7.13) or (7.14) that the impulse response of a stable system has finite L2
normandhence YJ(t) E 2 [0,oo).
The extension of the foregoing result to multi-input, multi-output systems is more
involved since, in this case, CeA 1B is a matrix referred to as the system's "weighting
matrix". This terminology is used in recognition of the fact that CeA 1B applies differing
weights to the components in the system's input vector in the calculation of the zero state
response

y(t)

=lot CeA(t-T) Bu(T)dT

Notice that the Laplace transform of a system's weighting matrix is the system's
transfer function. Notice also that the weighting matrix is the impulse response when the
system is single-input, single-output. In what follows we will use the sum of the energies in
each entry in the weighting matrix to generalize the L2 norm of the impulse response to
the multi-input, multi-output case.

Time Domain Spaces

171

l.[.

Suppose u(t) is a vector of length m and y(t) is a vector oflengthp, i.e., the system has
m scalar inputs and p scalar outputs and the weighting matrix, CeAt B, is p x m. Then if the
input is

u(t) = u~(t) = I~8(t)

where I~ is the ith column of them x m identity matrix, then the output is

where Ii is the i1h column of the n x m matrix B.

Now we want to calculate the sum of the energy in each of the outputs in the set of
outputs {y}(t) : i = 1, 2, m} resulting from the application of each input in the set
{u} : i = 1, 2, m} respectively. Clearly, we can write the desired energy as

(7.15)
where y1 (t) is the pm dimensional vector

y}(t)
yi{t)

y'J'(t)
Now we can rewrite (7.15) as

IIYI(t)ll2

(trace[fooc (CeA 1B)*CeAtBdtJY

(7.16)

(trace[B*W0 B])!

(7.17)

Alternatively, recall (Theorem 4.2) that for any matrix M we have

trace[M* M]
Therefore setting M

= trace[MM*]

CeAtB, we see that (7.16) can be rewritten as

(7.18)
Finally, since II.Y1 (t)ii 2 , (7.16), is the sum of the energy in each scalar signal in the p x m

172

Signal and System Spaces

matrix CeAr B, we define the L 2 norm of the p x m matrix CeAr B as

(7.19)
which we see from (7 .17, 7 .18) can be written in either of the following two forms
IICeA 1BII 2 = (trace[B*W 0 B])~

(7 .20)

= (trace[CWcC*])~

(7.21)

We will see in Section 7.3.6 that the foregoing time domain L 2 norm of the weighting
matrix equals an appropriately defined frequency domain L 2 norm of the system's
transfer function which is referred to as the H 2 norm.

7.2.3

Anticausal and antistable systems

We begin by considering the single-input, single-output case. Suppose u(t) E -2( -oo, 0].
Then substituting negative t in (7 .11) and considering t decreasing, i.e., reversing the
direction of integration in (7 .11 ), we see that the output for negative time is given as
t E

[O,oo)

or
t E ( -oo,

OJ

(7.22)

Notice that this system operates in negative time, i.e., runs backward in time. Thus an
impulse input at the time origin, u(t) = b(t), affects the output at earlier (negative) times.
This effect is contrary to nature where dynamic processes have the general behavioral
property that the output now is independent of the future input and is solely caused by
past initial conditions and the cumulative effect of the input over the past interval of time.
Natural systems having this behavior, where the cause precedes the effect, are referred to
as causal systems. Artificial systems having the predictive behavior exhibited in (7.22),
where the effect precedes the cause, are referred to as anticausal systems.
Now if we let u(t) = b(t) then we see from (7.22) that the impulse response of an
anticausal system is given by

YJ(t)

= -CeA 1B

Yr(t)

t E ( -oo,

OJ

(7.23)

for all t > 0

Notice that y r ( -oo) is zero only if- A is stable, i.e. only if A has all its eigenvalues in
the open right half-plane or equivalently, only if no eigenvalue of A is in the closed left
half-plane. Systems that have this property are referred to as being antistable. Recall that

173

Frequency Domain Hilbert Spaces

a system is unstable if at least one eigenvalue of its A matrix lies in the closed right halfplane. Thus we see that antistable systems are a special class of unstable systems.
The introduction of this class of system allows us to express the transfer function of an
unstable system, provided it has no imaginary axis poles, as the sum of two transfer
functions, one for a stable system and one for an antistable system. We can obtain this
sum decomposition by expanding the given transfer function in partial fractions.
The foregoing readily extends to the multi-input, multi-output case as follows. If an
anticausal system is antistable the system's weighting matrix satisfies
lim W(t)

l-----+-00

=0

and the system's A matrix has no eigenvalues in the closed left half-plane. Moreover, the
system's zero state response is given by

y(t)

=-lot

W(t- T)U(T)dT

W(t) = -CeA 1B

t E (-oo,O]

so that
0

II W(t)llz=

(trace[["" W*(t) W(t)dt])

Finally, in summary, we have shown that the weighting matrices for causal systems
which are stable satisfy

whereas the weighting matrices for anticausal systems which are antistable satisfy

W(t) E C2( -oo, OJ

7.3

Frequency Domain Hilbert Spaces

We begin this section by relating the time domain Hilbert space C2 ( -oo, oo) to the
frequency domain Hilbert space 2 . This relation is made possible by Parseval's theorem
in the general theory of the Fourier transform. Following this we introduce two frequency
domain Hilbert spaces, 1i2 and Ht which constitute an orthogonal decomposition of C2
and are known as Hardy spaces.

7.3.1

The Fourier transform

Recall that the Fourier transform and the inverse Fourier transform are defined as

F[.f(t)]
.f(t)

F(jw) = l:.f(t)e-Jwtdt

I
=27f

loo
-oo

F(jw)e1wt dw

(7.24)
(7.25)

174

Signal and System Spaces

Also recall the following well known sufficient condition for the convergence of the
integral defining the Fourier transform of[(t), (7.24)

1: [f*(t)f(t)]~dt

<

(7 .26)

where again [Pis to be interpreted as the positive square root. Notice that whenf(t) is a
scalar, (7.26) reduces to f(t) being absolute value integrable, i.e.,

1:

lf(t) I dt <X

(7 .27)

Signals satisfying (7 .26), or (7 .27) if appropriate, are said to be in the Lebesgue space
.C 1 ( -x, x).

Now we are interested in signalsf(t) E .C 2 ( -x, x ). Therefore we need to be concerned

about the convergence of the Fourier integral, (7.24) whenf(t) E .C 2 (-x,x),J(t)~
.C 1 ( -x, x). It turns out that the Fourier integral of signals in this latter subspace
converges for allmost all w except for, what is referred to as, a set of measure zero. This
means that Parseval's theorem

00

00

1!00

J*(t)f(t)dt = 2 71"

-oc

F*(jw)F(jw)dw

(7 .28)

applies to allf( t) E .C 2 ( -oo, oo) independent of the values assigned to F(jw) at points w
where the Fourier integral does not converge. This fact is discussed further in the next
subsection.
The important conclusion to be drawn from the foregoing is that the L2 norm of a
signal in the time domain, llf(t)/1 2, equals an appropriately defined L 2 norm in the
frequency domain, IIF(Jw) ll2,
llf(t) ll2

IIFUw) ll2

(7.29)

where
llf(t)ll2 = [J:!*(t)f(t)dtr
I
IIF(Jw)/12 = [ 271"

Joe F*(jw)F(jw)dw]~
oc

(7.30)

(7 .31)

Notice that the definition of the frequency domain L 2 norm, (7.31) includes the scaling
factor (27r)- 1 which is missing from the definition of the time domain L 2 norm, (7.31).
Any F(jw) satisfying

//F(Jw)// 2 <

is said to belong to the L 2 frequency domain space denoted .C2. Notice that the frequency
and time domain L 2 norms are denoted by .C 2 and .C 2 ( -x, oo) respectively.

Frequency Domain Hilbert Spaces

175

Now Parseval's theorem, (7.28), allows us to calculate the time domain inner product
of/1(t),J2 (t) E .C2(-oo, oo) in the frequency domain as

<JJ(t),J2(t) > = < F1(jw),F2 (jw) >

where F;(Jw)

(7.32)

= F[fi(t)] and
<!J (t),Ji(t) >= 1:ft(t)f2(t)dt

(7.33)

(7.34)

Thus .C2 is an inner product space which can be shown to be complete. Therefore .C2 is a
Hilbert space with inner product defined by (7.34).
The reason which make it possible for us to extend Parseval's theorem from
.C 1( -oo, oo) to .C2(-oo, oo) arises from the theory of Lebesgue integration which is
beyound the scope of this book. However the following subsection is included in an
attempt to provide some insight into this matter.

7.3:2

Convergence of the Fourier integral

Supposef(t) is specified as

f(t)

(t + l)- 1

t::=:o

=0

(7.35)

t<O

'Then

f(t) tt.CI ( -oo, oo)

or more specifically

lo

oo

--ldt =
t+

00

However we can determine the Fourier transform of (7.35) at all frequencies except
0. To see this we need to show that

w=

lo

oo

-1-sin(wt)dt < oo
+t

lo

and

oc

-1 -cos(wt)dt < oo
+t

We can do this by noting that

lo

oc

=L
00

-1 -sin(wt)dt
+t

ck

k=O

where fork= 0, 1, 2 we have

tk+l

ck =

krr

tk=-

-1 -sin(wt)dt

+t

(7.36)

176

Signal and System Spaces

Then since f( t) is a positive monotonical ly decreasing function, the foregoing series is

an alternating series satisfying
lim

ck =

0, 1, 2 ...

k---+XJ

which is known to converge. In a similar manner we can show that the integral involving
coswt converges. Therefore in this example (7 .24) defines F(jw) unambiguou sly for every
w except w = 0.
Now since F(jO) is undefined in this example, suppose we set it to n. Then we would
find that the L 2 norm IIF(Jwll~ is independent ofn. This fact is stated mathematica lly, in
this case, by saying that the Fourier transform of[(t) specified as (7.35) is uniquely
determined for all w E ( -oo, oo) except w = 0 which is a point of measure zero.
The foregoing example demonstrate s the fact that, in general, whenf(t)~.C 1 ( -oo, oo)
andf(t) E .C 2 ( -oo, oo) the Fourier transform F(jw) is defined uniquely as an element of
the Hilbert space .C2 but is only determined "almost everywhere" as a point function of
jw, i.e., is uniquely determined except on a set of measure zero. Therefore the Fourier
transform is a unitary operator between Hilbert spaces .C 2 ( -oo, oo) and .C 2 since for any
f 1 (t),J2 (t) E .C 2 ( -oo, oo) we have
(7 .37)
where the inner product on the left is defined by (7.34) and the inner product on the
right is defined by (7 .33). Hilbert spaces which are related in this way are said to be
isomorphic and the unitary operator between the Hilbert spaces is called an isomorphism. In the present case the isomorphism is the Fourier transform and .C 2 ( -oo, oo) is
isomorphic to .C 2 .

7.3.3

The Laplace transform

In order to characterize the properties of the Fourier transforms of causal and anti causal
signals in .C 2 [0, oo) and .C2 ( -oo, OJ, respectively, we need to define two Laplace transforms, the usual one denoted .C+[f(t)] for causal signals and one denoted .C_[f(t)] for
anticausal signals. These transforms are defined as
.C+[f(t)] =

~~1~ f(t)e-' dt

Re[s] >a+

(7 .38)

.C_[f(t)] =

~~1~f(t)e-' 1 dt

: Re[s] <a_

(7 .39)

where a, a+, and a_ are real scalars with a being positive.

Recall that the existence of the Laplace transform, .C+ [f(t)], requires that[( t) have a
lower bounded exponential order, i.e., there must exist a finite real scalar a+, called the
abscissa of convergence , such that
lim f(t)e-at = 0
t~oo

= 00

(7 .40)

I,
I

Freque,.cy Domain Hilbert Spaces

177

This implies that .C+[f(t)] has no poles to the right of a line called the axis of
convergence which is parallel to the imaginary axis and which cuts the real axis at the
abscissa of convergence, s =a+. Now it turns out thatf(t) E .C2 [0, oo) thenf(t) satisfies
(7.40) with a+ = 0. Therefore the integral (7.38) converges for Re[s] > 0 when
f(t) E .C2 [0,oo). This implies that .C+[f(t)] is devoid of poles in the open right halfplane whenf+(t) E .C2 [0, oo ).
We can use a similar argument to show that .C_ [/( t)] is devoid of poles in the open left
half-plane whenf(t) E .C2 ( -oo, 0]
Now suppose we are given a causal signalf+(t) E .C2 [0, oo) and an anticausal signal
f_(t) E .Cz(-oo,O]. Then we have

f+(t)e-a+t E .C 1 (-oo,oo)

a+ E (0, oo)

f_(t)e-a_t E .C,(-oo,oo)

a_ E (-oo,O)

and

F[f+(t)e-a+t]

= l:f+(t)e-a+te-Jwtd t = F+(a+ + jw)

F[f_(t)e-a_t] = l:f-(t)e-a_te-Jwtd t = F_(a_

+ jw)

for any real scalars a+ > 0, a_ < 0.

Recall that the Fourier integral, (7.24), converges for almost all w when
f(t) E .C2 ( -oo, oo ). Therefore if we take a+ and a_ zero we have the Fourier transforms
, forf+(t) andf_(t) as

= .C+[f+(t)]ls=jw
F_(jw) = .C_[f_(t)]ls=jw

F+(jw)

(7 .41)

with {F+(jw) :wE S+}, {F_(jw) :wE S_} being sets of arbitrary finite elements where
S+,S- are sets of measure zero. Then we see thatf(t) E .Cz[O, oo) (j(t) E .Cz(-oo,O]) has
a Fourier transform almost everywhere on the imaginary axis if and only if the Laplace
transform .C+[f(t)] (.C_ [f(t)]) has no poles in the closed right (left) half-plane. Functions
having these properties are said to belong to the Hardy space 'Hz ('H~-).

7.3.4

The Hardy spaces: 1{2 and

Hi

In this section we will introduce frequency domain spaces denoted 'Hz and 'Ht so that we
can write any F(jw) E .Cz as

where

These spaces are called Hardy spaces after the mathematician G. H. Hardy who
carried out extensive studies on these and other related spaces, e.g., 'H00 space.

Signal and System Spaces

178

The definition of the Hardy spaces H 2 and Ht involve ideas from the theory of
a
complex functions of a complex variable. A function of this sort is said to be analytic at
are
exist
not
does
e
derivativ
point, if it has a derivative at that point. Points where the
referred to as singular points. The only singular points a rational function has are its
poles.
H 2 consists of all complex functions of a complex variable which are analytic at every
point in the open right half-plane and have a finite L 2 norm, (7.31). Alternatively, the
Hardy space Ht consists of all complex functions which are analytic in the open left halfis
plane and have a finite L2 norm, (7.31). The analytic requirement in these definitions
either
in
function
No
spaces.
these
for
norm
needed to ensure that (7 .31) is a well defined
H 2 or Ht can have imaginary axis poles.
In the rational case, H 2 consists of all strictly proper functions which have no poles in
the closed right half-plane, whereas Ht consists of all strictly proper functions which
have no poles in the closed left half-plane. In summary, if F(s) is real rational then
(i) F(s) E H 2 if and only if F(s) is strictly proper and has no poles in the closed right
half-plane.
(ii) F(s) E Ht if and only if F(s) is strictly proper and has no poles in the closed left halfplane.
(iii) F(s) E 2 if and only if F(s) is strictly proper and has no poles on the imaginary axis.
Notice that if F(s) is the transfer function of some system, then in case (i) the system is
stable, and in case (ii) the system is antistable.

7.3.5

Decom posing 2 space

Recall that the L 2 time domain Hilbert spaces decompose as

Then since the Fourier transform maps 2[0, oo) onto H 2 and 2( -oo, OJ onto Ht and the
same inner product is used for 2 , H 2 and 1-if, we have

(7 .42)
and the Fourier transform is an isomorph ism such that

2 ( -oo, oo)
2[0,oo)
2( -oo, 0]

is isomorph ic to
is isomorph ic to

2
H2

is isomorph ic to

'Hf

As an illustration, suppose we are given the following signal

f(t)

where a, bare real and positive.

=e-at

t 2::0

Frequency Domain Hilbert Spaces

179

Then we can writef(t) as

f(t)

= f+(t) + f_(t)

where

f+(t) =e-at f_(t) = 0

f_(t) =it
f+(t) = 0

fort~

fort :s; 0

Since the Laplace transforms of these signals are given by

+[1'-..(t)]

= - 1-

Re[s]

s+a

1
C_[f_(t)] = - b
s-

>-a

Re[s] < b

we see that both C+[f+(t)] and C_[f_(t)] are free of poles on the imaginary axis.
Thereforef(t) has a Fourier transform which can be written as
1

F[f(t)] = -.- b + -:-------+a

JW-

JW

However +[f+(t)] and C_[f_(t)] are each zero at w = oo with +[/+(t)] being
analytic in the closed right half-plane and_ [J_ (t)] being analytic in the closed left halfplane. Therefore we have
C_[f_(t)] E Hf

F[f(t)]

The foregoing example demonstrate that, in general, systems with transfer function
,~
satisfying G(jw) E 2 can be written as

G(s) = G 1 (s)

+ G2 (s)

where

G2(s)

7.3.6

Hf

The H2 system norm

Recall from Section 7.2.2 that when the state model for a given system is minimal, the time
domain L 2 norm of the weighting matrix is finite only if A is stable and Dis null. Therefore
in this case the system's transfer function

G(s)
is strictly proper, i.e., G( oo)

.C+ [CeAt B]

= C(sl- A)-I B

0 and has no poles in the closed right half-plane. Thus we

Signal and System Spaces

180

have G(s)

7-i 2 and using Parseval's theorem we have

[[CeA 1B[[ 2

(trace[lx (ceA 1 B)*(ceA 1B)drJY

=(trace [L 1: G*(jw)G(jw)dw] Y= [[G(Jwll2

(7 .43)

Thus the time domain L 2 norm of the system's weighting matrix equals the frequency
domain L2 norm of the system's transfer function, G(s). Moreover, since in this case we
have G(s) E 7-i 2 , this norm is usually referred to as the system's H 2 norm.
Notice that if G(s), U(s) E 7-{ 2 with G(s) rational, then G(s) U(s) E 7-{ 2 since each term
in the product is analytic in the open right half-plane has no poles on the imaginary axis
with the product being zero at infinity. Thus the zero state response from a stable causal
system having a strictly proper transfer function satisfies Y(s) E 7-{ 2 or y(t) E 2[0, oo)
for any input u(t) E 2 [0, oo ).
In order to extend this result to the case where the transfer function is only proper we
need another Hardy space. In the next section we will see that transfer functions of stable
systems which are only proper, i.e., G(oo) -=f. 0, lie in a Hardy space denoted 7-iCXJ. Thus if
U(s) E 7-{ 2 and G(s) E 1ix then the product G(s)U(s) is analytic in the open right halfplane and has no poles on the imaginary axis. Clearly, in the case when U(s) is rational we
have G( oo) U( oo) = 0 and Y(s) E 7-i 2 or y(t) E 2 [0, oo) for any input u(t) E 2[0, oo ). It
turns out that this result holds in the more general case when U(s) is irrational. Thus we
have

Y(s)

G(s)U(s)

(7.44)

H2

when

G(s)

7-i 00

Just as there is a frequency domain norm for functions in either 1i2 or Ht, namely the
L 2 norm, we will show in the next section that there is a norm for functions in Hem called
the H 00 norm. Thus 7-{ 00 is a normed space. However, unlike 7-{ 2 or Ht there is no inner
product defined on 7-i 00 . Therefore H 00 is not a Hilbert space.
Now the space of rational transfer functions that have no poles in the closed left halfplane and are finite at infinity is denoted 7-i~. Notice that we do not use 1i~ to denote this
space since the concept of orthogonality is missing. Therefore in addition to (7.44) we
have the dual result

Y(s)

G(s)U(s)

Ht

(7.45)

when
G(s)

E 7-i~

U(s)

Ht

We will see in the next three chapters that the Hardy spaces, 1i00 , 1i~, 1i 2 and Ht, that

The H 00 Norm: SISO Systems

181

we are examining in this chapter are essential to the solution of the Hoo control problem.
This solution consists of a feedback controller which stabilizes the plant and constrains
the Hx norm of the plant's transfer function, from disturbance input to desired output, to
be less than a specified scalar. The Hx norm is introduced now as follows.

7.4

The H00 Norm: SISO Systems

At the end of the previous section we introduced the H 2 norm of a system's transfer
function, IIG(s)lb by relating it to the time domain L2 norm of the system's impulse
response, CeAt B. In this section we introduce another type of frequency domain
system norm referred to as the "H infinity norm" denoted IIG(s)lloo- We do this now
by relating IIG(s)llx to an equivalent time domain norm referred to as the "L2 system
gain".
Suppose we are given a stable single-input, single-output system having zero state
response y(t) when the input is u(t). Then we define the L 2 system gain, /'o, as
/'o

= sup lly(t)ll2
. u(t)E2[0,oo) llu(t)ll2

(7.46)

u(f),'O

where "sup" is the abbreviation for supremium or smallest upper bound of a set, which in
this case is the set of positive real numbers generated by the ratio of the L 2 norms in (7.46)
as u(t) is varied over 2 [0, oo). Notice that "sup" is used instead of"max" in case the set
over which the search is carried out does not contain a maximal element. In addition,
notice that we do not consider the null input to avoid dividing by zero.
Now the use of a signal norm to define a system norm as in (7.46) is referred to by
saying that the signal norm induces the system norm. Therefore the L 2 system gain, (7 .46),
is induced by the L 2 signal norm. Notice that, for a given system, this norm is, in general,
not equal to the L 2 norm of the system's impulse response, CeAt B.
Since we are assuming the system is stable, we saw in the previous Sction that
y(t) E 2[0,oo) when u(t) E 2[0,oo). In addition we can use the fact that 2[0,oo) is
isomorphic to 1t2 to rewrite the L 2 system gain, (7.46) in the frequency domain as
l'o =

sup II Y(s)ll2
U(s)EJt2ll U(s) ll2

(7.47)

U(.<),'O

Therefore comparing (7.46) and (7.47) we see that the L2 system gain in the time
domain is equivalent to the frequency domain system norm induced by the H 2 norm. This
norm is referred to as the "H infinity norm", denoted as IIG(s)ILx: Notice that this norm
and the H 2 norm (Section 7.3.6), for a given system are, in general, not equal.

7.4.1

Transfer function characterization of the H 00 norm

Suppose we are given a stable system and that we can choose the system input so that the
ratio in the definition of the L2 system gain, (7.46), is close to being maximized. More

Signal and System Spaces

182

specifically suppose the zero state response, Yopt(t), caused by the input

U 0 p1 (t)

satisfies

(7 .48)
where Eisa small positive scalar in the sense that E lo
Then recalling that linear systems satisfy the principle of superpositio n, we have
ay 0 p1 (t) when au 0 p1 (t) for any constant a. Therefore (7.48) can be written as

where
Ynor(t) = O:Yopt(t)
This shows that we can obtain the L 2 system gain, /o, (7.46), or H 00 system norm,
IIG(s)lloo, (7.47), by restricting u(t) or U(s) to satisfy the following conditions:

(i)

u(t)E.C 2 [0,oo)

or

U(s) E H 2

(ii)

llu(t)ll2 =I

or

IIV(s)ll 2 =I

Thus (7.46, 7.47) become

'Yo= sup II y(t)ll2 =

sup II Y(s)ll2

(7 .49)

Notice that

(7 .50)
where
O'max

sup IG(Jw)l
wE(-oo,x)

(7.51)

and 0' max is referred to as the Lx norm of the function G(jw).

However, since U(s) is restricted in (7.49) as
IIU(s)ll2= (

-1
1

27r

00

-oo

~
2
IU(Jw)l dw) = 1

(7 .52)

we see that (7.50) becomes

-1
27r

00

-oo

IG(Jw)U(jw) l 2 dw :S: O'~ax

(7.53)

The H 00 Norm: SISO Systems

183

Now ittums out that we can choose U(jw) in (7.53) subject to (7.52) so that the left side
of (7 .53) is arbitrarily close to a~ ax. This implies that the supremium in (7.49) is D'max and
we have
lo

O'max

IG(J.,.:)! = IG(Jw)llx

}UP

(7.54)

~ec.-x,x)

Notice that
sup IG(iw)l < oo
wE(-oo,oo)
provided G(s) has no imaginary axis poles. However we have
sup
u(t)E2[0ov)
u(t)#O

lly(t)ll2 <
llu( t) ll2

00

only if G(s) has no poles in the closed right half-plane, since y(t) is unbounded for some
u(t) E 2[0, oo) othenvise. Therefore IIG(Jw)lloo is referred to as:
(i) the L 00 norm of G(s) when the system is unstable and G(s) has no poles on the
imaginary axis including s = oo;
(ii) the H 00 norm of G(s) when the system is stable and G( :x) < oo.
In addition, transfer functions that have bounded H 00 (L 00 ) norm form a normed space
denoted 'H 00 ( 00 ). Notice that 'Hoo C 00 More is said about this in the next subsection.
Now there are two different ways of interpreting IIG(Jw) lloo One arises from defining
IIG(I.u)iloc as

IG(J..v)ll

=
00

sup II Y(jw)ll2
U(iw)E2II U(jw)ll2
U(Jw)#O

Then in this case we are considering G(jw) to be an operator which maps C(j;_,)) E 2 to
Y(jw) E 2 and IIG(Ju.:)lloo to be the operator norm induced by the frequency domain L2
norm or, if the system is stable. by the time domain L2 norm.
Alternatively, if we take (7.54) to be the definition of IIG(Jw)lloc then we are
considering G(jw) to be a function of w and IIG(Jw) lloo is a function norm.

7.4.2

Transfer function spaces

Since any proper rational transfer function with no imaginary axis poles has a finite Loo
norm, we can define three normed transfer function spaces depending on the location of
the transfer function's poles. Therefore assuming the rational function G(s) is proper we
have
(i) G(s) E Hoc when all poles of G(s) are in the open left half-plane
(ii) G(s) E 'H~ when all poles of G(s) are in the open right half-plane
(iii) G(s) E L 00 when G(s) has no poles on the imaginary axis.

184

Signal and System Spaces

Notice that when G(s) E Lx we can use the partial fraction expansion (after one cycle
of division) to write G(s) as

'Nhere

Then assuming some scheme for making G+(s), G-(s) unique, e.g., we can insist that
c- (s) be strictly proper, we can capture this spliting of a transfer function in L 00 into the
sum of stable and antistable transfer functions as a direct sum decomposition of the
spaces involved
(7 .55)
Alternatively, suppose two stable systems having transfer functions G1 (s), G2(s) E H)Q
form the components of a cascade connection. Then the transfer function of the
composite system, G 1 (s)G 2 (s), is stable since its poles are contained in the poles of the
component transfer functions. In addition, the transfer function of the composite system
is proper or strictly proper since the product of proper rational functions is proper
independent of any pole-zero cancellations that might occur. Therefore G1 (s)G 2(s) E Hoo
ifG 1 (s),G2(s) E rt 00
Now we can readily show, from the definition of the H 00 norm, (7.47), that the H 00
norm of the transfer function of the cascade connected system is related to the H 00 norms
of the transfer functions of the component systems as

(7.56)
This result plays an important role in control theory. An example of its use is given in
the following subsection.

7.4.3

The small gain theorem

Prior to the use of the H 00 norm in control theory, the H 00 norm appeared in a number of
different guises in various engineering applications. For example, the resonant gain used
to characterize the performance of an electric filter is the H:XJ norm of the transfer
function relating the filter's input and output voltages. Another instance of this occurs
when the Nyquist stability criterion is applied to determine the stability of a feedback
system having transfer function T(s)

G(s)
T(s) = 1 + G(s)H(s)
Then when G(s), H(s) E Hoc the distance from the origin to the point on the polar plot of
G(Jw)H(Jw) which is furthest from the origin is IIG(s)H(s)lloo This geometrical
observation leads to the following result which is known as the small gain theorem.
Recall that when G(s)H(s) E 1i 00 , the Nyquist criterion states that the feedback
system is stable if and only if the polar plot of G(Jw)H(jw) does not encircle the point

The H 00 Norm: NIIMO Systems

185

-1 + jO. Therefore since the foregoing geometrical interpretation of IIG(s)H(s)lloo

implies that the polar plot of G(jw)H(jw) cannot have any encirclements of -1 + jO if

IIG(s)H(s)l\x< I

(7.57)

we have (7 .57) as a sufficient condition for the feedback system to be stable when
G(s)H(s) E 'Hoc This result, referred to as the small gain theorem, is important in
connection with robust stabilization where a single fixed controller is required to stabilize
any one plant in a set of plants. A simplified example of the application of the small gain
theorem in this context is given as follows.
Suppose we are given a fixed controller having transfer function H(s) E 'H 00 which
satisfies

IIH(s)llx:s; a

(7.58)

Then using (7.56) we can show that (7.57) is satisfied if

1

IIG(s)lloo<a
.

(7.59)

Therefore any stable plant satisfying (7.59) is stabilized by the given fixed controller.
Notice that by interchanging the role of the controller and plant, the foregoing
development leads to a characterization of a set of stable controllers any one of which
can be used to stabilize a given fixed stable plant.
This completes the development of the H 00 (L 00 ) norm for single-input, single-output
systems. We need now to proceed to develop this norm for multi-input, multi-output
systems.

7.5

The H00 Norm: MIMO Systems

We have just seen that the H 00 norm of a single-input, single-output systefh is the H 2
induced norm of the system's transfer function, G(s). In order to extend this idea to the
multi-input, multi-output case, where G(s) is a matrix, we need to introduce the induced
2-norm for constant matrices, denoted as IIG(jw)ll (for fixed w). Then we can obtain the
H 00 norm of the matrix G(j w) as the largest induced 2-norm as w varies over (-oo, oo ),
i.e.,

IIG(jw)lloo=

7.5.1

sup (IIG(jw)ll)
wE(-oo,oo)

Singular value decomposition

Before we can develop the induced 2-norm for a constant matrix we need to introduce the
singular value decomposition (SVD) of a constant matrix. This is done as follows.
Recall, from Chapter 4, that Hermitian matrices have real eigenvalues and orthogonal
eigenvectors. Moreover notice that the product matrices M* M and MM* formed from
any complex matrix M, are Hermitian and nonnegative. Therefore these product

186

Signal and System Spaces

matrices have orthogonal eigenvectors and real nonnegative eigenvalues. This general
observation leads to the matrix analysis technique known as SVD. The following
theorem, which is proved in the appendix, defines the SVD of a matrix.
Theorem 7.1 Any p x m matrix of complex constants, M, which has rank r can be
decomposed as
M = UI:V*

(7.60)

where U, and V are p x p and m x m unitary matrices with

I:=[~ ~]
U1 is p

U2 is p

V1 ism x r

V2 ism

p- r
X

m- r

and I: 0 is an r x r, real, positive definite, diagonal matrix denoted as

a!

a2

a,

= diag[a 1, a 2, ,a,]

I:o =
0

with diagonal entries referred to as singular values and ordered so that

i = 1, 2, , r- 1
Proof See Appendix
The SVD of a matrix has become ever more useful since the establishment of practical
methods for its computation in the early 1970s. The SVD is of practical use wherever we
require the solution of problems involving matrices having small rank. One problem of
this sort, which we encountered in Chapter 4, involves the determination of a lower order
state model approximation of a state model having controllability and/or observability
matrices which are almost rank deficient, i.e., the state model is almost uncontrollable
and/or unobservable. We were able to obtain a solution to this problem without requiring
the SVD by using the state model's Gramians. Since the Gramians are nonnegative
Hermitian matrices, they enjoy the nice properties which make the SVD so useful,
namely, real and nonnegative eigenvalues and mutually orthogonal eigenvectors.

