0% found this document useful (0 votes)

196 views

A Mathematical Approach From Classical Control To Advanced Control

This book takes an untraditional approach to introductory control. The idea is to develop control theory, at an introductory classical level, as a rigorous subject. Control theory is an effective blend of many branches of mathematics.

Uploaded by

Mohammed Ahsan Fawad

Available Formats

Download as PDF, TXT or read online on Scribd

0% found this document useful (0 votes)

196 views

A Mathematical Approach From Classical Control To Advanced Control

Uploaded by

Mohammed Ahsan Fawad

Available Formats

Download as PDF, TXT or read online on Scribd

You are on page 1/ 346

ii

A Mathematical Approach
to Classical Control
Single-input, single-output,
time-invariant, continuous time,
finite-dimensional, deterministic,
linear systems

Andrew D. Lewis
JanuaryApril 2003
This version: 22/10/2004

This version: 22/10/2004

Preface
The purpose of this preface is two-fold: (1) to discuss the philosophy of the approach taken,
as it is nonstandard for an introductory course; (2) to discuss the content of the book.

The philosophy
Since this book takes an untraditional approach to introductory control, it is worth
outlining why I have taken the approach I have.
The goals Clearly a new text in classical control needs to have some justification for its
appearance, as there are already a large number of texts on the market, and these satisfy the
demands made by a typical introductory course in feedback control. The approach in this
book is not typical. The idea here is to develop control theory, at an introductory classical
level, as a rigorous subject. This is different, note, from presenting the mathematics needed
to understand control theory. One will often hear things like, Classical control is merely an
application of complex variable theory, or Linear control is merely an application of linear
algebra. While it is true that these parts of control theory do rely on the asserted branches
of mathematics, control theory is such an effective blend of many branches of mathematics
that to categorise it as a subset of one is a disservice. The subject of control theory, even
at an introductory level, has a mathematical life of its own, and it is this life that is being
exhibited here.
The main benefit of such an approach is that not just the mathematics behind the
subject, but the subject itself can be treated rigorously. The problems of control theory,
and these are actual practical problems, often have precise mathematical statements, and
the intent in this book is to give these wherever possible. The result is that a student
will be able to understand simple problems in a larger context. For some, at least, this is
useful. It also makes it possible to consider challenging control problems that cannot really
be considered in an exclusively ad hoc treatment. It would seem that many classical control
texts were written based upon the standard of control practice in, say, the early 1960s. This
practice, well laid out in the texts of Truxal [1955] and Horowitz [1963], had reached a point
where, for the problems to which it was applicable, it was finished. This was expressed
in one early paper as follows: The present state of the art is such that it is safe to assume
that, for linear single-loop feedback systems, almost no analysis or design problems of any
consequence remain. Such statements are seldom prophetic. Indeed, much has been done
since the date of publication of the cited paper (1961), even for linear single-loop systems.
Now we have means for handling problems that would be almost impossible to treat using
the ad hoc methods of classical design. And the methods all rely on a firm grasp of not just
the mathematics behind control theory, but the mathematics of the subject itself. This is
the reason for this book.
The mathematical approach With the above as backdrop, this book is provided for
students who can be relied upon to have a satisfactory background in linear algebra, differential equations (including the matrix exponential), basic complex analysis, and some

transform theory. The appendices contain a quick overview of necessary background material, so that an instructor or a student can determine whether the book is useful.
Apart from the above pedagogical concerns, I have also tried to write the book with
an eye towards its being a useful reference. In the book, I have tried to prove as many
statements as possible; even many that are not commonly proved, but often stated. I do
this not because I feel that all of these proofs should be delivered in lecturesI certainly do
not do this myself. Rather, my objectives here are scholarly. I do not feel that such lofty
goals clash with the rather more pedantic concerns of getting students to come to grips with
basic material. Students who find the course challenging may safely omit consideration of the
more technical proofs, provided that they understand the concepts behind the results. More
curious students, however, are rewarded by having for reference proofs that can be difficult
to find in the literature. Moreover, this approach has, in my experience, a pedagogical
byproduct. If one teaches an introductory course in a manner not completely method
oriented, natural questions will arise in the presentation. For example, if one even gets
around to posing the problem of finding a controller that stabilises a given plant in a unity
gain feedback loop, the natural question arises as to whether such controllers exist. The
answer is affirmative, but the determination of this answer is nontrivial. A traditional
approach to classical control masks the existence of the question, never mind providing the
answer. Again, the advantage of the approach taken here, at least for the curious student,
is that the answer to this more basic question may be found alongside the more standard ad
hoc methods for controller design.
The r
ole of control design A word needs to be said about control design. Greater
emphasis is being placed on engineering design in the engineering undergraduate curriculum,
and this is a by all means an appropriate tendency. When teaching a control course, one
faces a decision relative to design content. Should the design be integrated into the course at
every stage, or should it be separated from the analysis parts of the course? In this book, the
trend in engineering education is being bucked, and the latter course is taken. Indeed, care
has been taken to explicitly separate the book into three parts, with the design part coming
last. One can justly argue that this is a mistake, but it is the approach I have decided upon,
and it seems to work. My rationale for adopting the approach I do is that in control, there
is very simply a lot of analysis to learn before one can do design in a fulfilling way. Thus I
get all the tools in place before design is undertaken in the latter stages of the book.
How to use the book It is not possible to cover all of the topics in this book in a
single term; at least it is not advisable to attempt this. However, it is quite easy to break
the contents of the book into two courses, one at an introductory level, and another dealing
with advanced topics. Because this division is not readily made on a chapter-by-chapter
basis, it is perhaps worth suggesting two possible courses that can be taught from this book.
An introductory course for students with no control background might contain roughly
the following material:
1. Chapter 1;
2. Chapter 2, possibly omitting details about zero dynamics (Section 2.3.3), and going
lightly on some of the proofs in Section 2.3;
3. Chapter 3, certainly going lightly on the proofs in Section 3.3;
4. Chapter 4, probably omitting Bodes Gain/Phase Theorem (Section 4.4.2) and perhaps
material about plant uncertainty models (Section 4.5);

v
5. Chapter 5, omitting many of the details of signal and system norms in Section 5.3, omitting Liapunov stability (Section 5.4), and omitting the proofs of the Routh/Hurwitz
criteria;
6. Chapter 6, going lightly, perhaps, on the detailed account of signal flow graphs in
Sections 6.1 and 6.2, and covering as much of the material in Section 6.4 as deemed
appropriate; the material in Section 6.5 may form the core of the discussion about
feedback in a more traditional course;1
7. Chapter 7, probably omitting robust stability (Section 7.3);
8. Chapter 8;
9. maybe some of the material in Chapter 9, if the instructor is so inclined;
10. Chapter 11, although I rarely say much about root-locus in the course I teach;
11. Chapter 12, omitting Section 12.3 if robustness has not been covered in the earlier
material;
12. perhaps some of the advanced PID synthesis methods of Chapter 13.
When I teach the introductory course, it is offered with a companion lab class. The lab
course follows the lecture course in content, although it is somewhat more down to earth.
Labs start out with the objective of getting students familiar with the ideas introduced in
lectures, and by the end of the course, students are putting into practice these ideas to design
controllers.
A more advanced course, having as prerequisite the material from the basic course, could
be structured as follows:
1. thorough treatment of material in Chapter 2;
2. ditto for Chapter 3;
3. Bodes Gain/Phase Theorem (Section 4.4.2) and uncertainty models (Section 4.5);
4. thorough treatment of signal and system norms from Section 5.3, proofs of
Routh/Hurwitz criteria if one is so inclined, and Liapunov methods for stability (Section 5.4);
5. static state feedback, static output feedback, and dynamic output feedback (Section 6.4);
6. robust stability (Section 7.3);
7. design limitations in Chapter 9;
8. robust performance (Section 9.3);
9. Chapter 10, maybe omitting Section 10.4 on strong stabilisation;
10. basic loop shaping using robustness criterion (Section 12.3);
11. perhaps the advanced synthesis methods of Chapter 13;
12. Chapter 14;
13. Chapter 15.
1

Of course, someone teaching a traditional course is unlikely to be using this book.

The content
In Chapter 1 we engage in a loose discourse on ideas of a control theoretic nature. The
value of feedback is introduced via a simple DC servo motor example using proportional
feedback. Modelling and linearisation are also discussed in this chapter. From here, the
book breaks up into three parts (plus appendices), with the presentation taking a rather less
loose form.
Part I. System representations and their properties Linear systems are typically
represented in one of three ways: in the time domain using state space methods (Chapter 2);
in the Laplace transform domain using transfer functions (Chapter 3); and in the frequency
domain using the frequency response (Chapter 4). These representations are all related in
one way or another, and there exist vocal proponents of one or the other representation. I
do not get involved in any discussion over which representation is best, but treat each
with equal importance (as near as I can), pointing out the innate similarities shared by the
three models.
As is clear from the books subtitle, the treatment is single-input, single-output (SISO),
with a very few exceptions, all of them occurring near the beginning of Chapter 2. The focus
on SISO systems allows students to have in mind simple models. MIMO generalisations of
the results in the book typically fall into one of two categories, trivial and very difficult.
The former will cause no difficulty, and the latter serve to make the treatment more difficult
than is feasible in an introductory text. References are given to advanced material.
Specialised topics in this part of the book include a detailed description of zero dynamics
in both the state space and the transfer function representations. This material, along with
the discussion of the properties of the transfer function in Section 3.3, have a rather technical
nature. However, the essential ideas can be easily grasped independent of a comprehension
of the proofs. Another specialised topic is a full account of Bodes Gain/Phase Theorem in
Section 4.4.2. This is an interesting theorem; however, time does not normally permit me
to cover it in an introductory course.
A good understanding of the material in this first part of the book makes the remainder
of the book somewhat more easily digestible. It has been my experience that students find
this first material the most difficult.
Part II. System analysis Armed with a thorough understanding of the three representations of a linear system, the student is next guided through methods for analysing such
systems. The first concern in such a discussion should be, and here is, stability. A control
design cannot be considered in any way successful unless it has certain stability properties.
Stability for control systems has an ingredient that separates it from stability for simple dynamical systems. In control, one is often presented with a system that is nominally unstable,
and it is desired to stabilise it using feedback. Thus feedback is another central factor in our
discussion of control systems analysis. We are rather more systematic about this than is the
norm. The discussion of signal flow graphs in Sections Section 6.1 and 6.2 is quite detailed,
and some of this detail can be skimmed. However, the special notion of stability for interconnected systems, here called IBIBO stability, is important, and the notation associated with
it appears throughout the remainder of the book. The Nyquist criterion for IBIBO stability
is an important part of classical control. Indeed, in Section 7.3 the ideas of the Nyquist
plot motivate our discussion of robust stability. A final topic in control systems analysis is

vii
performance, and this is covered in two chapters, 8 and 9, the latter being concerned with
limitations on performance that arise due to features of the plant.
The latter of the two chapters on performance contains some specialised material concerning limitations on controller design that are covered in the excellent text of Seron, Braslavsky,
and Goodwin [1997]. Also in this chapter is presented the robust performance problem,
whose solution comprises Chapter 15. Thus Chapter 9 should certainly be thought of as one
of special topics, not likely to be covered in detail in a first course.
Part III. Controller design The final part of the text proper is a collection of control
design schemes. We have tried to present this material in as systematic a manner as possible.
This gives some emphasis to the fact that in modern linear control, there are well-developed
design methods based on a solid mathematical foundation. That said, an attempt has been
made to point out that there will always be an element of artistry to a good control
design. While an out of the box controller using some of the methods we present may be
a good starting point, a good control designer can always improve on such a design using
their experience as a guide. This sort of material is difficult to teach, of course. However,
an attempt has been made to give sufficient attention to this matter.
This part of the book starts off with a discussion of the stabilisation problem.
Part IV. Background and addenda There are appendices reviewing relevant material
in linear algebra, the matrix exponential, complex variables, and transforms. It is expected
that students will have seen all of the material in these appendices, but they can look here
to refamiliarise themselves with some basic concepts.
What is not in the book The major omission of the book is discrete time ideas. These
are quite important in our digital age. However, students familiar with the continuous time
ideas presented here will have no difficulty understanding their discrete time analogues.
That said, it should be understood that an important feature in control is missing with the
omission of digital control, and that instructors may wish to insert material of this nature.
This book is in its third go around. The version this year is significantly expanded from
previous years, so there are apt to be many errors. If you find an error, no matter how small,
let me know !
Andrew D. Lewis
Department of Mathematics & Statistics
Queens University
Kingston, ON K7L 3N6, Canada
[email protected]
(613) 533-2395
22/10/2004

viii

This version: 22/10/2004

3.6

Table of Contents
1

I
2

An
1.1
1.2
1.3

introduction to linear control theory

Some control theoretic terminology . . . . . . .
An introductory example . . . . . . . . . . . .
Linear differential equations for physical devices
1.3.1 Mechanical gadgets . . . . . . . . . . .
1.3.2 Electrical gadgets . . . . . . . . . . . .
1.3.3 Electro-mechanical gadgets . . . . . . .
1.4 Linearisation at equilibrium points . . . . . . .
1.5 What you are expected to know . . . . . . . . .
1.6 Summary . . . . . . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.

System representations and their properties

1
1
2
7
7
10
11
12
13
14

3.7
4

State-space representations (the time-domain)

2.1 Properties of finite-dimensional, time-invariant linear control systems . .
2.2 Obtaining linearised equations for nonlinear input/output systems . . .
2.3 Input/output response versus state behaviour . . . . . . . . . . . . . . .
2.3.1 Bad behaviour due to lack of observability . . . . . . . . . . . . .
2.3.2 Bad behaviour due to lack of controllability . . . . . . . . . . . .
2.3.3 Bad behaviour due to unstable zero dynamics . . . . . . . . . . .
2.3.4 A summary of what we have said in this section . . . . . . . . . .
2.4 The impulse response . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.4.1 The impulse response for causal systems . . . . . . . . . . . . . .
2.4.2 The impulse response for anticausal systems . . . . . . . . . . . .
2.5 Canonical forms for SISO systems . . . . . . . . . . . . . . . . . . . . .
2.5.1 Controller canonical form . . . . . . . . . . . . . . . . . . . . . .
2.5.2 Observer canonical form . . . . . . . . . . . . . . . . . . . . . . .
2.5.3 Canonical forms for uncontrollable and/or unobservable systems .
2.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

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

23
24
31
33
33
38
43
47
48
48
53
54
55
57
59
61

Transfer functions (the s-domain)

3.1 Block diagram algebra . . . . . . . . . . . . . . . . . . . . . .
3.2 The transfer function for a SISO linear system . . . . . . . .
3.3 Properties of the transfer function for SISO linear systems . .
3.3.1 Controllability and the transfer function . . . . . . . .
3.3.2 Observability and the transfer function . . . . . . . . .
3.3.3 Zero dynamics and the transfer function . . . . . . . .
3.4 Transfer functions presented in input/output form . . . . . .
3.5 The connection between the transfer function and the impulse
3.5.1 Properties of the causal impulse response . . . . . . .
3.5.2 Things anticausal . . . . . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.

75
76
78
80
81
85
87
90
95
95
98

. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
response
. . . . .
. . . . .

.
.
.
.
.
.
.
.
.
.

II
5

The matter of computing outputs . . . . . . . . . . . . . . . . . . . . . . . .

3.6.1 Computing outputs for SISO linear systems in input/output form
using the right causal Laplace transform . . . . . . . . . . . . . . . .
3.6.2 Computing outputs for SISO linear systems in input/output form
using the left causal Laplace transform . . . . . . . . . . . . . . . . .
3.6.3 Computing outputs for SISO linear systems in input/output form
using the causal impulse response . . . . . . . . . . . . . . . . . . . .
3.6.4 Computing outputs for SISO linear systems . . . . . . . . . . . . . .
3.6.5 Formulae for impulse, step, and ramp responses . . . . . . . . . . . .
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Frequency response (the frequency domain)

4.1 The frequency response of SISO linear systems . . . . . . . . . . .
4.2 The frequency response for systems in input/output form . . . . .
4.3 Graphical representations of the frequency response . . . . . . . . .
4.3.1 The Bode plot . . . . . . . . . . . . . . . . . . . . . . . . .
4.3.2 A quick and dirty plotting method for Bode plots . . . . . .
4.3.3 The polar frequency response plot . . . . . . . . . . . . . .
4.4 Properties of the frequency response . . . . . . . . . . . . . . . . .
4.4.1 Time-domain behaviour reflected in the frequency response
4.4.2 Bodes Gain/Phase Theorem . . . . . . . . . . . . . . . . .
4.5 Uncertainly in system models . . . . . . . . . . . . . . . . . . . . .
4.5.1 Structured and unstructured uncertainty . . . . . . . . . . .
4.5.2 Unstructured uncertainty models . . . . . . . . . . . . . . .
4.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

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

System analysis
Stability of control systems
5.1 Internal stability . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.2 Input/output stability . . . . . . . . . . . . . . . . . . . . . . . . .
5.2.1 BIBO stability of SISO linear systems . . . . . . . . . . . .
5.2.2 BIBO stability of SISO linear systems in input/output form
5.3 Norm interpretations of BIBO stability . . . . . . . . . . . . . . . .
5.3.1 Signal norms . . . . . . . . . . . . . . . . . . . . . . . . . .
5.3.2 Hardy spaces and transfer function norms . . . . . . . . . .
5.3.3 Stability interpretations of norms . . . . . . . . . . . . . . .
5.4 Liapunov methods . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.4.1 Background and terminology . . . . . . . . . . . . . . . . .
5.4.2 Liapunov functions for linear systems . . . . . . . . . . . .
5.5 Identifying polynomials with roots in C . . . . . . . . . . . . . . .
5.5.1 The Routh criterion . . . . . . . . . . . . . . . . . . . . . .
5.5.2 The Hurwitz criterion . . . . . . . . . . . . . . . . . . . . .
5.5.3 The Hermite criterion . . . . . . . . . . . . . . . . . . . . .
5.5.4 The Lienard-Chipart criterion . . . . . . . . . . . . . . . . .
5.5.5 Kharitonovs test . . . . . . . . . . . . . . . . . . . . . . . .
5.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

98
99
101
102
105
108
111
119
119
122
124
124
129
135
135
135
138
145
146
147
150

159
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

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

161
162
165
165
168
170
170
172
174
180
180
182
189
189
193
195
199
200
203

xi
6

Interconnections and feedback

6.1 Signal flow graphs . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.1.1 Definitions and examples . . . . . . . . . . . . . . . . . . .
6.1.2 Signal flow graphs and systems of equations . . . . . . . . .
6.1.3 Subgraphs, paths, and loops . . . . . . . . . . . . . . . . . .
6.1.4 Cofactors and the determinants . . . . . . . . . . . . . . . .
6.1.5 Masons Rule . . . . . . . . . . . . . . . . . . . . . . . . . .
6.1.6 Sensitivity, return difference, and loop transmittance . . . .
6.2 Interconnected SISO linear systems . . . . . . . . . . . . . . . . . .
6.2.1 Definitions and basic properties . . . . . . . . . . . . . . . .
6.2.2 Well-posedness . . . . . . . . . . . . . . . . . . . . . . . . .
6.2.3 Stability for interconnected systems . . . . . . . . . . . . .
6.3 Feedback for input/output systems with a single feedback loop . .
6.3.1 Open-loop versus closed-loop control . . . . . . . . . . . . .
6.3.2 Unity gain feedback loops . . . . . . . . . . . . . . . . . . .
6.3.3 Well-posedness and stability of single-loop interconnections
6.4 Feedback for SISO linear systems . . . . . . . . . . . . . . . . . . .
6.4.1 Static state feedback for SISO linear systems . . . . . . . .
6.4.2 Static output feedback for SISO linear systems . . . . . . .
6.4.3 Dynamic output feedback for SISO linear systems . . . . . .
6.5 The PID control law . . . . . . . . . . . . . . . . . . . . . . . . . .
6.5.1 Proportional control . . . . . . . . . . . . . . . . . . . . . .
6.5.2 Derivative control . . . . . . . . . . . . . . . . . . . . . . .
6.5.3 Integral control . . . . . . . . . . . . . . . . . . . . . . . . .
6.5.4 Characteristics of the PID control law . . . . . . . . . . . .
6.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

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

211
212
212
215
218
219
223
226
229
230
233
235
244
244
247
248
250
250
255
258
261
263
263
264
264
269

Frequency domain methods for stability

7.1 The Nyquist criterion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.1.1 The Principle of the Argument . . . . . . . . . . . . . . . . . . . . .
7.1.2 The Nyquist criterion for single-loop interconnections . . . . . . . . .
7.2 The relationship between the Nyquist contour and the Bode plot . . . . . .
7.2.1 Capturing the essential features of the Nyquist contour from the Bode
plot . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.2.2 Stability margins . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.3 Robust stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.3.1 Multiplicative uncertainty . . . . . . . . . . . . . . . . . . . . . . . .
7.3.2 Additive uncertainty . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

279
279
279
281
293

Performance of control systems

8.1 Time-domain performance specifications . . . . . . . . . . .
8.2 Performance for some classes of transfer functions . . . . . .
8.2.1 Simple first-order systems . . . . . . . . . . . . . . .
8.2.2 Simple second-order systems . . . . . . . . . . . . .
8.2.3 The addition of zeros and more poles to second-order
8.2.4 Summary . . . . . . . . . . . . . . . . . . . . . . . .
8.3 Steady-state error . . . . . . . . . . . . . . . . . . . . . . .

317
318
320
320
321
326
327
328

. . . . .
. . . . .
. . . . .
. . . . .
systems
. . . . .
. . . . .

.
.
.
.
.
.
.

293
294
303
304
308
311

xii

8.4
8.5

8.6

8.7
9

III

8.3.1 System type for SISO linear system in input/output form . . . . . . 329
8.3.2 System type for unity feedback closed-loop systems . . . . . . . . . . 333
8.3.3 Error indices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 336
8.3.4 The internal model principle . . . . . . . . . . . . . . . . . . . . . . 336
Disturbance rejection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 336
The sensitivity function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 342
8.5.1 Basic properties of the sensitivity function . . . . . . . . . . . . . . . 343
8.5.2 Quantitative performance measures . . . . . . . . . . . . . . . . . . . 344
Frequency-domain performance specifications . . . . . . . . . . . . . . . . . 346
8.6.1 Natural frequency-domain specifications . . . . . . . . . . . . . . . . 346
8.6.2 Turning time-domain specifications into frequency-domain specifications350
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 350

Design limitations for feedback control

9.1 Performance restrictions in the time-domain for general systems . . . . . . .
9.2 Performance restrictions in the frequency domain for general systems . . . .
9.2.1 Bode integral formulae . . . . . . . . . . . . . . . . . . . . . . . . . .
9.2.2 Bandwidth constraints . . . . . . . . . . . . . . . . . . . . . . . . . .
9.2.3 The waterbed effect . . . . . . . . . . . . . . . . . . . . . . . . . . .
9.2.4 Poisson integral formulae . . . . . . . . . . . . . . . . . . . . . . . .
9.3 The robust performance problem . . . . . . . . . . . . . . . . . . . . . . . .
9.3.1 Performance objectives in terms of sensitivity and transfer functions
9.3.2 Nominal and robust performance . . . . . . . . . . . . . . . . . . . .
9.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Controller design

10 Stabilisation and state estimation

10.1 Stabilisability and detectability . . . . . . . . . . . . . . . . . . . . . . . . .
10.1.1 Stabilisability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10.1.2 Detectablilty . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10.1.3 Transfer function characterisations of stabilisability and detectability
10.2 Methods for constructing stabilising control laws . . . . . . . . . . . . . . .
10.2.1 Stabilising static state feedback controllers . . . . . . . . . . . . . . .
10.2.2 Stabilising static output feedback controllers . . . . . . . . . . . . .
10.2.3 Stabilising dynamic output feedback controllers . . . . . . . . . . . .
10.3 Parameterisation of stabilising dynamic output feedback controllers . . . . .
10.3.1 More facts about RH+
. . . . . . . . . . . . . . . . . . . . . . . . .

10.3.2 The Youla parameterisation . . . . . . . . . . . . . . . . . . . . . . .

10.4 Strongly stabilising controllers . . . . . . . . . . . . . . . . . . . . . . . . .
10.5 State estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10.5.1 Observers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10.5.2 Luenberger observers . . . . . . . . . . . . . . . . . . . . . . . . . . .
10.5.3 Static state feedback, Luenberger observers, and dynamic output feedback . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

361
361
368
369
374
375
377
381
382
385
392

395
397
398
398
400
403
405
406
408
416
420
421
423
429
429
429
431
433
438

xiii
11 Ad hoc methods I: The root-locus method
11.1 The root-locus problem, and its role in control . . . . . . . . . . . . . .
11.1.1 A collection of problems in control . . . . . . . . . . . . . . . . .
11.1.2 Definitions and general properties . . . . . . . . . . . . . . . . .
11.2 Properties of the root-locus . . . . . . . . . . . . . . . . . . . . . . . . .
11.2.1 A rigorous discussion of the root-locus . . . . . . . . . . . . . . .
11.2.2 The graphical method of Evans . . . . . . . . . . . . . . . . . . .
11.3 Design based on the root-locus . . . . . . . . . . . . . . . . . . . . . . .
11.3.1 Location of closed-loop poles using root-locus . . . . . . . . . . .
11.3.2 Root sensitivity in root-locus . . . . . . . . . . . . . . . . . . . .
11.4 The relationship between the root-locus and the Nyquist contour . . . .
11.4.1 The symmetry between gain and frequency . . . . . . . . . . . .
11.4.2 The characteristic gain and the characteristic frequency functions

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

447
448
448
449
451
451
460
461
461
462
463
463
465

12 Ad hoc methods II: Simple frequency response methods for controller

469
design
12.1 Compensation in the frequency domain . . . . . . . . . . . . . . . . . . . . 469
12.1.1 Lead and lag compensation . . . . . . . . . . . . . . . . . . . . . . . 470
12.1.2 PID compensation in the frequency domain . . . . . . . . . . . . . . 473
12.2 Design using controllers of predetermined form . . . . . . . . . . . . . . . . 475
12.2.1 Using the Nyquist plot to choose a gain . . . . . . . . . . . . . . . . 476
12.2.2 A design methodology using lead and integrator compensation . . . . 477
12.2.3 A design methodology using PID compensation . . . . . . . . . . . . 483
12.2.4 A discussion of design methodologies . . . . . . . . . . . . . . . . . . 486
12.3 Design with open controller form . . . . . . . . . . . . . . . . . . . . . . . . 488
12.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 488
13 Advanced synthesis, including PID synthesis
13.1 Ziegler-Nichols tuning for PID controllers . . . . . . . .
13.1.1 First method . . . . . . . . . . . . . . . . . . . .
13.1.2 Second method . . . . . . . . . . . . . . . . . . .
13.1.3 An application of Ziegler-Nicols tuning . . . . . .
13.2 Synthesis using pole placement . . . . . . . . . . . . . .
13.2.1 Pole placement using polynomials . . . . . . . .
13.2.2 Enforcing design considerations . . . . . . . . . .
13.2.3 Achievable poles using PID control . . . . . . . .
13.3 Two controller configurations . . . . . . . . . . . . . . .
13.3.1 Implementable transfer functions . . . . . . . . .
13.3.2 Implementations that meet design considerations
13.4 Synthesis using controller parameterisation . . . . . . .
13.4.1 Properties of the Youla parameterisation . . . . .
13.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . .
14 An introduction to H2 optimal control
14.1 Problems in optimal control and optimal
14.1.1 Optimal feedback . . . . . . . .
14.1.2 Optimal state estimation . . . .
14.2 Tools for H2 optimisation . . . . . . . .

xiv

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

493
493
494
494
496
500
500
507
513
516
517
521
521
521
522

state estimation
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .

.
.
.
.

525
526
526
528
530

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

531
531
532
535
536
539
541
541
544
548
550
550
550
550
550
551
553
553

15 An introduction to H control theory

15.1 Reduction of robust performance problem to model matching problem
15.1.1 A modified robust performance problem . . . . . . . . . . . . .
15.1.2 Algorithm for reduction to model matching problem . . . . . .
15.1.3 Proof that reduction procedure works . . . . . . . . . . . . . .
15.2 Optimal model matching I. Nevanlinna-Pick theory . . . . . . . . . . .
15.2.1 Picks theorem . . . . . . . . . . . . . . . . . . . . . . . . . . .
15.2.2 An inductive algorithm for solving the interpolation problem .
15.2.3 Relationship to the model matching problem . . . . . . . . . .
15.3 Optimal model matching II. Neharis Theorem . . . . . . . . . . . . .
15.3.1 Hankel operators in the frequency-domain . . . . . . . . . . . .
15.3.2 Hankel operators in the time-domain . . . . . . . . . . . . . . .
15.3.3 Hankel singular values and Schmidt pairs . . . . . . . . . . . .
15.3.4 Neharis Theorem . . . . . . . . . . . . . . . . . . . . . . . . .
15.3.5 Relationship to the model matching problem . . . . . . . . . .
15.4 A robust performance example . . . . . . . . . . . . . . . . . . . . . .
15.5 Other problems involving H methods . . . . . . . . . . . . . . . . . .

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

557
558
558
560
562
566
566
568
570
570
571
573
576
578
580
580
580

14.3

14.4

14.5

14.6

14.2.1 An additive factorisation for rational functions . .

14.2.2 The inner-outer factorisation of a rational function
14.2.3 Spectral factorisation for polynomials . . . . . . .
14.2.4 Spectral factorisation for rational functions . . . .
14.2.5 A class of path independent integrals . . . . . . . .
14.2.6 H2 model matching . . . . . . . . . . . . . . . . .
Solutions of optimal control and state estimation problems
14.3.1 Optimal control results . . . . . . . . . . . . . . .
14.3.2 Relationship with the Riccati equation . . . . . . .
14.3.3 Optimal state estimation results . . . . . . . . . .
The linear quadratic Gaussian controller . . . . . . . . . .
14.4.1 LQR and pole placement . . . . . . . . . . . . . .
14.4.2 Frequency domain interpretations . . . . . . . . . .
14.4.3 H2 model matching and LQG . . . . . . . . . . . .
Stability margins for optimal feedback . . . . . . . . . . .
14.5.1 Stability margins for LQR . . . . . . . . . . . . . .
14.5.2 Stability margins for LQG . . . . . . . . . . . . . .
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . .

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

Background material

A Linear algebra
A.1 Vector spaces and subspaces . . . .
A.2 Linear independence and bases . .
A.3 Matrices and linear maps . . . . .
A.3.1 Matrices . . . . . . . . . .
A.3.2 Some useful matrix lemmas
A.3.3 Linear maps . . . . . . . .
A.4 Change of basis . . . . . . . . . . .

583
.
.
.
.
.
.
.

.
.
.
.
.
.
.

585
585
586
587
587
589
591
592

xv
A.5 Eigenvalues and eigenvectors . . . . . . . . . . . . . . . . . . . . . . . . . . 593
A.6 Inner products . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 594
B Ordinary differential equations
597
B.1 Scalar ordinary differential equations . . . . . . . . . . . . . . . . . . . . . . 597
B.2 Systems of ordinary differential equations . . . . . . . . . . . . . . . . . . . 600
C Polynomials and rational functions
609
C.1 Polynomials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 609
C.2 Rational functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 611
D Complex variable theory
D.1 The complex plane and its subsets . . . . .
D.2 Functions . . . . . . . . . . . . . . . . . . .
D.3 Integration . . . . . . . . . . . . . . . . . .
D.4 Applications of Cauchys Integral Theorem
D.5 Algebraic functions and Riemann surfaces .
E

.
.
.
.
.

617
617
618
620
621
623

Fourier and Laplace transforms

E.1 Delta-functions and distributions . . . . . . .
E.1.1 Test functions . . . . . . . . . . . . .
E.1.2 Distributions . . . . . . . . . . . . . .
E.2 The Fourier transform . . . . . . . . . . . . .
E.3 The Laplace transform . . . . . . . . . . . . .
E.3.1 Laplace transforms of various flavours
E.3.2 Properties of the Laplace transform .
E.3.3 Some useful Laplace transforms . . . .

.
.
.
.
.
.
.
.

627
627
627
629
630
632
632
635
638

xvi

This version: 22/10/2004

1 An introduction to linear control theory

22/10/2004

d(t)
e(t)

r(t)

Chapter 1

u(t)

controller

plant

y(t)

s(t)

An introduction to linear control theory

sensor

With this book we will introduce you to the basics ideas of control theory, and the setting
will be that of single-input, single-output (SISO), finite-dimensional, time-invariant, linear
systems. In this section we will begin to explore the meaning of this lingo, and look at
some simple physical systems which fit into this category. Traditional introductory texts in
control may contain some of this material in more detail [see, for example Dorf and Bishop
2001, Franklin, Powell, and Emani-Naeini 1994]. However, our presentation here is intended
to be more motivational than technical. For proper background in physics, one should look
to suitable references. A very good summary reference for the various methods of deriving
equations for physical systems is [Cannon, Jr. 1967].

Figure 1.1 A basic control system schematic

trajectory r(t). Another problem arises with plant uncertainties. One models the plant,
typically via differential equations, but these are always an idealisation of the plants actual
behaviour. The reason for the problems is that the open-loop control law has no idea what
the output is doing, and it marches on as if everything is working according to an idealised
model, a model which just might not be realistic. A good way to overcome these difficulties
is to use feedback . Here the output is read by sensors, which may themselves be modelled
by differential equations, which produce a signal s(t) which is subtracted from the reference
trajectory to produce the error e(t). The controller then make its decisions based on the
error signal, rather than just blindly considering the reference signal.

Contents
1.1

Some control theoretic terminology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1.2

An introductory example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1.3

Linear differential equations for physical devices . . . . . . . . . . . . . . . . . . . . . . .

1.3.1

Mechanical gadgets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1.3.2

Electrical gadgets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1.3.3

Electro-mechanical gadgets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1.4

Linearisation at equilibrium points . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1.5

What you are expected to know . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1.6

Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1.2 An introductory example

Lets see how this all plays out in a simple example. Suppose we have a DC servo
motor whose output is its angular velocity (t), the input is a voltage E(t), and there is a
disturbance torque T (t) resulting from, for example, an unknown external load being applied
to the output shaft of the motor. This external torque is something we cannot alter. A little
later in this section we will see some justification for the governing differential equations to
be given by
d(t) 1
+ (t) = kE E(t) + kT T (t).
dt
The schematic for the situation is shown in Figure 1.2. This schematic representation we
T (t)

1.1 Some control theoretic terminology

For this book, there should be from the outset a picture you have in mind of what you
are trying to accomplish. The picture is essentially given in Figure 1.1. The idea is that
you are given a plant, which is the basic system, which has an output y(t) that youd
like to do something with. For example, you may wish to track a reference trajectory
r(t). One way to do this would be to use an open-loop control design. In this case, one
would omit that part of the diagram in Figure 1.1 which is dashed, and use a controller
to read the reference signal r(t) and use this to specify an input u(t) to the plant which
should give the desired output. This open-loop control design may well work, but it has
some inherent problems. If there is a disturbance d(t) which you do not know about,
then this may well cause the output of the plant to deviate significantly from the reference

E(t)

u(t)

d
1
+
=u
dt

Figure 1.2 DC motor open-loop control schematic

(t)

1N,D (t)

(t)

give
herereplacements
is one we shall use frequently, and it is called a block diagram.1
PSfrag
Let us just try something nave and open-loop. The objective is to be able to drive the
ke(t)k constant velocity 0 . This constant desired output is then our reference
motor at a specified
e1 (t)
trajectory. You
decide to see what you might do by giving the motor some constant torques
e2 (t)
to see what happens.
Let us provide a constant input torque E(t) = E0 and suppose that
e3 (t)
the disturbancey(t)
torque T (t) = 0. We then have the differential equation
u(t)
h (t)
d 1
+ = kE E0 .
hN,D (t)
dt
1 (t)
N,D
Supposing, 1as
is(t)
reasonable, that the motor starts with zero initial velocity, i.e., (0) = 0,
the solution to the
initial value problem is
(t)

fj (t)
(t) = kE E0 1 et/ .
Re
Im
1
1
We give a numerical
x1 plot for kE = 2, E0 = 3, and = 2 in Figure 1.3. Well, we say, this all
x2
x1
x2 14
log 12
dB
deg 10
u = 0 ln
ln coth(|u| /2) 8
or 1
6
m

4
yos

2
tos 0

2
4
6
8
10
,0
t
0
Figure 1.3 Open-loop response of DC motor

looks too easy. To get the desired output velocity 0 after a sufficiently long time, we need
only provide the input voltage E0 = k0E .
However, there are decidedly problems lurking beneath the surface. For example, what
if there is a disturbance torque? Let us suppose this to be constant for the moment so
T (t) = T0 for some T0 > 0. The differential equation is then
d 1
+ = kE E0 kT T0 ,
dt
and if we again suppose that (0) = 0 the initial value problem has solution

(t) = (kE E0 kT T0 ) 1 et/ .
If we follow our simple rule of letting the input voltage E0 be determined by the desired final
angular velocity by our rule E0 = kE0 , then we will undershoot our desired final velocity
by error = kT T0 . In this event, the larger is the disturbance torque, the worse we doin

22/10/2004

(t)

1.2 An introductory example

22/10/2004

(t)
fj (t)
Re
Im
x1 1 An introduction to linear control theory
x2
x1
x2 14
log 12
dB
deg 10
u = 0 ln
ln coth(|u| /2) 8
or 1
6
m

4
yos

2
tos 0

2
4
6
8
,0
t
0

Figure 1.4 Open-loop response of DC motor with disturbance

fact, we can do pretty darn bad if the disturbance torque is large. The effect is illustrated
in Figure 1.4 with kT = 1 and T0 = 2.
Another problem arises when we have imperfect knowledge of the motors physical characteristics. For example, we may not know the time-constant as accurately as wed like.
While we estimate it to be , it might be some other value . In this case, the actual
differential equation governing behaviour in the absence of disturbances will be
d 1
+ = kE E0 ,
dt
which gives the solution to the initial value problem as

(t) = kE E0 1 et/ .
The final value will then be in error by the factor . This situation is shown in Figure 1.5
for 1 = 58 .
Okay, I hope now that you can see the problem with our open-loop control strategy. It
simply does not account for the inevitable imperfections we will have in our knowledge of
the system and of the environment in which it works. To take all this into account, let us
measure the output velocity of the motors shaft with a tachometer. The tachometer takes
the angular velocity and returns a voltage. This voltage, after being appropriately scaled by
a factor ks , is then compared to the voltage needed to generate the desired velocity by feeding
it back to our reference signal by subtracting it to get the error. The error we multiply by
some constant K, called the gain for the controller, to get the actual voltage input to the
system. The schematic now becomes that depicted in Figure 1.6. The differential equations
governing this system are
d
+ ( 1 + kE kS K) = kE Kref + kT T.
dt
We shall see how to systematically obtain equations such as this, so do not worry if you
think it in nontrivial to get these equations. Note that the input to the system is, in some
1
When we come to use block diagrams for real, you will see that the thing in the blocks are not differential
equations in the time-domain, but in the Laplace transform domain.

1N,D (t)

(t)
fj (t)
Re
Im
22/10/2004
x1
x2
x1
x2 14
log 12
dB
deg 10
u = 0 ln
ln coth(|u| /2) 8
or 1
6
m

4
yos

2
tos 0

Figure 1.5 Open-loop response of DC motor with actual motor

time-constant
T (t)
plant
kT

controller

u(t)

d
1
+
=u
dt

(t)

sensor
kS

Figure 1.6 DC motor closed-loop control schematic

(t)
significantly from
the open-loop controller (cf. Figure 1.3). It is possible to remove the final
f (t)
error by doing jsomething more sophisticated with the error than multiplying it by K, but
Re
we will get to that
Im only as the course progresses. Now lets see what happens when we
x1
add a constant disturbance
by setting T0 = 2. The result is displayed in Figure 1.8. We
x2
x1
x2 14
log 12
dB
deg 10
u = 0 ln
ln coth(|u| /2) 8
or 1
6
m

4
yos

2
tos 0

2
4
6
8
10
,0
t
0
(t)

,0
0

ref

22/10/2004

(t)

1.2 An introductory example

(t)
fj (t)
Re
Im
6
x1 1 An introduction to linear control theory
x2
x1
x2 14
log 12
dB
deg 10
u = 0 ln
ln coth(|u| /2) 8
or 1
PSfrag replacements
6
m
ke(t)k
4
os
e1y(t)

2
e2 (t)
teos3 (t)
0

2
4
6
8
y(t)
,0
u(t)
t
0
h (t)
hN,D (t)
1 (t)
Figure 1.7 Closed-loop response of DC motor
1N,D (t)

Figure 1.8 Closed-loop response of DC motor with disturbance

sense, no longer the voltage, but the reference signal ref . Let us suppose again a constant
disturbance torque T (t) = T0 and a constant reference voltage ref = 0 . The solution to
the differential equation, again supposing (0) = 0, is then
(t) =

1
kE K0 kT T0
1 e( +kE kS K)t .
1
+ kE kS K

Let us now investigate this closed-loop control law. As previously, let us first look at the
case when T0 = 0 and where we suppose perfect knowledge of our physical constants and
our model. In this case, we wish to achieve a final velocity of 0 = E0 kE as t , i.e., the
same velocity as we had attained with our open-loop strategy. We see the results of this
in Figure 1.7 where we have chosen kS = 1 and K = 5. Notice that the motor no longer
achieves the desired final speed! However, we have improved the response time for the system

see that the closed-loop controller reacts much better to the disturbance (cf. Figure 1.4),
although we still (unsurprisingly) cannot reach the desired final velocity. Finally we look at
the situation when we have imperfect knowledge of the physical constants for the plant. We
again consider having 1 = 58 rather than the guessed value of 12 . In this case the closed-loop
response is shown in Figure 1.9. Again, the performance is somewhat better than that of
the open-loop controller (cf. Figure 1.5), although we have incurred a largish final error in
the final velocity.
I hope this helps to convince you that feedback is a good thing! As mentioned above, we
shall see that it is possible to design a controller so that the steady-state error is zero, as this
is the major deficiency of our very basic controller designed above. This simple example,
however, does demonstrate that one can achieve improvements in some areas (response time

1N,D (t)

1.3 Linear differential equations for physical devices

,0
0

1 An introduction to linear control theory

22/10/2004

he would reply, Well, F = ma, now go think on it.2 After doing so youd arrive at
m
y (t) = F (t) ky(t) dy(t)

m
y (t) + dy(t)
+ ky(t) = F (t).

This is a second-order linear differential equation with constant coefficients and with inhomogeneous term F (t).
The same sort of thing happens with rotary devices. In Figure 1.11 is a rotor fixed to
a shaft moving with angular velocity . Viscous dissipation may be modelled with a force

(t)

(t)
fj (t)
Re
Im
22/10/2004
x1
x2
x1
x2 14
log 12
dB
deg 10
u = 0 ln
ln coth(|u| /2) 8
or 1
6
m

4
yos

2
tos 0

(t)
2

Figure 1.9 Closed-loop response of DC motor with actual motor

time-constant

in this case), although sometimes at the expense of deterioration in others (steady-state

error in this case).

1.3 Linear differential equations for physical devices

We will be considering control systems where the plant, the controller, and the sensors
are all modelled by linear differential equations. For this reason it makes sense to provide
some examples of devices whose behaviour is reasonably well-governed by such equations.
The problem of how to assemble such devices to, say, build a controller for a given plant is
something we will not be giving terribly much consideration to.
1.3.1 Mechanical gadgets In Figure 1.10 is a really feeble idealisation of a car suspension system. We suppose that at y = 0 the mass m is in equilibrium. The spring, as we
F (t)

y(t)

Figure 1.11 Rotor on a shaft

proportional to the angular velocity: Fd = d. In the presence of an external torque (t),

the governing differential equation is
J (t)

+ d(t) = (t)
where J is the moment of inertia of the rotor about its point of rotation. If one wishes
to include a rotary spring, then one must consider not the angular velocity (t) as the
dependent variable, but rather the angular displacement (t) = 0 + t. In either case, the
governing equations are linear differential equations.
Lets look at a simple pendulum (see Figure 1.12). If we sum moments about the pivot
we get
m`2 = mg` sin
=
+ g` sin = 0.
Now this equation, you will notice, is nonlinear. However, we are often interested in the
= (0 , 0) where 0
behaviour of the system near the equilibrium points which are (, )
{0, }. So, let us linearise the equations near these points, and see what we get. We write
the solution near the equilibrium as (t) = 0 + (t) where (t) is small. We then have
+ g` sin = + g` sin(0 + ) = + g` sin 0 cos + g` cos 0 sin .

Figure 1.10 Simplified automobile suspension

know, then supplies a restoring force Fk = ky and the dashpot supplies a force Fd = dy,

d
where means dt
. We also suppose there to be an externally applied force F (t). If we
ask Isaac Newton, Isaac, what are the equations governing the behaviour of this system?

Now note that sin 0 = 0 if 0 {0, }, and cos 0 = 1 if 0 = 0 and cos 0 = 1 if 0 = .

3
5
We also use the Taylor expansion for sin x around x = 0: sin x = x x3! + x5! + . . . . Keeping
2
This is in reference to the story, be it true or false, that when Newton was asked how hed arrived at
the inverse square law for gravity, he replied, I thought on it.

22/10/2004

1.3 Linear differential equations for physical devices

1 An introduction to linear control theory

22/10/2004

1.3.2 Electrical gadgets A resistor is a device across which the voltage drop is
proportional to the current through the device. A capacitor is a device across which the
voltage drop is proportional to the charge in the device. An inductor is a device across
which the voltage drop is proportional to the time rate of change of current through the
device. The three devices are typically given the symbols as in Figure 1.13. The quantity

I(t)

q(t)

E = RI

1
Cq

I(t)

E = L dI
dt

Figure 1.12 A simple pendulum

only the lowest order terms gives the following equations which should approximate the
behaviour of the system near the equilibrium (0 , 0):
+ g` = 0, 0 = 0
g` = 0, 0 = .
In each case, we have a linear differential equation which governs the behaviour near the equilibrium. This technique of linearisation is ubiquitous since there really are no linear physical
devices, but linear approximations seem to work well, and often very well, particularly in
control. We discuss linearisation properly in Section 1.4.
Let us recall the basic rules for deriving the equations of motion for a mechanical system.

Resistor

Capacitor

Inductor

Figure 1.13 Electrical devices

R is called the resistance of the resistor, the quantity C is called the capacitance of
the capacitor, and the quantity L is called the inductance of the inductor. Note that
the proportionality constant for the capacitor is not C but C1 . The current I is related to
. We can then imagine assembling these electrical components in
the charge q by I = dq
dt
some configuration and using Kirchhoffs laws3 to derive governing differential equations. In
Figure 1.14 we have a particularly simple configuration. The voltage drop around the circuit
R

1.1 Deriving equations for mechanical systems Given: an interconnection of point masses and
rigid bodies.
1. Define a reference frame from which to measure distances.
2. Choose a set of coordinates that determine the configuration of the system.
3. Separate the system into its mechanical components. Thus each component should be
either a single point mass or a single rigid body.
4. For each component determine all external forces and moments acting on it.
5. For each component, express the position of the centre of mass in terms of the chosen
coordinates.
6. The sum of forces in any direction on a component should equal the mass of the component times the component of acceleration of the component along the direction of the
force.
7. For each component, the sum of moments about a point that is either (a) the centre of
mass of the component or (b) a point in the component that is stationary should equal
the moment of inertia of the component about that point multiplied by the angular
acceleration.

This methodology has been applied to the examples above, although they are too simple
to be really representative. We refer to the exercises for examples that are somewhat more
interesting. Also, see [Cannon, Jr. 1967] for details on the method we outline, and other
methods.

+
E

Figure 1.14 A series RLC circuit

must be zero which gives the governing equations

+ 1 q(t)
E(t) = RI(t) + LI(t)
C

L
q (t) + Rq(t)
+ C1 q(t) = E(t)

where E(t) is an external voltage source. This may also be written as a current equation by
merely differentiating:
+ RI(t)
+ 1 I(t) = E(t).

LI(t)
C
In either case, we have a linear equation, and again one with constant coefficients.
Let us present a methodology for determining the differential equations for electric circuits. The methodology relies on the notion of a tree which is a connected collection of
branches containing no loops. For a given tree, a tree branch is a branch in the tree, and
a link is a branch not in the tree.
3
Kirchhoff s voltage law states that the sum of voltage drops around a closed loop must be zero and
Kirchhoff s current law states that the sum of the currents entering a node must be zero.

22/10/2004

1.3 Linear differential equations for physical devices

1.2 Deriving equations for electric circuits Given: an interconnection of ideal resistors, capacitors, and inductors, along with voltage and current sources.
1. Define a tree by collecting together a maximal number of branches to form a tree. Add
elements in the following order of preference: voltage sources, capacitors, resistors, inductors, and current sources. That is to say, one adds these elements in sequence until
one gets the largest possible tree.
2. The states of the system are taken to be the voltages across capacitors in the tree branches
for the tree of part 1 and the currents through inductors in the links for the tree from
part 1.
3. Use Kirchhoffs Laws to derive equations for the voltage and current in every tree branch
in terms of the state variables.
4. Write the Kirchhoff Voltage Law and the Kirchhoff Current Law for every loop and every
node corresponding to a branch assigned a state variable.

The exercises contain a few examples that can be used to test ones understanding of the
above method. We also refer to [Cannon, Jr. 1967] for further discussion of the equations
governing electrical networks.
1.3.3 Electro-mechanical gadgets If you really want to learn how electric motors
work, then read a book on the subject. For example, see [Cannon, Jr. 1967].
A DC servo motor works by running current through a rotary toroidal coil which sits in
a stationary magnetic field. As current is run through the coil, the induced magnetic field
induces the rotor to turn. The torque developed is proportional to the current through the
coil: T = Kt I where T is the torque supplied to the shaft, I is the current through the
coil, and Kt is the torque constant. The voltage drop across the motor is proportional
to the motors velocity; Em = Ke where Em is the voltage drop across the motor, Ke is a
constant, and is the angular position of the shaft. If one is using a set of consistent units
with velocity measured in rads/sec, then apparently Ke = Kt .
Now we suppose that the rotor has inertia J and that shaft friction is viscous and so the
Thus the motor will be governed by Newtons equations:
friction force is given by d.
J = d + Kt I

J + d = Kt I.

To complete the equations, one need to know the relationship between current and . This
is provided by the relation Em = Ke and the dynamics of the circuit which supplies current
to the motor. For example, if the circuit has resistance R and inductance L then we have
L

dI
+ RI = E Ke
dt

with E being the voltage supplied to the circuit. This gives us coupled equations
J + d = Kt I
dI
L + RI = E Ke
dt
which we can write in first-order system form as

0
1
0

K
d
t
v = 0
v + 0 E
J
J
1
I
0 KLe R
I
L
L

1 An introduction to linear control theory

22/10/2004

If the response of the circuit is much faster

where we define the dependent variable v = .
than that of the motor, e.g., if the inductance is small, then this gives E = Ke + RI and
so the equations reduce to

J + d + KtRKe = KRt E.
Thus the dynamics of a DC motor can be roughly described by a first-order linear differential
equation in the angular velocity. This is what we saw in our introductory example.
Hopefully this gives you a feeling that there are a large number of physical systems which
are modelled by linear differential equations, and it is these to which we will be restricting
our attention.

1.4 Linearisation at equilibrium points

When we derived the equations of motion for the pendulum, the equations we obtained
were nonlinear. We then decided that if we were only interested in looking at what is going
on near an equilibrium point, then we could content ourselves with linearising the equations.
We then did this in a sort of hacky way. Lets see how to do this methodically.
We suppose that we have vector differential equations of the form
x 1 = f1 (x1 , . . . , xn )
x 2 = f2 (x1 , . . . , xn )
..
.
x n = fn (x1 , . . . , xn ).
The n functions (f1 , . . . , fn ) of the n variables (x1 , . . . , xn ) are known smooth functions. Let
us denote x = (x1 , . . . , xn ) and f (x) = (f1 (x), . . . , fn (x)). The differential equation can
then be written as
x = f (x).
(1.1)
It really only makes sense to linearise about an equilibrium point. An equilibrium point
is a point x0 Rn for which f (x0 ) = 0. Note that the constant function x(t) = x0 is a
solution to the differential equation if x0 is an equilibrium point. For an equilibrium point
x0 define an n n matrix Df (x0 ) by

f1
f1
f1
(x0 )
(x0 ) x
(x0 ) x
x1
n
2
f2 (x0 ) f2 (x0 ) f2 (x0 )

x2
xn
.
Df (x0 ) = x1 .
..
..
..
..
.
.
.
fn
(x0 )
x1

fn
(x0 )
x2

fn
(x0 )
xn

This matrix is often called the Jacobian of f at x0 . The linearisation of (1.1) about an
equilibrium point x0 is then the linear differential equation
= Df (x0 ).
Lets see how this goes with our pendulum example.
1.3 Example The nonlinear differential equation we derived was
+ g` sin = 0.

22/10/2004

1.5 What you are expected to know

This is not in the form of (1.1) since it is a second-order equation. But we can put this into
The equations can then be
first-order form by introducing the variables x1 = and x2 = .
written
x 1 = = x2
x 2 = = g` sin = g` sin x1 .
Thus

g
f2 (x1 , x2 ) = sin x1 .
`
Note that at an equilibrium point we must have x2 = 0. This makes sense as it means
that the pendulum should not be moving. We must also have sin x1 = 0 which means that
x1 {0, }. This is what we determined previously.
Now let us linearise about each of these equilibrium points. For an arbitrary point
x = (x1 , x2 ) we compute

0
1
.
Df (x) =
g` cos x1 0
f1 (x1 , x2 ) = x2 ,

1 An introduction to linear control theory

22/10/2004

I have used Mathematica to do all the plotting in the book since it is what I am familiar
with. Also available are Maple and Matlab . Matlab has a control toolbox, and is the
most commonly used tool for control systems.4
You are encouraged to use symbolic manipulation packages for doing problems in this
book. Just make sure you let us know that you are doing so, and make sure you know what
you are doing and that you are not going too far into black box mode.

