CONTROL THEORY, TSRT09,

TSRT06
Exercises & solutions
23 mars 2021
Reading instructions
• The names and numbers of the chapters in this exercise collection are
consistent with the names and numbers of the chapters in the textbook.

• Starred (∗) exercises deals with discrete-time systems and are optional.
1 Introduction

1.1

Consider the linear feedback control system given by the figure below.

Σ G(s)

−F (s)

Show that if the small gain theorem (Swe: lågförstärkningssatsen) is fulfilled

the Nyquist criterion is also fulfilled.

1.2

Consider a static nonlinear system described by an ideal relay given by the

function  
 1, u>0
y(t) = f (u(t)) = 0, u = 0 .

−1, u<0
What is the gain of the relay?

1
1.3

Consider the system

2
Y (s) = G(s)U(s) G(s) =
s2 + 2s + 2

The control signal goes through a valve with saturation


1,
 if u(t) > 2
1
ũ(t) = 2 u(t), if |u(t)| ≤ 2


−1, if u(t) < −2

The output is thus

y(t) = G(p)ũ(t).

The system is controlled using proportional feedback, i.e. u(t) = −Ky(t). For
what values of K is the closed-loop system guaranteed to be stable according
to the small gain theorem?

1.4

Compute the norms k · k∞ and k · k2 of the continuous-time signals

(a)

a sin(t), t > 0
y(t) =
0, t≤0

(b)
 1
t
,t>1
y(t) =
0, t ≤ 1

(c)

e−t (1 − e−t ), t > 0
y(t) = .
0, t≤0

2
1.5

Consider the linear system

ω02
G(s) = 2 .
s + 2ζω0s + ω02

Compute the system gain kGk for all values of ω0 > 0 and ζ > 0.

1.6

Analyze the stability of the following system, first by using the small gain
theorem and then by computing the poles of the closed-loop system. Explain
possible differences.

a
Σ
s+a

1.7

Consider the feedback control system

Σ G(s)

−f (·)

where G(s) is a linear system with the magnitude plot

3
 |G(iω)|
1.5
1

10 ω [rad/s]

and f (·) s an amplifier with the following input-output relationship

f (x)
0.5

1 x

Is the closed-loop system stable?

1.8

Consider a DC motor given on state-space form

ẋ1 = x2
ẋ2 = −ax2 + au
y = x1

The inverse time constant a can vary as

a = 1 + ρ, |ρ| < δ.

The system is controlled using a proportional controller u(t) = −Ky(t).

Suppose that a is constant. Give a sufficient condition on K such that the
closed-loop system is stable for all a.

4
1.9

Once again consider the DC motor in exercise 1.8, but now assume that the
parameter a can vary arbitrarily fast with time

a = a(t) = 1 + ρ(t), |ρ(t)| < δ, ∀t

(a) Introduce a new, artificial input signal w and a new artificial output sig-
nal z such that the system can be described by the feedback connection
below

−K

u y
G
w z

ρ(t)

(b) Consider the time-varying, static system from z(t) to w(t):

w(t) = ρ(t) · z(t), |ρ(t)| < δ, ∀t.

Show that the gain of this system (according to Definition 1.1 in the
textbook) is at most δ.

(c) Give a sufficient condition on K, for instance an inequality that impli-

citly characterizes K, for the closed-loop system to be stable no matter
how a(t) varies with time.

5
6
2 Representation of Linear Systems

2.1

A simplified model of an alternating-current generator can be described as

follows. The input signals to the system are the magnetizing current Im ,
which is fed into the armature winding, and the driving torque M which is
applied to the rotor axis. The rotation speed of the generator is ω, and the
change in rotation speed is given by

J ω̇ = M − Me

where
Me = Ke · ω · If
is the electrical torque due to the emf. If is the current in the stator winding,
given by the relationship
e = R · If
where the voltage e is generated in the stator winding according to

e = C e · Im · ω

and R is the load resistance applied to the stator winding. Consider e and ω
as output signals, M, Im and R as input signals. Set Ke = Ce = J = 1 and
find a state-space representation for this sytem.

Linearize around the stationary point

ω0 = R0 = Im0 = M0 = 1

and derive the transfer function matrix from

 
∆M  
∆ω
u = ∆Im
  to y =
∆e
∆R

7
2.2

Consider the system consisting of two coupled tanks described in the figure
below.

u1 u2

h1
h2
y

The flow of water into the left and right halves of the tank are denoted u1
and u2 respectively. These flows are the input signals. The water levels in the
two halves are denoted h1 and h2 respectively. The flow y out from the tank
is assumed to be proportional to the water level in the right half of the tank

y(t) = αh2 (t)

The flow between the two halves is proportional to the difference between
the levels
f (t) = β(h1 (t) − h2 (t))

where a flow from left to right is considered positive. Let hi , ui and y be

deviations from nominal values. Thus, they can have negative values. Assume
that the area of the halves are A1 = A2 = 1.

(a) Derive the transfer function from u1 , u2 to y.

(b) Compute the maximum and minimum singular values of G(0) and give
an intuitive explanation to the corresponding input signals.

8
2.3

Find a state-space realization of the system

h i
1 s+3
G(s) = (s+1)(s+2) (s+1)(s2 +s+1)

2.4

Find a state-space realization of the system

p p−1
y(t) = u1 (t) + 2 u2 (t)
p2 + 4p + 4 p + 5p + 6

2.5

A system is described by the differential equation

ÿ + a1 ẏ + a2 y = b11 u̇1 + b12 u1 + b21 u̇2 + b22 u2 .

Find a state-space realization.

2.6

Consider the system


ẏ1 + y2 = u̇ + 2u
ẏ2 + y2 + y1 = u

Find a state-space realization.

9
10
3 Properties of Linear Systems

3.1

Consider the transfer function matrix

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

Derive the pole and the zero polynomials of the system? What is the dimen-
sion of a minimal state-space realization?

3.2

Find the poles and zeros of

 
1 1 0
G(s) =
(s + 1)(s + 3) −1 2(s + 1)2

3.3

Find the poles of

 
1 1 − s 31 − s
G(s) = .
(s + 1)2 2−s 1−s
What is the dimension of a minimal realization?

3.4

(a) Consider the system

s+5 1
!
s2 +3s+2 s+2
G(s) = 1 1
s+4 s+2

What is the dimension of a minimal state-space realization?

11
 (b) Consider the system
s+5 1
!
s2 +3s+2 s+2
G(s) = 1 1
s+4 s+4

What is the dimension of a minimal state-space realization?

3.5

A system has the following input-output relation


ẏ1 + y1 − ẏ2 = u1 − u2
.
ẏ2 + ẏ1 + y2 = u1 + u2

Find a matrix fraction description , y(t) = A(p)−1 B(p)u(t), and compute the
poles and zeros of the system.

3.6

Consider the MIMO system:

   
−2 1 0 0 1
ẋ(t) =  0 −1 0  x(t) + 1 0 u(t)
0 0 −3 1 0
 
1 0 0
y(t) = x(t)
0 1 0

Find a minimal relization of the system, i.e. a realization that is controllable

and observable.

12
3.7

Consider the multivariable system

Y (s) = G(s)U(s)

where  
1 3
G(s) = s+1 s+2
2 1
s+3 s+4

(a) Determine the maximum and minimum singular value of the frequency
reponse at the frequency ω = 2 rad/s.

(b) Determine also the input vectors, in terms of their Fourier transforms,
corresponding to the largest and smallest gain of the system at ω = 2.

(c) Generate, in Matlab, an input vector that corresponds to the largest

gain of the system and simulate the system using this input.
Hint: Use sinusoidal input signals and use the following properties of
a sinusoidal signal considered over a finite time interval.

– The Fourier transform is proportional to the amplitude of the

sinusoidal signal, i.e. for u1 (t) = A sin ωt the Fourier transform
U1 (iω) is proportional to A.
– Time delay of a signal corresponds to a change of the argument
of the Fourier transform, i.e. if u1 (t) = A sin ωt has the transform
U1 (iω) the signal A sin(ωt + φ) has the transform U1 (iω)eiφ .

(d) Verify that the obtained output signals correspond to largest gain of
the system.

3.8

Find the poles and zeros of

 1 s−1 
s+1
0 (s+1)(s+2)
G(s) = −1 1 1 .
s−1 s+2 s+2

13
14
5 Disturbance Models

5.1

A continuous-time stochastic process u(t) has the power spectrum Φu (ω).

For the power spectra below, find linear filters such that the processes can
be represented as white noise fed through those filters.

a2
(a) Φu (ω) =
ω 2 + a2

a2 b2
(b) Φu (ω) =
(ω 2 + a2 )(ω 2 + b2 )

5.2

A position sensor is mounted on a machine that vibrates with a frequency

around 5 Hz, and this causes that a disturbance n(t) affects the position mea-
surement. In order to include the properties of the measurement disturbance
in the control design one formulates a a model that describes the proper-
ties of the disturbance as filtered white noise V . The following models are
suggested
1
(i) N(s) = V (s)
s + 0.001

900
(ii) N(s) = V (s)
s2 + 6s + 900

25
(iii) N(s) = V (s)
s2 + s + 25

Which disturbance model is the best choice?

15
5.3

Consider a missile propelled by the thrust u. The missile’s position is z. A

simplified model for the air drag is

f = k1 · ż + v

where v are, more or less, random wind gusts.

(a) Derive a state-space representation and an input-output representation

for how the controlled output z depends on u and v.

(b) The system disturbance v has the spectral density

1
Φv (ω) = k0 ·
ω2 + a2
Modify the state-space representation in (a) to make it possible to
express the system disturbance using white noise. What is the corre-
sponding transfer function?

5.4

Assume, in exercise 5.3, that the position z is measured with an error

y(t) = z(t) + n(t)

Derive a state-space model for the missile if

(a)
Φn (ω) = 0.1

(b)
ω2
Φn (ω) = 0.1 .
ω 2 + b2
(c)
1
Φn (ω) = 0.1 .
ω 2 + b2

16
5.5

A system has the state-space representation

ẋ = Ax + Bu + Nw
y = Cx + n

We assume that the system disturbance w changes stepwise and that the
measurement noise is periodical with a frequency of about 2 Hz.

Modify the state-space representation to make it possible to model the distur-

bances.

5.6

In airplanes it is common to measure acceleration as well as speed. The

acceleration is measured using accelerometers and the speed is calculated
from measurements of air data, such as dynamical pressure et cetera. Thus,
the measurements are independent, but of course they are related to each
other.

(a) Derive a state-space model for the speed and acceleration. Let the me-
asured speed and acceleration be output signals and assume that the
derivative of the acceleration is white noise. Furthermore, assume that
the measurement errors in speed and acceleration are white noises, in-
dependent of each other.

(b) Discuss how we can get better estimates of the speed and acceleration
using Kalman filtering.

5.7

The depicted dynamical system is described by the differential equation

17
 x(t)

v(t)

ẍ(t) + x(t) = v(t)

The external force v(t) is white noise with

E v(t) = 0
E v(t)v(s) = δ(t − s)

We want to estimate the position x(t) and speed ẋ(t) at every time instant.
We have sensors for both speed and position but for economical reasons we
only want to use one sensor. We can choose between

Alternative I: The measured signal is

y1 (t) = x(t) + e1 (t)

Alternative II: The measured signal is

y2 (t) = ẋ(t) + e2 (t)

The measurement errors are e1 (t) and e2 (t). For simplicity we assume that
they are both white noises with

Ee1 (t) = Ee2 (t) = E[e1 (t)e2 (s)] = 0

Ee1 (t)e1 (s) = Ee2 (t)e2 (s) = δ(t − s)

For each alternative derive the linear filter that, in steady state, yields the
best estimate of x(t) and ẋ(t), in the sense of smallest variance of the esti-
mation error, from measurements up to and including time t. State, with an
explaination, which alternative you think is the best.

18
5.8

Consider the depicted radar antenna.

Θ
Antenna

Motor

From noisy measurments of the position of the antenna Θm we want to

estimate the true position Θ. To be able to do this we need a model of
the system. To this end, describe the dynamics of the antenna with
J Θ̈(t) + B Θ̇(t) = τ (t) + τd (t),
where J is the moment of inertia for the moving parts of the antenna, B is
the coefficient of viscous friction, τ (t) is the torque produced by the motor,
and τd (t) is the torque caused by the wind. Assume that τd (t) can be modeled
as white noise. Furthermore, assume that the torque τ (t) is proportional to
the motor voltage , µ(t), i.e.
τ (t) = kµ(t)
Finally, let us for simplicity, assume that the measurement error can be mo-
deled as additative white noise em (t). Hence, the output signal is
Θm (t) = Θ(t) + em (t).
Discuss how Θ(t) can be estimated from Θm (t) using a Kalman filter.

Technical data:
B/J = 4.6 s−1
k/J = 0.787 rad /Vs 2
J = 10 kg m2
E τd (t)τd (s) = vd δ(t − s) = 10 N 2 m2 · δ(t − s)
E em (t)em (s) = vm δ(t − s) = 10−7 rad 2 · δ(t − s)

19
5.9

Consider an electric motor with transfer operator

1
G(p) =
p(p + 1)

from input voltage to actual angular displacement. The motor operates in

two disturbance modes:

(i)
y(t) = G(p)(u(t) + w(t))

(ii)
y(t) = G(p)u(t) + w(t)

1
In both cases we have w(t) = p
v(t) where v(t) is a unit disturbance, for
example an impulse.

(a) Realize both cases on state-space form. For case (ii) it is assumed that
the states caused by the disturbance are separate from the ones descri-
bing the motor dynamics.

(b) For both cases, give examples of physical phenomena that can be mo-
deled with the disturbance w(t) .

(c) Study the two state-space realizations. Are all states controllable? Can
states corresponding to w(t) be made unobservable? Can the influence
of w(t) on y(t) be eliminated?

5.10

Consider the movement of a swing due to the wind. The swing is descibed
by the transfer operator
1
y(t) = u(t)
p2 +p+1

20
where the output signal y(t) is the angular displacement and the input signal
u(t) is the torque about the point of suspension. The influence of the wind
can be modeled as
u(t) = Kv(t)
where v(t) is a Gaussian distributed disturbance with the spectrum
2α
Φv (ω) = , α > 0.
α2 + ω2
K quantifies the strength of the wind and α quantifies the gustiness of the
wind.

(a) Does α increase or decrease when the gustiness increases, i.e. when the
wind changes direction more frequently?

(b) Derive and interprete conditions on α and K such that the swing has
an angular displacement of more than 1.15 at least a quarter of the
time. This is equivalent to the output having a variance greater than
1.

Hint:
Z ∞
1 |b2 (iω)2 + b1 iω + b0 |2
dω
2π −∞ |(iω)3 + a2 (iω)2 + a1 iω + a0 |2
b2 a0 a1 + (b21 − 2b0 b2 )a0 + b20 a2
= 2
2a0 (−a0 + a1 a2 )

21
22
6 The Closed-Loop System

6.1

For a given system G and a given controller F we have defined four transfer
functions as
Gwu u = (I + F G)−1 , Gwu = −(I + F G)−1 F
Gwu y = (I + GF )−1 G, Gwy = (I + GF )−1
All four transfer functions have to be stable for the closed-loop system to be
internally stable.

Show that    −1

Gwu u Gwu I F
=
Gwu y Gwy −G I

6.2

The system
s−1
G(s) =
s+1
and the controller
s+2
F (s) =
s−1
are are used in the feedback connection depicted below.

Σ G(s)

−F (s)

Compute Gc , T and S. Are they stable? Is the closed-loop system internally

stable?

23
24
7 Limitations in Control Design

7.1

Given the system

s−3
G(s) = .
s+1
we want the complementary sensitivity function to be
5
T (s) = .
s+5

(a) Compute a controller Fr = Fy = F which results in this T . Will this

controller really work?
(b) Suggest an alternative T , still having the bandwidth 5 rad/s, but re-
sulting in an internally stable system with Fr = Fy = F.
(c) A rule of thumb for control of non-minimum phase systems states that
the bandwidth of the closed-loop system cannot realistically be greater
than half the value of the non-minimum phase zero. In this case 1.5
rad/s. Have we cirumvented this rule of thumb in the above design or
does the closed-loop system have any disadvantages?

7.2

A continuous-time system has a zero at s = 3 and a time-delay of 1.0 second.

What is the upper limit of the realistic bandwith/crossover frequency if the
magnitude curve of the open-loop system decreases monotonically?

7.3

Give an example of a system for which there exists no controller having all
three properties: a stable closed-loop system, small magnitude of the sen-
sitivity function at low frequencies and small amplification of measurement
errors at high frequencies.

25
7.4

A multivariable system is supposed to attenuate system disturbances (w) at

least a factor 10 for frequencies under 0.1 rad/s. Furthermore, measurement
disturbances (n) should be attenuated at least a factor 10 for frequencies
above 2 rad/s. Constant system disturbances should be attenuated at least
a factor 100 in steady state.

(a) Formulate conditions on the singular values of S and T which will

guarantee that the requirements are fulfilled.
(b) Translate the specifications into requirements on the loop gain GFy .
(c) Formulate the requirements using k · k∞ and frequency weights WS och
WT .
(d) Which crossover frequency and phase margin would we expect, having
the weights i (b), had the system been a SISO system? What lower
bound on kT k∞ does this result in?
(e) Is this lower bound on kT k∞ consistent with the requirements in (c)?

7.5

A control system has the sensitivity function S, depicted below

log |S(iω)|

A2
0
A1

What can be stated about the open-loop system if the surface A2 is larger
than the surface A1 ?

26
7.6

For a certain feedback system we demand that:

(i) output disturbances, with frequencies under 2 rad/s, should be attenu-

ated at least a factor 1000.

(ii) the system should remain stable despite a model uncertainty

|∆G| ≤ 100|G|

for frequencies above 20 rad/s. G is the frequency response of the no-

minal system and ∆G is the absolute error in the frequency response.

Can this be accomplished using a linear, time-invariant controller?

7.7

We have the following specifications on a SISO system

|S(iω)| ≤ 10−3, ω≤1

|T (iω)| ≤ 10−3, ω ≥ 100

(a) State two non-constant frequency weights WS and WT which would

guarantee that the specifications are met.

(b) Trying to find a controller fulfilling the design criteria, for example
using the methods presented in Chapter 10 in the textbook, we fail.
Should this have been anticipated from the very beginning?

27
28
8 Controller Structure and Control Design

8.1

Let !
1 10
s+2 s+1
G(s) = 1 5
.
s+5 s+3

(a) Compute RGA(G(0)).

(b) Which input-output pairing should be avoided?

8.2

Given the multivariable system

  0.6
! 
y1 1 s+1
−0.4 u1
= .
y2 0.1s + 1 0.3 0.6 u2

Assume that we want the controller to be diagonal and that we use the
relative gain array (RGA) to decide what input should control what output.
Furthermore, assume that we want a crossover frequency of ωc = 10 rad/s.
Decide how the signals should be paired.