7.5.2

Induced 2-norm for constant matrices

We now proceed with the development of the constant matrix norm induced by the 2norm for constant vectors. This is done in the proof of the following theorem by
employing the SVD and its properties.

;~

The H 00 Norm: MIMO Systems

187

Theorem 7.2 Any constant p x m matrix J1 has a norm induced by the 2-norm for
constant vectors'' hich equals a 1 the largest singular value of .\1.
Proof The induced 2-norm for a matrix M, ' \1 , is defined as
IIMII =sup IIMxll
J!xll
xEC

(7.61)

xi=-

where em is the vector space of all vectors of length m having constant complex entries
and llxll equals the positive square root of the sum of the squared absolute value of each
entry in x, i.e.,
(7.62)

Notice that unlike the L 2 norm for a time varying vector which we used earlier, the 2norm for a constant vector has no subscript, i.e .. lll 2 denotes the L 2 norm whereas 1111
denotes the 2-norm.
Next since the transformation J;fx is linear. multiplying x by a constant scalar changes
both the numerator and the denominator in (7.61) by the same amount so that the ratio in
(7.61) remains unchanged. Therefore we have the alternative definition of the induced 2norm of M given by
IIJ1II =sup' ,Vfxll

(7.63)

Since them x m matrix V in the SVD of M, (7 .60), is invertible, we can use its columns
{ 1./ : i = l, 2, m} as a basis for em. Thus we can express any x E em as
m

x= La;v' = Va

(7.64)

i=O

where
with a E em.
Now, since Vis unitary we have

111

= a*a = Lla;l 2 = llall 2

(7 .65)

i=O

and the 2-norms of x and of a are equal. In addition since Vis invertible there is a unique
a for each x and vice versa so that
supiiMxll = supiiMVaJJ
xECm

cxECm

11<11~1

llall~l

(7.66)

188

\~

Signal and System Spaces

Then using the SVD of M, (7.60) we can write

IIMVall =

(a*V*M*MVa)~

* ~2 ~-(a
I:ma)--

where

2
2)
r
~O"iln;l

(7 .67)

I:; is them x m diagonal matrix given by

I:o2 = d"Jag [0"1,2 0"2,2 , O",2].

Therefore from (7.63), (7.66) and (7.67) we have

II

Mil~ b~t.aila;l 2 r

(7 .68)

Now the summation on right side of the foregoing equation satisfies the following
inequality
r

L O"flail

2 ::;

O"T Llail 2

i=l

(7.69)

i=l

where 0" 1 is the largest singular value of M. However since a is restricted in (7 .68) to have
unit norm, we see from (7.69) that
(7.70)
Notice that if the components of a satisfy
i?2
then

llall =

1 and
r

L O"flail = O"T
2

i=l

which determines the supremium in (7.68) as

II Mil= 0"1

Notice, in the foregoing proof, that the supremium is achieved for a = I 1, the first
column of them x m identity matrix. Therefore, letting a= I 1 in (7.64) gives the result
that the unit 2-norm x that gives the supremium in (7.63) is the right singular vector of M,
v 1, corresponding to the largest singular value of M, 0" 1 . Thus

IIMxll

~=0"]

,.

189

The H 00 Norm: MIMO Systems

when x = (3v 1 for any complex scalar, (3. Further consideration of this fact leads us to the
important conclusion that

!IMxll :S 0"1 llxll

for all

E Cm

(7.71)

with equality onlywhenxisa scalar multiple ofv 1 . Finally, notice that ' JIII =
M is a scalar.
1

7.5.3

'Jfl when

The Lx, H 00 norm for transfer function matrices

Having developed the idea of the induced 2-norm for constant matrices, we are now in a
position to give the definition of the Lee norm of a transfer function matrix, G(jw). This
norm is defined as the supremium of the induced 2-norm of G(j..,:) over all wE ( -oo, x)

IIG(j,.-.:) lloo=

sup

O"J (i~)

(7.72)

~'C(-oo.:x.

where 0" 1(jw) is the largest singular Yalue of G(j..v).

As in the SISO case, when an MI:\10 system is stable its transfer function matrix G(s)
has finite L 00 norm, which is referred to as the Hoc norm and G(s) is lies in the normed
space denoted by Hx. Alternatively. when the system is antistable, its transfer function
matrix G(s) has finite Lee norm, (7.72), and G(s) lies in the normed space denoted H:;,.
Finally notice that, as in the SISO case, all G(s) with finite Lx norm are denoted b:;
G(s) E 00 . Thus Hcc and H~ are each subspaces of 00 which are disjoint and complete
the 00 space so that (7.55) holds for MIMO systems.
In subsequent deYelopment we will need the following result which is an immediate
consequence of the foregoing.
Theorem 7.3 A real rational transfer function G(s) E Lx has an L 00 norm less than a
finite positive scalar ~., i.e.,

]G(j..v) lloo < !

if and only if F(jw) is positive definite for all w, i.e.,
'Vw E ( -oo, oo)

F(jw) > 0
where

F(jw) ="-/I- G*(jw)G(jw)

Proof Let the SVD (Theorem 7.I) of the p x m matrix G(j w 1 be given as

G(jw) = U(jw)L,(jw) V*(jw)

where U(jw). V(jw) are p x p and m x m unitary matrices respectively and

L,(jw) = [

L,o(.fw)
0

~]

L,0 (jw)

diag[O"J (jw), 0"2 (jw), . O",(jw)]

190

Signal and System Spaces

~..

with r = rank[G(jw] and

cr 1(jw) ~

crJjw)

>1

(7.73)

Then using the SVD of G(jw) we can write F(jw) as

(7.74)

where L:~,(jw) is them dimensional diagonal matrix

L:~7 (jw) = diag [crf(jw), cr~(jw), , cr~ (jw), 0, , 0]

Next pre and post multiplying (7.74) by V*(jw) and V(jw) respectively gives

V*(jw)F(jw) V(jw)

= L:~

(7.75)

where

Now since V(jw) is unitary and therefore invertible, we have

F(jw) > 0

V*(jw)F(jw)V(jw) > 0

if and only if

Thus we see from (7.75) that

F(jw) > 0

if and only if

[-?- crf(jw)] > 0

(7.76)

However, since the L 00 , norm for G(s) is defined as

IIG(Jw)lloo=

sup

wE( -oo,x)

cr1 (jw)

we see from (7.76) that

F(jw) > 0 for all wE ( -oo, oo)

if and only if

IIG(Jw)llx< I

and the theorem is proved.

In the next chapter we will use the foregoing theorem to obtain necessary and sufficient
conditions on the state model parameters, {A, B, C, D}, of a system which ensures that
the system's transfer function has L 00 norm less than a given finite real scalar.

7.6

Summary

We began by formalizing the idea, used earlier in Chapters 5 and 6, of relating the size of a
signal to the energy in the signal. This led to the time domain Hilbert space -2( -oo, oo)
and its orthogonal subspaces 2 ( -oo, OJ and 2 [0, oo ). Following this we used the Hilbert
space isomorphism from 2 ( -oo, oo) in the time domain to 2 in the frequency domain to
introduce the frequency domain spaces H 2 , H~ known as Hardy spaces.

Notes and References

191

Next we introduced the notion of the size of a system. This was done by using the
induced operator norm to characterize a system's ability to transfer energy from its input
to its output. As with signal norms we saw that there is an equivalence between a time
domain system norm known as the system's L 2 gain and a frequency domain norm
known as the Hx norm. Unlike the Hardy spaces, H 2 , H} for signals which are equipped
with an inner product, the Hardy spaces, 'H00 , 1{~ for systems are normed spaces only.

1.1

Notes and References

There are many texts on Hilbert space. A readily accessible treatment of this subject
which includes material related to the control problems being considered here can be
found in [46]. Unfortunately books on Hardy spaces require a considerable background
in functional analysis and complex analysis. Some of these references are given in [14, p.
13}. An interesting treatment of complex analysis as it applies to Laplace and Fourier
transforms is given in [26].
The example following (7.27) in Section 7.3.2 which is used to discuss the Fourier
transform of signals in 2 that are not in 1 is taken from [45, p. 272]. The convergence of
the series (7.36) is discussed in [36, p. 71]. The applicability of Parseval's theorem for
functions in 2 ( -oo, oo) is discussed in [37, p. 185]. An excellent reference for the SVD is
[17]:
The small gain theorem is used to solve a wide variety of robust control problems
including both linear, [47] [18] and nonlinear, [42] [21] [33], systems. Finally a good basic
introduction to the ideas in this and the next chapter can be found in [11].

8
System Alge bra

8.1

Introduction

In the previous chapter we were interested in characterizing signals and systems in terms
of normed spaces in both the time and frequency domains. This led to the introduction of
the frequency domain spaces H 2 , Hi and 2 for signals and 'H00 , H;;, and oc for systems.
In the present chapter we consider a number of operations involving systems in these
spaces. We will be especially interested in obtaining state models for the systems which
result from these operations. To facilitate this endeavour, we use the compact equivalence
relation

G(s)

[AC DB]

(8.1)

to denote that

G(s)

C(sl- A)- 1B + D

.,.
We begin the discussion of these operations by considering the effect of connecting
systems in parallel and in series.

8.1.1

Parallel connection

One of the simplest examples of operations with systems consists of connecting several
systems in parallel so that all systems have the same input and the output from the
connection is the sum of the outputs from each system. The connection of the first order
component systems resulting from a partial fraction expansion of the system's transfer
function is an example of the parallel connection of systems.
The input -output constraints which ensure that r component systems are connected in
parallel are
u(t)

= u;(t)

y(t)

= L:Y;(t)

i = 1, 2, r

(8.2)

r
i=l

(8.3)

194

System Algebra

u(t)

Figure 8.1

Parallel Connection

so that

Y(s) = G(s)U(s)
where
r

G(s)

L G;(s)
i=l

with G;(s) : i = 1, 2 r being the transfer functions of the component systems.

Thus assuming we know state models for each of the component systems as

G;(s) ~ [ A'
C;

Bl

'
D;

i = 1, 2, .. r

(8.4)

we want to determine state model parameters {A, B, C, D} for the composite system, i.e.,

G(s)

~ [;

;]

We do this as follows.
First we write the state equations for each component system as

+ B;u;(t)
= C;xi(t) + D;u;(t)

xi(t) = A;xi(t)
Y;(t)

(8.5)

where i = 1, 2, r. Next we use (8.5) together with the constraints imposed by (8.2, 8.3)
to give a state model for the composite system as
x(t)

y(t)

+ Bu(t)
Cx(t) + Du(t)
Ax(t)

(8.6)

Introduction

195

where
AI

A2

Il

A=
0

C=[Ct

c2

C,]

XI (t)
x 2(t)

x(t) =

B=

x'(t)

D=LD;
i=l

Notice that the dimension of the state model for the composite system is the sum of the
dimensions of each component system. MoreoYer. MIMO systems can be connected in
parallel if and only if each of the systems has: (i) same number of inputs, and (ii) the same
number of outputs. Notice in the case of the partial fraction expansion of G(s), the A;s are
the poles of G(s).
Finally notice that if we are given G(s) E Lx then we can use partial fraction
expansion to decompose G(s) into the parallel connection of two systems such that
G(s) = GI (s) + G:c (s J \\here GI(s) E n 00 and C:c (s) E H~ with the eigenvalues of AI (A 2 )
being in the open left (right) half plane. This fact was given in the previous chapter as the
direct sum decomposition of the spaces involved, i.e., 00 =Hoc EB 7-i~.

8.1.2

Series connection

Next consider a series connection of two systems, often referred to as system composition.
The constraints imposed by this connection are

YI(t)=r(t)

(8.7)

so that

Y(s) = G(s)U(s)

(8.8)

where

T\ow we can de\elop the state model for the composed system, by imposing the
constraints (8.7) on the state models for the component systems, (8.4), with r = 2. The
result is as follows.

xi(t)

= A 1x 1 (t)

x2 (t) =

A 2 x:c(t)

y(t) = C 1x 1 (t)

_u_(t)-------i>!)l

G2(s)

Figure 8.2

+BIC:cY 2

+ B2u(t)
+ DI C 2x 2(t) + D1D 2u(t)

1-----0>1)1

G,(s)

Series Connection

1--Y(_t)_

196

System Algebra

which can be rewritten as

y(t)

= [ C1

and we see that the composed system has state model

(8.9)
where

Notice that the constraint u1 (t) = y 2 (t) implies that the number of inputs to the system
labeled 1 must equal the number of outputs from the system labeled 2. Alternatively, we
could connect the systems in reverse order if the number of inputs to the system labeled 2
and the number of outputs from the system labeled 1 are equal. In this case the transfer
function for the composed system would be G2 (s)G 1 (s). Notice that we get the same
composed system independent of the order in which the component systems are
connected in series only if the matrices G 1 (s) and G2 (s) commute so that
G 1 (s)G 2 (s) = G2 (s)G 1 (s). Thus in the SISO case the result of composing systems is
independent of the order in which the component systems are connected.
In the remainder of this chapter we will take up the problem of determining the state
model for several other types of system operation. One of these problems concerns the
determinatio n of the state models for the systems needed in a series connection so that the
resulting composed system has a specified transfer function. This problem is known as the
system factorization problem. The requirement that the factors have certain properties
gives rise to a number of different classes of system factorizations, several of which are of
great importance to the development of control theory.

8.2

System Inversion

Before we begin considering system factorization we need to consider the related problem
of finding a system inverse.
Suppose we are given two SISO systems having transfer functions
b m-1
b m
+ + bm
G (s) = os + 1s
1
s n + a1sn-1 + +an

(8.1 0)

sn+a1sn- 1 ++an
G (s) - - - - - ; - - - - - - , -

(8.11)

boSm

+ b ]Sm-1 + + bm

er function which is unity, i.e.,

Then the series conne ction of these systems gives a transf
Thus it would appea r that the
input.
its
to
equal
the series connected system has its outpu t
system labeled 1 has the system labeled 2 as its inverse.
sinusoidal steady state gain
However, physical processes have the prope rty that their
G ( oo) and G2 ( oo) to be finite.
canno t be unbou nded at infinite. Therefore we require 1
m). There fore we can only have
Now G 1 ( oc) ( G2 (oo)) is finite if and only if m ~ n (n ~
a given physical process having
both G 1 ( oo) and G2 ( oo) finite if m = n. This means that
prope r (not strictly prope r) or
is
(s)
transf er function G 1 (s) has an inverse if and only if G1
D matrix.
ro
nonze
a
equivalently only if the state model for G1 (s) has
has an inverse if and only if
(s)
G
on
1
In the MIMO case, a system having transf er functi
systems having the same
only
Thus
its state models have aD matrix which is invertible.
for the inverse system of
model
state
numb er of inputs as outpu ts can have an inverse. The
s.
a given invertible system can be developed as follow

8.2.1

Inverse system state mod el

for an MIM O system

Suppose we are given the transfer function and state model

~ [~ ~]

G(s)

D- 1 exists

so that state equat ions for the system are

+ Bu(t)

(8.12)

= Cx(t) + Du(t)

(8.13)

x(t) = Ax(t)
y(t)

Then solving (8.13) for u(t) gives

u(t) = -D- 1Cx(t)

+ D- 1y(t)

(8.14)

which when substi tuted in (8.12) gives

(8.15)
outpu t u(t), input y(t), and
Thus we see from (8.14, 8.15) that the inverse system has
state model given as

G- 1 (s)b

[~:

Bx]
Dx

(8.16)

where
Ax =A- BD- 1C

ex=

-D- 1 C

= BD- 1
Dx = D- 1

Bx

useful in developing results in

As we will see, the state model for the system inverse is
contro l theory.

System Algebra

198

Notice that the transfer functions for aninvertible SISO system and its inverse (8.1 O,
8.11) are reciprocals of each other. Therefore the poles of the given system equal the zeros
of the inverse system and vice versa. A similar statement can be made for invertible
MIMO systems. To see this we need interpret the idea of a system zero in terms of a
system's state model. This will be done by first considering system zeros for SISO systems.

8.2.2

S/SO system zeros

Recall that an SISO system has a system zero at s = s 0 if its transfer function, G(s), is zero
for s = s0 or

U(s 0 )

Y(s 0 ) = 0

-=/=

(8.17)

where

Y(s) = G(s)U(s)
Notice that (8.17) implies that G(s 0 ) = 0.
Now in order to arrive at an interpretation of system zeros which we can use in the
MIMO case, we consider a system's zeros in terms of a state model for the system.
Suppose we are given a controllable and observable state model for the system,
s
G(s) =

[AC DB]

(8.18)

Then taking x(O) = 0 since system zeros are defined in terms of the system's zero state
response, the Laplace transform of the state equations for the system yields

[0 ]
-B] [X(s)]
Y(s)
U(s)

[sfC -A

(8.19)

Therefore if s0 is a system zero we have

[ sC0 ! - A ] X(so)

[-

B] U(so)

U(s 0 )

-=/=

(8.20)

which implies that the column

is dependent on the n columns in

Therefore s = s0 is a system zero if and only if the (n

side of (8.19) is not invertible or
rank [

s0 1- A

+ 1)

-B] < n+
D

x(n + 1) matrix on the left

(8.21)

199

System Inversion

As illustration, suppose we are given a second order system in controller form with D
zero. Then (8.19) is
[

s+ a 1
~1

which after eliminating X 1 (s) becomes

(s 2 + a 1s + a2)X2 (s) ~ U(s) = 0

(8.22)

(c 1 1 + c2 )X2 (s) = Y1s)

(8.23 I

Notice from (8.22) that if C(s0 ) i= 0 then X(s 0 ) 7'= 0. Consequently (8.23) implies that
= 0 for U(s 0 ) cJ 0 only if s0 is a zero of the polynomial

Y(s0 )

However

which shows that s0 is indeed a zero of G(s).

By proceeding in the same fashion. we can show. when D cJ 0. that the foregoing state
model has a system zero at s0 which is a root of the polynomial
D(s 2 -L a 1s- a2 ) + c 1 s + c2

which is the numerator of G(s).

The foregoing ideas are used now to extend the definition of a system zero
systems.

8.2.3

to

MIMO

MIMO system zeros

Suppose the system denoted by (8.18) has m inputs and p outputs \Yith m :.:; p and with all
m columns of B independent. Then a complex number s0 is a system zero if
G(so) C(s0 )

=o

for some Us 0 )

(8.24)

Notice that unlike the SISO case \Yhere (8. 17) implies that G(s 11 ) = 0. (8.24) does not
imply that G(s0 ) = 0.
Now with appropriate modification, (8.20) implies that s = s0 is a system zero if at least
one of the m columns of

[-:J

200

[1

System Algebra

\1
I

is dependent on the columns of

Therefore, recalling that the rank of a matrix cannot exceed its smallest dimension, we
see from (8.21) that s0 is a system zero if
rank [

-B] <n+m

s0 I- A

(8.25)

Alternatively, when m > p the matrix in the foregoing inequality cannot have rank
greater than n + p so that (8.25) is satisfied for all s0 and (8.25) is meaningless.
In order to overcome this problem we define s0 to be a system zero if
rank [

s0 I- A

-B]

<p

where
p =max ( rank [

sf- A

sEC

We want now to apply the foregoing ideas to invertible systems. Since invertible
systems are square, i.e., m = p, we can use (8.25) to define the system zeros for this class of
system.

8.2.4

Zeros of invertible systems

Suppose G(s) is invertible. Then we see, in this case, that the condition for s0 to be a
system zero, (8.25), is equivalent to
det [

s 0 I- A

-B] =0

(8.26)

Notice that if the state model is not controllable the foregoing condition is also
satisfied when s0 is an uncontrollable eigenvalue of the system since we have an
eigenvalue-left-eigenvector pair (.A, w) of A which satisfies
wT (M-

A) = 0 and

wT B

= 0

Therefore

(8.27)
and (8.26) holds for)\= s0 . Thus, in this case, s 0 is both a system zero and a system pole.
Alternatively, if (A, C) is an unobservable pair a similar argument can be used to show
that at least one of the eigenvalues of A is both a system zero and a system pole.

Coprime Factorizatio n

201

However, if the state model is both controllable and observable with Dis nonsingular,
then there is no system zero which equals a system pole. In addition, each system zero of
G(s) is a system pole of G- 1 (s) and vice versa. To see this suppose s0 is a system zero. Then
(8.26) is satisfied and is equivalent to
det [[

s0 l - A

(8.28)

where M is any nonsingular matrix of appropriate dimension.

Now suppose we choose M as

~]
Then the left side of (8.28) becomes
det [[

sol- A

-:]

c
=

det[s0 / - A+ BD- 1 C] det[D]

However since G(s) is invertible, we have det[D]

(8.29)

=I 0 and (8.29, 8.28) imply

where

and Ax is the system matrix for G- 1 (s), (8.16). Therefore s0 is a pole of G-l.(s) as well as
being a zero of G(s).
Finally, we will see that invertible systems which are stable with stable inverses play an
important role in control theory. These systems, which are referred to as "units", form a
subspace U 00 of 1{00 and are characterized in terms of their transfer functions G(s) as

G(s)

Uoc

if and only if

G(s), G- 1 (s) E 1i00

In the case of an SISO system, this definition implies that a system is a unit if it has a
transfer function which is proper with no poles or zeros in the closed right-half plane.

8.3

Coprime Factorization

Problems in system factorization are concerned with the determination of two systems,
called factors, whose series connection has the same input-outpu t behavior as ihe system
being factored. Additional constraints on the factors determine different classes of
factorization.

202

System Algebra

Coprime factorization can be characterized as a series connection of a stable system

and the inverse of a stable system with no unstable pole-zero cancellations between
factors. As an example consider the system with transfer function G(s) where
G(s) -

(s+l)(s-2)

_..:...._~..:...._-----'--_

- (s

+ 3 )(s -

4) (s

(8.30)

+ 5)

Then a coprime factorization of this system is indicated as follows:

where

Nt(s)=(s-2) (s+l)
(s + 3)a(s)

Mt(s) = (s-4)(s+5)
a(s)

with a(s) being any polynomial of degree 2 having no zeros in the closed right half plane.
Notice that this choice of degree for a(s) ensures that M 1 (s) is invertible whereas the
restriction on the location of the zeros of a(s) ensures that a(s) is not involved in any
unstable pole-zero cancellations between N 1 (s) and Mj 1 (s). Thus even if a(s) has zeros at
s = -1 and/or -5 the resulting pole-zero cancellations in N 1(s) and Mj 1(s) are allowed
since they are stable. More important, notice that the closed right-halfplane zero of N 1 (s)
at s = 2 and of M 1 (s) at s = 4 are different so that there are no unstable pole-zero
cancellation between N 1 (s) and Mj 1 (s). Stable systems which do not share any system
zeros in the closed right-half plane are said to be coprime.
Notice, from the foregoing example, that the form of coprime factorization,
N(s)M- 1 (s), bears a striking resemblance to the expression for a transfer function in
terms of its numerator polynomial N(s) and denominator polynomial M(s). Of course
the "numerator" and "denominator" in the present context are not polynomials but are
each rational and can therefore be thought of as transfer functions for "numerator" and
"denominator" systems.

8.3.1

Why coprime?

In the foregoing example the numerator N 1 (s) and denominator M 1 (s) are coprime. In
order to better appreciate this fact we give a second factorization of G(s) in which the
factors are not coprime
Suppose were-express G(s), (8.30), as
(8.31)

where
N 2 (s) = (s- 6)(s- 2)(s + 1)

(s

+ 3)f3(s)

() _ (s-4)(s+5)(s -6)

M2 s -

jJ(s)

with f3(s) restricted to be any degree 3 polynomial with no zeros in the closed right half
plane. Although this factorization has the desired property that N 2 (s), M 2 (s) E H 00 with

''"~-~-~~--.

c~~rin: ''ciOr:;iuon

203

(s) and Mz(s) be

Mz(s) being invei:tible, it does not satisfy the require ment that N 2
is not a coprim e
(8.31)
Thus
6.
=
s
at
zero
coprim e since they share a right-h alf plane
unstabl e polehaving
not
of
ance
import
the
factoriz ation of G(s). We can demon strate
.
follows
as
factors
zero cancell ations betwee n
having one of its n
Suppos e that we have a strictly proper scalar transfe r functio n G(s)
- 1 poles being in the
poles, n > 2, at s = s0 on the positive real axis with the remain ing n
as
open left-hal f plane. These specifications imply that G(s) can be written

G(s)

q(s)[s - (s0 +E)]

(s- soJP(s)

0 :::;

<

00,

So

>0

no zeros in the
where the degree of q(s) is less thanp( s) with q(s0 ) f. 0 and p(s) having
is stable only if E = 0
closed right half plane. Now the system having this transfe r functio n
inspect ion of the
so that there is an unstabl e pole-zero cancell ation. However, closer
rely, in practice, on
consequences of such a pole-zero cancell ation reveals that we cannot
This becomes
unstabl e pole-zero cancellations to make an unstabl e system stable. 1
e, :j: [G(s)] = g(t)
immediately evident when we conside r the system's impulse respons

g(t) = Koe501 + r(t)

where

Ko =

-E

q(so)
p(so)

with r(t) bounde d for 0 :::; E < oo.

e response is
Notice that if E = 0, K 0 esot is missing from g(t). Then the system's impuls
501 is present in g(t) and the
e
K
bounde d and the system is stable. Howev er, if E f. 0, then 0
is unstabl e. Thus the
system's impuls e respons e tends to infinity with time and the system
system's stability is catastro phicall y sensitive to E.
01
witlftim e and the
Conversely, if s0 is in the open left-ha lfplane then K 0 e" tends to zero
system is stable for all E.
ro cancellation,
Compa ring the effect on the system behavi or of the two types of pole-ze
y to unstabl e
stabilit
's
system
a
of
stable and unstabl e, we see that the lack of robustn ess
nt to reliable
detrime
a
ation
pole-zero cancell ation makes this type of pole-zero cancell
factors in a
the
of
y
stabilit
the
control system design. We will see that the coprimeness and
control lers
ining
determ
for
coprim e factoriz ation can be exploited to provide a means
additio n
in
but
stable
utput
which not only make a feedback control system input-o
that the
so
plant
and
ler
preven t unstabl e pole-zero cancellations betwee n the control
closed loop system is interna lly stable.
system zeros of
Return ing to coprim e factoriz ation, we saw earlier in this chapte r that
see now, that
we
re,
Therefo
.
an invertible system are poles of that system 's inverse
1
if and only if
ted
preven
are
(s)
Munstabl e pole-ze ro cancell ations betwee n N(s) and
i.e., if and
plane,
alf
right-h
closed
the
M(s) and N(s) have no commo n system zeros in
only if M(s) and N(s), are coprime.
1
G(s) is unstabl e.
In additio n, since N(s) is always stable, M- (s) is unstabl e if
are the only
G(s)
of
poles
e
unstabl
the
e,
coprim
Moreov er, since M(s) and N(s) are

204

System Algebra

unstable poles of M- 1 (s). Finally, since M(s) is stable, M- 1 (s) has no zeros in the closed
right-half plane. Therefore the closed right-half plane zeros of G(s) are the only closed
right-half plane zeros of N(s).

8.3.2

Coprime factorizatio n of MIMO systems

Moving on to MIMO systems, we need to distinguish between two types of coprime

factorization depending on the ordering of the factors. Thus a p x m real rational transfer
function matrix G(s) can be coprime factored as

G(s) = N(s)M- 1(s)

(8.32)

G(s) = M- 1 (s)N(s)

(8.33)

or as

where N(s), N(s) E 7-f. 00 are both p x m and M(s), M(s) E 7-f. 00 are m x m and p x p
respectively. In addition both pairs {N(s), M(s)}, and {N(s), M(s)} must be coprime.
Notice that the two varieties of factorization (8.32, 8.33) are referred to as right and left
coprime factorizations since the denominator is on the right in (8.32) and on the left in
(8.33). In the SISO case these two varieties coincide since products of scalars commute.
We give now, in the theorem to follow, an alternative characterization of coprimeness.
We will see later that this characterization is also useful in the problem of determining
controllers which make the closed loop control system internally stable.
Theorem 8.1 The factors M(s), N(s) E 7-f.DC {M(s), N(s) E 7-f.oo} are right {left}
coprime if and only if there exists X(s), Y(s) E 7-f. 00 {X(s), Y(s) E 7-f.oo} which satisfy
the following Bezout, (pronounced "bzoo") identity

X(s)M(s)
{M(s)X(s)

+ Y(s)N(s)

(8.34)

=I

+ N(s) Y(s) =I}

(8.35)

Proof (it) Consider the SISO case. Suppose M(s), N(s) are not coprime, i.e.,
M(s 0 ) = N(s 0 ) = 0 for some s0 in the closed right-half plane. Then since X(s) and Y(s)
are stable, X(s 0 ) < oo, Y(s 0 ) < oo. Therefore we have X(s0 )M(s0 ) = 0, Y(s 0 )N(s0 ) = 0
and (8.34) cannot be satisfied. This shows, in the SISO case, that satisfaction of the
Bezout identity is sufficient for M(s),N(s) to be coprime. The MIMO case is treated as
follows.
Rewrite the Bezout identity, (8.34), as

(8.36)

L(s)R(s) =I
where

L(s)
and L(s), R(s) E

7-{ 00

[X(s)

Y(s)]

are m x (m + p) and (m

R(s)

+ p)

[ M(s)]
N(s)

x m respectively.

I
I .

