1

AAE 666 Notes

Martin Corless
School of Aeronautics & Astronautics
Purdue University
West Lafayette, Indiana
[email protected]

January 3, 2017

Contents
1 Introduction

2 State space representation of dynamical systems

2.1 Examples . . . . . . . . . . . . . . . . . . . . . .
2.1.1 Continuous-time . . . . . . . . . . . . . .
2.1.2 Discrete-time . . . . . . . . . . . . . . . .
2.2 General representation . . . . . . . . . . . . . . .
2.2.1 Continuous-time . . . . . . . . . . . . . .
2.2.2 Discrete-time . . . . . . . . . . . . . . . .
2.3 Vectors . . . . . . . . . . . . . . . . . . . . . . . .
2.3.1 Vector spaces and IRn . . . . . . . . . . .
2.3.2 IR2 and pictures . . . . . . . . . . . . . . .
2.3.3 Derivatives . . . . . . . . . . . . . . . . .
2.4 Vector representation of dynamical systems . . . .
2.5 Solutions and equilibrium states . . . . . . . . . .
2.5.1 Continuous-time . . . . . . . . . . . . . .
2.5.2 Discrete-time . . . . . . . . . . . . . . . .
2.6 Numerical simulation . . . . . . . . . . . . . . . .
2.6.1 MATLAB . . . . . . . . . . . . . . . . . .
2.6.2 SIMULINK . . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

7
7
7
13
14
14
16
17
17
19
20
21
23
23
25
26
26
27

3 First order systems

3.1 Continuous time . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2 Solutions to scalar first order ODES . . . . . . . . . . . . . . . . . . . . . . .
3.3 Discrete time . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

29
29
31
32

4 Second order systems

4.1 Linear systems . . . . . . . . . .
4.2 Linearization . . . . . . . . . . .
4.3 Isocline method . . . . . . . . . .
4.4 Periodic solutions and limit cycles

.
.
.
.

33
33
35
38
38

5 Some general considerations

5.1 Existence of solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.2 Uniqueness of solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

41
41
41

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

CONTENTS
5.3

Indefinite extension of solutions . . . . . . . . . . . . . . . . . . . . . . . . .

6 Stability and boundedness

6.1 Boundedness of solutions . . . . . .
6.2 Stability of equilibrium states . . .
6.3 Asymptotic stability . . . . . . . .
6.3.1 Global asymptotic stability
6.3.2 Asymptotic stability . . . .
6.4 Exponential stability . . . . . . . .
6.5 LTI systems . . . . . . . . . . . . .
6.6 Linearization and stability . . . . .

.
.
.
.
.
.
.
.

7 Stability and boundedness: discrete

7.1 Boundedness of solutions . . . . . .
7.2 Stability of equilibrium states . . .
7.3 Asymptotic stability . . . . . . . .
7.3.1 Global asymptotic stability
7.3.2 Asymptotic stability . . . .
7.4 Exponential stability . . . . . . . .
7.5 LTI systems . . . . . . . . . . . . .
7.6 Linearization and stability . . . . .

time
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

42

.
.
.
.
.
.
.
.

43
43
46
48
48
48
50
52
54

.
.
.
.
.
.
.
.

57
57
58
59
59
60
60
62
63

8 Basic Lyapunov theory

8.1 Lyapunov conditions for stability . . . . . . .
8.1.1 Locally positive definite functions . . .
8.1.2 A stability result . . . . . . . . . . . .
8.2 Asymptotic stability . . . . . . . . . . . . . .
8.3 Boundedness . . . . . . . . . . . . . . . . . .
8.3.1 Radially unbounded functions . . . . .
8.3.2 A boundedness result . . . . . . . . . .
8.4 Global asymptotic stability . . . . . . . . . . .
8.4.1 Positive definite functions . . . . . . .
8.4.2 A result on global asymptotic stability
8.5 Exponential stability . . . . . . . . . . . . . .
8.5.1 Global exponential stability . . . . . .
8.5.2 Proof of theorem 6 . . . . . . . . . . .
8.5.3 Exponential stability . . . . . . . . . .
8.5.4 A special class of GES systems . . . .
8.5.5 Summary . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

65
67
67
69
72
74
74
74
77
77
78
81
81
82
83
84
86

9 Basic Lyapunov theory: discrete time

9.1 Stability . . . . . . . . . . . . . . . .
9.2 Asymptotic stability . . . . . . . . .
9.3 Boundedness . . . . . . . . . . . . .
9.4 Global asymptotic stability . . . . . .

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

87
88
89
90
91

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

CONTENTS
9.5

Exponential stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

10 Lyapunov theory for linear time-invariant systems

10.1 Positive and negative (semi)definite matrices . . . .
10.1.1 Definite matrices . . . . . . . . . . . . . . .
10.1.2 Semi-definite matrices . . . . . . . . . . . .
10.2 Lyapunov theory . . . . . . . . . . . . . . . . . . .
10.2.1 Asymptotic stability results . . . . . . . . .
10.2.2 MATLAB. . . . . . . . . . . . . . . . . . . .
10.2.3 Stability results . . . . . . . . . . . . . . . .
10.3 Mechanical systems . . . . . . . . . . . . . . . . . .
10.4 Rates of convergence for linear systems . . . . . . .
10.5 Linearization and exponential stability . . . . . . .

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

92
95
95
95
97
99
99
103
104
106
109
110

11 Quadratic stability
113
11.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
11.2 Polytopic nonlinear systems . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
11.2.1 LMI Control Toolbox . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
11.2.2 Generalization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
11.2.3 Robust stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
11.3 Another special class of nonlinear systems . . . . . . . . . . . . . . . . . . . 120
11.3.1 The circle criterion: a frequency domain condition . . . . . . . . . . . 124
11.3.2 Sector bounded nonlinearities . . . . . . . . . . . . . . . . . . . . . . 125
11.3.3 Generalization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
11.4 Yet another special class . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
11.4.1 Generalization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
11.4.2 Strictly positive real transfer functions and a frequency domain condition133
12 Invariance results
12.1 Invariant sets . . . . . . . . . . . . . . . . . .
12.2 Limit sets . . . . . . . . . . . . . . . . . . . .
12.3 LaSalles Theorem . . . . . . . . . . . . . . .
12.4 Asymptotic stability and LaSalle . . . . . . .
12.4.1 Global asymptotic stability . . . . . .
12.4.2 Asymptotic stability . . . . . . . . . .
12.5 LTI systems . . . . . . . . . . . . . . . . . . .
12.6 Applications . . . . . . . . . . . . . . . . . . .
12.6.1 A simple stabilizing adaptive controller
12.6.2 PI controllers for nonlinear systems . .

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

139
139
141
145
148
148
150
151
152
152
153

13 Stability of nonautonomous systems

13.1 Stability and boundedness . . . . .
13.1.1 Boundedness of solutions . .
13.1.2 Stability . . . . . . . . . . .
13.1.3 Asymptotic stability . . . .

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

155
156
156
156
158

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

CONTENTS
13.1.4 Global asymptotic stability . . . . . . . . . . . . . . . . . . . . . . . 158
13.1.5 Exponential stability . . . . . . . . . . . . . . . . . . . . . . . . . . . 159

14 Lyapunov theory for nonautonomous systems

14.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . .
14.2 Stability . . . . . . . . . . . . . . . . . . . . . . . . . .
14.2.1 Locally positive definite functions . . . . . . . .
14.2.2 A stability theorem . . . . . . . . . . . . . . . .
14.3 Uniform asymptotic stability . . . . . . . . . . . . . . .
14.4 Global uniform asymptotic stability . . . . . . . . . . .
14.4.1 Positive definite functions . . . . . . . . . . . .
14.4.2 A result on global uniform asymptotic stability
14.4.3 Proof of theorem ?? . . . . . . . . . . . . . . .
14.5 Exponential stability . . . . . . . . . . . . . . . . . . .
14.5.1 Exponential stability . . . . . . . . . . . . . . .
14.5.2 Global uniform exponential stability . . . . . .
14.6 Quadratic stability . . . . . . . . . . . . . . . . . . . .
14.7 Boundedness . . . . . . . . . . . . . . . . . . . . . . .
14.8 Attractive and invariant regions . . . . . . . . . . . . .
14.9 Systems with disturbance inputs . . . . . . . . . . . . .
14.9.1 Linear systems with bounded disturbance inputs
14.10Regions of attraction . . . . . . . . . . . . . . . . . . .
14.10.1 A special class of nonlinear systems . . . . . . .
14.11Linear time-varying systems . . . . . . . . . . . . . . .
14.11.1 Lyapunov functions . . . . . . . . . . . . . . . .
14.12Linearization . . . . . . . . . . . . . . . . . . . . . . .
15 The
15.1
15.2
15.3

describing function method

Periodic functions and Fourier series
Describing functions . . . . . . . . .
The describing function method . . .
15.3.1 Rational transfer functions . .
15.4 Exercises . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

161
161
163
163
163
164
165
165
165
166
168
168
168
168
169
171
172
174
176
177
179
179
181

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

183
184
185
189
193
194

16 Aerospace/Mechanical Systems
16.1 Aerospace/Mechanical systems . . . . . . . . . .
16.2 Equations of motion . . . . . . . . . . . . . . .
16.3 Some fundamental properties . . . . . . . . . .
16.3.1 An energy result . . . . . . . . . . . . .
16.3.2 Potential energy and total energy . . . .
16.3.3 A skew symmetry property . . . . . . . .
16.4 Linearization of aerospace/mechanical systems .
16.5 A class of GUAS aerospace/mechanical systems
16.6 Control of aerospace/mechanical systems . . . .
16.6.1 Computed torque method . . . . . . . .

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

199
199
200
203
203
204
206
208
209
211
211

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

CONTENTS

16.6.2 Linearization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211

16.6.3 Globally stabilizing PD controllers . . . . . . . . . . . . . . . . . . . 212
17 Quadratic stabilizability
17.1 Preliminaries . . . . . . . . . . . . . . . .
17.2 Some special classes of nonlinear systems .
17.2.1 A special class of nonlinear systems
17.2.2 Generalization . . . . . . . . . . . .
17.3 Another special class of nonlinear systems
17.3.1 Generalization . . . . . . . . . . . .
17.4 Yet another special class . . . . . . . . . .
17.4.1 Generalization . . . . . . . . . . . .

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

18 Singularly perturbed systems

18.1 Introduction and examples . . . . . . . . . . . . . .
18.1.1 High gain output feedback . . . . . . . . . .
18.1.2 Single-link manipulator with flexible joint .
18.2 Singularly perturbed systems . . . . . . . . . . . .
18.3 Linear systems . . . . . . . . . . . . . . . . . . . .
18.4 Exact slow and fast subsystems . . . . . . . . . . .
18.4.1 Approximations of the slow and fast systems
18.4.2 A special class of systems . . . . . . . . . .

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

19 Input output systems

19.1 Input-output systems . . . . . . . . . . . . . . . . . .
19.1.1 Causality . . . . . . . . . . . . . . . . . . . .
19.2 Signal norms . . . . . . . . . . . . . . . . . . . . . .
19.3 Input-output stability . . . . . . . . . . . . . . . . . .
19.4 General convolution systems . . . . . . . . . . . . . .
19.4.1 Systems with L1 impulse responses . . . . . .
19.4.2 Systems with impulses in impulse response . .
19.4.3 L2 gain and H norm . . . . . . . . . . . . .
19.4.4 Finite dimensional linear systems . . . . . . .
19.4.5 Linear differential delay systems . . . . . . . .
19.5 A class of Lp stable nonlinear systems . . . . . . . . .
19.6 Small gain theorem . . . . . . . . . . . . . . . . . . .
19.6.1 Stability of feedback systems . . . . . . . . . .
19.6.2 Truncations . . . . . . . . . . . . . . . . . . .
19.6.3 Small gain theorem . . . . . . . . . . . . . . .
19.7 Application of the small gain theorem: circle criterion
19.7.1 A more general circle criterion . . . . . . . . .

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

215
215
216
217
222
224
226
228
229

.
.
.
.
.
.
.
.

231
231
231
232
234
237
238
240
241

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

247
247
249
249
251
252
252
257
258
259
263
265
266
266
266
267
268
270

8
20 Nonlinear H
20.1 Analysis . . . . . . . . . . . . . .
20.1.1 The HJ Inequality . . . .
20.2 Control . . . . . . . . . . . . . .
20.3 Series solution to control problem
20.3.1 Linearized problem . . . .
20.3.2 Nonlinear problem . . . .

CONTENTS

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

273
273
273
276
277
277
278

21 Performance
21.1 Analysis . . . . . . . . . . . . . . . . . . . . . . .
21.2 Polytopic systems . . . . . . . . . . . . . . . . . .
21.2.1 Polytopic models . . . . . . . . . . . . . .
21.2.2 Performance analysis of polytopic systems
21.3 Control for performance . . . . . . . . . . . . . .
21.3.1 Linear time-invariant systems . . . . . . .
21.3.2 Control of polytopic systems . . . . . . . .
21.3.3 Multiple performance outputs . . . . . . .

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

281
281
282
282
285
286
286
287
288

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

22 Appendix
289
22.1 The Euclidean norm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 289

Chapter 1
Introduction

HI!

(How is it going?)

This is a course on the analysis and control of nonlinear systems.

Why study nonlinear systems?
Most physical systems are nonlinear.
Nonlinear systems exhibit many types of behavior which do not occur in linear systems:
several equilibrium conditions; isolated periodic solutions; chaos.
Some systems are not linearizable.
Linearization yields local behavior only.
Text: Khalil, K. H., Nonlinear Systems, 2nd Edition, Prentice Hall, 1996.
References
Slotine J.-J. E., and Li, W., Applied Nonlinear Control, Prentice Hall, 1991.
Vidyasagar, M., Nonlinear Systems Analysis, 2nd Edition, Prentice Hall, 1993.
Boyd, S., El Ghaoui, L., Feron, E., and Balakrishnan, V., Linear Matrix Inequalities in
System and Control Theory, SIAM, 1994.

10

CHAPTER 1. INTRODUCTION

Chapter 2
State space representation of
dynamical systems
2.1
2.1.1

Examples
Continuous-time

A first example

Figure 2.1: The first nonlinear system

Consider the motion of a small block of mass m moving along a straight horizontal line.
Suppose there is Coulomb (or dry) friction between the block and surface with coefficent of
friction > 0. Letting t IR represent time and v(t) IR represent the velocity of the
block at time t, the motion of the block can be described by the following first order ordinary
differential equation (ODE):
mv = mg sgm (v)
where the signum function, sgm, is defined by

1 if v < 0
0 if v = 0
sgm (v) :=

1 if v > 0
11

12

CHAPTER 2. STATE SPACE REPRESENTATION OF DYNAMICAL SYSTEMS

Figure 2.2: signum function

and g is the gravitational acceleration constant of the planet on which the block resides.
Introducing x := v results in
x = sgm (x)
and := g.
Planar pendulum

Figure 2.3: Planar pendulum

J + W l sin = 0
Introducing the state variables,
x1 :=

and

x2 := ,

results in the state space description:

x 1 = x2
x 2 = a sin x1
where a := W l/J > 0.
Attitude dynamics of a rigid body
The equations describing the rotational dynamics of a rigid body are given by Eulers equations of motion. If we choose the axes of a body-fixed reference frame along the principal

2.1. EXAMPLES

13

Figure 2.4: Attitude dynamics

axes of inertia of the rigid body with origin at the center of mass, Eulers equations of motion
take the simplified form
I1 1 = (I2 I3 )2 3
I2 2 = (I3 I1 )3 1
I3 3 = (I1 I2 )1 2
where 1 , 2 , 3 denote the components of the body angular velocity vector with respect
to the body principal axes, and the positive scalars I1 , I2 , I3 are the principal moments of
inertia of the body with respect to its mass center.
Body in central force motion

Figure 2.5: Central force motion

r r 2 + g(r) = 0
r + 2r

=0
For the simplest situation in orbit mechanics ( a satellite orbiting YFHB)
g(r) = /r 2

= GM

where G is the universal constant of gravitation and M is the mass of YFHB.

14

CHAPTER 2. STATE SPACE REPRESENTATION OF DYNAMICAL SYSTEMS

Ballistics in drag

Figure 2.6: Ballistics in drag

x
y
mv
mv

=
=
=
=

v cos
v sin
mg sin D(v)
mg cos

2.1. EXAMPLES

15

Double pendulum on cart

y

P2

m2

m0
P1

2
m1

Figure 2.7: Double pendulum on cart

The motion of the double pendulum on a cart can be described by
(m0 + m1 + m2 )
y m1 l1 cos 1 1 m2 l2 cos 2 2 + m1 l1 sin 1 12 + m2 l2 sin 2 22 = u
m1 l1 cos 1 y + m1 l12 1
+
m1 l1 g sin 1
= 0
2
m2 l2 cos 2 y + m2 l2 2
+
m2 l2 g sin 2
= 0

Figure 2.8: Double pendulum on cart

16

CHAPTER 2. STATE SPACE REPRESENTATION OF DYNAMICAL SYSTEMS

Two-link robotic manipulator

Payload

g^
lc 2

q2

q1

lc 1

u2

u1

Figure 2.9: A simplified model of a two link manipulator

The coordinates q1 and q2 denote the angular location of the first and second links relative
to the local vertical, respectively. The second link includes a payload located at its end. The
masses of the first and the second links are m1 and m2 , respectively. The moments of inertia
of the first and the second links about their centers of mass are I1 and I2 , respectively. The
locations of the center of mass of links one and two are determined by lc1 and lc2 , respectively;
l1 is the length of link 1. The equations of motion for the two arms are described by:
q1
[m1 lc21 + m2 l12 + I1 ]

+[m2 l1 lc2 cos(q1 q2 )]

q2 +m2 l1 lc2 sin(q1 q2 )q22 [m1 lc1 + m2 l1 ]g sin(q1 ) = u1

[m2 l1 lc2 cos(q1 q2 )]

q1 +[m2 lc22 + I2 ]
q2

m2 l1 lc2 sin(q1 q2 )q12 m2 glc2 sin(q2 )

= u2

2.1. EXAMPLES

17

Traffic flow
Two roads connecting two points. Let x1 and x2 be the traffic flow rate (number of cars per
hour) along the two routes.
x 1 = (c1 (x) c2 (x))x1 + (c2 (x) c1 (x))x2
x 2 = (c1 (x) c2 (x))x1 (c2 (x) c1 (x))x2
where
(y) =

y if y 0
0 if y < 0

ci (x) is the travel time along route i.

c1 (x) = x21 + 1

2.1.2

c2 (x) = 2x2

Discrete-time

Example 1 (Population dynamics)

x(k+1) = ax(k)
Example 2 (Computer control)
Example 3 (Iterative numerical algorithm) Recall Newtons algorithm for solving the
equation g(x) = 0:
g(x(k))
x(k+1) = x(k)
g (x(k))
Example 4 (Consensus algorithm)

18

CHAPTER 2. STATE SPACE REPRESENTATION OF DYNAMICAL SYSTEMS

2.2
2.2.1

General representation
Continuous-time

All of the preceding systems can be described by a bunch of first order ordinary differential
equations of the form
x 1 = f1 (x1 , x2 , . . . , xn )
x 2 = f2 (x1 , x2 , . . . , xn )
..
.
x n = fn (x1 , x2 , . . . , xn )
where the scalars xi (t), i = 1, 2, . . . , n, are called the state variables and the real scalar t is
called the time variable.
Higher order ODE descriptions
Single equation. (Recall the planar pendulum.) Consider a system described by a single
nth - order differential equation of the form
F (q, q,
. . . , q (n) ) = 0
n

where q(t) is a scalar and q (n) := ddtnq . To obtain a state space description of this system, we
need to assume that we can solve for q (n) as a function of q, q,
. . . , q (n1) . So suppose that
the above equation is equivalent to
q (n) = a(q, q,
. . . , q (n1) ) .
Now let
x1 := q
x2 := q
..
.
xn := q (n1)
to obtain the following state space description:
x 1 = x2
x 2 = x3
..
.
x n1 = xn
x n = a(x1 , x2 , . . . , xn )

2.2. GENERAL REPRESENTATION

19

Multiple equations. (Recall the two link manipulator.) Consider a system described by
m scalar differential equations in N variables:
(n1 )

F1 (q1 , q1 , . . . , q1

(N2 )

, q2 , q2 , . . . , q2

(n1 )

(n )

, . . . , qN , qN , . . . , qN N ) = 0

(n2 )

(n )

, . . . , qN , qN , . . . , qN N ) = 0
..
.
(n1 )
(n2 )
(nN )
FN (q1 , q1 , . . . , q1 , q2 , q2 , . . . , q2 , . . . , qN , qN , . . . , qN ) = 0
F2 (q1 , q1 , . . . , q1

, q2 , q2 , . . . , q2

(n )

where t, q1 (t), q2 (t), . . . , qN (t) are real scalars. Note that qi i is the highest order derivative
of qi which appears in the above equations.
(n )
(n )
(n )
First solve for the highest order derivatives, q1 1 , q2 2 , . . . , qN N , to obtain:
(n1 )

q1

(n )
q2 2

(n )

qN N

(n11)

= a1 ( q1 , q1 , . . . q1

(n21)

, q2 , q2 , . . . , q2

(n 1)
a2 ( q1 , q1 , . . . q1 1 ,

(n 1)

, . . . , qN , qN , . . . , qN N

(n 1)
q2 , q2 , . . . , q2 2 ,

(n 1)
qN , qN , . . . , qN N

=
...,
)
..
.
(n 1)
(n 1)
(n 1)
= aN ( q1 , q1 , . . . q1 1 , q2 , q2 , . . . , q2 2 , . . . , qN , qN , . . . , qN N )

Now let
(n 1)

x1 := q1
xn1 +1 := q2 ,

x2 := q1
xn1 +2 := q2

...
...
..
.

xn1 := q1 1
(n 1)
xn1 +n2 := q2 2

xn1 +...+nm1 +1 := qN

xn1 +...+nm1 +2 := qN

...

xn := qN N

where
n := n1 + n2 + . . . + nN
to obtain
x 1 = x2
x 2 = x3
..
.
x n11 = xn1
x n1 = a1 (x1 , x2 , . . . , xn )
x n1 +1 = xn1 +2
..
.
x n1 +n2 = a2 (x1 , x2 , . . . , xn )
..
.
x n = an (x1 , x2 , . . . , xn )
Example 5
q1 + q2 + 2q1 = 0

q1 + q1 + q2 + 4q2 = 0

(n 1)

20

2.2.2

CHAPTER 2. STATE SPACE REPRESENTATION OF DYNAMICAL SYSTEMS

Discrete-time

A general state space description of a discrete-time system consists of a bunch of first order
difference equations of the form
x1 (k+1) = f1 (x1 (k), x2 (k), . . . , xn (k))
x2 (k+1) = f2 (x1 (k), x2 (k), . . . , xn (k))
..
.
xn (k+1) = fn (x1 (k), x2 (k), . . . , xn (k))
where the scalars xi (k), i = 1, 2, . . . , n, are called the state variables and the integer k is
called the time variable.
Higher order DE descriptions
Story is similar to continuous-time.

2.3. VECTORS

2.3
2.3.1

21

Vectors
Vector spaces and IRn

A scalar is a real or a complex number. The symbols IR and C represent the set of real and
complex numbers, respectively.
In this section all the definitions and results are given for real scalars. However, they also
hold for complex scalars; to get the results for complex scalars, simply replace real with
complex and IR with C.
Consider any positive integer n. A real n-vector x is an ordered n-tuple of real numbers,
x1 , x2 , . . . , xn . This is is usually written as follows:

x1
x2

x = ..
or
x = x1 , x2 , , xn
.
xn

The real numbers x1 , x2 , . . . , xn are called the scalar components of x; xi is called the i-th
component. The symbol IRn represents the set of ordered n-tuples of real numbers.
Addition. The addition of any two elements x, y of IRn yields an element of IRn and is
defined by:

x1 + y1
y1
x1
x2 + y2
y2
x2

x + y = .. + .. :=

..

.
.
.
xn + yn
yn
xn
Zero element of IRn :

0 :=

0
0
..
.
0

Note that we are using the same symbol, 0, for a zero scalar and a zero vector.
The negative of x:

x :=

x1
x2
..
.
xn

22

CHAPTER 2. STATE SPACE REPRESENTATION OF DYNAMICAL SYSTEMS

Properties of addition
(a) (Commutative). For each pair x, y in IRn ,
x+y = y+x
(b) (Associative). For each x, y, z in IRn ,
(x + y) + z = x + (y + z)
(c) There is an element 0 in IRn such that for every x in IRn ,
x+0=x
(d) For each x in IRn , there is an element x in IRn such that
x + (x) = 0
Scalar multiplication. The multiplication of an element x of IRn by a real scalar yields
an element of IRn and is defined by:

x1
x1
x2
x2

x = .. := ..
.
.
xn
xn
Properties of scalar multiplication
(a) For each scalar and pair x, y in IRn
(x + y) = x + y
(b) For each pair of scalars , and x in IRn ,
( + )x = x + x
(c) For each pair of scalars , , and x in IRn ,
(x) = ()x
(d) For each x in IRn ,
1x = x

2.3. VECTORS

23

Vector space. Consider any set V equipped with an addition operation and a scalar
multiplication operation. Suppose the addition operation assigns to each pair of elements
x, y in V a unique element x + y in V and it satisfies the above four properties of addition
(with IRn replaced by V). Suppose the scalar multiplication operation assigns to each scalar
and element x in V a unique element x in V and it satisfies the above four properties
of scalar multiplication (with IRn replaced by V). Then this set (along with its addition
and scalar multiplication) is called a vector space. Thus IRn equipped with its definitions
of addition and scalar multiplication is a specific example of a vector space. We shall meet
other examples of vectors spaces later.
An element x of a vector space is called a vector.
A vector space with real (complex) scalars is called a real (complex) vector space.
Subtraction in a vector space is defined by:

x y := x + (y)
Hence in IRn ,

xy =

2.3.2

x1
x2
..
.
xn

y1
y2
..
.
yn

x1 y1
x2 y2
..
.
xn yn

IR2 and pictures

An element of IR2 can be represented in a plane by a point or a directed line segment.

24

2.3.3

CHAPTER 2. STATE SPACE REPRESENTATION OF DYNAMICAL SYSTEMS

Derivatives

Suppose x() is a function of a real variable t IR where x(t) is in IRn . Then

dx
1
x

1
dt

dx2 x 2

dt
dx

:=
x :=

=
.. ..
dt
. .

dxn
x n
dt

2.4. VECTOR REPRESENTATION OF DYNAMICAL SYSTEMS

2.4

25

Vector representation of dynamical systems

Recall the general description of a dynamical system given in Section 2.2.1. For a system
with n state variables, x1 , , xn , we define the state (vector) x to be the n-vector given by

x1
x2

x := .. .
.
xn

We also introduce the n-vector valued function f defined by

f1 (x1 , x2 , . . . , xn )
f2 (x1 , x2 , . . . , xn )

f (x) :=
.
..

.
fn (x1 , x2 , . . . , xn )

Continuous time The general representation of a continuous-time dynamical system can

now be compactly described by the following first order vector differential equation:
x = f (x)

(2.1)

where x(t) is an n-vector and the scalar t is the time variable. A system described by
the above equation is called autonomous or time-invariant because the right-hand side of the
equation does not depend explicitly on time t. For the first part of the course, we will only
concern ourselves with these systems.
However, one can have a system containing time-varying parameters or disturbance
inputs which are time-varying. In this case the system might be described by
x = f (t, x)
that is, the right-hand side of the differential depends explicitly on time. Such a system is
called non-autonomous or time-varying. We will look at them later.
Discrete time The general representation of a discrete-time dynamical system can now
be compactly described by the following first order vector differential equation:
x(k+1) = f (x(k))

(2.2)

where x(k) is an n-vector and the scalar k is the time variable. A system described by
the above equation is called autonomous or time-invariant because the right-hand side of the
equation does not depend explicitly on time k. For the first part of the course, we will only
concern ourselves with these systems.
However, one can have a system containing time-varying parameters or disturbance
inputs which are time-varying. In this case the system might be described by
x(k+1) = f (k, x(k))

26

CHAPTER 2. STATE SPACE REPRESENTATION OF DYNAMICAL SYSTEMS

that is, the right-hand side of the differential depends explicitly on time. Such a system is
called non-autonomous or time-varying. We will look at them later.

2.5. SOLUTIONS AND EQUILIBRIUM STATES

2.5
2.5.1

27

Solutions and equilibrium states

Continuous-time

A solution of (2.1) is any continuous function x() which is defined on some time interval and
which satisfies
x(t)

= f (x(t))
for all t in the time interval.
An equilibrium solution is the simplest type of solution; it is constant for all time, that is,
it satisfies
x(t) xe
for some fixed state vector xe . The state xe is called an equilibrium state. Since an equilibrium
solution must satisfy the above differential equation, all equilibrium states are given by:
f (xe ) = 0
or, in scalar terms,
f1 (xe1 , xe2 , . . . , xen ) = 0
f2 (xe1 , xe2 , . . . , xen ) = 0
..
.
fn (xe1 , xe2 , . . . , xen ) = 0
Sometimes an equilibrium state is referred to as an equilibrium point, a stationary point, or,
a singular point.
Example 6 (Planar pendulum) Here, all equilibrium states are given by


m
e
x =
0
where m is an arbitrary integer. Physically, there are only two distinct equilibrium states
 
 
0

e
e
x =
and
x =
0
0
An isolated equilibrium state is an equilibrium state with the property that there is a
neighborhood of that state which contains no other equilibrium states.

Figure 2.10: Isolated equilibrium state

28

CHAPTER 2. STATE SPACE REPRESENTATION OF DYNAMICAL SYSTEMS

Example 7 The system

x = x x3
has three equilibrium states: 0, 1, 1. Clearly, these are isolated equilibrium states.
Example 8 Every state of the system
x = 0
is an equilibrium state. Hence none of these equilibrium states are isolated.
Example 9 Consider the system described by

sin(1/x) if x 6= 0
x =
0
if x = 0

The origin is an equilibrium state, but, it is not isolated. All other equilibrium states are
isolated.
Higher order ODEs
Consider a system described by an ordinary differential equation of the form
F (y, y,
. . . , y (n) ) = 0
where y(t) is a scalar. An equilibrium solution is a solution y() which is constant for all time,
that is,
y(t) y e

for some constant y e . Hence,

F (y e, 0, . . . , 0) = 0

(2.3)

For the state space description of this system introduced earlier, all equilibrium states
are given by

ye
0

xe = ..
.
0

where y e solves (2.3).

2.5. SOLUTIONS AND EQUILIBRIUM STATES

29

Multiple higher order ODEs Equilibrium solutions

yi (t) yie ,
Hence

i = 1, 2, . . . , m

e
F1 (y1e , 0, . . . , y2e , 0, . . . , . . . , yN
,...,0 )

= 0

e
F2 (y1e , 0, . . . , y2e , 0, . . . , . . . , yN
,...,0 )

= 0

(2.4)

..
.
e
FN (y1e , 0, . . . , y2e , 0, . . . , . . . , yN
,...,0 ) = 0
For the state space description of this system
are given by
e
y1
0

..
.
e
y2

e
x = ..
.
.
..

ye
N
..
.
0
e
where y1e , y2e , . . . , yN
solve (2.4).

introduced earlier, all equilibrium states

Example 10 Central force motion in inverse square gravitational field

r r 2 + /r 2 = 0
r + 2r

=0
Equilibrium solutions
r(t) r e ,

Hence,

(t) e

r,
r, = 0
This yields
r e ( e )2 + /(r e )2 = 0
0
=0
Thus there are infinite number of equilibrium solutions given by:
p
e = /(r e )3

where r e is arbitrary. Note that, for this state space description, an equilibrium state corresponds to a circular orbit.

2.5.2

Discrete-time

30

CHAPTER 2. STATE SPACE REPRESENTATION OF DYNAMICAL SYSTEMS

2.6
2.6.1

Numerical simulation
MATLAB

>> help ode45

ODE45 Solve non-stiff differential equations, medium order method.
[T,Y] = ODE45(F,TSPAN,Y0) with TSPAN = [T0 TFINAL] integrates the
system of differential equations y = F(t,y) from time T0 to TFINAL with
initial conditions Y0. F is a string containing the name of an ODE
file. Function F(T,Y) must return a column vector. Each row in
solution array Y corresponds to a time returned in column vector T. To
obtain solutions at specific times T0, T1, ..., TFINAL (all increasing
or all decreasing), use TSPAN = [T0 T1 ... TFINAL].
[T,Y] = ODE45(F,TSPAN,Y0,OPTIONS) solves as above with default
integration parameters replaced by values in OPTIONS, an argument
created with the ODESET function. See ODESET for details. Commonly
used options are scalar relative error tolerance RelTol (1e-3 by
default) and vector of absolute error tolerances AbsTol (all
components 1e-6 by default).
[T,Y] = ODE45(F,TSPAN,Y0,OPTIONS,P1,P2,...) passes the additional
parameters P1,P2,... to the ODE file as F(T,Y,FLAG,P1,P2,...) (see
ODEFILE). Use OPTIONS = [] as a place holder if no options are set.
It is possible to specify TSPAN, Y0 and OPTIONS in the ODE file (see
ODEFILE). If TSPAN or Y0 is empty, then ODE45 calls the ODE file
[TSPAN,Y0,OPTIONS] = F([],[],init) to obtain any values not supplied
in the ODE45 argument list. Empty arguments at the end of the call list
may be omitted, e.g. ODE45(F).
As an example, the commands
options = odeset(RelTol,1e-4,AbsTol,[1e-4 1e-4 1e-5]);
ode45(rigidode,[0 12],[0 1 1],options);
solve the system y = rigidode(t,y) with relative error tolerance 1e-4
and absolute tolerances of 1e-4 for the first two components and 1e-5
for the third. When called with no output arguments, as in this
example, ODE45 calls the default output function ODEPLOT to plot the
solution as it is computed.
[T,Y,TE,YE,IE] = ODE45(F,TSPAN,Y0,OPTIONS) with the Events property in
OPTIONS set to on, solves as above while also locating zero crossings

2.6. NUMERICAL SIMULATION

31

of an event function defined in the ODE file. The ODE file must be
coded so that F(T,Y,events) returns appropriate information. See
ODEFILE for details. Output TE is a column vector of times at which
events occur, rows of YE are the corresponding solutions, and indices in
vector IE specify which event occurred.
See also ODEFILE and
other ODE solvers:
options handling:
output functions:
odefile examples:
>>

2.6.2

SIMULINK

ODE23, ODE113, ODE15S, ODE23S, ODE23T, ODE23TB

ODESET, ODEGET
ODEPLOT, ODEPHAS2, ODEPHAS3, ODEPRINT
ORBITODE, ORBT2ODE, RIGIDODE, VDPODE

32

CHAPTER 2. STATE SPACE REPRESENTATION OF DYNAMICAL SYSTEMS

Exercises
Exercise 1 By appropriate definition of state variables, obtain a first order state space
description of the following systems where y1 and y2 are real scalars.
(i)
2
y1 + y2 + sin y1 = 0
y1 + 2
y2 + sin y2 = 0
(ii)
y1 + y 2 + y13 = 0
y 1 + y 2 + y23 = 0

Exercise 2 By appropriate definition of state variables, obtain a first order state space
description of the following system where q1 and q2 are real scalars.
q1 (k+2) + q1 (k) + 2q2 (k+1) = 0
q1 (k+2) + q1 (k+1) + q2 (k) = 0
Exercise 3 Consider the Lorenz system described by
x 1 = (x2 x1 )
x 2 = rx1 x2 x1 x3
x 3 = bx3 + x1 x2
with = 10, b = 38 , and r = 28. Simulate this system with initial states

0
0
1
1 + eps
and
0
0

where eps is the floating point relative accuracy in MATLAB. Comment on your results for
the integration interval [0 60].

Chapter 3
First order systems
The simplest type of nonlinear systems is one whose state can be described by a single scalar.
We refer to such a system as a first order system.

3.1

Continuous time

A first order continuous-time system is described by

x = f (x)

(3.1)

where x(t) is a scalar. As an example, recall the first nonlinear system.

State portrait. As illustrated in Figure 3.1, one may readily determine the qualitative
behavior of a first order system from the graph of f .

Figure 3.1: Qualitative behavior of a first order system

Example 11 Consider the first order system described by

x = x x3
This system has three equilibrium states: 1, 0 and 1.
33

34

CHAPTER 3. FIRST ORDER SYSTEMS

Figure 3.2: State portrait for x = x x3

Linear systems. A linear time-invariant first order system is described by
x = ax

(3.2)

where a is a constant real scalar. All solutions of this system are of the form
x(t) = ceat
where c is an a constant real scalar. Thus the qualitative behavior of (3.2) is completely
determined by the sign of a.
Linearization. Suppose xe is an equilibrium state for system (3.1). Then
f (xe ) = 0 .
Suppose that the function f is differentable at xe with derivative f (xe ). When x is close
to xe , it follows from the definition of the derivative that
f (x) f (xe ) + f (xe )(x xe ) = f (xe )(x xe ) .
If we introduce the perturbed state,
x = x xe ,
then
x = f (x) f (xe )x .

So, the linearization of system (3.1) about an equilibrium state xe is defined to be the
following system:
x = ax
where
a = f (xe ) .
(3.3)
One can demonstrate the following result.
If f (xe ) 6= 0, then the local behavior of the nonlinear system (3.1) about xe is qualitatively
the same as that of the linearized system about the origin.
Example 12
x = x x3
Example 13
x = ax2

3.2. SOLUTIONS TO SCALAR FIRST ORDER ODES

3.2

35

Solutions to scalar first order ODES

Consider
x = f (x)
with initial condition x(0) = x0 . Suppose that f (x0 ) 6= 0; then f (x(t)) 6= 0 for some interval
[0, t1 ) and over this interval we have
1 dx
= 1.
f (x(t)) dt
Let
g(x) :=

x0

Then

1
d
f ()

d
d
dx
1 dx
(g(x(t))) =
(g(x))
=
= 1,
dt
dx
dt
f (x) dt

that is,
d
(g(x(t))) = 1 .
dt
Integrating over the interval [0, t] and using the initial condition x(0) = x0 yields g(x(t)) = t,
that is,
Z x(t)
1
d = t .
(3.4)
f ()
x0
One now solves the above equation for x(t).
Example 14 (Finite escape time) This simple example illustrates the concept of a finite
escape time.
x = x2
Here,
Z

x0

1
1
1
d
=

+
.
2
x x0

Hence, all nonzero solutions satisfy

1
1
+
= t.
x(t) x0

Then, provided x0 t 1, the above equation can be uniquely solved for x(t) to obtain
x(t) =

x0
.
1 x0 t

(3.5)

Suppose x0 > 0. Then x(t) goes to infinity as t approaches 1/x0 ; the solution blows up in a
finite time. This cannot happen in a linear system.

36

3.3

CHAPTER 3. FIRST ORDER SYSTEMS

Discrete time

A first order continuous-time system is described by

x(k+1) = f (x(k))
Example 15
x(k+1) = x(k)(1 x(k))
. Consider > 3.56994

Exercises
Exercise 4 Draw the state portrait of the first nonlinear system.
Exercise 5 Draw the state portrait for
x = x4 x2 .
Exercise 6 Obtain an explicit expression for all solutions of
x = x3 .

Chapter 4
Second order systems
In this section, we consider systems whose state can be described by two real scalar variables
x1 , x2 . We will refer to such systems as second order systems. They are described by
x 1 = f1 (x1 , x2 )
x 2 = f2 (x1 , x2 )

(4.1)

State plane, state portrait.

Vector field f
Quite often we consider second order systems described by
x 1 = x2
x 2 = f2 (x1 , x2 )
In this case, the state plane and the state portrait are sometimes referred to as the phase
plane and phase portrait, respectively.