8.3

Study the multivariable system

  1 −2
! 
y1 10s+1 2s+1 u1
= .
y2 1 s−1 u2
10s+1 2s+1

(a) Decide, using RGA analysis, which input signal should control which
output signal.

29
 (b) Assume that we want to use decentralized control, i.e. we want a con-
troller on the form
 
diag diag F11 (s) 0
F = W1 F (s)W2 , where F (s) = .
0 F22 (s)

Furthermore, assume that we do not want the steady-state error in

one channel to affect the steady-state error in the other channel. Gi-
ve the structure of a controller F (s), expressed in F diag (s), that will
accomplish this.

8.4

Design a controller, using the IMC method, for a stable first order process

K
G(s) = , τ > 0.
τs + 1
What type of controller do we get? Compute the sensitivity function and the
complementary sensitivity function and sketch the Bode plot of the sensitivity
function. What does Bode’s integral theorem state for this case?

8.5

Design a controller, using the IMC method, for the system

6 − 3s
G(s) = .
s2 + 5s + 6
What type of controller do we get?

8.6

Consider the DC motor

1
y= u
p(p + 1)

30
Compute an IMC based controller for this system. Write the controller on
the form u = −Fy (p)y, and sketch the Bode plot for Fy (p). Approximately
what type of controller do we get when we want a high bandwidth for the
closed-loop system?

8.7

Given the multivariable system

 9

1 s+1
2
G(s) = .
s/20 + 1 6 4

(a) What are the poles and zeros of G(s)?

(b) Compute an IMC based controller for the system.

8.8

Consider the system

 2 3

s+1 s+2
G(s) =  
1 1
s+1 s+1

(Example 1.1 in the textbook)

Show how an IMC based controller can be computed for this system. Give
an explicit expression for the corresponding sensitivity function.

8.9

Consider the multivariable system

Y (s) = G(s)U(s)

31
where  
2 3
 s+1 s+2 
 
G(s) =  
 α 1 
s+1 s+1
and α > 0.

(a) Determine the zero of the multivariable system. How does the zero
depend on the value of α?

(b) Assume that one would like to achive complete decouple of the system
G(s) such that
 1 
0
 (s + 1)2 
G(s)F (s) =  1 
0
(s + 1)2

Are there any cases when this is not a good idea? Motivate!

(c) Assume that one instead chooses to use a static decoupling such that
G(s)F (s) is decoupled for ω = 0. Are there any values of α for which
this is not a good idea? Motivate!

8.10

Consider the multivariable system

Y (s) = G(s)U(s)

where  
1 2
 s+2 s+4 
 
G(s) =  
 1 1 
s+1 s+2

(a) Determine the RGA at ω = 0.

32
(b) Assume that the system is going to be controlled by a diagonal regulator

U(s) = F (s)(R(s) − Y (s))

where  
K 0
F (s) =
0 K
Use the result from a) the judge how successful this will be. Determine
also the poles of the closed loop system for the case K = 5.

(c) How can the problem be modified such that a diagnonal F (s) can be
used? Verify that the closed loop system is stable for K = 5 for the
modified problem.

33
34
9 Minimization of Quadratic Criteria: LQG

9.1

Consider the system

1
G(s) =
s−1
represented on state-space form with noise as
ẋ(t) = x(t) + u(t) + v1 (t)
z(t) = x(t)
y(t) = x(t) + v2 (t)
The noises vi (t) are white with intensities Ri . We use the criterion
Z
V = Q1 x2 (t) + Q2 u2 (t) dt,

and want to find the LQG controller.

(a) Show that the controller is a function of α = Q1 /Q2 and β = R1 /R2

only.
(b) Compute the poles of the closed-loop system as a function of α and β.

9.2

Consider the system

1 1
z= u+ v
p+1 p+1
y =z+e
where v and e are unit disturbances with spectra
Φv (ω) ≡ r1 respektive Φe (ω) ≡ 1.
We minimize the criterion
Z
V = q1 z 2 (t) + u2 (t) dt

35
 (a) Compute the loop gain of the feedback connection.
(b) How do r1 and q1 influence the loop gain?
(c) Sketch the magnitude of the frequency response. What happens when
r1 → ∞ and when q1 → ∞ respectively?

9.3

Consider the double integrator

z̈(t) = u(t).
We want to find a controller such that the criterion
Z ∞
(z 2 (t) + η · u2 (t)) dt
0
is minimized for some η > 0. We assume that z(t) and ż(t) are both known
(and need not to be estimated).

Where are the poles of the optimal closed-loop system located? How is the
control signal affected when η is decreased?

9.4

Consider the antenna in Exercise 5.8. We want to control it and a suitable

measure on the performance of closed-loop system is given by the criterion

J = E Θ2 (t) + ρµ2 (t)
where ρ is a constant we can choose. Derive an optimal control signal and
discuss how it is to be combined with the Kalman filter.

9.5

Consider control of the DC motor

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

36
We want to use the motor together with a system that has a resonance
peak at approximately 0.5 rad/s. Other than that, we do not know much
about the system. Describe how we can compute an LQG controller with
good robustness qualities, i.e. small complementary sensitivity gain, at this
frequency.

9.6

A system has static gain G0 . It is influenced by system disturbances, with

all energy concentrated at zero frequency, i.e.
Φν (ω) = δ(ω)
The reference signal is zero, as is the measurement noise. We choose a con-
troller that minimizes 
E y 2 (t) + αu2 (t)
What is the value of the sensitivity function at zero frequency?

9.7

Consider the system

1 1
z= u+ ν
p+1 p+1
y =z+e
where ν is noise of very low frequency,
1
ν= v,
p+ε
v and e are noises with Φv (ω) ≡ Φe (ω) ≡ 1.

(a) Find a controller that minimizes


E z 2 + u2
when ε → 0.
What is the static gain of the sensitivity function?

37
 (b) Use output-LTR (LTR(y)) to compute L. What is the static gain of
the sensitivity function?

9.8

Consider a motor driving two rotating masses connected by a flexible shaft:

ϕ1 ϕ2

Motor

The angular displacements of the masses are ϕ1 and ϕ2 respectively and ω1

and ω2 are the angular velocities. The moments of inertia are 10 for both
masses. The spring rate of the shaft is k and the damping factor is 0.1. The
input is the voltage applied to the motor. With the states x1 = ϕ1 − ϕ2 ,
x2 = ω1 and x3 = ω2 we get the state-space representation
   
0 1 −1 0
 1 2   
ẋ = − 2 ω0 −0.01 0.01  x + ω0  u
1 2
ω
2 0
0.01 −0.01 0


z= 0 0 1 x

where
k
ω02 =
50
The Bode plot, when k = 1, is shown in the figure below. There is a resonance
peak at the frequency ω0 . The spring rate is not exactly known, but has a
value close to 1. We want to design a controller that yields a stable closed-loop
system despite variations in k.

How can the above model be extended with a model for the noise to assure
robustness for an uncertain value of k when we use LQG controller design?
Give an actual example of such an extended system.

38
 1
10

0
10

|G(iω)|

-1
10

-2
10

-3
10 -2 -1 0
10 10 10
ω rad/s

9.9

Consider the system

   
1 0 3
ẋ(t) = x(t) + u(t)
0 −2 2

Show that

u(t) = − 2 −3 x(t)
cannot be an optimal state feedback for any quadratic criterion on the form
Z
min (xT (t)Q1 x(t) + Q2 u2 (t)) dt

where Q1 is a positive definite matrix.

9.10

Consider the system

   
1 −1 −4
ẋ = x+ u
2 4 8
y = (1 1) x

39
We want to minimize the criterion
Z T
V (T ) = xT (t)x(t) + u2 (t)dt
0

Is it possible to find a state feedback u = −Lx such that V (T ) < ∞ when

T → ∞?

9.11

The figure below shows a simple electrical circuit.

VC [V]
i [A]

u [V]
R L
C

Introduce the state variables x1 = VC and x2 = i. With the component values

R = 5 Ω, L = 0.1 H, C = 1000 µF

we get the state-space representation

   
0 1000 0
ẋ(t) = x(t) + u(t)
−10 −50 10


y(t) = 1 0 x(t)

Compute a state feedback that minimizes

Z ∞

J= x22 (t) + 0.01u2(t) dt
0

This criterion aims at limiting the power loss without getting too large sig-
nals.

40
9.12

A system has the state-space representation

     
−2 −1 1 1 0
ẋ(t) =  1 0 0.5 x(t) + 0.5 u(t) + 0 v(t)
0 0 A 0 1


z(t) = 1 0 0 x(t)
y(t) = z(t) + e(t)

where e(t) are v(t) ı̈¿ 21 r unit disturbances.

The controller, a feedback from reconstructed states, minimizes

 
E z 2 (t) + u2 (t)

How does the value of A affect the sensitivity function?

9.13

A simplified model for how the elevator angle affects the movements of an
airplane is given by
   
−0.01 0.03 −10 4
ẋ =  0 −1 300  x +  −20  u
0 0 −0.5 −10

where  
roll angle
x =  yaw angle 
pitch-angle velocity
In particular we are interested in the control of the pitch-angle velocity and
choose the controlled variable to be
 
z= 0 0 1 x

All state variables are measured

y =x+e

41
We want to design a feedback from reconstructed states using LQG metho-
dology. It is especially important that the sensitivity function has a small
gain for frequencies around 1 rad/s. Show how to modify the model of the
airplane to achieve such a sensitivity function.

9.14

Consider the system

ẋ(t) = αx(t) + u(t) x(0) = x0 (1)

The system is controlled by the feedback

u(t) = −Lx(t) (2)

where L is chosen such that

Z ∞
J= x2 (t) + ρu2 (t))dt (3)
0

is minimized.

(a) Determine L as function of ρ and α.

(b) If it is desired to keep u(t) small, this can be achieved by choosing ρ

large. What is the resulting L when ρ → ∞? Consider, for example,
the cases α = 1 and α = −1, respectively. Why is it not optimal to
choose L = 0, i.e. u(t) = 0, in both cases?

9.15

An electrical motor has the transfer functions

1
Y (s) = U(s)
s(s + 1)
and it is controlled using state feedback

u(t) = −Lx(t) (r(t) = 0)

42
where x1 (t) = y(t) and x2 (t) = ẏ(t). The gain vector L is determined by
minimizing the criterion
Z ∞
J= xT (t)Q1 x(t) + Q2 u2 (t)dt
0

Figure 1 shows the simulation results when the system starts in the initial
condition x(0) = (1 1)T for some different choices of Q1 and Q2 . Combine
the figures with the choices of matrices.

(i)  
1 0
Q1 = Q2 = 0.1
0 0

(ii)  
1 0
Q1 = Q2 = 1
0 10

(iii)  
0.1 0
Q1 = Q2 = 0.1
0 0

(iv)  
1 0
Q1 = Q2 = 1
0 0

43
 A B
1.5 1.5

1
1
0.5
0.5
0
0
−0.5

−0.5 −1
0 5 10 0 5 10

C D
1.5 1.5

1 1

0.5 0.5

0 0

−0.5 −0.5
0 5 10 0 5 10

Figur 1:

9.16

Consider the simplified description of an aircraft in the figure below.

α v
θ

h
δ

Using the state space variables

44
 x1 (t) = α(t) angle of attack (rad)
x2 (t) = θ̇(t) pitch rate (rad/s)
x3 (t) = θ(t) pitch angle (rad)
x4 (t) = h(t) height (deviation from an operating point)

the input signal

u(t) = δ(t) control surface angle (rad)
and output signal
y(t) = h(t) height (hundreds of meters)
the system is decribed by the state space model
ẋ = Ax + Bu y = Cx
where
   
−0.17 1 0 0 0.011
 −0.56 0 0 0 
 , B =  0.56  , C = (0 0 0 1).
 
A=
 0 1 0 0   0 
−2.22 0 2.22 0 0

(a) Is the system asymptotically stable?

(b) Assume that the system has the initial state
x0 = (0 0 0.1 1)T
and that the system is controlled by the state feedback
u = −Lx
where the gain vector L is chosen such that the criterion
Z ∞
xT (t)Q1 x(t) + uT (t)Q2 u(t) dt
0

is minimized. Assume that the matrices are chosen as

 
0 0 0 0
 0 0 0 0 
 0 0 0 0  , Q2 = 1
Q1 =  

0 0 0 1
Determine the poles of the closed loop system. Simulate the closed loop
system.

45
 (c) Assume that Q2 is varied. How does that affect the location of the
closed loop poles and the properties of x and u?

(d) Assume now that the following conditions shall be fulfilled:

– |x1 | < 0.2 all the time.

– |x4 | < 0.1 after 25 seconds.
– |u| < 0.5 after one second.

Determine Q1 and Q2 such that these conditions are satisfied. What is

the resulting location of the closed loop poles?

9.17

The figure below illustrates a system consisting of a ball on a plane. The

variable r denotes the position of the ball relative to the center of the plane,
and α represents the angle of the plane. The input signal is the torque that
rotates the plane.

Figur 2: Ball on plane.

The system is represented by the state variables

x1 (t) - position, r(t)

x2 (t) - velocity, ṙ(t)
x3 (t) - plane angle, α(t)
x4 (t) - plane angular velocity, α̇(t)

and torque is the input signal u(t). The state space model is

ẋ(t) = Ax(t) + Bu(t)

46
where
   
0 1 0 0 0
0 0 −7 0 0

A=
0
 B=  C= 1 0 0 0
0 0 1 0
0 0 0 0 1

(a) Assume that the system starts in the initial state

x(0) = (0.1 0 − 0.1 0)T

i.e. the ball is positioned to the right of the center, and the plane leans
downwards on the right side. Assume that all state variables can be
measured. Determine a state feedback such that the following require-
ments are fulfilled:

– | x(t) |→ 0 when t → ∞.
– | x1 (t) |≤ 0.2 ∀ t.
– | u(t) |≤ 2.5 ∀ t

Determine also the absolute value of the poles of the closed loop system.

(b) Verify that all sensors that measure the states have to work in order to
obtain a stable closed loop system.
Hint: The characteristic equation of the closed loop system is given by

λ4 + l4 λ3 + l3 λ2 − 7l2 λ − 7l1 = 0

Missing a sensor is equivalent to setting the corresponding feedback li

to zero.

47
48
10 Loop Shaping

10.1

Consider the system

1
y= u
p+1
We want to create a closed-loop system with S, T and Gwu , such that
Z  
 S(iω) 2 2 2
 iω  + |0.5 T (iω)| + |5 Gwu (iω)| dω
 

is minimized. Compute the controller.

10.2

Consider the system

1
y= u
p+1
We want to create a closed-loop system with S, T and Gwu , such that
|S(iω)| < γω
|T (iω)| < 2γ
|Gwu (iω)| < 0.2γ
Write down the equations that determines the controller.

10.3

Consider the SISO system G(s) with state-space realization

ẋ = Ax + Bu
y = Cx
We want to use loop shaping with the weights
1
WS = , WT = 1, Wu = 1
s
49
 (a) State the equations that determine the optimal controller in H2 and
H∞ respectively.
(b) Explicitly write down the observer for the extended state vector and
show that the optimal controller can be written as
Z t
α
u(t) = y(τ ) dτ
1 + L(pI − A)−1 B 0
for some L, where α = 1 for the H2 controller and α > 1 for the H∞
controller. State the equation determining L.
(c) Show that the controller will have a pole at the origin unless the system
does itself has a pole at the origin.

10.4

Once again consider the system in Exercise 9.8.

(a) Suggest frequency weights WS , WT and Wu , for H2 and H∞ design,

such that we get robustness against uncertain values of k.
(b) State the extended system from u and w to z on state-space form.

10.5

A DC-motor has transfer function

1
G(s) =
s(s + 1)
and it going to be controlled using proportional feedback
U(s) = K(R(s) − Y (s))
The properties of the closed loop system is specified via the requirement
| S(iω) |<| WS−1 (iω) | ∀ω
The figure below shows three alternatives for the weight function WS−1 (iω).
Which alternative is the best? Motivate the answer.

50
 A B
1 1
10 10

0
0 10
10

−1
10
−1
10

−2 −1 0 1 −2 −1 0 1
10 10 10 10 10 10 10 10

C
1
10

0
10

−1
10
−2 −1 0 1
10 10 10 10

Figur 3: Suggestions for | WS−1 (iω) |.

10.6

The system
1
Y (s) =
U(s)
s+1
is going to be controlled by the proportional feedback

U(s) = K(R(s) − Y (s))

(a) Derive S(s), T (s) and Gru (s) respectively, i.e. the sensitivity function,
the complementary sensitivity function and the transfer function from
reference to input signal.

(b) The properties of the control system are specified using the weight
function according to

| S(iω)WS (iω) |< 1 ∀ω

| T (iω)WT (iω) |< 1 ∀ω

| Gru (iω)Wu (iω) |< 1 ∀ω

51
 The figures below show three suggestions for weight functions WS , WT
and WU . Two of the alternatives are unrealistically or incorrectly spe-
cified. Which are the two incorrect alternatives? Motivate the answer.

(c) Consider the alternative in b) that is realistically specified. Is it possible

to choose K such that all requirements are fulfilled?

1/WS 1/WT
1 0
10 10

0
10

−1 −1
10 10
−2 0 2 −1 0 1 2
10 10 10 10 10 10 10

1/WU
2
10

1
10

0
10
−2 0 2
10 10 10

Figur 4: Alternative I

52
 1/WS 1/WT
1 1
10 10

0 0
10 10

−1 −1
10 10
−2 0 2 −2 0 2
10 10 10 10 10 10

1/WU
2
10

1
10

0
10
−2 0 2
10 10 10

Figur 5: Alternative II

1/WS 1/WT
0 1
10 10

0
10

−1 −1
10 10
−1 0 1 2 −2 0 2
10 10 10 10 10 10 10

1/WU
2
10

1
10

0
10
−2 0 2
10 10 10

Figur 6: Alternative III

53
54
12 Stability of Nonlinear Systems

12.1

Given the nonlinear differential equation

ÿ + 0.2(1 + ẏ 2 )ẏ + y = 0

let the state variables be x1 = y and x2 = ẏ. Try to show that the origin is
a stable equilibrium by using the Lyapunov function candidate

1
V = (x21 + x22 ).
2

12.2

Consider the system

ẋ1 = sin x1 + x32
ẋ2 = x1 − x2

Is it possible to use the function

1 1
V (x1 , x2 ) = − x21 + x42
2 4

to prove Lyapunov stability for the above system? Motivate your answer.

12.3

A nonlinear function lies in the sector

55
 Slope 3

Slope 0.5

According to the circle criterion, what circle in the complex plane corresponds
to this nonlinearity?

12.4

A nonlinear system is described by the following block diagram

r
Σ G(s)

−1 f

where G(s) is a linear system and the static nonlinearity f is given in the
figure below (the saturations at −1 and 1 extends to −∞ and ∞).

56
 f
1

−1 −0.5 0.5 1

−1

What assumptions on G(s) must be fulfilled in order to prove that the feed-
back system is stable according to the circle criterion?

12.5

Consider the system below.

r u1 u2 1
Σ K f
s(s + 1)

−1

The nonlinearity f is such that u2 has the same sign as u1 but is otherwise
not known. For what values of K > 0 is the feedback system stable according
to the circle criterion?