1.6 Summary
Our objective in this chapter has been to introduce you to some basic control theoretic
ideas, especially through the use of feedback in the DC motor example. In the remainder of
these notes we look at linear systems, and to motivate such an investigation we presented
some physical devices whose behaviour is representable by linear differential equations, perhaps after linearisation about a desired operating point. We wrapped up the chapter with a
quick summary of the background required to proceed with reading these notes. Make sure
you are familiar with everything discussed here.

At the equilibrium point x1 = (0, 0) we thus have

0 1
Df (x1 ) =
,
g
` 0
and at the equilibrium point x2 = (0, ) we thus have

0 1
Df (x1 ) = g
.
0
`
With these matrices at hand, we may write the linearised equations at each equilibrium
point.

1.5 What you are expected to know

There are five essential areas of background that are assumed of a student using this
text. These are
1. linear algebra,
2. ordinary differential equations, including the matrix exponential,
3. basic facts about polynomials,
4. basic complex analysis, and
5. transform theory, especially Fourier and Laplace transforms.
Appendices review each of these in a cursory manner. Students are expected to have seen
this material in detail in previous courses, so there should be no need for anything but rapid
review in class.
Many of the systems we will look at in the exercises require in their analysis straightforward, but tedious, calculations. It should not be the point of the book to make you go
through such tedious calculations. You will be well served by learning to use a computer
package for doing such routine calculations, although you should try to make sure you are
asking the computer to do something which you in principle understand how to do yourself.

4
Mathematica and Maple packages have been made available on the world wide web for doing things
such as are done in this book. See http://mast.queensu.ca/~math332/.

Exercises for Chapter 1

1 An introduction to linear control theory

Exercises
E1.1 Probe your life for occurrences of things which can be described, perhaps roughly, by
a schematic like that of Figure 1.1. Identify the components in your system which
are the plant, output, input, sensor, controller, etc. Does your system have feedback?
Are there disturbances?
E1.2 Consider the DC servo motor example which we worked with in Section 1.2. Determine
conditions on the controller gain K so that the voltage E0 required to obtain a desired
steady-state velocity is greater for the closed-loop system than it is for the openloop system. You may neglect the disturbance torque, and assume that the motor
model is accurate. However, do not use the numerical values used in the notesleave
everything general.
E1.3 An amplifier is to be designed with an overall amplification factor of 2500 50. A
number of single amplifier stages is available and the gain of any single stage may
drift anywhere between 25 and 75. The configuration of the final amplifier is given in
Figure E1.1. In each of the blocks we get to insert an amplifier stage with the large

22/10/2004

Figure E1.2 Coupled masses

Vin

Vout

Figure E1.3 Pendulum on a cart

Figure E1.1 A multistage feedback amplifier

and unknown gain variation (in this case the gain variation is at most 50). Thus the
gain in the forward path is K1 K2 KN where N is the number of amplifier stages
and where Ki [25, 75]. The element in the feedback path is a constant 0 < < 1.
(a) For N amplifier stages and a given value for determine the relationship between
Vin and Vout .
The feedback gain is known precisely since it is much easier to design a circuit
which provides accurate voltage division (as opposed to amplification). Thus, we can
assume that can be exactly specified by the designer.
(b) Based on this information find a value of in the interval (0, 1) and, for that
value of , the minimal required number of amplifier stages, Nmin , so that the
final amplifier design meets the specification noted above.
E1.4 Derive the differential equations governing the behaviour of the coupled masses in
Figure E1.2. How do the equations change if viscous dissipation is added between
each mass and the ground? (Suppose that both masses are subject to the same
dissipative force.)
The following two exercises will recur as exercises in succeeding chapters. Since the computations can be a bit involvedcertainly they ought to be done with a symbolic manipulation
packageit is advisable to do the computations in an organised manner so that they may
be used for future calculations.

E1.5 Consider the pendulum on a cart pictured in Figure E1.3. Derive the full equations
which govern the motion of the system using coordinates (x, ) as in the figure. Here
M is the mass of the cart and m is the mass of the pendulum. The length of the
pendulum arm is `. You may assume that there is no friction in the system. Linearise
= (x0 , 0, 0, 0) and (x, , x,
= (x0 , , 0, 0),
the equations about the points (x, , x,
)
)
where x0 is arbitrary.
E1.6 Determine the full equations of motion for the double pendulum depicted in Figure E1.4. The first link (i.e., the one connected to ground) has length `1 and mass

Figure E1.4 Double pendulum

m1 , and the second link has length `2 and mass m2 . The links have a uniform mass
density, so their centres of mass are located at their midpoint. You may assume that

Exercises for Chapter 1

there is no friction in the system. What are the equilibrium points for the double
pendulum (there are four)? Linearise the equations about each of the equilibria.
E1.7 Consider the electric circuit of Figure E1.5. To write equations for this system we

1 An introduction to linear control theory

22/10/2004

E1.10 The mass flow rate from a tank of water with a uniform cross-section can be roughly
modelled as being proportional to the height of water in the tank which lies above
the exit nozzle. Suppose that two tanks are configured as in Figure E1.8 (the tanks

R
+
E

Figure E1.5 Electric circuit

need to select two system variables. Using IC , the current through the capacitor, and
IL , the current through the inductor, derive a first-order system of equations in two
variables governing the behaviour of the system.
E1.8 For the circuit of Figure E1.6, determine a differential equation for the current through

are not necessarily identical). Determine the equations of motion which give the mass
flow rate from the bottom tank given the mass flow rate into the top tank. In this
problem, you must define the necessary variables yourself.

R1
+
R2

Figure E1.8 Coupled water tanks

Figure E1.6 Another electric circuit

the resistor R1 in terms of the voltage E(t).

E1.9 For the circuit of Figure E1.7, As dependent variable for the circuit, use the volt-

+
E

Figure E1.7 Yet another electric circuit

age across the capacitor and the current through the inductor. Derive differential
equations for the system as a first-order system with two variables.

In the next exercise we will consider a more complex and realistic model of flow in coupled
tanks. Here we will use the Bernoulli equation for flow from small orifices. This says that
if a tank of uniform cross-section has fluid level h, thenthe velocity flowing from a small
nozzle at the bottom of the tank will be given by v = 2gh, where g is the gravitational
acceleration.
E1.11 Consider the coupled tanks shown in Figure E1.9. In this scenario, the input is the
volume flow rate Fin which gets divided between the two tanks proportionally to the
1
areas 1 and 2 of the two tubes. Let us denote = 1+
.
2
(a) Give an expression for the volume flow rates Fin,1 and Fin,2 in terms of Fin and .
Now suppose that the areas of the output nozzles for the tanks are a1 and a2 , and
that the cross-sectional areas of the tanks are A1 and A2 . Denote the water levels in
the tanks by h1 and h2 .
(b) Using the Bernoulli equation above, give an expression for the volume flow rates
Fout,1 and Fout,2 in terms of a1 , a2 , h1 , and h2 .
(c) Using mass balance (assume that the fluid is incompressible so that mass and
volume balance are equivalent), provide two coupled differential equations for the
heights h1 and h2 in the tanks. The equations should be in terms of Fin , , A1 ,
A2 , a1 , a2 , and g, as well as the dependent variables h1 and h2 .
Suppose that the system is in equilibrium (i.e., the heights in the tanks are constant)
with the equilibrium height in tank 1 being 1 .
(d) What is the equilibrium input flow rate ?
(e) What is the height 2 of fluid in tank 2?

Exercises for Chapter 1

Fin
1

Fin,1
A1
a1
Fout,1
Fin,2
A2
a2
Fout,2

Figure E1.9 Another coupled tank scenario

1 An introduction to linear control theory

22/10/2004

This version: 22/10/2004

Part I
System representations and their
properties

This version: 22/10/2004

2.6

Chapter 2
State-space representations (the time-domain)

Observer canonical form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

2.5.3

Canonical forms for uncontrollable and/or unobservable systems . . . . . . . . . .

Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

2.1 Definition A finite-dimensional, time-invariant linear control system is given

by a quadruple = (A, B, C, D) where A Rnn , B Rnm , C Rrn , and D Rrm .
The system is single-input, single-output (SISO) if m = r = 1 and is multi-input,
multi-output (MIMO) otherwise.

Er. . . how the heck is this a control system? Like this:

x(t)
= Ax(t) + Bu(t)
y(t) = Cx(t) + Du(t).

2.1

Properties of finite-dimensional, time-invariant linear control systems . . . . . . . . . . .

2.2

Obtaining linearised equations for nonlinear input/output systems . . . . . . . . . . . . .

2.3

Input/output response versus state behaviour . . . . . . . . . . . . . . . . . . . . . . . . .

2.3.1

Bad behaviour due to lack of observability . . . . . . . . . . . . . . . . . . . . . .

2.3.2

Bad behaviour due to lack of controllability . . . . . . . . . . . . . . . . . . . . . .

2.3.3

Bad behaviour due to unstable zero dynamics . . . . . . . . . . . . . . . . . . . .

2.3.4

A summary of what we have said in this section . . . . . . . . . . . . . . . . . . .

The impulse response . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

2.4.1

The impulse response for causal systems . . . . . . . . . . . . . . . . . . . . . . .

2.4.2

The impulse response for anticausal systems . . . . . . . . . . . . . . . . . . . . .

Canonical forms for SISO systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

2.5.1

Controller canonical form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

2.5.2

With that little bit of linear algebra behind us, we can have at our time-domain formulation. It is in this setting that many models are handed to us in practice, so it is in my
opinion the most basic way to discuss control systems. Here I differ in opinion with most
introductory control texts that place our discussion here late in the course, or do not have
it at all.
We begin by saying what we look at. Our definition here includes the multi-input, multioutput framework since it is easy to do so. However, we will quickly be specialising to the
single-input, single-output situation.

Contents

2.5

22/10/2004

2.1 Properties of finite-dimensional, time-invariant linear control

systems

With that little preamble behind us, let us introduce some mathematics into the subject.
We will approach the mathematical formulation of our class of control systems from three
points of view; (1) time-domain, (2) s-domain or Laplace transform domain, and (3) frequency domain. We will also talk about two kinds of systems; (1) those given to us in linear
system form, and (2) those given to us in input/output form. Each of these possible formulations has its advantages and disadvantages, and can be best utilised for certain types of
analysis or design. In this chapter we concentrate on the time-domain, and we only deal with
systems in linear system form. We will introduce the input/output form in Chapter 3.
Some of the material in this chapter, particularly the content of some of the proofs, is
pitched at a level that may be a bit high for an introductory control course. However, most
students should be able to grasp the content of all results, and understand their implications.
A good grasp of basic linear algebra is essential, and we provide some of the necessary material in Appendix A. The material in this chapter is covered in many texts, including [Brockett
1970, Chen 1984, Kailath 1980, Zadeh and Desoer 1979]. The majority of texts deal with
this material in multi-input, multi-output (MIMO) form. Our presentation is single-input,
single-output (SISO), mainly because this will be the emphasis in the analysis and design
portions of the book. Furthermore, MIMO generalisations to the majority of what we say in
this chapter are generally trivial. The exception is the canonical forms for controllable and
observable systems presented in Sections 2.5.1 and 2.5.2.

2.4

2 State-space representations (the time-domain)

(2.1)

Here x Rn is the state of the system, u Rm is the input, and y Rr is the output. We
call the system finite-dimensional because n < and time-invariant because the matrices
A, B, C, and D are constant. In the single-input, single-output case note that we may
write the equations (2.1) as

x(t)
= Ax(t) + bu(t)
(2.2)
y(t) = ct x(t) + Du(t)
for vectors b, c Rn . Here the matrix D is 1 1 and so is essentially a scalar, and ct
denotes the transpose of c. We will be coming back again and again to the equations (2.2).
They form a large part of what interests us in this book. Note that we will always reserve
the symbol n to denote the state dimension of a SISO linear system. Therefore, from now
on, if you see a seemingly undefined n floating around, it should be the state dimension
of whatever system is being discussed at that moment.
2.2 Example Lets look at a system that can be written in the above form. We consider the
mass-spring-damper system depicted in Figure 1.10. The differential equation governing the
system behaviour is
m
x + dx + kx = u

22/10/2004

2.1 Properties of finite-dimensional, time-invariant linear control systems

where we denote by u(t) the input force. To convert this into a set of equations of the
The governing equations are then
form (2.1) we define x1 = x and x2 = x.
x 1 = x = x2
k
x 2 = x = m
x

d
x
m

1
u
m

k
= m
x1

d
x
m 2

1
u.
m

We can write this in matrix/vector form as

0
1
x 1
x1
0
=
+ 1 u.
k
d
x 2
m m x2
m
So if we define

0
1
k
d ,
m m

0
1
m

1
,
0

2 State-space representations (the time-domain)

22/10/2004

(i) I = (, a),
(vi) I = [a, b),
(ii) I = (, a],
(vii) I = [a, b],
(iii) I = (a, b),
(viii) I = [a, ), or
(iv) I = (a, b],
(ix) I = R.
(v) I = (a, ),
Let I denote the set of intervals. If I I , a map f : I Rk is piecewise continuous
if it is continuous except at a discrete set of points in I,1 and at points of discontinuity, the
left and right limits of the function exist. An admissible input for (2.1) is a piecewise
continuous map u : I Rm where I I , and we denote the set of admissible controls by
U.

we have the first of equations (2.1).

We shall look at three ways in which the output equation may appear in this example.
1. Suppose that with a proximeter we measure the displacement of the mass. Thus we have
the output y = x = x1 that we can write in matrix form as

x1
y= 1 0
x2
so that

D= 0 .

2. The next scenario supposes that we have a means of measuring the velocity of the mass.
Thus we take y = x = x2 . In this case we have

x1
y= 0 1
x2

2.4 Remark All inputs we will encounter in this book will be in fact piecewise infinitely differentiable. However, we will also not be giving the issue too much serious considerationbe
advised, however, that when dealing with control theory at any level of seriousness, the specification of the class of inputs is important. Indeed, one might generally ask that the inputs
be, in the language of Lebesgue integration, essentially bounded and measurable.

Often when dealing with time-invariant systems one makes the assumption of causality
which means that inputs are zero for t < 0. In this book we will often tacitly make the
causality assumption. However, there are brief periods when we will require the opposite of
causality. Thus a system is anticausal when inputs are zero for t > 0.
The following result justifies our calling the system (2.1) linear.
2.5 Proposition Let I I and let u1 , u2 U be defined on I with x1 (t) and x2 (t) defined
as satisfying
x 1 = Ax1 + Bu1 , x 2 = Ax2 + Bu2 ,
and y 1 (t) and y 2 (t) defined by
y 1 (t) = Cx1 (t) + Du1 (t),

y 2 (t) = Cx1 (t) + Du2 (t).

For a1 , a2 R, define u(t) = a1 u1 (t) + a2 u2 (t). Then x(t) , a1 x1 (t) + a2 x2 (t) satisfies

so that
c=

0
,
1

x = Ax + Bu

D= 0 .

3. The final situation will arise when we mount an accelerometer atop the mass so we have
k
x md x + m1 u. In this case we have
y = x = m

k
x1

md
y = m
+ 1 u
x2
so that
c=

k
m
,
md

1
m

In order to be really clear on what we are doing, and in particular to state what we mean
by linearity, we should really specify the class of inputs we consider. Let us do this.
2.3 Definition An interval is a subset I of R of the form

and y(t) , a1 y 1 (t) + a2 y 2 (t) satisfies

y(t) = Cx(t) + Du(t).
Proof We compute
d
(a1 x1 + a2 x2 )
dt
= a1 x 1 + a2 x 2
= a1 (Ax1 + Bu1 ) + a2 (Ax2 + Bu2 )
= A(a1 x1 + a2 x2 ) + B(a1 u1 + a2 u2 )
= Ax + Bu

x =

1
You will recall the notion of a discrete set (or more precisely, it will be recalled for you). For I I , a
(possibly infinite) collection of distinct points P I is called discrete if there exists > 0 so that |x y|
for every x, y I. If I is a bounded set, one verifies that this implies that every discrete set is finite. If I is
not bounded, one wants to ensure that the points cannot get closer and closer together, and in so doing one
ensures that length of the intervals on which the function is continuous always have a lower bound.

22/10/2004

2.1 Properties of finite-dimensional, time-invariant linear control systems

as claimed. We also compute

2 State-space representations (the time-domain)

22/10/2004

We now note that

y = a1 y 1 + a2 y 2
= a1 (Cx1 + Du1 ) + a2 (Cx2 + Du2 )
= C(a1 x1 + a2 x2 ) + D(a1 u1 + a2 u2 )
= Cx + Du,
again, as claimed.

2z t (t)Az(t) = 2
2

n
X
i,j=1
n
X

zi (t)aij zj (t)
kz(t)k max |aij | kz(t)k
i,j

i,j=1

2n max |aij | kz(t)k2 .

The idea is that if we take as our new input a linear combination of old inputs, the same
linear combination of the old states satisfies the control equations, and also the same linear
combination of the old outputs satisfies the control equations.
In Proposition 2.5 we tacitly assumed that the solutions x1 (t) and x2 (t) existed for the
given inputs u1 (t) and u2 (t). Solutions do in fact exist, and we may represent them in a
convenient form.

i,j

Let = 2n2 maxi,j |aij | so that we have

d
(kz(t)k2 ) kz(t)k2 0.
dt
We write
et

2.6 Theorem For u U defined on I I , t0 I, and x0 Rn , there exists a unique

piecewise differentiable curve x : I Rn so that

(kz(t)k2 ) kz(t)k2 0

dt
d t
(e kz(t)k2 ) 0.
dt

x(t)
= Ax(t) + Bu(t),
This can be integrated to give

and x(t0 ) = x0 .
Proof We demonstrate existence by explicitly constructing a solution. Indeed, we claim
that the solution is
Z t
x(t) = eA(tt0 ) x0 +
eA(t ) Bu( ) d.
(2.3)
t0

First, note that the initial conditions are satisfied (just let t = t0 ). We also compute
Z t

x(t)
= AeA(tt0 ) x0 + Bu(t) +
AeA(t ) Bu( ) d
t0
Z t
= AeA(tt0 ) x0 + A
eA(t ) Bu( ) d + Bu(t)
t0

= Ax(t) + Bu(t).
Thus x(t) as defined by (2.3) is indeed a solution.
Now we show that x(t) as defined by (2.3) is the only solution with the initial condition
(t) is a solution to the same initial value problem. Therefore,
x(t0 ) = x0 . We suppose that x
(t) x(t) satisfies
z(t) = x

(t) x(t)

z(t)
=x
= A
x(t) Bu(t) Ax(t) Bu(t) = Az(t).
(t0 ) this means that z(t0 ) = 0. That is, z(t) is a solution to the initial value
Since x(t0 ) = x
problem

z(t)
= Az(t), z(t0 ) = 0.
(2.4)
Let us multiply the differential equation on each side by 2z t (t):

2z(t)t z(t)
=

d t
d
(z (t)z(t)) = (kz(t)k2 ) = 2z t (t)Az(t).
dt
dt

et kz(t)k2 et0 kz(t0 )k2 0

for all t I. Since z(t0 ) = 0 we must have
et kz(t)k2 0,

t I.

Since et > 0 this must mean that kz(t)k2 = 0 for all t I and so z(t) = 0 for all t I.
(t) = x(t), and so solutions are unique.
But this means that x

2.7 Remark As per Remark 2.4, if we suppose that u(t) is essentially bounded and measurable, then Theorem 2.6 still holds.

Of course, once we have the solution for the state variable x, it is a simple matter to
determine the output y:
Z t
y(t) = CeA(tt0 ) x0 +
CeA(t ) Bu( ) d + Du(t).
t0

Our aim in this book is to study the response of the output y(t) to various inputs u(t), and
to devise systematic ways to make the output do things we like.
Lets look at an example.
2.8 Example We return to our mass-spring-damper example. Let us be concrete for simplicity, and choose m = 1, k = 4, and d = 0. The system equations are then

x 1
0 1 x1
0
=
+
u(t).
x 2
4 0 x2
1

22/10/2004

2.1 Properties of finite-dimensional, time-invariant linear control systems

We compute (I use Mathematica )

cos 2t
=
2 sin 2t

1
2

sin 2t
.
cos 2t

(
1, t 0
u(t) =
0, otherwise.
Thus the input is a step function. Let us suppose we have zero initial condition x(0) = 0.
We then compute

1

Z t
1
PSfrag replacements
sin 2(t 0
(1 cos 2t)
x1 (t)
cos 2(t )
2
4
d
=
.
=
1
2 sin 2(t ) cos 2(t ) 1
sin 2t
x2ke(t)k
(t) t
0
2
e1 (t)

e2 (t) of this curve is shown in Figure 2.1.

The phase portrait
e3 (t)
y(t)
u(t)
h (t)
hN,D (t)
1 (t)
1N,D (t)
(t)
(t)
fj (t)
Re
Im

0.2

0.5
0.4
0.3
0.2
0.1
0
2

u(t)
h (t)
hN,D (t)velocity we have
2. If we measure
1 (t)

1N,D (t)

0
c=
, D= 0 .
(t)
1
(t)
f (t)
The output jhere is
Re
y(t) = 21 sin 2t
Im
x1
which we plotx in Figure 2.3.
2
x1
x2
log
0.4
dB
deg 0.2
u = 0 ln
ln coth(|u| /2)
0
or 1
m
-0.2
yos
-0.4
tos 0

2
4
6
,0
t
0
y(t)

-0.2

-0.4

-0.2

,0
0

0.2

0.4

Figure 2.1 Phase curve for step response of mass-spring system

As far as outputs are concerned, recall that we had in Example 2.2 considered three
cases. With the parameters we have chosen, these are as follows.
1. In the first case we measure displacement and so arrived at

1
c=
, D= 0 .
0
The output is then computed to be
y(t) =
which we plot in Figure 2.2.

22/10/2004

Figure 2.2 Displacement output for mass-spring system

0.4

x1
x2
log
dB
deg
u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

State-space representations (the time-domain)

y(t)

Let us suppose that

1N,D (t)
(t)
(t)
fj (t)
Re
Im
x12
30
x2
x1
x2
log
dB
deg
u = 0 ln
ln coth(|u| /2)
or 1
m

yos

PSfrag replacements
tos 0

ke(t)k
0
e1,
(t)
0
e2 (t)
e3 (t)

Figure 2.3 Velocity output for mass-spring system

3. The final case was when the output was acceleration, and we then derived

4
c=
, D= 1 .
0
One readily ascertains

1
(1
4

y(t) = cos 2t

cos 2t)
which we plot in Figure 2.4.

N,D

2 State-space representations (the time-domain)

f1
x1
f2
x1

A = Df (x0 , u0 ) = .
..

f1
x2
f2
x2

..
.

f1
xn
f2
xn

fn
x2

fn
xn

..
.

f2
f

(x0 , u0 ) = u
.
u
..
fn

c = Dh(x0 , u0 ) =
h
(x0 , u0 ),
D=
u

Figure 2.4 Acceleration output for mass-spring system

2.2 Obtaining linearised equations for nonlinear input/output systems

The example of the mass-spring-damper system is easy to put into the form of (2.2)
since the equations are already linear. For a nonlinear system, we have seen in Section 1.4
how to linearise nonlinear differential equations about an equilibrium. Now lets see how to
linearise a nonlinear input/output system. We first need to say what we mean by a nonlinear
input/output system. We shall only consider SISO versions of these.
2.9 Definition A SISO nonlinear system consists of a pair (f , h) where f : Rn R Rn
and h : Rn R R are smooth maps.

What are the control equations here? They are
x = f (x, u)
y = h(x, u).

22/10/2004

(A, b, ct , D) where

y(t)

(t)
(t)
fj (t)
Re
Im
x2.2
1
22/10/2004
Obtaining linearised equations for nonlinear input/output systems
x2
x1
1
x2
log
dB
deg 0.5
u = 0 ln
ln coth(|u| /2)
0
or 1
m

yos -0.5

tos 0
-1

2
4
6
8
10
,0
t
0

h
x2

h
xn

where all partial derivatives are evaluated at (x0 , u0 ).

2.11 Note Let us suppose for simplicity that all equilibrium points we consider will necessitate that u0 = 0.

Let us do this for the pendulum.
2.12 Example (Example 1.3 contd) We consider a torque applied at the pendulum pivot and
we take as output the pendulums angular velocity.
Let us first derive the form for the first of equations (2.5). We need to be careful in
deriving the vector function f . The forces should be added to the equation at the outset,
and then the equations put into first-order form and linearised. Recall that the equations
for the pendulum, just ascertained by force balance, are

(2.5)

This is a generalisation of the linear equations (2.2). For systems like this, is it no longer
obvious that solutions exist or are unique as we asserted in Theorem 2.6 for linear systems.
We do not really get into such issues here as they do not comprise an essential part of
what we are doing. We are interested in linearising the equations (2.5) about an equilibrium
point. Since the system now has controls, we should revise our notion of what an equilibrium
point means. To wit, an equilibrium point for a SISO nonlinear system (f , h) is a pair
(x0 , u0 ) Rn R so that f (x0 , u0 ) = 0. The idea is the same as the idea for an equilibrium
point for a differential equation, except that we now allow the control to enter the picture.
To linearise, we linearise with respect to both x and u, evaluating at the equilibrium point.
In doing this, we arrive at the following definition.

u
h

m`2 + mg` sin = 0.

It is to these equations, and not any others, that the external torque should be added since
the external torque should obviously appear in the force balance equation. If the external
torque is u, the forced equations are simply
m`2 + mg` sin = u.
We next need to put this into the form of the first of equations (2.5). We first divide by m`2
and get
u
+ g` sin =
.
m`2
We then have
To put this into first-order form we define, as usual, (x1 , x2 ) = (, ).

2.10 Definition Let (f , h) be a SISO nonlinear system and let (x0 , u0 ) be an equilibrium
point for the system. The linearisation of (2.5) about (x0 , u0 ) is the SISO linear system

x 1 = = x2
x 2 = = g sin +
`

so that

1
u
m`2

f (x, u) =

= g` sin x1 +

x2
sin x1 +

1
u
m`2

22/10/2004

2.3 Input/output response versus state behaviour

2 State-space representations (the time-domain)

22/10/2004

PSfrag replacements

The output is easy to handle in this example. We have h(x, u) = = x2 . Therefore

0
c=
, D = 01 .
1
Putting all this together gives us the linearisation as

0
x 1
0
1 x1
+ 1 u
=
g
x 2
` cos 0 0 x2
m`2

x1
y= 0 1
.
x2
We then substitute 0 = 0 or 0 = depending on whether the system is in the down or
up equilibrium.

