1

Corrections: ROBUST and OPTIMAL CONTROL

Convention: Page xx is denoted as Pxx. Line xxx from the top of the page is
denoted as Lxxx while line xx from the bottom of the page is denoted as Lxx.
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
16.
17.

P5, -L9, \be covered in an one or two" should be \be covered in a one or two"
P7, -L2, \Note that" should be \Note while"
P37, L2, C1 = and B1 = should be C = and B =.
P75, -L6, \decreasingly order numbers" should be \decreasingly ordered numbers"
P80, L8, \maximally possible" should be \maximum possible"
P89, L3, \Then for any nonzero vector u0 2 C m the output . . . " should be \Then
for any nonzero vector u0 2 C m such that G(z0 )u0 = 0 the output . . . "
P114, -L9, \Hence the 1 of .." should be \Hence the 1 norm of .."
(s+2)
s 1
s 1
P119, L11, u = (s+2)
3 (r n d) 3 d should be u = 3 (r n d) + 3 d.
P124, Corollary 5.6, add the following at the end of corollary: \or equivalently,
det(I P (s)K^ (s)) has no zeros in the closed right half plane."
P124, -L6, -L5, and -L3, \number of open rhp poles" should be "number of (closed)
rhp poles"
P124, change Theorem 5.7 (ii) to \(I P (s)K^ (s)) 1 is stable." This correction
does not aect any other results in the book.
P125, Theorem 5.8, a footnote on the Nyquist plot: the Nyquist contour is constructed so that all imaginary axis poles of P and K are counted as unstable
ones.
P149, L11, \Theorem 6.3 holds" should be \Theorem 6.3 hold"
P152, -L7, \applies to multivariable system" should be \applies to multivariable
systems"
P157, L5, \A11 is not asymptotically" should be \A11 is not asymptotically stable"
P159, -L6, \using the denition of B (s)" should be \using the denition of B~ (s)"
P162, Example can be simplied.
The above bound can be tight for some systems. For example, consider an n-th
order transfer function
n
X
G(s) = s +bi a ;
i
i=1

2
with ai > 0 and bi > 0. Then kG(s)k1 = G(0) =
following state space realization

Pn

i=1 bi =ai and G(s) has the

pb 3
2 a
1
pb1 7
66
a2
2 7
6
.
.
..
.. 77
G = 66
pb 75
4p p
a
n
n
p
b1 b2


bn 0

and the controllability and observability Gramians of the realization are given by

"p #
P = Q = a +bi baj :
i
j

It is easy to see that i = i (P ) = i (Q) and

n
X
i=1

i =

n
X
i=1

i (P ) = trace(P ) =

n b
X
i = 1 G(0) = 1 kGk :
2
a
2 1
i=1 i 2

In particular, let ai = bi = 2i , then P = Q ! 12 In , i.e., j ! 12 as  ! 1.

18. P218, -L7, delete \and kM k1 =  (M (j!0 )) = 1 ."

19. P222, -L17, \det(I + W1
W2 KS0) = (I + S0 W1
W2 K )" should be \det(I +
W1
W2 KS0 ) = det(I + S0 W1
W2 K )"

" should be \:=" for consistence.
20. P248, L5, \=
21. P248, -L8, u =
u y2 should be u2 =
u y2 .
22. P258, L2, \N =

p 

2I " should be \N = p I
2I
I
2I

pI

p 
2I
I "

23. P280, L13, \in simple cases" should be \in these simple cases"
24. P363, L7, L8, L12, B1 X + D1 C1 should be B1 X1 + D1 C1 .

25. P457 -L8, \ = 0" should be \Q = 0"

26. P488, -L1, [W1 PW2 1] should be [W1 KW2 1]
27. P489, Figure 18.5, W2 and W2 1 should be swapped.
28. P490, equation (18.7) should be (K1 K11 + I )  2 (K1 + Ps)M~ s M~ s (K11 + Ps ):
29. P492, Theorem 8.10,  (K1 ) should be \(K1 )", two times.

3
30. P494, L10,

 (M~ s ) = (Ms ) =

should be

 (M~ s ) = (Ms ) =




1=2

1=2

1 + 2 (W2 PW1 )
1 + 2 (W2 PW1 )

31. P494, L8 and L9

2 (M~ s ) = max (M~ s M~ s ) = 1 +  1 (P P  ) = 1 + 12 (P )

max s s
s
should be
and
should be

 2 (M~ s ) = max (M~ s M~ s ) = 1 +  1 (P P  ) = 1 + 12 (P )

min s s
s
2
2 (N~s ) = 1 2 (M~ s ) = 1 + (P2 (sP) ) :
s

2
2 (N~s ) = 1 2 (M~ s ) = 1 + (P2 (sP) ) :
s

32. P587, reference no. 214, add \pp. 5-16" at the end.

ROBUST AND OPTIMAL

CONTROL

ROBUST AND OPTIMAL

CONTROL

KEMIN ZHOU
with
JOHN C. DOYLE and KEITH GLOVER

PRENTICE HALL, Englewood Clis, New Jersey 07632

1.2.

How to Use This Book

loop shaping is described where the desired loop shape is specied together with a
requirement of robust stability, and has been found to be an intuitively appealing and
robust procedure with close connections to stabilization in the gap metric. Secondly
the method of  analysis (the structured singular value) is introduced. This approach
due to Doyle [1981] gives an eective analysis tool for assessing robust performance in
the presence of structured uncertainty. Note that the H1 norm alone can only give a
conservative prediction of robust performance. The synthesis of controllers that satisfy
such a criterion (the -synthesis problem), can be approached iteratively with the H1
synthesis for a scaled system as an intermediate step.
Finally it is interesting to consider the question of why the induced norm in L2 has
been used. Is it just for the mathematical convenience of L2 being a Hilbert space?
Apart from the relative simplicity of the solution the main advantage is probably its
easy interpretation in terms of the familiar frequency response considerations. However roughly in parallel with the development of the H1 theory has been the work on
L1 induced norms of Pearson and coworkers (see the book by Dahleh [1995]) where
analogous results on robustness and performance can be made.
1.2