Coprime Factorization

205

Next suppose that M(s), N(s) are not coprime with s0 being a zero of M(s) and of N(s)
which lies in the closed right-half plane. Then s0 is a system zero of R(s) and we see from
(8.24) that there is a complex vector U(.111 1 # 0 such that

However since L(s)

Hoc all elements of L(s0 ) are finite. Therefore we have

Uso)R(.Io) U(.1o I = 0

and the Bezout identity (8.36) is not satisfied. Therefore we have shown that, in the
MIMO case, if the Bezout identity, (8.36) or (8.34), is satisfied then M(s), N(s) are
coprime.
We can show, in a similar way. that (8.35) is sufficient for Jf(s),N(s) to be coprime.
Proof (only if) This will be done in the next section by using certain state models
we will determine for the transfer functions JJ(s), S(s), M(sl. N(s). X(s). Y(s), i(s),

w.

Before going on to the development of state models for the various systems involved in
coprime factorization, we need to consider the possibility of there being more than one
coprime factorization for a given system. This is done now as follows.

8.3.3

Relating coprime factorizations

The coprime factors in a coprime factorization of a given system are not unique.
Moreover the factors in two different coprime factorizations of the same transfer function
are related by a unit, (Section 8.2.4). This fact is shown in the following theorem.
Theorem 8.2 Suppose G(s) has right and left coprime factorizations

Then G(s) has right and left coprime factorizations given by

G = S 2 (s)M2 1 (s)
=

Jf2 1 (s)S.c(s)

if and only if there exists units R(s). L(s)

U00 such that

N 2 (s) = N 1 (s)R(s)

N 2 (s) = L(s)N 1(s)

Jf2(s) = M 1(siR(s)

jf2 (s) = L(spf 1(s)

(8.37)

Proof (if) If R(s) E Hx then we see from (8.37) that N 2 (s), M 2 (s) E Hoc since the
composition (series connection) of stable systems is stable. In addition we see from (8.3 7)
that
(8.38)

206

System Algebra

However since N 1 (s), M 1(s) are coprime the Bezout identity (8.34) is satisfied. Therefore pre and post-multiplying the Bezout identity by R- 1 (s) and R(s), and using (8.37) we
obtain
(8.39)
where

Finally if R- 1 (s) E Hoc then we have X2(s), Y2(s) E Hoc and M 2(s) and N 2(s) are
1
coprime from Theorem 8.1. Therefore we have shown that N 2 (s)M2 (s) is an additional
right coprime factorization of G(s) if R(s) E Ucc.
1
We can show, in a similar fashion, that if L(s) E Uoc then MJ. (s)N2 (s) is a second left
coprime factorization of G(s).
Proof(only if) Consideration of(8.38) shows that the invertability of R(s) is necessary.
1
Therefore suppose R(s) is invertible but R(s) '/:. U00 Then R(s) '/:.Hoc and/or R- (s) '/:. Hx.
Suppose R(s) '/:. H 00 Then since N 1 (s)M\ 1 (s) is a coprime factorization of G(s), we
see from (8.37) that either N 2 (s), M 2 (s) '/:. H 00 or N 2 (s), M 2 (s) E H 00 . Suppose N 2 (s),
M 2 (s) '/:. H 00 Then N 2 (s)M2 1 (s) is not a coprime factorization of G(s). Alternatively,
suppose N 2 (s), M 2 (s) E H 00 Then we see from (8.37) that all unstable poles of R(s) are
cancelled by zeros of N 1 (s) and M 1 (s). This implies that N 1 (s) and M 1 (s) share closed
1
right-half plane zeros and therefore, contrary to assumption, N 1 (s)M\ (s) is not a
coprime factorization of G(s). Therefore R(s) E H 00
Alternatively, if R- 1 (s) '/:. Hoo then R(s) has closed right-half plane zeros. Thus we see
from (8.37) that N 2 (s) and M 2 (s) share closed right-halfplane zeros and are therefore not
coprime. Thus N 2 (s)M2 1 (s) is not a coprime factorization of G(s).
The foregoing shows that N 2 (s)M2 1 (s) is an additional right coprime factorizations of
G(s) only if R(s) in (8.37) is a unit.
In the same way we can show that M2. 1(s)N2 (s) is a second left coprime factorization

of G(s) only if L(s) in (8.37) is a unit.

always
is
it
since
proper
always
is
M(s)
that
section
this
In summary, we have seen in
invertible. Alternatively, N(s) is proper when G(s) is proper and strictly proper when G(s)
is strictly proper. We will see in the next section that Y(s) is always strictly proper so that
Y(oo) = 0. Therefore the product Y(s)N(s) is always strictly proper so that
Y( oo )N( oo) = 0 no matter whether G(s) is proper or strictly proper. Thus we see that
1
a necessary condition for satisfying the Bezout identity, (8.34), is X( oo) = M- ( oo) which
implies that the state models for X(s) and M(s) must have D matrices which are inverses
of each other. Similarly, state models for X(s) and M(s) must have D matrices which are
inverses of each other.

8.4

State Models for Coprime Factorization

In what follows we develop, from a given minimal state model for the system being
factored, minimal state models for its coprime factors, M(s), N(s), M(s), N(s) as well as
the factors X(s), Y(s), X(s), Y(s) needed to satisfy the Bezout identities, (8.34, 8.35).

%07

8.4.1

Righ t and left copr ime factors

state model for the system

Suppo se we are given a contro llable and ob~ervable (minimal)
to be factore d
G(s) =s

B]

[A

(8.40)

C D

Then applyi ng state feedba ck

(8.41)

u(t) =Kx(t )--H(t )

ing state equati ons for the

with K chosen to make A + BK is stable. yields the follow
closed loop system

x(t) =(A+ BK)x( t) + Bv(t)

(8.42)

y(t) = (C 7 DK)x (t), Dv(r)

(8.43)

r functio ns S(s), .H(s)

Now we can interp ret (8.41- 8.43) as two systems haYing transfe
with each system having v(t) as input

Y(s) = N(s) V(s)

(8.44)

M(s) V(s)

(8.45)

U(s)

where

"V(s)

[AC +BK
+ DK

B]

M(s) =s

[A+ BK B]
K

(8.46)

n noti~ that Jf(s) is

Notice that N(s), M(s) E Hx since A+ BK is stable. In additio
y matrix . This allows us to
invertible since its state model has aD matrix which is an identit
solve (8.45) for V(s) and substit ute the result in (8.44) to obtain

Y(s) = N(s)A r 1 (s) U(s)

relatio n of the foregoing
The following simple example provid es an illustra tion of the
state model s, (8.46) to coprim e factori zation .
Suppo se we want to obtain a coprim e factori zation for
s- 1
G(s)

= s(s- 2)

in contro ller form. so that

We begin by obtain ing a minim al state model for G(s). say
G(s) =s

[AC

B]

208

System Algebra

where
A=

[~ ~]

C= [1

-1]

Next we need to choose K so that A + BK, is stable. Suppose we do this so that A

has a multiple eigenvalue at -1. Then K is
K = [ -4

and A+ BK, and C

-1]

+ DK in the state models for the factors,

-2

A +BK = [ l

+ BK

(8.46) are

C+DK=C=[1

-1]

Finally, we get the transfer functions for the coprime factors from the state models,
(8.46), as

M(s)

= K[sl- (A+ BK)r 1B +I=

s2

N(s)=(C+DK)[sl-(A+B K)r B=

(s + 1)

2s

(s + 1)

and we see that

N(s)M- 1 (s) = G(s)
The required coprimeness of N(s), M(s) in this example is seen by inspection.
We can show that the factors, M(s), N(s), (8.46), are coprime in general by using the
definition of a system zero given in Section 8.2.3 of this chapter. This is done as follows.
We begin by noting that the state models for the factors, (8.46), have the same system
matrix, A + BK, and the same input matrix, B. Therefore the states of these systems are
equal for all time if their inputs are equal and if their initial states are equal.
Recall from Section 8.2.3 that the system zeros are defined in terms of the zero state
response. Therefore in order to investigate the coprimeness of the factors specified by
(8.46), we take the initial states of these state models null, i.e., equal. Moreover, we note
from, (8.44, 8.45) that the inputs to these state models are the same being equal to v( t).
Therefore the states of the state models for the factors M(s) and N(s), (8.46), are assumed
equal in the following investigation of the system zeros of the factors.
Suppose s0 is a common system zero of the factors M(s) and N(s) given by (8.46). Then
we see from (8.20) that the following conditions must be satisfied
[sol- A - BK]X(s0 )
[C + DK]X(s 0 )

BU(s0 ) = 0

+ DU(s0 )

KX(so)

(8.47)

(8.49)

+ U(s 0 ) = 0

(8.49)

: i

~')

~.
..

State Models tor Coprime Factorization

209

Thus substituting for U(s 0 ) from (8.49) in (8.47, 8.48) yields

[so/- A]X(s0 ) = 0
CX(s0 ) = 0

which implies that the pair (A, C) is unobservable. This contradicts our assumption that
the state model for G(s) is minimal. Thus s0 cannot be both a zero of N(s) and a zero of
M(s) which is sufficient for N(s), M(s) to be coprime.
Notice, from the foregoing demonstration of the coprimeness of M(s) and N(s),
(8.46), that the detectability of the state model used for G(s), (8.40), is both necessary and
sufficient for the coprimeness of these factors, i.e., for there to be no common closed righthalfplane system zeros for M(s),N(s).
In the next section we will provide additional evidence for the coprimeness of the pair
M(s),N(s), (8.46), by finding state models for X(s), Y(s) to satisfy the Bezout identity,
(8.34). Before doing this we determine state models for some left coprime factors of G(s)
by utilizing the fact that the right coprime factorization of GT (s), i.e.,
(8.50)
becomes the desired left coprime factorization of G(s) after transposition, i.e.,
G(s)

M- 1(s)N(s)

We do this as follows.
From (8.40) we have

Then from (8.46) we see that the right coprime factors of GT (s) are given a~.

where L T replaces K in (8.46). Notice that as K was chosen to make A + BK stable in

(8.46), so must L T be chosen now to make AT + CT L T stable. Therefore after doing the
required transposition we obtain the left coprime factors for G(s) as
-

M(s) =
Notice that we have N(s), M(s)

8.4.2

[A+ LC ~]
c

(8.51)

1-00 with M(s) being invertible.

Solutions to the Bezout identities

In this section we will see that there is an important connection between the coprime
factors for a controller in an observer based feedback control system (Section 5.2), and

~, v

.,ysre m Atgeo ra

the pairs {X(s) , Y(s)} and {X(s), Y(s)} in the Bezou

t identities, (8.34, 8.35). This
connection is used here to establish state models for these
pairs. We will see later that this
connection can be used to characterize all contro llers which
impar t stability to a closed
loop contro l system.
In order to obtain a coprime factorization for the contro
ller, we need to recall that the
relevant equati ons defining an observer based contro ller
for the plant, (8.40) are
~ (t)

= A-~(t) + Bu(t) + L[Y(t )- y(t)]

y(t) = Cx(t)

u(t)

+ Du(t)

Kx(t)

Then we see from these equati ons that the state equations
for the contro ller are

x (t)

Ax(t)

u(t)

Cx(t)

+ By(t)
(8.52)

where

A = A+ BK + LC + LDK
so that the contro ller transf er function H(s) is given by

Notice that the contro ller is strictly prope r. Notice also that
since we are assuming that
the plant has m inputs and p outpu ts, the contro ller hasp
inputs and m outpu ts. Therefore
H(s) is an m x p matrix.
Next using (8.46, 8.51) we see that the right and left
coprime factorizations of the
contro ller are

where
N y(s) b [

Ny(s )b
with K and
choose

~ + BK

[~+ic

!l
!]

M y(s) b [

~ + BK ~

My(s)b[~+ic ~]

being chosen so that the factors are stable. One way of

doing this is to

K=

C+D K

-(B+ LD)

State Models for Coprime Factorization

211

where K, L were used in the previous section to obtain state models for the left and right
coprime factorization of G(s), (8.46, 8.51). Notice that this choice provides the required
stability since the A matrices for the state models of the right and left coprime factors of
H(s) are given as

Therefore using this choice forK and L we see that the coprime factor state models for
the controller become
Nn(s)
-

s
=

[A +BK

[A+ LC

Nn(s) =

Mn(s)

( ) .!_
Mn
s-

[A +BK
C +DK

[A + LC

(8.53)

;(B+LD)]

(8.54)

We show now that the Bezout identities, (8.34, 8.35), can be satisfied by making use of
the foregoing factorizations of the observer based controller for G(s).
Theorem 8.3 If a system has transfer function G(s) with minimal state model specified
as (8.40) and coprime factorizations specified as (8.46, 8.51), then the Bezout identities,
(8.34, 8.35), are satisfied by
X(s)

Mn(s)

Y(s) = -Nn(s)

(8.55)

X(s)

Mn(s)

Y(s) = -Nn(s)

(8.56)

where Mn(s),Nn(s ),Mn(s), and Nn(s) are given by (8.53, 8.54)

Proof We need to establish that
Mn(s)M(s) - Nn(s)N(s) =I

(8.57)

We can do this by recalling the state model for systems connected in series, (8.9) and
using the state models for Mn(s), M(s), N n(s), N(s) (8.54, 8.46) to obtain

[ [A +LC
Mn(s)M(s) ob
0

[ K

Nn(s)N(s) ob

[[A+LC
0
( K

-(B+LD)K ]
A+BK

[-(B;LD)l]

(8.58)

K
-L(C+DK )]
A+BK
0

[-fl]

(8.59)

Then changing coordinates for the state representation of Nn(s)N(s) by using the

212

System Algebra

coordinate transformation matrix T

T

~ ~]

yields

NH(s)N(s)cb

A+ LC
0
[ K

-(B+ LD)K]
A+BK

[[

(8.60)

Finally comparing the state models given by (8.60) and by (8.58) we see that

which shows that (8.57) is satisfied.

We can show that (8.56, 8.51, 8.53) satisfy the Bezout identity, (8.35) by proceeding in

a similar fashion.

8.4.3

Doubly-coprime factorization

The foregoing right and left coprime factorizations of G(s), (8.46, 8.51) constitute what is
referred to as a doubly-coprime factorization. This type of factorization is defined from
the following observations.
First recall, in general, that any right and left coprime factorizations of G(s) and of
H(s) satisfy

Therefore it follows that

M(s)N(s) - N(s)M(s) = 0

(8.61)

(8.62)

MH(s)NH(s)- NH(s)MH(s)

Next recall, from Theorem 8.3, that if we use the particular coprime factorizations for
G(s) and H(s) given by (8.46, 8.51, 8.53, 8.54) then we have the following Bezout identities

MH(s)M(s)- NH(s)N(s) =I

(8.63)

M(s)M H(s) - N(s)N H(s) =I

(8.64)

Finally, notice that (8.61-8.64) can be written as the single matrix equation

[ Mf!(s)
-N(s)

-i!_H(s)
M(s)

[M(s)
N(s)

N H(s)]
MH(s)

[I0 0]
I

Now we use this relation to define a doubly-coprime factorization as follows.

(8.65)

Stabilizing Controllers

213

Suppose a system with transfer function G(s) has left and right coprime factorizations

G(s)

= _M-I

(s)N(s)

= N(s)M- 1 (s)
Then these factorizations taken together constitute a doubly-coprime factorization of
G(s) if there exists a set of transfer functions, {MH(s),NH(s),MH(s),NH(s)} E 1ix,

which satisfies (8.65).

doubly-coprime
the
between
relation
foregoing
the
use
In the next section, we
factorization of a plant and the stabilizing controller for the plant to provide a means
of characterizing all controllers which stabilize a given plant, i.e., all controllers which
render the feedback control system internally stable.

8.5

Stabilizing Controllers

Recall, from the discussion in Section 8.3.1, that achieving stability through unstable
pole-zero cancellations is not robust. Therefore we cannot rely, in practice, on unstable
pole-zero cancellations between the m;1thematical models for the plant and controller to
lead to the implementation of a stable feedback control system. To avoid this pitfall we
need a controller with model H(s) which when used with a plant having model G(s)
makes the state model for the closed loop system internally stable. Controller models,
H(s), having this property are said to stabilize the plant model, G(s), or, alternatively, are
said to be stabilizing.
Alternatively, the internal stability of the state model for the closed loop system is
equivalent to the input-output stability from {u 1(t),u 2 (t)} as input to {y 1(t),y 2 (t)} as
output for the interconnection shown in Figure 8.3. Satisfaction of this condition implies
that 8(s) E 'H00 where
(8.66)
We will show that 8(s) E 7t00 is equivalent to W(s) E 7t00 where

[ E 1 (s)] = W(s)[Ut(s)]
U2 (s)
E 2 (s)

(8.67)

In order to establish conditions on G(s) and H(s) which insure that W(s) E 'Hoo we
need to determine relations between the blocks of W(s) and G(s), H(s). Before doing this
consider the following fundamental aspect of feedback control systems.
Recall that transfer functions of physical processes must be bounded at infinity.
Therefore to be physically meaningful G(s), H(s) and W(s) must be either proper or
strictly proper. Assuming that G(s) and H(s) are proper or strictly proper, the feedback
control system is said to be well-posed if W(s) is proper or strictly proper. We will see that
W(s) is proper or strictly proper if and only if

[/- G(=)H(oo)r'

exists

System Algebra

214

~----e_,(0__7 ~---Y~I(_t)__~

u,(t)

y,(t)

Setup for Plant Stabilization Criterion

Figure 8.3

Notice that if either G(oo) or H(oo) is null, i.e., strictly proper, this condition is
satisfied and the feedback control system is well-posed. Recall that H(s) is strictly proper,
independent of whether G(s) is proper or strictly proper, in the case of an observer based
controller, (8.52). Therefore observer based control systems are always well-posed.
We proceed now to establish expressions for the partitions of W(s) in terms of G(s)
and H(s).

8.5.1

Relating

W(s) to G(s), H(s)

We begin by noting from Figure 8.3 that

E 1 (s)

U1(s)+H(s)E2 (s)

(8.68)

E2 (s)

U2(s)

+ G(s)E 1(s)

(8.69)

or

U(s)

w- 1 (s)E(s)

where
W

I
-'()=[
-G(s)
s

-H(s)]
I

Recalling that G(s),H(s) arep x m and m xp respectively we see that w- 1 (s) and
W(s) are each (m + p) x (m + p). Now the determination of W(s), (8.67), in terms of
G(s), H(s) proceeds as follows.
First substituting (8.69) in (8.68) and solving for E 1 (s) as well as substituting (8.68) in
(8.69) and solving for E 2 (s) gives
E 1(s) = [I- H(s)G(s)t 1 U1 (s) +[I- H(s)G(s)t 1H(s) U2 (s)

(8.70)

E 2(s) =[I- G(s)H(s)t 1G(s)U 1 (s) +[I- G(s)H(s)t 1 U2 (s)

(8.71)

Alternatively, substituting for E 1 (s) from (8. 70) in (8.69) gives

E2(s)

G(s)[I- H(s)G(s)t 1 U1 (s) + [I+ G(s)[I- H(s)G(s)t 1H(s)] U2 (s)

(8.72)

Stabilizing Controllers

215

Then comparing (8.71, 8.72) yields the following equalities

[I- G(s)H(s)t 1G(s) = G(s)[I- H(s)G(s)t 1

(8.73)

[I- G(s)H(s)r' =I+ G(s)[I- H(s)G(s)t 1H(s)

(8.74)

Alternatively, by substituting for E 2 (s) from (8.71) in (8.68) and comparing the
resulting expression for E 1(s) with the expression for E 1(s) given by (8.70) leads to
equations (8.73, 8.74) with G(s) and H(s) interchanged.
The foregoing results, which can also be obtained using the matrix inversion lemma,
(Appendix), enable W(s), (8.67), to be expressed in terms of G(s), H(s) in the following
ways
W(s) = [ W 1 (s)
W 3 (s)

Wz(s)]
W4(s)

[I- H(s)G(s)r'
[I- H(s)G(s)r 1H(s)
G(s)[I- H(s)G(s)r 1 I+ G(s)[I- H(s)G(s)r 1H(s)

[I+ H(s)[I- G(s)H(s)r 1G(s)

[I- G(s)H(s)r'G(s)

H(s)[I- G(s)H(s)r']

(8.75 )

(8.76 )

[I- G(s)H(s)r 1

Thus we see from the dependency of the blocks of W(s) on [I- G(s)H(s)r 1 that W(s)
is proper or strictly proper if and only if [I- G(s)H(s)r' is proper or strictly proper
which requires that I- DcDn be invertible where G(oo) =De and H(oo) = Dn.
Finally, notice from (8.67, 8.75, 8.76) that we can write 8(s), (8.66), as

e(s)= [W3(s)
W 1(s)- I
Therefore we have 8(s) E Hoc is equivalent to W(s) E Hoc.

8.5.2

A criterion for stabilizing controllers

We showed, in the previous section, that W(s) E Hoc is equivalent to the corresponding
closed loop system being internally stable. In this section we give a more useful criterion
for deciding if a controller is stabilizing. We will use this criterion in the next section to
develop a characterization of all controllers which are capable of stabilizing a given plant.
The criterion to be developed here is stated in the following theorem.
Theorem 8.4 A controller having transfer function H(s) with right coprime factorization H(s) = N n(s)M!/ (s) stabilizes a plant having transfer function G(s) with left
coprime factorization G(s) = M(/ (s)Nc(s) if and only if
(8.77)
where

216

System Algebra

Proof Recall that 6(s) E Uoc is equivalent to 6(s), 6- 1 (s) E Hx. Since 6(s) is
defined as the sum and product of transfer functions in Hx, we have 6(s) E Hoc
irrespective of the internal stability of the closed loop system. Therefore we need to
show that 6- 1 (s) E Hx is necessary and sufficient for W(s) E H:c
In order to show that 6- 1 (s) E Hx is sufficient for W(s) E Hcxc we assume
6- 1 (s) E Hoc and proceed as follows.
From the given coprime factorization of the plant and controller we have
--1

-1

G(s)H(s) =Me (s)Nc(s)N H(s)MH (s)

or

Mc(s)G(s)H(s)M H(s)

Nc(s)N H(s)

Therefore 6(s), (8.77), can be written as

6(s)

Mc(s)[/- G(s)H(s)]MH(s)

This allows us to express[/- G(s)H(s)r 1as

Then we can use this expression to rewrite W(s), (8.75, 8.76) as follows

w( s )

= [/

+ N H(s)6 - 1(s)Nc(s)
1

M H(s)6- (s)Nc(s)

N H(s)6 - 1 (s)~c(s)
MH(s)6 - 1(s)Mc(s)

(8. 78)

Therefore, assuming that 6- 1 (s) E H 00 we see that W(s) E H 00 since each element of
W(s) consists of products and sums of transfer functions in H 00 This shows that (8.77) is
sufficient for the controller to be stabilizing. In order to show that 6 - 1 (s) E Hoc is
necessary for W(s) E Hoo we assume W(s) E Hoo and proceed as follows.
Since H(s) = N H(s)M}/ (s) is a coprime factorization, the following Bezout identity is
satisfied

for some X H(s), Y H(s) E H 00 . Then post-multiplying this Bezout identity by 6

Mc(s) and by 6 -I (s)Nc(s) gives the following two equations
X H(s)M H(s)6 -I (s)Mc(s)
X H(s)M H(s)6 -I (s)Nc(s)

+ Y H(s)N H(s)6 -I (s)Mc(s) =

+ Y H(s)N H(s)6 -I (s)Nc(s)

where
Q 1 (s) = 6- 1 (s)Mc(s)
Q2 (s)

= 6- 1 (s)Nc(s)

Q 1 (s)

= Q 2 (s)

-I

(s)

217

Stabilizing Controllers

However we can rewrite these equations using W(s), (8.78), as

+ Y H(s) W2 (s) = Q 1 (s)

(8.79)

+ Y H(s)[W1 (s)- I] = Q2 (s)

(8.80)

X H(s) W4(s)
X H(s) W3 (s)

Therefore, since all terms on the left sides of (8. 79, 8.80) belong to 'Hoc, we have
(8.81)
However the factors of an 'Hoo function need not be in 'Hoc Therefore we are not able to
use (8.81) to conclude that Ll- 1 (s) E 'H 00
Since G(s) = A-rc; 1 (s)NG(s) is a coprime factorization, the following Bezout identity is
satisfied

MG(s)XG(s)

+ NG(s) YG(s)

=I

for some XG(s), YG(s) E 'H 00 Then pre-multiplying this equation by Ll- 1 (s) yields

and we see that Ll- 1 (s) E 'H::xo since all terms on the left side of this equation belong to 1t00
This shows that (8. 77) is necessary for the controller to be stabilizing.

By interchanging the role of the plant and controller in the foregoing theorem, we see
that a left coprime factored controller, H(s) = Mii\s)NH(s), stabilizes a right coprime
factored plant, G(s) = NG(s)M(/(s), if and only if

Li(s) E U=

(8.82)

where

In summary, we see that we can use either (8.77) or (8.82) to determi;e if a given
controller stabilizes a given plant. We can use this fact together with a doubly-coprime
factorization of the plant to provide a parametrization of all the plant's stabilizing
controllers.

8.5.3

You/a parametrization of stabilizing controllers

We have just shown that if there are coprime factorizations of the transfer functions for
the plant and controller which makes Ll(s), (8.77), or b.(s), (8.82), a unit, i.e., Ll(s) E U 00 ,
or Ll(s) E U 00 , then the corresponding closed loop system is internally stable and the
controller is said to be stabilizing. However, if the coprime factorizations of the plant and
controller happen to also make both Ll(s) and b.(s) identity matrices, then not only is the
controller stabilizing but, in addition, the left and right coprime factorizations of the
plant constitute a doubly-coprime factorization, (Section 8.4.3), of the plant.
The foregoing observation is used now to show that we can construct a set of
stabilizing controllers by using a doubly-coprime factorization ofthe plant. The elements
of this set are generated by varying an m x p matrix, Q(s), over 1()().

218

System Algebra

Theorem 8.5 Given a doubly-coprime factorization of a p x m plant transfer function

matrix

G(s)

--1

Ma (s)Na(s)

-1

Na(s)Ma (s)

(8.83)

where

(8.84)

then either of the following forms for the m x p controller transfer function matrix
stabilize the given plant

H(s)

Nc(s)M-;: 1(s)

(8.85)

H(s) = M; 1(s)Nc(s)

(8.86)

for all m x p matrices Q(s) satisfying Q(s)

Hoc where

Mc(s)

+ Ma(s)Q(s)
MH(s) + Na(s)Q(s)

Nc(s)

NH(s) + Q(s)Ma(s)

Nc(s) = NH(s)

Mc(s) = MH(s)

(8.87)

+ Q(s)Na(s)

Proof Recall that for (8.85) to be a coprime factorization of H(s), we must have

(i)

(ii)

Nc(s), Mc(s) E Hoc

Nc(s), Mc(s) coprime

Since Nc(s), Mc(s), (8.87), are each dependent on sums and products of matrices in
Hoc, condition (i) is satisfied. In order to show that condition (ii) is satisfied, recall that
Nc(s),Mc(s) are coprime if X(s), Y(s) E Hoc satisfy the Bezout identity

X(s)Mc(s)

+ Y(s)Nc(s) =I

(8.88)

Now substituting for Mc(s), Nc(s) from (8.87) and choosing

X(s) = Ma(s)

Y(s)

-Na(s)

yields the following equation

where
81

= ifc;(s)MH(s)- Nc(s)NH(s)

82

[Ma(s)Nc;(s)- Nc(s)Ma(s)]Q(s)

(8.89)

Lossless Systems and Related Ideas

219

Then we see that 8 1 = I since the plant transfer function factorization is doublycoprime, (8.84). Moreover we have 8 2 = 0 from (8.83). Therefore (8.88) is satisfied and
condition (ii) is satisfied. Thus (8.85) is indeed a coprime factorization of H(s).
Finally, in order to show that the controllers, (8.85), are stabilizing. notice that we have
just shown that

Therefore, referring to Theorem 8.4, we have ~(s) = IE Ux so that the controllers given
by (8.85) are stabilizing.
This completes the proof that the controllers given by. (8.85), stabilize the given plant.
The proof that the controllers given by (8.86) are stabilizing can be done in a similar

fashion.

8.6

Lossless Systems and Related Ideas

A linear time-imariant system is said to be lossless if it is stable and if it has zero state
response y(t) corresponding to any u( t) E .C 2 [0, oc) such that
(8.90)
Notice, from Sections 7.4 and 7.5 of the previous chapter, that a lossless system has L 2
gain equal to one. However it is important to note that this condition is necessary but not
sufficient for a system to be lossless, i.e., for (8.90) to be satisfied for all u( t) E .C2 [0, oo ).
In contrast, recall from the previous chapter that .C2 fO. oo) is isomorphic to .C 2 so that
the L 2 norms in the time and frequency domains are equivalent. Therefore a system is
lossless if its transfer function G(s) E HCXJ maps U(s) E H 2 to Y(s) E H 2 such that
(8.91)
However since

I Y(Jwll~= -2j

!X

71'

U*(Jw)G*(jw)G(jw)U(jw)dw

-x

we see that (8.91) is satisfied for all U(s) E

G*(jw)G(jw)

Im

7-{ 2

if and only if

for all wE (-oc, oc)

(8.92)

where G(jw) is p x m. Notice that p ;:::: m is necessary for (8.92) to be satisfied.

Mathematically, we can consider lossless systems to be unitary operators. Recall, from
Chapter 7, that the Fourier transform is an isometry or unitary operator between the time
domain Hilbert space .C 2 (- oc, oc) and the frequency domain Hilbert space .C2 . In the
present instance (8.90) implies that an m-input p-output lossless system is an isometry or
unitary operator between the m-dimensional input Hilbert space, .C 2 [0. oc ), and pdimensional output Hilbert space, .C 2 [0, oc ). Alternati\d~, (8.91) implies that a lossless
system is a unitary operator between them-dimensional input Hardy space, 7( 2 , and the
p-dimensional output Hardy space, H 2

zzo
An important property of lossless systems can be seen as follows. Notice that if we
connect two systems with transfer function G 1 (s), G2 (s) E Hoo is series with G1 (s) being
the transfer function of a Iossless system, then we have IIG1 (s) lloo= 1 and from Section
7.4.2 we see that
(8.93)
However (8.92) implies that the L 2 norm of the output from the composite system is
independent of G1 (s). Therefore the inequality (8.93) becomes an equality

In the next chapter we will see that this type of invariance of the H 00 norm plays an
important role in the development of a solution to the H 00 control problem.
Now since the linear models of the time-invariant physical processes being considered
here, e.g., plants, have transfer function matrices whose elements are ratios of polynomials having real coefficients, we have

G*(jw) = GT(-jw)

for all w E ( -oo, oo)

and (8.92) is equivalent to

GT ( -s)G(s) = Im

for all complex s

(8.94)'

We can gain additional insight into these matters by reviewing the earlier development
of these ideas in connection with electric filters.