2.3 Input/output response versus state behaviour

In this section we will consider only SISO systems, and we will suppose that the 1 1
matrix D is zero. The first restriction is easy to relax, but the second may not be, depending
on what one wishes to do. However, often in applications D = 01 in any case.
We wish to reveal the problems that may be encountered when one focuses on input/output behaviour without thinking about the system states. That is to say, if we
restrict our attention to designing inputs u(t) that make the output y(t) behave in a desirable manner, problems may arise. If one has a state-space model like (2.2), then it is
possible that while you are making the outputs behave nicely, some of the states in x(t) may
be misbehaving badly. This is perhaps best illustrated with a sequence of simple examples.
Each of these examples will illustrate an important concept, and after each example, the
general idea will be discussed.
2.3.1 Bad behaviour due to lack of observability
will introduce us to the concept of observability.

We first look at an example that

2.13 Example We first consider the system

x 1
0 1 x1
0
=
+
u
x 2
1 0 x2
1

x1
y = 1 1
.
x2
We compute
eAt

(et + et )
= 12 t
(e
et )
2
1

1 t
(e
2
1 t
(e
2

et )
t
+e )

(2.6)

and so, if we use the initial condition x(0) = 0, and the input
ke(t)k
(
e1 (t)
1, t 0
e2 (t)
u(t)
=
e3 (t)
0, otherwise,
u(t)
h (t)
hN,D (t)

1 t

Z t 1 t
(e et+ ) 12 (et et+ ) 0
(e + et ) 1
2
x(t) 1= (t) 12 t
d
=
.
1 t
(e et+ ) 12 (et + et+ ) 1
(e et )
1N,D (t)
0
2
2
(t)
(t)
One also readily
computes the output as
fj (t)
Re
y(t) = et 1
Im
1
which we plot inx
x2 Figure 2.5. Note that the output is behaving quite nicely, thank you.
x1
x2
0
log
dB -0.2
deg
u = 0 ln
ln coth(|u| /2) -0.4
or 1
m -0.6

yos -0.8

-1
tos 0

2
4
6
8
10
,0
t
0

we get

y(t)

By a suitable choice for u0 , any point of the form (x1 , 0) can be rendered an equilibrium
point. Let us simply look at those for which u0 = 0, as we suggested we do in Note 2.11. We
determine that such equilibrium points are of the form (x0 , u0 ) = ((0 , 0), 0), 0 {0, }.
We then have the linearised state matrix A and the linearised input vector b given by

0
1
0
A=
, b= 1 .
g
` cos 0 0
m`2

Figure 2.5 Output response of (2.6) to a step input

However, the state is going nuts, blowing up to as t as shown in Figure 2.6.

What is the problem here? Well, looking at what is going on with the equations reveals
the problem. The poor state-space behaviour is obviously present in the equations for the
state variable x(t). However, when we compute the output, this bad behaviour gets killed
by the output vector ct . There is a mechanism to describe what is going on, and it is called
observability theory. We only talk in broad terms about this here.
2.14 Definition A pair (A, c) Rnn Rn is observable if the matrix

ct
ct A

O(A, c) =

.
n1
ct A
has full rank. If = (A, b, ct , D), then is observable if (A, c) is observable. The matrix
O(A, c) is called the observability matrix for (A, c).

PSfrag replacements

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)
h (t)
hN,D (t)
1 (t)
1N,D (t)
(t)
(t)
fj (t)
Re
Im

x1
x2
log
dB
deg
u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

2.3 Input/output response versus state behaviour

2 State-space representations (the time-domain)

22/10/2004

First suppose that (c, A) is observable, and let us suppose that z1 (t) = z2 (t), with z1
and z2 as above. Therefore we have

z1 (0)
z2 (0)
ct
ct
(1)
(1)
z (0) ct A
z (0) ct A
1

= .. x1 (0) =
= .. x2 (0).
..
..

.
.
.
(n1)
(n1)
ct An1
ct An1
z1
(0)
z2
(0)

22/10/2004

However, since (A, c) is observable, this gives

O(A, c)x1 (0) = O(A, c)x2 (0)

x1 (0) = x2 (0),

which is, as we have seen, equivalent to the assertion that u1 = u2 and y1 = y2 implies that
x1 = x2 .
Now suppose that rank(O(A, c)) 6= n. Then there exists a nonzero vector x0 Rn so
that O(A, c)x0 = 0. By the Cayley-Hamilton Theorem it follows that ct Ak x1 (t) = 0, k 1,
if x1 (t) is the solution to the initial value problem

-10

-20
-20

-10

,0
0

Figure 2.6 State-space behaviour of (2.6) with a step input

x(t)
= Ax(t),

x(0) = x0 .

Now the series representation for the matrix exponential gives z1 (t) = 0 where z1 (t) =
ct eAt x0 . However, we also have z2 (t) = 0 if z2 (t) = ct 0. However, x2 (t) = 0 is the solution
to the initial value problem

x(t)
= Ax(t), x(0) = 0,

The above definition carefully masks the real definition of observability. However, the
following result provides the necessary connection with things more readily visualised.

from which we infer that we cannot infer the initial conditions from the output with zero
input.

2.15 Theorem Let = (A, b, ct , 01 ) be a SISO linear system, let u1 , u2 U , and let x1 (t),
x2 (t), y1 (t), and y2 (t) be defined by

The idea of observability is, then, that one can infer the initial condition for the state
from the input and the output. Let us illustrate this with an example.

x i (t) = Axi (t) + bui (t)

yi (t) = ct xi (t),
i = 1, 2. The following statements are equivalent:
(i) (A, c) Rnn Rn is observable;
(ii) u1 (t) = u2 (t) and y1 (t) = y2 (t) for all t implies that x1 (t) = x2 (t) for all t.
Proof Let us first show that the second condition is equivalent to saying that the output with
zero input is in one-to-one correspondence with the initial condition. Indeed, for arbitrary
inputs u1 , u2 U with corresponding states x1 (t) and x2 (t), and outputs y1 (t) and y2 (t) we
have
Z t
yi (t) = ct eAt xi (0) +
ct eA(t ) bui ( ) d,
0

i = 1, 2. If we define zi (t) by

2.16 Example (Example 2.13 contd) We compute the observability matrix for the system in
Example 2.13 to be

1 1
O(A, c) =
1 1
which is not full rank (it has rank 1). Thus the system is not observable.
Now suppose that we start out the system (2.6) with not the zero initial condition, but
with the initial condition x(0) = (1, 1). A simple calculation shows that the output is then
y(t) = et 1, which is just as we saw with zero initial condition. Thus our output was
unable to see this change in initial condition, and so this justifies our words following
Definition 2.14. You might care to notice that the initial condition (1, 1) is in the kernel of
the matrix O(A, c)!

The following property of the observability matrix will be useful for us.

Z
zi (t) = yi (t)

ct eA(t ) bui ( ) d,

i = 1, 2, then we see that u1 = u2 and y1 = y2 is equivalent to z1 = z2 . However, since

zi (t) = ct eAt xi (0), this means that the second condition of the theorem is equivalent to the
statement that equal outputs for zero inputs implies equal initial conditions for the state.

2.17 Theorem The kernel of the matrix O(A, c) is the largest A-invariant subspace contained
in ker(ct ).

22/10/2004

2.3 Input/output response versus state behaviour

Proof First let us show that the kernel of O(A, c) is contained in ker(ct ).
ker(O(A, c)) then

ct
0
ct A
0

O(A, c)x =
x = .. ,
..

.
.
0
ct An1

37
If x

and in particular, ct x = 0that is, x ker(ct ).

Now we show that the kernel of O(A, c) is A-invariant. Let x ker(O(A, c)) and then
compute

cA
ct
ct A2
ct A

O(A, c)Ax =
Ax = .. x.
..

.
ct An1

ct An

Since x ker(O(A, c)) we have

ct Ax = 0, . . . , ct An1 x = 0.

This shows that

O(A, c)x = 0
or x ker(O(A, c)).
Finally, we show that if V is an A-invariant subspace contained in ker(ct ), then V is a
subspace of ker(O(A, c)). Given such a V and x V , ct x = 0. Since V is A-invariant,
Ax V , and since V is contained in ker(ct ), ct Ax = 0. Proceeding in this way we see that
ct A2 x = = ct An1 x = 0. But this means exactly that x is in ker(O(A, c)).

The subspace ker(O(A, c)) has a simple interpretation in terms of Theorem 2.15.
It turns out that if two state initial conditions x1 (0) and x2 (0) differ by a vector in
ker(O(A, c)), i.e., if x2 (0)x1 (0) ker(O(A, c)), then the same input will produce the same
output for these different initial conditions. This is exactly what we saw in Example 2.16.
2.18 Remark Although our discussion in this section has been for SISO systems =
(A, b, ct , D), it can be easily extended to MIMO systems. Indeed our characterisations
of observability in Theorems 2.15 and 2.17 are readily made for MIMO systems. Also, for a
MIMO system = (A, B, C, D) one can certainly define

C
CA

O(A, C) =
,
..

.
CAn1
and one may indeed verify that the appropriate versions of Theorems 2.15 and 2.17 hold in
this case.

22/10/2004

2.3.2 Bad behaviour due to lack of controllability Okay, so we believe that a lack
of observability may be the cause of problems in the state, regardless of the good behaviour
of the input/output map. Are there other ways in which things can go awry? Well, yes there
are. Let us look at a system that is observable, but that does not behave nicely.
2.19 Example Here we look at

x 1
1 0
x1
0
=
+
u
x 2
1 1 x2
1

x1
y= 0 1
.
x2

(2.7)

This system is observable as the observability matrix is

0 1
O(A, c) =
1 1
which has rank 2. We compute

Also, by the Cayley-Hamilton Theorem,

ct An x = pn1 ct An1 x p1 ct Ax p0 ct x = 0.

2 State-space representations (the time-domain)

eAt =

1 t
(e
2

et
0
et ) et

from which we ascertain that with zero initial conditions, and a unit step input,

0
x(t) =
y(t) = 1 et .
t ,
1e
Okay, this looks fine. Lets change the initial condition to x(0) = (1, 0). We then compute

et
x(t) =
, y(t) = 1 + 12 (et 3et ).
1 t
t
1 + 2 (e 3e )
Well, since the system is observable, it can sense this change of initial condition, and how!
As we see in Figure 2.7 (where we depict the output response) and Figure 2.8 (where we
depict the state behaviour), the system is now blowing up in both state and output.

Its not so hard to see what is happening here. We do not have the ability to get at
the unstable dynamics of the system with our input. Motivated by this, we come up with
another condition on the linear system, different from observability.
2.20 Definition A pair (A, b) Rnn Rn is controllable if the matrix

C(A, b) = b Ab An1 b
has full rank. If = (A, b, ct , D), then is controllable if (A, b) is controllable. The
matrix C(A, b) is called the controllability matrix for (A, b).

Let us state the result that gives the intuitive meaning for our definition for controllability.

N,D

2 State-space representations (the time-domain)

1 (t)
1N,D (t)
(t)
2.7 The output response of (2.7) with a step input and
Figure
(t)
fj (t)non-zero initial condition
Re
Im
5

,0
0

4
3

xt0 eAt bbt eA t x0 = 0.

Thus P (A, b)(t) is not invertible.
Now suppose that there exists T > 0 so that P (A, b)(T ) is not invertible. Therefore
there exists a nonzero x0 Rn so that xt0 eAt b = 0 for t [0, T ]. Differentiating this n 1
times with respect to t and evaluating at t = 0 gives
x0 b = x0 Ab = = x0 An1 b = 0.
This, however, infers the existence of a nonzero vector in ker(C(A, b)), giving us our initial
claim.
Let us now show how this claim gives the theorem. First suppose that C(A, b) is invertible so that P (A, b)(t) is positive-definite for all t > 0. One may then directly show, with a
slightly tedious computation, that if we define a control u : [0, T ] R by

t
u(t) = bt eA (T t) P (A, b)1 (T ) eAT x1 x2 ,
then the solution to the initial value problem

x1
x2
log
dB
deg
u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

22/10/2004

if and only if P (A, b)(t) is invertible. Suppose that C(A, b) is not invertible. Then there
exists a nonzero x0 Rn so that xt0 C(A, b)) = 0. By the Cayley-Hamilton Theorem, this
implies that xt0 Ak b = 0 for k 1. This in turn means that xt0 eAt b = 0 for t > 0. Therefore,
t
since eA t = (eAt )t , it follows that

y(t)

(t)
(t)
fj (t)
Re
Im
x1
22/10/2004
2.3 Input/output response versus state behaviour
x2
x1
x2 12000
log
dB 10000
deg
8000
u = 0 ln
ln coth(|u| /2)
PSfrag replacements
or 1t 6000
m
ke(t)k
4000

e1 (t)
e2y(t)
os
e3 (t)
2000

y(t)
tosu(t)
0

h,
(t)
2
4
6
8
10
0
hN,D(t)
t
0

x(t)
= Ax(t) + bu(t),

has the property that x(T ) = x2 .

Now suppose that C(A, b) is not invertible so that there exists T > 0 so that P (A, b)(T )
is not invertible. Thus there exists a nonzero x0 Rn so that
xt0 eAt b = 0,

Figure 2.8 The state-space behaviour of (2.7) with a step input

and non-zero initial condition

2.21 Theorem Let = (A, b, ct , 01 ) be a SISO linear system. The pair (A, b) is controllable
if and only if for each x1 , x2 Rn and for each T > 0, there exists an admissible input
u : [0, T ] R with the property that if x(t) is the solution to the initial value problem

x(t)
= Ax(t) + bu(t),

x(0) = x1 ,

then x(T ) = x2 .
Proof For t > 0 define the matrix P (A, b)(t) by
Z t
t
P (A, b)(t) =
eA bbt eA d.
0

Let us first show that C(A, b) is invertible if and only if P (A, b)(t) is positive-definite
for all t > 0 (we refer ahead to Section 5.4.1 for notions of definiteness of matrices). Since
P (A, b)(t) is clearly positive-semidefinite, this means we shall show that C(A, b) is invertible

x(0) = x1

t [0, T ].

(2.8)

Let x1 = e
x0 and let u be an admissible control. If the resulting state vector is x(t), we
then compute
Z T
x(T ) = eAT eAT x0 +
eA(T ) bu( ) d.
0

Using (2.8), we have

xt0 x(T ) = xt0 x0 .
Therefore, it is not possible to find a control for which x(T ) = 0.

This test of controllability for linear systems was obtained by Kalman, Ho, and Narendra
[1963]. The idea is quite simple to comprehend: controllability reflects that we can reach
any state from any other state. We can easily see how this comes up in an example.
2.22 Example (Example 2.19 contd) We compute the controllability matrix for Example 2.19 to be

0 0
C(A, b) =
1 1
which has rank 1 < 2 and so the system is not controllable.

22/10/2004

2.3 Input/output response versus state behaviour

Lets see how this meshes with what we said following Definition 2.20. Suppose we start
at x(0) = (0, 0). Since any solution to

x 1
1 0
x1
0
=
+
u
x 2
1 1 x2
1

x1
y= 0 1
x2
which has initial condition x1 (0) = 0 will have the property that x1 (t) = 0 for all t, the
control system is essentially governed by the x2 equation:
x 2 = x2 + u.
Therefore we can only move in the x2 -direction and all points with x1 6= 0 will not be
accessible to us. This is what we mean by controllability. You might note that the set of
points reachable are those in the columnspace of the matrix C(A, b).

Based on the above discussion, we say that a triple (A, b, c) Rnn Rn Rn is
complete if (A, b) is controllable and if (A, c) is observable.
We have a property of the controllability matrix that is sort of like that for the observability matrix in Theorem 2.17.
2.23 Theorem The columnspace of the matrix C(A, b) is the smallest A-invariant subspace
containing b.
Proof Obviously b is in the columnspace of C(A, b). We will show that this columnspace
is A-invariant. Let x be in the columnspace of C(A, b), i.e., in the range of the linear map
C(A, b). Then there is a vector y Rn with the property that

x = C(A, b)y = b Ab An1 b y.
We then have
Ax =

Ab A2 b An b y.

2 State-space representations (the time-domain)

There is a somewhat subtle thing happening here that should be understood. If a pair
(A, b) is controllable, this implies that one can steer between any two points in Rn with a
suitable control. It does not mean that one can follow any curve in Rn that connects the
two points. This then raises the question, What curves in Rn can be followed by solutions
of the differential equation

x(t)
= Ax(t) + bu(t)?
Let us explore the answer to this question, following Basile and Marro [1968]. Because we
will deal only with the single-input case, things are somewhat degenerate. Let T (A, b) Rn
be the subspace
(
span {b} , b is an eigenvector for A
T (A, b) =
{0},
otherwise.
The following lemma asserts the essential property of the subspace T (A, b).
2.24 Lemma T (A, b) is the largest subspace of Rn with the property that
A(T (A, b)) + T (A, b) span {b} .
Proof First we note that T (A, b) does indeed have the property that A(T (A, b)) +
T (A, b) span {b}. This is clear if T (A, b) = {0}. If T (A, b) = span {b} then it is
the case that Ab = b for some R. It then follows that if x1 = a1 b and x2 = a2 b for
a1 , a2 R then
Ax1 + x2 = (a1 + a2 )b span {b} .
Now we need to show that T (A, b) is the largest subspace with this property. Suppose
that V is a subspace of Rn with the property that A(V ) + V span {b}. Thus for each
x1 , x2 V we have
Ax1 + x2 span {b} .
In particular, if we choose x1 = 0 we see that if x2 V then x2 span {b}. Similarly, if
x2 = 0 we see that if x1 V then Ax1 span {b}. Thus we have shown that if V is a
subspace with the property that A(V ) + V span {b}, this implies that
V = span {b} A1 (span {b})

The result will follow if we can show that each of the vectors
Ab, . . . , An b
is in the columnspace of C(A, b). It is clear that
Ab, . . . , A

where
A1 (span {b}) = { v Rn | Av span {b}}
(note that we are not saying that A is invertible!). It now remains to show that T (A, b) =
span {b} A1 (span {b}). We consider the two cases where (1) b is an eigenvector for A
and (2) b is not an eigenvector for A. In the first case, b A1 (span {b}) so we clearly have

are in the columnspace of C(A, b). By the Cayley-Hamilton Theorem we have

A b = pn1 A

b p1 Ab p0 b,

which shows that An b is also in the columnspace of C(A, b).

Now we show that if V is an A-invariant subspace with b V then V contains the
columnspace of C(A, b). If V is such a subspace then b V . Since V is A-invariant,
Ab V . Proceeding in this way we ascertain that A2 b, . . . , An1 b V , and therefore the
columnspace of C(A, b) is contained in V .

22/10/2004

span {b} A1 (span {b}) = span {b} .

In the second case, b 6 A1 (span {b}) so that
span {b} A1 (span {b}) = {0}.
But this is our result.

Now we can use this lemma to describe the set of curves in R that can be followed
exactly by our control system.

2.3 Input/output response versus state behaviour

2.25 Proposition Let = (A, b, ct , 01 ) be a SISO linear system and let T (A, b) Rn be
the subspace defined above. If I R is an interval and if r : I T (A, b) is continuously
differentiable, then there exists a continuous function u : I R with the property that

r(t)
= Ar(t) + bu(t).
Furthermore, T (A, b) is the largest subspace of Rn with this property.

r(t + ) r(t)
T (A, b)

t I.

Therefore, for each t I there exists u(t) R so that

Ar(t) = u(t)b.
r(t)

The first part of the proposition now follows since the T (A, b)-valued function of t, r(t)
Ar(t) is continuous.
Now suppose that V is a subspace of Rn with the property that for every continuously
differentiable r : I V , there exists a continuous function u : I R with the property that

r(t)
= Ar(t) + bu(t).

Let x1 , x2 V and define r : R V by x1 +tx2 . Then we have r(0) = x1 and r(0)

= x2 .
Therefore there must exist a continuous u : R R so that
x2 = Ax1 + bu(0).
Since this construction can be made for any x1 , x2 V , we must have Ax1 + x2 span {b}

for every x1 , x2 V . By Lemma 2.24 this means that V = T (A, b).
Thus we see for single-input systems, the state trajectories we may exactly follow are
actually quite limited. Nevertheless, even though one cannot follow all state trajectories, it
is possible for a system to be controllable.
2.26 Remark As was the case for observability in Remark 2.18, it is easy to talk about
controllability in the MIMO setting. Indeed, if for a MIMO system = (A, B, C, D) we
define

C(A, B) = B AB An1 B ,
then the appropriate versions of Theorems 2.21 and 2.23 hold.

2.27 Example We take as our system

x 1
0
1
x1
0
=
+
u
x 2
2 3 x2
1

x1
y = 1 1
.
x2

22/10/2004

(2.9)

PSfrag replacements

since r(t + ), r(t) T (A, b). Therefore, by Lemma 2.24,

Ar(t) T (A, b),
r(t)

2 State-space representations (the time-domain)

First, lets see that the system is observable and controllable. The respective matrices are

0 1
1 1
O(A, c) =
, C(A, b) =
2 4
1 3

Proof For the first part of the proposition, we note that if r : I T (A, b) then

r(t)
= lim

2.3.3 Bad behaviour due to unstable zero dynamics Now you are doubtless thinking
that we must have ourselves covered. Surely if a system is complete then our state-space
behaviour will be nice if the output is nice. But this is in fact not true, as the following
example shows.

which both have rank 2. We compute

t
ke(t)k
2e e2t
et e2t
e1 (t)
eAt =
2t
t
2t
t .
2(e e ) 2e e
e2 (t)
e3 (t)
In this example, we do not use a step input, but rather a violent input:
u(t)
(
h (t)
et , t 0
hN,D (t)
u(t) =
0, otherwise.
1 (t)
1N,D (t)
Thus our input(t)
blows up as time increases. The usual calculations, using zero initial condi (t)
tions, give
1 t 1 2t 1 t
fj (t)
e + 3e 2e
6
x(t)
=
y(t) = e2t et .
Re
1 t
2 2t
1 t ,
e

e
+
e
6
3
2
Im
1 behaving nicely (see Figure 2.9) while the state is blowing up to infinity
Thus the outputxis
x2
x1
0.1
x2
log
0.05
dB
0
deg
u = 0 ln -0.05
ln coth(|u| /2)
or 1 -0.1
m -0.15

-0.2
yos
-0.25
tos 0

2
4
6
8
10
,0
t
0
y(t)

22/10/2004

Figure 2.9 The output response of (2.9) to an exponential input

(see Figure 2.10).

Things are a bit more subtle with this example. The problem is that the large input is
not being transmitted to the output. Describing the general scenario here is not altogether
easy, but we work through it so that you may know what is going on.

PSfrag replacements

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)
h (t)
hN,D (t)
1 (t)
1N,D (t)
(t)
(t)
fj (t)
Re
Im

x1
x2
log
dB
deg
u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

,0
0

2.3 Input/output response versus state behaviour

2 State-space representations (the time-domain)

22/10/2004

7. Define an ` ` matrix by
10

b11 b12 b1`

b21 b22 b2`

N = ..
.. . .
. .
.
. ..
.
b`1 b`2 b``

8. The linear differential equation

= N w
w

22/10/2004

is called the zero dynamics for .

This is plainly nontrivial! Lets illustrate what is going on with our example.
2.29 Example (Example 2.27 contd) We shall go through the algorithm step by step.
1. We take V0 = R2 as directed.
2. As per the instructions, we need to compute ker(ct ) and we easily see that

ker(ct ) = span {(1, 1)} .

Now we compute

Figure 2.10 The state-space behaviour of (2.9) with an exponential
input

2.28 Algorithm for determining zero dynamics We start with a SISO linear system =
(A, b, ct , D) with A Rnn and b, c Rn . We do not assume that (A, b) is controllable or
that (A, c) is observable.
1. Define Z0 = Rn .
2. Inductively define a sequence of subspaces of Rn by
Zk+1 = ker(ct ) {x Rn | Ax Zk + span {b}} .
3. This sequence will eventually stop, i.e., there will be a least K so that ZK+1 = ZK .
Denote Z = ZK , and suppose that dim(Z ) = `.
4. It turns out that is it possible to find f Rn with the property that Ab,f , A + bf t
Rnn has Z as an invariant subspace. Choose such an f .
5. Choose a basis {v 1 , . . . , v ` } for Z , and extend this to a basis {v 1 , . . . , v n } for Rn .
6. Write
Ab,f v 1 = b11 v 1 + + b`1 v `
..
.
Ab,f v ` = b1` v 1 + + b`` v `
Ab,f v `+1 = b1,`+1 v 1 + + b`,`+1 v ` + b`+1,`+1 v `+1 + + bn,`+1 v n
..
.
Ab,f v n = b1n v 1 + + b`n v ` + b`+1,n v `+1 + + bnn v n .

x R2 Ax Z0 + span {b} = R2

since Z0 = R2 . Therefore Z1 = ker(ct ). To compute Z2 we compute

x R2 Ax Z1 + span {b} = R2
since ker(ct ) and span {b} are complementary subspaces. Therefore Z2 = ker(ct ) and so
our sequence terminates at Z1 .
3. We have
Z = ker(ct ) = span {(1, 1)} .
4. Let f = (f1 , f2 ). We compute

0
1
0
0
1
f1 f2 =
Ab,f =
+
.
1
2 + f1 3 + f2
2 3
In order that this matrix leave Z invariant, it must map the basis vector (1, 1) for Z
to a multiple of itself. We compute

1
0
1
1
1
Ab,f
=
=
.
1
2 + f1 3 + f2 1
f1 + f2 5
In order that this vector be a multiple of (1, 1) we must have f1 +f2 5 = 1 or f1 +f2 = 6.
Let us choose f1 = f2 = 3.
5. We choose the basis {v 1 = (1, 1), v 2 = (1, 1)} for R2 , noting that v 1 is a basis for Z .
6. We compute
Ab,f (v 1 ) = (1, 1) = 1v 1 + 0v 2
Ab,f (v 2 ) = (1, 1) = 0v 1 1v 2 .

22/10/2004

2.3 Input/output response versus state behaviour

7. The matrix N is 1 1 and is given by

N = 1 .
8. The zero dynamics are then

w 1 = 1 w1
which is a scalar system.
Okay, so how is our bad behaviour reflected here? Well, note that the zero dynamics are
unstable! This, it turns out, is the problem.

2.30 Remarks 1. Systems with stable zero dynamics (i.e., all eigenvalues for the matrix N
have nonpositive real part) are sometimes called minimum phase systems. Note that
the response Figure 2.9 shows an output that initially does something opposite from
what it ends up eventually doingthe output decreases before it finally increases to its
final value. This, it turns out, is behaviour typical of a system that is not minimum
phase. We shall be investigating the properties of nonminimum phase systems as we
go along (see Theorem 3.15).
2. The zero dynamics as we construct them are not obviously independent of the choices
made in the algorithm. That is to say, it is not clear that, had we chosen a different
vector f , or a different basis {v 1 , . . . , v n }, that we would not arrive at an utterly different
matrix N . Nonetheless, it is true that if we were to have chosen a different vector f and
the same basis {v 1 , . . . , v n } that the resulting matrix N would be unchanged. A choice
of a different basis would only change the matrix N by a similarity transformation, and
so, in particular, its eigenvalues would be unchanged.

