Lecture 1

Introduction1
This first lecture has two parts. The first part gives an introduction and overview
of the course, starting from two examples of modern control, a DVD-reader and car
dynamics. The second part of the lecture is a brief review of linear input-output
models in continuous time used in the basic course. The concepts of signal norm
and system gain are introduced.

1.1 First example a DVD player

The appearance of cheap sensors, actuators and computing devices opens new application areas for feedback control all the time, even in mass produced consumer
products. The control technology is mostly hidden to the user, but still critical for
operation and performance. A prime example of this is positioning of the pick-up
head in a storage device as a DVD or CD-rom, where the speed of data recovery
is directly correlated to the control performance.
A DVD (Digital Versatile Disk) is a data disk of the same physical size as a
CD. Its use is mostly for video, but also for computer software as a large CD-ROM
disk. The storage technology is in principle the same as for the CD, but improved.
A CD holds about 650 megabytes of data whereas a DVD holds 4.7 gigabytes (for
single layer, single side).

Track

Pit

0.74 m
Figure 1.1

The right picture shows pits forming tracks on the DVD surface.

The disk surface is reflective, so that laser light is reflected back. Data bits
are represented by pits of different lengths in tracks on the disk. These pits make
the laser beam interfere destructively with itself, and therefore the pits look black
to the laser.
The surface velocity is constant (about 3.5 m/s), meaning that the disc should
rotate at different speeds depending on the current reading position. The challenge
of the control problem is related to the fact that only 0.022 m deviations from the
1 Written

by A. Rantzer with contributions by K.J. strm, B. Lincoln and B. Wittenmark

Lecture 1.

Introduction1

bit-track can be accepted. At the same time, a disk is always slightly asymmetric,
causing it to oscillate up to 100 m per rotation, and the rotation speed is up to 23
Hz (for single speed). The tracking controller must compensate for this oscillation.
A typical DVD player has a pick-up-head consisting of a laser, an astigmatic
lens, and a light detector with four fields see Figure 1.2. The lens is mounted
on springs in the axial (focus) and radial direction, and can be moved by electromagnets. This way, the laser spot can be moved very fast in a small range (a few
hundred tracks sideways). The lens and laser are mounted on the sledge, which
can move over the whole disk (in radial direction), but with much less precision
and speed.

Tracks

Radial electromagnet
Lens

Disk
A B
C D

Springs

Light detectors
Focus electromagnet

Laser

Sledge

Pickup head

Figure 1.2 The pick-up-head has two electromagnets for fast positioning of the lens (left).
Larger radial movements are taken care of by the sledge (right).

Four light detectors are available to estimate the focus error and radial error
of the lens. Measurements are taken with a sampling frequency of 40 kHz and
the DVD standard specifies that the speed of control (cross-over frequency) must
be at least 2.4 kHz.
It turns out that most of the main topics of this course are relevant for the
solution of the DVD control problem and we have therefore chosen to use it as a
demonstrator. The focus control will be treated in a lecture, while a lab exercise
towards the end of the course is devoted to disc track following.

Figure 1.3

The DVD reader used in lab 3.

1.2

Second example Control of car dynamics

1.2 Second example Control of car dynamics

A modern car contains numerous micro-processors devoted to feedback control. For
example, feedback from oxygen sensors in the exhaust gas are needed for proper
operation of the engine and catalyzer. This is essential for fuel efficiency and to
reduce the emission of polluting exhaust gases.
Other feedback loops are used to improve safety, by controlling the brakes to
prevent wheel-locks and to prevent skidding on slippery roads. A simplified model
for car dynamics is given by the state space description

 
V
r

V
r

0
b1

(u1 + u2 u3 u4 ) +

b2
b3

where V is lateral speed and r is angular velocity. There are five control signals,
the steering angle and the brake forces u1 , u2 , u3 and u4 on the four wheels.

U
r

Figure 1.4 A modern car relies on feedback control for comfort, safety and fuel efficiency.
The left picture shows a test-car used in a research project together with DaimlerChrysler.

The state is generally not available for direct measurement. Even if the angular
velocity of each wheel can be measured, there is always some discrepancy between
the rotational speed and the speed over ground. Hence the velocity of the car
must be estimated based on information from several sources and the remaining
uncertainty must be taken into account in the control algorithms.
A typical sampling frequency for speed measurements is a few milliseconds.
This may sound fast enough compared to typical car dynamics, but when the
purpose is to prevent wheel-lock or accidents, a delay of a few milliseconds can in
fact be a severe obstacle for proper control performance.
lateral velocity

brake forces
Vehicle
steering angle

Figure 1.5

yaw rate

Input-output diagram for car dynamics control.

Introduction1

Lecture 1.

1.3 Course overview

The objective of the course is that the students should learn how to state and solve
advanced design problems for computer controlled systems. A schematic picture
of such a system is given below.
y (t )

u (t )

u (t )

t
Process

Hold

Sampler
yk

uk
uk

y (t )

D-A

Computer

A-D

yk

The control signal u(t) is determined using measurements of y(t) to achieve

desired behaviour of the process. For the DVD player, y(t) would be a vector of
four variables, representing intensities in the four light detectors in Figure 1.2,
while u(t) would correspond to the two electromagnets.
It is important to note that the dynamics of a real process is never known
exactly. Neither is it possible to precisely state the true design objectives. It is
therefore necessary to maintain a broader perspective on the engineering design
problem. See Figure 1.6.
Matematical model
and
specification

Idea/Purpose

Analysis

Experiment

Synthesis

Implementation

Figure 1.6

Schematic overview of the design process

Everything starts with an idea about the purpose of the control task. In simple
cases, it is possible to directly come up with a solution proposal that can be tested
experimentally and be accepted, possibly after minor modifications. However, in a
vast number of applications costs and time can be reduced by analyzing or simulating a mathematical model before trying real experiments. The purpose of the
diagram is to illustrate this methodology. Note that the arrows point in two directions. Failure in the experimental phase could not only require reimplementation,
but also new analysis, more accurate models, or even redefinition of the control
purpose.
Imagine stepping through the diagram in order to design a controller for car
dynamics as in the previous example. Suppose that a controller has been synthe-

1.4

Linear input-output maps

sized based on the given two state model. Implementing a controller on a prototype
car is costly, so a second step would typically involve computer simulation. For this
purpose a more complex and accurate car model is needed, a model that is less
transparent from a synthesis perspective but better suited to reveal the deficiencies of a proposed controller. If the simulations fail, a reason could be that the
two state model was too simple and that additional features need to be taken
into account in the synthesis phase. After a sequence of attempts, one could hope
to find a solution ready for experimental tests. Alternatively, persisting failures
could be an indication that the original goal was overly optimistic and impossible
to achieve.
The main emphasis of this course is on the analysis/synthesis phase of the
diagram in Figure 1.6. However, to keep the big picture in mind, there will also
be lectures and laboratory sessions devoted to modelling, implementation and
experiments.
The main topics of the course are the following

Design of scalar controllers

Sampled systems
Discrete-time systems
Fundamental system limitations
Multiinput-multioutput systems
Synthesis by optimization
During the first four lectures we keep the continuous time perspective that was
used in the basic course. We build on material from that course, but make a deeper
study of robustness and performance evaluation in controller design. These four
lectures are supported by hand-out notes, but not by the book. The first lab exercise
is aimed to give practical training in scalar controller design by frequency domain
loop shaping.
Starting from lecture five, the textbook is used to introduce the discrete time
perspective of computer control. The theory for discrete time linear systems is
applied to sampled data control. This is the topic of lectures 5-9.
After this, in the fourth week, there is an intermezzo of two more lectures
supported by handout notes. The topic is fundamental limitations in controller
design and we also discuss how the previous ideas can be applied to multivariable
systems (i.e. systems with several inputs and outputs). This is the topic of the
second lab exercise.
The last three weeks of the course, lectures 12-16, continue along the lines
of the textbook and bring in the subject of optimization. The theory of linear
quadratic optimal control and Kalman filtering is a cornerstone of modern control.
It clarifies fundamental relationships between measurement accuracy, control authority and achievable performance. Multivariable systems also fit in very nicely.
Finally, the course is concluded by a lab exercise devoted to track following in the
DVD reader. Most of the main topics in the course are relevant for a successful
solution to this problem.

1.4 Linear input-output maps

In this section, we will review some different ways of specifying the input-output
relationship of a finite-dimensional linear time-invariant system. This is a system
that can be described by a state space equation
x = Ax + Bu
y = Cx + Du

Lecture 1.

