Control System Theory and Design: ECE 515 Class Notes

ECE 515 Class Notes
Lecture Notes on
arXiv:2007.01367v1 [math.OC] 2 Jul 2020
CONTROL SYSTEM THEORY

AND DESIGN
Tamer Basar, Sean P. Meyn, and William R. Perkins
University of Illinois at Urbana-Champaign
December, 2010
Preface
This is a collection of the lecture notes of the three authors for a first-
year graduate course on control system theory and design (ECE 515, for-
merly ECE 415) at the ECE Department of the University of Illinois at
Urbana-Champaign. This is a fundamental course on the modern theory
of dynamical systems and their control, and builds on a first-level course
in control that emphasizes frequency-domain methods (such as the course
ECE 486, formerly ECE 386, at UIUC). The emphasis in this graduate
course is on state space techniques, and it encompasses modeling, analysis
(of structural properties of systems, such as stability, controllability, and ob-
servability), synthesis (of observers/compensators and controllers) subject
to design specifications, and optimization. Accordingly, this set of lecture
notes is organized in four parts, with each part dealing with one of the issues
identified above. Concentration is on linear systems, with nonlinear systems
covered only in some specific contexts, such as stability and dynamic opti-
mization. Both continuous-time and discrete-time systems are covered, with
the former, however, in much greater depth than the latter.
The notions of “control” and “feedback”, in precisely the sense they will
be treated and interpreted in this text, pervade our everyday operations,
oftentimes without us being aware of it. Leaving aside the facts that the
human body is a large (and very complex) feedback mechanism, and an
economy without its build-in (and periodically fine-tuned) feedback loops
would instantaneously turn to chaos, the most common examples of control
systems in the average person’s everyday life is the thermostat in one’s living
room, or the cruise control in one’s their automobile. Control systems of a
similar nature can be found in almost any of today’s industries. The most
obvious examples are the aerospace industry, where control is required in
fly by wire positioning of ailerons, or in the optimal choice of trajectories in
space flight. The chemical industry requires good control designs to ensure
safe and accurate production of specific products, and the paper industry
requires accurate control to produce high quality paper. Even in applications
iii
iv
where control has not found any use, this may be a result of inertia within
industry rather than lack of need! For instance, in the manufacture of
semiconductors currently many of the fabrication steps are done without
the use of feedback, and only now are engineers seeing the difficulties that
such open-loop processing causes.
There are only a few fundamental ideas that are required to take this
course, other than a good background in linear algebra and differential equa-
tions. All of these ideas revolve around the concept of feedback, which is sim-
ply the act of using measurements as they become available to control the
system of interest. The main objective of this course is to teach the student
some fundamental principles within a solid conceptual framework, that will
enable her/him to design feedback loops compatible with the information
available on the “states” of the system to be controlled, and by taking into
account considerations such as stability, performance, energy conservation,
and even robustness. A second objective is to familiarize her/him with the
available modern computational, simulation, and general software tools that
facilitate the design of effective feedback loops.
TB, SPM, WRP

Urbana, January 2010
Contents
Preface iii
I System Modeling and Analysis 3
1 State Space Models 1

1.1 An electromechanical system . . . . . . . . . . . . . . . . . . 2
1.2 Linearization about an equilibrium state . . . . . . . . . . . . 3
1.3 Linearization about a trajectory . . . . . . . . . . . . . . . . 5
1.4 A two link inverted pendulum . . . . . . . . . . . . . . . . . . 6
1.5 An electrical circuit . . . . . . . . . . . . . . . . . . . . . . . 8
1.6 Transfer Functions & State Space Models . . . . . . . . . . . 10
1.7 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2 Vector Spaces 23
2.1 Fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.2 Vector Space . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.3 Bases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.4 Change of basis . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.5 Linear Operators . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.6 Linear operators and matrices . . . . . . . . . . . . . . . . . . 31
2.7 Eigenvalues and Eigenvectors . . . . . . . . . . . . . . . . . . 32
2.8 Inner Products . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.9 Orthogonal vectors and reciprocal basis vectors . . . . . . . . 37
2.10 Adjoint transformations . . . . . . . . . . . . . . . . . . . . . 38
2.11 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3 Solutions of State Equations 47

3.1 LTI state space models . . . . . . . . . . . . . . . . . . . . . . 47
3.2 Other descriptions of the state transition matrix . . . . . . . 48
v
vi CONTENTS
3.3 Change of state variables . . . . . . . . . . . . . . . . . . . . 50

3.4 Numerical examples . . . . . . . . . . . . . . . . . . . . . . . 51
3.5 Cayley-Hamilton Theorem . . . . . . . . . . . . . . . . . . . . 52
3.6 Linear Time-Varying Systems . . . . . . . . . . . . . . . . . . 55
3.7 Fundamental matrices . . . . . . . . . . . . . . . . . . . . . . 56
3.8 Peano-Baker Series . . . . . . . . . . . . . . . . . . . . . . . . 57
3.9 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
II System Structural Properties 63
4 Stability 65
4.1 Stability in the sense of Lyapunov . . . . . . . . . . . . . . . 66
4.2 Lyapunov’s direct method . . . . . . . . . . . . . . . . . . . . 71
4.3 Region of asymptotic stability . . . . . . . . . . . . . . . . . . 74
4.4 Stability of linear state space models . . . . . . . . . . . . . . 76
4.5 Stability subspaces . . . . . . . . . . . . . . . . . . . . . . . . 79
4.6 Linearization of nonlinear models and stability . . . . . . . . 80
4.7 Input-output stability . . . . . . . . . . . . . . . . . . . . . . 82
4.8 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
5 Controllability 89
5.1 A preview: The LTI discrete-time case . . . . . . . . . . . . . 92
5.2 The general LTV continuous-time case . . . . . . . . . . . . . 92
5.3 Controllability using the controllability matrix . . . . . . . . 95
5.4 Other tests for controllability . . . . . . . . . . . . . . . . . . 97
5.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
6 Observability, Duality and Minimality 105

6.1 The observability matrix . . . . . . . . . . . . . . . . . . . . . 107
6.2 LTV models and the observability grammian . . . . . . . . . 108
6.3 Duality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
6.4 Kalman canonical forms . . . . . . . . . . . . . . . . . . . . . 111
6.5 State space models and their transfer functions . . . . . . . . 113
6.6 Realization of MIMO transfer functions . . . . . . . . . . . . 114
6.7 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
CONTENTS vii
III Feedback 119
7 Pole Placement 121

7.1 State feedback . . . . . . . . . . . . . . . . . . . . . . . . . . 123
7.2 Observers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
7.3 Observer feedback . . . . . . . . . . . . . . . . . . . . . . . . 127
7.4 Reduced-order (Luenberger) observers . . . . . . . . . . . . . 129
7.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
8 Tracking and Disturbance Rejection 145

8.1 Internal model principle . . . . . . . . . . . . . . . . . . . . . 145
8.2 Transfer function approach . . . . . . . . . . . . . . . . . . . 150
8.3 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
9 Control Design Goals 155

9.1 Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
9.2 Measurements . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
9.3 Robustness and sensitivity . . . . . . . . . . . . . . . . . . . . 157
9.4 Zeros and sensitivity . . . . . . . . . . . . . . . . . . . . . . . 159
9.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
IV Optimal Control 165
10 Dynamic Programming and the HJB Equation 167

10.1 Problem formulation . . . . . . . . . . . . . . . . . . . . . . . 167
10.2 Hamilton-Jacobi-Bellman equations . . . . . . . . . . . . . . . 168
10.3 A solution to the LQR problem . . . . . . . . . . . . . . . . . 172
10.4 The Hamiltonian matrix . . . . . . . . . . . . . . . . . . . . . 174
10.5 Infinite horizon regulator . . . . . . . . . . . . . . . . . . . . 177
10.6 Return difference equation . . . . . . . . . . . . . . . . . . . . 187
10.7 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
11 An Introduction to the Minimum Principle 203

11.1 Minimum Principle and the HJB equations . . . . . . . . . . 204
11.2 Minimum Principle and Lagrange multipliers? . . . . . . . . . 206
11.3 The penalty approach? . . . . . . . . . . . . . . . . . . . . . . 210
11.4 Application to LQR . . . . . . . . . . . . . . . . . . . . . . . 211
11.5 Nonlinear examples . . . . . . . . . . . . . . . . . . . . . . . . 213
11.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221
viii CONTENTS
Index 224
List of Figures
1.1 Magnetically Suspended Ball . . . . . . . . . . . . . . . . . . 2

1.2 Trajectory of a 2D nonlinear state space model . . . . . . . . 4
1.3 The Pendubot . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.4 Coordinate description of the pendubot . . . . . . . . . . . . 7
1.5 Continuum of equilibrium positions for the Pendubot . . . . . 8
1.6 A simple RLC circuit . . . . . . . . . . . . . . . . . . . . . . . 9
1.7 Simulation diagram for the simple RLC circuit . . . . . . . . 9
1.8 Controllable Canonical Form . . . . . . . . . . . . . . . . . . 12
1.9 Observable Canonical Form . . . . . . . . . . . . . . . . . . . 12
1.10 Partial fraction expansion of a transfer function . . . . . . . . 15
4.1 Typical solutions of the linear state space model . . . . . . . 65

4.2 Is this model stable, or unstable? . . . . . . . . . . . . . . . . 67
4.3 If x0 ∈ Bδ (xe ), then x(t) = ϕ(t; x0 ) ∈ B (x0 ) for all t ≥ 0. . . 67
4.4 Frictionless pendulum . . . . . . . . . . . . . . . . . . . . . . 68
4.5 Stable and unstable equilibria. . . . . . . . . . . . . . . . . . 69
4.6 V (x(t)) represents the height of the ‘lifted’ trajectory . . . . 71
4.7 V (x(t)) is decreasing with time if ∇V (x) f (x) ≤ 0 . . . . . . 71
4.8 The vectors ∇V (x) and f (x) meet at ninety degrees . . . . . 74
4.9 The sub-level set of V . . . . . . . . . . . . . . . . . . . . . . 76
4.10 The graph of f remains in a tube of radius M . . . . . . . . . 83
5.1 The controlled model can be steered to any state . . . . . . . 90

5.2 The state can never leave this line, regardless of the control . 91
5.3 The modes of the model are decoupled . . . . . . . . . . . . . 98
6.1 If the input does not affect the internal temperature... . . . . 105
6.2 If the input to this circuit is zero... . . . . . . . . . . . . . . . 106
6.3 The Kalman Controllability Canonical Form . . . . . . . . . . 111
6.4 The Kalman Observability Canonical Form . . . . . . . . . . 113
ix
LIST OF FIGURES 1
6.5 A block diagram realization of P . . . . . . . . . . . . . . . . . 115
7.1 Observer design . . . . . . . . . . . . . . . . . . . . . . . . . . 126

7.2 The separation principle . . . . . . . . . . . . . . . . . . . . . 128
8.1 Observer/state feedback . . . . . . . . . . . . . . . . . . . . . 150
9.1 An error feedback configuration . . . . . . . . . . . . . . . . . 158

9.2 A Nyquist plot for the Pendubot . . . . . . . . . . . . . . . . 161
9.3 Sensitivity function for the Pendubot . . . . . . . . . . . . . . 162
10.1 If a better control existed... . . . . . . . . . . . . . . . . . . . 169

10.2 Solutions to the ARE for a scalar model . . . . . . . . . . . . 179
10.3 The loop transfer function . . . . . . . . . . . . . . . . . . . . 188
10.4 The symmetric root locus . . . . . . . . . . . . . . . . . . . . 189
10.5 The symmetric root locus . . . . . . . . . . . . . . . . . . . . 190
10.6 The loop transfer function . . . . . . . . . . . . . . . . . . . . 191
10.7 The symmetric root locus for the Pendubot . . . . . . . . . . 193
10.8 Nyquist plot for the Pendubot . . . . . . . . . . . . . . . . . . 193
11.1 An optimization problem on R2 with a single constraint . . . 207

11.2 A perturbation of the function z ∈ D[t0 , t1 ]. . . . . . . . . . . 208
11.3 Optimal state trajectory for the bilinear model . . . . . . . . 215
11.4 Optimal trajectories for the minimum time problem . . . . . 220
2 LIST OF FIGURES
Part I
System Modeling and

Analysis
3
Chapter 1
State Space Models
After a first course in control system design one learns that intuition is a
starting point in control design, but that intuition may fail for complex
systems such as those with multiple inputs and outputs, or systems with
nonlinearities. In order to augment our intuition to deal with problems as
complex as high speed flight control, or flow control for high speed commu-
nication networks, one must start with a useful mathematical model of the
system to be controlled. It is not necessary to “take the system apart” -
to model every screw, valve, and axle. In fact, to use the methods to be
developed in this book it is frequently more useful to find a simple model
which gives a reasonably accurate description of system behavior. It may
be difficult to verify a complex model, in which case it will be difficult to
trust that the model accurately describes the system to be controlled. More
crucial in the context of this course is that the control design methodology
developed in this text is model based. Consequently, a complex model of
the system will result in a complex control solution, which is usually highly
undesirable in practice.
Although this text treats nonlinear models in some detail, the most
far reaching approximation made is linearity. It is likely that no physical
system is truly linear. The reasons that control theory is effective in practice
even when this ubiquitous assumption fails are that (i) physical systems can
frequently be approximated by linear models; and (ii) control systems are
designed to be robust with respect to inaccuracies in the system model. It
is not simply luck that physical systems can be modeled with reasonable
accuracy using linear models. Newton’s laws, Kirchoff’s voltage and current
laws, and many other laws of physics give rise to linear models. Moreover,
in this chapter we show that a generalized Taylor series expansion allows
1
2 CHAPTER 1. STATE SPACE MODELS
the approximation of even a grossly nonlinear model by a linear one.

The linear models we consider are primarily of the state space form
ẋ = Ax + Bu
(1.1)
y = Cx + Du
where x is a vector signal in Rn , y and u are the output and input of the
system evolving in Rp and Rm , respectively, and A, B, C, D are matrices
of appropriate dimensions. If these matrices do not depend on the time
variable, t, the corresponding linear system will be referred to as linear
time invariant (LTI), whereas if any one of these matrices has time-varying
entries, then the underlying linear system will be called linear time varying
(LTV). If both the input and the output are scalar, then we refer to the
system as single input-single output (SISO); if either a control or output are
of dimension higher than one, then the system is multi input-multi output
(MIMO).
Our first example now is a simple nonlinear system which we approxi-
mate by a linear state space model of the form (1.1) through linearization.
u(t)
y(t)
Reference
Height r
Figure 1.1: Magnetically Suspended Ball
1.1 An electromechanical system

The magnetically suspended metallic ball illustrated in Figure 1.1 is a simple
example which illustrates some of the important modeling issues addressed
1.2. LINEARIZATION ABOUT AN EQUILIBRIUM STATE 3
in this chapter. The input u is the current applied to the electro-magnet,

and the output y is the distance between the center of the ball and some
reference height. Since positive and negative inputs are indistinguishable at
the output of this system, it follows that this cannot be a linear system.
The upward force due to the current input is approximately proportional to
u2 /y 2 , and hence from Newton’s law for translational motion we have
u2
ma = mÿ = mg − c ,
y2
where g is the gravitational constant and c is some constant depending on
the physical properties of the magnet and ball. This input-output model
can be converted to (nonlinear) state space form using x1 = y and x2 = ẏ:
c u2
ẋ1 = x2 , ẋ2 = g −
m x21
where the latter equation follows from the formula ẋ2 = ÿ. This pair of
equations forms a two-dimensional state space model
ẋ1 = x2 = f1 (x1 , x2 , u) (1.2)

c u2
ẋ2 = g − = f2 (x1 , x2 , u) (1.3)
m x21
It is nonlinear, since f2 is a nonlinear function of xx12 . Letting x = x1

x2 ,
and f = ff12 , the state equations may be written succinctly as

ẋ = f (x, u).
The motion of a typical solution to a nonlinear state space model in R2 is

illustrated in Figure 1.2.
1.2 Linearization about an equilibrium state

Suppose that a fixed value ue , say positive, is applied, and that the state
xe = ( xxe2
e1
) has the property that

f1 (xe1 ,xe2 ,ue )
f (xe1 , xe2 , ue ) = f2 (xe1 ,xe2 ,ue )
= ϑ.
From the definition of f we must have xe2 = 0, and

r
c
xe1 = ue
mg
x(0)
f(x(t),u(t))
x(t)
Figure 1.2: Trajectory of a nonlinear state space model in two dimensions:

ẋ = f (x, u)
which is unique when restricted to be positive. The state xe is called an

equilibrium state since the velocity vector f (x, u) vanishes when x = xe ,
and u = ue . If the signals x1 (t), x2 (t) and u(t) remain close to the fixed
point (xe1 , xe2 , ue ), then we may write
x1 (t) = xe1 + δx1 (t)

x2 (t) = xe2 + δx2 (t)
u(t) = ue + δu(t),
where δx1 (t), δx2 (t), and δu(t) are small-amplitude signals. From the state
equations (1.2) and (1.3) we then have
δ ẋ1 = xe2 + δx2 (t) = δx2 (t)

δ ẋ2 = f2 (xe1 + δx1 , xe2 + δx2 , ue + δu)
Applying a Taylor series expansion to the right hand side (RHS) of the
second equation above gives
∂f2 ∂f2
δ ẋ2 = f2 (xe1 , xe2 , ue ) + δx1 + δx2
∂x1 (xe1 ,xe2 ,ue ) ∂x2 (xe1 ,xe2 ,ue )

∂f2
+ δu + H.O.T.
∂u (xe1 ,xe2 ,ue )

After computing partial derivatives we obtain the formulae
δx˙ 1 = δx2 .
c u2e 2c ue
δ ẋ2 = 2 3 δx1 − δu + H.O.T.
m xe1 m x2e1
1.3. LINEARIZATION ABOUT A TRAJECTORY 5

x (t)
Letting x denote the bivariate signal x(t) = x12 (t) we may write the lin-
earized system in state space form:

0 1 0
δ ẋ = δx + δu
α 0 β
δy = δx1 ,
where
c u2e c ue
α=2 , β = −2 .
m x3e1 m x2e1
This linearized system is only an approximation, but one that is reasonable

and useful in control design as long as the state δx and the control δu remain
small. For example, we will see in Chapter 4 that “local” stability of the
nonlinear system is guaranteed if the simpler linear system is stable.
1.3 Linearization about a trajectory

Consider now a general nonlinear model
ẋ (t) = f (x(t), u(t), t)
where f is a continuously differentiable (C 1 ) function of its arguments. Sup-

pose that for an initial condition x0 given at time t0 (i.e., x(t0 ) = x0 ), and a
given input un (t), the solution of the nonlinear state equation above exists
and is denoted by xn (t), which we will refer to as the nominal trajectory
corresponding to the nominal input un (t), for the given initial state. For ex-
ample, in the control of a satellite orbiting the earth, the nominal trajectory
xn (t) might represent a desired orbit (see Exercise 7 below). We assume
that the input and state approximately follow these nominal trajectories,
and we again write
x(t) = xn (t) + δx(t)

u(t) = un (t) + δu(t)
From a Taylor series expansion we then have

∂f ∂f
ẋ = ẋn + δ ẋ = f (xn , un , t) + n n δx + n n δu + H.O.T.

∂x (x ,u ,t) ∂u (x ,u ,t)
| {z } | {z }
A(t) B(t)
Since we must have ẋn = f (xn , un , t), this gives the state space description
δ ẋ = A(t)δx + B(t)δu,
where the higher order terms have been ignored.
1.4 A two link inverted pendulum

Below is a photograph of the pendubot found at the robotics laboratory at
the University of Illinois, and a sketch indicating its component parts. The
pendubot consists of two rigid aluminum links: link 1 is directly coupled to
the shaft of a DC motor mounted to the end of a table. Link 1 also includes
the bearing housing for the second joint. Two optical encoders provide
position measurements: one is attached at the elbow joint and the other
is attached to the motor. Note that no motor is directly connected to link
2 - this makes vertical control of the system, as shown in the photograph,
extremely difficult!
Link 1
Encoder 1
Motor
Table
Link 1
Encoder 2
Encoder 2
Link 2
Link 2
Motor
and
Encoder 1
Figure 1.3: The Pendubot
Since the pendubot is a two link robot with a single actuator, its dynamic
equations can be derived using the so-called Euler-Lagrange equations found
in numerous robotics textbooks [11]. Referring to the figure, the equations
1.4. A TWO LINK INVERTED PENDULUM 7
of motion are
d11 q̈1 + d12 q̈2 + h1 + φ1 = τ (1.4)

d21 q̈1 + d22 q̈2 + h2 + φ2 = 0 (1.5)
where q1 , q2 are the joint angles and
d11 = m1 `2c1 + m2 (`21 + `2c2 + 2`1 `c2 cos(q2 )) + I1 + I2

d22 = m2 `2c2 + I2
d12 = d21 = m2 (`2c2 + `1 `c2 cos(q2 )) + I2
h1 = −m2 `1 `c2 sin(q2 )q̇22 − 2m2 `1 `c2 sin(q2 )q̇2 q̇1
h2 = m2 `1 `c2 sin(q2 )q̇12
φ1 = (m1 `c1 + m2 `1 )g cos(q1 ) + m2 `c2 g cos(q1 + q2 )
φ2 = m2 `c2 g cos(q1 + q2 )
The definitions of the variables qi , `1 , `ci can be deduced from Figure 1.4.
q
2
l c2
q
1
x
lc1
l1
Figure 1.4: Coordinate description of the pendubot: `1 is the length of the

first link, and `c1 , `c2 are the distances to the center of mass of the respective
links. The variables q1 , q2 are joint angles of the respective links.
This model may be written in state space form as a nonlinear vector

differential equation, ẋ = f (x, u), where x = (q1 , q2 , q̇1 , q̇2 )0 , and f is defined
from the above equations. For a range of different constant torque inputs τ ,
this model admits various equilibria. For example, when τ = 0, the vertical
downward position xe = (−π/2, 0, 0, 0) is an equilibrium, as illustrated in
the right hand side of Figure 1.3. When τ = 0 it follows from the equations
of motion that the upright vertical position xe = (+π/2, 0, 0, 0) is also an
equilibrium. It is clear from the photograph given in the left hand side of
Figure 1.3 that the upright equilibrium is strongly unstable in the sense
that with τ = 0, it is unlikely that the physical system will remain at rest.
Nevertheless, the velocity vector vanishes, f (xe , 0) = 0, so by definition the
upright position is an equilibrium when τ = 0. Although complex, we may
again linearize these equations about the vertical equilibrium. The control
design techniques introduced later in the book will provide the tools for
stabilization of the pendubot in this unstable vertical configuration via an
appropriate controller.
Figure 1.5: There is a continuum of different equilibrium positions for the

Pendubot corresponding to different constant torque inputs τ .
1.5 An electrical circuit

A simple electrical circuit is illustrated in Figure 1.6. Using Kirchoff’s volt-
age and current laws we may obtain a state space model in which the current
through the inductor and the capacitor voltage become state variables.
1.5. AN ELECTRICAL CIRCUIT 9
x1 (t)
+ - + -
C
L
+
u(t)
R
-
x2 (t)
Figure 1.6: A simple RLC circuit

u x2 x1
1 1
∑ ∑
L C
- -
1
RC
Figure 1.7: A simulation diagram for the corresponding state space model
with x1 = the voltage across the capacitor, and x2 = the current through
the inductor.
Kirchoff’s Current Law (KCL) gives
1
x2 = C ẋ1 + x1 ,
R
which may be written as
1 1
ẋ1 = − x1 + x2 (1.6)
RC C
From Kirchoff’s Voltage Law (KVL) we have
x1 + Lẋ2 = +u,
or,
1 1
ẋ2 = − x1 + u. (1.7)
L L
Equations (1.7) and (1.6) thus give the system of state space equations
1
− RC C1

0
ẋ = 1 x+ 1 u
−L 0 L
1.6 Transfer Functions & State Space Models

In the previous examples we began with a physical description of the system
to be controlled, and through physical laws obtained a state space model
of the system. In the first and second examples, a linear model could be
obtained through linearization, while in the last example, a linear model
was directly obtained through the KVL and KCL circuit laws. For a given
LTI system, a state space model which is constructed from a given transfer
function G(s) is called a realization of G. Realization of a given G is not
unique, and in fact there are infinitely many such realizations, all of which
are equivalent however from an input-output point of view. In this section
we show how to obtain some of the most common realizations starting from
a transfer function model of a system.
Consider the third-order model
...
y + a2 ÿ + a1 ẏ + a0 y = b2 ü + b1 u̇ + b0 u (1.8)
where the coefficients {ai , bi } are arbitrary real numbers. The corresponding
transfer function description is
Y (s) B(s) b2 s2 + b1 s + b0
G(s) = = = 3
U (s) A(s) s + a2 s2 + a1 s + a0
where Y and U are the Laplace transforms of the signals y and u, respec-
tively. The numerator term B(s) = b2 s2 + b1 s + b0 can create complexity
in determining a state space model. So, as a starting point, we consider the
“zero-free system” where B(s) ≡ 1, which results in the model
...
w + a2 ẅ + a1 ẇ + a0 w = u
An equivalent simulation diagram description for this system is:

... .. .
u w w w w
∑
- - -
a2
a1
a0
With zero initial conditions one has the relation Y (s) = B(s)W (s), so
that the signal y can be obtained by adding several interconnections to
the above simulation diagram, yielding the simulation diagram depicted in
1.6. TRANSFER FUNCTIONS & STATE SPACE MODELS 11
Figure 1.8. Letting the outputs of the integrators form states for the system
we obtain the state space model
x1 = w ẋ1 = x2
x2 = ẇ ẋ2 = x3
x3 = ẅ ẋ3 = −a2 x3 − a1 x2 − a0 x1 + u
and
y = b0 x1 + b1 x2 + b2 x3
This may be written in matrix form as
   
0 1 0 0
ẋ =  0 0 1 x +  0 u
−a0 −a1 −a2 1
y = [b0 b1 b2 ]x + [0]u.
This final state space model is called the controllable canonical form (CCF)
- one of the most important system descriptions from the point of view of
analysis.
Several alternative descriptions can be obtained by manipulating the
transfer function G(s) before converting to the time domain, or by defining
states in different ways. One possibility is to take the description
(s3 + a2 s2 + a1 s + a0 )Y (s) = (b0 + b1 s + b2 s2 )U (s) (1.9)
and divide throughout by s3 to obtain

a2 a1 a0 b0 b1 b2
1+ + 2 + 3 Y (s) = + + U (s)
s s s s3 s2 s
Rearranging terms then gives
1 1
Y = 3 (b0 U − a0 Y ) + 2 (b1 U − a1 Y )
s s
1
+ (b2 U − a2 Y )
s
We may again describe this equation using a simulation diagram, as given
in Figure 1.9. As before, by letting x1 , x2 , x3 denote the outputs of the
integrators we obtain a state space model which now takes the form
ẋ1 = x2 − a2 x1 + b2 u
ẋ2 = x3 − a1 x1 + b1 u
ẋ3 = −a0 x1 + b0 u
y = x1
b2
b1
b0
u w
∑ ∑
-
- -
y
a2
a1
a0
Figure 1.8: Controllable Canonical Form

a0
a1
a2
u
- - - y
b0 ∑ ∑ ∑
b1
b2
Figure 1.9: Observable Canonical Form

or in matrix form
   
−a2 1 0 b2
ẋ =  −a1 0 1  x +  b1  u
−a0 0 0 b0
y = [1 0 0]x + [ 0 ]u.
This final form is called the observable canonical form (OCF).

In the example above, the degree n0 of the denominator of G(s) is 3,
and the degree m0 of the numerator is 2, so that n0 > m0 . In this case
the model is called strictly proper . In the case where n0 = m0 , the “D”
matrix in (1.1) will be non-zero in any state space realization. To see this,
try adding the term b3 s3 to the right hand side of (1.9), or solve Exercise 10
of this chapter.
Both controllable and observable canonical forms admit natural gen-
eralizations to the n-dimensional case. For a SISO LTI system, let the
input-output transfer function be given by
Y (s) B(s) bn−1 sn−1 + · · · + b1 s + b0
G(s) = = = n
U (s) A(s) s + an−1 sn−1 + · · · + a2 s2 + a1 s + a0
Then, the n-dimensional state space realization in controllable canonical
form is identified by the following A, B, and C matrices:
     T
0 1 0 0 ··· 0 0 b0
 0 0 1 0 ··· 0   ..   b1 
A= . . . ; B= . ; C= .  .
     
. . . . .  .
 . . . .   0   . 
−a0 −a1 −a2 a3 · · · an−1 1 bn−1
The observable canonical form, on the other hand, will have the following
A, B, and C matrices:
     T
−an−1 1 0 0 ··· 0 bn−1 1
 −an−2 0 1 0 ··· 0   ..   0 
A= .. .. .. ; B= . ; C= ..  .
     
..
 . . . .   b1   . 
−a0 0 0 0 ··· 0 b0 0
Another alternative is obtained by applying a partial fraction expansion

to the transfer function G:
n
b(s) X ki
G(s) = =d+ ,
a(s) s − pi
i=1
where {pi : 1 ≤ i ≤ n} are the poles of G, which are simply the roots of
a. A partial expansion of this form is always possible if all of the poles are
distinct. In general, a more complex partial fraction expansion must be em-
ployed. When a simple partial fraction expansion is possible, as above, the
system may be viewed as a parallel network of simple first order simulation
diagrams; see Figure 1.10.
The significance of this form is that it yields a strikingly simple system
description:

ẋ1 = p1 x1 + k1 u 
..

. decoupled dynamic equations.

ẋn = pn xn + kn u

This gives the state space model

   
p1 0 · · · 0 k1
 0 p2 · · · 0   k2 
ẋ =  . x +  .. u
   
 .. ..
.   . 
0 0 · · · pn kn
y = [1, . . . , 1]x + [d]u.
This is often called the modal form, and the states xi (t) are then called
modes. It is important to note that this form is not always possible if the
roots of the denominator polynomial are not distinct. Exercises 12 and 13
below address some generalizations of the modal form.
Matlab Commands
Matlab is not well suited to nonlinear problems. However, the Matlab pro-
gram Simulink can be used for simulation of both linear and nonlinear
models. Some useful Matlab commands for system analysis of linear
systems are
RLOCUS calculates the root locus. e.g. rlocus(num,den), or rlocus(A,B,C,D).
STEP computes the step response of a linear system.
BODE computes the Bode plot.
NYQUIST produces the Nyquist plot.

x1
k1 ∑
p1
x2 y
k2 ∑ ∑
u
p2
x3
k3 ∑
p3
Figure 1.10: Partial fraction expansion of a transfer function
TF2SS gives the CCF state space representation of a transfer function

model, but in a different form than given here.
SS2TF computes the transfer function of a model, given any state space
representation.
RESIDUE may be used to obtain the partial fraction expansion of a trans-

fer function.
Summary and References
State space models of the form
ẋ = f (x, u)
y = g(x, u)
occur naturally in the mathematical description of many physical systems,

where u is the input to the system, y is the output, and x is the state variable.
In such a state space model, each of these signals evolves in Euclidean space.
The functions f and g are often linear, so that the model takes the form
ẋ = Ax + Bu
y = Cx + Du.
If not, the functions f, g may be approximated using a Taylor series expan-

sion which again yields a linear state space model. The reader is referred to
[7] for more details on modeling from a control perspective.
A given linear model will have many different state space representa-
tions. Three methods for constructing a state space model from its transfer
function have been illustrated in this chapter:
(a) Construct a simulation diagram description, and define the outputs

of the integrators as states. This approach was used to obtain the
controllable canonical form, but it is also more generally applicable.
(b) Manipulate the transfer function to obtain a description which is more
easily described in state space form. For instance, a simple division
by sn led to the observable canonical form.
(c) Express the transfer function as a sum of simpler transfer functions –
This approach was used to obtain a modal canonical form.
These three system descriptions, the modal, controllable, and observable

canonical forms, will be applied in control analysis in later chapters.
Much more detail on the synthesis of state space models for linear sys-
tems may be found in Chapter 6 of [6], and Chapter 3 of [5].
1.7. EXERCISES 17
1.7 Exercises
1.7.1 You are given a nonlinear input-output system which satisfies the non-
linear differential equation:
ÿ(t) = 2y − (y 2 + 1)(ẏ + 1) + u.
(a) Obtain a nonlinear state-space representation.

(b) Linearize this system of equations around its equilibrium output
trajectory when u(·) ≡ 0, and write it in state space form.
1.7.2 Repeat Exercise 1 with the new system
ÿ(t) = 2y − (y 2 + 1)(ẏ + 1) + u + 2u̇.
1.7.3 Obtain state equations for the following circuit. For the states, use the
voltage across the capacitor, and the current through the inductor.
i L (t)
vC (t)
+ - + -
C
+ L +
u (t) u2 (t)
1
- R -
1.7.4 In each circuit below,
(a) Obtain a transfer function and a state space realization.

(b) Sketch a frequency response.
(c) Use the step command in Matlab to obtain a step response.
+ +
R
(a) Lead network u(t) R y(t)
- -
+ +
R
(b) Lag network u(t) R y(t)
- C -
C C
+ +
R R
(c) Notch network u(t) y(t)
R/3 3C
- -
1.7.5 Consider the mass-spring system shown below

y(t)
u(t) = f(t)
k2 k1
m1 m2
k1 b
Assume that a force is acting on m1 , and let the horizontal position

of m2 represent the output of this system.
(a) Derive a set of differential equations which describes this input-

output system. To solve this problem you will require Newton’s
law of translational motion, and the following facts: (i) The force
exerted by a spring is proportional to its displacement, and (ii)
the force exerted by a frictional source is proportional to the
relative speed of the source and mass.
(b) Find the transfer function for the system.
(c) Obtain a state space description of the system.
1.7.6 The n-dimensional nonlinear vector differential equation ẋ = f (x) has

1.7. EXERCISES 19
a unique solution from any x ∈ Rn if the function f has continuous

partial derivatives. To see that just continuity of f is not sufficient for
uniqueness, and that some additional conditions are needed, consider
the scalar differential equation
p
ẋ = 1 − x2 , x(0) = 1.
Show that this differential equation with the given initial condition
has at least two solutions: One is x(t) ≡ 1, and another one is x(t) =
cos(t).
1.7.7 Consider a satellite in planar orbit about the earth. The situation is
modeled as a point mass m in an inverse square law force field, as
sketched below. The satellite is capable of thrusting (using gas jets,
for example) with a radial thrust u1 and a tangential (θ direction)
thrust u2 . Recalling that acceleration in polar coordinates has a radial
component (r̈ −rθ̇2 ), and a tangential component (rθ̈ +2ṙθ̇), Newton’s
Law gives
k
m(r̈ − rθ̇2 ) = − + u1
r2
m(rθ̈ + 2ṙθ̇) = u2 ,
where k is a gravitational constant.
(a) Convert these equations to (nonlinear) state space form using x1 =
r, x2 = ṙ, x3 = θ, x4 = θ̇.
(b) Consider a nominal circular trajectory r(t) = r0 ; θ(t) = ω0 t, where
r0 and ω0 are constants. Using u1 (t) = u2 (t) = 0, obtain expres-
sions for the nominal state variables corresponding to the circular
trajectory. How are k, r0 , m, and ω0 related?
(c) Linearize the state equation in (a) about the state trajectory in
(b). Express the equations in matrix form in terms of r0 , ω0 and
m (eliminate k).
r
θ
1.7.8 Using Matlab or Simulink, simulate the nonlinear model for the mag-
netically suspended ball.
(a) Using proportional feedback u = −k1 y + k2 r, can you stabilize
the ball to a given reference height r? Interpret your results by
examining a root locus diagram for the linearized system.
(b) Can you think of a better control law? Experiment with other
designs.
Transfer Functions & State Space Models

1.7.9 A SISO LTI system is described by the transfer function
s+4
G(s) =
(s + 1)(s + 2)(s + 3)
(a) Obtain a state space representation in the controllable canonical
form;
(b) Now obtain one in the observable canonical form;
(c) Use partial fractions to obtain a representation with a diagonal
state matrix A (modal form).
In each of (a)–(c) draw an all-integrator simulation diagram.
1.7.10 A SISO LTI system is described by the transfer function
s3 + 2
G(s) =
(s + 1)(s + 3)(s + 4)
(a) Obtain a state space representation in the controllable canonical
form;
(b) Now obtain one in the observable canonical form;
(c) Use partial fractions to obtain a representation of this model with
a diagonal state matrix A.
In each of (a)–(c) draw an all-integrator simulation diagram.
1.7.11 For the multiple input-multiple output (MIMO) system described by
the pair of differential equations
...
y 1 + 2ẏ1 + 3y2 = u1 + ü1 + u̇2
ÿ2 − 3ẏ2 + ẏ1 + y2 + y1 = u2 + u̇3 − u3
obtain a state space realization by choosing y1 and y2 as state variables.
Draw the corresponding simulation diagram.
1.7. EXERCISES 21
1.7.12 This exercise generalizes modal form to the case where some eigenval-
ues are repeated. For each of the following transfer functions, obtain a
state-space realization for the corresponding LTI system by breaking
H into simple additive terms, and drawing the corresponding simula-
tion diagrams for the sum. Choose the outputs of the integrators as
state variables.
2s2
(a) H(s) =
s3 − s2 + s − 1
s2 + s + 1
(b) H(s) = 3
s + 4s2 + 5s + 2
1.7.13 This exercise indicates that a useful generalization of the modal form
may be constructed when some eigenvalues are complex.
(a) Use partial fractions to obtain a diagonal state space representa-
tion for a SISO LTI system with transfer function
s+6
G(s) = .
s2+ 2s + 2
Note that complex gains appear in the corresponding all-integrator
diagram.
(b) Given a transfer function in the form
s+β
G(s) =
(s − λ1 )(s − λ2 )
and a corresponding state space realization with A matrix Com-
pute the eigenvalues λ1 , λ2 of the matrix

σ ω
A =
−ω σ
where σ ≥ 0, ω > 0, find the relationships between λ1 , λ2 , β
and σ, ω. In view of this, complete the state space realization
by obtaining the B and C matrices, and draw the corresponding
simulation diagram.
s2 + s + 1
(c) For the transfer function H(s) = 3 , obtain a state-
s + 4s2 + 5s + 2
space realization for the corresponding LTI system by breaking
H into simple additive terms, and drawing the corresponding
simulation diagrams for the sum.
(d) Apply your answer in (b) to obtain a state space realization of
G(s) in (a) with only real coefficients.
Chapter 2
Vector Spaces
Vectors and matrices, and the spaces where they belong, are fundamental to
the analysis and synthesis of multivariable control systems. The importance
of the theory of vector spaces in fact goes well beyond the subject of vectors
in finite-dimensional spaces, such as Rn . Input and output signals may be
viewed as vectors lying in infinite dimensional function spaces. A system is
then a mapping from one such vector to another, much like a matrix maps
one vector in a Euclidean space to another. Although abstract, this point of
view greatly simplifies the analysis of state space models and the synthesis
of control laws, and is the basis of much of current optimal control theory. In
this chapter we review the theory of vector spaces and matrices, and extend
this theory to the infinite-dimensional setting.
2.1 Fields
A field is any set of elements for which the operations of addition, subtrac-
tion, multiplication, and division are defined. It is also assumed that the
following axioms hold for any α, β, γ ∈ F
(a) α + β ∈ F and α · β ∈ F.
(b) Addition and multiplication are commutative:
α + β = β + α, α · β = β · α.
(c) Addition and multiplication are associative:
(α + β) + γ = α + (β + γ), (α · β) · γ = α · (β · γ).
23
24 CHAPTER 2. VECTOR SPACES
(d) Multiplication is distributive with respect to addition:
(α + β) · γ = α · γ + β · γ
(e) There exists a unique null element 0 such that 0 · α = 0 and 0 + α = α.