57
12.6

Consider the swing depicted below.

Center of gravity

The movement of the swing is described by the equation

d2 Φ
J 2 + mgℓ sin Φ = 0
dt
where m is the mass and J is the moment of inertia. The swing can be
controlled by alternating between bending and stretching the knees while
standing on the swing. The control signal is the location of the center of
gravity ℓ. We assume that J is constant.

Show that the control signal

ℓ = ℓ0 + εΦΦ̇, ε>0

will bring the swing to rest in Φ = 0.

12.7

The block diagram below is given.

58
 Relay

r=0 1
Σ G(s)
−1

−1 Σ

b H(s)

We have that

1
H(s) = s and G(s) = .
(s + 1)(s + 2)

How shall the feedback coefficients a and b be chosen to guarantee Lyapunov

stability?
Hint: Use a quadratic Lyapunov function candidate.

12.8

A servo system contains a nonlinearity where the relationship between the

input signal u and the output signal y is

y = u + arctan(u)

What requirements on the linear part of the servo system must be fulfilled
in order to prove stability using the circle criterion?

59
12.9

A simplified model for the movements of an airplane is given by

   
−0.01 0.03 −10 4
ẋ =  0 −1 300  x +  −20  u
0 0 −0.5 −10

All states are measured and the control signal is

ũ = −Lx + r̃

where L is the feedback that minimizes

Z
(xT (t)Q1 x(t) + u2 (t)) dt

for Q1 = 10 · I

The requested control signal ũ is different from the actual u, due to the
hydraulic servo dynamics. The relationship between ũ and u is

u
2.5

0.75

1 2 ũ

Will the closed-loop system be stable? Motivate your answer.

60
13 Phase Plane Analysis

13.1

Given the differential equation

10 2
ÿ − (0.1 − ẏ )ẏ + y + y 2 = 0
3

(a) Convert the model to a nonlinear state-space model.

(b) Find the stationary points.

(c) Compute linearizations around the stationary points and analyze sta-
bility.

(d) Draw the phase portrait of this system.

13.2

Draw the phase portrait of the depicted position servo.

K
Σ
s(s + B)

−a
−1 a

The position is measured using an E-transformer, which can be described as

2
a dead zone. Assume that K > B4 .

61
13.3

The following system is given

Relay

1 1
Σ
−1 s2

−1

(a) With zero input signal the output of the relay is +1 or −1, depending
on the history of the input signal. The relay does not switch until the
input signal has changed polarity.
Draw a phase portrait of the system.

(b) Due to imperfections the actual feedback loop is

−a −1
a

Draw the phase portrait of this system.

13.4

Linus is on his way home after an exam. On the highway outside of Linkı̈¿ 12 ping
a gust of wind makes the car drift from the desired path. Your task is to,

62
using phase plane analysis, decide how the movement of the car will pro-
gress. Will it return to the desired path? If the car has a constant speed in
the direction of travel the system can be described by the following block
diagram

u 1 −1 1
y
Σ s 1 s

−G(s)

The torque applied to the steering wheel is u. The backlash comes from a gear
unit in the steering. The output signal y is the deviation from the desired
path. G(s) is the transfer function from Linus’ visual perception to the torque
he applies to the steering wheel.

Distinguish between the cases:

(a) G(s) = 1 (there was a party after the exam)

(b) G(s) = 1 + s (there was not a party after the exam)

13.5

A simple ecological system consists of two species of fish. The first kind eats
algae and the second kind eats the first kind. Let x1 denote the number of
algae eating fish and x2 denote the number of predatory fish. Then we have

x1 x2
ẋ1 = 2x1 − − 0.2x21
1 + 61 x1
x1 x2
ẋ2 = −3x2 +
1 + 61 x1

63
 (a) From these equations, calculate the stationary points.

(b) Classify the stationary points and sketch the phase portraits in a sur-
rounding of them. It is sufficient to consider a linearised version of the
equations.

(c) Without any further calculations, merge the phase portraits you have
made around the stationary points in a fashion that seems reasonable.
Only consider x1 > x2 > 0.

An interpretation of the given equations is:

If the algae eating fish have an infinite amount of food and lack enemies,
their number will grow exponentially as

ẋ1 = 2x1

As there is a limited amount of algae the growth saturates according to

ẋ1 = 2x1 − 0.2x21 .

If there are predatory fish x2 present the algae eaters will be devoured at the
rate
x1 x2
1 + 61 x1
The interpretation of this term is that if x1 is large every predatory fish can
eat until it is full. This corresponds to 6 algae eating fish per time unit. On
the other hand, if the number x1 is relatively small the predatory fish will
eat less.

The second equation says that if the supply of food is unlimited (x1 = ∞)
the predatory fish will multiply according to

ẋ2 = 3x2 .

If food is lacking (x1 = 0) the predatory fish will expire as

ẋ2 = −3x2

64
13.6

A mass is suspended from a spring. Its position y(t) satisfies the differential
equation
ÿ(t) + y(t) = f (t)
where f (t) is an external force acting on the mass. Draw a phase portrait of
the system when 
−1 if ẏ(t) > 0
f (t) =
+1 if ẏ(t) < 0
Will the system reach an equilibrium?

13.7

Consider the system  

−x31 + u
ẋ =
x1

(a) Sketch a phase portrait when u = 0.

(b) Use the Lyapunov function

V (x) = x21 + x22

to compute a control signal

u = f (x1 , x2 )

which will make the origin globally asymptotically stable. Sketch a

phase portrait, in a neighborhood of the origin, for the closed-loop
system.

65
66
14 Oscillations and Describing Functions

14.1

Consider the feedback control system including an input saturation according

to the figure below.

1 u 10
Σ
1 s(s + 1)2

−1

(a) Investigate the stability of the system. If a periodical solution exists,

determine its frequency and amplitude.

(b) Build a simulation model of the control system and investigate the
validity of the results from a).

14.2

A temperature control system, depicted below, contains a relay with dead

zone.

r(t) = 0 u(t) y(t)

Σ G0 (s)

−1

67
 1
G0 (s) = s(1+s) 2 , ±D is the width of the dead zone and ±H is the output

level of the relay. The values of the dead zone and output level are such that
a stable oscillation just barely can exist. If H is increased or if D is decreased
an oscillation will not be possible.The amplitude of the oscillation is 2.5 units.
Compute D, H and the frequency of the oscillation. The describing function
for a relay with dead zone is

4H p
Re{YN (C)} = 1 − D 2 /C 2 , C≥D
πC
Im{YN (C)} ≡ 0

14.3

A relay servo is given by

θref u K θ
Σ
s(s + 1)2

−L(s)

The gain K is strictly positive.

(a) The feedback used is L(s) = 1. Show that there is an oscillation for all
values of K.

(b) To avoid too much wear on the system we do not want the amplitude
of the oscillation in θ to be greater than 0.1. For what values of K is
this fulfilled?

(c) We want to use a gain K that is larger than what is possible in (b). State
a feedback L(s) with L(0) = 1 that makes this feasible. No details are
necessary. Just motivate why the feedback should solve the problem.

68
14.4

Consider the nonlinear system

u 1 y
Σ
s(s + 1)(s + 2)

−H(s)

(a) If proportional control is used, i.e. H(s) = 1, a stable oscillation occurs.

Find the amplitude and frequency of the oscillation.
(b) To eliminate the oscillation we use proportional and derivative con-
trol, i.e. H(s) = 1 + Ks. Show how K can be chosen to eliminate the
oscillation.

14.5

Consider the feedback control system where a motor is controlled using a

relay with hysteresis.

θr = 0 1
u(t) 1 θ
Σ
−0.5
−1
0.5
s(s + 1)

−1

(a) Investigate the stability of the system using the describing function
method. If a periodical solution exists, determine its frequency and
amplitude.

69
 (b) Build a simulation model of the control system and investigate the
validity of the results from a).
(c) Introduce suitable state variables and sketch a phase portrait.

14.6

Consider the following servo system

r=0 e ũ u 1 y
Σ K(1 + 1
TI s
+ TD s)
s2
PID controller amplifier motor

−1

The PID controller has K = 2, TI = 2 and TD = 0.5.

(a) The tuning of the controller was done assuming that the amplifier has
the transfer function 1. Show that, if this assumption is true, this results
in an asymptotically stable closed-loop system.

(b) The actual amplifier contains a saturation

−1 1 ũ

−1

70
 State the amplitude, frequency and stability properties of possible oscil-
lations.

(c) Discuss, based on the results from (b), under what circumstances the
servo system will function as intended. Especially investigate the influ-
ence of different signal amplitudes.

14.7

Consider the feedback control system

r(t)=0 u y(t)
Σ G0(s)

-1

When the system is simulated a limit cycle occurs. Determine the amplitude
and frequency of the limit cycle. The Bode diagram for the linear part GO (s)
of the control system is given in the figure below.

G
|G| 20log|G| dB

20

10 20
90
5
|G0|
2
.1 2 5 1 2 5 1
1 0 0
Radianer/s
0.5 arg G0

0.2
0.1 -20 -90

0.05

0.02
-180
0.01 -40

71
The describing function of a relay with deadzone is given by
4 p
Yf (C) = 1 − 1/C 2 C≥1
πC

72
17 To Compensate Exactly for Nonlinearities

17.1

Find a feedback which makes the system

ẋ1 = −x1 + 7x2

ẋ2 = −x2 + cos x1 + u

linear.

17.2

Find an output feedback, u = f (y), which makes the system

ẋ1 = x3 + 8x2
ẋ2 = −x2 + x3
ẋ3 = −x3 + x41 − x21 + u
y = x1

linear.

17.3

Find a feedback which makes the system

ẋ1 = x21 + x2
ẋ2 = u
y = x1

linear.

73
17.4

Consider the two tank system:

x1

x2

The dynamics of the system is described by

√
ẋ1 = 1 + u − 1 + x1
√ √
ẋ2 = 1 + x1 − 1 + x2

Which state should be chosen as output to achieve a strong relative degree

2? Do a feedback linearization of the system.

17.5

A mass m is suspended from a spring:

m
F

74
The force F is generated by the control signal u fed through an actuator such
that
1
F = u
s+1
The position of the mass is y. The spring rate and the viscous damping are
nonlinear. Thus the force is

−k(y) − d(ẏ).

(a) Realize this system on state-space form. The input signal is u and the
output signal is y.

(b) Can the system in (a) be made linear using feedback? If so, compute
such a feedback.

75
76
Solutions

77
78
1 Introduction

1.1

The small gain theorem for linear systems can be stated as follows: Assume
that both G(s) and F (s) are stable transfer functions, and interconnected
according to the figure below.

Σ G(s)

−F (s)

Then the closed-loop system is stable if

|G(iω)| · |F (iω)| < 1, ∀ω.

The transfer function of the closed-loop system is

G(s)
Gc (s) =
1 + G(s)F (s)

Gc (s) is stable according to the Nyquist criterion if the Nyquist curve for
G(iω)F (iω) does not encircle the point −1. Since we know from the small
gain theorem that

|G(iω)F (iω)| ≤ |G(iω)| · |F (iω)| < 1,

the Nyquist curve can not encircle the point −1, and hence the Nyquist
criterion is fulfilled.

Note that input-output stability follows from asymptotic stability. Input-

output stability is the concept used in the general small gain theorem.

79
1.2

We have that y(t) = f (u(t)) where f (·) is the function describing the ideal
relay. The gain is defined as

kyk2
kf k = sup
u6=0 kuk2

We have that |f (u)| ≡ 1, ∀u(t) 6= 0, and this yields

Z ∞ Z T
kyk22 = 2
y (t)dt = lim [f (u(t))]2 dt = lim 2T = ∞
−∞ T →∞ −T T →∞

for all choices of u(t) 6= 0 such that 0 < kuk2 < ∞. Take for example
u(t) = 1t . This means that an ideal relay has infinite gain.

1.3
S
z }|1 {
ũ
Σ f (u) G(s)

−K
| {z }
S2

The system is stable according to the small gain theorem if kS1 k · kS2 k < 1.

kS1 k ≤ kf (u)k · kGk
We have that:
kS2 k = |K|

where
2 1
kGk = sup |G(iω)| = sup p = 1 (fı̈¿ r ω = 0)
ω ω (2 − ω 2)2 + 4ω 2 2

80
 R∞  
2 kũk22 −∞
(f (u(t))2 dt 1
kf (u)k = = ≤ |f (u(t))| ≤ |u(t)|
kuk22 kuk22 2
1 2
kuk 1 1
≤ 4 22 = ⇒ kf (u)k ≤
kuk2 4 2

1
kS1 k · kS2 k ≤ · |K| < 1
2
i.e. , we must choose |K| < 2 to be able to guarantee input-output stability.

1.4

(a) kyk∞ = |a|, kyk2 = ∞

(b) kyk∞ = 1, kyk2 = 1

1 1√
(c) kyk∞ = , kyk2 = 3
4 6

1.5

The gain of the system is

ω02
kGk = sup |G(iω)| = sup p
ω ω (ω02 − ω 2 )2 + 4ζ 2 ω02ω 2

By differentiating |G(iω)| we see that the magnitude of G(iω) has its maxi-
mum at ω = 0 if ζ > √12 . This results in the gain

kGk = 1
p
If 0 < ζ < √12 the maximum of |G(iω)| is attained at ω = ω0 1 − 2ζ 2. This
results in the gain
1
kGk = p
2ζ 1 − ζ 2

81
1.6

One has to distinguish between the cases a > 0 and a < 0 respectively.

(i) For a > 0 the system G(s) is stable and the small gain theorem is
applicable. The system G(s) has gain one, and the small gain theorem
hence gives the condition | K |< 1. The characteristic equation of the
closed loop system is given by (Note: positive feedback)
(s + a) − Ka = 0
which imples the pole s = (K − 1)a, which is located in the left half
plane for K < 1.
(ii) For a < 0 the system G(s) is not stable and the small gain theorem
is not applicable. The pole s = (K − 1)a is in the left half plane for
K > 1.

1.7

The linear part, represented by G(s), has gain 1.5 according to the figure.
For the nonlinear part we assume that f (x) is an odd function, such that
f (−x) = −f (x). The nonlinearity can hence be bounded by
| f (x) |≤ 0.5 | x |
and hence the gain is 0.5. Since 1.5 · 0.5 < 1 the closed loop system is stable
according to the small gain theorem.

1.8

Using proportional control u = −Ky = −Kx1 we get

ẋ1 = x2
ẋ2 = −aKx1 − ax2
The characteristic equation is s2 + as + aK = s2 + (1 + ρ)s + (1 + ρ)K = 0
The closed-loop system is stable if (1 + ρ) > 0 and (1 + ρ)K > 0, i.e. it is
stable for all K > 0 when δ < 1.

82
1.9

(a) With a = 1 + ρ it holds that

ẋ1 = x2
ẋ2 = −a(Kx1 + x2 ) = −Kx1 − x2 − ρ(Kx1 + x2 )
| {z }
w=ρz

The open system with input signal w and output signal z is given by
   
0 1 0
ẋ = x+ w
−K −1 −1


z = K 1 x,

that is
s+K
Gwz (s) = − .
s2 +s+K
(b) The relationship
w(t) = ρ(t)z(t)
gives Z ∞
kw(t)k22 = ρ2 (t)z 2 (t)dt < δ 2 kz(t)k22
∞

and
kw(t)k2
<δ
kz(t)k2
which implies that the gain is at most δ.

(c) The closed-loop system is depicted in the figure below.

w z
Gwz (s)

ρ(t)

A sufficient condition for stability, according to the low gain theorem,

is
kGwz k · kρk < 1 ⇔ kGwz k2 < 1/δ 2

83
The magnitude, of the linear systems frequency response, squared is

ω2 + K 2
|Gwz (iω)|2 =
(K − ω 2 )2 + ω 2

What is the maximum of this function?

√
x + K2 K 4 + 2K 3
sup 2
= √
x (K − x) + x 2K 4 + 4K 3 − (2K + 2K 2 − 1) K 4 + 2K 3
√
for x = −K 2 + K 4 + 2K 3 . Thus, an implicit condition on K to gua-
rantee stability of the closed-loop system is
√
K 4 + 2K 3
√ < 1/δ 2
4 3 2
2K + 4K − (2K + 2K − 1) K + 2K 4 3

84
2 Representation of Linear Systems

2.1

Putting x = ω, u1 = M, u2 = Im , u3 = R, y1 = ω and y2 = e give the state

equation
x2 u2
ẋ = u1 −
u3
and
y1 = x
y2 = u2 x
The input vector u1,0 = u2,0 = u2,0 = 1 and the state x0 is a stationary point,
giving the stationary output y1,0 = y2,0 = 1. Introducing

∆x = x − x0 ∆u = u − u0 ∆y = y − y0

and linarizing gives

d

∆x = −2∆x + 1 −1 1 ∆u
dt
   
1 0 0 0
∆y = ∆x + ∆u
1 0 1 0
Computing the transfer function and renaming the variables give
 
    ∆M
∆ω 1 1 −1 1 
= ∆Im 
∆e s+2 1 s+1 1
∆R

85
2.2

(a)

y = αh2 , f = β(h1 − h2 )
1 1
ḣ1 = (u1 − f ), ḣ2 = (u2 + f − y)
A1 A2

 1 1  1 
− A1 β A1
β 0
ḣ = 1 1 h + A1 1 u
A2
β − A2
(β + α) 0 A2


y= 0 α h

1

G(s) = αβ α(s + β)
s2 + (2β + α)s + αβ

(b) The result above gives

G(0) = 1 1

Singular values in ω = 0 (i.e., for constant input signals) are given by

the square roots of the largest and smallest eigenvalues of the matrix
   
T 1
1 1
G(0) G(0) = 1 1 =
1 1 1

Solving det(λI − G(0)T G(0)) = 0 yields eigenvalues 0 and 2, which

implies √
σ̄(G(0)) = 2 σ(G(0)) = 0
The maximum singluar value corresponds the input signal vector
 
1
umax =
1

which gives the steady state output signal y = 2, while the minumim
singular value corresponds the input signal vector
 
1
umin =
−1

which gives the steady state output signal y = 0.

