Exercises in

Nonlinear Control Systems

Mikael Johansson
Bo Bernhardsson
Karl Henrik Johansson
March 9, 1998

Latest update

February 11, 2004

1
Introduction
The exercises are divided into problem areas that roughly match the lecture
schedule. Exercises marked “PhD” are harder than the rest. Some exercises
require a computer with software such as Matlab and Simulink.
Many people have contributed to the material in this compendium.
Apart from the authors, exercises have been suggested by Lennart Ander-
sson, Anders Robertsson and Magnus Gäfvert. Exercises have also shame-
lessly been borrowed (=stolen) from other sources, mainly from Karl Johan
Åström’s compendium in Nonlinear Control.

The authors, Jan 1999

2
1. Nonlinear Models and
Simulation
EXERCISE 1.1[KHALIL, 1996]
The nonlinear dynamic equation for a pendulum is given by

mlθ¨ = −mg sin θ − klθ˙ ,

where l is the length of the pendulum, m is the mass and θ is the angle
subtended by the rod and the vertical axis through the pivot point, see
Figure 1.1.

Figure 1.1 The pendulum in Exercise 1.1

(a) Choose appropriate state variables and write down the state equa-
tions.
(b) Find all equilibria of the system.
(c) Linearize the system around the equilibrium points, and determine
if the system equilibria are locally asymptotically stable.

EXERCISE 1.2[KHALIL, 1996]

The nonlinear dynamic equations for a single-link manipulator, see Fig-
ure 1.2, with flexible joints, damping ignored, is given by

I q̈1 + M g sin q1 + k( q1 − q2 ) = 0
J q̈2 − k( q1 − q2 ) = u,

where q1 and q2 are angular positions, I and J are moments of inertia,

k is a spring constant, M is the total mass, L is a distance, and u is a
torque input. Choose state variables for this system and write down the
state equations.

3
 Chapter 1. Nonlinear Models and Simulation

q1

Figure 1.2 The flexible manipulator in Exercise 1.2

EXERCISE 1.3[KHALIL, 1996]

A synchronous generator connected to a strong electrical bus can be mod-
eled by

M δ¨ = P − Dδ˙ − η 1 Eq sin δ
τ Ėq = −η 2 Eq + η 3 cos δ + EF D ,

where δ is the rotor deflection angle in radians, Eq is voltage, P is me-

chanical input power, EF D is field voltage, D is damping coefficient, M is
inertial coefficient, τ is a time constant, and η 1 , η 2 , and η 3 are constant
parameters.
(a) Using E FD as input signal and δ , δ˙, and E as state variables, find
q
the state equation.
(b) Suppose that τ is relatively large so that Ėq  0. Show that assuming
Eq to be constant reduces the model to a pendulum equation.
(c) For the simplified model, derived in (b), find all equilibrium points.

EXERCISE 1.4

ψ ( t, y )

−
r u y
C(sI − A)−1 B

Figure 1.3 The feedback system in Exercise 1.4

Figure 1.3 shows a feedback connection of a linear time-invariant system

and a nonlinear time-varying element. The variables r, u and y are vectors
of the same dimension, and ψ (t, y) is a vector-valued function.
(a) Find a state-space model with r as input and y as output.
(b) Rewrite the pendulum model from Exercise 1.1 into the feedback
connection form described above.

4
 Chapter 1. Nonlinear Models and Simulation

EXERCISE 1.5[KHALIL, 1996]

+
θi sin(⋅) G ( s) y
−

θ0 1
s

Figure 1.4 The phase-locked loop in Exercise 1.5

A phase-locked loop can be represented by the block diagram of Figure 1.4.

Let { A, B , C} be a state-space representation of the transfer function G (s).
Assume that all eigenvalues of A have negative real parts, G (0) = 0 and
that θ i is constant. Let z be the state of the realization { A, B , C}.
(a) Show that

ż = Az + B sin e
ė = − Cz

is a state equation for the closed-loop system.

(b) Find all equilibrium points of the system.
(c) If G (s) = 1/(τ s + 1), the closed-loop model coincides with the model
of a pendulum with certain conditions on m, l, k, and g (as given in
Exercise 1.1), what conditions?

EXERCISE 1.6

F
Friction
−
xr x
u 1
PID Σ ms v
1
s

Figure 1.5 Control system with friction in Example 1.6.

Figure 1.5 shows a block diagram of a mechanical system with friction

under PID control. The friction block is given by

F (v) = F0 sign(v)

Let xr = 0 and rewrite the system equations into feedback connection form.

5
 Chapter 1. Nonlinear Models and Simulation

EXERCISE 1.7

Gaw
- +

r u v y
Gf f + Gp
-

Gf b

Figure 1.6 Anti-windup compensation in Example 1.7.

Figure 1.6 illustrates one approach to avoid integrator windup. Rewrite

the system into feedback connection form.

EXERCISE 1.8
Consider the model of a motor with a nonlinear valve in Figure 1.7. Assume

Motor Valve Process

r 1 y
Σ 1
s (s+1)2

−1

Figure 1.7 Block diagram of system in Exercise 1.8.

that the valve characteristic is given by f ( x) = x2 (this is unrealistic for

x < 0).
(a) Choose appropriate state variables and write down the state equa-
tions.
(b) For which constant input amplitudes r > 0 is the system locally
stable?
(c) What would be a more realistic valve model for x < 0?

EXERCISE 1.9
Is the following system (a controlled nonlinear spring) nonlinear locally
controllable around x = ẋ = u = 0?

ẍ = − k1 x − k2 x3 + u.

6
 Chapter 1. Nonlinear Models and Simulation

EXERCISE 1.10PHD
The equations for the unicycle in Figure 1.8 are given by

θ
( x, y)
Figure 1.8 The “unicycle used in Exercise 1.10.

ẋ = u1 cos θ
ẏ = u1 sin θ
θ˙ = u2 ,

where ( x, y) is the position and θ the angle of the wheel. Is the system
nonlinear locally controllable at (0, 0, 0)? (Hint: Linearization gives no in-
formation; use the definition directly).

EXERCISE 1.11PHD
The system in Figure 1.9 is known as the “rolling penny”. The equations

θ
( x, y)
Figure 1.9 The “rolling penny” used in Exercise 1.11.

are given by

ẋ = u1 cos θ
ẏ = u1 sin θ
θ˙ = u2
Ψ̇ = u1 .

Is the system nonlinear locally controllable at (0, 0, 0, 0)?

7
 Chapter 1. Nonlinear Models and Simulation

EXERCISE 1.12
Determine if the following system is nonlinear locally controllable at ( x0 , u0 ) =
(0, 0)

ẋ1 = cos( x1 ) − 1 + x22 + u

ẋ2 = sin( x1 ) + u2 .

EXERCISE 1.13
Simulate the system G (s) = 1/(s + 1) with a sinusoidal input u = sin ω t.
Find the amplitude of the stationary √ output for ω = 0.5, 1, 2. Compare with
the theoretical value h G (iω )h = 1/ 1 + ω 2 .

EXERCISE 1.14
Consider the pendulum model given in Exercise 1.1.
(a) Make a simulation model of the system in Simulink, using for in-
stance m = 1, g = 10, l = 1, k = 0.1. Simulate the system from
various initial states. Is the system stable? Is the equilibrium point
unique? Explain the physical intuition behind your findings.
(b) Use the function linmod in Matlab to find the linearized models for
the equilibrium points. Compare with the linearizations that you de-
rived in Exercise 1.1.
(c) Use a phase plane tool (such as pptool) to construct the phase plane
of the system. Compare with the results from (a).

EXERCISE 1.15
Simulate the example from the lecture with two tanks, using the models

ḣ = (u − q)/ A
p √
q = a 2g h,

where h is the liquid level, u is the inflow to the tank, q the outflow, A
the cross section area of the tank, a the area of the outflow and g the
acceleration due to gravity, see Figure 1.10. Use a step input flow. Make a
step change in u from u = 0 to u = c, where c is chosen in order to give a
stationary value of the height, h = 0.1. Make a step change from u = c to
u = 0. Is the process linear? Linearize the system around h1 = h2 = 0.1.
Use A1 = A2 = 3  10−3 , a1 = a2 = 7  10−6. What are the time constants
of the linearized system?

8
 Chapter 1. Nonlinear Models and Simulation

1 q
1/A f(u) 1
1 s
q q In
In Sum Gain Integrator Fcn h 1
1 In
h Out
2 In Subsystem2
h Subsystem

Figure 1.10 The flow system in Exercise 1.15

EXERCISE 1.16
Simulate the system with the the oscillating pivot point (the “electric hand-
saw”), see Figure 1.11. Use the equation

Figure 1.11 The inverted pendulum on an electric hand-saw in Exercise 1.16

1
θ¨ (t) = (g + aω 2 sin ω t)θ (t).
l
Assume a = 0.02m and ω = 2π ⋅ 50 for a hand-saw. Use simulation to find
for what length l the system is locally stable around θ = θ˙ = 0 (Note:
asymptotic stability is not required).

EXERCISE 1.17
The Lorentz equations
d
x1 = σ ( x2 − x1 )
dt
d
x2 = rx1 − x2 − x1 x3
dt
d
x3 = x1 x2 − bx3, σ , r, b > 0,
dt
where σ , r, b are constants, are often used as example of chaotic motion.
(a) Determine all equilibrium points.
(b) Linearize the equations around x = 0 and determine for what σ , r, b
this equilibrium is locally asymptotically stable.

9
2. Linearization and
Phase-Plane Analysis
EXERCISE 2.1[KHALIL, 1996]
For each of the following systems, find and classify all equilibrium points.

(a) ẋ1 = x2
ẋ2 = − x1 + x13 /6 − x2

(b) ẋ1 = − x1 + x2
ẋ2 = 0.1x1 − 2x2 − x12 − 0.1x13

(c) ẋ1 = (1 − x1 ) x1 − 2x1 x2 /(1 + x1 )

ẋ2 = (1 − x2 /(1 + x1 )) x2

(d) ẋ1 = x2
ẋ2 = − x1 + x2 (1 − 3x12 − 2x22 )

(e) ẋ1 = − x1 + x2 (1 + x1 )
ẋ2 = − x1 (1 + x1 )

(f) ẋ1 = ( x1 − x2 )( x12 + x22 − 1)

ẋ2 = ( x1 + x2 )( x12 + x22 − 1)

EXERCISE 2.2[ÅSTRÖM, 1968]

For all positive values of a, b and c, determine the equilibrium points of
the system

ẋ1 = ax1 − x1 x2
ẋ2 = bx12 − cx2

and determine the type of equilibrium.

10
 Chapter 2. Linearization and Phase-Plane Analysis

EXERCISE 2.3[KHALIL, 1996]

For each of the following systems, construct the phase portrait, preferably
using a computer program, and discuss the qualitative behaviour of the
system.

(a) ẋ1 = x2
ẋ2 = x1 − 2 tan−1 ( x1 + x2 )

(b) ẋ1 = x2
ẋ2 = − x1 + x2 (1 − 3x12 − 2x22 )

(c) ẋ1 = 2x1 − x1 x2

ẋ2 = 2x12 − x2

EXERCISE 2.4
Saturations constitute a severe restriction for stabilization of system. Fig-
ure 2.1 shows three phase portraits, each corresponding to one of the fol-
lowing linear systems under saturated feedback control.

(a) ẋ1 = x2
ẋ2 = x1 + x2 − sat(2x1 + 2x2 )

(b) ẋ1 = x2
ẋ2 = − x1 + 2x2 − sat(3x2 )

(c) ẋ1 = x2
ẋ2 = −2x1 − 2x2 − sat(− x1 − x2 )

Which phase portrait belongs to what system?

Phase plane Phase plane Phase plane

2 1 2

0.8
1.5 1.5

0.6
1 1
0.4

0.5 0.5
0.2
x2
x2

x2

0 0 0

−0.2
−0.5 −0.5

−0.4
−1 −1
−0.6

−1.5 −1.5
−0.8

−2 −1
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2 −1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2
x1 x1 x1

Figure 2.1 Phase portraits for saturated linear systems in Exercise 2.4

11
 Chapter 2. Linearization and Phase-Plane Analysis

EXERCISE 2.5[KHALIL, 1996]

The phase portraits of the following two systems are shown in Figures 2.2(a),
and 2.2(b), respectively. Mark the arrow heads and discuss the qualitative
behaviour of each system.

(a) ẋ1 = − x2
ẋ2 = x1 − x2 (1 − x12 + 0.1x14 )

(b) ẋ1 = x2
ẋ2 = x1 + x2 − 3 tan−1 ( x1 + x2 )

Phase plane Phase plane

8 8

6 6

4 4

2 2
x2

x2
0 0

−2 −2

−4 −4

−6 −6

−8 −8
−8 −6 −4 −2 0 2 4 6 8 −8 −6 −4 −2 0 2 4 6 8
x1 x1

Figure 2.2 Phase portraits for Exercise 2.5(a) to the left, and Exercise 2.5(b) to
the right.

EXERCISE 2.6
The following system

ẋ1 = (u − x1 )(1 + x22 )

ẋ2 = ( x1 − 2x2 )(1 + x12 )
y = x2

is controlled by the output feedback

u = −K y

(a) For all values of the gain K, determine the equilibrium points of the
closed loop system.
(b) Determine the equilibrium character of the origin for all values of
the parameter K. Determine in particular for what values the closed
loop system is (locally) asymptotically stable.

12
 Chapter 2. Linearization and Phase-Plane Analysis

EXERCISE 2.7[ÅSTRÖM, 1968]

As an application of phase plane analysis, consider the model of a syn-
chronous generator derived in Exercise 1.3(b):

ẋ1 = x2
P D η1
ẋ2 = − x2 − Eq sin x1 .
M M M

The equilibrium points of the system where derived in Exercise 1.3(c).Determine

the character of the equilibrium points.

EXERCISE 2.8
Linearize

ẋ1 = x2 + x1 (1 − x12 − x22 )

ẋ2 = − x1 + x2 (1 − x12 − x22 )

around the trajectory ( x1 , x2 ) = (sin t, cos t). Also determine whether this
limit cycle is stable or not.

EXERCISE 2.9
Linearize the ball-on-beam equation

7 2r
ẍ − xφ˙2 = g sin φ + φ¨,
5 5

around the trajectory

 

5g t2
φ (t), x(t) = φ 0 , sin(φ 0 ) ⋅
7 2

EXERCISE 2.10
Use a simple trigonometry identity to help find a nominal solution corre-
sponding to u(t) = sin (3t), y(0) = 0, ẏ(0) = 1 for the eqution

4 3 1
ÿ + y (t) = − u(t).
3 3

Linearize the equation around this nominal solution.

13
 Chapter 2. Linearization and Phase-Plane Analysis

EXERCISE 2.11
The equations for motion of a child on a swing are given by

d d
(ml 2 φ ) + mg l sin φ = 0
dt dt

Here φ (t) is the angle of the swing, m the mass, and l (t) the distance of the
child to the pivot of the swing. The child can excite the swing by changing
l (t) by moving its center of mass.
(a) Draw phase diagrams for two different constant lenghts l1 and l2 .
(b) Assume that it is possible to quickly change between the lenghts
l1 and l2. Show how to jump between the two different systems to
increase the amplitude of the swing.
Hint: During constant l the energy in the system is constant. When l (t)
changes quickly φ will be continuous but dt
d
φ (t) will change in such a way
that the angular momentum ml dt φ is continuous.
2 d

14
3. Lyapunov Stability
EXERCISE 3.1
Consider the scalar system

ẋ = ax3

(a) Show that Lyapunov’s linearization method fails to determine stabil-

ity of the origin.
(b) Use the Lyapunov function

V ( x) = x4

to show that the system is globally asymptotically stable for a < 0

and unstable for a > 0.
(c) What can you say about the system for a = 0?

EXERCISE 3.2
Consider the pendulum equation

ẋ1 = x2
g k
ẋ2 = − sin x1 − x2 .
l m

(a) Assume zero friction, i.e. let k = 0, and show that the origin is stable.
(Hint. Show that the energy of the pendulum is constant along all
system trajectories.)
(b) Can the pendulum energy be used to show asymptotic stability for the
pendulum with non-zero friction, k > 0? If not, modify the Lyapunov
function to show asymptotic stability of the origin.

EXERCISE 3.3
Consider the system

ẍ + d ẋ3 + kx = 0,

where d > 0, k > 0. Show that

1

V ( x) = kx2 + ẋ2
2

is a Lyapunov function. Is the system locally stable, locally asymptotically

stable, and globally asymptotically stable?