How to Use This Book

This book is intended to be used either as a graduate textbook or as a reference for

control engineers. With the second objective in mind, we have tried to balance the
broadness and the depth of the material covered in the book. In particular, some
chapters have been written suciently self-contained so that one may jump to those
special topics without going through all the preceding chapters, for example, Chapter
13 on algebraic Riccati equations. Some other topics may only require some basic linear
system theory, for instance, many readers may nd that it is not dicult to go directly
to Chapters 9  11. In some cases, we have tried to collect some most frequently
used formulas and results in one place for the convenience of reference although they
may not have any direct connection with the main results presented in the book. For
example, readers may nd those matrix formulas collected in Chapter 2 on linear algebra
convenient in their research. On the other hand, if the book is used as a textbook, it
may be advisable to skip those topics like Chapter 2 on the regular lectures and leave
them for students to read. It is obvious that only some selected topics in this book can
be covered in an one or two semester course. The specic choice of the topics depends
on the time allotted for the course and the preference of the instructor. The diagram
in Figure 1.2 shows roughly the relations among the chapters and should give the users
some idea for the selection of the topics. For example, the diagram shows that the only
prerequisite for Chapters 7 and 8 is Section 3.9 of Chapter 3 and, therefore, these two
chapters alone may be used as a short course on model reductions. Similarly, one only
needs the knowledge of Sections 13.2 and 13.6 of Chapter 13 to understand Chapter 14.
Hence one may only cover those related sections of Chapter 13 if time is the factor. The
book is separated roughly into the following subgroups:

INTRODUCTION









Basic Linear System Theory: Chapters 2  3.

Stability and Performance: Chapters 4  6.
Model Reduction: Chapters 7  8.
Robustness: Chapters 9  11.
H2 and H1 Control: Chapters 12  19.
Lagrange Method: Chapter 20.
Discrete Time Systems: Chapter 21.
In view of the above classication, one possible choice for an one-semester course
on robust control (analysis) would cover Chapters 4  5; 9  11 or 4  11 and an
one-semester advanced course on H2 and H1 control would cover (parts of) Chapters
12  19. Another possible choice for an one-semester course on H1 control may include
Chapter 4, parts of Chapter 5 (5:1  5:3; 5:5; 5:7), Chapter 10, Chapter 12 (except Section 12.6), parts of Chapter 13 (13:2; 13:4; 13:6), Chapter 15 and Chapter 16. Although
Chapters 7  8 are very much independent of other topics and can, in principle, be studied at any stage with the background of Section 3.9, they may serve as an introduction
to sources of model uncertainties and hence to robustness problems.
Robust Control H1 Control
Advanced Model & Controller
(Analysis)
H1 Control
Reductions
4
4
12
3.9
5
5.15.3,5.5,5.7 13.2,13.4,13.6
7
6*
10
14
8
7*
12
15
5.4,5.7
8*
13.2,13.4,13.6
16
10.1
9
14*
17*
16.1,16.2
10
15
18*
17.1
11
16
19*
19
Table 1.1: Possible choices for an one-semester course (* chapters may be omitted)
Table 1.1 lists several possible choices of topics for an one-semester course. A course
on model and controller reductions may only include the concept of H1 control and
the H1 controller formulas with the detailed proofs omitted as suggested in the above
table.
1.3

Highlights of The Book

The key results in each chapter are highlighted below. Note that some of the statements in this section are not precise, they are true under certain assumptions that are

1.3.





































 

Highlights of The Book

3:9

11



A
U
A

10
?

13

12
?

15
?

@
R
@


@
I
@

19

13:4

16
17

14

13:2
13:6

18

20

21

Figure 1.2: Relations among the chapters

INTRODUCTION

not explicitly stated. Readers should consult the corresponding chapters for the exact
statements and conditions.
Chapter 2 reviews some basic linear algebra facts and treats a special class of matrix
dilation problems. In particular, we show


X B

min
X C A

= max

C A

B
A

 

and characterize all optimal (suboptimal) X .

Chapter 3 reviews some system theoretical concepts: controllability, observability,
stabilizability, detectability, pole placement, observer theory, system poles and zeros,
and state space realizations. Particularly, the balanced state space realizations are
studied in some detail. We show that
for agiven stable transfer function G(s) there

A
B such that the controllability Gramian P
is a state space realization G(s) = C D
and the observability Gramian Q dened below are equal and diagonal: P = Q =  =
diag(1 ; 2 ; : : : ; n ) where
AP + PA + BB  = 0
A Q + QA + C  C = 0:
Chapter 4 denes several norms for signals and introduces the H2 spaces and the H1
spaces. The input/output gains of a stable linear system under various input signals
are characterized. We show that H2 and H1 norms come out naturally as measures of
the worst possible performance for many classes of input signals. For example, let


G(s) = CA B0 2 RH1 ; g(t) = CeAt B

Z 1
n
X
k
g

u
k
2
Then kGk1 = sup kuk and 1  kGk1 
kg(t)k dt  2 i . Some state
0
2
i=1
space methods of computing real rational H2 and H1 transfer matrix norms are also

presented:
and
where

kGk22 = trace(B  QB ) = trace(CPC  )

kGk1 = maxf : H has an eigenvalue on the imaginary axisg
H=

A BB  = 2 :
C C
A

Chapter 5 introduces the feedback structure and discusses its stability and performance properties.

1.3.