(f) There exists a unique identity element 1 such that 1 · α = α.
(g) For every α ∈ F there exists a unique element β ∈ F such that α + β =
0; this unique element is sometimes referred to as the additive inverse
or negative of α, and is denoted as −α.
(h) To every α ∈ F which is not the element 0 (i.e., α 6= 0), there corre-
sponds an element γ such that α · γ = 1; this element is referred to as
the multiplicative inverse of α, and is sometimes written as α−1 .
Fields are a generalization of R, the set of all real numbers. The next
example is the set of all complex numbers, denoted C. These are the only
examples of fields that will be used in the text, although we will identify
others in the exercises at the end of the chapter.
2.2 Vector Space

A set of vectors X is a set on which vector addition, and scalar multiplication
are defined, generalizing the concept of the vector space Rn . In this abstract
setting a vector space over a field F, denoted (X , F), is defined as follows:
(a) For every x1 , x2 ∈ X , the vector sum x1 + x2 ∈ X .

(b) Addition is commutative: For every x1 , x2 ∈ X , the sum x1 + x2 =
x2 + x1 .
(c) Addition is associative: For every x1 , x2 , x3 ∈ X ,
(x1 + x2 ) + x3 = x1 + (x2 + x3 )
(d) The set X contains a vector ϑ such that ϑ + x = x for all x ∈ X .

(e) For every x ∈ X , there is a vector y ∈ X such that x + y = ϑ.
(f) For every x ∈ X , α ∈ F, the scalar product α · x ∈ X .
(g) Scalar multiplication is associative: for every α, β ∈ F, and x ∈ X ,
α(βx) = (αβ)x.
2.3. BASES 25
Below is a list of some of the vector spaces which are most important in
applications
(Rn , R) – the real vector space of n dimensional real-valued vectors.
(Cn , C) – the complex vector space of n dimensional complex-valued vec-

tors.
(C n [a, b], R) – the vector space of real-valued continuous functions on the

interval [a, b], taking values in Rn .
(Dn [a, b], R) – the vector space of real-valued piecewise-continuous func-

tions on the interval [a, b], taking values in Rn .
(Lnp [a, b], R) – the vector space of functions on the interval [a, b], taking
values in Rn , which satisfy the bound
Z b
|f (t)|p dt < ∞, f ∈ Lp [a, b].
a
b(s)
(R(C), R) – the vector space of rational functions a(s) of a complex variable
s, with real coefficients.
A subspace Y of a vector space X is a subset of X which is itself a

vector space with respect to the operations of vector addition and scalar
multiplication. For example, the set of complex n-dimensional vectors whose
first component is zero is a subspace of (Cn , C), but Rn is not a subspace of
(Cn , C).
2.3 Bases
A set of vectors S = (x1 , . . . , xn ) in (X , F) is said to be linearly independent
if the following equality
α1 x1 + α2 x2 + · · · + αn xn = 0
holds for a set of n elements {αi : 1 ≤ i ≤ n} ⊂ F, then α1 = α2 = · · · αn =

0. If the set S contains an infinite number of vectors, then we call S linearly
independent if every finite subset of S is linearly independent, as defined
above.
In the case where (X , F) = (Rn , R), we will have linear independence of

{x1 , . . . , xn } if and only if
 1
xn1

x1
det[x1 x2 · · · xn ] = det  ... · · · ...  6= 0.
 
x1n xnn
The maximum number of linearly independent vectors in (X , F) is called

the dimension of (X , F). For example, (Rn , R) and (Cn , C) both have di-
mension n. What is the dimension of (Cn , R)? (see Exercise 6).
More interesting examples can be found in function spaces. If for exam-
ple (X , F) = (C[0, 1], R), where C[0, 1] is the set of real-valued continuous
functions on [0, 1], then we can easily find a set S of infinite size which is
linearly independent. One such set is the collection of simple polynomials
S = {t, t2 , t3 , . . .}. To see that S is linearly independent, note that for any
n,
n
X
αi ti = ϑ only if αi = 0, 1 ≤ i ≤ n,
i=1
where ϑ ∈ C[0, 1] is the function which is identically zero on [0, 1]. We have
thus shown that the dimension of (C[0, 1], R) is infinite.
A set of linearly independent vectors S = {e1 , . . . , en } in (X , F) is said
to be a basis of X if every vector in X can be expressed as a unique linear
combination of these vectors. That is, for any x ∈ X , one can find {βi , 1 ≤
i ≤ n} such that
x = β1 e1 + β2 e2 + · · · + βn en .
Because the set S is linearly independent, one can show that for any vector
x, the scalars {βi } are uniquely specified in F. The n-tuple {β1 , . . . , βn } is
often called the representation of x with respect to the basis {e1 , . . . , en }.
We typically denote a vector x ∈ Rn by
 
x1
x =  ...  .
 
xn
There are two interpretations of this equation:
(a) x is a vector (in Rn ), independent of basis.

2.4. CHANGE OF BASIS 27
(b) x is a representation of a vector with respect to the natural basis:

     
1 0 0
 0   1   0 
x = x1  ...  + x2  0  + · · · + xn  ...  .
     
     
   ..   
 0   .   0 
0 0 1
The following theorem is easily proven in Rn using matrix manipulations,

and the general proof is similar:
Theorem 2.1. In any n-dimensional vector space, any set of n linearly
independent vectors qualifies as a basis. t
u
In the case of Euclidean space (X , F) = (Rn , R), with {e1 , . . . , en } a
given basis, any vector x ∈ Rn may be expressed as
x = β1 e1 + · · · + βn en ,
where {βi } are all real scalars. This expression may be equivalently written
as x = Eβ, where
     
x1 e11 · · · e1n β1
 ..   .. .
..  , β =  ... 
x =  . , E =  . ,
 
xn en1 · · · enn βn
which through inversion shows that β is uniquely given by
β = E −1 x (2.1)
Here E −1 stands for the matrix inverse of E (i.e., E −1 E = EE −1 = I, where

I is the identity matrix), which exists since ei ’s are linearly independent.
Consider the numerical example with e1 = ( 11 ), e2 = ( 01 ), and x = ( 25 ).
Then we have x = β1 e1 + β2 e2 , with
−1
β1 1 0 2 1 0 2 2
= = =
β2 1 1 5 −1 1 5 3
2.4 Change of basis

Suppose now we are given two sets of basis vectors:
{e1 , · · · , en } {ē1 , · · · , ēn }

A vector x in X can be represented in two possible ways, depending on

which basis is chosen:
n
X
x = β1 e1 + · · · βn en = βk ek (2.2)
k=1
or,
n
X
x = β̄1 ē1 + · · · β̄n ēn = β̄k ēk . (2.3)
k=1
Since {ei } ⊂ X , there exist scalars {pki : 1 ≤ k, i ≤ n} such that for any i,
n
X
ei = p1i ē1 + · · · + pni ēn = pki ēk .
k=1
From (2.2) it then follows that a vector x may be represented as

n n n n
! !
X X X X
x= βi pki ēk = pki βi ēk . (2.4)
i=1 k=1 k=1 i=1
In view of (2.3) and (2.4) we have by subtraction

n
" n #
X X
pki βi − β̄k ēk = ϑ.
k=1 i=1
By linear independence of {ēk }, this implies that each coefficient in brackets

is 0. This gives a matrix relation between the coefficients {βi } and {β̄i }:
n
X
β̄k = pki βi , k = 1, . . . , n
i=1
or using compact matrix notation,

    
β̄1 p11 . . . p1n β1
 ..   ..   .. 
β̄ =  .  =  .   .  = P β.
β̄n pn1 . . . pnn βn
The transformation P maps F n → F n , and is one to one, and onto. It

therefore has an inverse P −1 , so that β can also be computed through β̄:
β = P −1 β̄.
2.5. LINEAR OPERATORS 29
For the special case where (X , F) = (Rn , R), the vectors {ei } can be
stacked to form a matrix to obtain as in (2.1),
x = Eβ = Ē β̄.
Hence the transformation P can be computed explicitly: β̄ = Ē −1 Eβ, so

that P = Ē −1 E. The inverse Ē −1 again exists by linear independence.
2.5 Linear Operators

A linear operator A is simply a function from one vector space (X , F) to
another (Y, F), which is linear . This means that for any scalars α1 , α2 , and
any vectors x1 , x2 ,
A(α1 x1 + α2 x2 ) = α1 A(x1 ) + α2 A(x2 ).
For a linear operator A or a general function from a set X into a set Y we

adopt the terminology
X : Domain of the mapping A

Y : Co-Domain of the mapping A
When A is applied to every x ∈ X , the resulting set of vectors in Y is

called the range (or image) of A, and is denoted by R(A):
[
R(A) := A(x)
x∈X
Pictorially, these notions are illustrated as follows:
R(A)
A(x)
X x Y
The rank of A is defined to be the dimension of R(A).

In the special case of a linear operator A : Rn → Rm defined as A(x) =

Ax for an m × n matrix A,
 
x1
y = Ax = [a1 . . . an ]  ... 
 
xn
1 n
= x1 a + · · · + xn a ,
the range of A is the set of all possible linear combinations of the columns
{ai } of A. That is, the space spanned by the columns of A. The dimension
of R(A) is then the maximum number of linearly independent columns of
A.
Theorem 2.2. For a linear operator A, the set R(A) is a subspace of Y.
Proof To prove the theorem it is enough to check closure under addition
and scalar multiplication. Suppose that y 1 , y 2 ∈ R(A), and that α1 , α2 ∈ F.
Then by definition of R(A) there are vectors x1 , x2 such that
A(x1 ) = y 1 , A(x2 ) = y 2 ,
and then by linearity of the mapping A,
A(α1 x1 + α2 x2 ) = α1 A(x1 ) + α2 A(x2 ) = α1 y 1 + α2 y 2 .
Hence, α1 y 1 + α2 y 2 ∈ R(A), which establishes the desired closure property.

t
u
By assumption, the set Y contains a zero vector, which could be mapped
from numerous x ∈ X . The set of all such x ∈ X is called the nullspace of
A, denoted N (A):
N (A) := {x ∈ X such that A(x) = 0} .
Theorem 2.3. For any linear operator A : X → Y, the nullspace N (A) is

a subspace of X .
Proof Again we check that N (A) is closed under addition and scalar
multiplication. Suppose that x1 , x2 ∈ N (A), and that α1 , α2 ∈ F. Then by
definition,
A(x1 ) = A(x2 ) = ϑ.
Again by linearity of A it is clear that A(α1 x1 + α2 x2 ) = ϑ, which proves

the theorem. t
u
2.6. LINEAR OPERATORS AND MATRICES 31
2.6 Linear operators and matrices

Suppose that V and W are two finite-dimensional vector spaces over the field
F, with bases {v 1 , . . . , v n } and {w1 , . . . , wm } respectively. If A : V → W,
then A may be represented by a matrix. To see this, take any v ∈ V, and
write
X n
v = αi v i
i=1
for scalars {αi } ⊂ F. By linearity we then have,

Xn X n
A(v) = A αi v i = αi A(vi )
i=1 i=1
But for any i we have that A(v i ) ∈ W, which implies that for some scalars
{aji },
m
X
i
A(v ) = aji wj , 1 ≤ i ≤ n.
j=1
From the form of v we must therefore have

X n m
X
A(v) = αi aji wj
i=1 j=1
n
X X m
= aji αi wj (2.5)
j=1 i=1
Recall that the vector w = A(v) in W has a unique representation

m
X
A(v) = βj wj
j=1
Consequently, the terms in parentheses in (2.5) are identical to the {βj }:

n
X
βj = aji αi j = 1, . . . , m.
i=1
From this we see how the representations of v and w are transformed through
the linear operator A:
    
β1 a11 . . . ain α1
β =  ...  =  ... ..   ..  = Aα
  
.  . 
βm ami . . . amn αn
so that A is the representation of A with respect to {v i } and {wi }.

The special case where A is a mapping of (X , F) into itself is of partic-
ular interest. A question that frequently arises is “if the linear operator is
represented by a matrix A of elements in F, how is A affected by a change
of basis?”. Let xb = A(xa ), and write
n
X n
X
xa = αi ei = ᾱi ēi
i=1 i=1
m
X m
X
xb = βi ei = β̄i ēi
i=1 i=1
where the α and β are related by
β = Aα β̄ = Āᾱ.
To see how A and Ā are related, recall that there is a matrix P such that
ᾱ = P α β̄ = P β
Combining these four equations gives
P Aα = P β = β̄ = Āᾱ = ĀP α.
Since α ∈ X is arbitrary, we conclude that P A = ĀP , and hence we can

also conclude that
Ā = P AP −1
A = P −1 ĀP
When these relationships hold, we say that the matrices A and Ā are similar .
2.7 Eigenvalues and Eigenvectors

For a linear operator A : X → X on an arbitrary vector space (X , F), a
scalar λ is called an eigenvalue of A if there exists a non-zero vector x for
which
A(x) = λx. (2.6)
The vector x in (2.6) is then called an eigenvector .
Let X = Cn , F = C, and A be a matrix representation for A. If an
eigenvalue λ of A does exist, then one may infer from the equation
(A − λI)x = 0,
2.7. EIGENVALUES AND EIGENVECTORS 33
that the matrix A − λI is singular. For nontrivial solutions, we must then

have
∆(λ) := det(λI − A) = 0. (2.7)
The function ∆( · ) is called the characteristic polynomial of the matrix A,
and (2.7) is known as the characteristic equation. The characteristic poly-
nomial is a polynomial of degree n, which must therefore have n roots. Any
root of ∆ is an eigenvalue, so at least in the case of operators on (Cn , C),
eigenvalues always exist. Note that if Ā is some other matrix representation
for A, since A and Ā are necessarily similar, Ā has the same characteris-
tic polynomial as A. Hence, the eigenvalues do not depend on the specific
representation picked.
If the roots of the characteristic polynomial are distinct, then the vector
space Cn admits a basis consisting entirely of eigenvectors:
Theorem 2.4. Suppose that λ1 , . . . , λn are the distinct eigenvalues of the

n × n matrix A, and let v 1 , . . . , v n be the associated eigenvectors. Then the
set {v i , i = 1 . . . n} is linearly independent over C.
When the eigenvalues of A are distinct, the modal matrix defined as
M := [v 1 . . . v n ]
is nonsingular. It satisfies the equation AM = M Λ, where

 
λ1 0
Λ=
 .. 
. 
0 λn
and therefore A is similar to the diagonal matrix Λ:
Λ = M −1 AM. (2.8)
Unfortunately, not every matrix can be diagonalized, and hence the eigen-
vectors do not span Cn in general. Consider the matrix
 
1 1 2
A =  0 1 3 
0 0 2
The characteristic polynomial for A is
∆(λ) = det(λI − A) = (λ − 1)(λ − 1)(λ − 2)

So, eigenvalues are λ1 = 1, λ2 = 1, λ3 = 2.

Considering the eigenvector equation
 
0 1 2
(A − λ1 I)x =  0 0 3  x = ϑ,
0 0 1
 
1
we see that x =  0  and its constant multiples are the only eigenvec-
0
tors associated with λ1 . It then follows that there does not exist a state
transformation P for which
 
1 0 0
Λ =  0 1 0  = P AP −1 .
0 0 2
To obtain a nearly diagonal representation of A we use generalized eigen-

vectors, defined as follows. Search for a solution to
(A − λ1 I)k x = 0
such that (A − λ1 I)k−1 x 6= 0. Then x is called a generalized eigenvector of

grade k.
Note that it then follows that the vector (A − λ1 I)x is a generalized
eigenvector of grade k − 1. In the example,
  
0 1 2 0 1 2
(A − λ1 I)2 x =  0 0 3   0 0 3  x
0 0 1 0 0 1
 
0 0 5
=  0 0 3  x.
0 0 1
 
0
So x =  1  is a generalized eigenvector of grade k.
0
Letting y = (A − λ1 I)x, we have
(A − λ1 I)y = (A − λ1 I)2 x = 0.
So y is an eigenvector to A, and x, y are linearly independent. In fact, y =

(1, 0, 0)0 is the eigenvector computed earlier. To obtain an approximately
2.8. INNER PRODUCTS 35
diagonal form, let x1 = y, x2 = x, x3 any eigenvector corresponding to

λ3 = 2. We then have
Ax1 = λ1 x1
Ax2 = Ax = y + λ1 x = x1 + λ1 x2
Ax3 = λ3 x3
Letting M = [x1 |x2 |x3 ] it follows that

        
λ1 1 0 λ1
1
0
AM = M  0 λ1 0  = M  0  M  λ1  M  0 

0 0 λ3 0 0 λ3
= MJ
where
 
λ1 1 0
J =  0 λ1 0  = M −1 AM
0 0 λ3
This representation of A with respect to a basis of generalized eigenvectors

is known as the Jordan form.
2.8 Inner Products

Inner products are frequently applied in the solution of optimization prob-
lems because they give a natural measure of distance between vectors. This
abstract notion of distance can often be interpreted as a cost in applications
to finance, or as energy in mechanical systems. Inner products also provide
a way of defining angles between vectors, and thereby introduce geometry to
even infinite dimensional models where any geometric structure is far from
obvious at first glance.
To define an inner product we restrict our attention to a vector space X
over the complex field C. An inner product is then a complex-valued function
of two vectors, denoted h · , · i, such that the following three properties hold:
(a) hx, yi = hy, xi (complex conjugate).

(b) hx, α1 y 1 +α2 y 2 i = α1 hx, y 1 i+α2 hx, y 2 i, for all x, y 1 , y 2 ∈ X , α1 , α2 ∈
C.
(c) hx, xi ≥ 0 for all x ∈ X , and hx, xi = 0 if and only if x = 0.
In the special case where X = Cn we typically define
hx, yi = x∗ y
where x∗ denotes the complex conjugate transpose of the vector x. Another

important example is the function space Lp [a, b] with p = 2. It can be shown
that the formula
Z b
hf, gi := f (t)g(t) dt, f, g ∈ L2 [a, b],
a
defines on inner product on L2 [a, b].

The most obvious application of the inner product is in the formulation
of a norm. In general, the norm of a vector x in a vector space (X , C),
denoted kxk, is a real-valued function of a vector x such that
1. kxk ≥ 0, and kxk = 0 if and only if x = 0.
2. kαxk = |α| kxk, for any α ∈ C.
3. kx + yk ≤ kxk + kyk.
The third defining property is known as the triangle inequality.

In Rn we usually define the norm as the usual Euclidean norm,
v
u n
√ uX
T
kxk = x x = t x2i ,
i=1
which we will henceforth write as |x|, and reserve the notation k · k for norm
of an infinite-dimensional vector. This Euclidean norm can also be defined
using the inner product: p
|x| = hx, xi (2.9)
In fact, one can show that the expression (2.9) defines a norm in an arbitrary
(finite- or infinite-dimensional) inner product space.
We define the norm of a vector f ∈ Lp [a, b] as
Z b 1/p
kf kLp := |f (t)|p dt .
a
In the case p = 2, this norm is derived from the inner product on L2 [a, b],
but for general p this norm is not consistent with any inner product.
2.9. ORTHOGONAL VECTORS AND RECIPROCAL BASIS VECTORS37
2.9 Orthogonal vectors and reciprocal basis vec-

tors
Two vectors x, y in an inner product space (X , C) are said to be orthogonal
if hx, yi = 0. This concept has many applications in optimization, and
orthogonality is also valuable in computing representations of vectors. To
see the latter point, write
n
X
x= αi v i ,
i=1
where {v i , i = 1 . . . n} is a basis for (X , C). By orthogonality we then have

n
X n
X
hv j , xi = hv j , αi v i i = αi hv j , v i i j = 1...n
i=1 i=1
This may be written explicitly as
hv 1 , xi = α1 hv 1 , v 1 i + α2 hv 1 , v 2 i + · · · + αn hv 1 , v n i
..
.
hv n , xi = α1 hv n , v 1 i + α2 hv 1 , v 2 i + · · · + αn hv n , v n i
or in matrix form
 1  1 1
hv , v i · · · hv 1 , v n i
  
hv , xi α1
.. ..   .. 
 = 
  
 . .  . 
hv n , xi hv n , v 1 i · · · hv n , v n i αn
| {z }
G
The n × n matrix G is called the Grammian. Its inverse gives a formula for
the representation α:
 1 
hv , xi
α = G−1 
 .. 
. 
hv n , xi
If the basis is orthogonal, then G is diagonal, in which case the computation
of the inverse G−1 is straightforward.
A basis is said to be orthonormal if

j i 1, i = j
hv , v i = δij =
0, i 6= j
In this case G = I (the identity matrix), so that G−1 = G.

A basis {ri } is said to be reciprocal to (or dual to) the basis {v i } if
hri , v j i = δij , i = 1, . . . , n, j = 1, . . . , n. (2.10)
If a dual basis is available, then again the representation of x with respect

to the basis {v i } is easily computed. For suppose that x is represented by
α:
n
X
x= αi v i
i=1
Then, by the dual property and linearity we have

n
X n
X
hrj , xi = hrj , αi v i i = αi hrj , v i i.
i=1 i=1
Pn
Since hrj , v i i = δij , this shows that x = i i
i=1 hr , xiv . Of course, to to
apply this formula we must have the reciprocal basis {ri }, which may be as
difficult to find as the inverse Grammian.
In the vector space (Cn , C) define the matrices
 1∗ 
r
1 n  .. 
M = [v · · · v ] R= . 
rn∗
From the defining property (2.10) of the dual basis, we must have RM = I,
so that R = M −1 .
2.10 Adjoint transformations

Suppose that A is a linear operator from the vector space (X , C) to another
vector space (Y, C), that is, A : X → Y. Then the adjoint is a linear
operator working in the reverse direction:
A∗ : Y → X .
Its definition is subtle since it is not directly defined through A. We say

that A∗ is the adjoint of A, if for any x ∈ X and any y ∈ Y,
hA(x), yi = hx, A∗ (y)i.

2.10. ADJOINT TRANSFORMATIONS 39
To illustrate this concept, let us begin with the finite dimensional case
X = Cn , Y = Cm
and suppose that A is defined through an m × n matrix A, so that A(x) =

Ax, x ∈ X . We may then compute the adjoint using the definition of the
inner products on X and Y as follows
hA(x), yi = (Ax)∗ y = x̄T ĀT y = hx, ĀT yi
Thus, the adjoint of A is defined through the complex conjugate transpose

of A:
A∗ (y) = ĀT y = A∗ y
Matlab Commands
Matlab is virtually designed to deal with the numerical aspects of the vector
space concepts described in this chapter. Some useful commands are
INV to compute the inverse of a matrix.
DET to compute the determinant.
EIG finds eigenvalues and eigenvectors.
RANK computes the rank of a matrix.

In this chapter we have provided a brief background on several different

topics in linear algebra, including
(a) Fields and vector spaces.

(b) Linear independence and bases.
(c) Representations of vectors, and how these representations change under
a change of basis.
(d) Linear operators and matrices.
(e) Inner products and norms.
(f) Adjoint operators.
(g) Eigenvectors.
Good surveys on linear algebra and matrix theory may be found in Chapter 2
of [6], or Chapters 4 and 5 of [5].
2.11. EXERCISES 41
2.11 Exercises
2.11.1 Determine conclusively which of the following are fields:
(a) The set of integers.
(b) The set of rational numbers.
(c) The set of polynomials of degree less than 3 with real coefficients.
(d) The set of all n × n nonsingular matrices.
(e) The set {0, 1} with addition being binary “exclusive-or” and mul-
tiplication being binary “and”.
2.11.2 Define rules of addition and multiplication such that the set consist-
ing of three elements {a, b, c} forms a field. Be sure to define the zero
and unit elements.
2.11.3 Let (X, F) be a vector space, and Y ⊂ X a subset of X. If Y satisfies
the closure property y1 , y2 ∈ Y, α1 , α2 ∈ F =⇒ α1 y1 + α2 y2 ∈ Y , show
carefully using the definitions that (Y, F) is a subspace of (X, F).
Linear independence and bases

2.11.4 Which of the following sets are linearly independent?
     
1 0 0
(a) 4 , 2 , 0 in (R3 , R).
2 3 1
       
1 4 7 7912
(b) 2 , 5 , 8 , −314 in (R3 , R).
3 6 9 0.098
     
1 0 j
(c) 4j , −2 , 0 in (C3 , C).
    
2 j 0
(d) sin(t), cos(t), t in (C[0, 1], R) - the set of real-valued continuous
functions on [0, 1] over the real field.
(e) ejt , sin(t), cos(t), t in (C[0, 1], C) - the set of complex-valued con-
tinuous functions on [0, 1] over the complex field.
2.11.5 Determine which of the following sets of vectors are linearly indepen-
dent in R3 by computing the determinant of an appropriate matrix.
     
1 2 0
(a) 0 , 0 , 5.
2 1 1
     
4 1 2
(b) 5 ,  2  , 1.
1 −1 3
2.11.6 For the vector space (Cn , R),
(a) Verify that this is indeed a vector space.

(b) What is its dimension?
(c) Find a basis for (Cn , R).
2.11.7 Let R2×2 be the set of all 2 × 2 real matrices.
(a) Briefly verify that R2×2 is a vector space under the usual matrix
addition and scalar multiplication.
(b) What is the dimension of R2×2 ?
(c) Find a basis for R2×2 .
2.11.8 Is the set I, A, A2 linearly dependent or independent in (R2×2 , R),

1 1
with A = ?
0 2
Representations of vectors
      
 1 2 1 
2.11.9 Given the basis v 1 = 1 , v 2 = 0 , v 3 = 0 and the vector
1 0 1
 
 
3
x = 2 = α1 v 1 + α2 v 2 + α3 v 3
1
(a) Compute the reciprocal basis.

(b) Compute the Grammian.
(c) Compute the representation of x with respect to {v i } using your
answer to (a).
(d) Compute the representation of x with respect to {v i } using your
answer to (b).
Linear operators and matrices

2.11.10 Compute the null space, range space, and rank of the following
matrices.
2.11. EXERCISES 43
 
1 1 2
(a) A = 0 2 2 .
0 3 3
 
1 3 2 1
(b) A =  2 0 1 −1 .
−1 1 0 1
2.11.11 Let b ∈ Rn and A ∈ Rn×m . Give necessary and sufficient conditions

on A and b in order that the linear system of equations Ax = b has a
solution x ∈ Rm .
2.11.12 Let A : R3 → R3 be a linear operator. Consider the two sets
B = {b1 , b2 , b3 } and C = {c1 , c2 , c3 } below
           
 1 0 0   1 0 1 
B = 0 , 1 , 0 , C = 1 , 1 , 0
0 0 1 0 1 1
   
It should be clear that these are bases for R3 .
(a) Find the transformation P relating the two bases.

(b) Suppose the linear operator A maps
     
2 0 0
A(b1 ) = −1 , A(b1 ) = 0 , A(b1 ) = 4
0 0 2
Write down the matrix representation of A with respect to the

basis B and also with respect to the basis C.
2.11.13 Find the inverse of the matrix A, where B is a matrix.

I B
A=
0 I
2.11.14 Consider the set Pn of all polynomials of degree strictly less than
n, with real coefficients, where x ∈ Pn may be written x = a0 + a1 t +
· · · + an−1 tn−1 .
(a) Verify that Pn is a vector space, with the usual definitions of poly-
nomial addition, and scalar multiplication.
(b) Explain why {1, t, . . . , tn−1 } is a basis, and thus why Pn is n-
dimensional.
(c) Suppose x = 10 − 2t + 2t2 − 3t3 . Find the representation α of x

with respect to the basis in (b) for n = 4.
d
(d) Consider differentiation, dt , as an operator A : Pn → Pn−1 . That
d
is, A(x) = dt x. Show that A is a linear operator, and compute
its null space and range space.
(e) Find A, the matrix representation of A for n = 4, using the basis
in (b). Use your A and α to compute the derivative of x in (c).
Inner products and norms

2.11.15 Let (V, C) be an inner product space.
(a) Let x, y ∈ V with x orthogonal to y. Prove the Pythagorean

theorem:
kx + yk2 = kxk2 + kyk2
(b) Prove that in an inner product space,
kx + yk2 + kx − yk2 = 2kxk2 + 2kyk2
where k·k is the norm induced by the inner product. This is called
the Parallelogram law. Can you give a geometric interpretation
of this law in R2 ?
Adjoint operators
2.11.16 If A : X → Y where X and Y are inner product spaces, the adjoint
A∗ is a mapping A∗ : Y → X . Hence, the composition Z = A∗ ◦ A is
a mapping from X to itself. Prove that N (Z) = N (A).
2.11.17 ForR 1 ≤ p < ∞, let Lp denote functions f : (−∞, ∞) → C such
∞
that −∞ |f (s)|p ds < ∞. For p = ∞, Lp denotes bounded functions
f : (−∞, ∞) → C. The set Lp is a vector space over the complex field.
Define the function A : Lp → Lp as A(f ) = a ∗ f , where “∗” denotes
convolution. We assume that for some constants C < ∞, c > 0, we
have the bound |a(t)| ≤ Ce−c|t| for all t ∈ R. This is sufficient to
ensure that A : Lp → Lp for any p.
(a) First consider the case where p = ∞, and let fω (t) = ejωt , where
ω ∈ R. Verify that fω ∈ L∞ is an eigenvector of A. What is the
corresponding eigenvalue?
2.11. EXERCISES 45
(b) In the special case p = 2, Lp is an inner product space with

Z
< f, g >L2 = f ∗ (s)g(s) ds, f, g ∈ L2 .
Compute the adjoint A∗ : L2 → L2 , and find conditions on a

under which A is self adjoint.
2.11.18 Let X = Rn with the usual inner product.

R∞ Let Y = Ln2 [0, ∞), the
set of functions f : [0, ∞) → R with 0 |f (s)|2 ds < ∞. We define
n
the inner product as before:

Z
< f, g >Y = f > (s)g(s) ds f, g ∈ Y
For an n × n Hurwitz matrix A, consider the differential equation

ẋ = Ax. By stability, for each initial condition x0 ∈ X, there exists a
unique solution x ∈ Y . Define A : X → Y to be the map which takes
the initial condition x0 ∈ X to the solution x ∈ Y .
(a) Explain why A is a linear operator.

(b) What is the null space N (A)? What is the rank of A?
(c) Compute the adjoint A∗ .
Eigenvectors
2.11.19 Find the eigenvalues of the matrix
 
1 1 0 0
0 1 0 0
A= 4 5 1

5
1 2 0 1
2.11.20 An n × n matrix A is called positive definite if it is symmetric,
A = A∗ = ĀT ,
and if for any x 6= ϑ, x ∈ Rn ,
x∗ Ax > 0.
The matrix A is positive semi-definite if the strict inequality in the

above equation is replaced by “≥”. Show that for a positive definite
matrix,
(a) Every eigenvalue is real and strictly positive.

(b) If v 1 and v 2 are eigenvectors corresponding to different eigenvalues
λ1 and λ2 , then v 1 and v 2 are orthogonal.
2.11.21 For a square matrix X suppose that (i) all of the eigenvalues of X
are strictly positive, and (ii) the domain of X possesses an orthogonal
basis consisting entirely of eigenvectors of X. Show that X is a pos-
itive definite matrix (and hence that these two properties completely
characterize positive definite matrices).
Hint: Make a change of basis using the modal matrix M = [v 1 · · · v n ],
where {vi } is an orthonormal basis of eigenvectors.
2.11.22 Left eigenvectors {ω i } of an n × n matrix A are defined by ω i A =
λi ω i , where {ω i } are row vectors (1 × n matrices) and the {λi } are the
eigenvalues of A. Assume that the eigenvalues are distinct.
(a) How are the {ω i } related to the ordinary (right) eigenvectors of

A∗ ?
(b) How are the {ω i } related to the reciprocal (dual) eigenvectors of
A?
Chapter 3
Solutions of State Equations
In this chapter we investigate solutions to the linear state space model (1.1)
in the general setting with x ∈ Rn , y ∈ Rp and u ∈ Rm . We begin with
the linear, time-invariant (LTI) case where the matrices (A, B, C, D) do not
depend upon time.
3.1 LTI state space models

Assume that an initial condition x0 is given at time t = 0. The LTI case
is then particularly simple to treat since we can take Laplace transforms in
(1.1) to obtain the transformed equations
sX(s) − x0 = AX(s) + BU (s)

Y (s) = CX(s) + DU (s)
Solving for X(s) gives [sI − A]X(s) = x0 + BU (s), and therefore
X(s) = [sI − A]−1 x0 + [sI − A]−1 BU (s) (3.1)
Letting
Φ(s) := [sI − A]−1
we finally obtain
X(s) = Φ(s)x0 + Φ(s)BU (s). (3.2)
To convert this into the time domain, we will take inverse Laplace trans-
forms. First define
φ(t) = L−1 (Φ(s)) = L−1 ([sI − A]−1 ) (3.3)
47
48 CHAPTER 3. SOLUTIONS OF STATE EQUATIONS
where the inverse Laplace transform of the matrix-valued function Φ(s) is

taken term by term. Since products in the transform domain are mapped to
convolutions in the time domain, the formula (3.2) may be inverted to give
Z t
x(t) = φ(t)x0 + φ(t − τ )Bu(τ ) dτ.
0
The function φ defined in (3.3) is evidently important in understanding the

solutions of (1.1). Since in the absence of inputs the state evolves according
to the equation
x(t) = φ(t)x0 , (3.4)
the function φ is known as the state transition matrix .
As for the output, in the transform domain we have
Y (s) = CX(s) + DU (s)
Assuming for simplicity that x(0) = 0, this gives
Y (s) = CX(s) + DU (s) = CΦ(s)BU (s) + DU (s)
Therefore the transfer function is a p × m matrix-valued function of s which

takes the form
G(s) = CΦ(s)B + D,
and the impulse response is the inverse Laplace transform of the transfer
function:
g(t) := L−1 {G(s)} = Cφ(t)B + Dδ(t).
In the time domain, we conclude that for zero initial conditions,

Z t Z t
y(t) = g ∗ u (t) := g(t − τ )u(τ )dτ = Cφ(t − τ )Bu(τ )dτ + Du(t)
0 0
3.2 Other descriptions of the state transition ma-

trix
We have seen that the solution of the differential equation
ẋ = Ax, x(0) = x0 (3.5)

3.2. OTHER DESCRIPTIONS OF THE STATE TRANSITION MATRIX49
is given by the formula (3.4). We also know that in the scalar case where
n = 1, the solution of (3.5) is given by the exponential x(t) = eAt x0 , so that
φ(t) = eAt in this special case. In the multivariable case, define the matrix
exponential through the power series
∞ k
X t t2 2
eAt := Ak = I + tA + A ··· ,
k! 2
k=0
so that eAt is a matrix-valued function of the matrix A. We now show that

we do indeed have φ(t) = eAt with this definition of the matrix exponential.
To see that x(t) = eAt x0 is a solution to (3.5), first note that the specified
initial conditions are satisfied:
x(0) = eA 0 x0 = Ix0 = x0 .
To see that the differential equation is satisfied, take derivatives term by

term to obtain the formula
d At
e = A + tA2 + · · · = AeAt
dt
d
This implies that dt (eAt x0 ) = A(eAt x0 ), so that the equation (3.5) is satis-
fied, and hence it is correct to assert that φ(t) = eAt .
The matrix exponential satisfies the following three properties, for any
t, s,
1. eA0 = I.
−1
2. eAt = e−At .
3. eAt eAs = eA(t+s) .
The third property, which is known as the semigroup property, is analogous

to the equation eA eB = eA+B when A and B are scalars. However, this
equation does not hold in the matrix case unless the matrices A and B
commute.
To summarize, we now have three descriptions of the state transition
matrix φ(t):
(a) φ(t) = L−1 {(sI − A)−1 }

tk k
(b) φ(t) = eAt = ∞
P
k=0 k! A
(c) φ̇(t) = Aφ(t), φ0 = I

3.3 Change of state variables

We now consider the effect of a change of basis on the state transition matrix
φ. Suppose that the n × n matrix P is used to make a change of basis, and
define the variable x̄ as
x̄ = P x x = P −1 x̄
After pre-multiplying by P , the state equation
ẋ = Ax + Bu
becomes a state equation for x̄:

−1
x̄˙ = P {z } x̄ + |{z}
| AP PB u
Ā B̄
The output equation y = Cx + Du can be similarly transformed:

−1
y = CP
| {z } x̄ + |{z}
D u
C̄ D̄
Hence for any invertible matrix P , we can form a new state space model
x̄˙ = Āx̄ + B̄u

y = C̄ x̄ + D̄u, (3.6)
with Ā = P AP −1 , B̄ = P B, C̄ = CP −1 , and D̄ = D. Moreover, if x̄(0) =

P x(0) and identical inputs are applied to both systems, then the outputs of
(1.1) and (3.6) will be identical.
The state transition matrix of the system (3.6) can be expressed in terms
of the state transition matrix of (1.1) as follows. Starting with the defining
formula Φ̄(s) = (sI − Ā)−1 = (sI − P AP −1 )−1 and noting that I = P P −1 ,
we have
Φ̄(s) = (P [sI − A]P −1 )−1 = P [sI − A]−1 P −1 .
It follows that Φ̄(s) = P Φ(s)P −1 , and hence that
φ̄(t) = P φ(t)P −1 (3.7)
To show how (3.7) may be applied, suppose that the eigenvalues of A are
distinct, so that the modal matrix M = [v 1 · · · v n ] exists and is non-singular.
Defining P = M −1 , we have seen in equation (2.8) that
Ā = P AP −1 = Λ
3.4. NUMERICAL EXAMPLES 51
The point is, when Ā is diagonal, the state transition matrix is given by the
direct formula
 λt 
e 1 0
φ̄(t) = eΛt = 
 .. 
. 
0 e λn t
and from (3.7) we have
φ(t) = M φ̄(t)M −1 ,
so that φ(t) is quickly computed via the modal matrix.
3.4 Numerical examples

Take the two-dimensional case, with A given by

0 1
A=
8 −2
We will compute φ(t) for the system (3.5) using the various methods de-
scribed in the previous sections. The simple-minded approach is to write
down the formula
t2

At 1 0 0 1 8 −2
φ(t) = e := +t + + ··· ,
0 1 8 −2 2 −16 12
but this does not give a closed form expression unless the infinite series can
be computed.
In the transform domain, we have

s+2 1
−1
8 s

−1 s −1
Φ(s) := (sI − A) = = 2
−8 s + 2 s + 2s − 8
Taking inverse Laplace transforms then gives
1 −4t 2 2t 1 −4t 1 2t
 
 e + e , − e + e 
φ(t) =  3 4 3
4 2
6
1
6  (3.8)
− e−4t + e2t , e−4t + e2t
3 3 3 3
Observe that φ(0) = I, as it should be.
A third approach is to convert the state space model to modal form. To

begin, we must compute the eigenvalues of the matrix A. The characteristic
polynomial is given by

s −1
∆(s) := det = (s + 4)(s − 2)
−8 s + 2
The roots of ∆, which are also the eigenvalues of A, are λ = −4, 2. Hence the
eigenvalues are distinct, and a modal form does exist. Solving the equation
Ax = λx for the eigenvector x in this example gives for λ = −4,

0 1 x1 x1
= −4
8 −2 x2 x2
which is solved by x2 = −4x1 . It follows that the vector

1 1
v =
−4
is an eigenvector for the eigenvalue λ1 = −4. Similarly, the vector v 2 = ( 12 )

can be shown to be an eigenvector for the eigenvalue
λ2 = 2.
1 1
The modal matrix becomes M = [v 1 | v 2 ] = , and the state
−4 2
transition matrix may be expressed as
−4t −1
−1 1 1 e 0 1 1
φ(t) = M φ̄(t)M = .
−4 2 0 e2t −4 2
After computing the matrix inverse and products, one again obtains the
expression (3.8).
3.5 Cayley-Hamilton Theorem

Another useful tool in systems theory which also gives a method for com-
puting φ is the following theorem attributed to Cayley and Hamilton. For
an n × n matrix A with real coefficients, we recall that the characteristic
polynomial defined as ∆(x) = det(xI −A) is a polynomial of degree n, which
therefore may also be expressed as
∆(x) = xn + α1 xn−1 + · · · + αn
for real coefficients {αi }.