Let us complete this section by giving a description of the subspace Z .
2.31 Theorem Let (A, b, ct , 01 ) be a SISO system and let Z be the set of all subspaces V of
Rn with the properties
(i) V ker(ct ) and
(ii) A(V ) V + span {b}.
The subspace Z constructed in Algorithm 2.28 is the largest subspace in Z .
Proof By the inductive procedure of Algorithm 2.28 it is clear that Z Z . We then need
only show that Z is the largest subspace in Z . Let V Z and let x V . This means
that x ker(ct ) and so x Z1 (since in Algorithm 2.28 we always have Z1 = ker(ct )). We
also have
Ax V + span {b} Z1 + span {b} .
Therefore x Z2 . Proceeding in this way we see that x Zi for i = 1, . . . , K, and so
x Z . This concludes the proof.

2 State-space representations (the time-domain)

22/10/2004

1. Unseen unstable dynamics due to lack of observability: This was illustrated in Example 2.13. The idea was that any input we gave the system leads to a nice output.
However, some inputs cause the states to blow up. The problem here is that lack of
observability causes the output to not recognise the nasty state-space behaviour.
2. Lurking unstable dynamics caused by lack of controllability: It is possible, as we saw in
Example 2.19, for the dynamics to be unstable, even though they are fine for some initial
conditions. And these unstable dynamics are not something we can get a handle on with
our inputs; this being the case because of the lack of controllability.
3. Very large inputs can cause no output due to the existence of unstable zero dynamics:
This is the situation illustrated in Example 2.27. The problem here is that all the input
energy can be soaked by the unstable modes of the zero dynamics, provided the input is
of the right type.
It is important to note that if we have any of the badness of the type listed above, there
aint nothing we can do about it. It is a limitation of the physical system, and so one has
to be aware of it, and cope as best one can.
We shall see these ideas arise in various ways when we discuss transfer functions in
Chapter 3. As we say, the connection here is a little deep, so if you really want to see what
is going on here, be prepared to invest some effortit is really a very neat story, however.

2.4 The impulse response

In this section we will only consider SISO systems, and we will suppose that the 1
1 matrix D is zero. Generalisations to cases where the first condition does not hold are
straightforward. Generalisation to the case where D is non-zero is essentially carried out in
Exercise E3.1.
2.4.1 The impulse response for causal systems Typically, we will use the impulse
response in situations where we are interested in positive times. Thus we consider everything
before t = 0 to be zero, and then at t = 0 the action starts. It is this standard situation
we deal with in this section.
Recall from Theorem 2.6 that the solution to the initial value problem

x(t)
= Ax(t) + bu(t),
is
x(t) = eAt x0 +

x(0) = x0

eA(t ) bu( ) d.

Therefore the output y(t) behaves like

y(t) = ct eAt x0 +

ct eA(t ) bu( ) d.

(2.10)

2.3.4 A summary of what we have said in this section We have covered a lot of
ground here with a few simple examples, and some general definitions. The material in this
section has touched upon some fairly deep concepts in linear control theory, and a recap is
probably a good idea. Let us outline the three things we have found that can go wrong, and
just how they go wrong.

We wish to determine the output when we start with zero initial condition, and at t0 = 0
give the system a sharp jolt. Let us argue intuitively for a moment. Our input will be
zero except for a short time near t = 0 where it will be large. One expects, therefore, that
the integration over in (2.10) will only take place over a very small interval near zero.
Outside this interval, u will vanish. With this feeble intuition in mind, we define the causal

t At
More succinctly, we may write h+
(t) = 1(t)c e b, where 1(t) is the unit step function. In
the next section we will define h
.
However,
since
we shall almost always be using h+

, let us
agree to simply write h for h+
, resorting to the more precise notation only in those special
circumstances where we need to be clear on which version of the impulse response we need.
The idea is that the only contribution from u in the integral is at = 0. A good question
is Does there exist u U so that the resulting output is h (t) with zero state initial
condition? The answer is, No there is not. So you will never see the impulse response if
you only allow yourself piecewise continuous inputs. In fact, you can allow inputs that are a
whole lot more general than piecewise continuous, and you will still not ever see the impulse
response. However, the impulse response is still an important ingredient in looking at the
input/output behaviour of the system. The following trivial result hints at why this is so.

2.32 Proposition For any u U the output of the system

x(t)
= Ax(t) + bu(t)
y(t) = ct x(t)
with the initial condition x = x0 is
y(t) = ct eAt x0 +

h (t )u( ) d.
0

That is to say, from the impulse response one can construct the solution associated with any
input by performing a convolution of the input with the impulse response. This despite
the fact that no input in U will ever produce the impulse response itself!
We compute the impulse response for the mass-spring-damper system.
2.33 Examples For this example we have

0
1
A=
,
k
m
md

0
.
1

Since the nature of e changes character depending on the choice of m, d, and k, lets choose
specific numbers to compute the impulse response. In all cases we take m = 1. We also
have the two cases of output to consider (we do not in this section consider the case when
D 6= 01 ).
1. We first take d = 3 and k = 2. The matrix exponential is then
t

2e 2 e2t et e2t
eAt =
2t
t
2t
t .
2(e e ) 2e e
(a) We first consider the case when c = (1, 0), i.e., when the output is displacement.
The impulse response is then
h (t) = 1(t)(et e2t )
which we show in the left plot in Figure 2.11.

,0
0

time-domain)

22/10/2004

h (t)

hN,D (t)
1 (t)
1N,D (t)
(t)
(t)
fj (t)
Re
Im (the
representations
x1
x2
x1
x2
1
log
dB 0.8
deg
u = 0 ln 0.6
ln coth(|u| /2)
or 1 0.4
m
0.2
yos

0
tos 0

8
10

,0
0

Figure 2.11 The displacement and velocity impulse response for a

mass-spring-damper system with m = 1, d = 3, and k = 2

(b) We next consider the case when c = (0, 1) so that the output is displacement. The
impulse response is then
h (t) = 1(t)(2e2t et ),
PSfrag replacements

PSfrag replacements

which we show in the right plot in Figure 2.11. This is the overdamped case when

ke(t)k
there are distinct real eigenvalues.ke(t)k
e1 (t)
e1 (t)
e2 (t)
e2 (t)
2.
e3 (t)
e3 (t)Next we take d = 2 and k = 1. We compute
y(t)
t y(t)
u(t)
e (1u(t)
+ t)
hN,D (t)
1 (t)
1N,D (t)
(t)
(t)
fj (t)
Re
Im
x1
x2
x1
x2
log
dB
deg
u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

eAt =

,0
0

te
hN,D (t)

tet
.
e (1 t)
t

1 (t)
1N,D (t)
(t)
(t)
t
j (t) = 1(t)(te
hf(t)
)
Re
Im
Figure 2.12. x1
x2
x1
x2
1
log
dB 0.8
deg
u = 0 ln 0.6
ln coth(|u| /2)
or 1 0.4
m
0.2
yos

0
tos 0

(a) Taking c = (1, 0) we compute

which is the left plot in

0.4
0.35
0.3
0.25

h (t)

impulse response, or simply the impulse response, of (2.1) to be

(
ct eAt b, t 0
+
h (t) =
0,
otherwise.

2 State-space

h (t)

2.4 The impulse response

22/10/2004

hN,D (t)
1 (t)
1N,D (t)
(t)
(t)
fj (t)
Re
50Im
x1
x2
x1
x2
log 0.25
dB
deg
0.2
u = 0 ln
ln coth(|u| /2)
0.15
1
or
m
0.1

yos
0.05

tos 0

0.2
0.15
0.1
0.05
2

10,0

Figure 2.12 The displacement and velocity impulse response for a

mass-spring-damper system with m = 1, d = 2, and k = 1

(b) If we let c = (0, 1) then we compute

h (t) = 1(t)(et (1 t))
which is the right plot in Figure 2.12. This is the critically damped case when
the eigenvalue is repeated.

2.4 The impulse response

ke(t)k
e1 (t)
underdamped
one when
e2 (t)
3 (t)
2 and key(t)
=
10 and compute
u(t)

at is the
we have complex roots with
take d =
t

1 t
e sin 3t
e (cos 3t + 13 sin 3t)
3
=
.
h (t)
10
et sinN,D
et (cos 3t 13 sin 3t)
13t
(t)
3

,0
0

22/10/2004

t
t

fj (t)
Re
Im 2.14 The displacement and velocity impulse response for a
Figure
x1mass-spring-damper system with m = 1, d = 0, and k = 1
x2
x1
x2
15
log
dB
deg 12.5
u = 0 ln
10
ln coth(|u| /2)
1
or
7.5
m

5
yos

2.5
tos 0

0.2
0.4
0.6
0.8
1
,0
t
0

u(t)

h (t)

e
hN,D (t)
1 (t)
1N,D (t)
1N,D (t)
(t) (a) Taking c = (1, 0) we compute
(t)
(t)
(t)
fj (t)
fj (t)
h (t) =Re1(t)( 31 et sin 3t)
Re
Im
Im
x1
which is the left plot in Figure 2.13. x1
x2
x2
x1
x1
0.25
x2
x2
1
log
log
0.2
dB
dB 0.8
deg 0.15
deg 0.6
u = 0 ln
u = 0 ln
0.4
ln coth(|u| /2)
ln coth(|u| /2)
0.1
or 1
or 1
0.2
0.05
m
m

0
0
yos
yos
-0.05
-0.2
tos 0
tos 0 -0.4

2
2
4
6
8
10,0
,0
t
0
0

State-space

hN,D (t)
1 (t)
1N,D (t)
(t)
(t)
fj (t)
Re
Im (the time-domain)
representations
x1
x2
x1
1
x2
log
dB
deg 0.5
u = 0 ln
ln coth(|u| /2)
0
or 1
m

yos -0.5

tos 0
-1

8
10,0
2
4
h (t)

PSfrag replacements

ke(t)k
e1 (t)
3.
e2 (t)The next case we look
e3 (t)
y(t)negative real part. We
u(t)

h (t)

PSfrag replacements
22/10/2004

hN,D (t)
1 (t)
1N,D (t)
(t)
(t)
fj (t)
PSfrag replacements
Re
52Im
x1
ke(t)k2
x2
e1 (t)
x1
1
e2 (t)
x2
log
e3 (t)
dB
y(t)
deg 0.5
u = 0 ln
ln coth(|u| /2)
h (t)
0
or 1
hN,D (t)
m

1 (t)
yos -0.5
1N,D (t)

(t)
tos 0
-1

2(t)

Figure 2.13 The displacement and velocity impulse response for a

mass-spring-damper system with m = 1, d = 2, and k = 10

(b) If we let c = (0, 1) then we compute

h (t) = 1(t) et (cos 3t 13 sin 3t)

Figure 2.15 A sequence of inputs giving the impulse response in

the limit

which is the right plot in Figure 2.13.

4. The final case we take is that when there is no damping: d = 0 and k = 1. Then we have

cos t sin t
eAt =
.
sin t cos t

2.34 Theorem If
Z
y (t) =

(a) Taking c = (1, 0) we compute

ct eA(t ) bu ( ) d

then

h (t) = 1(t) sin t

lim y (t) = h (t).

which is the left plot in Figure 2.14.

(b) If we let c = (0, 1) then we compute
h (t) = 1(t) cos t
which is the right plot in Figure 2.14.

Let us see if we can give some justification to the formula for the impulse response. For
> 0 define u U by
(
1
, t [0, ]
u (t) =
0, otherwise.
The behaviour of these inputs as shrinks is shown in Figure 2.15. It turns out that these
inputs in the limit 0 give the impulse response. Note, however, that in the limit we do
not get an input in U !

Proof We use the definition of the matrix exponential:

Z t
y (t) =
ct eA(t ) bu ( ) d
0
Z

1
A2 2
=
ct eAt I n A +
+ . . . b d.
0
2!
Since the sum for the matrix exponential converges uniformly and absolutely on [0, ] we
may distribute the integral over the sum:

1
A2 A2 3
y (t) = ct eAt I n
+
+ ... b

2!
3!

A A2 2
= ct eAt I n
+
+ . . . b.
2!
3!

2.4 The impulse response

22/10/2004

Clearly the result holds when we take the limit 0.

is given by

2 State-space representations (the time-domain)

x(t) = eAt x0

22/10/2004

eA(t ) bu( ) d,

t 0.

2.4.2 The impulse response for anticausal systems Now we turn our attention to a
situation that we will only have need to resort to in Section 15.3; the situation is one where
we deal with functions of time that end at t = 0. Thus functions are defined on the interval
(, 0]. The definition of the impulse response in these cases has the same motivation as
in the causal case. We shall use Theorem 2.34 for our motivation. For > 0, let us define
(
1
, t [, 0]
u (t) =
0, otherwise.

Clearly the final condition x(0) = x0 is satisfied. With x(t) so defined, we compute
Z 0

AeA(t ) bu( ) d
x(t)
= AeAt x0 + bu(t)
t
Z 0
eA(t ) bu( ) d + bu(t)
= AeAt x0 A

We then define

Thus x(t) also satisfies the differential equation.

= Ax(t) + bu(t).

t A(t )

y (t) =

bu ( ) d,

t 0,

and then h
= lim0 y . Let us determine the expression for h be performing the computations carried out in the proof of Theorem 2.34, but now for t 0:
Z 0
ct eA(t ) bu ( ) d
y (t) =
t
Z

1 0 t At
A2 2
=
ce
I n A +
+ b d

2!

1 t At
A2 A2 3
= ce

+ b
I n +

2!
3!

A A2 2
t At
= c e
In
+
+ b.
2!
3!

Thus the anticausal impulse response acts for anticausal inputs in much the same way as
the causal impulse response acts for causal inputs.
Note again that we shall only rarely require h
, so, again, whenever you see h , it
implicitly refers to h+
.

2.5 Canonical forms for SISO systems

In this section we look at the appearance of a typical SISO linear system of the
form (2.2). To do so, we shall take an arbitrary system of that form and make a linear
change of coordinates. So let us first make sure we understand what is a linear change
of coordinates, and how it manifests itself in the multi-input, multi-output system equations (2.1). We take as our state coordinates x, and define new state coordinates = T x
where T is an invertible n n matrix.2 We can easily derive the equations that govern the
behaviour of the state variables . The following result holds in the MIMO case.

Therefore, we conclude that

h
(t)

2.36 Proposition If u(t) Rm , x(t) Rn , and y(t) Rr satisfy

(
ct eAt b, t 0
=
0,
otherwise.

x(t)
= Ax(t) + Bu(t)
y(t) = Cx(t) + Du(t),

t At
This may be written as h
(t) = 1(t)c e b, which we call the anticausal impulse
response.
The anticausal impulse response is useful for solving a final value problem, as the following
result states.

2.35 Proposition Let = (A, b, ct , 01 ) be a SISO linear system. If u : (, 0] R is

admissible then the output of the final value problem

t 0.

Proof The result will follow if we can show that the solution to the final value problem

x(t)
= Ax(t) + bu(t) x(0) = x0

Proof We compute

(t)
= T x(t)
= T Ax(t) + T Bu(t) = T AT 1 (t) + T Bu(t),

One may consider more general changes of variable where one defines = Q y and =
R1 u, but since our interest is mainly in the SISO case, such transformations simply boil
down to scaling of the variables, and so constitute nothing of profound interest.

h
(t )u( ) d,

y(t) = CT 1 (t) + Du(t).

is given by
Z

(t)
= T AT 1 (t) + T Bu(t)

and y(t) = Cx(t) + Du(t) = CT 1 (t) + Du(t).

x(t)
= Ax(t) + Bu(t) x(0) = x0
y(t) = ct x(t),

y(t) = ct eAt x0 +

and if = T x, then u(t) Rm , (t) Rn , and y(t) Rr satisfy

2
Often T is arrived at as follows. One has n linearly independent vectors {f 1 , . . . , f n } in Rn which
therefore form a basis. If we assemble into the columns of a matrix T 1 the components of the vectors
f 1 , . . . , f n that is we take

T 1 = f 1 f n

then = T x are exactly the components of x Rn in the basis {f 1 , . . . , f n }.

22/10/2004

2.5 Canonical forms for SISO systems

2.5.1 Controller canonical form We now revert to the SISO setting, and prove a
normal form result for controllable SISO linear systems. Recall that a pair (A, b)
Rnn Rn is controllable if the vectors {b, Ab, A2 b, . . . , An1 b} form a basis for Rn .
2.37 Theorem If (A, b) Rnn Rn is controllable then there exists an invertible n n
matrix T with the property that

0
1
0
0
0
0
0
0
0
1
0
0

0
0
0
0
1
0

T AT 1 = ..
..
..
..
.. , T b = ..
..
.

.
.
.
.
.
.

0
0
0
0
0
1
p0 p1 p2 p3 pn1
1
Proof We begin with some seemingly unrelated polynomial constructions. Let the characteristic polynomial of A be
PA () = n + pn1 n1 + + p1 + p0 .
Define n + 1 polynomials in indeterminant by
Pi () =

ni
X

pk+i k ,

i = 0, . . . , n.

k=0

Note that P0 = PA and Pn () = 1 if we declare that pn = 1. These polynomials satisfy the

relation
Pi () = Pi1 () pi1 Pn ().
(2.11)
Indeed, we compute
Pi () =
=

ni
X
k=0
ni
X

pk+i k+1
pk+i k+1 + pi1 pi1

k=0
ni
X

pk+i k+1 pi1 Pn ()

k=1
n(i1)

pk0 +(i1) k pi1 Pn ()

k0 =0

= Pi1 () pi1 Pn (),

as asserted.
Now we define n + 1 vectors by
f i = Pi (A)b,

i = 0, . . . , n.

Note that Pi (A) is simply given by

Pi (A) =

ni
X
k=0

2 State-space representations (the time-domain)

By the Cayley-Hamilton Theorem, f 0 = 0, and we claim that the vectors {f 1 , . . . , f n } are

linearly independent. To see this latter assertion, note that since (A, b) is controllable the
vectors {g 1 = An1 b, g 2 = An2 b, . . . , g n = b} are linearly independent and so form a basis
for Rn . We also have
n
X
fi =
Tji g j
j=1

where T is the n n matrix

0
1

0
0
1
..
.

pn1

T = pn2 pn1
..
..
.
.
p1
p2 p3

0
0

. . . ..
.
1

which is clearly invertible. Therefore {f 1 , . . . , f n } are themselves linearly independent and

so form a basis.
We define the matrix T by asking that its inverse be given by

T 1 = f 1 f n
so that T AT 1 is simply the representation of the linear map A in the basis {f 1 , . . . , f n }.
The relation (2.11) gives
Af i = f i1 pi1 f n ,
from which we get the representation of A in the basis {f 1 , . . . , f n } as in the theorem
statement. It is trivially true that the coordinates of b in this basis are (0, 0, . . . , 0, 1) since
f n = b.

The pair (T AT 1 , T b) of the theorem are sometimes called the controller canonical
form for the pair (A, b). This is also sometimes known as the second LuenbergerBrunovsky canonical form for (A, b).
What is the import of this theorem? Well, let us suppose that we are handed an n n
matrix A and an n-vector b in the form of that in the theorem statement:

0
1
0
0
0
0
0
0
0
1
0
0

0
0
0
0
1
0

A = ..
..
..
..
.. , b = .. .
.
.
.

.
.
.
.
.
.

0
0
0
0
0
1
1
p0 p1 p2 p3 pn1
What does the system look like, really? Well, define a scalar variable x by x = x1 . We then

note that if x(t)

= Ax(t) + bu(t) then
x = x 1 = x2
x = x1 = x 2 = x3
..
.
x(n) = x n = p0 x1 p1 x2 pn1 xn

pk+i Ak ,

i = 0, . . . , n.

22/10/2004

= p0 x p1 x pn1 x(n1) + u.

22/10/2004

2.5 Canonical forms for SISO systems

Thus the vector equation x(t)

= Ax(t) + bu(t) reduces to the scalar equation
x(n) (t) + pn1 x(n1) (t) + + p1 x(1) (t) + p0 x(t) = u(t).
Therefore, when we study the controllable SISO linear system

x(t)
= Ax(t) + bu(t)
y(t) = ct x(t) + Du(t),

(2.12)

we can always make a change of coordinates that will render the system an nth order one
whose state variable is a scalar. This is important. It is also clear that if one conversely
starts with a scalar system
x(n) (t) + pn1 x(n1) (t) + + p1 x(1) (t) + p0 x(t) = bu(t)
y(t) = cn1 x(n1) (t) + cn2 x(n2) (t) + + c1 x(1) (t) + c0 x(t) + du(t),
one may place it in the form of (2.12) where

0
1
0
0
0
0
1
0

0
0
0
1

A = ..
..
..
..
.
.
.
.

0
0
0
0
p0 p1 p2 p3

0
c0
0
c1

b = 0 , c = c2 ,
..
..
.
.
1
cn1

..
.

0
0
0
..
.

1
pn1

2 State-space representations (the time-domain)

Proof We make the simple observation that (A, c) is observable if and only if (At , c) is
controllable. This follows from the easily seen fact that C(A, c) = O(At , c)t . Therefore, by
Theorem 2.37 there exists an invertible n n matrix T so that

0
1
0
0
0
0
0
0
0
1
0
0

0
0
0
0
1
0

T At T 1 = ..
..
..
..
.. , T c = .. .
..
.

.
.
.
.
.
.

0
0
0
0
0
1
p0 p1 p2 p3 pn1
1
Thus

0
1

T t AT t = 0
..
.

0
0

by rescaling variables (for example, if d = 0 we may choose u(t) = bu(t)).

We shall see that for controllable SISO systems, we may move easily from the linear
systems formulation, i.e., equation (2.2), to the scalar equation formulation in all settings
we examine, and here we have provided the time-domain setting for making this change.
We look at alternative canonical forms for controllable pairs (A, b) in Exercises E2.31 E2.32,
and E2.33.
2.5.2 Observer canonical form Now let us focus on the situation when the system is
observable. The proof here is simpler than for the controllable case, since we use a duality
between controllability and observability.
2.38 Theorem If (A, c) Rnn Rn is observable then there exists an invertible nn matrix
T so that

0 0 0 0 p0
1 0 0 0 p1

0 1 0 0 p2

T AT 1 = 0 0 1 0 p3 , ct T 1 = 0 0 0 0 1 .
.. .. .. . . ..
..
. . .
. .
.

0 0 0 0 pn2
0 0 0 1 pn1

0
0
1
0
..
.

0
0
0
1
..
.

..
.

0
0
0
0
..
.

p0
p1
p2
p3
..
.

0 0 0 pn2
0 0 1 pn1

ct T t = 0 0 0 0 1 .

The result now follows by letting T = T t .

D= 1

22/10/2004

The pair (T AT , T c) in the theorem statement are said to be in observer canonical

form or in second Luenberger-Brunovsky canonical form
Let us look at the value of this canonical form by expression the system equations for a
system that has this form. The differential equation x(t)

= Ax(t) + bu(t) reads

x 1 = p0 xn = b0 u
x 2 = x1 p1 xn = p0 x p1 xn = b1 u
..
.
x n = xn1 pn1 xn = bn1 u,
where we write b = (b0 , b1 , . . . , bn1 ). Now we differentiate the expression for x n with respect
to t n 1 times, and use the equations for x 1 , . . . , x n to get the equation
(n1)
(n1)
x(n)
+ + p1 x(1)
+ + b1 u(1) + b0 u.
n + pn1 xn
n + p0 xn = bn1 u

The equation y = ct x + Du simply reads y = xn + Du. The upshot, therefore, is that for
an observable system one can always make a change of coordinates so that the system is
effectively described by the equation
y (n) + pn1 y (n1) + + p1 y (1) + p0 y = bn u(n) + bn1 u(n1) + + b1 u(1) + b0 u,
where bn is defined by D = [bn ]. Thus an observable system can be immediately put into
the form of a differential equation for the output in terms of the output and its derivatives.
This is an essential observation for the discussion of input/output systems that is initiated
in Section 3.4. As with controllable pairs, in the exercises (See E2.34, E2.35, and E2.36) we
provide alternate canonical forms for observable pairs.

22/10/2004

2.5 Canonical forms for SISO systems

2.5.3 Canonical forms for uncontrollable and/or unobservable systems As we have

seen, for systems that are either controllable or observable, it is possible to find a set of
coordinates in which the system looks simple, in some sense. Now let us address the situation
when we know that the system is not both controllable and observable.
First we consider the situation when (A, b) is not controllable. The following result
expresses the simplest form such a pair may take.
2.39 Theorem If (A, b) Rnn Rn is not controllable then there exist an invertible matrix
T and a positive integer ` < n with the property that

A11 A12
b1
T AT 1 =
, Tb =
.
(2.13 )
0n`,` A22
0n`
Furthermore, T may be chosen so that (A11 , b1 ) R`` R` are in controller canonical
form.
Proof Let V be the smallest A-invariant subspace containing b, and suppose that dim(V ) =
`. Since (A, b) is not controllable, by Theorem 2.23, ` < n. Choose a basis {v 1 , . . . , v n } for
Rn with the property that {v 1 , . . . , v ` } is a basis for V . We define T so that

T 1 = v 1 v n .
Since V is A-invariant and since b V , the relations in (2.13) must hold. Note that A11
is simply the representation in the basis {v 1 , . . . , v ` } of the restriction of A to V , and that
b1 is the representation of b V in this same basis. Now we look at the final assertion. By
the very definition of V , the pair (A11 , b1 ) is controllable. Therefore, by Theorem 2.37 there
exists an ` ` invertible matrix T t so that
T t A11 T t ,

T t b1

is in controller canonical form.

Now let us do the same thing, except now we look at the situation when (A, c) is not
observable.
2.40 Theorem If (A, c) Rnn Rn is not observable then there exist an invertible matrix
T and a positive integer k < n with the property that

A11 0k,nk
T AT 1 =
, ct T 1 = ct1 0tnk .
(2.14 )
A21 A22
Furthermore, T may be chosen so that (A11 , c1 ) Rkk Rk are in observer canonical form.
Proof Since (A, c) is not observable, (At , c) is not controllable. Therefore, by Theorem 2.39, there exists an invertible matrix T so that

11 A
12
A
c = c1 ,
T At T 1 =
,
T
22
0nk
0nk,k A
11 , c1 ) in controller canonical form. Therefore,
with (A
T t AT t ,

ct T t

will have the form stated in the theorem, and thus the result follows by taking T = T 1 .