15
 Chapter 3. Lyapunov Stability

EXERCISE 3.4
Consider the linear system
 
0 −1
ẋ = Ax = x
1 −1

(a) Compute the eigenvalues of A and verify that the system is asymp-
totically stable
(b) From the lectures, we know that an equivalent characterization of
stability can be obtained by considering the Lyapunov equation

AT P + PA = − Q

where Q = Q T is any positive definite matrix. The system is asymp-

totically stable if and only if the solution P to the Lyapunov equation
is positive definite.
(i) Let
 
p11 p12
P=
p12 p22

Verify by completing squares that V ( x) = x T Px is a positive

definite function if and only if

p11 > 0, p11 p22 − p212 > 0

(ii) Solve the Lyapunov function with Q as the identity matrix. Is

the solution P a positive definite matrix?
(c) Solve the Lyapunov equation in Matlab.

EXERCISE 3.5[SLOTINE AND LI, 1991]

As you know, the system

ẋ (t) = Ax(t), t ≥ 0,

is asymptotically stable if all eigenvalues of A have negative real parts. It

might be tempted to conjecture that the time-varying system

ẋ (t) = A(t) x(t), t ≥ 0, (3.1)

is asymptotically stable if the eigenvalues of A(t) have negative real parts

for all t ≥ 0. This is not true.
(a) Show this by explicitly deriving the solution of
 
−1 e2t
ẋ = x, t ≥ 0.
0 −1

16
 Chapter 3. Lyapunov Stability

(b) The system (3.1) is however stable if the eigenvalues of A(t) + AT (t)
have negative real parts for all t ≥ 0. Prove this by showing that
V = x T x is a Lyapunov function.

EXERCISE 3.6[BOYD, 1997]

A student is confronted with the nonlinear differential equation

2x
ẍ + =0
(1 + x2 )2

and is asked to determine whether or not the equation is stable. The stu-
dents think “this is an undamped mass-spring system – the spring is non-
linear with a spring constant of 2/(1 + x2 )2 ”. The student re-writes the
system as

ẋ1 = x2
−2x1
ẋ2 =
(1 + x12 )2

and constructs the obvious Lyapunov function

Z x1
2ζ 1
V ( x) = dζ + x22 .
0 (1 + ζ 2)2 2

The student declares, “V is positive definite, because everywhere in IR2 ,

V ( x) ≥ 0, and V ( x) = 0 only if x = 0.” The student ascertains that V̇ ≤ 0
everywhere in IR2 and concludes, “the conditions for Lyapunov’s theorem
are satisfied, so the system is globally stable about x = 0”.
(a) Sadly, there is a mistake in the student’s reasoning. What is the
mistake?
(b) Perhaps the student has merely made a poor choice of Lyapunov
function, and the system really is globally stable. Is there some other
Lyapunov function that can be used to show global stability? Find
such a function, or show that no such function exists.

EXERCISE 3.7[SLOTINE AND LI, 1991]

Consider the system

ẋ1 = 4x12 x2 − f 1 ( x1 )( x12 + 2x22 − 4)

ẋ2 = −2x13 − f 2( x2 )( x12 + 2x22 − 4),

where the continuous functions f 1 and f 2 have the same sign as their
arguments, i.e. xi f i ( xi ) ≥ 0 (note that f i(0) = 0). Show that almost all
trajectories of the system tend towards the invariant set x12 + 2x22 = 4
independently of the explicit expressions of f 1 and f 2. However, this set
will NOT be a limit cycle as the system will have singular points belonging

17
 Chapter 3. Lyapunov Stability

to this set. What equilibrium points does the system have? Show that
x12 + 2x22 = 4 is an attractive invariant set. How will the trajectories of the
system behave?
Extra: Simulate the system.
(Remark. Compare with Example 3.13 in the book by Slotine and Li
where the invariant set will be an limit cycle.)

EXERCISE 3.8
Consider the system

ẋ1 = x2
ẋ2 = −2x1 − 2x2 − 4x13 .

Use the function

V ( x) = 4x12 + 2x22 + 4x14

to show that
(a) the system is globally stable around the origin.
(b) the origin is globally asymptotically stable.

EXERCISE 3.9
Consider the system

ÿ = sat(−3 ẏ − 2 y).

(a) Show that y(t) → 0 as t → 0.

(b) For PhD students. Is it possible to prove global asymptotic stability
using a Lyapunov function V ( x) that satisfies

α hh xhh22 ≤ V ( x) ≤ β hh xhh22 , V̇ ( x) ≤ −γ hh xhh22

for some positive scalars α and β ?

(c) For PhD students. Consider the system

ẍ = u

and show that all feedback laws u = k1 x + k2 ẋ that give an asymptot-

ically stable system, also give an asymptotically stable system when
the actuator saturates, i.e., when

ẍ = sat(u).

(d) For PhD students. Does the results in (c) hold for the triple integrator

d3 x
= sat(u)? (3.2)
dt3

18
 Chapter 3. Lyapunov Stability

EXERCISE 3.10[BOYD, 1997]

The set { z ∈ IRnh V̇ ( z) = 0} arising in invariance theorems (such as LaSalle’s
theorem) is usually a ‘thin’ hypersurface, but it need not be. Carefully prove
global asymptotic stability of

ẋ1 = x2
ẋ2 = − x1 − max(0, x1 ) ⋅ max(0, x2 )

using the Lyapunov function V ( x) = x T x.

EXERCISE 3.11
Consider the nonlinear system

ẋ1 = − x1 + x2
ẋ2 = − x1 − x2 + g( x)

(a) Show that V ( x) = 0.5x T x is a Lyapunov function for the system when
g( x) = 0.
(b) Use this Lyapunov function to show that the system is globally asymp-
totically stable for all g( x) that satisfy

g( x) = g( x2 )

and sign(g( x2 )) = −sign( x2 ).

(c) Let g( x) = x23 . This term does not satisfy the conditions in (a). How-
ever, we can apply Lyapunov’s linearzation method to show that the
origin is still locally asymptotically stable.
For large initial values, on the other hand, simulations reveal that
the system is unstable. It would therefore be interesting to find the
set of “safe” initial values, such that all trajectories that start in this
set tend to the origin. This set is called the region of attraction of the
origin. We will now illustrate how quadratic Lyapunov functions can
be used to estimate the region of attraction.
(i) Show that V̇ ( x) < 0 for h x2 h < 1. This means that V ( x) decreases
for all solutions that are confined in the strip h x2 (t)h ≤ 1 for all
t.
(ii) Recall that level sets for the Lyapunov function are invariant.
Thus, solutions that start inside a proper level set remain there
for all future times. Conclude that the region of attraction can
be estimated as the largest level set

Ω = { x : V ( x) ≤ γ }

for which h x2 h < 1. Compute the maxiumum value of γ such that

Ω is a region of attraction.

19
 Chapter 3. Lyapunov Stability

EXERCISE 3.12[KHALIL, 1996]

Consider the second order system

ẋ1 = − x2
ẋ2 = x1 + ( x12 − 1) x2 .

(This is the Van der Pol oscillator ran backwards in time.)

(a) Use linearization and a quadratic Lyapunov function to show that
the origin is asymptotically stable.
(b) Estimate the region of attraction for the nonlinear system using the
quadratic Lyapunov derived in (a). (Hint. Transform the system into
polar coordinates, x1 = r cos(θ ), x2 = r sin(θ ) and find the largest
radius r so that Lyapunov function is decreasing for all x in the ball
Br = { x ∈ IR2 : i xi2 ≤ r}.).
(c) The procedure used in (b) tends to give quite conservative estimates
of the region of attraction. Can you think of some method to get
better estimates of the region of attraction, or to enlarge the estimate
derived above?

EXERCISE 3.13[KHALIL, 1996]

Consider the system

ẋ1 = x2
ẋ2 = x1 − sat(2x1 + x2 ).

(a) Show that the origin is asymptotically stable.

(b) Show that all trajectories starting in the first quadrant to the right
of the curve

x1 x2 = c

for sufficiently large c, cannot reach the origin. (Hint: Consider V ( x) =

x1 x2 ; calculate V̇ ( x) and show that on the curve V ( x) = c, the deriva-
tive V̇ ( x) > 0 when c is sufficiently large.)
(c) Show that the origin is not globally asymptotically stable.

EXERCISE 3.14[BOYD, 1997]

So far, we have only considered stability of autonomous systems, i.e. sys-
tems without an external input. If we are faced with an open-loop unstable
system with an input that we can manipulate, a key question is whether
it is possible to find a control law that gives a stable closed-loop system.
We can use Lyapunov theory to give some answers to this question.
We say that the single-input system

ẋ = f ( x, u), x ∈ IRn, u ∈ IR

is stabilizable if there is a state feedback u = k( x) that results in a globally

asymptotically stable closed-loop system.

20
 Chapter 3. Lyapunov Stability

(a) Consider the special case when

ẋ = f ( x, u) = φ ( x) + ψ ( x)u,

and show that the system is stabilizable if there is positive definite

function V , such that the function
    2
VV VV
φ ( x) − ψ ( x)
Vx Vx

is negative definite. (Hint. Try to find a feedback in terms of V ( x),

φ ( x), and ψ ( x) that makes V ( x) a Lyapunov function for the closed
loop system.)
(b) For PhD students. Show that the linear system ẋ = Ax + bu is stabi-
lizable if and only if there is a P = P T such that

AP + PAT − bb T < 0.

(Hint. Some LQR theory may come handy when proving necessity. In
particular, if the system is stabilizable, what can you say about
R∞ the
feedback law u = − kx that you obtain from the LQR cost 0 x T x +
uT u dt?)

EXERCISE 3.15
It can sometimes be convenient to re-write nonlinearities in a way that
is more easy to manipulate. Consider the single input, open loop stable,
linear system under saturated feedback

ẋ = Ax + Bsat(u)
u = − K x.

(a) Show that this system can be re-written in the form

ẋ = Ax + µ ( x) B K x,

where 0 < µ ( x) ≤ 1.
(b) Assume P > 0 is such that

x T ( AT P + PA) x ≤ 0, ∀ x

. Show that all feedback gains K that satisfies

x T (( A − B K )T P + P( A − B K )) x ≤ 0, ∀ x

guarantees the closed loop system in (a) to be stable. (The nice thing
about this formulation is that it is possible to construct efficient nu-
merical methods for simultaneously finding both feedback gains K
and Lyapunov matrix P).

21
 Chapter 3. Lyapunov Stability

(c) For PhD students. Consider the nonlinear system

ẋ = Ax + f ( x) + Bsat(u)
u = − K x.

Assume that the perturbation term satisfies f (0) = 0 and that f ( x)

is globally Lipschitz with Lipschitz constant k f , i.e.,

h f ( x) − f ( y)h ≤ k f h x − yh, k f > 0.

Let Q be given by the Lyapunov equation AT P + PA = − Q, with

P > 0. Show that if the Lipschitz constant satisfies

λ min( Q)
kf < ,
2λ max ( P)

then the system is globally stabilizable by linear feedback. Also, sug-

gest a feedback law that stabilizes the system.

EXERCISE 3.16
In general, it is non-trivial to find a Lyapunov function for a given nonlinear
system. Several different methods have been derived for specific classes of
systems. In this exercise, we will investigate the following method, known
as Krasovskii’s method.
Consider systems on the form

ẋ = f ( x)

with f (0) = 0. Assume that f ( x) is continuously differentiable and that

its Jacobian, V f /V x, satisfies
 T
Vf Vf
P ( x) + ( x) P ≤ −I
Vx Vx

for all x ∈ IRn , and some matrix P = P T > 0. Then, the origin is globally
asymptotically stable with V ( x) = f T ( x) P f ( x) as Lyapunov function.
Prove the validity of the method in the following steps.
(a) Verify that f ( x) can be written as
Z 1
Vf
f ( x) = (σ x) ⋅ x dσ .
0 Vx

and use this representation to show that the assumptions imply

x T P f ( x) + f T ( x) Px ≤ − x T x, ∀ x ∈ IRn

(b) Show that V ( x) = f T ( x) P f ( x) is positive definite for all x ∈ IRn.

(c) Show that V ( x) is radially unbounded.

22
 Chapter 3. Lyapunov Stability

(d) Using V ( x) as a Lyapunov function candidate, show that the origin

is globally asymptotically stable.

EXERCISE 3.17
Use Krasovskii’s method to justify Lyapunov’s linearization method.

EXERCISE 3.18[ÅSTRÖM, 1968]

Consider the servo system in Figure 3.18. Introduce state variables x1 and

e x2 x1
K 1
+ g( e) s+1 s

−1

Figure 3.1 The pendulum in Exercise 1.1

x2 as indicated in the figure. Assume that the reference value is zero. The
system equations can then be written as

ẋ1 = x2
ẋ2 = − x2 + K g( e) = − x2 + K g(− x1 ).

Let the nonlinearity be on the form g( e) = e3 and investigate stability

of the closed loop system. (Hint: Use V ( x) = f ( x)T P f ( x) (Krassovskii’s
method) with suitable P.)

23
4. Input-Output Stability
EXERCISE 4.1
The norms used in the definitions of stability need not be the usual Euclid-
ian norm. If the state-space is of finite dimension n (i.e., the state vector
has n components), stability and its type are independent of the choice of
norm (all norms are “equivalent”), although a particular choice of norm
may make analysis easier. For n = 2, draw the unit balls corresponding to
the following norms.
(a) hh xhh2 = x12 + x22 (Euclidian norm)
2
(b) hh xhh = x12 + 5x22
(c) hh xhh = h x1 h + h x2 h
(d) hh xhh = sup(h x1 h, h x2 h)
Recall that a “ball” B ( x0 , R), of center x0 and radius R, is the set of x such
that hh x − x0 hh ≤ R, and that the unit ball is B (0, 1).

EXERCISE 4.2
Consider a linear time invariant system G (s) interconnected with a static
nonlinearity ψ ( y) in the standard form. Compare the Nyquist, Circle, Small
Gain, and Passivity Criterion with respect to the following issues.
(a) What are the restrictions that must be imposed on ψ ( y) in order to
apply the different stability criteria?
(b) What restrictions must be imposed on the Nyquist curve of the linear
system in order to apply the stability criteria above?

EXERCISE 4.3
1.5 1.5 1.5

1 1 1

0.5 0.5 0.5

0 0 0

−0.5 −0.5 −0.5

−1 −1 −1

−1.5 −1.5 −1.5

−2 −1.5 −1 −0.5 0 0.5 1 1.5 2 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2

Figure 4.1 Static nonlinearities in Exercise 4.3.

Consider the static nonlinearities shown in Figure 4.1. For each nonlinear-
ity,
(a) determine the minimal sector [α , β ],
(b) determine the gain of the nonlinearity,
(c) determine if the nonlinearity is passive.

24
 Chapter 4. Input-Output Stability

EXERCISE 4.4[KHALIL, 1996]

The Nyquist curve of
4
G ( s) =
(s + 1)(s/2 + 1)(s/3 + 1)

is shown in Figure 4.2 together with a circle with center in 1.5 and with
radius 2.85.
Nyquist Diagrams

1
Imaginary Axis

−1

−2

−3
−2 −1 0 1 2 3 4
Real Axis

Figure 4.2 The Nyquist-curve in Exercise 4.4

(a) Determine the maximal stability sector of the form (−α , α ).

(b) Use the circle in the figure to determine another stability sector.
(c) What is the maximal stability sector of the form (0, β )?

EXERCISE 4.5[KHALIL, 1996]

The Nyquist curve of

4
G ( s) =
(s − 1)(s/3 + 1)(s/5 + 1)

is shown in Figure 4.5 . Determine a possible stability sector (α , β ).

EXERCISE 4.6[KHALIL, 1996]

Using the circle criterion, for each of the scalar transfer functions below,
find a sector (α , β ) for which the system is BIBO stable.
(a)
s
G ( s) =
s2 −s+1

25
 Chapter 4. Input-Output Stability

Nyquist Diagrams

1.5

0.5

Imaginary Axis
0

−0.5

−1

−1.5

−2
−4 −3.5 −3 −2.5 −2 −1.5 −1 −0.5 0
Real Axis

Figure 4.3 The Nyquist-curve in Exercise 4.5

Bode Diagram 0.8

−0.1
10
0.6
−0.3
10
Magnitude (abs)

−0.5 0.4
10

−0.7
10 0.2

−0.9
10
Im

270
−0.2

225
−0.4
Phase (deg)

180
−0.6

135
−0.8

90
0 1
10 10 −1
−1.5 −1 −0.5 0 0.5
Frequency (rad/sec) Re

0.4

0.2
Im

−0.2

−0.4

−0.4 −0.2 0 0.2 0.4 0.6

Re

Figure 4.4 The Bode and Nyquist curves for the system in Exercise 4.6ab

(b)

1
G ( s) =
(s + 1)(s + 2)

26
 Chapter 4. Input-Output Stability

EXERCISE 4.7
Consider the linear time-varying system

ẋ(t) = ( A + B f (t) C) x,

(a) Show that the system can be written as a feedback connection of a

linear time invariant system with transfer function

G (s) = C(sI − A)−1 B

and the time-varying component f (t).