Highlights of The Book

w1

e1

+6

P
K^

++ w2
e2 ?
e

We show that the above closed-loop system is internally stable if and only if


I
P

K^
I


^ ) 1 K^ (I P K^ ) 1 
= P(I(I KP
^ ) 1 (I P K^ ) 1 2 RH1 :
KP

Alternative characterizations of internal stability using coprime factorizations are also

presented.
Chapter 6 introduces some multivariable versions of the Bode's sensitivity integral
relations and Poisson integral formula. The sensitivity integral relations are used to
study the design limitations imposed by bandwidth constraint and the open-loop unstable poles, while the Poisson integral formula is used to study the design constraints
imposed by the non-minimum phase zeros. For example, let S (s) be a sensitivity function, and let pi be the right half plane poles of the open-loop system and i be the
corresponding pole directions. Then we show that
1

ln (S (j!))d!  max

(Repi )i i :

This inequality shows that the design limitations in multivariable systems are dependent
on the directionality properties of the sensitivity function as well as those of the poles
(and zeros), in addition to the dependence upon pole (and zero) locations which is
known in single-input single-output systems.
Chapter 7 considers the problem of reducing the order of a linear multivariable
dynamical system using the balanced truncation method. Suppose
2

G(s) = 4

A11 A12
A21 A22
C1 C2

B1
B2
D

3
5

2 RH1

is a balanced realization with controllability and observability Gramians P = Q =  =

diag(1 ; 2 )
1 = diag(1 Is1 ; 2 Is2 ; : : : ; r Isr )
2 = diag(r+1 Isr+1 ; r+2 Isr+2 ; : : : ; N IsN ):

10

INTRODUCTION

Then the truncated system Gr (s) = AC11 BD1

1
error bound:

kG(s) Gr (s)k1  2

is stable and satises an additive

N
X
i=r+1

i :

On the other hand, if G 1 2 RH1 , and P and Q satisfy

PA + AP + BB  = 0
Q(A BD 1 C ) + (A BD 1 C ) Q + C  (D 1 ) D 1 C = 0
such that P = Q = diag(1 ; 2 ) with G partitioned compatibly as above, then


B
A
1
11
Gr (s) = C D
1
is stable and minimum phase, and satises respectively the following relative and multiplicative error bounds:

N
Y

G 1 (G Gr ) 1 

i=r+1
N 
Y

Gr 1 (G Gr ) 1 

i=r+1

1 + 2i ( 1 + i2 + i )

1 + 2i ( 1 + i2 + i )

1
1:

Chapter 8 deals with the optimal Hankel norm approximation and its applications
in L1 norm model reduction. We show that for a given G(s) of McMillan degree n
there is a G^ (s) of McMillan degree r < n such that

G(s) G^ (s) H = inf G(s) G^ (s) H = r+1 :

Moreover, there is a constant matrix D0 such that

G(s) G^ (s) D0 
1

N
X
i=r+1

i :

The well-known Nehari's theorem is also shown:

inf kG Qk1 = kGkH = 1 :

Q2RH1

Chapter 9 derives robust stability tests for systems under various modeling assumptions through the use of a small gain theorem. In particular, we show that an uncertain

1.3.

11

Highlights of The Book

system described below with an unstructured uncertainty
2 RH1 with k
k1 < 1 is
robustly stable if and only if the transfer function from w to z has H1 norm no greater
than 1.

nominal system

Chapter 10 introduces the linear fractional transformation (LFT). We show that

many control problems can be formulated and treated in the LFT framework. In particular, we show that every analysis problem can be put in an LFT form with some
structured
(s) and some interconnection matrix M (s) and every synthesis problem
can be put in an LFT form with a generalized plant G(s) and a controller K (s) to be
designed.

-

z

z



-K




w
u

Chapter 11 considers robust stability and performance for systems with multiple
sources of uncertainties. We show that an uncertain system is robustly stable for all

i 2 RH1 with k
i k1 < 1 if and only if the structured singular value () of the
corresponding interconnection model is no greater than 1.
1

nominal system

Chapter 12 characterizes all controllers that stabilize a given dynamical system G(s)
using the state space approach. The construction of the controller parameterization is

12

INTRODUCTION

done via separation theory and a sequence of special problems, which are so-called full
information (FI) problems, disturbance feedforward (DF) problems, full control (FC)
problems and output estimation (OE). The relations among these special problems are
established.
dual
- FC
FI 

equivalent

equivalent

DF
For a given generalized plant

dual

OE


A B1 B2
G
(
s
)
G
(
s
)
11
12
G(s) = G (s) G (s) = 4 C1 D11 D12
21
22
C2 D21 D22


3
5

we show that all stabilizing controllers can be parameterized as the transfer matrix from
y to u below where F and L are such that A + LC2 and A + B2 F are stable.

z

? ?
c

C2 

u1




D22 
R

-A

 
66
c

B2 


6
c

-F
y1

Chapter 13 studies the Algebraic Riccati Equation (ARE) and the related problems: the properties of its solutions, the methods to obtain the solutions, and some

1.3.

13

Highlights of The Book

applications. In particular, we study in detail the so-called stabilizing solution and its
applications in matrix factorizations. A solution to the following ARE

A X + XA + XRX + Q = 0
is said to be a stabilizing solution if A + RX is stable. Now let


H := AQ RA
and let X (H ) be the stable H invariant subspace and


X
1
X (H ) = Im X2
where X1 ; X2 2 C nn . If X1 is nonsingular, then X := X2 X1 1 is uniquely determined
by H , denoted by X = Ric(H ).
A key result of this chapter is the relationship between the spectral factorization of a
transfer function and the solution of a corresponding ARE. Suppose (A; B ) is stabilizable
and suppose either A has no eigenvalues on j!-axis or P is sign denite (i.e., P  0 or
P  0) and (P; A) has no unobservable modes on the j!-axis. Dene


(s) = B  ( sI

A ) 1 I

P S
S R



(sI A) 1 B :
I

Then

(j!) > 0 for all 0  !  1

() 9 a stabilizing solution X to

(A BR 1 S  ) X + X (A BR 1 S  ) XBR 1B  X + P SR 1S  = 0

() the Hamiltonian matrix

H=

has no j!-axis eigenvalues.

Similarly,

() 9 a solution X to

A BR 1 S 
(P SR 1S  )

BR 1B 
(A BR 1 S  )

(j!)  0 for all 0  !  1

(A BR 1 S  ) X + X (A BR 1 S  ) XBR 1B  X + P SR 1S  = 0

such that (A BR 1 S  BR 1 B  X )  C .