2 State-space representations (the time-domain)

22/10/2004

Finally, we look at the case where (A, b) is not controllable and where (A, c) is not
observable.
2.41 Theorem Suppose that = (A, b, c0 , D) is neither controllable nor observable. Then
there exists integers j, k, ` > 0 and an invertible matrix T so that

A11 A12 A13 A14
b1

b2

0
A
0
A
k,j
22
j,`
24
1
t 1

T AT =
0`,j 0`,k A33 A34 , T b = 0` , c T = 0j c2 0` c4 ,
0m,j 0m,k 0m,` A44
0m
where m = n j k `, and where the pair

A11 A12
,
0k,j A22

b1
b2

is controllable and the pair

A22 A24
,
0m,k A44

c2
c4

is observable.
Proof Choose a basis {v 1 , . . . , v n } for Rn with the property that
1. {v 1 , . . . , v j } is a basis for image(C(A, b)) ker(O(A, c)),
2. {v 1 , . . . , v j , v j+1 , . . . , v j+k } is a basis for image(C(A, b)), and
3. {v 1 , . . . , v j , v j+k+1 , . . . , v j+k+` } is a basis for ker(O(A, c)).
Now define T by

T 1 = v 1 v j v j+1 v j+k v j+k+1 v j+k+` v j+k+`+1 v n .
From the properties of the basis vectors it follows that
b span {v 1 , . . . , v j , v j+1 , . . . , v j+k }
and that
c span {v j+1 , . . . , v j+k , v j+k+`+1 , . . . , v n } .
From these two observations follow the form of T b and ct T 1 in the theorem statement. Furthermore, since image(C(A, b)) and ker(O(A, c)) are A-invariant (Theorems 2.17 and 2.23),
it follows that image(C(A, b)) ker(O(A, c)) is A-invariant and that image(C(A, b)) +
ker(O(A, c)) is A-invariant. From these observations we conclude the following:
1. Av i span {v 1 , . . . , v j+k } for i {1, . . . , j + k};
2. Av i span {v 1 , . . . , v j , v j+k+1 , . . . , v j+k+` } for i {1, . . . , j, j + k + 1, . . . , j + k + `};
3. Av i span {v 1 , . . . , v j } for i {1, . . . , j};
4. Av i span {v 1 , . . . , v j , v j+1 , . . . , v j+k , v j+k+1 , . . . , v j+k+` }, i {1, . . . , j + k + `}.

2.6 Summary

22/10/2004

From these observations follow the form of T AT 1 in the theorem statement.

Now let us show that the pair

1 = b1
1 = A11 A12 , b
A
0k,j A22
b2

2 State-space representations (the time-domain)

22/10/2004

4. You should be able to determine whether a SISO linear system is observable or controllable, and know how the lack of observability or controllability affects a system.
5. You should know roughly the import of ker(O(A, c)) and of the columnspace of C(A, b).
6. You should know that there is a thing called zero dynamics, and you should convince
yourself that you can work through Algorithm 2.28 to determine this, at least if you had
some time to work it out. We will revisit zero dynamics in Section 3.3, and there you will
be given an easy way to determine whether the zero dynamics are stable or unstable.
7. You should be able to determine, by hand and with the computer, the impulse response of
a SISO linear system. You should also understand that the impulse response is somehow
basic in describing the behaviour of the systemthis will be amply borne out as we
progress through the book.
8. You should know that a controllable pair (A, b) has associated to it a canonical form, and
you should be able to write down this canonical form given the characteristic polynomial
for A.

is controllable. First, by direct calculation, we have

1 A
1 A
1
1b
n1 b
b
C(T AT 1 , T b) =
.
0`+m 0`+m
0`+m
Now, by our choice of basis vectors we also have
image(C(T AT 1 , T b)) = span {v 1 , . . . , v j , v j+1 , . . . , v j+k } .
Thus the matrix

1 A
1 A
1
1b
n1 b
b
must have maximal rank. However, by the Cayley-Hamilton Theorem it follows that the
matrix

1
1 A
1 A
1b
j+k1 b
b
is controllable.
b)
also has full rank, showing that (A,
That the pair

A22 A24
c2
,
c4
0m,k A44
is observable follows in the same manner as the previous step, noting that
ker(O(T AT 1 , T t c)) = span {v 1 , . . . , v j , v j+k+1 , . . . , v j+k+` } .

If we write the state vector as x = (x1 , x2 , x3 , x4 ) in the decomposition given by the

theorem then we may roughly say that
1. x1 represents the states that are controllable but not observable,
2. x2 represents the states that are controllable and observable,
3. x3 represents the states that are neither controllable nor observable,
4. x4 represents the states that are observable but not controllable.

2.6 Summary
This is, as we mentioned in the introduction to the chapter, a difficult bit of material. Heres what you should take away with you, and make sure you are clear on before
proceeding.
1. You should know exactly what we mean when we say SISO linear system. This terminology will be used constantly in the remainder of the book.
2. You should be able to take a physical system and put it into the form of an SISO linear
system if requested to do so. To do this, linearisation may be required.
3. Given a SISO linear system with a specified input u(t), you should know how to determine, both on paper and with the computer, the output y(t) given an initial value x(0)
for the state.

Exercises for Chapter 2

Exercises
The next three exercises are concerned with interconnections of SISO linear systems. We
shall be discussing system interconnections briefly at the beginning of Chapter 3, and thoroughly in Chapter 6. Indeed system interconnections are essential to the notion of feedback.
E2.1 Consider two SISO linear systems governed by the differential equations
(
x 1 (t) = A1 x1 (t) + b1 u1 (t)
System 1 equations
y1 (t) = ct1 x1 (t)
(
x 2 (t) = A2 x2 (t) + b2 u2 (t)
System 2 equations
y2 (t) = ct2 x2 (t),
where x1 Rn1 and x2 Rn2 . The input and output signals of System 1, denoted
u1 (t) and y1 (t), respectively, are both scalar. The input and output signals of System 2,
denoted u2 (t) and y2 (t), respectively, are also both scalar. The matrices A1 and A2
and the vectors b1 , b2 , and c1 , and c2 are of appropriate dimension.
Since each system is single-input, single-output we may connect them as shown
in Figure E2.1. The output of System 1 is fed into the input of System 2 and so the
u1 (t)

System 1

y1 (t) = u2 (t)

System 2

y2 (t)

Figure E2.1 SISO linear systems connected in series

interconnected system becomes a single-input, single-output system with input u1 (t)

and output y2 (t).
(a) Write the state-space equations for the combined system in the form

x
x 1
= A 1 + bu1
x 2
x2

t x1
y2 = c
+ Du1 ,
x2
where you must determine the expressions for A, b, c, and D. Note that the
combined state vector is in Rn1 +n2 .
(b) What is the characteristic polynomial of the interconnected system A matrix?
Does the interconnected system share any eigenvalues with either of the two
component systems?
E2.2 Consider again two SISO linear systems governed by the differential equations
(
x 1 (t) = A1 x1 (t) + b1 u1 (t)
System 1 equations
y1 (t) = ct1 x1 (t)
(
x 2 (t) = A2 x2 (t) + b2 u2 (t)
System 2 equations
y2 (t) = ct2 x2 (t),

2 State-space representations (the time-domain)

22/10/2004

where x1 Rn1 and x2 Rn2 . The input and output signals of System 1, denoted
u1 (t) and y1 (t), respectively, are both scalar. The input and output signals of System 2,
denoted u2 (t) and y2 (t), respectively, are also both scalar. The matrices A1 and A2
and the vectors b1 , b2 , and c1 , and c2 are of appropriate dimension.
Since each system is single-input, single-output we may connect them as shown
in Figure E2.2. The input to both systems is the same, and their outputs are added
System 1

y1 (t)

u1 (t) = u2 (t) = u(t)

y(t) = y1 (t) + y2 (t)

System 2

y2 (t)

Figure E2.2 SISO linear systems connected in parallel

to get the new output.

Exercises for Chapter 2

u(t)

u1 (t) =
u(t) y2 (t)

System 1

y1 (t) = u2 (t)

System 2

y(t) = y2 (t)

Figure E2.3 SISO linear systems connected in a negative feedback

loop

(a) Write the state-space equations for the combined system in the form

x 1
x
= A 1 + bu
x 2
x2

x
y = ct 1 + Du,
x2
where you must determine the expressions for A, b, c, and D. Note that the
combined state vector is in Rn1 +n2 .
(b) What is the characteristic polynomial of the interconnected system A matrix?
Does the interconnected system share any eigenvalues with either of the two
component systems?
Hint: See Exercise E3.7.
E2.4 Consider the pendulum/cart system of Exercise E1.5. If one adds a force that is
applied horizontally to the cart, this leads to a natural input for the system. As
output, there are (at least) four natural possibilities: the position of the cart, the
velocity of the cart, the pendulum angle, and the pendulum angular velocity. For
each of the following eight cases, determine the linearised equations of the form (2.2)
for the linearisations:
(a) the equilibrium point (0, 0) with cart position as output;
(b) the equilibrium point (0, 0) with cart velocity as output;
(c) the equilibrium point (0, 0) with pendulum angle as output;
(d) the equilibrium point (0, 0) with pendulum angular velocity as output;
(e) the equilibrium point (0, ) with cart position as output;
(f) the equilibrium point (0, ) with cart velocity as output;
(g) the equilibrium point (0, ) with pendulum angle as output;
(h) the equilibrium point (0, ) with pendulum angular velocity as output.
In this problem you first need to determine the nonlinear equations of the form (2.5),
and then linearise.
E2.5 Consider the double pendulum of Exercise E1.6. There are at least two ways in which
one can provide a single input to the system. The two we consider are
1. a torque at the base of the bottom link relative to the ground (we call this the
pendubot configuration), and
2. a torque applied to top link from the bottom link (we call this the acrobot
configuration).

2 State-space representations (the time-domain)

22/10/2004

There are various outputs we can consider, but let us choose the angle 2 of the top
pendulum arm.
For each of the following cases, determine the equations in the form (2.2) for the
linearisations:
(a) the equilibrium point (0, 0, 0, 0) with the pendubot input;
(b) the equilibrium point (0, , 0, 0) with the pendubot input;
(c) the equilibrium point (, 0, 0, 0) with the pendubot input;
(d) the equilibrium point (, , 0, 0) with the pendubot input;
(e) the equilibrium point (0, 0, 0, 0) with the acrobot input;
(f) the equilibrium point (0, , 0, 0) with the acrobot input;
(g) the equilibrium point (, 0, 0, 0) with the acrobot input;
(h) the equilibrium point (, , 0, 0) with the acrobot input.
The equilibrium points are written using coordinates (1 , 2 , 1 , 2 ). In this problem
you first need to determine the nonlinear equations of the form (2.5), and then linearise.
E2.6 Consider the coupled tanks of Exercise E1.11. Take the input u to be Fin . Suppose
the system is at equilibrium with the height in tank 1 denoted 1 , and the input flow
and height in tank 2 as determined in parts (d) and (e) of Exercise E1.11. Obtain the
linearised equations for the system at this equilibrium.
E2.7 Obtain the output y(t) for the SISO linear system

0
x1
x 1
+
u
=
1
x 2
x2

x1
y= 1 0
,
x2
for R and > 0 when u(t) = cos t and when x(0) = 0.
E2.8 Use a computer package to determine the output response of the SISO linear system
(A, b, ct , D) to a unit step input when

2 3
1
0
0
1
3 2 0

0
1
0
, b = , c = , D = 01 .
A=
0
0
0
0 2 3
0
0 3 2
1
0
E2.9 (a) Come up with (A, c) R33 R3 so that (A, c) is observable.
(b) Come up with (A, c) R33 R3 so that (A, c) is not observable. Choosing
either A or c to be zero is not acceptable.
E2.10 (a) Come up with (A, b) R44 R4 so that (A, b) is controllable.
(b) Come up with (A, b) R44 R4 so that (A, b) is not controllable. Choosing
either A or b to be zero is not acceptable.
E2.11 Define (A, b) Rnn Rn by

0
1
0
0
0
0
0
0
0
1
0
0

0
0
0
0
1
0

A = ..
..
..
..
.. , b = ..
...
.

.
.
.
.
.

0
0
0
0
0
1
p0 p1 p2 p3 pn1
1

Exercises for Chapter 2

for p0 , p1 , . . . , pn1 R. Show that (A, b) is controllable by verifying that the controllability matrix has full rank.
E2.12 Define (A, c) Rnn Rn by

0 0 0 0 p0
1 0 0 0 p1

0 1 0 0 p2

A = 0 0 1 0 p3 , ct = 0 0 0 0 1 .
.. .. .. . . ..
..
. . .
. .
.

0 0 0 0 pn2
0 0 0 1 pn1

E2.20

for p0 , p1 , . . . , pn1 R. Show that (A, c) is observable by verifying that the observability matrix has full rank.
The next two exercises give conditions for controllability and observability called the PopovBelevitch-Hautus conditions [see Hautus 1969].
E2.13 Show that (A, b) Rnn Rn is controllable if and only if the matrix

sI n A b

E2.21

has rank n for all s C.

E2.14 Show that (A, c) Rnn Rn is observable if and only if the matrix

sI n A
t
c
has rank n for all s C.
E2.15 Show that the definitions of controllability and observability are invariant under linear
changes of state variable.
E2.16 Consider the circuit of Exercise E1.7. Take as output the current through the resistor.
(a) Give conditions under which the system is observable.
(b) Give conditions under which the system is controllable.
E2.17 Consider the circuit of Exercise E1.8. Take as output the current through the resistor
R1 .
(a) Give conditions under which the system is observable.
(b) Give conditions under which the system is controllable.
E2.18 Consider the circuit of Exercise E1.9. Take as output the current emanating from
the voltage source (by the Kirchhoff current law, this is also the sum of the currents
through the two resistors).
(a) Give conditions under which the system is observable.
(b) Give conditions under which the system is controllable.
E2.19 For the coupled masses of Exercise E1.4 (assume no damping), suppose you apply a
force to the leftmost mass of magnitude F1 = u(t). You also apply a force to the
rightmost mass that is proportional to F1 ; thus you take F2 = u(t) for some R.
The system is still single-input since the two forces are essentially determined by u(t).
As an output for the system take the displacement of the rightmost mass.

E2.22

E2.23

2 State-space representations (the time-domain)

22/10/2004

(a) Determine for which values of the system is observable. For those cases when
the system is not observable, can you give a physical interpretation of why it is
not by looking at ker(O(A, c))?
(b) Determine for which values of the system is controllable. For those cases when
the system is not controllable, can you give a physical interpretation of why it is
not by looking at image(C(A, b))?
For the pendulum/cart system of Exercises E1.5 and E2.4, determine whether the
linearisations in the following cases are observable and/or controllable:
(a) the equilibrium point (0, 0) with cart position as output;
(b) the equilibrium point (0, 0) with cart velocity as output;
(c) the equilibrium point (0, 0) with pendulum angle as output;
(d) the equilibrium point (0, 0) with pendulum angular velocity as output;
(e) the equilibrium point (0, ) with cart position as output;
(f) the equilibrium point (0, ) with cart velocity as output;
(g) the equilibrium point (0, ) with pendulum angle as output;
(h) the equilibrium point (0, ) with pendulum angular velocity as output.
Make sense of your answers by examining ker(C(A, b)) and image(O(A, c)).
Consider the the double pendulum of Exercises E1.6 and E2.5. For each of the following cases, determine whether the linearisation is controllable and/or observable:
(a) the equilibrium point (0, 0, 0, 0) with the pendubot input;
(b) the equilibrium point (0, , 0, 0) with the pendubot input;
(c) the equilibrium point (, 0, 0, 0) with the pendubot input;
(d) the equilibrium point (, , 0, 0) with the pendubot input;
(e) the equilibrium point (0, 0, 0, 0) with the acrobot input;
(f) the equilibrium point (0, , 0, 0) with the acrobot input;
(g) the equilibrium point (, 0, 0, 0) with the acrobot input;
(h) the equilibrium point (, , 0, 0) with the acrobot input.
In each case, the output is the angle of the second link.
Consider the coupled tank system of Exercises E1.11 and E2.6. Determine whether
the system is controllable and/or observable for the following outputs:
(a) the output is the level in tank 1;
(b) the output is the level in tank 2;
(c) the output is the difference in the levels.
Determine the zero dynamics for the pendulum/cart system of Exercises E1.5 and E2.4
for each of the following linearisations:
(a) the equilibrium point (0, 0) with cart position as output;
(b) the equilibrium point (0, 0) with cart velocity as output;
(c) the equilibrium point (0, 0) with pendulum angle as output;
(d) the equilibrium point (0, 0) with pendulum angular velocity as output;
(e) the equilibrium point (0, ) with cart position as output;
(f) the equilibrium point (0, ) with cart velocity as output;
(g) the equilibrium point (0, ) with pendulum angle as output;
(h) the equilibrium point (0, ) with pendulum angular velocity as output.

Exercises for Chapter 2

E2.24 For the double pendulum of Exercises E1.6 and E2.5, and for each of the following
cases, determine the zero dynamics:
(a) the equilibrium point (0, 0, 0, 0) with the pendubot input;
(b) the equilibrium point (0, , 0, 0) with the pendubot input;
(c) the equilibrium point (, 0, 0, 0) with the pendubot input;
(d) the equilibrium point (, , 0, 0) with the pendubot input;
(e) the equilibrium point (0, 0, 0, 0) with the acrobot input;
(f) the equilibrium point (0, , 0, 0) with the acrobot input;
(g) the equilibrium point (, 0, 0, 0) with the acrobot input;
(h) the equilibrium point (, , 0, 0) with the acrobot input.
In each case, use the angle of the second link as output.
E2.25 Determine the linearised zero dynamics of for the coupled tank system of Exercises E1.11 and E2.6 for the following outputs:
(a) the height in tank 1;
(b) the height in tank 2;
(c) the difference of the heights in the tanks.
E2.26 Define a SISO linear system = (A, b, ct , 01 ) with

0
1
A=
, b=
, c=

1
0
for R and > 0. Determine the impulse response h (t). Plot the impulse
response for various values of .
E2.27 Consider the pendulum/cart system of Exercises E1.5 and E2.4, and determine the
impulse response of the system for the following linearisations:
(a) the equilibrium point (0, 0) with cart position as output;
(b) the equilibrium point (0, 0) with cart velocity as output;
(c) the equilibrium point (0, 0) with pendulum angle as output;
(d) the equilibrium point (0, 0) with pendulum angular velocity as output;
(e) the equilibrium point (0, ) with cart position as output;
(f) the equilibrium point (0, ) with cart velocity as output;
(g) the equilibrium point (0, ) with pendulum angle as output;
(h) the equilibrium point (0, ) with pendulum angular velocity as output.
E2.28 Select values for the parameters of the double pendulum system of Exercises E1.6
and E2.5. For each of the following cases, determine the impulse response for the
linearisation:
(a) the equilibrium point (0, 0, 0, 0) with the pendubot input;
(b) the equilibrium point (0, , 0, 0) with the pendubot input;
(c) the equilibrium point (, 0, 0, 0) with the pendubot input;
(d) the equilibrium point (, , 0, 0) with the pendubot input;
(e) the equilibrium point (0, 0, 0, 0) with the acrobot input;
(f) the equilibrium point (0, , 0, 0) with the acrobot input;
(g) the equilibrium point (, 0, 0, 0) with the acrobot input;
(h) the equilibrium point (, , 0, 0) with the acrobot input.

2 State-space representations (the time-domain)

22/10/2004

In each case, use the angle of the second link as output.

E2.29 Determine the linearised impulse response for the coupled tank system of Exercises E1.11 and E2.6 for the following outputs:
(a) the height in tank 1;
(b) the height in tank 2;
(c) the difference of the heights in the tanks.
E2.30 Let

2 1 3
0
A = 0 2 1 , b = 1 .
0
0 1
1
(a) Verify that (A, b) is controllable.
(b) Find the controller canonical form for (A, b).
The next few exercises deal with alternative canonical forms for controllable pairs (A, b) and
for observable pairs (A, c).
E2.31 Let (A, b) be a controllable pair. Show
basis {b, Ab, . . . , An1 b} are given by

0 0
1 0

0 1

T AT 1 = .. .. . .
. .
.

0 0
0 0

that the representations of A and b in the

0
0
0
..
.

p0
p1
p2
..
.

0 pn2
1 pn1

1
0

0

T b = .. ,
.

0
0

where
T 1 =

b Ab An1 b .

This is the controllability canonical form or the first Luenberger-Brunovsky

canonical form for (A, b).
E2.32 Again let (A, b) be controllable. Recall from Exercise E2.31 that if

b Ab An1 b ,
T 1
1 =
then

T 1 AT 1
1

0
1

= 0
..
.

0 p0
0 p1

0 p2
,
..
. . ..
. .
.
0 0 1 pn1
0
0
1
..
.

Now answer the following questions.

(a) Let

1 pn1
0
0
1
pn1

..
..
T 2 = ...
.
.

0
0
0
0
0
0

1
0

T 1 b = 0 ,
..
.
0

0
0
0
0

.. .
. . ..
. .
.

1 pn1
0
1

Exercises for Chapter 2

Use mathematical induction to show that

1 pn1 p2n1
0
1
pn1

..
..
..
T 1
.
.
2 = .

0
0
0
0
0
0
(b) Define T = T 2 T 1 and show that

pn1 pn2
1
0

1
T AT 1 = 0
..
..
.
.
0
0

...

..
.
1
0

pn1
1

p1 p0

0
0

0
0
,
..
..
..
.
.
.

1
0

1
0

T b = 0 .
..
.
0

are given by

T AT 1

1
0
0
..
.

0
1
0
..
.

0
0
1
..
.

..
.

0
0
0
..
.

=
,

0
0
0
0
1
p0 p1 p2 p3 pn1

pn1
n1

pn2
n1

E2.34 Let (A, c) be an observable pair. Show that the representations of A and c in the
basis formed from the columns of the matrix

1
c At c (At )n1 c

0
0
0
..
.

2 State-space representations (the time-domain)

22/10/2004

then
n2
pn1
n3
pn1

E2.33 Let (A, b) be controllable. Find an invertible matrix T so that

pn1 1 0 0
0
pn2 0 1 0
0

0
pn3 0 0 0

T AT 1 = ..
.. .. . . .. , T b = .. .
.
.
. .
.
.

p1 0 0 1
0
p0 0 0 0
1

ct T 1 = 1 0 0 ,

where

T = c At c (At )n1 c .
This is the observability canonical form or the first Luenberger-Brunovsky
canonical form for (A, c).
E2.35 Again let (A, c) be observable. Recall from Exercise E2.34 that if

T 1 = c At c (At )n1 c ,

T 1 AT 1
1

0
0
0
..
.

1
0
0
..
.

0
1
0
..
.

...

0
0
1
..
.

0
0
0
0
p0 p1 p2 p3

1
pn1

Now answer the following questions.

(a) Let

1
0
pn1
1

0
pn1

1
T 2 = ..
..
.
.

0
0
0
0
and use mathematical induction to

1
pn1

p2
n1
T 2 = ..
.
n2
pn2
pn1
n1
(b) Define T = T 2 T 1 and show that

pn1
pn2

pn3

T AT 1 = ..
.

p1
p0

0
0
0
..
.

0
0
1
..
.

..
.

0
0
0
..
.

0
0

1
pn1

0
0

..
.

0
1

show that
0
1
pn1
..
.

..
.

0
0
0
..
.

pn3
1
n1
pn2
n1 pn1

0
0

,
.
.. .
. .

0 0 1
0 0 0

1
0
0
..
.

ct T 1
1 = 1 0 0 .

0
1
0
..
.

0
0

.. .
.

0
1

ct T 1

1
0

= 0 .
..
.
0

E2.36 Let (A, c) be observable. Find an invertible matrix T so that

pn1 pn2 p1 p0
1
0

0
0

0
1
1

0
0
T AT =
, ct T 1 = 0 0 0 1 .
..

..
.
.
.
..
..
..
.
.
0
0

1
0
E2.37 Consider the SISO linear system = (A, b, ct , 01 ) of Example 2.19:

1 0
0
0
A=
, b=
, c=
.
1 1
1
1

Exercises for Chapter 2

Does there exist an invertible matrix P for which

0
1
0
T AT 1 =
, Tb =
p0 p1
1
for some p0 , p1 R?
E2.38 Let (A, b) Rnn Rn be controllable. Show that there exists a unique invertible
T Rnn for which (T AT 1 , T b) is in controller canonical form.
For a given matrix A Rnn it may not be possible to find a vector b Rn so that (A, b)
is controllable. This is related to the Jordan canonical form for A, and in the next two
exercises you are asked to look into this a little.
E2.39 Let = (A, b, ct , D) be a SISO linear system where A has repeated eigenvalues and
is diagonalisable.
(a) Is (A, b) controllable?
(b) Is (A, c) observable?
(c) When is the matrix

0
1
0
0
0
0
0
1
0
0

0
0
0
1
0

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

.
.
.
.
.

0
0
0
0
1
p0 p1 p2 p3 pn1
diagonalisable?
E2.40 Define

A1 =
0
0

0
0
0
0

0
0
,

A2 =
0
0

1
0
0
0

0
1

(a) Show that there is no vector b R4 so that (A1 , b) is controllable.

(b) Let V R4 be the subspace spanned by {(1, 0, 0, 0), (0, 1, 0, 0)}. Show that
(A2 , b) is controllable if and only if b 6 V .

2 State-space representations (the time-domain)

22/10/2004

This version: 22/10/2004

3.7

Chapter 3

3 Transfer functions (the s-domain)

22/10/2004

3.6.4

Computing outputs for SISO linear systems

3.6.5

Formulae for impulse, step, and ramp responses . . . . . . . . . . . . . . . . . . . 108

. . . . . . . . . . . . . . . . . . . . . 105

Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111

Transfer functions (the s-domain)

3.1 Block diagram algebra

We have informally drawn some block diagrams, and it is pretty plain how to handle
them. However, let us make sure we are clear on how to do things with block diagrams.
In Section 6.1 we will be looking at a more systematic way of handling system with interconnected blocks of rational functions, but our discussion here will serve for what we need
immediately, and actually serves for a great deal of what we need to do.
The blocks in a diagram will contain rational functions with indeterminate being the
Laplace transform variable s. Thus when you see a block like the one in Figure 3.1 where

Although in the previous chapter we occasionally dealt with MIMO systems, from now
on we will deal exclusively with SISO systems unless otherwise stated. Certain aspects of
what we say can be generalised to the MIMO setting, but it is not our intention to do this
here.
Much of what we do in this book revolves around looking at things in the sdomain, i.e., the complex plane. This domain comes about via the use of the Laplace
transform. It is assumed that the reader is familiar with the Laplace transform, but we
review some pertinent aspects in Section E.3. We are a little more careful with how we
use the Laplace transform than seems to be the norm. This necessitates the use of some
Laplace transform terminology that may not be familiar to all students. This may make
more desirable than usual a review of the material in Section E.3. In the s-domain, the
things we are interested in appear as quotients of polynomials in s, and so in Appendix C we
provide a review of some polynomial things you have likely seen before, but perhaps not as
systematically as we shall require. The transfer function that we introduce in this chapter
will be an essential tool in what we do subsequently.

x
1 (s)

x
2 (s)

R(s)

Figure 3.1 The basic element in a block diagram

R R(s) is given by
R(s) =

pn sn + pn1 sn1 + + p1 s + p0
qk sk + qk1 sk1 + + q1 s + q0

means that x2 (s) = R(s)

x1 (s) or that x1 and x2 are related in the time-domain by

Contents

(n)

(n1)

pn x1 + pn1 x1
3.1

Block diagram algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

3.2

The transfer function for a SISO linear system . . . . . . . . . . . . . . . . . . . . . . . .