(b) Let A be Hurwitz (i.e. asymptotically stable), let G (s) have one input
and one output, and let hh f (t)hh ≤ 1 ∀t. Show that if

sup h G (iω )h < 1

ω ∈IR

then the system is BIBO stable.

(c) Figure 4.5 shows the Nyquist curves for different transfer functions
G (s). Which transfer functions will give a BIBO stable closed loop
according to the criteria in (b)?

1.5 1.5 1.5

1 1 1

0.5 0.5 0.5

0 0 0

−0.5 −0.5 −0.5

−1 −1 −1

−1.5 −1.5 −1.5

−1.5 −1 −0.5 0 0.5 1 1.5 −1.5 −1 −0.5 0 0.5 1 1.5 −1.5 −1 −0.5 0 0.5 1 1.5

Figure 4.5 Nyquist curves for transfer function G (s) in Exercise 4.7.

(d) For PhD students. Let G (s) be a transfer function matrix with m
inputs and n outputs. Show that if A is Hurwitz, hh f (t)hh ≤ 1 ∀t, and
supω ∈IR σ max [ C( jω I − A)−1 B ] < 1, then the system is BIBO stable.

EXERCISE 4.8
The singular values of a matrix A are denoted σ i( A).
(a) Use Matlab to compute σ ( A) for
 
1 10
A= .
0 1

(b) The maximal singular value is defined by

h Axh
σ 1 ( A) = sup .
x h xh

Show that σ 1( AB ) ≤ σ 1( A)σ 1 ( B ).

27
 Chapter 4. Input-Output Stability

EXERCISE 4.9
In the previous chapter, we have seen how we can use Lyapunov functions
to prove stability of systems. In this exercise, we shall see how another
type of auxillary functions, called storage functions, can be used to assess
passivity of a system.
Consider the nonlinear system

ẋ = f ( x, u)
y = g( x, u) (4.1)

with zero initial conditions, x(0) = 0. Show that if we can find a storage
function V ( x, u) with the following properties
• V ( x, u) is continuously differentiable.
• V (0) = 0 and V ( x, u) ≥ 0 for x = 0.
• uT y ≥ V̇ ( x, u).
then, the system (4.1) is passive.

28
 Chapter 4. Input-Output Stability

EXERCISE 4.10
Let P be the solution to
AT P + PA = − I ,
where A is an asymptotically stable matrix. Show that G (s) = B T P(sI −
A)−1 B is passive. (Hint. Use the function V ( x) = x T Px.)

EXERCISE 4.11[BOYD, 1997]

Consider the dynamic controller

ẏ = −2 y + sat( y) + u, y(0) = y0.

(a) Show that the system is passive.

(b) Is the system strictly passive ? Justify your answer. (Hint. What can
you say about the system when it operates in the linear region, i.e.,
sat( y) = y?)
(c) A DC motor is characterized by

θ˙ = ω
ω̇ = −ω + η ,

where θ is the shaft angle and η is the input voltage. The dynamic
controller

ż = 2(θ − z) − sat(θ − z)
η = z − 2θ

is used to control the shaft position. Use any method you like to prove
that θ (t) and ω (t) converge to zero as t → ∞.

EXERCISE 4.12
(a) Let uc (t) be an arbitrary function of time and let H (⋅) be a passive
system. Show that

y(t) = uc (t) ⋅ H (uc (t)u(t))

is passive from u to y.
(b) Show that the following adaptive system is stable


e(t) = G (s) θ (t) − θ 0 uc (t)
θ˙(t) = −γ uc (t) e(t),

if γ > 0 and G (s) is strictly passive.

29
 Chapter 4. Input-Output Stability

EXERCISE 4.13PHD
Let f be a static nonlinearity in the sector (0, ∞).
(a) Show that the system ẋ = γ x + e, y = f ( x) is passive from e to y if
γ ≥ 0.
(b) Show that if the Popov criterion

Re (1 + γ iw) G (iω ) > 0, ∀ω ,

is satisfied, with γ ≥ 0, then the closed loop system in Figure 4.6 is

absolutely stable.

1
1+γ s f (⋅)

1 +γ s G ( s)

Figure 4.6 Proof of the Popov criterion in Exercise 4.13.

(c) How does the Popov criterion change if f is in the sector (α , β ) in-
stead?
(d) Figure 4.7 shows the Nyquist curve and the Popov curve (Re G (iω ), ω Im G (iω ))
for the system

s+1
G ( s) = .
s(s + 0.1)(s2 + 0.5s + 9)

Determine a stability sector (0, β ) using the line in the figure.

30
 Chapter 4. Input-Output Stability

0.5

−0.5

−1

−1.5

−2
−10 −9 −8 −7 −6 −5 −4 −3 −2 −1 0

Figure 4.7 Nyquist (dash-dot) and Popov curve (solid) for the system in Exer-
cise 4.13d. The Popov curve is to the right of the dashed line for all ω .

31
5. Describing Function
Analysis, Limit Cycles
EXERCISE 5.1
Match each of the odd, static nonlinearities in Figure 5.1 with one of the
describing functions in Figure 5.2. (Hint. Use the interpretation of the
describing function N ( A) as “equivalent gain” for sinusoidal inputs with
amplitude A.)

Nonlinearity a b
15 15

10 10

5 5

0 0
0 5 10 0 5 10

c d
15 15

10 10

5 5

0 0
0 5 10 0 5 10

Figure 5.1 Nonlinearities in Exercise 5.1.

Describing functions nr 1 nr 2
4 4

3 3

2 2

1 1

0 0
0 5 10 0 5 10

nr 3 nr 4
4 4

3 3

2 2

1 1

0 0
0 5 10 0 5 10

Figure 5.2 Describing functions in Exercise 5.1.

32
 Chapter 5. Describing Function Analysis, Limit Cycles

EXERCISE 5.2
Compute the describing functions for
(a) the saturation,
(b) the deadzone, and
(c) the piece-wise linear function
in Figure 5.3. (Hint: Use (a) in (b) and (c).)

H H β
−D −D −D α
D D 2D −α D
−H −H −β

Figure 5.3 The static nonlinearities in Exercise 5.2

EXERCISE 5.3
Show that the describing function for a relay with hysteresis in Figure 5.4
satisfies  
 2 !1/2
1 πA  D D
− =− 1− +i .
N ( A) 4H A A

H Im
Re
πA
4H
−D D −π D
4H
−H − N (1A)

Figure 5.4 The relay with hysteresis in Exercise 5.3.

EXERCISE 5.4
If the describing function for the static nonlinearity f ( x) is YN ( C), then
show that the describing function for D f ( x/ D ) equals YN ( C/ D ), where D
is a constant.

EXERCISE 5.5

33
 Chapter 5. Describing Function Analysis, Limit Cycles

r k y
Σ G ( s)
1

Figure 5.5 Electronic oscillator.

Show that all odd, static nonlinearities f such that

d f ( x) d2 f ( x)
> 0, > 0,
dx dx2

for x > 0, have a real describing function Ψ(⋅) that satisfies the inequalities

f (a)
Ψ(a) < , a > 0.
a

EXERCISE 5.6[ÅSTRÖM, 1968]

Compute the describing function for a static nonlinearity of the form

f ( x) = k1 x + k2 x2 + k3 x3 .

How does the term k2 x2 influence the analysis?

EXERCISE 5.7[SLOTINE AND LI, 1991]

Consider the system in Figure 5.5, which is typical of the dynamics of
electronic oscillators used in laboratories. Let

−5s
G ( s) =
s2 + s + 25

(a) Assess intuitively the possibility of a limit cycle, by assuming that the
system is started at some small initial state, and notice that the sys-
tem can neither stay small (because of instability) nor at saturation
values (by applying the final value theorem of linear control).
(b) Use the describing function method to predict whether the system
exhibits a limit cycle, depending on the saturation level H. In such
cases, determine the frequency and amplitude of the limit cycle. The
describing function of a saturaion is plotted in Figure 5.6.
(c) Use the extended Nyquist criterion to assess whether the limit cycle
is stable or unstable.

EXERCISE 5.8

34
 Chapter 5. Describing Function Analysis, Limit Cycles

0.8

0.6

N(A)
0.4

0.2

0
0 2 4 6 8 10
A

Figure 5.6 Normalized describing function.

1
+ −a G (s)
+ −1 a 0
−

Figure 5.7 Position servo in Exercise 5.8.

Consider a servo motor with transfer function

4
G0 (s) =
s(s + 1)(s + 2)

controlled by a relay with a dead-zone a as shown in Figure 5.7.

(a) Show that the describing function for the relay with dead-zone a is
given by

0
 A<a
N ( A) = r
 4
 a2
1− A≥a
πA A2

(b) How should the parameter a be choosen so that the describing func-
tion method predicts that sustained oscillations are avoided in the
closed loop system?

EXERCISE 5.9
The Ziegler-Nichols frequency response method suggest PID parameters
based on a system’s ultimate gain Ku and ultimate period Tu according to
the following table. The method provides a convenient method for tuning

Parameter Value
K 0.6Ku
Ti 0.5Tu
Td 0.125Tu

Table 5.1 Tuning rules for Ziegler-Nichol’s method.

35
 Chapter 5. Describing Function Analysis, Limit Cycles

r e u y
G ( s)
−

Figure 5.8 An auto-tuning experiment: linear system under relay feedback.

0.5

−0.5

−1
0 5 10 15

Figure 5.9 Input and output of system under relay feedback.

PID controllers, since Ku and Tu can be estimated through simple experi-

ments. Once Ku and Tu have been determined, the controller parameters
are directly given by the formulas above.

(a) Show that the parameters Ku and Tu can be determined from the
sustained oscillations that may occur in the process under relay feed-
back. Use the describing function method to give a formula for com-
puting Ku and Tu based on oscillation data. (amplitude A and angular
frequency ω of the oscillation). Let the relay amplitude be D.
Recall that the ultimate gain and ultimate period are defined in the
following way. Let G (s) be the systems transfer function, and ω u be
the frequency where the system transfer function has a phase lag of
−180 degrees. Then we have

Tu = 2π /ω u
Ku = 1/h G (iω u )h

(b) What parameters would the relay method give for the process

50
G ( s) =
s(s + 1)(s + 10)

which is simulated in Figure 5.9 with D = 1? Compare what you

obtain from analytical computations ( Ku = 2.20, Tu = 1.99)

36
 Chapter 5. Describing Function Analysis, Limit Cycles

EXERCISE 5.10[SLOTINE AND LI, 1991]

In many cases, it is desirable to limit the high frequency content in a
signal. Usually, such filtering is performed using a linear low-pass filter.
Unfortunately, this type of filtering introduces phase lags. If the limiter is
positioned within a feedback loop, this phase lag can be harmful for system
stability.

r k y
τ 1 s+1 τ 2 s+1
τ 2 s+1 τ 1 s+1
α
τ1 > τ2

Figure 5.10 The nonlinear lowpass filter suggested in Exercise 5.10.

Figure 5.10 shows an alternative way of limiting the high frequency con-
tent of a signal. The system is composed of a high pass filter, a saturation,
and a lowpass filter. Show that the system can be viewed as a nonlinear
lowpass filter that attenuates high-frequency inputs without introducing a
phase lag.

EXERCISE 5.11[KHALIL, 1996]

Consider the second order system

ẋ1 = x1 − x2 − x1 ( x12 + x22 )

ẋ2 = x1 + x2 − x2 ( x12 + x22 )

(a) Show that the unit circle is a periodic orbit.

(b) Use a Poincaré map to show that the periodic orbit derived in (a) is
asymptotically stable.
(b) Use the Lyapunov function candidate V = r2 − 1 (together with La
Salle’s Theorem) and show that the limit cycle derived in (a) is glob-
ally asymptotically stable.

EXERCISE 5.12PHD
Show that the system

1
G ( s) =
s(s + 1)2

with relay feedback has a locally stable limit cycle.

37
 Chapter 5. Describing Function Analysis, Limit Cycles

EXERCISE 5.13PHD
Consider a linear system with relay feedback:

ẋ = Ax + Bu,
y = Cx,
u = −sgn y,

where A is assumed to be non-singular. In this exercise we will study limit

cycles for this system. The Jacobian of the Poincaré map will be derived.
It gives a condition for local stability of the limit cycles.
(a) Draw a figure that illustrates a limit cycle for the system if the linear
dynamics is of third order.
(b) Consider an initial point x(0) = z that satisfies Cz = 0, that is, z
lies in the switch plane. Show that the Poincaré map from z to next
switch plane intersection is

g( z) = eAh( z) z − ( eAh( z) − I ) A−1 B , (5.1)

where h( z) is the time it takes from z to next switch plane intersec-

tion.
(c) A limit cycle corresponds to a fix point z∗ of −g, that is, a point such
that z∗ = −g( z∗ ). Let h( z) in (5.1) be equal to h and solve the equation
z∗ = −g( z∗ ). (The solution is a function of h.)
(d) Consider a perturbed initial point z + δ z and a perturbed switch time
h + δ h. Derive a first-order Taylor expansion of g and show that

g( z + δ z) = − z + eAhδ z − ( Az + B )δ h + O (δ 2), (5.2)

where O (δ 2 ) represents second-order and higher-order terms in δ z

and δ h.
(e) Let the perturbations δ z and δ h be such that both the perturbed
initial point and the Poincaré mapping of it lie in the switch plane.
This is equivalent to that Cδ z = 0 and Cg( z + δ z) = 0. Show by using
(5.2) and Cg( z + δ z) = 0 that asymptotically (when δ z → 0)

CeAh
δh= δ z.
C( Az + B )

(f) Use (d) and (e) to show that

( Az + B ) C Ah
g( z + δ z) = − z + ( I − ) e δ z.
C( Az + B )

We have now shown that the Jacobian of the Poincaré map for a
linear system with relay feedback is equal to the matrix

( Az + B ) C Ah
e .
C( Az + B )

The limit cycle is locally stable if and only if this matrix has all
eigenvalues in the unit disc.

38
 Chapter 5. Describing Function Analysis, Limit Cycles

(g) [Matlab exercise] Choose a state-space representation so that

1
C(sI − A)−1 B = .
(s + 1)3

In c you derived the solution z∗ = z∗ (h). Plot the scalar function

Cz∗ (h) as a function of h. The zero crossing at h > 0 indicate a
possible stable limit cycle. What is h∗ ?
Let z = z∗ (h∗ ) and h = h∗ . Derive the eigenvalues of

( Az + B ) C Ah
e .
C( Az + B )

39
6. Anti-windup, Friction,
Backlash, Quantization
EXERCISE 6.1
Consider the antiwindup scheme in polynomial form described in the figure
together with a process A(s) y = B (s)u. Put uc = 0. Make a block trans-
(a) (b)
uc
uc T
T

1 u y 1 v u
∑ −S ∑
R A aw
y
−S

A aw− R

formation to write the system in standard feedback form with lower block
AAaw − 1. Use the circle criterion (or passivity) to conclude that the
+BS
P = AR
system is globally asymptotically stable if A is stable and the following
condition holds:
 
AR + B S
Re (iω ) ≥ ε > 0, ∀ω .
AAaw

EXERCISE 6.2
The following model for friction is described in a recent PhD thesis

dz hvh
= v− z
dt g(v)
dz
F = σ 0 z + σ 1(v) + Fvv,
dt

where σ 0 , Fv are positive constants and g(v) and σ 1(v) are positive func-
tions of velocity.
(a) What friction force does the model give for constant velocity?
(b) Prove that if 0 < g(v) ≤ a and h z(0)h ≤ a then

h z(t)h ≤ a, t≥0

(Hint: Use the function V ( z) = z2 )

(c) Prove that the map from v to z is passive if z(0) = 0.
(d) Prove that the map from v to F is passive if z(0) = 0 and 0 ≤ σ 1(v) ≤
4σ 0g(v)/hvh.

40
 Chapter 6. Anti-windup, Friction, Backlash, Quantization

EXERCISE 6.3
Derive the describing function (v input, F output) for
(a) Coulomb friction, F = F0 sign (v)
(b) Coulomb + linear viscous friction F = F0 sign (v) + Fvv
(c) as in b) but with stiction for v = 0.

EXERCISE 6.4
If v is not directly measurable the adaptive friction compensation scheme
in the lectures must be changed. Consider the following double observer
scheme:

Fb = ( zF + K F hb
vh)sign(b
v)
żF = − K F (u − F b)sign(b
v)
v = zv + Kv x
b
żv = − F v.
b + u − Kv b

Show that linearisation of the error equations

eḃv = v̇ − vḃ = gv( ev, eF , v, F )

eḃF = F = g F ( ev , eF , v, F )
Ḟ − ċ

gives Jacobian " #

− Kv −1
A= .
− Kv K F 0
Conclude that local error convergence is achieved if Kv > 0 and K F < 0.