3.5. CAYLEY-HAMILTON THEOREM 53
Theorem 3.1 (Cayley-Hamilton Theorem). The matrix A satisfies the ma-

trix characteristic equation
∆(A) = An + α1 An−1 + · · · + αn I = 0n×n (3.9)
Proof We provide a proof for the case of distinct eigenvalues only - for
the general case, see [5], pages 286, 294-5.
When the eigenvalues are distinct, we can work with the transformed
matrix
 
λ1 0
Λ=
 .. −1
 = M AM

.
0 λm
where M is the modal matrix, M = [v 1 · · · v n ]. It follows that
A = M ΛM −1 ,
and that for any k,
Ak = M Λk M −1
Applying this to (3.9) gives
∆(A) = M (Λn + α1 Λn−1 · · · + αn I)M −1

 n   n−1   
λ1 0 λ1 0 1 0
= M 
 . ..  + α1 
  ..  + · · · + αn 
  ..  −1
 M
. .
0 λnn 0 λn−1
n 0 1
λn1 + α1 λn−1
 
1 . . . + αn 0
 λn2 + α1 λ2n−1 + · · · + αn 
 −1
= M M

..
 . 
0 λnn + α1 λnn−1 + · · · + αn
For each i, the eigenvalue λi satisfies the characteristic equation ∆(λi ) = 0.

So, ∆(A) = 0, as claimed. t
u
Example 3.5.1. Consider the 2 × 2 matrix

0 1
A=
−2 −3
The characteristic polynomial in this example is

λ −1
∆(λ) = det(λI − A) = det = λ2 + 3λ + 2
2 λ+3
We therefore have

2 −2 −3 0 1 1 0
∆(A) = A + 3A + 2I = +3 +2
6 7 −2 −3 0 1
which is evidently equal to 0 (the 2 × 2 null matrix). t

u
In view of Theorem 3.1, the infinite series used to define the transition
matrix may be written as a finite sum
∞ k n−1
At
X t k
X
φ(t) = e = A = βk (t)Ak
k!
k=0 k=0
To see this, apply the Cayley-Hamilton theorem to replace any Ak for k ≥ n

by a linear combination of I, A, . . . , An−1 .
To compute φ(t), we must then find βk (t). We illustrate this with a
numerical example, using the matrix A defined above.
Example 3.5.2. Our goal is to compute
φ(t) = eAt = β0 (t)I + β1 (t)A (3.10)
by computing the scalar functions β0 , β1 . Solving the characteristic equation

λ2 + 3λ + 2 = 0 then gives the distinct eigenvalues
λ1 = −1, λ2 = −2
From the definition of the matrix exponential we have for i = 1, 2,

t2 2 i t2
eAt v i = Iv i + tAv i + A v · · · = v i + tλi v i + λ2i v i · · · = eλi t v i .
2 2
That is, the eigenvectors {v i } for the matrix A are also eigenvectors for the
matrix eAt . We also have
[β0 (t)I + β1 (t)A]v i = [β0 (t) + β1 (t)λi ]v i
and since eAt = [β0 (t)I + β1 (t)A], this gives the system of equations
eλ1 t = e−t = β0 (t) + β1 (t)(−1)

eλ2 t = e−2t = β0 (t) + β1 (t)(−2)
3.6. LINEAR TIME-VARYING SYSTEMS 55
Rewriting this as
e−t = β0 (t) − β1 (t)

e−2t = β0 (t) − 2β1 (t)
we may solve to obtain
β0 (t) = 2e−t − e−2t

β1 (t) = e−t − e−2t
Applying the formula (3.10) then gives

At β0 0 0 β1
φ(t) = e = +
0 β0 −2β1 −3β1

β0 β1
=
−2β1 β0 − 3β1
2e−t − e−2t e−t − e−2t

=
−2e−t + 2e−2t −e−t + 2e−2t
t
u
3.6 Linear Time-Varying Systems

We conclude this chapter with a development of the linear time-varying
(LTV) case
ẋ = A(t)x, x(t0 ) = x0 (3.11)
Given our experience with the LTI case, oneR might presume that the response
t
to (3.11) may be expressed as x(t) = exp( t0 A(s) ds)x0 . While this expres-
sion is valid in the scalar case, it is unfortunately not correct in general.
This is a more complicated situation since we can no longer take Laplace
transforms, and in general we do not even know if (3.11) has a solution.
However, we do have the following result.
Theorem 3.2. If A(t) is a piecewise-continuous function of t, then:
(a) For any x0 ∈ Rn , the differential equation (3.11) possesses a unique

solution x(t) defined on (−∞, ∞);
(b) The set of all possible solutions to ẋ = A(t)x forms an n-dimensional
vector space.
Proof The proof of (a) is outside of the scope of this book. To see (b),
let {xi0 } be a basis for Rn , and define for each 1 ≤ i ≤ Pn thei function ψ i to
be the solution to (3.11) with i n
P x0i = x0 . For any x0 = αi x0 ∈ R one may
easily verify that x(t) = αi ψ (t) is the unique solution to (3.11). Hence
the span of {ψ i } contains all solutions to (3.11). We now show that the {ψ i }
are linearly independent, and hence provide a basis for the solution set.
To show that the {ψ i (t)} are linearly independent time functions, assume
that there exist α1 , . . . , αn such that
α1 ψ 1 (t) + · · · + αn ψ n (t) = ϑ, t ∈ R.
αi xi0 = ϑ, and since {xi0 } was assumed to be a basis,

P
Setting t = t0 gives
we must have αi = 0 for all i. By definition, the set {ψ i (t), 1 ≤ i ≤ n} is
linearly independent, and hence forms a basis for the solution set of (3.11).
t
u
3.7 Fundamental matrices

Using Theorem 3.2 we may define an analog of the state transition matrix
φ(t) = eAt constructed earlier in the LTI case. Let {ψ i (t), 1 ≤ i ≤ n} be
any set of solutions to (3.11) which are linearly independent as vectors in
C n (−∞, ∞), and define the n × n fundamental matrix U as
U (t) = [ψ 1 (t) · · · ψ n (t)] (3.12)
Since each ψi is a solution to (3.11), the fundamental matrix satisfies the

matrix differential equation
U̇ (t) = A(t)U (t), (3.13)
just as the state transition matrix does in the LTI case. To proceed, we need
to invert the fundamental matrix.
Lemma 3.3. The inverse U −1 (t) exists for all −∞ < t < ∞.
Proof Fix t0 , and suppose that for some vector x ∈ Rn ,
U (t0 )x = ϑ ∈ Rn .
We will show that x is necessarily equal to ϑ, and from this it will follow
that U (t0 ) is invertible as claimed.
xi ψ i (t) is evidently a solution to (3.11)
P
The function x(t) = U (t)x =
with x(t0 ) = ϑ. By uniqueness of solutions, we must then have x(t) = ϑ ∈
3.8. PEANO-BAKER SERIES 57
Rn for all t, and hence xi ψ i = ϑ as a vector in C n (−∞, ∞). Since the

P
{ψ i (t), 1 ≤ i ≤ n} are linearly independent, it follows that x = ϑ. t
u
Given Lemma 3.3, we may define
φ(t, τ ) := U (t)U −1 (τ ), t, τ ∈ (−∞, ∞)
The matrix valued function φ is again called the state transition matrix
of the differential equation ẋ = A(t)x. The state transition matrix shares
many of the properties of the matrix eAt defined earlier, although we do not
have φ(t) = eA(t)t in the LTV case! We do have, however, the following
properties, which hold even in the LTV case: For any t0 , t1 , t2 ,
1. φ(t0 , t0 ) = I.
2. φ−1 (t0 , t1 ) = φ(t1 , t0 ).
3. φ(t2 , t0 ) = φ(t2 , t1 )φ(t1 , t0 ).
The third property is again known as the semigroup property.

The solution to (3.11) can now be conveniently expressed in terms of the
state transition matrix
x(t) = φ(t, t0 )x(t0 ).
To see this, use property (1) above, and the matrix differential equation
(3.13). The state transition matrix also gives the solution to the controlled
system (1.1) in the time varying case. We have that for any control u, initial
time t0 , and any initial condition x(t0 ) = x0 ,
Z t
x(t) = φ(t, t0 )x0 + φ(t, τ )B(τ )u(τ )dτ
t0
To see this, check the equation ẋ = Ax + Bu by differentiating both sides

of the equation above.
3.8 Peano-Baker Series

There are few tools available for computing the state transition matrix in
the LTV case. One approach is by iteration, which is related to the Taylor
series expansion used in the LTI case. From the differential equation
d
φ(t, t0 ) = A(t)φ(t, t0 ), φ(t0 , t0 ) = I.
dt
and the fundamental theorem of calculus, we have the integral equation

Z t
φ(t, t0 ) = I + A(σ)φ(σ, t0 ) dσ
t0
This suggests the following iterative scheme for finding φ: Set φ0 (t, t0 ) ≡ I,
and
Z t
φk+1 (t, t0 ) = I + A(σ)φk (σ, t0 ) dσ, k ≥ 0.
t0
It can be shown that if A is piecewise continuous, then this Peano-Baker

series does converge to the state transition matrix.
We can show this explicitly in the LTI case where A(t) ≡ A. The Peano-
Baker series then becomes
φ0 (t, t0 ) = I
Z t
φ1 (t, t0 ) = I + A I dσ = I + A(t − t0 )
t0
Z t
φ2 (t, t0 ) = I + A (I + A(σ − t0 )) dσ
t0
(t − t0 )2 2
= I + (t − t0 )A + A
2
Thus, we see again that in the LTI case,
k
X (t − t0 )k
φ(t, t0 ) = Ak .
k!
i=0
Matlab Commands
EXPM computes the matrix exponential.

3.8. PEANO-BAKER SERIES 59
We have provided in this chapter a detailed description of solutions to

(1.1) through the construction of the state transition matrix. Chapter 9
of [5] contains a similar exposition on the solution of state equations; for a
more detailed exposition, the reader is referred to [12].
3.9 Exercises
3.9.1 Use the (matrix) exponential series to evaluate eAt for

0 1 1 0 0 π
(a) A = (b) A = (c) A = .
0 0 1 1 −π 0
3.9.2 If A is an n × n matrix of full rank, show that

Z t
eAσ dσ = A−1 [eAt − I] .
0
Hint: This is easy once you realize that both functions of t solve the
same differential equation.
Using this result, obtain the solution to the linear time-invariant equa-
tion
ẋ = Ax + B ū , x(0) = x0 ,
where ū is a constant m-dimensional vector and B is an (n × m)-

dimensional matrix.
3.9.3 Given the single-input single-output system

0 1 0 1
ẋ = x+ ; x(0) = , y= 1 0 x,
1 0 1 0
let φ(t, s) denote the corresponding state transition matrix.
(a) Evaluate φ(t, s) using the transfer function approach (i.e. (sI −
A)−1 ).
(b) Evaluate φ(t, s) by transforming the system into diagonal form
(i.e. using modal decomposition).
(c) With u(t) ≡ 1, compute y(1).
3.9.4 Given the second-order system

0 1
ẋ = x ; x(0) = x0 ,
−4 −5
find the conditions on the initial state vector x0 such that only the
mode corresponding to the smaller (in absolute value) eigenvalue is
excited.
3.9. EXERCISES 61
3.9.5 Given the matrix

 
−1 2 0
A= 1 1 0 ,
2 −1 2
(a) Compute its inverse using the Cayley-Hamilton Theorem.
(b) Compute A6 .
3.9.6 Revisit Exercise 8 in Chapter 2, and using the Cayley-Hamilton The-
orem, explain why these three matrices must be linearly dependent.
3.9.7 Suppose that A and B are constant square matrices. Show that the
state transition matrix for the time-varying system described by
ẋ(t) = e−At BeAt x(t)
is
Φ(t, s) = e−At e(A+B)(t−s) eAs .
3.9.8 Suppose that A and Q are n × n matrices, and consider the differential
equation
Ż = AZ + ZA∗ , Z(0) = Q. (3.14)
(a) Show using the product rule that the unique solution to (3.14) is
∗t
Z(t) = eAt QeA
(b) Integrating both sides of (3.14) from 0 to tf , and applying the

fundamental theorem of calculus, conclude that if eAt → 0, then
Z tf
P = lim Z(t) dt
tf →∞ 0
is a solution to the Lyapunov equation
AP + P A∗ + Q = 0.
This fact will be very useful when we study stability, and it will
come up again in our study of optimization.
3.9.9 If φ denotes the state transition matrix for the system ẋ (t) = A(t)x(t),
show that V (t) = φ(t, t0 )Qφ∗ (t, t0 ) is a solution to the equation
d
V (t) = A(t)V (t) + V (t)A∗ (t), V (t0 ) = Q.
dt
3.9.10 Associated with the control free differential equation (3.11) is the
so-called adjoint equation
ż = −A∗ z,
where A∗ = ĀT . Show that the state transition matrix φa for the
adjoint equation satisfies
φa (t1 , t0 ) = φ∗ (t0 , t1 ) ,
where φ(t, t0 ) is the state transition matrix for (3.11).

3.9.11 The LTV differential equation (1.1), with D = 0, has the adjoint
equations
ż = −A∗ z + C ∗ v
w = B∗z .
(i) Show that the impulse response matrix ga for the adjoint model
satisfies
ga (t1 , t0 ) = g ∗ (t0 , t1 ).
(ii) In the linear case, this relation becomes
ga (s) = g ∗ (−s).
Comment on this result, in view of Exercise 17 of the previous

chapter.
3.9.12 Derive the three properties of the state transition matrix given on
page 57,
(a) In the LTI case, using properties of the matrix exponential.

(b) In the LTV case, using the definition of φ in terms of the funda-
mental matrix.
Part II
System Structural Properties
63
Chapter 4
Stability
Depicted in Figure 4.1 are results from several simulations conducted using
Matlab. Three different linear state space models of the form (1.1) have been
simulated using three different initial conditions.
1
Initial Conditions:
0.8
0
0 1
0.6
(a) x = x x01 =
1
-2 -3
0.4
0.5
0.2 x02 =
1
0
1
-0.2 x03 =
15 1
-0.4 4 x 10
3.5
0.2 0.4 0.6 0.8 1 1.2
3
2.5
1 2
0 1
0.8
1.5 (b) x = x
0.6
2 3
1
0.4
0.2 0.5
0
2 4 6 8 10 12
-0.2 14
-0.4 x 10
-1 4
-0.6 (c) x = x
-4 -1
-0.8
-0.4 -0.2 0 0.2 0.4 0.6 0.8 1 1.2
Figure 4.1: Typical solutions of the linear state space model
One can see in these plots three substantially different modes of behavior.
In (a) the solutions are attracted to the origin along the line x2 = −x1 , while
65
66 CHAPTER 4. STABILITY
in (c) solutions are also attracted to the origin, but they move there in a
cyclical fashion. In (b) solutions explode to infinity approximately along a
straight line.
These simulations suggest that some solidarity among different states ex-
ists, but this is not true in general. Consider for example the interconnection
shown in Figure 4.2. From the state equations
ẋ1 = x2
ẋ2 = 2x1 + x2 + u
we have, in the Laplace transform domain,
1 s
X1 (s) = U (s) X2 (s) = U (s)
s2 − s − 2 s2 − s − 2
Thus the output may be expressed in terms of the input as
s−2 1
Y (s) = X2 (s) − 2X1 (s) = U (s) = U (s)
s2 −s−2 s+1
Since the single pole p = −1 is in the left hand plane, this is a BIBO
stable model, as defined in any undergraduate signals and systems course.
However, the state equations also give

0 1
A=
2 1
Solving for the eigenvalues of A yields
det(sI − A) = 0 =⇒ λ1 = −1, λ2 = 2
Thus, we have an unstable root λ2 = 2, but it is cancelled by the numerator
of the transfer function due to the feed forward of x2 . For certain initial
conditions one would expect the state to diverge to infinity, even if the input
and output remain bounded.
In this chapter we present an introduction to stability theory for non-
linear models, and in the process we give a complete characterization of the
types of behavior possible for a linear state space model.
4.1 Stability in the sense of Lyapunov

Once a feedback control law has been specified, it is frequently possible to
express the closed-loop system in the form of a nonlinear state space model
ẋ = f (x); x(0) = x0 , (4.1)
4.1. STABILITY IN THE SENSE OF LYAPUNOV 67
∑
y
-
x2 x1
u
∑ 2
Figure 4.2: Is this model stable, or unstable?
n
IR
ε
x(t)
δ
xo
xe
Figure 4.3: If x0 ∈ Bδ (xe ), then x(t) = ϕ(t; x0 ) ∈ B (x0 ) for all t ≥ 0.
where x(t) evolves in Rn . We assume throughout this chapter that the

function f : Rn → Rn has continuous first and second partial derivatives
(denoted f ∈ C 2 (Rn )). The solution is denoted by x(t), or ϕ(t; x0 ) when we
wish to emphasize the dependence of the state on its initial condition. Recall
that xe is an equilibrium of (4.1) if f (xe ) = ϑ, in which case ϕ(t; xe ) = xe
for all t. Our most basic, and also the weakest form of stability considered
in this book is the following: The equilibrium xe is stable in the sense of
Lyapunov if for all > 0, there exists δ > 0 such that if kx0 − xe k < δ, then
kϕ(t; x0 ) − ϕ(t; xe )k < for all t ≥ 0.
In words, if an initial condition is close to the equilibrium, then it will stay

close forever. An illustration is provided in Figure 4.3.
Example 4.1.1. Consider the frictionless pendulum illustrated in Figure 4.4.

L
.. g
θ=- mL
sin(θ)
θ
Figure 4.4: Frictionless pendulum
Letting x1 = θ, x2 = θ̇ we obtain the state space model

ẋ1 x2
= g .
ẋ2 − mL sin(x1 )
In Figure 4.5 some sample trajectories of x(t) are illustrated. As Figure 4.5
shows, the equilibrium xe = ( π0 ) is not stable in any sense, which agrees with
physical intuition. Trajectories which begin near the equilibrium xe = ϑ will
remain near this equilibrium thereafter. Consequently, the origin is stable
in the sense of Lyapunov. t
u
Although the equilibrium xe = ϑ is stable, there is no reason to expect

that trajectories x(t) will converge to xe . Since it is frequently a goal in
control design to make the model fall into a specific regime starting from any
initial condition, we often require the following stronger forms of stability.
(a) An equilibrium xe is said to be asymptotically stable if (i) xe is stable in

the sense of Lyapunov; and (ii) for some δ0 > 0, whenever kx0 − xe k <
δ0 ,
lim ϕ(t; x0 ) = xe .
t→∞
(b) An equilibrium xe is globally asymptotically stable if it is asymptotically

stable with δ0 = ∞. Hence x(t) → xe from any initial condition.
Consider now the case where f is linear, so that ẋ = Ax. When is this
LTI model asymptotically stable? First note that by the homogeneity prop-
erty of linear systems, asymptotic stability and global asymptotic stability
are equivalent concepts. In the linear case, xe is an equilibrium if and only
4.1. STABILITY IN THE SENSE OF LYAPUNOV 69
.
x2 = θ
−π π
x1 = θ
0
0
x = 0 stable
π
x = 0 unstable
Figure 4.5: Stable and unstable equilibria.
if xe ∈ N (A). Hence the origin is always an equilibrium. For asymptotic

stability we must in fact have N (A) = {0} since, by global asymptotic sta-
bility, we cannot have more than one equilibrium. Unstable modes are also
ruled out. The following result can be demonstrated using these facts:
Theorem 4.1. An LTI model is asymptotically stable if and only if Re (λ) <
0 for every eigenvalue λ of A. t
u
A matrix A satisfying the conditions of Theorem 4.1 is said to be Hur-
witz .
Stability in the sense of Lyapunov is more subtle. For example, consider
the two matrices

0 0 0 1
A1 = A2 =
0 0 0 0
The eigenvalues of both matrices are the same: λ1 = λ2 = 0. In the first

case,

−1 1/s 0
(sI − A1 ) = ,
0 1/s
so that the state transition matrix becomes φ1 (t) = I. Since state trajec-
tories are constant, it follows that every state is an equilibrium, and that
every equilibrium is stable in the sense of Lyapunov.
However for A2 ,
1/s 1/s2

−1
(sI − A2 ) =
0 1/s
Therefore,

1 t
φ2 (t) =
0 1
which is growing with time. This implies that x(t) → ∞ for certain initial
conditions, and so this system is not stable in any sense.
Thus, the eigenvalues alone do not foretell stability in the sense of Lya-
punov. Looking at the eigenvectors, we see in the first instance from the
equation

0 0 x1 x1
=0
0 0 x2 x2
that we may take as eigenvectors

1 1 2 0
x = x =
0 1
However, in the second case, the eigenvector equation becomes

0 1 x1 x1
=0
0 0 x2 x2
This has a solution

1 1
x = ,
0
but in this example it is not possible to choose another linearly independent

eigenvector. This is a special case of the more general result:
Theorem 4.2. The LTI model is stable in the sense of Lyapunov if and
only if
(a) Re (λ) ≤ 0 for every eigenvalue λ.

(b) If Re (λ) = 0, where λ is an eigenvalue of multiplicity m, then there
exist m linearly independent corresponding eigenvectors.
Proof This follows from Theorem 8-14 of [6]. t

u
4.2. LYAPUNOV’S DIRECT METHOD 71
4.2 Lyapunov’s direct method

Lyapunov’s direct method is a general approach to verifying stability for
linear or nonlinear models. This method allows us to determine stability
without solving an ordinary differential equation. The idea is to search for a
“bowl-shaped” function V on the state space Rn . If V (x(t)) is decreasing in
time, then we can qualitatively describe the behavior of the trajectory x(t),
and determine if the model is stable in the sense of Lyapunov.
V(x(t))
x(t)
Rn
Figure 4.6: V (x(t)) represents the height of the ‘lifted’ trajectory from the
state space Rn
∆ V(x(t))
x(t)
f(x(t))
A level set of V:
{ x : V(x) = V(x(t)) }
x(0) A solution to
Rn the state equations
Figure 4.7: V (x(t)) is decreasing with time if ∇V (x) f (x) ≤ 0
To capture what we mean by “bowl-shaped”, we give the following defi-

nition. A scalar-valued function V : Rn → R with continuous partial deriva-
tives is called positive definite on an (open) region Ω about the equilibrium

xe if V (xe ) = 0, and V (x) > 0 for all x ∈ Ω, x 6= xe .
Theorem 4.3. Suppose xe ∈ Rn is an equilibrium, and that Ω ⊂ Rn is an

open set containing xe .
(a) xe is stable in the sense of Lyapunov if there exists a positive definite

function V such that
d
V (x(t)) ≤ 0 whenever x(t) ∈ Ω
dt
(b) xe is asymptotically stable if there exists a positive definite function

V such that
d
V (x(t)) < 0 whenever x(t) ∈ Ω, x(t) 6= xe
dt
(c) xe is globally asymptotically stable if there exists a positive definite

function V on Rn such that V (x) → ∞ as x → ∞, and
d
V (x(t)) < 0 for all x(t), x(t) 6= xe .
dt
Proof See for example [9]. t

u
4.2. LYAPUNOV’S DIRECT METHOD 73
Using the formula

d
V (x(t)) = ∇V (x(t)) f (x(t)),
dt
the condition for asymptotic stability becomes
∇V (x) f (x) < 0 for x ∈ Ω, x 6= xe .
This is illustrated geometrically in Figure 4.7.

As an application of Theorem 4.3 consider the Van der Pol oscillator
described by the pair of equations
ẋ1 = x2
(4.2)
ẋ2 = −(1 − x21 )x2 − x1 .
We first search for equilibria: If f (x) = ϑ, it follows that x2 = 0 and
−(1 − x21 )x2 − x1 = 0. The unique equilibrium is therefore xe = ϑ.
To see if xe is stable in any sense, try the Lyapunov function V (x) =
1 2 . We then have
2 |x|

T x2
∇V (x) = x , f (x) =
−(1 − x21 )x2 − x1
and so
∇V (x) f (x) = x1 x2 − (1 − x21 )x22 − x1 x2 = −(1 − x21 )x22 ≤ 0 If |x1 | < 1.
The set Ω = {x ∈ R2 : |x1 | < 1} contains xe and is open. Hence, xe is stable

in the sense of Lyapunov.
Is the origin asymptotically stable? Considering Figure 4.8, it is apparent
that we need to be more careful in choosing V (x). To obtain elliptical level
sets of the form shown in the second illustration of Figure 4.8, set
V (x) = 21 (x21 + x22 ) + x1 x2
where is small. If || < 1, the function V is then positive definite. The
derivative is now
T
x1 + x2
∇V (x) =
x2 + x1
and the equation to be analyzed is
∇V (x) f (x) = x1 x2 + x22 − (1 − x21 )x22 + x1 x2 + (1 − x21 )x1 x2 + x21

= − (1 − x21 ) − x22 − (1 − x21 )x1 x2 − x21 .

{x : V(x) = 1/2}
x2 x2
∆
∆ V(x)
x V(x) x
f(x) f(x)
V(x) = 1/2 |x|2 V(x) = a more general quadratic
Figure 4.8: The vectors ∇V (x) and f (x) meet at ninety degrees when x =
(γ, 0)T , and V (x) = |x|2 . By distorting V , and hence also its level set
{x : V (x) = 1/2}, one can increase this angle beyond ninety degrees.
1
If we take x21 < 2 this then gives
1
− x22 + |x1 x2 | − x21

∇V (x) f (x) ≤ − 2
Now use the bound |x1 x2 | ≤ 21 (x21 + x22 ):
∇V (x) f (x) ≤ − 21 − x22 + 12 (x21 + x22 ) − x21

= − 12 − − 12 x22 − x21 .

2
Finally, take = 16 . Then for any x 6= ϑ,

1 1 1
∇V (x) f (x) ≤ − 12 − − x22 − x21
6 12 12
1 1
= − x22 − x21 < 0.
4 12
With Ω = {x : x21 < 21 }, we see that
∇V (x) f (x) < 0 for x ∈ Ω, x 6= xe = 0.
It follows that the equilibrium xe is asymptotically stable.
4.3 Region of asymptotic stability

Let V (x) be a positive-valued function on Rn with continuous first partial
derivatives, and let Ω` designate the open region within which V (x) is less
4.3. REGION OF ASYMPTOTIC STABILITY 75
than `:
Ω` := {x : V (x) < `}.
Suppose that, within Ω` , V (x(t)) is always decreasing. This will be the case
if and only if
∇V (x) f (x) ≤ 0 for all x ∈ Ω` . (4.3)
Since under (4.3) we have V (x(t)) < ` for all t if V (x(0)) < `, it follows
that the set Ω` is invariant: if x0 ∈ Ω` , then x(t) ∈ Ω` for all t > 0. Under
a stronger set of conditions, the set Ω` becomes a region of asymptotic
stability.
Theorem 4.4. Suppose that V : Rn → R is C 1 , and that the nonlinear
model (4.1) satisfies for some ` > 0,
The set Ω` is bounded. (4.4)

V (x) > 0 for all x 6= ϑ, x ∈ Ω` . (4.5)
∇V (x) f (x) < 0 for all x 6= ϑ, x ∈ Ω` . (4.6)
Then xe = ϑ is an asymptotically stable equilibrium, and every solution

beginning in Ω` tends to ϑ as t → ∞. That is, Ω` is a region of asymptotic
stability.
Proof Since V (x(t)) is decreasing in t and is bounded from below, it
must have a limit v0 ≥ 0. If v0 = 0, then by (4.5) it follows that x(t) → ϑ,
as claimed. To prove the theorem we will show that it is impossible to have
v0 > 0.
We proceed by contradiction: If v0 > 0 then for all t, x(t) lies in the
closed and bounded set
S = {x : v0 ≤ V (x) ≤ `}.
Since ∇V (x) f (x) is strictly negative on S, by continuity it follows that for

some ε > 0,
∇V (x) f (x) ≤ −ε, x ∈ S.
Hence, from this bound, the bound V (x(0)) ≤ `, and the fundamental the-
orem of calculus, for any T > 0,
Z tf
−` ≤ V (x(tf )) − V (x(0)) = ∇V (x(t)) f (x(t)) dt ≤ −εtf .
0
Letting tf → ∞, we obtain a contradiction, and this proves the theorem. t

u
Example 4.3.1. Consider the nonlinear state space model
ẋ1 = −x1 − x2
ẋ2 = x1 − x2 + x32
With the Lyapunov function V (x) = x21 + x22 , we have for all x 6= ϑ
∇V (x) f (x) = 2(x1 , x2 ) · (−x1 − x2 , x1 − x2 + x32 )

= −2[x21 + x22 (1 − x22 )]
which will be negative for x 6= ϑ so long as x22 < 1. Considering Figure 4.9,
it follows from Theorem 4.4 that Ω1 = {x : x21 + x22 < 1} is a region of
asymptotic stability. t
u
x2
Ω1
x1 V(x) = | x |
2
Ω 1 = { x : V(x) < 1 }
{x : x < 1}
2
2
Figure 4.9: The sub-level set of V shown as a grey disk in the figure is a
region of asymptotic stability
4.4 Stability of linear state space models

We have already seen in Theorem 4.1 that to determine asymptotic stability
of a linear state space model, it is sufficient to compute the eigenvalues of
an n × n matrix. By specializing Theorem 4.3 to the LTI case, we derive
here an algebraic characterization of stability which is simpler to apply than
the eigenvalue test. Before we proceed, we need a better understanding of
positive definite quadratic functions. An n×n matrix M is called Hermitian
if M = M ∗ (complex-conjugate transpose). A matrix M is said to be positive
definite if M is Hermitian, and
x∗ M x > 0 for all x 6= ϑ, x ∈ C.

4.4. STABILITY OF LINEAR STATE SPACE MODELS 77
If the strict inequality above is replaced by x∗ M x ≥ 0, then the matrix

M is called positive semi-definite. We adopt the notation M > 0, M ≥ 0,
respectively. If V (x) = xT M x for a real matrix M > 0, then the function
V is positive definite on Rn .
There are several tests for positive definiteness which avoid computation
of x∗ M x for every x ∈ Rn . Suppose for instance that M is Hermitian, and
is partitioned as shown below:
m11 m12 m13

m11 m12 m13 m14 ∆1 = det m 11 ∆3 = det m m m
21 22 23
m31 m32 m33

m21 m22 m23 m24
M = m31 m32 m33 m34
m11 m12 m13 m14
m41 m42 m43 m44
m11 m12 m21 m22 m23 m24
∆2 = det m21 m22 ∆4 = det m31 m32 m33 m34
m41 m42 m43 m44
In this 4 × 4 example, we have taken four submatrices found on the di-

agonal, and indicated their respective determinants, ∆1 , ∆2 , ∆3 , ∆4 . These
are called the (leading) principle minors of M . One of the most readily
verified tests for positivity is given in the following theorem.
Theorem 4.5 (Sylvester-Principal Minors Test). A Hermitian matrix M is
positive definite if and only if each of its leading principle minors is strictly
positive.
Proof See [3]. t
u
Consider now the LTI model
ẋ = Ax
Choose V (x) = xT P x, where we can assume without loss of generality that
P is real and Hermitian (P = P T = P ∗ ). If P is a positive definite matrix,
then V is also positive definite.
The gradient ∇V (x), which is a row vector, may be computed using the
product rule,
∇V (x)T = P x + P T x = 2P x.
Also, f (x) = Ax, so the condition for asymptotic stability is
∇V (x) f (x) = 2xT P T Ax < 0, x 6= 0.
Since xT P T Ax = xT AT P x, we have that 2xT P T Ax = xT [P A + AT P ]x.
Thus, we need to find P > 0 such that the Lyapunov equation
AT P + P A = −Q (4.7)
is satisfied for some positive definite matrix Q. This approach characterizes

stability for LTI models:
Theorem 4.6. The LTI model ẋ = Ax is asymptotically stable if and only
if for some Q > 0, there exists a solution P > 0 to the Lyapunov equation.
Proof If for one Q > 0 we can find a P > 0 which solves the Lya-
punov equation (4.7), then we have seen that that V (x) = xT P x is a valid
Lyapunov function. Hence the model is globally asymptotically stable by
Theorem 4.3.
Conversely, if the model is asymptotically stable, then Re (λ) < 0 for
any λ, so eAt → 0 as t → ∞ at an exponential rate. Let
Z ∞
T
P = eA t QeAt dt
0
where Q > 0 is given. Given the form of the state transition matrix φ, it
follows that the matrix P has the following interpretation:
Z ∞
xT P x = x(t)T Qx(t) dt when x(0) = x.
0
T
To see that P solves the Lyapunov equation, let F (t) = eA t QeAt , and
apply the product rule for differentiation:
d T T
F (t) = AT eA t QeAt + eA t QeAt A = AT F (t) + F (t)A.
dt
So, by the Fundamental Theorem of Calculus,
Z s Z s Z s
d T

F (s) − F (0) = F (t) dt = A F (t)dt + F (t) dt A.
0 dt 0 0
Rs
We have F (0) = Q, and letting s → ∞ we have F (s) → 0, and 0 F (t)dt →
P . In the limit we thus see that the identity −Q = AT P + P A holds, so the
Lyapunov equation is satisfied. t
u
Example 4.4.1. Consider the two-dimensional model

0 1
ẋ = Ax = x.
−2 −2
To see if this model is stable we will attempt to solve the Lyapunov equation
with Q = ( 10 01 ):

0 −2 p11 p12 p11 p12 0 1 −1 0
+ =
1 −2 p12 p22 p12 p22 −2 −2 0 −1
4.5. STABILITY SUBSPACES 79
Solving the resulting three equations gives:
−2p11 − 2p12 = −1
−2p22 + p11 − 2p12 = 0
2p12 − 4p22 = −1,

5/4 1/4
which has the solution P = .
1/4 3/8
To see if the system is stable, we must now check to see if the matrix P
is positive definite. The principle minors can be computed as
5/4 > 0, (5/4)(3/8) − (1/4)(1/4) > 0
Hence the Sylvester-Principal Minors test is positive. From this we can

conclude that P > 0, and that the model is globally asymptotically stable.
t
u
4.5 Stability subspaces

The stable subspace Σs of ẋ = Ax is the set of all x ∈ Rn such that
eAt x → 0 as t → ∞.
In the special case where the matrix A has distinct eigenvalues, the stable
subspace is easily characterized using the modal form. Letting {λi } denote
the distinct eigenvalues, and {v i } the corresponding eigenvectors, we may
write any solution x as
n
X
x(t) = αi eλi t v i
i=1
where {αi } is the representation of x(0) with respect to the basis {v i }. From
this equation it is evident that
Σs = Span {v i : Re(λi ) < 0}.
In this case we may also define an unstable subspace as
Σu = Span {v i : Re(λi ) ≥ 0}.

Strictly speaking, if the stable subspace is to be real, then these formulae

will not be valid if there are complex eigenvalues. Consider for example the
three-dimensional model with
 
−2 1 −1
A =  −2 −5 6 
−1 −3 4
The eigenvalues of A are distinct, λ = 1, −2 ± j, and hence a modal form of

the state equations is possible with a complex matrix Λ.
The unstable subspace is the span of the eigenvector v 1 corresponding
to λ1 = 1:
Σu : {x : (λ1 I − A)x = 0},
from which we compute x1 = 0, x2 = x3 . Therefore,

 
 0 
Σu = Span  1 
1
 
The following vectors are eigenvectors corresponding to λ2 , λ3 :

   
1 1
v 2 =  2j  , v 3 =  −2j  .
j −j
Since the stable subspace is real, it follows that

   
1 1
Σs : x = k  j2  + k̄  −j2  , k ∈ C.
j1 −j1
where k̄ denotes the complex conjugate of k. The formula for the stable
subspace may be equivalently expressed as
   
 1 0 
Σs = Span  0 , 2  .
 
0 1
 
4.6 Linearization of nonlinear models and stability

We now return to the nonlinear case, and show that the simply stated charac-
terization of stability given in Theorem 4.6 for linear models has remarkably
4.6. LINEARIZATION OF NONLINEAR MODELS AND STABILITY81
strong implications to general nonlinear state equations. Consider again the

d
nonlinear model (4.1) with xe = ϑ. With A = dx f (ϑ), the linearized model
may be expressed
˙ = A δx.
δx (4.8)
Theorem 4.7. For the nonlinear state space model with xe = ϑ,
(a) If xe is asymptotically stable for (4.8), then xe is also asymptotically

stable for (4.1).
(b) If A has an eigenvalue with Re (λ) > 0, then xe is not stable in the
sense of Lyapunov for (4.1).
(c) If Re (λ) ≤ 0 for all λ, but Re (λ) = 0 for some λ, nothing can be
said about the stability of the nonlinear model.
Proof (sketch) For (a), the proof proceeds by constructing a Lyapunov

function for the linear model, and then applying this to the nonlinear model.
The result then follows from Theorem 4.3.
The proof of (b) is similar: If the linear model is strictly unstable in this
sense, then there exists an initial condition x0 for which the correspond-
ing solution to the linearization tends to infinity. One may then construct
ε > 0, δ > 0 such that the solution to the nonlinear equations with initial
condition x(0) = cx0 will eventually leave the ball Bε (ϑ) centered at the
origin with radius ε, whenever c < δ. This follows from the Taylor series
expansion. Hence, the origin cannot be stable in the sense of Lyapunov.
For (c), there is nothing to prove, but we may give an example. Consider
the two state space models on R1 :
(a) ẋ = −x3 (b) ẋ = +x3
In the first instance, using the Lyapunov function V (x) = x2 we see that
∇V (x) f1 (x) = −2x4 , and we conclude that the model is globally asymptot-
ically stable. In the second case, we have ∇V (x) f2 (x) = +2x4 , which shows
that trajectories explode to infinity from any initial condition. Although
the behavior of these two models is very different, both have the same lin-
earization with A = 0. t
u
We can now revisit the Van der Pol equations defined in (4.2). The
corresponding matrix A can be computed as

d 0 1
A= f (x)
=
dx x=ϑ
−1 −1
s −1
= s2 + s + 1, which gives λi =

Taking determinants, |Is − A| = det 1 s+1
√ √
−1± 1−4
2 = − 12 ± 23 j. So, the origin is asymptotically stable. Previously
we found that the Lyapunov function
V (x) = 21 x21 + 12 x22 + x1 x2
is effective for the nonlinear model. The function V may be written
1 1
T
V (x) = x P x where P = 12 21
2 2
This also works for the linearized model, as can be seen from the following
Lyapunov equation:

0 −1 1 1 0 1
AT P + P A = 12 + 12
1 −1 1 1 −1 −1

− −1 − 1 −
= 12 + 12
1− −1 −1 − 1
1

2
= − 1 = −Q.
2 1 −
Since the matrix Q is positive definite for < 4/5, we again see that the
origin is an asymptotically stable for the linearized model.
Although this is obviously a powerful method, it does not in general
tell the complete story about the stability of a nonlinear system. In some
applications we are in the situation of part (c) where the linearized model is
not asymptotically stable, and in this case the nonlinear equations must be
tackled directly. Another shortcoming of this method is that when using a
linearized model it is not possible to estimate a region of asymptotic stability
for the nonlinear model.
4.7 Input-output stability

A function f : [0, ∞) → Rn is called bounded if there exists a constant M
such that |f (t)| ≤ M for all t. A typical bounded function is illustrated
in Figure 4.10. A state space model with input u and output y is bounded
input/bounded output stable (BIBO stable) if for any t0 , and any bounded
input u(t), t ≥ t0 , the output is bounded on t0 ≤ t < ∞ if x(t0 ) = ϑ.
Consider for example the LTI model whose state space and transfer
function descriptions are expressed as
ẋ = Ax + Bu
G(s) = C[Is − A]−1 B + D.
y = Cx + Du,
4.7. INPUT-OUTPUT STABILITY 83
f (t)
-M
Figure 4.10: A typical bounded function: The graph of f remains in a tube

of radius M .
This model is BIBO stable if and only if each pole of every entry of G(s)
has a strictly negative real part.
For an LTI model, we see that asymptotic stability of ẋ = Ax implies
BIBO stability. The converse is false since pole-zero cancellations may occur.
The connections between BIBO stability and asymptotic stability may be
strengthened, but first we must develop controllability and observability for
LTI models.
Matlab Commands
EIG finds eigenvalues and eigenvectors.
ROOTS finds the roots of a polynomial.
LYAP solves the Lyapunov matrix equation (4.7).