8.6.1

All pass filters

Recall that a stable time-invariant linear SISO system having transfer function G(s) has a
steady-state output given as

y(t) = A 0 sin(wt + 8o)

when its input is given by

u(t)

= A 1 sin(wt)

where

A 0 = A;yiG*(jw)G(jw)

IG(jwiAi

e 0 =tan _1 [Im[G(jw)]]
Re[G(jw)]

This system is referred to as an all pass filter if

IG(jwl =I

for all w

E ( -oo,

oo)

(8.95)

Lossless Systems and Related Ideas

221

or if
G( -s)G(s) = 1

for all complex s

(8.96)

so that the amplitude A 0 of the steady state output equals the amplitude A; of the input for
all frequencies. Thus comparing (8.95, 8.96) with (8.92. 8.94) we see that all pass filters are
lossless systems. Moreover if G(s) is the transfer function of an all pass filter, then we see
from (8.96) that G(s) is proper, with G(O) = 1, and has poles which are mirror images of
its zeros across the imaginary axis, i.e., ifP; is a pole of G(s) then -p; must be a zero of G(s ).
For example, the first order system having transfer function

(s- 1)
G(s) = (s+ 1)
is an all pass filter
An all pass transfer function is a generalization of the concept of an all pass filter. This
generalization is done by relaxing the requirement of stability. Thus G(s) is said to be an
all pass transfer function if (8.96) is satisfied. Notice that the mirror imaging of poles and
zeros implied by (8.96) means that when G(s) is an all pass transfer function it has no
imaginary axis poles, i.e., G(s) E 00

8.6.2

Inner transfer functions and adjoint systems

In order to distinguish between stable and unstable all pass transfer functions, stable all
pass transfer functions are referred to as being inner. Thus a transfer function is inner
when it is the model for a lossless system.
In the case of a transfer function matrix the following generalization is made. If
G(s) E 7t00 is p x m, then G(s) is said to be inner if (8.97) is satisfied or co-inner if (8.98) is
satisfied where
(8.97)
G~ (s)G(s) = Im
if p 2m
G(s)G~ (s) = IP

if p::::; m

(8.98)

for all complex s with

Notice that p 2 m (p ::::; m) is necessary for G(s) to be inner (co-inner). Also notice that
if the transfer function is square and inner then G- 1 (s) = G ~ (s) E 7t~. Moreover, since
dimensionally square systems have the property that the poles of G- 1 (s) are the zeros of
G(s). (Section 8.2.4), we see that in this case that all system zeros of G(s) lie "in" the open
right half plane, i.e., lie in the region of the complex plane where G(s) is analytic. This is
the origin of the terminology "inner". However in the nonsquare case this concentration
of zeros in the open right half plane is not necessary for a transfer function to be inner,
e.g ..

G(s)

.1

=--

s+Vi

[s+l]
1

is inner without having any zeros in the right half plane.

The system whose transfer function is G- (s) is referred to as the adjoint sys!em
(relative to a system having transfer function G(s)). Notice that the state model for G (s)
is related to the state model for G(s) as

G(s) =s

[AC DB]

(8.99)

since

Properties which the parameters of a minimal state model must satisfy in order for a
system to have an inner transfer function are given in the following theorem.
Theorem 8.6 G(s) E Hoc is inner if it has a minimal state model

G(s) =s

[AC DB]

with parameters that satisfy the following conditions:

DTD=l
BTWO +DTC = 0
AT Wa

+ WoA + cT c

Proof From the state model for G- (s), (8.99), and the rule for composing two systems,
(8.9), we obtain

Next transforming coordinates using the coordinate transformation matrix

gives

- Ar
G-(s)G(s) de [ [ 0
[ C,
where A2 = -A 7 X- XA- CT C and

AA2 ]
C2]

[BB21 ]

DTD

Jl
f.

-~

Summary

223

Finally, applying the conditions given in the theorem yields

X= W

A2 = 0

Bl =0

C2=0

DTD=l

and the state model for G (s)G(s) becomes

~]

C
G- (s)G(s) =s [A
where

A=

r-~T ~]

C=[

c1

0]

B=
jj

[!]

=I

Thus we have
G- (s)G(s)

= C(sl- A)- 1B + jj
=I

and the theorem is shown.

In developing a solution to the Hoc state feedback control problem in the next chapter,
we will extend the foregoing characterization of state models so that the resulting system
is J-lossless with J-inner transfer function.

8.7

Summary

In this chapter we have established state models for certain interconnections and
operations with systems. This included the ideas of system inversion, system zeros,
system coprimeness and lossless systems. We will see in the next two chapters that the
operation of coprime factorization coupled with the requirements on a system's state
model for that system to be lossless, plays an important role in developing a solution to
the H 00 feedback control problem.

8.8

Notes and References

The treatise by Vidyasagar [44], on the consequent ramifications of the coprime

factorization approach to feedback control places coprime factorization in the mathematical setting of algebra known as ring theory. The term "unit" refers to elements in a
ring that have an inverse in the ring. The development of state models for the coprime
factors used here is treated in [14].

9
H 00 State Feedb ack and
Estimation

9.1

Introduction

In the previous two chapters we developed ideas for characterizing signals and systems. In
the next two chapters we use these ideas to develop the design equations for a controller
which is required to
(i) stabilize the closed-loop control system and
(ii) constrain the L 00 norm of the closed-loop transfer function from disturbance input,
u 1(t), to desired output, y 1 (t), to be no greater than/, a specified positive scalar.
Figure 9.1 gives the setup for this problem. Before going on to consider this setup in
more detaiL we make the following more general observations on this control problem.
First, we should recall that to avoid having a system's input-output stability
dependent on the physically unrealistic requirement of pole-zero cancellation, we have
used the term "stability" to refer to a system's internal stability. Thus when requirement
(i) is satisfied so that the closed-loop system is internally stable, the L 00 norm mentioned
in requirement (ii) is an H 00 norm. Therefore controllers satisfying (i, ii) are said to solve
the H 00 control problem.
AlternatiYely. recall from Chapter 7, that the H 00 system norm is induced by the H 2
signal norm in the frequency domain or the L 2 signal norm in the time domain. Therefore
when requirement (i) is satisfied, (ii) can be interpreted as the requirement that the closedloop control system's L 2 gain from u 1(t) to y 1 (t) be no greater than f. Thus the H 00
control problem is sometimes referred to as the L 2 gain control problem.
In addition, notice that satisfaction of requirement (ii) does not imply that requirement
(i) is satisfied since the L 00 norm is defined for both stable and unstable systems provided
only that the system's transfer function have no imaginary axis poles. This differs with
both the quadratic and LQG control problems since the performance indices, lQc and
lGc, are infinite when the closed-loop system is unstable.
Concerning the setup for the Hx control problem, it has become customary to replace
the actual plant in the feedback control configuration, Figure 9.1, by a composite system
referred to as a generalized plant. The generalized plant consists of the plant to be
controlled plus any other systems which the control system designer may want to connect

226

H 00 State Feedba ck and Estimat ion

u 1(t)

u,(t)

y,(t)
GENERALIZED
PLANT

CON1ROLL ER

y,(t)

Setup for Hex, Control Problem

disturbance input
y 1 (t) = desired output
controlled input
y 2 (t) = measure d output
Figure 9.1

u 1 (t)

u2 (t)

where u 1 ( t), u2 ( t), y 1 ( t), y 2 (t) have dimensions m 1 m 2 . p 1 . p 2 respectively.

in cascade with the plant (referred to as weights or filters). For instance. the
filters can be
used to take into account any known spectral characteristics of the disturba
nce signals,
e.g., narrow band noise. Then when the feedback control system is impleme
nted, the
controller determi ned by applying the control algorith m to the generalized plant
is used
with the actual plant.
In what follows the signals {u;(t). Y;(t) : i = 1, 2} shown in Figure 9.1 are assumed
to
be the various inputs and outputs associated with the generalized plant which
has realrational transfer function G(s) with given real-par ameter state model

D~~]
1

D22

(9.1)

As in the LQG and quadrat ic control problems, a solution to the H output

feedback
00
control problem takes the form of an observer-based controll er and involves
the
stabilizing solutions to two algebraic Riccati equations. Thus in order to
develop a
solution to the Hoc output feedback control problem we need first to consider
two simpler
problems namely, the Rx state feedback control problem, and the Hoc state
estimati on
problem. These basic Hx problems are taken up in this chapter with the
R)C output
feedback control problem being addressed in the next chapter.
As we progress through the development of solutions to these problems. we
will see
that the generalized plant's D matrix plays a pivotal role. However, only certain
parts of
the D matrix play a role in each problem . For instance, the existence of a solution
to the
Hx state feedback control problem requires D 12 to have indepen dent columns
and only
the first p 1 rows of D, denoted by D 1, are involved in the solution to this problem
where

However, the existence of a solution to the H.cx. state estimati on problem requires
Del
to have indepen dent rows and only the first m 1 columns of D, denoted by D
2 are involved

Hoo State Feedback Control Problem

227

in the solution to this problem where

, are involved in
In the next chapte r we will see that the last p 2 rows of D, denote d by D 3
where
problem
the solution to the H 00 output feedback control

9.2

H00 State Feedback Control Problem

want to choose the

Assuming that the state of the generalized plant, x(t), is known we
state feedback matrix, K 2 , so that setting the controlled input as
(9.2)

n from u 1(t) to
makes the closed-loop system internally stable and the transfer functio
'Y
t
constan
positive
d
specifie
a
y 1(t) have H 00 norm no greater than
the plant as the
Since the measured output, Y2(t), is not needed in this problem, we take
of G 1(s) are
ters
parame
the
assume
we
Thus
.
missing
generalized plant, (9.1), with Y2(t)
known where

(9.3)
with

where

k control problem
The relation between the signals and systems in the H 00 state feedbac
going to be used
are
x(t)
K
and
(t)
1
are shown in Figure 9.2. where the additio nal signals u1
.
problem
control
k
feedbac
in the development of a solution, K 2 , to the Hx state
Now after implementing state feedback we have
(9.4)

where

and we want to find K 2 such that

and

(ii): IITc(s)lloo:C:: 'Y

(9.5)

228

H 00 State Feedbac k and Estimatio n

ii, (t)

u,(t)

y,(t)

PLANT

u,(t)

K,x(t)
STATE
CONTROLLE R

K1 x(t)

Figure 9.2

~J

Setup for the Hx State Feedback Control Problem

Next recall from Chapter 7 that the H 00 norm is defined by

1/Tc(s)//CX)=

O"max[Tc(jw)]

sup
w;o(-oo,x)

where O"rnaxfTc(jw)] is the largest singular value of Tc(jw). Now since

lim Tc(jw)

w-.oo

D11

we see that it is not possible to satisfy condition (ii), (9.5), if

where

In subseque nt sections we will see that the more restrictive condition

(9.6)

must be satisfied in order to solve the H 00 state feedback control problem. In addition, we
will see that this condition is also needed to solve the H 00 state estimatio n problem.
We can summari ze the strategy we will employ now to develop the design equation s for
K2 as follows.
l. Start by assuming stabilizing feedback, i.e., assume K 2 achieves condition (i), (9.5).
2. Define a scalar quadratic performa nce index, P.1 , in terms of y 1 (t) and u (t) such that
1
P 1 :S: 0 implies condition (ii), (9.5), is satisfied.
3. Use G1 (s), (9.3), to introduce a transfer function G1(s) relating { u ( t), y (()}as output
1
1
to { u1 ( t), u2 ( t)} as input under the constrain t of stabilizing sta.te feedback. i.e.,
u2(t) = K2x(t) such that G1(s) E H 00 . This enables P1 to be rewritten as a quadratic
in u1 (r) and u2 (t) with U 1(s) E H 2 and U2 (s) restricted to stabilize G1 (s).
4. Convert P1 to a quadratic in { V1(s). V 2 (s)} E H 2 where { V (s). V (s)} arise as
1
a
2
consequence of using J-inner coprime factoriza tion on G1 (s).
We will see that the design equation s for K 2 result from carrying out the last step.

Hoo State

9.2.1

Feedback Control Problem

229

Introduction of P1

see from Chapte r 7

Suppose K 2 is chosen so that conditi on (i), (9.5), is satisfied. Then we
that
if

YI(t) E .Cz(O,oo)

and we can interpr et conditi on (ii), (9.5), in the time domain as a constra

int on the L 2 gain

(9.7)

Next define P1 as

(9.8)
Then conditi on given in (9.7) is equival ent to
(9.9)
Now since 2 [0, oo) is an inner produc t space we can rewrite P1 as

(9.10)

where

, we see that P 1 ,
However, recalling from Chapte r 7 that 2 [0, oo) is isomor phic to 1t2
as
(9.10), can be rewritte n in the frequency domain

P,

= j [ U1 (s)] ,J"'mp [ U1(s)] )

\

Y1 (s)

Y1 (s)

-1' 2 IIUI(s)ii~+IIYI(s)ii~

(9.11)

and conditi on (ii), (9.5), is satisfied if

(9.12)

s stabilizing state
It is import ant to remem ber that the foregoing result assume
feedback.

9.2.2

Introduction of G1 ( s)

that we can relate

Recalling from (9.3) that Y 1 (s) depend s on U 1 (s) and U 2 (s), we see

z;ro

H00 "Blll'e~'i!CIBiil:Tsit'a"'tlf/ihi;\ton .

{ Ur (s), Yr (s)} to { U 1 (s), U2 (s)} throu gh the use of a

transf er functi on Gr (s) as

(9.13)
where

Now under the constr aint that the state feedback is stabili
zing we have
for

(i): U1 (s) E Hz
(ii): U2 (s) = K2 X(s)

where A.[A + B 2 K 2 ] are all in the open left half plane, and
X(s) is the Lapla ce transf orm of
the closed-loop system state.
There fore we can use (9.13) to rewrite the inner produ ct
defining P1 , (9.11 ), in terms of
{U1(s), U2 (s)} as

(9.14)
where
U 2 (s)

= K 2 X(s)

In what follows, it is impor tant to remem ber that even

thoug h G1 (s) may possibly be
unstable, the restric tion that the contro lled input, U (s)
= K2 X(s), must be stabilizing
2
maint ains P1 , given as either (9.8) or (9.14), finite.

9.2.3

Introduction of J-in ner copr ime facto rizat ion

Recall, from Chapt er 8, that G1 (s) E Lx has a right coprim
e factor izatio n given by
-

--1

G1 (s) = N(s)M

(s)

(9.15)

where N(s), M(s) E H~ with H(s), i\f(s) being coprime.

Now we are going to make use of
a special type of right coprime factor izatio n in which lUs)
in (9.15) satisfies
(9.16)
j

.
(li1f'

[ (l)-~,2
I

/,,,

A system having transf er function N(s) E fix which satisfi

es (9.16) is said to be J-inne r
and a right coprim e factorization (9 .15) which has H(s)
J-inne r is referred to here as a J-

H 00 State Feedback Control Problem

231

inner coprime factorization . The utility of using this type of coprime factorization here
can be seen immediately as follows.
Let G\ (s) have a J-inner coprime factorization as denoted by (9.15, 9.16). Then using
the fact that N(s) is J-inner we see that P'Y' (9.14), can be rewritten as

where

U2 (s)

= K2X(s)

Next recall, from Chapter 8, that since M(s), N(s) E 7t00 we can obtain { U 1(s), U2(s)}
and {U1 (s), Y1 (s)} from V1 (s), V2 (s) E 'H2 as
(9.18)

(9.19)
Then since N(s) E 1-loo we have

(9.20)
Finally, by using (9.18) in (9.17) we can rewrite P7 as

p'Y = ( [
=

~:~;~

,J'Ymm [

~:~;~

l)

(9.21)

--/I IV, (s) II~+IIV2(s) II~

Notice that unlike U2 (s), which is restricted to a subspace of7t 2 since it must provide
stabilizing state feedback, V1 (s), V2 (s) are not restricted to any subspace of7t 2 . We will
use this freedom in the choice of the V,.s to ensure that P1 :::; 0.

9.2.4

Consequences of J-inner coprime factorization

Suppose we have a J-inner coprime factorized for G1 (s), (9.15, 9.16). Then to proceed we
need to consider state models for these coprime factors. Therefore, recalling (8.46) and
Section 8.3.3 we see that state models for the coprime factors, M(s), N(s), (9.15), can be
determined from the state model for G1 (s), (9.13), as
M(s)

I'

[A+BK

Bfl-']

fi.- 1

(9.22)

(9.23)

232

~.,,.~,-'~""""'""'"''.,...."'~-"''~~"'~~~"~-,.,.~

H 00 State Feedbac k and Estimatio n

where K, 1:1 are chosen so that M(s), N(s) E H_ 00 with N(s) satisfying (9.16); 1:1
is a
constant nonsingu lar (m 1 + m 2 ) x (m 1 + m 2 ) matrix which along with K is partition ed
so
as to conform with the partition ing of Band 15 1 , viz.,

We show now that the Hoc state feedback control problem is solved when K and 6. are
chosen so that the foregoing coprime factoriza tion is J-inner. The equation s needed
to
determin e K and Do are develope d in subseque nt sections.
We begin by using (9.18) to express V1(s), V 2 (s) in terms of U 1 (s), [h(s) as
(9.24)

Then using the state model for the inverse system develope d in Chapter 8, we see that
the
state model for the inverse of M(s), (9.22), is

~]

(9.25)

Notice that the system and input matrices for the generaliz ed plant, G (s), (9.3), and
1
for M- 1 (s), (9.24), are the same. Therefor e the states of these systems are equal for
all
time if their initial states are equal. The initial state of each system is the same since each
is
null. This is because we are basing our developm ent of the Hx state feedback controlle
r
on transfer functions which relate the system output to the system input under
the
assumpti on the initial state is null.
The foregoing observati on together with the state model for M- 1 (s). (9.25). allows us
to write the V;s as

(9.26)
where X(s) is the Laplace transform of the state of the state model for G (s), (9.3).
1
Now suppose we choose U1(s) and U2 (s) as
(9.27)
where D1 (s) E 7-{ 2 is an external disturban ce input whose effect will be seen to
be
equivalen t to the actual disturban ce input U1 (s) E H 2 . Then substituti ng (9.27)
in
(9.26) yields
(9.28)

We show now that P,! <:; 0 for U 1 (s)

E 7-{ 2 ,

(9.12). This is done by first showing that (9.28)

implies that P'Y :S 0 for 0 1 (s)

U1 (s) E 1-lz.
Recall that

for
1-2 and then showing that this implies P7 :S 0

{oo

II Vi(s)lli= 27r 1-oo V7(Jw )V;(jw )dw

so that using (9.28) in P'Y, (9.21) can be writte n as
(9.29)

Therefore we see from (9.29) that

if

[-1' 2 ~i~l +~3~3] < 0

then

P'Y < 0

for ~11 non-n ull 0 1 (s) E 1-2

p'Y = 0

for

ul (s) =

}
(9.30)

(/)

that the H 00 state feedback contr ol

Howe ver recall, from the beginning of this section,
will show in Secti on 9.3.3 that
probl em has a soluti on only if Du satisfies (9.6). We

if O"max[DIJ] < '/'

then

-1' 2 ~t~~

(9.31)

+ ~3~3 < o

There fore (9.31, 9.30) imply that

if O"max[Dn] < '/'
then P7 < 0 for a~l non- null

p'Y = 0 for

ul (s) =

0 1(s)

E 1-{2

(9.32)

(/)

see this notice from (9.27) that

We need next to show that (9.32) implies (9.12). To
(9.33)

ack, u2 (t) = K 2 x(t) stabilizes the

Now we will show in Section 9.3.3 that the state feedb
for any
K
=
(t)
u
1x(t). Thus X(s) E 1-{ 2
generalized plant by itself witho ut requi ring 1
E H2
(s)
0
that
1
(9.33)
from
see
we
fore
input Ut (s) E 1-{2 provi ded u2 (t) = K 2 x(t). There
i.e.,
ied,
satisf
is
(9.12)
that
s
implie
for all U1 (s) E 1-{ 2 . This fact toget her with (9.32)
and U1 (s) -1- K 1X(s) }

for

U 1(s)

for

U1 (s) = K 1X(s)

E 1-{2

(9.34)

the soluti on of the H"" state feedback

as is requi red to satisfy requi remen t (ii), (9.5), for
contr ol probl em.

Notic e that the distur bance input U (s) = K X(s)

maximizes P.,. There fore U1 (s)
1
1
K1X( s) is referr ed to as the worst distur bance
input for the Hoc state feedb

ack contr ol
system.
In summ ary, we have show n that a soluti on to
the H= state feedback probl em is
obtai ned when (9.6) is satisfied and when we choos
e K and L1. so that the right copri me
factor izatio n of G1 (s), (9.13), has nume rator N(s),
(9.23), which is J-inner. The equat ions
forK which make N(s) J-inn er are the design
equat ions for a soluti on to the H 00 state
feedback probl em. These equat ions are developed
in the next section.

9.3

H oo State Feedback Controller

In order to obtai n a J-inn er coprime facto rizati on of

G1 (s), (9.13), we need condi tions on
the state mode l param eters of a given system which
ensure that the system is J-inner.
These condi tions are given in Theo rem 9.1. A simila
r set of condi tions was given in the
previous chapt er in conne ction with inner system
s, (Theo rem 8.6). Once we have
Theo rem 9.1 we can apply it to the state mode l param
eters for S(s), (9.23). In this way
we determ ine an equat ion relating 6. to D and
an expression for K involving the
1
stabilizing~solution Xoc to an algeb raic
Riccati equat ion.

9.3.1

Des ign equ atio ns for K

Notic efrom (9.l3 ,9.l5) that.! V(s)i san(m +pi) x

(m 1 +m 2 )matr ix.No wfor lV(s) tobe
1
J-inner, the condi tions given in the following theor
em must be satisfied.
Theorem 9.1 A system having an (m + pJ) x(m
1
1 + m 2 ) transf er funct ion N(s) is Jinner if N(s) E Hx and the param eters of a minim
al state mode l for the system

satisfy

(i)
'

where

J,mp' l.ymm

J ")lnYfl

(ii)

(iii)

are defined as

Proof The proof is similar to the proof of Theo

rem 8. 7. Here we need to obtai n a
state mode l for NT ( -s)J., 111pN(s) and chang e coord
inates with coord inate transf orma tion
matri x T given by

T=lII x1 ]
LO

H 00 State Feedback Controller

235

The resulting state model then reveals that in order for NT ( -s)J-ympN(s) = J-ymm we need

to impose the conditions in the theorem.

Jis
N(s)
that
so
(9.23),
N(s),
for
Now we want to choose~ and Kin the state model
9.1.
Theorem
applying
and
inner. We do this by equating N(s) to N(s)
To begin, recall that the D matrix in a state model is unchanged by a change of
coordinate s. Therefore from N(s) = N(s) we have
(9.35)

Then using condition (i) in Theorem 9.1, we obtain

(9.36)

where

Notice that D Jc is square. We will show later in this section that D Jc is invertible if D 12 has
independe nt columns and D 11 satisfies (9.6). This inverse is needed to determine K.
Continuing with the development, we see from Theorem 9.1 and (9.23) that further
consequences of setting N(s) = N(s) are
AN= A+BK
BN

= B~-1

eN=

(9.37)
(9.38)

c, + i5 K
1

(9.39)

Then using (9.35, 9.38, 9.39) in condition (ii) ofTheorem 9.1 we obtain

which after multiplying on the left by ~ T becomes

(9.40)

Now we see from the definitions of C1 and

15 1, (9.13). that
(9.41)
(9.42)

Thus using (9.41, 9.36) and assuming D 1, is invertible enables us to solve (9.40) forK
as
(9.43)

236

H= State Feedback and Estimation

Notice that K depends on the state model parameters for the given generalized plant,
(9.3), and on X. We will show now that X is a solution to an algebraic Riccati equation.
To begin, we substitute from (9.37, 9.39) in condition (iii) of Theorem 9.1 and use
(9.41, 9.42) to obtain
ATX +XA +KTBTX +XBK +KTDfe!

+ eTD!K +KTDJcK +ere!= 0

(9.44)

Next grouping terms involving K and KT and using (9.43) to substitute -DJcK for
BT X+ Di e 1 enables (9 .44) to be rewritten as

(9 .45)
Finally substituting forK from (9.43) gives the equation

(9 .46)
where
Ax! =A - BD]c1Df el
Roc!= BD]/BT

Qool

ef(I- DJD]/DT)e l

with

Notice that (9.46) is an algebraic Riccati equation like the GCARE (Section 6.2). We
refer here to (9.46) as the HCARE. Since A001 - R 001 X =A+ BK we have K as a
stabilizing state feedback matrix if the solution X to (9.46) makes A 001 - Roc 1X stable.
As in the quadratic and LQG state feedback control problems, we reserve a special
notation for these solutions which in this case is X oc Notice that X 00 ?': 0.
In addition, notice that unlike Df2 D 12 , D Jc is not sign definite, i.e., neither nonnegative nor non-positive. Thus unlike R 1 in the GCARR (section 6.2.2), K" 1 in the
HCARE is not sign definite. This makes the determinatio n of conditions on the state
model for the generalized plant which ensure the existence of a stabilizing solution to the
HCARE more involved than in the case of the QCARE.

9.3.2

On the stability of A+ B2 K2

We have just developed equations for the determinatio n of K to effect a J-inner coprime
factorization of G1(s), a system which is made up from the given generalized plant. Thus
A + BK is a stability matrix. However the H:xJ state feedback controller, (9.2), affects only
u 2 (t) through the state feedback equation u 2 (t) = K 2 x(t). In addition, recall from Section
9.2.4, that condition (ii), (9.5), is satisfied provided K 2 is the last m 2 rows of the matrix K

'
..I

I
JI
.

of G1(s). There fore it remains to

which is neede d in the J-inne r coprim e factor izatio n
is determ ined in the foregoing
K
when
ed
establish that condi tion (i), (9.5), is satisfi
manne r.
has a stabilizing soluti on X 00 ,
More specifically, assuming that the HCAR E, (10.96),
to show that A+ B 2 K 2 has all its
and that K, (9.43), is determ ined using X= X 0 , , we need
m rows of K. This is done in the
eigenvalues in the open left half plane when K 2 is the last 2
proof of the following theorem.
0 is a stabilizing soluti on to
Theorem 9.2 Ifthe pair (A, B 2 ) is stabilizable and X 00 2:
half plane where
left
open
the HCAR E, (10.96), then >.[A+ B 2K 2] are all in the

with K given by (9.43).

A
Proof Suppo se X 00 is a stabilizing soluti on to (9.46). Then 001
where

R 001 X 00 is stable

(9.47)
with 12 = A + B2K2.
B ) is necessary for 1 2 to be
Notice, from Chapt er 2, that the stabilizabilty of (A, 2
is detect able if we can find L
C)
(A,
stable. In additi on, recall from Chapt er 3, that a pair
is detect able all unstab le
C)
(A,
such that A+ LC is stable. Alternatively, when
then Cv -I 0. There fore
>.v,
=
Av
eigenvalues of A are observable, i.e., if Re[>.] 2: 0 when
Thus if K 1v = 0 where
able.
detect
since A2 +B1Kh (9.47), is stable, we have (A 2 ,Kt)
.
proof
the
in
1 2v = >.v then Re[.>.] < 0. We will use this fact later
to and from (9 .45) to obtain
Next we add and subtra ct both X 00 B 2 K 2 and its transp ose

(9.48)
where

8 = XooB2K2

+ K[ BfXoo + KrD, cK- CfCt

that this fact plus the

Now we are going to show that e ~ 0. Then we will see
(A2, K 1) is detect able implies that 1 2 is stable.
We begin by expan ding KT D1cK

KTD .K = [KT1
Jc

KT] [Mt
2

fact that

M2] [Kt]
K2

M[ M4

= K{ MtKt + K{ M2K2 + K[ Mf Kt + K[ M4K2

where we see from (9 .46) that the M;s are given as

Then 8, (9.48), becomes

2- Cf C1
= (XocB2 + K{ M2)K2 + K{ (Bf Xoc + M[ Kt) + K{ MtKt + K{ M4K

(9.49)

238

H= State Feedback and Estimation

Next after rearranging (9.43) as

we see that
(9.50)
Therefore using (9.50) in 8, (9.49), gives

+ Df2C1)T K2- K!(M4K2 + Df2C1)

-(M4K2

-KJ M4K2- C{ D12K2- K'{ D{2C1

-t-

+ K[ M1K1

K[ivf1K1

+ K'{M4K2- C{ C1

cf C1
(9.51)

Finally, r-ecall from (9.6) that we require all the eigenvalues of D{1D 11 to be less that 1 2
since this is a necessary condition for the solution of the H 00 state feedback control
problem. We can readily show that under this condition we have A1 1 < 0. Thus vve see
from (9.51) that 8 :S 0. We can now proceed to show that A2 is stable.
Pre and post multiplying (9.48) by,,* and t' where A2 v = .Xv gives
2Re[.X]v*X00 v = v*8v

(9.52)

Now since we have just shown that 8 :S 0 we have either v*8v < 0 or v*8v = 0.
If v*8v < 0 it follows from the condition X 00 2: 0 that v* Xocv > 0 and therefore
Re[.X] < 0
If v* 8v

0 we see from (9 .51) that

(9.53)

and since M 1 < 0 we see that (9.53) requires

and

(9.54)

Finally, recalling from the beginning of the proof that the pair (A 2 , K 1) is detectable, we
see that if K 1v = 0 then Re[.X] < 0.

To recap, we have shown that the state feedback matrix K 2 which solves the H 00 state
feedback control problem can be determined as the last m 2 rows of the matrix K, (9.43),
when X= Xx the stabilizing solution to the HCARE, (9.46). A remarkable feature of
this solution, shown in Theorem 9.2, is that while X)C makes A+ BK stable we also have
A + BK2 stable. Before discussing conditions on the generalized plant which ensure the
existence of a stabilizing solution to the HCARE. we pose and solve the H x state
estimation problem.

. !1.

9.3.3

Dete rmin ation of ~

Recall that in Section 9.2.4 we made use of the condit ion

We will shmv now that this condit ion is satisfied when

O"max[Dil] <I

and

Df2 D 12 is invertible

of D 11 , D 12 so that (9.36) is
We do this by first developing an expression for~ in terms
satisfied. This is done in the following theore m.
Theorem 9.3 If
and
then 6 can be taken as

where

with
(9.55)

where ],1mm is giYen in (9 .16) and

giYen by
Proof Consid er the consta nt symmetric matrix factori zation

~]

(9.56)

from (9.56) that

Then identifying the blocks of D Jc with the Mis, (9.55), we see
(9.57)

where

r' e are defined in the theorem.