3.3

Properties of the transfer function for SISO linear systems . . . . . . . . . . . . . . . . .

3.3.1

Controllability and the transfer function . . . . . . . . . . . . . . . . . . . . . . .

3.3.2

Observability and the transfer function . . . . . . . . . . . . . . . . . . . . . . . .

3.3.3

Zero dynamics and the transfer function . . . . . . . . . . . . . . . . . . . . . . .

3.4

Transfer functions presented in input/output form . . . . . . . . . . . . . . . . . . . . . .

3.5

The connection between the transfer function and the impulse response . . . . . . . . . .

3.5.1

Properties of the causal impulse response . . . . . . . . . . . . . . . . . . . . . . .

3.5.2

Things anticausal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

The matter of computing outputs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

3.6

3.6.1

Computing outputs for SISO linear systems in input/output form using the right
causal Laplace transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

3.6.2

Computing outputs for SISO linear systems in input/output form using the left
causal Laplace transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101

3.6.3

Computing outputs for SISO linear systems in input/output form using the causal
impulse response . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102

(1)

(k)

(k1)

+ + p1 x1 + p0 x1 = qk x2 + qk1 x2

(1)

+ + q1 x 2 + q0 x 2

(ignoring initial conditions). We shall shortly see just why this should form the basic element
for the block diagrams we construct.
Now let us see how to assemble blocks and obtain relations. First lets look at two blocks
in series as in Figure 3.2. If one wanted, one could introduce a variable x that represents
x
1 (s)

R1 (s)

R2 (s)

x
2 (s)

Figure 3.2 Blocks in series

the signal between the blocks and then one has

x(s) = R1 (s)
x1 (s), x2 (x) = R2 (s)x(s)
x2 (s) = R1 (s)R2 (s)
x1 (s).

Since multiplication of rational functions is commutative, it does not matter whether we

write R1 (s)R2 (s) or R2 (s)R1 (s).
We can also assemble blocks in parallel as in Figure 3.3. If one introduces temporary

3.1 Block diagram algebra

22/10/2004

78
x
1 (s)

R1 (s)
x
1 (s)

3 Transfer functions (the s-domain)

R21 (s)

R1 (s)

R2 (s)

22/10/2004

x
2 (s)

x
2 (s)
R2 (s)

Figure 3.5 A unity feedback equivalent for Figure 3.4

Figure 3.3 Blocks in parallel

signals x1 and x2 for what comes out of the upper and lower block respectively, then we have
x1 (s) = R1 (s)
x1 (s),

x2 (s) = R2 (s)
x1 (s).

Notice that when we just split a signal like we did before piping x1 into both R1 and R2 ,
the signal does not change. The temporary signals x1 and x2 go into the little circle that is
a summer . This does what its name implies and sums the signals. That is

We shall sometimes consider the case where we have unity feedback (i.e., R2 (s) = 1)
and to do so, we need to show that the situation in Figure 3.4 can be captured with unity
feedback, perhaps with other modifications to the block diagram. Indeed, one can check
that the relation between x2 and x1 is the same for the block diagram of Figure 3.5 as it is
for the block diagram of Figure 3.4.
In Section 6.1 we will look at a compact way to represent block diagrams, and one that
enables one to prove some general structure results on how to interconnect blocks with
rational functions.

x2 (s) = x1 (s) + x2 (s).

3.2 The transfer function for a SISO linear system

Unless it is otherwise depicted, the summer always adds the signals. Well see shortly what
one has to do to the diagram to subtract. We can now solve for x2 in terms of x1 :
x2 (s) = (R1 (s) + R2 (s))
x1 (s).
The final configuration we examine is the negative feedback configuration depicted
in Figure 3.4. Observe the minus sign attributed to the signal coming out of R2 into the
x
1 (s)

R1 (s)

The first thing we do is look at our linear systems formalism of Chapter 2 and see how
it appears in the Laplace transform scheme.
We suppose we are given a SISO linear system = (A, b, ct , D), and we fiddle with
Laplace transforms a bit for such systems. Note that one takes the Laplace transform of a
vector function of time by taking the Laplace transform of each component. Thus we can
take the left causal Laplace transform of the linear system

x(t)
= Ax(t) + bu(t)
y(t) = ct x(t) + Du(t)

x
2 (s)

(3.1)

to get
R2 (s)

Figure 3.4 Blocks in negative feedback configuration

summer. This means that the signal going into the R1 block is x1 (s) R2 (s)
x2 (s). This
then gives

x2 (s) = R1 (s)(
x1 (s) R2 (s)
x2 (s))
R1 (s)
x2 (s) =
x1 (s).
1 + R1 (s)R2 (s)

We emphasise that when doing block diagram algebra, one need not get upset when dividing
by a rational function unless the rational function is identically zero. That is, dont be
thinking to yourself, But what if this blows up when s = 3? because this is just not
something to be concerned about for rational function arithmetic (see Appendix C).

+
+
+
sL0
(x)(s) = AL0
(x)(s) + bL0
(u)(s)
+
t
+
+
L0 (y)(s) = c L0 (x)(s) + DL0
(u)(s).

It is convenient to write this in the form

+
+
(u)(s)
(x)(s) = (sI n A)1 bL0
L0
+
+
+
L0
(y)(s) = ct L0
(x)(s) + DL0
(u)(s).

(3.2)

We should be careful how we interpret the inverse of the matrix sI n A. What one does
is think of the entries of the matrix as being polynomials, so the matrix will be invertible
provided that its determinant is not the zero polynomial. However, the determinant is simply
the characteristic polynomial which is never the zero polynomial. In any case, you should
not really think of the entries as being real numbers that you evaluate depending on the
value of s. This is best illustrated with an example.
3.1 Example Consider the mass-spring-damper A matrix:

0
1
s
1
A=
=
sI 2 A = k
k
d
d .
m m
s+ m
m

3.2 The transfer function for a SISO linear system

22/10/2004

3 Transfer functions (the s-domain)

To compute (sI 2 A)1 we use the formula (A.2):

(sI 2 A)1

22/10/2004

1
adj(sI 2 A),
=
det(sI 2 A)

u
(s)

(sI n A)1

(s)
x

y(s)

where adj is the adjugate defined in Section A.3.1. We compute

det(sI 2 A) = s2 +

d
s
m

k
.
m

This is, of course, the characteristic polynomial for A with indeterminant s! Now we use
the cofactor formulas to ascertain

s + md 1
adj(sI 2 A) =
k
s
m
and so
(sI 2 A)1 =

1
d
s
m

s2 +

k
m

2. If we take the velocity of the mass as output, then c = (0, 1) and D = 01 and with this
we compute

s + md 1
.
k
m s

k
Note that we do not worry whether s2 + md s + m
vanishes for certain values of s because
we are only thinking of it as a polynomial, and so as long as it is not the zero polynomial,
we are okay. And since the characteristic polynomial is never the zero polynomial, we are
always in fact okay.

Back to the generalities for the moment. We note that we may, in the Laplace transform
+
+
domain, solve explicitly for the output L0
(y) in terms of the input L0
(u) to get
+
L0
(y)(s)

= c (sI n A)

Note we may write

T (s) ,

+
bL0
(u)(s)

and we call T the transfer function for the linear system = (A, b, c , D). Clearly if
we put everything over a common denominator, we have
ct adj(sI n A)b + DPA (s)
.
PA (s)

It is convenient to think of the relations (3.2) in terms of a block diagram, and we show
just such a thing in Figure 3.6. One can see in the figure why the term corresponding to the
D matrix is called a feedforward term, as opposed to a feedback term. We have not yet
included feedback, so it does not show up in our block diagram.
Lets see how this transfer function looks for some examples.
3.2 Examples We carry on with our mass-spring-damper example, but now considering the
various outputs. Thus we take

0
1
0
A=
b= 1 .
k
d ,
m m
m
1. The first case is when we have the position of the mass as output. Thus c = (1, 0) and
D = 01 , and we compute
T (s) =

s2 +

k
m

s
m
d
s
m

s2 +

k
m

k
3. The final case was acceleration output, and here we had c = ( m
, md ) and D = I 1 .
We compute in this case

T (s) =

s2
m
d
s
m

s2 +

k
m

To top off this section, lets give an alternate representation for ct adj(sI n A)b.
3.3 Lemma ct adj(sI n A)b = det

1
m
d
s
m

T (s) =

+
DL0
(u)(s).

+
L0
(y)(s)
= ct (sI n A)1 b + D
+
L0 (u)(s)

T (s) =

Figure 3.6 The block diagram representation of (3.2)

sI n A b
.
t
c
0

Proof By Lemma A.1 we have

sI n A b
det
= det(sI n A) det(ct (sI n A)1 b)
t
c
0

sI n A b
det
ct
0
= ct (sI n A)1 b =
det(sI n A)
Since we also have
ct (sI n A)1 b =

ct adj(sI n A)b
,
det(sI n A)

we may conclude that

ct adj(sI n A)b = det

sI n A b
,
ct
0

as desired.

3.3 Properties of the transfer function for SISO linear systems

Now that we have constructed the transfer function as a rational function T , let us
look at some properties of this transfer function. For the most part, we will relate these
properties to those of linear systems as discussed in Section 2.3. It is interesting that we

22/10/2004

3.3 Properties of the transfer function for SISO linear systems

can infer from the transfer function some of the input/output behaviour we have discussed
in the time-domain.
It is important that the transfer function be invariant under linear changes of state
variablewed like the transfer function to be saying something about the system rather
than just the set of coordinates we are using. The following result is an obvious one.
3.4 Proposition Let = (A, b, ct , D) be a SISO linear system and let T be an invertible
n n matrix (where A is also in Rnn ). If 0 = (T AT 1 , T b, ct T 1 , D) then T0 = T .
By Proposition 2.5 this means that if we make a change of coordinate = T 1 x for the
SISO linear system (3.1), then the transfer function remains unchanged.

are coprime as elements of R[s] if and only if (A, c) is observable.

Proof Although A, b, and c are real, let us for the moment think of them as being complex.
This means that we think of b, c Cn and A as being a linear map from Cn to itself. We
also think of P1 , P2 C[s].
Since (A, b) is controllable, by Theorem 2.37 and Proposition 3.4 we may without loss
of generality suppose that

0
1
0
0
0
0
0
0
0
1
0
0

0
0
0
0
1
0

A = ..
(3.3)
..
..
..
.. , b = .. .
.
.
.

.
.
.
.
.
.

0
0
0
0
0
1
p0 p1 p2 p3 pn1
1
Let us first of all determine T with A and b of this form. Since the first n 1 entries
of b are zero, we only need the last column of adj(sI n A). By definition of adj, this
means we only need to compute the cofactors for the last row of sI n A. A tedious but
straightforward calculation shows that

1

s

adj(sI n A) = .. . . ..
.. .
.
. .
.
sn1
Thus, if c = (c0 , c1 , . . . , cn1 ) then it is readily seen that

1
s

adj(sI n A)b = ..
.
sn1
=

P1 (s) = ct adj(sI n A)b = cn1 sn1 + cn2 sn2 + + c1 s + c0 .

(3.4)

22/10/2004

With these preliminaries out of the way, we are ready to proceed with the proof proper.
First suppose that (A, c) is not observable. Then there exists a nontrivial subspace
V Cn with the property that A(V ) V and V ker(ct ). Furthermore, we know by
Theorem 2.17 that V is contained in the kernel of O(A, c). Since V is a C-vector space and
since A restricts to a linear map on V , there is a nonzero vector z V with the property
Az = z for some C. This is a consequence of the fact that the characteristic polynomial
of A restricted to V will have a root by the fundamental theorem of algebra. Since z is an
eigenvector for A with eigenvalue , we may use (3.3) to ascertain that the components of
z satisfy
z2 = z1
z3 = z2 = 2 z1
..
.
zn1 = zn2 = n2 z1
p0 z1 p1 z2 pn1 zn = zn .

3.3.1 Controllability and the transfer function We will first concern ourselves with
cases when the GCD of the numerator and denominator polynomials is not 1.
3.5 Theorem Let = (A, b, ct , 01 ) be a SISO linear system. If (A, b) is controllable, then
the polynomials
P1 (s) = ct adj(sI n A)b, P2 (s) = PA (s)

3 Transfer functions (the s-domain)

The last of these equations then reads

zn + pn1 zn + n2 pn2 z1 + + p1 z1 + p0 z1 = 0.
Using the fact that is a root of the characteristic polynomial P2 we arrive at
zn + pn1 zn = n1 ( + pn1 )z1
from which we see that zn = n1 z1 provided 6= pn1 . If = pn1 then zn is left free.
Thus the eigenspace for the eigenvalue is

span (1, , . . . , n1 )
if 6= pn1 and

span (1, , . . . , n1 ), (0, . . . , 0, 1)
if = pn1 . In either case, the vector z 0 , (1, , . . . , n1 ) is an eigenvector for the
eigenvalue . Thus z 0 V ker(ct ). Thus means that
ct z 0 = c0 + c1 + + cn1 n1 = 0,
and so is a root of P1 (by (3.4)) as well as being a root of the characteristic polynomial
P2 . Thus P1 and P2 are not coprime.
Now suppose that P1 and P2 are not coprime. Since these are complex polynomials, this
means that there exists C so that P1 (s) = (s )Q1 (s) and P2 (s) = (s )Q2 (s) for
some Q1 , Q2 C[s]. We claim that the vector z 0 = (1, , . . . , n1 ) is an eigenvector for A.
Indeed the components of w = Az 0 are
w1 =
w2 = 2
..
.
wn1 = n2
wn = p0 p1 pn1 n1 .

22/10/2004

3.3 Properties of the transfer function for SISO linear systems

However, since is a root of P2 the right-hand side of the last of these equations is simply
n . This shows that Az 0 = w = z 0 and so z 0 is an eigenvector as claimed. Now we claim
that z 0 ker(O(A, c)). Indeed, since is a root of P1 , by (3.4) we have
ct z 0 = c0 + c1 s + + cn1 n1 = 0.
Therefore, z 0 ker(ct ). Since Ak z 0 = k z 0 we also have z 0 ker(ct Ak ) for any k 1.
But this ensures that z 0 ker(O(A, c)) as claimed. Thus we have found a nonzero vector
in ker(O(A, c)) which means that (A, c) is not observable.
To complete the proof we must now take into account the fact that, in using the Fundamental Theorem of Algebra in some of the arguments above, we have constructed a proof
that only works when A, b, and c are thought of as complex. Suppose now that they are
real, and first assume that (A, c) is not observable. The proof above shows that there is
either a one-dimensional real subspace V of Rn with the property that Av = v for some
nonzero v V and some R, or that there exists a two-dimensional real subspace V of
Rn with vectors v 1 , v 2 V with the property that
Av 1 = v 1 v 2 ,

Av 2 = v 1 + v 2

for some , R with 6= 0. In the first case we follow the above proof and see that R
is a root of both P1 and P2 , and in the second case we see that + i is a root of both P1
and P2 . In either case, P1 and P2 are not coprime.
Finally, in the real case we suppose that P1 and P2 are not coprime. If the root they share
is R then the nonzero vector (1, , . . . , n1 ) is shown as above to be in ker(O(A, c)).
If the root they share is = + i then the two nonzero vectors Re(1, , . . . , n1 ) and
Im(1, , . . . , n1 ) are shown to be in ker(O(A, c)), and so (A, c) is not observable.

I hope you agree that this is a non-obvious result! That one should be able to infer observability merely by looking at the transfer function is interesting indeed. Let us see that this
works in an example.
3.6 Example (Example 2.13 contd) We consider a slight modification of the example Example 2.13 that, you will recall, was not observable. We take

0 1
0
1
A=
, b=
, c=
,
1
1
1
from which we compute
ct adj(sI 2 A)b = 1 s
det(sI 2 A) = s2 s 1.
Note that when = 0 we have exactly the situation of Example 2.13. The controllability
matrix is

0 1
C(A, b) =
1
and so the system is controllable. The roots of the characteristic polynomial are

4 + 2
s=
2

3 Transfer functions (the s-domain)

22/10/2004

and ct adj(sI 2 A)b has the single root s = 1. The characteristic polynomial has a root of
1 when and only when = 0. Therefore, from Theorem 3.5 (which applies since (A, b) is
controllable) we see that the system is observable if and only if 6= 0. This can also be seen
by computing the observability matrix:

1
1
O(A, c) =
.
1 1
This matrix has full rank except when = 0, and this is as it should be.

Note that Theorem 3.5 holds only when (A, b) is controllable. When they are not controllable, the situation is somewhat disastrous, as the following result describes.
3.7 Theorem If (A, b) Rnn Rn is not controllable, then for any c Rn the polynomials
P1 (s) = ct adj(sI n A)b,

P2 (s) = PA (s)

are not coprime.

Proof By Theorem 2.39 we may suppose that A and b are given by

A11 A12
b1
A=
, b=
,
0n`,` A22
0n`
for some ` < n. Therefore,

sI ` A11
A12
sI n A =
0n`,` sI n` A22

(sI n A)1 =

(sI ` A11 )1

,
0n`,`
(sI n` A1
22

where the denotes a term that will not matter to us. Thus we have

(sI ` A11 )1 b1
(sI n A)1 b =
0n` .
This means that if we write c = (c1 , c2 ) R` Rn` we must have
ct (sI n A)1 b = ct1 (sI ` A11 )1 b1 .
This shows that

ct adj(sI n A)b
ct adj(sI ` A11 )b1
= 1
.
det(sI n A)
det(sI ` A11 )

The denominator on the left is monic of degree n and the denominator on the right is monic
and degree `. This must mean that there is a monic polynomial P of degree n ` so that
ct adj(sI n A)b
P (s)ct1 adj(sI ` A11 )b1
=
,
det(sI n A)
P (s) det(sI ` A11 )
which means that the polynomials ct adj(sI n A)b and det(sI n A) are not coprime.
This result shows that when (A, b) is not controllable, the order of the denominator in T ,
after performing pole/zero cancellations, will be strictly less than the state dimension. Thus
the transfer function for an uncontrollable system, is never representing the complete state
information.
Lets see how this works out in our uncontrollable example.

22/10/2004

3.3 Properties of the transfer function for SISO linear systems

3.8 Example (Example 2.19 contd) We consider a slight modification of the system in Example 2.19, and consider the system

1
0
A=
, b=
.
1 1
1

3 Transfer functions (the s-domain)

22/10/2004

and from this it follows that

bt adj(sI n At )c = ct adj(sI n A)b.
Now, since (A, c) is observable, (At , c) is controllable (cf. the proof of Theorem 2.38).
Therefore, by Theorem 3.5, the polynomials

The controllability matrix is given by

P1 (s) = bt adj(sI n At )c,

0
C(A, b) =
,
1 1

are coprime if and only if (At , b) is observable. Thus the polynomials P1 and P2 are coprime
if and only if (A, b) is controllable. However, by (3.5) P1 = P1 and P2 = P2 , and the result
now follows.

which has full rank except when = 0. We compute

s+1

adj(sI 2 A) =
=
adj(sI 2 A)b =
.
1
s1
s1

Let us illustrate this result with an example.

3.10 Example (Example 2.19 contd) We shall revise slightly Example 2.19 by taking

1
0
0
A=
, b=
, c=
.
1 1
1
1

Therefore, for c = (c1 , c2 ) we have

ct adj(sI 2 A)b = c2 (s 1) + c1 .
We also have det(sI 2 A) = s2 1 = (s + 1)(s 1) which means that there will always
be a pole/zero cancellation in T precisely when = 0. This is precisely when (A, b) is not

controllable, just as Theorem 3.7 predicts.
3.3.2 Observability and the transfer function The above relationship between observability and pole/zero cancellations in the numerator and denominator of T relies on
(A, b) being controllable. There is a similar story when (A, c) is observable, and this is told
by the following theorem.
3.9 Theorem Let = (A, b, ct , 01 ) be a SISO linear system. If (A, c) is observable, then
the polynomials
P1 (s) = ct adj(sI n A)b, P2 (s) = PA (s)
are coprime as elements of R[s] if and only if (A, b) is controllable.
Proof First we claim that
bt adj(sI n At )c = ct adj(sI n A)b
det(sI n At ) = det(sI n A).

(3.5)

Indeed, since the transpose of a 1 1 matrix, i.e., a scalar, is simply the matrix itself, and
since matrix inversion and transposition commute, we have
t

P2 (s) = PAt (s)

We determine that
ct adj(sI 2 A)b = s 1
det(sI 2 A) = s2 1.
The observability matrix is computed as
O(A, c) =

0 1
,
1 1

so the system is observable for all . On the other hand, the controllability matrix is

0
C(A, b) =
,
1 1
so the (A, b) is controllable
if and only if = 0. Whats more, the roots of the characteristic

polynomial are s = 1 + . Therefore, the polynomials ct adj(sI 2 A)b and det(sI 2 A)

are coprime if and only if = 0, just as predicted by Theorem 3.9.

We also have the following analogue with Theorem 3.7.
3.11 Theorem If (A, c) Rnn Rn is not observable, then for any b Rn the polynomials

t 1

c (sI n A) b = b (sI n sA ) c.
This implies, therefore, that

P1 (s) = ct adj(sI n A)b,

P2 (s) = PA (s)

are not coprime.

ct adj(sI n A)b
bt adj(sI n At )c
=
.
det(sI n A)
det(sI n At )
Since the eigenvalues of A and At agree,
det(sI n At ) = det(sI n A),

Proof This follows immediately from Theorem 3.7, (3.5) and the fact that (A, c) is observable if and only if (At , b) is controllable.

It is, of course, possible to illustrate this in an example, so let us do so.

22/10/2004

3.3 Properties of the transfer function for SISO linear systems

3.12 Example (Example 2.13 contd) Here we work with a slight modification of Example 2.13 by taking

0 1
1
A=
, c=
.
1
1
As the observability matrix is

3 Transfer functions (the s-domain)

Conversely, suppose that ct adj(sI n A)b = 0. By Exercise EE.4 this means that ct eAt b =
0 for t 0. If we Taylor expand eAt about t = 0 we get
c

X
tk
k=0

O(A, c) =

1
1
,
1 1 +

the system is observable if and only if = 0. If b = (b1 , b2 ) then we compute

ct adj(sI 2 A)b = (b1 b2 )(s 1) + b1 .
We also have det(sI 2 A) = s2 +s1. Thus we see that indeed the polynomials ct adj(sI 2
A)b and det(sI 2 A) are not coprime for every b exactly when = 0, i.e., exactly when
the system is not observable.

The following corollary summarises the strongest statement one may make concerning
the relationship between controllability and observability and pole/zero cancellations in the
transfer functions.
3.13 Corollary Let = (A, b, ct , 01 ) be a SISO linear system, and define the polynomials
P1 (s) = ct adj(sI n A)b,

P2 (s) = det(sI n A).

The following statements are equivalent:

(i) (A, b) is controllable and (A, c) is observable;
(ii) the polynomials P1 and P2 are coprime.
Note that if you are only handed a numerator polynomial and a denominator polynomial that are not coprime, you can only conclude that the system is not both controllable
and observable. From the polynomials alone, one cannot conclude that the system is, say,
controllable but not observable (see Exercise E3.9).
3.3.3 Zero dynamics and the transfer function It turns out that there is another
interesting interpretation of the transfer function as it relates to the zero dynamics. The following result is of general interest, and is also an essential part of the proof of Theorem 3.15.
We have already seen that det(sI n A) is never the zero polynomial. This result tells us
exactly when ct adj(sI n A)b is the zero polynomial.
3.14 Lemma Let = (A, b, ct , D) be a SISO linear system, and let Z be the subspace
constructed in Algorithm 2.28. Then ct adj(sI n A)b is the zero polynomial if and only if
b Z .
Proof Suppose that b Z . By Theorem 2.31 this means that Z is A-invariant, and so
also is (sI n A)-invariant. Furthermore, from the expression (A.3) we may ascertain that
Z is (sI n A)1 -invariant, or equivalently, that Z is adj(sI n A)-invariant. Thus we
must have adj(sI n A)b Z . Since Z ker(ct ), we must have ct adj(sI n A)b = 0.

22/10/2004

Ak b = 0.

Evaluating the kth derivative of this expression with respect to t at t = 0 gives cAk b = 0,
k = 0, 1, 2, . . . . Given these relations, we claim that the subspaces Zk , k = 1, 2, . . . of
Algorithm 2.28 are given by Zk = ker(ct Ak1 ). Since Z1 = ker(ct ), the claim holds for
k = 1. Now suppose that the claim holds for k = m > 1. Thus we have Zm = ker(ct Am1 ).
By Algorithm 2.28 we have
Zm+1 = { x Rn | Ax Zm + span {b}}

= x Rn | Ax ker(ct Am1 ) + span {b}

= x Rn | Ax ker(ct Am1 )

= x Rn | x ker(ct Am )
= ker(ct Am ),
where, on the third line, we have used the fact that b ker(ct Am1 ). Since our claim
follows, and since b ker(ct Ak ) = Zk1 for k = 0, 1, . . . , it follows that b Z .

