Professional Documents
Culture Documents
Typo (4 Files Merged)
Typo (4 Files Merged)
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 denition of B (s)" should be \using the denition of B~ (s)"
P162, Example can be simplied.
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
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
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
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 .
3
30. P494, L10,
(M~ s ) = (Ms ) =
should be
(M~ s ) = (Ms ) =
1=2
1=2
1 + 2 (W2 PW1 )
1 + 2 (W2 PW1 )
2
2 (N~s ) = 1 2 (M~ s ) = 1 + (P2 (sP) ) :
s
32. P587, reference no. 214, add \pp. 5-16" at the end.
KEMIN ZHOU
with
JOHN C. DOYLE and KEITH GLOVER
1.2.
loop shaping is described where the desired loop shape is specied 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
INTRODUCTION
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.
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
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
presented:
and
where
A BB =
2 :
C C
A
Chapter 5 introduces the feedback structure and discusses its stability and performance properties.
1.3.
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
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
10
INTRODUCTION
kG(s) Gr (s)k1 2
N
X
i=r+1
i :
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 satises 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) D0
1
N
X
i=r+1
i :
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
system described below with an unstructured uncertainty 2 RH1 with kk1 < 1 is
robustly stable if and only if the transfer function from w to z has H1 norm no greater
than 1.
nominal system
-
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 ki 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
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 denite (i.e., P 0 or
P 0) and (P; A) has no unobservable modes on the j!-axis. Dene
(s) = B ( sI
A ) 1 I
P S
S R
(sI A) 1 B :
I
Then
(A BR 1 S ) X + X (A BR 1 S ) XBR 1B X + P SR 1S = 0
() 9 a solution X to
A BR 1 S
(P SR 1S )
BR 1B
(A BR 1 S )
(A BR 1 S ) X + X (A BR 1 S ) XBR 1B X + P SR 1S = 0
= 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
Dene
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
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
u
y
M1
M1 (s) = 4
A^1
F1
C2
Z1 L1 Z1 B2
0
I
I
0
3
5
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
~ 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
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 : : : : : : : : : : :
Semidenite Matrices : : : : : : : : : : :
Matrix Dilation Problems* : : : : : : : :
Notes and References : : : : : : : : : : :
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
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
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 : : : : : :
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
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
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 : : : : : : : : : : : : : :
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
Model Uncertainty : : : : : : : : : : : : : : : : : :
Small Gain Theorem : : : : : : : : : : : : : : : : :
Stability under Stable Unstructured Uncertainties :
Unstructured Robust Performance : : : : : : : : :
Gain Margin and Phase Margin : : : : : : : : : : :
Deciency of Classical Control for MIMO Systems
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
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
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
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
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 : : : : : : : :
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
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
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
523
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
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
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 unied 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
2
[
\
belong to
subset
union
intersection
2
3
~
end of proof
end of example
end of remark
:=
'
/
dened 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
xvi
ij
1+ (t)
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?
L2 ( 1; 1)
L2 [0; 1)
L2 ( 1; 0]
L2+
L2
prex B or B
prex Bo
prex R
R[s]
Rp (s)
polynomial ring
rational proper transfer matrices
G (s)
G
(z )
A B
C D
F` (M; Q)
Fu (M; Q)
S (M; N )
lower LFT
upper LFT
star product
xvii
xviii
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
LIST OF ACRONYMS
xx
SPR
SSV
SVD
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, specied
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
z
y
-K
w
u
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 kk1 < 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 specication hence gives a welldened 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
signicant 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, innite horizon to nite horizon, and nite dimensional
to innite 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 specications 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