240

H 00 State Feedback and Estimation

Next assume for the time being that r, 8 can be factored in the manner specified in the
theorem. Then we can write

where Jymm was given in (9.16) and

Finally refering to (9.57) we have

so that D Jc can be factored as (9.55) provided 8, r have square root factors.

Recall fmm the proof of Theorem 6.1 that a matrix has a square root factor provided it
is symmetric and non-negative. Now 8 and r are seen to be symmetric by inspection.
However unlike 8 it is not evident that r is non-negative.
To show that r is non-negative notice that it can be expressed as
(9.58)

where De was used in Section 6.5.2

Next recall from Chapter 7 that the singular value

vectors vi, u; for D 11 are related as

a;

and corresponding singular

where

Therefore noting that

T

Dllu = a;v

we see from (9.58) that

(9.59)

However we saw, following the proof of Theorem 6.1, that De is a contraction so that
(9.60)

Hoo State Feedback Controller

241

Therefore from (9.59, 9.60) we have

v

iTr ;

v ?."!-a;

and

for all eigenvectors, vi : i = 1, 2 p 1 of Df1D 11 provided the singular values of D 11 satisfy

i

= 1, 2 p 1

Finally, since Df1D 11 is symmetric, its eigenvectors, v\ are complete and it follows that
if

then

for any p 1 dimension al vector v. Therefore r ?. 0 when the necessary condition (9.6) is

satisfied, and we can factor r in the manner stated in the theorem.

from
see
we
Thus
(9.31).
prove
to
used
be
now
can
The foregoing expression for 6
Theorem 9.3 that

and (9 .31) follows.

Finally, recall that K, (9.43), requires D 1c to be invertible. We can show that D 1c is
indeed invertible, by using the factorizati on of D 1 c given in Theorem 9.3 to obtain an
expression D]c1 as
-! = 6
D Jc

-lj-1

A-T

rmmL.l.

(9 .61)

Then we see, from the lower triangular structure of 6, (9.55), that

.11-1 =

(9.62)

where r. e are given in (9.55).

In the next chapter we will use the form for 6 given in Theorem 9.3 to develop the
solution to the Hx output feedback control problem. Before we can do this we need to
develop a solution to the H 00 state estimation problem.

242

9.4

H 00 State Feedback and Estimation

H 00 State Estimation Problem

Recall from the introduction to this chapter that our ultimate goal is the determination of
an Rx controller which uses the measured output to determine the controlled input to the
generalized plant so that the resulting closed loop system is internally stable and has
transfer function from the disturbance input to the desired output, T0 (s), with H 00 norm
no greater than r.
In this section we treat the problem of determining an estimator or observer for the
generalized plant state, x(t), from knowledge of the measured output, y2 (t), and
controlled input, u2 (t). We need to do this so that we can use this state estimator in
conjunction with the RXJ state feedback controller determined in the previous two
sections to obtain the desired Hoc controller. Notice that since the state of the generalized
plant depends on the disturbance input uJ(t), \vhich is unknown. it is not possible to
design an observer to give an asymptotic estimate of the generalized plant state. Instead
we design the observer so that the L2 gain from the disturbance input to desired output
estimation error, _v 1(t), is no greater than{, i.e., so that
(9.63)

where
(9.64)

and .h (t) is the estimate of the desired output,

(9.65)

Alternatively, we must determine the observer so that the transfer function Te(s) from
u 1 (t) to ji 1 (t) satisfies

and

(9.66)

As in the quadratic and LQG state estimation problems where the solution was
obtained by dualizing the solution to the corresponding state feedback control problem,
we will see now that the solution to the Hx state estimation problem can be obtained by
dualizing the solution we obtained in previous sections for the Hoc state feedback control
problem.

9.4.1

Determination ofT e ( s)

Recalling the development of observers given in Chapter 3. we see that an observer which
estimates the state of the generalized plant (9 .1) using u21r) and .1 2 (r) is constructed as
(9.67)

where

243

H 00 State Estim ation Prob lem

d plant , (9.1), we see that

However, from the state mode l for the generalize
(9.68)
alized plant state estim ation error ,
Thus using (9.67, 9.68) we see that the gener
l equa tion
i(t) = x(t) - x(t). is the solution to the differentia
(9.69)
the desired outp ut estim ation error ,
In addit ion, we see from (9.1, 9.64, 9.65) that
distu rbanc e input as
y1 (t) = )' 1 (t)- y 1 (t), depends on the state estimation error and
(9.70)
fer funct ion Te(s) from u1(t) to
There fore we see from (9.69, 9.70) that the trans
has state mode l given as
(B! -+- L2D21)]
-Dll

y1 (t)

(9.71)

from this state mode l for Te(s) that

Recalling the discussion following (9.5), we see
condition (ii). (9.66), cann ot be satisfied if

h we requi red to solve the H 00 state

Thus the same necessary cond ition, (9.6), whic
solve the Hx state estim ation probl em
feedback contr ol probl em. is needed now to
namely

(9.72)
for the deter mina tion of L 2 so that
We will see now that the design equa tions
ned by using the fact that the prese nt
conditions (i, ii), (9.66). are satisfied can be obtai
the Hx state feedb ack probl em solved
H')(; state estim ation probl em is the dual of
previously in this chap ter.

9.4.2

Dua lity

Te(s) is defined as
Recall, from Chap ter 7, that the HOG norm of

))Te(s))lx=

sup

amaxUw)

c<:E(-X .e<c)

Teljw).
where a 111 a,(jw ) is the largest singular value of
the positive squa re root of the largest
is
(jw)
More specifically, recall that a max
(ju.-) Te(j;.,.,) or Te(I.;.;) T,~ (jw). Ther efore
eigem alue of either of the Herm itian matri ces r;
largest singu lar value of r;(Jw ). As
the largest singu lar value of Te(jw) equa ls the
spon ding to a system having real-r ation al
discussed in Chap ter 8, the adjoi nt system corre

244

H= State Feedback and Estimation

transfer function Te(s) has transfer function T[ ( -s) and satisfies

r;(s)

ls=jw=

T[(-s)

ls=jw

Therefore condition (ii), (9.66), is equivalent to the condition

IIT[(-s)llx ~I

(9.73)

where we see from (9.71) that

(9.74)
Notice that condition (i), (9.66), is satisfied if and only if T[ (-s) is antistable since
Te(s) E 1i00 (s) is satisfied if and only if T[ ( -s) E 1i~(s).
Thus we see from the foregoing that the H 00 state estimation problem and the H 00 state
feedback con.trol problem are dual problems in the sense that a solution to the H 00 state
estimation problem requires that

(i): T[(-s)

1i~

and

(ii): IIT[(-s)iloo~ I

(9.75)

whereas a solution to the H 00 state feedback control problem requires that

(9.76)

9.4.3

Design Equations for L2

In order to exploit the foregoing dual nature of the H 00 state feedback control problem
and H 00 state estimation problem, (9.75, 9.76), we need to establish a symbol correspondence between these problems. This can be done by comparing the state models for Tc(s)
and TT ( -s), (9.4, 9.74). This gives the following correspondences between their parameters.
Recall that we determine K 2 in the H"" state feedback problem by first solving the
HCARE, (10.96), for its stabilizing solution, X 00 Then K 2 is the lastm 2 rows of the matrix
K which we determine by setting X= X 00 in (9.43). Therefore, proceeding in an
analogous manner we determine L 2 as the last p 2 columns of the matrix L obtained by
using the stabilizing solution Y oo to an algebraic Riccati equation. More specifically, the
design equations for determining L are obtained by replacing parameters in (9.46) and
(9.43) using the parameter equivalences in Table 9.1. Thus doing this and also replacing
-X with Y gives
(9.77)

and
(9.78)

Table 9.1

State mode l Para mete r Corr espo nden ces

Control

Estimation

-AT

-c!

Be

LI

K2

-ci

BI

-Bf

cl

-D~

Dl2

-Df1

D11

where
A 00 2

=A - B1D f DJ,} C

Roo2

CT D}ol C

Qoo2

B1

[I- D[ D]01D2JBf

with

)
equa tion like the GFA RE, (Section 6.3.2
Noti ce that (9.77) is an algebraic Riccati
es
mak
h
whic
Y
tion
HF ARE . The solu
which is refer red to here as the
and
as the stabilizing solu tion to the HFA RE
to
red
A 2 - YR 2 =A + LC stabl e is refer
from
ed
rmin
dete
is
K
n
as A + B 2 K 2 is stable whe
is deno ted Y oo and satisfies Y 00 2 0. Just
g Y oo
le when Lis dete rmin ed from (9.78) usin
stab
C
(9.43) using XX) so too is A+ L 2 2
\Vhere
c is
of D 12 and cond ition (9.72) ensu re that D 1
Just as the inde pend ence of the colu mns
re
ensu
)
(9.72
ition
cond
of the rows of D 21 and
invertible so too does the inde pend ence
that D Je. is invertible.
be
back prob lem we requ ired (A, B2 ) to
Finally, recall that in the H')C state feed
be
to
K
B
2
A+
2
for
r
solu tion X)() 2 0 in orde
stabilizable and the HCA RE to have a
to
RE
HFA
the
need
state estim ation prob lem we
stable. We see by anal ogy that in the Hex,
e.
stabl
be
to
C
L
A+
2
2
dete ctab le in orde r for
have a solu tion Y 00 2 0 and (A, C 2 ) to be

9.5

Sufficient Conditions

s in
ralized plan t mus t satisfy certa in cond ition
Recall that a given state model for the gene
6).
pter
Cha
of
stabilizing solu tions (Section 5
orde r for the GCA RE and GF ARE to have
the
to
tions
solu
re the existence of stabilizing
A similar set of cond ition s which ensu

246

H= State Feedback and Estimation

HCARE and HF ARE are

(Al)
(A2)

(A, B 2 ) and (A, C2 ) are respectively stabilizable and detectable pairs.

(A3)

(A- jwl)

cl

and

(A- jwl)

c2

have respectively independent columns and independent rows for all real w.
D 12 and D 21 have respectively independent columns and independent rows.

Taken together, (Al-A3) constitute a sufficient condition for the solution of both the H 00
state feedback control problem and the H 00 state estimation problem.
Notice that (Al,A2) ensure that the Hamiltonian matrices corresponding to the
HCARE and HF ARE have no imaginary axis eigenvalues. This can be seen by noting
the similarity of the HCARE with the GCARE in Chapter 6 and referring to Theorem 6.1
in Section 6.5.2.

9.6

Summary

In this chapter we have given a derivation of the design equations for solving the H 00 state
feedback control and state estimation problems for a given linear time-invariant,
continuous-time plant. Unlike the solutions to the corresponding quadratic or LQG
problems obtained in Chapters 5 and 6, solutions obtained here for the H 00 problems do
not minimize any performance criterion or cost. Rather the solutions to the H 00 problems
achieve a specified upper bound on the L 2 gain from disturbance input to desired output
with the resulting closed loop systems being stable. As in the quadratic and LQG control
problems, algebraic Riccati equations still play a central role in the solution. In the next
chapter we will see how the solutions obtained in this chapter can be used to provide
controllers which use the measured plant output instead of the plant state to achieve the
just mentioned Hoc norm bounding and closed loop stability.

9.7

Notes and References

A great deal of work has been done on the Hx control problem since its inception over a
decade ago. One of the first approaches solved this problem by beginning with a Youla
parametrization of all stabilizing controllers for the given plant and finished by having to
solve a model matching problem. This work is the subject of [14]. This and other early
approaches which are reviewed in [8] and [20] are quite complicated and have controllers
with unnecessarily high dimension. Indications for a simpler approach came from
research on stability robustness [35], which lead to the two-Riccati-equation solution
given in this chapter [12, 16]. Following this similar results were obtained using J type
factorizations, e.g., J spectral, Jlossless coprime, [19], [41], [20], [40]. Discussion of this
approach in connection with the chain-scattering formulation of the generalized plant is
given in [24].

10
H 00 Output Fe ed ba ck an d
Control

10.1

Introduction

state
prob lems of state feedback cont rol and
Having seen how to solve the basic Hoc
a
ing
rmin
we come now to the prob lem of dete
estimation for the given generalized plan t,
that:
inpu t from the measured outp ut such
controller which generates the controlled
e and
(i) the closed loop system is internally stabl
to desired outp ut is no grea ter than a given
t
inpu
ce
rban
(ii) the H 00 norm from the distu
positive scalar J.
ut feedback cont rol problem. The setup for
This prob lem is referred to as the H 00 outp
9.
given previously at the start of Cha pter
the outp ut feedback cont rol problem was
fer
trans
that the generalized plan t has know n
Again, as in that chap ter, we are assuming
function G(s) and state model, i.e.,

, [A B]

G(s) "=

(10.1)

where
l./1 (s)

C2(s)

disturbance inpu t
controlled inpu t

desired outp ut
mea sure d outp ut

that there are man y controllers which solve

In the development to follow, we will see
rs
Moreover. we will show that these controlle
the H:x. outp ut feedback control problem.
ng
havi
m
syste
the
for
rs
meterized controlle
constitute a cons train ed subset ofY oula para
Theorem 8.5, that Youla parameterized
from
ll,
Reca
ut.
outp
Ll:,(sl as inpu t and Y 2 (s) as
to
ition (i) is satisfied. We will show how
controllers are stabilizing. Therefore cond
(ii).
constrain this set so as to achieve cond ition

248

H 00 Output Feedback and Control

10.2

Development

Recall, from Section 9.3.2, that the partition K 2 of K which effects a J-inner coprime
factorization of G\ (s) stabilizes the generalized plant. In addition, recall from Section
9.2.4 that P.., ~ 0 for all 0 1 (s) E H 2 and that this implies that P1 ~ 0 for all [' 1 (s) E 7{2
where
I 10.2)

This fact is used now to develop a solution to the H 00 output feedback control problem.
The plan for doing this is given as follows.
1. We begin by assuming that the disturbance input U1(s) is given by (10.2) with K 1
determined in Section 9. 3.1.
2. Then we show that the signals V1 (s). V2 (s) can be replaced in the expression for the
quadratic performance index P1 , introduced in Section 9.2.1, by the "external"
disturbance input 0 1 (s) and an output-like signal, Y0 (s). This gives P. = P. 0 where
P 10 is us.\(d to denote the new expression of L. The output-like signal, Y 0 (s). depends
on 0 1(s), U2 (s) and the state of the given state model for the generalized plant.
3. Next we determine an estimator for the generalized plant state which uses y 0 (t) and
achieves Y 10 ~ 0 for all 0 1 (s) E H 2 . This is accomplished by adapting the H 00 state
estimator obtained in the previous chapter.
4. Following this we replace the actual generalized plant state in the Hoc state feedback
controller determined in Chapter 9, by the estimate of this state which is generated by
the state estimator developed in the previous step. This leads to a controller which
achieves P., 0 ~ 0 for all 0 1 (s) E 7{ 2 but requires J 0 (t) as an input in addition to the
measured output y 2 (t).
5. Since y 0 (t) is not available for use as an input to the controller, the remainder of the
development concentrates on eliminating the need for y 0 (t) as an input to the
controller.

10.2.1

Reformulation of P1

Recall from the development of the solution to the H 00 state feedback control problem
(Section 9.2.1) that P., ~ 0 for all u 1 (t) E 2 [0, :x) implies that condition (ii) is satisfied.
We are going to use this fact to solve the Hx output feedback control problem. Also
recall, (Section 9.2.3), that P,1 was expressed in terms of the signals Vds). V 2 (s), which
resulted from a J-inner coprime factorization. i.e.,

(10.3)
and recalling (9.25) we have

need
to the H 00 state estim atio n prob lem we
Now in orde r to inco rpor ate the solu tion
state
the
g
min
assu
ed plan t state. We do this by
tore -exp ress P 1 in term s of the generaliz
can
L\
that
and
9.3.1
ion
man ner specified in Sect
feedback mat rix K is dete rmin ed in the
be taken, from The orem 9.3, as
( 10.4)

where

8 =

eT/ 2 8 1/ 2

= D{2D12

[ = rTI 2[ 112 =-/ Im,

- D{1(Ip, - D12 8- 1Df2)D11

generalized
) we see that the state mod els for the
The n subs titut ing for U1- (s)1 from (10.2
become
plan t G(s), (10.1), and Ar (s), (10.3),

(1 0.5)

(10.6)

where

els for
s and inpu ts to the foregoing state mod
::\otice that the state differential equa tion
these
of
each
of
s
notice that the initial state
G(s) and i;j-l (s) are identical. In addi tion
uts
outp
to
ts
g tran sfer functions to relate inpu
models are identical because we are usin
e
thes
of
s
state
each mod el be null. The refo re the
which requires that the initial state of
the
n
whe
t
plan
l to the state of the generalized
models are identical for all time and equa
equa tion
obse rvat ion toge ther with the outp ut
distu rban ce inpu t is given by (10.2). This
t on the
nden
depe
1
), reveals that V 1 (s), V2 (s) are
of the state mod el for iU- (s), (10.6
n as
generalized plan t state X(s) and are give
(10.7)
V 1 (s) = i:\ 1 G\(s)
(10.8)
U2(s)
V2 (s) = -L\4 K2X (s) + L\3U1 C1) + L\ 4
the L\is.
Vi(s)s together with the relat ions for
We can use these expressions for the
.
rem
ner given in the following theo
(10.4) tore -exp ress PT (10.3), in the man

250

H 00 Output Feedback and Control

Theorem 10.1 If the state feedback matrix K is determined as in the previous chapter
so that its partition K 2 solves the H 00 state feedback control problem, then P'Y, (10.3), can
be re-expressed as P'Y = P'Yo where

(10.9)
with

and X(s) being the state of the generalized plant.

Proof From (10.7) we see that

-;

2'

2i

J I (s)llz= -~ ~~~~ U1 (s)ll2= -( \~1 U1 (s). ~~

~.

[I

(s))

(10.10)

where

Similarly from (10.8) we see that

II V2(s) II~ = II~4K2X(s) II~+II~4U2(s) 11~+11~3 ()I (s) 11~-2(~4K2X(s), ~4 U2(s))

- 2(~4KzX(s), ~3 ()I (s)) + 2(~4U2(s), ~3 ()I (s))
However from the dependency of the
T

~~ ~~ = lm,-

-2

D11Dll

~is

(10.11)

on D 11 and D 12 , (10.4), we have

T
( T
)~! T
+ "Y -2 D11D12
D12D12
D12D11

~f ~3

Dt!D12(Df2D12r 1DfzD11

~r ~3

Df2D11

~r ~4

Df2D12

Therefore using these relations in (10.10, 10.11) gives

( 10.12)
and

- 2(D 12 K2X(s), D 12 U2(s))- 2(D 12 K2X(s), D11 ()! (s))

+2(D 12 U2

.D 11 U1 (s)i

Finally substituting (10.12) and (10.13) in P. ,(10.3), gives P.

il0.13)
=

P'Yo (10.9).

ralized plan t state is available for state

Notice that, in the ideal case when the gene
Y (s) = D 11 0 1 (s) and P"~0 ' (10.9 ), becomes
feedback, we can set U2(s) = K2 X(s) so that 0

P"~o = (Oh (-"-/ lm, +Df 1Du )U1)

we are assuming that O"max[DIJ] ~ "'(.
which satisfies P10 ~ 0 for all U1 (s) E 'H 2 since
prese nt case, we use its estimate, X(s), to
Alternatively, since X(s) is unkn own in the
replacing U2 (s) by K2 X(s) in Y0 (s) and
generate the cont rolle d input. Therefore
substituting the result in P10 , (10.9), yield
(10.14)

where

Y0 (s) = Y0 (s) - Y0 (s)

= D12K2X(s) -

D11

[h (s)

X(s) = X(s) - X(s)

P ~ 0 for all 0 1 (s)
Thus we see from (10.14) that in orde r for 10
develop an estim ator for x(t) so that

(i) : Tac(s) E 'Hoo

(ii): IIToc(s)lloo~ "!

'H 2 we need to
(10.15)

where
(10.16)

We obta in a solut ion to this problem by

which is an H 00 state estimation problem.
ation prob lem obta in in Section 4 of
adapting the solut ion to the H 00 state estim
Chap ter 9.

10.2.2

An H 00 state est ima tor

prob lem just pose d with the H 00 state

A comp ariso n of the Hoc state estim ation
ls that the role played by the desired
estimation prob lem solved in Chap ter 9, revea
d by y0 (t), the error in estim ating y0 (t).
outp ut estim ation error , y1(t), is now being playe
using u1(t) in place of u 1(t).
In addition, in the present situation, we are
the Hoc state estim ation prob lem given in
to
ion
Therefore in orde r to use the solut
a system which relates u1 ( t), u2 ( t) as inpu t
Chap ter 9 we need to develop a state model for
as the generalized plan t. We can obta in
to Yo(t),y 2 (t) as outp ut and has the same state
ut, Y1 (s) from G(s), (10.5), by Y 0 (s), (10.9).
such a system by replacing the desired outp
fer function by G0 (s), we obta in a system
Thus doing this and denoting the resulting trans
which is specified as
(10.17)

252

H 00 Output Feedback and Control

where

Now we could develop a solution to the present H 00 state estimation problem by

obtaining T0 c(s), (1 0.15), from the state model for G0 (s), (1 0.17) and proceeding as we did
with T0 in the previous chapter in the development of an Hx state estimator for the
generalized plant. However, since our goal here is to determine a stable estimator for the
state of G0 (s) subject to the requirement that
(10.18)

we can obtain the design equations for the desired estimator by replacing the parameters
in the design equations for the H::xJ state estimator for the generalized plant state model,
G(s), (10.1), by the corresponding parameters in the state model for Ga. (10.17).
Inspection of the state models for G(s) and G0 (s) shows correspondences between the
state modelparameters as given in Table 10.1.
Therefortthe H 00 state estimator for G0 (s), (10.17), has differential equation
(10.19)

Notice that we are assuming temporarily that y 0 (t) is known and can be used in the
state estimator, (10.19). This assumption will be removed later since the final controller
we are seeking has only y 2 (t) as input and zdt) = K 2.-\:(t) as output. Notice also that L 0 in
(10.19) corresponds to Lin the H::xJ state estimator for the generalized plant given in
Section 9.4.1.
Now the design equations for L 0 are obtained by replacing the parameters in the
HFARE and the equation for L which we obtained in Section 9.4.3 by the corresponding
parameters as indicated in Table 10.1. Thus L 0 is obtained (when it exists) by finding the
solution Yx 3 to the following ARE such that A 0 + L 0 C0 is stable
Ax3 Y oo3

+ Yoo3A~3

- Y oo3Roo3 Yoo3 + Qcc3 = 0

La= -[Yoc3ci + BtDJ]D}ri

where
Ax3

Aa- BtDf D]ot Co

Rod= C"[;DjjCu
Q:xc3

with

Qx2

Bt (I- Df D.!a1 D2)Bf

(10.20)

Table 10.1

the gen
Par ame ter corr espo nde nce s for
(s).
G
ized plan t, G(s), and 0

A
B

C
C,
C2

eral-

A0
B

C0
-D1 2K2
C2 + D 21K1

tion, Y 003 , to the

blem treated in Chapter 9, the solu
As in the H 00 state estimation pro
s Y003 ~ 0 and is
sfie
sati
le,
A 0 + L 0 C0 stab
Y
003 R 003 =
A
es
003
mak
ch
whi
,
ameters of
ARE, (10.20)
ation (10.20). Notice tha t the par
equ
to
tion
solu
ing
iliz
stab
the
problem.
referred to as
the H 00 state feedback control
to
tion
solu
the
e
olv
inv
,
.20)
ends on the
the ARE, (10
solution to equation (10.20) dep
ing
iliz
stab
a
of
e
tenc
exis
the
Therefore
ation suggests tha t the
k control problem. This observ
solution to the H 00 state feedbac
m and the H 00 state
ble
pro
back control
H
the
00 stat e feed
h
bot
to
s
tion
solu
tence of a
existence of
plant, does not guarantee the exis
ized
eral
gen
n
give
a
for
m
ble
trasts with
estimation pro
problem for tha t plant. This con
trol
con
k
bac
feed
put
out
H
h the LQ G
solution to the 00
which has a solution whenever bot
m,
ble
pro
trol
con
k
bac
feed
put
the LQ G out
solutions.
state estimation problems have
state feedback control and LQG
H 00 out put feedback
the
e of a solution to
tenc
exis
the
t
tha
10.4
tion
Sec
in
e stabilizing
We will see
h the HC AR E and HF AR E hav
bot
t
tha
s
uire
req
y
onl
not
m
control proble
solutions satisfies the
addition the pro duc t of these
solutions, X 00 and Yom but in
following inequality

10.2.3

feedback
Introducing estimated state

pla nt under the

state estimator for the generalized
erate the state
So far we have developed an H 00
gen
to
generalized pla nt is being used
assumption tha t the state of the
(t)
j! (t)\u 2 and of Y2(t),
In this case the estimate of y 0 (t), 0
feedback signal, u2 (t) = K2 x(t).
are seen from (1 0.17) to
the generalized plant state, x(t),
Yz(t)iuz(t) based on the estimate of
be
(10.21)
)
Yo(t)lu 2 (t) = -Dl2Kz.X(t) + D12uz(t
(10.22)
D22u 2 (t)
Yz(t)lu 2 (t) = (C2 + D21K1)x(t) +
is generated as
However, if the controlled input

(10.23)

254

H 00 Output Feedback and Control

we see from (10.21, 10.22) that

(10.24)

Yo(t) = Yo(t)- Yo(t) = -yo(t)

h(t) = Y2(t)- Y2(t) = (C2 + D21K1

+ D22Kc)i( t)- Y:>(t)

(10.25)

where we have y0 (t), (10.21), null when the controlled input satisfies (10.23). This fact
plays an important role in the developme nt of the controller.
Continuing , we see that use of(l0.24, 10.25) in (10.19) yields
(10.26)

where

and the controlled input is generated by a controller with transfer function f(s) as
(10.27)
However, since y 0 (t) is unknown we need to eliminate it as part of the input to this
controller in a manner which preserves P'Y0 ::; 0 for all U 1 (s) E H 2 . We will do this in the
next section.
Before we go on to develop these J 0 (t)-free controllers, we summarize what has been
achieved so far in the development of a solution to the H 00 output feedback control
problem.
1. We began by incorporat ing the solution to the Hoc state feedback control problem in
the performan ce index, P,, used in the previous chapter. The expression we obtained
was denoted P 1 ()' It was shown that P 10 depends on YoU), ( 10.9). an output from a
system, (1 0.1 7), whose state equals the state of the generalized plant.
2. Next, in Section 10.2.2, we used the solution to the H 00 state estimation problem
obtained in Chapter 9 to obtain an H 00 state estimator based on the output y 0 (t) so
that the error in estimating Yo(t), i.e., Yo(t), satisfies IIYa(t)ll2< 1llu1 (t)ll2
3. Finally, in Section 10.2.3, we showed that by using the estimated state to generate the
state feedback signal as u2 (t) = K 2 .x(t) and the H::x: state estimator based ony0 (t) gave
E H2 as
a controller which satisfied p 'YO ::; 0 for all cl (s) E H2 or equiYa1ently all ul
problem.
control
feedback
output
Rx;
the
solve
to
required