EXERCISE 6.5
Show that the describing function for quantization is given by

 0
D
 A< 2
s  2
N ( A) = 4D Pn 2i − 1
 2n−1 2n+1
 1− D 2 D < A< 2 D
π A i=1 2A

(Hint: Use one of the nonlinearities from Lecture 6 and superposition.)

EXERCISE 6.6
Show that a saturation is a passive element.

EXERCISE 6.7[KHALIL, 1996]

41
 Chapter 6. Anti-windup, Friction, Backlash, Quantization

Consider the mass-spring system with dry friction

ÿ + c ẏ + ky + η ( y, ẏ) = 0

where η is defined as


 µ mgsign( ẏ) for h ẏh > 0
 k
η ( y, ẏ) = − ky for ẏ = 0 and h yh ≤ µ s mg/ k



−µ s mgsign( y) for ẏ = 0 and h yh > µ s mg/ k

Construct the phase potrait and discuss its qualitative behavior. (Hint: Use
piecewise linear analysis.)

EXERCISE 6.8
The accuracy of a crude A/D converter can be improved by adding a high-
frequency dither signal before quantization and lowpass filtering the dis-
cretized signal, see Figure 6.1 Compute the stationary value y0 of the out-

+ A/ D filter decim.

Figure 6.1 The static nonlinearities in Exercise 6.8

put if the input is a constant u0 . The dither signal is a triangle wave with
zero mean and amplitude D /2 where D is the quantization level in the
A/D converter.

EXERCISE 6.9
For PhD students. Show that the antiwindup scheme in observer form is
equivalent to the antiwindup scheme in polynomial form with A e equal to
the observer polynomial (see CCS for definitions).

EXERCISE 6.10
For PhD students. Show that the equilibrium point of an unstable linear
system preceded with a saturation can not be made globally asymptotically
stable with any control law.

42
7. Nonlinear Controller
Design
EXERCISE 7.1
In some cases, the main nonlinearity of a system can be isolated to a static
nonlinearity on the input. This is, for example, the case when a linear pro-
cess is controlled using a actuator with a nonlinear characteristic. A simple
design methodology is then to design a controller C(s) for the linear process
and cancel the effect of the actuator nonlinearity by feeding the computed
control through the inverse of the actuator nonlinearity, see Figure 7.1.
Compute the inverse of the following common actuator characteristics
Controller

C ( s) f −1(⋅) f (⋅) G ( s)
−

Figure 7.1 Compensation of input nonlinearity by inversion.

(a) The quadratic (common in valves)

f (v) = v2, v≥0

(b) The piecewise linear characteristic

(
k1 v hvh ≤ d
f (v) =
( k1 − k2 )d + k2 v hvh > d

Use your result to derive the inverse of the important special case of
a dead zone.
(c) A backlash nonlinearity.

EXERCISE 7.2
An important class of nonlinear systems can be written on the form

ẋ1 = x2
ẋ2 = x3
..
.
ẋn = f ( x) + g( x)u

Assume that the full state x is available for measurement.

43
 Chapter 7. Nonlinear Controller Design

(a) Find a feedback

u = h( x, v)

that renders the closed loop system from the new input v to the state
linear. What conditions do you have to impose on f ( x) and g( x) in
order to make the procedure well posed?
(b) Apply this procedure to design a feedback for the inverted pendulum

ẋ1 = x2
ẋ2 = a sin( x1 ) + b cos( x2 )u

that makes the closed loop system behave as a linear system with a
double pole in s = −1. Is the control well defined for all x? Can you
explain this intuitively?
(c) One drawback with the above procedure is that it is very sensitive to
modelling errors. Show that this is the case by designing a linearizing
feedback for the system

ẋ = x2 + u

that makes the closed loop system linear with a pole in −1. Apply
the suggested control to the system

ẋ = (1 + ε ) x2 + u

and show that some solutions are unbounded irrespectively of ε = 0.

EXERCISE 7.3
Consider a linear system

ẋ1 = ax2 + bu
ẋ2 = x1

with nominal parameter values a = 1, b = 1. The system equations were

obtained by linearization of a nonlinear system, which has the consequence
that the parameters a and b vary with operating region.
(a) One of the design parameters in the design of a sliding mode con-
troller is the choice of sliding surface. Which of the following sliding
surfaces will result in a stable sliding mode for the above system?
(i) σ ( x) = 2x1 − x2
(ii) σ ( x) = x1 + 2x2
(iii) σ ( x) = x1
(b) Let the sliding mode be σ ( x) = x1 + x2 . Construct a sliding mode
controller for the system.

44
 Chapter 7. Nonlinear Controller Design

(c) How large variations in the parameters a and b can the controller
designed in (b) tolerate in order to still guarantee a stable closed
loop system?

EXERCISE 7.4
Consider concentration control for a fluid that flows through a pipe, with
no mixing, and through a tank, with perfect mixing. A schematic diagram
of the process is shown in Figure 7.2(left). The concentration at the inlet
of the pipe is cin. Let the pipe volume be Vd and let the tank volume be
Vm. Furthermore, let the flow be q and let the concentration in the tank
at the outlet be c. A mass balance gives

dc(t)
Vm = q(t)(cin(t − L) − c(t))
dt

where L = Vd/ q(t).

Vd k
0.63k
c in

Time
Vm a

c
L T

Figure 7.2 Schematic of the concentration control system (left). Parameters in

Ziegler-Nichols step response method (right).

(a) Show that for fixed q(t), the system from input cin to output c can be
reprented by a linear transfer function

K
G ( s) = e−sL
sT + 1

where L and T depend on q(t).

(b) Use Ziegler-Nichols time response method and your model knowl-
edge from (a) to determine a gain scheduled PI-controller from the
step response in Figure 7.3. Recall that the Ziegler-Nichols step re-
sponse method relies on the parameters L and a = K L/T defined in
Figure 7.2(right) . (The line is tangent to the point where the step
response has maximum slope).
Given these process parameters, the method suggest PID controller
gains according to Table 7.1.
(c) Compare, using simulations, the performance of your gain scheduled
controller with a controller designed for fixed q = q0 = 1.

45
 Chapter 7. Nonlinear Controller Design

Step response

0.5

Amplitude
0

−0.5

−1

0 2 4 6 8 10
Time [s]

Figure 7.3 Experimental step response for q(t) = 1.

Controller Kp Ti Td
P 1/ a
PI 0.9/a 3L
PID 1.2/a 2L L /2

Table 7.1 PID parameters suggested by Ziegler-Nichols step response method.

EXERCISE 7.5
We have seen how it in many cases can be of interest to control the system
into a set, rather than to the origin. One example of this is sliding mode
control, where the system state is forced into an invariant set, chosen
in such a way that if the state is forced onto this set, the closed loop
dynamics are exponentially stable. In this example, we will use similar
ideas to design a controller that “swings” up an inverted pendulum from
its stable equilibrium (hanging downwards) to its upright position.
Let the pendulum dynamics be given by

ẋ1 = x2
mg l ml
ẋ2 = sin( x1 ) − cos( x1 )u
Jp Jp

(a) Denote the total energy of the pendulum by E and determine the
value E0 corresponding to the pendulum standing in the upright po-
sition.
(b) Investigate whether the control strategy

u = k( E − E0 )sign( x2 cos( x1 ))

forces the value of E towards E0 . (Hint. Use a Lyapunov function

argument with V ( x) = ( E( x) − E0)2
(c) Draw a phase portrait of the system and discuss the qualitative be-
haviour of the closed loop system. In particular, will the suggested
control stabilize the unstable equilibrium point?

46
 Chapter 7. Nonlinear Controller Design

EXERCISE 7.6
Consider the system

ẋ1 = x1 + u
ẋ2 = x1
y = x2

Show that the control law

u = −2x1 − sign( x1 + x2 )

will make σ ( x) = x1 + x2 = 0 into a sliding mode (Hint: use V ( x) =

σ 2( x)/2). Determine the equivalent dynamics on the sliding plane σ ( x) =
0.

EXERCISE 7.7
R1
Minimize 0 x2 (t) + u2 (t) dt when

ẋ (t) = u(t)
x(0) = 1
x(1) = 0

EXERCISE 7.8
Neglecting air resistance and the curvature of the earth the launching of
a satellite is described with the following equations

ẋ1 = x3
ẋ2 = x4
F
ẋ3 = cos u
m
F
ẋ4 = α sin u − g
m

Here x1 is the horizontal and x2 the vertical coordinate and x3 and x4

are the corresponding velocities. The signal u is the controlled angle. The
criterion is to maximize e.g. height or velocity at the end point, under
constraints on the other variables. Show that the optimal control signal
has the form
At + B
tan u =
Ct + D
where A, B , C, D are constants (which you do not have to determine).

EXERCISE 7.9

47
 Chapter 7. Nonlinear Controller Design

Suppose more realistically that m and F vary. Let F = u2 (t) be a control

signal with limitations
0 ≤ u2 (t) ≤ umax
and let the mass m = x5 (t) vary as

ẋ5 = −γ u2

Show that 
λ1  umax σ < 0

tan u1 = and u2 = 0 σ >0
λ4 

L σ = 0,
where L means that the solution is unknown. Determine equations for λ
and σ . (You do not have to solve these equations).

EXERCISE 7.10
Consider the system
(
ẋ1 = x2
ẋ2 = − x1 − x23 + (1 + x1 )u

with initial conditions x1 (0) = 1, x2 (0) = 1 and let the criterion be

Z 1
2
min ex1 + x22 + u2 dt.
0

Is the problem normal ? Show that extremals satisfy

ẋ1 = f 1 ( x1 , x2 , λ 1 , λ 2 )
ẋ2 = f 2 ( x1 , x2 , λ 1 , λ 2 )
λ̇ 1 = f 3 ( x1 , x2 , λ 1 , λ 2 )
λ̇ 2 = f 4 ( x1 , x2 , λ 1 , λ 2 )

Determine f 1, f 2 , f 3 , f 4. What conditions must λ 1 , λ 2 satisfy at the end

point?

EXERCISE 7.11
Consider the double integrator

ẋ1 = x2
ẋ2 = u, huh ≤ 1

with initial value x(0) = x0 . We are interested in finding the control that
brings the system to rest ( x(t f ) = 0) in minum time. (You may think of this
as a way of designing a controller that reacts quickly on set-point changes)

48
 Chapter 7. Nonlinear Controller Design

(a) Show that the optimal control is of “bang-bang type” with at most one
switch. (In other words, show the optimal control is first maximal in
one direction for some time, and then possibly changes sign to become
maximal in the other direction for the remaining time)
(b) Show that the control can be expressed as the feedback law
(
umax σ ( x) > 0
u=
−umax σ ( x) < 0

Derive an explicit formula for the switch function σ ( x).

(c) Using a phase plane tool, construct the phase portrait of the closed
loop system under the optimal control.

EXERCISE 7.12
Consider the problem of controlling the double integrator
(
ẋ1 = x2
ẋ2 = u, huh ≤ 1

from an arbitrary intitial condition x(0) to the origin so that the criterion
Z tf
(1 + huh) dt
0

is minimized (t f is the first time so that x(t f ) = 0). Show that all extremals
are of the form 
 −1 0 ≤ t ≤ t1

u(t) = 0 t1 ≤ t ≤ t2


1 t2 ≤ t ≤ t f
or 
 1 0 ≤ t ≤ t1

u(t) = 0 t1 ≤ t ≤ t2


−1 t2 ≤ t ≤ t f
for some t1, t2 with 0 ≤ t1 ≤ t2 ≤ t f . Some time interval can have the
length 0. Assume that the problem is normal.

EXERCISE 7.13
Consider the system
   
 −5 2  0

ẋ =  x+
   u
−6 2 1

49
 Chapter 7. Nonlinear Controller Design

 
1
 
from x0 = 0 to x(t f ) =   in minimum time with hu(t)h ≤ 3. Show that
1
the optimal controller is either
(
−3 0 ≤ t ≤ t1
u(t) =
+3 t1 ≤ t ≤ t f

or (
+3 0 ≤ t ≤ t1
u(t) =
−3 t1 ≤ t ≤ t f
for some t1 .

EXERCISE 7.14
Show that minimum time control of a linear system

ẋ = Ax + Bu, y = Cx, huh ≤ 1, x ( t f ) = 0,

leads to extremals with

u(t) = sign σ (t)
Determine σ (t).

EXERCISE 7.15
What is the conclusion from the maximum principle for the problem
Z 1
min u dt,
0
ẋ1 = u
x1 (0) = 0
x1 (1) = 1

Explain.

50
Solutions to Chapter 1
SOLUTION 1.1
(a) Choose the angular position and velocity as state variables, i.e., let

x1 = θ
x2 = θ˙

We obtain

ẋ1 = x2
g k
ẋ2 = − sin( x1 ) − x2
l m

(b) By setting the state derivatives to zero, we obtain

0 = x2
g k
0 = − sin( x1 ) − x2
l m

We find the equilibrium points ( x1 , x2 ) = (nπ , 0) with n = 0, ±1, ±2, . . . .

The equilibrium points correspond to the pendulum hanging down (n
even), or the pendulum balancing in the upright position (n odd).
Linearization gives
 
d 0 1
∆x = g ∆x (7.1)
dt − l (−1)n k
−m

The linearized system is stable for even n, and unstable for odd n.
We can use Lyapunov’s linearization method to conclude that the
pendulum is LAS around the lower equilibrium point, and unstable
around the upper equilibrium point.

SOLUTION 1.2
We choose angular positions and velocities as state variables. Letting x1 =
q1 , x2 = q̇1 , x3 = q2 , x4 = q̇2, we obtain

ẋ1 = x2
MgL k
ẋ2 = − sin x1 − ( x1 − x3 )
I I
ẋ3 = x4
k 1
ẋ4 = ( x1 − x3 ) + u
J J

51
 Solutions to Chapter 1

SOLUTION 1.3
(a) Let x1 = δ , x2 = δ˙, x3 = Eq and u = EF D . We obtain

ẋ1 = x2
P D η1
ẋ2 = − x2 − x3 sin x1
M M M
η2 η3 1
ẋ3 = − x3 + cos x1 + u
τ τ τ

(b) With Eq being constant, the model reduces to

ẋ1 = x2
P D η1
ẋ2 = − x2 − Eq sin x1
M M M

which is the pendulum equation with input torque P/ M.

(c) The equilibrium points of the simplified equation are given by sin x1 =
η1 Eq , x2 = 0.
P

SOLUTION 1.4
(a) Let

ẋ = Ax + Bu, y = Cx

be a state-space realization of the linear system. We have

u = r − ψ (t, y) = r − ψ (t, Cx)

and hence

ẋ = Ax − Bψ (t, Cx) + Br, y = Cx

(b) To separate the linear dynamics from the nonlinearities, write the
pendulum state equations as

ẋ1 = x2
k g
ẋ2 = − x2 − sin( x1 )
m l

and view sin( x1 ) as an input, generated by a nonlinear feedback from

y = x1 (Compare with Figure 1.3). Introduce the state vector x =
( x1 , x2 )T , and re-write the equations as
   
0 1 0
ẋ = x+ u (7.2)
0
− k/m g/l
y = [1 0] x (7.3)
u = − sin( y), (7.4)

which is on the requested form.

52
 Solutions to Chapter 1

SOLUTION 1.5
(a) Hint: ė = − y = − Cx.
(b) The equilibrium points are given by

0 = G (0) sin e

from which we obtain

e = ±nπ , n = 0, 1, 2, . . .

(c) For G (s) = 1/(τ s + 1), we take A = −1/τ , B = 1/τ and C = 1. Then

1 1
ż = − z + sin e
τ τ
ė = − z

Now, let x1 = e and x2 = − z, so that

ẋ1 = x2
1 1
ẋ2 = − x2 − sin x1 ,
τ τ
which is the pendulum model with g/l = k/m = 1/τ .

SOLUTION 1.6
Let G P I D (s) be the transfer function for the PID controller. The requested
form

V ( s) = − G l ( s) F ( s)
F (v) = F0 sign(v)

is obtained with
s
G l ( s) =
ms2 − G P I D (s)

SOLUTION 1.7
The requested form

U ( s) = − G l ( s) V ( s)
v(t) = sat(u)

is obtained with
G f b G p − Gaw
G l ( s) = .
1 + Gaw

53
 Solutions to Chapter 1

SOLUTION 1.8

(a) Introduce x1 = y, x2 = ẏ, x3 = “valve input”, then

ẋ1 = x2
ẋ2 = −2x2 − x1 + f ( x3 )
ẋ3 = r − x1
√
(b) For a constant input r the equilibrium point is given by x = (r, 0, r).
The linearization has
 
0 1 0
√
A =  −1 −2 2 r  .
−1 0 0

The characteristic equation is given by

√
λ 2 (λ + 2) + 2 r + λ = 0.