In this chapter we have introduced a powerful approach to understanding the

qualitative behavior of dynamical systems. The main tool is the Lyapunov
function, which is interpreted geometrically as “lifting” the state trajectory
onto the surface defined by the Lyapunov function, and checking to see if the
trajectory travels “down hill”. Starting with these ideas, we then developed
(a) Criteria for stability and instability of nonlinear state space models.
(b) Strong connections between stability of a nonlinear state space model,
and stability of its linearization.
(c) An algebraic characterization of stability for linear systems, based on
the matrix Lyapunov equation.
There are many excellent texts on stability and the Lyapunov function ap-
proach. See in particular [9]. A historical development of stability and the
Lyapunov equation is contained in [8].
4.8. EXERCISES 85
4.8 Exercises
4.8.1 Explain the behavior exhibited in Figure 4.1.
(a) In this first simulation, the trajectories appear to converge to a

line within the two dimensional state space. Show that this is
the case for the given initial conditions. Will this occur for every
nonzero initial condition? Is the model stable in any sense?
(b) Explain the behavior in simulation (b). Will the trajectories ap-
proximately follow a line from each initial condition?
(c) Explain the behavior in the last simulation. Will the trajectories
cycle to zero from each nonzero initial condition?
4.8.2 This exercise shows how to use state space methods to solve numerical
equations and optimization problems.
(a) Suppose that one wishes to compute numerically a solution x? to

f (x) = g(x), where f, g : Rn → Rn are C 1 functions. Find con-
ditions on f and g which ensure that solutions to the differential
equation
ẋ = A(f (x) − g(x))
will converge to x? for initial conditions near x? . In this equation

A is an n×n matrix which can be chosen by the user. Experiment
with the algorithm using Simulink with f (x) = ex , and g(x) =
sin(x), for a suitable scalar A.
(b) Suppose that one wants to minimize a C 2 function V : Rn → R+ .
A necessary condition for a point x? ∈ Rn to be a minimum is
that it be a stationary point; that is, ∇V (x? ) = ϑ.
Consider the differential equation
ẋ = −P ∇V (x)T ,
where P is a positive definite matrix. Find conditions on the

function V to ensure that a given stationary point x? will be
asymptotically stable for this equation. Hint: Find conditions
under which the function V is a Lyapunov function. Try this
algorithm out using Simulink with a function of your choice with
n = 2. For example, try V (x) = x61 − 2x21 + x1 x2 + 6 + |x2 |3 .
4.8.3 Consider the pendulum described in Figure 4.4, with the addition of
an applied torque u. This is a nonlinear system which is described by
the differential equation θ̈ = − sin(θ) + u.
Find a PD controller (state feedback) which stabilizes the system

about θ = π. That is, for controlled state space model with x = θθ̇ ,
the equilibrium ( π0 ) is asymptotically stable.
Is ( π0 ) globally asymptotically stable with your control law?
4.8.4 Suppose that V (x) = xT Ax for a real matrix A, and suppose that the
function V is positive definite. Prove that the matrix M = 21 (A + AT )
is positive definite. Can you find an example in which the matrix A is
not symmetric?
4.8.5 Study the stability of the equilibrium state xe = 0 (x := (y ẏ)T ) for
the nonlinear system
ÿ + ẏ + y 3 = 0
using the function
V (y, ẏ) = y 4 + y 2 + 2y ẏ + 2ẏ 2
as a candidate Lyapunov function.

4.8.6 For the LTI system described by

−1 1
ẋ = x,
−2 3
investigate asymptotic stability using the Lyapunov equation
AT P + P A = −Q with Q = I.
Can you arrive at any definitive conclusion?

4.8.7 Consider the linear model

1 1
ẋ = x,
−1 − 2α + α2 −1 − α + α2
where α is a scalar parameter. For what values of α is the origin (i)

asymptotically stable? (ii) stable in the sense of Lyapunov?
4.8. EXERCISES 87
4.8.8 Consider the simple scalar model
ẏ = α0 u.
We would like to use feedback of the form u = −Ky + N r so that

the closed-loop pole is at −1, where K and N are scalar constants.
However, the parameter α0 is not known!
To solve this problem, we estimate α0 using the “observer”:
d
α̂ = `(ẏ − α̂u)u.
dt
Here ` is a positive constant.
d
(a) Verify that dt (α̂ − α0 )2 ≤ 0 for all time. Hence the absolute
identification error |α̂ − α0 | can never increase, regardless of the
control. Find a control signal u(t) which forces the error to go to
zero.
Assuming that α̂(t) does converge to α0 , a natural control law is

1
u(t) = (r(t) − y(t))
α̂(t)
We now consider the regulation problem, where r(t) = r0 ∈ R for all t.
(b) Obtain a nonlinear state space model for the closed-loop system,
with states y and α̂.
(c) Show that y = r0 ; α̂ = β is an equilibrium for any β 6= 0. Deduce
that the equilibrium ( αr00 ) cannot be asymptotically stable.
(d) Linearize the nonlinear model about an equilibrium. What can
r0
you conclude about the stability of an equilibrium β ?
(e) Show that V (y, α̂) = 21 (y−r0 )2 + 21 (α̂−α0 )2 is a Lyapunov function
for this model. Can you conclude that the equilibrium (r0 , α0 ) is
stable in the sense of Lyapunov?
These ideas are the starting point of adaptive control; an approach

based on simultaneous system identification and control.
4.8.9 We have seen that stability of an LTI model may be characterized in
terms of the existence of a Lyapunov function, and we found that the
Lyapunov function in the stable case can be taken as
Z ∞
T
V (x) = x φ(τ )T Qφ(τ ) dτ x = xT P x with Q > 0.
0
A similar result is possible in the nonlinear case under suitable stability

conditions, although the function V will not necessarily be a quadratic
if the model is not linear.
Suppose that for the nonlinear model ẋ = f (x), the origin xe = 0 is
globally asymptotically stable, and suppose moreover that the stability
is such that the solutions φ(t; x) are square
R ∞ integrable over time. Hence
T
for any Q > 0 the function V (x) = 0 φ(τ ; x) Qφ(τ ; x) dτ is finite
valued.
(a) Compute V (x(t)) for t > 0, and simplify using properties of the
state transition function, e.g.,
φ(0; x) = x, φ(t; φ(s; x)) = φ(t + s; x)
d
(b) Using (a), show that dt V (x(t))|t=0 = −x(0)T Qx(0). Conclude
that V is an effective Lyapunov function for this model.
Chapter 5
Controllability
In this chapter we consider exclusively the LTV state space model
ẋ = A(t)x + B(t)u, x(t0 ) = x0 , (5.1)
and as a special case its LTI version. Our interest lies in the effectiveness
of the input u to control the state x to arbitrary points in Rn . Consider for
example the two-dimensional LTI model illustrated in Figure 5.1.
From the state equation

0 1 b
ẋ = x + 1 u.
−2 −3 b2
we see that when b1 = b2 = 0, this is the example illustrated in Figure 4.1 (a)
at the beginning of the previous chapter. Obviously, in this case the control
cannot affect the state, so when b = ϑ the model is uncontrollable in the
strongest sense. Consider now b = ( 01 ), and suppose that it is desired to
choose a control which takes the initial state x0 = ( 11 ) to a final state
xf = ( 0.5
0.5 ) at some time tf . One approach is illustrated in Figure 5.1:
Simply follow the control-free trajectory until a time ts at which the state
is nearly “below” the desired final state xf . Then, apply a large positive
value so that the vector bu dominates the state derivative Ax + bu. This
procedure will place x(tf ) at xf for some tf , and a similar strategy can be
used to drive any state to any other state.
A substantially different situation occurs with b = −1

1 . In this case the
vector b is parallel to the subspace indicated by the dotted line Span {v 1 } =
Span {b}, where v 1 is an eigenvector for A. It can be shown in this case that
no control will take a state from one side of this line to the other, and that if
89
90 CHAPTER 5. CONTROLLABILITY
1 1
x0 = 1
0.8
0.6
0 1
0.4 x = x
-2 -3
0.2
0
(control-free)
-0.2
-0.4
0.2 0.4 0.6 0.8 1 1.2
0.8
0.6
0 1 0
0.5
0.4 xf = 0.5
x = x + u
-2 -3 1
0.2
0
0 ; t < ts
u(t) =
-0.2 10 15 ; t s < t < ts + ε
-0.4 ts
0.2 0.4 0.6 0.8 1 1.2
Figure 5.1: The controlled model can be steered to any state in Rn using
impulsive control.
the initial state x0 lies on this line, then it will stay there forever, regardless
of the control.
The general LTV model (5.1) is called controllable at time t0 if there
exists a finite time, tf , such that for any initial state x0 , and any final
state xf , there is a control u defined on [t0 , tf ] such that x(tf ) = xf . This
concept is not obviously directly useful in control design, except possibly in a
ballistics problem where xf denotes a target! However, we will see that this
is a central idea in the theory of state space models, and that controllability
will be an underlying assumption in many of the control design methods to
be developed. The purpose of this chapter is to characterize controllability
to better understand examples such as these simple models. We will see in
Theorem 5.6 that the pathology illustrated in Figure 5.2 is the only way
that a lack of controllability can occur for an LTI model.
91
1
x0 = 1
0.6
0.4
0.2
-0.2
-0.4
0.2 0.4 0.6 0.8 1 1.2
0.6 0 1 -1
0.4 x = x + u
-2 -3 1
0.2
0
(any control)
-0.2
-0.4
0.2 0.4 0.6 0.8 1 1.2

1
x0 = -1
Span {v 1 = 1 }
-1
Figure 5.2: If the initial condition satisfies x1 = −x2 , so that the state lies on
the indicated diagonal line, then the state can never leave this line, regardless
of the control. If the initial condition does not start on the diagonal, then
it may never reach it, though it will come arbitrarily close to it for large t.
5.1 A preview: The LTI discrete-time case

Controllability is most easily understood in the LTI discrete-time case
x(k + 1) = Ax(k) + Bu(k), k = 0, 1, 2, 3, . . . ,
where again x is of dimension n and u is of dimension m. Assume that we
start at x(0) = ϑ, and that it is desired to move to another state xf in finite
time. We have x(1) = Bu(0), and in general
k=2 x(2) = Ax(1) + Bu(1) = ABu(0) + Bu(1)
k=3 x(3) = Ax(2) + Bu(2)
= A2 Bu(0) + ABu(1) + Bu(2)
..
.
k=r x(r) = Ar−1 Bu(0) + Ar−2 Bu(1) + . . . Bu(r − 1).
This can be written compactly as
 
u(r − 1)
 u(r − 2) 
x(r) = [B AB · · · Ar−1 B]  ..
 

 . 
u(0)
which shows that we can reach arbitrary xf ∈ Rn at time tf = r if and
only if R([B AB · · · Ar−1 B]) = Rn , which holds if and only if the rank of
[B AB · · · Ar−1 B] is equal to n.
How do we know when to stop iterating? The Cayley-Hamilton Theo-
rem 3.1 tells us that Ak can be written as a linear combination of Ai , i < n.
Hence the rank of the matrix [B | AB | · · · | Ar−1 B] cannot increase for
r ≥ n. The controllability matrix is the n × (nm) matrix
C := [B | AB | · · · | An−1 B].
From this discussion it follows that in the discrete time model it is possible
to steer from the origin to any other state in Rn if and only if rank (C) = n.
5.2 The general LTV continuous-time case

We saw in Section 3.7 that the solution to the state space model (5.1) may
be expressed
Z t
x(t) = φ(t, t0 )x0 + φ(t, τ )B(τ )u(τ ) dτ,
t0
5.2. THE GENERAL LTV CONTINUOUS-TIME CASE 93
Controllability is thus determined by the range of the linear operator A : D[t0 , tf ] →

Rn defined as Z tf
A(u) = φ(tf , τ )B(τ )u(τ ) dτ. (5.2)
t0
The input space D[t0 , tf ] is the set of all piecewise continuous time functions
on [t0 , tf ]. We leave the proof of the following as an exercise.
Theorem 5.1. The following are equivalent for the LTV model (5.1)
(a) The model is controllable at time t0 ;

(b) R(A) = Rn for some tf > t0 .
(c) There exists a time tf > t0 such that for any x0 ∈ Rn , there is a
control u defined on [t0 , tf ] such that x(tf ) = ϑ when x(t0 ) = x0 .
The controllability grammian W (t0 , tf ) for the LTV model (5.1) is the
positive definite matrix
Z tf
W (t0 , tf ) := φ(t0 , τ )B(τ )B ∗ (τ )φ∗ (t0 , τ ) dτ. (5.3)
t0
The following deeper characterization of controllability gives a finite dimen-

sional test in terms of the n × n matrix W . The essence of the proof is
this: the grammian is full rank if and only if the n rows of the function
φ(t0 , · )B( · ) are linearly independent as functions in D[t0 , tf ]. This then is
equivalent to controllability of the model.
Theorem 5.2. The LTV model (5.1) is controllable at time t0 if and only
if there exists a finite time tf > t0 such that W (t0 , tf ) is nonsingular.
Proof
(i) Assume W (t0 , tf ) is nonsingular, so that W (t0 , tf )−1 exists. We will
show that the model is controllable by showing that it is possible to control
any x0 to ϑ at time tf . Consider the control
u(t) = −B ∗ (t)φ∗ (t0 , t)W −1 (t0 , tf )x0 .
Given the formula

Z tf
x(tf ) = φ(tf , t0 )x0 + φ(tf , τ )B(τ )u(τ ) dτ
t0
for the final value of the state, we may substitute the formula for the control
to give
Z tf
φ(tf , τ )B(τ ) −B ∗ (τ )φ∗ (t0 , τ )W −1 (t0 , tf )x0 dτ

x(tf ) = φ(tf , t0 )x0 +
t0
Z tf
= φ(tf , t0 )x0 − φ(tf , τ )B(τ )B ∗ (τ )φ∗ (t0 , τ ) dτ W −1 (t0 , tf )x0 .
t0
The term W −1 (t0 , tf )x0 is taken outside of the integral since it does not
depend on τ . By the semigroup property
φ(tf , t0 )φ(t0 , τ ) = φ(tf , τ ),
the integral becomes φ(tf , t0 )W (t0 , tf ). Therefore

x(tf ) = φ(tf , t0 )x0 − φ(tf , t0 )W (t0 , tf ) W −1 (t0 , tf )x0
which gives x(tf ) = ϑ. Hence by Theorem 5.1 the model is controllable at

time t0 .
(ii) We now show that controllability of the LTV model implies that
W (t0 , tf )−1 exists. We prove the contrapositive: Assuming W (t0 , tf ) is sin-
gular, we show that the model is not controllable.
If W is singular, there exists an α 6= ϑ such that
W (t0 , tf )α = ϑ and hence also α∗ W (t0 , tf )α = 0. (5.4)
From the definition of W it then follows that

Z tf
α∗ φ(t0 , τ )B(τ )B ∗ (τ )φ∗ (t0 , τ )α dτ = 0
t0
This can be equivalently expressed as

Z tf
hB ∗ (τ )φ∗ (t0 , τ )α, B ∗ (τ )φ∗ (t0 , τ )αi dτ = 0,
t0
or
Z tf
|B ∗ (τ )φ∗ (t0 , τ )α|2 dτ = 0.
t0
Since the integral of the square norm is zero, and all of these functions are
assumed to be piecewise continuous, we must have B ∗ (τ )φ∗ (t0 , τ )α = ϑ and
hence
α∗ φ(t0 , τ )B(τ ) = ϑ∗ (5.5)
5.3. CONTROLLABILITY USING THE CONTROLLABILITY MATRIX95
for all τ ∈ [t0 , tf ].

If the model is controllable at time t0 , then starting at the particular
state x0 = α, there is a control u which makes x(tf ) = ϑ:
Z tf
x(tf ) = ϑ = φ(tf , t0 )α + φ(tf , τ )B(τ )u(τ ) dτ.
t0
From the property φ−1 (tf , t0 ) = φ(t0 , tf ), we can multiply this equation
through by by φ−1 to obtain
Z tf
ϑ=α+ φ(t0 , tf )φ(tf , τ )B(τ )u(τ ) dτ.
t0
From the semigroup property we must therefore have

Z tf
α=− φ(t0 , τ )B(τ )u(τ ) dτ.
t0
Multiplying both sides on the left by α∗ gives

Z tf
2
|α| = − α∗ φ(t0 , τ )B(τ )u(τ ) dτ
t0
which by (5.5) must be zero. However, |α|2 6= 0 by assumption. Hence the

model cannot be controllable if W is singular, and the proof is complete. t
u
5.3 Controllability using the controllability ma-

trix
In the LTI discrete-time case we have shown that the controllability matrix
is key to understanding controllability. To show that this is still true for the
continuous-time LTI model
ẋ = Ax + Bu
(5.6)
y = Cx + Du,
we consider how the controllability grammian W is related to the controlla-

bility matrix
C = [B | AB | A2 B | · · · | An−1 B].
Using this approach we will prove the following theorem.

Theorem 5.3. For the LTI model (5.6), the model is controllable if and
only if the n × (nm) matrix C has rank n.
Proof To begin, note that in the LTI case we have for t0 = 0,
Z tf
W (0, tf ) = φ(−τ )BB ∗ φ∗ (−τ ) dτ
0
Z tf
∗
= e−Aτ BB ∗ e−A τ dτ. (5.7)
0
To proceed, we show that
rank (C) < n ⇐⇒ W is singular.
(i) We first show that assuming rank (C) < n, we must also have that
W is singular.
If rank (C) < n, there exists α 6= ϑ such that
α∗ [B AB · · · An−1 B] = ϑ∗
which can be equivalently expressed as
α∗ Ak B = ϑ∗ , k = 0, . . . , n − 1.
From the Cayley-Hamilton Theorem 3.1, the state transition matrix can be
expressed as
n−1
X
φ(−τ ) = e−Aτ = βk (−τ )Ak
k=0
from which it follows that α∗ φ(−τ )B = ϑ∗ for all τ . From (5.7) we see that
α∗ W (0, tf ) = ϑ∗ , which shows that W is singular.
(ii) Assume now that W is singular. We will show that rank (C) < n.
We have seen in equation (5.5) that if W (0, tf ) is singular, then there
exists an α ∈ Rn such that
α∗ φ(0, τ )B(τ ) = ϑ∗ , 0 ≤ τ ≤ tf .
In the LTI case, it follows that α∗ e−Aτ B = ϑ∗ for all τ . By setting t = 0

this shows that α∗ B = ϑ∗ . Moreover, if a function is zero over an interval,
so is its derivative:

d ∗ −At T −At

(α e B) = α (−Ae B), =0
dt t=0
5.4. OTHER TESTS FOR CONTROLLABILITY 97
which shows that also −α∗ AB = ϑ∗ . Continuing this procedure, we see that
d2 ∗ −At

(α e B) = α ∗ A2 B = ϑ ∗ ,
dt2 t=0
and for any k
(−1)k−1 α∗ Ak−1 B = ϑ∗ .
It then follows that
α∗ [B|AB|A2 B | · · · | An−1 B] = ϑ∗ ,
which establishes the desired result that rank (C) < n. t

u
A careful examination of the proof shows that we in fact have established
the stronger result
R(C) = R(A).
Hence, the subspace of Rn
Σc := R(C) = {Cz : z ∈ Rmn }
is equal to the set of all states which can be reached from the origin. The
set Σc is called the controllable subspace.
5.4 Other tests for controllability

The controllability of an LTI state space model is much more apparent when
the model is in modal form. To see this, we must construct the modal form
using a similarity transformation, so we begin with the following:
Theorem 5.4. Controllability of the LTI model
ẋ = Ax + Bu
is invariant under any equivalence transformation x̄ = P x
Proof For the transformed model x̄˙ = Āx̄ + B̄u we have
C¯ = [B̄ | ĀB̄ | · · · | Ān−1 B̄]

= [P B | (P AP −1 )P B | (P AP −1 )(P AP −1 )P B | · · · | (P AP −1 ) · · · (P AP −1 )P B].
1 x1
u
B1
s-p
1
1 x2
u
B2 s-p
2
1 xn
u
Bn s - pn
Figure 5.3: The modes of the model are decoupled, and hence the model
cannot be controllable if some B̄i is zero.
Therefore, the controllability matrix for the transformed model is
C¯ = [P B | P AB | P A2 B | · · · | P An−1 B],
or C¯ = P C. Since the matrix P is nonsingular, it follows that rank C¯ =

rank C. t
u
Suppose that the eigenvalues of A are distinct, so that A is similar to a
diagonal matrix Λ:
Ā = Λ, P = M −1 ,
where M = [v 1 · · · v n ]. This gives the modal form
x̄˙ = Λx̄ + B̄u,
where the modes of the model are decoupled. Figure 5.3 illustrates the
structure of the transformed model. If some B̄i is zero, which means that
a row of the matrix B̄ is zero, it is apparent that the model cannot be
controllable. It turns out that the converse is also true:
Theorem 5.5. Suppose that the matrix A has distinct eigenvalues. Then,
(A, B) is controllable if and only if B̄ = M −1 B has no zero rows.
Proof We prove the theorem in two steps.

(i) If row i of of B̄ = M −1 B is zero, it follows that if the initial state

satisfies xi (0) = 0, then xi (t) = 0 for all t. Obviously then the model is not
controllable.
(ii) Conversely, if the model is not controllable, there exists an α ∈ Rn ,
α 6= 0, such that
Z tf
∗ ∗
0 = α W̄ α = α∗ e−Λτ B̄ B̄ ∗ e−Λ τ α dτ.
0
Equivalently, we must have

n
X
αi B̄i∗ e−λi τ ≡ ϑ∗ .
i=1
If none of the {B̄i } are zero, choose k such that αk 6= 0. We then have
n
1 X
e−λk τ = αi B̄k B̄i∗ e−λi τ .
αk B̄k B̄k∗ i=1
i6=k
This is impossible, since the functions {eλi t } are linearly independent in

C[t0 , tf ]. Hence, some row Bk must be zero, which thus proves the theorem.
t
u
Theorem 5.5 may also be understood in terms of reciprocal eigenvectors.
Recall that the matrix B̄ may be expressed as
 1∗ 
r
−1  .. 
B̄ = M B =  .  B
rn∗
The jth row of B̄ is rj∗ B. From Theorem 5.5 we see that (A, B) is control-
lable if and only if there is no rj within the null space of B ∗ .
Since rj∗ is also a left eigenvector (see Exercise 22 of Chapter 2), we
always have
rj∗ (λj I − A) = ϑ∗ .
Then, (A, B) is uncontrollable if and only if there is a j such that
rj∗ (λj I − A) = ϑ∗ and rj∗ B = ϑ∗ ,
which can be written compactly as
rj∗ [ λj I − A| B] = ϑ∗ .
To summarize, we see that the pair (A, B) in uncontrollable if and only if

for some j,
rank [λj I − A | B] < n.
This final test was derived using a modal form, but it turns out to be true
in general:
Theorem 5.6 (Hautus-Rosenbrock test). The pair (A, B) is controllable if

and only if for every s ∈ C,
rank [sI − A | B] = n. (5.8)
Proof One direction is easy: If r∗ [λI − A | B] = ϑ∗ , then we must have

for any input u,
d ∗
(r x (t)) = r∗ Ax(t) + r∗ Bu(t) = λr∗ x(t),
dt
so that r∗ x(t) = eλt r∗ x(0), regardless of the control! Hence, if (5.8) is
violated for s = λ, then the model is not controllable.
We have seen that the converse is a direct consequence of Theorem 5.5 in
the case of distinct eigenvalues, but the proof is more subtle in full generality.
The reader is referred to page 184 of [6] for details. t
u
Note that when s is not an eigenvalue of A the matrix sI − A already
has rank n, so that we also have rank ([sI − A | B]) = n. Hence the rank of
the matrix (5.8) need only be checked for s equal to the eigenvalues of A.
Example 5.4.1. We illustrate some of the controllability tests developed

so far using the simple model

−2 0 1
ẋ = x+ u.
−1 −1 1
(i) The controllability matrix for this model is

1 −2
C = [B | AB] = .
1 −2
Since the rank of C is 1, the model is not controllable. The controllable

subspace Σc is equal to the range space of C, which in this case is the
diagonal in R2 :
Σc = {α ( 11 ) : α ∈ R}.
(ii) The eigenvalues

of Aare λ1 = −1, λ2 −2, and the modal matrix
0 1 −1 1
can be taken as M = , M −1 = . We can then compute
1 1 1 0

−1 0 −1 0
Λ= B̄ = M B= ,
0 −2 1
and again we conclude that the model is not controllable. Since the first
row is zero, it follows that the first mode (λ1 = −1) is not connected to the
input in modal form.
(iii) We have

s+2 0 1
[sI − A | B] = .
1 s+1 1
Evaluating this matrix with s = −1 gives

1 0 −2
.
1 0 −2
Since the rank of this matrix is one, we again see from the Hautus-Rosenbrock
test that the model is not controllable. t
u
5.5 Exercises
5.5.1 Investigate the controllability properties of the LTI model ẋ = Ax +
Bu, for the three pairs of (A, B) matrices given below.

−5 1 1
(a) A = , B= .
0 4 1
   
3 3 6 0
(b) A = 1 1 2 , B = 0
  
2 2 4 1
   
0 1 0 0
(c) A = 0 0 1 , B = 0
  
0 0 0 1

−5 1 1
5.5.2 For the system model corresponding to A = ,B= , obtain
0 4 1
a control that drives the system from x = (1 0)T at t = 0, to x =
(0 1)T at t = 1. Simulate your control using Simulink.
5.5.3 Given the linear time-invariant model
   
−7 −2 6 1 1
ẋ =  2 −3 −2 x + 1 −1 u = Ax + Bu
−2 −2 1 1 0

−1 −1 2
y = x = Cx ,
1 1 −1
check controllability using
(a) the controllability matrix

(b) the rows of B̄ = M −1 B, where M is chosen such that M −1 AM is
diagonal
(c) The Hautus-Rosenbrock condition.
5.5.4 Transform the state space model below into CCF, and from the re-
sulting equations compute its transfer function.
   
0 0 −2 1
ẋ = 1 0 1 x + 1 u
   y = (0 0 1)x
0 1 2 1
5.5. EXERCISES 103
5.5.5 Consider the LTI model

   
−3 1 0 0
ẋ =  0 −3 0 x + 1 u
0 0 4 0
y = [1 0 1] x.
For what initial states x(0) = x0 is it possible to choose a control u so

that y(t) = te−3t for t > 1?
5.5.6 Let A be an n × n matrix and B be an n × r matrix, both with real
entries. Assume that the pair (A, B) is controllable. Prove or disprove
the following statements. (If the statement is false, then producing a
counterexample will suffice.)
(a) The pair (A2 , B) is controllable.
(b) Let k(·) be an r-dimensional (known) function, piecewise continu-
ous in t. Then, the model described by
ẋ = Ax + Bu + k(t)
is completely controllable, in the sense that any state can be

reached from any other state using some control.
(c) Given that the model ẋ = Ax + Bu has the initial condition
x(0) = x0 6= ϑ, it is possible to find a piecewise continuous con-
trol, defined on the interval [0, ∞), so that the model is brought
to rest at t = 1 (that is, x(t) = ϑ for all t ≥ 1).
(d) Assume now that the model above is initially (at t = 0) at rest,
and we wish to find a piecewise continuous control which will
bring the state to x̄ ∈ Rn by time t = 1 and maintain that value
for all t ≥ 1. Such a control can always be found.
5.5.7 Consider the LTI model ẋ = Ax + Bu. Suppose that A has n distinct
eigenvalues, and corresponding eigenvectors {v 1 , . . . , v n }.
Show that if R(B) ⊂ Span (v 1 , . . . , v j−1 , v j+1 , . . . , v n ) for some j then
the model is not controllable.
Hint: Look at the reciprocal basis vectors: How are rj and R(B) re-
lated?
5.5.8 Compute the adjoint A∗ of the operator A defined in (5.2), and then
compute the composition
V = A ◦ A∗ : Rn → Rn
as an n × n matrix. How are the controllability grammian W and the

operator V related? Comment on this result in view of Theorem 5.2
and Exercise 16 of Chapter 2.
5.5.9 Consider the m-input, p-output LTI model described by the convolu-
tion
Z ∞
y(t) = g(t − τ )u(τ ) dτ,
−∞
where g is a p × m matrix-valued function.

Call the model output controllable if for any y0 ∈ Rp , t0 ∈ R, there
exists a control u0 such that y(t0 ) = y0 .
(a) Formulate output controllability as a range space dimension prob-

lem for a specific linear operator.
(b) Compute the adjoint of the linear operator determined in (a).
(c) Derive a finite dimensional test for output controllability, based
on the rank of some p × p matrix (see the previous problem).
5.5.10 Prove Theorem 5.1.

Chapter 6
Observability, Duality and

Minimality
In many control problems we obtain measurements y(t), but we may not

be able to measure directly all of the system states. For a complex system
with a large number of states, it may be expensive to position a sensor to
measure every state, and it may even be impossible to do so. In the light of
this, the question we address in this chapter is the following: Is it possible
to determine the state using only the input-output information {u(t), y(t)}?
Informally, we say that some states are observable if one can determine
them unambiguously based on these input-output measurements. This is
obviously a very important property for a physical system.
T(t) = rod temperature
u(t) = rod levels y(t) = power
Figure 6.1: If the input to this plant does not affect the internal temperature,
another actuator is needed. If the temperature cannot be estimated based
on input-output measurements, another sensor is needed to directly measure
the temperature.
To give a simple, if artificial example, consider the academic model of
105
106 CHAPTER 6. OBSERVABILITY, DUALITY AND MINIMALITY
+ xc
+ -
u(t) y(t)
- C
Figure 6.2: If the input to this circuit is zero, then the output is zero. Hence
the charge across the capacitor is not observable.
a nuclear reactor given in Figure 6.1. Suppose that there is a state space
model for this plant with the input equal to the rod levels, the output equal
to the power generated, and one of the states equal to the rod temperature.
Suppose first that using this control it is not possible to influence the tem-
perature T . Then the system is not controllable, and for safe operation it
will be necessary to add another actuator to raise or lower the internal tem-
perature. This situation seems highly unlikely for a real reactor. Suppose
now that given the input-output measurements, it is not possible to estimate
the internal temperature T . We then say that the system is not observable.
This again is a highly undesirable situation, though it can be resolved by
adding another measurement - a sensor to directly measure T . Hence, an
uncontrollable or unobservable plant can be “repaired” by expanding the
number of inputs or outputs for the plant.
To take a more realistic example, consider the circuit given in Figure 6.2.
It is obvious here that if the input voltage is zero, then the output voltage
is zero, regardless of the voltage across the capacitor. Hence, the voltage xc
is not observable.
To give a formal definition of observability, consider the LTV model with
an output equation y of the form
ẋ(t) = A(t)x + B(t)u;

x(t0 ) = x0 . (6.1)
y(t) = C(t)x + D(t)u.
Since we are assuming that the system is defined by a linear state space
model, to determine x(t) for all t it is sufficient to compute the initial state
x0 = x(t0 ).
6.1. THE OBSERVABILITY MATRIX 107
The model (6.1) is said to be observable at time t0 if there exists a finite

time t1 > t0 such that for any initial x0 at t0 , knowledge of the input u(t)
and the output y(t) for t0 ≤ t ≤ t1 suffices to determine x0 . Given the
expression for the state
Z t
x(t) = φ(t, t0 )x0 + φ(t, τ )B(τ )u(τ ) dτ
t0
the output may be expressed as

Z t
y(t) = C(t)φ(t, t0 )x0 + C(t) φ(t, τ )B(τ )u(τ ) dτ + D(t)u(t) .
t0
| {z } | {z }
unknown ū(t) known
Since ū is known, we can simplify matters by considering the known quantity
ȳ(t) = y(t) − ū(t) = C(t)φ(t, t0 )x0 . (6.2)
Equivalently, when analyzing observability we may assume without loss of

generality that the input is zero. Defining the linear operator B : Rn →
D[t0 , t1 ] as
B(x) = f, where f (t) = C(t)φ(t, t0 )x,
we see that the model is observable if, for some t1 , the operator B is one to
one.
6.1 The observability matrix

In the LTI case we have
ȳ(t) = CeAt x0 t ∈ [0, t1 ],
where in this case we have taken t0 = 0. We have some information about

x0 through the equation y(0) = Cx0 . However, typically C is not square,
so inversion cannot be used directly to solve this equation for x0 . Taking
derivatives yields
˙
ȳ(t) = CAeAt x0
˙
ȳ(0) = CAx0 ,
and this process can be continued until sufficient information has been gen-
erated to find x0 :
   
ȳ(0) C
 ˙
ȳ(0)   CA 
   
 ¨
ȳ (0)   CA2 
=  x0 .
..
  
   .. 
 .   . 
(n−1) n−1
ȳ (0) CA
Defining the observability matrix as the (np) × n-matrix

 
C
 CA 
 2 

O := 
 CA  ,
 ··· 
CAn−1
where p is the dimension of y, we can now prove the following theorem.
Theorem 6.1. The LTI model (5.6) is observable if and only if
rank O = n.
Proof If O is full rank, then we have seen that x0 can be computed using
y and its derivatives.
Conversely, suppose that there is an x0 ∈ Rn , x 6= ϑ, such that Ox = ϑ.
We then have CAk x0 = ϑ for any k, and hence also CeAt x0 = ϑ for all t.
Since x0 is indistinguishable from the state ϑ, it follows that the model is
not observable. t
u
The unobservable subspace Σō is defined to be the null space of the ob-
servability matrix. From the proof above, we see that for an initial condition
x(0) ∈ Σō , the output y is identically zero if the input is equal to zero.
6.2 LTV models and the observability grammian

In the LTV case we define the observability grammian as
Z t1
H(t1 , t0 ) := φ∗ (τ, t0 )C ∗ (τ )C(τ )φ(τ, t0 ) dτ. (6.3)
t0
Theorem 6.2. The model
ẋ = A(t)x, y = C(t)x
6.3. DUALITY 109
is observable at time t0 if and only if there exists a finite time t1 such that
the observability grammian H(t1 , t0 ) has rank n.
Proof We first prove that non-singularity of the observability grammian
is sufficient for observability:
(i) From (6.2) we have
Z t Z t
∗ ∗
φ (τ, t0 )C (τ )ȳ(τ ) dτ = φ∗ (τ, t0 )C ∗ (τ )C(τ )φ(τ, t0 ) dτ x0 .
t0 t0
Thus Z t1
φ∗ (τ, t0 )C ∗ (τ )ȳ(τ ) dτ = H(t1 , t0 ) x0 . (6.4)
t0
| {z } | {z }
known n × 1 vector known n × n
matrix
If H is nonsingular, x0 is uniquely determined by (6.4), and we conclude

that the model is observable. We now prove the converse:
(ii) Assuming the model is observable, we prove that H is nonsingular
by contradiction. If H is singular, then there exists α 6= 0 such that
H(t1 , t0 )α = ϑ and hence α∗ H(t1 , t0 )α = ϑ.
This may also be written as

Z t1
α∗ φ∗ (τ )C ∗ (τ )C(τ )φ(τ ) dτ α = 0
t0
or equivalently
hCφα, CφαiL2 = kCφαk2L2 = 0.
This implies that C(t)φ(t)α = 0 for all t0 ≤ t ≤ t1 . Since the two initial
conditions x(t0 ) = x0 + α and x(t0 ) = x0 yield the same output ȳ(t), we
conclude that the model is not observable. t
u
6.3 Duality
By exploiting the similarities between controllability and observability, we
can generate further tests for observability. This approach is known as
duality: observability and controllability are dual concepts. This can be
made precise by defining a model which is “dual” to the model under study.
Given two models
ẋ = A(t)x + B(t)u
(I)
y = C(t)x + D(t)u, x(t0 ) ∈ Rn .
ż =− A∗ (t)z + C ∗ (t)v
(II)
γ = B ∗ (t)z + D∗ (t)v, z(t0 ) ∈ Rn .
The second model is called the dual of the first.
In the LTI case we know that rank (CII ) = n if and only if (II) is con-
trollable, where
CII = [C ∗ | A∗ C ∗ | A∗2 C ∗ | · · · | A∗n−1 C ∗ ].
Since we have
 
C
∗
 CA 
CII =  = observability matrix for model I,
 
CA2
..
 
.
it is obvious that the controllability test for model II is equivalent to the

observability test for I in the LTI case. This is a special case of the following
result.
Theorem 6.3. For the general LTV model,
(a) The LTV model (I) is controllable at t0 if and only if (II) is observ-
able at t0 .
(b) The model (I) is observable at t0 if and only if (II) is controllable at
t0 .
Proof We just prove (a) since the proof of (b) is identical.