86
2.3

Find the common denominator of the system

(s2 + s + 1) (s + 2)(s + 3)
Y (s) = U1 (s) + U2 (s)
(s + 1)(s + 2)(s2 + s + 1) (s + 1)(s + 2)(s2 + s + 1)
(s2 + s + 1) (s2 + 5s + 6)
= 4 U1 (s) + 4 U2 (s)
(s + 4s3 + 6s2 + 5s + 2) (s + 4s3 + 6s2 + 5s + 2)
It is now straightforward to realize the system on observer canonical form
   
−4 1 0 0 0 0
 −6 0 1 0   x(t) +  1 1  u(t)
 
ẋ(t) = 
 −5 0 0 1   1 5 
−2 0 0 0 1 6
y(t) = (1 0 0 0) x(t)

2.4

Find the common denominator

 
1
u 1 (t)
y(t) = 3 p2 + 3p p2 + p − 2
p + 7p2 + 16p + 12 u2 (t)

Observer canonical form yields

   
−7 1 0 1 1  
u 1 (t)
ẋ(t) = −16 0 1 x(t) + 3 1 
u2 (t)
−12 0 0 0 −2


y(t) = 1 0 0 x(t)

2.5

Laplace transformation yields

(b11 s + b12 ) (b21 s + b22 )
Y (s) = 2
U1 (s) + 2 U2 (s)
(s + a1 s + a2 ) (s + a1 s + a2 )

87
The system on observer canonical form is
   
−a1 1 b11 b21
ẋ(t) = x(t) + u(t)
−a2 0 b12 b22


y(t) = 1 0 x(t)

2.6

Laplace transformation yields

A(s)Y (s) = B(s)U(s)

where    
s 1 s+2
A(s) = B(s) =
1 (s + 1) 1
Multiplication by A−1 (s) results in

Y (s) = A−1 (s)B(s)U(s)

i.e.    
s2 +3s+1 2s+2
s2 +s−1 s2 +s−1
+ 1
Y (s) = −2 U(s) = −2 U(s)
s2 +s−1 s2 +s−1

The system on controller canonical form

   
−1 1 1
ẋ(t) = x(t) + u(t)
1 0 0
   
2 2 1
y(t) = x(t) + u(t)
0 −2 0

88
3 Properties of Linear Systems

3.1

The transfer function matrix has the minors

1 (s + 1) 1
− 2
− 2
=−
(s + 2) (s + 2) s+2

when the first column is deleted,

1 1
2
− =0
(s + 2) (s + 2)2

when the second column is deleted and

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

when the third column is deleted. In addition, the elements of the transfer
function are themselves minors. The pole polynomial, i.e. the least common
denominator to all minors is thus

p(s) = (s + 2)

The system has a pole in s = −2. Hence, the system can be realized as a
state-space sytem of order one.

The maximal minors are

1 1
− , 0,
s+2 (s + 2)

Thus, the zero polynomial is a constant. The system lacks zeros.

89
3.2

The transfer function matrix has the determinant

2
det G(s) =
(s + 3)2

and the minors

1 −1 2(s + 1)
, ,
(s + 1)(s + 3) (s + 1)(s + 3) (s + 3)

The pole polynomial, i.e. the least common denominator of the minors, is

p(s) = (s + 1)(s + 3)2 ,

Hence the poles are −1, −3 and −3. The maximal minor is
2
(s + 3)2

If we normalize with the pole polynomial we get

2(s + 1)
.
(s + 1)(s + 3)2

The zero polynomial is thus

n(s) = (s + 1)

There is a zero in −1.

3.3

Minors:
1−s 2−s 1/3 − s 1/3
, , ,
(s + 1)2 (s + 1)2 (s + 1)2 (s + 1)3
| {z }
2 st

There are 3 poles in s = −1 as the least common denominator is (s + 1)3 .

Thus a minimal realization must be of order three.

90
3.4

(a) The determinant of the transfer function matrix is

(s + 5) 1 6(s + 3)
− =
(s2 + 3s + 2)(s + 2) (s + 2)(s + 4) (s + 1)(s + 2)(s + 2)(s + 4)
and the minors are
(s + 5) 1 1 1
, , ,
s2 + 3s + 2 (s + 2) (s + 4) (s + 2)
Thus, the pole polynomial is
p(s) = (s + 1)(s + 2)(s + 2)(s + 4)
which means that the poles are located at −1, −2, −2 and −4. We need
four states to realize the system.
(b) The determinant of the transfer function matrix is
(s + 5) 1 4
2
− =
(s + 4)(s + 3s + 2) (s + 2)(s + 4) (s + 1)(s + 2)(s + 4)
The pole polynomial is
p(s) = (s + 1)(s + 2)(s + 4)
which means that the poles are located at −1, −2 and −4. We need
three states to realize the system.

3.5

The system can be written on the form

A(p)y(t) = B(p)u(t)
where    
(p + 1) −p 1 −1
A(p) = B(p) =
p (p + 1) 1 1
Multiplication with A−1 (p) yields
 
−1 1 (2p + 1) −1
G(p) = A (p)B(p) = 2
2p + 2p + 1 1 (2p + 1)

91
The transfer function matrix, G(s), has the determinant
(2s + 1)2 1 2
det G(s) = 2 2
+ 2 2
= 2
(2s + 2s + 1) (2s + 2s + 1) (2s + 2s + 1)
and the minors
(2s + 1) −1 1
(2s2 + 2s + 1) (2s2 + 2s + 1) (2s2 + 2s + 1)
This results in the pole polynomial
p(s) = 2s2 + 2s + 1
Hence, the poles are located at − 12 ± i 12 .

The maximal minor is

2
(2s2
+ 2s + 1)
Thus, there are no zeros of the system.

3.6

Alt. 1: The output signal y only depends on x1 and x2 . The states x1 , x2 do

not depend on x3 due to the structure of the matrix A. Hence, the state
x3 is unobservable and can be eliminated from the state-space form:
     
˙ −2 1 0 1 1 0
x̃(t) = x̃(t) + u(t), y(t) = x̃(t),
0 −1 1 0 0 1
where x̃ = (x1 , x2 )T . The controllability matrix and the observability
matrix both have full rank and hence the realization is minimal.
Alt. 2: The transfer function matrix can be computed as G(s) = C(sI −
A)−1 B  1 1 
(s+2)(s+1) s+2
G(s) = 1 ,
s+1
0
i.e.
1 1
y1 = u1 + u2
(s + 2)(s + 1) s+2
1
y2 = u1
s+1
This results in the block diagram:

92
 y2

1 1
u1 s+1 Σ s+2 y1

u2

Introduce a state after each block, for example x1 = y1 and x2 = y2 .

This results in the same minimal realization as in Alt.1.

3.7

(a) The singular values at ω = 2 can be determined in two ways, and for
both alternatives we start by entering the transfer function matrix in
Matlab.

>> s=tf(’s’);
>> G=[1/(s+1) 3/(s+2); 2/(s+3) 1/(s+4)];

Alternative (i): The frequency function G(iω) at the angular frequency

ω = 2 can be computed according to

>> G2 = freqresp(G,2)

G2 =

0.2000 - 0.4000i 0.7500 - 0.7500i

0.4615 - 0.3077i 0.2000 - 0.1000i

The eigenvalues and eigenvector of G(iω)∗G(iω) are now obtained from

>> [V,D]=eig(G2’*G2)

V =

0.8288 + 0.2392i 0.4860 + 0.1403i

-0.5059 0.8626

93
 D =

0.1579 0
0 1.5248
√
The smallest singular value
√ is hence σ(G(i2)) = 0.1579 ≈ 0.40 and
the largest is σ̄(G(i2)) = 1.5248 ≈ 1.24.
Alternative (ii): The singular values can be detemined graphically using
the command

>> sigma(G)

which gives the result

Singular Values

1
10

0
10
Singular Values (abs)

−1
10

−2
10
−2 −1 0 1 2
10 10 10 10 10
Frequency (rad/sec)

(b) The second column of the matrix V defines the Fourier transform of the
input vector that corresponds to the largest gain of the system, i.e. the
input vector is such that the input components fulfill
√
| U1 (iω) |= 0.4862 + 0.14032 ≈ 0.51
arg U1 (iω) = arctan(0.1403/0.486) ≈ 0.28 rad
and
| U2 (iω) |≈ 0.86 arg U2 (iω) = 0

94
(c) Using the hints an input vector can be generated using the command
sequence

>> t=(0:0.01:50).’;
>> u12=0.51*sin(2*t+0.28);
>> u22=0.86*sin(2*t);

The system can then be simulated using

>> y2=lsim(G,[u12 u22],t);

>> plot(t,y2)

This gives the result below showing that the output components have
amplitudes 1.14 and 0.48.

1.5

0.5

−0.5

−1

−1.5
0 5 10 15 20 25 30 35 40 45 50

(d) Using the fact that the Fourier transform at the studied frequency
are proportional to the signal amplitude the ratio of the norms of the
output and input vectors becomes
√
1.142 + 0.482
√ ≈ 1.24
0.512 + 0.862

which corresponds to the largest singular value.

95
3.8

The minors of order 1 are

1 s−1 −1 1
, , ,
s + 1 (s + 1)(s + 2) s − 1 s + 2

The minors of order 2 are

−(s − 1) 2 1
2
, , .
(s + 1)(s + 2) (s + 1)(s + 2) (s + 1)(s + 2)

The least common denominator yields the pole polynomial

p(s) = (s + 1)(s + 2)2 (s − 1),

and the poles are therefore −1, −2, −2, 1. The maximal minors, normalized
with the pole polynomial, are then given by

−(s − 1)2 2(s − 1)(s + 2) (s − 1)(s + 2)

, , ,
(s + 1)(s + 2)2 (s − 1) (s + 1)(s + 2)2 (s − 1) (s + 1)(s + 2)2 (s − 1)

and the gcd of the numerators is thus z(s) = s − 1 and the only zero is 1.

96
5 Disturbance Models

5.1

Φu (ω) is an even function. Do the decomposition Φu (ω) = G(iω)G(−iω)Φe (ω)

where G(s) has all poles and zeros in the left-half plane and Φe = 1.

(a)
a2 a a
Φu (ω) = 2 2
Φe (ω) = ·
ω +a iω + |a| −iω + |a|
Thus the linear filter is
a
G(s) = , a 6= 0.
s + |a|

(b) Analogously we get

a2 b2
Φu (ω) = Φe (ω)
(ω 2 + a2 )(ω 2 + b2 )
ab ab
= ·
(iω + |a|)(iω + |b|) (−iω + |a|)(−iω + |b|)

ab
⇒ G(s) =
(s + |a|)(s + |b|)

5.2

Consider the disturbance model

N(s) = H(s)V (s)

where V denotes white noise. In (i) the transfer function is of low pass charac-
ter, which means that N will be of low frequency character. The disturbance
is located around 5 Hz, i.e. 10π rad/s. The magnitude curve of model (ii) has
a peak around this angular frequency, which means that this model is the
most appropriate one. In model (iii) the peak is located around 5 rad/s.

97
5.3

(a) We are given

f = k1 ż + v

The force is mz̈ = u − f , where m is the mass of the missile and u is

the thrust.
On input-output form:

k1 1
z̈ + ż = (u − v)
m m

State-space form: Let x1 = z, x2 = ż ⇒ ẋ1 = x2 ,

1 1
ẋ2 = (u − f ) = (u − k1 x2 − v)
m m

That is      
0 1 0 0
ẋ = x+ u+ v
0 − km1 1
m
− m1


z= 1 0 x

(b) Description of v:
Φv (ω) = |H(iω)|2Φe (ω)
√
k0
√
Thus H(s) = s+|a| , i.e. v̇+|a|v = k0 e. Introduce an extra state x3 = v
which results in a new state-space form with ẋ3 = −|a|x3 + e:
     
0 1 0 0 0
ẋ = 0 − km1 − m1  x +  m1  u + √0  e
0 0 −|a| 0 k0


z= 1 0 0 x

The input-output form is

 √ 
k1 p 1 k0
(p2 + )z = u− e
m m p + |a|

98
5.4

(a) With {A, B, C, N} according to exercise 5.3 we get

ẋ = Ax + Bu + Ne
y = Cx + n

where n has spectral density Φn ≡ 0.1.

(b) A noise signal with the desired spectral density can be generated by a
s
system with transfer function Gn (s) = s+|b| . The input is white noise
with spectral density Φwn = 0.1. On state-space form we get

ẋ4 = −|b|x4 + |b|wn

n = −x4 + wn

The extended state-space form is

      
A 0 B N 0 e
ẋ = x+ u+
0 −|b| 0 0 |b| wn


y = C −1 x + wn

(c) Following the same procedure as in (b) we get a transfer function

1
Gn (s) = s+|b| . The input is white noise with spectral density Φwn = 0.1.
On state-space form we get

ẋ4 + |b|x4 = wn .

The extended state-space form is

      
A 0 B N 0 e
ẋ = x+ u+
0 −|b| 0 0 1 wn


y= C 1 x

5.5

(1) A model for w: A stepwise change results from w = 1s v where v is a

number of impulses.
Introduce the state xw , ẋw = v.

99
 (2) A model for n: Use a second order system with a resonance peak at
ω0 = 2π · 2 = 4π rad/s and damping ξ = 0.01
ω02
n= 2 e
p + 2ξω0 p + ω02
Introduce the states xn1 = n and xn2 = ṅ
     
ẋn1 0 1 0
ẋn = = x + e
ẋn2 −ω02 −2ξω0 n 1
| {z } |{z}
An Bn

We get the extended model

       
ẋ A N 0 B 0 0  
v
ẋu = ẋw  =  0 0 0  xu +  0  u + 1 0 
e
ẋn 0 0 An 0 0 Bn

5.6

(a) Choose the states x1 = acceleration and x2 = speed. This results in the
state-space form    
0 0 1
ẋ = x+ e
1 0 0
   
1 0 v1
y= x+
0 1 v2
(b)      
0 0 k11 k12 k11 k12
x̂˙ = − x̂ + y
1 0 k21 k22 k21 k22
The matrix K is determined from the algebraic Riccati equation.

5.7

We get the state-space form

      
ẋ1 0 1 x1 0
= + v
ẋ2 −1 0 x2 1 1
| {z } |{z}
=A =N

100
where x1 = x, x2 = ẋ.

The Kalman filter:

   
0 1 k1
x̂˙ = x̂ + (y − C x̂)
−1 0 k2
where C = (1 0) (Case I) or C = (0 1) (Case II). The noise intensity is
 
R1 R12
R=
R21 R2
 "" # # " #

 v(t) h i 1 0

 E v(t) e1 (t) = in case I, and


 e1 (t) 0 1
=

 "" # # " #

 v(t) h i 1 0
E v(t) e2 (t) = in case II.


e2 (t) 0 1

The Kalman gain is  

k1
K= = P C T R2−1 ,
k2
where P = E[x̃(t)x̃T (t)] is given by the ARE
AP + P AT − P C T R2−1 CP + NR1 N T = 0.
Case I:  
0.910 0.414
P =
0.414 1.287
Case II:  
1 0
P =
0 1
The position x1 is estimated more accurately in case I and the speed x2 is
estimated more accurately in case II.

5.8

Introduce the states  

Θ(t)
x(t) =
Θ̇(t)

101
and denote α = B/J, H = k/J and γ = 1/J. A state space model of the
system is
     
0 1 0 0
ẋ(t) = x(t) + µ(t) + τ (t)
0 −α H γ d
| {z } | {z } | {z }
A B N


y(t) = 1 0 x(t) + em (t)
| {z }
C

The Riccati equation used used to compute the Kalman gain K is

AP + P AT + NR1 N T − (P C T + NR12 )R2−1 (P C T + NR12 ) = 0.
Using R12 = 0 (process- and measurement noise independent), A, N and C
as above, R1 = vd and R2 = vm gives
       
0 1 0 0 0
1 −1

P +P + vd 0 γ − P vm 1 0 P.
0 −α 1 −α γ 0
With  
p11 p12
P = ,
p12 p22
the components of this matrix equation become
p211
2p12 − = 0
vm
p11 p12
p22 − αp12 − = 0
vm
p2
−2αp22 + γ 2 vd − 12 = 0
vm
If we eliminate p12 and p22 we get
p411 αp311 α2 p211
3
+ 2
+ − γ 2 vd = 0
4vm vm vm
Now introduce
p11 = vm · p′11
which yields
4 3 2 vd
p′ 11 + 4αp′ 11 + 4α2p′ 11 − 4γ 2 =0
vm
2 vd
(p′ 11 + 2αp′11 )2 − 4γ 2 =0
vm
r
′2 ′ vd
p 11 + 2αp11 − 2γ =0
vm

102
 q
vd
Define β = γ vm
This results in
p
p′ 11 = −α + α2 + 2β
The solution is
 p p 
−α + α2 p + 2β α2 + β − α α2 + 2β p
P = vm
α2 + β − α α2 + 2β −α3 − 2αβ + (α2 + β) α2 + 2β
The steady state Kalman gain K = P C T R2−1 becomes
 p 
−α + α2 p + 2β
K=
α2 + β − α α2 + 2β
and using the numerical values given we get
 
40.36
K=
814.3
The covariance matrix for the estimation error is
 
40.36 · 10−7 814.3 · 10−7
P =
814.3 · 10−7 366.1 · 10−5
Hence the filter for estimating Θ is
   
˙x̂ = 0 1 0


x̂ + µ(t) + K y − 1 0 x̂
0 −α H
with K as above.

5.9

(i)

1
p

w
u 1
Σ y
p(p + 1)

103
(ii)

1
p

w
u 1
Σ y
p(p + 1)

v(t) unit disturbance

(a) (i)
A B
z
 }| { z}|{  
0 1 0 0 0
ẋ = 0 −1 1 x + 1 u + 0 v
    
0 0 0 0 1


y = 1 0 0 x.
| {z }
C

(ii)

A B
z
 }| { z}|{  
0 1 0 0 0
ẋ = 0 −1 0 x + 1 u + 0 v
0 0 0 0 1


y = 1 0 1 x.
| {z }
C

(b) (i) Offset in the motor voltage, step disturbance in the load
(ii) Measurement disturbance – error in the sensor for angular displa-
cement

104
 (c) (i)
 

0 1 −1
S = B AB A2 B = 1 −1 1  not full rank
0 0 0

(ii)
 
0 1 −1
S = 1 −1 1  not full rank
0 0 0

In (i) we can make x3 unobservable by chosing u = −Lx with ℓ3 = 1.

This is not possible in (ii).

5.10

(a) The spectrum of the wind has low pass characteristics with bandwidth
α. When α increases v(t) behaves more and more like white noise, i.e.
the gustiness increases. This can also be seen by studying the covariance
function
Z ∞
1
Rv (τ ) = Φv (ω)eiωτ dω = e−α|τ | , α > 0.
2π −∞

The covariance function gets more narrow when α increases, i.e. the
correlation with neighboring values of v(t) decreases and the gustiness
increases.

(b) Using spectral factorization, the influence from the wind can be descri-
bed as white noise e(t) with intensity 1 filtered through a linear system
with transfer function
p
2/α
H(s) =
1 + s/α

. We get y = G(s) H(s)e where

√ √
K 2α K 2α
G(s) H(s) = = 3 .
(α + s)(s2 + s + 1) s + (1 + α)s2 + (1 + α)s + α

105
The variance of the output signal is
Z ∞
1
Var(y) = |G(iω)H(iω)|2dω
2π −∞
Z ∞  √ 2

1  K 2α 
=   dω
2π −∞  (iω)3 + (1 + α)(iω)2 + (1 + α)iω + α 
K 2 (1 + α)
= .
1 + α + α2

K 2 (1+α)
Thus the requirement can be formulated as 1+α+α2
> 1.