Introduction1

The differential equation has the solution formula

y(t) = CeAt x(0) +

CeA(t ) Bu( )d + Du(t)

Note that the formula remains valid for multivariable systems, i.e. when both u(t)
and y(t) are vector valued.
The map from u to y is linear provided that x(0) = 0. Introducing the impulse
response (t) as

(t) =

CeA(t ) B ( )d + D (t) = CeAt B + D (t)

the input-output map can be written as a convolution

y(t) =

(t )u( )d = [ u](t)

In frequency domain, the convolution becomes multiplication

Y ( s) = G ( s) U ( s)
and the Laplace transform of the impulse response is equal to the transfer function
G (s) = C (sI A)1 B + D. For multivariable systems, both (t) and G (s) are
matrices. The term impulse response is of course motivated by the fact that the
matrix element i j (t) is the value of output i obtained when input j is an impulse
(Dirac function) at t = 0. This is sometimes used to determine (t) experimentally.
yi (t) = i j (t)

0.5
0.4
0.3
0.2
0.1
0

0.1
1

10

10

u j (t) = (t)

1
0.8
0.6
0.4
0.2
0
1

A more common experiment in process industry is the step response. Assuming

that the (possibly vector valued) input is a step

u(t) =

t<0

u0

t0

the output becomes

y(t) =

(t s)u0 ds =

Z

t
0

( )d

u0

Accordingly, the Laplace transform of the step response is G (s)u0 / s.

6

1.4

Linear input-output maps

yi (t)

0.8
0.6
0.4
0.2
0
1

10

10

u j (t)

1
0.8
0.6
0.4
0.2
0
1

The main use of the Laplace transform is however to characterize the frequency
response. The input u(t) = u0 sin t gives

y(t) =

( )u(t )d = Im

Z

t
0

( ) ei d ei t u0

The integral approaches G (i ) as t , so
after a transient, also the output

becomes sinusoidal and y(t) = Im G (i ) ei t u0 . To summarize, a linear timeinvariant system always gives a sinusodal response to a sinusodal input. For a
scalar system, the gain and phase shifts are determined by the amplitude and
phase of the complex number G (i ).
0.3

yi (t)

0.2
0.1
0
0.1
0.2

10

12

14

16

18

20

10

12

14

16

18

20

1.5

u j (t)

1
0.5
0

0.5
1
1.5

There are several ways to graphically illustrate the transfer function G (i ).

One is to plot the amplitude and phase separately versus the frequency. This is
called the Bode diagram:

Lecture 1.

Introduction1

Magnitude

10

10

10

Phase

45
90
135
180

10

10

Frequency (rad/sec)

It should be noted that each additional factor in the transfer function contributes
additively to the Bode plots:
log p G1 G2 G3 p = log p G1 p + log p G2 p + log p G3 p
arg G1 G2 G3 = arg G1 + arg G2 + arg G3
The Nyquist diagram is obtained by plotting G (i ) directly in the complex
plane for different values of :
0.6

Im

0.4
0.2

Re
arg G (i )

0
0.2

p G (i )p

0.4
0.6
1

0.5

0.5

If instead log G (i ) = log p G (i )p + i arg G (i ) is plotted in the complex plane,

the Nichols diagram is obtained:
40
0 dB

30

0.25 dB
0.5 dB

20

1 dB

Amplitude

1 dB

3 dB

10

3 dB

6 dB

6 dB

10

12 dB

20

20 dB

30
40
360

40 dB

315

270

225

180

Phase

135

90

45

The level curves of p G /(1 + G )p and arg G /(1 + G ) are plotted as dotted lines to
support use of the diagram in controller design.

1.5

Signal norm and system gain

1.5 Signal norm and system gain

In order to efficiently analyze and optimize dynamical systems, it is useful to have
mathematical notions that measure the size of a signal and the gain of a system.
This is the reason for the following definitions.
The size of a signal y(t) Rn can be measured by the L2 -norm, defined as

q yq2 :=

sZ

p y(t)p2 dt

According to a theorem known as Parsevals formula, the same norm can be defined in frequency domain as

q yq2 =

1
2

pLy(i )p2 d

For a system S with input u, output S (u) and zero initial state, the L2 -gain
is defined as the largest possible fraction between the input norm and the output
norm

qS q := sup
u

qS (u)q
quq

The system is called input-output stable (or L2 -stable) if its L2 -gain is finite. For
example, a time delay does not change the signal norm, so it has gain one. However,
an integrator has infinite gain, since an input u(t) that is identically zero for t 1,
can give an output y(t) that is a nonzero constant for t 1. Hence, the fraction
q yq2 /quq2 can be arbitrarily large.
More generally, the L2 -gain of a system can be obtained as the maximum
amplitude in the Bode diagram:
THEOREM 1.1
A stable system with transfer function G (s) has the L2 -gain

q G q := sup p G (i )p

Remark. For multivariable systems the p G (i )p should be interpreted as the matrix norm (the largest singular value) of G (i ). This case will be studied more
carefully later.
Proof. Let y be the output corresponding to the input u. Then

q yq2 =

1
2

pLy(i )p2 d

1
2

p G (i )p2 pLu(i )p2 d q G q2 quq2

The inequality is arbitrarily tight when u(t) is a sinusoid near the maximizing
frequency.
2
Example 1
a. For a time delay G (s) = esT we have p G (i )p " 1.
b. For an integrator p G (i )p = p i1 p = 1 which is unbounded = 0.
c. The Bode diagram plotted in the previous section has a peak magnitude
about 0.5 at the frequency 2 rad/sec. Hence, the L2 -gain of the corresponding
system is smaller than one and the highest gain is obtained for an input sinusoid
of this frequency.
2

Lecture 2

Stability and Robustness1

This lecture discusses the role of stability in feedback design. The emphasis is
not on yes/no tests for stability, but rather on how to measure the distance to
instability. The small gain theorem is introduced as a means to verify stability in
presence of model uncertainty.

2.1 Stability of feedback systems

Feedback and stability are closely connected issues. On one hand, the introduction
of feedback may potentially create instability. On the other hand, properly applied
feedback is often the best way to get rid of instabilities.
There are several well known examples in the history. One is the construction
of airplanes. The first airplanes in the early 1900s were depending on the pilot to
stabilize the dynamics manually using the control-stick. The modern fighter JASGripen was built unstable for the sake of maneuverability and relies on computer
control for stabilization.

Figure 2.1 Lawrence Sperry demonstrates a stabilizing gyroscopic controller. He waves his
hand in the air, while his mechanic is walking on the wing.

Another striking example was the construction of the first electronic feedback
amplifiers, that were necessary to build long distance telephone connections in
the 1930s. In this case, high gain feedback was needed to reduce the nonlinear
signal distorsion. Stability problems became a major issue, and the development
of frequency domain stability criteria was critical for successful implementation.
A modern example is the maneuver test for Mercesdes A-class, that created
unstable oscillations severe enough to turn the car over. The problem was solved
by introducing electronic feedback control.
Recall from the previous lecture that a system is called input-output stable
(or L2 -stable) if its L2 -gain is bounded. A transfer function is called stable if it
corresponds to an input-output stable system. The following stability criterion is
available for linear time-invariant systems.
THEOREM 2.1
A rational transfer function G (s) is stable if and only if all poles of G have negative
real part. In particular, if G (s) = C (sI A)1 B + D, it is sufficient that all
eigenvalues of A have negative real part.
1 Written

10

by A. Rantzer with contributions by K.J. strm

2.1

Figure 2.2

Stability of feedback systems

The stability problems of Mercedes A-class were solved by electronic feedback

generally, a system with impulse response (t) is stable provided that G (i ) =

RMore
i t
(t)dt is well defined and there is a c with p G (i )p c for all .
0 e
Example 1 Let us determine input-output stability of the following systems


  

1 2
1

x =
+
u
3 2
0

y = [1 1] x + u

( I)

( I I)

y(t) =



e t u( ) y( ) d

( I ) The first system is input-output stable due to stable eigenvalues of the system matrix. A two-by-two matrix like this is stable if and only if the trace is
negative (here 3) and the determinant is positive (here 8). This is because
the trace is the sum of the eigenvalues and the determinant is the product.
In general, eigenvalues can be computed by the matlab command eig(A):
>> eig([-1 2; -3 -2])
ans =
-1.5000 + 2.3979i
-1.5000 - 2.3979i
Note that the coefficients of the characteristic polynomial should not be computed, at least for high order systems, since this generally leads to numerical
difficulties.