Exercise 10 of Chapter 3 gives the relation
φII (t1 , t0 ) = φI (t0 , t1 )∗ .
Hence the controllability grammian for I and the observability grammian for
II are identical:
Z tf
WI := φI (t0 , τ )BI (τ )BI∗ (τ )φ∗I (t0 , τ ) dτ
t0
Z tf
= φ∗II (τ, t0 )CII
∗
(τ )CII (τ )φII (τ, t0 ) dτ
t0
= HII .
6.4. KALMAN CANONICAL FORMS 111
u xc
Bc ∑
xc Ac
A 12
Ac
Figure 6.3: The Kalman Controllability Canonical Form
In view of Theorem 5.2 and Theorem 6.2 the proof of (a) is complete. t
u
As a direct application of Theorem 6.3 and Theorem 5.6 we have the
following dual of the Hautus-Rosenbrock test.
Theorem 6.4. The LTI model is observable if and only if
rank [sI − A∗ | C ∗ ] = n for any s ∈ C.
t
u
We then call a mode λ unobservable if rank [λI − A∗ | C ∗ ] < n.
6.4 Kalman canonical forms

If the model is not controllable, then we can construct a state transformation
P in which the state x̄ = P x has the form x̄ = x̄x̄cc̄ , as shown in Figure 6.3.
After this transformation, the state space model takes the form
c
Ac A12 x̄ Bc
x̄˙ = + u
0 Ac̄ x̄c̄ 0
c
x̄
y = [Cc Cc̄ ] .
x̄c̄
This is called the Kalman Controllability Canonical Form (KCCF). In the

case of distinct eigenvalues, the eigenvalues of (Ac ) are precisely those for
which rank [sI − A | B] = n. The matrix P can be defined as

 
← n1 →
← (n − n1 ) →
linear independent set of vectors
P −1 = 
 

 columns linearly independent of

of C columns of C
where rank (C) = n1 < n (see Brogan, §11.7).

Example 6.4.1. Consider the LTI model with

−2 0 1
A= , B=
−1 −1 1

1 −2
The controllability matrix is C = , which has rank one. Hence the
1 −2
model is not controllable. Define the matrix P by

−1 1 1
P =
1 0
where the column ( 11 ) was taken as the first column of the controllability
matrix, and the column ( 10 ) was chosen to make P −1 full rank. Thus,

0 1
P =
1 −1
The Kalman controllability canonical form can be written as

−1 −2 −1 1
Ā = P AP = B̄ = P B =
0 −1 0
t
u
For general LTI models, the observability canonical form can be obtained
using duality: Using a state transformation, the transformed state may be
written x̄ = x̄x̄oō with

AO 0 x̄o Bo
ẋ = + u
A21 Aō x̄ō Bō
o
x̄
y = [Co 0]
x̄ō
This Kalman Observability Canonical Form (KOCF) is illustrated pictorially

in Figure 6.4.
6.5. STATE SPACE MODELS AND THEIR TRANSFER FUNCTIONS113
xo y
u Bo ∑ Co
Ao
A 21
xo
Bo ∑
Ao
Figure 6.4: The Kalman Observability Canonical Form
6.5 State space models and their transfer func-

tions
It should now be clear that given a transfer function, there are many possible
state space descriptions. Moreover, the dimension of the realization is not
unique since we can always append uncontrollable or unobservable dynamics.
A state space realization of a transfer function P (s) is called minimal if it
has minimum dimension over all state space models for which
C(sI − A)−1 B + D = P (s).
Theorem 6.5. A realization (A, B, C, D) of a transfer function P (s) is
minimal if and only if it is both controllable and observable.
Proof For a general proof, see [5], #12.17, page. 439. We prove the
theorem here in the special case where the eigenvalues of A are distinct.
When the eigenvalues are distinct, we have the partial fraction expansion
u
X ki
P (s) = D + (6.5)
s − λi
i=1
where in the MIMO case, the {ki } are p × m matrices. By considering the
modal form, the transfer function can also be written as
P (s) = D̄ + C̄(sI − Ā)−1 B̄ (6.6)
where
Ā = Λ B̄ = M −1 B
C̄ = CM D̄ = D
This latter form will allow us to compute the gains {ki } in (6.5). First break
up the matrices C̄, B̄ into their component columns and rows, respectively:
C̄ = [γ 1 · · · γ n ]
β 1T
 
B̄ = [β 1 · · · β n ]T =  ... 
 
β nT
Then the expansion (6.6) can be written explicitly as

n
X γ i β iT
P (s) = D̄ + .
s − λi
i=1
The model is controllable if and only if each βi is non-zero, and it is observ-

able if and only if each γi is non-zero. From the equation above, it follows
that the modal form is controllable and observable if and only if none of the
poles λi are cancelled in this sum. That is, there are no pole/zero cancella-
tions in the expression C(sI − A)−1 B + D. Thus, the transfer function has
n poles, and hence any realization must have dimension no less than n. t u
Consider for example
s+1 s+1
P (s) = = .
s2 + 6s + 5 (s + 5)(s + 1)
The second order modal realization is not minimal, but the first order real-
ization is.
6.6 Realization of MIMO transfer functions

For a SISO model with transfer function P it is straightforward to obtain
a minimal realization. First, cancel any common pole-zero pairs, count the
number of poles remaining, say n, and construct any n-dimensional state
space model. One can for example use the CCF. For a MIMO transfer
function the situation is more complex. We refer the reader to [6] for details
- here we provide an example to illustrate some of the issues involved.
6.6. REALIZATION OF MIMO TRANSFER FUNCTIONS 115
Consider the transfer function P which describes a 2-input, 2-output

model (m = p = 2):
" #
1 2
s+1 s+1
P (s) = −1 1 .
(s+1)(s+2) s+2
By examining each entry of the matrix P we see that there are no direct
pole-zero cancellations. Since there are two poles, one might expect that
a realization can be constructed of degree two. On the other hand, since
P consists of three first-order, and one second-order transfer functions, one
might suspect that a degree four state space model will be required. It turns
out that for this model both predictions are incorrect. The degree of any
minimal realization is three, as we now show.
1
s+1 Y1
2 ∑ ∑
-
1
U1 s+1
-1 ∑
-
1
U2 s+2 Y2
∑ ∑
-
Figure 6.5: A block diagram realization of P .
The input-output description Y (s) = P (s)U (s) may be written as

1
Y1 (s) = [U1 (s) + 2U2 (s)]
s+1
1 1
Y2 (s) = [− U1 (s) + U2 (s)]
s+2 s+1
Hence, using just three first order filters we may simulate this model using
the block diagram description shown in Figure 6.5. Taking the outputs of
the integrators as states, we obtain the following state space realization:
   
−1 0 0 1 2
ẋ =  0 −1 0  x +  −1 0  u
0 1 −2 0 1

1 0 0
y = x.
0 0 1
By computing the controllability and observability matrices one can show

that this model is both controllable and observable. Hence, this is a minimal
realization.
A general approach to obtaining a minimal realization for a MIMO trans-
fer function is to mimic the partial fractions expansion used in the SISO case
to form the modal form. This will again yield a diagonal, minimal realiza-
tion. For the transfer function P given above we may write

1 s + 2 2(s + 2)
P (s) = .
(s + 1)(s + 2) −1 s+1
Performing a partial fractions expansion on each term we obtain

1 1 2 1 0 0 1 1
P (s) = + := R1 + R2 .
(s + 1) −1 0 (s + 2) 1 1 (s + 1) (s + 2)
The rank of R1 is two, and the rank of R2 is one. The sum of these ranks
gives the degree 3 of the minimal realization.
To compute the diagonal realization let
R1 = C1 B1 R2 = C2 B2
|{z} |{z} |{z} |{z}
2×2 2×2 2×1 1×2
For instance, we can take

1 2 1 2 1 0
R1 = =
−1 0 −1 0 0 1

0 0 0
R2 = = 1 1
1 1 1
6.6. REALIZATION OF MIMO TRANSFER FUNCTIONS 117
With this factorization, the transfer function P may be expressed

1 1
P (s) = C1 B1 + C2 B 2
(s + 1) (s + 2)
" #
1
0

(s+1) 1
= C1 1 B1 + C2 B2
0 (s+1) (s + 2)
 1 
(s+1) 0 0
1  B1
= [C1 | C2 ]  0 0

(s+1) 
B2
1
0 0 (s+2)
 1 
0 0

(s+1) 1 0
1 2 0  1
=  0 (s+1) 0   0 1 
−1 0 1 1
0 0 (s+2)
1 1
= C(Is − A)−1 B.
These values of A, B and C define a minimal realization of the transfer

function.
In summary, when a simple partial fractions expansion is possible, then a
minimal realization is obtained by a factorization of the matrix gains found
in this expansion. In general, this approach might fail if one of the de-
nominator polynomials in P possesses a repeated root. However, a minimal
realization may still be found by a more general factorization. Write the
transfer function P as
P (s) = NR (s)DR (s)−1
where NR , DR are matrices whose entries are polynomial functions of s.

This is known as a right factorization. A left factorization can be defined
analogously: P (s) = DL (s)−1 NL (s). Using the right factorization one may
construct a generalization of the controllable canonical form, and using the
left factorization an observable canonical form may be constructed. To en-
sure minimality, one must impose a co-primeness condition on the matrices
N and D - for details, the reader is referred to [6].
6.7 Exercises
6.7.1 Consider the transfer function
s+1
P (s) =
s2 + 3s + 2
(a) Obtain a second-order state space realization in controllable canon-
ical form. Is your realization controllable? Observable?
(b) Obtain a second-order state space realization in observable canon-
ical form. Is your realization controllable? Observable?
6.7.2 Given the linear time-invariant model

   
−7 −2 6 1 1
ẋ =  2 −3 −2 x + 1 −1 u = Ax + Bu
−2 −2 1 1 0

−1 −1 2
y = x = Cx
1 1 −1
check observability using
(a) The observability matrix

(b) columns of C̄ = CM , where M is chosen such that M −1 AM is
diagonal
(c) the Hautus-Rosenbrock test.
6.7.3 Prove the following assertion: If the state space model
ẋ = Ax + Bu, y = Cx + Du
is a minimal realization of the transfer function G(s), and if the model

is BIBO stable, then the state space model ẋ = Ax is asymptotically
stable.
For example: Take the BIBO stable model Y (s) = (s − 1)/[(s − 1)(s +
1)]U (s). Any second order realization will have characteristic poly-
nomial s2 − 1, and hence the second order model ẋ = Ax cannot be
asymptotically stable. A minimal realization will be first order with
A = −1, so that ẋ = Ax is asymptotically stable.
6.7.4 Show that the observability grammian given by (6.3) is equal to B ∗ ◦B,
where B(x) = C(t)φ(t, t0 )x.
Part III
Feedback
119
Chapter 7
Pole Placement
In this chapter, we discuss a number of feedback control issues for the LTI
state space model
ẋ = Ax + Bu
(7.1)
y = Cx + Du.
We consider initially the regulation problem: By translating the state space,
we may assume that we are regulating to the value x = ϑ. To perform this
translation we must assume that the DC gain of the plant is known exactly,
and that there are no disturbances. We will develop more “robust” designs
which work in spite of disturbances later in Section 8.1.
The following example serves to illustrate most of the concepts intro-
duced in this chapter. Consider the special case of (7.1),
   
1 0 0 1
ẋ = Ax + Bu = 0 2 0  x + 1 u
(7.2)
0 0 −1 0
y = Cx = 1 0 1 x
The system has three modes:
λ1 = 1 controllable and observable.

λ2 = 2 controllable but not observable.
λ3 = −1 observable but not controllable.
Define the control as u = −Kx + r. This is linear state feedback , with an

auxiliary input r. With this controller, the closed-loop system becomes
ẋ = (A − BK)x + Br = Acl x + Br
(7.3)
y = Cx
121
122 CHAPTER 7. POLE PLACEMENT
where Acl , the closed-loop system matrix, can be explicitly written as

 
1 − k1 −k2 −k3
Acl =  −k1 2 − k2 −k3 
0 0 −1

with K = k1 k2 k3 . We now study properties of the closed-loop system
(7.3), such as controllability, observability, stability, as a function of K.
Fact 1. The controllability matrix
C = [B | Acl B | A2cl B]
has rank 2 for all k1 , k2 , k3 . Hence,
The rank of C is invariant under state feedback. The control-

lable subspace is invariant under state feedback.
Fact 2. Using the Hautus-Rosenbrock test, one may show that λ = −1 is

the eigenvalue corresponding to the uncontrollable mode of Acl , independent
of K. The other two modes of Acl are controllable for all K, but their values
depend explicitly on K. Hence,
Uncontrollable modes are invariant under state feedback.
Fact 3. The observability matrix for the controlled system is

   
C 1 0 1
O = CAcl  =  1 − k1 −k2 −k3 − 1 .
2
CAcl 2
(1 − k1 ) + k1 k2 −k2 (3 − k1 − k2 ) 1 + k3 (k1 + k2 )
This matrix does not have a fixed rank. It could be
one (corresponding to k1 = 1, k2 = 0, k3 = −1),
two (corresponding to k1 = k2 = k3 = 0), or
three (corresponding to k1 = k3 = 0, k2 = −1)
Hence,
Observability is not invariant under state feedback.

7.1. STATE FEEDBACK 123
Fact 4. The eigenvalues of Acl can be found by solving the characteristic

equation
det(λI − Acl ) = 0 ⇒ (λ + 1)(λ2 − (3 − k1 − k2 )λ + 2 − 2k1 − k2 ) = 0.
One eigenvalue is λ = −1, the other two depend on k1 and k2 . Since the two
terms k1 + k2 and 2k1 + k2 are linearly independent, these two eigenvalues
can be chosen arbitrarily (subject to the restriction that complex eigenvalues
can appear only as conjugate pairs) by proper choices of real scalars k1 and
k2 . Hence,
Arbitrary pole placement is possible by linear state feedback, in

the controllable subspace.
The LTI model (7.2) is stabilizable by linear state feedback, since the
uncontrollable mode is asymptotically stable.
After gaining an understanding of the state feedback control problem

we will turn to output feedback of the form u = −Gy + r. Even if an LTI
system is both controllable and observable, it is in general not possible to
assign the closed-loop poles arbitrarily by static output feedback. We will
see shortly that all poles can be assigned arbitrarily (subject to the complex
conjugate pair restriction) if we use dynamic output feedback, where G is a
transfer function rather than a static matricial gain.
7.1 State feedback

Returning to the regulation problem, assuming that we can measure all of
the states, the simplest controller is given by the state feedback control law
u = −Kx,
where K is a n × m matrix. This gives rise to the closed-loop system
ẋ = (A − BK)x = Acl x.
To determine whether or not x(t) → ϑ from any initial condition, we must

consider the eigenvalues of the closed-loop matrix Acl .
Consider the SISO case, where we initially assume that the pair (A, B)
is in CCF (controllable canonical form):

   
0 1 0 ··· 0 0
 0 0 1 ··· 0   0 
 .. . .. ..
   
A= . .. .. .. ,

B=

.

 . . .   . 
 0 0 ··· 0 1   0 
−αn −αn−1 · · · · · · −α1 1
Recall that
(a) The characteristic polynomial of this model is
∆(s) = det(sI − A) = sn + α1 sn−1 + · · · + αn
(b) (A, B) is a controllable pair.
Writing the control as u = −k1 x1 − k2 x2 · · · − kn xn , it is immediate that

the closed-loop system has the form
 
0 1 0 ··· 0
 0 0 1 ··· 0 
. . .
 
A − BK = Acl = 
 .. .. .. .. .. ,

 . . 
 0 0 ··· 0 1 
−(αn + k1 ) −(αn−1 + k2 ) · · · · · · −(α1 + kn )
so that the closed-loop characteristic polynomial is
∆(s) = det(sI − [A − BK]) = sn + (α1 + kn )sn−1 + · · · + (αn + k1 ). (7.4)
We can assign the coefficients of this polynomial, and hence the closed-loop
poles can be assigned arbitrarily.
This analysis can be applied to any controllable SISO system. It is
performed on a system in CCF, but we can transform to CCF for any time
invariant SISO state space model, provided the model is controllable. The
question is, what is the state transformation P that is required to ensure
that x̄ = P x is described by a model in CCF? The controllability matrices
C and C¯ are related by the formula
C¯ = P C.
¯ −1 . The inverse exists if the original
Hence the matrix P is given by P = CC
state space model is controllable. System eigenvalues do not change with
7.1. STATE FEEDBACK 125
a state transformation, so by converting to CCF and then applying pole

placement to the transformed model, one can place the poles as desired.
This is summarized in the following theorem. While this result was derived
for SISO systems, it is also true in the MIMO case.
Theorem 7.1. The eigenvalues of (A − BK) can be placed arbitrarily, re-
specting complex conjugate constraints, if and only if (A, B) is a controllable
pair.
All of this is applicable if (A, B) is controllable. What do we do if (A, B)
is not controllable? Consider the Kalman controllability canonical form
introduced previously in Figure 6.3, and given by the equations below

Ac A12 Bc
ẋ = x+ u
0 Ac̄ 0

xc
x = .
xc̄
We can still apply a control law of the form

xc
u = −[K1 K2 ]
xc̄
to obtain the closed-loop system

Ac − Bc K1 A12 − Bc K2
ẋ = x (7.5)
0 Ac̄
From the form of the closed-loop system, we find that the eigenvalues are
dependent only upon (Ac − Bc K1 ) and not on Ac̄ . In fact, the characteristic
polynomial for the closed-loop system becomes
∆(s) = det(sI − Acl ) = det(sI − (Ac − Bc K1 )) det(sI − Ac̄ )
| {z }| {z }
arbitrary independent
Based on K of K
In other words, one can place the controllable modes arbitrarily, but the
uncontrollable modes remain unchanged.
The pair (A, B) is said to be stabilizable if there exists a K such that
(A − BK) is a Hurwitz matrix. From these arguments we see that
Theorem 7.2. If the eigenvalues are distinct, then (A, B) is stabilizable if
and only if the eigenvalues of the uncontrollable modes are in the strict LHP.
For a general LTI model, (A, B) is stabilizable if and only if the eigen-
values of Ac̄ lie in the strict LHP, where Ac̄ is the matrix defined in the
KCCF.
Geometrically, Theorem 7.2 implies that the model is stabilizable if and

only if the unstable subspace is contained in the controllable subspace.
7.2 Observers
The state feedback approach can be generalized to the situation where only
partial measurements of the state are available. Consider again the LTI
model
ẋ = Ax + Bu,
y = Cx.
We would like to define an observer, of the form illustrated in Figure 7.1,

in which input-output measurements are collected on-line to give estimates
b(t) of the state x(t). To mimic the behavior of the system one can try
x
b˙ (t) = Ab
x x(t) + Bu(t).
Defining the error as x

e(t) = x(t) − x e˙ (t) =
b(t), this gives the error equation x
x(t), from which we deduce that
Ae
e(t) = eAt x
x e(0).
This is a poor approach since there is no flexibility in design. If for example

the open-loop system is unstable, then for some initial conditions the error
will not converge to zero, and may diverge to infinity.
Plant Observer
^
U Y X
P(s) O(s)
Figure 7.1: Can the observer O be designed so that x̂(t) − x(t) → 0, as

t → ∞, at a predesignated rate?
The observers we consider are of a similar form, but we adjoin an output

error term:
b˙ = Ab
x x + Bu + L(y − yb), b(0) ∈ Rn ,
x
7.3. OBSERVER FEEDBACK 127
where yb = C x
b. For any fixed n × p matrix L one obtains
b˙ = Ax + Bu − Ab
ė = ẋ − x x − Bu − L(Cx − C x
b) = (A − LC)e.
To ensure that x
e(t) → 0 as t → ∞, we must compute the eigenvalues of the
matrix (A − LC) as a function of L. But note:
eig(A − LC) = eig(A∗ − C ∗ L∗ )
We are thus exactly in the same position as when we considered pole place-
ment. It follows that the eigenvalues of (A − LC) can be placed arbitrarily,
provided that the matrix pair (A∗ , C ∗ ) is controllable. Based on duality, this
is simply observability of the pair (A, C). Thus, we can place the observer
poles arbitrarily if and only if (A, C) is observable.
This again raises the question, what do we do if this basic assumption
fails? The pair (A, C) is said to be detectable if there exists some L such that
(A − LC) is a Hurwitz matrix. This property is the dual of stabilizability,
in the sense that (A, C) is detectable if and only if (A∗ , C ∗ ) is stabilizable.
We then obtain the following dual statement to Theorem 7.2:
Theorem 7.3. If the eigenvalues are distinct, then (A, C) is detectable if

and only if the eigenvalues corresponding to the unobservable modes lie in
the strict LHP.
For a general LTI model, (A, C) is detectable if and only if the eigenval-
ues of Aō lie in the strict LHP, where Aō is the defined in the KOCF.
A state space model is thus detectable if and only if the unobservable

subspace is contained in the stable subspace.
7.3 Observer feedback

The objective in the previous section was to construct a useful state estima-
tor for a state space model, but this of course was rooted in the desired to
control the system. Suppose that we ignore that x b is an estimate, and we
apply the observer feedback control law
u(t) = −K x
b(t).
This control law is designed based on the faith that x e(t) will converge to
zero fast enough so that this control law is essentially equivalent to full state
feedback. To see if this faith is in vain, we must first see if the overall system
Plant Observer
+ ^
U Y X
∑ P(s) O(s)
-
Figure 7.2: The separation principle justifies the use of state estimates in
place of the true state values in the feedback control law. However, full state
feedback is not completely equivalent to feedback based on state estimation.
is asymptotically stable. The overall system is linear, so we can check this

by computing the closed-loop eigenvalues.
First apply the following state transformation for the overall state

x I 0 x e = x−x
=
b
e I −I x
b b = x − e.
x
| {z }
P

x
The transformed state is described by the closed-loop equations
e

ẋ A − BK BK x
=
ė 0 A − LC e
| {z }
Acl
Thus, the closed-loop poles of the overall system are defined by the charac-
teristic polynomial
det(sI − Acl ) = det(sI − [A − BK]) · det(sI − [A − LC]) (7.6)
where
(a) det(sI − [A − BK]) ≡ state feedback eigenvalues – arbitrary placement

if and only if (A, B) is controllable.
(b) det(sI − [A − LC]) ≡ observer eigenvalues – arbitrary placement if and
only if (A, C) is observable.
7.4. REDUCED-ORDER (LUENBERGER) OBSERVERS 129
Equation (7.6) is known as the separation principle. Its obvious consequence

to control design is that the design of the feedback gain can be conducted
independently of the design of an observer. One must be cautious when
interpreting this result however, since we are only considering stability and
the placement of closed-loop poles. In particular, we are ignoring such issues
as sensitivity to plant uncertainty.
The separation principle allows us to separately place the state feedback
eigenvalues for good dynamic response, and the observer-eigenvalues so they
are faster than the desired closed-loop response. This still leaves open the
question, where do we place the observer poles? A typical rule of thumb is
that the observer poles should be faster than the “slowest” state feedback
pole by a factor of 2 to 5. This however is very problem specific. In some
applications, the location of the observer poles is not very important, so
long as the poles are faster than those of the state feedback poles. In other
examples, the observer must be designed with care.
7.4 Reduced-order (Luenberger) observers

Considering the state space equations (7.1) we see that at least part of the
state is directly observed through y. Assume for simplicity that D = 0, and
construct a state transformation x̄ = P x so that y is explicitly a part of the
state. This can be accomplished if C has rank p, since then we can define
the matrix P so that
" #
y C
x̄ = = anything for x.
x̄2 linear independence
Then we do have a transformed model in the desired form:

x̄˙ 1

A11 A12 x̄1 B1
= + u.
x̄˙ 2 A21 A22 x̄2 B2
| {z } | {z }
P AP −1 PB
Since by definition we have x̄1 = y, the output equation is evidently

x̄1
y = [I 0] .
x̄2
The state estimation problem has been reduced in complexity through

this transformation: rather than construct an estimator of dimension n, we
only have to estimate the (n−p) dimensional vector x̄2 . The next step then is
to construct an observer for x̄2 alone. To see how this can be accomplished,
write the transformed state equations as follows. The equation for x̄2 may
be written in the suggestive form,
x̄˙ 2 = A22 x̄2 + A21 y + B2 u . (7.7)

| {z }
known from
measurements
The equation for x̄1 can be written
ẏ − A11 y − B1 u = A12 x̄2 (7.8)

| {z }
known from
measurements
Defining ū = A21 y + B2 u and ȳ = ẏ − A11 y − B1 u, we obtain the state space

model
x̄˙ 2 = A22 x̄2 + ū

ȳ = A12 x̄2
We can now write down the corresponding observer to estimate x̄2 . The
only technicality is that ȳ contains the derivative ẏ, which strictly speaking
is not known. We will assume for now that we know ẏ, but through some
manipulations of the observer equations we will relax this assumption below.
An observer for x̄2 takes the form
b˙ 2 = A22 x
x b2 + ū + L(ȳ − ȳˆ) (7.9)
where ȳˆ = A12 x

b2 . Since ȳ = A12 x̄2 , the error equation for e := x̄2 − x
b2 takes
the form
b˙ 2 = (A22 − LA12 )e
ė = x̄˙ 2 − x
So we can place the observer poles arbitrarily if and only if the pair (A22 , A12 )
is observable.
We now show that it is unnecessary to differentiate the output measure-
ments. First write the observer equations in the expanded form
b˙ 2 = A22 x
x b2 + A21 y + B2 u + L(ẏ − A11 y − B1 u − A12 x
b2 ).
These equations can be represented through the block diagram

Y d
A21 - LA11
dt
U ^
X2
B2 - LB1 ∑
A22 - LA12
It is evident in this diagram that the term Lẏ is immediately integrated.

To eliminate the derivative, we simply cancel the derivative and integral to
obtain the equivalent block diagram
Y
A21 - LA11
U ^
X2
B2 - LB1 ∑ ∑
A22 - LA12
We then obtain an estimate x

b of the original state through
Y ^
X
^ P -1
X2
In conclusion, provided (A22 , A12 ) is an observable pair, the state can

be estimated based upon an (n − p)-dimensional observer using the mea-
surements u, y, and it is not necessary to differentiate these measurements.
This is important, since if the measurement y is noisy, its derivative may be
much worse.
If we define

A11 A12
Ā =
A21 A22
C̄ = [I 0]
then by the Hautus-Rosenbrock test we know that (Ā, C̄) is observable if

and only if

sI − Ā
rank =n for any s ∈ C.
C̄
This matrix may be written as

 
sI − A11 −A12
 −A21 sI − A22  .
I 0
The first p columns are automatically independent because of the identity

matrix I. Because of the 0, the matrix is full rank if the last n − p columns
are independent of each other. That is, we must have

−A12
rank = (n − p).
sI − A22
This condition holds if (A22 , A12 ) is observable, since this is the Hautus
test on the reduced complexity observer problem. We conclude that the
condition for observability of the reduced order observer is equivalent to
observability of (Ā, C̄), which is equivalent to observability of the original
pair (A, C).
Example 7.4.1. Consider the magnetically suspended ball, whose linearized

and normalized state space model is given by

0 1 0
ẋ = x+ u, y = 1 0 x.
1 0 1
Full state feedback is simply proportional-derivative (PD) feedback:
u = −Kx = −K1 y − K2 ẏ.
If the derivative cannot be measured directly, then there are several possi-
bilities:
(a) The second state x2 = ẏ could be obtained by differentiating y, but

this will greatly amplify any high-frequency noise.
(b) An approximate derivative could be obtained through a filter:
x̂˙ 2 + `x̂2 = `ẏ,
or in the frequency domain
˙ `s
X̂2 (s) = Y (s),
s+`
where ` is a large constant. The trouble with this approach is that it
does not make use of the system model, and for this reason one cannot
expect that x̂2 (t) − ẏ(t) → 0 as t → 0.
(c) A reduced order observer can be constructed to estimate ẏ. If there is
noise, the observer can be designed to obtain a good trade off between
tracking of ẏ, and rejection of this noise. In the noise free case we have
that x̂2 (t) − ẏ(t) → 0, as long as the observer gain is chosen to give a
stable observer.
The reduced order observer is defined by the equations
x̂˙ 2 = x1 + u + `(ẋ1 − x̂2 ).
Taking transforms gives a formula which is similar to the estimator (b) given
above:
˙ `s + 1 1
X̂2 (s) = X1 (s) + U (s).
`+s `+s
However, this observer does make use of system information, and hence does
exhibit desirable convergence properties.
The controller can be written in the frequency domain as
`s + 1 1
U (s) = −K1 X1 (s) − K2 X1 (s) − K2 U (s).
`+s `+s
Solving for U , it is readily seen that for large ` this is a lead compensator.
That is, the above expression may be written as
s−z
U (s) = −g Y (s)
s−p
where p < z < 0. In conclusion, we see that in simple models, observer
based feedback generalizes classical control design. t
u
Matlab Commands
In Matlab there are two methods for assigning eigenvalues. These can be
used to obtain the state feedback gain K, or in observer design by
placing the eigenvalues of (A∗ − C ∗ L∗ ) to compute the observer gain
L.
ACKER This command is named after Ackerman, who derived a formula

for explicit solution of the feedback gain K using the controller canon-
ical form. This is only suitable for SISO plants.
PLACE A more numerically robust algorithm which can be used for MIMO
systems. This command cannot be used with repeated roots.
7.5. EXERCISES 135
7.5 Exercises
7.5.1 For the LTI model

−1 2 1
ẋ = x+ u
0 3 2
(a) Find the controllable subspace.

(b) Show that the system is unstable, and compute the unstable modes
and corresponding eigenvector.
(c) Is the system stabilizable? Explain carefully using (a) and (b). If
the system is stabilizable, find a stabilizing state feedback control
law.
7.5.2 Consider the SISO, LTI system

11 1
ẋ = x+ u; y = [7 8]x
03 1
(a) There is a matrix P such that (Ā, B̄) = (P AP −1 , P B) is in con-

trollable canonical form. Compute P and P −1 using the formula
P = CC¯ −1 .
(b) Compute by hand a feedback control law K̄ = [k̄1 , k̄2 ] which places
the eigenvalues of Ā − B̄ K̄ at −2 (twice).
(c) Let K = K̄P . Verify that the eigenvalues of A − BK are at −2.
7.5.3 Given the system

−1 4 2
ẋ = x+ u,
1 −1 1
(a) Show that when u ≡ ϑ the origin is unstable.

(b) Determine the stable subspace Σs , the span of the eigenvectors cor-
responding to unstable modes, Σu , and the controllable subspace
Σc . Plot all on the same graph.
(c) Now choose the feedback control
u = (1 β)x + r
where β is a scalar, and r is a reference input. Show that the

closed-loop system will be asymptotically stable for some value(s)
of β.
(d) What are the closed-loop poles? Do different choices of β affect

the stable mode?
7.5.4 Consider the single input/single output second-order linear time-invariant

system
ẋ = Ax − [ a0 ] u
y = (1, b)x
where a and b are scalar real parameters.

It is known that for some constant c > 0,
−2
−2
AT [ 1c ] =

−2c , and A [ 11 ] = −2 ,
where AT denotes the transpose of A. Also, it is known that λ, one of

the eigenvalues of A, is strictly positive.
(a) Compute the left eigenvector and the right eigenvector of A corre-
sponding to the positive eigenvalue λ. Hint: left eigenvectors are
reciprocal basis vectors for the basis consisting of right eigenvec-
tors.
(b) For what values of (a, b) is the system observable?
(c) For what values of (a, b), if any, is this system BIBO (bounded-
input bounded-output) stable?
Explain your answers using appropriate system structural concepts.

It may help to draw pictures of left and right eigenvectors: Think
geometrically!
7.5.5 This problem is a follow-up to Exercise 5 of Chapter 5. Consider again
the LTI model
   
−3 1 0 0
ẋ =  0 −3 0 x + 1 u
0 0 4 0
y = [1 0 1] x.
Suppose now that the initial condition of the model is not known, and
that x(t) is not observed directly. Is it possible to choose a control so
that y(t) = te−3t for t > 1?
7.5. EXERCISES 137
7.5.6 The system ẋ = Ax + Bu; y = Cx is controlled using static output

feedback u = −Hy + v. Show that the resulting closed-loop system
ẋ = (A − BHC)x + Bv; y = Cx
has the same controllability/observability properties as the original

system. Hint: use the Hautus-Rosenbrock test.
7.5.7 In this problem you will see that the feedback gain K may not be
unique for multivariate models. Consider the state space model defined
by the pair of matrices
   
0 1 0 0 0
A= 0 1 1  B= 1 0  C= 1 0 0
0 0 0 0 1
(a) Find vectors wa , wb ∈ R2 such that each pair (A, Ba ), (A, Bb ) is

controllable, where
Ba = Bwa Bb = Bwb .
Try to pick the vectors so that the two 3 × 1 matrices Ba , Bb are

significantly different.
(b) Now, find two controllers of the form
ua = −wa Ka x ub = −wb Kb x
so that in each case, the closed-loop poles are placed at −1 and

−1 ± j. These designs should be performed by placing the poles
of A − Ba Ka , A − Bb Kb , respectively.
(c) Provide simultations of the step response u(t) = −Kx(t) + (1, 1)T
with zero initial conditions for the two designs. Include in your
plots both x1 (t) and the two dimensional input u(t), for 0 ≤ t ≤ 1.
7.5.8 A schematic of an active suspension system is illustrated below.

Car M
u
Suspension
u
Wheel m
x 1 + x o1
x2 + x2o
yR
Road
A state space model in scientific units is given below, where vi = x0i .

The states xi have been normalized so that their nominal values are
0.
        
x1 0 0 1 0 x1 0 0
d x
 2 = 
  0 0 0 1 x
  2 + 
    0  u +  0  yR
  
dt v1
  −10 10 −2 2   v1  3.34  0 
v2 60 −660 12 −12 v2 −20 600
(a) Obtain an open-loop simulation of x1 and x2 for 0 ≤ t ≤ 5 s., with

x1 (0) = 0.5 m; x2 (0) = 0 m; v1 (0) = v1 (0) = 0; u(t) = yR (t) ≡ 0.
(b) Repeat(a), but let yR be a square wave of amplitude 0.2 m, and
fundamental frequency of 0.2 Hz. You may take zero initial con-
ditions.
(c) Obtain a state feedback control law for a comfortable ride, but
make sure that the car does not bottom-out - Given the normal-
ized state variables, this means that x1 (t) − x2 (t) > −0.5m. You
should obtain a critically damped response for X1 (s)/YR (s), and
you should also speed up the settling time to approximately half
the open loop response.
Show simulations under the conditions of (a) and (b) above.
(d) Full state feedback is not feasible, since it is not practical to mea-
sure the distance from the car to the road. Letting y = x1 − x2
denote the actual measurement available to the controller, is the
resulting system observable?
(e) Repeat (c) with a full order observer, where the measurement is
taken to be y.
7.5. EXERCISES 139
7.5.9 The linearized (and normalized) magnetically suspended ball is de-

scribed by

0 1 0
ẋ = x+ u.
1 0 1
(a) Show that the system is unstable with u = 0.

(b) Explain why it is possible to place the poles of the system at
arbitrary locations (with the restriction of conjugate pairs) by
linear state feedback.
(c) Find a state feedback which would place the poles of the closed-
loop system at −1 ± j1.
(d) Simulate your controller for the nonlinear model ÿ = 1 − y 2 /u2 ,
with the nominal values u0 = y0 = 1, and experiment with ini-
tial conditions to find a region of asymptotic stability for the
controlled system.
7.5.10 In the previous problem, the ball position x1 can be measured using
a photocell, but the velocity x2 is more difficult to obtain. Suppose,
therefore, that the output is y = x1 .
(a) Design a full-order observer having eigenvalues at −5, −6, and

use the observer feedback to produce closed-loop eigenvalues at
−1 ± j1, −5, −6.
(b) Simulate your controller for the nonlinear model ÿ = 1 − y 2 /u2 ,
as in the previous problem, and again experiment with initial
conditions to find a region of asymptotic stability.
7.5.11 You will now construct a reduced-order observer for the previous
model.
(a) Repeat Exercise 10 using a reduced-order observer to yield closed-

loop eigenvalues at −1 ± j1 and −5.
(b) Letting z = ẏ in (a), compute the transform Ẑ(s), and show that it
is approximately equal to sY (s) when the observer gain is large.
7.5.12 Below is a drawing of a cart of mass M with a uniform stick of mass

m pivoted on top:
In appropriate dimensionless units, the equations of motion may be

written as
θ̈ = θ + u , ẍ = −βθ − u,
where β := 34 [m/(M + m)] is a parameter of the system, and u is the

torque applied to the wheels of the cart by an electric motor. We
wish to find a linear feedback control that will balance the stick (i.e.,
keep θ ≈ 0) and keep the cart near x = 0. To do this, find the gains
k1 , k2 , k3 and k4 in the state-variable feedback
u = k1 θ + k2 θ̇ + k3 x + k4 ẋ
such that the closed-loop system has a double pole at s = −1 and a

pair of complex poles at s = −1 ± j1
7.5.13 On the line connecting the center of the earth to the center of the
moon, there is a so-called libration point where the pull of the earth
on a satellite (in an orbit about the earth with the same period as the
moon’s orbit) exactly equals the pull of the moon plus the centrifugal
force. The dynamic equations for small deviations in position away
from the libration point can be shown to be:
ẍ − 2ω ẏ − 9ω 2 x = 0
ÿ + 2ω ẋ + 4ω 2 y = u
where x:= radial position perturbation, y:= azimuthal position pertur-

bation, u = F/mω 2 (control exerted by a small reaction engine), F :=
engine thrust in the y direction, m:= satellite mass, and ω := 2π/29
rad/day.
(a) With u = 0, show that the equilibrium point x ≡ y ≡ 0 is unstable.

7.5. EXERCISES 141
(b) To stabilize the position, one can use state feedback
u = k1 x + k2 ẋ + k3 y + k4 ẏ.
Show that it is possible to stabilize the system with a control

law of this form. Determine the constants k1 , ..., k4 such that
the closed-loop system has poles at s = −3ω, s = −4ω, and
s = (−3 ± j3)ω.
7.5.14 Given the LTI system: ẋ = Ax + Bu, y = Cx, where

   
0 0 −2 0 β
A = 1 0 1  , B = 1 0  , C = 0 0 1 .
0 1 2 0 1
(a) Determine the range of values of the scalar real parameter β for
which we can place 2 poles arbitrarily, using static output feed-
back (as usual, we assume that if the two poles are complex then
they should be in conjugate pairs).
(b) for β = 1, determine static output feedback so that two of the
eigenvalues of the closed-loop system will be λ1,2 = −1 ± j1.
7.5.15 Consider the linear system with transfer function
s + 28
P (s) =
(s + 27)(s + 29)
We would like to find a control law which places the poles at (−29, −28),
so that the zero is canceled.
(a) Place the system in modal canonical form, and apply the feedback
control law u = −Km x + r. Compute Km for the desired closed-
loop pole locations.
(b) Place the system in observable canonical form, and apply the feed-
back control law u = −Ko x + r. Compute Ko for the desired
closed-loop pole locations.
(c) Why do you think the gains are larger in the second case.
7.5.16 Design a first-order controller of the form
(s + a)
G(s) = K
(s + b)
(K, a, b free parameters) for the plant with transfer function

s + 28
P (s) =
(s + 27)(s + 29)
to place poles at (−30.5, −31, −31.5). Repeat this for poles at (−31, −32, −33).
Note that the two controllers are quite different, even though the
closed-loop poles differ only slightly.
7.5.17 Design a first-order controller of the form
(s + a)
G(s) = K
(s + b)
(K, a, b free parameters) for the plant with transfer function
s−α
P (s) =
(s + 4)
to place poles at −10 ± 3j, where α > 0 is a scalar parameter. Plot
the magnitude of the sensitivity function S = P1G for several values of
α ranging from 1 to 100. For what values of α is the controlled system
most sensitive to plant uncertainty?
7.5.18 Below is a diagram of a flexible structure found in the undergraduate
controls lab at the University of Illinois
θ2
θ1
τ
+
u(t)
-
An accurate state space model can be found, which is defined by the

following matrices:
   
0 1 0 0 0
−500.3513 −2.626 500.3513 0.3972  299.535
A= ; B= 
 0 0 0 1   0 
558.964 0 −558.964 −2.1267 0
(7.10)
The input is the voltage to a DC motor, and the states of this model
are
x1 = θ1 , x2 = θ̇1 , x3 = θ2 , x4 = θ̇2 .
7.5. EXERCISES 143
(a) Make a Simulink model for this system and plot the open-loop
response θ2 (t) to a non-zero initial condition with θ1 (0) 6= θ2 (0).
(b) Design a full state feedback controller u = −Kx + gr to place the
poles in the region −10 ≤ σ ≤ −5, and choose the gain g so that
the DC gain of the closed-loop system from r to θ2 is one. Plot
step responses for your controlled system, with r = π/2 radians.
In one step response, use zero initial conditions, in the other,
choose the same initial conditions as in (a). Design K so that the
overshoot is less than 20%.
(c) Make a Simulink model for an observer, assuming only θ1 is avail-
able through direct measurements. Combine the observer, your
full state feedback controller, and plant as on page 129 of the
lecture notes. After designing appropriate observer gains, obtain
two step responses of the controlled system, as in (b).
(d) Repeat (c) with a reduced order observer, assuming only θ1 is
directly measured.
Chapter 8
Tracking and Disturbance

Rejection
8.1 Internal model principle

Tracking and disturbance rejection are two of the basic goals in control
design. This chapter addresses each of these issues by applying the state
space theory developed so far.
Assume that one wishes to make the output y track a constant reference
input r. One approach is to modify the state feedback control law
u = −Kx + N r
where N is chosen based upon the DC gain of the system. Using the final
value theorem
y(∞) = lim sY (s)

s→0
a formula for N is easily obtained. This is a reasonable approach if the DC

gain of the plant is known. However, the scaling matrix N does not account
for disturbances and parameter variations that will hamper tracking.
To obtain a solution which is robust to DC disturbances we explicitly
model the unknown constant disturbance as follows:
ẋ = Ax + Bu + Ew
y = Cx + F w,
where w ∈ Rn is a constant. Our goal then is to make e(t) = y(t) − r → 0,

as t → ∞, regardless of the value of the disturbance w, or the reference
145
146 CHAPTER 8. TRACKING AND DISTURBANCE REJECTION
input r. The approach taken allows us to achieve these three objectives

simultaneously:
(a) Asymptotic tracking: y(t) → r as t → ∞.