4.1

Linear systems

A linear time-invariant second order system is described by

x 1 = a11 x1 + a12 x2
x 2 = a21 x1 + a22 x2
where each aij is a constant real scalar. The qualitative behavior of this system is determined
by the eigenvalues 1 , 2 of the matrix
A=

a11 a12
a21 a22

37

38

CHAPTER 4. SECOND ORDER SYSTEMS

stable node
unstable node
saddle
stable focus
unstable focus
center
See Khalil.

real
real
real
complex
complex
complex

1 , 2 < 0
0 < 1 , 2
1 < 0 < 2
(1 ) < 0
(1 ) > 0
(1 ) = 0

4.2. LINEARIZATION

4.2

39

Linearization

Suppose xe = (xe1 , xe2 ) is an equilibrium state of system (4.1). Then the linearization of (4.1)
about xe is given by:
x 1 = a11 x1 + a12 x2
x 2 = a21 x1 + a22 x2
where
aij =

(4.2a)
(4.2b)

fi e e
(x , x )
xj 1 2

for i, j = 1, 2.
The behavior of a second order nonlinear system near an equilibrium state is qualitatively the
same as the behavior of the corresponding linearized system about 0, provided the eigenvalues
of the linearized system have non-zero real part.
Example 16 Van der Pol oscillator. Consider
w + (w)
+w =0
where is a nonlinear function. If one lets y = w and differentiates the above equation, one
obtains
y + (y)y + y = 0 .
Considering (y) = (y 3/3 y) yields
y (1 y 2)y + y = 0
Introducing state variables x1 = y and x2 = y,
this system can be described by
x 1 = x2
x 2 = x1 + (1 x21 )x2
This system has a single equilibrium state at the origin. Linearization about the origin
results in
x 1 = x2
x 2 = x1 + x2 .
The eigenvalues 1 , 2 of this linearized system are given by
r
r

2
2
and
2 = + 1
.
1 = 1
2
4
2
4
If we consider 0 < < 2, then 1 and 2 are complex numbers with positive real parts.
Hence, the origin is an unstable focus of the linearized system. Thus, the origin is an unstable
focus of the the original nonlinear system. Figure 4.1 contains some state trajectories of this
system for = 0.3.
Note that, although the system has a single equilibrium state and this is unstable, all
solutions of the system are bounded. This behavior does not occur in linear systems. If the
origin of a linear system is unstable, then the system will have some unbounded solutions.

40

CHAPTER 4. SECOND ORDER SYSTEMS

Phase portrait of Van der Pol oscillator
3

ydot

3
3

Figure 4.1: Van der Pol oscillator

Example 17 Duffings equation Consider
y + (y) = 0
where is a nonlinear function. With (y) = y + y 3 , we have
y y + y 3 = 0
The above differential equation has three equilibrium solutions:
y e = 0, 1, 1 .
Linearization about the zero solution yields

y y = 0 .
A state space description of this linearized system is given by
x 1 = x2
x 2 = x1
The eigenvalues 1 , 2 and corresponding eigenvectors v1 , v2 associated with this linear system are given by
 


1
1
1 = 1 ,
v1 =
,
2 = 1 ,
v1 =
.
1
1
So the origin of the linearized system is a saddle; see Figure 4.2. Hence the origin is a saddle
for the original nonlinear system.
Linearization about either of the two non-zero equilibrium solutions results in

y + 2y = 0 .
The eigenvalues associated with any state space realization of this system are

1 = 2 .
1 = 2

4.2. LINEARIZATION

41

Figure 4.2: Duffing saddle

Phase portrait of ddot y = y y3
3

ydot

3
3

0
y

10

15

Phase portrait of ddot y = y y^3

15

10

ydot

10

15
15

10

0
y

Figure 4.3: Phase portraits of Duffing system

42

4.3

CHAPTER 4. SECOND ORDER SYSTEMS

Isocline method

See Khalil.

4.4

Periodic solutions and limit cycles

A solution x() of x = f (x) is a periodic solution if it is non-constant and there is a T > 0

such that
x(t + T ) = x(t)
for all t
The smallest T for which the above identity holds is called the period of the solution. Periodic
solutions are sometimes called limit cycles.
Note that a limit cycle corresponds to a closed trajectory in the state plane. If a linear
system has a periodic solution with period T , it will have an infinite number of periodic
solutions with that period. However, an nonlinear system can have just a single periodic
solution; recall the Van der Pol oscillator. Also, a nonlinear system may have an infinite
number of periodic solutions, but, the period may be different for different amplitudes; this
occurs for Duffings equation.

Figure 4.4: limit cycle

In the analysis of second order systems, it is sometimes useful to introduce polar coordinates r, which are implicity defined by
x1 = r cos

and

Note that

x2 = r sin .
1

r = (x21 + x22 ) 2
and, when (x1 , x2 ) 6= (0, 0),

= tan1 (x2 /x1 )

The scalar is not defined when (x1 , x2 ) = (0, 0).

Figure 4.5: Polar coordinates

4.4. PERIODIC SOLUTIONS AND LIMIT CYCLES

43

One may show that

r = (x1 x 1 + x2 x 2 )/r
= (x1 x 2 x 1 x2 )/r 2

(4.3a)
(4.3b)

Example 18 The simple harmonic oscillator.

x 1 = x2
x 2 = x1
Here
r = 0
= 1

Figure 4.6: Simple harmonic oscillator

Example 19 An isolated periodic solution
x 1 = x2 + x1 (1 x21 x22 )
x 2 = x1 + x2 (1 x21 x22 )
Here
r = r(1 r)2
= 1

Exercises
Exercise 7 Determine the nature of each equilibrium state of the damped duffing system
y + 0.1y y + y 3 = 0
Numerically obtain the phase portrait.

44

CHAPTER 4. SECOND ORDER SYSTEMS

Exercise 8 Determine the nature (if possible) of each equilibrium state of the simple pendulum.
y + sin y = 0
Numerically obtain the phase portrait.
Exercise 9 Numerically obtain a state portrait of the following system:
x 1 = x22
x 2 = x1 x2
Based on the state portrait, predict the stability properties of each equilibrium state.

Chapter 5
Some general considerations
Consider the differential equation
x = f (x)

(5.1)

where x is an n-vector. By a solution of (5.1) we mean any continuous function x() : [0, t1 )
IRn , with t1 > 0, which satisfies (5.1). Consider an initial condition
x(0) = x0 ,

(5.2)

where x0 is some specified initial state. The corresponding initial value problem associated
with (5.1) is that of finding a solution to (5.1) which satisfies the initial condition (5.2). If
the system is linear then, for every initial state x0 , there is a solution to the corresponding
initial value problem, this solution is unique and is defined for all t 0, that is, t1 = .
One cannot make the same statement for nonlinear systems in general.

5.1

Existence of solutions

Example 20 Example of nonexistence

x = s(x) ,
where
s(x) =

x(0) = 0

1 if x 0
1 if x < 0

Fact. Continuity of f guarantees existence.

The above function is not continuous at 0.

5.2

Uniqueness of solutions

Example 21 Example of nonuniqueness: Im falling

p
x = 2g|x| ,
x(0) = 0
45

46

CHAPTER 5. SOME GENERAL CONSIDERATIONS

The above initial value problem has an infinite number of solutions. They are given by

0
if 0 t T
x(t) =
g
2
(t

T
)
if
T t
2
where T 0 is arbitrary.
Fact.Differentiability of f guarantees uniqueness.
The f above is not differentiable at 0.

5.3

Indefinite extension of solutions

Example 22 Finite escape time

x = 1 + x2 ,

x(0) = 0

Fact. If a solution cannot be extended indefinitely, that is over [0, ), then it must have
a finite escape time Te , that is, Te is finite and
lim x(t) = .

tTe

Hence, if a solution is bounded on bounded time intervals then, it cannot have a finite
escape time and, hence can be extended over [0, ).
The following condition guarantees that solutions can be extended indefinitely. There
are constants and such that
kf (x)k kxk +
for all x. Here one can show that when 6= 0
kx(t)k e(tt0 ) kx(t0 )k +

t
(e 1)

When = 0,
kx(t)k kx(t0 )k + (t t0 )

Chapter 6
Stability and boundedness
Consider a general nonlinear system described by
x = f (x)

(6.1)

where x(t) is a real n-vector and t is a real scalar. By a solution of (6.1) we mean any
continuous function x() : [0, t1 ) IRn with t1 > 0, which satisfies x(t)

= f (x(t)) for
0 t < t1 .

6.1

Boundedness of solutions

DEFN. (Boundedness) A solution x() is bounded if there exists 0 such that

||x(t)||

for all

t0

A solution is unbounded if it is not bounded.

It should be clear from the above definitions that a solution x() is bounded if and only
if each component xi () of x() is bounded. Also, a solution is unbounded if and only if
at least one of its components is unbounded. So, x(t) = [ et e2t ]T is bounded while
x(t) = [ et et ]T is unbounded.
Example 23 All solutions of
x = 0
are bounded.
Example 24 All solutions of
x = x x3
are bounded.
47

48

CHAPTER 6. STABILITY AND BOUNDEDNESS

Example 25 All solutions (except the zero solution) of

x = x
are unbounded.
Example 26 Consider
x = x2 ,

x(0) = x0

If x0 > 0, the corresponding solution has a finite escape time and is unbounded. If x0 < 0,
the corresponding solution is bounded.
Example 27 Undamped oscillator.
x 1 = x2
x 2 = x1
All solutions are bounded.

Boundedness and linear time-invariant systems. Consider a general LTI (linear timeinvariant) system
x = Ax
(6.2)
Recall that every solution of this system has the form
x(t) =

l n
i 1
X
X

tj ei t v ij

i=1 j=0

where 1 , . . . , l are the eigenvalues of A, ni is the index of i , and the constant vectors v ij
depend on initial state.
We say that an eigenvalue of A is non-defective if its index is one; this means that the
algebraic multiplicity and the geometric multiplicity of are the same. Otherwise we say
is defective.
Hence we conclude that all solutions of (6.2) are bounded if and only if for each eigenvalue
i of A:
(b1) (i ) 0 and
(b2) if (i ) = 0 then i is non-defective.
If there is an eigenvalue i of A such that either
(u1) (i ) > 0 or
(u2) (i ) = 0 and i is defective
then, the system has some unbounded solutions.

6.1. BOUNDEDNESS OF SOLUTIONS

49

Example 28 Unattached mass

x 1 = x2
x 2 = 0
Here
A=

0 1
0 0

has a single eigenvalue 0. This eigenvalue has algebraic multiplicity 2 but geometric multiplicity 1; hence some of the solutions of the system are unbounded. One example is
 
t
x(t) =
1
Example 29
x 1 = 0
x 2 = 0
Here
A=

0 0
0 0

has a single eigenvalue 0. This eigenvalue has algebraic multiplicity 2 and geometric multiplicity 2. Hence all the solutions of the system are bounded. Actually every state is an
equilibrium state and every solution is constant.
Example 30 (Resonance) Consider a simple linear oscillator subject to a sinusoidal input
of amplitude W :
q + q = W sin(t+)
Resonance occurs when = 1. To see this, let
x1 := q ,

x2 := q ,

x3 := W sin(t+) ,

x4 := W cos(t+)

to yield
x = Ax
where

0
1
A=
0
0

1 0
0 1
0 0
0 2

0
0

1
0

If = 1 then, A has eigenvalues and . These eigenvalues have algebraic multiplicity two
but geometric multiplicity one; hence the system has unbounded solutions.

50

6.2

CHAPTER 6. STABILITY AND BOUNDEDNESS

Stability of equilibrium states

Suppose xe is an equilibrium state of the system x = f (x). Then, whenever x(0) = xe , we

have (assuming uniqueness of solutions) x(t) = xe for all t 0. Roughly speaking, we say
that xe is a stable equilibrium state for the system if the following holds. If the initial state
of the system is close to xe , then the resulting solution is close to xe . The formal definition
is as follows.
DEFN. (Stability) An equilibrium state xe is stable if for each > 0 there exists > 0 such
that whenever ||x(0)xe || < one has ||x(t)xe || < for all t 0.

Figure 6.1: Stability

An equilibrium state xe is said to be unstable if it is not stable.
Example 31
x = 0
Every equilibrium state is stable. (Choose = .)
Example 32
x = x
The origin is unstable.
Example 33
x = x x3
The origin is unstable; the remaining equilibrium states 1 and 1 are stable.
Example 34 Undamped oscillator
The origin is stable.

6.2. STABILITY OF EQUILIBRIUM STATES

51

Example 35 Simple pendulum

x 1 = x2
x 2 = sin x1
(0, 0) is stable; (, 0) is unstable.
Example 36 Van der Pol oscillator
x 1 = x2
x 2 = (1 x21 )x2 x1

Figure 6.2: Van der Pol oscillator

The origin is unstable. However, all solutions are bounded.

Stability of linear time-invariant systems. It can be shown that every equilibrium

state of a LTI system (6.2) is stable if and only if all eigenvalues i of A satisfy conditions
(b1) and (b2) above. Hence every equilibrium state of a LTI system is stable if and only if
all solutions are bounded.
It can also be shown that every equilibrium state is unstable if and only if there is an
eigenvalue i of A which satisfies condition (u1) or (u2) above. Hence every equilibrium
state of a LTI system is unstable if and only if there are unbounded solutions.

52

6.3
6.3.1

CHAPTER 6. STABILITY AND BOUNDEDNESS

Asymptotic stability
Global asymptotic stability

DEFN. (Global asymptotic stability) An equilibrium state xe is globally asymptotically stable

(GAS) if
(a) It is stable
(b) Every solution x() converges to xe with increasing time, that is,
lim x(t) = xe

If xe is a globally asymptotically stable equilibrium state, then there are no other equilibrium states and all solutions are bounded. In this case we say that the system x = f (x)
is globally asymptotically stable.
Example 37 The system
x = x
is GAS.
Example 38 The system
x = x3
is GAS.

6.3.2

Asymptotic stability

In asymptotic stability, we do not require that all solutions converge to the equilibrium state;
we only require that all solutions which originate in some neighborhood of the equilibrium
state converge to the equilibrium state.
DEFN. (Asymptotic stability) An equilibrium state xe is asymptotically stable (AS) if
(a) It is stable.
(b) There exists R > 0 such that whenever ||x(0)xe || < R one has
lim x(t) = xe .

(6.3)

The region of attraction of an equilibrium state xe which is AS is the set of initial states
which result in (13.3), that is it is the set of initial states which are attracted to xe . Thus, the
region of attraction of a globally asymptotically equilibrium state is the whole state space.

6.3. ASYMPTOTIC STABILITY

53

Figure 6.3: Asymptotic stability

Example 39
x = x

Example 40

x = x x3

The equilibrium states 1 and 1 are AS with regions of attraction (, 0) and (0, ),
respectively.
Example 41 Damped simple pendulum

The zero state is AS but not GAS.

x 1 = x2
x 2 = sin x1 x2

Example 42 Reverse Van der Pol oscillator

x 1 = x2
x 2 = (1 x21 )x2 + x1

The zero state is AS but not GAS. Also, the system has unbounded solutions.
LTI systems. For LTI systems, it should be clear from the general form of the solution
that the zero state is AS if and only if all the eigenvalues i of A have negative real parts,
that is,
(i ) < 0
Also AS is equivalent to GAS.
Example 43 The system

is GAS.

x 1 = x1 + x2
x 2 = x2

Example 44 The system

is GAS.

x 1 = x1 + x2
x 2 = x1 x2

54

CHAPTER 6. STABILITY AND BOUNDEDNESS

Figure 6.4: Reverse Van der Pol oscillator

6.4

Exponential stability

We now present the strongest form of stability considered in this section.

DEFN. (Global exponential stability) An equilibrium state xe is globally exponentially stable
(GES) with rate of convergence > 0 if there exists > 0 such that every solution satisfies
||x(t)xe ||

||x(0)xe|| exp(t)

for all t 0

Example 45
x = x
GES with rate = 1.
Note that global exponential stability implies global asymptotic stability, but, in general,
the converse is not true. This is illustrated in the next example. For linear time-invariant
systems, GAS and GES are equivalent.
Example 46
x = x3
Solutions satisfy

GAS but not GES

x(t) = p

x0
1 + 2x20 t

where

x0 = x(0) .

Example 47 (General Example) Consider a general system

x = f (x)

6.4. EXPONENTIAL STABILITY

55

and suppose that there exists > 0 such that

x f (x) kxk2
for all x. Then this system is GES about zero with rate of convergence .

As a specific example, consider

x = 2x 7x3
As another specific example, consider

x 1 = x1 + cos(x1 )x2
x 2 = 2x2 cos(x1 )x1

DEFN. (Exponential stability) An equilibrium state xe is exponentially stable (ES) with rate of
convergence > 0 if there exists R > 0 and > 0 such that whenever ||x(0)xe || < R one
has
||x(t)xe || ||x(0)xe|| exp(t)
for all t 0
Note that exponential stability implies asymptotic stability, but, in general, the converse is
not true.
Example 48
x = x3
GAS but not even ES
Example 49
x =
GAS, ES, but not GES

x
1 + x2

56

CHAPTER 6. STABILITY AND BOUNDEDNESS

Figure 6.5: Exponential stability

6.5

LTI systems

Consider a LTI system

x = Ax

(6.4)

Recall that every solution of this system has the form

x(t) =

l n
i 1
X
X

tj ei t v ij

i=1 j=0

where 1 , . . . , l are the eigenvalues of A, the integer ni is the index of i , and the constant
vectors v ij depend on initial state. From this it follows that the stability properties of this
system are completely determined by the location of its eigenvalues; this is summarized in
the table below.
The following table summarizes the relationhip between the stability properties of a LTI
system and the eigenproperties of its A-matrix. In the table, unless otherwise stated, a
property involving must hold for all eigenvalues of A.
Stability properties

Eigenproperties

Global exponential stability

and boundedness

() < 0

Stability and
boundedness

() 0
If () = 0 then is non-defective.

Instability and some

unbounded solutions

There is an eigenvalue with () > 0 or

() = 0 and is defective.

Example 50 The system

x 1 = x1 + x2
x 2 = x2

6.5. LTI SYSTEMS

57

is GES.
Example 51 The system
x 1 = x1 + x2
x 2 = x1 x2
is GES.
Example 52 The system
x 1 = 0
x 2 = 0
is stable about every equilibrium point.
Example 53 The system
x 1 = x2
x 2 = 0
is unstable about every equilibrium point.

Figure 6.6: The big picture. The concepts in each dashed box are equivalent for linear
systems.

58

6.6

CHAPTER 6. STABILITY AND BOUNDEDNESS

Linearization and stability

Consider a nonlinear time-invariant system described by

x = f (x)
where x(t) is an n-vector at each time t. Suppose xe is an equilibrium state for this system,
that is, f (xe ) = 0, and consider the linearization of this system about xe :
x = Ax

where A =

f e
(x ) .
x

The following results can be demonstrated using nonlinear Lyapunov stability theory.
Stability. If all the eigenvalues of the A matrix of the linearized system have negative real
parts, then the nonlinear system is exponentially stable about xe .
Instability. If at least one eigenvalue of the A matrix of the linearized system has a positive
real part, then the nonlinear system is unstable about xe .
Undetermined. Suppose all the eigenvalues of the A matrix of the linearized system have
non-positive real parts and at least one eigenvalue of A has zero real part. Then, based on
the linearized system alone, one cannot predict the stability properties of the nonlinear system
about xe .
Note that the first statement above is equivalent to the following statement. If the
linearized system is exponentially stable, then the nonlinear system is exponentially stable
about xe .
Example 54 (Damped simple pendulum.) Physically, the system
x 1 = x2
x 2 = sin x1 x2
has two distinct equilibrium states: (0, 0) and (, 0). The A matrix for the linearization of
this system about (0, 0) is


0
1
A=
1 1

Since all the eigenvalues of this matrix have negative real parts, the nonlinear system is
exponentially stable about (0, 0). The A matrix corresponding to linearization about (, 0)
is


0
1
A=
1 1
Since this matrix has an eigenvalue with a positive real part, the nonlinear system is unstable
about (, 0)

The following example illustrates the fact that if the eigenvalues of the A matrix have nonpositive real parts and there is at least one eigenvalue with zero real part, then, one cannot
make any conclusions on the stability of the nonlinear system based on the linearization.

6.6. LINEARIZATION AND STABILITY

59

Example 55 Consider the scalar nonlinear system:

x = ax3
This origin is GAS if a < 0, unstable if a > 0 and stable if a = 0. However, the linearization
of this system about zero, given by
x = 0
is independent of a and is stable.
The following example illustrates that instability of the linearized system does not imply
instability of the nonlinear system.
Example 56 Using nonlinear techniques, one can show that the following system is GAS
about the origin.
x 1 = x2
x 2 = x31 x32
However, the linearization of this system, given by
x 1 = x2
x 2 = 0 ,
is unstable about the origin.

60

CHAPTER 6. STABILITY AND BOUNDEDNESS

Exercises
Exercise 10 For each of the following systems, determine (from the state portrait) the
stability properties of each equilibrium state. For AS equilibrium states, give the region of
attraction.
(a)
x = x x3
(b)
x = x + x3
(c)
x = x 2x2 + x3
Exercise 11 If possible, use linearization to determine the stability properties of each of
the following systems about the zero equilibrium state.
(i)
x 1 = (1 + x21 )x2
x 2 = x31
(ii)
x 1 = sin x2
x 2 = (cos x1 )x3
x 3 = ex1 x2
Exercise 12 If possible, use linearization to determine the stability properties of each equilibrium state of the Lorenz system.

Chapter 7
Stability and boundedness: discrete
time
7.1

Boundedness of solutions

Consider a general discrete-time nonlinear system described by

x(k+1) = f (x(k))

(7.1)

where x(k) is an n-vector and k is an integer. By a solution of (7.1) we mean a sequence

x() = ( x(0), x(1), x(2), . . . ) which satisfies (7.1) for all k 0.
DEFN. (Boundedness of solutions) A solution x() is bounded if there exists 0 such that
||x(k)||

for all

k 0.

A solution is unbounded if it is not bounded.

It should be clear from the above definitions that a solution x() is bounded if and only
if each component xi () of x() is bounded. Also, a solution is unbounded if and only if at
least one of its components is unbounded. So, x(k) = [ (0.5)k (0.5)k ]T is bounded while
x(k) = [ (0.5)k 2k ]T is unbounded.
Example 57
x(k+1) = 0
All solutions are bounded.
Example 58
x(k+1) = x(k)
All solutions are bounded.
Example 59
x(k+1) = 2x(k)
All solutions (except the zero solution) are unbounded.

61

62

CHAPTER 7. STABILITY AND BOUNDEDNESS: DISCRETE TIME

Linear time invariant (LTI) systems. All solutions of the LTI system
x(k+1) = Ax(k)

(7.2)

are bounded if and only if for each eigenvalue of A:

(b1) || 1 and
(b2) if || = 1 then is non-defective.
If there is an eigenvalue of A such that either
(u1) || > 1 or
(u2) || = 1 and is defective.
then the system has some unbounded solutions.
Example 60 Discrete unattached mass. Here


1 1
A=
0 1
has a single eigenvalue 1. This eigenvalue has algebraic multiplicity 2 but geometric multiplicity 1; hence this eigenvalue is defective. So, some of the solutions of the system
x(k+1) = Ax(k) are unbounded. One example is
 
k
x(k) =
1

7.2

Stability of equilibrium states

Suppose xe is an equilibrium state of the system x(k + 1) = f (x(k)). Then, whenever

x(0) = xe , we have x(k) = xe for all k 0. Roughly speaking, we say that xe is a stable
equilibrium state for the system if the following holds. If the initial state of the system is
close to xe , then the resulting solution is close to xe . The formal definition is as follows.
DEFN. (Stability) An equilibrium state xe is stable if for each > 0 there exists > 0 such
that whenever ||x(0)xe || < , one has ||x(k)xe || < for all k 0.
xe is unstable if it is not stable.
Example 61
x(k+1) = x(k) .
The origin is stable. (Consider = .)

7.3. ASYMPTOTIC STABILITY

63

Example 62
x(k+1) = 2x(k) .
The origin is unstable.
Example 63
x(k+1) = x(k)3 .
The single equilibrium at the origin is stable, but, the system has unbounded solutions.

LTI systems. Every equilibrium state of a LTI system (7.2) is stable if and only if all
eigenvalues of A satisfy conditions (b1) and (b2) above. Hence every equilibrium state of
a LTI system is stable if and only if all solutions are bounded.
Every equilibrium state is unstable if and only if there is an eigenvalue of A which
satisfies condition (u1) or (u2) above. Hence every equilibrium state of a LTI system is
unstable if and only if there are unbounded solutions.

7.3
7.3.1

Asymptotic stability
Global asymptotic stability

DEFN. (Global asymptotic stability (GAS)) An equilibrium state xe is globally asymptotically

stable (GAS) if
(a) It is stable
(b) Every solution x() converges to xe , that is,
lim x(k) = xe .

If xe is a globally asymptotically stable equilibrium state, then there are no other equilibrium states. In this case we say the system (7.1) is globally asymptotically stable.
Example 64
1
x(k+1) = x(k)
2
GAS
Example 65
x(k+1) =

x(k)
2 + x(k)2

64

7.3.2

CHAPTER 7. STABILITY AND BOUNDEDNESS: DISCRETE TIME

Asymptotic stability

DEFN. (Asymptotic stability (AS)) An equilibrium state xe is asymptotically stable (AS) if

(a) It is stable
(b) There exists R > 0 such that whenever ||x(0) xe || < R one has
lim x(k) = xe

(7.3)

The region of attraction of an equilibrium state xe which is AS is the set of initial states
which result in (7.3), that is it is the set of initial states which are attracted to xe . Thus, the
region of attraction of a globally asymptotically equilibrium state is the whole state space.
Example 66
x(k+1) = x(k)3
Origin is AS with region of attraction [1 1].
LTI systems. For LTI systems, it should be clear from the general form of the solution
that the zero state is AS if and only if all the eigenvalues i of A have magnitude less than
one, that is,
|i | < 1 .
Also AS is equivalent to GAS.

7.4

Exponential stability

We now present the strongest form of stability considered in this section.

DEFN. (Global exponential stability) An equilibrium state xe is globally exponentially stable
(GES) with there exists 0 < 1 and > 0 such that every solution satisfies
||x(k)xe ||

k ||x(0)xe ||

Example 67
1
x(k+1) = x(k)
2
GES with = 21 .

for all k 0

7.4. EXPONENTIAL STABILITY

65

Note that global exponential stability implies global asymptotic stability, but, in general,
the converse is not true. This is illustrated in the next example. For linear time-invariant
systems, GAS and GES are equivalent.
Lemma 1 Consider a system described by x(k + 1) = f (x(k)) and suppose that for some
scalar 0,
kf (x)k kxk
for all x. Then, every solution x() of the system satisfies
kx(k)k k kx(0)k
for all k 0. In particular, if < 1 then, the system is globally exponentially stable.
Example 68
x(k+1) = p

Solutions satisfy
x(k) = p

GAS but not GES

x0
1 + 2kx20

x(k)
1 + 2x(k)2

where

x0 = x(0) .

DEFN. (Exponential stability) An equilibrium state xe is exponentially stable (ES) if there

exists R > 0, 0 < 1 and > 0 such that whenever ||x(0)xe || < R one has
||x(k)xe ||

k ||x(0)xe ||

for all k 1

Note that exponential stability implies asymptotic stability, but, in general, the converse
is not true.
Example 69
x(k+1) = p

Solutions satisfy

GAS but not even ES

x(k) = p

x0
1 + 2kx20

x(k)
1 + 2x(k)2

where

x0 = x(0) .

66

CHAPTER 7. STABILITY AND BOUNDEDNESS: DISCRETE TIME

7.5

LTI systems

The following table summarizes the relationhip between the stability properties of a LTI
system and the eigenproperties of its A-matrix. In the table, unless otherwise stated, a
property involving must hold for all eigenvalues of A.
Stability properties

eigenproperties

Asymptotic stability
and boundedness

|| < 1

Stability and
boundedness

|| 1
If || = 1 then is non-defective

Instability and some

unbounded solutions

There is an eigenvalue of A with

|| > 1 or
|| = 1 and is defective

7.6. LINEARIZATION AND STABILITY

7.6

67

Linearization and stability

Consider a nonlinear time-invariant system described by

x(k+1) = f (x(k))
where x(k) is an n-vector at each time t. Suppose xe is an equilibrium state for this system,
that is, f (xe ) = xe , and consider the linearization of this system about xe :
x(k+1) = Ax(k)

where A =

f e
(x ) .
x

The following results can be demonstrated using nonlinear Lyapunov stability theory.
Stability. If all the eigenvalues of the A matrix of the linearized system have magnitude less
than one, then the nonlinear system is exponentially stable about xe .
Instability. If at least one eigenvalue of the A matrix of the linearized system has magnitude
greater than one, then the nonlinear system is unstable about xe .
Undetermined. Suppose all the eigenvalues of the A matrix of the linearized system have
magnitude less than or equal to one and at least one eigenvalue of A has magnitude one.
Then, based on the linearized system alone, one cannot predict the stability properties of the
nonlinear system about xe .
Note that the first statement above is equivalent to the following statement. If the
linearized system is exponentially stable, then the nonlinear system is exponentially stable
about xe .
Example 70 (Newtons method) Recall that Newtons method for a scalar function can
be described by
g(x(k))
x(k+1) = x(k)
g (x(k))
Here
f (x) = x
So,
f (x) = 1 1 +

g(x)
g (x)

g(x)g (x)
g(x)g (x)
=
.
g (x)2
g (x)2

At an equilibrium state xe , we have g(xe ) = 0; hence f (xe ) = 0 and the linearization about
any equilibrium state is given by
x(k+1) = 0 .
Thus, every equilibrium state is exponentially stable.

68

CHAPTER 7. STABILITY AND BOUNDEDNESS: DISCRETE TIME

Chapter 8
Basic Lyapunov theory
Suppose we are interested in the stability properties of the system
x = f (x)

(8.1)

where x(t) is a real n-vector at time t. If the system is linear, we can determine its stability
properties from the properties of the eigenvalues of the system matrix. What do we do for
a nonlinear system? We could linearize about each equilibrium state and determine the
stability properties of the resulting linearizations. Under certain conditions this will tell us
something about the local stability properties of the nonlinear system about its equilibrium
states. However there are situations where linearization cannot be used to deduce even the
local stability properties of the nonlinear system. Also, linearization tells us nothing about
the global stability properties of the nonlinear system.
In general, we cannot explicitly obtain solutions for nonlinear systems. However, Lyapunov theory allows to say something about the stability properties of a system without
knowing the form or structure of its solutions. Lyapunov theory is based on Lyapunov functions which are scalar-valued functions of the system state.
Suppose V is a scalar-valued function of the state, that is V : IRn IR. If V is
continuously differentiable then, at any time t, the derivative of V along a solution x() of
system (8.1) is given by
dV
(x(t)) = DV (x(t))x(t)

dt
= DV (x(t))f (x(t))
where DV (x) is the derivative of V at x and is given by
DV (x) =
Note that
DV (x)f (x) =

V
(x)
x1

V
(x)
x2

...

V
(x)
xn

V
V
V
(x)f1 (x) +
(x)f2 (x) + . . . +
(x)fn (x)
x1
x2
xn

In what follows, if a condition involves DV , then it is implicitly assumed that V is continuously differentiable. Sometimes DV is denoted by
V
or
V (x)T .
x
69

70

CHAPTER 8. BASIC LYAPUNOV THEORY

Also, when the system under consideration is fixed, DV f is sometimes denoted by

V
This a slight but convenient abuse of notation.
In the following sections, we use Lyapunov functions to present sufficient conditions for
the stability and boundedness concepts previously introduced. To guarantee a specific type
of stability using a Lyapunov function, one has to obtain a Lyapunov function V which has
certain properties and whose time-derivative V also has certain properties.

8.1. LYAPUNOV CONDITIONS FOR STABILITY

8.1
8.1.1

71

Lyapunov conditions for stability

Locally positive definite functions

DEFN. (Locally positive definite function) A function V is locally positive definite (lpd) about
a point xe if
V (xe ) = 0
and there is a scalar R > 0 such that
V (x) > 0

whenever x 6= xe

and ||x xe || < R

Basically, a function is lpd about a point xe if it is zero at xe has a strict local minimum
at xe .

Figure 8.1: A locally positive definite function

Example 71 (Scalar x) The following functions are lpd about zero.
V (x) = x2
2

V (x) = 1 ex
V (x) = 1 cos x
V (x) = x2 x4
Example 72
V (x) = ||x||2
= x21 + x22 + . . . + x2n
Lpd about the origin.
Example 73
V (x) = ||x||1
= |x1 | + |x2 | + . . . + |xn |
Lpd about the origin.

72

CHAPTER 8. BASIC LYAPUNOV THEORY

Quadratic forms. Suppose P is a real n n symmetric matrix and is positive definite.

Consider the quadratic form defined by
T

V (x) = x P x =

n X
n
X

Pij xi xj .

i=1 j=1

Clearly V (0) = 0. Recalling the definition of a positive definite matrix, it follows that
V (x) = xT P x > 0 for all nonzero x. Hence V is locally positive definite about the origin.
The second derivative of V at x is the square symmetric matrix given by:

that is,

2
D V (x) :=

2V
(x)
2 x1

2V
(x)
x1 x2

...

2V
(x)
x1 xn

2V
2V
2V

(x)
(x)
(x)
.
.
.
x2 x1
2 x2
x2 xn

..
..
..

.
.
.

2V
2V
2V
(x) xn x2 (x) . . . 2 xn (x)
xn x1
2V
(x)
D V (x)ij =
xi xj
2

Sometimes D 2 V (x) is written as

2V
x2

(x) and is referred to as the Hessian of V .

The following lemma is sometimes useful in demonstrating that a function is lpd.

Lemma 2 Suppose V is twice continuously differentiable and
V (xe ) = 0
DV (xe ) = 0
D 2 V (xe ) > 0
Then V is a locally positive definite about xe .
Example 74 Consider V (x) = x2 x4 where x is a scalar. Since V (0) = DV (0) = 0 and
D 2 V (0) = 2 > 0, V is lpd about zero.
Example 75 Consider

1
V (x) = 1 cos x1 + x22
2

Clearly
V (0) = 0
Since
DV (x) =

sin x1 x2

8.1. LYAPUNOV CONDITIONS FOR STABILITY

73

we have
DV (0) = 0
Also,
2

D V (x) =
Hence,

cos x1 0
0
1

1 0
0 1

D V (0) =

>0

Since V satisfies the hypotheses of the previous lemma with xe = 0, it is lpd about zero.

8.1.2

A stability result

If the equilibrium state of a nonlinear system is stable but not asymptotically stable, then
one cannot deduce the stability properties of the equilibrium state of the nonlinear system
from the linearization of the nonlinear system about that equilibrium state. The following
Lyapunov result is useful in demonstrating stability of an equilibrium state for a nonlinear
system.
Theorem 1 (Stability) Suppose there exists a function V and a scalar R > 0 such that V is
locally positive definite about xe and
for ||x xe || < R

DV (x)f (x) 0
Then xe is a stable equilibrium state.

If V satisfies the hypotheses of the above theorem, then V is said to be a Lyapunov

function which guarantees the stability of xe .
Example 76
x = 0
Consider
V (x) = x2
as a candidate Lyapunov function. Then V is lpd about 0 and
DV (x)f (x) = 0
Hence (it follows from Theorem 1 that) the origin is stable.
Example 77 (Undamped linear oscillator.)
x 1 = x2
x 2 = kx1

k>0

74

CHAPTER 8. BASIC LYAPUNOV THEORY

Consider the total energy,

1
1
V (x) = kx21 + x22
2
2
as a candidate Lyapunov function. Then V is lpd about the origin and
DV (x)f (x) = 0
Hence the origin is stable.
Example 78 (Simple pendulum.)
x 1 = x2
x 2 = sin x1
Consider the total energy,

1
V (x) = 1 cos x1 + x22
2
as a candidate Lyapunov function. Then, as we have allready shown, V is lpd about the origin;
also
DV (x)f (x) = 0
Hence the origin is stable.
Example 79 (Stability of origin for attitude dynamics system.) Recall
(I2 I3 )
x2 x3
I1
(I3 I1 )
=
x3 x1
I2
(I1 I2 )
x1 x2
=
I3

x 1 =
x 2
x 3
where

I1 , I2 , I3 > 0
Consider the kinetic energy
V (x) =

1
I1 x21 + I2 x22 + I3 x23
2

as a candidate Lyapunov function. Then V is lpd about the origin and

DV (x)f (x) = 0
Hence the origin is stable.
Example 80 (Undamped Duffing system)
x 1 = x2
x 2 = x1 x31

8.1. LYAPUNOV CONDITIONS FOR STABILITY

75

As candidate Lyapunov function for the equilibrium state xe = [ 1 0 ]T consider the total
energy
1
1
1
1
V (x) = x41 x21 + x22 + .
4
2
2
4
Since

 2


3x1 1 0
2
3
DV (x) = x1 x1 x2
and
D V (x) =
0
1
we have V (xe ) = 0, DV (xe ) = 0 and D 2 V (xe ) > 0, and it follows that V is lpd about xe .
One can readily verify that DV (x)f (x) = 0. Hence, xe is stable.

Exercises
Exercise 13 Determine whether or not the following functions are lpd. (a)
V (x) = x21 x41 + x22
(b)
V (x) = x1 + x22
(c)
V (x) = 2x21 x31 + x1 x2 + x22
Exercise 14 (Simple pendulum with Coulomb damping.) By appropriate choice of Lyapunov function, show that the origin is a stable equilibrium state for
x 1 = x2
x 2 = sin x1 c sgm (x2 )
where c > 0.
Exercise 15 By appropriate choice of Lyapunov function, show that the origin is a stable
equilibrium state for
x 1 = x2
x 2 = x31
Note that the linearization of this system about the origin is unstable.
Exercise 16 By appropriate choice of Lyapunov function, show that the origin is a stable
equilibrium state for
x 1 = x2
x 2 = x1 + x31

76

8.2

CHAPTER 8. BASIC LYAPUNOV THEORY

Asymptotic stability

The following result presents conditions which guarantee that an equilibrium state is asymptotically stable.
Theorem 2 (Asymptotic stability) Suppose there exists a function V and a scalar R > 0
such that V is locally positive definite about xe and
DV (x)f (x) < 0

for x 6= xe

and ||x xe || < R

Then xe is an asymptotically stable equilibrium state for x = f (x).

Example 81
x = x3
Consider
V (x) = x2
Then V is lpd about zero and
DV (x)f (x) = 2x4 < 0

for x 6= 0

Hence the origin is AS.

Example 82
x = x + x3
Consider
V (x) = x2
Then V is lpd about zero and
DV (x)f (x) = 2x2 (1 x2 ) < 0

for |x| < 1, x 6= 0

Hence the origin is AS. Although the origin is AS, there are solutions which go unbounded
in a finite time.
Example 83
x = sin x
Consider
V (x) = x2
Then V is lpd about zero and
DV (x)f (x) = 2x sin(x) < 0
Hence the origin is AS.

for |x| < , x 6= 0

8.2. ASYMPTOTIC STABILITY

77

Example 84 (Simple pendulum with viscous damping.) Intuitively, we expect the origin to
be an asymptotically stable equilibrium state for the damped simple pendulum:
x 1 = x2
x 2 = sin x1 cx2
where c > 0 is the damping coefficient. If we consider the total mechanical energy
1
V (x) = 1 cos x1 + x22
2
as a candidate Lyapunov function, we obtain
DV (x)f (x) = cx22 .
Since DV (x)f (x) 0 for all x, we have stability of the origin. Since DV (x)f (x) = 0
whenever x2 = 0, it follows that DV (x)f (x) = 0 for points arbitrarily close to the origin;
hence V does not satisfy the requirements of the above theorem for asymptotic stability.
Suppose we modify V to
1
1
V (x) = c2 x21 + cx1 x2 + x22 + 1 cos x1
2
2
where is any scalar with 0 < < 1. Letting


1 c2 c
P =
c 1
2
note that P > 0 and
V (x) = xT P x + 1 cos x1
xT P x
Hence V is lpd about zero and we obtain
DV (x)f (x) = cx1 sin x1 (1 )cx22 < 0

for ||x|| < , x 6= 0

to satisfy the requirements of above theorem; hence the origin is AS.

Exercises
Exercise 17 By appropriate choice of Lyapunov function, show that the origin is an asymptotically stable equilibrium state for
x 1 = x2
x 2 = x51 x2
Exercise 18 By appropriate choice of Lyapunov function, show that the origin is a asymptotically stable equilibrium state for
x 1 = x2
x 2 = x1 + x31 x2

78

CHAPTER 8. BASIC LYAPUNOV THEORY

8.3
8.3.1

Boundedness
Radially unbounded functions

DEFN. A scalar valued function V of the state x is said to be radially unbounded if

lim V (x) =

Example 85
V (x)
V (x)
V (x)
V (x)
V (x)
V (x)

=
=
=
=
=
=

x2
2
1 ex
x2 x
x4 x2
x sin x
x2 (1 cos x)

yes
no
yes
yes
no
no

Example 86 Suppose P is a real n n matrix and is positive definite symmetric and

consider the quadratic form defined by
V (x) = xT P x
Since P is real symmetric,
xT P x min (P )||x||2
for all x, where min(P ) is the minimum eigenvalue of P . Since P is positive definite,
min (P ) > 0. From this it should be clear that V is radially unbounded.
The following lemma can be useful for guaranteeing radial unboundedness.
Lemma 3 Suppose V is twice continuously differentiable, and there is a positive definite
symmetric matrix P and a scalar R 0 such that
D 2 V (x) P

for

||x|| R

Then, V is radially unbounded.