( I I ) The input-output relationship can be written

y = (u y)
where (t) = et , t 0. After Laplace transformation, this gives
1
[ U (s) Y (s)]
s+1
1
Y ( s) =
U ( s)
s+2

Y ( s) =

The transfer function 1/(s + 2) shows that the system is stable, since the
only pole 2 is negative.

2
Our main objective is to study stability of feedback loops. From the basic course,
we recall the Nyquist criterion, which supports understanding by graphical illustrations.

11

Lecture 2.

Stability and Robustness1

1
Am
0.5

G ( s)

Im

- n 6

1 

0.5
c

1
1

0.5

0
Re

0.5

Figure 2.3 The closed loop system remains stable as long as the Nyquist diagram does
not encircle 1. The amplitude margin Am and phase margin m measure the distance from
instability.

THEOREM 2.2THE NYQUIST CRITERION

Suppose that G (s) is stable and that the Nyquist plot G (i ), R does not
encircle 1. Then (1 + G (s))1 is also stable.
The following example illustrates the use of the theorem.
Example 2 As motivating example, consider a position control in a mechanical
system with damping coefficient c. The controller contains a time delay:
x + c x + x = u

u(t) = k[ x(t T ) + x (t T )]

The feedback loop is illustrated in the figure below. Nominal values of the parameters are k = 1, c = 1 and T = 0. Let us investigate how much margin there is in
each of the parameters before the system becomes unstable?

- h
6

k(s+1)
s2 +cs+1

esT

1 
For this purpose, we plot the Nyquist and Bode digrams of the nominal transfer
function (s + 1)/(s2 + s + 1). The Matlab command margin gives numerical values
for the amplitude- and phase-margins.
Gm = Inf, Pm = 109.47 deg (at 1.4142 rad/sec)

Nyquist diagram

Magnitude

1.5

Im

0.5

0.5

0
0
45

0.5

Re

1.5

Phase

1.5
0

90
135
180

Frequency
10

Figure 2.4

12

Nyquist and bode plots for the nominal transfer function (s + 1)/(s2 + s + 1)

10

2.2

Sensitivity

For k = c = 1 the open loop transfer function is

s2

s+1
esT
+s+1

For small values of T the Nyquist plot will not encircle 1, so the systems remains
stable. To find out exactly how large values of T are needed for instability, note that
the phase margin 109 degrees (or 109 /180 radians) is obtained at the frequency
1.41 rad/sec. Hence a time delay of
109
= 1.35 seconds
180 1.41
can be tolerated for k = c = 1.
To investigate the robustness to variations in the parameters c and k, we note
that the closed loop characteristic polynomial for T = 0 is

(s2 + cs + 1) + k(s + 1) = s2 + (c + k)s + 1 + k

The stability condition for a second degree polynomial is positive coefficients, so
stability is maintained as long as c + k > 0 and 1 + k > 0.
2
Having analyzed the example with respect to parametric uncertainty above,
it is natural to ask about robustness to unmodelled dynamics. For example, this
would be relevant to capture the difference between a rigid body and an elastic
one. This is also the subject for the remaining part of the lecture.

2.2 Sensitivity
Two transfer functions are of particular interest in the study of the feedback loop
below.
n

d
e

u
C ( s)

P ( s)

Figure 2.5

A simple control loop

These are
1
1 + C ( s) P ( s)
C ( s) P ( s)
T ( s) =
1 + C ( s) P ( s)
S ( s) =

(the sensitivity function)

(the complementary sensitivity function)

The term complementary refers to the fact that

S ( s) + T ( s) " 1
Note that T (s) is the transfer function from reference signal r to the the plant
output x. Another name for T (s) is therefore the closed loop transfer function of
13

Lecture 2.

Stability and Robustness1

Im





1


= sup 
1 + C (i ) P(i ) 

Re

C (i ) P(i )
Figure 2.6 The L2 -gain of the sensitivity function is the inverse of the distance from the
Nyquist plot to 1.

the system, while C (s) P(s) is called the open loop transfer function or just the
loop transfer function.
The name sensitivity function refers to the fact that S measures how small
relative errors in P are mapped into relative errors in T. This is verified by a
simple calculation:
d
dT
=
dP
dP


1

1
1 + CP

C
TS
=
(1 + CP)2
P

dT / T
=S
dP/ P

Notice that T is a nonlinear function of P. Hence the sensitivity calculation

only says something about the response to small pertubations in P. For larger
perturbations, the system could have a drastically different behaviour and even
become unstable.
Given the Nyquist criterion, it is natural to conjecture that the robustness to
unmodelled dynamics should somehow be related to the distance from the Nyquist
plot to the point 1. It is therefore striking to note that the L2 -gain of the sensitivity function turns out to be exactly the inverse of this distance. Consequences
are investigated in the next section.

2.3 Robustness via the small gain theorem

r1

e1

S1

e2

r2

S2
In this section, we will investigate stability robustness using the following
theorem, based on the notion of input-output L2 -gain. For simplicity, calculations
are done assuming zero initial conditions.
THEOREM 2.3THE SMALL GAIN THEOREM
Assume that S 1 and S 2 are input-output stable systems with L2 -gain qS 1 q and
qS 2 q. If qS 1 q qS 2 q < 1, then the L2 -gain from (r1 , r2 ) to ( e1 , e2 ) in the closed loop
14

2.3

Robustness via the small gain theorem

system
e1 = S 2 ( e2 ) + r1
e2 = S 1 ( e1 ) + r2
is finite.
Proof. Define q yqT =

qR
T
0

p y(t)p2 dt. Then qS ( y)qT qS q q yqT .

e1 = r1 + S 2 (r2 + S 1 ( e1 ))


q e1 qT qr1 qT + qS 2 q qr2 qT + qS 1 q q e1 qT

q e1 qT

q r 1 q T + qS 2 q q r 2 q T
1 qS 1 q qS 2 q

Bounded gain from (r1 , r2 ) to e1 follows as T . The gain to e2 is bounded in

the same way.
2
To demonstrate how the small gain theorem can be used for robustness analysis, consider the feedback loop in Figure 2.3, where the plant P(s) has been
replaced by [1 + (s)] P(s).
v

C ( s)

(s)

-?
f

P ( s)

Figure 2.7

Loop diagram with perturbed plant [1 + (s)] P(s)

The transfer function from w to v is equal to T (s), the complementary sensitivity function. Hence, by the small gain theorem, the feedback system remains
stable as long as

qq qT q < 1
Note that the small gain theorem does not assume linearity or time-invariance.
Hence the closed loop system will remain stable for all plants of the form P(s)[1 +
(s)] where has L2 -gain smaller than [sup pT (i )p]1 , even for that are
nonlinear or time-varying.
As a second example, let us derive a stability criterion for the case that the
perturbation appears additively, i.e. P(s) is replaced by P(s) + (s).
Then the transfer function from w to v is equal to C (s) S(s), so the small gain
theorem shows stability for all perturbations satisfying

qq q CSq < 1
For linear time-invariant perturbations, this criterion can be nicely illustrated in
the Nyquist diagram. Clearly, the condition

p C p < p1 + PC p
guarantees that the Nyquist plot of ( P + ) C does not encircle 1.
15

Lecture 2.

Stability and Robustness1

C ( s)

(s)

P ( s)

-?
f

1 

Figure 2.8

Loop diagram with perturbed plant P(s) + (s)

1 + PC

16

Lecture 3

Scalar control synthesis1

The lectures reviews the main aspects in synthesis of scalar feedback systems.
Another name for such systems is single-input-single-output (SISO) systems. The
specifications include ability to follow reference signals, to attenuate load disturbances and measurement noise and to reduce the effects of process variations. In
the presentation, we separate the solution into feedback control and feedforward
control.

3.1 Specifications
Recall from Lecture 1 the illustration of the design process shown in Figure 3.1.
While Lecture 2 was mainly concerned with analysis, we are now focusing on the
three neighboring blocks: Specification, Analysis and Synthesis.

Matematical model
and
specification

Idea/Purpose

Analysis

Experiment

Synthesis

Implementation

Figure 3.1

Schematic overview of the design process

We will restrict attention to the following structure (Figure 3.2), with a scalar
transfer function for the plant. This setup was studied in the basic course and is
sufficient for many practical situations.
The controller consists of two transfer functions, the feedback part C (s) and
the feedforward part F (s). The control objective is to keep the process output x
close to the reference signal r, in spite of load disturbances d. The measurement
y is corrupted by noise n.
1 Written

by A. Rantzer with contributions by K.J. strm

17

Lecture 3.