(b) Complete insensitivity to w.
(c) Tuned transient response through pole placement.

r
Conceptually, the approach taken is to note that the exogenous signal
w
is generated by a state space model
ż = Am z z(0) ∈ Rq ;

r
= Cm z
w
where in this special case of constant disturbances and a constant reference

input, the matrix Am is equal to zero. This reference/disturbance model is
simply an integrator, and based upon this we incorporate an integrator in
the control law. Thus, the control law we adopt is of the general form
u = −K1 x − K2 η, η̇ = e = y − r.
This procedure is known as the internal model principle.

The internal model principle is most easily understood when viewed
in the frequency domain. The controller transfer function (from x to u)
possesses a pole at the origin, which in the frequency domain is equivalent
to demanding infinite gain at DC. A look at the sensitivity function then
shows that sensitivity with respect to plant uncertainty or disturbances at
DC will be zero, provided the plant itself does not possess a zero at DC.
However in this course we remain in the time domain, and so our approach
is to show that a desired equilibrium is asymptotically stable.
First note that since the external inputs r and w are constant, assum-
ing stability one can expect that all signals will converge to some constant
values, which will form an equilibrium for the controlled system. If the
integrated error η converges in this sense, then the error itself e will neces-
sarily converge to ϑ. To make this precise, consider the closed-loop system
equations, given by

ẋ A − BK1 −BK2 x 0 E
= + r+ w.
η̇ C 0 η −I F
8.1. INTERNAL MODEL PRINCIPLE 147

ẋ
At an equilibrium we have = ϑ, from which we conclude that
η̇
ϑ = η̇ = Cx − r + F w = y − r.
Thus, we will have asymptotic disturbance rejection and asymptotic tracking

if the controlled system is stable. To determine if this is the case, we must
consider the eigenvalues of the matrix

A − BK1 −BK2
Acl = .
C 0
Writing the closed-loop matrix as

A − BK1 −BK2 A 0 B
= − [K K ]
C 0 C 0 0 | 1{z 2}
K
| {z } | {z } e
A
e B
e
we conclude that we can achieve arbitrary pole placement if and only if the
pair (A, e is controllable. If this is the case then the two gains (K1 , K2 ) =
e B)
K can be designed simultaneously using a single Matlab command.
e
We use the Hautus-Rosenbrock test to better understand the control-
lablility of (A, e We have for any complex s,
e B).

sI − A 0 B
rank (sI − A | B) = rank
e e .
−C sI 0
For controllability, we must have for all s,

sI − A 0 B
rank = n + p. (8.1)
−C sI 0
That is, all (n + p) rows must be linearly independent.
First suppose that s 6= 0. Then the bottom p rows (given by [−C sI |
0]) are linearly independent because of the presence of the rank p matrix
sI. The remaining upper n rows will be linearly independent if (A, B) is
controllable (the 0 term does not affect the rank of these rows). Thus if
(A, B) is controllable, the rank condition is met for s 6= 0. For s = 0, the
rank condition (8.1) becomes

−A 0 B
rank = n + p.
−C 0 0
To summarize, the following two conditions are equivalent to controllability
of the pair (A,
e B):
e
(a) (A, B) is controllable

−A B
(b) rank = n + p.
−C 0
For m ≥ p, condition (b) will be true if and only if no transmission zero

exists at s = s0 = 0 (see page 162).
Note that if the rank condition (b) fails then there exists a non-zero pair
(x0 , u0 ) ∈ Rn+p such that

A B x0 ϑ
= .
C 0 u0 ϑ
If the constant control u ≡ u0 is applied when the initial condition is x0

then we have
ẋ = Ax0 + Bu0 = ϑ.
That is, the state x0 is an equilibrium when this constant input is applied.
The output equation is y = Cx0 = ϑ, so that the initial state/input signal
is invisible to the output. It is this complete lack of response to a DC input
that must be ruled out if integral control, or any other approach to DC
tracking is to be successful.
If the states are not available, we can still use an observer. If there are
disturbances then the estimates may not converge, but one can still show
that the unique equilibrium will satisfy e = 0. In Exercise 3 you will see
how the estimator behaves in the presence of a disturbance.
Example 8.1.1. Consider the first order model
ẋ = −2x + u + w
y = x.
To regulate this system to track a constant reference r, the control law will
be of the form illustrated below.
8.1. INTERNAL MODEL PRINCIPLE 149
Plant
w
r - - u y
∑ k2 ∑ ∑
- -
k1
To place the closed-loop poles we consider the closed-loop system matrix

−2 − k1 −k2
Acl = A − B K =
e e e .
1 0
For example, to place the closed-loop poles at −2 ± j2 we set
det[sI − (A
e−B e = s2 + (2 + k1 )s + k2 = s2 + 4s + 8.
e K)]
This gives k1 = 2, k2 = 8, so that we obtain the “PI” controller

Z t
u = −2y − 8 e dt.
0
t
u
The internal model principle can be extended to the more general sit-
uation where the disturbance and reference signal are the sum of periodic
signals of known frequency. If this is the case, then we may again find a
matrix Am such that
ż = Am z z(0) ∈ Rq ;

r
= Cm z,
w
where the eigenvalues of the matrix Am lie on the jω-axis. To design a
controller one replicates this model to define the signal η:
ẇ = Am w + B1 e w(0) ∈ Rq ;
η = C1 w.
As before, the control can then be defined as u = −K1 x − K2 η.
8.2 Transfer function approach

When we use an observer together with state feedback to control a plant,
we are actually designing a form of dynamic feedback compensation, similar
to what is obtained through loop shaping in a frequency domain design. We
show here that equivalent dynamic compensation may be conducted entirely
in the Laplace transform domain by manipulating transfer functions. We
concern ourselves only with the SISO case, and for simplicity we treat only
the regulator problem. Integral control or reduced order observer based
feedback can be analyzed in the transform domain in the same manner.
Controller: U(s) = - G(s) Y(s) Plant

U Y
P(s)
Observer
-K O(s)
Figure 8.1: The observer/state feedback combination can be regarded as

dynamic compensation
The regulator problem using a full order observer is illustrated in Fig-

ure 8.1. To compute the controller transfer function G(s), we first construct
the transfer function from Y to X̂. The state estimates are described by the
equation
b˙ = Ab
x x + Bu + L(y − C x
b) = (A − BK − LC)b
x + Ly
where the second equation follows from the control law
u = −K x
b.
Taking transforms, we see that
U (s) = −K X̂(s) = −K(sI − [A − BK − LC])−1 LY (s).
This results in the formula
G(s) = K(sI − [A − BK − LC])−1 L. (8.2)

8.2. TRANSFER FUNCTION APPROACH 151
Since there is no direct feedthrough term (no “D” matrix) in the formula
(8.2), the transfer function G is strictly proper.
This transfer function can be directly designed, without consideration of
a state space model. To see this, first write the plant and controller transfer
functions as a ratio of polynomials:
b(s)
Plant transfer function = P (s) =
a(s)
n(s)
Controller transfer function = G(s) = .
d(s)
If the control U (s) = −G(s)Y (s) + R(s) is applied, then the closed loop
transfer function becomes
b(s)
Y (s) a(s) b(s)d(s)
Pcl (s) = = n(s) b(s)
= .
R(s) 1 + d(s) a(s) a(s)d(s) + b(s)n(s)
The poles of the closed-loop system are determined by the corresponding

characteristic equation
∆cl (s) = a(s)d(s) + b(s)n(s) = 0.
Thus, if we let αc , αo denote the polynomials which define the state feedback
poles, and observer poles, respectively:

αc (s) = det(sI − (A − BK))
= known,
αo (s) = det(sI − (A − LC))
then the polynomials n(s), d(s) must satisfy the following Diophantine equa-
tion:
a(s)d(s) + b(s)n(s) = αc (s)αo (s). (8.3)
We assume that both a(s) and d(s) are monic, so that each of these
polynomials has degree n. We have already noted that G is strictly proper,
which means that the degree of n(s) is strictly less than n. The product
αc (s)αo (s) should also be monic. With this information, the unknown poly-
nomials d(s) and n(s) can be computed. However, some conditions must
be imposed. When we treated this problem using state space arguments
we found that the system must be controllable and observable to compute
the gains K and L. This is equivalent to minimality, which means that the
transfer function P (s) = C(Is − A)−1 B cannot have any common pole-zero
pairs, or that the polynomials a and b do not have common roots.
Example 8.2.1. Consider the transfer function description of the magnet-

ically suspended ball
1
P (s) = .
s2 −1
There can be no pole zero cancellations since b(s) = 1, and hence any two
dimensional state space model is minimal. We conclude that the degree of
n(s) is less than or equal to 1, and the degree of d(s) is equal to 2. So, the
controller can be written
n1 s + n0
U (s) = Y (s)
s2 + d1 s + d0
From equation (8.3) we must solve the following equation to place the two
state feedback poles, and the two observer poles:
(s2 − 1)(s2 + d1 s + d0 ) + (1)(n1 s + n0 ) = (s − p1 )(s − p2 )(s − `1 )(s − `2 ),
which gives four equations and four unknowns through equating of the co-
efficients. These equations are called the Sylvester equations. t
u
Based on our criteria for solving the Diophantine equation, it follows

that for the Sylvester equations to be independent and solvable for any set
of observer poles and state feedback poles, the plant transfer function can-
not have common pole-zero pairs. For example, consider the plant transfer
function
s+1
P (s) = .
(s + 1)(s + 2)
The corresponding Diophantine equation is
(s + 1)(s + 2)d(s) + (s + 1)n(s) = αc (s)αo (s)
Since the left hand side of this equation vanishes when s = −1, obviously
this equation cannot be solved if neither αc nor αo have a root at −1. Pole
placement is not possible in this example because the transfer function P
does contain a common pole-zero pair.
8.3. EXERCISES 153
8.3 Exercises
8.3.1 Consider the servo motor with transfer function
1
Y (s) = U (s),
s(1 + s)
where u is the voltage input to the motor, and y is its position. In
this exercise you will compare closed-loop sensitivity with and without
integral control.
(a) Design a first order dynamic compensator
s−z
U (s) = k (gR(s) − Y (s))
s−p
so that the closed-loop poles are at −3, −2 ± j. Choose g so that
the closed-loop transfer function Y /R has unity gain at DC.
(b) Plot the step response, and a Bode plot of the sensitivity transfer
function S1 .
8.3.2 Consider the LTI system
ẋ1 = x2
y = x1 , (8.4)
ẋ2 = u + w
where w is a constant but unknown disturbance. Rather than intro-
duce integral control, if we could obtain estimates of w, then we could
design a control law to cancel the effect of this disturbance.
One way to attempt to do just this is to consider x3 = w as an addi-
tional state variable:
ẋ1 = x2
ẋ2 = u + x3 y = x1 .
ẋ3 = 0
(a) Carefully show that this third-order “augmented” system is ob-
servable.
(b) Design a full-order observer for this system. Draw the correspond-
ing all-integrator diagram of the observer in full detail, including
all integrators.
(c) Based upon this observer, design a controller for this system to
place the closed-loop poles at −1, −2, −3, −4, −5. Explain care-
fully why the poles will be placed, and the state estimates will
converge.
(d) Is it possible to design the observer-based feedback control law in

(c) so that the closed-loop system matrix is Hurwitz?
8.3.3 Consider again the LTI system given in (8.4).
(a) Design a feedback compensator of the form u = −K1 x b − K2 η to

regulate the output to zero. Take η̇ = y, and x b2 as an estimate
of x2 based on a reduced order observer, assuming w ≡ 0.
(b) With this controller, compute x b∞
2 = limt→∞ x
b2 (t), as a function of
w. Note: you do not need to compute the entire transfer function.
(c) Provide simulations of your controller using Simulink for various
initial conditions, and values of w.
8.3.4 An electromagnetic suspension system for a magnetic ball is modeled

by
mÿ = mg − ci2 /y 2
where m is the mass of the ball (= 10−3 kg.), g = 9.8m./sec2 , c =

9.8 × 10−9 newton-meter2 /amp2 , y is the distance between the magnet
and the ball, and i is the current through the electromagnet. Suppose
that transducers are available for measuring x1 = y, x2 = ẏ. It is
desired that in the steady state, x1 → v, where v is a constant reference
input.
To accomplish this, it is proposed to feed back the integral of the error
x1 − v in addition to the states x1 and x2 . That is, the control i is to
be implemented as
i = k0 x0 + k1 x1 + k2 x2
where x0 is the integral of x1 − v, and k0 , k1 , k2 are constant gains,

chosen by the designer as k0 = 100, k1 = 1200, k2 = 1. Find the
equilibria of the closed-loop system when v = 0.025 m, and determine
whether or not they are asymptotically stable.
Chapter 9
Control Design Goals
9.1 Performance
Control can be broadly defined as the use of feedback to obtain satisfactory
performance in the presence of uncertainty. Feedback is simply the act of
incorporating measurements as they become available to define the inputs to
the plant. The definition of performance depends of course on the nature of
the particular control application, but there are some universal performance
issues. Among these are:
(i) Stability of the controlled system is of course the most basic goal. Any-
one who has held a microphone too close to its speaker system knows
that feedback can result in instability, but when properly designed,
feedback can prevent unstable behavior.
(ii) Frequently one desires reasonably accurate tracking of some desired tra-
jectory. If the trajectory is a constant reference value, this is known as
the regulation problem. In the production of chemicals, particularly in
the pharmacuetical industry, yields must be controlled within stringent
specifications. In the automatic control of an aircraft in flight, again
one objective is to keep the flight at a level altitude and a constant
speed.
(iii) Another time domain specification concerns speed of response. For

example, if the pilot of an aircraft desires a change in altitude, this
should occur quickly, but not at the expense of the comfort of the
passengers.
155
156 CHAPTER 9. CONTROL DESIGN GOALS
(iv) Rarely do we have available a perfect model of the plant to be con-

trolled. Even if an accurate model is available, it may be convenient
to linearize the nonlinear model, which can introduce significant in-
accuracies. Such inaccuracies combined with unforeseen influences on
the plant are collectively known as disturbances, and we must always
consider robustness to such disturbances when designing a control sys-
tem.
All of these issues are typically addressed in the design of a control system in
the face of physical and economic constraints, and in most current industries,
time constraints as well!
One cannot expect good performance if the physical system is poorly
designed. In particular, it is important from the start that the measurement
devices be reasonably precise, and that actuation - the final synthesis of the
control action - be performed accurately. However, in this course we assume
that these issues were considered in the construction of the plant, so that
what is left for the control engineer can be largely completed with a pen,
paper, and a computer.
9.2 Measurements
This course concerns state space models, which in many of our examples are
naturally constructed based upon signals which have physical significance,
and which can be measured if sensors are adjoined to the plant. For instance,
in the pendubot illustrated in Figure 1.3, the states are the link angles, and
their velocities. The link angles can be measured with high precision using
an optical encoder. The velocities are more difficult to measure, but they
can be estimated using finite differences. If, as in this example, an accurate
state space model is available together with accurate state measurements,
then control based upon pole placement or optimal control is frequently
simple and very effective.
However, there are times when full state information is not feasible. For
instance, in many applications the state space model is constructed by first
finding a frequency domain model (i.e. a Bode plot) based upon input-output
measurements. A state space model can then be constructed so that its fre-
quency response accurately matches this empirical Bode plot. In this case,
the state has no physical significance, and the only information available
for controlling the system may be the input and output. In Section 7.2 we
circumvent this difficulty by constructing observers to estimate the state as
input-output measurements are collected. In principle at least, one can then
9.3. ROBUSTNESS AND SENSITIVITY 157
treat the estimates as if they are perfect measurements of the state, and then
apply one of the state feedback control techniques. This gives a mathemat-
ically elegant approach to control design which is also easily analyzed. In
particular, one can easily verify that the goals (i)–(iii) outlined above will be
met if the observer and state feedback controller are appropriately designed.
The fourth goal, robustness, can be easily forgotten in a state space de-
sign, and this can lead to absurd control designs! To see this, consider once
again the pendubot shown in Figure 1.3. With the input u taken as the
torque applied to the lower link, and with the output y taken to be the
lower link angle, the linearized plant in the vertical position is both control-
lable and observable. Theoretically then, it is easy to design a controller
which will keep the pendubot balanced in the vertical position based upon
measurements of the lower angle alone. Physically, this seems very unlikely,
and in fact such a controller has never been successfully implemented on this
plant. The reason for this divergence between theory and practice is that
when evaluating a controller one must also consider robustness to plant un-
certainty. Using a frequency domain analysis, one can more easily see that
any controller for this plant based upon the measurements u, y alone will
be highly sensitive to model inaccuracy. An insensitive design is possible if
both link angles are used in the feedback compensator.
9.3 Robustness and sensitivity

The Bode sensitivity function is a measure of the sensitivity of the closed-
loop transfer function to a deviation in the plant model. For the control
system illustrated in Figure 9.1, the closed-loop transfer function may be
expressed
Y GP
T = =
R 1 + GP
If the transfer function P changes to form a new transfer function P 0 , the
controller will not function precisely as it was intended, since the transfer
function T will undergo a corresponding change. Rather than consider ab-
solute changes, it is more appropriate to analyze the affect of the relative
change in the plant
∆P
,
P
where ∆P = P 0 − P denotes the difference between the true transfer func-
tion and the nominal one. For small changes, this can be interpreted as
a differential. The sensitivity is then the ratio of this differential, and the
corresponding differential in T :
dT /T P dT
S := = . (9.1)
dP/P T dP
For the error feedback configuration given in Figure 9.1, the derivative can
be computed to give
1 1
S= =
1 + PG 1+L
where L(s) is the loop gain P (s)G(s). The quantity 1 + L(s) is known as
the return difference.
In the MIMO case we can again write for the error feedback configuration
Y = (I + P G)−1 P G R,

and the sensitivity function is then defined to be the matrix-valued function

of ω,
S(jω) = (I + P (jω)G(jω))−1 .
Controller Plant
R + U Y
∑ G(s) P(s)
-
Figure 9.1: An error feedback configuration
We can conclude from this formula that if we want a system which is

insensitive to changes in the plant, then we should make the loop gain L(s)
as large as possible. For SISO systems the size of the sensitivity function is
measured by means of the magnitude of its frequency response. We might
require for example that
|S(jω)| < 1, ω ≥ 0,
which does correspond to a “large” loop gain. In making the sensitivity

small however, we must keep all of the previous design goals in mind, and
9.4. ZEROS AND SENSITIVITY 159
there are also other constraints which are absolute. For instance, for the
error feedback configuration we will always have
S + T = 1.
Moreover, if the transfer function L is relative degree at least two, and all
poles lie in the left half plane in C, we have the constraint
Z ∞
log |S(jω)|dω = 0.
0
In this case it is impossible to make |S(jω)| < 1 for all ω, since then
log |S(jω)| is always negative.
9.4 Zeros and sensitivity

It is clear that the structure of the open-loop poles will play an important
role in the design of an effective controller. What is less obvious at first
glance is the importance of the open-loop zeros.
To illustrate the importance of zeros, we again take the pendubot intro-
duced in Chapter 1. The pendubot can be linearized in the upright position
illustrated in Figure 1.3. With the input u equal to the applied torque, and
the output y equal to the lower link angle, the resulting state space model
is defined by the following set of matrices:
   
0 1.0000 0 0 0
 51.9243 0 −13.9700 0  B =  15.9549 
 
A =   0 0 0 1.0000  0 
−52.8376 068.4187 0 0 −29.3596

C = 1 0 0 0 D = 0. (9.2)
Using Matlab, the corresponding transfer function model is found to be
Y (s) (s − 6.5354)(s + 6.5354)

P (s) = = 15.9549
U (s) (s − 9.4109)(s + 9.4109)(s − 5.6372)(s + 5.6372)
This is of the general form
(s − γ)(s + γ)
P (s) = k ,
(s − α)(s + α)(s − β)(s + β)
where 0 < α < γ < β.

Because P contains no common pole-zero pairs, the state space model

must be both controllable and observable, i.e., minimal. In this chapter
we will show that it is therefore possible to design a fourth order dynamic
compensator U (s) = −G(s)Y (s) which stabilizes this model at x = ϑ. We
will find in Chapter 10 that a set of closed-loop poles which will perform
well using full state feedback are
p1 = −36 p2 = −2.9 p3 = −5.3 + 0.84j p4 = −5.3 − 0.84j. (9.3)
The compensation will introduce four poles, which we will place at
p5 = −3.9932 p6 = −7.4167 p7 = −8.4862 p8 = −11.3012.
All of these poles are based upon an “optimal” design. By examining a

root locus plot it may be shown that any compensator which stabilizes this
system will have poles in the right half plane in C. In this example, we find
that the compensator which places these poles is given by
(s + 9.4)(s + 5.64)(s − 7.41)

G(s) = 24, 424 .
(s + 60.24 + 63.5j)(s + 60.24 − 63.5j)(s − 46.37)(s + 6.54)
As expected, this compensator has an unstable pole at s = 46.37.

A Nyquist plot of the loop transfer function L(s) = G(s)P (s) is given in
Figure 9.2. If the plant is changed slightly, then the number of encirclements
of −1 will change, resulting in an unstable closed-loop system. A Bode plot
of S = 1/(1 + GP ) is shown in Figure 9.3. Here it is seen that the controlled
system is highly sensitive to plant uncertainty in the low frequency range.
By applying higher feedback gain, it is possible to reduce sensitivity in this
range, but then one will find high sensitivity at higher frequencies.
One way to explain these difficulties is the presence of zeros in the right
half plane, which impose strong limits on achievable sensitivity. For the zero
free model with transfer function
1
P (s) = ,
(s − α)(s + α)(s − β)(s + β)
it is is easy to obtain a low sensitivity design. Another explanation is again

common sense - it is difficult to imagine that this plant can be controlled
using only information at the lower link. By adding another sensor, an
insensitive design may be obtained using state space methods.
For a SISO transfer function P (s) = b(s)/a(s), the poles are the roots
of a, and the zeros are the roots of b, so that we have P (s0 ) = 0 for a
9.4. ZEROS AND SENSITIVITY 161
-1
Imag Axis
ω >0 0.2
15
Real Axis
Figure 9.2: The Nyquist plot of G(jω)P (jω). The compensated open-loop
system with transfer function GP possesses two poles in the right half plane.
The three encirclements of −1 ensure that the controlled system is asymp-
totically stable, from the Nyquist Stability Criterion, but this stability is
very fragile.
zero s0 . In the MIMO case, we define the characteristic polynomial as

∆(s) = det(Is − A), and the roots of ∆ are then the poles of P . We define
zeros similarly, but it would be far too strong to suppose that P (s0 ) = ϑ at
a zero s0 . For the state space model
ẋ = Ax + Bu, y = Cx, x0 = ϑ
we can take Laplace transforms to obtain
(sI − A)X(s) = BU (s), Y (s) = Cx(s).
Combining these equations in one matrix expression gives

sI − A −B X(s) ϑ
= .
C ϑ U (s) Y (s)
| {z }
(n+p)×(n+m)
Assume that there are as many plant inputs as outputs (p = m), so that the
plant is square. A natural definition of a zero can be formulated as follows:
60
40
Gain dB
20
0 -1 0 1 2
10 10 10 10
Frequency (rad/sec)
Figure 9.3: The magnitude plot of the frequency response for the sensitivity
function S = 1/(1 + GP ). The controlled system is highly sensitive to plant
uncertainty at low frequencies.
At s0 , can I find a fixed U (s0 ) and X(s0 ) so that the resulting Y (s0 ) is zero?
That is,

s0 I − A −B X(s0 ) ϑ
= .
C 0 U (s0 ) ϑ
This is possible if and only if the matrix above is rank deficient:

s0 I − A −B
ρ < n + m.
C 0
Provided that s0 is not also a pole of P , this is equivalent to rank deficiency

of the transfer function P at s0 :
ρ (P (s0 )) = ρ C(Is0 − A)−1 B < m,

which does generalize the definition given in the SISO case. The complex
number s0 is called a transmission zero. If all of the transmission zeros lie
within the open left half plane in C then the model is called minimum phase.
In the multivariate case, the location of the transmission zeros can play an
important role in the achievable sensitivity of the controlled plant, just as
in the SISO case.
9.5. EXERCISES 163
9.5 Exercises
9.5.1 For the Pendubot described by the state space model (9.2), you will
design two controllers. The design can be done using the PLACE com-
mands within Matlab, and the formula (8.2).
(a) Design a fifth order dynamic compensator U = −GY of the form
n(s)
G(s) = ,
sd(s)
where d is a fourth order, monic polynomial. The dominant closed

loop poles should lie to the left of the line Re(s) = −10 in the
complex plane. Obtain a Bode plot of the sensitivity function for
your design.
(b) You will now design a compensator based on the two angle mea-
surements. We then have
n1 (s) n2 (s)
U (s) = Y1 (s) + Y2 (s)
sd1 (s) sd2 (s)
where both d1 and d2 are fourth order polynomials. See if you

can find an observer based dynamic feedback controller of this
form which has significantly improved sensitivity. Obtain a Bode
plot of the sensitivity function for your design.
Part IV
Optimal Control
165
Chapter 10
Dynamic Programming and

the HJB Equation
10.1 Problem formulation

This chapter concerns optimal control of dynamical systems. Most of this
development concerns linear models with a particularly simple notion of
optimality. However, to understand the most basic concepts in optimal
control, and not become lost in complex notation, it is most convenient to
consider first the general model given in nonlinear state space form
ẋ = f (x, u, t), x(t0 ) = x0 . (10.1)
The cost V of a particular control input u is defined by the following integral

Z t1
V (u) = `(x, u, τ ) dτ + m(x(t1 )) (10.2)
t0
where
(a) t1 is the final time of the control problem.

(b) ` is a scalar-valued function of x, u, and t
(c) m is a function of x. It is called the terminal penalty function.
We assume that x0 , t0 , and t1 are known, fixed values, and x(t1 ) is free. Our
goal is to choose the control u[t0 ,t1 ] to minimize V .
A case which is typical in applications is where m and ` are quadratic
functions of their arguments. For an LTV model, the system description
167
168CHAPTER 10. DYNAMIC PROGRAMMING AND THE HJB EQUATION
and cost are then given by
ẋ = A(t)x + B(t)u, x(t0 ) = x0

Z t1
(10.3)
V (u) = (xT Q(t)x + uT R(t)u) dt + xT (t1 )M x(t1 )
t0
where M , Q and R are positive semidefinite matrix-valued functions of time.

These matrices can be chosen by the designer to obtain desirable closed-loop
response. The minimization of the quadratic cost V for a linear system is
known as the linear quadratic regulator (LQR) problem.
We will study the LQR problem in detail, but first we develop some
general results for the nonlinear state space model (10.1).
10.2 Hamilton-Jacobi-Bellman equations

The value function V ◦ = V ◦ (x0 , t0 ) is defined to be the minimum value of
V over all controls. This is a function of the two variables x and t which
can be written explicitly as
hZ t1 i
V ◦ (x, t) = min `(x(τ ), u(τ ), τ ) dτ + m(x(t1 )) .
u[t,t1 ] t
Under very general conditions, the value function satisfies a partial differ-
ential equation (PDE) known as the Hamilton-Jacobi-Bellman (HJB) equa-
tion. To derive this result, let x and t be an arbitrary initial time and initial
state, and let tm be an intermediate time, t < tm < t1 . Assuming that x(τ ),
t ≤ τ ≤ t1 , is a solution to the state equations with x(t) = x, we must have
hZ tm Z t1 i
◦
V (x, t) = min `(x(τ ), u(τ ), τ ) dτ + `(x(τ ), u(τ ), τ ) dτ + m(x(t1 ))
u[t,t1 ] t tm
hZ tm Z t1 i
= min `(x(τ ), u(τ ), τ ) dτ + min `(x(τ ), u(τ ), τ ) dτ + m(x(t1 )) .
u[t,tm ] t u[t,t1 ] tm
| {z }
V ◦ (x(tm ),tm )
This gives the functional equation

hZ tm i
V ◦ (x, t) = min `(x(τ ), u(τ ), τ ) dτ + V ◦ (x(tm ), tm ) . (10.4)
u[t,tm ] t
As a consequence, the optimal control over the whole interval has the prop-
erty illustrated in Figure 10.1: If the optimal trajectory passes through the
10.2. HAMILTON-JACOBI-BELLMAN EQUATIONS 169
state xm at time x(tm ) using the control u◦ , then the control u◦[tm ,t1 ] must
be optimal for the system starting at xm at time tm . If a better u∗ existed
on [tm , t1 ], we would have chosen it. This concept is called the principle of
optimality.
Optimal trajectory starting from x at time t
Optimal trajectory starting from xm at time t m
x(t m ) = xm x(t1 )
x(t)
Figure 10.1: If a better control existed on [tm , t1 ], we would have chosen it.
By letting tm approach t, we can derive a partial differential equation for

the value function V ◦ . Let ∆t denote a small positive number, and define
tm = t + ∆t
xm = x(tm ) = x(t + ∆t) = x(t) + ∆x.
Assuming that the value function is sufficiently smooth, we may perform a

Taylor series expansion on V ◦ using the optimality equation (10.4) to obtain
∂V ◦ ∂V ◦

◦ ◦
V (x, t) = min `(x(t), u(t), t)∆t + V (x, t) + (x(t), t)∆x + (x(t), t)∆t
u[t,tm ] ∂x ∂t
+p(∆t).
Dividing through by ∆t and recalling that x(t) = x then gives
∆t ∂V ◦ ∆x ∂V ◦

∆t
0 = min `(x, u(t), t) + (x, t) + (x, t) .
u[t,tm ] ∆t ∂x ∆t ∂t ∆t
Letting ∆t → 0, the ratio ∆x/∆t can be replaced by a derivative to give
∂V ◦ ∂V ◦

0 = min `(x, u, t) + (x, t)ẋ(t) +
u ∂x ∂t
where
∂V ◦ ∂V ◦ ∂V ◦

= ... = (∇x V ◦ )T .
∂x ∂x1 ∂xn
Thus, we have obtained the following partial differential equation which the
value function must satisfy if it is smooth. The resulting equation (10.5) is
the Hamilton-Jacobi-Bellman (HJB) equation,
∂V ◦ ∂V ◦

− (x, t) = min `(x, u, t) + (x, t)f (x, u, t) . (10.5)
∂t u ∂x
The terminal penalty term gives a boundary condition for the HJB equation
V ◦ (x(t1 ), t1 ) = m(x(t1 )). (10.6)
The term in brackets in (10.5) is called the Hamiltonian, and is denoted
H:
H(x, p, u, t) := `(x, u, t) + pT f (x, u, t), (10.7)
where p = ∇x V ◦ . We thus arrive at the following
Theorem 10.1. If the value function V ◦ has continuous partial derivatives,
then it satisfies the following partial differential equation
∂V ◦
− (x, t) = min H(x, ∇x V ◦ (x, t), u, t),
∂t u
and the optimal control u◦ (t) and corresponding state trajectory x◦ (t) must
satisfy
min H(x◦ (t), ∇x V ◦ (x◦ (t), t), u, t) = H(x◦ (t), ∇x V ◦ (x◦ (t), t), u◦ (t), t).
u
(10.8)
Conversely, if there exists a value function V ◦ (x, t) solving (10.5) subject to
(10.6), and a control u◦ obtained from (10.8), then u◦ is the optimal con-
troller minimizing (10.2) for the system (10.1) and V ◦ (x0 , t0 ) is the mini-
mum value of (10.2). t
u
An important consequence of Theorem 10.1 is that the optimal control
can be written in state feedback form u◦ (t) = ū(x◦ (t), t), where the function
ū is defined through the minimization in (10.8).
Proof Proof of Theorem 10.2.1 The first part of the statement of the
theorem has already been proved through the derivation that led to (10.5).
For the second part of the statement (the sufficiently part), let us first note
that for an arbitrary control u, from (10.5):
∂V ◦ ∂V ◦
− (x, t) = `(x, u◦ , t) + (x, t)f (x, u◦ , t)
∂t ∂x
∂V ◦
≤ `(x, u, t) + (x, t)f (x, u, t)
∂x
10.2. HAMILTON-JACOBI-BELLMAN EQUATIONS 171
which is equivalent to
0 = `(x, u◦ , t) + d ◦
dt V (x, y) ≤ `(x, u, t) + d ◦
dt V (x, t) (10.9)
where the total derivative (with respect to t) is evaluated on the trajectory

generated by (10.1), with u = u◦ on the left hand-side and arbitrary u on
the right hand-side. Now integrate both side of the inequality in (10.9) from
t0 to t1 , the left hand-side over the trajectory of (10.1) with u = u◦ , and
the right hand-side also over the trajectory of (10.1) but for an arbitrary u.
Further use (10.6) to arrive at
Z t1
◦ ◦
V (x0 , t0 ) = m(x (t1 )) + `(x◦ , u◦ , t) dt
t0
Z t1
≤ m(x(t1 )) + `(x, u‘, t) dt
t0
where x◦ is the trajectory from (10.1) corresponding to u◦ . The inequality

above shows that u◦ is indeed optimal, and V ◦ (x0 , t0 ) is the optimal value
of the cost. t
u
Example 10.2.1. Consider the simple integrator model, with the polyno-
mial cost criterion
Z t1
ẋ = u V (u) = [u2 + x4 ] dt
0
Here we have f (x, u, t) = u, `(x, u, t) = u2 + x4 , and m(x, t1 ) ≡ 0. The

Hamiltonian is thus
H(x, p, u, t) = pu + u2 + x4 ,
and the HJB equation becomes

∂V ◦ ∂V ◦
u + u2 + x4 .

− = min
∂t u ∂x
Minimizing with respect to u gives
∂V ◦
u◦ = − 21 (x, t),
∂x
which is a form of state feedback. The closed-loop system has the appealing
form
∂V ◦ ◦
ẋ◦ (t) = − 12 (x (t), t)
∂x
This equation shows that the control forces the state to move in the direction
in which the “cost to go” V ◦ decreases.
Substituting the formula for u◦ back into the HJB equation gives the
PDE
2
∂V ◦ 1 ∂V ◦

− (x, t) = − (x, t) + x4 ,
∂t 4 ∂x
with the boundary condition V ◦ (x, t1 ) = 0. This is as far as we can go,

since we do not have available methods to solve a PDE of this form. If a
solution is required, it may be found numerically. However, a simpler set
of equations is obtained in the limit as t1 → ∞. This simpler problem is
treated in Exercise 1 below. t
u
We now leave the general nonlinear model and concentrate on linear

systems with quadratic cost. We will return to the more general problem in
Chapter 11.
10.3 A solution to the LQR problem

For the remainder of this chapter we consider the LQR problem whose sys-
tem description and cost are given in (10.3). For this control problem, x0 ,
t0 and t1 are given, and x(t1 ) is free. To ensure that this problem has a
solution we assume that R is strictly positive definite (R > 0).
To begin, we now compute the optimal control u◦ by solving the HJB
partial differential equation. The Hamiltonian for this problem is given by
H(x, p, u, t) = ` + pT f = xT Qx + uT Ru + pT (Ax + Bu).
To minimize H with respect to u we compute the derivative
∇u H = 0 + 2Ru + B T p.
Setting this equal to zero, we see that the optimal control is given by
1
u◦ = − R−1 B T p (10.10)
2
where we recall that p = ∇x V ◦ . Hence the controlled state evolves according
to the equation
1
ẋ (t) = A(t)x(t) − BR−1 B T ∇x V ◦ (x, t).
2
10.3. A SOLUTION TO THE LQR PROBLEM 173
Since the matrix BR−1 B T is positive definite, we see once again that the
control tends to force x in a direction in which V ◦ is decreasing.
Equation (10.10) shows that the optimal control may be written in state
feedback form. However, we cannot compute the optimal control law unless
we can compute the value function. Substituting the formula for the optimal
control into the HJB equation gives
∂V ◦
− = H(x, ∇x V ◦ , u◦ , t)
∂t
1
= xT Qx + (∇x V ◦ )T BR−1 RR−1 B T ∇x V ◦
4
1
+(∇x V ◦ )T Ax − (∇x V ◦ )T BR−1 B T (∇x V ◦ ).
2
This yields the partial differential equation
∂V ◦ 1
− = xT Qx + (∇x V ◦ )T Ax◦ − (∇x V ◦ )T BR−1 B T (∇x V ◦ ). (10.11)
∂t 4
To solve this PDE, start at time t1 when the value function is known:
V ◦ (x, t1 ) = xT M x.
The terminal V ◦ is quadratic, which suggests that we try a quadratic solu-

tion for V ◦ over the entire interval
V ◦ (x, t) = xT P (t)x
where P is a positive semidefinite, n × n matrix. If V ◦ is of this form then

we have
∂V ◦
= xT Ṗ x ∇x V ◦ = 2P (t)x.
∂t
Thus, the PDE (10.11) becomes
−xT Ṗ x = xT Qx + 2xT P Ax − xT P BR−1 B T P x.
This can be equivalently expressed
−xT Ṗ x = xT Qx + xT (P A + AT P )x − xT P BR−1 B T P x.
Since this equation must hold for any x, the matrix P must satisfy the
following matrix differential equation:
− Ṗ = Q + P A + AT P − P BR−1 B T P. (10.12)
This is an ordinary differential equation for P , but it is time-varying, and

nonlinear due to the term which is quadratic in P . Adding the boundary
condition
V (x, t1 ) = xT M x = xT (t1 )P x(t1 )
gives
P (t1 ) = M. (10.13)
Equation (10.12) is called the Riccati Differential Equation (RDE), and its
boundary condition is given by (10.13). Since the boundary condition is
given at the final time, the solution to this differential equation may be
viewed as starting at time t1 , and then traveling backwards in time until
time t0 . The Riccati equation possesses a unique solution, provided that the
matrices A, B, Q and R are piecewise continuous in t. Further note that
this unique solution is symmetric, because P T satisfies the same differential
equation (to see this simply take transpose of both side of (10.12)).
The solution to the LQR problem is now summarized in the following
Theorem 10.2. For the LQR optimal control problem (10.3), assuming
that A, B, Q and R are piecewise continuous in t, and that R(t) > 0 for all
t,
(a) The optimal control u◦ is given in linear state feedback form u◦ (t) =
−K(t)x◦ (t), where K(t) is a matrix of time-varying gains:
u◦ = − 12 R−1 B T ∇x V ◦ = −R−1 (t)B T (t)P (t)x◦ (t) = −K(t)x◦ (t).
(b) The matrix-valued function P (t) is positive semidefinite for each t,

and is defined by the RDE (10.12) with boundary condition (10.13).
(c) The value function V ◦ is quadratic:
V ◦ (x0 , t0 ) = xT0 P (t0 )x0 .

t
u
10.4 The Hamiltonian matrix

We now show how the Riccati equation can be solved by computing the
solution to a linear ODE. In addition to providing a complete solution to
the LQR problem, this will provide a solution to the infinite time horizon
10.4. THE HAMILTONIAN MATRIX 175
optimal control problem where t1 = ∞. Consider the 2n dimensional LTV

model, where the state (X(t), Y (t))0 is a 2n × n matrix-valued function of t.
A −BR−1 B T

Ẋ(t) X(t) X(t1 ) I
= , = (10.14)
Ẏ (t) −Q −AT Y (t) Y (t1 ) M
| {z } | {z }
2n×2n 2n×n
The 2n × 2n system-matrix of (10.14) is called the Hamiltonian matrix ,

denoted by H.
Theorem 10.3. The solution to the RDE (10.12) with boundary condition
(10.13) is given by the formula
P (t) = Y (t)X −1 (t), t0 ≤ t ≤ t1 .
Proof To prove the theorem, observe that the desired initial conditions
are satisfied:
Y (t1 ) X −1 (t1 ) = M = P (t1 ).

| {z } | {z }
M I
We next verify that the Riccati differential equation is also satisfied. Using
the product rule and the matricial quotient rule dt d
(X −1 ) = −X −1 ẊX −1 ,
we have
d d
(Y X −1 ) = Y (X −1 ) + Ẏ X −1 = −Y X −1 ẊX −1 + Ẏ X −1 .
dt dt
Substituting the definitions of Ẋ and Ẏ then gives
d
(Y X −1 ) = −Y X −1 (AX − BR−1 B T Y )X −1 + (−QX − AT Y )X −1
dt
= −(Y X −1 )A + (Y X −1 )BR−1 B T (Y X −1 ) − Q − AT (Y X −1 )
which is precisely Riccati equation for P = Y X −1 . Note that this immedi-

ately tells us that Y X −1 is symmetric and postive semidefinite. t
u
In the LTI case we may extend this analysis further to explicitly solve the
RDE. We first require the following result, which shows that the eigenvalues
of the Hamiltonian matrix possess symmetry about the complex axis in C.
Lemma 10.4. For an LTI model, if λ is an eigenvalue of H, so is −λ.