The condition for stability of λ 3 + aλ 2 + bλ + c is a, b, c > 0 and ab > c.

Hence we get local stability if 0 < r < 1. An alternative approach is to
compute that the gain margin of s(s+1 1)2 is 2. Since the linearization of the
√
valve has gain f P ( x3 ) = 2x3 = 2 r we get stability for r < 1.
(c) More realistic is that the flow is zero.

SOLUTION 1.9
The linearization is given by

ẍ = − k1 x + u,

which is controllable. Hence the system is nonlinear locally controllable.

SOLUTION 1.10
The linearized system is not controllable. The system is however nonlinear
locally controllable. This can be seen directly from the definition as follows:
We must show that we can drive the system from (0, 0, 0) to a near by state
( xT , yT , θ T ) using small control signals u1 and u2 . By the sequence u =
(u1 , u2 ) = (0, ε 1 ),u = (ε 1 , 0),u = (0, −ε 1 ), u = (−ε 2 , 0) (or in words: "turn
left, forward, turn right, backwards") one can move to the state (0, yT , 0).
Then apply (ε 3 , 0) and then (0, ε 4 ) to end up in ( xT , yT , θ T ). For any time
T > 0 this movement can be done with small ε i if xT , yT and θ T are small.

SOLUTION 1.11
Same solution as in 1.10, except that you have to find a movement af-
terwards that changes Ψ without changing the other states. This can be
done by the sequence: L-F-R-B-R-F-L-B where F=forward, B=backwards,
L=turn left, R=turn right.

54
 Solutions to Chapter 1

SOLUTION 1.12
The linearized system at ( x0 , u0 ) is

ẋ1 = u
ẋ2 = x1

The controllability matrix

 
1 0
Wc =
0 1

has full rank. Since the linearized system is controllable the nonlinear
system is also locally controllable at ( x0 , u0 ).

SOLUTION 1.13
See lecture slides. Use max(y(:)) to compute the amplitude.

SOLUTION 1.14
See lecture slides.

SOLUTION 1.15
See lecture slides.

SOLUTION 1.16
With a = 0.02 and w = 100π we get local stability for l ∈ [0.044, 1.9].

SOLUTION 1.17
p p
(a) x =
p 0, and if rp> 1 also x = ( b ( r − 1 ), b(r − 1), r − 1) and x =
(− b(r − 1), − b(r − 1), r − 1).
(b) The linearization around x = 0 is
 
−σ σ 0
ẋ =  r −1 0 
0 0 −b

with characteristic polynomial (s + b)(s2 +(σ + 1)s + σ (1 − r). A second

order monic polynomial has all roots in the left half plane iff it has
positive coefficients, hence x = 0 is LAS when 0 < r < 1, (σ , b, r > 0
by default).

55
Solutions to Chapter 2
SOLUTION 2.1
(a) The equilibrium points are
√ √
( x1 , x2 ) = (0, 0), ( 6, 0), (− 6, 0),

which are stable focus, saddle, and saddle, respectively.

(b) The equilibrium points are

( x1 , x2 ) = (0, 0), (−2.5505, −2.5505), (−7.4495, −7.4495),

which are stable node, saddle point, and stable focus, respectively.
(c) The equilibrium points are

( x1 , x2 ) = (0, 0), (1, 0), (0, 1),

which are unstable node, saddle, and stable node, respectively.

(d) The equilibrium is

( x1 , x2 ) = (0, 0),

which is an unstable focus.

(e) The equilibrium point is

( x1 , x2 ) = (0, 0),

which is a stable focus.

(f) The system has an equilibrium set

x12 + x22 = 1

and an equilibrium point

( x1 , x2 ) = (0, 0),

which is a stable focus.

SOLUTION 2.2
The three equilibrium points are
p p
( x1 , x2 ) = (0, 0), ( (ac/b), a), (− (ac/b)).

The first equilibrium point is a saddle. The other equilibria are stable nodes
if 8a < c and stable focuses if 8a > c.

SOLUTION 2.3

56
 Solutions to Chapter 2

(a) The system has three equilibrium points

( x1 , x2 ) = (0, 0), (a, 0), (−a, 0)

where a is the first positive root of

a
a − tan( ) = 0
2

given, approximately, by a = 2.33. The origin is a stable node, while

the other two equilibria are saddles.
(b) The system has the origin as a unique equilibrium point, being an
unstable focus.
(c) The system has the equilibrium points

( x1 , x2 ) = (0, 0), (1, 2), (−1, 2),

which are saddle, stable focus, and stable focus, respectively.

SOLUTION 2.4
Close to the origin, the saturation element opens in the linear region, and
all system are assigned the same closed loop dynamics. Far away from the
origin, the influence of the saturated control can be neglected, and the open
loop dynamics governs the behaviour.
(a) System (a) has one stable and one unstable eigenvalue. For initial
values close to the stable eigenvector, the state will move towards
the origin. For initial values close to the unstable eigenvector, the
system diverges towards infinity. This corresponds to the rightmost
phase portrait.
(b) All eigenvalues of system (b) are unstable. Thus, for initial values
sufficiently far from the origin, the system state will diverge. This
corresponds to the leftmost phase portrait. Note how the region of
attraction (the set of initial states, for which the state converges to
the origin) is severely limited.
(c) System (c) is stable also in open loop. This corresponds to the phase
portrait in the middle.

SOLUTION 2.5
(a) The system has the origin as a unique equilibrium point, being a
stable focus. The direction of the arrow heads can be determined by
inspection of the vector fields. In particular, notice that since f 1 ( x) =
− x2 the function f 1 is negative in the upper half of the plane, and
negative in the lower half.
(b) The system has three equilibrium points

( x1 , x2 ) = (0, 0), (a, 0), (−a, 0),

57
 Solutions to Chapter 2

where a is the first root of the equation

a
a − tan( ) = 0
3

which is approximately given by a = 3.97. The origin is a stable focus,

while the other equilibria are saddles. Similarly as in (a), the fact that
f 1 = x2 makes it particularly simple to determine the directions of
the arrow heads.

SOLUTION 2.6
(a) The equilibrium points are obtained by setting ẋ = 0. For K = −2,
the origin is the unique equilibrium point. When K = −2, the line
x1 = 2x2 is an equilibrium set.
(b) The Jacobian is given by
 
Vf −1 −K
(0) =
Vx 1 −2

with eigenvalues
r
3 1
λ =− ± − K.
2 4

Thus, the closed loop system is asymptotically stable about the origin
for K > −2. Depending on the value of K, we can origin has the
following character

1
<K stable focus
4
1
−2 < K < stable node
4
K < −2 saddle.

SOLUTION 2.7
The equilibria are given by sin x10 = η Eq ,
P
x20 = 0. The characteristic equa-
tion for the linearization becomes

λ 2 + α λ + β = 0,
ηE
where α = M D
> 0 and β = Mq cos x10 . Depending on α , β the equilibria are
stable focus, stable nodes or saddle points.

58
 Solutions to Chapter 2

SOLUTION 2.8
Introduce ∆ x as the deviation from the nominal trajectory:
 
sin t + ∆ x1 (t)
x ( t) =
cos t + ∆ x2 (t)
This gives for the state derivative
 
dx(t) cos t + ∆ ˙x1 (t)
=
dt − sin t + ∆ ˙x2 (t)
 
cos t + ∆ x2 (t) + (sin t + ∆ x1 (t))(−2 sin t∆ x1 (t) − 2 cos t∆ x2 (t) − ∆ x1 (t)2 − ∆ x2 (t)2 )
= 2 2
− sin t − ∆ x1 (t) + (cos t + ∆ x2 (t))(−2 sin t∆ x1 (t) − 2 cos t∆ x2 (t) − ∆ x1 (t) − ∆ x2 (t) )

Subtracting the nominal trajectory gives:

 
d ∆ x ( t) ∆ x2 (t) − 2 sin2 (t)∆ x1 (t) − 2 sin t cos t ∆ x2 (t) + O (∆ x2 )
=
dt −∆ x1 (t) − 2 sin t cos t ∆ x1 (t) − 2 cos2 (t)∆ x2 (t) + O (∆ x2 )
Neglecting second order effects gives
 2 
d ∆ x ( t)  −2 sin (t) 1 − sin(2t) 
= A(t)∆ x(t) = 
  ∆ x(t).

dt −1 − sin(2t) −2 cos2 (t)
To determine stability of the limit cycle, the following transformation
of variables is useful.

 2 2 1/2
 r(t) = ( x1 (t) + x2 (t) )
x2 (t)
 θ (t) = arctan

x1 (t)

This gives

dr(t)

 = r(1 − r2 )
dt
 dθ (t) = −1

dt
which implies that the limit cycle is stable, but not asymptotically
stable.

SOLUTION 2.9

7¨ 2r
x̃ = g cos(φ 0 )φ˜ + φ¨˜
5 5

SOLUTION 2.10
Using the identity
3 1
(sin t)3 = sin t − sin 3t
4 4
we see that u0 (t) = sin (3t), y0(t) = sin t is a nominal solution. The
linearization is given by
1
ỹ¨ + 4 sin2 t ⋅ ỹ = − ũ.
3

59
 Solutions to Chapter 2

SOLUTION 2.11
No solution yet.

60
Solutions to Chapter 3
SOLUTION 3.1
(a) Linearization about the system around the origin yields

Vf
A= = 3ax2
Vx
Thus, at the origin we have A = 0. Since the linearization has one
eigenvalue on the imaginary axis, linearization fails to determine
stability of the origin.
(b) V (0) = 0, V ( x) = 0 for x = 0, and V ( x) → ∞ as x → ∞.
Thus, V ( x) satisfies the conditions for being a Lyapunov function
candidate. Its time derivative is
VV
V̇ ( x) = f ( x) = 4ax6 (7.5)
Vx
which is negative definite for a < 0, and positive definite for
a > 0. The desired results now follow from Lyapunov’s stability
and instability theorems.
(c) For a = 0, the system is linear and given by

ẋ = 0

The system has solutions x(t) = x0 for all t. Thus, the system
is stable. A similar conclusion can be drawn from the Lyapunov
function used in (b).

SOLUTION 3.2
(a) We use the pendulum’s total energy

l2 2
V ( x) = g l (1 − cos( x1 )) + x
2 2
as Lyapunov function. We see that V is positive, and compute the
time derivative

dV ( x) X V V
= ẋi = g l sin( x1 ) x2 + x2 (−g l sin( x1 )) = 0
dt V xi
i

From Lyapunov’s theorem, we conclude that the origin is a stable

equilibrium point. Since V̇ ( x) = 0, we can also conclude that the ori-
gin is not asymptotically stable; trajectories starting at a level surface
V ( x) = c, remain on this surface for all future times.
(b) For k = 0, using V ( x) as above, similar calculations give

dV ( x) kl 2 2
=− x
dt m 2

61
 Solutions to Chapter 3

V̇ ( x) is negative semidefinite. It is not negative definite because

V̇ ( x) = 0 for x2 = 0. In other words V̇ ( x) = 0 along the x1 axis.
Now, consider a modified Lyapunov function

g 1
V ( x) = (1 − cos( x1 )) + x T Px
l 2

for some positive definite matrix P. In order for P to be positive def-

inite, the elements of P must satisfy

p11 > 0; p22 > 0; p11 p22 − p212 > 0

The derivative V̇ ( x) is given by

g g
V̇ ( x) = (1 − p22 ) x2 sin x1 − p12 x1 sin x1 +
l l
k k
+ ( p11 − p12 ) x1 x2 + ( p12 − p22 ) x22
m m

We now want to choose p11 , p12 and p22 such that V̇ ( x) is negative
definite. Since the cross terms x2 sin x1 and x1 x2 are sign indefinite,
we cancel them out by letting p22 = 1, p11 = p12 k/m. With these
choices, p12 must satisfy 0 < p12 < k/m. We let p12 = 0.5k/m, and
obtain

1g k 1 k 2
V̇ ( x) = − x1 sin x1 − x .
2lm 2m 2

The term x1 sin x1 > 0 for all 0 < h x1 h < π . Within this strip, V ( x) is
positive definite, and V̇ ( x) is negative definite. We conclude that the
origin is asymptotically stable.

SOLUTION 3.3
With V = kx2 /2 + ẋ2 /2 we get V̇ = −d ẋ4 ≤ 0. Since V̇ = 0 only when
ẋ = 0 and the system equation then gives ẍ = − kx = 0 unless also x = 0,
we conclude that x = ẋ = 0 is the only invariant set. The origin is globally
asymptotically stable since the Lyapunov function is radially unbounded.

SOLUTION 3.4
√
(a) The eigenvalues of A are λ = −1/2 ± i 3/2.
(b) (i) We have

V ( x) = p11 x12 + 2p12 x1 x2 + p22 x22 = (if p11 = 0)

p12 p2
= p11 ( x1 + x2 )2 + ( p22 − 12 ) x22
p11 p11

If p11 > 0 and p11 p22 − p212 > 0, both terms are non-negative.
Moreover, V ( x) → ∞ as x → ∞, and V ( x) = 0 ; x1 = x2 = 0
(This proves the "if"-part). If the conditions on pi j do not hold, it
is easy to find x such that V ( x) < 0 (proving the "only if"-part).

62
 Solutions to Chapter 3

(ii) We want to solve

       
0 1 p11 p12 p11 p12 0 −1 −1 0
+ =
−1 −1 p12 p22 p12 p22 1 −1 0 −1

Reading off the elements, we see that



2p12 = −1

p22 − p11 − p12 =0



−2p12 − 2p22 = −1

which has the solution p11 = 1.5, p12 = −0.5 and p22 = 1. P is a
positive definite matrix.
(c) Use the Matlab command lyap(A’,eye(2)).

SOLUTION 3.6
(a) The mistake is that V is not radially unbounded. The student has
forgotten to check that lim x→∞ V ( x) = ∞. In fact,

x12 1
V ( x1, x2 ) = 2
+ x22
1 + x1 2

so that lim x→∞ = 1. Consequently, V is not radially unbounded.

(b) No, the problem is not that the student was not clever enough to find
a Lyapunov function. There is no Lyapunov function, since the system
is not globally stable. Let’s show this now. In part (a), you may have
noticed that V̇ = 0 for all x. In other words, V is an “integral” of the
motion; the trajectories lie on the curves of constant value of V , i.e.,
we have

1 2 x12
V ( x) = x2 + = V ( x0 ) = c
2 1 + x12

If c > 1 then x(t) cannot change sign, since

x12
x22 = c − ≥ c−1
1 + x12
√
In this case, we have h x2 h ≥ c − 1. Since ẋ1 = x2 , it follows that
h x1 h → ∞ as t → ∞. Roughly speaking, if the system starts with
more initial stored energy than can possibly be stored as potential
energy in the spring, the trajectories will diverge.

SOLUTION 3.7
Find the equilibrium points for the system.

ẋ1 = 0 = 4x12 x2 − f 1 ( x1 )( x12 + 2x22 − 4)

ẋ2 = 0 = −2x13 − f 2( x2 )( x12 + 2x22 − 4),

63
 Solutions to Chapter 3

For general functions f i, the only possibility is that

4x12 x2 = 0, f i( xi ) ⋅ ( x12 + 2x22 − 4) = 0 and − 2x13 = 0

√
at the same time. We thus get that ( x1 , x2 ) = (0, 0) and ( x1 , x2 ) = (0, ± 2)
are the equilibrium points.
The set x12 + 2x22 − 4 = 0 is invariant since on trajectories

d 2
( x + 2x22 − 4) = 2x13 4x2 + 4x2 (−2x13 ) = 0.
dt 1
The motion on this invariant set is given by

ẋ1 = 4x2 ⋅ x12

ẋ2 = −2x1 ⋅ x12

The set √ is however NOT a limit cycle, as the two equilibrium points ( x1 , x2 ) =
(0, ± 2) belongs to this set. Any trajectory moving along the invariant
set (the state vector
√ moves clockwise√ ) will eventually end up in either
( x1 , x2 ) = (0, + 2) or ( x1 , x2 ) = (0, − 2) .
Is the invariant set asymptotically stable?
Use the squared distance as Lyapunov candidate

V ( x) = ( x12 + 2x22 − 4)2 .

Since V̇ = . . . = 4( x12 + 2x22 − 4)2(− x1 f ( x1 ) − 2x2 f ( x2 )) we conclude that

V̇ ≤ 0. We have V̇ < 0 unless we are on the invariant set or if xi f ( xi ) = 0.
If we assume xi f i ( xi ) > 0 for xi = 0 we see that the invariant set is the set
x12 + 2x22 − 4 = 0 and the origin x1 = x2 = 0.
The
√ state converges to either√ of the two equilibrium points ( x1 , x2 ) =
(0, + 2) or ( x1 , x2 ) = (0, − 2) from any initial condition except from the
origin.