Scalar control synthesis1

n

d
e

F ( s)

u
C ( s)

P ( s)

1
Controller

Figure 3.2

Process
A Controller with two degrees of freedom

Several types of specifications could be relevant for this control loop.

A: Reduce the effects of load disturbances
B: Control the effects of measurement noise
C: Reduce sensitivity to process variations
D: Make the output follow command signals
A useful synthesis approach is to first design C (s) to meet the specifications
A, B, and C, then design F (s), to deal with the response to reference changes, D.
However, the two steps are not completely independent: A poor feedback design
will have a negative influence also on the response to reference signals.
The following relations hold between the Laplace transforms of the signals in
the closed loop system.
PC
P
PC F
R( s )
N ( s) +
D ( s)
1 + PC
1 + PC
1 + PC
CF
C
1
V ( s) =
R( s )
N ( s) +
D ( s)
1 + PC
1 + PC
1 + PC
PC F
1
P
Y ( s) =
R( s ) +
N ( s) +
D ( s)
1 + PC
1 + PC
1 + PC

X ( s) =

Several observations can be made:

The signals in the feedback loop are characterized by four transfer functions
(sometimes called The Gang of Four)
1
1 + P ( s) C ( s)

P ( s)
1 + P ( s) C ( s)

C ( s)
1 + P ( s) C ( s)

P ( s) C ( s)
1 + P ( s) C ( s)

In particular, we recognize the first one as the sensitivity function and the
last one as the complementary sensitivity.

The total system with a controller having two degrees of freedom is characterized by six transfer functions (The Gang of Six).
To fully understand the properties of the closed loop system, it is necessary to look
at all the transfer functions. It can be strongly misleading to only show properties
of a few input-output maps, for example a step response from reference signal to
process output. This is a common mistake in the literature.
The properties of the different transfer functions can be illustrated in several
ways, by time- or frequency-responses. For a particular example, we show below
first the six frequency response amplitudes, then the corresponding six step responses.
It is worthwhile to compare the frequency plots and the step responses and to
relate their shape to the specifications A-D:

18

3.1
PC F /(1 + PC )
0

10

10

P/(1 + PC )

PC /(1 + PC )

10

10

10

10

C F /(1 + PC )

10

10

Specifications

10

10

C /(1 + PC )

10

10

10
1

10

1/(1 + PC )

10

10
1

10

10

10

10

10
1

10

10

10

10

10

10

10

10

10

Figure 3.3 Frequency response amplitudes for P(s) = (s + 1)4 , C(s) = 0.775(s1 /2.05 + 1)
when F (s) is designed to give PCF /(1 + PC) = (0.5s + 1)4

PC F /(1 + PC )

P/(1 + PC )

PC /(1 + PC )

1.5

1.5

1.5

0.5

0.5

0.5

10

20

C F /(1 + PC )

30

10

20

C /(1 + PC )

30

1.5

1.5

1.5

0.5

0.5

0.5

10

20

30

10

20

30

10

20

30

10

20

30

1/(1 + PC )

Figure 3.4 Step responses for P(s) = (s + 1)4 , C(s) = 0.775(s1 /2.05 + 1) when F (s) is
designed to give PCF /(1 + PC) = (0.5s + 1)4

Disturbance rejection The two upper right plots show the effect of the disturbance d in process output x and input v respectively. The resulting process error
should not be too large and should settle to zero quickly enough. The control input
would cancel the disturbance exactly if the mid upper step response would be an
ideal step. In a short time-scale this is impossible, since the control input will not
change until the effect of the disturbance has appeared in the process output and
been available for measurement. However, slow disturbances should normally be
cancelled by u. Equivalently, the sensitivity function 1/(1 + PC ) should be small
for low frequencies. This specification is usually corresponds to an integrator in
the controller.
Supression of measurement noise The second specification was to limit the
effect of measurement noise, typically a high frequency phenomenon. The mid upper frequency plot shows good attenuation of measurement noise above the cut
off frequency of 1 Hz. In this example, this is mainly an effect of the process dynamics. A more interesting question is maybe the gain from measurement noise
to control input, since fast oscillations in the control actuator are usually undesir-

19

Lecture 3.

Scalar control synthesis1

able. For this aspect, the mid lower frequency plot, showing the Bode amplitude
from n to v, is of interest.
Robustness to process variations As shown in the previous lecture, the robustness to process variations is determined by the sensitivity functions. In this
example, the lower right frequency plot has a maximal value of 2, which shows
that a small relative error in the process can give rise to a relative error of double
size in the closed loop transfer function. The maximal amplitude of the frequency
plot for the complementary sensitivity function is 1.35, so the small gain theorem
proves stability of the closed loop system as long as the relative error in the process model is below 74% = 1/1.35. In fact, most process models are inaccurate at
high frequencies, so the complementary sensitivity function PC /(1 + PC ) should
be small for high frequencies.
Command response The upper left corner plot shows the map from reference
signal r to process output x. Using the prefilter F, it is possible to get a better
step response here than in the upper mid plot. The prize to pay is that the corresponding response in the control signal gets higher amplitude. This can be seen
by comparing the lower left plot, showing the map from r to v, to the lower mid
plot, which shows the corresponding map when F " 1.

3.2 Loop shaping

The closed loop performance depends critically on the loop transfer function
L( s ) = P ( s ) C ( s )
In particular, the sensistivity functions can be written as S = (1 + L)1 and
T = L(1 + L)1 respectively. A popular approach to control synthesis, known
as loop shaping, is to focus on the shape of the loop transfer function and keep
modifying C (s) until the desired shape is obtained.
Recall that proper disturbance rejection requires small sensitivity S (large L)
for for small frequencies, while process uncertainty requires the complementary
sensitivity function to be small (small L) for high frequencies. On the other hand,
if the amplitude of L decreases very rapidly, the phase tends to become lower
than 180 and the system becomes unstable. Loop shaping is therefore a tradeoff between different kinds of specifications.
Many control problems can be adequately solved by PID controllers, which can
be viewed a combination of one lag compensator and one lead compensator. For
more advanced applications, like resonant systems, higher order controllers are
desirable. An example of such a system is the flexible servo treated in lab 1.
Graphical illustrations in Bode- or Nichols- diagrams are typically used to
support the design. These diagrams are convenient because of the logarithmic
scale, where the controller contributes additively to the loop transfer function:
log p L(i )p = log p P(i )p + log p C (i )p
arg p L(i )p = arg P(i ) + arg C (i )
From the basic course, recall the following essential properties of lead and lag
compensators, illustrated in Figure 3.5:
Lag compensator

Increases low frequency gain: Can be used to reduce stationary errors

Decreases phase, which may reduce stability margins

20

Phase

Magnitude

3.2

20

20

10

10

0
0

0
60

30

30

60

10

Figure 3.5
(right)

10

10

0 2
10
s+10
s+1

Bode diagrams for lag compensator

Loop shaping

10

10

10

(left) and lead compensator

10s+1
s+1

Lead compensator

Increases high frequency gain: Can be used for faster closed loop response
Increases phase, which may improve stability margins

Phase

Magnitude

Loop shaping design of high order controllers will be exercised in lab 1. We will
first design a controller C1 (s) for low frequencies, then keep adding compensator
links C2 (s), C3 (s), . . . to modify the dynamics at higher and higher frequencies
until a satisfactory controller C (s) = C1 (s) Cm (s) is obtained. Lead/lag links
are often sufficient, but occasionally it is useful to also consider controllers with
poles or zeros outside the real axis. The figure below shows the Bode diagram for
cases with stable complex zeros (left) and complex poles (right).
0

30

10

20

20

10

30
90

0
90

45

45

45

45

90

10

Figure 3.6

Notch compensator

10
s2 +0.1s+1
(s+1)2

90

10

(left) and resonant compensator

10
(s+1)2
s2 +0.1s+1

(right)

21

Lecture 3.