Proof Define the 2n × 2n matrix J as

0 I
J := ,
−I 0
where each element is n × n. The inverse of J is equal to

−1 0 −I
J = = JT .
I 0
Using the definition of H gives
JHJ T = −HT .
By the definition of similarity transformations, we have for any eigenvalue

λ of H,
0 = det(λI − H) = det(λI − JHJ −1 ).
Moreover, for this special case
det(λI − JHJ −1 ) = det(λI + HT ) = (−1)n det((−λ)I − JHJ −1 ).
These two equations combined show that when λ is equal to an eigenvalue

of H then so is −λ, as claimed. tu
We will see below that the matrix H cannot have eigenvalues on the
jω axis, provided the model is stabilizable and detectable. Consider the
special case where the eigenvalues of H are distinct. In this case we can
solve the RDE by solving a linear differential equation. We first diagonalize
the Hamiltonian matrix:

−1 Λs 0
U HU =
0 −Λs
where the diagonal elements of Λs lie in the strict left hand plane in C. Note
that if the eigenvalues of H are distinct then, by Lemma 10.4, none can lie
on the jω axis. The matrix of eigenvectors U can be written

U11 U12
U=
U21 U22
where U11 , U21 correspond to the stable eigenvalues. The solution to the
differential equation (10.14) can be computed in this case:
Λ (t−t )
X(t) e s 1 0 −1 I
= U U (10.15)
Y (t) 0 e−Λs (t−t1 ) M
10.5. INFINITE HORIZON REGULATOR 177
This is a special case of the solution to an LTI model

x(t) = φ(t − t1 )x(t1 )
using the state transition matrix φ.
With considerable algebra, equation (10.15) can be solved to give closed-
form expressions for X(t) and Y (t), which can then be combined to form
P (t, t1 ):
P (t, t1 ) = [U21 + U22 e−Λs (t−t1 ) Ge−Λs (t−t1 ) ][U11 + U12 e−Λs (t−t1 ) Ge−Λs (t−t1 ) ]−1
(10.16)
where
G = −[U22 − M U12 ]−1 [U21 − M U11 ].
Hence the RDE can be solved explicitly for an LTI model.
10.5 Infinite horizon regulator

We now consider the infinite-horizon or steady-state control problem. We
assume that the model is LTI, and since the time horizon is infinite, the cost
only depends upon the initial condition of the model:
ẋ = Ax + Bu, x(0) = x0
Z ∞
(10.17)
V (x) = (xT (t)Qx(t) + uT (t)Ru(t)) dt
0
There is no need to include a terminal penalty function because the assump-

tions we impose will imply that x(t) → ϑ as t → ∞.
Our approach is to consider an optimal control problem on the finite
interval [0, t1 ], and then let t1 → ∞. Let P (t, t1 ) denote the solution to
the RDE with time horizon t1 , so that the optimal control may be written
u(t) = −K(t, t1 )x(t) = −R−1 B T P (t, t1 )x(t). Suppose that P (t, t1 ) → P̄
as t1 → ∞, where P̄ is an n × n matrix independent of time. Then the
limiting control law u(t) = −Kx(t) = −R−1 B T P̄ x(t) is optimal for the
infinite horizon control problem. To see this, observe that for any control u
on [0, ∞), since xT P (0, t1 )x is equal to the optimal cost for the finite horizon
problem,
Z t1
T
x P (0, t1 )x ≤ (xT (t)Qx(t) + uT (t)Ru(t)) dt
0
Z ∞
≤ (xT (t)Qx(t) + uT (t)Ru(t)) dt.
0
Letting V ◦ denote the optimal cost for the infinite horizon problem, it follows
that xT P (0, t1 )x ≤ V ◦ (x) for all x and t1 , and hence
lim xT P (0, t1 )x ≤ V ◦ (x).

t1 →∞
This gives a lower bound on the cost V ◦ (x), and this lower bound is attained
with the limiting controller.
Under certain conditions we can show that the limit does exist, and
thereby solve the infinite horizon control problem.
Example 10.5.1. Consider the simple model
ẋ = x + u
Z t1
V (u) = (x2 + u2 ) dt + 5x(t1 )2 .
0
The RDE becomes

d
− P (t, t1 ) = 2P (t, t1 ) + 1 − P (t, t1 )2 , P (t1 , t1 ) = 5.
dt
A solution to this differential equation is shown in Figure 10.2. These plots
suggest that for any fixed t, P (t, t1 ) converges to a constant P̄ as t1 → ∞.
If this is the case, then the limit P̄ must be an equilibrium for the ARE.
That is,
0 = 2P̄ + 1 − P̄ 2 .
√
Solving this equation gives P̄ = 1 + 2, which is consistent with the limiting
value shown
√ in the plots. We will see that the limiting control law u =
−(1 + 2)x is optimal for the infinite horizon cost with t1 = ∞. t
u
P(t,2)
4.5
3.5
2.5
t
0 1 2
P(t,10)
5
4.5
3.5
2.5
t
0 2 4 6 8 10
Time (second)
Figure 10.2: Solutions to the ARE for a scalar model with t1 = 2 and
t1 = 10. It appears that for any fixed t, P (t, t1 ) converges to a constant P̄
as t1 → ∞.
Using the formula (10.16) it is easy to show that the limiting matrix
P̄ does exist in the special case where the eigenvalues of the Hamiltonian
are distinct. Because the matrix Λs is Hurwitz, in this case it follows from
(10.16) that
−1
lim P (t, t1 ) = U21 U11 := P̄ (10.18)
t1 →∞
where P̄ is an n × n symmetric matrix which does not depend on t. The

optimal control for (10.17) is given by the limiting formula
u◦ (t) = −R−1 B T P̄ x◦ (t).
Note that u◦ is given by a state feedback control law of the form u◦ = −Kx,
which is a form of pole placement. We will see in Section 10.6 that the
location of the closed-loop poles possess desirable properties.
We now use the Hamiltonian matrix to characterize the closed-loop poles
under optimal control. From the eigenvalue equation HU = U Λ we have
A −BR−1 B T

U11 U11 U11
H = = Λ . (10.19)
U21 −Q −AT U21 U21 s
The upper set of equations defined by (10.19) may be written U11 Λs =

AU11 − BR−1 B T U21 , and this can be transformed to obtain
−1
U11 Λs = AU11 − BR−1 B T U21 (U11 U11 )
−1
= (A − BR−1 B T U21 U11 )U11 .
−1
Since K = R−1 B T P̄ and P̄ = U21 U11 , this gives the formula
(A − BK)U11 = U11 Λs .
The matrix U11 is thus a modal matrix for the optimal closed-loop system
matrix Acl = (A − BK). That is,
Theorem 10.5. Assume that the controlled system is asymptotically stable,

so that the closed-loop poles determined by
∆(s) = det[Is − (A − BR−1 B T P̄ )]
lie in the strict left half plane. Then,
(a) The stable eigenvalues Λs for H are also the optimal closed-loop
poles.
(b) The columns of U11 are the eigenvectors for Acl = A − BR−1 B T P̄ .
t
u
Using the Hamiltonian we now give a quadratic algebraic equation which
the matrix P̄ must satisfy. Considering both the first and second rows of
(10.19) gives
AU11 − BR−1 B T U21 = U11 Λs (10.20)

T
−QU11 − A U21 = U21 Λs (10.21)
−1 −1
Multiplying the first equation by (U21 U11 ) on the left hand side and by U11
on the right gives
−1 −1 −1 −1 −1 −1
(U21 U11 )AU11 U11 − (U21 U11 )BR−1 B T U21 U11 = (U21 U11 )U11 Λs U11
or
−1
P̄ A − P̄ BR−1 B T P̄ = U21 Λs U11 (10.22)
−1
Multiplying (10.21) by U11 on the right we obtain
−1 −1 −1
−QU11 U11 − AT U21 U11 = U21 Λs U11
or
−1
− Q − AT P̄ = U21 Λs U11 . (10.23)
Combining (10.22) and (10.23) gives P̄ A − P̄ BR−1 B T P̄ = −Q − AT P̄ , or
AT P̄ + P̄ A + Q − P̄ BR−1 B T P̄ = 0 (10.24)
This final formula is known as the algebraic Riccati equation (ARE). It

contains n(n+1)
2 equations to solve for the components of P̄ , since P̄ is sym-
metric.
Example 10.5.2. Consider the system/cost combination
ẋ = Z
x+u
∞
V = u2 dt.
0
We have R = 1, which is strictly positive, and Q = 0 ≥ 0, so our conditions

on the cost function are satisfied. However, it is obvious that the optimal
control is to set u equal to zero so that V is also zero. The resulting closed-
loop system is unstable, so this is obviously a poor design! Clearly, some
additional conditions on the cost function must be imposed. t
u
Example 10.5.3. The difficulty in the previous example is that the cost is
not a good measure of the internal stability of the model. This is reminiscent
of detectability, a connection that we further explore in two more examples.
Consider first the LTI model with quadratic cost

−3 −2 0
ẋ = x+ u;
1 0 1
Z ∞ (10.25)
2 2

V (u) = (x1 + x2 ) + u dt
0
In this example

−3 −2 0 1 1
A= , B= , Q= ≥ 0, R = 1.
1 0 1 1 1
The matrix A is Hurwitz, with eigenvalues at −2, −1.

Letting y = Cx = x1 + x2 , the cost becomes V (u) = (y 2 + u2 ) dt.
R
Consider a very special case where the initial condition x(0) belongs to the
unobservable subspace Σō . What then is the optimal control? Recall that
if u ≡ 0 then the output becomes
y(t) = CeAt x(0), t ≥ 0.
Since x(0) ∈ Σō we then have y ≡ 0. Obviously then, this is the optimal
control since it yields V (u) = 0! It follows from the definition of the optimal
cost that
V ◦ (x) = 0 for x ∈ Σō .
This can be seen explicitly for this example by solving the ARE, and
computing Σō . The observability matrix for this example is

C 1 1
O= = .
CA −2 −2
Hence the unobservable subspace becomes

1
Σō = N (O) = Span .
−1
Using the lqr command in Matlab, the solution to the ARE is approximately

1 1
P = 0.24
1 1
Hence we do indeed find that V ◦ (x) = xT P x = 0 for x ∈ Σō .

Setting u = 0 may be reasonable for a stable system such as (10.25).
Consider however the reversed situation where the matrix A is replaced by
−A:

3 2 0
ẋ = x+ u.
−1 0 1
The new state matrix has eigenvalues at +2, +1. Hence for any non-zero
initial condition the state explodes as t → ∞ when the input is set to zero.
For this system can we again conclude that a zero input is optimal? For
this cost criterion, the answer is yes, for precisely the same reason as before.
The observability matrix for this model is

C 1 1
O= = ,
CA 2 2
so that the unobservable subspace is again equal to Σō is again equal to

the span of the single vector (1, −1)T . If x(0) ∈ Σō , and u ≡ 0, then
y = Cx = x1 + x2 is identically zero.
We conclude that V (u) = 0, even though the state x(t) → ∞ if x(0) 6= 0
lies in the unobservable subspace. This unstable behavior of the optimized
system is a consequence of a lack of detectability. t
u
For a general LQR problem, suppose that the matrix Q is factored as

Q = C T C for a p × n matrix C. As in the previous example, we then define
y = Cx so that the cost function becomes
Z ∞ Z ∞
T T
V = (x Qx + u Ru) dt = (y T y + uT Ru) dt.
0 0
If the pair (A, C) is observable, then for any control u, and any initial state
x(0) 6= ϑ, the cost V is strictly positive. To see Rthis, note that R if u is
identically zero, then by observability we must have xT Qx dt = |y|2 dt >
0. If u is not identically zero, then V > 0 since R > 0. Since for an
initial condition x0 the value function may be expressed V ◦ (x0 ) = xT0 P̄ x0 ,
it follows that the matrix P̄ is positive definite (P̄ > 0). From the previous
example it is clear that P will be singular if the system is not observable.
If the system is not detectable, then the last example shows that we might
expect disaster. These results can be refined to give the following
Theorem 10.6. For the LTI system (10.17) with R > 0 and Q = C T C ≥ 0,
(i) If (A, B) is stabilizable, and (A, C) is detectable then

(a) There exists a positive semi-definite solution P̄ to the ARE (10.24),
which is unique in the class of positive semi-definite matrices.
(b) The closed-loop system matrix Acl = A − BR−1 B T P̄ is Hurwitz.
(c) The infinite horizon optimal control is u◦ = −R−1 B T P̄ x, and the
infinite horizon optimal cost is V ◦ (x0 ) = xT0 P̄ x0 .
(d) If P (t, t1 ) is the solution to the RDE (10.12) subject to P (t1 , t1 ) =
0, then limt1 →∞ P (t, t1 ) = P̄ .
(ii) If (A, B) is stabilizable, and (A, C) is observable, then (a)–(d) hold, and
in addition the matrix P̄ is positive definite.
t
u
Example 10.5.4. Consider the multivariate model

0 −1 1 0
ẋ = x+ u
0 0 0 1
Z ∞
V = (4x21 + 4x1 x2 + x22 + u21 + u22 ) dt.
0
From these equations we may deduce that the weighting matrices are

4 2 1 0
Q = , R= .
2 1 0 1
The matrix Q has rank one, and Q ≥ 0. It follows from the spectral repre-
sentation of a positive definite matrix that there is a 1 × 2 matrix C such
that Q = C T C. In fact C = (2, 1), so we define y = Cx = 2x1 + x2 . Thus,
the controllability and observability matrices become

1 0 0 −1 2 1
C= O= .
0 1 0 0 0 −2
Since both of these matrices have rank 2, the conditions of Theorem 10.6 (ii)
are satisfied.
We now check directly to see if the controlled system is stable by com-
puting the closed-loop system poles. The Hamiltonian matrix for this model
is
 
0 −1 −1 0
 0 0 0 −1 
H =   −4 −2
.
0 0 
−2 −1 1 0
The eigenvalues of H are found by solving the characteristic equation

det[sI − H] = s4 − 5s2 + 4
= (s − 2)(s + 2)(s − 1)(s + 1).
Note that the roots of this equation possess the symmetry required by
Lemma 10.4. The eigenvectors {wi } of the stable modes {λi } are
   
1 1
0  −1 
λ1 = −1, w1 =  2
, λ2 = −2, w 2
=  2 .
 
0 −1
Thus, the matrices {Uij } may be computed:
 
1 1
U11  0 −1 
= 2
.
U21 2 
0 −1
Applying (10.18) we can compute the solution to the ARE:
−1
2 2 1 1 2 0
P̄ = = ,
0 −1 0 −1 0 1
and the optimal feedback gain matrix is

−1 T 2 0
K = R B P̄ = .
0 1
The resulting optimal control

◦ −2x1
u = .
−x2
places the closed-loop eigenvalues at (−2, −1), which are the stable eigen-
values of H. The closed-loop eigenvectors are contained in the columns of
U11 . t
u
Example 10.5.5. To show how the ARE can be solved directly without
constructing the Hamiltonian, consider the model

0 1 0
ẋ = x+ u; y = x1
0 −1 1
Z ∞ (10.26)
2 2
V = (x1 + u ) dt
0
In this example we have R = 1, and Q = C T C, where C = (1 0). The

controllability and observability matrices become

0 1 1 0
C= O= .
1 −1 0 1
Again we see that the conditions of Theorem 10.6 (ii) are satisfied, and it
follows that the solution P̄ to the ARE will be positive definite. We now
verify this directly.
The ARE for this model is
AT P̄ + P̄ A − P̄ BR−1 B T P̄ + Q = 0.

` m
Writing P̄ as P̄ = , this gives the equation
m n
m2 mn

0 0 0 `−m 1 0
+ − + = 0.
`−m m−n 0 m−n mn n2 0 0
Examining the (1, 1) entry of each matrix gives
−m2 = −1 ⇒ m = +1 (since P̄ ≥ 0).
Examining the (1, 2) entry of the ARE we see that ` − 1 − n = 0, and from
the (2, 2) entry n2 +2n−2 = 0. These equations result in the unique positive
definite solution
√
3 √ 1
P̄ =
1 −1 + 3
√
The optimal feedback control law is u◦ = −R−1 B T P x = −[1 3 − 1]x.
To find the closed-loop poles, we must compute the eigenvalues of

0 1 0 √ 0 0 √ 1
Acl = A − BK = − = .
0 −1 1 3−1 −1 − 3
√
3
The characteristic equation is ∆(s) = s2 + 2 2 s + 1 = 0, which gives
√
pi = − 12 ( 3 ± j).
The closed-loop response will be somewhat under-damped, but of course it

is stable. t
u
10.6. RETURN DIFFERENCE EQUATION 187
10.6 Return difference equation

The LQR solution has some striking and highly desirable properties when
viewed in the frequency domain. These results have many applications to
design and analysis of linear control systems. In particular,
(a) For a SISO system and certain MIMO systems, a generalization of the
root locus method can be used to find the optimal state feedback
controller.
(b) For an LQR design based upon full state feedback, the closed-loop
system is insensitive to plant uncertainty, regardless of the choice of
weighting matrices Q and R. In particular, the gain margin is always
infinite, and the phase margin at least 60%.
The conclusion (b) is surprising, given the purely time domain approach that
we have taken. In the end this serves as perhaps the greatest motivation for
using LQR methods in control design.
For the LTI model with infinite horizon cost criterion expressed in Equa-
tion (10.17), we have the following identity for any s ∈ C, known as the
return difference equation:
R + B T (−sI − A)−T Q(sI − A)−1 B

= [I + K(−sI − A)−1 B]T R[I + K(sI − A)−1 B]. (10.27)
The vector K is the optimal feedback gain
K = R−1 B T P̄ .
The left hand side of (10.27) involves only the open-loop system, while the
right hand side depends on the feedback gain K. The return difference
equation is derived by simply expanding both sides of the equation, and
substituting the definition of K and the algebraic Riccati equation.
The return difference equation appears complex, but after the matrices
are given proper interpretation it has a relatively simple form. First, factor
the state weighting matrix as Q = C T C. By defining y = Cx we obtain the
controlled model illustrated in Figure 10.3. With P (s) = C(Is − A)−1 B,
the left hand side of (10.27) becomes
R + P T (−s)P (s).
From the feedback from u = −Kx we define the loop transfer function
described in Chapter 9 as L(s) = K(Is − A)−1 B, so that the right hand side
of (10.27) becomes
[I + L(−s)]T R[I + L(s)].
Hence the return difference equation can be expressed succinctly as
R + P T (−s)P (s) = [I + L(−s)]T R[I + L(s)]. (10.28)
On the complex axis we have H T (−jω) = H ∗ (jω) for any transfer function
H, so that (10.28) may be written
R + P ∗ (jω)P (jω) = [I + L(jω)]∗ R[I + L(jω)]. (10.29)
r u x y
∑ B ∑ C
-
Figure 10.3: The loop transfer function of the LTI system under full state
feedback u = −Kx + r is L(s) = K(Is − A)−1 B.
In the SISO case with R = r > 0 a scalar, the return difference equation
becomes
r + P (−s)P (s) = r[1 + L(−s)][1 + L(s)]. (10.30)
The poles of the closed-loop system are found by solving the equation
1 + L(s) = 0.
In view of (10.30), the return difference equation gives the closed-loop poles
as a function of the weighting parameter r and the open-loop transfer func-
tion P (s) through the formula
1
1 + P (−s)P (s) = 0. (10.31)
r
The set of all solutions to (10.31), with r ranging from 0 to ∞, is called
the symmetric root locus. For fixed r, the roots of this equation are the
eigenvalues of the Hamiltonian matrix H with R = r. We have seen in
Section 10.5 that the stable (left hand plane) eigenvalues of the Hamiltonian
matrix are the poles of the closed-loop system.
As an illustration consider the general second order system with transfer
function P (s) = (s − z)/(s2 + 2ζs + 1), where z < 0, and ζ > 0. The
symmetric root locus is illustrated in Figure 10.5. Note the symmetry of
the poles and zeros with respect to the jω axis. It is seen here that when
r ≈ ∞, the closed-loop poles are approximately equal to the open loop poles
of the system. For small r, one closed-loop pole converges to ∞ along the
negative real axis, and the other converges to the zero z.
Imag Axis
p -p
1 1
z -z Real Axis
p -p
2 2
Left half plane
Figure 10.4: The symmetric root locus for the system with transfer function
P (s) = (s − z)/(s2 + 2ζs + 1). The roots of 1 + 1r P (−s)P (s) = 0 give the
optimal closed-loop poles, where r is the control weighting in the LQR cost.
As r varies from 0 to ∞, this gives the symmetric root locus.
Example 10.6.1. We now revisit (10.26) to show how a control design

may be accomplished using the symmetric root locus. The point is that
everything on the left hand side of the return difference equation is known.
The desired closed-loop feedback poles are unknown, but are defined through
the right hand side of this equation. In the SISO case, this can be exploited
to solve for the feedback gain K directly, without solving a Riccati equation.
We must first compute the transfer function P (s). From the state equa-
tions we have
sX1 (s) = X2 (s)

sX2 (s) = −X2 (s) + U (s)
Y (s) = X1 (s).
Combining these equations gives s2 Y (s) = −sY (s) + U (s), or

Y (s) 1
P (s) = = .
U (s) s(s + 1)
To compute the symmetric root locus we solve for the roots of the equation
h 1 ih 1 i h 1 i
0 = 1 + r−1 = 1 + r−1 4 .
s(s + 1) −s(−s + 1) s − s2
Imag Axis
r=1
-1 0 (twice) +1 Real Axis
r=1
Left half plane
Figure 10.5: TheR symmetric root locus for the model ÿ = −ẏ + u, with cost
criterion V = y 2 + ru2 dt. A more highly damped closed-loop response can
be obtained by increasing r below the nominal value of 1, while decreasing
r will result in greater closed-loop bandwidth.
A sketch of the symmetric root locus is provided in Figure 10.5.√ When
r = 1 we have already seen that the optimal ◦
√ control is u = −[1, 3 − 1]x,
and that the closed-loop poles are − 12 ( 3 ± j). From Figure 10.5 it is
evident that a more highly damped closed loop response can be obtained by
increasing r, while reducing r will result in greater closed-loop bandwidth.
t
u
The return difference equation leads to an elegant approach to pole place-
ment. The resulting symmetric root locus gives the designer a ‘good’ set of
poles to choose from, which can then be narrowed down based upon other
constraints, such as closed-loop bandwidth or time domain specifications.
The most striking implication of the return difference equation is perhaps
its application to sensitivity analysis. In the SISO case, we have seen that
by letting s = jω we obtain
1
|1 + L(jω)|2 = 1 + |P (jω)|2 .
r
This implies Kalman’s inequality,

|1 + L(jω)|2 ≥ 1, ω ∈ R.
If we define the sensitivity function as S = 1/(1 + L), then we also have
|S(jω)| ≤ 1, ω ∈ R. (10.32)
It follows from (9.1) that the controlled system is insensitive to plant un-
certainty at all frequencies. A typical plot of the loop transfer function L is
shown in Figure 10.6. We see from this plot that the controlled system has
a gain margin of infinity and a phase margin of ±60 degrees. These results
hold for any full state LQR design.
Imag Axis
ω = 10
ω=1
-1 Real Axis
L(j ω)
Figure 10.6: The loop transfer function L(jω) = K(Ijω − A)−1 B for the
LTI system under optimal control cannot enter the shaded disk centered at
−1.
Example 10.6.2. As an illustration of the frequency domain properties of

a state space model under optimal control, we now return to the Pendubot
model linearized in the upright position. To control the system we define
the cost as
Z ∞
|x(t)|2 + ru(t)2 dt.

V =
0
where x = (q1 , q̇1 , q2 , q̇2 )T . We therefore have Q = C T C, where Q = C = I.

Since the matrix Q has rank n, so that C is square, the transfer function
defined as P (s) = C(Is − A)−1 B is an n × 1 vector-valued function of s. The

return difference equation for the Pendubot with this cost criterion becomes
4
X
r+ Pi (−s)Pi (s) = r[1 + L(−s)][1 + L(s)],
i=1
so that a generalization of the symmetric root locus is still possible, since

the closed-loop poles are still equal to the roots of 1 + L(s). For any i, the
poles of Pi are equal to the eigenvalues of A. These are simply the open-loop
poles, which are roughly located at {−9.5, −5.5, 5.5, 9.5}. The numerator of
Pi depends upon i, but the sum on the left hand side is readily computable
using Matlab. The resulting symmetric root locus is shown in Figure 10.7.
t
u
For r = 1 the closed-loop poles are placed at (9.3), as described in

the beginning of Chapter 9. We saw then that when this controller was
combined with an observer based on measurements at the lower link alone,
the controlled system is highly sensitive to modeling errors. For instance,
the gain margin is found to be less than 10−2 . With full state feedback we
know that the controlled system is far more robust. A Nyquist plot of the
loop transfer function L(jω) is shown in Figure 10.8. We see in this figure
that, as expected, the gain margin with this design is infinite, and that the
gain can also be reduced by one half, and the system will remain stable.
Imag Axis
(2) (2) (2) (2) Real Axis
Figure 10.7: The symmetric root locus for the Pendubot with Q = I. Each
of the poles and zeros in this diagram are indicated with crosses and zeros,
respectively. Each pole appears twice because of the symmetry of the open-
loop transfer function.
Imag Axis
L(j ω)
-1
Real Axis
Figure 10.8: The Nyquist plot for the Pendubot controlled with full state
feedback, based on an LQR design. If all of the states can be measured,
a far more robust design is possible compared with what is found using a
single output measurement.
This chapter has provided only a brief survey on the LQR problem. For a
more detailed discussion, see the book [1]. Some historic papers on the topic
are:
R. E. Kalman, “Contributions to the Theory of Optimal Con-

trol,” Bol. Soc. Matem. Mex., (1960), pp. 102–119.
R. E. Kalman, “When is a Linear Control System Optimal?”,
Trans. ASME Ser. D: J. Basic Eng., Vol. 86 (March 1964),
pp. 1–10.
Another good source for survey/tutorial papers on the linear quadratic the-
ory is the December 1971 special issue of the IEEE Transactions on Auto-
matic Control.
Matlab Commands
LQG Solves the LQR problem using the data A, B, Q, and R.
LQG2 A different numerical algorithm for solving the LQR problem.
ARE Gives solution to the algebraic Riccati equation.
RLOCUS Used together with the convolution command CONV, this can
be used to graph the symmetric root locus.
10.7. EXERCISES 195
10.7 Exercises
10.7.1 The purpose of this exercise is to review the derivation of the HJB
equation, and to provide an introduction to a general theory of infinite
horizon optimal control. Consider the time invariant nonlinear system
ẋ = f (x, u), x(0) = x0 ∈ Rn .
We assume that there is an equilibrium at ϑ ∈ Rn , in the sense that

f (ϑ, ϑ) = ϑ. Assuming that the goal is to force the state to zero,
consider the cost criterion
Z ∞
J(u) = `(x, u) dt,
0
where ` : Rn × Rm → R+ is smooth, with `(ϑ, ϑ) = 0. We typically

also assume that ` is positive definite, as defined in Section 4.2. We
let J ◦ (x) denote the minimum of J over all controls, when the initial
condition of the model is x.
(a) Derive a “principle of optimality” for this problem by considering

an optimal trajectory (x◦ , u◦ ), as done in the derivation of the
HJB equations.
(b) Using (a), derive a differential equation that J ◦ must satisfy.
(c) Apply your solution in (b) to solve the optimal control problem de-
fined by the simple integrator model with polynomial cost defined
below:
Z ∞
ẋ = u J(u) = [u2 + x4 ] dt
0
10.7.2 Solve by hand the ARE with

0 1 0 1 0
A= , B= , Q= , R = 1.
0 0 1 0 0
10.7.3 For the double integrator ÿ = u, what is the minimal value of the
cost
Z ∞
V (u) = y(t)2 + u(t)2 dt,
0
over all controls u(·), when y(0) = 1, ẏ(0) = 0? Explain why the
minimum exists.
10.7.4 For the scalar time-invariant plant ẋ = ax(t) + u(t), and under the
quadratic performance index
Z ∞
V = [x2 (t) + ru2 (t)]dt ,
0
(a) Obtain the optimal feedback control as a function of the scalar

parameters a and r;
(b) Study the eigenvalues of the optimal closed-loop system as
(i) r → ∞ (expensive control)
(ii) r → 0 (cheap control)
10.7.5 The linearized (and normalized) magnetically suspended ball is de-

scribed by ẋ1 = x2 , ẋ2 = x1 + u, where the initial state x(0) is
specified. Answer the following three questions by constructing the
Hamiltonian matrix. You may check your work using Matlab.
(a) Obtain the feedback control u(x) that minimizes the infinite-horizon
Z ∞
V = (x21 + 2x1 x2 + x22 + 4u2 ) dt.
0
(b) Determine the eigenvalues of the closed-loop system matrix Acl .

(c) Find the minimum value of V as a function of x(0). Explain why
in this example we can have V (x0 ) = 0 even though x0 6= ϑ.
10.7.6 The following is the so-called generalized regulator problem: The sys-
tem equation is
ẋ(t) = A(t)x(t) + b(t)u(t) + c(t),
where c(·) is an n-dimensional vector function defined on the finite

interval [t0 , t1 ], with continuous entries; and the performance index is
Z t1
V = xT (t1 )Qf x(t1 ) + [xT (t)R1 (t)x(t) + 2xT (t)R12 (t)u(t) + uT (t)R2 (t)u(t)] dt
t0
where all matrices have continuous entries, and

R1 (·) R12 (·)
R2 (·) > 0, Qf ≥ 0, T (·) ≥0.
R12 R2 (·)
10.7. EXERCISES 197
(a) Using the Hamilton-Jacobi-Bellman equation, show that the opti-

mal control for this problem can be written as
u◦ (t) = γ ◦ (x, t) = −R2−1 (t) [B T (t)P (t) + R12

T
(t)]x + B T (t)k(t) ,

where P (·) is a matrix function (≥ 0) and k(·) is a vector function.

Obtain the expressions (differential equations) satisfied by P and
k.
(b) Obtain an expression for the minimum value of V .
10.7.7 In the previous problem, assume that all matrices and the vector c(·)
are time-invariant, and t1 → ∞. Obtain an expression for the steady-
state optimal feedback control under appropriate controllability and
observability assumptions. Is this a stabilizing control?
10.7.8 The Pendubot arranged in the vertical position may be approxi-
mately described by the linear system ẋ = Ax + Bu; y = x1 where
   
0 1.0000 0 0 0
 42.0933 0 −4.3893 0   6.2427 
A=  B= .
 0 0 0 1.0000   0 
−29.4557 0 43.3215 0 −9.7741
There exists a feedback control law u = −K(r)x which minimizes the

Z ∞
y 2 + αẏ 2 + ru2 dτ

V (u) =
0
Plot the resulting closed-loop poles as r is varied from 0 to ∞, and dis-

cuss your results. How does the introduction of ẏ in the cost functional
affect the root locus? Note that this will require a generalization of the
symmetric root locus derived in these notes. Include the derivation of
this result with your plots.
10.7.9 LQR theory can also be extended to problems where also the deriva-
tive of the control variable appears in the performance index (i.e. we
have a soft constraint not only on the control variable but also on its
derivative). Specifically, the LQR problem with cost
Z ∞
V (u) = [xT (t)Qx(t) + uT (t)R1 u(t) + u̇T (t)R2 u̇(t)] dt
0
subject to ẋ = Ax + Bu can be solved by introducing the additional

state variable, z = u, and treating v = u̇ as the control variable.
(a) By following this procedure, assuming that Q ≥ 0, R1 ≥ 0, R2 > 0,

and x(0) = x0 , u(0) = u0 are specified, obtain the general solution
of the above by simply applying the LQR theory developed in
these notes.
(b) Can you formulate a generalization of the symmetric root locus
for this more general problem?
(c) After obtaining the general solution, apply it to the specific prob-
lem:

0 1 0
ẋ = x+ u; x1 (0) = x2 (0) = 1, u(0) = 1
0 0 1
Z ∞
V (u) = [x21 (t) + ρu̇2 (t)] dt.
0
Provide a root locus plot for 0 < ρ < ∞, and compute the optimal
control for ρ = 2.
Note: For this numerical example, your solution will be in the
form v ≡ u̇ = k1 x1 + k2 x2 + k0 u, where k1 , k2 , k0 are scalar
quantities, whose values will have to be determined as part of the
solution process.
10.7.10 This problem deals with optimal control problems in which the
closed-loop system is required to have a prescribed degree of stability.
Optimal control is an expression of the tradeoff between the desire
to place closed-loop poles deep into the left-half-plane and the desire
not to use excessive control (because of possible actuator saturation)
or have large excursions of the state trajectories (because of modeling
uncertainty). Sometimes we want to guarantee that the closed-loop
poles are stable by a “margin” of at least α. That is, the poles lie in
the region Ωα = {s : Re(s) ≤ −α}.
To this end, consider the plant ẋ = Ax+Bu, and the cost functional
Z ∞
e2αt xT Qx + uT Ru dt,

V (u) = α>0
0
√
where Q ≥ 0 and R > 0. Observe that if V (u◦ ) < ∞ then Qx◦ (t)
must decay faster than e−αt which, under appropriate conditions,
means the closed-loop poles lie in the left-half plane as desired. In
this problem you will rigorously establish this assertion.
10.7. EXERCISES 199
(a) Set z(t) = eαt x(t) and v(t) = eαt u(t), and denote the resulting
as a function v by Ṽ (v). Compute Ṽ (v) and the corresponding
state equations.
(b) Carefully write down a complete solution to this optimal control
problem including necessary and sufficient conditions under which
solutions exist and conditions under which these solutions are sta-
bilizing. In the case where the eigenvalues are distinct, interpret
your conditions using the Hautus-Rosenbrock criteria.
(c) Show that the closed-loop system has all its poles inside the left
hand plane as prescribed under appropriate conditions on the
open-loop system.
(d) With A = B = Q = R = x0 = 1 find the optimal control law and
the optimal cost, and observe that the poles are indeed placed
within Ωα .
10.7.11 Consider again the flexible structure defined in (7.10). The input
is the voltage to a DC motor, and the states of this model are
x1 = θ1 , x2 = θ̇1 , x3 = θ2 , x4 = θ̇2 .
(a) Plot the loop transfer function with your previous design. You
should use the Matlab command NYQUIST, as follows
nyquist(A-.00001*eye(4),B,K,0)
where K is the feedback gain. This will give a plot of L(jω) =
K(Ijω − A + I)−1 B. The term is needed since the open-loop
system has a pole at the origin. Does your plot enter the region
dist(s, −1) < 1?
(b) Plot the optimal closed-loop poles for the cost function
Z ∞
V (u) = θ22 + β(θ̇2 )2 + ρu2 dτ
0
where ρ varies from 0 to ∞. Try β = 0, and then some β > 1.

(c) Obtain an optimal design for which the dominant poles satisfy
−10 ≤ σ ≤ −5, and plot the loop transfer function with this
design.
10.7.12 Consider the input-output system described in the frequency do-
main by
s−1
Y (s) = U (s) + W (s)
s2 + s + 1
where y is the measurement, u is the input, and w is a disturbance.