106
6 The Closed-Loop System

6.1

Consider the block diagram

✓✏
w✲
u u✲
Σ G
✒✑
✻

y ✓✏
❄ w
−F ✛ Σ
✒✑

We have the relationships

y = (I + GF )−1 (w + Gwu ) = Gwy w + Gwu y wu
and
u = (I + F G)−1 (wu − F w) = Gwu u wu + Gwu w
which results in the input-output model
    
u Gwu u Gwu wu
= .
y Gwu y Gwy w
We also have
wu = u + F y
and
w = y − Gu
which results in the transfer function matrix
    
wu I F u
=
w −G I y

Thus, we have shown that

 −1  
Gwu u Gwu I F
=
Gwu y Gwy −G I

Alternative solution: Show that the matrix product is the identity matrix.

107
6.2

r Σ G Σ y

Σ n

−F

With the transfer functions

s−1 s+2
G= , F =
s+2 s−1

we get

Y = G(R − F (Y + N)) + W ⇒ (1 + GF )Y = GR − GF N + W

⇒ Y = (1 + GF )−1GR − (1 + GF )−1 GF N + (1 + GF )−1 W

The closed-loop system, the sensitivity function and the comlementary sen-
sitivity function are

s−1
Gc = Gry = (1 + GF )−1 G =
2s + 3
s+1
S = Gwy = (1 + GF )−1 =
2s + 3
s+2
T = 1−S =
2s + 3

and are all stable.

Internal stability?

108
Check the following transfer functions
s+1
Gwu u = (1 + F G)−1 =
2s + 3
(s + 2)(s + 1)
Gwu = −(1 + F G)−1 F = −
(s − 1)(2s + 3)
s−1
Gw u y = (1 + GF )−1 G =
2s + 3
s +1
Gwy = (1 + GF )−1 =
2s + 3
The systemet is not internally stable as Gwu is unstable.

109
7 Limitations in Control Design

7.1

(a) The complementary sensitivity function is given by

F (s)G(s)
T (s) =
1 + F (s)G(s)
To compute the controller F (s) which results in the desired T (s) we
express F (s) as a function of T (s) and G(s) as follows. The above
expression yields
T (s)
F (s) = G−1 (s)
1 − T (s)
With the given T (s) and G(s) this results in
5(s + 1)
F (s) =
s(s − 3)
The zero located in s = 3 will be cancelled. However, the transfer
function from reference signal to control signal
5(s + 1)
U(s) = R(s)
(s − 3)(s + 5)
has a pole located in s = 3 and is unstable.

(b) We can get a bandwidth of 5 rad/s if we keep the right-half plane zero
and add a pole in s = −3, i.e.
5 3−s
T (s) = ·
s+5 3+s
In this case, the relationship
T (s)
F (s) = G−1 (s)
1 − T (s)
yields
5(s + 1)
F (s) = −
s(s + 13)
No pole-zer cancellation occurs and all closed-loop system transfer fun-
ctions are stable.

110
(c) With Fr = Fy = F we get

5 3−s
Gc (s) = T (s) = ·
s+5 3+s

The sensitivity function is

s(s + 13)
S(s) = 1 − T (s) =
(s + 3)(s + 5)

The Bode diagrams for those transfer functions are shown in the figure
below.

Bode Diagrams

From: U(1)
10

−10

−20
Phase (deg); Magnitude (dB)

−30

−40 T
S
−50

100

50

−50
To: Y(1)

−100

−150

−200

−250 T
S
−300
−2 −1 0 1 2
10 10 10 10 10

Frequency (rad/sec)

The bandwidth for Gc (i.e. T ) is 5 rad/s as desired. However, the phase

is −90◦ already at 2 rad/s. This means that the output signal will
not follow the reference signal for frequencies above this. In addition,
|S(iω)| > 1 for ω ≥ 1.3 rad/s, i.e. system disturbances are amplified
for those frequencies.

111
7.2

According to the rules of thumb presented in the textbook the bandwidth

of the closed-loop system cannot be greater than (1) half the magnitude of
the right-half plane zero, in this case 3/2 = 1.5 rad/s, (2) the inverse of the
time-delay, here 1/1 = 1 rad/s.

As a comparison we can study the crossover frequency. The crossover fre-

quency is often close to the bandwidth. The transfer function for a system
with a zero in s = 3 and a time-delay of 1 second can be expressed as

G(s) = e−s (3 − s)Ḡ(s)

or
(3 − s)
G(s) = e−s (3 + s)Ḡ(s)
(3 + s)
The argument of the frequency response is
ω
arg G(iω) = −ω − 2 arctan + arg((3 + iω)Ḡ(iω))
3
According to the assumptions the magnitude curve decreases monotonically
and according to Bode’s relation we get

arg((3 + iω)Ḡ(iω)) ≤ 0

This implies that

ω
arg G(iω) ≤ −ω − 2 arctan
3
and the phase margin is
ωc
ϕm = π + arg G(iωc ) ≤ π − ωc − 2 arctan
3
Let us study the case ϕm = 0 under the assumption that equality holds in
the above inequality. Then
ωc
0 = π − ωc − 2 arctan
3
i.e.
ωc ≈ 2
Hence, the crossover frequency cannot be greater than 2 rad/s.

112
7.3

Assume that we have a zero close to the origin and a pole far from the origin
in, both in the right-half plane. For example, we could have (ǫ << 1)
−ǫ + s
G(s) =
− 1ǫ + s
According to Theorem 7.4 in the textbook, the magnitude of the sensitivity
function must have a peak in a neighborhood of ǫ. In addition, Theorem 7.6
says that the magnitude of the complementary sensitivity function must have
a peak in the neighborhood of 1/ǫ.

7.4

(a) The requirements on |S(iω)| = σ̄(S(iω)) and |T (iω)| = σ̄(T (iω)) can
be formulated as
1 1
|S(iω)| ≤ 10
, ω ≤ 0.1, |T (iω)| ≤ 10
, ω≥2
1
|S(0)| ≤ 100

(b) The corresponding requirements on the loop gain GFy is

|G(0)Fy (0)| > 100
|G(iω)Fy (iω)| > 10, ω ≤ 0.1
1
|G(iω)Fy (iω)| < , ω ≥ 2
10
(c) The requirements in (a) can be reformulated using weighting functions
WS and WT such that
|S(iω)| ≤ |WS−1 (iω)|, ∀ω
|T (iω)| ≤ |WT−1 (iω)|, ∀ω
If WS−1 and WT−1 are first order transfer functions
   
−1 s −1 a2 s
WS (s) = a1 1 + , WT (s) = 1+
b1 s b2
we get, for example,
√ 
1 √ 2 s
WS−1 (s) = (1 + 30 11s), WT−1 (s) = 1+
100 10s 2
113
 (d) The minimal slope of the magnitude of the loop gain in the interval
[0.1, 2] is approximately given by the line tangent to the forbidden
regions in (b).

log 0.1 − log 10

Slope in the Bode plot: ≈ −1.53
log 2 − log 0.1
This implies
log 1 − log 10
= −1.53 ⇒ ωc = 0.45 rad/s
log ωc − log 0.1
From Bode’s relation we get
< π
arg GFy ≈ −1.53 · = −138◦
2
which results in a phase margin of approximately 40◦ .
A lower bound on kT k∞ ?

G(iωc )Fy (iωc ) = 1 · e−i·138 = −0.743 + 0.669i

◦

 
 G(iωc )Fy (iωc ) 
|T (iωc )| = 
  ≈ 1.4
1 + G(iωc )Fy (iωc ) 
kT k∞ = sup |T (iω)| ⇒ kT k∞ ≥ |T (iω)|, ∀ω
ω

⇒ kT k∞ ≥ 1.4

(e)

|T (iωc )| = 1.4
r
W (iωc ) = 0.14
 −1  0.452
T 1+ = 0.32
0.45 22
It is impossible to find a feasible solution using this choice of weighting
functions. Try weighting functions of higher order.

7.5

R∞
If the surface A2 is greater than the surface A1 we have that 0 log |S(iω)|dω >
0. According to Theorem 7.3, the loop gain G(s)Fy (s) has unstable poles.

114
7.6

The first requirement implies that

|S(iω)| < 10−3 ω≤2
where
1
S(s) =
1 + F (s)G(s)
When |F (iω)G(iω)| is large we approximately have
1
|S(iω)| ≈
|F (iω)G(iω)|
which results in
|F (iω)G(iω)| > 103 ω≤2
Furthermore, the system should be stable in spite of the model uncertainty
|∆G(iω)| ≤ 100|G(iω)| ω ≥ 20
where ∆G(s) is the absolute model error in G(s). Thus, the relative model
error fulfills the inequality
 
 ∆G(iω) 
 G(iω)  ≤ 100
 

To preserve stability we must have

|T (iω)| < 10−2 ω ≥ 20
where
F (s)G(s)
T (s) =
1 + F (s)G(s)
When |F (iω)G(iω)| is small we approximately have
|T (iω)| ≈ |F (iω)G(iω)|
and this results in
|F (iω)G(iω)| < 10−2 ω ≥ 20
To fulfill this requirement, the loop gain must decrease from 103 to 10−2
between the frequencies ω = 2 to ω = 20, i.e. 100 dB in a decade (slope
−5). According to Bode’s relation we have arg G(iω) ≈ −5 · 90◦ in this
interval. This results in an unstable closed-loop system. We can not fulfill
the requirements.

115
7.7

8 Controller Structures and Control Design

8.1

(a) !
− 57 12
7
RGA(G(0)) = G(0) . ∗ G−T (0) = 12
7
− 57

(b) Avoid the pairs u1 ↔ y1 and u2 ↔ y2 .

8.2

3 s+1
!
s+4 s+4
RGA(G(s)) = s+1 3
s+4 s+4

For the frequency zero we get

 
3/4 1/4
RGA(G(0)) =
1/4 3/4

As all elements in the RGA(G(0)) are positive all combinations are possible.
At the crossover frequency we get
12−30i 104+30i
!
116 116
RGA(G(10i)) = 104+30i 12−30i
.
116 116

We have elements close to 1 if u1 controls y2 and u2 controls y1 .

8.3

(a)  s−1 
2
RGA(G(s)) = s+1 s+1
2 s−1
s+1 s+1

116
 yields  
−1 2
RGA(G(0)) = .
2 −1
As we want to avoid pairing corresponding to negative elements in the
RGA(0) we have to choose u1 ↔ y2 and u2 ↔ y1 .

(b) As  
1 −2
G(0) =
1 −1
we choose W1 = G−1 (0) and W2 = I. A controller that decouples the
system in steady state is
 
diag −F11 (s) 2F22 (s)
F (s) = W1 F (s)W2 = .
−F11 (s) F22 (s)

8.4

Ideally, we want Q(s) to be Q(s) = G−1 (s). A realizable choice is:

τs + 1
Q(s) = .
K(λs + 1)
This results in
 
Q(s) τ 1
Fy (s) = = 1+ .
1 − Q(s)G(s) Kλ τs
τ
This is a PI controller with gain KP I = Kλ
and integration time TI = τ . The
sensitivity function is
λs
S(s) = 1 − G(s)Q(s) =
λs + 1
and the complementary sensitivity function is
1
T (s) = G(s)Q(s) = .
λs + 1
Thus, we have
|S(iω)| ≤ 1 ∀ω
which seems to disagree with Bode’s integral theorem. However, as the loop
gain GFy decreases as 1/ω Bode’s integral theorem is not applicable.

117
8.5

The system is a nonminimum phase system. We chose to replace the right-

half plane zero with a zero mirrored in the imaginary axis.
s2 + 5s + 6
Q(s) = .
(6 + 3s)(λs + 1)
This results in the controller
s2 + 5s + 6
F (s) = ,
s(3λs + 6(λ + 1))
which can be rewritten as
 
5 6 s 1
F (s) = 1+ + 3λ
.
6(1 + λ) 5s 5 6(λ+1)
s +1
This is a PID controller with a filter added to make it realizable.

8.6

1 p(p + 1)
Q(p) = G−1 (p) 2
=
(λp + 1) (λp + 1)2
p+1 1 1+p
Fy (p) = (1 − Q(p)G(p))−1 Q(p) = 2 = · p
λ p + 2λ 2λ 1 + 2/λ
1 1+p
⇒ u=− · p y
2λ 1 + 2/λ

|Fy (iω)|

1
λ2

1
2λ

1 ω
2/λ

1+p
High bandwidth ⇒ λ small ⇒ Fy (p) ≈ ⇒ PD controller
2λ
118
8.7

(a) We have the transfer function matrix

9
!
1 s+1
2
G(s) = .
s/20 + 1 6 4

The poles are the least common denominator of the minors

9 2
g11 (s) = (s/20+1)(s+1)
g12 (s) = s/20+1
6 4
g21 (s) = s/20+1
g22 (s) = s/20+1

and the minor

 
1 36 24(1 − s/2)
det G(s) = − 12 = .
(s/20 + 1)2 s+1 (s/20 + 1)2 (s + 1)

This results in the poles −20, −20 and −1. The zeros are given by
det G(s) normalized with the pole polynomial. This yields a zero located
at s = 2. We have to take proper care of the right-half plane zero in
the IMC design.

(b)

9
!−1
s+1
2
G−1 (s) = (s/20 + 1)
6 4
!
(s/20 + 1)(s + 1) 4 −2
= 9
24(−s/2 + 1) −6 s+1

Mirror the right-half plane zero of G(s) in the imaginary axis and add
the factor (λs + 1).
!
(s/20 + 1)(s + 1) 4 −2
Q(s) = 9
24(λs + 1)(s/2 + 1) −6 s+1

This results in the controller

!
(s/20 + 1)(s + 1) 4 −2
Fy (s) = (I − Q(s)G(s))−1 Q(s) = 9
24s(λs/2 + λ + 1) −6 s+1

119
8.8

The inverse of the system is:

 
−1 1 (s + 1)(s + 2) −3(s + 1)2
G (s) =
−s + 1 −(s + 1)(s + 2) 2(s + 1)(s + 2)
Mirror the right-half plane zero in the imaginary axis and form Q(s)
 
1 −s + 1 −1 1 s+2 −3(s + 1)
Q(s) = G (s) =
λs + 1 s + 1 λs + 1 −(s + 2) 2(s + 2)
The controller is given by Fy = (I − QG)−1 Q. The corresponding sensitivity
function is
s(λs+2+λ)
!
1 −s + 1 0
S = I − GQ = I − I = (s+1)(λs+1) s(λs+2+λ)
λs + 1 s + 1 0 (s+1)(λs+1)

|S(iω)| → 0 dı̈¿ 21 ω → 0 ⇒ The controller has integral action.

8.9

(a) Since system is quadratic, i.e the number of inputs equals the number
of outputs, the zeros can be determined as the poles of G−1 (s). This
gives  
1 −3
1  s+1 s+2 
G−1 (s) =
 
 =
det G(s)  −α 2 
s+1 s+1
 
(s + 1)(s + 2) −3(s + 1)2
1
=  
s(2 − 3α) + 4 − 3α
−α(s + 1)(s + 2) 2(s + 1)(s + 2)
The pole polynomial becomes
s(2 − 3α) + 4 − 3α
and hence the zero polynomial is given by
3α − 4
n(α) = .
2 − 3α
The location of the zero is shown in the figure below.

120
 10

n(α) 2

−2

−4

−6

−8

−10
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
α

Figur 7: Location of the zero as function of α.

(b) In order to obtain that F (s)G(s) is diagonal, F (s) has to contain the
inverse of G(s). For 2/3 < α < 4/3 we have n(α) > 0, i.e. the zero is
located in the right half plane. This means that F (s) is unstable, which
should be avoided if possible.

(c) Using static decoupling F is a constant matrix containing G(0)−1 . For

α = 4/3 the inverse does not exist, since the rows of G(0) becomes
linear dependent.

8.10

(a) The RGA can be computed as follows.

>> s = tf(’s’);
G = [1/(s+2) 2/(s+4); 1/(s+1) 1/(s+2)];
G0 = freqresp(G,0);
RGA = G0.*inv(G0).’

RGA =

-1 2
2 -1

121
 Negative elements in the diagonal of the RGA indicate that it will not
be possible the control system using a diagonal regulator.

(b) The transfer function matrix of the closed loop system is given by

Gc (s) = (I + G(s)F (s))−1G(s)F (s)

and using Matlab it can be computed as

>> F = diag([5 5]);

Gc = feedback(G*F,eye(2));
pole(Gc)

ans =

-14.4265
-2.3060 + 1.3712i
-2.3060 - 1.3712i
0.0385
-14.4265
-2.3060 + 1.3712i
-2.3060 - 1.3712i
0.0385

which means that the poles of the closed loop system are −14.4, −2.3 ±
1.37i and 0.04. Since there is a pole in the right half plane the system
is unstable.
Note: G(s) has 4 poles and a constant regulator does not add any
poles. The number of poles are doubled by the functions feedback and
tf. The extra poles are removed by using state space representation
and the command minreal.

>> pole(minreal(ss(Gc)))

ans =

-14.4265
0.0385
-2.3060 + 1.3712i
-2.3060 - 1.3712i

122
(c) The problem can be modified by renumbering the output signals, i.e.
switch the columns in G(s). This gives the modified transfer function
matrix  
1 1
Ḡ(s) = s+1 s+2
1 2
s+2 s+4

The RGA becomes

Gb = [1/(s+1) 1/(s+2); 1/(s+2) 2/(s+4)];

G0 = freqresp(Gb,0);
RGA = G0.*inv(G0).’

RGA =

2 -1
-1 2

and the closed loop system becomes

Ḡc (s) = (I + Ḡ(s)F (s))−1 Ḡ(s)F (s)

and this gives the poles

>> Gc = feedback(Gb*F,eye(2));
pole(Gc)

ans =

-16.2449
-4.8756
-1.4398 + 0.9522i
-1.4398 - 0.9522i
-16.2449
-4.8756
-1.4398 + 0.9522i
-1.4398 - 0.9522i

i.e. the closed loop is stable.

123
9 Minimization of Quadratic Criteria: LQG

9.1

(a) We have A = B = C = N = M = 1. Thus, the Riccati equation for

the Kalman filter is
P2
2P + R1 − = 0,
R2
q
R1
The positive semidefinite solution is P = R2 + R2 1+ R2
. Hence, the
Kalman gain is
r
1 R1 p
K= P =1+ 1+ = 1 + 1 + β.
R2 R2

Analogously we get for the state feedback

s
1 Q1 √
L= S =1+ 1+ = 1 + 1 + α.
Q2 Q2

This results in the controller

√ √
−1(1 + 1 + α)(1 + 1 + β)
Fy (p) = L(p − 1 + L + K) K = √ √ .
p+1+ 1+α+ 1+β

(b) The poles of the transfer functions of the closed loop system are given
by the eigenvalues of A − BL and A − KC respectively. Thus, the poles
are
√ p
− 1 + α and − 1 + β.

A small penalty on u (α large) results in a pole far from the origin in

the left-half plane. A large penalty on u results in a pole close to -1, i.e.
the pole of the
√ original system mirrored in the imaginary axis. The pole
located in − 1 + β can be affected in a similar way using the Kalman
filter design parameters.