Scalar control synthesis1

3.3 Feedforward synthesis

Let us finally consider the design of F (s) to shape the response to reference signals.
For this purpose, it is useful to also consider the second diagram below.
r

Mu

My

um
ym

u
C

1
The two configurations are mathematically equivalent provided that
C F = Mu + CMy
The transfer functions Mu and My can be viewed as generators of the desired
output ym and the corresponding input um .
The transfer function from r to the error signal e = ym y is ( My PMu ) S.
The error is zero provided that
Mu = My / P
Notice that this condition does not depend on C! Since Mu = My / P should be
stable, causal and not include derivatives we find that

Unstable process zeros must be zeros of My

Time delays of the process must be time delays of My
The pole excess of My must be greater than the pole excess of P
Apparently all three conditions put limitations on the achievable command response.
Example 1 If
P ( s) =

1
(s + 1)4

My (s) =

1
(sT + 1)4

then
Mu (s) =

(s + 1)4
My (s)
=
P ( s)
(sT + 1)4

Mu ()
1
= 4
Mu (0)
T

Fast response (T small) requires high gain of Mu . Bounds on the control signal
therefore limit how fast response we can obtain.
2

22

Lecture 4

DVD focus control1

Imagine the following: You are traveling at half the speed of light, along a line
from which you may only deviate 1 m. The line is not straight but oscillates up to
4.5 km sideways, 23 times per second.
This is a scaled version of the control task in a DVD player, where the pick-up
head needs to follow the bit-track. The real numbers are 3.5 m/s, with maximally
0.022 m deviations from the track. A disk is always slightly asymmetric, causing
it to oscillate up to 100 m per rotation, and the rotation speed is up to 23 Hz
(for single speed).

Track

Disk

Pit

Sledge

Pickup head

trackpitch
Figure 4.1 Pits forming tracks on DVD surface (left). Larger radial movements are taken
care of by the sledge (right).

The surface velocity is constant (about 3.5 m/s), meaning that the disc should
rotate at different speeds depending on the current reading position.

4.1 The DVD player

A typical DVD player has a pick-up-head consisting of a laser, an astigmatic lens,
and a light detector with four fields. See Figure 4.1.
The lens is mounted on springs in the axial (focus) and radial direction, and
can be moved by electromagnets. This way, the laser spot can be moved very fast
in a small range (a few hundred tracks sideways). The lens and laser are mounted
on the sledge, which can move over the whole disk (in radial direction), but with
much less precision and speed. As the disk rotates, the track moves both radially
and axially because of asymmetries, so feedback control is needed for both the
radial and the axial position of the lens.
Light is emitted by the laser, focused through the lens onto the disk surface.
The disk surface is reflective, so that laser light is reflected back. Data bits are
1 This

text is based on an original manuscript by Bo Lincoln

23

Lecture 4.

DVD focus control

represented by pits of different lengths in tracks on the disk. The image of the
surface is reflected back through the lens and read by a set of four photo detectors.
These pits make the laser beam interfere destructively with itself, and therefore
the pits look black to the detector.

Tracks

Radial electromagnet

A B
C D

Lens

A B
C D

Springs

Light detectors
Focus electromagnet

Laser

Figure 4.2 The pick-up-head has two electromagnets for fast positioning of the lens (left).
The four photo detectors A D (right).

The photo detectors

The photo detectors sense the reflected laser light from the disk. It can be thought
of as a very simple camera with only four pixels (see Figure 4.2). This information
is enough to sense position radially (are we to the left or right of the track?),
axially (thanks to the astigmatic lens) as well as the current bit. There are no
other sensors in the pick-up head to help keep the laser in the track.
Focus error
The pick-up lens is astigmatic diagonally, so bad focus results in brighter light in
either detectors A + D (too high) or B + C (too low). Therefore, the focus error
(FE) can be calculated as
F E = A + D ( B + C)
which gives a surprisingly good result. Sweeping the focus lens from low to high
usually results in an F E as in Figure 4.3. There is a linear range in the center
where the F E signal is useful to control the lens around the correct focus. The
slope of the curve depends on the reflectivity of the disc.
Possible to see tracks

Focus Error

Correct focus

Lens height
Too low

Figure 4.3

24

Too high

Sweeping the lens axially from low to high results in a curve like this.

4.1

The DVD player

Radial error
The radial error ( RE), meaning sideways deviation from the track center, can be
measured in two ways using the photo detectors. The simple way is to let
RE = A + C ( B + D ),
i.e., use the difference in light from the left and right pair of detectors. For example,
if the reflected light is brighter to the left, the radial error is positive, and we
should move right. This measurement method is called radial push-pull (PP). See
Figure 4.4.

A B
C D

Pit

Figure 4.4

Push-pull: Calculating radial error as RE = A + C ( B + D )..

There is a second measurement method (DPD), which usually gives better result,
but requires the disk to have pits. This is not true for a non-written DVD-R
(writable), for example. The signals f1 = A + D and f2 = B + C are created,
and phase compared (see Figure 4.5). For example, if f1 comes before f2 the lens
is too far to the right. The time difference forms the error signal RE.

A B
C D

f1
f2
Radial error
Pit

Figure 4.5

DPD: Radial error from phase difference between f1 and f2 .

According to the DVD specification, DPD should be used whenever possible. Sweeping the lens radially over the disk creates an RE as in Figure 4.6 for PP and DPD.
1 track

DPD

Correct radial position

PP

Radial position

Figure 4.6 DPD- and PP-signals as the disk rotates and the lens is swept over the tracks
radially. As can be seen, DPD is linear in a larger range.

25

Lecture 4.

DVD focus control

4.2 Dynamics
The DVD player system can be viewed as a two-input, two-output dynamical system. The main dynamics are due to the springs and masses in the lens system.
The inputs are voltages to the electromagnets moving the lens, and the outputs
are voltages corresponding to F E and RE. See Figure 4.7.
ufocus

FE
Pick-up & Disc

uradial

RE

The pick-up head and disk seen as a two-input, two-output dynamical system.

Figure 4.7

The department has a raw DVD player without any controller. Using system
identification techniques, the transfer function P f (s) from ufocus to F E and the
transfer function Pr (s) from uradial to RE have been estimated. The cross-coupling
between inputs and outputs have been ignored for simplicity. The resulting Bode
diagrams can be seen in Figure 4.8.
Bode Diagram

Bode Diagram

100

Magnitude (dB)

50

50

50

50

90

90

Phase (deg)

Phase (deg)

Magnitude (dB)

100

180

270
2
10

10

10
Frequency (rad/sec)

10

10

180

270
1
10

10

10
Frequency (rad/sec)

10

10

Figure 4.8 Left: Transfer function estimate for the focus servo. The model is of second
order. Right: Transfer function estimate for the radial servo.

4.3 Specifications and control structure

The control structure is simple: the focus and radial parts are separated (see
Figure 4.9). The task of the focus controller C f is to keep the disk sufficiently in
focus for the system to see the tracks, and the task of the radial controller Cr is
to keep the laser spot in the track.
In the DVD specification, bounds on the open-loop transfer functions for both
the focus and radial loops are given. The open-loop transfer function is (as you
know) the product of all elements in the control loop: Lr = Cr Pr , L f = C f P f . The
control specifications in the DVD standard are as follows:

p C (i ) P(i )p 1000

for 23.1 Hz

p C (i ) P(i )p 1

for > 2 kHz

The first specification is chosen to remove enough of the disk oscillation disturbances to stay in track. The second specification is chosen to reduce the effects of
measurement noise and small disturbances (dirt and scratches).
26

4.4

Design of the Focus Controller

Focus controller

ufocus
uradial

FE
PUH & Disk

RE

Radial controller

Figure 4.9

The control structure.

Meeting these specifications is not a trivial task. In fact, most DVD readers
probably dont! Instead, the manufacturers modify the specifications according
to the circumstances. For example, a CD player for a car would need very good
disturbance rejection (high gain at low frequencies) and this is obtained at the
expense of high gain also at high frequncies, which gives less robustness to disc
scratches.
The goal of this lecture is to design the focus controller according to the specifications and try it on the experimental setup. In Lab 3 in this course, you will
design your own radial controller to keep the DVD player in track!

4.4 Design of the Focus Controller

The specifications require that the magnitude of the loop transfer function decreases by a factor 1000 in the frequency interval between 23.1 Hz and 2.0 kHz.
Without compensation, the magnitude decreases only by a factor 200. See left
plot in Figure 4.10. A proportional controller with gain 0.4 satisfies the amplitude
specification at 2.0 kHz (right plot). However, the phase at 2.0 kHz is 184 , so
the gain has to be even smaller in order to keep the closed loop stable. Hence, a
dynamic compensator is needed.

Figure 4.10

Uncompensated system (left) and with proportional gain 0.4 (right).

It is natural to introduce lag compensation to increase the gain at low frequencies. However, the break points need to be at frequencies well below 2 kHz
in order to avoid additional phase lag at the cut-off frequency. Using a lag filter
C1 (s) = 0.4 s+s600 gives the modified plots in Figure 4.11 (left).
27

Lecture 4.