Furthermore, there exists a M 2 Rp such that

 = M RM:

14

INTRODUCTION

with

M=

A B ; F = R 1(S  + B  X ):
F I

Chapter 14 treats the optimal control of linear time-invariant systems with quadratic
performance criteria, i.e., LQR and H2 problems. We consider a dynamical system
described by an LFT with
3

A B1 B2
G(s) = 4 C1 0 D12 5 :
C2 D21 0

z
y
Dene

w


u



A
0
B2
 C1 B2 
D



12
C1 C1 A
C1 D12

 

 

A
0
C
2
D21 B1 C2
J2 := B B  A

B1D21
1 1
X2 := Ric(H2)  0; Y2 := Ric(J2 )  0
 C1 ); L2 := (Y2 C  + B1 D ):
F2 := (B2 X2 + D12
2
21
Then the H2 optimal controller, i.e. the controller that minimizes kTzw k2 , is given by


Kopt (s) := A + B2 FF2 + L2C2 0L2 :
2

H2 :=

Chapter 15 solves a max-min problem, i.e., a full information (or state feedback)

H1 control problem, which is the key to the H1 theory considered in the next chapter.
Consider a dynamical system
x_ = Ax + B1 w + B2 u 
 C1 D12  =  0 I  :
z = C1 x + D12 u; D12
Then we show that sup umin
kz k2 < if and only if H1 2 dom(Ric) and X1 =
w2BL2+ 2L2+
Ric(H1 )  0 where

H1 :=

A
2B1 B1 B2 B2

C1 C1
A

1.3.

15

Highlights of The Book

Furthermore, u = F1 x with F1 := B2 X1 is an optimal control.

Chapter 16 considers a simplied H1 control problem with the generalized plant

G(s) as given in Chapter 14. We show that there exists an admissible controller such
that kTzw k1 < i the following three conditions hold:
(i) H1 2 dom(Ric) and X1 := Ric(H1 )  0;
(ii) J1 2 dom(Ric) and Y1 := Ric(J1 )  0 where
J1 :=

A 2 C1 C1 C2 C2 :
B1 B1
A

(iii) (X1 Y1 ) < 2 .

Moreover, the set of all admissible controllers such that kTzw k1 < equals the set of
all transfer matrices from y to u in

u

y
M1 

M1 (s) = 4

A^1
F1
C2

Z1 L1 Z1 B2
0
I
I
0

3
5

where Q 2 RH1 , kQk1 < and

A^1 := A + 2B1 B1 X1 + B2 F1 + Z1 L1 C2

F1 := B2 X1 ;

L1 := Y1 C2 ;

Z1 := (I 2Y1 X1 ) 1 :

Chapter 17 considers again the standard H1 control problem but with some assumptions in the last chapter relaxed. We indicate how the assumptions can be relaxed
to accommodate other more complicated problems such as singular control problems.
We also consider the integral control in the H2 and H1 theory and show how the
general H1 solution can be used to solve the H1 ltering problem. The conventional
Youla parameterization approach to the H2 and H1 problems is also outlined. Finally,
the general state feedback H1 control problem and its relations with full information
control are discussed.
Chapter 18 rst solves a gap metric minimization problem. Let P = M~ 1 N~ be a
normalized left coprime factorization. Then we show that
K




stabilizing

inf

K (I + PK )
I

1

I P

16

INTRODUCTION

K

= K stabilizing
inf
(I + PK ) 1M~
I

q

N~ M~

 2

This implies that there is a robustly stabilizing controller for

P
= (M~ +
~ M ) 1 (N~ +
~ N )
with
if and only if

~ N
~ M

 1

1<

N~ M~

 2

H:

Using this stabilization result, a loop shaping design technique is proposed. The proposed technique uses only the basic concept of loop shaping methods and then a robust
stabilization controller for the normalized coprime factor perturbed system is used to
construct the nal controller.
Chapter 19 considers the design of reduced order controllers by means of controller
reduction. Special attention is paid to the controller reduction methods that preserve
the closed-loop stability and performance. Methods are presented that give sucient
conditions in terms of frequency weighted model reduction problems for which some
numerical methods are suggested.
Chapter 20 brie y introduces the Lagrange multiplier method for the design of xed
order controllers.
Chapter 21 discusses discrete time Riccati equations and some of their applications in
discrete time control. Finally, the discrete time balanced model reduction is considered.

Contents
Preface

xiii

Notation and Symbols

xv

List of Acronyms

xix

1 Introduction

1.1
1.2
1.3

Historical Perspective : : : : : : : : : : : : : : : : : : : : : : : : : :
How to Use This Book : : : : : : : : : : : : : : : : : : : : : : : : :
Highlights of The Book : : : : : : : : : : : : : : : : : : : : : : : : :

2 Linear Algebra
2.1
2.2
2.3
2.4
2.5
2.6
2.7
2.8
2.9
2.10
2.11
2.12

Linear Subspaces : : : : : : : : : : : : :
Eigenvalues and Eigenvectors : : : : : :
Matrix Inversion Formulas : : : : : : : :
Matrix Calculus : : : : : : : : : : : : : :
Kronecker Product and Kronecker Sum :
Invariant Subspaces : : : : : : : : : : : :
Vector Norms and Matrix Norms : : : :
Singular Value Decomposition : : : : : :
Generalized Inverses : : : : : : : : : : :
Semidenite Matrices : : : : : : : : : : :
Matrix Dilation Problems* : : : : : : : :
Notes and References : : : : : : : : : : :

3 Linear Dynamical Systems

3.1

:
:
:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:

Descriptions of Linear Dynamical Systems : : : : : : : : : : : : : :

vii

1
5
6

17

17
20
22
24
25
26
28
32
36
36
38
44

45

45

CONTENTS

viii
3.2
3.3
3.4
3.5
3.6
3.7
3.8
3.9
3.10
3.11
3.12

Controllability and Observability : : : : : : : :

Kalman Canonical Decomposition : : : : : : : :
Pole Placement and Canonical Forms : : : : : :
Observers and Observer-Based Controllers : : :
Operations on Systems : : : : : : : : : : : : : :
State Space Realizations for Transfer Matrices :
Lyapunov Equations : : : : : : : : : : : : : : :
Balanced Realizations : : : : : : : : : : : : : :
Hidden Modes and Pole-Zero Cancelation : : :
Multivariable System Poles and Zeros : : : : : :
Notes and References : : : : : : : : : : : : : : :