124
9.2

(a) Introduce the state x = z. Then the system can be written as

ẋ = −x + u + v,
y = x + e,
z = x,
i.e. A = −1, B = 1, M = 1, Q1 = q1 and Q2 = 1. Q12 = 0. We
cannot measure the state but we mesure y. According to the separation
theorem V is minimized if we
(i) Estimate the state x̂(t) using a Kalman filter.
(ii) Use feedback according to u(t) = −Lx̂(t), where L is computed
using LQ theory.
Thus,
x̂˙ = Ax̂ + Bu + K(y − C x̂)
where K = P C T R2−1 and P is the positive semidefinite solution to
AP + P AT + NR1 N T − P C T R2−1 CP = 0.
In our case this is a scalar equation
P 2 + 2P − r1 = 0
with the solution √
P = −1 + 1 + r1 ,
i.e. √
K = −1 + 1 + r1 .
Use the feedback u = −Lx̂ with L = Q−1 T
2 B S, where S is the solution
to
AT S + SA + M T Q1 M − SBQ−1 T
2 B S = 0

As M = 1, Q1 = q1 and Q2 = 1 we get
p
L = S = −1 + 1 + q1 .
The loop gain is
1 1
G(s)Fy (s) = L K=
s + 1 1√ +s+L+K
√
(−1 + 1 + r1 )(−1 + 1 + q1 )
= √ √
(s + 1)(s − 1 + 1 + r1 + 1 + q1

125
 (b) The parameters r1 and q1 influence the loop gain in the same way due
to symmetry.

(c)
√ √
(−1 + 1 + r1 )(−1 + 1 + q1 )
G(s)Fy (s) = √ √
(s + 1)(s − 1 + 1 + r1 + 1 + q1 )
What happens when r1 or q1 → ∞?
√ √
(−1 + 1 + q1 ) −1 + 1 + q1
r1 → ∞ ⇒ G(s)Fy (s) = √ √ → .
(s + 1) (s−1+−1+
1+r1 + 1+q1 )
√ s+1
1+r1

Analogously we get
√
−1 + 1 + r1
lim G(s)Fy (s) =
q1 →∞ s+1

By varying q1 and/or r1 we can shape the loop gain according to the

sketch below:

|GFy |

The level is affected by r1 or q1

1 ω

9.3

State-space form:    
0 1 0
ẋ = x+ u
0 0 1


z =  1 0 x
1 0
y= x
0 1

126
The weighting matrices are Q1 = 1 and Q2 = η. The Riccati equation is:

AT S + SA + M T Q1 M − SBQ−1 T
2 B S = 0

Define  
s1 s2
S= ,
s2 s3

This results in
       2 
0 0 0 s1 1 0 1 s2 s2 s3
+ + − · =0
s1 s2 0 s2 0 0 η s2 s3 s23

The positive definite solution is

√
s1 = 2 · η 1/4
η 1/2
s2 = √
s3 = 2 · η 3/4

This yields the optimal feedback

 √ 1/4 1/2

1
2η η
L= Q−1 T
2 B S = · 0 1 √
η η 1/2 2 · η 3/4
1 √ 3/4 √
= · (η 1/2 2η ) = (η −1/2 2 · η −1/4 )
η

−1/4
The poles
√ are

the eigenvalues of A − BL. Define µ = η ⇒ L =
2
µ 2 · µ . This results in
 
s −1 √
0 = det 2
√ = s2 + 2µs + µ2 ,
µ s+ 2·µ

i.e. r
µ µ2 µ µ
s = −√ ± − µ2 = − √ ± i · √ =
2 2 2 2

µ 1
= − √ · (1 ± i) = − √ · (1 ± i)
2 2 · η 1/4
If η decreases the poles will be placed further away from the origin. This
results in an increased input signal u(t). Compare this result to the criterion.

127
9.4

Description of the system:

     
0 1 0 0
ẋ(t) = x(t) + µ(t) + τ (t)
0 −α H ν d


y(t) = 1 0 x(t) + vm (t)
The steady state Riccati equation is
       
0 1 0 0 1 0 0 1

0=S + S+ +S 0 H S
0 −α 1 −α 0 0 H ρ
Component by component we get
H2 2
0=1− s
ρ 12
H2
0 = − s12 s22 + s11 − αs12
ρ
H2
0 = − s222 + 2s12 − 2αs22
ρ
The positive definite solution is
√ s
ρ 2H
s11 = α2 + √
H ρ
√
ρ
s12 =
H s !
ρ 2H
s22 = − 2 α − α2 + √
H ρ
why the feedback gain L is given by
 p q √ 
ρ 2H ρ
1
H
α2 + √
ρ H
−1 T
L = Q2 B S = 0 H  √
ρ ρ
q 
ρ (−α + α2 + 2H
√ )
H H2 ρ
  q 
= √1ρ H1 −α + α2 + 2H √
ρ

The separation theorem states that it is optimal to use the estimated states
in the feedback. Thus, the optimal feedback is
s !!
1 1 2H
µ(t) = − √ −α + α2 + √ x̂
ρ H ρ

128
9.5

First find a state-space realization of the system

1
G(s) = ⇔ ÿ + ẏ = u
s(s + 1)
Let x1 = y, x2 = ẏ ⇒
   
0 1 0
ẋ(t) = x(t) + u(t)
0 −1 1


y(t) = 1 0 x(t) + n(t)

We want good robustness properties around the frequency ω = 0.5 rad/s, i.e.
we want the magnitude of the complementary sensitivity function T (s) to be
small at this frequency. As T (s) is the transfer function from the measurement
noise n(t) to the output signal y(t) we can proceed as follows:

If we estimate x̂(t) using the Kalman filter we will minimize the covariance
matrix of the estimation error. The model we use for n(t) will tell us for
which frequencies the measurments of y(t) are inaccurate. The Kalman filter
will suppress measurements at those frequencies, i.e |T (iω)| will be small.

As the Kalman gain K does not influence the closed-loop system Gc (s), we
can choose Q1 and Q2 to get a desired Gc (s).

If we study the transfer function of the Kalman filter, i.e. the transfer function
from y(t) to ŷ(t), we get an indication of how the mesurement noise affects
the suppression.

We want the noise model to have much energy around the frequency ω = 0.5
rad/s. One such model is n(t) = H(p)w(t) where w(t) is white noise, the
poles of H(s) are located at s = −0.01 ± 0.5i and there is a zero at s = 0,
i.e.
Kn s
H(s) = 2 .
s + 0.02s + 0.2501
Using controllable canonical form we get
   
−0.02 −0.2501 1
ẋn (t) = xn (t) + w(t)
1 0 0


n(t) = Kn 0 xn (t)

129
Extending the original state-space form with the noise model yields
     
0 1 0 0 0 0
 0 −1 0 0  x̄(t) +  1  u(t) +  0  w(t)
    
˙
x̄(t) =  0 0 −0.02 −0.2501   0   1 
0 0 1 0 0 0


y(t) = 1 0 Kn 0 x̄(t).
If this model, with an appropriate value of Kn , is used to compute the Kalman
gain K, the magnitude curve of the transfer function from y(t) to ŷ(t) = x̂1 (t)
will look as in the figure below. Signals at frequencies around ω = 0.5 rad/s
are heavily attenuated.

1
10

0
10
Amplitude

−1
10

−2
10 −2 −1 0 1 2
10 10 10 10 10
ω [rad/s]

9.6

Let G be the system, F the controller, y the output signal and v the distur-
bance. This results in
1
y= v = Sv
1 + GF
F
u= v
1 + GF
where S is the sensitivity function.

F is chosen such as the criterion

J(F ) = E{y(t)2 + αu(t)2}, α > 0

130
is minimized given that Φv (ω) = δ(ω).

The criterion can be written as

Z ∞
1 F02
J(F ) = Φy (ω) + αΦu (ω) dω = + α .
−∞ (1 + G0 F0 )2 (1 + G0 F0 )2
where F0 and G0 are the stationary gains of the controller and the system
respectively. As F minimizes J(F ) we have
∂V 2αF0 − 2G0
=0=
∂F (1 + G0 F0 )3
which yields F0 = G0 /α. Thus, the sensitivity function at ω = 0 is
1 1
S0 = = .
1 + G0 F0 1 + G20 /α

9.7

(a) How to solve the problem:

(i) Compute the Kalman filter: x̂˙ = Ax̂ + Bu + K(y − C x̂)

(ii) Use feedback: u = −Lx̂, where L is computed using LQ the-
ory.

(i) + (ii) results in the controller Fy = L(sI−A+BL+KC−KDL)−1 K.

(i) State-space description: Let x1 = z, x2 = ν, v1 = v, v2 = e and

x = (x1 , x2 )T . This results in
     
−1 1 1 0
ẋ = x+ u+ v1
0 −ǫ 0 1
| {z } | {z } | {z }
A B N


y = 1 0 x + v2
| {z }
C

Furthermore, we have

R1 = Φv1 (ω) = Φv (ω) = 1

R2 = Φv2 (ω) = Φe (ω) = 1
R12 = Φv1 v2 = 0

131
 The Kalman filter is given by: K = P C T R2−1 with P according to

AP + P AT + NR1 N T − P C T R2−1 CP = 0
 
p1 p2
Define P = .
p2 p3
 √ 
3 − 1 √1
This results in limǫ→0 P = and
1 3
 √ 
3−1
K = P CT R2−1 =
1
(ii) Compute L such that
Z ∞ Z ∞
2 2
min x1 (t) + u (t) dt = min y T Q1 y + uT Q2 u dt
L 0 L 0

where Q1 = Q2 = 1.
The optimal L is given by: L = Q−1 T
2 B S where S is the positive
semidefinite solution of

AT S + SA + C T Q1 C − SBQ−1 T
2 B S = 0.
 
s1 s2
Define S = .
s2 s3
√ !
2 − 1 1 − √12
This results in limǫ→0 S = and
1 − √12 ∗
 √ 
L = Q−1 B T
S = 2 − 1 1 − √1
2 2
The LQG controller is

x̂˙ = Ax̂ + Bu + K(y − C x̂)
u = −Lx̂

with K and L as above.

The static gain of the sensitivity function (ǫ = 0):

(i) and (ii) yields
 
−1 D=0
Fy = L(sI − A + BL + KC − KDL) K = =1
s=0

1 1 1
S(0) = = = .
1 + Fy (0)G(0) 1+1 2

132
 (b) Compute L using LTR(y):

Lltr = Q−1 T
2 B S
S : AT S + SA + C T ρQ2 C − SBQ−1 T
2 B S = 0

√ √ !
1+ρ−1 √1+ρ−1  √ √ 
⇒S= √ 1+ρ+ǫ
⇒ Lltr = 1+ρ−1 √1+ρ−1
√1+ρ−1 ∗ 1+ρ+ǫ
1+ρ+ǫ

The static gain of the sensitivity function:

ǫ
S(0) → √
3ǫ + 1

when ρ → ∞. It is necessary to let ǫ → 0 to get S(0) → 0, i.e.

introducing an integrator into the system.

9.8

Define the matrices

A B
z }| { z }| {

0 1 −1 0
1 2
ẋ = − 2 w 0
 −0.01 0.01 x + w0  u
 
1 2
w
2 0
0.01 −0.01 0


z = 0 0 1 x,
| {z }
M

where
k
w 20 =
.
50
The Bode plot for k = 1 is given. There is a resonance peak at w0 ≈ 0.14.

Introduce measurement noise: Let y = z+v2 , where v2 is colored measurement

noise. Using the feedback u = −Lx̂ + p̃r we can write z as

z = Gc r − T v2 + s̃v1 .

The robustness criterion implies that |T (iω0 )| should be small to handle large
errors in k. A large spectrum for v2 at w = w0 will force T to be small at this

133
frequency. Let v2 be colored noise with a peak in the spectrum at w0 . This
can be achieved by chosing poles in −0.01 ± 0.14i and a zero in 0, i.e.

k2 p
v2 = w,
p2 + 0.021p + 0.02

where w is white noise.

State-space representation of v2 :
   
−0.02 −0.02 1
ẋv = xv + w
1 0 0
| {z } |{z}
Av Bv

v2 = (k2 0) xv
| {z }
Cv

The extended model is

       
A 0 B 0 x
x̄˙ = x̄ + u+ w, x̄ =
0 Av 0 Bv xv

y = (M Cv ) x̄
z = (M 0) x̄
v2 = (0 Cv ) x̄

with A, B, M, Av , Bv , Cv as above.

9.9

The loop gain is

18
L(sI − A)−1 B =
(s − 1)(s + 2)

The Nyquist curve will approach the origin with the angle −180◦ . An LQ
controller always approaches the origin with −90◦ .

134
9.10

The system has the following controllability and observability matrices

   
−4 −12 1 1
C= , O= ,
8 24 3 3

respectively. Thus, the system is neither controllable nor observable. Since

V (T ) tends to the quadratic norm of the LQG problem as T → ∞, we must
have V (∞) = ∞.

9.11

9.12

9.13

9.14

(a) According to Equation (9.7a) in the text book

L = Q−1 T
2 B S

where
0 = AT S + SA + M T Q1 M − SBQ−1 T
2 B S

Here A = α, B = 1 M = 1 Q1 = 1 and Q2 = ρ. This gives

S2
0 = 2αS + 1 −
ρ
which implies
S 2 − 2ραS − ρ = 0
with solution
+ p
S = ρα (−) (αρ)2 + ρ
This gives p
L=α+ α2 + 1/ρ

135
 (b) Using the result from above gives for the case α = 1
p
L = 1 + 1 + 1/ρ
i.e. L → 2, while the case α = −1 gives
p
L = −1 + 1 + 1/ρ

i.e. L → 0. For α = 1 the open loop system is unstable and has to be

stabilized using feedback, and hence u = 0 will not work.

9.15

The choices (iii) and (iv) give the same gain vector L since J(iii) = 0.1 J(iv) .
The L that minimizes J(iii) will also minimize J(iv) . The figures (A) and (C)
show the same simulation results. The matrices in (i) put less weight on the
input u which implies a faster settling, i.e. (B). The choice (ii) puts a weight
on the velocity which implies a slower response, i.e. (D).

Answer: (i) – B, (ii) – D, (iii) – A and C, (iv) – A and C

9.16

(a) The system has the poles

0 0 − 0.0850 + 0.7435i − 0.0850 − 0.7435i

Since the system has two poles in the origin the system is not asymp-
totically stable.
(b) The feedback gain becomes

L = (−7.0973 2.0419 8.1531 1)

and the poles of the closed loop system becomes, i.e. the eigenvalues of
A − BL, become
−0.1809 + 0.8271i
−0.1809 − 0.8271i
−0.4368 + 0.2915i
−0.4368 − 0.2915i

136
 x’ = Ax+Bu
y = Cx+Du
State−Space Auto−Scale
Graph2
K
Matrix
Gain

Auto−Scale
Graph1

1.5

1
x4

0.5

−0.5
0 5 10 15 20 25 30 35 40 45 50
Sekunder

0.5

−0.5
u

−1

−1.5

−2
0 5 10 15 20 25 30 35 40 45 50
Sekunder

The closed loop system can be simulated using the model. The state
x4 and u are given by the figure and it can be seen that all signals
tend to zero. The other states are given in the figure below, where
x1 ↔ solid line
x2 ↔ dashed line
x3 ↔ dash-dotted line
0.6

0.4

0.2

−0.2
x

−0.4

−0.6

−0.8

−1
0 5 10 15 20 25 30 35 40 45 50
Sekunder

137
 (c) Increasing Q2 reduces u and decreasing Q2 increases u.

(d) An example of matrices that gives a feedback such that the conditions
are fulfilled is given by
 
250 0 0 0
 0 0 0 0 
Q1 = 
 0 0
 , Q2 = 1
0 0 
0 0 0 1

The feedback gain becomes

L = (−6.8105 7.5015 22.0860 1)

9.17

(a) The requirements are fulfilled by the state feedback

u(t) = −Lx(t)

where

L = −4.4721 −4.9405 19.1028 6.1811
which is achieved from LQ-minimization using
 
20 0 0 0
 0 0 0 0
 0 0 0 0 Q2 = 1
Q1 =  

0 0 0 0

The poles of the closed loop system, i.e the eigenvalues of A − BL, all
have the absolute value 2.36.

(b) The closed loop system has the characteristic equation

λ4 + l4 λ3 + l3 λ2 − 7l2 λ − 7l1 = 0

The closed loop system is stable if all roots are located (strictly) in the
left half of the complex plane. A necessary condition (but not sufficient)
is that all coefficients in the polynomial are strictly positive. A loss of
a measurement of state variable can be interpreted as li = 0 for some
i, and this violates the condition.

138
10 Loop Shaping

10.1

The criterion to minimize is the H2 norm of Gec . The system on state-space

form:

ẋ1 = −x1 + u
y = x1

Weighting functions

1
Wu (s) = 5, WT (s) = 0.5, WS (s) =
s
Form the extended system G0 :

z1 = Wu u = 5u
z2 = WT Gu = 0.5x1
z3 = WS (Gu + w) = x2

where x2 is a new state defined as

1
x2 = (Gu + w) ⇔ ẋ2 = x1 + w
p

This yields
     
−1 0 1 0
ẋ = x+ u+ w
1 0 0 1
   
0 0 5
z = 0.5 0 x + 0 u
0 1 0


y = 1 0 x+w

Check M and D




0 0 5


DT M D = 5 0 0 0.5 0 0 = 0 0 25 6= 0 0 1
0 1 0

139
Hence, define a new input signal ũ as

1
ũ = (D T D)1/2 u + (D T D)−1/2 D T Mx = 5u ⇔ u = ũ
5
This is just a scaling of the original input signal. Thus we get a new B matrix

1
B̃ = B
5

Solve the Riccati equation: AT S + SA + M T M − S B̃ B̃ T S = 0

Define  
s1 s2
S=
s2 s3
which yields
   
−s1 + s2 −s2 + s3 −s1 + s2 0
+
0 0 −s2 + s3 0
   2 
0.25 0 1 s1 s1 s2
+ − =0
0 1 25 s1 s2 s22

Hence,
 1 2
 −2s1 + 2s2 + 0.25 − 25 s1 = 0

1
−s2 + s3 − 25 s1 s2 = 0

 1 2
1 − 25 s2 = 0
52
which has the positive semidefinite solution s1 = 4.686, s2 = 5 and s3 ≥ 4.686
.
Thus, the state feedback for the scaled system is



L̃ = B̃ T S = 1
s
5 1
1
s
5 2 = 0.937 1

For the original system we get

1

L = L̃ = 0.187 0.2
5
The controller is

x̂˙ = Ax̂ + Bu + N(y − C x̂)

u = −Lx̂

140
10.2

The criterion to minimize is the H∞ norm of Gec . The extended system is the
same as in the previous exercise. The controller is L = B T S, for the smallest
value of γ for which
AT S + SA + M T M + S(γ −2 NN T − BB T )S = 0
has a positive semidefinite solution. If we solve this numerically using Mat-
lab we see that γ ≥ 5.12 produce positive definite solutions. For γ = 5.2 we
get

L = 2.6873 2.7632

10.3
1
(a) The frequency weights WS = s
and WT = Wu = 1 result in
z1 = Wu u = u
z2 = WT Gu = Cx
z3 = WS (Gu + w) ⇔ ż3 = Cx + w = y
and
        
ẋ A 0 x B 0
= + u+ w
ż3 C 0 z3 0 1
 

x
y= C 0 +w
z3
Controllers for the H2 and H∞ cases can be computed using the ex-
pressions in the textbook.
(b) The observer is given by
x̂˙ = Ax̂ + Bu ⇔ x̂ = (pI − A)−1 Bu
R
ẑ˙3 = C x̂ + (y − C x̂) = y ⇔ ẑ3 = y dτ
The state feedback is
 

x̂ R
u = − L −α = −Lx̂ + αẑ3 = −L(pI − A)−1 Bu + α y dτ
ẑ3
By solving for u we get the desired controller structure
α R
u= y dτ
1 + L(pI − A)−1 B

141
 (c) If the system contains an integrator we have det(pI − A) = p · ξ(p)
which implies that
α 1 αξ(p)
u= 1 · y= y
1+ pξ(p)
L(pI − A)a B p pξ(p) + L(pI − A)a B

i.e. the integral part of the controller is cancelled.