DVD focus control

However, the closed loop system is still unstable, so further compensation is

needed. Hence, we add a lead filter to increase the phase near 2 kHz; C2 (s) =
1+s/5000
1+s/50000 C1 ( s) (right plot).

Figure 4.11 A lag filter has been used to increase the gain at low frequencies (left). A lead
filter improves stability by increasing the phase near 2 kHz

The gain needs to be adjusted to keep the cross-over frequency unchanged.

See Figure 4.12 (left). Now the closed loop system is stable with good margins,
but the gain at 23.1 Hz is still too low, just 100 instead of 1000. This can be
corrected by modifying the break point of the lag filter to get the final controller
5000
. See Figure 4.12 (right).
C (s) = 0.15 s(s+11600) 11++ss//50000

Figure 4.12 Bode and Nichols plots with three lag filters and one lead filter (dashed) and
without compensation (solid).

The controller is verified by experiments done in the lecture. Notice

that the fi

sTd
1
nal design is very similar to a PID controller of the form C (s) = K sTi + 1+sT
.
d/ N

28

Lecture 10

Fundamental limitations1
10.1 Introduction
Maybe the cardinal mistake in control engineering would be to consider the process
as fixed once and for all. In fact, the control system specifications could very well
be impossible to meet once the process is constructed and fixed. A striking example
of this is given in the following citation from F. R. Whitt and D. G. Wilson (MIT
Press, 1974), Bicycling Science - Ergonomics and Mechanics:
Many people have seen theoretical advantages in the fact that front-drive,
rear-steered recumbent bicycles would have simpler transmissions than
rear-driven recumbents and could have the center of mass nearer the front
wheel than the rear. The U.S. Department of Transportation commissioned
the construction of a safe motorcycle with this configuration. It turned out
to be safe in an unexpected way: No one could ride it.
This lecture is devoted to the fundamental limitations that are inherited from
properties of the controlled plant and will address questions like the following
two:

Why are some bicycles impossible to ride?

How short inverted pendulums can be balanced by hand?
One of the practically most important restrictions to control performance is the
fact that actuators have limited capacity and may saturate. However, saturation
is a non-linear effect and will not be studied further here. Instead, the focus will
be on limitations caused by unstable zeros, unstable poles and time-delays.
An unstable pole p means that the response to a disturbance grows exponentially as ept . It is intuitively clear that in order to stabilize such a system, the
feedback loop must be faster than the time constant 1/ p. A formal argument for
this will be given in this lecture.
In case the unstable system also includes a time-delay, the control problem
could become impossible. A time-delay T means that control action at time t does
not have any effect until time t + T. Hence, it is intuitively clear that an unstable
pole can not be stabilized unless T is small compared to 1/ p. The argument applies
to the question about pendulum balancing above, since the human feedback system
through the eyes always involves a time delay.
A more intricate performance limitation is imposed by unstable zeros. It is
well known that an unstable zero results in a step response that initially goes in
the wrong direction. In fact, this can be seen directly in the expression for the
Laplace transform Y (s) of the step response y(t), where the zero at z means that
0 = Y ( z) =
1 Written

y(t) e zt dt

by A. Rantzer with contributions by K.J. strm

29

Lecture 10.

Fundamental limitations1

Figure 10.1 Schematic picture of a bicycle. The top view is shown on the left and the rear
view on the right.

Clearly the integral cannot be zero unless y(t) takes both positive and negative
values. The time duration of such dynamics is approximately 1/ z and limits the
achievable rate of control. Hence, while an unstable pole requires a fast feedback
loop, an unstable zero gives an upper bound on how fast it can be. A combination
of the two phenomena can make the system impossible to control.
Example 1 A tourque balance for a bicycle can be written as


mV0 {
d
d2
V0 + a
J 2 = m{ +
dt
b
dt
where the physical parameters have typical values as follows:
Mass:

m = 70 kg

Distance rear-to-center:

a = 0.3m

Height over ground:

Distance center-to-front:

{ = 1.2 m
b = 0.7 m

Moment of inertia:

J = 120 kgm2
V0 = 5 ms1

Speed:
Acceleration of gravity:

= 9.81 ms2

The transfer function from to is

P ( s) =

mV0 { as + V0
b Js2 m{

The system has an unstable pole p with time-constant

s
J
1
( 0.4 s
p =
m{
The closed loop system must be at least as fast as this. Moreover, the transfer
function has a zero z with
a
z1 =
( 0.06s
V0

30

10.2

The Maximum Modulus Theorem

Riding the bicycle at this speed, the zero is not really an obstacle for control.
However, with a rear-wheel-steered bicycle, the speed gets a negative sign and
the zero becomes unstable. In particular, for slow speed (( 0.7m/s) there is an
unstable pole-zero cancellation, which is impossible to stabilize.
2

10.2 The Maximum Modulus Theorem

More formal arguments about fundamental limitations can be obtained using theory for analytic functions. It is natural that analytic functions appear, since we
have a seen that a controller is stabilizing if and only if the closed loop transfer
function is analytic in the right half plane (all poles in the left half plane). The
main mathematical theorem to be used is the following:
THEOREM 10.1THE MAXIMUM MODULUS THEOREM
Suppose that the function f is analytic in a set containing the unit disc. Then
max p f ( z)p = max p f ( z)p
p zp1

p zp=1

In Laplace transform applications, the stability boundary will be the imaginary

axis. It is therefore convenient to note that for every stable rational transfer function G (s), analytic in the right half plane, the function
f ( z) = G

1+z
1z

is analytic in the unit disc. Hence the Maximum Modulus Theorem can be applied
to give the following corollary:
COROLLARY 10.1
Suppose that all poles of the rational function G (s) have negative real part. Then
max p G (s)p = max p G (i )p
R
Re s0

10.3 Sensitivity bounds from unstable zeros and poles

It is easy to see that the sensitivity function must be equal to one at an unstable
zero of the transfer function:
P( z) = 0

S( z) :=

1
=1
1 + C ( z) P( z)

Notice that the unstable zero in the plant can not be cancelled by an unstable pole
in the controller, since this would give an unstable transfer function C /(1 + CP)
from measurement noise to control input.
Similarly, the complimentary sensitivity must be one at an unstable pole:
P( p) =

T ( p) :=

C ( p) P( p)
=1
1 + C ( p) P( p)

31

Lecture 10.

Fundamental limitations1

log p Wa p

log p W b p

0
a

log

log

Figure 10.2 Amplitude plots for weighting functions. The left weighting function is used
to bound the sensitivity at small frequencies, while the right function is used to boud the
complementary sensitivity at high frequencies

In this case, cancellation by an unstable zero in the controller would give an

unstable transfer function P/(1 + CP) from input disturbance to plant output.
Combining the first constraint S( z) = 1 with Corollary 10.1 immediately gives
a lower bound on the norm of the sensitivity function:
max p S(i )p = max p S(s)p p S( z)p = 1
R
Re s0

q S q 1

This bound is however not particularly interesting, since usually S(i ) ( 1 for high
frequencies anyway. A much more interesting conclusion will next be obtained by
using a weighting function.
Recall that disturbance rejection requires small sensitivity for small frequencies. One way to formalize this condition is to define
Wa (s) =

s+a
2s

and require that

sup p Wa (i ) S(i )p 1

(10.1)

for some value of a. See Figure 10.2, left. Satisfying (10.1) with a high value of a
means fast disturbance rejection.
The specification requires that S(s) has a zero in the origin. This is often
obtained by an integrator in the controller. Moreover, Corollary 10.1 implies that
sup p Wa (i ) S(i )p = sup p Wa (s) S(s)p p Wa ( zi )p

Re s0

for every unstable zero zi of the plant P. In particular, the specification (10.1)
is impossible to satisfy unless p Wa ( zi )p 1, or in other words a zi , for every
unstable zero zi . Hence the unstable zeros give an upper bound on the achievable
bandwidth. In the following theorem, this discussion is summarized together with
a corresponding argument for unstable poles:
THEOREM 10.2
Suppose that the plant P(s) has unstable zeros zi and unstable poles p j . Define
the weighting functions Wa = (s + a)/(2s) and W b (s) = (s + b)/(2b). Then the
specifications
sup p Wa (i ) S(i )p 1

32



sup  W b (i )T (i ) 1

10.4

Bodes integral formula

are impossible to meet with a stabilizing controller unless

a min zi
i

b max p j
j

Proof. The statement about the sensitivity function was proved above, and the
statment about the complementary sensitivity function is analogous.
2
Example 2 Let us see what Theorem 10.2 has to say
p about the bicycle example.
The unstable pole gives a bound q W b T q 1 for b m{/ J. This shows that the
closed loop transfer function from measurement
noise to process output can not
p
be forced small for frequencies below m{/ J. A loose interpretation is that it
is impossible to ride the bicycle and keep the eyes shut except for a sample every
second. This applies for the bicycle with normal steering regardless of speed.
For a rear wheel steering bike, there is the second complication of an unstable
zero at V0 / a, which gives a bound on how fast disturbances one can reject. For
low speed, only slow disturbances can be rejected.
The special difficulties corresponding to a combination of an unstable pole and
an unstable zero nearby are however not apparent in Theorem 10.2. Such problems
will be treated next.
2
The following theorem gives simple expressions for the limitations caused by
an unstable pole/zero pair.
THEOREM 10.3
If P(s) has an unstable pole p and an unstable zero z, then






1
 z + p

 z p
1 + CP

for every stabilizing C (s).