SOLUTION 3.8
Verify that V (0) = 0, V ( x) > 0 for x = 0 and V ( x) → ∞ for hh xhh → ∞.
Now,
(a) We have

d
V ( x1 , x2 ) = 8x1 ẋ1 + 4x2 ẋ2 + 16x13 ẋ1 =
dt
= 8x1 x2 + 4x2 (−2x1 − 2x2 − 4x13 ) + 16x13 x2 =
= −8x22

Since V̇ ( x) ≤ 0, we conclude global stability.

(b) The Lyapunov function has V̇ ( x) = 0 for x2 = 0. For x2 = 0, we obtain

ẋ2 = −2x1 (2 + x12 ).

which implies that if x2 should remain zero, then x1 has also to be

zero. The invariance theorem from the lectures can now be used to
conclude global asymptotic stability of the origin.

64
 Solutions to Chapter 3

SOLUTION 3.9
(a) Introduce the state vector x = ( x1 , x2 )T = ( y, ẏ)T . The system dynam-
ics can now be written as

ẋ1 = x2
ẋ2 = −sat(2x1 + 3x2 )

Consider the Lyapunov function

Z 2x1 +3x2
1 1
V ( x1, x2 ) = x22 + sat( z)dz
2 2 0

It is straight forward to show that V is positive definite. (Can x go

to infinity without V going to infinity?). Now,

d 1
V ( x1, x2 ) = x2 ẋ2 + sat(2x1 + 3x2 )(2x1 + 3x2 )
dt 2
3
= − (sat(2x1 + 3x2 ))2
2
≤0

By Lyapunov theory, the system is globally stable. Further,

V̇ ( x) = 0 ; ẋ2 = sat(2x1 + 3x2 ) = 0

; x1 (t) = x2 (t) = 0, ∀t

which implies global asymptotic stability.

(b) No. These conditions would imply global exponential stability. This
can not be achieved by the system, since in the saturated regions we
have

ÿ = ±1.

(c) Try a slight modification of the procedure suggested in (a).

(d) No. So don’t work too long on this one.

SOLUTION 3.10
The derivative of the suggested Lyapunov function is

V̇ ( x) = −2x2 max{0, x1 } max{0, x2 } ≥ 0

with equality for x1 ≤ 0, x2 ≤ 0 or both. Thus, the Lyapunov function

derivative is negative in the first quadrant and zero elsewhere in the plane.
When the Lyapunov function derivative is zero, we have

ẋ1 = x2
ẋ2 = − x1

65
 Solutions to Chapter 3

This system has solutions

x1 (t) = A cos(t) + B sin(t)

x2 (t) = B cos(t) − A sin(t)

The trace of x1 , x2 is a circle, then if A and B are both nonzero, x1 (t) > 0 and
x2 (t) > 0 for some t. This implies that the only solution of V̇ ( x) = 0 is x(t) =
0 for all t. By LaSalle’s theorem, this system is globally asymptotically
stable.

SOLUTION 3.11
(a)  
1 0
P = 0.5
0 1
solves the Lyapunov equation with Q as the identity matrix.
(b) We have

V̇ ( x) = x T ( AT P + PA) x + 2x T Pg( x2 ) =
= − x12 − x22 + x2 g( x2 ) < 0

since the contribution from the x2 g( x2 )-term is non-positive under

the stated conditions.
(c) We have

V̇ ( x) = − x12 − x22 + x24

which is negative for x22 ≤ 1. Since the level sets 0.5( x12 + x22 ) = γ
are circles, we conclude that the unit circle is a guaranteed region of
attraction. A couple of simulations are shown in Figure 7.4. Notice, in
particular, that the entire strip h x2 h ≤ 1 is not a region of attraction.

Phase plane
2

1.5

0.5
x2

−0.5

−1

−1.5

−2
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2
x1

Figure 7.4 Trajectories of the nonlinear system and level surfaces of V ( x) =

0.5xT x. The region of attraction is the unit circle.

66
 Solutions to Chapter 3

SOLUTION 3.12
(a) The origin is locally asymptotically stable, since the linearization
 
d 0 −1
x̃ = x̃
dt 1 −1

is asymptotically stable. A Lyapunov function for the system can be

found by solving the Lyapunov equation

AT P + PA = − I ,

which has the unique solution

 
1.5 −0.5
P= (7.6)
−0.5 1

(You can solve Lyapunov equations in Matlab using the command

lyap(A.’,-eye(2));. Note the transpose on the A matrix, due to
Matlab’s definition of the command lyap.
(b) Since the Lyapunov function in (a) is positive for all x, we just have
to find the largest domain in which the derivative of the Lyapunov
function is negative. Introduce the polar coordinates (r, θ ) by

x1 = r cos θ
x2 = r sin θ

We get

V̇ (r, θ ) = −r2 + r4 cos2 θ sin θ (2 sin θ − cos θ ) ≤ −r2 + 0.861r4

which is negative for r2 < 1/0.861. Using this, together with λ min( P) ≥
0.69, we choose

0.69
c = 0.8 < = 0.801
0.861

The set { xh x T Px ≤ c} is an estimate of the region of attraction.

SOLUTION 3.13
(a) For h2x1 + x2 h ≤ 1, we have
 
0 1
ẋ = x. (7.7)
−1 −1

The system matrix is as. stable. Hence, the origin is locally asymp-
totically stable.
(b) We have V ( x) > 0 in the first and third quadrant.

V̇ ( x) = ẋ1 x2 + x1 ẋ2 = x12 − x1 sat(2x1 + x2 ) + x22 .

67
 Solutions to Chapter 3

Now, let us evaluate V̇ ( x) on the strip x1 x2 = c and suppose that c > 0

is chosen so large that the saturation is active, i.e., h2x1 + x2 h > 1.
By inspection of the dynamics, we see that for sufficiently large c, no
trajectories can escape from the first quadrant. We can therefore use
the following Lyapunov argument. Consider the Lyapunov function
derivative

c2
V̇ ( x) = x12 − x1 + .
x12

If c ≥ 1, V̇ ( x) is positive for all x1 ≥ 0. Hence, all trajectories starting

in the first quadrant to the right of the curve x1 x2 = c cannot cross
the curve. Consequently, they cannot reach the origin.
(c) It follows from (b) that the origin is not globally asymptotically stable.

68
 Solutions to Chapter 3

SOLUTION 3.14
(a) Use V as a Lyapunov function candidate and let u be generated by
the nonlinear state feedback

VV
u = −( ψ ( x))
Vx

(b) Intentionally left blank.

SOLUTION 3.5
Use convexity wrt K.

SOLUTION 3.16
(a) Integration of the equality d
dσ f (σ x) = V x (σ x) ⋅
Vf
x gives the equation
Z 1
Vf
f ( x) = (σ x) ⋅ x dσ .
0 Vx

We get
Z 1 Z 1  T
Vf T Vf
T T
x P f ( x) + f ( x) Px = x PT
(σ x) xdσ + x (σ x) dσ Px
0 Vx 0 Vx
Z 1(  T )
Vf Vf
=x T
P (σ x) + (σ x) P dσ x ≤ − x T x
0 V x V x

(b) Since P is positive definite, V ( x) is clearly positive semidefinite. To

show that it is positive definite, we need to show that f ( x) = 0 only
when x = 0. But the inequality proved in (a) shows that if f ( p) = 0
then
0 ≤ − pT p.

(c) Suppose that f is bounded, i.e. that i f ( x)i ≤ c for all x. Then

i x T P f + f T Pxi ≤ 2ci Pii xi.

But this contradicts the inequality in (a) as i xi → ∞.

(d) We have shown that V is positive definite and radially unbounded.
Moreover
 T "  T #
T Vf T Vf Vf Vf
V̇ = ẋ Pf +f P ẋ = f T
P ( x) + ( x) P f ≤ −i f ( x)i2 .
Vx Vx Vx Vx

Hence V̇ ( x) < 0 for all x = 0. Thus, the origin is globally asymptoti-

cally stable.

69
 Solutions to Chapter 3

SOLUTION 3.17
Assume the linearization A = VV xf of f is asymptotically stable. Then the
equation
PA + AT P = − I ,
R∞ T
has a solution P > 0. (To prove that P = 0 eA s eAs ds > 0 is such a
solution integrate both sides of

d AT s As T T
e e = AT eA s eAs + eA s eAs A
ds

from 0 to ∞.) All conditions of Krasovskii’s method are then satisfied and
we conclude that the nonlinear system is asymptotically stable. The unsta-
bility result is harder.

SOLUTION 3.18
The system is given by

ẋ1 = x2 =: f 1
ẋ2 = − x2 + K g( e) = − x2 + K g(− x1 ) =: f 2.

Following the hint we put V = f T ( x) P f ( x). Direct calculations give

 
 −6p12 K x12 p11 − p12 − 3K p22 x12 
V̇ = f 
T
  f.

p11 − p12 − 3K p22 x12 2( p12 − p22 ).

With  
 1 1

P= 

1 2
we get V̇ ≤ 0 if 3K x12 < 1. Hence the system is locally stable. Actually
one gets V̇ < 0 if 3K x12 < 1 unless x1 = 0. The invariant set is x1 = x2 =
0. From LaSalle’s theorem the origin is hence also locally asymptotically
stable.

70
Solutions to Chapter 4
SOLUTION 4.1
See the Figure 7.5.
1.5 1.5

1 1

0.5 0.5

0 0

−0.5 −0.5

−1 −1

−1.5 −1.5
−1 0 1 −1 0 1

1.5 1.5

1 1

0.5 0.5

0 0

−0.5 −0.5

−1 −1

−1.5 −1.5
−1 0 1 −1 0 1

Figure 7.5 The balls B (0, 1) in Exercise 4.1

SOLUTION 4.2
Note that the Small Gain Theorem and the Passivity Theorem can also
cope with dynamic nonlinearities. For simplicity, we will only consider the
case when the linear system is asymptotically stable.

(a) What are the restrictions that we must impose on the nonlinearities
so that we can apply the various stability theorems?
The Nyquist Criterion ψ ( y) must be a linear function of y, i.e.,
ψ ( y) = k1 y for some constant k1 .
The Circle Criterion ψ ( y) must be contained in some sector [ k1 , k2 ].
Small Gain Theorem ψ ( y) should be contained in a symmetric sec-
tor [− k2 , k2 ]. The gain of the nonlinearity is then k2 .
The Passivity Theorem ψ ( y) is strictly passive for all nonlineari-
ties in the sector (ε , 1/ε ) for some small ε .
These conditions are illustrated in Figure 7.6.
(b) If the above restrictions hold, we get the following conditions on the
Nyquist curve
The Nyquist Criterion The Nyquist curve should not encircle the
point −1/ k1
The Circle Criterion The Nyquist curve should not intersect the
disc D ( k1 , k2 ).
Small Gain Theorem The Nyquist curve has to be contained in a
disc centered at the origin, with radius 1/ k2 .

71
 Solutions to Chapter 4

The Passivity Theorem The Nyquist curve has to stay in the right
half-plane, Re( G (iω )) ≥ 0.
These conditions are illustrated in Figure 7.7.

Nyquist criterion Circle Criterion

1 1
k2

0.5 k1 0.5

k1
ψ(y)

ψ(y)
0 0

−0.5 −0.5

−1 −1
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
y y

Small Gain Theorem Passivity theorem

1 1
k2

0.5 0.5
ψ(y)

ψ(y)
0 0

−k2
−0.5 −0.5

−1 −1
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
y y

Figure 7.6 Sector conditions on memoryless nonlinearity.

Nyquist criterion Circle Criterion

1 1

0.5 0.5
ψ(y)
ψ(y)

0 0

−0.5 −0.5

−1 −1
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
y y

Small Gain Theorem Passivity theorem

1 1

0.5 0.5
ψ(y)

ψ(y)

0 0

−0.5 −0.5

−1 −1
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
y y

Figure 7.7 Conditions on Nyquist curve matched to sector conditions on memo-

ryless nonlinearity.

72
 Solutions to Chapter 4

SOLUTION 4.3
(a) The systems belong to the sectors [0, 1], [0, ∞] and [−1, ∞] respec-
tively.
(b) Only the saturation nonlinearity (the leftmost nonlinearity) has finite
gain, which is equal to one. The other two nonlinearities have infinite
gain.
(c) The saturation and the sign nonlinearity are passive. The rightmost
nonlinearity is not passive.

SOLUTION 4.4
Note that the circle theorem in Slotine and Li is stated erroneously. The
second and third case must require that ρ = 0, i.e., that the open loop
system is Hurwitz. A correct version of the theorem in Slotine and Li would
read

THEOREM 7.1—4.13 CIRCLE CRITERION

If the system

ẋ = Ax − Bψ ( y)
y = Cx

satisfies the conditions

1) the matrix A has no eigenvalues on the jω axis, and has ρ eigenval-
ues strictly in the right half-plane;
2) the nonlinearity ψ ( y) belongs to the sector [ k1 , k2 ].
and, in addition, one of the following holds
3a) 0 < k1 ≤ k2 , the Nyquist plot of G (iω ) does not enter the disk
D ( k1 , k2 ) and encircles it ρ times counter-clockwise.
3b) 0 = k1 < k2 , ρ = 0, and the Nyquist plot of G (iω ) stays in the
half-plane Rep > −1/ k2
3c) k1 < 0 < k2 , ρ = 0, and the Nyquist plot of G (iω ) stays in the interior
of the disk D ( k1 , k2 )
3d) k1 < k2 < 0, the Nyquist plot of − G (iω ) does not enter the disk
D (− k1 , − k2 ) and encircles it ρ times counter-clockwise
then the equilibrium point 0 of the system is globally asymptotically stable.

Now, since the linear part of the system is Hurwitz, we are free to use all
versions of the circle criterion.
(a) In order to guarantee stability of a nonlinearity belonging to a sym-
metric sector [−α , α ], the Nyquist curve has to stay strictly inside a
disk centered at the origin with radius 1/α . We may, for instance,
take α = 0.25 − ε for some small ε > 0.
(b) The Nyquist curve lies inside the disk D (−1.35, 4.35). Thus, stability
can be guaranteed for all nonlinearities in the sector −0.23, 0.74.

73
 Solutions to Chapter 4

Nyquist Diagrams

0.5

0.4

0.3

0.2

0.1

Imaginary Axis
0

−0.1

−0.2

−0.3

−0.4

−0.5
−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5
Real Axis

Figure 7.8 The Nyquist curves for the system in Exercise 4.6b, and the circle
corresponding to k1 = −2, k2 = 7.

(c) We must find β such that the Nyquist plot lies outside of a half-plane
Re( G (iω )) < −1/β . A rough estimate from the plot is β = 1.1.

SOLUTION 4.5
The open loop system has one unstable pole, and we are restricted to apply
the first or fourth version of the circle criterion. In this example, we can
place a disk with center in −3 and with radius 0.75, and apply the first
version of the Nyquist criterion to conclude stability for all nonlinearities
in the sector [0.27, 0.44].

SOLUTION 4.6
(a) The Nyquist diagram is a circle with midpoint in −0.5 and radius
0.5, see Figure 4.4. Since the open system is unstable the Nyquist
curve should encircle the disc twice. Choosing the circle that passes
through −1/ k1 = −1 + ε and −1/ k2 = −ε we see that the loop is
1
stable for the sector ( , 1/ε ).
1−ε
(b) The circle with k1 = −2, k2 = 7 does not intersect the Nyquist curve.
Hence the sector (−2, 7) suffices. As always there are many other
circles that can be used (The lower limit can be traded against the
upper limit).

SOLUTION 4.7
The system is written on the standard feedback connection form that we
know from the lectures. Since the gain of the time-varying element is
bounded, hh f (t)hh2 ≤ 1, it is appropriate to apply the Small Gain Theorem.
(a) Introduce y = Cx and u = f (t) y, then
ẋ = Ax + Bu
y = Cx

74
 Solutions to Chapter 4

Y (s) = C X (s) = C(sI − A)−1 B U (s) = G (s) U (s)

(b) Let the gain of the linear system be γ . Then, the feedback connection
is BIBO-stable if γ < 1. Since the gain of a linear system is given by

γ = sup h G (iω )h = sup C(iω I − A)−1 B . (7.8)

ω ∈(0,∞) ω ∈(0,∞)

the desired results follow directly.

(c) Only the leftmost Nyquist curve shows a system with gain greater
than one. Thus, systems corresponding to the middle and the right-
most Nyquist curve are guaranteed to give a BIBO stable feedback
loop.
(d) Follows from the definition of gain for linear time invariant MIMO
systems.