1 0.3

Hex Output Feedback Controllers

As mentioned before, the controller given by equation ( 10.27) cannot be used as it has the
unknown output-lik e signal, y 0 (t), as an input. In this section we will shO\\ hO\\ this
unknown signal can be removed from the input to the controller, (10.27).

~~~

nw n .
7f~biiijARHNOIMOH vw

to as the cen tral

s to a single con trol ler referred
One way of doi ng this which lead
urb
dist anc e, (Section
dist urb anc e inp ut is the wor st
controller, is to assu me tha t the
9.2.4), so tha t 0 1 (s) is null.
em. This app roa ch
relate y 0 (t) to u 1 (t) thro ugh a syst
Ano ther way of doi ng this is to
feedback con trol
put
out
Hxof which solves the
one
any
lers
trol
con
of
ily
fam
yields a
ame triz ed stabilizing
ily is a sub set of the You la par
fam
this
t
tha
see
will
We
.
problem
d in Cha pte r 8.
ized plan t which we enc oun tere
controllers for the given general

10.3.1

Ce ntr al Co ntr oll er

Yo(t) satisfies
tha t the erro r in esti mat ing
ion,
sect
ious
prev
the
Recall, from
e inde x P'Y 0 ' (10.14) as
re we can write the per form anc
51 0 (t) = -y 0 (t), (10.24). The refo
(10.28)
and the rela tion (10.16) as
(10.29)
t
ned the esti mat or, (10.19). so tha
However recall tha t we dete rmi
(10.30)
10.30) tha t
Thus we see from (10.28, 10.29,

put feedback con trol

(ii) in the stat eme nt of the H 00 out
as is requ ired to satisfy con diti on
problem.
led inp ut, u2 (t),
r sup pos e we generate the con trol
Now 0 1(s) is unk now n. Howeve
anc e inp ut is the wor st
er the assu mpt ion tha t the dist urb
und
),
0.27
(1
ler,
trol
con
the
from
is null. The n in this
und er the assu mpt ion tha t u1(t)
i.e.,
4),
9.2.
n
ctio
(Se
e,
anc
urb
dist
out put , (10.27), is
is null and ther efo re the con trol led
29),
(10.
).
YJ~
t
tha
see
we
n
atio
situ
given as

Lo2]

(10.31)

where

the cen tral controller.

This con trol ler is referred to as
eralized plan t und er
is an Hx stat e esti mat or of the gen
In essence the cen tral con trol ler
the con trol led inp ut
e inp ut is the worst dist urb anc e and
the restriction that the dist urb anc
the Bx state feedback
u (t) = K 2.'{(t), where K 1 solves
depends on the stat e estimate as 2
is free from Yo(t).
of obt aini ng a con trol ler which
con trol pro blem . Thi s is one way

H 00 Output Feedback and Control

10.3.2

Controller parametrization

Suppose thP unknown output Y 0 (s) which is present in the controller, (10.27), is replaced
by Yq,(s) where

Yq,(s)

= ~(s)U 1

(s)

(10.32)

with
(i) : ~(s) E

Hoc

(ii) : jj~(s) 1\oc::; I

Notice that when U1(s) E 1i 2 we have Yq,(s) E Hz and jjYq,(s)jjz::; ri\UJ(s)llz Thus
when Y0 (s) is replaced by Yq,(s) in (10.28) we have

which implies that P10 ::; 0 for U1 (s) E 1i2 as required to satisfy condition (ii) in the
statement of the Hoc output feedback control problem.
Next supp(;se we have W(s) E Hrx; such that
(10.33)
Then replacing Y 0 (s) in (10.33) by

and solving (10.33) for U2 (s) gives the desired controller as

(10.34)

where

with ~(s) constrained as specified in (10.32). The specification of W(s) is given in the
following theorem
Theorem 10.2 The state model for the transfer function W(s), (10.33), satisfies
W(s) E 1ix and has state model

( 10.35)

where

. . 'JtilB .

feed bac k con trol

in the solu tion of the H 00 stat e
Pro of Since K was dete rmi ned

H
E
)
00
W(s
(Section 9.3.1), we hav e
pro blem so that A+ BK is stable,
34) becomes
(10.
that
so
null
)
W (s) we cho ose <P(s
In ord er to dete rmi ne W2 (s) and 4
(10.36)
(10.31) yields
and com pari ng this equ atio n with
(10.37)

cop rim e
dete rmi ne W 2 (s) and W4 (s) by
H
E
00 , we can
(s)
W
(s),
4
W
e
2
sinc
Now
and W 4 (s) = Mr(s)
lly, we cho ose W2(s) = Nr(s)
factorizing f 2 (s). Mo re specifica
s) cop rim e.
Mr(
s),
1
Nr(s), Mr(s) E H 00 and Nr(
where f 2(s) = Nr( s)M i (s) with
as
oted
den
is
(s)
stat e mod el for r 2
Recall from Sec tion 8.4.1 that if the

rim e fact ors off 2 (s) are given as

then stat e mod els for the righ t cop

Nr(s) ~ [Ar +B rKr

Cr +D rKr

Br]
Dr

Mr(s) E H 00
where Kr is cho sen so that Nr(s),
Now from (10.31) we hav e
Dr =0
gives the cop rim e fact ors as
so that cho osin g Kr = (C2 + D 3 K)

s
Nr(s) = Wz(s) =

[A+ BK

-Lo2]

Kz

[ A + BK

s
Mr(s) = W 4 (s) = Cz

+ D3K

(10.38)

Lo2]
I

of the Hoc stat e

K was dete rmi ned in the solu tion
where Nr(s), Mr(s) E 'Hoo since
s com plet es the
Thi
le.
stab
9.3.1) so that A+ BK is
feedback con trol pro blem , (Section
W (s) and W 4 (s).
dete rmi nati on of stat e mod els for 2
that
we recall from (10.17) and (10.33)
(s)
Nex t in ord er to dete rmi ne W 1
(10.39)
Y0 (s) = Gal(s)Dl (s) + GoZ(s)Uz(s)
(10.40)
s)U (s)

Uz(s)

WJ(s)Ya(s)

+ W2(

258

H 00 Output Feedback and Control

Now if the system inverse,

6;;21 (s), were to exist we could solve (10.39) for U2 (s) as

(10.41)

and we would have W1 (s) = 6;;21 (s) from a comparison of (10.41) with (10.40). However
we see from (10.17) that 6a2 (s) is not invertible in general since its state model,

has aD matrix, D 12 , which is not invertible in general, (Section 8.2). This difficulty can be
overcome by recalling that we are assuming D[2 D 12 is invertible since this is a necessary
condition for the existence of a solution to the H 00 state feedback control problem. Thus
G02 (s) = D'[2 6 02 (s) is invertible where
(10.42)

Therefore we rewrite (10.39, 10.40) as

Y (s) =Go! (s)Ut(s) + Go2(s)Uz(s)

(10.43)

WI (s) Y0 (s) + Wz(s) 0 1 (s)

(10.44)

U2(s)

where

Then solving (10.43) for U2 (s)

(10.45)

and comparing (1 0.44) with (10.45) yields

( 10.46)

Finally, using the form for the state model of a system inverse, (Section 8.2), we see
from (I 0.42) that
( 10.47)

and (10.47) yields

( 10.48)

DLD 12 =I, given as

where DL is a left inverse of D 12 , i.e.,

(10.49)

T
D t12 = (DT12D12 )-l D12

(s).
This completes the determination of W 1
and in
the second block equa tion s in (10.33)
ite
rewr
(s),
W
e
3
In order to determin
(l0.1 7)as
(10.50)
Yz(s) = W3(s)Y0 (s) + W4(s)01(s)
(10.51)
Yz(s) = G0 3(s)01(s) + G0 4(s)U2(s)
0) in (10.51)
Then substituting for U2 (s) from (10.4
( 10.52)
als that
and comparing (10.50) with (10.52) reve

the
(10.17), and W 1(s), ( 10.48) toge ther with
The n using the state models for G04 (s),
in
tion 8.1 ), we obta
rule for combining systems in series, (Sec

ugh the
out affecting W 3 (s). We can do this thro
Nex t this state model can reduced with
dina te
coor
the
let
dinates. Therefore if we
use of an appr opri ate change of coor
tran sfor mati on matr ix be

dina tes is
the state model for W3 (s) in the new coor
[

T- 1 A w T

r- 1 B w

CwT

Dw

[A +0 BK

r [ C + D3K
2

where

l1

= ~ w J3 w
LCw

Dw J

~2~2KJ

[B2;j l]

D22K2;

D=zDi 2

260

H 00 Output Feedbac k and Control

Finally, notice that W3 (s) is given by

W3 (s)

l -

= Cw(sl- Aw)- Bw + Dw
= (C2 + D3K)[si - A- BKr 1B2DL + D22 DL

which implies that W 3 (s) has a state model given as

In summary , in this section we have determin ed a family of controlle rs, H(s), (1034.
next
10.35), any one of which solves the H 00 output feedback control problem. In the
ized
parametr
oula
Y
the
of
subset
a
to
rs
controlle
section we relate this family of
controllers.

1 0.3.3

Relatio n to You/a param etrizat ion

rs.
Recall, from Theorem 8.5, that the Y oula parametr ization of stabilizing controlle
8.4.3).
(Section
H(s), for a plant, G(s), involves the doubly coprime factoriza tion of G(s),
These controlle rs are given by
(10.53)

where

with

-~H(s)l [~G(s)
MG(s)

;\G(s)

Now in the present situation , the H'Xl output feedback controlle r is connecte d between
the generalized plant's measured output, Y2 (t) and controlle d input, u 2 (t). Moreove r, this
controlle r must stabilize the generalized plant for all disturban ce inputs satisfying
(Section 9.2.4). Therefore.
ul (s) E H2 including the worst disturban ce, Uj (t) = Klx(t)
assuming the disturban ce input is the worst disturban ce, we see that the controlle r must
stabilize a plant G(s) = Gn(s), where from (10.5) we have
(10.54)

systems involved in a doub ly coprime

Thus recalling the state models for the
has a doub ly coprime factorization given by
factorization, (8.46, 8.53), we see that G22 (s)

Nn(s )

[ Ae

~=eKe ~ J

Nc(s) ~ [ Ac + BeK e
Ce+ DeK e

Be]
De

Mn( s) d: [ Ae + BeK e
Ce+ DeK e
Mc(s )

L1c]
(10.55)

~e]

[ AG +;:GKG

as specified in (10.54)
Then substituting the state model para mete rs
and Le = -L02 in (10.55) gives

and setting Ke

= K2

(10.56)

in the general
, we can repla ce Q6 (s)
However, since 2 Q6 (s) E 1(::o if Qe(s) E 'H 00
(s) so that
DL<P
by
controller, (10.53),
expression for the parametrized stabilizing

D1

<P(s) E 'H 00

(10.57)

where

the generalized plan t and have the same

Thus controllers satisfying (10.57) stabilize
determined in the previous section, (10.34,
form as the H 00 outp ut feedback controllers
n <P(s) E 1-lx which ensures that the control
10.35). Notice that in addition to the restrictio
lem we also require II<P(s)llooS "(in orde r
system is internally stable, in the present prob
l)y (t))b s 'YIIu1 (t))lz.
that the cont rol system satisfies the condition 1

10.4

H00 Separation Principle

the controller which solves LQG outp ut

Recall, from Chap ter 6, that we obta ined
corresponding solutions to the basic LQG
feedback cont rol prob lem directly from the
cont rol problems. We referred to this fact
plant state estimation and plan t state feedback
case, being able to solve the correspondas the separation principle. However in the Hx
estimation problems is only a necessary
ing basic Hoc state feedback cont rol and state
ut feedback cont rol problem.
condition for being able to solve the H 00 outp
l condition needed to ensure the existence
In this section we will show that the addi tiona
rol problem is
of the solution to the Hoc outp ut feedback cont
(10.58)

262

H 00 Output Feedback and Control

'

..0

where X 00 , Y 00 are stabilizing solutions to the HCARE and HFARE, (9.46, 9.77). Before
doing this we show now that Y = 3 and L 0 , (10.20), needed to construct the controllers,
(10.34, HU5) can be calculated directly from the stabilizing solutions to the HCARE and
HF ARE as follows:
( 10.59)

ZL 2 ]

(10.60)

where

Z=(I- 1 - 2 Y DC X CXJ )- 1

with L being the solution to the H 00 state estimation problem for the giwn generalized
plant (9.78).
The approach used to show these results relies on a similarity relation between the
Hamiltonian matrices associated with the HFARE. (9.77), and the ARE. (10.20).

10.4.1

A relation between Hamiltonians

We are going to make use of the following ideas from Chapter 5.

The Hamiltonian matrix Hp associated with the ARE

AP+PA 7

PRP+ Q = 0

110.61)

1s g1ven as

(10.62)
The eigenvalues of H p exhibit mirror image symmetry across the imaginary axis.

A[Hp]

A[(A- PR{]

A -(A-PR)]

This fact is seen from the following similarity transformation of H p

fl

=
P

rl H T
P

= [

(A-PR)
AP+PA*-PRP+Q

-(A~ PR)]

where

T
with

-P

Hp being block-upper triangular when

0]I

P satisfies (10.61).

(10.63)

<

263

H00 Separation Principle

Notice that equating the first n columns on either side of (10.63) yields
(10.64)
so that if Pis a stabilizing solution to (10.61) then

is the stable subspace of H P

Now the relation between the Hamiltonian matrices associated with the HFARE,
(9.77), and the ARE, (10.20) which is given in the following theorem will be used to show
(10.59, 10.60).
Theorem 10.3 The Hamiltonian matrix H 2 correspondin g to the HFARE (9.77) and
the Hamiltonian matrix H 3 correspondin g to the ARE (10.20) are similar and related as
(10.65)
where
i

= 2,3

Proof Suppose (10.65) is valid. Then carrying out the matrix multiplicatio n which is
indicated on the right side of (10.65) enables (10.65) to be rewritten as

H2

= [

(Aoo3

-~- 2 Qoo3Xoc) T

r- 2 (A;;:,3xoo + XocAoo3) -~- 4 XooQx3Xoo + Rx3]

-(Aoc3 -I- 2 Qoo3X"")

Qoo3

(10.66)
Then equating like positioned blocks on either side of (10.66) yields
(10.67)

Q""2 = Qoc3
Aoo2 =

Aoo3 -!- 2 Q""3Xoo

R""2

I- 2 (A~3Xao

(10.68)

+ XooAoc3)

-~- 4 XxQoo3Xoc

+ Roo3

(10.69)

Now we show that (10.65) is satisfied by showing that (10.67-10.69) are satisfied.
Recall from (10.20) and the HFARE, (9.77), that

A=2

A - B1 Df D.!a1 C

C~D7t}C0

R=2

c 7 v--Jo1c

B1(I- D[Dj}D2)B f

Q=2

B, (I- D2T D}a1 D2 ) B,T

Ax3

A0

Rx3
Qoo3

81 Df D_j} Co

(10.70)

H= Output Feedback and Control

264

where
Ao

D2

[Dll]

D;o

[ -"/IP~ + D 11Df1
D 21 D 11

A+B 1K 1

D21

DllD~l

Co

D21D21

[ -D12K2

C2 +D2,KI

and (10.67) is satisfied.

Next in order to show (10.68) we use (10.70) so that
(10.71)

Now we are going tore-express the right side of this equation using results from the H 00
state feedback control problem.
Recall from the solution to the H 00 state feedback control problem given in Chapter 9,
(9.43), that
(10.72)

Now writing D Jc as

~]

(10.73)

where

enables (10.72) to be rewritten as

DIITc IK =I2 K,- B,Txoo

(10.74)
(10.75)

where
C,K=D 1K+C 1

(10.76)

Then K 1 is obtained from (10.74) as

(10.77)

Next we can readily show that

H 00 Separ ation Princi ple

265

where

which after substit uting for K 1 from (1 0. 77) becomes

( 10.78)

Howev er writing D 1a. (10.70), as

gtves

and (10.78) can be rewritt en as

(10.79)
right side of (10.71) gives
Return ing to (10.71) we see that use of (10.79, 10. 77) on the

AX!,-

Ao.c2 =; - 2B1 (Bf Xoo + Df1 elK)

-~- 2 B1Df ( [~]elK+ D]a1D2Bf Xoo)

= ~~- 2 (B1Bf X:xc- B1Df Dj} D2Bf Xoc)

is satisfied.
and we see from the definition of Qoc 3 , (10.70), that (10.68)
Finally it remain s to show that (1 0.69) is satisfied, i.e., that
(10.80)
where

We begin by using (10.67, 10.68) to express

n as

H 00 Output Feedback and Control

266

Then using the definitions for Ax 2 and Qx 2 , (10,70), enables \1 to be written as

!.1

~r- 2 (AT Xx:

+ Xx:A-

eT MT Xoc- XxMe)

+ 1-4 (X""B1BfXoo- XxMD2BfXoo)

(10,81)

where

Next we are going tore-express ATX::x: + X 00 A for use in (10,81),

Recall from the development of the HCARE given in the previous chapter that

which we can rewrite using the expression forD In (10,73), as

(10,82)

However, from the definition of elK (10,76), we see that expanding e[KeiK and ef elK
gives

which when substituted in (10,82) yields

Finally we can expand this relation by substituting for K 1 from (1 0. 77) to obtain
(10.83)

where
T

Fcx = eiKeiK-

elT elK-

eiKel

Fx = XocBIDfl elK+ eT~D]]Bf XX+

efKDllDfl elK

Returning to (10.81), we can use (10.83) to rewrite (10.81) as

(10.84)

This completes the expansion of n for use in showing (10.80)

Finally using ( 10. 79) yields
(10.85)

Hoc Separati on Principle

267

and from R003 , (10.70), we have

Roc3

= Roo2- , -2 [FcK-

cT MTXoo- XocMCJ

+ ,-4 [Fx + XooBtDzD]j DzBf Xoo]

(10.86)

where M,FCK and Fx are given in (10.81, 10.83)

Finally we see from (10.86) and (10.84) that (10.80) is satisfied so that (10.69) is

wW.

10.4.2

Relating stabilizing solutions

The relation between Hamilton ian matrices, (10.65), is used now to show that (10.59) is
satisfied and that condition (10.58) is required in order to solve the Hoc output feedback
control problem.
Recall from (10.64) that the Hamilton ian matrix, H 2 , corresponding to the HFARE
satisfies
(10.87)
so that

is the stable subspace of H 2 .

1
Therefore premultiplying (10.87) by <P, inserting <P- <P following H 2 and using (1 0.65)
yields

which indicates that

range [

/ _

-2 X

~'_ y 0000

y 00

is the stable subspace of H 3 , since A 002 - Y00 R 002 is stable.

Now since the range of a matrix is unchanged by post-multiplication by a nonsingular
matrix, we have
2x

range [ I_,- oo
-Yx
provided I-

yx

l [
=range

]
I_
) -I
(
2
-Y00 I-1' XxYxo

,-z X"" Yx is invertible. Therefore we see that

H 00 Output Feedba ck and Control

268

is the stable subspace of H 3 . However we saw in the previous subsection that

is the stable subspace of H 3 . Thus we see that

(10.88)
ence of
and (10.59) is established. Notice in (10.88) that the rightmost equality is a consequ
the
ensure
to
the matrix inversion lemma, (Appendix). We show next that ( 10.58) needed
from
follows
existence of the solution to the H 00 output feedback control problem
(10.88).
we have
Recall that Y:xJ 3 and Yex; are each stabilizing solutions to AREs so that
)
of(l0.88
on
Y 003 , Yoo 2: 0. Thus we see from (10.88) that Z 2': 0. However in the derivati
the
for
required
we required Z to be invertible. Therefore we can conclude that Z > 0 is
n on
existence of a solution to the H 00 output feedback control problem. Now the conditio
Y 00 , X 00 which ensure that Z > 0 is given in the following theorem.
Theorem 10.4 Z > 0 if and only if
(10.89)
where
Z= (1- ' / - 2 Y 00 X 00 )- 1

> 0 if and only if z- 1 > 0. In addition recall that

z- 1 > 0 if and only if Az- > 0 for all Az-1 E .x[z- 1].
Proof Recall from Chapter 4, that Z

Let Az-1 be any eigenvalue of z- 1 Then

(10.90)
and from the dependency of z- 1 on Y xXx. we have
(10.91)

Then solving for Az-' gives

(10.92)

and we see that Az-' > 0 if and only if /\YX <

largest
Finally we see from (10.92) that ,\ 2 > 0 for all .\ 2 E .X[Zi if and only if the

.
theorem
the
in
gi~en
n
conditio
the
satisfies
ooX:xJ],
eigenvalue of Y00 X 00 , Amax[Y

Summary

10.4.3

zn

Dete rmin atio n of L0

ished directly form (10.88).

The relation between L 0 and (L, Z), (10.60), can now be establ
calculated directly from the
As mentioned earlier, this relation enables Y 003 to be
since L depends on Y 00 and Z
stabilizing solutions to the HCA RE and the HFARE
depends on both X 00 and Y00
in (10.20) to be
In order to establish (10.60), recall that Lm was determined
(10.93)
(10.85) yields
Therefore substituting for Y 003 and C0 from (10.59) and

where

However from the definition of Z, (10.88), we have

so that (10.94) becomes

2
1
L 0 = -Z(Yo oCT +B1D I)Dj0 -"Y- ZY00 C[K[I

=ZL -"(- 2 ZYooC [K[l

0]

0]
(10.95)

where L was obtain ed in Section 9.4.3.

10.5

Summary

ARE developed in the previous

In this chapt er we have shown that if the H CARE and HF
these solutions satisfy
chapter have stabilizing solutions, Xxo, Yoo and provided

m. In additi on we have shown

then we can solve the HQO outpu t feedback contro l proble
which solve the Hoo outpu t
H(s),
llers,
contro
that when these conditions are satisfied the
feedback contro l proble m are given by

where

Nc(s) = W 1 (s)<P(s) + W2(s) }

Mc(s) = W3 (s)<P(s) + W4(s)

ci>(s) E 1-{""

270

H 00 Output Feedback and Control

with W(s) having state model

and with K, L being determined from

+ Df CI)

-Dj} (BT Xoo

-(Y XCT + B!DI)Djj

where X00 , Y 00 are stabilizing solution of the HCARE and HFARE, viz.,

A~1X + XAxl- XRoo1X + Qool = 0

with

A - BD}c1Df C 1
BD}/ BT

Ax2 =A- B1DJ D}a1C

R:x2 = CT D"Jol C

Cf (I- DID)c1Df)Cl

Qoo2 = B1 [I- D[ D"J01D2JBf

Aool

Roo I

Qool

and

D;c

D;o

[-lIm, + Dft Du

D~D12]

Df2D11

D12D12

7
-''(lp,

+ D11D1tT
T

D21D11

D11D~
D21D21

Conditions on the generalized plant state model parameters which ensure that the
HCARE and the HF ARE have stabilizing solutions were given at the end of Chapter 9.

10.6

Notes and References

Hx control can be used in conjunction with the small gain theorem (Section 7.4.3) to
provide a means for designing controllers to stabilize plants haYing inexactly specified
models [47, pp. 221~229]. These ideas were extended recently to plants having a class of
nonlinear state models [31]. Additional material on the ideas presented in this chapter can
be found in [18].

Appendix A:
Linear Algebra

Most of the current "first" books on control systems with terms like "automatic control"
or .. feedback control" in their titles have an appendix containing a summary of matrix
theory. These can be consulted for this purpose. The topics treated in this appendix are
selected to give the reader additional insight into issues alluded to in the text.

A.1

Multiple Eigenvalues and Controllability

When some of the roots of the characteristic polynomiaL det[.A/- A] for ann x n matrix
A are repeated, A may not have a complete set ofn independent eigenvectors. Notice that
if
(A.l)
where the .Aks are unique A; # )...i i # j i,j E [I, m], then we say that )...k has multiplicity
or is repeated nk times. Since the degree of the characteristic polynomial, (A. I), is n, the
dimension of the A matrix, and since an nth degree polynomial always has n roots,
counting multiplicities, we have

nk

:\ow there is at least one eigenvector, vk 1, corresponding to each eigenvalue

>..k : k E [1, m]. However since )...k is repeated nk times. we obtain a complete set of
eigenvectors only if there are nk independent eigenvectors, vki : i = 1, 2, nkl corresponding to each >..k, k E [1, m], i.e., only if

has nk independent solutions v = vki, i = 1, 2, nk for each k E [l, rn].

Any square matrix of size n having a complete set of n eigenvectors gives rise to a

272

Appendix A: Linear Algebra

diagonal matrix under the transformati on

where then columns of V are eigenvectors of A, i.e.,

and A is a block diagonal matrix having n1 x n1 diagonal blocks, A 1 : i

).. 1 at each diagonal position, i.e.,

1.0

A,~

1, 2, m, with

r :,0 ;,0]

[>..,

Matrices having a complete set of n eigenvectors are referred to as diagonalizable

matrices. Any matrix whose eigenvalues are simple, i.e .. do not repeat, is a diagonalizable
matrix. Alternatively, a matrix A that does not have a complete set of eigenvectors cannot
1
be diagonalized by any invertible matrix V in the transformati on v- A V. In this case A is
said to be defective, i.e., defective matrices are not diagonalizable matrices. For more
detail the reader is referred to [17, pp. 338-339].
Concerning the controllability of a state model having an A matrix with multiple
eigenvalues, it turns out that when A has more than one eigenvector correspondin g to any
eigenvalue it is not possible to control the state from one input, i.e., the pair (A, B) is
uncontrollable for all B having only one column.
The foregoing fact can be shown as follows. Whenever A has two independent righteigenvectors correspondin g an eigenvalue >.., A has two independent left-eigenvectors,
w 1 , w2 correspondin g to this eigenvalue. Then any linear combination of these lefteigenvectors is also a left-eigenvector of A, i.e.,

where

with the a 1s being arbitrary, constant, non-zero scalars. Therefore we can always choose
the a 1s for any vector B of length n so that
wTB= 0

showing that >.. is an uncontrollable eigenvalue for all vectors B.

A.2

Block Upper Triangular Matrices

Block upper triangular matrices play an important role in many different situations in
control theory. They are directly related to invariant subspaces. A subspace W of Vis
said to be an A-invariant subspace if for all x E W we have Ax E W.

'

rj

r
l

Block Upper Triangular Matrices

273

Suppose we are given a square matrix of size n which has a block upper triangular
structure, e.g.,
(A.2)

where All A 2 , A 3 , and A 4 are n 1 x

any x of the form

nb

n1 x

n2, n2

n~>

and n 2 x

n2

with n

n 1 + n 2 . Then

where x 1 is any vector of length nh lies in an A-invariant subspace, since x E W

In this case W is called an eigenspace corresponding to the eigenvalues of A which are also
eigenvalues of A 1 To see this consider the eigenvalue-eigenvector equation for A. We can
rewrite this equation in the following way
A 1v\

+ A2v~ = >.;v~

(A.3)

= >.;v~

(A.4)

A4v~
where

is the partitioned form for the eigenvector vi corresponding to eigenvalue>.;.

Suppose that v~ is null. Then we see from (A.3) that must be an eigenvector of A 1
corresponding to eigenvalue >.;. Thus the eigenvalues of A1 are eigenvalues of A i.e.
>.(AI) C >.(A) with the corresponding eigenvectors of A being

vi

with A 1v\ =>.;vi

Real Schur decomposition [17, p. 362], provides a computationally robust method for
obtaining an orthogonal matrix Q such that

has the form (A.2). In addition there are ways of doing this so that the eigenvalues of A 1
are a specified subset of the eigenvalues of A.

274

A.3

Appendix A: Linear Algebra

Singular Value Decomposition (SVD)

In this section we give a proof of Theorem 7.1, viz.,

Theor.:m 7.1 Any p x m matrix of complex constants, M, which has rank r can be
decomposed as
Jf

U:EV*

(A.5)

where U, and V are p x p and m x m unitary matrices with

U 1 is p x r

U2 is p

Vl ism

V2 ism

X I'

X jJ- I'

m- r

and 'Eo is an r x r, real, positive definite, diagonal matrix denoted as

O'J

0'2

0
= diag[0" 1 , 0'2 , , 0',.]

'Eo=
0

O'r

with diagonal entries referred to as singular values and ordered so that

i = 1. 2, .. , r

Proof Let (O'T, v;) denote an eigenvalue-eigenvec tor pair forM* Man m x m Hermitian non-negatve matrix. Thus we have
M* Mv; = O'JV;

(A.6)

with O'T real and positive so that we can always order the eigenvalues as
2
O'J

::0.

2
O'r+l

2
0'2

::0. .. ::0. O'r2

2

O'r+2

= =

>0
2
O'm

= 0

Notice that since rank[M] = r, only r of the eigenvalues of M* Mare non-zero.

In addition, since M* l'vf is Hermitian, we can always choose the eigemectors so that
they are orthonormal
=

=0

i = 1, 2, m

(A.7)

1- j

(A.8)

for i

Next we form a matrix V from the foregoing orthonormal eigenvectors of M* }.1 so

.ll

215

that (A.6) can be written as the matrix equations

(A.9)

M*MVI = V 1 I;6

(A.lO)

M*MV2 = 0

where
V = [VI
V1

= [v 1

V2]
v2

,..2
"-'O =

V2

v,]

d"'
Iag [a 21 o Cf 22' ... '

[vr+l

Vr+2

2]

rJ r

...

Vm]

Notice that the orthonormal nature of the eigenvectors, (A.7, A.S), makes them x r
matrix V 1 satisfy Vj V 1 =I, and the m x m- r matrix V 2 satisfy Vi V 2 =I. Thus we
have V* V = VV* =I, i.e., them x m matrix Vis unitary.
)Jext, pre-multiplying (A.9) by Vj and then pre- and post-multiplying the resulting
equation by I; 01 gives

which we can interpret as

(All)
where

Notice from (A.11) that the p x r matrix U 1 has orthonormal columns. Therefore we
can always choose a p x p- r matrix U 2 having orthonormal columns so that the p x p
matrix U

is unitary. i.e., U' U = UU* = I.

We show next that the foregoing unitary matrices. U, V satisfy (A.5) in the statement
of Theorem 7.1.
We begin by expanding U* MV in terms of its blocks
U*MV= [

U*MV
I

Ui"o\t!VI

U*.1MV2]
U2MVc.

(A.12)

Then we see that the 1,1 and 1,2 blocks can be reduced by substituting for U 1 from
(All) and using (A.9, A.10). Doing this yields
I;r) 1 VjM*A'fV 1 = I; 01

UjMV1

U)MV 2

= I; 01 V) M* MV 2 =

I::6 = I;o

I; 0 Vj V2

=0

p;ext the reduction of the 2,1 block in (A.l2) relies on using the

expres~ion

for Uh

-------

---~--u

(A. ll) to ena ble MV 1 to be repl

aced by U,E o so tha t

u;M v, = u;u ,E 0 = 0
Fiually in ord er to reduce the 2,2
block in (A.12) we use the fact tha
t MV 2
seen by not ing , from the con stru
ctio n of V b (A.IO), tha t
M* Mv ; =

= 0. This is

where v; is any colu mn of V . Thi

s implies tha t
2
v;M *M v;= O

and letting z;

= Mv ; we see tha t

which is only possible if z; is null

. Thu s ivfv ; mu st be null and sinc
e v; is any colu mn of V 2
we have MV 2 = 0. The refo re we
see tha t the 2,2 block of (A. l2)
becomes

Thu s from the foregoing we see

tha t (A. l2) is reduced to
U'M V

[~ ~]

and we hav e the form for the SVD

of M which is given in the theo rem
.

A.4

Different Forms for the SVD

The re are occasions where con

stra ints imp ose d on the SVD by
the dimensions of the
mat rix being dec omp ose d can be
tak en into acc oun t so as to pro vid
e a mor e explicit fom1
for the SVD. Thi s possibility aris
es from the fact tha t the ran k of
a rect ang ular mat rix
can not exceed the sma ller of its
dimensions. The refo re since Af
is p x m, we hav e
rank[M]

= r:::;

min (p, m)

and the SVD of M, (A.5), can be

expressed in one of the following
form s dep end ing on
which of the dim ens ion s of M is
larger.
If p < m, i.e., M has mor e colu mns
tha n rows, then (A.5) can be wri
tten as

~
'\;i

where U, V in (A.5) equal UP

and [ Vrl

Vr 2 ] respectively. In add itio

n Vr 1 , Vp 2 are

Matrix Inversion Lemma (MIL)

277

m x p and m x (m- p) with :EP being a p x p diagonal matrix

:EP = diag[a 1 , a 2, ,a, 0, , 0]

In this case it is customary to refer to the p diagonal entries in :EP as the singular values of
M and the p columns of UP and of VP 1 as the left and right singular vectors of M
respectively.
Similarly, if p > m, i.e., M has more rows than columns, then (A.5) can be written as
V m2 ] [

:E.m]
0 V*m

where U, V in (A.5) equal [ U mi U m 2 ] and V m respectively. In addition U 1111 , U m 2 are

p x m and p x (p - m) with L:m being the m x m diagonal matrix

Again it is customary to refer to the m diagonal entries in :Em as the singular values of M
and the m columns of Umi and of V m as the left and right singular vectors of M
respectively.

A.5

Matrix Inversion Lemma (MIL)

The following relation has been of considerable importance in the development of control
theory.
Lemma If D and I: are nonsingular matrices then

where
L = r2
R

+ \li I:<I>

= ~.r~-

o-I\li(<I>o-1\li- I:-Ir'<I>o-1

Proof Expanding the product RL yields

(A.l3)
where
=

o-1\li(<I>o-'\li

+ I;-~~-~<I>

r2=

o-'\li ( <I>rl- 1\li

+ I::- 1) -I <I>o- 1wE <I>

r,

278

Appendix A: Linear Algebra

Then we see that

r1

r2

D~ 1 W(<PD~ 1 w + :E~ 1 )~ 1 [I + <PD~ 1 W:E]<P

= 0 ~1\lf(<PD~1\lf
=

+ I:-1)-1 [I:~ I+ <PD-IW]I:<P

n- 1WI: <I>

Finally substituting (A.14) in (A.13) gives RL =I.

(A.14)

_,_j
~~~~

Appendix 8: Reduced
Or de r M od el Stability
order model is stable, we require the
As mentioned at the end of Section 4.6, since the full
the input -outp ut behavior of the
reduced order model to also be stable. Otherwise
model behavior as the difference
reduced order model canno t approximate the full order
ls diverges in response to a
mode
between the outpu ts of the full and reduced order
a system matrix A 1 which is a
has
l
common boun ded input. Since the reduced order mode
inates,
coord
ced
partition of the full order system matrix Ab, in balan

A has all its eigenvalues in the

we need a criterion which can be applied to ensure that 1
es this prope rty was given, without
open left half plane. A practical criterion which ensur
we are going to prove this theorem.
proof, as Theorem 4.3 in Section 4.6. In this appendix
entry in 'Ebl is the same as any
Theorem 4.3 If in balanced coordinates, no diagonal
diagonal entry in Eb 2 then A 1 will be stable where

Proof Suppose A 1 is not stable. Then we have..\

..\[Ad such that

Re[..\];::: 0

(B.l)

of the theorem, Ab is unstable.