8.3.2

A boundedness result

Theorem 3 Suppose there exists a radially unbounded function V and a scalar R 0 such
that
DV (x)f (x) 0
for ||x|| R
Then all solutions of x = f (x) are bounded.

8.3. BOUNDEDNESS

79

Note that, in the above theorem, V does not have to be positive away from the origin; it
only has to be radially unbounded.
Example 87 Recall
x = x x3 .
Consider
V (x) = x2
Since V is radially unbounded and
DV (x)f (x) = 2x2 (x2 1)
the hypotheses of the above theorem are satisfied with R = 1; hence all solutions are bounded.
Note that the origin is unstable.
Example 88 Duffings equation
x 1 = x2
x 2 = x1 x31
Consider
1
1
1
V (x) = x22 x21 + x41
2
2
4
It should be clear that V is radially unbounded; also
DV (x)f (x) = 0 0

for all x

So, the hypotheses of the above theorem are satisfied with any R; hence all solutions are
bounded.
Example 89
x = x x3 + w
where |w| .
Example 90
x = Ax + Bw
where A is Hurwitz and |w| .

80

CHAPTER 8. BASIC LYAPUNOV THEORY

Exercises
Exercise 19 Determine whether or not the the following function is radially unbounded.
V (x) = x1 x31 + x41 x22 + x42
Exercise 20 (Forced Duffings equation with damping.) Show that all solutions of the
system
x 1 = x2
x 2 = x1 x31 cx2 + 1

c>0

are bounded.
Hint: Consider

1
1
1
1
V (x) = c2 x21 + cx1 x2 + x22 x21 + x41
2
2
2
4
where 0 < < 1. Letting


1 c2 c
P =
c 1
2
note that P > 0 and

1
1
V (x) = xT P x x21 + x41
2
4
1
T
x Px
4
Exercise 21 Recall the Lorenz system
x 1 = (x2 x1 )
x 2 = rx1 x2 x1 x3
x 3 = bx3 + x1 x2
with b > 0. Prove that all solutions of this system are bounded. (Hint: Consider V (x) =
rx21 + x22 + (x3 2r)2 .)

8.4. GLOBAL ASYMPTOTIC STABILITY

8.4
8.4.1

81

Global asymptotic stability

Positive definite functions

DEFN. (Positive definite function.) A function V is positive definite (pd) if

V (0) = 0
V (x) > 0
lim V (x) =

for all x 6= 0

In the above definition, note the requirement that V be radially unbounded.

Example 91 Scalar x
V (x) = x2
2
V (x) = 1 ex
V (x) = 1 cos x

pd
lpd but not pd
lpd but not pd

Example 92
V (x) = x41 + x22
Example 93
V (x) = ||x||2
= x21 + x22 + . . . + x2n
Example 94 Suppose P is a real n n matrix and is positive definite symmetric and
consider the quadratic form defined by
V (x) = xT P x
V is a positive definite function.
The following lemma can be useful in demonstrating that a function has the first two
properties needed for a function to be pd.
Lemma 4 Suppose V is twice continuously differentiable and
V (0) = 0
DV (0) = 0
D 2 V (x) > 0
Then V (x) > 0 for all x.

for all x

82

CHAPTER 8. BASIC LYAPUNOV THEORY

If V satisfies the hypotheses of the above lemma, it is not guaranteed to be radially unbounded, hence it is not guaranteed to be positive definite. Lemma 3 can be useful for
guaranteeing radial unboundedness. We also have the following lemma.
Lemma 5 Suppose V is twice continuously differentiable,
V (0) = 0
DV (0) = 0
and there is a positive definite symmetric matrix P such that
D 2 V (x) P
Then

for all x

1
V (x) xT P x
2

for all x.

8.4.2

A result on global asymptotic stability

Theorem 4 (Global asymptotic stability) Suppose there exists a positive definite function V
such that
DV (x)f (x) < 0
for all x 6= 0
Then the origin is a globally asymptotically stable equilibrium state for x = f (x).
Example 95
x = x3

V (x) = x2
DV (x)f (x) = 2x4
< 0
for all x 6= 0
We have GAS. Note that linearization of this system about the origin cannot be used to
deduce the asymptotic stability of this system.
Example 96 The first nonlinear system
x = sgm (x) .
This system is not linearizable about its unique equilibrium state at the origin. Considering
V (x) = x2
we obtain
DV (x)f (x) = 2|x|
< 0
for all x 6= 0 .
Hence, we have GAS.

8.4. GLOBAL ASYMPTOTIC STABILITY

83

Example 97 Consider
x 1 = x1 + sin(x1 x2 )
x 1 = 2x2 sin(x1 x2 )
We will show that this system is GAS provided is small enough. Considering the positive
definite function
V (x) = x21 + x22 .
as a candidate Lyapunov function, we have
V

= 2x1 x 1 + 2x2 x 2

Example 98 (Linear globally stabilizing controller for inverted pendulum.)

x 1 = x2
x 2 = sin x1 + u
Consider
u = k1 x1 k2 x2

with k1 > 1, k2 > 0

Closed loop system:

x 1 = x2
x 2 = k1 x1 + sin x1 k2 x2
Consider

1
1
1
V (x) = k22 x21 + k2 x1 x2 + x22 + k1 x21 + cos x1 1
2
2
2
where 0 < < 1. Then V is pd (apply lemma 5) and
DV (x)f (x) = k2 (k1 x21 + x1 sin x1 ) (1 )k2 x22
Since
| sin x1 | |x1 |

for all x1

84

CHAPTER 8. BASIC LYAPUNOV THEORY

it follows that
x1 sin x1 x21

for all x1

hence
DV (x)f (x) k2 (k1 1)x21 (1 )k2 x22

<0

for all x 6= 0

The closed loop system is GAS.

Example 99 (Stabilization of origin for attitude dynamics system.)
(I2 I3 )
x2 x3
I1
(I3 I1 )
x3 x1
=
I2
(I1 I2 )
=
x1 x2
I3

x 1 =

x 2

x 3

u1
I1
u2
I2
u3
I3

where
I1 , I2 , I3 > 0
Consider any linear controller of the form
ui = kxi

i = 1, 2, 3,

k>0

Closed loop system:

x 1
x 2
x 3


I2 I3
x2 x3
=
I1


I3 I1
=
x3 x1
I2


I1 I2
x1 x2
=
I3


kx1
I1
kx2
I2
kx3
I3

Consider the kinetic energy

V (x) =

1
I1 x21 + I2 x22 + I3 x23
2

as a candidate Lyapunov function. Then V is pd and

DV (x)f (x) = k(x21 + x22 + x23 )

< 0
for all x 6= 0
Hence the origin is GAS for the closed loop system.

8.5. EXPONENTIAL STABILITY

85

Exercises
Exercise 22 Determine whether or not the following function is positive definite.
V (x) = x41 x21 x2 + x22
Exercise 23 Consider any scalar system described by
x = g(x)
where g has the following properties:
g(x) > 0 for x > 0
g(x) < 0 for x < 0
Show that this system is GAS.
Exercise 24 Recall the Lorenz system
x 1 = (x2 x1 )
x 2 = rx1 x2 x1 x3
x 3 = bx3 + x1 x2
Prove that if
b>0

and

0 r < 1,

then this system is GAS about the origin. (Hint: Consider V (x) = x21 + x22 + x23 .)
Exercise 25 (Stabilization of the Duffing system.) Consider the Duffing system with a
scalar control input u(t):
x 1 = x2
x 2 = x1 x31 + u
Obtain a linear controller of the form
u = k1 x1 k2 x2
which results in a closed loop system which is GAS about the origin. Numerically simulate
the open loop system (u = 0) and the closed loop system for several initial conditions.

8.5
8.5.1

Exponential stability
Global exponential stability

Theorem 5 (Global exponential stability) Suppose there exists a function V and positive
scalars , 1 , 2 such that for all x,
1 ||xxe ||2 V (x) 2 ||xxe ||2

86

CHAPTER 8. BASIC LYAPUNOV THEORY

and
DV (x)f (x) 2V (x) .
Then, for the system x = f (x), the state xe is a globally exponentially stable equilibrium
state with rate of convergence . In particular, all solutions x() of the system satisfy
kx(t)xe k

2 /1 kx(0)xe k et

for all

Proof. See below.

Example 100
x = x
Considering
V (x) = x2
we have
DV (x)f (x) = 2x2
= 2V (x)
Hence, we have GES with rate of convergence 1.
Example 101
x = x x3
Considering
V (x) = x2
we have
DV (x)f (x) = 2x2 2x4
2V (x)
Hence, we have GES with rate of convergence 1.

8.5.2

Proof of theorem 6

Consider any solution x() and let v(t) = V (x(t)). Then

v 2v .

t 0.

8.5. EXPONENTIAL STABILITY

8.5.3

87

Exponential stability

Theorem 6 (Exponential stability) Suppose there exists a function V and positive scalars R,
, 1 , 2 such that, whenever kxxe k R, one has
1 ||xxe ||2 V (x) 2 ||xxe ||2
and
DV (x)f (x) 2V (x)

Then, for system x = f (x), the state xe is an exponentially stable equilibrium state with rate
of convergence .
Exercise 26 Consider the scalar system
x = x + x3
As a candidate Lyapunov function for exponential stability, consider V (x) = x2 . Clearly,
the condition on V is satisfied with 1 = 2 = 1. Noting that
V = 2x2 + 2x4 = 2(1 x2 )x2 ,
and considering R = 1/2, we obtain that whenever |x| R, we have V 2V where
= 3/4. Hence, we have ES.

88

8.5.4

CHAPTER 8. BASIC LYAPUNOV THEORY

A special class of GES systems

Consider a system described by

x = f (x)

(8.2)

and suppose that there exist two positive definite symmetric matrices P and Q such that
xT P f (x) xT Qx .
We will show that the origin is GES with rate
:= min (P 1 Q)
where min(P 1 Q) > 0 is the smallest eigenvalue of P 1 Q.
As a candidate Lyapunov function, consider
V (x) = xT P x .
Then
min (P )||x||2 V (x) max (P )||x||2

that is

1 ||x||2 V (x) 2 ||x||2

with

1 = min(P ) > 0

and

2 = max (P ) > 0 .

Now note that

DV (x)f (x) = 2xT P f (x)
2xT Qx
For any two positive-definite matrices P and Q, one can show that all the eigenvalues of
P 1 Q are real positive and
xT Qx min (P 1 Q) xT P x

where min(P 1 Q) > 0 is the smallest eigenvalue of P 1 Q. Thus,

DV (x)f (x) 2min (P 1 Q)xT P x
= 2V (x)
Hence GES with rate .
Example 102 Recall example 99.
Closed loop system:
x 1
x 2
x 3


I2 I3
=
x2 x3
I1


I3 I1
x3 x1
=
I2


I1 I2
x1 x2
=
I3

kx1
I1
kx2
I2
kx3
I3

8.5. EXPONENTIAL STABILITY

89

Considering

we have

I1 0 0
P = 0 I2 0
0 0 I3
xT P f (x) = kxT x

that is, Q = kI. Hence, GES with rate

= min (P 1 Q)
= min (kP 1 )
= kmin (P 1 )
k
=
max (P )
k
=
max{I1 , I2 , I3 }

90

8.5.5

CHAPTER 8. BASIC LYAPUNOV THEORY

Summary

The following table summarizes the results of this chapter for stability about the origin.

stability properties

lpd

0 for ||x|| R

lpd

< 0 for ||x|| R, x 6= 0

AS

ru

0 for ||x|| R

pd

< 0 for x 6= 0

GAS

1 ||x||2 V (x) 2 ||x||2

2V (x)

GES

1 ||x||2 V (x) 2 ||x||2 for ||x|| R

2V (x) for ||x|| R

ES

Figure 8.2: Lyapunov table

Chapter 9
Basic Lyapunov theory: discrete time
Suppose we are interested in the stability properties of the system,
x(k + 1) = f (x(k))

(9.1)

where x(k) IRn and k IR. If the system is linear, we can determine its stability properties
from the properties of the eigenvalues of the system matrix. What do we for a nonlinear
system? We could linearize about each equilibrium state and determine the stability properties of the resulting linearizations. Under certain conditions (see later) this will tell us
something about the local stability properties of the nonlinear system about its equilibrium
states. However there are situations where linearization cannot be used to deduce even the
local stability properties of the nonlinear system. Also, linearization tells us nothing about
the global stability properties of the nonlinear system.
In general, we cannot explicitly obtain solutions for nonlinear systems. Lyapunov theory
allows to say something about the stability properties of a system without knowing the form
or structure of the solutions.
In this chapter, V is a scalar-valued function of the state, i.e. V : IRn IR. At any time
k, the one step change in V along a solution x() of system (9.1) is given by
V (x(k + 1)) V (x(k)) = V (x(k))
where
V (x) := V (f (x)) V (x)

91

92

9.1

CHAPTER 9. BASIC LYAPUNOV THEORY: DISCRETE TIME

Stability

Theorem 7 (Stability) Suppose there exists a locally positive definite function V and a
scalar R > 0 such that
V (x) 0
for ||x|| < R
Then the origin is a stable equilibrium state.
If V satisfies the hypotheses of the above theorem, then V is said to be a Lyapunov function
which guarantees the stability of origin.
Example 103
x(k + 1) = x(k)
Consider
V (x) = x2
as a candidate Lyapunov function. Then V is a lpdf and
V (x) = 0
Hence (it follows from theorem 7 that) the origin is stable.

9.2. ASYMPTOTIC STABILITY

9.2

93

Asymptotic stability

Theorem 8 (Asymptotic stability) Suppose there exists a locally positive definite function
V and a scalar R > 0 such that
V (x) < 0

for

x 6= 0

and ||x|| < R

Then the origin is an asymptotically stable equilibrium state.

Example 104
1
x(k + 1) = x(k)
2
Consider
V (x) = x2
Then V is a lpdf and
1
V (x) = ( x)2 x2
2
3
= x2
4
< 0
for x 6= 0
Hence the origin is AS.
Example 105
1
x(k + 1) = x(k) + x(k)2
2
Consider
V (x) = x2
Then V is a lpdf and
1
V (x) = ( x + x2 )2 x2
2
1
3
= x2 ( + x)( x)
2
2
1
< 0
for |x| < , x 6= 0
2
Hence the origin is AS.

94

9.3

CHAPTER 9. BASIC LYAPUNOV THEORY: DISCRETE TIME

Boundedness

Theorem 9 Suppose there exists a radially unbounded function V and a scalar R 0 such
that
V (x) 0
for ||x|| R
Then all solutions of (9.1) are bounded.
Note that, in the above theorem, V does not have to positive away from the origin; it only
has to be radially unbounded.
Example 106
x(k + 1) =

2x(k)
1 + x(k)2

Consider
V (x) = x2
Since V is radially unbounded and
V (x) =

2x
1 + x2

2

x2

x2 (x2 + 3)(x2 1)
(x2 + 1)2

for |x| 1

the hypotheses of the above theorem are satisfied with R = 1; hence all solutions are bounded.
Note that the origin is unstable.

9.4. GLOBAL ASYMPTOTIC STABILITY

9.4

95

Global asymptotic stability

Theorem 10 (Global asymptotic stability) Suppose there exists a positive definite function
V such that
V (x) < 0
for all x 6= 0
Then the origin is a globally asymptotically stable equilibrium state.

Example 107
1
x(k + 1) = x(k)
2
Example 108
x(k + 1) =

x(k)
1 + x(k)2

Consider
V (x) = x2
Then
V (x) =

=
< 0

x
1 + x2

2

x2

2x4 + x6
(1 + x2 )2
for all x 6= 0

Hence this system is GAS about zero.

Exercise 27 Consider any scalar system described by
x(k + 1) = g(x(k))
where g has the following properties:
|g(x)| < |x| for x 6= 0
Show that this system is GAS.

96

9.5

CHAPTER 9. BASIC LYAPUNOV THEORY: DISCRETE TIME

Exponential stability

Theorem 11 (Global exponential stability.) Suppose there exists a function V and scalars
, 1 , 2 such that for all x,
1 ||x||2 V (x) 2 ||x||2

1 , 2 > 0

V (f (x)) 2 V (x)

0<1

and

Then, every solution satisfies

||x(k)||

2 k
||x(0)||
1

for

k0

Hence, the origin is a globally exponentially stable equilibrium state with rate of convergence
.
Proof.

Example 109
1
x(k + 1) = x(k)
2

9.5. EXPONENTIAL STABILITY

97

Considering
V (x) = x2
we have
1 2
x
4
1
= ( )2 V (x)
2

V (f (x)) =

Hence, we have GES with rate of convergence = 21 .

Example 110
x(k + 1) =

1
sin(x(k))
2

Considering
V (x) = x2
we have
1
V (f (x)) = ( sin x)2
2
1 2
= ( ) | sin x|2
2
1 2 2
( ) |x|
2
1 2
= ( ) V (x)
2
Hence, we have GES with rate of convergence =

1
2

98

CHAPTER 9. BASIC LYAPUNOV THEORY: DISCRETE TIME

Chapter 10
Lyapunov theory for linear
time-invariant systems
The main result of this section is contained in Theorem 12.

10.1

Positive and negative (semi)definite matrices

10.1.1

Definite matrices

For any square matrix P we can define an associated quadratic form:

x Px =

n X
n
X

pij xi xj

i=1 j=1

If P is hermitian (P = P ) then the scalar x P x is real for all x Cn .

DEFN. A hermitian matrix P is positive definite (pd) if
x P x > 0
for all nonzero x. We denote this by P > 0. The matrix P is negative definite (nd) if P is
positive definite; we denote this by P < 0.
Example 111 For
P =

1 1
1
2

we have (note the completion of squares trick)

x P x = x1 x1 x1 x2 x2 x1 + 2x2 x2
= (x1 x2 ) (x1 x2 ) + x2 x2
= |x1 x2 |2 + |x2 |2
Clearly, x P x 0 for all x. If x P x = 0, then x1 x2 = 0 and x2 = 0; hence x = 0. So,
P > 0.
99

100CHAPTER 10. LYAPUNOV THEORY FOR LINEAR TIME-INVARIANT SYSTEMS

Fact 1 The following statements are equivalent for any hermitian matrix P .
(a) P is positive definite.
(b) All the eigenvalues of P are positive.
(c) All the leading principal minors of P are positive.

10.1. POSITIVE AND NEGATIVE (SEMI)DEFINITE MATRICES

101

Example 112 Consider

P =

1 1
1
2

Since p11 = 1 > 0 and det(P ) = 1 > 0, we must have P > 0.

>> eig(P)
ans =
0.3820
2.6180
Note the positive eigenvalues.

10.1.2

Semi-definite matrices

DEFN. A hermitian matrix P is positive semi-definite (psd) if

x P x 0
for all non-zero x. We denote this by P 0
P is negative semi-definite (nsd) if P is positive semi-definite; we denote this by P 0

Fact 2 The following statements are equivalent for any hermitian matrix P .
(a) P is positive semi-definite.
(b) All the eigenvalues of P are non-negative.
(c) All the leading minors of P are positive.
Example 113 This example illustrates that non-negativity of the leading principal minors
of P is not sufficient for P 0.


0
0
P =
0 1
We have p11 = 0 and det(P ) = 0. However,
x P x = |x2 |2
hence, P is not psd. Actually, P is nsd.
Fact 3 Consider any m n complex matrix M and let P = M M. Then
(a) P is hermitian and P 0
(b) P > 0 iff rank M = n.

102CHAPTER 10. LYAPUNOV THEORY FOR LINEAR TIME-INVARIANT SYSTEMS

Example 114
P =

1
1

Since
P =
and

rank
P 0 but P is not pd.

1 1
1 1


1 1

1 1

=1

Exercise 28 (Optional)
Suppose P is hermitian and T is invertible. Show that P > 0 iff T P T > 0.
Exercise 29 (Optional)
Suppose P and Q are two hermitian matrices with P > 0. Show that P + Q > 0 for
> 0 such that whenever || < ,
one has
all real sufficiently small; i.e., there exists
P + Q > 0.

10.2. LYAPUNOV THEORY

103

10.2

Lyapunov theory

10.2.1

Asymptotic stability results

The scalar system

x = ax
is asymptotically stable if and only if a+a < 0. Consider now a general linear time-invariant
system described by
x = Ax
(10.1)
The generalization of a + a < 0 to this system is A + A < 0. We will see shortly that this
condition is sufficient for asymptotic stability; however, as the following example illustrates,
it is not necessary.
Example 115 The matrix
A=

1
3
0 1

is asymptotically stable. However, the matrix



2
3

A+A =
3 2
is not negative definite.
Since stability is invariant under a similarity transformation T , i.e., the stability of A and
T AT are equivalent for any nonsingular T , we should consider the more general condition
1

T 1 AT + T A T < 0
Introducing the hermitian matrix P := T T 1 , and pre- and post-multiplying the above
inequality by T and T 1 , respectively, yields
P > 0
PA + A P < 0

(10.2a)
(10.2b)

We now show that the existence of a hermitian matrix P satisfying these conditions guarantees asymptotic stability.
Lemma 6 Suppose there is a hermitian matrix P satisfying (10.2). Then system (10.1) is
asymptotically stable
Proof. Suppose there exists a hermititian matrix P which satisfies inequalities (10.2).
Consider any eigenvalue of A. Let v 6= 0 be an eigenvector corresponding to , i.e.,
Av = v
Then
v P Av = v P v ;

P v
v A P v = v

104CHAPTER 10. LYAPUNOV THEORY FOR LINEAR TIME-INVARIANT SYSTEMS

Pre- and post-multiplying inequality (10.2b) by v and v, respectively, yields
P v < 0
v P v + v
i.e.,
2Re()v P v < 0
Since P > 0, we must have v P v > 0; hence Re() < 0. Since the above holds for every
eigenvalue of A, system (10.1) is asymptotically stable.
Remark 1 A hermitian matrix P which satisfies inequalities (10.2) will be referred to as a
Lyapunov matrix for (10.1) or A.
Remark 2 Conditions (10.2) are referred to as linear matrix inequalities (LMIs). In recent
years, efficient numerical algorithms have been developed to solve LMIs or determine that
a solution does not exist.
Lyapunov functions. To obtain an alternative interpretation of inequalities (10.2), consider the quadratic function V of the state defined by
V (x) := x P x
Since P > 0, this function has a strict global minimum at zero. Using inequality (10.2b), it
follows that along any non-zero solution x() of (10.1),
dV (x(t))
= x P x + x P x
dt
= x (A P + P A) x
< 0
i.e., V (x()) is strictly decreasing along any non-zero solution. Intuitively, one then expects
every solution to asymptotically approach the origin. The function V is called a Lyapunov
function for the system. The concept of a Lyapunov function readily generalizes to nonlinear
systems and has proved a very useful tool in the analysis of the stability properties of
nonlinear systems.
Remark 3 Defining the hermitian matrix S := P 1 , inequalities (10.2) become
S > 0
AS + SA < 0

(10.3a)
(10.3b)

Hence the above lemma can be stated with S replacing P and the preceding inequalities
replacing (10.2).
So far we have shown that if a LTI system has a Lyapunov matrix, then it is AS. Is the
converse true? That is, does every AS LTI system have a Lyapunov matrix? And if this is

10.2. LYAPUNOV THEORY

105

true how does one find a Lyapunov matrix? To answer this question note that satisfaction
of inequality (10.2b) is equivalent to
P A + A P + Q = 0

(10.4)

where Q is a hermitian positive definite matrix. This linear matrix equation is known as the
Lyapunov equation. So one approach to looking for Lyapunov matrices could be to choose a
pd hermitian Q and determine whether the Lyapunov equation has a pd hermitian solution
for P .
We first show that if the system x = Ax is asymptotically stable and the Lyapunov
equation (10.4) has a solution then, the solution is unique. Suppose P1 and P2 are two
solutions to (10.4). Then,
(P2 P1 )A + A (P2 P1 ) = 0 .
Hence,

eA t (P2 P1 )AeAt + eA t A (P2 P1 )eAt = 0

that is,

This implies that, for all t,

d eA t (P2 P1 )eAt
=0
dt

eA t (P2 P1 )eAt = eA 0 (P2 P1 )eA0 = P2 P1 .

Since x = Ax is asymptotically stable, limt eAt = 0. Hence

P2 P1 = lim eA t (P2 P1 )eAt = 0 .

t

From this it follows that P2 = P1 .

The following lemma tells us that every AS LTI system has a Lyapunov matrix and a
Lyapunov matrix can be obtained by solving the Lyapunov equation with any pd hermitian
Q.
Lemma 7 Suppose system (10.1) is asymptotically stable. Then for every matrix Q, the
Lyapunov equation (10.4) has a unique solution for P . If Q is positive definite hermitian
then this solution is positive definite hermitian.
Proof. Suppose system (10.1) is asymptotically stable and consider any matrix Q. Let
P :=

eA t QeAt dt

This integral exists because each element of eAt is exponentially decreasing.

To show that P satisfies the Lyapunov equation (10.4), use the following properties of
eAt ,

deA t
deAt
A t
At
Ae =
e A=
dt
dt

106CHAPTER 10. LYAPUNOV THEORY FOR LINEAR TIME-INVARIANT SYSTEMS

to obtain

PA + A P =

eA t QeAt A + A eA t QeAt

dt




At
deA t At
A t de
=
dt
e Q
+
Qe
dt
dt
0


Z

d eA t QeAt
=
dt
dt
0
Z t A t At
de Qe
dt
= lim
t 0
dt
Z

= lim eA t QeAt Q
t

= Q
We have already demonstrated uniqueness of solutions to (10.4),
Suppose Q is pd hermitian. Then it should be clear that P is pd hermitian.

Using the above two lemmas, we can now state the main result of this section.
Theorem 12 The following statements are equivalent.
(a) The system x = Ax is asymptotically stable.
(b) There exist positive definite hermitian matrices P and Q satisfying the Lyapunov equation (10.4).
(c) For every positive definite hermitian matrix Q, the Lyapunov equation (10.4) has a
unique solution for P and this solution is hermitian positive definite.
Proof. The first lemma yields (b) = (a). The second lemma says that (a) = (c).
Hence, (b) = (c).
To see that (c) = (b), pick any positive definite hermitian Q. So, (b) is equivalent to
(c). Also, (c) = (a); hence (a) and (c) are equivalent.
.
Example 116
x1 = x2
x2 = x1 + cx2
Here
A=

0 1
1 c

10.2. LYAPUNOV THEORY

107

Choosing
Q=

1 0
0 1

and letting
P =

p11 p12
p12 p22

(note we have taken p21 = p12 because we are looking for symmetric solutions) the Lyapunov
equation results in
2p12 + 1 = 0
p11 + cp12 p22 = 0
2p12 + 2cp22 + 1 = 0

We consider three cases:

Case i) c = 0. No solution; hence no GAS.
Case ii) c = 1. A single solution; this solution is pd; hence GAS.
Case iii) c = 1. A single solution; this solution is not pd; hence no GAS.
Remark 4 One can state the above theorem replacing P with S and replacing (10.4) with
AS + SA + Q = 0

10.2.2

(10.5)

MATLAB.

>> help lyap

LYAP

Lyapunov equation.
X = LYAP(A,C) solves the special form of the Lyapunov matrix
equation:
A*X + X*A = -C
X = LYAP(A,B,C) solves the general form of the Lyapunov matrix
equation:
A*X + X*B = -C
See also DLYAP.

Note. In MATLAB, A is the complex conjugate transpose of A (i.e. A ). Hence, to solve

(10.4) one must use P= lyap(A, Q).

108CHAPTER 10. LYAPUNOV THEORY FOR LINEAR TIME-INVARIANT SYSTEMS

>> q=eye(2);
>> a=[0 1; -1 -1];
>> p=lyap(a, q)
p =
1.5000
0.5000
0.5000
1.0000
>> p*a + a*p
ans =
-1.0000
-0.0000
0
-1.0000
>> det(p)
ans =
1.2500
>> eig(p)
ans =
1.8090
0.6910
Exercise 30 Consider the Lyapunov equation
P A + A P + 2P + Q = 0
where A is a square matrix, Q is some positive definite hermitian matrix and is some
positive scalar. Show that this equation has a unique solution for P and this solution is
positive definite hermitian if and only if for every eigenvalue of A,
Re() <

10.2.3

Stability results

For stability, the following lemma is the analog of lemma 10.2.

Lemma 8 Suppose there is a hermitian matrix P satisfying
P > 0
PA + A P 0

(10.6a)
(10.6b)

Then system (10.1) is stable.

We do not have the exact analog of lemma 7. However the following can be shown.

10.2. LYAPUNOV THEORY

109

Lemma 9 Suppose system (10.1) is stable. Then there exist a positive definite hermitian
matrix P and a positive semi-definite hermitian matrix Q satisfying the Lyapunov equation
(10.4).
The above lemma does not state that for every positive semi-definite hermitian Q the
Lyapunov equation has a solution for P . Also, when the Lyapunov equation has a solution,
it is not unique. This is illustrated in the following example.
Example 117
A=

With
P =
we obtain

0 1
1 0
1 0
0 1




P A + A P = 0
In this example the Lyapunov equation has a pd solution with Q = 0; this solution is
not unique; any matrix of the form


p 0
P =
0 p
(where p is arbitrary) is also a solution.
If we consider the psd matrix
Q=
the Lyapunov equation has no solution.

1 0
0 0

Combining the above two lemmas results in the following theorem.

Theorem 13 The following statements are equivalent.
(a) The system x = Ax is stable.
(b) There exist a positive definite hermitian matrix P and a positive semi-definite matrix
Q satisfying the Lyapunov equation (10.4).

110CHAPTER 10. LYAPUNOV THEORY FOR LINEAR TIME-INVARIANT SYSTEMS

10.3

Mechanical systems

Example 118 A simple structure

Figure 10.1: A simple structure

m1 q1 + (c1 + c2 )q1 c2 q2 + (k1 + k2 )q1 k2 q2 = 0
m2 q2
c2 q1
+ c2 q2
k2 q1
+ k2 q2 = 0
Letting
q=

q1
q2

this system can be described by the following second order vector differential equation:
M q + C q + Kq = 0
where the symmetric matrices M, C, K are given by




c1 + c2 c2
m1 0
C=
M=
c2
c2
0 m2

K=

k1 + k2 k2
k2
k2

Since m1 , m2 > 0, we have M > 0.

If k1 , k2 > 0 then K > 0 (why?) and if c1 , c2 > 0 then C > 0.
Note also that
kinetic energy

potential energy =

1
m q2
2 1 1
1
k q2
2 1 1

+ 21 m2 q22

1
q M q
2

+ 21 k2 (q2 q1 )2 =

1
q Kq
2

10.3. MECHANICAL SYSTEMS

111

Consider now a general mechanical system described by

M q + C q + Kq = 0
where q(t) is a real N-vector of generalized coordinates which describes the configuraton of
the system. The real matrix M is the inertia matrix and satisfies
M = M > 0
We call K and C the stiffness matrix and damping matrix respectively. The kinetic energy
of the system is given by
1
q M q
2
With x = (q , q ) , this system has a state space description of the form x = Ax with


0
I
A=
M 1 K M 1 C
A first Lyapunov matrix. Suppose
K = K > 0

and

C = C 0 .

The potential energy of the system is given by

1
q Kq
2
Consider the following candidate Lyapunov matrix


1 K 0
P =
0 M
2

Then P = P > 0 and

P A + A P + Q = 0
where
Q=
It should be clear that

0 0
0 C

P A + A P 0

iff
Hence,

C0

K = K > 0 and C = C 0 imply stability.

Note that
1
1
V (x) = x P x = q M q + q Kq = total energy
2
2

V (x) = q C q

112CHAPTER 10. LYAPUNOV THEORY FOR LINEAR TIME-INVARIANT SYSTEMS

A second Lyapunov matrix.


1 K + C M
P =
M
M
2

1
2

K 0
0 M

+
2

C M
M 0

For sufficiently small , the matrix P is positive definite.

Now,
P A + A P + Q = 0
where
Q=

K
0
0 C M

For sufficiently small > 0, the matrix C M is pd and, hence, Q is pd. So,
K = K > 0 and C = C > 0 imply asymptotic stability.

10.4. RATES OF CONVERGENCE FOR LINEAR SYSTEMS

10.4

113

Rates of convergence for linear systems

Theorem 14 Consider an asymptotically stable linear system described by x = Ax and let

be any real number satisfying
0<<

where

:= max{() : is an eigenvalue of A } .
Then x = Ax is GES with rate .
Proof: Consider any satisfying 0 < <
. As a consequence of the definition of
, all
the eigenvalues of the matrix A+I have negative parts. Hence, the Lyapunov equation
P (A+ I) + (A+I)P + I = 0 .
has a unique solution for P and P = P > 0. As a candidate Lyapunov for the system
x = Ax, consider V (x) = x P x. Then,
V

= x P x + x P x
= x P Ax + (Ax) P x
= x (P A + A P )x .

From the above Lyapunov matrix equation, we obtain that

P A + A P = 2P I ;
hence,
V

= 2x P x x x
2V (x) .

Hence the system is globally exponentially stable with rate of convergence .

114CHAPTER 10. LYAPUNOV THEORY FOR LINEAR TIME-INVARIANT SYSTEMS

10.5

Linearization and exponential stability

Consider a nonlinear system

x = f (x)

(10.7)

and suppose that xe is an equilibrium state for this system. We assume that f is differential
at xe and let

f1 e f1 e
f1 e
x1 (x ) x2 (x ) xn (x )

f2 e
f2 e f2 e

(x )
(x )
(x )

x1
x2
xn
e

Df (x ) =

.
.
.

..
..
..

fn e fn e
fn e
(x )
(x )
(x )
x1
x2
xn

where x = (x1 , x2 , , xn ). Then the linearization of x = f (x) about xe is the linear system
defined by
x = Ax
where A = Df (xe ) .
(10.8)

Theorem 15 Suppose xe is an equilibrium state of a nonlinear system of the form x = f (x)

and the corresponding linearization is exponentially stable. Then the nonlinear system is
exponentially stable about xe .
Proof. Suppose the linearization is exponentially stable and let

:= max{() : is an eigenvalue of A } .
Since all the eigenvalues have negative real parts, we have
> 0. Consider now any
satisfying 0 < < .
As a consequence of the definition of
, all the eigenvalues of the
matrix A+I have negative parts. Hence, the Lyapunov equation
P (A+ I) + (A+I)P + I = 0 .
has a unique solution for P and P = P > 0. Without loss of generality consider xe = 0. As
a candidate Lyapunov function for the nonlinear system, consider
V (x) = x P x .
Recall that
f (x) = f (0) + Df (0)x + o(x)
where the remainder term has the following property:
o(x)
= 0.
x0, x6=0 kxk
lim

10.5. LINEARIZATION AND EXPONENTIAL STABILITY

115

Hence
f (x) = Ax + o(x)
and
V

=
=
=

2x P x
2x P f (x)
2x P Ax + 2x P o(x)
x (P A + A P )x + 2kxkkP kko(x)k .

From the above Lyapunov matrix equation, we obtain that

P A + A P = 2P I ;
hence,
V

= 2x P x x x + 2kxkkP kko(x)k
= 2V (x) kxk2 + 2kxkkP kko(x)k .

As a consequence of the properties of o, there is a scalar R > 0 such that

2kP kko(x)k kxk when kxk R .
Hence, whenever kxk R we obtain that
V 2V (x) .
Hence the nonlinear system is exponentially stable about xe with rate of convergence .

116CHAPTER 10. LYAPUNOV THEORY FOR LINEAR TIME-INVARIANT SYSTEMS

Chapter 11
Quadratic stability
11.1

Introduction

[3] In this chapter, we introduce numerical techniques which are useful in finding Lyapunov
functions for certain classes of globally exponentially stable nonlinear systems. We restrict
consideration to quadratic Lyapunov functions. For specific classes of nonlinear systems, we
reduce the search for a Lyapunov function to that of solving LMIs (Linear matrix inequalities).
The results in this chapter are also useful in proving stability of switching linear systems.
To establish the results in this chapter, recall that a system
x = f (x)

(11.1)

is globally exponentially stable about the origin if there exist positive definite symmetric
matrices P and Q which satisfy
xT P f (x) xT Qx
for all x. When this is the case we call P a Lyapunov matrix for system (11.1).

11.2

Polytopic nonlinear systems

Here we consider nonlinear systems described by

x = A(x)x

(11.2)

where the n-vector x is the state. The state dependent matrix A(x) has the following
structure
A(x) = A0 + (x)A
(11.3)
where A0 and A are constant n n matrices and is a scalar valued function of the state
x which is bounded above and below, that is
a (x) b

(11.4)

for some constants a and b. Examples of functions satisfying the above conditions are given
2
by (x) = g(c(x)) where c(x) is a scalar and g(y) is given by sin y, cos y, ey or sgm (y).
The signum function is useful for modeling switching systems.
117

118

CHAPTER 11. QUADRATIC STABILITY

Example 119 Inverted pendulum under linear feedback. Consider an inverted pendulum
under linear control described by
x 1 = x2
x 2 = 2x1 x2 + sin x1

>0

(11.5a)
(11.5b)

This system can be described by (11.2)-(11.4) with





0 0
0
1
,
A =
A0 =
1 0
2 1
and
(x) =
Since | sin x1 | |x1 |, we have

sin x1 /x1 if x1 =
6 0

if x1 = 0

(x) ;

hence a = and b = .
The following theorem provides a sufficient condition for the global exponential stability of
system (11.2)-(11.4). These conditions are stated in terms of linear matrix inequalities and
the two matrices corresponding to the extreme values of (x), namely
A1 := A0 + aA

and

A2 := A0 + bA

Theorem 16 Suppose there exists a positive-definite symmetric matrix P which satisfies the
following linear matrix inequalities:
P A1 + AT1 P < 0
P A2 + AT2 P < 0

(11.6)

Then system (11.2)-(11.4) is globally exponentially stable (GES) about the origin with Lyapunov matrix P .
Proof. As a candidate Lyapunov function for GES, consider V (x) = xT P x. Then
V

= 2xT P x
= 2xT P A0 x + 2(x)xT P Ax .

For each fixed x, the above expression for V is a linear affine function of the scalar (x).
Hence an upper bound for V occurs when (x) = a or (x) = b which results in
V 2xT P A1 x = xT (P A1 + AT1 P )x
or

V 2xT P A2 x = xT (P A2 + AT2 P )x

11.2. POLYTOPIC NONLINEAR SYSTEMS

119

respectively. As a consequence of the matrix inequalities (11.6), there exist positive scalars
1 and 2 such that
P A1 + AT1 P 21 P
P A2 + AT2 P 22 P
Letting = min{1 , 2 }, we now obtain that
V 2xT P x = 2V .
This guarantees the system is GES with rate .
Exercise 31 Prove the following result: Suppose there exists a positive-definite symmetric
matrix P and a positive scalar which satisfy
P A1 + AT1 P + 2P 0
P A2 + AT2 P + 2P 0

(11.7a)
(11.7b)

where A1 := A0 + aA and A2 := A0 + bA. Then system (11.2)-(11.4) is globally exponentially stable about the origin with rate of convergence .

11.2.1

LMI Control Toolbox

Recall the pendulum of Example 119. We will use the Matlab LMI Control Toolbox to see if
we can show exponential stability of this system. First note that the existence of a positive
definite symmetric P satisfying inequalities (11.6) is equivalent to the existence of another
symmetric matrix P satisfying
P A1 + AT1 P < 0
P A2 + AT2 P < 0
P > I
For = 1, we determine the feasibility of these LMIs using the following Matlab program.
% Quadratic stability of inverted pendulum
%
gamma=1;
A0 = [0 1; -2 -1];
DelA = [0 0; 1 0];
A1 = A0 - gamma*DelA;
A2 = A0 + gamma*DelA;
%
%
setlmis([])
%
p=lmivar(1, [2,1]);
%

under linear PD control.