Two control designs will be obtained below by setting w ≡ 0. In
practice however, w 6= 0, and hence its affect on the output must be
understood.
(a) Obtain a state space model in observable canonical form: Let x1 =

y − w, and ẋ2 = −x1 − u, so that y = x1 + w.
(b) Derive a state feedback control law u = −Kx for an approximately
critically damped response which minimizes V (u) = y 2 +ru2 dτ
R
for some r. Your closed-loop poles should be constrained in mag-

nitude between 1 and 5.
Sketch the symmetric root locus by hand - use Matlab to verify
your plot, and to determine an appropriate feedback gain.
Y
What is the frequency response W (jω) for the closed-loop system
when the state x is perfectly observed?
(c) Design a reduced order observer, and combine this with the control
law obtained above so that now u = −K1 y − K2 x̂2 . Note that the
observer will have to use the approximation y ≈ x1 since neither
x1 nor x2 are directly measurable.
Your observer pole should be at least three times as large in
magnitude than the magnitude of the largest closed-loop pole
obtained in (b).
Y
Plot the frequency response W (jω) using the BODE command, and
discuss your findings.
10.7.13 In this problem you will design a controller for a model of a Boeing
747 aircraft.
A linearized model for the longitudinal motion of this plant when
operating shortly after take-off, at sea level, with a speed of 190 miles
per hour is given by
      
δV −0.0188 11.5959 0 −32.2 δV 0
 δα  =  −0.0007 −0.5357
d  1 0    δα  +  0
     
 δe
dt q
   0.000048 −0.4944 −0.4935 0   q  −0.5632
δθ 0 0 1 0 δθ 0
where δV is the deviation from the nominal speed of 190 mph, δα is the
angle of attack, which is the longitudinal angle of the plane with respect
to the longitudinal velocity vector, δθ is the pitch angle, which is the
longitudinal angle of the plane with respect to the ground, and q = δ̇θ.
10.7. EXERCISES 201
The input δe is the deviation of the elevators with respect to some

nominal value. Of course, the throttle position is also a significant
control, but we will hold the throttle value constant in this design.
The altitude of the plane h is the variable that we would like to control.
Linearizing, we find that a reasonable approximation to the rate of
change of h is given by
δ̇h = 290(δθ − δα)
(a) Adjoin the variable h to the state space model above to obtain a
fifth order system with state x. Obtain an LQR design to obtain
a feedback control u = −Kx + r for this system using symmetric
root locus. Plot the step response of both the altitude h and the
pitch angle θ of the controlled system. Try to obtain a rise time
of 250 seconds with reasonable undershoot.
Note that the response is always slow due to one closed-loop pole
close to the origin.
(b) Can you obtain a better controller with a more general choice of Q
and R? For your best design, where are the resulting closed-loop
poles?
(c) Using your best choice of K in (b), improve the response by modi-
fying the controller as follows: Pass the reference signal r through
a first order filter G(s). Hence, in the frequency domain, your
controller is of the form
s−z
U (s) = −KX(s) + k R(s).
s−p
You should place the zero z at the location of the slow pole (ap-
proximately), and place the pole p further from the origin. This
should cancel the effect of the slow closed-loop pole in the closed-
loop response. The gain k should be chosen to achieve good
steady state tracking for d.c. reference inputs.
Again plot the step response of both the altitude h and the pitch
angle θ of the controlled system. Can you choose k, z and p to
improve your response over that obtained in (b)?
Chapter 11
An Introduction to the
Minimum Principle
We now return to the general nonlinear state space model, with general cost
criterion of the form
ẋ(t) = f (x(t), u(t), t), x(t0 ) = x0 ∈ Rn

Z t1
(11.1)
V (u) = `(x(t), u(t), t) dt + m(x(t1 ))
t0
We have already shown under certain conditions that if the input u mini-
mizes V (u), then a partial differential equation known as the HJB equation
must be satisfied. One consequence of this result is that the optimal control
can be written in state feedback form through the derivative of the value
function V ◦ . The biggest drawback to this result is that one must solve a
PDE which in general can be very complex. Even for the relatively sim-
ple LQR problem, the solution of the HJB equations required significant
ingenuity.
The minimum principle is again a method for obtaining necessary con-
ditions under which a control u is optimal. This result is based upon the
solution to an ordinary differential equation in 2n dimensions. Because it
is an ordinary differential equation, in many instances it may be solved
even though the HJB equation is intractable. Unfortunately, this simplicity
comes with a price. The solution u◦ to this ordinary differential equation is
in open-loop form, rather than state space form. The two view points, the
Minimum Principle and the HJB equations, each have value in nonlinear
optimization, and neither approach can tell the whole story.
203
204CHAPTER 11. AN INTRODUCTION TO THE MINIMUM PRINCIPLE
11.1 Minimum Principle and the HJB equations

In a course at this level, it is not possible to give a complete proof of the
Minimum Principle. We can however give some heuristic arguments to make
the result seem plausible, and to gain some insight. We initially consider
the optimization problem (11.1) where x0 , t0 , t1 are fixed, and x(t1 ) is free.
Our first approach is through the HJB equation
∂ ◦ ∂V ◦
− V (x, t) = min(`(x, u, t) + f (x, u, t))
∂t u ∂x
∂V ◦
u◦ (t) = arg min(`(x, u, t) + f (x, u, t)).
u ∂x
For x, t fixed, let ū(x, t) denote the value of u which attains the minimum
above, so that u◦ = ū(x◦ (t), t). In the derivation below we assume that ū
is a smooth function of x. This assumption is false in many models, which
indicates that another derivation of the Minimum Principle which does not
rely on the HJB equation is required to create a general theory. With ū so
defined, the HJB equation becomes
∂ ◦ ∂V ◦
− V (x, t) = `(x, ū(x, t), t) + (x, t)f (x, ū(x, t), t).
∂t ∂x
∂V ◦
Taking partial derivatives of both sides with respect to x gives the term ∂x
on both sides of the resulting equation:
∂2 ∂`
− V ◦ (x, t) = (x, ū(x, t), t)
∂x∂t ∂x
∂2V ◦
+ 2 (x, t)f (x, ū(x, t), t)
∂ x
∂V ◦ ∂f
+ (x, t) (x, ū(x, t), t)
∂x ∂x
∂ ∂V ◦ ∂ ū(x, t)
+ `+ f (x, ū(x, t), t) .
∂u ∂x ∂x
| {z }
derivative vanishes at ū
◦
As indicated above, since ū is the unconstrained minimum of H = ` + ∂V
∂x f ,
the partial derivative with respect to u must vanish.
This PDE holds for any state x and time t - they are treated as inde-
pendent variables. Consider now the optimal trajectory x◦ (t) with optimal
11.1. MINIMUM PRINCIPLE AND THE HJB EQUATIONS 205
input u◦ (t) = ū(x◦ (t), t). By a simple substitution we obtain

∂2 ∂` ◦
0= V ◦ (x◦ (t), t) + (x (t), u◦ (t), t)
∂x∂t ∂x
∂2V ◦ ◦
+ (x (t), t)f (x◦ (t), u◦ (t), t)
∂x2
∂V ◦ ◦ ∂f
+ (x (t), t) (x◦ (t), u◦ (t), t). (11.2)
∂x ∂x
We can now convert this PDE into an ODE. The main trick is to define
∂V ◦ ◦
p(t) := (x (t), t).
∂x
This function is interpreted as the sensitivity of the cost with respect to
current state, and takes values in Rn . The derivative of p with respect to t
can be computed as follows:
d ∂2V ◦ ◦ ◦ ∂2V ◦ ◦
p(t) = (x (t), t)ẋ (t) + (x (t), t)
dt ∂x2 ∂x∂t
∂2V ◦ ◦ ◦ ◦ ∂2V ◦ ◦
= (x (t), t)f (x (t), u (t), t) + (x (t), t).
∂x2 ∂x∂t
The two mixed partial terms in this equation are also included in (11.2).
Combining these two equations, we thereby eliminate these terms to obtain
d ∂` ◦
0= p(t) + (x (t), u◦ (t), t)
dt ∂x
∂V ◦ ◦ ∂f ◦ (11.3)
+ (x (t), t) (x (t), u◦ (t), t).
|∂x {z } ∂x
pT (t)
From the form of the Hamiltonian

H(x, p, u, t) = pT f (x, u, t) + `(x, u, t),
the differential equation (11.3) may be written
ṗ(t) = −∇x H(x◦ (t), p(t), u◦ (t), t)
with the boundary condition
p(t1 ) = ∇x V ◦ (x◦ (t1 ), t1 ) = ∇x m(x◦ (t1 ), t1 ).
This is not a proof, since for example we do not know if the value func-
tion V ◦ will be smooth. However, it does make the following result seem
plausible. For a proof see [10].
Theorem 11.1 (Minimum Principle). Suppose that t1 is fixed, x(t1 ) is free,

and suppose that u◦ is a solution to the optimal control problem (11.1). Then
(a) There exists a costate vector p(t) such that for t0 ≤ t ≤ t1 ,

u◦ (t) = arg min H(x◦ (t), p(t), u, t).
u
(b) The pair (p, x◦ ) satisfy the 2-point boundary value problem:

ẋ◦ (t) = ∇p H(x◦ (t), p(t), u◦ (t), t) = f (x◦ (t), u◦ (t), t)
(11.4)
ṗ(t) = −∇x H(x◦ (t), p(t), u◦ (t), t)
with the two boundary conditions
x(t0 ) = x0 ; p(t1 ) = ∇x m(x(t1 ), t1 ).
t
u
11.2 Minimum Principle and Lagrange multipliers?

Optimal control involves a functional minimization which is similar in form
to ordinary optimization in Rm as described in a second year calculus course.
For unconstrained optimization problems in Rm , the main idea is to look
at the derivative of the function V to be minimized, and find points in Rm
at which the derivative is zero, so that ∇V (x) = ϑ. Such an x is called a
stationary point of the optimization problem. By examining all stationary
points of the function to be minimized one can frequently find among these
the optimal solution. The calculus of variations is the infinite dimensional
generalization of ordinary optimization in Rm . Conceptually, it is no more
complex than its finite dimensional counterpart.
In more advanced calculus courses, solutions to constrained optimiza-
tion problems in Rm are addressed using Lagrange multipliers. Suppose
for example that one desires to minimize the function V (x) subject to the
constraint g(x) = ϑ, x ∈ Rm , where g : Rm → Rd . Consider the new cost
function Vb (x, p) = V (x) + pT g(x), where p ∈ Rd . The vector p is known as
a Lagrange multiplier. The point of extending the state space in this way
is that we can now solve an unconstrained optimization problem, and very
often this gives a solution to the constrained problem. Indeed, if x0 , p0 is a
stationary point for Vb then we must have
ϑ = ∇x Vb (x0 , p0 ) = ∇V (x0 ) + ∇g(x0 )p0 (11.5)
ϑ = ∇p Vb (x0 , p0 ) = g(x0 ). (11.6)
11.2. MINIMUM PRINCIPLE AND LAGRANGE MULTIPLIERS? 207
Equation (11.5) implies that the gradient of V is a linear combination of

the gradients of the components of g at x0 , which is a necessary condition
for optimality under very general conditions. Figure 11.1 illustrates this
with m = 2 and d = 1. Equation (11.6) is simply a restatement of the
constraint g = ϑ. This Lagrange multiplier approach can also be generalized
to infinite dimensional problems of the form (11.1), and this then gives a
direct derivation of the Minimum Principle which does not rely on the HJB
equations.
g(x1 )
x1 g(x2 )
V(x1)
x2
V(x2 ) {x : V(x) = -2}
{x : V(x) = -1}
{x : g(x) = 0} {x : V(x) = 0}
{x : V(x) = 1}
Figure 11.1: An optimization problem on R2 with a single constraint g = 0.

The point x1 is not optimal: By moving to the right along the constraint set,
the function V (x) will decrease. The point x2 is a minimum of V subject
to the constraint g(x) = 0 since at this point we have V (x2 ) = −1. V can
get no lower - for example, V (x) is never equal to −2 since the level set
{x : V (x) = −2} does not intersect the constraint set {x : g(x) = 0}. At x2 ,
the gradient of V and the gradient of g are parallel.
To generalize the Lagrange multiplier approach we must first general-

ize the concept of a stationary point. Suppose that F is a functional on
Dr [t0 , t1 ]. That is, for any function z ∈ Dr [t0 , t1 ], F (z) is a real number.
For any η ∈ Dr [t0 , t1 ] we can define a directional derivative as follows:
F (z + η) − F (z)
Dη F (z) = lim ,
→0
whenever the limit exists. The function z(t, ) = z(t) + η(t) may be viewed
as a perturbation of z, as shown in Figure 11.2
z(t)
z(t,ε) = z(t) + ε η(t)
Figure 11.2: A perturbation of the function z ∈ D[t0 , t1 ].
We call z0 a stationary point of F if Dη F (z0 ) = 0 for any η ∈ Dr [t0 , t1 ].

If z0 is a minimum of F , then a perturbation cannot decrease the cost, and
hence we must have for any η,
F (z0 + η) ≥ F (z0 ), ∈ R.
From the definition of the derivative it then follows that an optimal solution
must be a stationary point, just as in the finite dimensional case!
The problem at hand can be cast as a constrained functional minimiza-
tion problem,
Z
Minimize V (x, u) = ` dτ + m
Subject to ẋ − f = ϑ, x ∈ Dn [t0 , t1 ], u ∈ Dm [t0 , t1 ].
To obtain necessary conditions for optimality for constrained problems of

this form we can extend the Lagrange multiplier method. We outline this
approach in five steps:
Step 1. Append the state equations to obtain the new cost functional
Z t1 Z t1
V (x, u) =
b ` dt + m(x(t1 )) + pT (f − ẋ) dt.
t0 t0
The Lagrange multiplier vector p lies in Dn [t0 , t1 ].

The purpose of this is to
gain an expression for the cost in which x and u can be varied independently.
Step 2. Use integration by parts to eliminate the derivative of x in the
definition of Vb :
Z t1 Z t1 Z t1
T T T
t1
p ẋ dt = p dx = p x t0 −
ṗT x dt.
t0 t0 t0
Step 3. Recall the definition of the Hamiltonian,
H(x, p, u, t) := `(x, u, t) + pT f (x, u, t).

11.2. MINIMUM PRINCIPLE AND LAGRANGE MULTIPLIERS? 209
Combining this with the formulas given in the previous steps gives
Z t1 Z t1
V (x, u) =
b H(x, p, u, t) dt + ṗT x dt
t0 t0
T T
+p (t0 )x(t0 ) − p (t1 )x(t1 ) + m(x(t1 )). (11.7)
Step 4. Suppose that u◦ is an optimal control, and that x◦ is the

corresponding optimal state trajectory. The Lagrange multiplier theorem
asserts that the pair (x◦ , u◦ ) is a stationary point of Vb for some p◦ . Perform
perturbations of the optimal control and state trajectories to form
u(t, δ) = u◦ (t) + δψ(t) x(t, ) = x◦ (t) + η(t). (11.8)
Since we are insisting that x(t0 ) = x0 , we may assume that η(t0 ) = ϑ.

Consider first variations in , with δ set equal to zero. Letting Vb () =
Vb (x◦ + η, u◦ ), we must have
d b
V () = 0.
d
Using (11.7) to compute the derivative gives
Z t1 Z t1
∂ ◦ ◦
0= H(x , p, u , t)η(t) dt + ṗT η dt
t0 ∂x t0
∂
−pT (t1 )η(t1 ) + pT (t0 )η(t0 ) + m(x◦ (t1 ))η(t1 ). (11.9)
∂x
Similarly, by considering perturbations in u◦ we obtain for any ψ ∈
Dm [t0 , t1 ],
Z t1
∂
ϑ= H(x◦ , p, u◦ , t)ψ(t) dt. (11.10)
t0 ∂u
This simpler expression is obtained because only the first term in (11.7)
depends upon u.
Step 5. We can choose η(t) freely in (11.9). From this it follows that
∂H
+ ṗT = ϑT ⇒ ṗ = −∇x H
∂x
and since η(t0 ) = ϑ, and η(t1 ) is free,
∂m
−pT (t1 ) + (t1 ) = ϑT ⇒ p(t1 ) = ∇x m(x(t1 )).
∂x
Similarly, by (11.10) we have

∂H
= ϑT ⇒ ∇u H = ϑ.
∂u
In fact, if u is to be a minimum of V̂ , then in fact it must minimize H point-
wise. These final equations then give the Minimum Principle Theorem 11.1.
From this proof it is clear that many generalizations of the Minimum
Principle are possible. Suppose for instance that the final state x(t1 ) = x1
is specified. Then the perturbation η will satisfy η(t1 ) = ϑ, and hence using
(11.9), it is impossible to find a boundary condition for p. None is needed
in this case, since to solve the 2n-dimensional coupled state and costate
equations, it is enough to know the initial and final conditions of x◦ .
11.3 The penalty approach?

A third heuristic approach to the Minimum Principle involves relaxing the
hard constraint ẋ − f = ϑ, but instead impose a large, yet “soft” constraint
by defining the cost
Z t1
k t1
Z
V (x, u) =
b `(x(t), u(t), t) dt + |ẋ(t) − f (x(t), u(t), t)|2 dt + m(x(t1 )).
t0 2 t0
The constant k in this equation is assumed large, so that ẋ(t)−f (x(t), u(t), t) ≈
ϑ.
We assume that a pair (xk , uk ) exists which minimizes Vbk . Letting
(x , u◦ ) denote a solution to the original optimization problem, we have
◦
by optimality,
Vbk (xk , uk ) ≤ Vbk (x◦ , u◦ ) = V ◦ .
Assuming ` and m are positive, this gives the following uniform bound
Z t1
2
|ẋ(t) − f (x(t), u(t), t)|2 dt ≤ V ◦ .
t0 k
Hence, for large k, the pair (xk , uk ) will indeed approximately satisfy the
differential equation ẋ = f .
If we perturb xk to form xk + η and define Vb () = Vb (xk + η, uk ) then
we must have d/d Vb () = 0 when = 0. Using the definition of V̂ gives
Z t1
V () =
b `(xk (t) + η(t), uk (t), t) dt + m(xk (t1 + η(t1 ))
t0
Z t1
k
+ |ẋk (t) + η̇(t) − f (xk (t) + η(t), uk (t), t)|2 dt.
2 t0
11.4. APPLICATION TO LQR 211
The derivative of this expression with respect to can be computed as

follows:
Z t1
d d
0 = Vb (0) = `(x(t), u(t), t)η(t) dt
d t0 dx
Z t1
∂
+k (ẋ(t) − f (x(t), u(t), t))T [η̇ − f (x(t), u(t), t)η(t)] dt
t0 ∂x
∂
+ m(x(t1 ))η(t1 ).
∂x
To eliminate the derivative term η̇ we integrate by parts, and this gives the
expression
Z t1 n
d ∂ d o
0= `(xk (t), uk (t), t) dt + pk (t)T f (xk (t), uk (t), t) + pk (t)T η(t) dt
t0 dx ∂x dt
∂
− pk (t1 ) + m(xk (t1 )η(t1 ),
∂x
(11.11)
where we have set pk (t) = −k(ẋk (t) − f (xk (t), uk (t), t)). Since η is arbitrary,
we see that
d ∂` ∂
ϑ= pk (t)T + (xk (t), uk (t), t) + pk (t)T f (xk (t), uk (t), t),
dt ∂x ∂x
and we again obtain the boundary condition
∂
pk (t1 ) = m(x(t1 ).
∂x
Considering perturbations in u gives the equation ∇u H (xk (t), pk (t), uk (t), t) =
0, which is a weak form of the Minimum Principle for the perturbed problem.
By letting k → ∞, this penalty approach can be used to prove both the
complete Minimum Principle and the Lagrange Multiplier Theorem.
11.4 Application to LQR

The LQR problem is a good test case to see if the Minimum Principle is
a useful tool for the construction of optimal policies. Consider again the
general LTI model with quadratic cost:
ẋ = Ax + Bu
Z t1
V = (xT Qx + uT Ru) dt + xT (t1 )M x(t1 ).
t0
The Hamiltonian is written
H = xT Qx + uT Ru + pT (Ax + Bu)
and hence the control can be computed through
∇u H = 0 = 2Ru + B T p =⇒ u = − 12 R−1 B T p.
This then gives the first set of differential equations:
ẋ = Ax + Bu = Ax − 21 BR−1 B T p. (11.12)
Through the expression ∇x H = 2Qx + AT p, we find that the derivative of

p is
ṗ = −∇x H = −2Qx − AT p. (11.13)
Equations (11.12) and (11.13) form the coupled set of differential equations
− 12 BR−1 B T

ẋ A x
= T
ṗ −2Q −A p
with boundary conditions
x(t0 ) = x0
p(t1 ) = 2M x(t1 ).
If we scale p we again arrive at the Hamiltonian defined in Section 10.4:

With λ = 21 p,
A −BR−1 B T

ẋ x ẋ
= T H , (11.14)
λ̇ −Q −A λ λ̇
with u◦ (t) = −R−1 B T λ(t).

This ODE can be solved using the “sweep method”. Suppose that λ(t) =
P (t)x(t). Then
λ̇ = Ṗ x + P ẋ = Ṗ x + P (Ax + Bu),
and substituting u◦ = −R−1 B T λ = −R−1 B T P x gives
λ̇ = Ṗ x + P (Ax − BR−1 B T P x).

11.5. NONLINEAR EXAMPLES 213
From (11.14) we also have
λ̇ = −Qx − AT λ = −Qx − AT P x.
Equating the two expressions for λ̇ gives
−Qx − AT P x = Ṗ x + P Ax − P BR−1 B T x
which yields the RDE given in (10.12). The boundary condition for λ is
λ = 12 p = 21 m0 (x1 ), which from the definitions of p and m implies that
P (t1 ) = M.
Solving for P (t) gives λ(t), which in turn gives p(t).
11.5 Nonlinear examples

We now solve some nonlinear problems where it is possible to obtain an
explicit solution to the coupled state-costate equations given in the Minimum
Principle. At the same time, we also give several extensions of this result.
Our first extension involves constraints on the input. Suppose that U is
a subset of Rm , and that we require that u(t) ∈ U for all t. Since this is a
hard constraint, Theorem 11.1 is not directly applicable, but we have the
following simple extension:
Theorem 11.2 (Minimum Principle with constraints). Suppose that x(t1 )

is free, t1 is fixed, and suppose that u◦ is a solution to the optimal control
problem (11.1) under the constraint u ∈ U. That is,
V ◦ = min{V (u) : u(t) ∈ U for all t}.
We then have
(a) There exists a costate vector p(t) such that

u∈U
(b) The pair (p, x◦ ) satisfy the 2-point boundary value problem (11.4),
∂
x(t0 ) = x0 ; p(t1 ) = m(x(t1 ), t).
∂x
t
u
Example 11.5.1. To illustrate Theorem 11.2 consider the control of a bi-

linear system, defined by the differential equation
ẋ = ux, x(0) = x0 > 0, 0 ≤ u(t) ≤ 1.
Suppose that the goal is to make x large, while keeping the derivative of x
small on average. Then a reasonable cost criterion is
Z t1 Z t1
V (u) = ẋ(τ ) − x(τ ) dτ = [u(τ ) − 1]x(τ ) dτ.
0 0
We assume that t1 ≥ 1 is fixed, and that x(t1 ) is free.

The Hamiltonian becomes
H(x, p, u, t) = pf + ` = p(ux) + [u − 1]x

= x{u(p + 1) − 1}
and by the Minimum Principle, the optimal control takes the form

u∈U
Since x◦ (t) > 0 for all t, the minimization leads to


1
 if p(t) + 1 < 0
◦
u (t) = 0 if p(t) + 1 > 0

unknown if p(t) = −1

The costate variable and state variable satisfy the differential equations
∂H
ṗ = − = −(p + 1)u◦ + 1
∂x
∂H
x˙◦ = = u◦ x◦
∂p
with the boundary conditions
∂m
x◦ (0) = x0 , p(t1 ) = = 0.
∂x
How do we solve these coupled differential equations to compute u◦ ?
First note that if t = t1 then p(t) + 1 = 1 > 0 so u◦ (t) = 0. By continuity,
if t ≈ t1 , then p(t) > 0, so we still must have u◦ (t) = 0. Thus ṗ(t) = +1 for
t ≈ t1 , which implies
p(t) = t − t1 , t ≈ t1 .
As we move backwards in time from time t1 , the costate variable will con-
tinue to have this form up until the first time that p(t) + 1 changes sign.
This will occur when 0 = t − t1 + 1, or t = t1 − 1.
For t < t1 − 1 we then have p(t) + 1 < 0 so that u◦ (t) = 1. For these
time values, the costate vector is described by the equation
ṗ = −(p + 1) + 1 = −p.
Since p(t1 − 1) = −1, we obtain
p(t) = −e−t+t1 −1 , t < t1 − 1.
The costate trajectory is sketched in Figure 11.3.
u(t) = 1 u(t) = 0 u(t) = 1 u(t) = 0
-1
t
t
t1 - 1 t1 t1 - 1 t1
x(t)
0
p(t)
Figure 11.3: The costate trajectory, optimal control, and optimal state tra-
jectory for the bilinear model.
We conclude that the optimal control is bang-bang. That is, it takes on

only two values, which are the upper and lower limits:

1 t < t1 − 1
u(t) =
0 t > t1 − 1.
The time t1 − 1 is called the switching point. The optimal controlled state
trajectory is defined by the differential equation
(
◦ x◦ (t) t < t1 − 1
ẋ (t) =
0 t > t1 − 1
so the resulting state trajectory is expressed
x0 e t

◦ t < t1 − 1
x (t) =
x0 et1 −1 t > t1 − 1.
This is also illustrated in Figure 11.3. t
u
The next example we treat has a hard constraint on the final state value,
so that t0 , t1 , x(t0 ) = x0 , and x(t1 ) = x1 are all fixed and prespecified. In this
situation the Minimum Principle is the same, but the boundary conditions
for differential equations must be modified.
Theorem 11.3 (Minimum Principle with final value constraints). Suppose

that t0 , t1 , x(t0 ) = x0 , and x(t1 ) = x1 are prespecified, and suppose that u◦ is
a solution to the optimal control problem (11.1), subject to these constraints.
Then

u
x(t0 ) = x0 ; x(t1 ) = x1 .
t
u
Example 11.5.2. To illustrate the application of Theorem 11.3 we consider

the LQR problem where the terminal state has been specified. Note that in
general we cannot make the state stay at the terminal state x1 . The cost
criterion is
Z t1
V = 2 1
(xT Qx + uT Ru)dt R > 0, Q ≥ 0
0
H = p(Ax + Bu) + 21 (xT Qx + uT Ru).
The Minimum Principle again implies that the optimal control has the form
u◦ (t) = −R−1 B T p(t),
and that the costate vector satisfies the differential equation

∂H
ṗ(t) = − = −pT A − xT Q
∂x
Thus we again arrive at the linear differential equation based on the
Hamiltonian:
−BR−1 B T

ẋ A x x(t0 ) = x0
= T ; .
ṗ −Q −A p x(t1 ) = x1
Let ψ(t, τ ) denote the state transition matrix, so that

d
ψ(t, τ ) = H(t)ψ(t, τ ); ψ(t, t) = I.
dt
This is a 2n × 2n matrix which we decompose as

ψ11 (t, τ ) ψ12 (t, τ )
ψ(t, τ ) = .
ψ21 (t, τ ) ψ22 (t, τ )
To solve for the optimal control, we will compute the initial condition
p(t0 ) = p0 . This will allow us to compute p(t) for all t, and hence also
the optimal control expressed in terms of p above as u◦ = −R−1 B T p. The
unknown term p0 can be expressed in terms of the other known quantities
as follows:
x1 = x(t1 ) = ψ11 (t1 , t0 )x0 + ψ12 (t1 , t0 )p0
Assuming that ψ12 (t1 , t0 ) is invertible, this gives the formula
p0 = ψ12 (t1 , t0 )−1 {x1 − ψ11 (t1 , t0 )x0 }.
For all t we then have
p(t) = ψ21 (t, t0 )x0 + ψ22 (t, t0 )p0 .
The optimal control is
u◦ (t) = −R−1 B T p(t).
and the optimal state trajectory becomes
x◦ (t) = ψ11 (t, t0 )x0 + ψ12 (t, t0 )p0 , t ≥ t0 .
t
u
Our last extension concerns the case where the terminal time is not fixed.
At the same time, suppose that some components of x(t1 ) are specified. We
assume that t0 and x(t0 ) = x0 are fixed.
Theorem 11.4 (Minimum Principle with free terminal time). Suppose that
t0 and x(t0 ) = x0 are fixed, and that for some index set I ⊂ {1, . . . , n},
xi (t1 ) = x1i , if i ∈ I.
Suppose that u◦ is a solution to the optimal control problem (11.1), subject

to these constraints. Then

u
with the boundary conditions
xi (ti ) = x1i , i ∈ I
∂m ◦
pj (t1 ) = (x (t1 ), t1 ), i ∈ I c.
∂xj
The unspecified terminal time t1 satisfies

∂m ◦
(x (t1 ), t) + H(x◦ (t1 ), p(t1 ), u◦ (t1 ), t1 ) = 0. (11.15)
∂t
t
u
The most common example in which the terminal time is free is a mini-
mum time problem. As an example consider a single input linear state space
model
ẋ = Ax + bu.
We wish to find the input u◦ which drives x from x(0) = x0 to x(t1 ) = x1

in minimum time, under the constraint |u(t)| ≤ 1 for all t. Hence the cost
criterion we consider is
Z t1
V (u) = t1 = 1 dt,
0
so that ` ≡ 1, and m ≡ 0.
The Hamiltonian for this problem is
H = ` + pT f = 1 + pT Ax + pT bu.
Minimizing over all H again gives a bang-bang control law
p(t)T b < 0

◦ 1
u (t) =
−1 p(t)T b > 0
If A, b are time invariant and the eigenvalues {λi } of A are real, distinct,
and negative, then since the modes are all decreasing, the quantity p(t)T b
changes sign at most n − 1 times. Hence, the number of switching times for
the optimal control is also bounded above by n − 1.
The costate equation is
ṗ = −∇x H = −AT p.
Using the final time boundary condition (11.15), or H|t=t1 = 0, we obtain
1 + pT (t1 )Ax◦ (t1 ) + pT (t1 )bu◦ (t1 ) = 1 + pT (t1 )Ax◦ (t1 ) − |pT (t1 )b| = 0.
We now show how these equations may be solved using a second order
example.
Example 11.5.3. Consider the special case where the model is the double
integrator ÿ = u, which is described by the state space model with

0 1 0
A= b=
0 0 1
so that ẋ1 = x2 , ẋ2 = u. The optimal control has the form u◦ = −sgn(pT b) =
−sgn(p2 (t)). From the costate equations

ṗ1 T ṗ1
= −A
ṗ2 ṗ2
we find that for constants c1 and c2 ,
p1 (t) = c1 p2 (t) = −c1 t + c2 .
It follows that p2 (t) changes sign on (0, t1 ) at most once.

If p2 (t) < 0 then u◦ (t) = +1, so that by solving the differential equations
ẋ◦1 = x◦2 , ẋ◦2 = 1,
we find that x◦1 (t) = 21 (x◦2 (t))2 + K1 for some constant K1 . If p2 (t) > 0
so that u◦ (t) = −1, then for another constant, x◦1 (t) = − 21 (x◦2 (t))2 + K2 .
Hence the optimal state trajectories follow the quadratic paths illustrated
in Figure 11.4. t
u
x (t)
u >0
x0
u <0
y
x1
Figure 11.4: The optimal trajectories for the minimum time problem in this
two dimensional example follow quadratic paths, and may switch paths at
most once. A typical optimal path is illustrated as a union of two dark line
segments in this drawing.
11.6. EXERCISES 221
11.6 Exercises
11.6.1 Consider the system/performance index
T
1
Z
q
ẋ = u V (u) = x2 (T ) + u2 dt,
2 2 0
with x(0) = x0 , T given, and q ≥ 0. Find u◦ using the Minimum

Principle. First consider the case where ∞ > q > 0. Then, obtain the
control for “q = ∞”. That is, impose the constraint that x(T ) = 0.
Do the controllers converge as q → ∞?
11.6.2 A simplified model for a reactor is given by the bilinear system ẋ =
xu, where x is the concentration, and u is the rate, taken as the control
variable. Suppose that x(0) = 0.5, and the terminal time is t1 = 1.
Using the Minimum Principle, find the control function that minimizes
the performance index
Z 1
2
V = [x(1) − 1] + u2 (t) dt.
0
11.6.3 The temperature of a room is described by ẋ = ax+u , where x is

the room temperature, a is a negative constant, and u is the heat input
rate. Suppose that x(0) = x0 is given and it is desired to find a control
function u(·) so that x(t1 ) = q, where q > x0 , and the performance
index
Z t1
V = u2 (t) dt
0
is minimized. Note that, in this problem, the state trajectory x(·) is

specified at two end points, while the terminal time t1 is free. Using the
Minimum Principle, determine the optimal open-loop control function
and the corresponding value of t1 , in terms of a, x0 and q.
Bibliography
[1] B. D. O. Anderson and J. B. Moore. Optimal Control: Linear

Quadratic Methods. Prentice-Hall, Englewood Cliffs, NJ, 1990.
[2] T. Basar (edt). Control Theory: Twenty-Five Seminal Papers (1932-
1981). IEEE Press, 2001.
[3] R. Bellman. Introduction to Matrix Analysis. McGraw-Hill, New York,
NY, 2nd edition, 1970.
[4] R. W. Brockett. Finite Dimensional Linear Systems. Wiley, New York,
1970.
[5] W. L. Brogan. Modern Control Theory. Prentice Hall, Englewood
Cliffs, NJ, 1991. 3rd ed.
[6] C.-T. Chen. Linear System Theory and Design. Oxford University
Press, Oxford, 1999. 3rd ed.
[7] G. F. Franklin, J. D. Powell, and A. Emami-Naeini. Feedback Control
of Dynamic Systems. Addison-Wesley, 1991.
[8] R. E. Kalman and J. E. Bertram. Control system analysis and design
by the second method of Lyapunov. Trans. ASME Ser. D: J. Basic
Eng., 82, pp. 371–400, 1960.
[9] J. LaSalle and S. Lefshetz. Stability by Liapunov’s Direct Method.
Academic, New York, NY, 1961.
[10] D. G. Luenberger. Optimization by Vector Space Methods. John Wiley
and Sons, New York, NY, 1969.
[11] M. W. Spong and M. Vidyasagar. Robot Dynamics and Control. John
Wiley and Sons, New York, NY, 1989.
[12] L. A. Zadeh and C. A. Desoer. Linear System Theory: The State
Space Approach. McGraw-Hill, 1969.
223
Index
adjoint, 38 generalized eigenvectors, 34

adjoint equation, 62 globally asymptotically stable, 68
algebraic Riccati equation (ARE), 181 Grammian, 37
asymptotically stable, 68
Hamilton-Jacobi-Bellman (HJB) equa-
bang-bang, 215 tion, 168, 170
basis, 26 Hamiltonian, 170
bounded, 82 Hamiltonian matrix, 175
bounded input/bounded output sta- Hermitian, 76
ble, 82 Hurwitz, 69
characteristic equation, 33 image, 29

characteristic polynomial, 33 internal model principle, 146
closed-loop system, 121 invariant, 75
controllability grammian, 93
controllability matrix, 92 Jordan form, 35
controllable, 90
controllable canonical form, 11 Kalman Controllability Canonical Form,
111
controllable subspace, 97
Kalman Observability Canonical Form,
detectable, 127 112
dimension, 26 Kalman’s inequality, 191
Diophantine equation, 151
Lagrange multipliers, 206
dual, 38, 110
linear, 29
eigenvalue, 32 linear operator, 29
eigenvector, 32 linear quadratic regulator, 168
equilibrium, 67 linear state feedback, 121
equilibrium state, 4 linearity, 1
linearly independent, 25
Feedback, 155 loop gain, 158
field, 23 loop transfer function, 187
fundamental matrix, 56 Lyapunov equation, 61, 77, 84
224
INDEX 225
Lyapunov function, 84 stable in the sense of Lyapunov, 67

stable subspace, 79
minimal, 113 state transition matrix, 48, 57
minimum phase, 162 stationary point, 85, 208
modal form, 14 strictly proper, 13
modes, 14 subspace, 25
Sylvester equations, 152
natural basis, 27 symmetric root locus, 188
norm, 36
nullspace, 30 tracking, 155
transmission zero, 162
observability grammian, 108
observability matrix, 108 value function, 168
observable, 107 vector space, 24
observable canonical form, 13
orthogonal, 37
orthonormal, 37
Partial fraction expansion, 13

Peano-Baker series, 58
performance, 155
positive definite, 45, 72, 76
positive semi-definite, 45, 77
principle minors, 77
principle of optimality, 169
range, 29
rank, 29
realization, 10
reciprocal, 38
regulation, 155
return difference, 158
return difference equation, 187
Riccati Differential Equation, 174
semigroup property, 49, 57

sensitivity, 158
separation principle, 129
similar, 32
square, 161
stabilizable, 123, 125

Control System Theory and Design: ECE 515 Class Notes

Uploaded by

Copyright:

Available Formats

Control System Theory and Design: ECE 515 Class Notes

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Control System Theory and Design: ECE 515 Class Notes

Uploaded by

Copyright:

Available Formats

ECE 515 Class Notes

CONTROL SYSTEM THEORY

Tamer Basar, Sean P. Meyn, and William R. Perkins

University of Illinois at Urbana-Champaign

TB, SPM, WRP

I System Modeling and Analysis 3

1 State Space Models 1

3 Solutions of State Equations 47

3.3 Change of state variables . . . . . . . . . . . . . . . . . . . . 50

II System Structural Properties 63

6 Observability, Duality and Minimality 105

III Feedback 119

7 Pole Placement 121

8 Tracking and Disturbance Rejection 145

9 Control Design Goals 155

IV Optimal Control 165

10 Dynamic Programming and the HJB Equation 167

11 An Introduction to the Minimum Principle 203

1.1 Magnetically Suspended Ball . . . . . . . . . . . . . . . . . . 2

4.1 Typical solutions of the linear state space model . . . . . . . 65

5.1 The controlled model can be steered to any state . . . . . . . 90

6.5 A block diagram realization of P . . . . . . . . . . . . . . . . . 115

7.1 Observer design . . . . . . . . . . . . . . . . . . . . . . . . . . 126

8.1 Observer/state feedback . . . . . . . . . . . . . . . . . . . . . 150

9.1 An error feedback configuration . . . . . . . . . . . . . . . . . 158

10.1 If a better control existed... . . . . . . . . . . . . . . . . . . . 169

11.1 An optimization problem on R2 with a single constraint . . . 207

System Modeling and

State Space Models

the approximation of even a grossly nonlinear model by a linear one.

Figure 1.1: Magnetically Suspended Ball

1.1 An electromechanical system

in this chapter. The input u is the current applied to the electro-magnet,

ẋ1 = x2 = f1 (x1 , x2 , u) (1.2)

It is nonlinear, since f2 is a nonlinear function of xx12 . Letting x = x1

The motion of a typical solution to a nonlinear state space model in R2 is

1.2 Linearization about an equilibrium state

From the definition of f we must have xe2 = 0, and

Figure 1.2: Trajectory of a nonlinear state space model in two dimensions:

which is unique when restricted to be positive. The state xe is called an

x1 (t) = xe1 + δx1 (t)

δ ẋ1 = xe2 + δx2 (t) = δx2 (t)

After computing partial derivatives we obtain the formulae

This linearized system is only an approximation, but one that is reasonable

1.3 Linearization about a trajectory

ẋ (t) = f (x(t), u(t), t)

where f is a continuously differentiable (C 1 ) function of its arguments. Sup-

x(t) = xn (t) + δx(t)

From a Taylor series expansion we then have

where the higher order terms have been ignored.

1.4 A two link inverted pendulum

Figure 1.3: The Pendubot

d11 q̈1 + d12 q̈2 + h1 + φ1 = τ (1.4)

where q1 , q2 are the joint angles and

d11 = m1 `2c1 + m2 (`21 + `2c2 + 2`1 `c2 cos(q2 )) + I1 + I2

Figure 1.4: Coordinate description of the pendubot: `1 is the length of the

This model may be written in state space form as a nonlinear vector

Figure 1.5: There is a continuum of different equilibrium positions for the

1.5 An electrical circuit