10.4

10.5

A normal requirement is to have a small sensitivity function for low frequen-

cies, which means that C is excluded. Bode’s integral theorem states that it
is impossible to achieve that | S(iω) |< 1 ∀ ω, which excludes B. Therefore
A is the best alternative.

10.6

(a) Straightforward calculations give

1 1 s+1
S(s) = = 1 = ,
1 + G(s)F (s) 1 + s+1 K s+1+K
1
G(s)F (s) K K
T (s) = = s+1 1 = and
1 + G(s)F (s) 1 + s+1 K s+1+K
F (s) K K(s + 1)
Gru (s) = = 1 =
1 + G(s)F (s) 1 + s+1 K s+1+K

(b) – In Alternative II it is required that both S and T are small for

low frequencies. Since S + T = 1 always holds this requirement is
not realistic.
– In alternative III the requirement on T is the opposite to a correct
specification for the complementary sensitivity function. Normally
one requires, for robustness and measurement noise reasons, that
T is small for high frequencies.

142
 – Alternative I is realistically specified with small S for low frequen-
cies and small T for high frequencies.

(c) The requirements can be fulfilled by e.g choosing K = 9.

143
12 Stability of Nonlinear Systems

12.1

The state variables x1 = y and x2 = ẏ yield

 
x2
ẋ = = f (x),
−0.2(1 + x22 )x2 − x1

and
V̇ = Vx f (x) = x1 ẋ1 + x2 ẋ2 = −0.2x22 (1 + x22 ).
Hence, V̇ < 0 except when x2 = 0. If x2 ≡ 0 we have that x1 = constant = 0.
Thus the zero solution is asymptotically stable.

12.2

No, V (x) ≥ 0, ∀x is not fulfilled.

12.3

The slopes 
k1 = 0.5
⇒
k2 = 3
result in a circle going through the points −1/3 and −2.

12.4

The nonlinearity is bounded by two lines with slope k1 = 0 and k2 = 1

respectively. According to the circle criterion the closed loop system is stable
if the Nyquist curve lies to the right of the line s = −1.

144
12.5

The circle criteria applies to a model with one nonlinear static block, and
one linear dynamic block. Here, we have two linear blocks, so our first step
is to convert it to the standard case.

To begin with, we note that we cannot simply move the constant K straight
through the nonlinear block and put it next to the linear dynamics, since
Kf (u1 ) 6= f (Ku1 ). Nevertheless, we can move it to the linear dynamics
block, but we have to go the other way around.

Let z denote the output from the linear dynamics and call the linear dynamics
H(s). We have Z(s) = H(s)U2 (s), u1 = K(r − z), u2 = f (u1 ). This can be
written as u1 = Kr − Kz, i.e., u1 = r̃ − y if we define r̃ = Kr and Y (s) =
KZ(s) = (KH(s))U2 (s) = G(s)U2 (s). In this form, we have u2 = f (r̃ − y),
Y (s) = G(s)U2 (s) which is exactly the type of loop analyzed using the circle
criteria.

This was just a long-winded way of proving that you can pull K along the
loop backwards and multiply it with the linear dynamics, remembering that
when you pull it through the loop, you are redefining the external variable.

The nonlinearity is bounded by two lines with slope k1 = 0 and k2 = ∞

respectively. The linear dynamics must have a strictly positive real part.
K K(−iω)(−iω+1)
After simplifying the expression G(iω) = (iω)(ıω+1) = (iω)(iω+1)(−iω)(−ıω+1) to
find the real part, we arrive at

K
Re G(iω) = − < 0, ∀ω
ω2 + 1

Notice that there is really no need to simplify as far as done here, as you
know the denominator is a positive real number as it is an absolute value
zz ∗ = |z|2 . All you have to do is to study the real part of the numerator after
having multiplied with the conjugate of the denominator, since all you need
is the sign of this term.

Hence, stability cannot be guaranteed for any K > 0.

145
12.6

The state variables x1 = Φ, x2 = Φ̇ yield

ẋ1 = x2
mg
ẋ2 = − sin x1 l
J
The controller
l = l0 + εΦΦ̇
results in

ẋ1 = x2
mg
ẋ2 = − sin x1 (l0 + εΦΦ̇)
J
As a candidate Lyapunov function we use
1
V (x) = Jx22 + mgl0 (1 − cos x1 )
2
which corresponds to the energy of the system. We get

V̇ = Jx2 ẋ2 + mgl0 sin x1 ẋ1 = −εmgx22 x1 sin x1 ≤ 0 (−π/2 < x1 π/2)

with V̇ = 0 only when x1 ≡ 0 or x2 ≡ 0. x1 ≡ 0 ⇒ x2 = 0 and x2 ≡ 0 ⇒

x1 = 0.

Answer: V̇ (x) < 0 for allx1 6= 0, x2 6= 0 implies that x → 0.

12.7

The states x1 = y, x2 = ẏ and the controller

u = − sgn(ax1 + bx2 )

yield

ẋ1 = x2
ẋ2 = −2x1 − 3x2 − sgn(ax1 + bx2 )

146
The Lyapunov function candidate
α β
V (x) = ( x21 + x22 )
2 2
results in
V̇ = (α − 2β)x1 x2 − 3βx22 − βx2 sgn(ax1 + bx2 )
Take, for example, α = 2, β = 1, a = 0, b = 1, which result in

V̇ = −3x22 − |x2 | ≤ 0

12.8

The nonlinearity is
f (u) = u + arctan(u)
The derivative of f (u) is
1
f ′ (u) = 1 +
1 + u2
and has its maximum f ′ (0) = 2 for u = 0. f ′ (u) → 1 dı̈¿ 21 u → ∞. This
implies
u + arctan(u)
1≤ ≤2
u
which means that the Nyquist curve of the linear part of the system must lie
outside and not encircle the circle passing through −1 and −1/2.

12.9

According to the circle criterion the system is stable if the Nyquist curve lies
outside and does not encircle the circle passing through−4/3 and −4/7.

According to the textbook the loop gain for an LQ controller lies outside
and does not encircle the circle passing through 0 and −2. As this circle
encompasses the above smaller circle the system is stable.

147
13 Phase Plane Analysis

13.1

a) From
10 2
ẏ )ẏ + y + y 2 = 0
ÿ − (0.1 −
3
a reasonable guess is to try the states x1 = y and x2 = ẏ. With these,
we can write the system on state-space form

ẋ1 = x2 = f1 (x1 , x2 )
ẋ2 = −x1 (1 + x1 ) + x2 (0.1 − 10x22 /3) = f2 (x1 , x2 )

b) To find the stationary points, we must find solutions to ẋ(t). We call

the stationary points x̄ here.
f (x̄) = 0 ⇒ x̄2 = 0 and x̄1 (1 + x̄1 ) = 0.
 
x̄1 = 0 x̄1 = −1
SP I : , SP II :
x̄2 = 0 x̄2 = 0

c) To linearize the system around the stationary point we use

f (x) ≈ f (x̄) + fx (x̄)(x − x̄) = fx (x̄)(x − x̄)

as f (x̄) = 0. The matrix fx (x) is the Jacobian off . The ij entry is

∂fi
∂xj
(x):
∂f1 ∂f1
∂x1
= 0, ∂x2
= 1,
∂f2 ∂f2
∂x1
= −1 − 2x1 , ∂x2
= 0.1 − 10x22 .
Make the change of variables z = x − x̄.
SP I:
Linear approximation ż = Az, with
 
0 1
A=
−1 0.1

The eigenvalues of the matrix A are given by

0 = det(λI − A) = λ(λ − 0.1) + 1,

148
 i.e. we have unstable dynamics since we have postive real part on both
eigenvalues. √
λ = 0.05 ± 0.052 − 1
SP II:
Linear approximation ż = Bz, where
 
0 1
B=
1 0.1

The eigenvalues of B,

0 = det(λI − B) = λ(λ − 0.1) − 1

√
λ = 0.05 ± 0.052 + 1, λ1 = −0.95, λ2 = 1.05

Once again unstable, since one of the eigenvalues is positive.

d) SP I :

The linear approximation has an unstable focus at (0,0). For unstable

nodes the the nonlinear differential equation has a stationary of the
same type as the linear approximation. Note that the linear approxi-
mation is only valid close to the stationary point.

linear
0.5

0.4

0.3

0.2

0.1

−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

nonlinear

149
 0.5

0.4

0.3

0.2

0.1

−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

SP II:
The linearized approximation has a saddle point at (−1, 0). This is
true for the nonlinear differential equation too. The eigenvector cor-
responding to the stable eigenvalue is (1, −0.95), and the eigenvalue
corresponding to the unstable eigenvalue is (1, 1.05).
0.4

0.3

0.2

0.1
x2

−0.1

−0.2

−0.3

−0.4
−1.3 −1.2 −1.1 −1 −0.9 −0.8 −0.7 −0.6
x1

Far away from the stationary points:

How do the trajectories behave far from the origin? Form the derivative,
dx2 ẋ2 −x1 (1 + x1 ) + x2 (0.1 − 10x22 /3)
= =
dx1 ẋ1 x2
When x1 is bounded and x2 → ±∞, we have that ẋ2 /ẋ1 → ∞. Hence,
the trajectories ar vertical when |x2 | grows and also when x2 → 0.

150
 3

1
x2

−1

−2

−3
−4 −3 −2 −1 0 1 2 3 4
x1

13.2

The nonlinearity is described by


 x + a, x < −a
f (x) = 0, −a ≤ x ≤ a

x − a, x > a

The relationship between x and e is given by

p(p + B)x(t) = Ke(t)

i.e.
ẍ + B ẋ = Ke
In addition we have that

e = u − f (x) = −f (x)

which yields
ẍ + B ẋ + Kf (x) = 0

151
Introduce the states x1 = x and x2 = ẋ. The state-space form is

ẋ1 = x2 (4)
ẋ2 = −Kf (x1 ) − Bx2 (5)

Partition into regions where f (x) is linear.

1. The region x1 < −a:

The stationary point is x1 = −a, x2 = 0. The cange of variables z1 =
x1 + a, z2 = x2 results in linear state-space equations , ż = Az, where
 
0 1
A=
−K −B

The eigenvalues of A are

r
B B2
λ=− ± −K
2 4
Thus, the point x = (−a, 0) is a stable node or a stable focus. Sketching
the phase portrait when 4K > B 2 , results in a stable focus.

−3a −2a −a 0 a

2. The region−a ≤ x1 ≤ a:
The stationary points are line segments : −a ≤ x1 ≤ a and x2 = 0. The
dynamic equations are

ẋ1 = x2
ẋ2 = −Bx2

152
 Form the derivative
dx2 ẋ2
= = −B
dx1 ẋ1
In the entire region the trajectories have the slope −B.

−a 0 a

3. The region x1 > a:

stationary point x1 = a, x2 = 0. Make the change of variables z1 =
x1 − a, z2 = x2 . This results in the same state-space equation for z as
for the case x1 < −a, ż = Az. If 4K > B 2 , we have the unstable focus
x = (a, 0) analogously to the case x1 < −a.

To get the phase portrait we need to join the three partial solutions found
in 1, 2 and 3.
x2

−a 0 a
x1

153
13.3

(a) Introduce x1 = y, x2 = ẏ, which yields

ẋ1 = x2
ẋ2 = −sgn x1

There are no stationary points. The phase portrait is constructed in

two steps depending on sgn x1 . When x1 > 0 we have
dx2 1
=−
dx1 x2
This separable differential equation has the solution
1 2
x + x1 = constant
2 2
i.e. x1 as a function of x2 is a number of parabolas. When x1 < 0 we
get analogously
1 2
x − x1 = constant
2 2
The phase portrait:

1.5

0.5
x2

−0.5

−1

−1.5

−2
−1.5 −1 −0.5 0 0.5 1 1.5
x1

(b) For x1 > a we have 21 x22 + x1 = konst and for x1 < −a we have
1 2
x − x1 = konst. For the case |x1 | ≤ a the relay will have the same
2 2
output as before it entered teh region, i.e. the parabola is continuing.
The phase portrait when a = 0.5:

154
 4

x2 0

−1

−2

−3

−4
−6 −4 −2 0 2 4 6 8
x1

13.4

(a) Let x1 = y and let x2 be the input to the nonlinearity. This results in
the state-space form

ẋ1 = f (x2 )
ẋ2 = −x1

where


 x + 1, x < −1
f (x) = 0, −1 ≤ x ≤ 1

x − 1, x>1

We get centers in the stationary points x = (0, −1) (for x2 ≤ −1) and
x = (0, 1) (for x2 ≥ 1 ). When −1 < x2 < 1 we have x1 = constant.
This results in the phase portrait

155
 x2
0

0
x1

The car will not return to the desired position with proportional con-
trol.

(b) Now, we have the state-space form

ẋ1 = f (x2 )
ẋ2 = −x1 − f (x2 )

The difference from (a) is that the stationary points are stable focuses.
Hence, we get the phase portrait
x2

0
x1

The car returns to the desired position.

156
13.5

1 1
ẋ2 = 0 ⇒ 0 = −3x2 (1 + x1 ) + x1 x2 = (x1 − 6)x2
6 2

i.e.
x1 = 6 or x2 = 0.

Two cases:

(i) x2 = 0 and ẋ1 = 0 ⇒ 0 = 2x1 − 0.2x21 = 0.2(10 − x1 )x1 ,

i.e. x1 = 0 or x1 = 10.

(ii) x1 = 6 and ẋ1 = 0 ⇒

1 1
0 = 2 · 6(1 + · 6) − 6 · x2 − 0.2 · 62 (1 + · 6) = 24 − 6x2 − 14.4
6 6

Hence, the stationary points (SP) are

  
x1 = 0 x1 = 10 x1 = 6
SP I : , SP II : , SP III :
x2 = 0 x2 = 0 x2 = 1.6

The Jacobian is
 
2 − 0.4x1 − x2 /(1 + x1 /6)2 −x1 /(1 + x1 /6)
H(x) = fx (x) = 2
x2 /(1 + x1 /6) −3 + x1 /(1 + x1 /6)

SP I:
x1 = x2 = 0 yield
 
2 0
H1 =
0 −3

The origin is a saddle point with trajectories according to

157
 1

0.8

0.6

0.4

0.2
x2

−0.2

−0.4

−0.6

−0.8

−1
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2
x1

SP II:
When x1 = 10 and x2 = 0 the Jacobian is
 
−2 −3.75
H2 =
0 0.75
This is also a saddle point. The eigenvector corresponding to the unstable
eigenvalue is (3.75, −2.75), and the eigenvalue corresponding to the stable
eigenvalue is (1, 0). The phase portrait is

0.8

0.6

0.4

0.2
x2

−0.2

−0.4

−0.6

−0.8
9 9.2 9.4 9.6 9.8 10 10.2 10.4 10.6 10.8 11
x1

SP III:
x1 = 6, x2 = 1.6 yield  
−0.8 −3
H3 =
0.4 0

158
The eigenvalues are −0.4 ± 1.02i. We have a stable focus with the phase
portrait

1.9

1.8

1.7
x2

1.6

1.5

1.4

1.3

1.2
5 5.2 5.4 5.6 5.8 6 6.2 6.4 6.6 6.8 7
x1

IF we join these phase portraits together we get:

1.8

1.6

1.4

1.2
x2

0.8

0.6

0.4

0.2

0
0 2 4 6 8 10 12
x1

13.6

Introduce the states x1 = y and x2 = ẏ ⇒.

ẋ1 = x2
ẋ2 = −x1 + f (x2 )

159
For x2 > 0 we have

ẋ1 = x2
⇒ stationary point (−1, 0)
ẋ2 = −x1 − 1

The Jacobian is  
0 1
−1 0

As the eigenvalues are ±i ⇒ x = (−1, 0) is a center.

For x2 < 0 we get analogously that x = (1, 0) is a center. If we join the phase
portraits together we get

1
x2

−1

−2

−3
−4 −3 −2 −1 0 1 2 3 4
x1

The system will tend to a point on the x1 -axis, i.e. ẏ will tend to zero.

13.7

(a) We have stationary points for x1 = 0, i.e. the entire x2 -axis, when
u = 0. The trajectories are described by

dx2 1 1
=− 2 ⇔ x2 = + C.
dx1 x1 x1

160
 u=0
5

x2
0

−1

−2

−3

−4

−5
−5 −4 −3 −2 −1 0 1 2 3 4 5
x
1

(b)

1
V̇ = Vx ẋ = −2x41 + 2x1 u + 2x1 x2 ⇒ Vı̈¿ lj u = −x1 − x2
2

This results in V̇ = −2x41 − 2x21 < 0 and the stationary point x1 = x2 =

0. The corresponding linearized system is

  √
−1 −1 1 3
ẋ = x, with the eigenvalues λ = − ± i .
1 0 2 2

The stationary point is a stable focus.

u = −x −x
1 2
1

0.8

0.6

0.4

0.2
2

0
x

−0.2

−0.4

−0.6

−0.8

−1
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
x
1

161
14 Oscillations and Describing Functions

14.1

The describing function of a saturation is given in the textbook, together

with describing functions for some other nonlinearities. We will derive it here
anyway to demonstrate how to do it.