120

CHAPTER 11. QUADRATIC STABILITY

lmi1=newlmi;
lmiterm([lmi1,1,1,p],1,A1,s)
%
lmi2=newlmi;
lmiterm([lmi2,1,1,p],1,A2,s)
%
Plmi= newlmi;
lmiterm([-Plmi,1,1,p],1,1)
lmiterm([Plmi,1,1,0],1)
%
lmis = getlmis;
%
[tfeas, xfeas] = feasp(lmis)
%
P = dec2mat(lmis,xfeas,p)
Running this program yields the following output
Solver for LMI feasibility problems L(x) < R(x)
This solver minimizes t subject to L(x) < R(x) + t*I
The best value of t should be negative for feasibility
Iteration

Best value of t so far

1
2
Result:

tfeas =

0.202923
-0.567788
best value of t:
-0.567788
f-radius saturation: 0.000% of R =

1.00e+09

-0.5678

xfeas = 5.3942
1.0907
2.8129
P = 5.3942
1.0907

1.0907
2.8129

Hence the above LMIs are feasible and the pendulum is exponentially stable with Lyapunov
matrix


5.3942 1.0907
P =
1.0907 2.8129

11.2. POLYTOPIC NONLINEAR SYSTEMS

121

By iterating on , the largest value for which LMIs (11.6) were feasible was found to
be 1.3229. However, from previous considerations, we have shown that this system is
exponentially stable for < 2. Why the difference?
Exercise 32 What is the supremal value of > 0 for which Theorem 16 guarantees that
the following system is guaranteed to be stable about the origin?
2

x 1 = 2x1 + x2 + ex1 x2
2
x 2 = x1 3x2 ex1 x1
Exercise 33 Consider the pendulum system of Example 119 with = 1. Obtain the largest
rate of exponential convergence that can be obtained using the results of Exercise 31 and
the LMI toolbox.

122

11.2.2

CHAPTER 11. QUADRATIC STABILITY

Generalization

One can readily generalize the results of this section to systems described by
x = A(x)x

(11.8)

where the state dependent matrix A(x) has the following structure
A(x) = A0 + 1 (x)A1 + + l (x)Al ,

(11.9)

where A0 , A1 , , Al are constant n n matrices and each i is a scalar valued function

of the state x which satisfies
ai i (x) bi
(11.10)
for some known constants ai and bi .
To obtain a sufficient condition for exponential stability, introduce the following set of 2l
extreme matrices:
A = {A0 + 1 A1 + + l Al : i = ai or bi for i = 1, , l}
For example, with l = 2, A consists of the following four matrices:
A0 + a1 A1 + a2 A2
A0 + a1 A1 + b2 A2
A0 + b1 A1 + a2 A2
A0 + b1 A1 + b2 A2

(11.11)

We have now the following result.

Theorem 17 Suppose there exists a positive-definite symmetric matrix P which satisfies the
following linear matrix inequalities:
P A + AT P < 0

for all A in A

(11.12)

Then system (11.8)-(11.10) is globally exponentially stable about the origin with Lyapunov
matrix P .
Exercise 34 Consider the double inverted pendulum described by
1 + 21 2 + 2k1 k2 sin 1 = 0
2 1 + 2 k1 + k2 sin 2 = 0
Using the results of Theorem 17, obtain a value of the spring constant k which guarantees
that this system is globally exponentially stable about the zero solution.

11.2. POLYTOPIC NONLINEAR SYSTEMS

11.2.3

123

Robust stability

Consider an uncertain system described by

x = f (x, )

(11.13)

where is some uncertain parameter vector. suppose that the only information we have on
is its set of possible values, that is,

where the set is known. So we do not know what is, we only know a set of possible
values of . We say that the above system is robustly stable if it is stable for all allowable
values of , that is, for all .
We say that the uncertain system is quadratically stable if there are positive definite
symmetric matrices P and Q such
xT P f (x, ) xT Qx

(11.14)

for x and . Clearly, quadratic stability guarantees robust exponentially stability about
zero.
One of the useful features of the above concept, is that one can sometimes guarantee that
(11.14) holds for all by just checking that is holds for a finite number of . For, example
consider the linear uncertain system described by
x = A()x .
Suppose that the uncertain matrix A() has the following structure
A() = A0 + 1 A1 + + Al ,

(11.15)

where A0 , A1 , , Al are constant n n matrices and each i is a scalar which satisfies

ai i bi

(11.16)

Introduce the following set of 2l extreme matrices:

A = {A0 + 1 A1 + + l Al : i = ai or bi for i = 1, , l}
Suppose there exists a positive-definite symmetric matrix P which satisfies the following
linear matrix inequalities:
P A + AT P < 0

for all A in A

(11.17)

Then, for all allowable , the uncertain system is globally exponentially stable about the
origin with Lyapunov matrix P .

124

11.3

CHAPTER 11. QUADRATIC STABILITY

Another special class of nonlinear systems

Here we consider nonlinear systems which have the following structure:

x = Ax + B(Cx)

(11.18)

where the n-vector x(t) is the state; B and C are constant matrices with dimensions n m
and p n, respectively; and is a function which satisfies
||(z)|| ||z||

(11.19)

for some 0.
Note that if we introduce a fictitious output z = Cx and a fictitious input w = (Cx),
the nonlinear system (11.18) can be described as a feedback combination of a linear time
invariant system
x = Ax + Bw
z = Cx
and a memoryless nonlinearity
w = (z)
This is illustrated in Figure 11.1.

Figure 11.1: A feedback system

Example 120 Inverted pendulum under linear feedback. Recall the inverted pendulum of
the previous example described by (11.5). This system can be described by the above general
description with

 



0
0
1
C= 1 0
(z) = sin z
B=
A=
1
2 1
Also,

||(z)|| = | sin z| |z| = ||z||

The following theorem provides a sufficient condition for the global exponential stability
of system (11.18).
Theorem 18 Suppose there exists a positive-definite symmetric matrix P which satisfies the
following quadratic matrix inequality (QMI):
P A + AT P + 2 P BB T P + C T C < 0

(11.20)

Then system (11.18)-(11.19) is globally exponentially stable about the origin with Lyapunov
matrix P .

11.3. ANOTHER SPECIAL CLASS OF NONLINEAR SYSTEMS

125

Proof. Along any solution of system (11.18)-(11.19) we have

2xT P x = 2xT P Ax + 2xT P B = xT (P A + AT P )x + 2xT P B .

(11.21)

Note that for any two real numbers a and b,

2ab = a2 + b2 (a b)2 ;
hence
2ab a2 + b2 .
Considering the last term in (11.21), we now obtain
2xT P B

2||B T P x||||||
2||B T P x||||Cx||
2 ||B T P x||2 + ||Cx||2
2 xT P BB T P x + xT C T Cx .

(11.22)

Combining (11.21) and (11.22), it now follows that

2xT P x 2xT Qx
where 2Q = P A + AT P + 2 P BB T P + C T C. By assumption, the matrix Q is positive
definite. Hence system (11.18)-(11.19) is globally exponentially stable about the origin with
Lyapunov matrix P .
A linear matrix inequality (LMI). Using a Schur complement result (see Appendix at
the end of this chapter), one can show that satisfaction of quadratic matrix inequality (11.20)
is equivalent to satisfaction of the following matrix inequality:


P A + AT P + C T C
PB
T
B P
2 I

<0

(11.23)

Note that this inequality is linear in P and 2 . Suppose one wishes to compute the supremal
value of which guarantees satisfaction of the hypotheses of Theorem 18 and hence stability
of system (11.32). This can be achieved by solving the following optimization problem:
minimize subject to


P A + AT P + C T C P B
BT P
I

< 0

0 < P
0 <

1
and then letting = 2 where is the infimal value of . The following program uses the
LMI toolbox to compute for the inverted pendulum example.

126

CHAPTER 11. QUADRATIC STABILITY

%Quadratic stability radius of inverted pendulum

%
A = [0 1; -2 -1];
B=[0;1];
C=[1 0];
%
setlmis([])
%
pvar=lmivar(1,[2,1]);
%P
betavar=lmivar(1,[1,1]);
%beta
%
lmi1=newlmi;
lmiterm([lmi1 1 1 pvar],1,A,s);
%PA+AP
lmiterm([lmi1 1 1 0],C*C);
%CC
lmiterm([lmi1 1 2 pvar],1,B);
%PB
lmiterm([lmi1 2 2 betavar],-1,1);
%-beta I
%
lmi2=newlmi;
lmiterm([-lmi2 1 1 pvar],1,1)
%P
%
lmi3=newlmi;
lmiterm([-lmi3 1 1 betavar],1,1);
%beta
%
lmis=getlmis;
%
c=mat2dec(lmis,0,1);
%specify weighting
options=[1e-5 0 0 0 0];
[copt,xopt]=mincx(lmis,c,options)
%optimize
%
gamma=1/sqrt(copt)
Running this program yields the following output:
Solver for linear objective minimization under LMI constraints
Iterations
1
2
3
***
4
***
5
***

Best objective value so far

0.862772
0.807723
0.675663
new lower bound:
0.645705
new lower bound:
0.626163
new lower bound:

-0.031634
0.148578
0.284167

11.3. ANOTHER SPECIAL CLASS OF NONLINEAR SYSTEMS

6
***

new
7

***

new
8

***

new
9

***

new
10

***

new
11

***

new
12

***

new
13

***

new
14

***
Result:

new

0.598055
lower bound:
0.580987
lower bound:
0.572209
lower bound:
0.571637
lower bound:
0.571450
lower bound:
0.571438
lower bound:
0.571429
lower bound:
0.571429
lower bound:
0.571429
lower bound:

127

0.394840
0.535135
0.551666
0.567553
0.570618
0.571248
0.571391
0.571420
0.571426

feasible solution of required accuracy

best objective value:
0.571429
guaranteed absolute accuracy: 2.86e-06
f-radius saturation: 0.000% of R = 1.00e+09

copt = 0.5714
xopt = 1.1429
0.2857
0.5714
0.5714
gamma = 1.3229
Thus = 1.3229 which is the same as that achieved previously.
A Riccati equation. It should be clear that P satisfies the above QMI if and only if it
satisfies the following Riccati equation for some positive definite symmetric matrix Q:
P A + AT P + 2 P BB T P + C T C + Q = 0

(11.24)

Using properties of QMI (11.20) (see Ran and Vreugdenhil, 1988), one can demonstrate the
following.
Lemma 10 There exists a positive-definite symmetric matrix P which satisfies the quadratic
matrix inequality (11.20) if and only if for any positive-definite symmetric matrix Q there

128

CHAPTER 11. QUADRATIC STABILITY

is an > 0 such that for all (0, ] the following Riccati equation has a positive-definite
symmetric solution for P
P A + AT P + 2 P BB T P + C T C + Q = 0

(11.25)

Using this corollary, the search for a Lyapunov matrix is reduced to a one parameter search.
Exercise 35 Prove the following result: Suppose there exist a positive-definite symmetric
matrix P and a positive scalar which satisfy


P A + AT P + C T C + 2P
PB
0.
(11.26)
BT P
2 I
Then system (11.18)-(11.19) is globally exponentially stable about the origin with rate of
convergence .

11.3.1

The circle criterion: a frequency domain condition

In many situations, one is only interested in whether a given nonlinear system is stable or
not; one may not actually care what the Lyapunov matrix is. To this end, the following
frequency domain characterization of quadratic stability is useful. Consider system (11.18)
by
and define the transfer function G

and let

G(s)
= C(sI A)1 B

(11.27)

:= sup kG(j)k

kGk

(11.28)

IR

Then we have the following result from the control literature.

Lemma 11 (KYP1) [6, 4] There exists a positive-definite symmetric matrix P which satisfies the linear matrix inequality (11.23) if and only if
(i) A is asymptotically stable and
(ii)
<1
kGk

(11.29)

The last lemma and Theorem 18 tell us that , the supremal value of for stability of
system (11.18)-(11.19) via the method of this section, is given by

= 1/||G||
Example 121 Consider Example 120 again. Here the matrix A is asymptotically stable
and

G(s)
= 1/(s2 + s + 2)
One may readily compute that

= 2/ 7 < 1
kGk

=
Hence, this nonlinear system is exponentially stable. Also, = 1/||G||
This is the same as before.

7/2 1.3229.

11.3. ANOTHER SPECIAL CLASS OF NONLINEAR SYSTEMS

129

0.8

0.6

0.4

Imag Axis

0.2

-0.2

-0.4

-0.6

-0.8
-0.8

-0.6

-0.4

-0.2

0
Real Axis

0.2

0.4

0.6

0.8

Figure 11.2: A nice figure

11.3.2

Sector bounded nonlinearities

Suppose is a scalar valued function of a scalar variable (m = p = 1). Then the bound
(11.19) is equivalent to
z 2 z(z) z 2
That is, is a function whose graph lies in the sector bordered by the lines passing through
the origin and having slopes and ; see Figure 11.3. Consider now a function whose

Figure 11.3: A norm bounded nonlinearity

graph lies in the sector bordered by the lines passing through the origin and having slopes a
and b, with a < b; see Figure 11.4. Analytically, this can be expressed as

Figure 11.4: A general sector bounded nonlinearity

az 2 z(z) bz 2

(11.30)

130

CHAPTER 11. QUADRATIC STABILITY

Example 122 The saturation function which is defined below is illustrated in Figure 11.5.

(z) =

z if |z| 1
1 if z < 1

1 if z > 1

Since 0 z(z) z 2 , the saturation function is sector bounded with a = 0 and b = 1.

Figure 11.5: Saturation

The following result is sometimes useful in showing that a particular nonlinear function
is sector bounded.
Fact 4 Suppose is a differentiable scalar valued function function of a scalar variable,
(0) = 0 and for all z
a (z) b
for some scalars a and b. Then satisfies (11.30) for all z.
We can treat sector bounded nonlinearities using the results of the previous section by
introducing some transformations. Specifically, if we introduce

(z)
:= (z) kz

where

k := (a + b)/2 ,

then,

z 2 z (z)
z 2

with

:= (b a)/2

and system (11.32) is described by

+ B (Cx)
x = Ax

(11.31)

with A = A + kBC. Since |(z)|

|z|, we can apply the results of the previous section to
system (11.31) to obtain stability results for the original nonlinear system, which contains
the nonlinearity satisfying (11.30).

11.3. ANOTHER SPECIAL CLASS OF NONLINEAR SYSTEMS

11.3.3

131

Generalization

Here we consider systems whose nonlinearity is characterized by several nonlinear functions,

i : IRpi IRmi , i = 1, 2, . . . , l :
x = Ax + B1 (C1 x) + . . . + Bl l (Cl x)

(11.32)

where A0 and Bi , Ci , i = 1, 2, , l are constant matrices of appropriate dimensions. Each

i satisfies
||i (zi )|| ||zi ||
(11.33)
for some 0.
If we introduce fictitious outputs zi = Ci x and a fictitious inputs wi = i (Ci x), the above
nonlinear system can be described as a feedback combination of a linear time invariant system
x = Ax + B1 w1 + . . . + Bl wl
z1 = C1 x
..
.
zl = Cl x
and multiple memoryless nonlinearities:
wi = i (zi )

i = 1, 2, , l

This is illustrated in Figure 11.6.

Figure 11.6: A multi-loop feedback system

We will use the results of the previous section to obtain results for the systems of this
section. Letting m = m1 + + ml and p = p1 + + pl , suppose we introduce a function
: Rp Rm defined by

z1
1 (z1 )
z2
2 (z2 )

(11.34)
where
z = .. with zi IRpi
(z) =

..
.

.
zl
l (zl )

132

CHAPTER 11. QUADRATIC STABILITY

Then system (11.32) can also be described by (11.18) with

B1 B2 . . . Bl

B :=

C :=

Also,
||(z)||2 =

l
X
i=1

||i (zi )||2

l
X
i=1

C1
C2
..
.
Cl

2 ||zi ||2 = 2 ||z||2

that is ||(z)|| ||z||. We can now apply the results of the previous section to obtain
sufficient conditions for global exponential stability. However, these results will be overly
conservative since they do not take into account the special structure of ; the function is
not just any function satisfying the bound k(z)k kzk; it also has the structure indicated
in (11.34). To take this structure into account, consider any l positive scalars 1 , 2 , . . . , l
and introduce the nonlinear function

1 1 (1
z1
1 z1 )
2 2 (1 z2 )
z2
2

(z) =
where
z = .. with zi IRpi

..

.
1
l l (l zl )
zl

Then system (11.32) can also be described by (11.18), specifically, it can be described by
Cx)
(

x = Ax + B
(11.35)

with
:=
B

Noting that

1
1 B1

||i i (1
i zi )||

1
2 B2

...

i ||(1
i zi )||

||(z)||
=

l
X
i=1

1
l Bl

i ||1
i zi ||

2
||i (1
i zi )||

l
X
i=1

C :=

1 C 1
2 C 2
..
.
l C l

= ||zi ||, we obtain

||zi ||2 = 2 ||z||2

that is, ||(z)||

||z||.
To obtain sufficient conditions for stability of (11.32), we simply apply the results of the
previous section to (11.35). Using Theorem 18 one can readily obtain the following result:
Theorem 19 Suppose there exists a positive-definite symmetric matrix P and l positive
scalars 1 , 2 , . . . , l which satisfy the following matrix inequality:
T

PA + A P +

l
X
i=1

T
2
T
C
1

P
B
B
P
+

C
<0
i
i
i
i
i
i

(11.36)

Then system (11.32) is globally exponentially stable about the origin with Lyapunov matrix
P.

11.3. ANOTHER SPECIAL CLASS OF NONLINEAR SYSTEMS

Proof. If we let i =

133

i for i = 1, 2, , l, then (11.36) can be written as

B
T P + C T C < 0
P A + AT P + 2 P B

Since ||(z)||
||z||, it follows from representation (11.35) and Theorem 18 that system
(11.32) is globally exponentially stable about the origin with Lyapunov matrix P .
An LMI. Note that, using a Schur complement result, inequality (11.36) is equivalent to
the following inequality which is linear in P and the scaling parameters 1 , , l :

P
P A + AT P + li=1 i Ci Ci P B1 P B2 P Bl

B1T P
1 I
0

T
B
P
0

0
2
<0

.
.
.
.
..
..
..
..

T
Bl P
0
0
l I

It should be clear that one may also obtain a sufficient condition involving a Riccati
equation with scaling parameters i using Corollary 10 and a H sufficient condition using
Lemma 11.
Exercise 36 Recall the double inverted pendulum of Example 34. Using the results of this
section, obtain a value of the spring constant k which guarantees that this system is globally
exponentially stable about the zero solution.

134

11.4

CHAPTER 11. QUADRATIC STABILITY

Yet another special class

Here we consider systems described by

x = Ax B(Cx)

(11.37)

z (z) 0

(11.38)

where
for all z. Examples of include (z) = z, z 3 , z 5 , sat(z), sgm (z).

Figure 11.7: A first and third quadrant nonlinearity

Note that such a system can also be regarded as a feedback combination of the LTI
system
x = Ax + Bw
z = Cx
and a memoryless nonlinear system
w = (z) .
Theorem 20 Suppose there exists a positive-definite symmetric matrix P which satisfies
P A + A P < 0
BP = C

(11.39)

Then system (11.37)-(11.38) is globally exponentially stable about the origin with Lyapunov
matrix P .
Proof.
2x P x = x (P A + A P )x 2x P B(Cx)
= x (P A + A P )x 2(Cx) (Cx)
x (P A + A P )x .
Hence
2x P x 2xT Qx

11.4. YET ANOTHER SPECIAL CLASS

135

where 2Q = P A AT P . Since Q is positive definite, we have global exponential stability

about zero.
Now note that P > 0 satisfies (11.39) if and only if the following optimization problem
has a minimum of zero:
Minimize subject to
P A + A P < 0
P > 0


I
BP C
P B C
I

Example 123 Consider

x 1 = 3x1 +x2
x 2 =
x1 3x2 x32

Here
A=

3
1
1 3

B=

0
1

C=

0 1

and (z) = z 3 . Clearly z(z) 0 for all z. Also, conditions (11.39) are assured with P = I.
Hence, the system of this example is globally exponentially stable. Performing the above
optimization with the LMI toolbox yields a minimum of to equal to zero and


22.8241 0
P =
0
1
This P also satisfies conditions (11.39).
Exercise 37 Prove the following result: Suppose there exists a positive-definite symmetric
matrix P and a positive scalar which satisfy which satisfies
P A + A P + 2P 0
BP = C

(11.40a)
(11.40b)

Then system (11.37)-(11.38) is globally exponentially stable about the origin with rate
and with Lyapunov matrix P .

11.4.1

Generalization

Here we consider systems described by

x = Ax + Bw
z = Cx + Dw
w = (z)

(11.41)

z (z) 0

(11.42)

where

136

CHAPTER 11. QUADRATIC STABILITY

for all z. In order for the above system to be well-defined, one must have a unique solution
to the equation
z + D(z) = Cx
for each x.
Theorem 21 Consider a system described by (11.41) and satisfying (11.42). Suppose there
exist a positive-definite symmetric matrix P and a positive scalar which satisfy


P A + A P + 2P P B C
BP C
(D + D )

(11.43)

Then system (11.41) is globally exponentially stable about the origin with rate and Lyapunov
matrix P .
Proof.

2x P x = x (P A + A P )x + 2x P Bw

x
w

 

P A + A P P B
BP
0



We also have

2z (z) = 2(Cx + Dw) w

x
w

 

0
C
C D + D



x
w

It now follows from (11.43) that

x (P A + A P + 2P )x + 2x P Bw 2(Cx + Dw)w 0 ,
that is,
2xT P x + 2xT P x + z (z) 0 .
Hence
2x P x 2x P x z (z)
2x P x .

Note that (11.43) is linear in P .

If D + D > 0, then (11.43) is equivalent to
P A + A P + 2P + (P B C )(D + D )1 (B P C) 0

x
w

11.4. YET ANOTHER SPECIAL CLASS

11.4.2

137

Strictly positive real transfer functions and a frequency

domain condition

given by
We consider here square transfer functions G

G(s)
= C(sI A)1 B + D

(11.44)

where A, B, C, D are real matrices of dimensions n n, n m, m n and m m. We will

is stable if all its poles have negative real parts. Clearly, if A is Hurwitz (all its
say the G
is stable.
eigenvalues have negative real parts) then G
is strictly positive real (SPR) if there
Definition 1 (SPR) A square transfer function G

exists a real scalar > 0 such that G has no poles in a region containing the set {s C :
(s) } and

G()
+ G()
0
for all
IR .
(11.45)

is regular if det[G()+

We say that G
G()
] is not identically zero for all IR.

Requirement (11.45) is not very satisfactory; it involves an apriori unknown parameter

. How do we get rid of ?
is SPR and regular if and only if the following conLemma 12 [5] A transfer function G
ditions hold.
is stable.
(a) [Stability] G
(b) [Dissipativity]

G()
+ G()
>0

for all

IR

(11.46)

(c) [Asymptotic side condition]

lim 2 det[G()
+ G()
] 6= 0

||

(11.47)

where is the nullity of D + D .

In either case, the above limit is positive.

Note that D = G()

:= lim|| G()
and the nullity of D + D is the dimension of
the null space of D + D .

Remark 5 Suppose G(s)

= C(sI A)1 B + D is a scalar valued transfer function. Then
= 0 if D 6= 0 and = 1 is D = 0. Hence, the above side condition reduces to
D 6= 0
When D = 0,

or

lim 2[G()
+ G()
] 6= 0

||

lim 2 [G()
+ G()
] = CAB

||

138

CHAPTER 11. QUADRATIC STABILITY

Example 124 Consider the scalar transfer function

g(s) =

s2

s+1
.
+ bs + 1

We claim that this transfer function is SPR if and only if b > 1. This is illustrated in Figures
11.8 and 11.9 for b = 2 and b = 0.5 respectively.
Nyquist Diagram
0.5
0.4
0.3

Imaginary Axis

0.2
0.1
0
0.1
0.2
0.3
0.4
0.5
1

0.8

0.6

0.4

0.2

0
Real Axis

0.2

0.4

0.6

Figure 11.8: Nyquist plot of

0.8

s+1
s2 +2s+1

Nyquist Diagram
2.5
2
1.5

Imaginary Axis

1
0.5
0
0.5
1
1.5
2
2.5
1

0.5

0.5

1
Real Axis

1.5

Figure 11.9: Nyquist plot of

2.5

s+1
s2 +0.5s+1

To prove the above claim, we first note that g is stable if and only if b > 0. Now note
that
1 +
g() =
1 2 + b
Hence
2[1 + (b 1) 2]
g() + g() =
(1 2 )2 + (b)2

It should be clear from the above expression that g() + g() > 0 for all finite if and
only if b 1. Also,
lim 2[
g () + g()] = b 1 .

Thus, in order for the above limit to be positive, it is necessary and sufficient that b > 1.

11.4. YET ANOTHER SPECIAL CLASS

139

The next lemma tells us why SPR transfer functions are important in stability analysis
of the systems under consideration in the last section.
Lemma 13 (KYPSPR) Suppose (A, B) is controllable and (C, A) is observable. Then
there exists a matrix P = P > 0 and scalar > 0 such that LMI (11.43) holds if and only

if the transfer function G(s)

= D + C(sI A)1 B is strictly positive real.
Corollary 1 Consider a system described by (11.41) and satisfying (11.42) and suppose that
(A, B) is controllable, (C, A) is observable and the transfer function C(sI A)1 B + D is
SPR. Then this system is globally exponentially stable about the origin.
Remark 6 Sometimes the KYPSPR Lemma is stated as follows: Suppose (A, B) is control
lable and (C, A) is observable. Then G(s)
= D + C(sI A)1 B is strictly positive real if and

only if there exists a matrix P = P > 0, a positive scalar and matrices W and L such that
P A + A P = L L P
P B = C L W
W W = D + D

(11.48)
(11.49)
(11.50)

To see that this result is equivalent to Lemma 13 above, let = /2 and rewrite the
above equations as
P A + A P + 2P = L L
P B C = L W
(D + D ) = W W ,
that is,
M :=

P A + A P + 2P P B C
BP C
(D + D )

L L L W
W L W W

L
W

 

L
W

The existence of matrices L and W satisfying


 

L
L
M =
W
W
is equivalent to M 0.
Exercise 38 Consider the transfer function
g(s) =

s + 1
s2 + s + 2

Using Lemma 12, determine the range of for which this transfer function is SPR. Verify
your results with the KYPSPR lemma.

140

CHAPTER 11. QUADRATIC STABILITY

Appendix: A Schur complement result

The following result is useful in manipulating matrix inequalities. Suppose


Q11 Q12
Q=
Q21 Q22
is a hermitian 2 2 block matrix. Then
Q>0
if and only if
Q22 > 0

and

Q11 Q12 Q1
22 Q21 > 0

To see this, note that Q22 = Q22 , Q21 = Q12 and



 
 

I
0
Q11 Q12 Q1
0
I
0
Q11 Q12
22 Q21
=
Q1
0
Q22
Q1
Q21 Q22
22 Q21 I
22 Q21 I
Note that Q11 and Q22 do not have to have the same dimensions.
Clearly, we can readily obtain the following result:
Q<0
if and only if
Q22 < 0

and

Q11 Q12 Q1
22 Q21 < 0

The following result also follows from (11.51). Suppose Q22 > 0. Then
Q0
if and only if
Q11 Q12 Q1
22 Q21 0

(11.51)

Bibliography
[1] Boyd, S. and El Ghaoui, L. and Feron, E. and Balakrishnan, V., Linear Matrix Inequalities in System and Control Theory, SIAM, Philadelphia, 1994.
[2] Gahinet, P., Nemirovski, A., Laub, A.J., and Chilali, M., LMI Contol Toolbox Users
Guide, The MathWorks Inc., Natick, Massachussets, 1995.
[3] Corless, M., Robust Stability Analysis and Controller Design With Quadratic Lyapunov Functions, in Variable Structure and Lyapunov Control, A. Zinober, ed.,
Springer-Verlag, 1993.
[4] Ackmese, A.B., and Corless, M., Stability Analysis with Quadratic Lyapunov Functions: Some Necessary and Sufficient Multiplier Conditions, Systems and Control Letters, Vol. 57, No. 1, pp. 78-94, January 2008.
[5] Corless, M. and Shorten, R., A correct characterization of strict positive realness,
submitted for publication.
[6] A. Rantzer, On the Kalman-Yakubovich-Popov lemma, Systems & Control Letters,
vol. 28, pp. 7-10, 1996.
[7] Shorten, R., Corless, M., Wulff, K., Klinge, S. and R. Middleton, R., Quadratic Stability and Singular SISO Switching Systems, submitted for publication.
[8] Ran, A.C.M., R. Vreugdenhil, R. 1988, Existence and comparison theorems for algebraic
Riccati equations for continuous- and discrete-time systems, Linear Alg. Appl. Vol. 99,
pp. 6383, 1988

141

142

BIBLIOGRAPHY

Chapter 12
Invariance results
Recall the simple damped oscillator described by
x 1 = x2
x 2 = (k/m)x1 (c/m)x2
with m, c, k positive. This system is globally asymptotically stable about the origin. If we
consider the total mechanical energy, V (x) = 21 kx21 + 21 mx22 , of this system as a candidate
Lyapunov function, then along any solution we obtain
V (x) = cx22 0 .
Using our Lyapunov results so far, this will only guarantee stability; it will not guarantee
asymptotic stability, because, regardless of the value of x1 , we have V (x) = 0 whenever
x2 = 0. However, one can readily show that there are no non-zero solutions for which x2 (t)
is identically zero. Is this sufficient to guarantee asymptotic stability? We shall shortly see
that it is sufficient; hence we can use the energy of this system to prove asymptotic stability.
Consider a system described by
x = f (x)
(12.1)
By a solution of system (12.1) we mean a continuous function x() : [0, ) IRn which
identically satisfies x(t)

= f (x(t)).

12.1

Invariant sets

A subset M of the state space IRn is an invariant set for system (12.1) if it has the following
property.
Every solution which originates in M remains in M for all future time, that is, for every
solution x(),
x(0) M
implies
x(t) M for all t 0
The simplest example of an invariant set is a set consisting of a single equilibrium point.
143

144

CHAPTER 12. INVARIANCE RESULTS

As another example of an invariant set, consider any linear time-invariant system x = Ax

and let v be any real eigenvector of the matrix A. Then the line consisting of all vectors of
the form cv where c is any real number is an invariant set for the linear system. If v = u + w
is a complex eigenvector with u and w real then, the 2-dimensionsal subspace spanned by u
and w is an invariant set for the linear system.

Figure 12.1: Invariant sets

Lemma 14 Suppose
DV (x)f (x) 0

when

V (x) < c

Then
M = {x : V (x) < c}
is an invariant set.

Figure 12.2: Invariant set lemma

Example 125
x 1 = x1 + x31
Considering V (x) = x2 and c = 1, one can readily show that the interval (1, 1) is an
invariant set.

12.2. LIMIT SETS

145

Largest invariant set. It should be clear that the union and intersection of two invariant
sets is also invariant. Hence, given any subset S of IRn , we can talk about the largest invariant
set M contained in S, that is M is an invariant set and contains every other invariant set
which is contained in S. As an example, consider the Duffing system
x 1 = x2
x 2 = x1 x31
and let S correspond the the set of all points of the form (x1 , 0) where x1 is arbitrary. If a
solution starts at one of the equilibrium states
(1, 0)

(0, 0)

(1, 0)

it remains there. Hence, the set

{(1, 0), (0, 0), (1, 0)}
is an invariant set for the system under consideration. We claim that this is the largest
invariant set in S. To see this, let M be the largest invariant set in S and consider any
solution (x1 (), x2 ()) which lies in M. Then this solution lies in S from which it follows that
x2 (t) is zero for all t. This implies that x 2 (t) is zero for all t. From the system description,
we must have x1 (t) x31 (t) = 0. Hence x1 (t) must be 1, 0 , or 1.

Figure 12.3: Largest invariant set

12.2

Limit sets

A state x is called a positive limit point of a solution x() if there exists a sequence {tk }
k=0
of times such that:
lim tk =

lim x(tk ) = x

The set of all positive limit points of a solution is called the positive limit set of the solution.
The solution x(t) = et has only zero as a positive limit point, whereas x(t) = cos t has
the interval [1, 1] as its positive limit set. What about x(t) = et cos t? The positive limit

146

CHAPTER 12. INVARIANCE RESULTS

set of x(t) = et is empty. Note that the positive limit set of a periodic solution consists of
all the points of the solution.
Suppose a solution converges to a single state x ; then the positive limit set of this solution
is the set consisting of that single state, that is, {x }. For another example, consider the
Van der Pol oscillator. Except for the zero solution, the positive limit set of every solution
is the set consisting of all the states in the limit cycle.
Exercise 39 What are the positive limit sets of the following solutions?
(a) x(t) = sin(t2 )
(b) x(t) = et sin(t)
Distance of a point x from a set M:
d(x, M) := inf{||x y|| : y M}

Figure 12.4: Distance of a point from a set

Convergence to a set: A solution x() converges to a set M if

lim d(x(t), M) = 0

Figure 12.5: Convergence to a set

12.2. LIMIT SETS

147

Lemma 15 (Positive limit set of a bounded solution.) The positive limit set L+ of a
bounded solution to (12.1) is nonempty and has the following properties.
1. It is closed and bounded.
2. It is an invariant set.
3. The solution converges to L+ .
Proof. See Khalil.

148
Example 126 (Van der Pol oscillator)

CHAPTER 12. INVARIANCE RESULTS

12.3. LASALLES THEOREM

12.3

149

LaSalles Theorem

In the following results, V is a scalar valued function of the state, that is, V : IRn IR, and
is continuously differentiable.
Lemma 16 (A convergence result for bounded solutions.) Suppose x() is a bounded
solution of (12.1) and there exists a function V such that
DV (x(t))f (x(t)) 0
for all t 0. Then x() converges to the largest invariant set M contained in the set
S := {x IRn | DV (x)f (x) = 0}
Proof. Since x() is a bounded solution of (12.1), it follows from the previous lemma that
it has a nonempty positive limit set L+ . Also, L+ is an invariant set for (12.1) and x()
converges to this set.
We claim that V is constant on L+ . To see this, we first note that
d
(V (x(t)) = DV (x(t))f (x(t) 0 ,
dt
hence, V (x(t)) does not increase with t. Since x() is bounded and V is continuous, it follows
that V (x()) is bounded below. It now follows that there is a constant c such that
lim V (x(t)) = c .

Consider now any member x of L+ . Since L+ is the positive limit set of x(), there is a
sequence {tk }
k=0 such that
lim tk =

and

lim x(tk ) = x .

Since V is continuous,
V (x ) = lim V (x(tk )) = c .
k

We have now shown that V (x ) = c for all x in L+ . Thus V is constant on L+ .

Since L+ is an invariant set for (12.1) and V is constant on L+ , it follows that DV (x )f (x )
is zero for all x in L+ . Hence L+ is contained in S and, since L+ is an invariant set, it must
be contained in the largest invariant set M in S.
Finally, since x() converges to L+ and L+ is contained in M, the solution x() also
converges to M.
The above result tells us that the solution converges to largest invariant set which is
contained in the set in which V is zero. Thus, the solution also converges to the set in which
V is zero.
Remark 7 Note that the above result does not require that V have any special properties
such as positive definiteness or radial unboundedness.

150

CHAPTER 12. INVARIANCE RESULTS

Example 127 Consider the damped simple pendulum described by

x 1 = x2
x 2 = sin x1 x2 .
Suppose V is the total mechanical energy of this system, that is,
1
V (x) = 1 cos x1 + x22 .
2
Then V (x) = x22 0. Since V (x) = 0 is equivalent to x2 = 0, it follows that every bounded
solution of this system converges to the largest invariant set M in
S = {(x1 , 0) : x1 is an arbitrary real number } .
Consider any solution x() contained in M. Since this solution is in S, we must have x2 (t) = 0
for all t. Hence x 2 (t) is zero for all t. It now follows from the system description that
sin x1 (t) = 0

for all t .

Thus, x1 (t) must belong to the set E of equilibrium states, that is,
E = {(n, 0) : n is an integer } .
Thus M is contained in E. Since E is invariant, E = M. Thus all bounded solutions converge
to E.
Theorem 22 (LaSalles Theorem.) Suppose there is a radially unbounded function V
such that
DV (x)f (x) 0
for all x. Then all solutions of (12.1) are bounded and converge to the largest invariant set
M contained in the set
S := {x IRn | DV (x)f (x) = 0}
Proof. We have already seen that the hypothesis of the theorem guarantee that all
solutions are bounded. The result now follows from Lemma 16.

Example 128 The damped mass.

x 1 = x2
x 2 = x2
Consider

1
V (x) = x21 + x1 x2 + x22
2

12.3. LASALLES THEOREM

151

Figure 12.6: The damped mass

Since V is a positive definite quadratic form, it is radially unbounded. Also,
V (x) = x22 0
Hence all solutions approach the set

T
M = S = { x1 0
: x1 IR}

that is, the x1 axis.

Figure 12.7: Phase portrait of the damped mass

Example 129 (Damped Duffing oscillator.)

x 1 = x2
x 2 = x1 x31 cx2
with c > 0.

Exercise 40 Using LaSalles Theorem, show that all solutions of the system
x 1 = x22
x 2 = x1 x2
must approach the x1 axis.

152

CHAPTER 12. INVARIANCE RESULTS

Exercise 41 Consider the scalar nonlinear mechanical system

q + c(q)
+ k(q) = 0
If the term c(q)
is due to damping forces it is reasonable to assume that c(0) = 0 and
c(q)
q > 0

for all q 6= 0

Suppose the term k(q) is due to conservative forces and define the potential energy by
Z q
P (q) =
k() d
0

Show that if limq P (q) = , then all motions of this system must approach one of its
equilibrium positions.

12.4

Asymptotic stability and LaSalle

12.4.1

Global asymptotic stability

Using LaSalles Theorem, we can readily obtain the following result on global asymptotic
stability. This result does not require V to be negative for all nonzero states.
Theorem 23 (Global asymptotic stability) Suppose there is a positive definite function
V with
DV (x)f (x) 0
for all x and the only solution for which
DV (x(t))f (x(t)) 0
is the zero solution. Then the origin is a globally asymptotically stable equilibrium state.
Proof.
Example 130 A damped nonlinear oscillator
x 1 = x2
x 2 = x31 cx2
with c > 0.
Example 131 A nonlinearly damped oscillator
x 1 = x2
x 2 = x1 cx32
with c > 0.

12.4. ASYMPTOTIC STABILITY AND LASALLE

153

Exercise 42 Consider a nonlinear mechanical system described by

m
q + cq + k(q) = 0
where q is scalar, m, c > 0 and k is a continuous function which satisfies
k(0) = 0
k(q)q > 0
lim

for all q 6= 0

k() d =

(a) Obtain a state space description of this system.

(b) Prove that the state space model is GUAS about the state corresponding to the equilibrium position q = 0.
(i) Use a La Salle type result.
(ii) Do not use a La Salle type result.

Figure 12.8: Single link manipulator

Exercise 43 Consider the inverted pendulum (or one link manipulator) as illustrated where
u is a control torque. This system can be described by
q a sin q = bu
with a = mgl/I, b = 1/I, where m is the mass of B, I is the moment of inertia of B about
its axis of rotation through O, l is the distance between O and the mass center of B, and g
is the gravitational acceleration constant of YFHB. We wish to stabilize this system about
the position corresponding to q = 0 by a linear feedback controller of the form
u = kp q kd q
Using the results of the last problem, obtain the least restrictive conditions on the controller gains kp , kd which assure that the closed loop system is GUAS about the state corresponding to q(t) 0. Illustrate your results with numerical simulations.

154

12.4.2

CHAPTER 12. INVARIANCE RESULTS

Asymptotic stability

Theorem 24 (Asymptotic stability.) Suppose there is a locally positive definite function

V and a scalar R > 0 such that
DV (x)f (x) 0

for ||x|| R

and the only solution for which

DV (x(t))f (x(t)) 0

and ||x(t)|| R

is the zero solution. Then the origin is an asymptotically stable equilibrium state.
Example 132 Damped simple pendulum.
x 1 = x2
mgl
c
x 2 =
sin x1 x2
I
I
with c > 0.

12.5. LTI SYSTEMS

12.5

155

LTI systems
x = Ax

Theorem 25 The following statements are equivalent.

(a) The system x = Ax is asymptotically stable.
(b) There exist a positive definite hermitian matrix P and a matrix C with (C, A) observable which satisfy the Lyapunov equation
P A + A P + C C = 0

(12.2)

(c) For every matrix C with (C, A) observable, the Lyapunov equation (12.2) has a unique
solution for P and this solution is hermitian positive-definite.