4 Performance Specications

4.1
4.2
4.3
4.4
4.5
4.6
4.7
4.8

Normed Spaces : : : : : : : : :
Hilbert Spaces : : : : : : : : : :
Hardy Spaces H2 and H1 : : :
Power and Spectral Signals : :
Induced System Gains : : : : :
Computing L2 and H2 Norms :
Computing L1 and H1 Norms
Notes and References : : : : : :

:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:

5 Stability and Performance of Feedback Systems

5.1
5.2
5.3
5.4
5.5
5.6
5.7
5.8

Feedback Structure : : : : : : : : :
Well-Posedness of Feedback Loop :
Internal Stability : : : : : : : : : :
Coprime Factorization over RH1 :
Feedback Properties : : : : : : : :
The Concept of Loop Shaping : : :
Weighted H2 and H1 Performance
Notes and References : : : : : : : :

6 Performance Limitations
6.1
6.2
6.3
6.4
6.5

:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:

Introduction : : : : : : : : : : : : : : : : : :
Integral Relations : : : : : : : : : : : : : : :
Design Limitations and Sensitivity Bounds :
Bode's Gain and Phase Relation : : : : : :
Notes and References : : : : : : : : : : : : :

:
:
:
:
:
:
:
:

:
:
:
:
:

:
:
:
:
:
:
:
:

:
:
:
:
:

:
:
:
:
:
:
:
:

:
:
:
:
:

:
:
:
:
:
:
:
:

:
:
:
:
:

:
:
:
:
:
:
:
:

:
:
:
:
:

:
:
:
:
:
:
:
:

:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:

47
53
58
63
65
68
71
72
78
80
89

91
91
93
97
102
104
112
114
116

117
117
119
121
126
130
134
137
141

143
143
145
148
151
152

CONTENTS

ix

7 Model Reduction by Balanced Truncation

7.1
7.2
7.3
7.4

Model Reduction by Balanced Truncation : : :

Frequency-Weighted Balanced Model Reduction
Relative and Multiplicative Model Reductions :
Notes and References : : : : : : : : : : : : : : :

8 Hankel Norm Approximation

8.1
8.2
8.3
8.4
8.5
8.6
8.7
8.8
8.9

Hankel Operator : : : : : : : : : : : : : : : :
All-pass Dilations : : : : : : : : : : : : : : : :
Optimal Hankel Norm Approximation : : : :
L1 Bounds for Hankel Norm Approximation
Bounds for Balanced Truncation : : : : : : :
Toeplitz Operators : : : : : : : : : : : : : : :
Hankel and Toeplitz Operators on the Disk* :
Nehari's Theorem* : : : : : : : : : : : : : : :
Notes and References : : : : : : : : : : : : : :

:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:
:

9 Model Uncertainty and Robustness

9.1
9.2
9.3
9.4
9.5
9.6
9.7

Model Uncertainty : : : : : : : : : : : : : : : : : :
Small Gain Theorem : : : : : : : : : : : : : : : : :
Stability under Stable Unstructured Uncertainties :
Unstructured Robust Performance : : : : : : : : :
Gain Margin and Phase Margin : : : : : : : : : : :
Deciency of Classical Control for MIMO Systems
Notes and References : : : : : : : : : : : : : : : : :

10 Linear Fractional Transformation

10.1
10.2
10.3
10.4
10.5

Linear Fractional Transformations :

Examples of LFTs : : : : : : : : :
Basic Principle : : : : : : : : : : :
Redheer Star-Products : : : : : :
Notes and References : : : : : : : :

11 Structured Singular Value

11.1
11.2
11.3
11.4
11.5

:
:
:
:
:

:
:
:
:
:

:
:
:
:
:

:
:
:
:
:

:
:
:
:
:

:
:
:
:
:

General Framework for System Robustness : :

Structured Singular Value : : : : : : : : : : :
Structured Robust Stability and Performance
Overview on  Synthesis : : : : : : : : : : : :
Notes and References : : : : : : : : : : : : : :

:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:

155

156
163
165
171

173

174
179
188
192
196
198
199
205
211

213

213
217
221
229
237
240
245

247

247
254
265
266
269

271

272
276
286
297
300

CONTENTS

12 Parameterization of Stabilizing Controllers

12.1
12.2
12.3
12.4
12.5
12.6
12.7

Existence of Stabilizing Controllers : : : : : : : : : :

Duality and Special Problems : : : : : : : : : : : : :
Parameterization of All Stabilizing Controllers : : : :
Structure of Controller Parameterization : : : : : : :
Closed-Loop Transfer Matrix : : : : : : : : : : : : :
Youla Parameterization via Coprime Factorization* :
Notes and References : : : : : : : : : : : : : : : : : :

13 Algebraic Riccati Equations

13.1
13.2
13.3
13.4
13.5
13.6
13.7
13.8
13.9

All Solutions of A Riccati Equation : : : :

Stabilizing Solution and Riccati Operator
Extreme Solutions and Matrix Inequalities
Spectral Factorizations : : : : : : : : : : :
Positive Real Functions : : : : : : : : : : :
Inner Functions : : : : : : : : : : : : : : :
Inner-Outer Factorizations : : : : : : : : :
Normalized Coprime Factorizations : : : :
Notes and References : : : : : : : : : : : :

14 H2 Optimal Control
14.1
14.2
14.3
14.4
14.5
14.6
14.7
14.8
14.9
14.10
14.11

:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:

Introduction to Regulator Problem : : : : : : : : :

Standard LQR Problem : : : : : : : : : : : : : : :
Extended LQR Problem : : : : : : : : : : : : : : :
Guaranteed Stability Margins of LQR : : : : : : :
Standard H2 Problem : : : : : : : : : : : : : : : :
Optimal Controlled System : : : : : : : : : : : : :
H2 Control with Direct Disturbance Feedforward*
Special Problems : : : : : : : : : : : : : : : : : : :
Separation Theory : : : : : : : : : : : : : : : : : :
Stability Margins of H2 Controllers : : : : : : : : :
Notes and References : : : : : : : : : : : : : : : : :

15 Linear Quadratic Optimization

15.1
15.2
15.3
15.4
15.5
15.6

:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:

Hankel Operators : : : : : : : : : : : : : : : : : : : : :
Toeplitz Operators : : : : : : : : : : : : : : : : : : : :
Mixed Hankel-Toeplitz Operators : : : : : : : : : : : :
Mixed Hankel-Toeplitz Operators: The General Case*
Linear Quadratic Max-Min Problem : : : : : : : : : :
Notes and References : : : : : : : : : : : : : : : : : : :

:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:

301

302
305
311
320
322
323
326

327

328
333
341
350
361
365
366
370
372

373

373
375
380
381
383
386
387
389
396
398
399

401

401
403
405
407
409
412

CONTENTS

16 H1 Control: Simple Case

16.1
16.2
16.3
16.4
16.5
16.6
16.7
16.8
16.9
16.10
16.11
16.12

Problem Formulation : : : : : : : :
Output Feedback H1 Control : : :
Motivation for Special Problems : :
Full Information Control : : : : : :
Full Control : : : : : : : : : : : : :
Disturbance Feedforward : : : : :
Output Estimation : : : : : : : : :
Separation Theory : : : : : : : : :
Optimality and Limiting Behavior
Controller Interpretations : : : : :
An Optimal Controller : : : : : : :
Notes and References : : : : : : : :

17 H1 Control: General Case

17.1
17.2
17.3
17.4
17.5
17.6
17.7
17.8
17.9

General H1 Solutions : : : : : : :
Loop Shifting : : : : : : : : : : : :
Relaxing Assumptions : : : : : : :
H2 and H1 Integral Control : : : :
H1 Filtering : : : : : : : : : : : :
Youla Parameterization Approach*
Connections : : : : : : : : : : : : :
State Feedback : : : : : : : : : : :
Notes and References : : : : : : : :

xi

:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:

18 H1 Loop Shaping
18.1
18.2
18.3
18.4

Robust Stabilization of Coprime Factors : : : : : : : :

Loop Shaping Using Normalized Coprime Stabilization
Theoretical Justication for H1 Loop Shaping : : : :
Notes and References : : : : : : : : : : : : : : : : : : :

19 Controller Order Reduction

19.1
19.2
19.3
19.4
19.5

Controller Reduction with Stability Criteria : :

H1 Controller Reductions : : : : : : : : : : : :
Frequency-Weighted L1 Norm Approximations
An Example : : : : : : : : : : : : : : : : : : : :
Notes and References : : : : : : : : : : : : : : :

:
:
:
:
:

:
:
:
:
:

:
:
:
:
:

:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:

413
413
414
420
424
430
431
433
434
438
443
445
446

449
449
453
457
459
462
464
468
470
475

477
477
486
489
495

497
498
505
512
517
521

CONTENTS

xii

20 Fixed Structure Controllers

523

21 Discrete Time Control

535

20.1
20.2
20.3

21.1
21.2
21.3
21.4
21.5
21.6
21.7
21.8
21.9
21.10

Lagrange Multiplier Method : : : : : : : : : : : : : : : : : : : : : : 523

Fixed Order Controllers : : : : : : : : : : : : : : : : : : : : : : : : 528
Notes and References : : : : : : : : : : : : : : : : : : : : : : : : : : 533
Discrete Time Systems : : : : : : : : : : :
Discrete Lyapunov Equations : : : : : : :
Discrete Riccati Equations : : : : : : : : :
Bounded Real Functions : : : : : : : : : :
Matrix Factorizations : : : : : : : : : : : :
Relations with the Continuous Time Case
Discrete Time H2 Control : : : : : : : : :
Discrete Balanced Model Reduction : : : :
Model Reduction Using Coprime Factors :
Notes and References : : : : : : : : : : : :

:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:

535
537
538
547
552
558
560
564
569
572

Bibliography

573

Index

591

Preface
This book is inspired by the recent development in the robust and H1 control theory, particularly the state-space H1 control theory. We give a fairly comprehensive
and step-by-step treatment of the state-space H1 control theory in the style of Doyle,
Glover, Khargonekar, and Francis [1989]. We have tried to make this derivation as selfcontained as possible and for reference have included many background results on linear
systems, the theory and application of Riccati equations and model reduction. We also
treat the robust control problems with unstructured and structured uncertainties. The
linear fractional transformation (LFT) and the structured singular value (known as )
are introduced as the unied tools for robust stability and performance analysis and
synthesis. Chapter 1 contains a more detailed chapter-by-chapter review of the topics
and results presented in this book. We have not included any exercises in this edition.
However, exercises and their solutions will be available through anonymous ftp on Internet: /pub/kemin/ZDG at hot.caltech.edu. Like any book, there are inevitably errors
and we therefore encourage all readers to write us about errors and suggestions. We
also encourage readers to send us exercise problems so that they can be shared by all
readers.
We would like to thank Professor Bruce A. Francis at University of Toronto for his
helpful comments and suggestions on early versions of the manuscript. As a matter
of fact, this manuscript was inspired by his lectures given at Caltech in 1987 and his
masterpiece { A Course in H1 Control Theory. We are grateful to Professor Andre Tits
at University of Maryland who has made numerous helpful comments and suggestions
that have greatly improved the quality of the manuscript. Professor Jakob Stoustrup,
Professor Hans Henrik Niemann, and their students at The Technical University of
Denmark have read various versions of this manuscript and have made many helpful
comments and suggestions. We are grateful to their help. Special thanks go to Professor
Andrew Packard at University of California-Berkeley for his help during the preparation
of the early versions of this manuscript. We are also grateful to Professor Jie Chen at
University of California-Riverside for providing material used in Chapter 6. We would
also like to thank Professor Kang-Zhi Liu at Chiba University (Japan) and Professor
Tongwen Chen at University of Calgary for their valuable comments and suggestions.
In addition, we would like to thank Gary Balas, Carolyn Beck, Dennis S. Bernstein,
Bobby Bodenheimer, Guoxiang Gu, Weimin Lu, John Morris, Matt Newlin, Li Qiu,
xiii

xiv

PREFACE

Hector P. Rotstein, Malcolm Smith, and many other people for their comments and
suggestions. The rst author is especially grateful to Professor Pramod P. Khargonekar
at The University of Michigan for introducing him to robust and H1 control and to
Professor Tryphon Georgiou at University of Minnesota for encouraging him to complete
this work. The trailing authors are pleased to acknowledge that the major contribution
to the writing of this book has been made by the rst author.
Finally, the rst author wants to thank his family for their support and encouragement.
Kemin Zhou
John C. Doyle
Keith Glover