SOLUTION 4.8
(a) >> A=[1 10; 0 1];svd(A)
ans =
10.0990
0.0990
(b)
 
i AB xi i AB xi i B xi
σ 1 ( AB ) = sup = sup ⋅
x i xi x i B xi ix
 
i Ayi i B xi
sup ⋅ sup = σ 1 ( A)σ 1 ( B )
y i yi x i xi

SOLUTION 4.9
The proof follows directly from the definition of passivity, since, according
to the definition of a storage function
Z T
〈u, y〉T = uT y dt
0
Z T
≥ V̇ ( x)dt = V ( x(T )) − V ( x(0)) = V ( x(T ))
0

which is non-negative since V ( x(T )) ≥ 0. Passivity follows.

75
 Solutions to Chapter 4

SOLUTION 4.10
The linear system G (s) corresponds to

ẋ = Ax + Bu, y = B T Px, x(0) = 0.

Let V = x T Px. Then

V̇ = ẋ T Px + x T P ẋ
= x T ( AT P + PA) x + 2x T P Bu = − x T x + 2 y T u ≤ 2 y T u

Integrate and use the fact that V (0) = 0, then

Z T
y T udt ≥ V (T ) − V (0) ≥ 0,
0

which proves passivity.

SOLUTION 4.11
(a) We will use the storage function V ( y, u) = y2 /2. First note that
d
V ( y, u) = y(−2 y + sat( y) + u)
dt
and

y2 ≥ ysat( y) ≥ 0.

which gives
d
V ( y, u) ≤ yu − y2 ≤ yu
dt
Hence, F is passive.
(b) Pick u such that h yh ≤ 1 for some amount of time T. (After the time
T, you can set the input u(t) = 0) Then, for t ∈ [0, T ], the system
behaves like the linear time invariant system
1
G ( s) = ,
s+1
which is (strictly) passive. Hence the nonlinear system is (strictly)
passive.
(c) You can solve this problem in many ways. Here we will base our proof
on the circle criterion. With the obvious state vector x = (θ , ω , z)P , we
rewrite the system in the feedback connection form
   
0 1 0 0
ẋ = Ax − Bψ ( y) = −2 −1 1 x − 0  sat([ 1 0 −1 ] x)
  
2 0 −2 1
The Nyquist curve of the linear system is illustrated in Figure 7.9.
Since the Nyquist curve does not intersect the half plane Re( G (iω )) <
−1/2, we conclude stability for all ψ in the sector [0, 2]. Since the
saturation element lies in the sector [0, 1], we conclude stability of
the closed loop.

76
 Solutions to Chapter 4

Nyquist Diagrams

0.8

0.6

0.4

0.2

Imaginary Axis
0

−0.2

−0.4

−0.6

−0.8

−1
−1.5 −1 −0.5 0 0.5
Real Axis

Figure 7.9 Nyquist curve of linear subsystem.

SOLUTION 4.12
(a) We have
Z T
〈 y, u〉 = y(t)u(t)dt =
0
Z T
= {u(t)uc (t)}{ H (u(t)uc (t))}dt =
0
Z T
= w(t) H (w(t))dt = 〈w, H (w)〉
0

where w = uc u. Since H is passive, the result follows.

(b) We will only consider the case where θ 0 = 0. The case θ 0 is a little
tricker, and the discussion about this case is postponed to the course
in Adaptive Control.
If θ 0 = 0, the system equations read

e(t) = G ( p)θ uc (t)

θ˙ (t) = −γ uc (t) e(t)

In light of exercise (a), we can identify the second equation modeling

the signal w(t) = uc (t) e(t) sent into an integrator with gain γ and
postmultiplied by uc (See the lecture slides for an illustration). This
system is passive, and interconnected in a negative feedback loop
with the strictly passive system G. Stability now follows from the
passivity theorem.

SOLUTION 4.13
No solution yet.

77
Solutions to Chapter 5
SOLUTION 5.1
Use the interpretation of describing function as “equivalent gain”. We have
1-b, 2-c, 3-a, 4-d.

SOLUTION 5.2
Denote the nonlinearity by f . For memoryless, static nonlinearities, the
describing function does not depend on ω , and the describing function in
Slotine and Li reduces to
b1( A) + ia1 ( A)
N ( A) =
A
where a1 and b1 can be computed as
Z 2π
1
a1 = f ( A sin(φ )) cos(φ ) dφ (7.9)
π 0
Z 2π
1
b1 = f ( A sin(φ )) sin(φ ) dφ . (7.10)
π 0

(a) First, we notice that the saturation is a odd function, which implies
that a1 = 0. In order to simplify the computations of b1, we set H = 1
and note that the saturation can be described as
(
A/ D sin(φ ) 0 ≤ φ ≤ φl
f ( A sin(φ )) =
1 φ l < φ < π /2

Here, φ l = arcsin( D / A) denotes the value of φ where so that f satu-

rates. Now,
Z 2π
1
b1 = u(φ ) sin(φ )dφ =
π 0
Z π /2
4
= u(φ ) sin(φ )dφ =
π 0
Z Z !
φl π /2
4 2
= A/ D sin (φ )dφ + sin(φ )dφ =
π 0 φl
Z Z !
φl π /2
4
= A/(2D )(1 − cos(2φ ))dφ + sin(φ )dφ =
π 0 φl
4
= ( A/(2D )(φ l − sin(φ l ) cos(φ l )) + cos(φ l )) =
π  
2A D
= φ l + cos(φ l )
Dπ A

Thus, the describing function for the normalized saturation is

2 D
N ( A) = (φ l + cos(φ l ))
Dπ A

78
 Solutions to Chapter 5

Now, using the calculation rule Nα f ( A) = α N f ( A), we find that for

the saturation under consideration we have
2H D
N ( A) = (φ l + cos(φ l ))
Dπ A

(b) We see that the nonlinearity is a superposition of a linear function

H
g( e) = e
D
and the nonlinearity − f ( e) with f ( e) as in (a). Using the fact that a
linear function g( e) = ke has describing function N ( A) = k, and the
superposition rule N f +g ( A) = N f ( A) + Ng ( A), we find
  
H 2 D
N ( A) = 1− φ l + cos(φ l )
D π A

(c) Noting that this nonlinearity can be written as the sum of the two
nonlinearities in (a) and (b), we arrive at the describing function
 
2(α − β ) D
N ( A) = φ l + cos(φ l ) + β .
π A

SOLUTION 5.3
Let the input to the relay be

u(t) = A sin(ω t) = A sin(φ )

The output of the relay is then



 −H 0 < φ < φ0

y(φ ) = H φ0 < φ < π + φ0



−H π + φ 0 < φ < 2π

where φ 0 = arcsin( D / A). We get

Z
1 2π
a1 = y(φ ) cos(φ )dφ
π 0
Z φ0 Z π +φ 0 Z 2π
1 1 1
= (− H ) cos(φ )dφ + H cos(φ )dφ + (− H ) cos(φ )dφ
π 0 π φ0 π π +φ 0
4H
=− sin(φ 0 )
π
and
Z 2π
1
b1 = y(φ ) sin(φ )dφ
π 0
Z φ0 Z π +φ 0 Z 2π
1 1 1
= (− H ) sin(φ )dφ + H sin(φ )dφ + (− H ) sin(φ )dφ
π 0 π φ0 π π +φ 0
4H
= cos(φ 0 )
π

79
 Solutions to Chapter 5

We obtain

4H
N ( A) = (cos(φ 0 ) − i sin(φ 0 ))
πA
q
The identity cos( z) = 1 − sin2 ( z) gives the desired result.

SOLUTION 5.4
Follows from the integration rule
Z
1
f (ax)dx = F (ax)
a
R
where F ( x) = f ( x)dx.

SOLUTION 5.5
We have
φ ( x) φ (a)
< , x < a.
x a
and thus
Zπ
2
Φ(a) = φ (a sin(θ )) sin(θ )dθ
aπ 0
Zπ
2 φ (a)
< a sin(θ ) sin(θ )dθ
aπ 0 a
Zπ
2
= φ (a) sin2 (θ )dθ = φ (a)/a
aπ 0

SOLUTION 5.6
The describing function is

N ( A) = k1 + 3A2 k3 /4

Note, however, that the output y(T ) of the nonlinearity for the input e(t) =
A sin(φ ) is

y(t) = A2 k2 /2 + ( k1 A + 3A3 k2 /4) sin(φ )

− A2 k2 /2 ⋅ cos(2φ ) − A3 k3 /4 ⋅ sin(3φ )

We conclude that the term k2 x22 does not influence N ( A). Still, we can not
just apply the describing function method, since there is a bias term. If the
linear system has integral action, the presence of a constant offset on the
input will have a very big influence after some time.

SOLUTION 5.7

80
 Solutions to Chapter 5

Nyquist Diagrams

Imaginary Axis
0

−1

−2

−3

−4

−8 −7 −6 −5 −4 −3 −2 −1 0 1
Real Axis

Figure 7.10 Nyquist curve and −1/ N ( A) for oscillator example.

(a) When the saturation works in the linear range, we have the closed
loop dynamics

−5ks
G ( s) =
s2 + (1 − 5k)s + 25

which is unstable for k > 0.2. Thus, the state can not remain small.
In saturation, on the other hand, the nonlinearity generates a con-
stant(“step”) input to the system. The final value theorem then gives

−5s
lim y(t) = lim =0
t→∞ s→0 s2 + s + 25

The observation that y(t) → 0 contradicts the assumption that the

nonlinearity remains saturated.
(b) We should investigate intersection of the Nyquist curve and −1/ N ( A).
Since A ∈ [ k, ∞), −1/ N ( A) lies in the interval (−∞, −1/ k).
The frequency response of the system is

−i5ω
G (iω ) =
25 − ω 2 + iω

which intersects the negative real axis for ω P = 5 rad/s. The value
of G (iω P ) = −5. Thus, there will be an intersection if k > 0.2. The
frequency of the oscillation is estimated to 5 rad/s, and for fixed k,
the amplitude can be determined directly from Figure 7.10.
(c) The Nyquist curve of the system is shown in Figure 7.10. The func-
tion −1/ N ( A) is also displayed, with an arrow in the direction of in-
creasing A. The Nyquist curve encircles the points Re( G (iω )) > −5,
indicating increased oscillation amplitude. The points to the left of
the intersection are not encircled, indicating stability and a decaying
oscillation amplitude. We can thus expect a stable limit cycle.

81
 Solutions to Chapter 5

SOLUTION 5.8
(a) Introduce θ 0 = arcsin(a/ A) and proceed similarly to the saturation
nonlinearity.
(b) The describing function has maximum for
√
A∗ = 2a

which gives

2
N ( A∗ ) =
πa
√
√ crosses the negative real axis for ω = 2, for which
The Nyquist curve
the gain is G (i 2) = −2/3. Thus, we should expect no oscillations if

4
a> .
3π

SOLUTION 5.9
(a) The describing function for a relay with amplitude D is given by

4D
N ( A) =
πA
−1/ N ( A) lies on the negative real axis. If the Nyquist curve intersects
the negative real axis, the describing function methods will predict a
sustained oscillation

4D
− h G (iω u )h = −1
πA
Thus, given the amplitude A of the oscillation, we estimate the ulti-
mate gain as

4D
Ku = 1/h G (iω u )h =
πA
The ultimate period is the period time of the oscillations

Tu = 2π /ω

(b) From the simulation, we estimate the amplitude A = 0.6 which gives
Ku  2.12. The ultimate period can be estimated directly from the
plot to be Tu  2. Note that the estimates have good correspondence
with the analytical results (which require a full process model)

SOLUTION 5.10
No solution yet.

82
 Solutions to Chapter 5

SOLUTION 5.11
No solution yet.

SOLUTION 5.12
No solution yet.

SOLUTION 5.13
No solution yet.

83
Solutions to Chapter 6
SOLUTION 6.1
We would like to write the system equations as

v = G (s)(−u)
u = φ (v)
where φ (⋅) denotes the saturation. Block diagram manipulations give
 
AR BS
v=u− + u
Aw A Aw A
 
AR + B S
= − 1 (−u) = G (s)(−u)
AAw
Since the saturation element belongs to the sector [0, 1], we invoke the
circle criterion and conclude stability if the Nyquist curve of G (iω ) does
not enter the half plane Re( G (iω )) < −1. This gives the desired condition.

SOLUTION 6.2
The model is given by
dz hvh
= v− z (7.11)
dt g(v)
dz
F = σ 0 z + σ 1(v) + Fvv (7.12)
dt
(a) For any constant velocity, v, (7.11) converges to the value
g(v)
z= v = g(v)sign(v)
hvh
and F therefore converges to
F = σ 0 g(v)sign(v) + Fvv

(b) Following the hint, we consider V ( z) = z2 . Along trajectories of the

model, we have
hvh
V̇ = 2z(v − z)
g(v)
h zh
≤ 2h zhhvh(1 − )
g(v)

which is non-positive if h zh/g(v) ≥ 1. Since 0 ≤ g(v) ≤ a, we see that for

h z(t)h ≥ a

we have V̇ ≤ 0. We conclude that

Ω = { zh zT z < a2 }
is invariant. In other words, all trajectories that start within Ω remain
there for all future times. The set Ω provides a bound on the state z(t).

84
 Solutions to Chapter 6

(c) Consider the storage function V ( z) = z2 /2. We have

dz hvh 2 dz
zv = z + z ≥z = V̇ (t)
dt g(v) dt

and passivity follows from the discussion in (+).

(d) Consider the storage function V ( z) = σ 0 z2 . We have

hvh
Fv = Fvv2 + (σ 1 ż + σ 0 z)( ż + z) (7.13)
g(v)
hvh hvh
≥ σ 1 ż2 + σ 0 z2 + ( σ 1 + σ 0 ) zż (7.14)
g(v) g(v)

Next, we separate out the storage function derivative and make a comple-
tion of squares to estimate the additional terms

hvh 2 hvh
Fv ≥ σ 0 zż + σ 1 ż2 + σ 0 z + σ1 zż (7.15)
g(v) g(v)
 2  2 !
hvh hvh hvh
= V̇ + σ 1 ż + z + σ0 − σ1 z2 (7.16)
2g(v) g(v) 2g(v)

Since the second term is non-negative, we have

Fv ≥ V̇

and thus passivity if

hvh
σ0 − σ1 >0
4g(v)

This concludes the proof.

SOLUTION 6.3
(a) The describing function for a relay has been derived on Lecture 6 to
be
4F0
N ( A) =
πA

(b) Using the superposition property of describing functions for static

nonlinearities N f +g = N f + Ng , and the fact that for a scalar gain
y = ku the describing function is N ( A) = k, we obtain

4F0
N ( A) = Fv +
πA

(c) Stiction is a point-wise phenomenon (occurring for v = 0) with finite

amplitude, and has no influence on the integral calculations involved
in the describing function computation. The describing function is
therefore the same as in (b).

85
 Solutions to Chapter 6

SOLUTION 6.4
Recall from the lecture slides that the process is given by

ẋ = v
v̇ = − F + u

The velocity is observed through

v = zv + Kv x
b
żv = − Fb + u − Kv vb

b denotes the estimated friction force, estimated by the observer

where F

b = ( zF + K F hb
F vh)sign(b
v)
b)sign(b
żF = − K F (u − F v)

Defining the observer errors

ev = v − vb
b
eF = F − F

we obtain the observer dynamics

b + u − Kv vb) − Kv v = − eF − Kv ev
v = − F + u − (− F
ėv = v̇ − ḃ

ėF = Ḟ − ċ b) − K f (− F̂ + u − Kv b
F = Ḟ − (− K F (u − F v) − Kv v) =
= Ḟ − K f Kv ev

The term Ḟ is zero (except at zero velocity where it is not well defined).
Putting Ḟ = 0, we obtain
    
ėv − Kv −1 ev
= (7.17)
ėF − Kv K f 0 eF

with the characteristic equation

λ ( s ) = s 2 + Kv s − Kv K f

We conclude that the error dynamics are locally asymptotically stable if

Kv > 0 ,
− Kv K f > 0

which implies the desired conditions.

86
 Solutions to Chapter 6

D
− D /2
D /2
−D

Figure 7.11 The function in 6.5.

SOLUTION 6.5
We have already (lecture) seen that the describing function for the function
in Figure 7.11 is given by

 0 A < D /2
N D ( A) = q
 4D 1 − ( D )2 A > D /2
πA 2A

Superposition ( N f 1 + f 2 = N f 1 + N f 2 ) gives

NQ = N D + N3D + N5D + . . . + N2i+1

which gives the stated describing function NQ ( A) for the quantizer.

SOLUTION 6.6
We have
Z T Z T
〈u, y〉T = u y dt = usat(u) dt ≥ 0
0 0

We conclude passivity from to the definition given in the lecture slides.