156

CHAPTER 12. INVARIANCE RESULTS

12.6

Applications

12.6.1

A simple stabilizing adaptive controller

Consider a simple scalar system described by

x = ax + bu

(12.3)

where the state x and the control input u are scalars. The scalars a and b are unknown but
constant. The only information we have on these parameters is that
b > 0.
We wish to obtain a controller which guarantees that, regardless of the values of a and b, all
solutions of the resulting closed loop system satisfy
lim x(t) = 0 .

To achieve the above goal, we propose the following adaptive controller

u = kx

k = x2

(12.4a)
(12.4b)

where is some positive scalar.

We now show that the above controller results in the desired goal and in addition, the
adaptive parameter k is bounded. Let k any scalar for which
a
:= a bk < 0 .
Letting k = k k, the closed loop system resulting from the proposed adaptive controller
can be described by
x = ax bkx
k = x2

(12.5a)
(12.5b)

As a candidate Lyapunov function, consider,

V (x, k) = x2 + (b/)(k)2
Along any solution of the closed loop system (12.5), we have
V

= 2xx + 2(b/)k k
= 2ax2 2x2 bk + 2bkx2
= 2ax2 0 .

Since b and are positive, V is radially unbounded; also V 0. It now follows that x and k
are bounded. Also, all solutions approach the largest invariant set in the set corresponding
to x = 0. This implies that x(t) approaches zero as t goes infinity.

12.6. APPLICATIONS

157

Exercise 44 (a) Consider an uncertain system described by

x = 3x + sin x + u

(12.6)

where t IR is time, x(t) IR is the state, and u(t) IR is the control input. The
uncertainty in the system is due to the unknown constant real scalar parameter . We wish
to design a controller generating u(t) which assures that, for all initial conditions and for all
IR, the closed-loop system satisfies
lim x(t) = 0

(12.7)

If were known, one could let

u = sin x
to obtain a GUAS closed-loop system and, hence, the desired behavior. Since is unknown,
we propose the following parameter adaptive controller
u = sin x

= x sin x

(12.8a)
(12.8b)

where > 0 is a constant.

(i) Prove that the closed-loop system (12.6),(12.8) has property (12.7).
(ii) Numerically simulate the open loop system (system (12.6) with u = 0) and the closed
loop system ((12.6),(12.8)) for different initial conditions, and for different values of
and .
(b) Consider now an uncertain described by
x = f (x) + g(x) + u

(12.9)

where t, x, u, are as before and f and g are continuous functions with f satisfying
f (x) > 0 if x < 0
f (x) < 0 if x > 0
Design a controller generating u which assures that, for all initial conditions and for all ,
the resulting closed-loop system has property (12.7). Prove that your design works.

12.6.2

PI controllers for nonlinear systems

Here we consider systems described by

x = f (x) + g(x)(u + w)

(12.10)

where the n-vector x(t) is the state, the m-vector u(t) is the control input while the constant
m-vector w is an unknown constant disturbance input.

158

CHAPTER 12. INVARIANCE RESULTS

We assume that there is a positive definite Lyapunov function V such that

DV (x)f (x) < 0

for all x 6= 0 .

This implies that the nominal uncontrolled undisturbed system

x = f (x)
is GAS about zero.
We wish to design a controller for u which assures that, for every constant disturbance
input, one has
lim x(t) = 0 .
t

To this end, we propose the following PI-type controllers:

Z t
u(t) = KP y(t) KI
y( ) d
where y = g(x) DV (x)

(12.11)

and the gain matrices KP and KI are arbitrary positive definite symmetric matrices. We
claim that these controllers achieve the desired behavior.
To prove our claim, note that the closed loop system can be described by
x = f (x) g(x)KP y + g(x)(w xc )
x c = KI y
y = g(x) DV (x)
and consider the candidate Lyapunov function
1
W (x, xc ) = V (x) + (xc w) KI1 (xc w) .
2

Exercise 45 Consider the system of exercise 43 subject to an unknown constant disturbance

torque w, that is,
q a sin q = b(u + w) .
Design a controller which assures that, for any initial conditions, the angle q approaches
some desired constant angle q e as time goes to infinity. Consider a controller structure of
the form:
Z
e
u = kp (q q ) kd q kI (q q e ) .

Prove your design works and illustrate your results with numerical simulations.

Chapter 13
Stability of nonautonomous systems
Here, we are interested in the stability properties of systems described by
x = f (t, x)

(13.1)

where the state x(t) is an n-vector at each time t. By a solution of the above system we mean
any continuous function x() : [t0 , t1 ) IRn (with t0 < t1 ) which satisfies x(t)

= f (t, x(t))
for t0 t < t1 . We refer to t0 as the initial time associated with the solution.
We will mainly focus on stability about the origin. However, suppose one is interested in
stability about some nonzero solution x(). Introducing the new state e(t) = x(t) x(t), its
evolution is described by
e = f(t, e)
(13.2)
with f (t, e) = f (t, x(t) + e) x (t). Since e(t) = 0 corresponds to x(t) = x(t), one can study
the stability of the original system (13.1) about x() by studying the stability of the error
system (13.2) about the origin.
Scalar linear systems. All solutions of
x = a(t)x
satisfy

Rt

x(t) = e

t0

a( ) d

x(t0 ) .

Consider
x = et x

Here a(t) = et . Since a(t) < 0 for all t, one might expect that all solutions converge to
zero. This does not happen. Since,
x(t) = e(e

t et0 )

we have
lim x(t) = ee

t0

x(t0 ) .

Hence, whenever x(t0 ) 6= 0, the solution does not converge to zero.

159

160

CHAPTER 13. STABILITY OF NONAUTONOMOUS SYSTEMS

13.1

Stability and boundedness

13.1.1

Boundedness of solutions

DEFN. Boundedness. A solution x() is bounded if there exists 0 such that

||x(t)||

for all

t t0

A solution is unbounded if it is not bounded.

DEFN. Global uniform boundedness (GUB). The solutions of (13.1) are globally uniformly
bounded if for each initial state x0 there exists a scalar (x0 ) such that for all t0
x(t0 ) = x0

||x(t)|| (x0 )

for all

t t0

Note that, in the above definition, the bound (x0 ) is independent of the initial time t0 .

13.1.2

Stability

Suppose the origin is is an equilibrium state of system (13.1), that is,

f (t, 0) = 0
for all t.
DEFN. Stability (S). The origin is stable if for each > 0 and each t0 there exists > 0 such
that
||x(t0 )|| < = ||x(t)|| < for all t t0

Figure 13.1: Stability

The origin is unstable if it is not stable.

13.1. STABILITY AND BOUNDEDNESS

161

Note that, in the above definition, the scalar may depend on the initial time t0 . When
can be chosen independent of t0 , we say that the stability is uniform.
DEFN. Uniform stability (US). The origin is uniformly stable if for each > 0 there exists
> 0 such that for all t0 ,
||x(t0 )|| <

||x(t)|| <

for all t t0

Example 133 Stable but not uniformly stable. Consider the scalar linear time-varying system:
x = (6t sin t 2t)x

We will analytically show that this system is stable about the origin. We will also show that
not only is it not uniformly stable about zero, the solutions are not uniformly bounded. If
one simulates this system with initial condition x(t0 ) = x0 for different initial times t0 while
keeping x0 the same one will obtain a sequence of solutions whose peaks go to infinity as t0
goes to infinity. This is illustrated in Figure 13.2.

Figure 13.2: Growing peaks

All solutions x() satisfy
x(t) = x(t0 )e[(t)(t0 )]

where

(t) = 6 sin t 6t cos t t2

Clearly, for each initial time t0 , there is a bound (t0 ) such that
e[(t)(t0 )] (t0 )

for all t t0

Hence every solution x() satisfies |x(t)| (t0 )|x(t0 )| for t t0 and we can demonstrate
stability of the origin by choosing = /(t0 ) for any > 0.
To see that this system is not uniformly stable consider t0 = 2n for any nonnegative
integer n. Then (t0 ) = 12n 4n2 2 and for t = (2n + 1) we have (t) = 6(2n + 1)
(2n + 1)2 2 . Hence
(t) (t0 ) = (6 )(4n + 1)
So for any solution x() and any nonnegative integer n, we have
x(2n + ) = x(t0 )e(6)(4n+1)
So, regardless of how small a nonzero bound one places on x(t0 ), one cannot place a bound
on x(t) which is independent of t0 . Hence the solutions of this system are not uniformly
bounded. Also, this system in not uniformly stable about the origin.

162

CHAPTER 13. STABILITY OF NONAUTONOMOUS SYSTEMS

13.1.3

Asymptotic stability

DEFN. Asymptotic stability (AS). The origin is asymptotically stable if

(a) It is stable.
(b) For each t0 , there exists R(t0 ) > 0 such that
||x(t0 )|| < R(t0 )

lim x(t) = 0

(13.3)

Figure 13.3: Asymptotic stability

Note that, in the above definition, R may depend on t0 .
DEFN. Uniform asymptotic stability (UAS). The origin is uniformly asymptotically stable if
(a) It is uniformly stable.
(b) There exists R > 0 such that the following holds for each initial state x0 with ||x0 || < R.
i) There exists (x0 ) such that for all t0 ,
x(t0 ) = x0

||x(t)|| (x0 )

for all t t0

ii) For each > 0, there exists T (, x0 ) 0 such that for all t0 ,
x(t0 ) = x0

13.1.4

||x(t)|| <

for all

t > t0 + T (, x0 )

Global asymptotic stability

DEFN. Global asymptotic stability (GAS). The origin is globally asymptotically stable if
(a) It is stable.
(b) Every solution satisfies
lim x(t) = 0

13.1. STABILITY AND BOUNDEDNESS

163

DEFN. Global uniform asymptotic stability (GUAS). The origin is globally uniformly asymptotically stable if
(a) It is uniformly stable.
(b) The solutions are globally uniformly bounded.
(c) For each initial state x0 and each each > 0, there exists T (, x0 ) 0 such that for all
t0 ,
x(t0 ) = x0 = ||x(t)|| <
for all t > t0 + T (, x0 )

13.1.5

Exponential stability

DEFN. Uniform exponential stability (UES). The origin is uniformly exponentially stable with
rate of convergence > 0 if there exists R > 0 and > 0 such that whenever ||x(t0 )|| < R
one has
||x(t)|| ||x(t0 )||e(tt0 )
for all t t0
Note that exponential stability implies asymptotic stability, but, in general, the converse is
not true.
DEFN. Global uniform exponential stability (GUES). The origin is globally exponentially stable
with rate of convergence > 0 if there exists > 0 such that every solution satisfies
||x(t)||

||x(t0 )||e(tt0 )

for all t t0

164

CHAPTER 13. STABILITY OF NONAUTONOMOUS SYSTEMS

Chapter 14
Lyapunov theory for nonautonomous
systems
14.1

Introduction

Here, we are interested in the stability properties of the system,

x = f (t, x)

(14.1)

where x(t) IRn and t IR. We will focus on stability about the origin. However, suppose
one is interested in stability about some nonzero solution x(). Introducing the new state
e(t) = x(t) x(t), its evolution is described by
e = f(t, e)

(14.2)

with f (t, e) = f (t, x(t) + e) x (t). Since e(t) = 0 corresponds to x(t) = x(t), one can study
the stability of the original system (14.1) about x() by studying the stability of the error
system (14.2) about the origin.
In this chapter, V is a scalar-valued function of the time and state, that is, V : IRIRn
IR. Suppose V is continuously differentiable. Then, at any time t, the derivative of V along
a solution x() of system (14.1) is given by
dV (t, x(t))
V
V
=
(t, x(t)) +
(t, x(t))x(t)

dt
t
x
=

V
V
(t, x(t)) +
(t, x(t))f (t, x(t))
t
x

is the partial derivative of V with respect to t and V

is the partial derivative of
where V
t
x
V with respect to x and is given by


V
V
V
V
=
...
x
x1 x2
xn
Note that

V
V
+
f
t
x

V
V
V
V
+
f1 +
f2 + . . . +
fn
t
x1
x2
xn
165

166

CHAPTER 14. LYAPUNOV THEORY FOR NONAUTONOMOUS SYSTEMS

In what follows, if a condition involves

continuously differentiable.

V
t

or

V
,
x

then it is implicitly assumed that V is

14.2. STABILITY

167

14.2

Stability

14.2.1

Locally positive definite functions

A function V is said to be locally positive definite (lpd) if there is a locally positive definite
function V1 : IRn IR and a scalar R > 0 such that
V1 (x) V (t, x)

for all t and for ||x|| < R

A function V is said to be locally decresent (ld) if there is a locally positive definite function
V2 : IRn IR and a scalar R > 0 such that
V (t, x) V2 (x)

for all t and for ||x|| < R

Example 134
(a)
V (t, x) = (2 + cos t)x2
lpd and ld
(b)
2

V (t, x) = et x2
ld but not lpd
(c)
V (t, x) = (1 + t2 )x2
lpd but not ld

14.2.2

A stability theorem

Theorem 26 (Uniform stability) Suppose there exists a function V with the following
properties.
(a) V is locally positive definite and locally decresent.
(b) There is a scalar R > 0 such that
V
V
(t, x) +
f (t, x) 0
t
x

for all t and for ||x|| < R

Then the origin is a uniformly stable equilibrium state for x = f (t, x).
If V satisfies the hypotheses of the above theorem, then V is said to be a Lyapunov
function which guarantees the stability of origin for system (14.1).
Example 135
x = a(t)x
where a(t) 0 for all t.

168

CHAPTER 14. LYAPUNOV THEORY FOR NONAUTONOMOUS SYSTEMS

14.3

Uniform asymptotic stability

Theorem 27 (Uniform asymptotic stability) Suppose there exists a function V with

the following properties.
(a) V is locally positive definite and locally decresent.
(b) There is a continuous function W and a scalar R > 0 such that for all t and all nonzero
x with ||x|| < R,
V
V
(t, x) +
(t, x)f (t, x) W (x) < 0
t
x
Then the origin is an uniformly asymptotically stable equilibrium state for system (14.1).
Example 136 To see that V < 0 is not sufficient, consider
x = et x
With V (x) = x2 , we have V = 2et x2 < 0 for all nonzero x. However, with x(0) = x0 6= 0,
we have
t
x(t) = e(e 1) x0
Hence limt x(t) = e1 x0 6= 0 and this system is not asymptotically stable about zero.
Example 137
x = (2 + sin t)x + x2

14.4. GLOBAL UNIFORM ASYMPTOTIC STABILITY

14.4

Global uniform asymptotic stability

14.4.1

Positive definite functions

169

A function V is said to be positive definite (pd) if there is a positive definite function V1 :

IRn IR such that
V1 (x) V (t, x)
for all t and x
A function V is said to be decresent (d) if there is a positive definite function V2 such that
V (t, x) V2 (x)

for all t and x

Example 138 Recall Example 134.

14.4.2

A result on global uniform asymptotic stability

Theorem 28 (Global uniform asymptotic stability) Suppose there exists a function V

with the following properties.
(a) V is positive definite and decresent.
(b) There is a continuous function W such that for all t and all nonzero x,
V
V
(t, x) +
f (t, x) W (x) < 0
t
x
Then the origin is a globally uniformly asymptotically stable equilibrium state for x = f (t, x).
Example 139
1
x 1 = x1 (sin t)x2
4
x 2 = x1 x2

As a candidate Lyapunov function, consider

V (t, x) = x21 + (1 +

1
sin t)x22
4

Since

5
3
x21 + x22 V (t, x) x21 + x22
4
4
it follows that V is positive definite and decresent. Along any solution of the above system,
we have
1
1
1
V =
cos t x22 + 2x1 (x1 sin t x2 ) + (2 + sin t)x2 (x1 x2 )
4
4
2
1
1
= 2x21 (2 + sin t cos t)x22 + 2x1 x2
2
4
W (x)

where

5
W (x) = 2x21 2x1 x2 + x22
4
Since W (x) is positive for all nonzero x, it follows from the previous theorem that the above
system is globally uniformly stable about the origin.

170

CHAPTER 14. LYAPUNOV THEORY FOR NONAUTONOMOUS SYSTEMS

14.4.3

Proof of theorem 28

In what follows x() represents any solution and t0 is its initial time.
Uniform stability. Consider any > 0. We need to show that there exists > 0 such
that whenever ||x(t0 )|| < , one has ||x(t)|| < for all t t0 .
Since V is positive definite, there exists a positive definite function V1 such that
V1 (x) V (t, x)
for all t and x. Since V1 is positive definite, there exists c > 0 such that
V1 (x) < c

||x|| <

Since V is decresent, there exists a positive definite function V2 such that

V (t, x) V2 (x)
for all t and x. Since V2 is continuous and V (0) = 0, there exists > 0 such that
||x|| <

V2 (x) < c

Along any solution x(), we have

dV (t, x(t))
0
dt
for all t t0 ; hence
V1 (x(t)) V (t, x(t)) V (t0 , x(t0 )) V2 (x(t0 ))

for t t0 .

that is,
V1 (x(t)) V2 (x(t0 )) t t0

It now follows that if ||x(t0 )|| < , then for all t t0 :

V1 (x(t)) V2 (x(t0 )) < c

that is, V1 (x(t)) < c; hence, ||x(t)|| < .
Global uniform boundedness. Consider any initial state x0 . We need to show that
there is a bound such that whenever x(t0 ) = x0 , one has ||x(t)|| for all t t0 .
Since V1 is a radially unbounded function, there exists > 0 such that
V1 (x) V2 (x0 )

||x||

We have already seen that along any solution x() we have

V1 (x(t)) V2 (x(t0 )) t t0
It now follows that if x(t0 ) = x0 , then for all t t0 :
V1 (x(t)) V2 (x0 ) ;
hence ||x(t)|| .

14.4. GLOBAL UNIFORM ASYMPTOTIC STABILITY

171

Global uniform convergence. Consider any initial state x0 and any > 0. We need
to show that there exists T 0 such that whenever x(t0 ) = x0 , one has ||x(t)|| < for all
t t0 + T .
Consider any solution with x(t0 ) = x0 . From uniform stability, we know that there exists
> 0 such that for any t1 t0 ,
||x(t1 )|| <

||x(t)|| < for t t1

We have shown that there exists 0 such that

||x(t)|| for t t0
We also have

for all t t0 . Let

dV (t, x(t))
W (x(t))
dt
:= min{W (x) : ||x|| }

Since W is continuous and W (x) > 0 for x 6= 0, the above minimum exists and > 0. Let
T := V2 (x0 )/
We now show by contradiction, there exists t1 with t0 t1 t2 , t2 := t0 + T , such that
||x(t1 )|| < ; from this it will follow that ||x(t)|| < for all t t1 and hence for t t0 + T .
So, suppose, on the contrary, that
||x(t)||

for t0 t t2

Then V1 (x(t2 )) > 0, and for t0 t t2 ,

||x(t)|| ;
hence

dV (t, x(t))

dt

So,
V1 (x(t2 )) V (x(t2 ), t2 )
= V (t0 , x(t0 )) +

t2

t0

V2 (x0 ) T
0

dV (t, x(t))
dt
dt

that is,
V1 (x(t2 )) 0
This contradicts V1 (x(t2 )) > 0. Hence, there must exist t1 with t0 t1 t0 + T such that
||x(t1 )|| < .

172

CHAPTER 14. LYAPUNOV THEORY FOR NONAUTONOMOUS SYSTEMS

14.5

Exponential stability

14.5.1

Exponential stability

Theorem 29 (Uniform exponential stability) Suppose there exists a function V and

positive scalars , 1 , 2 , and R such that for all t and all x with kxk < R,
1 ||x||2 V (t, x) 2 ||x||2
and
V
V
(t, x) +
f (t, x) 2V (t, x)
t
x
Then, for system (14.1), the origin is a uniformly exponentially stable equilibrium state with
rate of convergence .

14.5.2

Global uniform exponential stability

Theorem 30 (Global uniform exponential stability) Suppose there exists a function

V and positive scalars , 1 , 2 such that for all t, x,
1 ||x||2 V (t, x) 2 ||x||2
and
V
V
(t, x) +
f (t, x) 2V (t, x)
t
x
Then, for system (14.1), the origin is a globally exponentially uniformly stable equilibrium
state with rate of convergence .
Proof.
Example 140
x = (2 + cos t)x x3

14.6

Quadratic stability

From the above it should be clear that all the results in the section on quadratic stability
also hold for time-varying systems.

14.7. BOUNDEDNESS

14.7

173

Boundedness

A scalar valued function V of time and state is said to be uniformly radially unbounded if
there are radially unbounded functions V1 and V2 of state only such that
V1 (x) V (t, x) V2 (x)
for all t and x.
Example 141
V (t, x)
V (t, x)
V (t, x)
V (t, x)

=
=
=
=

(2 + sin t)x2
2
(1 + et )x2 x
2
et x2
(1 + t2 )x2

yes
yes
no
no

Theorem 31 Suppose there exists a uniformly radially unbounded function V and a scalar
R 0 such that that for all t and x with ||x|| R,
V
V
(t, x) +
f (t, x) 0
t
x
Then the solutions of x = f (t, x) are globally uniformly bounded.
Note that, in the above theorem, V does not have to be positive away from the origin.
Example 142 Consider the disturbed nonlinear system,
x = x3 + w(t)
where w is a bounded disturbance input. We will show that the solutions of this system are
GUB.
Let be a bound on the magnitude of w, that is, |w(t)| for all t. Consider V (x) = x2 .
Since V is (uniformly) radially unbounded and
V

= 2x4 + 2xw(t)
2|x|4 + 2|x|
1

the hypotheses of the above theorem are satisfied with R = 3 ; hence GUB.
Exercise 46 (Forced Duffings equation with damping.) Show that all solutions of the system
x 1 = x2
x 2 = x1 x31 cx2 + w(t)
|w(t)|

c>0

174

CHAPTER 14. LYAPUNOV THEORY FOR NONAUTONOMOUS SYSTEMS

are bounded.
Hint: Consider

1
1
1
1
V (x) = c2 x21 + cx1 x2 + x22 x21 + x41
2
2
2
4
where 0 < < 1. Letting


1 c2 c
P =
2 c 1
note that P > 0 and

1
1
V (x) = xT P x x21 + x41
2
4
1
T
x Px
4

14.8. ATTRACTIVE AND INVARIANT REGIONS

14.8

175

Attractive and invariant regions

Suppose E is a nonempty subset of the state space IRn . Recall that E is invariant for the
system
x = f (t, x)
(14.3)
if E has the property that, whenever a solution starts in E, it remains therein thereafter,
that is, if x(t0 ) is in E, then x(t) is in E for all t t0 .
We say that the set E is attractive for the above system if every solution x() of (14.3)
converges to E, that is,
lim d(x(t), E) = 0 .
t

Theorem 32 Suppose there is a continuously differentiable function V , a continuous function W and a scalar c such that the following hold.
1) V is radially unbounded.
2) Whenever V (x) > c, we have
DV (x)f (t, x) W (x) < 0

(14.4)

for all t.
Then, the solutions of the system x = f (t, x) are globally uniformly bounded and the set
{x IRn : V (x) c}
is an invariant and attractive set.

Figure 14.1: An invariant and attractive set

Example 143
x = x3 + cos t

176

CHAPTER 14. LYAPUNOV THEORY FOR NONAUTONOMOUS SYSTEMS

14.9

Systems with disturbance inputs

Here we consider disturbed systems described by

x = F (t, x, w)

(14.5)

where the m-vector w(t) is the disturbance input at time t. As a measure of the size of a
disturbance w() we shall consider its peak norm:
kw()k = sup kw(t)k .
t0

Figure 14.2: The peak norm of a signal

We wish to consider the following type of problem: Given the peak norm of the disturbance, obtain an ultimate bound on the norm of the state or some performance output
defined by
z = H(t, x, w) .
(14.6)
We first have the following result.
Theorem 33 Consider the disturbed system (14.5) and suppose there exists a continuously
differentiable function V , a scalar 1 > 0 and a continuous function U such that the following
hold.
1) V is radially unbounded.
2) For all t,
DV (x)F (t, x, w) U(x, w) < 0

whenever V (x) > 1 kwk2

Then, for every bounded disturbance w(), the state x() is bounded and the set


x IRn : V (x) 1 kw()k2
is invariance and attractive for system (14.5).
Example 144
x = x x3 + w

(14.7)

14.9. SYSTEMS WITH DISTURBANCE INPUTS

177

Considering V (x) = x2 , we have

DV (x)x = x2 x4 + xw
|x|2 |x|4 + |x||w|


= |x| |x| + |x|3 |w| = U(x, w)

where U(x, w) = |x| (|x| + |x|3 |w|). Clearly U(x, w) > 0 when |x| > |w|, that is V (x) >
|w|2. Hence, taking 1 = 1, it follows from the above theorem that the following interval is
invariant and attractive for this system:
[ ]

where = kw()k .

Proof of Theorem 33 Consider the disturbed system (14.5) subject to a specific bounded
disturbance w(). This disturbed system can be described by x = f (t, x) with f (t, x) =
F (t, x, w(t)). Let
= kw()k = sup kw(t)k .
t0

We shall apply Theorem 32 with c = 1 2 .

Consider any x which satisfies V (x) > c. Then, whenever kwk , we have V (x) >
1 kwk2 . If we define
W (x) = min {U(x, w) : kwk }
then, it follows from condition (14.7) that W (x) > 0 whenever V (x) > c.
Since kw(t)k for all t 0, it follows that whenever V (x) > c, we have
DV (x)f (t, x) = DV (x)F (t, x, w(t))
U(x, w(t))
W (x) < 0
It now follows from Theorem (32) that the state x() is bounded and the set


x IRn : V (x) 1 kw()k2
is invariance and attractive.

178

CHAPTER 14. LYAPUNOV THEORY FOR NONAUTONOMOUS SYSTEMS

Corollary 2 Consider the disturbed system (14.5) equipped with a performance output specified in (14.6). Suppose the hypotheses of Theorem 33 are satisfied and there exists a scalar
2 such that for all t, x and w,
kH(t, x, w)k2 V (x) + 2 kwk2

(14.8)

Then for every disturbance bounded disturbance w(), the output z() is bounded and satisfies
lim sup kz(t)k kw()k
t

where
=

1 + 2 .

(14.9)

If in addition, V (0) = 0 then

kz()k kw()k
when x(0) = 0.

14.9.1

Linear systems with bounded disturbance inputs

Consider a disturbed linear system:

x = Ax + Bw
z = Cx + Dw
where all the eigenvalues of A have negative real part and w is a bounded input.
Suppose there exists a positive real scalar such that



P A + AT P + P
PB
< 0
BT P
1 I

 T
C CP
CT D
0
DT C
D T D 2 I

(14.10a)
(14.10b)

Then
lim supt kz(t)k kw()k
where
=

1 + 2 .

Also, when x(0) = 0, we have

kz()k kw()k
Proof.

(14.11)

14.9. SYSTEMS WITH DISTURBANCE INPUTS

179

For each fixed we have LMIs in P , 1 and 2 . One can minimize 1 + 2 for each
and do a line search over .
The above results can be generalized to certain classes of nonlinear systems like polytopic
systems. Above results can also be used for control design.

180

CHAPTER 14. LYAPUNOV THEORY FOR NONAUTONOMOUS SYSTEMS

14.10

Regions of attraction

Suppose that the origin is an asymptotically stable equilibrium state of the system
x = f (t, x)

(14.12)

where the state x(t) is an n-vector. We say that a non-empty subset A of the state space
IRn is a region of attraction for the origin if every solution which originates in A converges to
the origin, that is,
x(t0 ) A
implies
lim x(t) = 0 .
t

Figure 14.3: Region of attraction

The following result can be useful in computing regions of attraction.
Theorem 34 Consider system (14.12) and suppose there is a continuously differentiable
function V , a continuous function W and a scalar c with the following properties.
1) V is positive definite.
2) Whenever x 6= 0 and V (x) < c, we have
DV (x)f (t, x) W (x) < 0
for all t.
Then, the origin is a uniformly asymptotically stable equilibrium state with
A = {x IRn : V (x) < c}
as a region of attraction. Also, A is invariant.

Figure 14.4:
Example 145
x = x(1 x2 )

14.10. REGIONS OF ATTRACTION

181

Example 146
x 1 = x1 (1 x21 x22 ) + x2
x 2 = x1 x2 (1 x21 x22 )

Considering V (x) = x21 + x22 , we have

V = 2V (1 V )

14.10.1

A special class of nonlinear systems

We consider here nonlinear systems described by

x = A(t, x)x

(14.13)

where
A(t, x) = A0 + (t, x)A

and

a(||Cx||) (t, x) b(||Cx||)

(14.14)

while a, b are non-decreasing functions and C is a matrix. We have previously considered

these systems where a and b are constants and obtained sufficient conditions for global
uniform exponential stability using quadratic Lyapunov functions.
Example 147 Consider
x 1 = x2
x 2 = x1 (1 x21 )x2
This system can be expressed as

 


 

0 0
x1
x 1
0
1
2
=
+ x1
0 1
x2
x 2
1 1

(14.15)

Here

C=

1 0

b() = 2 .

a() = 0 ,

Theorem 35 Suppose there is a positive definite symmetric P and a positive scalar such
that
P A0 + AT0 P a()(P A + AT P ) < 0
P A0 + AT0 P + b()(P A + AT P ) < 0
CT C P 0

(14.16a)
(14.16b)
(14.16c)

Then the origin is a uniformly asymptotically stable equilibrium state with

A = {x IRn : xT P x < 2 }
as a region of attraction. Also A is invariant. If, in addition,
P I 0

(14.17)

for positive number , then the ball

1

{x IRn : ||x|| < 2 }

is a region of attraction.

182

CHAPTER 14. LYAPUNOV THEORY FOR NONAUTONOMOUS SYSTEMS

Proof. First note that the inequality, C T C P 0, implies that

kCxk2 xT P x
for all x. So, consider any x for which xT P x < 2 . Then
kCxk < .
Using the nondecreasing property of b, we now obtain that
(t, x) b(kCxk) b() .
In a similar fashion, we also obtain that
(t, x) a(kCxk) a() .
Hence, whenever xT P x < 2 , we have
a() (t, x) b() .
Considering V (x) = xT P x as a candidate Lyapunov function, one may use inequalities
(14.16a)-(14.16b) to show that whenever xT P x < 2 and x 6= 0, one has
DV (x)x W (x) < 0 .
Hence the set
A = {x IRn : xT P x < 2 }
is invariant and a region of attraction for the origin.
Now note that the inequality, P I 0, implies that
xT P x 1 kxk2
1

for all x. So whenever, kxk < 2 , we must have xT P x < 2 , that is x is in A. Hence the
1
set of states of norm less than 2 is a region of attraction for the origin.

14.11. LINEAR TIME-VARYING SYSTEMS

14.11

183

Linear time-varying systems

Here we consider systems described by

x = A(t)x

(14.18)

All solutions satisfy

x(t) = (t, )x( )
where is the state transition matrix.
Computation of the state transition matrix. Let X be the solution of

X(t)
= A(t)X(t)
X(0) = I

(14.19a)
(14.19b)

where I is the n n identity matrix. Then

(t, ) = X(t)X( )1

(14.20)

Example 148 (Markus and Yamabe, 1960.) This is an example of an unstable second
order system whose time-varying A matrix has constant eigenvalues with negative real parts.
Consider


1 + a cos2 t
1 a sin t cos t
A(t) =
1 a sin t cos t 1 + a sin2 t
with 1 < a < 2. Here,
(t, 0) =

e(a1)t cos t et sin t

e(a1)t sin t et cos t

Since a > 1, the system corresponding to this A(t) matrix has unbounded solutions. However,
for all t, the characteristic polynomial of A(t) is given by
p(s) = s2 + (2 a)s + 2
Since 2 a > 0, the eigenvalues of A(t) have negative real parts.

14.11.1

Lyapunov functions

As a candidate Lyapunov function for a linear time-varying system consider a time-varying

quadratic form V given by:
V (t, x) = xT P (t)x
where P () is continuously differentiable and P (t) is symmetric for all t.
To guarantee that V is positive definite and decrescent, we need to assume that there
exists positive real scalars c1 and c2 such that for all t,
0 < c1 min (P (t))

and

max (P (t)) c2

(14.21)

184

CHAPTER 14. LYAPUNOV THEORY FOR NONAUTONOMOUS SYSTEMS

This guarantees that

c1 kx||2 V (t, x) c2 kx||2 .

Considering the rate of change of V along a solution, we have

V
V
(t, x) +
(t, x)x
V (t, x) =
t
x
= x P (t)x + 2x P (t)A(t)x
= x Q(t)x
where Q(t) is given by
P + P A + A P + Q = 0

(14.22)

To satisfy the requirements on V (t, x) for GUAS and GUES we need

0 < c3 min(Q(t))
for all t. This guarantees that
Hence

(14.23)

V (t, x) c3 kx||2 .

V (t, x) 2V (t, x)

where

= c3 /2c2 .

So, if there are matrix functions P () and Q() which satisfy (14.21)-(14.23) , we can conclude
that the LTV system (14.18) is GUES. The following theorem provides the converse result.
Theorem 36 Suppose A() is continuous and bounded and the corresponding linear timevarying system (14.18) is globally uniformly asymptotically stable. Then, for all t, the matrix
Z
P (t) =
(, t)T (, t) d
(14.24)
t

is well defined, where is the state transition matrix associated with (14.18), and
P + P A + A P + I = 0 .

(14.25)

Also, there exists positive real scalars c1 and c2 such that for all t,
c1 min(P (t))

and

max (P (t)) c2

Proof. See Khalil or Vidyasagar.

A consequence of the above theorem is that GUES and GUAS are equivalent for systems
where A() is continuous and bounded.
Example 149 The following system has a time varying quadratic Lyapunov function but
does not have a time-invariant one.
x = (1 + 2 sin t)x
Here

(, t) = e
Hence,
P (t) =

(1+2 sin s) ds

= e(t +2 cos t2 cos ) = e(t ) e2(cos tcos )

2(t ) 4(cos tcos )

d =

e2 e4(cos tcos (t+ )) d .

14.12. LINEARIZATION

14.12

185

Linearization
x = f (t, x)

Suppose that x() is a specific solution.

Let
x = x x
Linearization about x()
x = A(t)x
where
A(t) =

f
(t, x(t)) .
x

(14.26)

186

CHAPTER 14. LYAPUNOV THEORY FOR NONAUTONOMOUS SYSTEMS

Chapter 15
The describing function method
Here we consider the problem of determining whether or not a nonlinear system has a limit
cycle (periodic solution) and approximately finding this solution.
We consider systems which can be described by the negative feedback combination of
a SISO LTI (linear time invariant) system and a SISO memoryless nonlinear system; see
and the nonlinear
Figure 15.2. If the LTI system is represented by its transfer function G

Figure 15.1: Systems under consideration

system is represented by the function , then the system can be described by
u(s)
y(s) = G(s)
u(t) = (y(t))

(15.1a)
(15.1b)

where u and y are the input and output, respectively for the LTI system and their respective
Laplace transforms are u and y.
As a general example, consider any system described by
x = Ax B(Cx)
where x(t) is an n-vector. In this example

G(s)
= C(sI A)1 B
A solution y() is periodic with period T > 0 if
y(t + T ) = y(t)
We are looking for periodic solutions to (15.1).
187

for all t .

(15.2)

188

15.1

CHAPTER 15. THE DESCRIBING FUNCTION METHOD

Periodic functions and Fourier series

Consider any piecewise continuous signal s : IR IR which is periodic with period T . Then
s has a Fourier series expansion, that is, s can be expressed as
s(t) = a0 +

ak cos(kt) +

bk sin(kt)

k=1

k=1

where = 2/T is called the natural frequency of s. The Fourier coefficients a0 , a1 , . . . and
b1 , b2 , . . . are uniquely given by
Z T
2
1
a0 =
s(t) dt
T T2
Z T
Z T
2
2
2
2
ak =
s(t) cos (kt) dt
and
bk =
s(t) sin (kt) dt
for k = 1, 2, . . .
T T2
T T2
If s is odd, that is, s(t) = s(t) then,
ak = 0
and
4
bk =
T

T
2

for k = 0, 1, 2, . . .

s(t) sin (kt) dt

for k = 1, 2, . . .

Example 150 Sometimes it is easier to compute Fourier coefficients without using the above
integrals. Consider
s(t) = sin3 t .
Clearly, this is an odd periodic signal with period T = 2. Since
sin t =

ejt ejt
2j

we have
3

sin (t) =

ejt ejt
2j

3

e3jt 3ejt + 3ejt e3jt

8j

3(ejt ejt ) e3jt e3jt

8j
8j
3
1
=
sin t sin 3t .
4
4

So,

3
b1 = ,
4
and all other Fourier coefficients are zero.

b3 =

1
4

Exercise 47 Obtain the Fourier series for s(t) = sin5 t

15.2. DESCRIBING FUNCTIONS

15.2

189

Describing functions

The describing function for a static nonlinearity is an approximate description of its frequency
response. Suppose : IR IR is a piecewise continuous function and consider the signal
s(t) = (a sin t)
for a, > 0. This signal is the output of the nonlinear system defined by and subject to
input a sin t.
The signal s is piecewise continuous and periodic with period T = 2/. Hence it has a
Fourier series expansion and its Fourier coefficients are given by
1
a0 (a) =
T

T
2

(a sin t) dt

1
2

T2

and for k = 1, 2, . . .,
Z
1
ak (a) =
(a sin ) cos k d

(a sin ) d

( = t = 2t/T )

1
bk (a) =

and

(a sin ) sin k d

Real describing functions. Suppose that the function : IR IR is piecewise continuous

and odd, that is,
(y) = (y)
If is odd, then s is odd; hence
ak (a) = 0
and

2
bk (a) =

for k = 0, 1, . . .
Z

(a sin ) sin k d .

If we approximate s by neglecting its higher harmonics, that is,

s(t) b1 (a) sin(t)
then
s(t) N(a)a sin t
where N is called the describing function for and is given by
N(a) = b1 (a)/a

2
a

(a sin ) sin d

Basically, N(a) is an approximate system gain for the nonlinear system when subject to a
sinusoidal input of amplitude a. Notice that, unlike a dynamic system, this gain is independent of the frequency of the input signal. For a linear system, the gain is independent of
amplitude.

190

CHAPTER 15. THE DESCRIBING FUNCTION METHOD

Example 151 Cubic nonlinearity. Consider

(y) = y 3
Then, recalling the last example, we see that
a3
3a3
sin
sin 3
(a sin ) = a sin =
4
4
3

Hence b1 (a) = 3a3 /4 and

N(a) =

b1 (a)
3a2
=
.
a
4

Example 152 Signum function. Recall the signum function given by

1 if y < 0
0 if y = 0
(y) =

1 if y > 0

Its describing function is given by

Z
Z
2
2
4
N(a) =
(a sin ) sin d =
sin d =
.
a 0
a 0
a

Example 153

Consider the following approximation to the signum switching function.

Figure 15.2: Approximation to signum function

1 if y < e
0 if e y e
(y) =

1 if y > e

(15.3)

where e is some positive real number. This is an odd function. Clearly, N(a) = 0 if a e.
If a > e, let e be the unique number between 0 and /2 satisfying e = arcsin(e/a) that is,
a sin(e ) = e. Then,
Z

(a sin ) sin d =

sin d = cos( e ) + cos e = 2 cos e

15.2. DESCRIBING FUNCTIONS

191

p
Since sin e = e/a and 0 e < /2, we have cos e = 1 (e/a)2 . Thus

4 a2 e2
b1 (a) =
a
Hence,
N(a) =

if a e

4 a2 e2

if a > e
a2

(15.4)

One can readily show that the maximum value of N(a) is 2/(e) and this occurs at a =

2e.

Example 154 Consider a sector bounded nonlinearity satisfying

y 2 y(y) y 2
for some constants and . We will show that
N(a)

Complex describing functions. If the function is not odd then, its describing function
may be complex valued; see [1] for an example. Assuming a0 (a) = 0, in this case, we let
N(a) = b1 (a)/a + a1 (a)/a

(15.5)

where
1
a1 (a) =

(a sin ) cos d

and

1
b1 (a) =

(a sin ) sin d

Note that
N(a) = |N(a)|ej(a)
where

a1 (a)2 + b1 (a)2
a
and (a) is the unique angle in [0, 2) given by
|N(a)| =

cos((a)) = p

b1 (a)
a1

(a)2

+ b1

(a)2

sin((a)) = p

a1 (a)
a1 (a)2 + b1 (a)2

192

CHAPTER 15. THE DESCRIBING FUNCTION METHOD

Here, the approximate output of the nonlinearity due to input a sin(t) is given by
s(t) a1 (a) cos(t) + b1 (a) sin(t)
= |N(a)| a [sin((a)) cos(t) + cos((a)) sin(t)]
= |N(a)| a sin(t + (a))
that is,
s(t) |N(a)|a sin(t + (a))
where (a) is the argument of N(a). Notice the phase shift in s due to (a).