Notation and Symbols

R
C
F
R+
C and C
C + and C +
C 0 , jR
D
D
@D

eld of real numbers

eld of complex numbers
eld, either R or C
nonnegative real numbers
open and closed left-half plane
open and closed right-half plane
imaginary axis
open unit disk
closed unit disk
unit circle

2

[
\

belong to
subset
union
intersection

2
3
~

end of proof
end of example
end of remark

:=

'
/



dened as
asymptotically greater than
asymptotically less than
much greater than
much less than


jj

Re ()

complex conjugate of  2 C
absolute value of  2 C
real part of  2 C

(t)

unit impulse
xv

NOTATION AND SYMBOLS

xvi

ij

1+ (t)

Kronecker delta function, ii = 1 and ij = 0 if i 6= j

unit step function

In
n  n identity matrix
[aij ]
a matrix with aij as its i-th row and j -th column element
diag(a1 ; : : : ; an ) an n  n diagonal matrix with ai as its i-th diagonal element
AT
transpose
A
adjoint operator of A or complex conjugate transpose of A
A 1
inverse of A
A+
pseudo inverse of A
A 
shorthand for (A 1 )
det(A)
determinant of A
Trace(A)
trace of A
(A)
eigenvalue of A
(A)
spectral radius of A
(A)
the set of spectrum of A
 (A)
largest singular value of A
 (A)
smallest singular value of A
i (A)
i-th singular value of A
(A)
condition number of A
kAk
spectral norm of A: kAk =  (A)
Im(A), R(A)
image (or range) space of A
Ker(A), N(A)
kernel (or null) space of A
X (A)
stable invariant subspace of A
X+ (A)
antistable invariant subspace of A
Ric(H )
gf


\
h; i
x?y
D?
S?

the stabilizing solution of an ARE

convolution of g and f
Kronecker product
direct sum or Kronecker sum
angle
inner product
orthogonal, hx; yi = 0


orthogonal
complement
of D, i.e., D D?


or DD is unitary
?
orthogonal complement of subspace S , e.g., H2?

L2 ( 1; 1)
L2 [0; 1)
L2 ( 1; 0]
L2+
L2

time domain Lebesgue space

subspace of L2 ( 1; 1)
subspace of L2 ( 1; 1)
shorthand for L2 [0; 1)
shorthand for L2 ( 1; 0]

NOTATION AND SYMBOLS

l2+
l2
L2 (j R)
L2 (@ D )
H2 (j R)
H2 (@ D )
H2? (j R)
H2? (@ D )
L1 (j R)
L1 (@ D )
H1 (j R)
H1 (@ D )
H1 (j R)
H1 (@ D )

shorthand for l2 [0; 1)

shorthand for l2 ( 1; 0)
square integrable functions on C 0 including at 1
square integrable functions on @ D
subspace of L2 (j R) with functions analytic in Re(s) > 0
subspace of L2 (@ D ) with functions analytic in jz j < 1
subspace of L2 (j R) with functions analytic in Re(s) < 0
subspace of L2 (@ D ) with functions analytic in jz j > 1
functions bounded on Re(s) = 0 including at 1
functions bounded on @ D
the set of L1 (j R) functions analytic in Re(s) > 0
the set of L1 (@ D ) functions analytic in jz j < 1
the set of L1 (j R) functions analytic in Re(s) < 0
the set of L1 (@ D ) functions analytic in jz j > 1

prex B or B
prex Bo
prex R

closed unit ball, e.g. BH1 and B

open unit ball
real rational, e.g., RH1 and RH2 , etc

R[s]

Rp (s)

polynomial ring
rational proper transfer matrices

G (s)

G
 (z )
A B
C D

shorthand for GT ( s) (continuous time)

shorthand for GT (z 1) (discrete time)
shorthand for state space realization C (sI A) 1 B + D
or C (zI A) 1 B + D

F` (M; Q)
Fu (M; Q)
S (M; N )

lower LFT
upper LFT
star product

xvii

xviii

NOTATION AND SYMBOLS

List of Acronyms
ARE
BR
CIF
DF
FC
FDLTI
FI
HF
i
lcf
LF
LFT
lhp or LHP
LQG
LQR
LTI
LTR
MIMO
nlcf
NP
nrcf
NS
OE
OF
OI
PR
rcf
rhp or RHP
RP
RS
SF
SISO

algebraic Riccati equation

bounded real
complementary inner factor
disturbance feedforward
full control
nite dimensional linear time invariant
full information
high frequency
if and only if
left coprime factorization
low frequency
linear fractional transformation
left-half plane Re(s) < 0
linear quadratic Gaussian
linear quadratic regulator
linear time invariant
loop transfer recovery
multi-input multi-output
normalized left coprime factorization
nominal performance
normalized right coprime factorization
nominal stability
output estimation
output feedback
output injection
positive real
right coprime factorization
right-half plane Re(s) > 0
robust performance
robust stability
state feedback
single-input single-output
xix

LIST OF ACRONYMS

xx
SPR
SSV
SVD

strictly positive real

structured singular value ()
singular value decomposition

Introduction
1.1

Historical Perspective

This book gives a comprehensive treatment of optimal H2 and H1 control theory and
an introduction to the more general subject of robust control. Since the central subject
of this book is state-space H1 optimal control, in contrast to the approach adopted
in the famous book by Francis [1987]: A Course in H1 Control Theory, it may be
helpful to provide some historical perspective of the state-space H1 control theory to
be presented in this book. This section is not intended as a review of the literature
in H1 theory or robust control, but rather only an attempt to outline some of the
work that most closely touches on our approach to state-space H1 . Hopefully our lack
of written historical material will be somewhat made up for by the pictorial history of
control shown in Figure 1.1. Here we see how the practical but classical methods yielded
to the more sophisticated modern theory. Robust control sought to blend the best of
both worlds. The strange creature that resulted is the main topic of this book.
The general area of methods for multivariable feedback design is well covered in the
book by Maciejowski [1989], where the importance of frequency response interpretations
is emphasized, whether the method explicitly involves the graphical manipulation of
frequency responses (e.g. the inverse Nyquist array of Rosenbrock, characteristic locus
method of MacFarlane or quantitative feedback theory of Horowitz) or is based of time
domain criteria such as linear quadratic methods. One of the motivations for the original
introduction of H1 methods by Zames [1981] was to bring plant uncertainty, specied
in the frequency domain, back into the centre-stage, as it had been in classical control in
contrast to analytic methods such as LQG. The H1 norm was found to be appropriate
1