SOLUTION 6.7
No solution yet.

SOLUTION 6.8
Assume without loss of generality that 0 < u0 < D /2. The input to the
quantizer is u0 + d(t) where d(t) is the dither signal. The output y from
the quantizer is
(
0 u0 + d(t) < D /2
y(t) = Q(u0 + d(t)) =
D u0 + d(t) < D /2

It is easy to see that y = D during a time interval of length uD0 T, where T

is the time period of the dither. The average value of y becomes

1 u0
y0 = T ⋅ D = u0 .
T D

87
 Solutions to Chapter 6

Hence the dither signal gives increased accuracy, at least if the signal y
can be treated as constant compared to the frequency of the dither signal.
The method does not work for high-frequency signals y.

SOLUTION 6.9
No solution yet.

SOLUTION 6.10
No solution yet.

88
Solutions to Chapter 7
SOLUTION 7.1
Let the output of the nonlinearity be u, so that u = f (v).
(a) We have

u = v2, v≥0

which implies that

√
v= u, u≥0

(b) The piecewise linear characteristic

(
k1 v, hvh ≤ d
u=
sign( k1 − k2 )d + k2 v hvh > d

gives the inverse

(
u/ k1 , huh ≤ k1 d
v=
(u − sign(u)( k1 − k2 )d)/ k2 huh > k1 d

Consider a (unit) dead-zone with slope k1 = ε in the interval hvh ≤ d,

and slope k2 = 1 otherwise. We obtain the inverse
(
u/ε , huh ≤ ε d
v=
u + sign(u)(1 − ε )d, huh > ε d

The dead-zone nonlinearity and its inverse are shown in Figure 7.12.

Dead zone Dead zone inverse

2 2

1 1
v=f−1(u)
u=f(v)

0 0

−1 −1

−2 −2
−2 0 2 −1 0 1
v u

Figure 7.12 Deadzone and its inverse.

(c) See the slides from the lecture of backlash.

89
 Solutions to Chapter 7

SOLUTION 7.2
(a) We notice that all state equation but the last one are linear. The last
state equation reads

ẋn = f ( x) + g( x)u

If we assume that g( x) = 0 for all x, we can apply the control

1
u = h( x, v) = (− f ( x) + Lx + v)
g( x)

renders the last state equation linear

ẋn = Lx + v

The response from v to x is linear, and the closed loop dynamics is

given by
   
0 1 0 0 0 0
0
 0 1 0 ...
 
0

ẋ =  ..  x +  ..  v
0 0 0 . 0  .
l1 l2 l3 ... ln 1

(You may recognize this as the controller form from the basic con-
trol course). For the control to be well defined, we must require that
g( x) = 0 for all x.
(b) The above procedure suggest the control

1
u= (−a sin( x1 ) + l1 x1 + l2 x2 + v)
b cos( x1 )

which results in the closed loop system

   
0 1 0
ẋ = x+ v
l1 l2 1

The system matrix has a double eigenvalue in s = −1 if we let

l1 = −1, l2 = −2

The control law is well defined for x1 = π /2. This corresponds to the
pendulum being horizontal. For x1 = π /2, u has no influence on the
system. Notice how the control “blows up” nearby this singularity.
Extra. You may want to verify by simulations the behaviour of the
modified control

u = sat(h( x, v))

for different values of the saturation level.

90
 Solutions to Chapter 7

(c) The above procedure suggest the control

u = − x2 − x + v

Letting v = 0, we apply the control to the perturbed system

ẋ = (1 + ε ) x2 − x2 − x = ε x2 − x

and note that for x > 1/ε , we have ẋ > 0, which implies that the
trajectories tend to infinity. Thus, global cancellation is non-robust in
the sense that it may require a very precise mathematical model.

SOLUTION 7.3
(a) The sliding surface in a sliding mode design is invariant, i.e., if x(t s )
belongs to the sliding surface σ ( x) = 0, at time t s, then it belongs to
the set σ ( x) = 0 for all future times t ≥ t s. Thus, it must hold that

σ ( x) = σ̇ ( x) = 0

which yields the dynamics on the sliding surface.

(i) We have

σ̇ ( x) = 2 ẋ1 − ẋ2 = 2 ẋ1 − x1 = 0

The third equality implies that x1 (t) → ±∞ on the sliding sur-

face. Thus, forcing this surface to be a sliding mode would give
unstable solutions.
(ii) Similarly as above

σ̇ ( x) = ẋ1 + 2 ẋ2 = ẋ1 + 2x1 = 0

Thus, the equivalent dynamics along this surface satisfy ẋ1 =

−2x1 and is hence stable.
(iii) We have

σ̇ ( x) = ẋ1 = 0

The dynamics on this sliding surface would thus be stable, but

not asymptotically stable.
(b) According to the lecture slides, the sliding mode control law is

pT Ax µ
u=− T
− T sign(σ ( x))
p B p B

Where the sliding surface is given by

σ ( x) = pT x = 0

Thus, in this example we have pT = [ 1 1 ] and

u = −( x1 + x2 ) − µ sign( x1 + x2 )

91
 Solutions to Chapter 7

(c) According to the robustness result of the sliding mode controller

presented on the lecture, the above controller will force the sys-
tem toward the sliding mode if µ is chosen large enough, and if
sign( pT B̂ ) = sign( pT B ), which implies sign(b̂) = sign(b). Since the
nominal design has b̂ = 1, we must have

b>0 (7.18)

It remains to check that the dynamics of the sliding mode remains

stable. (Otherwise, we could have a situation where the controller
forces the state onto the sliding mode, but where the sliding mode
dynamics are unstable. The state would then tend toward infinity
along the sliding mode.) In this case, we can verify that on the sliding
mode, we have σ̇ ( x) = 0 for all values of the parameter a.

SOLUTION 7.4
(a) Straightforward manipulations give

K 1
G ( s) = e−sL = e−sVd / q(t)
sT + 1 sVm/ q(t) + 1

(b) The step response gives parameters a = 0.9, L = 1. Using the results
from (a) and a = K L/T we obtain

a = Vd / Vm
L = Vd/ q(t)

Since the experiment was performed for q(t) = 1, we see that L =

1/ q(t). Now, a gain scheduled PI controller can be constructed using
Ziegler-Nichols recommendations as

K p = 0.9/a = 1
Ti = 3L = 3/ q(t)

SOLUTION 7.5
(a) The pendulum energy is given by

Jp 2
E( x) = mg l cos( x1 ) + x
2 2

If the energy of the pendulum hanging downwards is taken to be

E(0) = 0, the energy for x1 = π , x2 = 0 is E0 = 2mg l.
(b) The time derivative of the Lyapunov function candidate reads

d
V̇ ( x) = 2( E( x) − E0 ) E( x) =
dt
= 2( E( x) − E0 )(−mg l sin( x1 ) ẋ1 + J p x2 ẋ2 ) =
= 2( E( x) − E0 )(−mlx2 cos( x1 )u)

92
 Solutions to Chapter 7

Applying the suggested control, we obtain

V̇ ( x) = −2kml ( E( x) − E0 )2 x2 cos( x1 )sign( x2 cos( x1 )) ≤ 0

with equality attained for E( x) = E0, or x2 = 0 or x1 = π /2. The only

unwanted invariant manifold is x1 = x2 = 0.
(c) The phase portrait of the closed loop system is shown in Figure 7.13.
We notice how the state is driven to the set E( x) = E0 , and that
this set contains no stable equilibrium points. Note that the velocity
approaches zero as the pendulum approaches the upright position.
Since the equilibrium point is unstable, and the control for this state
is zero, the pendulum does not remain in the upright position.
Extra. Feel free to design a stabilizing controller for the upright posi-
tion (using, for example the results from Exercise 7.2). In particular,
how should you switch between the two control strategies to make
the system stable? (Some Lyapunov theory will help you on this one)

Phase plane
10

2
x2

−2

−4

−6

−8

−10
−6 −4 −2 0 2 4 6
x1

Figure 7.13 Phase plane for pendulum under energy control.

SOLUTION 7.6
We get

V̇ = σ ( x)σ̇ ( x) = ( x1 + x2 )(2x1 + u) = ( x1 + x2 )(−sign)( x1 + x2 ) = −h x1 + x2 h

and therefore σ ( x) = x1 + x2 → 0. The equivalent dynamics is easiest

determine by
d
0 = σ ( x) = ẋ1 + ẋ2 = x1 + u + x1
dt
which gives u = −2x1 and hence ẋ1 = − x1 on the sliding plane.

93
 Solutions to Chapter 7

SOLUTION 7.7

H = n0 ( x12 + u2 ) + λ 1 u
Let us first assume n0 = 1. Then the optimal control signal is

λ1
0 = Hu = 2n0 u + λ 1 ; u=−
2n0
The adjoint equation is

λ̇ 1 = − H x1 = −2n0 x1 , λ 1 (1) = free

Hence
λ̇ 1
ẍ1 = u̇ = − = x1
2n0
Therefore
x(t) = c1 et + c2 e−t
The boundary conditions gives

c1 + c2 = 1
c1 e + c2 e−1 = 0

This gives c1 = − e−2 /(1 − e−2), c2 = 1/(1 − e−2 ) and the control signal is

u = ẋ = c1 et − c2 e−t

What about the case n0 = 0? Then λ 1 is constant and λ (1) = µ = 0.

Hence H = λ 1 u has no minima in u, so this case gives no candidates.

SOLUTION 7.8
Suppose φ ( x(t f )) is the criterion to be minimized and that Ψ( x(t f )) = 0
is the end constraints to be satisfied at t = t f . Note that L = 0. We have
(with α = F /m)

H = λ 1 x3 + λ 2 x4 + λ 3α cos u + λ 4 (α sin u − g)

Minimizing H wrt u gives

λ4
tan u =
λ3
The adjoint equations are

λ̇ 1 = 0
λ̇ 2 = 0
λ̇ 3 = −λ 1
λ̇ 4 = −λ 2

This gives λ 3 (t) = Ct + D and λ 4 = At + B for some constants A, B , C, D.

How could one determine A, B , C, D? They of course depend on φ and Ψ .
The optimal trajectories can be generated by ẋ = f ( x, u∗ ), and one chooses
the trajectory that satisfies the boundary conditions on x and λ at t = t f .
It is generally a hard problem.

94
 Solutions to Chapter 7

SOLUTION 7.9
We get

u2 u2
H = λ 1 x3 + λ 2 x4 + λ 3 (cos u1 ) + λ 4 ( sin u1 − g) − λ 5γ u2
x5 x5
= σ (t, u1)u2 + terms independent of u

λ3 λ4
where σ (t, u1) = x5 cos u1 + x5 sin u1 − λ 5γ . We get minimum with

 umax σ < 0

u2 = L σ =0


0 σ >0

and (
λ4
λ3 u2 > 0
tan u1 =
L u2 = 0

SOLUTION 7.10
The problem is normal, can use n0 = 1. We have
2
H = ex1 + x22 + u2 + λ 1 x2 + λ 2 (− x1 − x23 + (1 + x1 )u)
2
λ̇ 1 = − H x1 = −2x1 ex1 − λ 2 (−1 + u)
λ̇ 2 = − H x2 = −2x2 − λ 1 + 3x22 λ 2
λ (1) = 0

Minimization of H wrt u gives

VH λ2
=0 ; 2u + λ 2 (1 + x1 ) = 0 ; u=− (1 + x1 )
Vu 2
2
( VV uH2 = 2 > 0 hence minimum). This gives

ẋ1 = f 1 = x2
λ2
ẋ2 = f 2 = − x1 − x23 − (1 + x1 )2
2
2
λ̇ 1 = f 3 = −2x1 ex1 − λ 2 (−1 + u)
λ̇ 2 = f 4 = −2x2 − λ 1 + 3x22 λ 2
λ 1 (1) = λ 2 (1) = 0

SOLUTION 7.11

95
 Solutions to Chapter 7

 
 x1 (t f ) 
a) We have L = 1, φ = 0, Ψ(t f ) = 
 
 = 0 and t f free. We get
x2 (t f )

H = n0 + λ 1 x2 + λ 2 u

Hence 
 1
 λ 2 ( t) < 0
u(t) = ? λ 2 ( t) = 0


−1 λ 2 (t) > 0
The adjoint equations are

λ̇ 1 = 0
λ̇ 2 = λ 1

This gives λ 1 = µ 1 , λ 2 (t) = µ 1 t + B. Now µ 1 = B = 0 is impossible

since this leads to n0 = 0 (since we should have H  0), and we can
not have [ n0 µ 1 µ 2 ] = 0. Therefore λ 2 (t) is linear and nonzero. This
shows a).
b) We know that u = ±1 with at most one switch. The simplest way to
get σ ( x) (note that σ should be a function of x not of t) is to solve
the equations for such input signals. For u(t) = 1 we get

x1 + C1 = x22 /2

This gives the phase plane in the figure For u = −1 we get

x1 + C2 = − x22 /2

This gives the phase plane in the figure Now we know that we get to
5

−1

−2

−3

−4

−5
−5 −4 −3 −2 −1 0 1 2 3 4 5

Figure 7.14 Phase plane for u = 1

the origin with u = ±1 with at most one switch. Consider especially

the two curves for u = ±1 that pass through the origin. From the
plots we see that the control law is given by
 
1 2
u(t) = −sign x1 (t) + sign{ x2 (t)} x2 (t)
2

since then u(t) = −1 above the switch curve and u(t) = 1 below it.

96
 Solutions to Chapter 7

−1

−2

−3

−4

−5
−5 −4 −3 −2 −1 0 1 2 3 4 5

Figure 7.15 Phase plane for u = −1

Phase plane
5

1
x2

−1

−2

−3

−4

−5
−5 −4 −3 −2 −1 0 1 2 3 4
x1

Figure 7.16 Phase plane with switching curve

SOLUTION 7.12
Since we assume the problem is normal (t f is free so this is not obvious)
we have
H = 1 + huh + λ 1 x2 + λ 2 u.
Minimization wrt huh ≤ 1 gives

λ 2 > 1 ; u = −1
hλ 2 h < 1 ; u = 0
λ 2 < −1 ; u = 1

We also have

λ̇ 1 = − H x1 = 0 ; λ 1 = B
λ̇ 2 = − H x2 = −λ 1 ; λ 2 = A − Bt

97
 Solutions to Chapter 7

for some constants A, B. If B < 0 we see that λ 2 increases (linearly) and

hence u(t) passes through the sequence 1 → 0 → −1, or a subsequence of
this. If B > 0 the (sub-) sequence is passed in the other direction −1 →
0 → −1.
If B = 0 then u is constant: either u = −1, u = 0 or u = 1. The cases
λ 2  1 and λ 2  −1 are then impossible since the condition H  0 (since
t f is free) then can not be satisfied.

SOLUTION 7.13
Alternative 1 Use the Bang-bang
  theorem (p. 472). Note that ( A, B ) is
 1 0
controllable and Ψ x =  
 has full rank, hence u(t) is bang-bang.
0 1
From “sats 18.6” we know that there are at most n − 1 = 1 switches in u
(the eigenvalues of A are −1, −2 and are hence real).
Alternative 2 Direct calculation shows

H = σ (t)u + terms independent of u

Minimization wrt u shows that huh = 3 where the sign is given by the sign
of σ (t). From λ̇ = − AT λ and λ (t f ) = Ψ Tx µ = µ we get

σ (t) = λ T B == µ T e− A(t−t f ) B = c1 e−t + c2 e−2t

for some constants c1, c2. Since σ (t) = e−t(c1 + c2 e−t ) can have at most one
sign change and there will be only one switch in u. (It is easy to check that
the case σ (t)  0 is impossible).

SOLUTION 7.14
See the solution of Problem 7.13. We have that

u(t) = sign λ T (t f ) e− A(t−t f ) B

where λ (t f ) depends on x(0).

SOLUTION 7.15
Minimization of
H = (1 + λ 1 )u
gives

1 + λ 1 = 0 : no minimum in u
1 + λ1 = 0 : all u give minima

This does not prove that all u in fact give minima. It only says that all u(t)
are so far possible minima and we need more information.
But in fact since
Z 1 Z 1
u dt = ẋ1 dt = x(1) − x(0) = 1
0 0

all u that give x1 (1) = 1 are minimizers.

98
8. Bibliography
Åström, K. J. (1968): Reglerteknik – Olinjära System. TLTH/VBV.
Boyd, S. P. (1997): “Homework assignments in ee375 – advanced analysis of
feedback.” Available from http://www-leland.stanford.edu/class/ee375/.
Khalil, H. K. (1996): Nonlinear Systems, 2nd edition. Prentice Hall, Upper
Saddle River, N.J.
Slotine, J.-J. E. and W. LI (1991): Applied Nonlinear Control. Prentice Hall,
Englewood Cliffs, N.J.

99