Exercise 48 Compute the describing function for (y) = y 5 .

Exercise 49 Consider the saturation function sat described by

1 if y < 1
y if 1 y 1
sat(y) =

1 if y > 1

Show that its describing function is given by

N(a) =
h

2 sin1 ( 1 ) +

1 if 0 a 1

a2 1
a2

if 1 < a

(15.6)

15.3. THE DESCRIBING FUNCTION METHOD

15.3

193

The describing function method

Recall the nonlinear system described in (15.1) Suppose we are looking for a periodic solution
for y which can be approximately described by
y(t) a sin(t)
where a > 0.
Real describing functions. Suppose is odd,
u(t) (a sin(t)) =

bk (a) sin(kt)

k=1

where
2
bk (a) =

(a sin ) sin(k) d

So,
u(t) b1 (a) sin(t) = N(a)a sin(t)
where N is the describing function for . Since u(t) N(a)a sin(t) and y(t) a sin(t)
u, it follows that
and y = G

G()
1/N(a) .
So, if for some pair a, > 0, the condition

1 + G()N(a)
=0

(15.7)

is satisfied, it is likely that the nonlinear system will have a periodic solution with
y(t) a sin t
When N(a) is real the above condition can be expressed

(G())
= 0

1 + (G())N(a) = 0

(15.8a)
(15.8b)

The first condition in (15.8) simply states that the imaginary part of G()
is zero, that is,

G() is real. This condition is independent of a and can be solved for values of . The
values of a are then determined by the second condition in (15.8).
Example 155
y + y 3 = 0

194

CHAPTER 15. THE DESCRIBING FUNCTION METHOD

With (y) := y 3 , this system is described by

y = u
u = (y)
Hence

G(s)
= 2
s

and
N(a) =

3a2
4

Here condition (15.7) is

3a2
1 2 = 0
4
Hence, we predict that this system has periodic solutions of all amplitudes a and with
approximate periods
2
2
4
=
T =
=

3a/2
3a
Example 156 Consider the nonlinear system,
x 1 = x2
x 2 = 4x1 5x2 sgm (x1 x2 )
This can be represented as the LTI system
x 1 = x2
x 2 = 4x1 5x2 + u
y = x1 x2
subject to negative feedback from the memoryless nonlinearity u = sgm (y). Hence

G(s)
=

s2

s + 1
+ 5s + 4

and N(a) = 4/a. The condition 1 + G()N(a)

= 0 results in
1+

+ 1
N(a) = 0
2 + 4 + 5

that is
2 + 4 + 5 + ( + 1)N(a) = 0
or
4 2 + N(a) = 0
5 N(a) = 0

15.3. THE DESCRIBING FUNCTION METHOD

195

0.8

0.6

0.4

0.2

0.2

0.4
0

10

12

14

16

18

20

Figure 15.3: Limit cycle for system in example 156

Solving yields = 3 and N(a) = 5. This results in
a=

4
= 0.2546 ,
5

T =

2
= 2.0944

The y-response of this system subject to initial conditions x1 (0) = 0.5 and x2 (0) = 0 is
illustrated in Figure 15.3
Example 157 Consider the system of the previous example with the signum function replaced with an approximation as described in Example 153, that is,
x 1 = x2
x 2 = 4x1 5x2 (x1 x2 )
where is given by (15.3) for some e > 0. Proceeding as in the previous example, the
describing function method still gives the following conditions for a limit cycle
=3

and

N(a) = 5 .

Recalling the expression for N(a) in 15.4, we must have a > e and

4 a2 e2
=5
a2
Solving this for a2 yields two solutions
p
8 4 4 (5e)2
a =
(5)2
2

For these solutions to be real, we must have 4 (5e)2 0 that is

e

2
= 0.1273 .
5

196

CHAPTER 15. THE DESCRIBING FUNCTION METHOD

Complex describing functions. We still have the criterion (15.7) for a limit cycle, that
is,

1 + G()N(a)
=0
However, we cannot usually simplify things as in (15.8).

Example 158 (Van der Pol oscillator) Here we demonstrate a generalization of the method.
y + (y 2 1)y + y = 0
We can describe this system by
y y + y = u
u = y 2 y
Here

G(s)
=
If y(t) = a sin t then

s2

s + 1

a3
a3
cos t
sin(t + /2)
y y = a sin t cos t
4
4
2

So here
N(a, ) =

a2
a2 e/2
=
.
4
4

Solving 1 + G()N(a,
) = 0 yields
=1

a = 2.

15.3. THE DESCRIBING FUNCTION METHOD

15.3.1

197

Rational transfer functions

When the describing function of the nonlinearity is real, the approximate frequencies at

which a limit cycle occurs are a subset of those frequencies for which G()
is real, that

is, the imaginary part of G()

is zero. Here we consider strictly proper rational transfer
and show how one can easily determine the frequencies for which G()

functions G
is

real. This involves a state space representation (A, B, C, 0) of G.

Suppose G(s)
= C(sI A)1 B is a scalar transfer function with real A, B, C and A has
no imaginary eigenvalues. Noting that
(I A)1 = (I A)(I A)1 (I A)1
= (I A)( 2I + A2 )1
= A( 2 I + A2 )1 ( 2I + A2 )1 ,
we obtain that

G()
= CA( 2 I + A2 )1 B C( 2I + A2 )1 B .

(15.9)

The above expression splits G()

into its real and imaginary part, in particular the imagi
nary part of G()
is given by



G()
= C( 2 I + A2 )1 B
(15.10)

To determine the nonzero values (if any) of for which the imaginary part of G()
is zero,
we note that






2 I A2 B
det
= det 2 I A2 det C( 2 I + A2 )1 B
C
0


= det 2 I A2 C( 2I + A2 )1 B .
Introducing the matrix pencil given by

P () =

A2 I B
C
0

(15.11)

we have shown that, for 6= 0,




det P ( 2) = det [ 2 I A2 ] G()

/

Since it is assumed that A does not have any purely imaginary eigenvalues, A2 does not have
any negative real eigenvalues. This implies that det( 2 I A2 ) is non-zero for all . Hence,
for nonzero ,



G()
=0
if and only if
det P ( 2) = 0
(15.12)
Note that

P () =

A2 B
C
0

I 0
0 0

198

CHAPTER 15. THE DESCRIBING FUNCTION METHOD

The values of for which det P () = 0 are the finite generalized eigenvalues of the matrix
pencil P . You can use the Matlab command eig to compute generalized eigenvalues. Thus
we have the following conclusion:

G()
is real if and only if = and is a positive, real, finite, generalized eigenvalue
of the matrix pencil P
Example 159 For example 156 we have

 

0
0
1
B=
A=
1
4 5
Hence
P () =

A I B
C
0

C=

1 1

4
5
0
= 20 21 1
1
1
0

and det P () = + 9. Hence 9 is the only finite generalized eigenvalue of P and it is

positive real. Thus, there is a single nonzero value of for which G()
is real and this is
given by = 3.

15.4

Exercises

Exercise 50 Obtain the describing function of the nonlinear function

(y) = y 5
Exercise 51 Determine whether or not the following Duffing system has a periodic solution.
Determine the approximate amplitude and period of all periodic solutions.
y y + y 3 = 0 .
Exercise 52 Determine whether or not the following system has a periodic solution. Determine the approximate amplitude and period of all periodic solutions.
y + (y 3 /3 y)
+y = 0.
where is a positive real number.
Exercise 53 Use the describing function method to predict period solutions to
y + sgm (y) = 0
Illustrate your results with numerical simulations.
Exercise 54 Use the describing function method to predict period solutions to
x(t)

= x(t) 2 sgm (x(t h))

Illustrate your results with numerical simulations.

15.4. EXERCISES

199

Exercise 55 Consider the double integrator

q = u
subject to a saturating PID controller
u = sat(
u)

where

u = kP q kI

q kD q

(a) For kP = 1 and kD = 2 determine the largest of kI 0 for which the closed loop
system is stable about q(t) 0.
(b) For kP = 1 and kD = 2, use the describing function method to determine the smallest
value kI 0 for which the closed loop system has a periodic solution.

200

CHAPTER 15. THE DESCRIBING FUNCTION METHOD

Bibliography
[1] Boniolo, I., Bolzern, P., Colaneri,P., Corless, M., Shorten, R. Limit Cycles in Switching
Capacitor Systems: A Lure Approach

201

202

BIBLIOGRAPHY

Chapter 16
Aerospace/Mechanical Systems

16.1

Aerospace/Mechanical systems

Let the real scalar t represent time. At each instant of time, the configuration of the
aerospace/mechanical systems under consideration can be described by a real N-vector q(t).
We call this the vector of generalized coordinates; each component is usually an angle, a
length, or a displacement. It is assumed that there are no constraints, either holonomic or
non-holonomic, on q. So, N is the number of degrees-of-freedom of the system. We let the
real scalar T represent the kinetic energy of the system. It is usually given by
N
N X
X
1
T = qT M(q)q =
Mjk (q)qj qk
2
j=1
k=1

where the symmetric N N matrix M(q) is called the system mass matrix. Usually, it
satisfies
M(q) > 0
for all q
The real N-vector Q is the sum of the generalized forces acting on the system. It includes
conservative and non-conservative forces. It can depend on t, q, q.

Example 160 Two cart system.

Here

1
1
T = m1 q12 + m2 q22
2
2

So,
M=

m1 0
0 m2
203

204

CHAPTER 16. AEROSPACE/MECHANICAL SYSTEMS

Example 161 Single link manipulator (inverted pendulum).

q=
1
T = J 2
2
So, M = J. Also, Q = W l sin .
Example 162 Particle undergoing central force motion.

Figure 16.1: Central force motion

Consider a particle P of mass m undergoing a central force motion. The force center is
O which is fixed in inertial reference frame i. Here,
q = (r

)T

and the kinetic energy is given by

1
T = m(r 2 + r 2 2 )
2
Hence
M(q) =
Also, the generalized forces are given by

Q=

16.2

m 0
0 mr 2

F
0

Equations of motion

La Granges Equation
 
T
d T

= Qi
dt qi
qi
In vector form:

d
dt

T
q

for i = 1, 2, , N


=Q
q

16.2. EQUATIONS OF MOTION

205

Example 163 Single link manipulator (inverted pendulum). Application of La Granges

equation yields:
J W l sin = 0
(16.1)
Example 164 Particle undergoing central force motion. Application of La Granges equation yields
m
r mr 2 F = 0
mr 2 + 2mr r
=0
The following result makes a statement about the structure of the differential equations
obtained from the application of La Granges equation.
Proposition 1 Application of La Granges equation yields
M(q)
q + c(q, q)
Q=0

(16.2)

where the i-th component of c is a quadratic function of q and is given by

ci (q, q)
=

N X
N 
X
Mij

qk

j=1 k=1

Proof. Since

1 Mjk

2 qi

qj qk

1
T = qT M(q)q
2

it follows that

T
= M(q)q
q

Hence
d
dt

T
q

= M q + M q

and application of La Granges equation yields

M q + c Q = 0
where

Since

it follows that

T
c = M q
.
q
N

(16.3)

1 XX
T =
Mjk (q)qj qk ,
2 j=1 k=1
N

T
1 X X Mjk
=
qj qk
qi
2 j=1 k=1 qi

(16.4)

206

CHAPTER 16. AEROSPACE/MECHANICAL SYSTEMS

Noting that
N
X

N
N
X
X
Mij

(M q)
i=
Mij qj =
qk
qk
j=1
j=1
k=1

qj =

N X
N
X
Mij
j=1 k=1

we have
ci


N X
N 
X
Mij
1 Mjk
T

qj qk
=

= (M q)
i
qi
qk
2 qi
j=1 k=1

qk

qj qk

(16.5)

16.3. SOME FUNDAMENTAL PROPERTIES

16.3

207

Some fundamental properties

In this section, we present some fundamental properties of aerospace/mechanical systems

which are useful in the analysis and control design for these systems.

16.3.1

An energy result

First we have the following result.

Proposition 2


1
q c M q = 0
2

(16.6)

Proof. It follows from (16.3) that





1
1
T
1
T

q c M q = q
M q
= q M q
q
2
2
q
2
q

(16.7)

Using (16.4),
N

X T
X
T
q =
qi =
q
qi
i=1
i=1

1 X X Mjk
qj qk
2 j=1 k=1 qi

1 X X X Mjk
qi =
qi qj qk .
2 i=1 j=1 k=1 qi

(16.8)

Using (16.5),
N
N
N
N
1X
1 X X X Mij
1

q M q =
qj qk
qi (M q)
i=
2
2 i=1
2 i=1 j=1
qk
k=1

1 X X X Mij
qi =
qi qj qk
2 i=1 j=1
qk
k=1

By interchangining the indices i and k and using the symmetry of M, we obtain that
N

1
1 X X X Mjk
q M q =
qi qj qk .
2
2 i=1 j=1 k=1 qi
Substituting (16.8) and (16.9) into (16.7) yields


1

q c M q = 0
2

(16.9)

208

CHAPTER 16. AEROSPACE/MECHANICAL SYSTEMS

Now we have the main result of this section.

Theorem 37 The time rate of change of the kinetic energy equals the power of the generalized forces, that is,
dT
= q Q
(16.10)
dt
Proof. Since the kinetic energy is given by
1
T = q M q
2
application of (16.2) and (16.6) yields:
dT
dt

16.3.2

1
= q M q + q M q
2


1
1

= q (c + Q) + q M q = q Q q c M q
2
2

= q Q .

Potential energy and total energy

A generalized force is said to be conservative if it satisfies the following two properties:

(a) It depends only on q.
(b) It can be expressed as the negative of the gradient of some scalar valued function of
q. This function is called the potential energy associated with the force.
Example 165 Consider
Q = W l sin

and

q=

Here Q = U
where
q

U = W l(cos 1) .

Example 166 Consider

Q=

F
0

q=

where
Here Q = U
q
U =

m
.
r

F =

m
r2

16.3. SOME FUNDAMENTAL PROPERTIES

209

Suppose one splits the generalized force vector Q into a conservative piece K and a
non-conservative piece D, thus
Q = K(q) D(q, q)

(16.11)

where K is the sum of the conservative forces and D is the sum of the non-conservative
forces. Let U be the total potential energy associated with the conservative forces, then
U
.
q

(16.12)

L=T U

(16.13)

K=
Introducing the Lagrangian

Lagranges equation can now be written as



d L
L

+D = 0.
dt q
q
In particular, if all the forces are conservative, we have


L
d L

= 0.
dt q
q
Introducing the total mechanical energy
E =T +U

(16.14)

dE
= q D
dt

(16.15)

we obtain that

that is, the time rate of change of the total energy equals the power of the non-conservative
forces. To see this:
dE
dT
dU
=
+
dt
dt
dt
U
q
= q Q +
q
U
U
= q
q D(q, q)
+
q
q
q
= q D(q, q)
.
In particular, if all the forces are conservative, we have
dE
=0
dt
thus E is constant, that is, energy is conserved.

(16.16)

210

16.3.3

CHAPTER 16. AEROSPACE/MECHANICAL SYSTEMS

A skew symmetry property

Proposition 3 For each q, q IRN , there is a skew-symmetric matrix S(q, q)

such that

1
M S q
c=
2

Proof. Using (16.3),

(16.17)

1
1
T
c = M q + M q
2
2
q

From (16.5),
(M q)
i=

N X
N
X
Mij

qk

j=1 k=1

qj qk =

N X
N
X
Mik

qj

j=1 k=1

qj qk

Thus, using (16.4),




N
N 
T
1
T 1 X X Mik Mjk
1
M q
=
=

qj qk
(M q)
i
2
q i
2
qi
2 j=1 k=1 qj
qi
=

N
X
1
j=1

where
Sij :=

Sij qj

N 
X
Mjk
k=1

Hence Sji = Sij ,

qi

Mik

qj

qk

(16.18)

T
1
1
M q
= S q
2
q
2

and
c=
with S skew-symmetric.


1
M S q
2

Corollary 3 For each q, q IRN , there is a matrix C(q, q)

such that
c = C q

(16.19)

and M 2C is skew-symmetric.
Proof. Let

1
C = (M S)
2
where S is skew-symmetric as given by previous result. Then c = C q and
M 2C = S

(16.20)

16.3. SOME FUNDAMENTAL PROPERTIES

211

Remark 8 Substituting (16.19) into (16.2) yields

M(q)
q + C(q, q)
q Q = 0,

(16.21)

and using (16.20) to (16.18) we obtain

N

1X
Cij =
2 k=1

Mij
Mik Mjk
+

qk
qj
qi

qk .

(16.22)

212

16.4

CHAPTER 16. AEROSPACE/MECHANICAL SYSTEMS

Linearization of aerospace/mechanical systems

Here we consider aerospace/mechanical systems described by (16.2), that is,

M(q)
q + c(q, q)
Q(q, q)
=0

(16.23)

The equilibrium configurations q(t) q e of the above system are given by

Q(q e , 0) = 0
Linearization about q(t) q e results in
q + D
q + Kq
=0
M

(16.24)

where

= Q (q e , 0),
= Q (q e , 0)
D
K
q
q
Note that the c term does not contribute anything to the linearization; the contribution of
M(q)
q is simply M(q e ) q.
We have now the following stability result.
= M(q e ),
M

Theorem 38 Consider the linearization (16.24) of system (16.23) about q(t) q e and
and K
are symmetric and positive definite. Suppose also that for all q,
suppose that M

q 0
q D
and

q(t)
q(t)
D
0 q(t)
0

then the nonlinear system (16.23) is asymptotically stable about the solution q(t) q e .
Proof. As a candidate Lyapunov function for the linearized system (16.24), consider
1
q + 1 q Kq

V (q, q)
= q M
2
2

Example 167
m1 q1 + dq1 +(k1 + k2 )q1 k2 q2
m2 q2
k2 q1
+k2 q2

= 0
= 0

This system is asymptotically stable.

Example 168

m
q1 + dq1 dq2 +(k1 + k2 )q1 k2 q2
m
q2 dq1 + dq2 k2 q1
+(k1 + k2 )q2
This system is not asymptotically stable.

= 0
= 0

16.5. A CLASS OF GUAS AEROSPACE/MECHANICAL SYSTEMS

213

Figure 16.2: An example

Figure 16.3: Another example

16.5

A class of GUAS aerospace/mechanical systems

Here we consider aerospace/mechanical systems described by

M(q)
q + c(q, q)
+ D(q, q)
+ K(q) = 0

(16.25)

where q(t) is a real N-vector and

M(q) = M(q) > 0
for all q. So, here Q = D(q, q)
K(q)
Assumption 1 There is an N-vector q e such that:
(i)
K(q e ) = 0

and

K(q) 6= 0 for q 6= q e

(ii)
D(q, 0) = 0

for all q

This first assumption guarantees that q(t) q e is a unique equilibrium solution of the above
system. The next assumption requires that the generalized force term K(q) be due to
potential forces and that the total potential energy associated with these forces is a positive
definite function.

214

CHAPTER 16. AEROSPACE/MECHANICAL SYSTEMS

Assumption 2 There is a function U, which is positive definite about q e , such that

K(q) =

U
(q)
q

for all q.
The next assumption requires that the generalized force term D(q, q)
be due to damping
forces and that these forces dissipate energy along every solution except constant solutions.
Assumption 3 For all q, q,

q D(q, q)
0

If q() is any solution for which

q(t)
D(q(t), q(t))

0
then
q(t)
0
Theorem 39 Suppose the mechanical system (16.25) satisfies assumptions 1-3. Then the
system is GUAS about the equilibrium solution q(t) q e .
Proof Consider the total mechanical energy (kinetic + potential)
1
V = q M(q)q + U(q)
2
as a candidate Lyapunov function and use a LaSalle type result.
Example 169
I1 q1 + dq1 +(k1 + k2 )q1 k2 q2
= 0
I2 q2
k2 q1
+k2 q2 W l sin(q2 ) = 0
where all parameters are positive and
Wl <

k1 k2
=
k1 + k2

1
k1

1
+

1
k2

Exercise 56 Consider the sprung pendulum subject to aerodynamic drag described by

+ W l sin + k = 0
I + d||
where k > W l and d > 0.
(i) Linearize this system about the zero solution. What are the stability properties of the
linearized system? Based on this linearization, what can you say about the stability of
the nonlinear system about the zero solution?
(ii) Use Theorem 39 to show that the nonlinear system is GUAS about the zero solution.

16.6. CONTROL OF AEROSPACE/MECHANICAL SYSTEMS

16.6

215

Control of aerospace/mechanical systems

Here we consider aerospace/mechanical systems with control inputs described by

M(q)
q + c(q,
q) + F (q, q)
=u

(16.26)

where q(t) and the control input u(t) are real N-vectors. So, Q = F + u. The salient
feature of the systems considerd here is that the number of scalar control inputs is the same
as the number of degrees of freedom of the system.

16.6.1

Computed torque method

This is a special case of the more general technique of feedback linearization. Suppose we
let
u = c(q, q)
+ F (q, q)
+ M(q)v
(16.27)
where v is to be regarded as a new control input. Then, M(q)
q = M(q)v and since M(q) is
nonsingular, we obtain
q = v
(16.28)
This can also be expressed as a bunch of N decoupled double integrators:
qi = vi

for i = 1, 2, , N

Now use YFLTI (your favorite linear time invariant) control design method to design controllers for (16.28).
Disadvantages: Since this method requires exact cancellation of F and exact knowledge of
M, it may not be robust with respect to uncertainty in M and F .

16.6.2

Linearization

Consider any desired controlled equilibrium configuration

q(t) q e
for the system
M(q)
q + c(q,
q) + F (q, q)
=u
This can be achieved with the constant input
u(t) ue := F (q e , 0)
If we linearize the above system about (q(t), u(t)) (q e , ue ), we obtain
q + D
q + Kq
= u
M

(16.29)

where

= F (q e , 0)
= F (q e , 0),
K
D
q
q
Now use YFLTI control design method to design controllers for u and let
= M(q e ),
M

u = ue + u

(16.30)

216

16.6.3

CHAPTER 16. AEROSPACE/MECHANICAL SYSTEMS

Globally stabilizing PD controllers

Here we consider mechanical systems described by

M(q)
q + c(q,
q) + D(q, q)
+ K(q) = u

(16.31)

So, F (q, q)
= D(q, q)
+ K(q).
Assumption 4 There is an N-vector q e uch that
K(q e ) = 0

and

D(q, 0) = 0

for all q

This first assumption guarantees that q(t) q e is an equilibrium solution of the uncontrolled
(u = 0) system. It does not rule out the possibility of other equilibrium solutions.
Assumption 5 There is a function U and a scalar such that for all q
K(q) =
and

U
(q)
q

K
(q) I
q

Assumption 6 For all q and q,

q D(q, q)
0
PD controllers

with

u = Kp (q q e ) Kd q
Kp = Kp > I ,

Kd + Kd > 0

(16.32a)
(16.32b)

Theorem 40 Consider a system described by (16.31) which satisies assumptions 4-6 and is
subject to PD controller (16.32). Then the resulting closed system is GUAS about q(t) q e .
Proof. We show that the closed loop system
M(q)
q + c(q, q)
+ D(q, q)
+ Kd q + K(q) + Kp (q q e ) = 0
satisfies the hypotheses of Theorem 39.
Example 170 Single link manipulator with drag.
|
W l sin = u
I + d|
with d 0. Here q = ,
K() = W l sin

and

= d|
|

D(, )

(16.33)

16.6. CONTROL OF AEROSPACE/MECHANICAL SYSTEMS

Hence

217

K
(q) = W l cos q W l
q

So = W l and a globally stabilizing controller is given by

u = kp q kd q
kp > W l ,

kd > 0

link angle (rad)

0.5

q1(0) =1,

q1dot(0) = 0

-0.5
0

10
time (sec)

12

14

16

18

20

16

18

20

0.6

link angle (rad)

q1(0) =0,

q1dot(0) = 1

0.4
0.2
0
-0.2
0

10
time (sec)

12

14

Figure 16.4: Response of closed-loop, rigid, single-link manipulator

218

CHAPTER 16. AEROSPACE/MECHANICAL SYSTEMS

Exercise 57 Consider the two link manipulator with two control input torques, u1 and u2 :
[m1 lc21 + m2 l12 + I1 ]
q1

+[m2 l1 lc2 cos(q1 q2 )]

q2 +m2 l1 lc2 sin(q1 q2 )q22 [m1 lc1 + m2 l1 ]g sin(q1 ) = u1

q2
[m2 l1 lc2 cos(q1 q2 )]
q1 +[m2 lc22 + I2 ]

m2 l1 lc2 sin(q1 q2 )q12 m2 glc2 sin(q2 )

Using each of the 3 control design methods of this section, obtain controllers which asymptotically stabilize this system about the upward vertical configuration. Design your controllers
using the following data:
m1
kg
10

l1
m
1

lc1
I1
m kg m2
0.5 10/12

m2
kg
10

l2
m
1

lc2
I2
m kg m2
0.5
5/12

Table 16.1: Parameters of the two-link manipulator

Simulate the closed loop system with each of your three controller and for the following
values of m2 : 5 kg and 10 kg.

= u2

Chapter 17
Quadratic stabilizability
17.1

Preliminaries

Here we consider the problem of designing stabilizing state feedback controllers for specific
classes of nonlinear systems. Controller design is based on quadratic Lyapunov functions.
First recall that a system
x = f (t, x)
(17.1)
is quadratically stable if there exist positive definite symmetric matrices P and Q which satisfy
2x P f (t, x) 2x Qx

(17.2)

for all t and x. When this is the case, the system is globally uniformly exponentially stable
(GUES) about the origin and the scalar = min (P 1 Q) is a rate of exponential convergence.
Also, we call P a Lyapunov matrix for (17.1).
Consider now a system with control input u where u(t) is an m-vector:
x = F (t, x, u) .

(17.3)

Suppose u is given by a static linear state feedback controller

u = Kx ;

(17.4)

then the resulting closed loop system is described by

x = F (t, x, Kx) .

(17.5)

We refer to K as a state feedback gain matrix.

We say that system (17.3) is quadratically stabilizable via linear state feedback if there exists
a matrix K such that the closed loop system (17.5) is quadratically stable, that is, exist a
matrix K and positive definite symmetric matrices P and Q which satisfy
2x P F (t, x, Kx) 2x Qx
for all t and x.
219

(17.6)

220

CHAPTER 17. QUADRATIC STABILIZABILITY

= P 1 QP 1 , the above requirement

If we let x = P x, S = P 1 and L = KP 1 and Q
can we expressed as:
x
2
x F (t, S x, L
x) 2
x Q
(17.7)
for all t and x. Then
P = S 1

17.2

K = LS 1 .

(17.8)

Some special classes of nonlinear systems

Here we consider nonlinear systems with control inputs described by

x = A(t, x)x + B(t, x)u

(17.9)

where the real n-vector x(t) is the state and the real m-vector u(t) is the control input. The
following observation is sometimes useful in reducing control design problems to the solution
of linear matrix inequalities (LMIs).
which
Suppose there exist a matrix L and positive definite symmetric matrices S and Q
satisfy

A(t, x)S + B(t, x)L + SA(t, x) + L B(t, x) Q

(17.10)
for all t and x. Then the linear state feedback controller
u = LS 1 x

(17.11)

renders system (17.9) GUES about the origin with Lyapunov matrix P = S 1 .
To see this, suppose (17.10) holds and let P = S 1 and K = LS 1 . Pre- and postmultiply inequality (17.10) by S 1 to obtain

P [A(t, x) + B(t, x)K] + [A(t, x) + B(t, x)K] P P QP

(17.12)

Now notice that the closed loop system resulting from controller (17.11) is described by
x = [A(t, x) + B(t, x)K]x

(17.13)

that is, it is described by (17.1) with f (t, x) = [A(t, x) + B(t, x)K]x. Thus, using inequality
(17.12), the closed loop system satisfies
2x P f (t, x) = 2x P [A(t, x) + B(t, x)K]x
= x P [A(t, x) + B(t, x)K] + [A(t, x) + B(t, x)K] P x
x.
x P QP
replacing Q. Therefore,
Thus, the closed loop system satisfies inequality (17.6) with P QP
the closed loop system (17.13) is GUES with Lyapunov matrix P = S 1 .

17.2. SOME SPECIAL CLASSES OF NONLINEAR SYSTEMS

17.2.1

221

A special class of nonlinear systems

We first consider systems described by (17.9) where the time/state dependent matrices
A(t, x) and B(t, x) have the following structure:
A(t, x) = A0 + (t, x)A ,

B(t, x) = B0 + (t, x)B

(17.14)

The matrices A0 , B0 and A, B are constant and is a scalar valued function of time t
the state x which is bounded above and below, that is,
a (t, x) b

(17.15)
2

for some constants a and b. Examples of such functions are given by sin x, cos x, ex and
sin t.
Example 171 (Single-link manipulator) A single-link manipulator (or inverted pendulum) subject to a control torque u is described by
J W l sin = u
with J, W l > 0. Letting x1 = and x2 = this system has the following state space
description:
x 1 = x2
J x 2 = W l sin x1 + u
This description is given by (17.9) and (17.14) with






0 0
0
0 1
,
,
A =
,
B0 =
A0 =
1 0
1/J
0 0
and
(t, x) =
Since | sin x1 | |x1 |, we have

(17.16a)
(17.16b)

B = 0

(W l/J) sin x1 /x1 if x1 6= 0

(W l/J)
if x1 = 0

W l/J (t, x) W l/J;

hence satisfies inequalities (17.15) with a = W l/J and b = W l/J.
The following theorem yields controllers for the global exponential stabilization of systems
described by (17.9) and (17.14) and satisfying (17.15).
Theorem 41 Consider a system described by (17.9) and (17.14) and satisfying (17.15).
Suppose there exist a matrix L and a positive-definite symmetric matrix S which satisfy the
following linear matrix inequalities:
A1 S + B1 L + SA1 + L B1 < 0
A2 S + B2 L + SA2 + L B2 < 0

(17.17a)
(17.17b)

222

CHAPTER 17. QUADRATIC STABILIZABILITY

with
A1 := A0 + aA ,
A2 := A0 + bA ,

B1 := B0 + aB
B2 := B0 + bB .

(17.18a)
(17.18b)

Then the controller given by (17.11) renders system (17.9),(17.14) globally uniformly exponentially stable about the origin with Lyapunov matrix S 1 .
be any positive definite
Proof. Suppose S and L satisfy inequalities (17.17) and let Q
matrix which satisfies

A1 S + B1 L + SA1 + L B1 Q
.
A2 S + B2 L + SA2 + L B2 Q

(17.19a)
(17.19b)

For each t and x,

N(t, x) := A(t, x)S + B(t, x)L + SA(t, x) + L B(t, x)
= N0 + (t, x)N
where
N0 = A0 S + B0 L + SA0 + L B0

and

N = AS + BL + SA + L B

Since N(t, x) = N0 + (t, x)N and a (t, x) b, it follows that

N(t, x) Q
for all t and x if

N0 + aN Q

and

N0 + bN Q

These last two inequalities are precisely inequalities (17.19).

Exercise 58 Prove the following result: Consider a system described by (17.9),(17.14) and
satisfying (17.15). Suppose there exists a positive-definite symmetric matrix S, a matrix L
and a positive scalar which satisfy
A1 S + B1 L + SA1 + L B1 + 2S 0
A2 S + B2 L + SA2 + L B2 + 2S 0

(17.20a)
(17.20b)

where A1 , B1 , A2 , B2 are given by (17.18). Then controller (17.11) renders system (17.9),
(17.14), GUES about the origin with rate of convergence and Lyapunov matrix S = P 1 .
Note that the existence of a positive definite symmetric S and a matrix L satisfying
inequalities (17.17) or (17.20) is equivalent to the existence of another symmetric matrix S
and another matrix L satisfying S I and (17.17) or (17.20), respectively.
Suppose B1 = B2 and (S, L) is a solution to (17.17) or (17.20), then (S, LB1 ) is also
a solution for all 0. Since the feedback gain matrix corresponding to the latter pair is

17.2. SOME SPECIAL CLASSES OF NONLINEAR SYSTEMS

223

K = L0 S 1 B1 S 1 , this results in large gain matrices for large . So how do we restrict

the size of K? One approach is to minimize subject to
KSK I

(17.21)

If we also require I S, we obtain that KK KSK I. Thus, is an upper bound on

||K||2. If we minimize subject to inequalities (17.17), the resulting controller may yield
an unacceptably small rate of convergence. To avoid this, we can choose a specific desired
rate of convergence > 0 and minimize subject to (17.20). Finally, note that inequality
(17.21) can be written as LS 1 L I, or,


I L
0
L S
To summarize, we can globally stabilize system (17.9),(17.14),(17.15) with rate of convergence by solving the following optimization problem:
Minimize subject to
A1 S + B1 L + SA1 + L B1 + 2S
A2 S + B2 L + SA2 + L B2 + 2S
I


I L
L S

0
0
S

(17.22)

and letting u = LS 1 x.
Example 172 Recall the single-link manipulator of Example 171. For J = W l = 1, the
following Matlab program uses the results of this section and the LMI toolbox to obtain
stabilizing controllers.
% Quadratic stabilization of
%
clear all
% Data and specs
alpha = 0.1;
J = 1;
Wl = 1;
%
% Description of manipulator
A0 = [0 1; 0 0];
DelA = [0 0; 1 0];
B0 = [0; 1/J];
DelB = 0;
a= -Wl/J
b= Wl/J

single-link manipulator

%The specified rate of convergence

224

CHAPTER 17. QUADRATIC STABILIZABILITY

A1 = A0 + a*DelA;
A2 = A0 + b*DelA;
B1 = B0 + a*DelB;
B2 = B0 + b*DelB;
%
% Form the system of LMIs
setlmis([])
%
S=lmivar(1, [2,1]);
L=lmivar(2, [1,2]);
beta=lmivar(1, [1,1]);
%
lmi1=newlmi;
lmiterm([lmi1,1,1,S], A1,1,s)
lmiterm([lmi1,1,1,L], B1,1,s)
lmiterm([lmi1,1,1,S], 2*alpha, 1)
%
lmi2=newlmi;
lmiterm([lmi2,1,1,S], A2,1,s)
lmiterm([lmi2,1,1,L], B2,1,s)
lmiterm([lmi2,1,1,S], 2*alpha, 1)
%
Slmi= newlmi;
lmiterm([-Slmi,1,1,S],1,1)
lmiterm([Slmi,1,1,0],1)
%
lmi4=newlmi;
lmiterm([-lmi4,1,1,beta],1,1)
lmiterm([-lmi4,1,2,L],1,1)
lmiterm([-lmi4,2,2,S],1,1)
lmis = getlmis;
%
c=mat2dec(lmis,0,0,1);
options=[1e-5 0 0 0 0];
[copt,xopt]=mincx(lmis,c,options)
%
S = dec2mat(lmis,xopt,S)
L = dec2mat(lmis,xopt,L)
%
K =L*inv(S)

%A1*S + S*A1
%B1*L + L*B1
%2*alpha*S

%A2*S + S*A2
%B2*L + L*B2
%2*alpha*S

%beta*I
%L
%S

%specify weighting
%Minimize

%Controller gain matrix

Running this program yields:

K =
K =

1.3036 1.2479


3.6578 2.8144

for
for

= 0.1
=1

17.2. SOME SPECIAL CLASSES OF NONLINEAR SYSTEMS

K =

152.9481 20.1673

For initial conditions

for

225

= 10

(0)
= 0,

(0) =

Figure 17.1 illustrates the closed loop behavior of corresponding to = 1 and = 10.
3.5

2.5

theta

1.5

0.5

0.5
0

3
t

Figure 17.1: Closed loop single-link manipulator

Exercise 59 (PID regulation of single-link manipulator) Consider a single-link manipulator subject to a unknown constant disturbance torque w:
J W l sin = w + u .
Suppose we wish to stabilize this system about some desired constant motion (t) d .
The corresponding steady state value of u is unknown and is given by ud = W l sin d w.
Consider now the following PID controller:
Z t
u = kp ( d ) kI
( d ) d kd
(17.23)
0

The resulting closed loop system is described by

J + kd + kp ( d ) W l sin w + kI = 0
= d
The equilibrium solutions (t) d , (t) d of this system satisfy
= d
u d + kI = 0

226

CHAPTER 17. QUADRATIC STABILIZABILITY

So, when ud 6= 0, one needs kI 6= 0 and the equilibrium d value of is given by d = ud /kI .
Introduce states
x1 = d
x2 =
x3 = d
to obtain the following state space description:
x 1 = x2
J x 2 = kp x1 kd x2 kI x3 + W l[sin(d + x1 ) sin(d )]
x 3 = x1

(17.24a)
(17.24b)
(17.24c)

If one now obtains a matrix

K=

kp kd kI

so that system (17.24) is GUES about the origin, then the PID controller (17.23) will guarantee that the angle of the manipulator will exponentially approach the desired value d
and all other variables will be bounded. Note that a stabilizing K will have kI 6= 0.
Using Theorem 41 and the LMI toolbox, obtain a stabilizing K for J = W l = 1. Simulate
the resulting closed loop system for several initial conditions and values of d and w.

17.2.2

Generalization

One can readily generalize the results of this section to systems described by (17.9) where
the time/state dependent matrices A(t, x) and B(t, x) have the following structure
A(t, x) = A0 + 1 (t, x)A1 + + l (t, x)Al
B(t, x) = B0 + 1 (t, x)B1 + + l (t, x)Bl

(17.25a)
(17.25b)

Here, A0 , A1 , , Al are constant nn matrices and B0 , B1 , , Bl are constant nm

matrices. Also, each i is a scalar valued function of the time t and state x which satisfies
ai i (t, x) bi

(17.26)

for some known constants ai and bi .

Example 173 (Robust control of single-link manipulator) Recall from Example 171
that the motion of a single-link manipulator (or inverted pendulum) subject to a control
torque u can be described by
x 1 = x2
x 2 = (W l/J) sin x1 + (1/J)u

(17.27a)
(17.27b)

with J, W l > 0. Suppose that we do not know the parameters J and W l exactly and only
have knowledge on their upper and lower bounds, that is,
0<J J J

and

Wl Wl Wl

17.2. SOME SPECIAL CLASSES OF NONLINEAR SYSTEMS

where the bounds J, J, W l, W l are known. Introducing

(W l/J) sin x1 /x1 if x1 6= 0
1 (t, x) =
and
(W l/J)
if x1 = 0

227

2 (t, x) = 1/J ,

the above state space description is given by (17.9) and (17.25) with l = 2 and


 


0 1
0
0 0
A0 =
, B0 =
, A1 =
, B1 = 0 , A2 = 0 ,
0 0
0
1 0

B2 =

0
1

Since | sin x1 | |x1 |, we have

W l/J 1 (t, x) W l/J
1/J 2 (t, x) 1/J ;
hence 1 and 2 satisfy inequalities (17.26) with a1 = W l/J, b1 = W l/J and a2 =
1/J, b2 = 1/J.
To obtain a sufficient condition for a system given by (17.9), (17.25) and (17.26), we
introduce the following set of 2l extreme matrix pairs:
AB = { (A0 + 1 A1 + + l Al ,

B0 + 1 B1 + + l Bl ) : i = ai or bi for i = 1, , l}

For example, with l = 2, AB consists of the following four matrix pairs:

(A0 + a1 A1 + a2 A2 , B0 + a1 B1 + a2 B2 )
(A0 + a1 A1 + b2 A2 , B0 + a1 B1 + b2 B2 )
(A0 + b1 A1 + a2 A2 , B0 + b1 B1 + a2 B2 )
(A0 + b1 A1 + b2 A2 , B0 + b1 B1 + b2 B2 )
We have now the following result.
Theorem 42 Consider a system described by (17.9),(17.25) and satisfying (17.26). Suppose
there exist a matrix L and a positive-definite symmetric matrix S which satisfy
AS + BL + SA + L B < 0

(17.28)

for every pair (A, B) in AB. Then the controller given by (17.11) renders system (17.9),(17.25)
globally uniformly exponentially stable about the origin with Lyapunov matrix S 1 .
Recalling the discussion in the previous section, we can globally stabilize system (17.9),(17.14),(17.26)
with rate of convergence by solving the following optimization problem:
Minimize subject to
AS + BL + SA + L B + 2S 0
S I


I L
0
L S
and letting u = LS 1 x.

for all (A, B) in AB

(17.29)

228

CHAPTER 17. QUADRATIC STABILIZABILITY

Exercise 60 Consider the double inverted pendulum described by

1 + k1 k2 sin 1 = u
2 k1 + k2 sin 2 = 0
with k = 1. Using the above results, obtain a state feedback controller which renders this
system globally exponentially stable about the zero solution.