INTRODUCTION

Figure 1.1: A picture history of control

for specifying both the level of plant uncertainty and the signal gain from disturbance
inputs to error outputs in the controlled system.
The `standard' H1 optimal control problem is concerned with the following block
diagram:

z
y

-K




w
u

where w represents an external disturbance, y is the measurement available to the

controller, u is the output from the controller, and z is an error signal that it is desired
to keep small. The transfer function matrix G represents not only the conventional
plant to be controlled but also any weighting functions included to specify the desired

1.1. Historical Perspective

performance. The H1 optimal control problem is then to design a stabilizing controller,

K , so as to minimize the closed loop transfer function from w to z , Tzw , in the H1
norm, where
kTzw k1 = sup  (Tzw (j!)) :
!

The H1 norm gives the maximum energy gain (the induced L2 system gain), or sinusoidal gain of the system. This is in contrast to the H2 norm, kTzw k2 , which for example
gives the variance of the output given white noise disturbances. The important property
of the H1 norm comes from the application of the small gain theorem, which states
that if kTzw k1  then the system with block diagram,

z
y

-

G

-K




w
u

will be stable for all stable
with k
k1 < 1= . It is probably the case that this robust
stability consequence was the main motivation for the development of H1 methods
rather than the worst case signal gain.
The synthesis of controllers that achieve an H1 norm specication hence gives a welldened mathematical problem whose solution we will now discuss. Most of the original
solution techniques were in an input-output setting and involved analytic functions
(Nevanlinna-Pick interpolation) or operator-theoretic methods [Sarason, 1967; Adamjan
et al., 1978; Ball and Helton, 1983], and such derivations involved a fruitful collaboration
between Operator Theorists and Control Engineers (see Dym [1994] for some historical
remarks). Indeed, H1 theory seemed to many to signal the beginning of the end
for the state-space methods which had dominated control for the previous 20 years.
Unfortunately, the standard frequency-domain approaches to H1 started running into
signicant obstacles in dealing with multi-input multi-output (MIMO) systems, both
mathematically and computationally, much as the H2 (or LQG) theory of the 1950's
had.
Not surprisingly, the rst solution to a general rational MIMO H1 optimal control
problem, presented in Doyle [1984], relied heavily on state-space methods, although
more as a computational tool than in any essential way. The procedure involved statespace inner/outer and coprime factorizations of transfer function matrices which reduced
the problem to a Nehari/Hankel norm problem solvable by the state-space method in
Glover [1984]. Both [Francis, 1987] and [Francis and Doyle, 1987] give expositions of
this approach, which in a mathematical sense \solved" the general rational problem but
in fact suered from severe problems with the high order of the Riccati equations to be
solved.

INTRODUCTION

It is interesting to observe at this point that the above techniques related to Hankel
operators were simultaneously being exploited and developed in the model reduction
literature. In particularly the striking result of Adamjan, Arov and Krein [1978] on
rational approximation in the Hankel norm, (which had been communicated to the
Western Systems and Control community by Helton), had led to the state space multivariable results in Kung and Lin [1981] and Glover [1984]. The latter paper gave a
self-contained state-space treatment exploiting the balanced realizations proposed for
model reduction by Moore [1981] which are also of independent interest.
The simple state space H1 controller formulae to be presented in this book were rst
announced in Glover and Doyle [1988] (after some sustained manipulation). However
the very simplicity of the new formulae and their similarity with the H2 ones suggested
a more direct approach. Independent encouragement for a simpler approach to the H1
problem came from papers by Khargonekar, Petersen, Rotea, and Zhou [1987,1988],
They showed that the state-feedback H1 problem can be solved by solving an algebraic
Riccati equation and completing the square.
Derivations of the controller formulae in Glover and Doyle [1988] using derivations
more akin to the above state feedback results were given in Doyle, Glover, Khargonekar
and Francis [1989] and will form the basis of the developments in this book. The
operator theory still plays a central role (as does Redheer's work [Redheer, 1960] on
linear fractional transformations), but its use is more straightforward. The key to this
was a return to simple and familiar state-space tools, in the style of Willems [1971], such
as completing the square, and the connection between frequency domain inequalities (e.g
kGk1 < 1), Riccati equations, and spectral factorizations.
This has been a brief and personal account of these developments and more extensive,
but still inadequate, comments are made in section 16.12. Relations between H1 have
now been established with many other topics in control: e.g. risk sensitive control of
Whittle [1981, 1990]; dierential games (see Basar and Bernhard [1991], Limebeer et al
[1992], Green and Limebeer [1995]); J-lossless factorization (Green [1992]); maximum
entropy methods of Dym and Gohberg [1986] (see Mustafa and Glover [1990]). The
state-space theory of H1 has been also been carried much further, by generalizing
time-invariant to time-varying, innite horizon to nite horizon, and nite dimensional
to innite dimensional and even to some nonlinear results. It should be mentioned that
in these generalizations the term H1 has come to be (mis-)used to mean the induced
norm in L2 . Some of these developments also provided alternative derivations of the
standard H1 results. Indeed one of the attractions of this area to researchers is that it
can be approached from such a diverse technical backgrounds with each providing its
own interpretations. These developments are beyond the scope of the present book.
Having established that the above H1 control problem can be relatively easily solved
and can represent specications for performance and robustness let us return to the
question of whether this gives a suitable robust control design tool. There is no question
that the algorithm can be used to provide poor controllers due to poorly chosen problem
descriptions resulting in, for example, very high bandwidth controllers. Two approaches
mentioned in this book attempt to satisfy this requirement. Firstly the method of H1