Now we are going to show that under the conditions
only possible if the given system
is
ation
However this is impossible since a balanced realiz
is stable.
ions for the controllability and
To begin, consider expanding the Lyapunov equat

280

Appendix B: Reduced Order Model Stability

observability Gramians as

1 A2] [~bl

[A
A3

A4

Then we see that

~bl

0.

~b2

[~bl

~b2

[Aj.

A)]

A2

A4

=-

[B1]
B [Bj

Bi]

(B.2)

satisfies the following Lyapunov equations

(B.4)
(B.5)
Now since ~bl is diagonal with all diagonal entries being positive and real we have
> 0. Therefore we see from the proofs of Theorems 4.7 and 4.8 that the assumed
unstable eigenvalue, A, of A 1 must lie on the imaginary axis and be an unobservable and
uncontromtble eigenvalue of the pairs (A 1 , C 1) and (A 1, B 1) respectively. Therefore the
corresponding right and left eigenvectors, (B.l ), satisfy
~bl

(B.6)
C1v=O

(B.7)

Next we show that there are many eigenvectors corresponding to this eigenvalue. In
what follows we show how these eigenvectors are related.
To begin, postmultiply (B.5) by the eigenvector v, and use (B.7) to obtain
or

(B.8)

However, A is on the imaginary axis. Therefore -A* = A and we see from (B.8) that
is an additional left eigenvector of A 1 corresponding to the uncontrollable
eigenvalue A and therefore we have
~b 1 v

(B.9)
Next premultiply (B.4) by v* ~b 1 , and use (B.8, B.9) to obtain
\ *"2
AV
uhl = -V *'<'2
ub14*I

or

Thus we see that both v and ~; 1 v are (independent) right eigenvectors of A 1 corresponding to the eigenvalue A.
Continuing in this fashion we can generate sets of left and right eigenvectors
corresponding to A as

s"

= {~ /,J'L'

:i

1. 3, s.... , 2k + 1}

sv = {~hi v : i = 0, 2, 4,

0
,

2k}

(B.IO)
(B.ll)

.l.~x.
,,

'

281

Appendix B: Reduced Order Model Stability

where k is an integer which is chosen large enough so that we can find a matrix Q to satisfy
(B.l2)
where M is the n 1 x (k + I) matrix

Notice from (B.l2) that under the transform ation :Eb 1, the range of M is unchanged. This
fact plays an importan t role in what is to follow.
Now we saw in Theorem 5.5, that (Bl2) implies that every eigenvalue of Q is an
eigenvalue of :Eb1

(B.13)
This importan t fact will be used to show that A 3M= 0. Notice that since the elements in
lSw, Sv are left and right eigenvectors, respectively, of A 1 corresponding to the uncontro
have
we
lable and unobservable eigenvalue.>..,
v*:E~ 1 B 1 = 0

CI:E~Iv

=0

i=l,3, .. ,2k+l

(B.14)

i = 0,2, ... ,2k

(B.15)

~BIB;

(B.16)

~C2,CI

(B.17)

Returning to (B2, B.3) we have

+ :Eb1A:i =
A2,:Ebl + l:b2A3 =

A2l:b2

Then premultiplying (B.l6) by

(B.l4, B.15) gives
"'
2ubiV =
ub2 A*"';
A *"'i+l
2L..bl V

v*:E

hi> and postmultiplying (B.l7) by :Ei 1v

"'i+l V
A )L.Jbl

"'

~L.Jb2

i = 1, 3, ... '2k

"';
A 3uhJV

+1

i = 0,2, ... ,2k

and using

(B.l8)
(B.l9)

Next we can rewrite (B.l8) as

(B.20)

and (Bl9) as
I
"'
2 :Eblv
L..b2 A*[

3
:Eblv

"'2 AM
2k+l v ] = ~ub2
3
:Ebl

(B.21)

where :Eb2 was used to premultiply (B.l9) and M was defined previously in (B.l2).
Now since the left sides of (B.20, B.21) are equal we have
(B.22)

282

Appendix 8: Reduced Order Model Stability

Moreover we can use (B.l2) to rewrite the expression on the left side of this equation.
Doing this gives rise to the following Lyapunov equation
~Q- 2:;~2~

=0

(B.23)

where
Now it turns out if Q and 2:;~ 2 do not have any eigenvalues in common then~= 0 is
the only solution to (B.23). Indeed Q and 2:;~2 do not share any eigenvalues since the
eigenvalues of Q were shown to be a subset of the eigenvalues of 2::~ 1 , (B.l3), and from the
statement of the theorem we have that no eigenvalue of 2:;~ 1 is an eigenvalue of 2::~ 2 . Thus
we have A 3M = 0.
Since the columns of M consist of all the right-eigenvectors of A 1 corresponding to the
imaginary axis eigenvalue,\ of A 1 , we have that (.X, q) is an eigenvalue right-eigenvector
pair for the full system matrix A 6 where
q

[~]

smce

However, since ,\ lies on the imaginary axis, A 6 is unstable. This contradicts the fact
that A 6 is stable. Therefore our assumption at the start of the proof that A 1 is unstable is
false and the proof is complete.

l..

...i

Ap pen dix C:
Pro ble ms

of the
The following problem s illustrate the basic ideas introduc ed in the earlier chapters
ment.
develop
under
are
es
principl
d
advance
more
book. Problem s dealing with

C.1
1.

Problems Relating to Chapter 1

matrix
Use the fact that any matrices M 2 and M 3 have the same eigenvalues as the
M 1 if

ied
for any nonsing ular matrix S, to determi ne relations between the unspecif
of
set
same
the
has
matrix
each
that
so
4,
3,
2,
1,
=
k
elements in the matrices Ak :
akjs.
the
for
values
the
specify
3}
eigenvalues. If this set is { 1, -2,

AI=

A3=

2.

["i'
[!

a12
0

0
0

Tl
""]

A2 =

["'' !]
a21

a31

A,

a23
a33

0
0

a31

a32

matrices
Find the eigenvalues {Aki : i = 1, 2} for each of the following system
for
tors,
Ak : k E ~1, 6;. Then calculate the correspo nding right eigenvec
one.
Ak : k = 1, 2 letting the first compon ent of each eigenvector be

vi

AI= [

A4=

~3 ~1]

[~~

j~]

A2

As=

[-1
O

-2j

[-5
-6 ~]

A3

= [ ~2

A6=

~1]

[:2 !s J

284

3.

4.

Appendix C: Problems

Using Laplace transforms, and without calculating numerical values, determine

which of the transition matrices corresponding to the Aks. k E [1, 6], given in the
previous question have entries dependent on only one mode, e:u.
A certain system has a zero-input state response of

x(t)

[x1(t)]
x 2 (t)

where

5.

6.

7.

Find, without using Laplace transforms, the eigenvalues and eigenvectors of the A
matrix for the system's state model.
Using the trajectories obtained when the initial state is an eigenvector, sketch the
state trajectory for the given initial states x;(O) : i = L 2 and system matrix A. Use
your result to determine a relation between the elements of C, namely, c 1, and c2 , so
that the system output is bounded for all initial states.

Calculate the eigenvalue left-eigenvector pairs,{.\. 11i: i E [L 2]}, for A given in the
previous question. From this result determine an output matrix C such that for any
initial state the zero-input response y(t) depends only on the mode e~ 21 What is the
relation between the C matrix you obtain here and the C matrix you obtained in the
previous question.
Suppose a state model has system matrix, A, which is partially specified as

If the last component of each right-eigenvector is 1, find:

(i) the initial state, x(O), which gives a zero input response of

y(t)

when C = [ -1

4e- 1

(ii) the output matrix C which gives a zero input response of

y(t)
8.

when x(O) = [ 1 ]

4e- 1

-2

Recall that if (t) is a transition matrix then

(i)

0(0) =I

(ii)

O(t) = Ao(r)

If M(t) is a transition matrix

Aflt) = ~]

' ,
find the

et;s

+2

-21

ole
e
ct3et - e-2r

a 2ei a4e'

-,

~e

-c l

+ e-2t J

and the corresponding system matrix A.

functions
Determ ine diagon al form state models for systems having transfe r

9.

GI (

3s- 1
s) = -,s2'+---:3,--s_+_2

Gz(s)

s 2 - 3s+ 2
2
3
2
s + s+

10. If the transfe r functio n for a certain system is

s+4
G(s)

= (s + l)(s + 1 + j)(s +I- j)

system where
find the real parame ters a 1, a2 , a 3 , c1 , c2 , c3 in the state model for this

D=O

that the state goes to

11. Use the eigenvectors of A to find B (other than the null vector) so
the initial state is
and
the origin with time when the input is a unit impulse, u( t) = c5( t)
null. It is known that one of the eigenvalues of A is at + 1 and

, u(t)
12. Ifthe steady-state output is zero when the input is a unit impulse
when
b
and
b
for
3
values
2
(withou t using Laplac e transfo rms) the

-1

4/3]
8/3

-3/2

c = [1
C.2

1]

c5(t), find

D=O

Problems Relating to Chapter 2

a system matrix,
A certain plant is known to have a control ler form state model with
for the plant is
A, which has eigenvalues at {-1, 1,2,3}. Furthe r, this state model
steady state
plant's
known to have direct feedthrough matrix, D, which is zero. If the
x(O) = 0,
and
1
=
output is y( oo) = 5 when the plant input and initial state are u( t)
find the control ler form state model for the plant.
If the state model for
2. Suppose we are given the transfe r functio n G(s) for some plant.
in each case, so that
xi(O),
state,
the plant is in control ler form, determine the initial

1.

286

Appendix C: Problems

the correspondin g output, y;(t), is as specified. The input is zero.

(s+4)(s+5 )

G (s) = . .,.s-+-'-1-)
.(
. .,.s-'. (---!--'-2..,...). .,.s-'---~--3...,-)
.(

3.

Given that .\[A]= {I, 2, -3} determine left-eigenvectors of A, w;, assuming the first
entry in each left-eigenvector is one. Use your result to find a and (3 in B so that x( oo)
is null when u(t) = 8(t) and x(O) = 0

A= [ 07 0l 0]1
-6

4.

0 0

Suppose there is a set of three state models each having the same A matrix but
different B and C matrices as given. Determine which models if any can be
transf2rmed to controller form. Do this without using Laplace transforms.
case(i)

A=[; ~]
5.

6.

~1]

c1 = [1

1J

1]

case(ii)

B2 = [

~]

c2 =

case(iii)

B3 = [

~]

C3 = [ -4

[2

3]

Specifyfore achstatemod el,{(A,B;,C ;): i= 1,2,3},given intheprevio usproblem:

(i) the controllable eigenvalues, (ii) the uncontrollable eigenvalues. In addition.
specify which of these three state models is stabilizable.
Given the state models

CI=[1

7.

B! = [

0 0

l]

find, without using Laplace transforms, a coordinate transformati on matrix T in

each case so that the state models in the new coordinates are each in controller form.
Specify T and the controller form state model, C4, B, C) in each case.
Use your knowledge of the controllable decomposed form and the eigenvalues of
lower triangular matrices to rapidly determine the polynomials
PI (s)

= det[s/- A]

P2(s)

= Cadj[s/-

A]B

Pro'bfiiiiilflfiilatlng 16 Cftapter 3
if the state model paramet ers are
0

-6

-2

-3

-2

A=

-4 6

8.

D=O

-1

C= [3

B=

zeros, i.e.,
Determi ne the transfer function which is free from coincide nt poles and
whose numera tor and denomi nator are coprime .
= 5 for all
Determi ne B so that the steady state output is y(oo) = 10 when u(t)
positive t and

+2
A=

r+1

0 0]

C=[1

-2

C.3
1.

Problems Relating to Chapter 3

Den(s), of
Find the numera tor polynom ial, Num(s), and denomi nator polynom ial,
model
state
g
follmvin
the
to
nding
the transfer function G(s) correspo

-5
A=

-10

-10

-4

OJ

C= [1

B=

D= 1

where

G(s) = Num(s)
Den(s)
matrix but
Suppose there is a set of three state models each having the same A
can be
any
if
models
which
ne
Determi
given.
as
matrices
C
differen t B and

288

Appendix c:

Problems

transformed to observer form. Do this without using Laplace transforms.

A=[: ~]

case(i)

Bl=

case(ii)

B2

case(iii)
3.

4.

[-;1]

c1

[!]

c2 =

B3 = [

~]

c3

=[I
[2

1]
1j

= [ -4 3]

Specify for each state model, {(A, B;, C;) : i = 1, 2, 3}, given in the previous problem:
(i) the observable eigenvalues, (ii) the unobservable eigenvalues. Specify which of
these state models allow the design of an observer to estimate their state .
Suppose a plant state model is in controller form and is 3 dimensional with A having
eigenvalues at 1,2,-3. If the C matrix in this model is partially specified as

-2 q]

C =[I

determine values for q which would make it impossible to design an observer which
could be used to obtain an asymptotic estimate of the plant state.

C.4
1.
2.

Problems Relating to Chapter 4

If every entry in a square matrix is either 1/2 or -1/2 determine the size of the matrix
and the sign pattern of its entries if the matrix is orthogonal.
Find real numbers a, b, c 1 , c2 , and c3 such that M is orthogonal where
a

= [a

a
3.

4.

b
0
-b

c1 ]
c2

c3

M 1 and M 2 are square matrices having eigenvalue- right eigenvector pairs as follows:

forM 1:

{0,1}

and

0,2(.[2-1 (}

for M 2 :

{0, 1}

and

{[1,2]T.[2

1]T}

respectively. Is either of these matrices symmetric? If so which one. Justify your

answer.
For each matrix, specify which of the following properties holds: positive definite,
non-negative, indefinite, i.e., not positive definite. not nonnegative. not negative
definite. (Hint: use the distinguishing properties of the eigenvalues of each type of
matrix)
l'vfl =

[ 1 -2]
[] 0 0]
-2

0
0

2
0

0
3

M2 =

[~

~]
[01 00 02]
2

.
Jl
--~~

5.

Which of the following matrices could be either a controllability or an

Gramia n. Give reasons for your answer in each case.

[1-2]

MI=[l 2]
2

observability

-1

-1

The observability Gramia n for a system is determined to be

= [

2 -4]

-4

the input,
Is there an initial state vector, x(O) other than the null vector, which when
justify. If
If"no"
zero?
always
is
which
y(t),
u(i), is always zero, produces an output,
has this
which
vector,
null
the
than
other
"yes", give an example, i.e., give an x(O),
property.
ed such that
7. If, in the previous problem, the initial state vector is restrict
e for the
possibl
value
largest
the
find
zero),
xT(O)x(O) = 1, (and u(t) is always
ion.
restrict
this
under
y(t)
signal
output
the
integral from 0 to oo of the square of
value.
this
e
produc
would
which
Specify x(O) (subject to the restriction)
te the observability
8. Determine, in each case, if the pairs (A, C) are observable. Calcula
n in each case.
equatio
ov
Lyapun
riate
Gramia n manually by solving the approp
definite. Relate
positive
is
n
Gramia
onding
Determine, in each case, if the corresp
bility in each
observa
the
on
s
finding
your
to
your findings on the positive definiteness
case

9.

case(i):

A=[-10 -22]

C =[I

I]

case(ii):

A=[-10 -22]

C= [0

1]

that the r;s are

Determine the observability Gramia n for the given system assuming
real.

c=

[0

1]

D=O

r;s make the

What values of the r;s make the Gramia n singular? What values of the
.
explain
so
if
related?
system unobservable? Are these values

Appendix C: Problems

290

10. A certain plant has state model matrices (A, B) and observability Gramian W 0 given
as

[31 21]

Given that D = 0 and recalling that the integral defining W o does not exist when A is
unstable, find the transfer function of the plant using the Lyapunov equation relating
A, C, W 0 to find a and C.
11. Suppose that a certain system is known to have maximum output energy equal to 2
0. Suppose the
for all possible initial states satisfying x~(O) + x~ (0) = I, \vhen u(t)
this system
that
suppose
addition
In
~].
[~,
=
0)
(
xT
is
this
achieves
initial state that
satisfies
also
which
state
has minimum output energy equal to I for some initial
xt(O) + x~(O) = 1 when u(t) = 0. Find the system's observability Gramian and the
initial state that gives this minimum output energy.

C.5
I.

Problems Relating to Chapter 5

An observer based controller (Fig. 5.1) generates a feedback signal,f(t), from

f(t)

Hz(t)

i(0 =Fz(0 +G 1v(0+G2y(0

where

2.

If the plant output equals the first component of the plant state \ector, find: (i) the
closed loop transfer function from v(t) to y(t), (ii) the characteristic polynomial for
the observer.
If in the previous problem the plant state model is in controller form and the plant
output is the sum of the first three components of the plant state, determine the closed
loop transfer function from v to y for the observer based feedback control system
where
4

3.

-4

-13

13

If in the observer based feedback control system shown in the Figure 5.1 we are given
the transfer functions from v(t) to y(t), Gvy(s), and from y(t) to f(t), Gyr(s), where

f(t) is the feedb ack signal, deter mine the contr oller
when
GV\(s)
.

~-= -3~~-2
+ 2s + 1

+ 3s

state mode l matri ces (Q, R, S, T)

-2
' 3s-? + 2 s + 3
Gyr(s
) = s,

with plant mode l param eters given as

and contr oller equa tions

i(t) = Qz(t)

+ Rv(t) + Sy(t)

f(t) = T::(t)

Appe ndix D:
MATL AB Exper iment s

0.1

State Models and State Response

The intention of this lab is to introduce some of the ways MATLAB can be used in
connection with state models and to provide further experience in thinking in terms of the
state space.

D.1.1

Controller form

Given the transfer function model for some system, the command

[a, b, c, d]

= tf2ss(num. den)

can be invoked to find an equivalent state model, (use the help facility), where num and
den are entered as row vectors containing coefficients of the numerator and denominator
of the transfer function. Try the following examples
1

Gl (s) = s2
G2 (s ) =

+ 3s + 2

s-7

+ 2s +-i

---co-----

3s2

+ 2s + 1

Check your result by using the command "ss2tf', (use the help facility), to convert
back to the transfer function models.

D.1.2

Second order linear behavior

Recall that the state at any time is represented by a point in ann-dimension al space and
the path traced out by this point is referred to as a state trajectory. Unlike cases where the
system order is greater that 2 , we can readily display the trajectory for a second order
system as a graph with the two components of the state forming the axes of the graph.

~~-"'-""'""

294

~- '-_

...,.,-.

Appendix D: MATLAB Experiments

Consider the system whose behavior is governed by the following second order linear
differential equation

where

Then the controller state model is

= [ - 2~wn

-;~ ]

C= [0 w~]

= [ ~]

D=O

Now we are going to investigate the zero-input state response by obtaining the state
trajectories for one value of natural frequency, say wn = 1 and different values of the
damping ratio, say

( = 0.3, I, 10,-0.3
each for some initial state, say xT (0) = [2 0]
One way to do this is to invoke the command, (use the help facility),
initial(A, B, C, D, xO)

Then from the resulting graph of the output vs. time that is obtained on the screen, you
can decide over what interval of time you will need the state.
Suppose you decide you need to know the state in the time interval [0. 5] seconds every
0.5 seconds. We can create a vector, t, of these times by incorporating this data in a vector
as start:increment: end. This is done here by typing
t =

0: 0.5: 5

and

[y, x, t]

initial(A, B. C, D, xO)

This produces a column y, followed by two columns. one for each component of the
state. The entries in these columns give the values of the output, and components of the
state at 0.5 s intervals from 0 to 5 s. Now you can obtain the graph of the state trajectory
by entering

plot(xi:.l),x(:. 2))
where x(:, i) is all rows in the ith column of x. Your graph can be labelled using the

295

same line
procedu re (on line help) where you enter each line (there are four) on the
separate d by commas .
Explain why these plots are markedl y different for different signs on(.

D.1.3

Secon d order nonlin ear behav ior

nonlinea r
Consider the system whose behavio r is governe d by the second order
,
equation
Pol
der
van
a
as
to
differential equation , referred
/

l(t)- (1 -/(t))y (ll(t)

+ y(t) = 0

manner
Then if \Ve assign compon ets of the state to derivates of the output in the same
of the
ents
compon
the
of
r
behavio
the
case,
linear
the
in
form
er
as done to get the controll
equation s
state are seen to be governe d by the following first order nonline ar differential

x1 (t) = (1- x~(t))x 1 (t)- x 2 (t)

x2(t) = x 1 (t)
ned using
Now the state trajecto ry obtained for a specific initial state can be determi
and type
editor
an
open
file
this
create
MATLA B by creating an m-file called vdpol.m. To

functionyp
yp

= vdpol(t, q);
= [(1- q(2) * q(2)) * q(l)- q(2); q(l)];

solvers, say
Save the file and invoke one of the ordinary (vector) differential equatio n
ODE23

[t, q] = ode23('vdpol', [tO,({], qO)

as a column
where tO is the start time, if is the finish time and qO is the initial state, entered
vector by typing

of q.
The sample values of the compon ents of the state are returned as the columns
states
initial
the
of
each
for
state
the
getting
try
Use tO= 0, if= 25 and

x(O)

[0.2]
0

and

[~]

Obtain a graph of the state trajecto ry in each case by typing

plot(q(:, I). q(:. 2))

on as
Notice that q( :, 1) is all the elements in the first column of q produce d by the simulati
state.
the
of
ent
compon
first
the values of the

296

Appendix D: MATLAB Experiments

In the second case where xT (0) = [2 0] it is instructive to concentrate on the latter part
of the trajectory, say from time step 40 to the end of the simulation. To find out what this
end valt:e is type

size(q)
Then if this end value is 200 type

plot(q(40: 200, l),q(40: 200,2))

to obtain the trajectory from time 40 to time 200.
Generalize the nature of the trajectories in terms of their dependency on the initial
state. Compare the state trajectories you obtain for the present nonlinear system with
those you obtained for the linear system in the previous section.

D.1.4

Diagonal form

So far we have only considered the controller form state model. There are several other
important canonical forms for the state model of a linear dynamic system. One of these is
the diagonal or normal form. This form is characterized by having its system matrix A
which is diagonal. This simplifies solving the state differential equation since the
derivative of each component of the state is dependent on itself only.
Recall that if [A, B, C, D] is a state model for a given system so is [Q- 1AQ, Q- 1B, CQ,
D] for any nonsingular matrix Q of appropriate size. where Q is a coordinate transformation matrix. Also recall that if the eigenvalues of A are distinct, we can change coordinates
so that A in the new coordinates is diagonal. This is done by letting Q have columns equal
to the eigenvectors of A. We are going to use these facts.
Use MATLAB to determine the parameters, (A. B, C, D), of the controller form state
model for the system having transfer function

G(s)=

s -s+20
s 4 + 5s 3 + 5s 2 - Ss - 6

Use the command 'roots" to decide if all the roots of the denominator are poles of
G(s). Calculate the eigenvalues and eigenvectors of A by typing

[Q, E]

eig(A)

and the characteristic polynomial of A by typing

poly( A)
How do the poles of G(s) compare with the eigenvalues of A, i.e .. diagonal entries in E?
Should each of these sets be the same and if not why not?
Use Q to transform the controller form state model to a diagonal form state model.

..j. )
~/'

Feed back and Cont rolla bility

297

This can be done by typin g

*A *Q
inv(Q ) * B

aa = inv(Q )

bb =

cc = CA

*Q

dd= D

the poles of G(s)?

How do the diago nal entries of aa comp are with
mine the residues, poles and any
deter
to
due"
"resi
Finally, inYoke the comm and
r). This is done by typin g
const ant term D (nonz ero only if G(s) is prope
[res,poles, d]

= residue(num, den)

s in bb and cc?
How do the residues of G(s) relate to the entrie

0.2

Feedback and Controllability

with the tasks described in what follows,

In order for you to proce ed as rapid ly as possible
us matrices you will be enter ing in the
you shou ld plan a subsc riptin g scheme for the vario
computer.

D.2.1

Controllable state mod els

Enter the state mode l

5

-6

-3
0

3
2

a44

b2
b3
b4

-6

hs

2
0

A=

c=

[1

B=

D=O

1]

s
with the following assignment to the variable entrie

bi = 1 : i = 2, 3, 4, 5

and

function using the comm and "ss2z p".

Obta in the poles and the zeros of the trans fer
that part of the state space is superfluous
Are there any pole-zero cancellations indicating
to the input -outp ut beha vior?
comm and "ctrb " and check its rank
Next obtai n the contr ollab ility matri x using the
using the comm and "rank ".
of A. Then find the state feedback gain
Using the comm and "eig" find the eigenvalues
s to -1. Try doing this with both the
matrix K which shifts the unsta ble eigenvalue
Check using the comm and "eig" to see
comm and '"place'' and the comm and "ack er".
m matrix A + BK are those you specified.
that the eigem alues of the state feedback syste

298

Appendi x D: MA TLAB Experim ents

Obtain the poles and zeros of the state feedback system using the comman d "ss2zp". Are
there any pole-zero cancellations? What effect does the state feedback have on the
transfer function zeros? Do other experiments if need be so that you can give a general
reasons for your answers to the foregoing questions.
Obtain the feedback matrix which produces a closed loop transfer function which has
no finite zeros (they are all cancelled) and poles which are all in the open left half-plane.
Does the rank of the controllability matrix change. Explain your answer by using
eigenvectors to identify uncontro llable eigenvalues.

D.2.2

Uncontrollable state model s

In this section we are going to examine the use of state feedback when the given state
model for the plant is not controllable. One way to obtain an uncontro llable system to
enable this study, is to form the input matrix B so that it is linearly dependen t on a subset
of the eigenvectors of A. We do this as follows.
Invoke "eig" to get the eigenvectors of A. These will be displayed in a square matrix
you specify in the argumen t for "eig". Suppose this matrix is V and Dis used as the
diagonal .matrix displaying the eigenvalues. Create a new input matrix which we call Bl
by combining a subset of the columns of V. i.e., some of the eigenvectors of A. Suppose
we do this by combinin g all the eigenvectors correspo nding to the unstable eigenvalues of
A with all the weights used to combine these eigenvectors equal to 1. This can
be done
using the MATLA B subscripting rules (note the use of:).
Obtain the controllability matrix ("'ctrb") for (A, B1) as well as the rank of this
controllability matrix. Obtain the transfer function for the state model (A, Bl, C, D) and
comment , with an explanation, on any pole-zero cancellations. What is the order of the
system indicated by the transfer function and does the transfer function indicate that this
system is stable?
To gain more insight into the nature of uncontro llable state models. invoke "ctrbf' to
transform coordina tes so that the state model is in controlla ble decompo sed form,
(Section 2.5). In this case we see that the block matrix consisting of the last two rows and
columns of A and the last two elements of Bin these coordina tes has special significance.
If(Aban Bhar) are the system and input matrices in the controlla ble decompo sed form then
we can extract the subsystem correspo nding to these rows and columns by typing
AC = AhaA4: 5,4: 5)
BC = Bhar(4: 5, 1)
CC = Cbm(1,4: 5)

Then calculate the transfer function using the command 'ss2zp" for the system
(AC, BC, CC) and compare it with the transfer function you obtained earlier for
the
system in the original coordinates. Explain what you see.
Next can you find a state feedback matrix Kin these coordina tes so that the resulting
system is internally stable?
Repeat the foregoing by making up a new B matrix from a linear combina tion of the
eigenvectors correspo nding to the stable eigenvalues of A. Notice that the new AC matrix
will be a 3 x 3.

Obse rver Based Control Systems

299

Controllability and Repeated Eigenvalues

s are uncontrollable for any

Recall that we showed in Appendix A that single-input system
ector associated with
eigenv
one
than
input matrix B when the system matrix A has more
s.
follow
as
d
procee
we
fact
any repeated eigenvalue. To investigate this
A. Then use
typing
by
A
recall
can
We
lab.
this
of
Assume A is as given at the beginning
eigenvectors of A. Are the
the comm and "eig" to determine the eigenvalues and
typing
eigenvalues distinct? Next change the 4,4 element in A by
A(4,4 ) = A(4,4 )

+3

matrix. Are the eigenvalues

Find the eigenvectors and eigenvalues of the new A
input matrix B so that the
an
ruct
Const
?
distinct? Is there a complete set of eigenvectors
new (A, B) pair is controllable. Explain.

0.3

Observer Based Control Systems

the plant state from measureRecall that.a n observer is used to provide an estimate of
will examine the design and
ments of the plant' s input and outpu t. In this lab 'YOU
estimate of the plant state in
behavior of observers and the effect of using an observer's
l system.
place of the actual plant state in a state feedback contro
the tasks described in what
In order for you to proceed as rapidly as possible with
e you will use in connection
follows, read the lab throug h and plan the subscripting schem
computer. This is especially
with the various matrices you will need to enter in the
you will need to make a block
impor tant in connection with the second section where
t numb ered for use with the
diagram of the contro l system with each input and outpu
commands "blkb uild" and "conn ect".
Observer Estimates of the Plant State

[To Be Done Prior to Using Computer]

Suppose you are given the unstable second order plant

for the plant and

with a 1 = 0 and a2 = -1. Write the observer form state model
observer (Section 3.3)

design an

i(t) = Fz(t) + c[u(t )]

y(t)
where
F=A +LC

G= [B+L D

-L]

with F having eigenvalues at -1, -2.