17.3

Another special class of nonlinear systems

Here we consider nonlinear systems which have the following structure:

x = Ax + B1 (t, Cx + Du) + B2 u

(17.30)

where the real n-vector x(t) is the state and the real m2 -vector u(t) is the control input. The
matrices B1 , B2 and C are constant with dimensions n m1 , n m2 , and p n, respectively.
The function : IR IRp IRm1 satisfies
||(t, z)|| ||z||

(17.31)

Note that if we introduce a fictitious output z = Cx + Du and a fictitious input w =

(t, Cx + Du), the nonlinear system (17.30) can be described as a feedback combination of
a linear time invariant system
x = Ax + B1 w + B2 u
z = Cx + Du
and a memoryless nonlinearity
w = (t, z)
This is illustrated in Figure 17.2.

Figure 17.2: A feedback system

Example 174 Single link manipulator. Recall the single-link of Example 171 described by
(17.16). This system can be described by the above general description where








0
0
0 1
C= 1 0
D=0
B2 =
B1 =
A=
1/J
W l/J
0 0

and (t, z) = sin z. Also,

||(t, z)|| = | sin z| |z| = ||z||

17.3. ANOTHER SPECIAL CLASS OF NONLINEAR SYSTEMS

229

The following theorem yields controllers for for the global exponential stabilization of
system (17.30).
Theorem 43 Consider a system described by (17.30) and satisfying (17.31). Suppose there
exist a matrix L, a positive-definite symmetric matrix S and a positive scalar which satisfy
the following matrix inequality :
AS + B2 L + SA + L B2 + 2 B1 B1 + 1 (CS + DL) (CS + DL) < 0

(17.32)

Then the controller given by (17.11) renders system (17.30) globally uniformly exponentially
stable about the origin with Lyapunov matrix S 1 .
Proof. Consider any K and u = Kx. Then
x = (A + B2 K)x + B1 (t, (C + DK)x)
Recalling quadratic stability results, this system is GUES with Lyapunov matrix P if there
exists > 0 such that
P (A + B2 K) + (A + B2 K) P + 2 P B1 B1 P + 1 (C + DK) (C + DK) < 0 = 0
Letting S = P 1 and L = KS, this is equivalent to
AS + B2 L + SA + L B2 + 2 B1 B1 + 1 (CS + DL) (CS + DL) < 0
Exercise 61 Consider a system described by (17.30) and satisfying (17.31). Suppose there
exist a matrix L, a positive-definite symmetric matrix S and a positive scalar which satisfy
AS + B2 L + SA + L B2 + 2 B1 B1 + 2S + 1 (CS + DL) (CS + DL) 0

(17.33)

Then the controller given by (17.11) renders system (17.30) globally uniformly exponentially
stable about the origin with rate of convergence and Lyapunov matrix S 1
Using a Schur complement result, one can show that satisfaction of quadratic matrix
inequality (17.32) is equivalent to satisfaction of the following matrix inequality :


AS + B2 L + SA + L B2 + 2 B1 B1 (CS + DL)
<0
(17.34)
CS + DL
I
Note that this inequality is linear in S, L, and . Recalling the discussion in Section 17.2,
we can globally stabilize system (17.30) with rate of convergence by solving the following
optimization problem:
Minimize subject to


AS + B2 L + SA + L B2 + 2 B1 B1 + 2S (CS + DL)
CS + DL
I

0
(17.35)

S I

and letting u = LS 1 x.

I L
L S

230

17.3.1

CHAPTER 17. QUADRATIC STABILIZABILITY

Generalization

Here we consider systems whose nonlinearity is characterized by several nonlinear functions,

i : IRpi IRmi , i = 1, 2, . . . , l :
x = Ax + B11 1 (t, C1 x+D1u) + . . . + B1l l (t, Cl x+D1 u) + B2 u

(17.36)

where A, B2 and B1i , Ci , Di (i = 1, 2, , l) are constant matrices of appropriate dimensions.

Each i satisfies
||i (t, zi )|| ||zi ||
(17.37)
If we introduce fictitious outputs zi = Ci x+Di u and fictitious inputs wi = i (t, Ci x+Di u),
the above nonlinear system can be described as a feedback combination of a linear time
invariant system
x = Ax + B11 w1 + . . . + B1l wl + B2 u
z1 = C1 x + D1 u
..
.
zl = Cl x + Dl u
and multiple memoryless nonlinearities:
wi = i (t, zi )

for i = 1, 2, , l

This is illustrated in Figure 17.3.

Figure 17.3: A multi-loop feedback system

One can readily obtain the following result:
Theorem 44 Consider a system described by (17.36) and satisfying (17.37). Suppose there
exist a matrix L, a positive-definite symmetric matrix S and l positive scalars 1 , 2 , . . . , l
which satisfy the following matrix inequality :

AS + B2 L + SA + L

B2

l
X
i=1

i 2 B1i B1i + 1
i (Ci S + Di L) (Ci S + Di L) < 0 (17.38)

Then the controller given by (17.11) renders system (17.36) globally uniformly exponentially
stable about the origin with Lyapunov matrix S 1 .

17.3. ANOTHER SPECIAL CLASS OF NONLINEAR SYSTEMS

231

Proof. Similar to proof of last theorem.

Note that, using a Schur complement result, inequality (17.38) is equivalent to the following inequality which is linear in L, P and the scaling parameters 1 , , l :

AS + B2 L + SA + L B2 + li=1 i 2 B1i B1i

(C1 S + D1 L) (C2 S + D2 L) (Cl S + Dl L)

C1 S + D1 L
1 I
0

C2 S + D2 L
0
2 I

..
..
..
..
.
.

.
.
.
.
.
Cl S + Dl L
0
0

l I
Recalling previous discussions, we can globally stabilize system (17.36) with rate of convergence by solving the following optimization problem:
Minimize subject to
P

AS + B2 L + SA + L B2
+ li=1 i 2 B1i B1i
+ 2S

C
S
+
D
1
1L

C2 S + D 2 L

.
Cl S + D l L

(C1 S + D1 L)
1 I
0
.
.
.
0

(C2 S + D2 L)
0
2 I
.
.
.
0

(Cl S + Dl L)
0
0
.
.
.
l I

I
L

L
S

(17.39)
and letting u = LS 1 x.

<0

232

17.4

CHAPTER 17. QUADRATIC STABILIZABILITY

Yet another special class

x = Ax + B1 (t, Cx + Du) + B2 u

(17.40)

z (t, z) 0

(17.41)

where
for all z.

Figure 17.4: A first and third quadrant nonlinearity

Theorem 45 Consider a system described by (17.40) and satisfying (17.41). Suppose there
exist a matrix L, a positive-definite symmetric matrix S and a positive scalar which satisfy
AS + B2 L + SA + L B2 < 0
B1 = CS + DL

(17.42a)
(17.42b)

Then controller (17.11) renders system (17.40) globally exponentially stable about the origin
with Lyapunov matrix S 1 .
Proof. Apply corresponding analysis result to closed loop system.
Note that (17.42) has a solution with S, > 0 if and only if the following optimization
problem has a minimum of zero:
Minimize subject to
AS + B2 L + SA + L B2

<

>0


I
B1 CS DL

B1 SC L D
I

17.4. YET ANOTHER SPECIAL CLASS

233

Exercise 62 Prove the following result: Consider a system described by (17.40) and satisfying (17.41). Suppose there exist a matrix L, a positive-definite symmetric matrix S, and
positive scalars , which satisfy
AS + B2 L + SA + L B2 + 2S 0
B1 = CS + DL

(17.43a)
(17.43b)

Then controller (17.11) renders system (17.40) globally uniformly exponentially stable about
the origin with rate and with Lyapunov matrix S 1 .

17.4.1

Generalization

Here we consider systems described by

x = Ax + B1 w + B2 u
z = Cx + D1 w + D2 u
w = (t, z)

(17.44)

z (t, z) 0

(17.45)

where
for all z.
Theorem 46 Consider a system described by (17.44) and satisfying (17.45). Suppose there
exist a matrix L, a positive-definite symmetric matrix S and positive scalars and which
satisfy


AS + SA + B2 L + L B2 + 2S
(CS + D2 L) B1
0
(17.46)
CS + D2 L B1
(D1 + D1 )
Then controller (17.11) renders system (17.40) globally uniformly exponentially stable about
the origin with rate and with Lyapunov matrix S 1 .
Proof.
Note that (17.46) is linear in L, S, .
If D + D > 0, then (17.46) is equivalent to
AS + SA + B2 L + L B2 + 2S + 1 (B1 SC L D2 )(D1 + D1 )1 (B1 CS D2 L) 0

234

CHAPTER 17. QUADRATIC STABILIZABILITY

Chapter 18
Singularly perturbed systems
18.1

Introduction and examples

Roughly speaking, a singularly perturbed system is a system which depends on a scalar

parameter in such a fashion that for a particular value of the parameter (usually taken to
be zero) the system order reduces.

18.1.1

High gain output feedback

Consider a scalar-input scalar-output system described by

x 1 = x2
x 2 = u
y = 2x1 + x2
This system has the following transfer function
s+2

G(s)
= 2
s
Suppose we subject this system to high-gain static output feedback, that is, we let
u = ky
where k is large. Letting = k 1 , the closed loop system is described by
x 1 = x2
x 2 = 2x1 x2
We are interested in the behavior of this system for > 0 small.
The eigenvalues of the closed loop system matrix


0
1
21 1
235

(18.1)
(18.2)

236

CHAPTER 18. SINGULARLY PERTURBED SYSTEMS

are s () and 1 f () where

4

() =
,
1 + 1 8
s

() =

1+

1 8
2

and
lim s () = 2 ,

lim f () = 1

s ()t

Thus, for small > 0, system responses are characterized by a slow mode e
f
fast mode e ()t/ .
For = 0, the differential equation (18.2) reduces to the algebraic equation

and a

2x1 x2 = 0
Substitution for x2 into (18.1) yields
x 1 = 2x1
Note that the dynamics of this system are described by the limiting slow mode.

18.1.2

Single-link manipulator with flexible joint

Consider the single-link manipulator illustrated in Figure 18.1. There is a flexible joint or
shaft between the link and the rotor of the actuating motor; this motor exerts a control
torque u. The equations of motion for this system are:

q1
m,I

q2

k
u

Figure 18.1: Single link manipulator with flexible joint

18.1. INTRODUCTION AND EXAMPLES

237

I q1 = W l sin q1 c(q1 q2 ) k(q1 q2 )

J q2 = u
+ c(q1 q2 ) + k(q1 q2 )
where W is the weight of the link and k > 0 and c > 0 represent the stiffness and damping
coefficients, respectively, of the shaft. We want to design a feedback controller for u which
guarantees that the resulting closed loop system is globally asymptotically stable about
q1 (t) q2 (t) 0. Suppose for the purpose of controller design, we simplify the problem and
model the shaft as rigid. This yields the following rigid model:
(I + J)
q1 = W l sin q1 + u
A controller which globally exponentially stabilizes this model about q1 (t) 0 is given by
u = p(q1 , q1 ) = kp q1 kd q1
with
kp > W l,

kd > 0

Will this controller also stabilize the flexible model provided, the flexible elements are
sufficiently stiff? To answer this question we let
k = 2 k0 ,

c = 1 c0

where k0 , c0 > 0 and consider system behavior as 0. Introducing state variables

x1 = q1 ,

x2 = q1 ,

y1 = 2 (q2 q1 ) ,

y2 = 1 (q2 q1 )

the closed loop flexible model is described by

x 1
x 2
y 1
y 2

= x2
= I 1 W l sin x1
+ I 1 (k0 y1 + c0 y2 )
=
y2
1
1
= I W l sin x1 + J p(x) I1 (k0 y1 + c0 y2 )

where I1 := I 1 + J 1 .
Suppose we let = 0 in this description. Then the last two equations are no longer
differential equations; they are equivalent to
y2 = 0
1

y1 = k01 I(I
W l sin x1 + J 1 p(x))
Substituting these two equations into the first two equations of the above state space description yields
x 1 = x2
x 2 = (I + J)1 (W l sin x1 + p(x))
Note that this is a state space description of the closed loop rigid model.

238

18.2

CHAPTER 18. SINGULARLY PERTURBED SYSTEMS

Singularly perturbed systems

The state space description of the above two examples are specific examples of singularly
perturbed systems. In general a singularly perturbed system is described by
x = f (x, y, )
y = g(x, y, )

(18.3a)
(18.3b)

where x(t) IRn and y(t) IRm describe the state of the system at time t IR and > 0
is the singular perturbation parameter. We suppose 0 <
.
Reduced order system. Letting = 0 results in
x = f (x, y, 0)
0 = g(x, y, 0)
If we assume that there is a continuously differentiable function h0 such that
g(x, h0 (x), 0) 0 ,
then y = h0 (x). This yields the reduced order system:
x = f (x, h0 (x), 0)
A standard question is whether the behavior of the original full order system (18.3) can be
predicted by looking at the behavior of the reduced order system. To answer this question
one must also look at another system called the boundary layer system.
Boundary layer system. First introduce fast time described by
= t/
and let
( ) := x( ) = x(t)
( ) := y( ) h0 (x( )) = y(t) h0 (x(t))
to obtain a regularly perturbed system:
= f (, h0() + , )
= g(, h0() + , )

h0
()f (, h0() + , )
x

) 0; hence ( ) = (0) = x(0) =: x0 and

Setting = 0 results in (
= g(x0 , h0 (x0 ) + , 0)

18.2. SINGULARLY PERTURBED SYSTEMS

239

This is a dynamical system parameterized by x0 . It has a unique equilibrium state = 0.

We will refer to it as the boundary layer system.
A standard result is the following: Suppose the boundary layer system is exponentially
stable and this stability is uniform with respect to x0 ; then on any compact time interval not
including 0, the behavior of x for the full order system approaches that of x for the reduced
order system as 0.
Another result is the following: Suppose the reduced order system is exponentially stable
about 0 and the boundary layer is exponentially stable for all x0 . Then, under certain
regularity conditions, the full order system is exponentially stable for sufficiently small.
Recalling the first example of the last section, the boundary layer system is given by the
exponentially stable system
=
Since the reduced order system is exponentially stable, the full order system is exponentially
stable for > 0 small, that is, for k > 0 large.
The boundary layer system for the second example is given by
1 = 2
2 = I1 (k0 1 + c0 2 )
Clearly this system is exponentially stable. Since the reduced order system is exponentially
stable, the closed loop flexible system is exponentially stable for > 0 sufficiently small,
that is, for k and c sufficiently large. This is illustrated by numerical simulations in Figure
18.2.

CHAPTER 18. SINGULARLY PERTURBED SYSTEMS

link angle (rad)

link angle (rad)

link angle (rad)

link angle (rad)

240

1
rigid
0
-1
0

10
time (sec)

12

14

16

18

20

16

18

20

16

18

20

16

18

20

1
flexible: k=900, c=30

0
-1
0

10
time (sec)

12

14

1
flexible: k=9, c=3

0
-1
0

10
time (sec)

12

14

10
flexible: k=1, c=1
0
-10
0

10
time (sec)

12

14

Figure 18.2: Comparison of rigid case and flexible case with large and small stiffnesses

18.3. LINEAR SYSTEMS

18.3

241

Linear systems

A linear time invariant singularly perturbed system is described by

x = A()x + B()y
y = C()x + D()y

(18.4)

where A(), B(), C(), and D() are differentiable matrix valued functions defined on some
interval [0,
). For > 0, this is a LTI system with A-matrix


A()
B()
.
(18.5)
A() =
1 C() 1 D()
We assume that
D(0) is invertible
Letting = 0 results in the reduced order system:

x = Ax
where
A := A(0) B(0)D(0)1 C(0)
The boundary layer system is described by:
= D(0)
Let
r1 , r2 , . . . , rn
and
bl
bl
bl
1 , 2 , . . . , m

be the eigenvalues of A and D(0), respectively, where an eigenvalue is included p times if its
algebraic multiplicity is p. Then we have the following result.
Theorem 47 If D(0) is invertible, then for each > 0 sufficiently small, A() has n
eigenvalues
s1 (), s2 (), . . . , sn ()
with
lim si () = ri ,

i = 1, 2, . . . , n

and m eigenvalues
f1 ()/,

f2 ()/,

...,

fm ()/

with
lim fi () = bl
i ,

i = 1, 2, . . . , m

242

CHAPTER 18. SINGULARLY PERTURBED SYSTEMS

Recalling the high gain feedback example we have

r1 = 2 ,

bl
1 = 1

Hence the closed loop system has two eigenvalues s1 () and f1 ()/ with the following
properties
lim s1 () = 2 ,
lim f1 () = 1
0

The above theorem has the following corollary: If all the eigenvalues of A and D(0) have
negative real parts then, for sufficiently small, all the eigenvalues of A() have negative real
parts. If either A or D(0) has an eigenvalue with a positive real part, then for sufficiently
small, A() has an eigenvalue with positive real part.

18.4

Exact slow and fast subsystems

Suppose a singularly perturbed system has a boundary layer system which is just marginally
stable, that is, just stable but not exponentially stable. Then we cannot apply our previous
results. Further analysis is necessary.
The slow system. Consider a singularly perturbed system described by (18.3). A subset
M of IRn IRm is an invariant set for (18.3) if every solution originating in M remains there.
A parameterized set M , is a slow manifold for (18.3), if for each [0,
), M is an invariant
n
m
set for (18.3) and there is a continuous function h : IR [0,
) IR such that
M = {(x, y) IRn IRm : y = h(x, )} .

(18.6)

When system (18.3) has a slow manifold described by (18.6), then the behavior of the system
on this manifold is governed by
x = f (x, h(x, ), )
y = h(x, ) .

(18.7a)
(18.7b)

We refer to system (18.7a) as the slow system associated with singularly perturbed system
(18.3). Thus the slow system describes the motion of (18.3) restricted to its slow invariant
manifold.
The slow manifold condition. Now note that if M , given by (18.6), is an invariant
manifold for (18.3), then h must satisfy the following slow manifold condition:
g(x, h(x, ), )

h
(x, )f (x, h(x, ), ) = 0
x

(18.8)

For each > 0, this condition is a partial differential equation for the function h(, ); for
= 0, it reduces to a condition we have seen before:
g(x, h(x, 0), 0) = 0 .

18.4. EXACT SLOW AND FAST SUBSYSTEMS

243

If we introduce the boundary layer state z(t) given by

z := y h(x, ) ,

(18.9)

then the original system (18.3) is equivalently described by

x = f(x, z, )
z = g(x, z, )

(18.10a)
(18.10b)

with
f(x, z, ) := f (x, h(x, ) + z, )
g(x, z, ) := g(x, h(x, ) + z, )

(18.11a)
h
(x, )f (x, h(x, ) + z, )
x

(18.11b)

Note that M is an invariant manifold for system (18.3) if and only if the manifold
{(x, z) IRn IRm : z = 0}

(18.12)

is invariant for (18.10). Also, the slow manifold condition is equivalent to

g(x, 0, ) = 0 .

(18.13)

From (18.10) and (18.13), it should be clear that if the slow manifold condition holds then,
for any x(0) = x0 and z(0) = 0, the resulting system behavior satisfies (assuming uniqueness
of solutions), z(t) 0. Hence the manifold described by (18.12) is an invariant manifold. So,
the slow manifold condition is necessary and sufficient for the existence of a slow invariant
manifold.
The fast system. Introducing the fast time variable
:= t/
and letting
( ) := x(t) = x( )
( ) := z(t) = z( )
system (18.10) is described by the following regularly perturbed system:
= f(, , )
= g(, , )
We define
to be the fast system.

= g(, , )

244

CHAPTER 18. SINGULARLY PERTURBED SYSTEMS

18.4.1

Approximations of the slow and fast systems

Suppose
f (x, h(x, ), ) = f0 (x) + f1 (x) + o(; x)
where
lim o(; x)/ = 0 .

Then
x = f0 (x)
and
x = f0 (x) + f1 (x)
are called the zero and first order slow systems, respectively. Note that the zero order slow
system is the reduced order system.
To compute these approximations, we suppose that
h(x, ) = h0 (x) + h1 (x) + o(; x)
Then,
f0 (x) = f (x, h0 (x), 0)
f1 (x) =

f
f
(x, h0 (x), 0)h1 (x) +
(x, h0 (x), 0)
y

The following expressions for h0 and h1 can be obtained from the slow manifold condition
g(x, h0 (x), 0) = 0

(18.14a)

h0
g
g
(x, h0 (x), 0)h1 (x)
(x)f0 (x) +
(x, h0 (x), 0) = 0
y
x

(18.14b)

To approximate the fast system, suppose that

g(, , ) = g0 (, ) + g1 (, ) + o(; , )
Then
= g0 (, )
and
= g0 (, ) + g1 (, )

(18.15)

are called the zero and first order fast systems, respectively. Note that = 0 is an equilibrium
state for both. Using (18.11b) one can obtain
g0 (, ) = g(, h0() + , 0)
g
h0
g
g1 (, ) =
(, h0 () + , 0)h1()
()f (, h0() + , 0) +
(, h0 () + , 0)
y
x

Note that the zero order fast system is the boundary layer system.
A reasonable conjecture is the following: if the reduced order system is exponentially
stable and the first order fast system is exponentially stable for > 0 sufficiently small, then
the full order system is exponentially stable for small > 0.

18.4. EXACT SLOW AND FAST SUBSYSTEMS

18.4.2

245

A special class of systems

Here we consider a special class of singularly perturbed systems described by

x = f(x, ) + B()y
y = g(x, ) + D()y

(18.16)

with D(0) invertible.

Utilizing (18.14a) the function h0 is given by
g(x, 0) + D(0)h0 (x) = 0 ;
hence
h0 (x) = D(0)1 g(x, 0)

and

f0 (x) = f(x, 0) B(0)D(0)1 g(x, 0) .

Utilizing (18.14b) the function h1 is given by

D(0)h1 (x)

h0

g
D
(x)f0 (x) +
(x, 0) +
(0)h0 (x) = 0
x

The function g0 is given by

g0 (, ) = D(0)
and the function g1 is given by

g
D
h0
()[f0 () + B(0)] +
(x, 0) +
(0)[h0 () + ]
x

h0

g
D
h0
D
= D(0)h1 ()
()f0 () +
(x, 0) +
(0)h0 () + [
()B(0) +
(0)]
x

= D1 ()

g1 (, ) = D(0)h1 ()

where
D1 () = D(0)1

g
D
(, 0)B(0) +
(0)
x

So the first order fast system is described by

= [D(0) + D1 ()]
Example 175 Consider now the closed loop flexible one link manipulator with
k = 2 k0 ,

c = c0

This is described by
x 1
x 2
y 1
y 2

=
=
=
=

x2
I 1 W l sin x1
+ I 1 (k0 y1 + c0 y2 )
y2
I 1 W l sin x1 + J 1 p(x) I1 (k0 y1 + c0 y2 )

246

CHAPTER 18. SINGULARLY PERTURBED SYSTEMS

where I1 = I 1 + J 1 . This is in the form of (18.16) with





0
1
0
0
B() =
,
D() =
I 1 k0 I 1 c0
I1 k0 I1 c0
and
g(x, ) =

0
1
I W l sin x1 + J 1 p(x)

D1 =

(I + J)1 kd
0
1

0
I c0

Hence

and
D(0) + D1 =

(I + J)1 kd
I1 k0

1
I1 c0

A necessary and sufficient condition for this sytem to be exponentially for small > 0 is
kd < (I + J)I1 c0
Figures 18.3 and 18.4 illustrate numerical simulations with the above condition unsatisfied
and satisfied, respectively.

247

1
rigid
0
-1
0

10
time (sec)

12

14

16

18

20

14

16

18

20

14

16

18

20

14

16

18

20

20
flexible: k=900, c=0.1
0
-20
0

10
time (sec)

12

50
flexible: k=9, c=0.1
0
-50
0

link angle (rad)

link angle (rad)

link angle (rad)

link angle (rad)

18.4. EXACT SLOW AND FAST SUBSYSTEMS

10
time (sec)

12

x 10

flexible: k=1, c=0.1

0
-5
0

10
time (sec)

12

Figure 18.3: Flexible case with low damping

CHAPTER 18. SINGULARLY PERTURBED SYSTEMS

link angle (rad)

link angle (rad)

link angle (rad)

link angle (rad)

248

1
rigid
0
-1
0

10
time (sec)

12

14

16

18

20

14

16

18

20

14

16

18

20

14

16

18

20

2
flexible: k=900, c=0.5
0
-2
0

10
time (sec)

12

1
flexible: k=9, c=0.5
0
-1
0

10
time (sec)

12

500
flexible: k=1, c=0.5
0
-500
0

10
time (sec)

12

Figure 18.4: Flexible case with medium damping

18.4. EXACT SLOW AND FAST SUBSYSTEMS

249

Suppose we colocate the rate sensor with the actuator. Then

u =
=
=
=
and

Now

x 1
x 2
y 1
y 2

=
=
=
=

p(q1 , q2 )
kp q1 kd q2
kp x1 kd x2 kd y2
p(x) kd y2

x2
I 1 W l sin x1
+ I 1 (k0 y1 + c0 y2 )
y2
I 1 W l sin x1 + J 1 p(x) I1 k0 y1 (I1 c0 + J 1 kd )y2
D() =

0
1
1
1

I k0 (I c0 + J 1 kd )

D(0) + D1 =

(I + J)1 kd
I1 k0

This yields

1
(I1 c0 + J 1 kd )

This is exponentially stable for > 0 small.

Figure 18.5 illustrates numerical simulations with c0 = 0.

CHAPTER 18. SINGULARLY PERTURBED SYSTEMS

link angle (rad)

link angle (rad)

link angle (rad)

link angle (rad)

250

1
rigid
0
-1
0

10
time (sec)

12

14

16

18

20

14

16

18

20

14

16

18

20

14

16

18

20

1
flexible: k=900, c=0
0
-1
0

10
time (sec)

12

1
flexible: k=10, c=0
0
-1
0

10
time (sec)

12

5000
flexible: k=1, c=0
0
-5000
0

10
time (sec)

12

Figure 18.5: Rigid and flexible case with colocated rate feedback

Chapter 19
Input output systems
19.1

Input-output systems

To describe an input-output system, we need several ingredients. First we have the time set
T . Usually
T = [0, )
for continuous-time systems or

T = { 0, 1, 2, . . . , }

for discrete-time systems. We have the set U of input signals u : T U where U is the set
of input values. Common examples are U = IRm and U is the set of piecewise continuous
functions u : [0, ) IRm . We also have the set Y of output signals y : T Y where U
is the set of output values. Common examples are Y = IRo and Y is the set of piecewise
continuous functions y : [0, ) IRo .
By saying that a function s : [0, ) IRn is piecewise continuous we will mean that on every
finite interval [0, T ], the function is continuous except (possibly) at isolated points.
Example 176 Two examples of piecewise continuous functions.
(i) Square wave function.

N t < N + 21
1 if
s(t) =

1 if N + 12 t N + 1

N = 0, 1, 2, . . .

(ii)

s(t) =

(1 t)1 if 0 t < 1
0
if 1 t

An input-output system is a map

G:U Y
251

252

CHAPTER 19. INPUT OUTPUT SYSTEMS

Figure 19.1: Input-output system

that is, for each input signal u U the system produces an output signal y = G(u) with
y Y.
Example 177 Consider a linear time-invariant state space system described by
x = Ax + Bu
y = Cx + Du
x(0) = x0
where x0 IRn is fixed and t [0, ), u(t) IRm , y(t) IRo . For each piecewise continuous input signal u, this system produces a unique piecewise continuous output signal y,
specifically
Z t
At
y(t) = Ce x0 +
CeA(t ) Bu( ) d + Du(t)
for t 0 .
0

Example 178 Consider a discrete-time linear time-invariant system described by

x(k + 1) = Ax(k) + Bu(k)
y(k) = Cx(k) + Du(k)
x(0) = x0
where k { 0, 1, 2, . . . , }, u(k) IRm , y(k) IRo , and x0 IRn is fixed.
k

y(k) = CA x0 +

k1
X

CA(k1j) Bu(j) + Du(k)

j=0

Example 179 Consider a linear time-invariant system described by a delay-differential equation:

x(t)

= A0 x(t) + A1 x(th) + Bu(t)

y = Cx(t) + Du(t)
x(t) = x0 (t)
for
0th
where the continuous function x0 () is fixed and t [0, ), u(t) IRm , y(t) IRo . For each
piecewise continuous input signal u, this system produces a unique piecewise continuous
output signal y,

19.2. SIGNAL NORMS

19.1.1

253

Causality

Basically, an input-output system is causal if the current value of the output depends only
on the previous values of the input. Mathematically, we express this as follows. A system G
is causal, if
u1 (t) = u2 (t)
for t T
implies

G(u1 )(t) = G(u2 )(t)

for t T .

The systems in the above examples are causal. As examples of noncausal systems, consider
y(t) = u(t + 1)
Z
y(t) =
u( ) d
t

19.2

Signal norms

To talk about the size of signals, we introduce signal norms. Consider any real scalar p 1.
We say that a piecewise continuous signal s : [0, ) IRn is an Lp signal if
Z
||s(t)||p dt <
0

where ||s(t)|| is the Euclidean norm of the n-vector s(t). If S is any linear space of Lp signals,
then the scalar valued function || ||p defined by
||s||p :=

Z

 p1
||s(t)|| dt
p

is a norm on this space. We call this the p-norm of s. Common choices of p are one and two.
For p = 2 we have
Z
 12
2
||s||2 =
||s(t)|| dt
.
0

This is sometimes called the rms (root mean square) value of the signal. For p = 1,
Z
||s||1 =
ks(t)k dt .
0

We say that s is an L

signal if
ess sup ||s(t)|| < .
t0

By ess sup we mean that supremum is taken over the set of times t where s is continuous. If
S is a any linear space of L signals, then the scalar valued function || || defined by
||s|| := ess sup ||s(t)||
t0

254

CHAPTER 19. INPUT OUTPUT SYSTEMS

is a norm on this space. We call this the -norm of s. Also ||s(t)|| ||s|| a.e (almost
everywhere) that is, everywhere except(possibly) where s is not continuous.
Example 180 Lp or not Lp
(i) Lp for all p.

s(t) = et

with

> 0.

Since |et
| 1 for all t 0, this signal is an L signal; also ksk = 1. For 1 p < ,
R
we have 0 |et |p dt = 1/p. Hence, this is an Lp signal with kskp = (1/p)1/p .

(ii) Not Lp for all p.

s(t) = et

(iii) L but not Lp for any other p.

s(t) 1

(iv) Not L but Lp for any other p

s(t) =

(v) Not L1 but Lp for any p > 1

(1 t) 2 for 0 t < 1
0
for
1t
s(t) = (1 + t)1

For 1 p < , we say that a piecewise continuous signal s is an extended Lp signal or

an Lpe signal if
Z T
||s(t)||p dt <
0

for all T 0. A piecewise continuous signal s is an extended L signal or an L

e signal if
ess sup ||s(t)|| <
0tT

for all T 0. Clearly, if s is Lp , then it is Lpe .

Example 181 Lpe or not Lpe
(i)

s(t) = et
Lpe for all p.
(iv)
s(t) =

(1 t) 2 for 0 t < 1
0
for
1t

p
Not L
e but L for any other p
p
Note that if a signal is L
e then it is Le for all p.

19.3. INPUT-OUTPUT STABILITY

19.3

255

Input-output stability

Suppose U is a set of Lpe signals u : [0, ) IRm and Y is a set of Lpe signals y : [0, ) IRo .
Consider an input-output system
G:U Y.
DEFN. The system G is Lp stable if it has the following properties.
(i) Whenever u is an Lp signal, the output y = G(u) is an Lp signal.
(ii) There are scalars and such that, for every Lp signal u U, we have
||G(u)||p + ||u||p

(19.1)

The Lp gain of a system is the infimum of all such that there is a which guarantees (19.1)
for all Lp signals u U.
Example 182 Consider the memoryless nonlinear SISO system defined by
y(t) = sin(u(t)) + 1
Since
|y(t)| = 1 + | sin(u(t))| 1 + |u(t)|
it follows that for any 1 p and any Lp signal u we have
||y||p 1 + ||u||p
Hence this system is Lp stable with gain 1 for 1 p .
Example 183 Consider the simple delay system with time delay h 0:

0
for 0 t h
y(t) =
u(t h) for
ht
One can readily show that for any 1 p and any Lp signal u we have
||y||p = ||u||p
Hence this system is Lp stable for 1 p .
Example 184 (Simple integrator is not io stable.)
x = u
y = x
x(0) = 0

256

CHAPTER 19. INPUT OUTPUT SYSTEMS

For 1 p < consider

u(t) =

1 for 0 t 1
0 for
1t

y(t) =

t for 0 t 1
1 for
1t

Then

Since u is an Lp signal and y is not, this system is not Lp stable.

For p = consider
u(t) 1
Then

y(t) = t
Since u is an L signal and y is not, this system is not L stable.

19.4

General convolution systems

Suppose U is a set of Lpe signals u : [0, ) IRm and Y is a set of Lpe signals y : [0, ) IRo .
Consider the input-output system defined by the convolution integral
Z t
y(t) =
(19.2)
H(t )u( ) d
0

where H : [0, ) IRom is a piecewise continuous matrix valued function of time. H is

called the impulse response of the system.
As a specific example, consider a finite-dimensional linear system described by
x = Ax + Bu
y = Cx
x(0) = 0
Then y(t) =

19.4.1

Rt
0

CeA(t ) Bu( ) d . Hence this is a convolution system with

H(t) = CeAt B .

Systems with L1 impulse responses

Consider a convolution system with impulse response H. We say that H is L1 if

Z
||H(t)|| dt <
0

where

and we define

||H(t)|| = max [H(t)]

||H||1 :=

||H(t)|| dt .

19.4. GENERAL CONVOLUTION SYSTEMS

257

Example 185 Consider the SISO system

x = x + u
y = x
x(0) = 0
with > 0. Here
y(t) =

H(t )u( ) d

with

H(t) = et
For any T > 0,

Hence H is L and

T
0

|H(t)| dt =

1
(1 eT )

||H||1 =

Remark 9 Suppose there are constants > 0 and 0 such that

kH(t)k et
for all t 0. Then H is L1 .

(19.3)

258

CHAPTER 19. INPUT OUTPUT SYSTEMS

Remark 10 Consider
x = Ax + Bu
y = Cx
where A is Hurwitz. Then H(t) = CeAt B is L1 . To see this, first choose any > 0 satisfying
<
:= max {() : is an eigenvalue of A} .

(19.4)

Then there exists > 0 such that kH(t)k et .

Theorem 48 Consider a convolution system (19.2) where the impulse response H is L1 .

For every p [1, ], this system is Lp stable and for every Lp signal u U,
||y||p ||H||1||u||p
Remark 11 The above result is a nice general result. However, except for p = , it usually
yields a conservative estimate for the Lp gain of a convolution system. Below we present an
exact characterization of the L2 norm of a convolution system.
Remark 12 The main problem in applying the above result is in calculating kHk1.
Before proving Theorem 48, we need the following result.
Holders inequality. Suppose f, g : [0, ) IR are piecewise continuous and consider
any positive real scalars p and q which satisfy
1 1
+ =1
p q
Then, for all T > 0,
Z

T
0

|f (t)g(t)| dt

Z

 p1 Z
|f (t)|p dt

 1q
|g(t)|q dt
.

19.4. GENERAL CONVOLUTION SYSTEMS

259

Proof of Theorem 48. First note that, for any input signal u and any time t, we have
Z t




||y(t)|| =
H(t
)u(
)
dt


0
Z t

||H(t )u( )|| dt

0
Z t

||H(t )|| ||u( )|| dt .

0

Case : p = . Suppose u is L , then ||u( )|| ||u|| a.e. and

||y(t)||

||H(t )|| ||u|| dt

||H(t )|| d ||u||

||H(
)|| d
||u||
Z0

||H(
)|| d
||u||
0

= ||H||1||u|| .

Hence y is L and

||y|| ||H||1||u|| .

Case : p < . Suppose u is Lp . Choose q such that

1 1
+ = 1.
p q
Then, using Holders inequality, we obtain
||y(t)||

||H(t )||1/q ||H(t )||1/p ||u( )|| dt

Z

||H(t )||d

1/q
||H||1

Z

1/q Z

1/p
||H(t )||||u( )|| dt
p

1/p
||H(t )||||u( )|| dt
.
p

Hence, for any T > 0, we have

Z

T
p

||y(t)|| dt

p/q
||H||1

||H(t )||||u( )||p d dt .

260

CHAPTER 19. INPUT OUTPUT SYSTEMS

By changing the order of integration, we can write

Z TZ t
Z TZ T
p
||H(t )||||u( )|| d dt =
||H(t )||||u( )||p dtd
0
0
0

Z TZ T
=
||H(t )|| dt ||u( )||p d
0

Z T

||H||1||u( )||p d
0
Z T
= ||H||1
||u( )||p d .
0

Since 1 + p/q = p, we now obtain

Z T
Z
p
p
||y(t)|| ||H||1
0

If u is Lp then

R
0

||u( )||p d .

||u( )||p is finite and

Z T
Z
p
p
||y(t)|| d ||H||1
0

||u( )||p d ;

hence y is L and

||y||p ||H||1||u||p .

Remark 13 Here we provide a way of obtaining an upper bound on kHk1 for single output
systems using a Lyapunov equation. Consider a scalar output system described by
x = Ax + Bu
y = Cx
where A is Hurwitz, that is, all eigenvalues of A have negative real parts. Consider any
satisfying (19.4) and let S be the solution to
AS + SA + 2S + BB = 0 .

(19.5)

Then
kHk1

1
CSC
2

 21

(19.6)

To see this first note that the above Lyapunov equation can be written as
(A + I)S + S(A + I) + BB = 0
Since A + I is Hurwitz, S is uniquely given by
Z
Z
(A+I)t
(A+I) t
S=
e
BB e
dt =
0

e2t eAt BB eAt dt

19.4. GENERAL CONVOLUTION SYSTEMS

261

Hence

CSC =

2t

At

At

Ce BB e C dt =

2t

H(t)H(t) dt =

We now use Holders inequality to obtain:

Z
Z
kH(t)k dt =
et et kH(t)kdt
0

e2t kH(t)k2 dt .

Z

Z

 21 Z
) dt

(e

 12 Z

e2t dt
 21

 21
(e kH(t)k) dt
t

 21
e2t kH(t)k2 dt

This implies that



1
CSC
2

kHk1

t 2

19.4.2

1
CSC
2

 21

Systems with impulses in impulse response

H(t) = H0 (t) + (t1 )H1 + (t2 )H2 +

(19.7)

where H0 is piecewise continuous, the matrices H1 , H2 , . . . are constant matrices, is the

unit impulse function and 1 , 2 , . . . 0 are constants.
Example 186
x = Ax + Bu
y = Cx + Du
Here
H(t) = CeAt B + (t)D .
Note that when H is given by (19.7),
Z t
y(t) =
H0 (t )u( ) d + H1 u(t1 ) + H2 u(t2 ) +
0

Here we generalize the L1 norm to

kHkA = kH0 k1 + kH1 k + kH2 k +
Also, whenever kHkA is bounded, we have
kykp kHkA kukp .