1. Apply the signal e(t) = C sin Φ, where Φ = ωt, to the input of the
saturation. If C ≤ 1 the output signal after the saturation will beu(t) =
e(t). If C > 1 the output signal u(t) will be the solid curve in the figure
below:

1
u(Φ)

Φ1

−1

0 π 2π
Φ
Here Φ1 is given by C sin Φ1 = 1, i.e. Φ1 = arcsin(1/C).
2. Compute the Fourier coefficients a1 and b1 as
Z Z
1 2π 1 2π
a1 = u(Φ) cos Φ dΦ , b1 = u(Φ) sin Φ dΦ
π 0 π 0
As u(Φ) is an odd function and cos Φ is even a1 = 0. Utilizing symmetry
we can write b1 as
Z
4 π/2
b1 = u(Φ) sin Φ dΦ
π 0
Z Φ1 Z π/2 !
4
= C sin2 Φ dΦ + sin Φ dΦ
π 0 Φ1
 
4C Φ1 sin 2Φ1 cos Φ1
= − +
π 2 4 C

162
 √
As sin 2Φ1 = 2 sin Φ1 cos Φ1 , sin Φ1 = 1/C and cos Φ1 = C 2 − 1/C we
get  √ 
2C 1 C2 − 1
b1 = arcsin +
π C C2
3. The describing function is given by Yf (C) = (b1 + ia1 )/C
 √ 
2 1 C2 − 1
Yf (C) = arcsin +
π C C2

This is valid for C > 1, Yf (C) = 1 when C ≤ 1)

(a) The problem can be solved either by hand or by using Matlab and both
alternatives will be presented here.
Alternative (i): According to the calculations above the describing fun-
ction of the satuaration is real, starts in 1 for C ≤ 1 and tends to zero
when C tends to infinity. This means that −1/Yf will start in −1 and
tend to −∞ when C grows.
The transfer function of the linear part is
10
G(s) =
s(s + 1)2
Since the system contains an intergrator the argument of G(iω) will
start at −90◦ for low frequencies, and since the system has relative
degree three the argument will tend to −270◦ . This implies that G(iω)
will cross the negative axis and there is a possibility that it will cross
−1/Yf . Using the fact that

arg G(iω) = arg 10 − arg(iω(iω + 1))2 = 0 − 90◦ − 2 arctan ω

we find that arg G(iω) = −180◦ (i.e. it crosses the negative real axis)
for ω = 1. Using also that

|G(iω)| = 10/(ω(1 + ω 2))

we find that | G(i1) |= 5, i.e. G(iω) crosses the negative real axis in the
point −5, and there will hence be an intersection with −1/Yf . In order
to find the corresponding value of C we need to solve the equation
 √ 
2 1 C2 − 1
arcsin + = 0.2
π C C2

163
which has the approximate solution C = 6.3. For oscillations with
C < 6.3 the curve G(iω) will encircle −1/Yf and hence the amplitude of
the oscillations will grow. Correspondingly, for oscillations with C > 6.3
the curve G(iω) will not encircle −1/Yf and hence the amplitude of the
oscillations will decay. Hence the describing function method predicts
that there will be a limit cycle with angular frequency ω = 1 and
amplitude C = 6.3.
Alternative (ii): The Nyquist curve can be plotted in Matlab using

>> s=tf(’s’);
>> G=10/(s*(s+1)^2);
>> nyquist(G)
>> axis([-8 0 -2 2])

and by clicking in the plot one gets

Nyquist Diagram

1.5

0.5 System: G
Imaginary Axis

Real: −4.92
Imag: −0.00704
Frequency (rad/sec): −1.01
0

−0.5

−1

−1.5

−2
−8 −7 −6 −5 −4 −3 −2 −1 0
Real Axis

Note: The curve does not pass exactly through −5, which is the correct
value according to the analytical calculation, and this is caused by the
automaticaly selection of frequency points in the Matlab function.
The describing function is real and can hence be plotted according to

>> C=1:0.01:10;
>> Yf=2/pi*(asin(1./C)+1./C.*sqrt(1-C.^(-2)));
>> plot(C,Yf)

which gives

164
 1

0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1
1 2 3 4 5 6 7 8 9 10

By zooming in one find that Yf = 0.2 for C = 6.35

(b) An example of a Simulink model of the control system is shown in

the figure below. A step with small amplitude is sufficient to start the
oscillations. The simulation results agree very well with the theoretical
values from a).

10

Step s3 +2s2 +s
Add Saturation Transfer Fcn Scope

165
14.2

The Nyquist curve and the describing function are plotted below

1.5

0.5
Im

A B
0

−0.5

−1

−1.5

−2
−2 −1 0 1 2

Re

The describing function changes direction in a point A, which corresponds

to when Yf (C) takes it maximum value. The corresponding value of C can
be found by differentiating Yf (C) with respect to C. Differentiation of
4H p
Yf (C) = 1 − D 2 /C 2
πC
√
with respect to C gives that the derivative is zero for C = 2D and that
A = − πD2H
. A possible intersection occurs when the Nyquist curve crosses
the negative real axis. We have that arg G(iω) = −π when ω = 1, and
|G(i1)| = 1/2. The point B is thus −1/2. That the oscillation barely
√ can exist
means that B ≈√A. The amplitude of √ the oscillation is 2.5 yields 2D = 2.5.
Hence, D = 5 · 2/4 and H = π · 5 · 2/4. The frequency of the oscillation
is ω = 1.

14.3

(a) The describing function of the relay is

Yf (C) = 4/(πC), ⇒ −1/Yf (C) = −πC/4

166
 The curve −1/Yf (C) covers the entire negative real axis. The frequency
response is
K K(1 − iω)2 · (−iω)
G(iω) = =
iω(iω + 1)2 ω 2(1 + ω 2 )2
K(1 − ω 2 − 2iω)(−iω) −2Kω − iK(1 − ω 2)
= =
ω 2(1 + ω 2 )2 ω(1 + ω 2 )2
arg G(iω) = arg(K) − arg(iω) − 2 arg(1 + iω)
= 0 − π/2 − 2atan(ω)
Nyquist curve
4

1 . .
. . . .
.
. .
. .
. .
. .
. .
Im

0 .
. .
.
. .
. .
. .
. .
. . .
-1 . . . .

-2

-3

-4
-4 -3 -2 -1 0 1 2 3 4

Re

We will always have an oscillation as the Nyquist curve intersects with

−1/Yf (C) for all values of K.
(b) At the intersection point we have that arg G(iω) = −π, or alternatively
that the imaginary part is 0. This occurs for ω = 1. As |G(i1)| = K/2
the amplitude of the oscillation is given by
K πC
−=−
2 4
The requirement that C < 0.1 results in K < π/20.
(c) With a possibly dynamic feedback L(s), the phase of the linear loop-
gain will be arg L(iω)G(iω) = arg L(iω) + arg G(iω). A controller yiel-
ding a phase lead (positive phase, rotating the Nyquist curve counter-
clockwise) at ω ≥ 1 will thus allow us to use an increased gain K. One

167
 such controller is a PD-controller 1 + TD s which will have the phase
atan(TD ) at ω = 1. Note that the phase of L(iω)G(iω) now asymptoti-
cally tends to −π instead of −3π/2 when w → ∞, and for sufficiently
large TD the Nyquist curve does not even cross the real axis.

>> s = tf(’s’);
>> G = 1/(s*(1+s)^2);
>> L1 = 1;L2 = 1 + 0.1*s; L3 = 1 + 0.25*s;L4 = 1+2*s;
>> nyquist(L1*G,L2*G,L3*G,L4*G);
>> axis([-1 0 -1 1]);
>> figure
>> bode(L1*G,L2*G,L3*G,L4*G);

14.4

The describing function of an ideal relay:

4 1 π
Yf (C) = ⇒ − =− ·C
π·C Yf (C) 4

(a) Plot the Nyquist curve of G(s)H(s) = G(s)

1 −i(1 − iω)(2 − iω)

G(iω) = =
iω(iω + 1)(iω + 2) ω(ω 2 + 1)(ω 2 + 4)
3 2 − ω2
= − 2 − i
(ω + 1)(ω 2 + 4) ω(ω 2 + 1)(ω 2 + 4)

168
 Nyquist curve
2

1.5

1 . . . . .
. .
. .
. .
0.5 . .
. .
. .
Im

0 . .
. .
. .
-0.5 . .
. .
. .
. .
. . . .
-1 .

-1.5

-2
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2

Re

If the point −1/YN (C) is encircled by the Nyquist curve the amplitude
of the oscillation will increase and otherwise it will decrease. This results
in a stable oscillation. The frequency and amplitude can be determined
from the intersection
√ of the curves
√ which occurs when Im G(iω) = 0,
i.e. when ω = 2. As Re G(i 2) = −1/6, we get
πC 2
−1/6 = − ⇒ C=
4 3π
Hence, the oscillation has the amplitude 2/(3π) and the frequency ω =
√
2.

(b) Study G(iω)H(iω)

−i(1 − iω)(2 − iω)(1 + Kiω)
G(iω) · H(iω) =
ω(ω 2 + 1)(ω 2 + 4)
−3 + 2K − Kω 2 −2 + ω 2 − 3Kω 2
= + i
(ω 2 + 1)(ω 2 + 4) ω(ω 2 + 1)(ω 2 + 4)
According to (a), we will avoid oscillations if Im G(iω)H(iω) < 0, ∀ω.

−2 + ω 2 − 3Kω 2 < 0 ⇒
ω2 − 2
K>
3ω 2
As (ω 2 − 2)/(3ω 2 ) < 1/3, ∀ω we can choose any K > 1/3.

169
 Nyquist curve
0.4

0.3

0.2

0.1
Im

-0.1

-0.2

-0.3

-0.4
-0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4

Re

14.5

(a) Alternative (i): The describing function of a relay with hysteresis is

given by
4 p 2

Yf (C) = 1 − 1/(2C) − i/(2C) , C ≥ 0.5
πC
πC p π
−1/Yf (C) = − 1 − 1/(2C)2 − i
4 8
which means that the imaginary part of −1/Yf will be −π/8 indepen-
dent of C and the real part will start at zero and tend to −∞.
The transfer function of the linear part is
1
G(s) =
s(s + 1)
Since the transfer function contains an integrator the argument will
start at −90◦ and since the relative degree is two the argument will
tend to −180◦ . This indicates that there will be an intersection between
G(iω) and −1/Yf . From the transfer function we have
1 −w − i
G(iω) = =
iω(1 + iω) ω(1 + ω 2 )

170
 Putting the real and imaginary parts of G(iω) and −1/Yf equal to each
other gives

1 π
=
ω(1 + ω 2 ) 8
πC p 1
1 − 1/(2C)2 =
4 1 + ω2
The first equation has the approximate solution ω = 1.125, which in-
serted in the second equation implies the solution C = 0.75.
Plot the Nyquist curve and the describing function.

1.5

0.5
Im

−0.5

−1

−1.5

−2
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2

Re

The curves intersect when ω = 1.235, C = 0.75. This result can be

found by looking at the plot or by solving the system of equations
C small ⇒ −1/Yf (C) is encircled ⇒ the amplitude of the oscillation
increases
C large ⇒ −1/Yf (C) is not encircled ⇒ the amplitude of the oscillation
decreases.
Thus, the oscillation is stable.

(b) Build a model in Simulink and verify the result.

(c) x1 = θ, x2 = θ̇ yield
(
ẋ1 = x2 1, x1 < −0.5
, u=
ẋ2 = −x2 + u −1, x1 > 0.5

171
 2

1.5

0.5
x2

−0.5

−1

−1.5

−2
−1.5 −1 −0.5 0 0.5 1 1.5
x1

172
14.6

Inserting the numerical values for the PID coefficients gives the transfer fun-
ction
s2 + 2s + 1
G(s) =
s3
for the controller together with the motor. Evaluating G for s = iω gives

−2ω + i(1 − ω 2 )
G(iω) =
ω3
It follows that G crosses the negative real axis at ω = ±1 with G(i) = −2.
A plot of the Nyquist curve is given below.

Nyquist Diagram

1.5

0.5
Imaginary Axis

−0.5

−1

−1.5

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

(a) Since the point −1 is not encircled by the Nyquist curve the closed loop
system is asymptoticaldly stable when the amplifier is linear.

(b) The describing function for the saturation is

r !
2 1 1 1
Yf = arcsin + 1− 2
π C C C

The condition GYf = −1 gives Yf = 0.5 which in turn gives C ≈ 2.5.

For values of C less than ≈ 2.5 the point −1/Yf (C) is not encircled
so the amplitude ought to decrease, while for values of C greater than
≈ 2.5 the point −1/Yf (C) is encircled which indicates an increasing
amplitude. The oscillation with ω = 1 and C ≈ 2.5 therefore probably

173
 has un unstable amplitude. This is confirmed by simulation. Below
the output of the linear part is plotted for different initial amplitudes,
showing a decreasing and an increasing oscillation.
3 4

3
2

2
1

1
0
y

y
0

−1
−1

−2
−2

−3 −3
0 2 4 6 8 10 0 2 4 6 8 10
t t

It is clear that the control system will work well as long as there is
no disturbance large enough to start an oscillation with an amplitude
above the critical limit. (The growing oscillations that are created by
large disturbances can be seen as a windup phenomenon of the integ-
rator part of the regulator. When controlling a double integrator using
a PID controller it is therefore very important to have some form of
anti-windup compensation of the integral part.)

14.7

The describing function is real. The Bode diagram shows that arg GO (iω) =
−180◦ and | GO (iω) |= 2 at ω = 2. This implies that the Nyquist curve
crosses the negative real axis in the point −2 for ω = 2. We hence have to
solve the equation
−1
−2 =
Yf (C)
which implies r
4 1 1
Yf (C) = 1− 2
=
πC C 2
and
64 2 64
C4 − C + 2 =0
π2 π
This gives
+ +
C =(−) 2.29 resp (−) 1.11
By inserting some values of C one realizes that Yf (C) looks like the figure

174
 0.8

0.6

0.4

0.2

0
0 2 4 6 8 10 12 14 16 18 20

The function −1/Yf (C) thus moves along the real axis from −∞ towards 0
when C increases, but stops at rougly −1/.6 and starts moving back towards
−∞. Hence, the curve −1/Yf (C) will intersect the Nyquist curve twice, as
the computations indicate.

Im

Re
-2 -1

Analysis of the two candidate solutions gives

(I). C=1.11. For fixed amplitudes smaller than this value, when we think of
the nonlinearity as a static gain with gain Yf (C), the point −1/Yf (C) will
act as the point −1 in linear stability analysis, and tells us that the closed-
loop system in a linear analysis would be asymptotically stable since it is
not encircled. That means that any oscillation would decay, and C would
not be constant as assumed. Instead it must decrease, and a new thought
fixed value of C would once again indicate asymptotic stability. Hence, if
initial oscillations are small, we suspect they will die out. (The relay here
has a dead-zone which zeroes out everything between −1 and 1 so the result
is reasonable, as a sinusoidal with amplitude 1.1 will almost completely be

175
zeroed out and almost no energy enters the system. If the open-loop system
G0 is stable it is reasonable that the output will go to zero if the input almost
always is zero)

Im

Re
minskande ökande
c
c

(II). C=2.29. For fixed amplitudes larger than this value, when we think of
the nonlinearity as a static gain with gain Yf (C), the point −1/Yf (C) will
act as the point −1 in linear stability analysis, and tells us that the closed-
loop system in a linear analysis would be asymptotically stable. That means
that any oscillation would decay, and C would not be constant as assumed.
Instead it would decrease, and a new thought fixed value of C would once
again indicate asymptotic stability and C would decrease. However, if it
decreases below 2.29, the point −1/Yf (C) is encircled by the Nyquist curve,
and linear analysis tells us the system would be unstable and C would have
to increase. At 2.29, we reach a stationary case were we neither increase nor
decrease C according to linear theory, and we should suspect we will have
oscillations with this amplitude. The limit cycle will have amplitude C = 2.29

Im

Re
ökande
minskande c
c

and angular frequency ω = 2.

176
17 To Compensate Exactly for Nonlinearities

17.1

If we let
u = r − cos x1

we get a linear closed-loop system.

17.2

The control signal

u = −y 4 + y 2 + r = −x41 + x21 + r

results in an exact feedback linearization.

17.3

The system is defined by


 ẋ1 = x21 + x2
(∗) ẋ2 = u

y = x1

Make the change of variables:


z1 = y
z2 = ẏ

This results in
ż1 = ẏ = z2

177
 d
ż2 = ÿ = (ẋ1 )
dt
d 2

= x1 + x2
dt
= 2x1 ẋ1 + ẋ2 = [according to (∗)]


= 2x1 x21 + x2 + u
 
(∗) ⇒ x2 = ẋ1 − x21
=  x1 = y = z1 
ẋ1 = ẏ = z2


= 2z1 z12 + z2 − z12 + u
= 2z1 z2 + u = α(z) + β(z)u

An exact feedback linearization results from

−α(z) + ū
u= = −2z1 z2 + ū.
β(z)

 2 
r ū u x1 + x2 x y
✲ +✐ ✲ +✐ ✲ ẋ = u ✲ T (x) ✲
✻ ✻ y = x1

z
−2z1 z2 ✛ q

−L ✛

17.4

As ẋ1 depends on u we cannot choose y to be x1 . Hence, choose y = x2 .

√ √
ẏ = ẋ2 = 1 + x1 − 1 + x2
d √ √

ÿ = ẍ2 = 1 + x1 − 1 + x2 =
dt
1 1
= √ ẋ1 − √ ẋ2 = · · · =
2 1 + x1 2 1 + x2

178
  √ 
1 1 1 + x1 u
= √ −√ + √
2 1 + x1 1 + x2 2 1 + x1

Thus, the relative degree is 2. Now, do the change of variables z1 = y, z2 =

ẏ ⇒

ż1 = z2
 √ 
1 1 1 + x1 u
ż2 = /from above/ = √ −√ + √ =
2 1 + x1 1 + x2 2 1 + x1
 √ 
1 1 z2 + 1 + z1 1 1
= √ − √ + √ u=
2 z2 + 1 + z1 1 + z1 2 z2 + 1 + z1
= α(z) + β(z)u
1
Choose u = β(z) (ū − α(z)) to get an exact feedback linearization. What are
the poles of the system?

17.5

(a) The force is

mÿ = F − k(y) − d(ẏ)
Do the change of variables x1 = y, x2 = ẏ and x3 = F which results in
the state-space form

ẋ1 = x2
1
ẋ2 = (−k(x1 ) − d(x2 ) + x3 )
m
ẋ3 = −x3 + u
y = x1

(b) Relative degree ν? Differentiate y with respect to time

ẏ = ẋ1 = x2
1
ÿ = ẋ2 = (−k(x1 ) − d(x2 ) + x3 )
m
1
y (3) = (−k ′ (x1 )ẋ1 − d′ (x2 )ẋ2 + ẋ3 )
m
1
= (−k ′ (x1 )ẋ1 − d′ (x2 )ẋ2 − x3 + u)
m

179
As ν = n = 3 we can make an exact feedback linearization. Make the
change of variables
 
 z1 = y  x1 = z1
z2 = ẏ ⇔ x2 = z2
 
z3 = ÿ x3 = k(z1 ) + d(z2 ) + mz3

which results in the state-space form

ż1 = z2
ż2 = z3
1
ż3 = (−k ′ (z1 )z2 − d′ (z2 )z3 − k(z1 ) − d(z2 ) − mz3 + u)
m
y = z1

The control signal

u = mũ + k ′ (z1 )z2 + d′ (z2 )z3 + k(z1 ) + d(z2 ) + mz3

results in a linear system from ũ y.

180