TLAB that the observer state
Now you will attem pt to verify experimentally using MA
alues of the observer and
eigenv
and plant state appro ach each other with time. Use the

300

Appendi x D: MATLAB Experime nts

plant state model to determine a suitable length of time over which to let the simulatio
n
operate.
[To Be Done Using Compute r]
Begin by entering the observer form plant state model matrices: A. B, C, D. Then use
the comman d "acker" or "place" to determine the observer matrix L so that the observer
eigenvalues are at -1 ,-2. This is done by typing

where

pis a column vector of the desired observer eigenvalue locations. Construc t F and
G and check the eigenvalues ofF using the comman d "eig". Now we are going
to check

this observer's ability to track the compone nts of the state of the plant by generating both
the plant and observer states when the plant input is a unit step.
Invoke the comman d "step" to obtain the plant output, y and state x. Try using a time
interval of ten seconds in steps of 0.1 s. In order to do this you will need to create a row
vector t having elements equal to the computa tion times. This can be done by typing
t=O:O. l: 10

You will need t in the last section of this lab.

Next use the comman d "1sim" to obtain the state of the observer when the plant input
is a step andy is the plant output in response to a step. This can be done by typing
[yl,zl]

= !sim(F,G ,CI>DB, UB,t)

where
UB = [uu,y]

DB=[O;O]

and uu is a column vector of 1's of the same length as y withy being a column vector of
samples of the plant output you just compute d. The input sequence. uu. could be
generated by typing
uu

ones(IOL 1)

Once you have z l, the state of the observer, (a matrix having two columns of samples
of the compone nts of the observer's state) you can compare the observer state, zl. just
obtained , with the plant state x by generating the error
err=(z l-x)

However recall that since the plant is in observer form, the plant output is the first
compone nt of the plant state. Therefore we need only compare the second compone nts of
the plant and observer states in order to characterize the performa nce of the observer as
a
state estimator . In addition in attemptin g to commen t on this performa nce, you might
find it helpful to examine the relative error, i.e.,
z1 (:. 2)- x(:, 2)/x(:, 2)

Obser ver Based Contro l System s

301

type
For instanc e if we want to compu te this for time 10 we would
[zl(lO l, 2)- x(IOI, 2)]/x(I OI, 2)
the plant output is null.
Now the initial state of the plant in the foregoing genera tion of
plant. Theref ore to test the
However in practic e we do not know the initial state of the
state of the plant and
observer's ability to estimate the plant state when the initial
state was xT(O) = [1, 1].
observer are not equal, suppos e you though t the initial plant
ption, use the comm and
Then setting the initial state of the observ er equal to your assum
"lsim" with the observer, i.e., type
zO

= [1, 1]

[y2, z2] = lsim(F , G, C, DB, UB, t, zO)

tion time and obtain a
Next obtain the history of the estima tion error over the simula
n by typing
divisio
t
elemen
by
t
elemen
plot of this history. This can be done using
rele

= (z2- x).jx

ts in the first row of the

Notice that since the actual initial plant state is null. the elemen
(not a numbe r) or oo as its
matrix xs are zero, so that the first listing in "rele" has NaN
state, e.g., set the initial
first entry. Try other guesses for the "unkn own" initial plant
ng relative error in the
observer state to zO = [100, 100] and obtain a plot of the resulti
estimate of the plant state.

D.3.1

Obse rver base d cont rolle rs

[To Be Done Prior to Using Compu ter]

section to implem ent a
Suppose you want to use the observer studied in the previous
study the interac tion which
state feedback contro ller. In this section you will begin to
connec ted via a feedback
takes place between the plant and the observ er when they are
matrix K in a closed loop configuration.
were used the closed loop
Choos e the feedback matrix K so that if exact state feedback
model in observer form
system would have eigenvalues at -3, -4. Do this for the plant state
which you used in previo us sections.
the feedback system, i.e.,
Next referring to Section 5.2, obtain the system equati ons for
q for the closed loop system
let the plant input u be given as u = Kz + v. The state vector
has four entries, viz.,

state for the observer you

where xis the state for the plant (in observer form) and:: is the
designed in the previous section.
[To Be Done Using Compu ter]
the plant state model
Use the comm and "acker " to find the feedback matri:-; K so that
the observer based
denote
Then
-4.
-3,
(in observer form) has its eigenvalues shifted to

302

Appendix D: MA TLAB Experiments

controller state model as

where F,G, were determined in the previous section and K was just determined.
Now you are going to use the commands "blkbuild" and .. connect" to form the
observer based feedback control system by interconnecting the various subsystems
involved. This is done in two steps. First enter each subsystem, either as transfer
functions, (nidi) or as state models (ai. bi, ci, di). In each case i should identify a different
subsystem and should be the integers starting with one. In the present case there are three
subsystems: the input-controller (i = 1), the plant (i = 2) and the observer-controller
(i = 3). The input-controller is trivial and is done by entering its transfer function as
nl = 1 and dl = 1. Then the plant and observer-controller can be entered in terms of
their state models as [a2,b2,c2,d2] and [a3,b3,c3,d3] respectively.
Once you have entered all the subsystems, type
nblocks = 3
blkbuild

A message appears "state model [a. b, c, d] of the block diagram has 4 inputs and 3
outputs". You are now ready to proceed to the second step which involves specifying how
to interconnect the subsystems you have just identified. This is done through the use of a
connection matrix, Q.
The connection matrix, Q, specifies how the subsystems inputs and outputs are to be
connected to form the observer based control system. Notice that the outputs r 1, y2, and
y3 are the outputs from the input-controller. the plant and the observer-controller
respectively. The inputs ul, u2, u3 and u4 are the input to the input-controller, the plant
and the two inputs to the observer. It is important to keep the ordering of the inputs to the
observer consistent with the ordering used in the previous section in forming the state
differential equation for the observer.
Draw a block diagram showing how all inputs and outputs relate to one another.
The connection matrix, Q, has rows corresponding to subsystem inputs with the integer in
the first position of each row identifying the subsystem input. Subsequent entries in a
given row are signed integers which identify how the subsystem outputs are to be
combined to make up the subsystem input. Zero entries are used when a subsystem
output does not connect to a subsystem input. The sign associated with each integer entry
in Q indicates the sign on the corresponding output in the linear combination of
subsystem outputs making up a given input. Since there are three inputs which depend
on the subsystem outputs, Q will have three rows.
Using your block diagram, enter the matrix Q.
Once Q is entered, the inputs and outputs of the feedback control system must be
declared by typing
inputs= 1
outputs = [2]

rol
er, i = 1, is the obse rver base d feed back cont
where here the inpu t to the inpu t-con troll
rol
cont
t, i = 2, is the outp ut from the feedback
system inpu t and the outp ut from the plan
system.
ectio n info rmat ion in the Q matr ix to form
At this stage you are read y to use the conn
m
m from the mult i-inp ut mult i-ou tput syste
the observer base d feedback cont rol syste
the
i.e.,
de],
cc,
be,
[ac,
m
The conn ected syste
[a, b, c, d) you form ed using "blk buil d".
obta ined by typing
is
m,
syste
rol
cont
observer based feedback

ts, outputs)
[ac, be, cc, de] = eonneet(a, b, c, d, Q, inpu
obta ined in the penc il-an d-pa per work you
Check your result with the state mod el you
did before start ing this com putin g session.
model. calculate the poles and zeros of the
Once you have the corr ect closed-loop state
side
m by using 'ss2 zp'' (don 't use k on the left
transfer func tion for the closed-loop syste
s
matrix). Are there any com mon pole and
if you have alrea dy used it for the feedback
of
trans fer function with the trans fer function
zeros indicated? If so, why? Com pare this
uss
Disc
state feedback (wit hout the observer).
the closed loop system assuming exact
ms.
syste
rol
cont
d
base
rYer
obse
of
theo ry
your obse rvati ons by relating them to the
el by
mod
state
loop
d
close
the
of
ility
rvab
Finally, check the cont rolla bilit y and obse
and com putin g rank s by using "ran k".
using the com man ds"c trb" and "obs v"

0.3 .2

avi or
Ob ser ver bas ed con tro l system beh

(To Be Don e Prio r To Using Com pute r]

for
on, and assuming the state s are available
For the plan t intro duce d in the first secti
state
with
m
syste
inpu t for the closed loop
feedback, calcu late the outp ut for a step
l
has poles at -3, -4. Do this for the initia
m
syste
feedback such that the closed loop
or
asef
achc
tine
earo ugh sket chof theo utpu
conditionxO = [O,O]andxO = [10, lO]. Mak
using MAT LAB .
in
com paris on with the results you will obta
[To Be Don e Usin g Com pute r]
d
ut of the state mod el of the obse rver base
Using "lsim " obta in the state and outp
d"
buil
"blk
g
usin
prev ious section as a resu lt of
cont rol system that you obta ined in the
ofO. l
Do this over a ten second inter val in steps
followed by 'con nect ", i.e., [ac, be, ec, de].
of
part
first
the
uu, which you created for use in
s (use the t vect or and step inpu t vector,
when
10]
0,
[1
the plan t state is x(O) = [0, 0] and
this lab). Obta in a plot of the outp ut when
the observer state is zO = [0, OJ in each case.

0.4

State Mo del Reduction

estimation we saw that cont rolla bilit y and

In previous labs on state feedback and state
n
to (i) assign feedback system poles (ii) assig
observability play a key role in being able
the
tion model of a system has orde r less than
observer poles. Recall that the trans fer func
ntro lstate model is unobservable and/ or unco
dimension of the state model when the
ost"
vior of a given state model which is ''alm
lable. Ther efor e the inpu t-ou tput beha
state
a
ld be able to be appr oxim ated using
unco ntro llabl e and; or unobservable shou
dimension of the given state model. This
model whose dime nsio n is less than the
to do
in Cha pter 4. In this lab we are going
possibility was intro duce d theoretically

some simu lation studies to furth er investigate

the effectiveness of this model order
reduc tion technique.

D.4.1

Dec omp osit ion of unc ont roll able and /or
uno bse rvab le systems

Recall that we can always trans form the coord inate

s so that a given state model that is not
contr ollab le is trans form ed to contr ollab le deco
mpos ed form
Ac= [ Acl
Ac3

cc =

[eel

0]

Ac4

Be=

[:J

c,2J

where (Ac2, Bc2 ) is contr ollab le (done using "ctrb

f"). Alternatively if the given model is
unobservable we can chan ge coord inate s so the
state mode l becomes

where (A 04 , C02 ) is obser vable (done using "obsv

f"')
Ther efore when a given state mode l is both unco
ntrol lable and unob serva ble we can
use the foregoing deco mpos itions to obtai n a state
mode l which has a state space which is
divided into four subspaces. In the following
list of these subspaces C, C indicates
contr ollab le and unco ntrol lable , respectively with
a similar notat ion for observable and
unobservable. Thus the state space splits up into
1 a subsp ace CO of dimension n 1
2 a subspace CO of dimension n2
3 a subsp ace CO of dimension n 3
4 a subsp ace CO of dimension n4
Notice that if the dime nsion of the given state
model is n then

An example demo nstra ting the use of MA TLA

B to do this deco mpos ition can be
carrie d out as follows. Ente r the system state mode
l matrices

[
AI "
Cl= [O

I
:

-2

-3

0
0

!.]
1l

Iii

[J'

Dl = 0

State Model Reduction

305

and determine the corresponding transfer function by entering

[zl,pl,kl] = ss2zp(Al,Bl, Cl,Dl)

f_

What is the order of the system? What can be said about the controllability and
observability of the state model? We can use the eigenvector tests for controllability and
observability to answer this question by interpreting the result of making the following
entries
[Vl,DRIJ = eig(AI)
Cl *VI

and
[WI,DLl] = eig(AI')
Bl' *WI

As a further check on your results, obtain the rank of the controllability and
observability matrices using the commands "'ctrb". "obsY" and 'rank". Use your results
to specify basis vectors for each subspace in the decomposition of the state space.

0.4.2

Weak controllabi lity and/or observabil ity

Enter the following system state model matrices

A2~

-1

-2
0
0

0
-3
0

I]

.01

C2= [ 0

D2 = 0

and repeat the preyious section for this state model. Contrast the results you obtain now
with those you obtained in the previous section.
Next obtain the controllability and observability Gramians, (fVc, W 0 ), for the system
you just entered. Do this by typing

QO = C2'

* C2

QC

= B2*B2'

WO

= lyap(A2'. QO)

we= lyap(A2, QC)

(Use "'help" for information on the command 'lyap".l Obtain the eigenvalues of the
product of the Gramians. Record these values for future reference.
Next use the command "balreal" to obtain a balanced realization by typing
[A2b, B2b, C2b,g2, t2]

= balreal(A2, B2, C2)

306

Appendix D: MATLAB Experiments

Determine the eigenvalues of the product of the Gramians you just entered and
compare them with the entries in g2. Comment on any correspondences.
Next obtain a reduced order model by discarding parts of the balanced realization
correspondin g to elements of g2 that are relatively small. For example suppose
g2(3) < < g2(2). Then since the elements of g2 are arranged in descending size we can
obtain a reduced order model, (A2r, B2r, C2r) by typing
A2r

= A2b(l : 2, I : 2)

B2r = B2b(l : 2)
C2r = C2b(l : 2)

Use use the command "ss2zp" to determine the poles and zeros of the transfer function
for the reduced order system. Is there any obvious relation between the original system
transfer function and the reduced order system transfer function?

0.4.3

Energy interpret ation of the controlla bility and

observab ility Gramian s

Recall, from Chapter 4, that the sum of the diagonal entries in the controllabili ty
Gramian equals the energy transferred into the state over all positive time, from the
input when the input is a unit impulse at time zero. We can check to see if this is the case
here.
Recall from Chapter 1 that the zero state response to a unit impulse is the same as the
zero input response to an initial state x(O) =B. Therefore we can use "initial" with the
initial state set to B. As before we need to decide on the time interval and sample times for
doing this. Suppose we use an interval of 10 s with samples taken every 0.1 s. Then we
create a vector of sample times by typing
ts=O:O.l: 10
Notice that ts has 10 I elements.
Therefore using the command "initial" with B2 in the position reserved for the initial
state, we can obtain the resulting state by typing
[Y2, X2, ts] = initial(A2, B2, C2, D2, ts)

where X2 has four columns with each column correspondin g to a different component of
the state and each row correspondin g to a sample time.
Recall that the energy transferred to the state is given by

where

State Model Reduct ion

307

fact we can
is the energy transfe rred to the i1h compo nent of the state. Using this
an approx imation , EAc, to Ec by enterin g

compu te

eei=X 2(1: 10l,i/* X2(1: lOl,i)

with i = 1, 2, 3, 4 in succession and then enter
1

EAc

-L:ee i
10 i=l

by typing
Compa re your result with the trace of the control lability Gramia n

ETc= trace(WC)
output from an
Recall, from Chapte r 4, that the energy, E 0 , which is transfe rred to the
initial state can be determ ined using the observability Gramia n as

te E01 when the

Now we can use the proced ure just described for compu ting Ec to compu
Y2. Compa re your
initial conditi on x(O) = BI by using the data stored previously in
result with the result obtaine d by using the observability Gramia n
(B2)'

D.4.4

* WO * B2

Design of reduced orde r models

ch to model order
In this part of the lab you will examine the balance d realization approa
filter. This
analog
orth
reduction. The system to be reduced is a degree eight Butterw
of degree
inator
denom
a
and
system has a transfe r function with constan t numera tor
origin. A
the
from
e
distanc
eight. The poles are located in the open left half-plane at unit
list of the poles can be obtaine d by enterin g

[zbw,pbw, kbw] = buttap(8)

ing the
We can check that the poles are equidis tant from the origin by examin
elements of the matrix obtaine d by enterin g

diagon al

pbw * pbw'
is obtaine d by first
A balance d realiza tion (state model) for the Butterw orth filter
to balance d form.
it
rming
transfo
then
and
calculating a control ler form state model
Therefo re enter
[A3, B3, C3, D3] = zp2ss(z bw,pbw , kbw)
and
[A3b, B3b, Cb3,g3 , t3]

= balreal(A3, B3, C3)

308

Appendix D: MA TLAB Experiments

Recall that the reduced order model is obtained by discarding the latter part of the
state vector in balanced coordinates so that the corresponding discarded part of g3 is
negligible in the sum of the elements in g3. Notice that there is a significant decrease in the
size of the elements in g3 in going from the fifth to the sixth element. Therefore a reduced
order model obtained by discarding the last three elements in the state should give a good
approximation to the full order system. The state model for the reduced order approximation can therefore be obtained by entering
A3b5 = A3b(l : 5.1 : 5)
B3b5 = B3b(l : 5)
C3b5

= C3b( 1 : 5)

Notice that this reduced order state model approximation is also balance. This can be
seen by generating the controllability and observability Gramians by typing
QObS

= C3b5' * C3b5

QCbS = B3b5

* B3b5'

Wob5

= lyap(A3b5'. QOb5i

WebS

= lyap(A3b5. QCbS)

The final part of this lab concerns the performance of the reduced order model. One
possible test would be to compare the step responses of the reduced and full order models.
[y3b, x3b] = step(A3b, B3b, C3b, D3, 1, ts)
[y3b5, x3b5]

= step(A3b5, B3b5, C3b5, D3, I, rs)

Note any marked differences in the plots on the screen and proceed to compare the
error between the reduced order model output and the full order model output. This can
be done using element by element division, i.e.,
rete =(y3b5- y3b).jy3b
plot( rs, rele)

Repeat the foregoing by deleting more than the last three elements of the state to
obtain a reduced order model and use the step response to compare the performance of
this approximation with the performance of the fifth order approximation just studied.

References

[I]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[II]
[12]

Anderson, B. D. 0. and Moore, J. R Optimal Control: Linear Quadratic Methods. PrenticeHall, Englewood Cliffs, New Jersey, 1990.
Anderson, B. D. 0. and Liu, Y. "Controller reduction: Concepts and Approaches", IEEE
Trans. Automatic Control, TAC-34, pp. 802-812, 1989.
Bittanti, S., Laub, A. J. and Willems, J. C. (Eds.) The Riccati Equation. Springer-Verl ag,
Berlin, 1991.
Blackman, P. F. Introduction to State-variable Analysis. The Macmillan Press, London, 1977.
Brockett, R. W. Finite Dimensional Linear Systems, John Wiley and Sons. New York, 1970.
Brogan, W. L. Modern Control Theory. 3rded., Prentice-Hall , Englewood Cliffs, N.J., 1991.
Brown, R. G. Introduction to Random Signal Analysis and Kalman Filtering. John Wiley and
Sons, Chichester, 1983.
Chandrasekha ran, P. C. Robust Control of Dynamic Systems. Academic Press, London,
1996.
Chen, C. T. Introduction to Linear Systems Theor. Holt Rinhart and Winston, New York,
1970.
Dorato, P., Abdallah, C. and Cerone, V. Linear Quadratic Control, An Introduction.
Prentice-Hall , Englewood Cliffs, N.J., 1995.
Doyle, J. C., Francis, B. A. and Tannenbaum, A. R. Feedback Control Theory. Maxwell
Macmillan Canada, Toronto, 1992.
Doyle, J. C., Glover, K., Khargonekar, P. P. and Francis, B. A. "State space solutions to
standard H 2 and Rx: control problems", IEEE Trans. Automatic Control, AC-34, pp. 831-

847, 1989.
Fairman, F. W., Danylchuk, G. J., Louie, J. and Zarowski, C. J. "A state-space approach to
discrete-time spectral factorization" , IEEE Transactions on Circuits and Systems-ll Analog
and Digital Signal Processing, CAS-39, pp. 161-170, 1992.
[14] Francis, B. A. A Course in Hx Control Theory, Springer-Verl ag, Berlin, 1987.
[15] Furuta, K., Sana, S. and Atherton, D. State Variable Methods in Automatic Control, John
Wiley and Sons, Chichester, 1988.
[16] Glover, K. and Doyle, J. C. "A state space approach to H"" optimal control", Lecture Notes
in Control and Information Science, 135, pp. 179-218, 1989.
[17] Golub, G. H. and Van Loan, C. F. Matrix Computations, 2nd ed., The Johns Hopkins
University Press, Baltimore, Maryland, 1989.
[18] Green, M. and Limebeer, D. J. N. Linear Robust Control. Prentice--Hall, Englewood Cliffs.
New Jersey, 1995.

[13]

310

[19]
[20]
[21]
[22]

[23]
[24]
[25]
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
[34]
[35]
[36]
[37]
[38]
[39]
[40]
[41]
[42]
[43]
[44]

References

Green, M., Glover, K., Lime beer, D. J. N. and Doyle, J. "A J spectral factorization approach
to Hoc control", SIAM Journal on Control and Optimi::.ation, 28, pp. 1350-1371, 1990.
Green, M. "Hx controller synthesis by J lossless coprime factorization", SIAM Journal on
Control and Optimization, 30, pp. 522-547, 1992.
Isidori, A. "Hx control via measurement feedback for affine nonlinear systems", Int. J.
Robust and Nonlinear Control, 4, pp. 553-574, 1994.
Jonckheere, E. A. and Silverman, L. M. '"A new set of imariants for linear systemsapplications to reduced order compensator design", IEEE Trans. Automatic Control, AC28, pp. 953-964, 1983.
Kailath, T. Linear Systems, Prentice-Hall, Englewood Cliffs, l\. J .. 1980.
Kimura, H. Chain-Scattering Approach to Hx Control, Birkhausser Boston, 1997.
Kwakernaak, H. and Sivan, R. Linear Optimal Control Systems, Wiley Interscience, New
York, N.Y .. 1972.
LePage, W. R. Complex Variables and the Laplace Tran.sformj!Jr Engineers, Dover, 1980.
Moore, B. C. "Principal component analysis in linear systems: controllability, observability,
and model reduction". IEEE Trans. Automatic Control. AC-26. pp. 17-32, 1981.
Missaghie, M. M. and Fairman, F. W. "Sensitivity reducing obsen ers for optimal feedback
cont.!;_ol", IEEE Transactions on Automatic Control, AC-22, pp. 952-957, 1977.
Mullis, C. T. and Roberts, R. A. "Synthesis of minimum roundoff noise in fixed point digital
filters", IEEE Trans. Circuits and Systems. CAS-23. 551-562, 1976.
Mustafa, D. and Glover, K. "Controller reduction by Hac balanced truncation". IEEE
Trans. Automatic Control, AC-36, pp. 668-682, 1991.
Pavel, L. and Fairman, F. W. "Robust stabilization of nonlinear plants-an L 2 approach",
Int. J. Robust and Nonlinear Control, 6, pp. 691-726. 1996.
Pavel, L. and Fairman, F. W. "Controller reduction for nonlinear plants-an L 2 approach",
Int. J. Robust and Nonlinear Control, 7, pp. 475-505, 1997.
Pavel, L. and Fairman, F. W. (1998). "Nonlinear H" control: a J-dissipatiYe approach",
IEEE Trans. Automatic Control, AC-42, pp. 1636-1653, 1997.
Pernebo, L. and Silverman, L. M. "Model reduction via balanced state space representation", IEEE Trans. Automatic Control, AC-27, pp. 382-387, 1982.
Petersen, I. R. "Disturbance attenuation and Hex- optimization: a design method based on the
algebraic Riccati equation", IEEE Trans, Automatic Control, AC-32, pp. 427-429, 1987.
Rudin, W. Principles of Mathematical Analysis, 3rd ed., McGraw-Hill Book Company, New
York, 1976.
Rudin, W. Real and Complex Analysis, 3rd ed., McGraw-Hill Book Company, Ne\\ York,
1987.
Scherpen, J. M. A. "Hx balancing for nonlinear systems'', Int. J. Rohust and Nonlinear
Control, 6, pp. 645-668. 1996.
Strang, G. Linear Algebra and irs Applications, 2nd ed., Academic Press, Ne\\ York, 1980.
Sveinsson, J. R. and Fairman, F. W. "Simplifying basic Hx problems", Proc. Conference on
Decision and Control, 33,2257-2258, 1994.
Tsai, M. C. and Postlethwaite. I. ''On 1 lossless coprime factorization and Hx control", Int.
J. Robust and Nonlinear Control, 1, pp. 47-68, 1991.
van der Shaft, A. J. "L 2 gain analysis of nonlinear systems and nonlinear state feedback Hoc
control", IEEE Trans. Automaric Control, AC-37, pp. 770-784. 1992.
Varga, A. "A multishift Hessenbcrg method for pole assignment of single-input system,;",
IEEE Trans. Automatic. Control, AC-41, pp. 1795-1799, 1996.
Vidyasagar, M. Control System Synthesis: A Factori::.ation Approach, The MIT Press,
Cambridge. Massachusetts, 1985.

Referen ces
[45]
[-+6]
[47]

311

onal, EngelVidyasag ar, M. Nonlinear Systems Analysis, 2nd ed., Prentice -Hall Internati
wood Cliffs, 1993.
Cambrid ge, 1988.
Young. N. An Introduction To Hilbert Space, Cambrid ge Universit y Press,
Prentice -Hall,
Control,
Optimal
and
Zhou, K. with Doyle. J. C. and Glover, K. Robust
Upper Saddle River, New Jersey, 1996.

Ind ex

Ackerm ann formul a 52, 64, 74

adjoint system 221
adjugate matrix 11, 18
algebraic Riccati equatio n (ARE) 115, 119,
1-!5, !66, 224, 246
GCAR E 153
GFAR E 157
HCAR E 236
HFAR E 244
QCAR E 126
QFAR E 142
system 221
pass
all
anticausal, antistab le systems 172
asymptotic state estimat ion 69
average power 147
balanced realization I04
Bezout identity 204, 209
causality constra int 31
causal system, signal 172
Cayle:- Hamilt on theorem 52
characteristic polyno mial 9
co inner functio n 22 I
compan ion matrix 9
comput er determ ination of state 37, 66
contrac tion 162
controllability 34, 55
Gramia n 101, 112
matrix 5 L 57, 63
control lable
decomp osed form 60, 93, 118
eigenvalue 4-t
subspace 56

control ler
form 47
Hoc-central, parame trized 255
LQG 157
quadra tic 120
convol ution integral 29
coordin ation transfo rmation 12, 28
coprim e factoriz ation 201, 204,
doubly 212
J-inner 230
state models 206
covariance 147
decomp osition of a space 169, 178
detectable system 72
direct sum 169
disturb ance 115, 120, 138, 147, 154, 225, 232,
242,24 8
dom(R ic) 145
doubly coprime factoriz ation 212
dual system 72
eigenvalue 17, 96
assignment 43, 64, 74, 82, 86
controllable, observa ble 44, 72
invariance 20
eigenvector 1'7. 96
left-eigenvector 19
right-eigenvector 16
tests (controllability, observability) 43, 71
energy 94, 102. 110, 167
feedback 41
filtering 68

314

Index

filtering (contd.)
H.00 242
LQG (Kalman) 90, 155, 166
quadratic 138
Fourier transform 173
Gaussian random vector 147
Gramian
controllability 10 I, 112
observability 94, 109
Hamiltonian matrices 130, 158
Hx relation 262
Hankel norm Ill
Hardy spaces
H2, H.~ 177
H,X)) H.~ 183
Hermitian 111atrix 95
Hilbert space 167
induced norm 181
initial condition response
(see zero-input response)
inner function 221
inner product space 169
input-output response
(see zero-state response)
invariant subspace 133
inverse
left matrix 259
right matrix 83
system 197
isomorphic, isomorphism 178, 229
Laplace transform 176
Lebesgue spaces
time domain: ..CE(-x,oo), ")()(-x,x)
167, 174
frequency domain: ", Lx: 174
L 2 gain 181
linear combination 3, 18
lossless system 219
LQG control
requirement for solution 165
state estimation !53
state feedback 149
Lyapunov equation in
balanced realization l 07
Hex, control 234

LQG control 151, 155

quadratic control 123-129, 140
matrix
Hermitian 95
nonnegative 96
positive definite 98
matrix exponential 9
matrix fraction description
(see coprime factorization)
matrix square root 160
measurement noise 138
minimal realization 34
reduction to 91
nonnegative matrix 96
norms
2-norm for vectors 94
H 2 signal norm 167
H 2 system norm 172, I 79
H 00 system norm 181, 191
Hankel norm 112
induced 2-norm for matrices 186
L2 induced system norm 181
L 2 time domain signal norm 95, 167
L 00 function norm 182
null space of a matrix 78
null vector. matrix 17. 36
observability 34, 76
Gramian 94, I 09
matrix 73
observable
decomposed form 82, 94
eigenvalue 72
subspace 78
observer 70
form 72
minimal order 82
observer-based controllers 116
orthogonal 21
compliment 169, 178
matrix 98
spaces, signals 169
output feedback 149. 157. 254
output injection 90
Parsevars theorem 174
PBH (Popov-Belevich-Hautus) tests

Index
(see eigenvector tests)
performance index
Gaussian 148, 153

H 00

224, 233, 242-252

quadratic 121, 139

phase plane 5
pole placement
(see eigenvalue assignment)
pole-zero cancellation 34
positive definite matrix 98
projector 161
proper 12
quadratic control 115
requirement for solution
state estimation 13 7
state feedback 119

147

range of a matrix (operator) 57, 77, 163

rank of a matrix 51, 59, 76
rational function 12
realization
balanced 104
minimal 34,35
Ric(H) 145
robustness
(see stability)
separation principle 118, 149, 157, 261
singular value decomposition 185, 274
small gain theorem 184, 191
spaces
Hardy 179, 185
Hilbert 168, 173
Lebesgue 167
stability
internal, external 25
reduced order model 279
robustness 26, 69, 203
stabilizable system 43
stabilizing controllers 213, 215

315

stabilizing solution 128, 135, 158, 227, 267

stable system 6, 109
state
computation 37, 66
estimation 67, 153, 242, 251
feedback 41, 120, 151, 227
state model forms
controllable decomposed 60
controller 48
diagonal (normal) 32
observable decomposed 78
observer 72
state trajectory 5, 21
strictly proper 12
symmetric matrix
(see Hermitian)
system factorization 201
system interconnections 193
system inverse 196
system zero 198
trace of a matrix 102, 150
transfer function 34
proper, strictly proper 12
transformatio n to
controllable decomposed form 64
controller form 49
observable decomposed form ,81
observer form 73
transition matrix 9, 31
unitary matrix

98

weighting matrix 170

well posed feedback control 214
worst disturbance 234, 259
Y oula parametrization

217, 260

zero-input (initial condition) response

1, 25,

32

zero-state (input output) response

I, 32