Note that if S is very large, then the same is true for T, since S + T " 1. Hence, if
p( z + p)/( z p)p is significantly larger than one, the system is impossible to control
because of poor robustness to model errors and amplification of measurement
noise.
b(s), with P
b proper and P
b( p) ,= 0. Then
Proof. Assume that P(s) = (s z)(s p)1 P
the sensitivity function satisfies






 1 
1


 = sup 
q Sq = sup 


b(i z)(i p)1 
1 + CP

 1 + CP








i p
i + p




= sup 
 = sup 

b(i z) 
b(i z) 
 i p + C P
 i p + C P
 

  z + p 

s+ p
 


= sup 

b

z p
Re s0 s p + C P( s z) 
p
The fourth inequality uses that pi pp = 2 + p2 = pi + pp and the fifth
inequality is Corollary 10.1.
2
A similar argument can be applied to a system involving a time delay but
application of the maximum modulus theorem is less straightforward in this case.

10.4 Bodes integral formula

Another striking performance limitation, known as Bodes integral formula, shows
that the effort to make the sensitivity function small is always a trade-off between
different frequency regions:

33

Lecture 10.

Fundamental limitations1

0.1

magnitude

10

10

0.1

10

0.2

10

0.3

10

frequency
Figure 10.3 The amplitude curve of the sensitivity function always enclose the same area
below the level p Sp = 1 as above.

THEOREM 10.4
If P(s), C (s) and S(s) = [1 + C (s) P(s)]1 are stable and s2 C (s) P(s) is bounded,
then
Z
log p S(i )p d = 0
0

Proof. Proof sketch. From the theory of analytic functions, recall that Cauchys
formula states that
Z
f ( z)dz = 0

for every closed path in the region where the function f is analytic. Bodes
integral formula follows by application of Cauchys formula to
f ( z) = log S( z)
The stability of C and P guarantee that f is well-defined and analytic in the
whole right half plane. Integration along the imaginary axis can be extended to
integration along a closed path by adding a large half-circle in the right half
plane. The condition that s2 C (s) P(s) is bounded is needed to make sure that the
contribution from the half-circle vanishes as the radius tends to infinity.
2
The invariance of Bodes integral is sometimes referred to as the water-bed
effect: If the designer tries to push the magnitude of the sensitivity function down
at some point, it will inevitably pop up somewhere else!
The assumptions behind Bodes integral formula deserve some discussion. The
expression s2 C (s) P(s) is always bounded whenever C (s) and P(s) correspond to
real sensor/actuator interconnections, since direct terms are not physically implementable. With unstable poles in C (s) P(s) the integral formula changes into

log p S(i )p d =

Re pi

which makes it even harder to push down the sensitivty magnitude! The faster
unstable modes, the harder it is. In fact, this can be used as an argument why
unstable controllers should in general be avoided.

34

Lecture 11

Multivariable control1
There is a clear trend in modern engineering toward systems of higher and higher
complexity. One reason is that demands for efficiency give tighter interconnections
between subsystems. For example, to get minimally pollutive emissions in the
exhaust gas of a car, it is necessary that engine, carburettor, catalyst, gearbox, etc.
all cooperate in an optimal manner. Similarly, the demand for efficient production
and distribution of electrical power has led to tighter coupling between production
units in different geographical regions and more complex large scale dynamics.
A system with several inputs and several outputs is sometimes called a MIMO
(Multiple-Input-Multiple-Output) system. Control theory for such systems is a
highly active research area and a review of the available methods is outside the
scope of this course. However, many of the ideas that were developed for scalar
systems can be easily adapted also to a multivariable setting. This lecture will
present a few such items:

Multivariable performance specifications

Limitations due to unstable multivariable zeros
Decentralized control by pairing of signals

Figure 11.1 A modern car, a power plant and an oil refinery all make extensive use of
multivariable control systems

11.1 Multivariable specifications

Consider again the feedback loop illustrated in Figure 11.2 but assume that all
signals are vector valued. Then P(s), F (s) and C (s) are matrices, so the closed
loop transfer functions need to be derived with some more care:
X (s) = PC F R(s) + P D (s) PC [ N + X ](s)
[ I + PC ] X (s) = PC F R(s) + P D (s) PC N (s)
X (s) = [ I + PC ]1 PC F R(s) + [ I + PC ]1 P D (s) [ I + PC ]1 PC N (s)
1 Written by A. Rantzer. The examples come from the book by Glad and Ljung, Control Theory
Multivariable and Nonlinear Methods, Taylor & Francis, 2000

35

Lecture 11.

Multivariable control1
n

d
r

e
F ( s)

C ( s)

P ( s)

I
Controller

Process
Figure 11.2

where a dot denotes matrix-vector multiplication.

Similarly
V (s) = [ I + CP]1 C F R(s) [ I + CP]1 C N (s) + [ I + CP]1 D (s)
Notice that S = [ I + PC ]1 is generally not the same as [ I + CP]1 . The first is
called the sensitivity function and has a matrix size determined by the number
of outputs. The second is called input sensitivity function and its size corresponds
to the number of inputs. The convention is to call T = [ I + PC ]1 PC the complementary sensitivity function. Notice the following identities:

[ I + PC ]1 P = P[ I + CP]1
C [ I + PC ]1 = [ I + CP]1 C
T = P[ I + CP]1 C = PC [ I + PC ]1
S+T = I
The first equality follows by multiplication with I + CP from the right and I + PC
from the left. The second one is analogous. Using the first two equalities, we
immediately get the third. The last one is straight from definitions as well.
Also for multivariable systems, it is common to require S to be small at low
frequencies and T to be small at high frequencies. The first specification means
that y follows Fr well at small frequencies, while the second means that high
frequency measurement noise n does not influence x significantly. Another way
to state these requirements is to say that the loop transfer matrix
P(i ) C (i )
should have small norm q P(i ) C (i )q at high frequencies, while at low the frequencies instead q[ P(i ) C (i )]1 q should be small. See Figure 11.3.

11.2 Limitations due to unstable zeros

Just like for scalar systems, there are fundamental limitations on the achievable
performance in a multivariable system. For a multivariable system with square
tranfer matrix P(s), i.e. the same number of inputs and outputs, the zeros can be
defined as the poles of P(s)1 . The following theorem captures the influence of an
unstable zero:
THEOREM 11.1
Let WS (s) be stable and let S(s) = [ I + P(s) C (s)]1 be the sensitivity function of
a stable closed loop system. Then, the specification

q WS Sq 1
is impossible to satisfy unless q WS ( z)q 1 for every unstable zero z of P(s).

36

11.2

Limitations due to unstable zeros

Magnitude

20

P(i ) C (i )

10
0
10
20
30

Robustness

40
50
60
w0

70

w1

10
Frequency
0

10

Disturbance rejection
Figure 11.3 Specifications on the singular values of the loop transfer function often have
this form. A lower bound on the singular values for low frequencies is needed for disturbance
rejection. This means that q[ P(i ) C(i )]1 q should be small. An upper bound on the singular
values for high frequencies, making q P(i ) C(i )q small enough, is needed for robustness to
model errors and measurement noise.