(19.8)

262

19.4.3

CHAPTER 19. INPUT OUTPUT SYSTEMS

L2 gain and H norm

Consider a convolution system described by (19.2) and suppose the impulse response H is
piecewise continuous. The Laplace Transform of H is a matrix valued function of a complex
variable s and is given by
Z

H(s) =
H(t)est dt .
0

for all complex s where the integral is defined. If u and y are the Laplace transforms of u
and y, respectively, then the convolution system can be described by
u(s)
y(s) = H(s)

(19.9)

is called the transfer function of the system.

and H
Suppose H has no poles in the closed right half of the complex plane. If H is L1 , we have

||H(s)||
||H||1
for all s with (s) 0.
: C Cmo . Suppose
H norm of a transfer function. Consider a transfer function H
is analytic in the open right half complex plane. Then, the H -norm of H
is defined by
H
= sup ||H()||

||H||
IR

where
Note that if H is L1 then

||H()||
= max [H()]
.
||H||1
||H||

Theorem 49 Consider a convolution system (19.2) where the impulse response H is piece be the Laplace transform of H. Then, this system is L2 stable if
wise continuous and let H
is defined for all s with (s) > 0 and ||H||
is finite. In this case, whenever
and only if H
u is L2 ,
||u||2
||y||2 ||H||
(19.10)
is the L2 gain of the system in the sense that
Actually, one can show that ||H||
=
kHk

kyk2
uL2 , u6=0 kuk2
sup

Example 187 Consider the SISO system,

x = x + u
y = x
x(0) = 0

19.4. GENERAL CONVOLUTION SYSTEMS

with > 0. Here

H(s)
=

263

1
s+

Since H(j)
= 1/( + ), it follows that |H()|
= 1/( 2 + 2 ) 2 . Hence

=
kHk

1
.

Thus, the L2 gain of this system is 1/.

19.4.4

Finite dimensional linear systems

x = Ax + Bu
y = Cx + Du
x(0) = x0

Here,
At

y(t) = Ce x0 +

(19.11a)
(19.11b)
(19.11c)

CeA(t ) Bu( ) d + Du(t) .

0

The transfer function of this system is given by

= C(sI A)1 B + D
H
Theorem 50 Suppose (A, B) is stabilizable and (C, A) is detectable. Then the following
statements are equivalent.
(i) All the eigenvalues of the matrix A have negative real parts.
have negative real parts.
(ii) All the poles of the transfer function H
(iii) The input-output system (19.11) is Lp stable for some p [1, ].
(iv) The input-output system (19.11) is Lp stable for all p [1, ].
LMI characterization of L2 gain
Here we provide an upper bound on the L2 gain of a finite-dimensional linear time-invariant
system described by (19.11).
Suppose there is a symmetric matrix P 0 and a scalar 0 such that


P A + A P + C C P B + C D
0.
(19.12)
B P + DC
2 I + D D
Then,
Z

T
2

ky(t)k dt

x0 P x0

T
0

ku(t)k2 dt

for all T 0. Hence, whenever u is L then so is y and

kyk2 + kuk2

where

x0 P x0

264

CHAPTER 19. INPUT OUTPUT SYSTEMS

Remark 14 Inequality (19.12) is an LMI in the variables P and 2 . The scalar is an

upper bound on the L2 gain of the system.
Proof. Consider any solution x() of (19.11) and let v(t) = x(t) P x(t). Then v0 := v(0) =
x0 P x0 and
v = 2x P x = 2x P Ax + 2x P Bu .
It now follows from the matrix inequality (19.12) that
x (P A + A P + CC )x + x (P B + C D)u + u (B P + D C)x + u( 2 I + DD )u 0 ,
that is,
2x P Ax + 2x P Bu + kyk2 2 kuk2 0 .
Recalling the definition of v and the expression for v,
we now obtain that
v(t)
+ kyk2 2 kuk2 0
for all t 0. Hence, for any T 0,
Z
v(t) v(0) +

T
2

ky(t)k dt

ku(t)k dt 0 .

Since v(t) 0 and v(0) = x0 P x0 , we obtain that

Z

T
2

ky(t)k dt

x0 P x0

ku(t)k dt

19.4. GENERAL CONVOLUTION SYSTEMS

265

LMI characterization of L gain

Here we provide an upper bound on the L gain of a finite-dimensional linear time-invariant
system described by (19.11).
Suppose there is a symmetric matrix P 0 and scalars > 0, 1 , 2 0 such that


P A + A P + 2P
PB
0
(19.13)
BP
21 I
and

and let

P + C C
C D

DC
2 I + D D

(19.14)

= (1 + 2 ) 2 .
Then, for all t 0,

ky(t)k et + kuk

(19.15)

where = (x0 P x0 ) 2 . Hence, whenever u is L then so is y and

kyk + kuk

(19.16)

Remark 15 For a fixed , inequalities (19.13) and (19.14) are LMIs in the variables P , 1
and 2 . The scalar is an upper bound on the L gain of the system.
Proof. Consider any solution x() of (19.11) and let v(t) = x(t) P x(t). Then v0 := v(0) =
x0 P x0 and
v = 2x P x = 2x P Ax + 2x P Bu .
It now follows from the matrix inequality (19.13) that
x (P A + A P + 2P )x + x P Bu + u B P x 21 u u 0 ,
that is,
2x P Ax + 2x P Bu 2x P x + 21 kuk2 .

Recalling the definition of v and the expression for v,

we now obtain that
v(t)
2v(t) + 21 ku(t)k2
for all t 0. Suppose now u() is L . Then ku(t)k kuk for all t 0 and
v(t)
2v(t) + 21 kuk2 .
for all t 0. It now follows that
2t

v(t) v(0)e

e2(t ) 21 kuk2 d

= v0 e2t + 1 kuk2 (1 e2t )

v0 e2t + 1 kuk2 .

266

CHAPTER 19. INPUT OUTPUT SYSTEMS

Thus,
v(t) v0 e2t + 1 kuk2 .
Recalling the matrix inequality (19.14) we obtain that
x (P + C C)x + x CDu + uD Cx + u (2 I + D D)u 0 ,
that is,
kyk2 = x CCx + 2x CDu + u DDu x P x + 2 kuk2 .
Hence, for all t 0,
ky(t)k2 v(t) + 2 ku(t)k2 v0 e2t + 1 kuk2 + 2 kuk2 ,
that is,
ky(t)k2 v0 e2t + 2 kuk2 .
Taking the square root of both sides on this inequality yields
ky(t)k et + kuk
1

where = v02 = (x0 P x0 ) 2 . This is the desired result.

(19.17)

19.4. GENERAL CONVOLUTION SYSTEMS

19.4.5

267

Linear differential delay systems

Consider a system with a delay described by

x(t)

= A0 x(t) + A1 x(th) + Bu(t)

y(t) = Cx(t)
x(t) = 0
ht0

(19.18a)
(19.18b)
(19.18c)

Suppose is the unique solution to

(t)
= A0 (t) + A1 (th)
(0) = I
(t) = 0 for h t < 0 .
Then,
y(t) =

H(t )u( ) d

where

H(t) = C(t)B .

(19.19)

Proof: Consider
x(t) =

(t )Bu( ) d

for t h. We will show that this is the unique solution to

x(t)

= A0 x(t) + A1 x(th) + Bu(t)

x(t) = 0
ht0
Clearly x(t) = 0 for h t 0 . Now note that for t 0,
Z th
Z t
x(th) =
(th )Bu( ) d =
(th )Bu( ) d
0

Now differentiate x to obtain

Z

x(t)

= (0)Bu(t) +
(t
)Bu( ) d
0
Z t
= Bu(t) +
(A0 (t ) + A1 (t h))Bu( ) d
0
Z t
Z t
= Bu(t) + A0
(t )Bu( ) d + A1
(t h)Bu( ) d
0

= A0 x(t) + A1 x(th) + Bu(t) ,

that is, x satisfies (19.20). Thus

y(t) = Cx(t) = C

where H(t) = C(t)B.

(t )Bu( ) d =

H(t )Bu( ) d
0

(19.20)
(19.21)

268

CHAPTER 19. INPUT OUTPUT SYSTEMS

Example 188 Consider a system with delay h described by

x(t)

= ax(t) bx(th) + u
y(t) = x(t)
x(t) = 0
for h t 0
with
|b| < a ,

h0

The transfer function of this system is given by

H(s)
=

1
d(s)

where

d(s) = s + a + behs

is analytic for (s) 0. Suppose = + is a pole of H;

then is
We first show that H
a zero of d, that is,
+ + a + beh(+) = 0
Hence,
= a beh cos(h) a + |b|eh a + |b| < 0 .

We now obtain an upper bound on the H norm of H.

19.5. A CLASS OF LP STABLE NONLINEAR SYSTEMS

19.5

269

A class of Lp stable nonlinear systems

Consider an input-output system described by

x = F (x, u)
y = H(x, u)
x(0) = x0
We make the following assumptions.
There is a continuously differentiable function V and a positive scalars 1 , 2 , 3 and
such that
1 kx||2 V (x) 2 ||x||2
DV (x)F (x, 0) 2V (x)
||DV (x)|| 3 ||x||
There are non-negative scalars 4 , 5 , and 6 such that
||F (x, u) F (x, 0)|| 4 ||u||
||H(x, u)|| 5 ||x|| + 6 ||u|| .
Claim: The above system is Lp stable for all p [1, ].
Example 189
x = x x3 + u
y = sin x
Proof of claim:

270

CHAPTER 19. INPUT OUTPUT SYSTEMS

19.6

Small gain theorem

19.6.1

Stability of feedback systems

Figure 19.2: A Feedback system

Figure 19.3: Feedback system for stability definition

19.6.2

Truncations

Suppose s : T IRn is a signal and consider any time T . The corresponding truncation of s
is defined by

s(t) for 0 t < T
sT (t) :=
0
for T t
Note that a system G is causal if and only if for every input u and every time T , we have
G(u)T = G(uT )T .

19.6. SMALL GAIN THEOREM

271

For any Lp signal s, we have ksT kp kskp . A signal s is Lpe if and only if sT is Lp for all
T . Also, an Lpe signal s is an Lp signal if there exists a constant such that ksT kp for
all T .

19.6.3

Small gain theorem

Suppose U1 and U2 are vector spaces of Lpe signals with the property that if they contain
a signal, they also contain every truncation of the signal. Suppose G1 : U1 U2 and
G2 : U2 U1 are causal Lp -stable systems which satisfy
||G1 (u1 )||p 1 + 1 ||u1 ||p
||G2 (u2 )||p 2 + 2 ||u2 ||p

(19.22)

Consider the feedback system described by

y1
y2
u1
u2

=
=
=
=

G1 (u1 )
G2 (u2 )
r1 + y 2
r2 + y 1

(19.23)

We assume that for each pair (r1 , r2 ) of signals with r1 U1 and r2 U2 , the above
relations uniquely define a pair (y1 , y2 ) of outputs with y1 U2 and y2 U1 .
Theorem 51 (Small gain theorem) Consider a feedback system satisfying the above conditions and suppose
1 2 < 1
Then, whenever r1 and r2 are Lp signals, the signals u1 u2, y1 , y2 are Lp and
||u1||p (1 1 2 )1 (2 + 2 1 + ||r1||p + 2 ||r2||p )
||u2||p (1 1 2 )1 (1 + 1 2 + ||r2||p + 1 ||r1||p )
Hence,
||y1 ||p 1 + 1 (1 1 2 )1 (2 + 2 1 + ||r1 ||p + 2 ||r2 ||p )
||y2 ||p 2 + 2 (1 1 2 )1 (1 + 1 2 + ||r2 ||p + 1 ||r1 ||p )
and the feedback system is Lp stable.
Proof. Suppose r1 and r2 are Lp signals. Since r1 , r2 and y1 , y2 are Lpe signals, it follows
that u1 and u2 are Lpe signals. Consider any time T > 0. Since all of the signals r1 , r2 , u1 ,
u2 , y1 , y2 are Lpe , it follows that their respective truncations r1T , r2T , u1T , u2T , y1T , y2T are
Lp signals. Since G1 is causal,
y1T = G1 (u1 )T = G1 (u1T )T
Hence,
ky1T kp = kG1 (u1T )T kp kG1 (u1T )kp 1 + 1 ku1T kp

272

CHAPTER 19. INPUT OUTPUT SYSTEMS

Since u2T = r2T + y1T , it now follows that

ku2T kp kr2T kp + ky1T kp
1 + kr2T kp + 1 ku1T kp
In a similar fashion we can show that
ku1T kp 2 + kr1T kp + 2 ku2T kp
Hence,
ku1T kp 2 + kr1T kp + 2 (1 + kr2T kp + 1 ku1T kp )
Rearranging yields
ku1T kp (1 1 2 )1 (2 + 2 1 + kr1T kp + 2 kr2T kp )
(1 1 2 )1 (2 + 2 1 + kr1 kp + 2 kr2 kp )
Since the above inequality holds for all T > 0, it follows that u1 is an Lp signal and
ku1 kp (1 1 2 )1 (2 + 2 1 + kr1 kp + 2 kr2 kp )
In a similar fashion we can show that
ku2 kp (1 1 2 )1 (1 + 1 2 + kr2 kp + 1 kr1 kp )

19.7

Application of the small gain theorem: circle criterion

Figure 19.4: A Feedback system

Consider a nonlinear system which is a feedback combination of a convolution system
Z t
y(t) =
H(t )u( ) d
0

and a memoryless nonlinear system

u(t) = (t, y(t)) .

19.7. APPLICATION OF THE SMALL GAIN THEOREM: CIRCLE CRITERION

273

Suppose that the impulse response H of the convolution system is L1 and, for some 0,
the nonlinearity satisfies
k(t, y)k kyk
(19.24)
for all t and y.
Since the impulse response H of the convolution system is L1 , the L2 gain of the con where H
is the system transfer function. We now claim that the
volution system is kHk
2
nonlinear system is L stable with gain . To see this, consider u(t) = (t, y(t)) and note
that, for any T > 0,
Z T
Z T
Z T
Z T
2
2
2
2
ku(t)k dt =
k(t, y(t))k dt
(ky(t)k) dt =
ky(t)k2 dt .
0

If y is L2 then

for all T 0; hence

T
2

ku(t)k dt

ku(t)k dt

ky(t)k2 dt
ky(t)k2 dt

This implies that u is L2 and kuk2 kyk2. Hence the nonlinear system is L2 stable with
gain . From the small gain theorem we obtain that the feedback system is L2 stable if
<1
kHk

(19.25)

For a SISO system, the above condition is equivalent to the requirement that the Nyquist
plot of the system transfer function lie within the circle of radius 1/ centered at the origin.
As a consequence, the above condition is sometimes referred to as the circle criterion.
Example 190 Consider
x = ax(t) + b sin(x(th))

where a > |b|. Letting y(t) = x(th) this system can be described by
x(t)

= ax(t) + u(t)
y(t) = x(th)
u(t) = b sin(y(t))
Here

H(s)
= esh /(s + a)

Thus,
so, kHk = 1/a. Also,

and

(t, y) = b sin(y) .

|H()|
= |e |2 /( 2 + a2 ) = 1/( 2 + a2 ) ;

|b sin(y)| |b||y| ;

hence the condition (19.24) is satisfied with = |b|. Since a > |b|, we obtain that
kHk = |b|/a < 1.

274

CHAPTER 19. INPUT OUTPUT SYSTEMS

Figure 19.5: Another feedback system

19.7.1

A more general circle criterion

Consider a nonlinear system which is a feedback combination of a SISO convolution system

Z t
y(t) =
H(t )u( ) d
0

and a SISO memoryless nonlinear system

u(t) = (t, y(t)) .
Suppose that the impulse response H of the convolution system is L1 and, for some scalars
a and b, the nonlinearity satisfies the sector condition:
ay 2 y(t, y) by 2

(19.26)

for all t and y.

We will convert this system into one considered in the previous section. To this end,
introduce a new input
a+b
.
w = u + ky
where k =
2
Then
y)
w = (t,
(19.27)
y) := (t, y) + ky and, it follows from (19.26) that
where (t,
y) y 2
y 2 y (t,
Thus,

where =

y)| |y|
|(t,

ba
2
(19.28)

u and w = u + k y, we obtain that

Since y = H

w k y)
y = H(
Hence
where

w
y = G
= (1 + k H)
1 H
.
G

(19.29)

19.7. APPLICATION OF THE SMALL GAIN THEOREM: CIRCLE CRITERION

275

Thus the systems under consideration here can be described by (19.29) and (19.27) where
satisfies (19.28). Using the circle criterion of the last section, this system is L2 stable
have negative real parts and
provided all the poles of G
<1
kGk

(19.30)

This last condition can be expressed as

2 G()
G() < 1 .

for all IR {}, that is,

h
i h
i
1
1

H() [1 + k H()]
H() < 1
2 [1 + k H()]

for all IR {}, Recalling the definitions of and k, the above inequality can be
rewritten as

abH()
H() k[H()
+ H()
]<1
(19.31)
We have three cases to consider:
Case 1: ab > 0 In this case, inequality (19.31) can be rewritten as

[H()
+ c] [H()
+ c] > R2
where

This is equivalent to

1
c=
2

1 1
+
a b



1  1 1 
R=  
2 a b

|H() + c| > R .

must lie outside the circle of radius R centered at c. Note

that is, the Nyquist plot of H
that this is the circle which intersects the real axis at 1/a and 1/b.

Figure 19.6: Circle criterion for 0 < a < b

Case 2: ab < 0: In this case (19.31) can be rewritten as

[H()
+ c] [H()
+ c] < R2
This is equivalent to
|H() + c| < R .

must lie inside the circle of radius R centered at c. Again,

that is, the Nyquist plot of H
this is the circle which intersects the real axis at 1/a and 1/b.

276

CHAPTER 19. INPUT OUTPUT SYSTEMS

Figure 19.7: Circle criterion for a < 0 < b

Case 3: ab = 0: Suppose a = 0 and b > 0. In this case k = b/2 and (19.31) can be
rewritten as

(b/2)(H()
+ H()
)<1
that is,



1

H()
>
(19.32)
b
must lie to the right of the vertical line which
This means that the Nyquist plot of H
intersects the real axis at 1/b.

Figure 19.8: Circle criterion for a = 0 and 0 < b

References
1. Desoer, C.A. and Vidyasagar, M., Feedback Systems: Input-Output Properties, Academic Press, 1975.

Chapter 20
Nonlinear H
20.1

Analysis

Consider a system,
x = F (x) + G(x)w
z = H(x)
with initial condition
x(0) = x0
where w : [0, ) IRm is a disturbance input and z is a performance output with z(t) IRp .
Desired performance. We wish that the system has the following property for some
> 0. For every initial state x0 there is a real number 0 such that for every disturbance
input w one has
Z
Z
2
2
||z(t)|| dt
||w||2 dt + 0
0

20.1.1

The HJ Inequality

Theorem 52 Suppose there is a function V which for all x, satisfies

DV (x)F (x) +

1
DV (x)G(x)G(x)T DV T (x) + H(x)T H(x) 0
4 2
V (x) 0

(20.1)

for some > 0. Then for every initial state x0 and for every disturbance input w one has
Z
Z
2
2
||z(t)|| dt
||w||2 dt + V (x0 )
(20.2)
0

277

278

CHAPTER 20. NONLINEAR H

Proof. Consider any initial state x0 and any disturbance input w. Along any resulting
solution x(), we have
dV (x(t))
= L(x(t), w(t))
dt
where
L(x, w) = DV (x)F (x) + DV (x)G(x)w
1
1
= DV (x)F (x) + 2 ||G(x)T DV (x)T ||2 + 2 ||w||2 ||w G(x)T DV (x)T ||2
4
2
1
DV (x)F (x) + 2 ||G(x)T DV (x)T ||2 + 2 ||w||2
4
Using inequality (20.1) we have:
DV (x)F (x) +

1
||G(x)T DV (x)T ||2 ||H(x)||2
2
4

Hence
L(x, w) ||H(x)||2 + 2 ||w||2

i.e.,

dV (x(t))
||z(t)||2 + 2 ||w(t)||2
dt
Integrating from 0 to T yields:
Z T
Z T
2
2
V (x(T )) V (x0 )
||w(t)||2 dt
||z(t)|| dt +

(20.3)

Rearranging terms and recalling that V (x(T )) 0, we obtain

Z T
Z T
2
2
||z(t)|| dt
||w(t)||2 dt + V (x0 ) V (x(T ))
0
0
Z T
||w(t)||2 dt + V (x0 )
2
0

2
If
R w is not2L2 , then 0 ||w(t)||
R dt = 2 and (20.2) trivially holds. If w is L2 , the integral
||w(t)|| dt is finite, hence 0 ||z(t)|| dt is finite and satisfies (20.2).
0

Hamilton Jacobi equation

DV (x)F (x) +

1
DV (x)G(x)G(x)T DV T (x) + H(x)T H(x) = 0
2
4

The worst case w is given in feedback form as

w = w (x) =

1
G(x)T DV (x)T
2
2

(20.4)

20.1. ANALYSIS

279

Quadratic V . Suppose
V (x) = xT P x
where P is a symmetric real matrix. Then DV (x) = 2xT P and the HJ inequality is
2xT P F (x) + 2 xT P G(x)G(x)T P x + H(x)T H(x) 0

LTI system. Suppose

x = Ax + Bw
z = Cx
Then the HJ is
DV (x)Ax +

1
DV (x)BB T DV (x) + xT C T Cx 0
4 2

If V (x) = xT P x, this results in

2xT P Ax + 2 xT P BB T P x + xT C T Cx 0
This is satisfied for all x iff P satisfies
P A + AT P + 2 P BB T P + C T C 0

280

20.2

CHAPTER 20. NONLINEAR H

Control
x = F (x) + G1 (x)w + G2 (x)u
z =

H(x)
u

with initial condition

x(0) = x0

Desired performance. We wish to choose a state feedback controller

u = k(x)
such that the resulting closed loop system
x = F (x) + G2 (x)k(x) + G1 (x)w


H(x)
z =
k(x)
has the following property for some > 0. For every initial state x0 there is a real number
0 such that for every disturbance input w one has
Z
Z
2
2
||w||2 dt + 0
||z(t)|| dt
0

We can guarantee the desired performance if there is a function V 0 satisfying

DV (F + G2 k) +

1
DV G1 GT1 DV T + C T C + k T k 0
4 2

Suppose V 0 satisfies The Hamilton-Jacobi Issacs (HJI) inequality

1
DV F + DV ( 2 G1 GT1 G2 GT2 )DV T + C T C 0
4
Then letting
k = 21 GT2 DV T
the desired inequality holds.

(20.5)

20.3. SERIES SOLUTION TO CONTROL PROBLEM

20.3

281

Series solution to control problem

We wish to solve the HJI equation,

DV F + 41 DV ( 2 G1 GT1 G2 GT2 )DV T + C T C = 0
for V 0. Letting
T

G(x) := [G1 (x) G2 (x)]

Q(x) := H(x) H(x),

R :=

(20.6)

2 0
0 1

the HJI equation (20.6) can be rewritten as

1
DV F DV GR1 B T DV T + Q = 0
4

(20.7)

Also letting
v :=
we have

20.3.1

u
w

w =

1 T
G1 DV T
2
2

1
v = R1 GT DV T
2

(20.8)

Linearized problem
x = Ax + Bv


Cx
z =
u

(20.9a)
(20.9b)

with
A = DF (0) ,

B = G(0),

C = DH(0)

Here the HJI equation (20.6) becomes

1
DV (x)Ax DV (x)BR1 B T DV (x) + xT C T Cx = 0
4
Considering a quadratic form
V (x) = xT P x
with P symmetric, as a candidate solution to the above HJE we obtain
xT [P A + AT P P BR1 B T P + C T C]x = 0
This is satisified for all x iff the matrix P solves the following Algebraic Riccati Equation
(ARE)
P A + AT P P BR1 B T P + C T C = 0
(20.10)
From standard linear H theory, a necessary and sufficient condition for the linear system to
have L2 -gain less than is that the above Algebraic Riccati Equation (ARE) has a solution
P with ABR1 BP Hurwitz. When (C, A) is observable, this solution P is positive definite.
In this case v is given by
v (x) = R1 B T P x
(20.11)

282

CHAPTER 20. NONLINEAR H

20.3.2

Nonlinear problem

First note that the HJI equation (20.7) can be written as

DV F vT Rv + Q = 0
1
v + R1 GT DV T = 0
2

(20.12a)
(20.12b)

Suppose
F = F [1] + F [2] + F [3] +
G = G[0] + G[1] + G[2] +
H = H [1] + H [2] + H [3] +
where F [k], G[k] and H [k] are homogeneous functions of order k. Note that
F [1] (x) = Ax,

G[0] (x) = B,

H [1] (x) = Cx

We consider a series expansion for V of the form

V = V [2] + V [3] + V [4] +

(20.13)

where V [k] is a homogeneous function of order k. Substituting (20.13) in equation (20.12b)

one obtains that
v = v[1] + v[2] + v[3] +
(20.14)
[k]

where v is the homogeneous function of order k given by

k1

v[k]

X
1
= R1
G[j] T DV [k+1j] T
2
j=0

(20.15)

Substitution into (20.12a) and equating terms of order m 2 to zero yields

m2
X

DV

[mk]

[k+1]

k=0

m1
X
k=1

v[mk] T Rv[k]

For m = 2 equation (20.16) simplifies to

DV [2] F [1]

v[1] Rv[1]

Q[2] = 0

Since F [1] (x) = Ax,

1
v[1] = R1 G[0] T DV [2] T
2
1 1 T
= R B DV [2] T
2
and
Q[2] (x) = xT C T Cx

Q[m] = 0

(20.16)

20.3. SERIES SOLUTION TO CONTROL PROBLEM

283

we obtain

1
DV [2] (x)Ax DV [2] T (x)BR1 B T DV [2] (x) + xT C T Cx = 0
4
which is the HJI equation for the linearized problem. Hence
V [2] (x) = xT P x
where P T = P > 0 solves the ARE with A := A BR1 B T P Hurwitz. Hence
v[1] (x) = R1 B T P x
Consider now any m 3. Using
m2

X
1
G[k] T DV [mk] T
v[m1] = R1
2
k=0

the sum of the terms in (20.16) depending on V [m] is given by:

DV [m] F [1] v[m1] T Rv[1] v[1] T Rv[m1] = DV [m] F [1] 2v[m1] T Rv[1]
m2
X
DV [mk] G[k] v[1]
= DV [m] F [1] + DV [m] G[0] v[1] +
k=1

= DV [m] f [1] +

m2
X
k=1

DV [mk] G[k] v[1]

where
f = F + Gv[1]

(20.17)

f [1] = A x

(20.18)

and
The sum of the terms on the lefthandside of (20.16) can now be written as:
DV [m] f [1] +

m2
X

DV [mk] F [k+1] +

k=1

k=1

= DV [m] f [1] +

m2
X
k=1

m2
X

DV [mk] f [k+1]

m2
X
k=2

DV [mk] G[k] v[1]

m2
X
k=2

v[mk] T Rv[k] + Q[m]

v[mk] T Rv[k] + Q[m]

Hence, using
f [k+1] = F [k+1] + G[k] v [1]
(20.16) can be written as
DV [m] f [1]

m2
X
k=1

DV [mk] f [k+1]

m2
X
k=2

v[mk] T Rv[k]

Q[m]

284

CHAPTER 20. NONLINEAR H

For example, with Q[m] = 0,

DV [3] f [1] = DV [2] f [2]
= DV [2] (F [2] + G[1] v[1] )
hence

1
v[2] = R1 (B T DV [3] + G[1] T DV [2] )
2

For m = 4:
DV [4] f [1] = DV [3] f [2] DV [2] f [3] + v[2] T Rv[2]
and

1
v[3] = R1 (B T DV [4] + G[1] T DV [3] + G[2] T DV [2] )
2

If
F [2] , G[1] = 0
Then
V [3] , v[2] = 0
and
v[3]

DV [4] f [1] = DV [2] f [3]

1
= R1 (B T DV [4] + G[2] T DV [2] )
2

Example 191
x = x x3 + w + u
z = u
HJI equation

Chapter 21
Performance
21.1

Analysis

Consider a system with output described by

x = f (t, x)
z = h(t, x)

(21.1a)
(21.1b)

where z is a performance output.

Lemma 17 Suppose that there exists a differentiable function V such that V (x) 0 and
DV (x)f (t, x) + h(t, x) h(t, x) 0
for all t and x. Then every solution of system (21.1) satisfies
Z
kz(t)k2 dt V (x0 )
t0

where x0 = x(t0 ).

285

(21.2)

(21.3)

286

CHAPTER 21. PERFORMANCE

Example 192 Consider

x = x
z = x
where > 0. Let V (x) = x2 /2.
Example 193
x = x3
z = x2
Let V (x) = x2 /2.
Example 194
x = 2x + sin x
z = x
Let V (x) = x2 /2.
Linear time-invariant systems. Consider a LTI system described by
x = Ax
z = Cx

(21.4)
(21.5)

where all the eigenvalues of A have negative real part. The Lyapunov equation
P A + A P + C C = 0

(21.6)

has a unique solution for P ; moreover P is symmetric and positive semi-definite. Post- and
pre-multiplying this equation by x and its transpose yields
2x P Ax + (Cx) Cx = 0 ;
thus, (21.2) holds with V (x) = x P x. Actually, in this case one can show that
Z
kz(t)k2 dt = V (x0 ) .
t0

21.2

Polytopic systems

21.2.1

Polytopic models

Here V is a vector space; so V could be IRn or the set of real n m matrices or some more
general space. Suppose v1 , . . . , vl is a finite number of elements of V. A vector v is a convex
combination of v1 , . . . , vl if it can be expressed as
v = 1 v1 + + l vN

where the real scalars 1 , . . . , N satisfy j 0 for j = 1, . . . , N and

1 + + l = 1 .

The set of all convex combinations of v1 , . . . , vN is called a polytope. The vectors v1 , . . . , vN

are called the vertices of the polytope.

21.2. POLYTOPIC SYSTEMS

287

Polytopes and LMIs. Suppose that

L(v) 0
is an LMI in v. Then this LMI holds for all v in a polytope if and only if it holds for the
vertices v1 , . . . , vN of the polytope, that is,
L(vj ) 0

for j = 1, . . . , N .

(21.7)

To see this, suppose that (21.7) holds. Since L(v) is affine in v, we can express it as
L(v) = L0 + L1 (v)
where
1 (v) is linear in v. Suppose v is in the polytope; then v =
PL

and j=1 j = 1. Hence

j=1

j vj where j 0

L(v) = L0 + L1 (
j=1 j vj ) = j=1 j L0 + j=1 j L1 (vj ) = j=1 L(vj ) 0 .

Obtaining polytopic models. Consider a parameter which satisfies

.
Then, can be expressed as
= + ( )
where = ( )/( ). Clearly 0 1. If we let 1 = 1 and 2 = then,
1 , 2 0, 1 + 2 = 1 and
= 1 + 2 ,
that is can be written as a convex combination of its two extreme values and . This is
a special case of the general result contained in the next remark.
Remark 16 This remark is useful in obtaining polytopic models. Suppose G is a matrix
valued function of some variable and for all ,

G() = G(())
depends in a
where is some parameter vector (which can depend on ). Suppose that G
multi-affine fashion on the components of the l-vector and each component of is bounded,
that is,
for
k = 1, . . . , l
k k () k
where 1 , . . . , l and 1 , . . . , l are known constants. Then, for all , the vector G() can be
expressed as a convex combination of the N = 2l matrices G1 , . . . , GN corresponding to the
extreme values of the components of ; these vertices Gj are given by

G()
where k = k or k for k = 1, . . . , l .

(21.8)

In applying to show satisfaction of Condition 1 below, let = (t, x), G() = (A(t, x) C(t, x)).
Then (Aj Cj ) = Gj for j = 1, . . . , N.

288

CHAPTER 21. PERFORMANCE

Example 195 Consider

G() =

cos 1 sin 2 + cos 1 1 + cos 1 + sin 2

For all , the matrix G() can be expressed as a convex combination of the following four
matrices:








0 1 ,
2 1 ,
0 1 ,
2 3 .
Remark 17 The following remark is also useful for obtaining polytopic models. Consider
an n-vector-valued function g of two variables t and which is differential with respect to
its second variable and satisfies the following two conditions:
(a)
g(t, 0) 0 .
(b) There are matrices, G1 , . . . , GN such that, for all t and , the derivative matrix
g
(t, )

can be expressed as a convex combination of G1 , . . . , GN .

Then, for each t and ,
g(t, ) = G
where G is some convex combination of G1 , . . . , GN .
In applying this to show satisfaction of Condition 1, let = x, g(t, ) = (f (t, x) h(t, x)).
Then


g
h
(t, ) = f
(t,
x)
(t,
x)
x
x

and (Aj Cj ) = Gj for j = 1, . . . , N.

Example 196 Consider

sin 1

cos 1
0

g(t, ) =

2
1+|2 |

Here g(t, 0) = 0 and

g
=

1
(1+|2 |)2

Hence, g(t, ) = G where G can be expressed as a convex combination of the following four
matrices:








1 0
1 0
1 0
1 0
.
,
,
,
0 1
0 0
0 1
0 0

21.2. POLYTOPIC SYSTEMS

21.2.2

289

Performance analysis of polytopic systems

In this section, we consider a specific type of uncertain/nonlinear system described by (21.1).

Specifically, we consider systems which satisfies the following condition.
Condition 1 There are matrices, Aj , Cj , j = 1, . . . , N so that for each t and x,
f (t, x) = Ax
h(t, x) = Cx

(21.9)
(21.10)

where the matrix (A C) (which can depend on t, x) is a convex combination of (A1 C1 ) , . . . , (AN CN ).
So, when Condition 1 is satisfied, system 21.1 can be described by
x = A(t, x)x
z = C(t, x)x

(21.11a)
(21.11b)

where (A(t, x) C(t, x)) is contained in the polytope whose vertices are (A1 C1 ), . . . , (AN CN ).
Hence, we sometimes refer to a system satisfying Condition 1 as a polytopic uncertain/nonlinear
system.
Suppose that
P Aj + Aj P + Cj Cj 0
for j = 1, , N .
(21.12)
Using a Schur complement result, the above inequalities can be expressed as


P Aj + Aj P Cj
0
for j = 1, , N .
Cj
I

(21.13)

Since the above inequalities are affine in (Aj Cj ), and (A(t, x) C(t, x)) is a convex combination
of (A1 C1 ), . . . , (AN CN ) it now follows that


P A(t, x) + A(t, x) P C(t, x)
0
C(t, x)
I
for all t and x. Reusing the Schur complement now results in
P A(t, x) + A(t, x) P + C(t, x) C(t, x) 0
for all t and x. Post- and pre-multiplying this equation by x and its transpose yields
2x P A(t, x)x + (C(t, x)x) C(t, x)x 0
that is,
2x P f (t, x) + h(t, x) h(t, x) 0

for all t and x; thus, (21.2) holds with V (x) = x P x. Hence every solution of the system
satisfies
Z
kz(t)k2 dt x0 P x0
(21.14)
t0

290

CHAPTER 21. PERFORMANCE

where x0 = x(t0 ).
To obtain a performance estimate for a fixed initial state one could minimize subject
to LMIs (21.13) and
x0 P x0 0 .
(21.15)
Note that, if we let S = P 1 then, (21.13) can be expressed as


Aj S + Aj S SCj
0
for j = 1, , N
Cj S
I

Also, using Schur complements, inequality (21.15) can be expressed as



S x0
0
for j = 1, , N
x0

(21.16)

(21.17)

Suppose x0 lies in a polytope with vertices x1 , . . . , xM . Then we minimize subject to

(21.16) and


S xk
0
for k = 1, , M
(21.18)
xk

21.3

Control for performance

21.3.1

Linear time-invariant systems

Consider
x = Ax + Bu
z = Cx + Du
In this case
Z

kz(t)k dt =

(21.19a)
(21.19b)

x(t) C Cx(t) + 2x(t) C Du(t) + u(t)D Du(t) dt .

When C D = 0, this reduces to the usual performance encountered in LQR control, that is,
Z
Z
2
kz(t)k dt =
x(t) Qx(t) + u(t) Ru(t) dt ;
0

here
Q = C C

and

R = DD .

Suppose
u = Kx

(21.20)

The resulting closed loop system is described by

x = (A + BK)x
z = (C + DK)x

(21.21a)
(21.21b)

21.3. CONTROL FOR PERFORMANCE

291

and inequality (21.2) holds for all x if and only if

P (A + BK) + (A + BK) P + (C + DK) (C + DK) 0 ,
that is,
P A + A P + C C + (P B + C D)K + K (B P + D C) + K D DK 0 .
Assuming that D D is invertible, this can be expressed as
0
P A + A P + C C (P B + C D)(D D)1 (B P + D C) + Q

(21.22)

where P = P and

= [K + (D D)1 (B P + D C)] D D[K + (D D)1 (B P + D C)] 0 .

Q
Suppose P is a stabilizing solution to the following LQR Riccati equation:
P A + A P + C C (P B + C D)(D D)1 (B P + D C) = 0 .

(21.23)

K = (D D)1 (B P + D C) .

(21.24)

Then it can be shown that P P where P is any matrix satisfying inequality (21.22) with
a stabilizing K. In this case

This is the LQR control gain matrix. Note that P is a stabilizing solution if AB(D D)1 (B P +
D C) is asymptotically stable.

21.3.2

Control of polytopic systems

In this section, we consider a specific type of uncertain/nonlinear system described by

x = F (t, x, u)
z = H(t, x, u)

(21.25a)
(21.25b)

Specifically, we consider systems which satisfies the following condition.

Condition 2 There are matrices, Aj , Bj , Cj , Dj , j = 1, . . . , N so that for each t, x and u,
F (t, x) = Ax + Bu
H(t, x) = Cx + Du

(21.26)
(21.27)

where the matrix (A B C D) (which can depend on t, x, u) is a convex combination of

(A1 B1 C1 D1 ) , . . . , (AN BN CN DN ).
So, when Condition 3 is satisfied, a system under consideration can be described by
x = A(t, x)x + B(t, x)u
z = C(t, x)x + D(t, x)u

(21.28a)
(21.28b)

where (A(t, x) B(t, x) C(t, x) D(t, x)) is contained in the polytope whose vertices are
(A1 B1 C1 D1 ), . . . , (AN BN CN DN ). Hence, we sometimes refer to a system satisfying
Condition 3 as a polytopic uncertain/nonlinear system.

292

CHAPTER 21. PERFORMANCE

Linear state feedback controllers. We consider

u = Kx

(21.29)

This results in the following closed loop system

x = [A(t, x) + B(t, x)K]x
z = [C(t, x) + D(t, x)K]x

(21.30a)
(21.30b)

Note that this system satisfies Condition 1 with vertex matrices (A1 +B1 K, C1 +D1 K), . . . , (AN +
BN K). Hence
Z

t0

where S = P


kz(t)k2 dt x0 P x0

Aj S + Bj L + SAj + L Bj SCj + L Dj
Cj S + Dj L
I

for

j = 1, . . . , N

(21.31)

and
K = LS 1

21.3.3

(21.32)

Multiple performance outputs

x = F (t, x, u)
zi = Hi (t, x, u) for i = 1, . . . , p

(21.33a)
(21.33b)

Condition 3 There are matrices, Aj , Bj , C1j , D1j . . . , Cpj , Dpj , j = 1, . . . , N so that for each
t, x and u,
F (t, x) = Ax + Bu
Hi (t, x) = Ci x + Di u for i = 1, . . . , p

(21.34)
(21.35)

where the matrix (A B C1 D1 , . . . , Cp Dp ) (which can depend on t, x, u) is a convex combination of (A1 B1 Ci1 Di1 , . . . , Cp1 Dp1 ) , (AN BN CiN DiN , . . . , CpN DpN ) .
Z

t0

kzi (t)k2 dt x0 P x0

for

i = 1, . . . , p

where S = P 1 ,


Aj S + Bj L + SAj + L Bj SCij + L Dij

Cij S + Dij L
I

for

i = 1, . . . , p and j = 1, . . . ,(21.36)
N

and
K = LS 1

(21.37)

Chapter 22
Appendix
22.1

The Euclidean norm

What is the size or magnitude of a vector? Meet norm.

Consider any complex n vector x.
The Euclidean norm or 2-norm of x is the nonegative real number given by
1

||x|| := (|x1 |2 + . . . + |xn |2 ) 2

Note that

||x|| = (
x1 x1 + . . . + xn xn ) 2
1

= (x x) 2

If x is real, these expressions become

1

||x|| = (x21 + . . . + x2n ) 2

1

= (xT x) 2

>>

norm([3; 4])

ans = 5
>>

norm([1; j])

ans = 1.4142

Note that in the last example xT x = 0, but x x = 2.

Properties of || ||. The Euclidean norm has the following properties.
293

294

CHAPTER 22. APPENDIX

(i) For every vector x,

||x|| 0
and
||x|| = 0

iff

x=0

(ii) (Triangle Inequality.) For every pair of fectors x, y,

||x + y|| ||x|| + ||y||
(iii) For every vector x and every scalar .
||x|| = ||||x||
Any real valued function on a vector space which has the above three properties is defined
to be a norm.