The proof is analogous to the result for scalar systems in Lecture 4. Instead of
giving the details, we turn our attention to an example.
Example 1 (Non-minimum phase MIMO System) Consider a feedback system
y = ( I + PC )1 r with the multivariable process

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

3
s+2

1
s+1

Computing the determinant by hand

det P(s) =

2
3
s + 1

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

shows that the process has an unstable zero at s = 1, which will limit the achievable performance. For further understanding of the limitation, consider the following three different control structures:
Controller 1

The controller

K 1 (s + 1)

s
C1 (s) =
K 1 (s + 1)

3K 2 (s + 0.5)
s(s + 2)

2K 2 (s + 0.5)

s(s + 1)

gives the diagonal loop transfer matrix

K 1 (s + 1)
s(s + 2)
P(s) C1 (s) =

K 2 (s + 0.5)(s + 1)
s(s + 1)(s + 2)

Hence the system is decoupled into to scalar loops, each with an unstable zero
at s = 1 that limits the bandwidth. The closed loop step responses are shown in
Figure 11.4.

37

Lecture 11.

Multivariable control1

Step Response
1.5

Amplitude

1
0.5
0
0.5

5
Time (sec)

10

10

Step Response
1.5

Amplitude

1
0.5
0
0.5

5
Time (sec)

Figure 11.4 Closed loop step responses with decoupling controller C1 (s) for the two outputs
y1 (solid) and y2 (dashed). The upper plot is for a reference step for y1 . The lower plot is for
a reference step for y2 .

Controller 2

The controller

K (s + 1)
1
K2

s
C2 (s) =

K 1 (s + 1)
K2

s
gives the upper triangular loop transfer matrix

K 1 (s + 1)
K 2 (5s + 7)
s(s + 2)
(s + 2)(s + 1)

P(s) C2 (s) =

2K 2
0
s+1

Now the decoupling is only partial: Output y2 is not affected by r1 . Moreover,

there is no unstable zero that limits the rate of response in y2 ! The closed loop
step responses for K 1 = 1, K 2 = 10 are shown in Figure 11.5.
Controller 3

The controller

C3 (s) =

K1
K1

K 2 (s + 0.5)

s(s + 2)

2K 2 (s + 0.5)
s(s + 1)

gives the lower triangular loop transfer matrix

K 1 (5s + 7)
0
(s + 1)(s + 2)

P(s) C3 (s) =

K 2 (1 + s)(s + 0.5)
2K 1
s+1
s(s + 1)2 (s + 2)

In this case y1 is decoupled from r2 and can respond arbitrarily fast for high values
of K 1 , at the expense of bad behavior in y2 . Step responses for K 1 = 10, K 2 = 1
are shown in Figure 11.6.
To summarize, the example shows that even though a multivariable unstable
zero always gives a performance limitation, it is possible to influence where the
effects should show up.
2

38

11.3

Pairing of signals

Step Response
1.5

Amplitude

1
0.5
0
0.5

5
Time (sec)

10

10

Step Response
3

Amplitude

2.5
2
1.5
1
0.5
0
0.5

5
Time (sec)

Figure 11.5 Closed loop step responses with controller C2 (s) for the two outputs y1 (solid)
and y2 (dashed). The right half plane zero does not prevent a fast y2 -response to r 2 but at the
price of a simultaneous undesired response in y1 .
Step Response
1.5

Amplitude

1
0.5
0
0.5

5
Time (sec)

10

10

Step Response

Amplitude

0.5

0.5

5
Time (sec)

Figure 11.6 Closed loop step responses with controller C3 (s) for the two outputs y1 (solid)
and y2 (dashed). The right half plane zero does not prevent a fast y1 -response to r 1 but at the
price of a simultaneous undesired response in y2 .

11.3 Pairing of signals

The simplest way to deal with a multivariable control problem is to select an equal
number of inputs and outputs and make a pairing, so that each input is responsible
for control of one particular output. The corresponding transfer function from input
to output of the process is determined for each pair, and the scalar feedback loops
are designed ignoring the coupling between the loops inside the plant.
When connecting all the feedback loops simultaneously, the cross-coupling may
potentially lead to performance degradation or even instability. The approach
therefore works much better if there is a good way to select input and output
variables for the scalar loops. One way to do this is to use the Relative Gain Array

39

Lecture 11.

Multivariable control1

(RGA) that will be studied next.

For an full rank matrix A Cm$n , define the Relative Gain Array as
RGA( A) := A. ( A1 )T
where .* denotes element-by-element multiplication. For non-square matrices,
the inverse A1 is replaced by the pseudo-inverse A (in Matlab pinv(A)). If m > n,
then A = ( A A)1 A and if m < n then A = A ( AA )1 . The Relative Gain Array
has several interesting properties:

The sum of all elements in a column or row is one.

Permutations of rows or columns in A give the same permutations in RGA( A)
RGA( A) = RGA( D1 AD2 ) if D1 and D2 are diagonal, i.e. RGA( A) is independent of scaling
If A is triangular, then RGA( A) is the unit matrix I.
Furthermore, the Relative Gain Array has an interpretation related to control
theory: Let P(s) be the transfer matrix from u to y. Then

The ( k, j ) element of P is the transfer function u j yk when ui = 0 for i ,= j

(open loop control )
The ( j, k) element of P1 is the inverse of the transfer function u j yk when
the other inputs are such that yi = 0 for i ,= k (closed loop control )
Hence, if the ( k, j ) element of RGA( P(s)) is equal to one, then the value of transfer
function from u j to yk does not depend on whether the remaining inputs operate
in open loop or closed loop. This indicates that the cross-coupling with other loops
is weak.
The following rules of thumb use RGA to identify input-output pairings that
have small cross-coupling. However, the outcome must always be evaluated using
other tools.
1. Find a permutation of inputs and outputs that brings RGA( P(i c )) as close
as possible to the identity matrix. Here c is the closed loop bandwidth of
the system.
2. Avoid pairings that give negative diagonal elements of RGA( P(0))
The second rule is motivated by the desire to have the same sign of the static
gain in one loop regardless if the other loops are closed or not. We illustrate the
rules with an example.
Example: A Distillation Column
A distillation column is used in chemical process industry to separate different
components in a chemical product. This example comes from Shell and describes a
column where raw oil is separated into various petro-chemical derivatives. Heated
raw oil is inserted at the bottom of the column and evaporates. Subcomponents
then condense and get extracted at different levels of the column.
Outputs:

Inputs:

y1 = top draw composition

u1 = top draw flowrate

y2 = side draw composition

u2 = side draw flowrate

u3 = bottom temperature control input

40

4
e27s
Y1 (s)
50s + 1
=
5.4
Y2 (s)
e18s
50s
+1
|

1.8
e28s
60s + 1
5.7
e14s
60s +{z1
P ( s)

5.9
U1 ( s )
e27s

50s + 1
U2 ( s )
6.9
e15s
U3 ( s )
40s + 1
}

11.3

Figure 11.7

Pairing of signals

Schematic picture of a distillation column

Computing the RGA for the distillation column model gives



4.0 1.8 5.9
P(0) =
5.4 5.7 6.9


0.6871 2.7437i 0.1548 1.1419i 1.0135 4.0469i
P(i/50) =
1.5758 3.4781i 1.4704 3.3397i 3.0247 4.4589i


0.6111 1.3285
RGA( P(0)) =
0.0134 1.5827 0.5962


0.4355 0.3667i 0.6536 0.0171i 1.2181 + 0.3839i
RGA( P(i/50)) =
0.0906 + 0.3667i 1.5933 + 0.0171i 0.5027 0.3839i
0.2827

To choose control signal for y1 , we apply the rules of thumb to the top row. This
suggests the bottom temperature u3 for control of the top draw composition y1 ,
since the third column has values slightly closer to 1 than the first column.
Based on the bottom row, we choose the side draw flowrate u2 to control the
side draw composition y2 . The top draw flow rate u1 is left unused. The matrix
transfer function P (s) from (u3 , u2 ) to ( y1 , y2 ) is now to be controlled by a diagonal
controller C (s), say a PI controller:

60s + 1

5.9e27s 1.8e28s
0
50s + 1 60s + 1

C (s) = 50s
P (s) =

6.9e15s 5.7e14s
60s + 1
0
50s
40s + 1 60s + 1
Without feedforward, the closed loop transfer matrix from reference to output
becomes
PC ( I + PC )1

41

Lecture 11.

Multivariable control1

Bode Magnitude Diagram

From: U(1)

From: U(2)

20

To: Y(1)

0
20
40

Magnitude (dB)

60
80
20

To: Y(2)

0
20
40
60 2
10

10

2 2

1010

10

10

Frequency (rad/sec)

Figure 11.8

Magnitude plots for the closed loop transfer function of the distillation column

and the Bode magnitude plots for the four transfer functions are given in Figure 11.8. As seen in the plots, the resulting cross-coupling is generally small and
there is no static error.

42