The lemma, note, gives us conditions on so-called invertibility of the transfer function. In
this case we have invertibility if and only if the transfer function is non-zero.
With this, we may now prove the following.
3.15 Theorem Consider a SISO control system of the form = (A, b, ct , 01 ). If ct (sI n A)b
is not the zero polynomial then the zeros of ct adj(sI n A)b are exactly the spectrum for
the zero dynamics of .
Proof Since ct adj(sI n A)b 6= 0, by Lemma 3.14 we have b 6 Z . We can therefore choose
a basis B = {v 1 , . . . , v n } for Rn with the property that {v 1 , . . . , v ` } is a basis for Z and
v `+1 = b. With respect to this basis we can write c = (0, c2 ) R` Rn` since Z ker(ct ).
We can also write b = (0, (1, 0, . . . , 0)) R` Rn` , and we denote b2 = (1, 0, . . . , 0) Rn` .
We write the matrix for the linear map A in this basis as

A11 A12
.
A21 A22
Since A(Z ) = Z + span {b}, for k = 1, . . . ` we must have Av k = uk + k v `+1 for some
uk Z and for some k R. This means that A21 must have the form

1 2 `
0 0 0

A21 = ..
(3.6)
.. . .
. .
.
. ..
.
0 0 0
Therefore f 1 = (1 , . . . , ` ) is the unique vector for which b2 f t1 = A21 . We then define
f = (f 1 , 0) R` Rn` and determine the matrix for A + bf t in the basis B to be

A11 A12
.
0n`,` A22

22/10/2004

3.3 Properties of the transfer function for SISO linear systems

3 Transfer functions (the s-domain)

22/10/2004

Thus, by (A.3), A + bf t has Z as an invariant subspace. Furthermore, by Algorithm 2.28,

we know that the matrix N describing the zero dynamics is exactly A11 .
We now claim that for all s C the matrix

sI n` A22 b2
(3.7)
0
ct2

This theorem is very important as it allows us to inferat least in those cases where the
transfer function is invertiblethe nature of the zero dynamics from the transfer function.
If there are zeros, for example, with positive real part we know our system has unstable zero
dynamics, and we ought to be careful.
To further illustrate this connection between the transfer function and the zero dynamics,
we give an example.

is invertible. To show this, suppose that there exists a vector (x2 , u) Rn` R with the
property that

sI n` A22 b2 x2
(sI n` A22 )x2 + b2 u
0
=
=
.
(3.8)
ct2
0
u
ct2 x2
0

3.16 Example (Example 2.27 contd) Here we look again at Example 2.27. We have

0
1
0
1
A=
, b=
, c=
.
2 3
1
1

Define

We had computed Z = span {(1, 1)}, and so b 6 Z . Thus ct adj(sI 2 A)b is not the zero
polynomial by Lemma 3.14. Well, for pitys sake, we can just compute it:

Z = Z + span {(0, x2 )} .
Since Z ker(ct ) and since ct2 x2 = 0 we conclude that Z ker(ct ). Given the form of
A21 in (3.6), we see that if v Z , then Av Z + span {b}. This shows that Z Z , and
from this we conclude that (0, x2 ) Z and so x2 must be zero. It then follows from (3.8)
that u = 0, and this shows that the kernel of the matrix (3.7) contains only the zero vector,
and so the matrix must be invertible.
Next we note that

sI n A bf t b
sI n A b
In 0
=
t
t
t
c
0
c
0 f 1
and so

sI n A bf t b
sI n A b
det
= det
.
t
t
c
0
c
0

(3.9)

We now rearrange the matrix on the left-hand side corresponding to our decomposition. The
matrix for the linear map corresponding to this matrix in the basis B is

sI ` A11
A12
0
0n`,`
sI n` A22 b2 .
0
0t
ct2
The determinant of this matrix is therefore exactly the determinant on the left-hand side
of (3.9). This means that

sI ` A11
A12
0
sI n A b
sI n` A22 b2 .
det
= det 0n`,`
ct
0
0t
ct2
0
t

By Lemma 3.3 we see that the left-hand determinant is exactly c adj(sI n A)b. Therefore,
the values of s for which the left-hand side is zero are exactly the roots of the numerator of
the transfer function. On the other hand, since the matrix (3.7) is invertible for all s C,
the values of s for which the right-hand side vanish must be those values of s for which
det(sI ` A11 ) = 0, i.e., the eigenvalues of A11 . But we have already decided that A11 is
the matrix that represents the zero dynamics, so this completes the proof.

ct adj(sI 2 A)b = 1 s.
Since this is non-zero, we can apply Theorem 3.15 and conclude that the spectrum for the
zero dynamics is {1}. This agrees with our computation in Example 2.29 where we computed

N = 1 .
Since spec(N ) C+ 6= , the system is not minimum phase.

3.17 Remark We close with an important remark. This section contains some technically demanding mathematics. If you can understand this, then that is really great, and I encourage
you to try to do this. However, it is more important that you get the punchline here which
is:
The transfer function contains a great deal of information about the behaviour of
the system, and it does so in a deceptively simple manner.
We will be seeing further implications of this as things go along.

3.4 Transfer functions presented in input/output form

The discussion of the previous section supposes that we are given a state-space model
for our system. However, this is sometimes not the case. Sometimes, all we are given is a
scalar differential equation that describes how a scalar output behaves when given a scalar
input. We suppose that we are handed an equation of the form
y (n) (t) + pn1 y (n1) (t) + + p1 y (1) (t) + p0 y(t) =
cn1 u(n1) (t) + cn1 u(n2) (t) + + c1 u(1) (t) + c0 u(t) (3.10)
for real constants p0 , . . . , pn1 and c0 , . . . , cn1 . How might such a model be arrived at? Well,
one might perform measurements on the system given certain inputs, and figure out that a
differential equation of the above form is one that seems to accurately model what you are
seeing. This is not a topic for this book, and is referred to as model identification. For
now, we will just suppose that we are given a system of the form (3.10). Note here that
there are no states in this model! All there is is the input u(t) and the output y(t). Our

3.4 Transfer functions presented in input/output form

22/10/2004

3 Transfer functions (the s-domain)

22/10/2004

system may possess states, but the model of the form (3.10) does not know about them.
As we have already seen in the discussion following Theorem 2.37, there is a relationship
between the systems we discuss in this section, and SISO linear systems. We shall further
develop this relationship in this section.
For the moment, let us alleviate the nuisance of having to ever again write the expression (3.10). Given the differential equation (3.10) we define two polynomials in R[] by

If we take the causal left Laplace transform of the differential equation (3.10) we get
+
+
simply D(s)L0
(y)(s) = N (s)L0
(u)(s), provided that we suppose that both the input u
and the output y are causal signals. Therefore we have

D() = n + pn1 n1 + + p1 + p0
N () = cn1 n1 + cn2 n2 + + c1 + c0 .

Block diagrammatically, the situation is illustrated in Figure 3.7. We should be very clear

Note that if we let =

we can write

d
dt

TN,D (s) =

+
L0
(y)(s)
cn1 sn1 + cn1 sn2 + + c1 s + c0
N (s)
=
.
=
+
D(s)
sn + pn1 sn1 + + p1 s + p0
L0 (u)(s)

u
(s)

d
d
then we think of D( dt
) and N ( dt
) as a differential operator, and

d
D( dt
)(y) = y (n) (t) + pn1 y (n1) (t) + + p1 y (1) (t) + p0 y(t).

N (s)
D(s)

y(s)

Figure 3.7 The block diagram representation of (3.10)

d
In like manner we can think of N ( dt
) as a differential operator, and so we write
d
N ( dt
)(u) = cn1 u(n1) (t) + cn1 u(n2) (t) + + c1 u(1) (t) + c0 u(t).
d
d
In this notation the differential equation (3.10) reads D( dt
)(y) = N ( dt
)(u). With this little
bit of notation in mind, we make some definitions.

3.18 Definition A SISO linear system in input/output form is a pair of polynomials

(N, D) in R[s] with the properties
(i) D is monic and
(ii) D and N are coprime.
The relative degree of (N, D) is deg(D) deg(N ). The system is proper (resp. strictly
proper ) if its relative degree is nonnegative (resp. positive). If (N, D) is not proper, it is
improper . A SISO linear system (N, D) in input/output form is stable if D has no roots
in C+ and minimum phase if N has no roots in C+ . If (N, D) is not minimum phase, it
is nonminimum phase. The transfer function associated with the SISO linear system
(s)
.

(N, D) in input/output form is the rational function TN,D (s) = N
D(s)
3.19 Remarks 1. Note that in the definition we allow for the numerator to have degree
greater than that of the denominator, even though this is not the case when the input/output system is derived from a differential equation (3.10). Our reason for doing
this is that occasionally one does encounter transfer functions that are improper, or
maybe situations where a transfer function, even though proper itself, is a product of
rational functions, at least one of which is not proper. This will happen, for example, in
Section 6.5 with the derivative part of PID control. Nevertheless, we shall for the most
part be thinking of proper, or even strictly proper SISO linear systems in input/output
form.
2. At this point, it is not quite clear what is the motivation behind calling a system (N, D)
stable or minimum phase. However, this will certainly be clear as we discuss properties
of transfer functions. This being said, a realisation of just what stable might mean
will not be made fully until Chapter 5.

on why the diagrams Figure 3.6 and Figure 3.7 are different: there are no state variables x
in the differential equations (3.10). All we have is an input/output relation. This raises the
question of whether there is a connection between the equations in the form (3.1) and those
in the form (3.10). Following the proof of Theorem 2.37 we illustrated how one can take
differential equations of the form (3.1) and produce an input/output differential equation
like (3.10), provided (A, b) is controllable. To go from differential equations of the form (3.10)
and produce differential equations of the form (3.1) is in some sense artificial, as we would
have to invent states that are not present in (3.10). Indeed, there are infinitely many
ways to introduce states into a given input/output relation. We shall look at the one that is
related to Theorem 2.37. It turns out that the best way to think of making the connection
from (3.10) to (3.1) is to use transfer functions.
3.20 Theorem Let (N, D) be a proper SISO linear system in input/output form. There exists
a complete SISO linear control system = (A, b, ct , D) with A Rnn so that T = TN,D .
Proof Let us write
D(s) = sn + pn1 sn1 + + p1 s + p0
N (s) = cn sn + cn1 sn1 + cn2 sn2 + + c1 s + c0 .
We may write

N (s)
D(s)

N (s)
cn sn + cn1 sn1 + cn2 sn2 + + c1 s + c0
=
D(s)
sn + pn1 sn1 + + p1 s + p0
sn + pn1 sn1 + + p1 s + p0
= cn n
+
s + pn1 sn1 + + p1 s + p0
n1
(
cn1 cn pn1 )s
+ (
cn2 cn pn2 )sn2 + + (
c1 cn p1 )s + (
c0 cn p0 )
sn + pn1 sn1 + + p1 s + p0
cn1 sn1 + cn2 sn2 + + c1 s + c0
= cn +
,
sn + pn1 sn1 + + p1 s + p0

22/10/2004

3.4 Transfer functions presented in input/output form

where ci = ci cn pi , i = 0, . . . , n 1. Now define

0
1
0
0
0
0
0
0
0
1
0
0

0
0
0
0
1
0

A = ..
..
..
..
.. , b = .. ,
..
.

.
.
.
.
.
.

0
0
0
0
0
1
p0 p1 p2 p3 pn1
1

3 Transfer functions (the s-domain)

22/10/2004

cn1

c0
c1
c2
..
.

c=
,

cn2
cn1

cn2

D = cn .
c1

(3.11)
By Exercise E2.11 we know that (A, b) is controllable. Since D and N are by definition
coprime, by Theorem 3.5 (A, c) is observable. In the proof of Theorem 3.5 we showed that
cn1 sn1 + cn2 sn2 + + c1 s + c0
ct adj(sI n A)b
,
=
det(sI n A)
sn + pn1 sn1 + + p1 s + p0
(see equation (3.4)), and from this follows our result.

1
s

x
n

1
s

x
n1

...

x
2

1
s

x
1

pn1

We shall denote the SISO linear control system of the theorem, i.e., that one given
by (3.11), by N,D to make explicit that it comes from a SISO linear system in input/output
form. We call N,D the canonical minimal realisation of the transfer function TN,D .
Note that condition (ii) of Definition 3.18 and Theorem 3.5 ensure that (A, c) is observable.
This establishes a way of getting a linear system from one in input/output form. However,
it not the case that the linear system N,D should be thought of as representing the physical
states of the system, but rather it represents only the input/output relation. There are
consequences of this that you need to be aware of (see, for example, Exercise E3.20).
It is possible to represent the above relation with a block diagram with each of the
states x1 , . . . , xn appearing as a signal. Indeed, you should verify that the block diagram of
Figure 3.8 provides a transfer function which is exactly
+
L0
(y)s)
cn1 sn1 + + c1 s + c0
+ D.
= n
+
s + pn1 sn1 + + p1 s + p0
L0 (u)(s)

Note that this also provides us with a way of constructing a block diagram corresponding
to a transfer function, even though the transfer function may have been obtained from a
different block diagram. The block diagram of Figure 3.8 is particularly useful if you live in
mediaeval times, and have access to an analogue computer . . .
3.21 Remark Note that we have provided a system in controller canonical form corresponding
to a system in input/output form. Of course, it is also possible to give a system in observer
canonical form. This is left as Exercise E3.19 for the reader.

Theorem 3.20 allows us to borrow some concepts that we have developed for linear systems of the type (3.1), but which are not obviously applicable to systems in the form (3.10).
This can be a useful thing to do. For example, motivated by Theorem 3.15, our notion that
a SISO linear system (N, D) in input/output form is minimum phase if all roots of N lie in
C+ , and nonminimum phase otherwise, makes some sense.
Also, we can also use the correspondence of Theorem 3.20 to make a sensible notion
of impulse response for SISO systems in input/output form. The problem with a direct
definition is that if we take u(t) to be a limit of inputs from U as described in Theorem 2.34,
it is not clear what we should take for u(k) (t) for k 1. However, from the transfer function

pn2

p0
D

Figure 3.8 A block diagram for the SISO linear system of Theorem 3.20

point of view, this is not a problem. To wit, if (N, D) is a strictly proper SISO linear system
in input/output form, its impulse response is given by
hN,D (t) = 1(t)ct eAt b
where (A, b, c, 01 ) = N,D . As with SISO linear systems, we may define the causal impulse

response h+
N,D : [0, ) R and the anticausal impulse response hN,D : (, 0] R. Also,
as with SISO linear systems, it is the causal impulse response we will most often use, so we
will frequently just write hN,D , as we have already done, for h+
N,D .
We note that it is not a simple matter to define the impulse response for a SISO linear
system in input/output form that is proper but not strictly proper. The reason for this is
that the impulse response is not realisable by a piecewise continuous input u(t). However, if
one is willing to accept the notion of a delta-function, then one may form a suitable notion
of impulse response. How this may be done without explicit recourse to delta-functions is
outlined in Exercise E3.1.

22/10/2004

3.5 The connection between the transfer function and the impulse response

3.5 The connection between the transfer function and the impulse
response
We can make some statements about how the transfer function impinges upon the impulse
response. We made some off the cuff remarks about how the impulse response contained
the essential time-domain character of the system = (A, b, ct , 01 ). We justified this in
some way by stating Proposition 2.32. We will now see that the information contained in
the impulse response is also directly contained in the transfer function, and further in about
the simplest way possible.
3.5.1 Properties of the causal impulse response We begin by looking at the case
that is the most interesting for us; the causal impulse response. Our discussions in this
section will be important for the development of certain fundamental aspects of, for example,
input/output stability.
We begin by making the essential connection between the impulse and the transfer function. We state the result for both the left and right causal Laplace transform of the impulse
response.
3.22 Theorem For a SISO linear system = (A, b, ct , 01 ), we have
+
+
L0+
(h )(s) = L0
(h )(s) = T (s)
1
provided that Re(s) > min (h+
).

Proof Since h+
has no delta-function at t = 0 as we are supposing D = 01 , we have
Z
+
+
st
L0+
(h )(s) = L0
h+
(h )(s) =
dt.
(t)e

3 Transfer functions (the s-domain)

22/10/2004

Now we note that

d
(sI n + A)1 e(sI n +A)t = e(sI n +A)t ,
dt
from which we ascertain that

+
(h )(s) = ct (sI n + A)1 e(sI n +A)t b .
L0
0

(sI n +A)t

Since the terms in the matrix e

the integral is zero so we have

satisfy an inequality like (E.2), the upper limit on

+
L0
(h )(s) = ct (sI n + A)1 b = ct (sI n A)1 b = T (s)

as claimed.

3.23 Remark Note that we ask that the feedforward term D be zero in the theorem. This
can be relaxed provided that one is willing to think about the delta-function. The manner
in which this can be done is the point of Exercise E3.1. When one does this, the theorem
no longer holds for both the left and right causal Laplace transforms, but only holds for the
left causal Laplace transform.

We should, then, be able to glean the behaviour of the impulse response by looking only
at the transfer function. This is indeed the case, and this will now be elucidated. You will
recall that if f (t) is a positive real-valued function then

lim sup f (t) = lim sup f ( ) .
t >t

The idea is that lim supt will exist for bounded functions that oscillate whereas limt
may not exist for such functions. With this notation we have the following result.

Thus clearly it does not matter in this case whether we use the left or right Laplace transform.
We have
Z
Z
Z
+
st
t At
st
L0
(h )(s) =
h+
(t)e
dt
=
c
e
be
dt
=
ct est eAt b dt.

Recall that the matrix exponential has the property that

ea x = eaI n x

where (N, D) is the c.f.r. of T . The following statements hold.

(i) If D has a root in C+ then

lim h+
(t) = .
t

(ii) If all roots of D are in C then

for a R and x Rn . Therefore we have

+
(h )(s)
L0

3.24 Proposition Let = (A, b, ct , 01 ) be a SISO linear system with impulse response h+

and transfer function T . Write

N (s)
T (s) =
D(s)

Z
=

t stI n At

e b dt.

Now recall that if B, C Rnn have the property that BC = CB then eB+C = eB eC .
Noting that the n n matrices stI n and At commute, we then have
Z
+
L0
(h )(s) =
ct e(sI n +A)t b dt.
0
1
Here is an occasion where we actually think of T as a function of the complex variable s rather than
as a rational function.

lim h+
(t) = 0.

(iii) If D has no roots in C+ , but has distinct roots on iR, then

lim sup h+
(t) = M
t

for some M > 0.

(iv) If D has repeated roots on iR then

lim h+
(t) = .

22/10/2004

3.5 The connection between the transfer function and the impulse response

Proof (i) Corresponding to a root with positive real part will be a term in the partial
fraction expansion of T of the form

(s )k

with > 0. By Proposition E.11, associated with such a term will be a term in the impulse
response that is a linear combination of functions of the form
t` et

3.25 Theorem For a SISO linear system = (A, b, ct , 01 ) we have

(h
L0+
)(s) = L0 (h )(s) = T (s),

or t` et cos t or t` et sin t.

provided that Re(s) < max (h

Such terms will clearly blow up to infinity as t increases.

(ii) If all roots lie in the negative half-plane then all terms in the partial fraction expansion
for T will have the form

1 s + 0
or
k
(s + )k
(s + )2 + 2
for > 0. Again by Proposition E.11, associated with these terms are terms in the impulse
response that are linear combinations of functions of the form
t` et

or t` et cos t or t` et sin t.

All such functions decay to zero at t increases.

(iii) The roots in the negative half-plane will give terms in the impulse response that
decay to zero, as we saw in the proof for part (ii). For a distinct complex conjugate pair of
roots i on the imaginary axis, the corresponding terms in the partial fraction expansion
for T will be of the form
1 s + 0
s2 +
which, by Proposition E.11, lead to terms in the impulse response that are linear combinations of cos t and sin t. This will give a bounded oscillatory time response as t , and
so the resulting lim sup will be positive.
(iv) If there are repeated imaginary roots, these will lead to terms in the partial fraction
expansion for T of the form
1 s + 0
.
(s2 + 2 )k
The corresponding terms in the impulse response will be linear combinations of functions
like t` cos t or t` sin t. Such functions clearly are unbounded at t .

This result is important because it tells us that we can understand a great deal about
the stability of a system by examining the transfer function. Indeed, this is the whole
idea behind classical control: one tries to make the transfer function look a certain way. In
this section we have tried to ensure that we fully understand the relationship between the
time-domain and the transfer function, as this relationship is a complete one most of the
time (youll notice that we have made controllability and/or observability assumptions quite
often, but many systems you encounter will be both controllable and observable).

22/10/2004

3.5.2 Things anticausal We will have occasion to make a few computations using
the anticausal impulse response, and the causal impulse response in relation to anticausal
inputs. In this section we collect the relevant results.
The first result is a natural analogue of Theorem 3.22. Since the proof is exactly the
same with the exception of some signs, it is omitted.

1 s + 0
(s )2 + 2

3 Transfer functions (the s-domain)

Of course, we also have an analogue of Proposition 3.24 for the anticausal impulse response, where the statement is altered by replacing C+ with C , and vice versa, and by
replacing limt with limt . However, we shall not make use of this result, so do not
state it. Instead, let us state a few results about causal inputs and the anticausal impulse
response, and vice versa.
For NP suff proof

3.26 Proposition Let = (A, b, ct , 01 ) be a SISO linear system.

(i)
Proof

3.6 The matter of computing outputs

Given a SISO linear system = (A, b, ct , D), Theorem 2.6 gives us a manner of determining the state, and therefore the output, using the matrix exponential. While this is
certainly a valid way to compute the output, it is really only useful in simple cases, where
the matrix exponential formalism provides a way of determining a nice symbolic expression
for the output. Similarly, using the right causal Laplace transform, it is a simple matter to
compute outputs in cases where the Laplace transform can be inverted. However, this is
often not the case. Furthermore, the matrix exponential and Laplace transform techniques
are not readily implemented numerically. Therefore, in this section we look at ways of writing scalar differential equations for determining output that can be readily implemented
numerically.
In the course of this discussion we shall be forced to come to grips with the difference
between the left and right causal Laplace transform. To be in any way accurate, this must be
done. That this is normally glossed over is a reflection of the willingness to oversimplify the
use of the Laplace transform. What we shall essentially see is a dichotomy of the following
type for the use of the two forms of causal transform.
1. For solving differential equations whose right-hand side involves no delta-function, the
right causal Laplace transform is the most convenient tool. Note that this precludes
the convenient use of the right causal Laplace transform with the impulse response
since the impulse response involves a delta-function on the right-hand side.
2. For general control theoretic discussions, the left causal Laplace transform is more convenient since it eliminates the appearance of the pesky initial condition terms present
in the expressions for derivatives. Indeed, in the places in the discussion above where
the Laplace transform was used, it was the left causal transform that was used.

3.6 The matter of computing outputs

22/10/2004

Thus we see that there is a natural tension between whether to use the left or right causal
Laplace transform. It is important to realise that the two things are different, and that their
differences sometimes make one preferable over the other.
3.6.1 Computing outputs for SISO linear systems in input/output form using the
right causal Laplace transform We begin looking at systems that are given to us in
input/output form. Thus we have a SISO linear system (N, D) in input/output form with
deg(D) = n, and we are concerned with obtaining solutions to the initial value problem

d
d
y(t) = N dt
u(t), y(0+) = y0 , y (1) (0+) = y1 , . . . , y (n1) (0+) = yn1 ,
(3.12)
D dt
for a known function u(t). We shall assume that u is sufficiently differentiable that the
needed derivatives exist at t = 0+, and that u possesses a right causal Laplace transform.
Then, roughly speaking, if one wishes to use Laplace transforms one determines the right
+
causal Laplace transform L0+
(u), and then inverse Laplace transforms the function
+
L0+
(y)(s) =

100

3 Transfer functions (the s-domain)

22/10/2004

3.28 Remark The assumptions of differentiability on the input can be relaxed as one can
see by looking at the proof that one really only needs for u to satisfy the conditions of
the Laplace transform derivative theorem, Theorem E.7, for a sufficiently large number of
derivatives. This can be made true for a large class of functions by using the definition of
the Laplace transform on distributions. This same observation holds for results that follow
and possess the same hypotheses.

One may, if it is possible, use Proposition 3.27 to obtain the solution y(t) by using the
inverse Laplace transform. Lets see how this works in an example.
3.29 Example We have (N (s), D(s)) = (1 s, s2 + 2s + 2) and we take as input u(t) = 1(t)t.
The initial value problem we solve is
y(t) + 2y(t)
+ 2y(t) = t 1,

y(0+) = 1, y(0+)

= 0.

Taking the Laplace transform of the left-hand side of this equation gives

N (s) +
L (u)(s)
D(s) 0+

+
+
(y)(s) (s + 2)y(0+) y(0+)

= (s2 + 2s + 2)L0+
(y)(s) s 2.
(s2 + 2s + 2)L0+

to get y(t). This is indeed very rough, of course, because we have lost track of the initial
conditions in doing this. The following result tells us how to use the Laplace transform
method to obtain the solution to (3.12), properly keeping track on initial conditions.
3.27 Proposition Let (N, D) be a proper SISO linear system in input/output form,
let u : [0, ) R possess a right causal Laplace transform, and assume that
u, u(1) , . . . , u(deg(N )1) are continuous on [0, ) and that u(deg(N )) is piecewise continuous
on [0, ). Then the right causal Laplace transform of the solution to the initial value problem (3.12) is
k1
n X
X

1
+
+
pk sj y (kj1) (0+) ck sj u(kj1) (0+) ,
L0+
(y)(s) =
N (s)L0+
(u)(s) +
D(s)
k=1 j=0

One may verify that (s + 2)y(0+) + y(0+)

is exactly the expression

k1
n X
X

pk sj y (kj1) (0+)

k=1 j=0

in the statement of Proposition 3.27 with D as given. The Laplace transform of the righthand side of the differential equation is
+
(1 s)L0+
(u)(s) + u(0+) =

1s
,
s2

using the fact that the Laplace transform of u(t) is s12 . As with the expression involving the
left-hand side of the equation, we note that u(0+) is exactly the expression

where
N (s) = cn sn + cn1 sn1 + cn2 sn2 + + c1 s + c0
D(s) = sn + pn1 sn1 + + p1 s + p0 ,

and where we take pn = 1.

k1
X

sj y (kj1) (0+),

ck sj u(kj1) (0+)

k=1 j=0

Proof Note that by Corollary E.8 the right causal Laplace transform of y (k) (t) is given by
+
(y)(s)
sk L0+

n X
k1
X

in the statement of Proposition 3.27. Combining our work, the Laplace transform of the
differential equation gives
+
(y)(s) =
L0+

(3.13)

1s
s+2
+
.
s2 (s2 + 2s + 2) s2 + 2s + 2

j=0

with a similar formula holding, of course, for u(k) (t). Thus, if we take the Laplace transform
of the differential equation in (3.12) and use the initial conditions given there we get the
relation
n
X
k=0

k1
n
k1

X

X
X
+
+
pk sk L0+
(y)(s)
sj y (kj1) (0+) =
ck sk L0+
(u)(s)
sj u(kj1) (0+) ,
j=0

k=0

Thus we obtain
+
L0+
(y)(s) =

j=0

where we take pn = 1. The result now follows by a simple manipulation.

To compute the inverse Laplace transform we perform the partial fraction expansion for the
first term to get
s + 32
1s
1
1
= 2 + 2
.
2
2
s (s + 2s + 2)
2s
s s + 2s + 2

2s + 72
1
1
+ 2
2
2s
s s + 2s + 2

22/10/2004

3.6 The matter of computing outputs

101

102

The inverse Laplace transform of the first term is 21 t, and of the second term is 1. To
determine the Laplace transform of the third term we note that the inverse Laplace of

22/10/2004

3.30 Proposition Let (N, D) be a proper SISO linear system in input/output form with input
u as given by (3.14). The solution y of the initial value problem

d
d
D dt
y(t) = N dt
u(t)

1
s2 + 2s + 2

with initial conditions (3.15) has the form y(t) = y0 (t) + d(t) where d = and where y0 is
the solution of the initial value problem

(1)
(n1)
d
d
y0 (t) = N dt
u0 (t), y0 (0+) =, y0 (0+) =, . . . , y0
(0+) =,
D dt

is et sin t, and the inverse Laplace transform of

s+1
s2 + 2s + 2

where

is et cos t. Putting this all together we have obtained the solution

y(t) = et 23 sin t + 2 cos t + 2t 1
to our initial value problem.

3 Transfer functions (the s-domain)

N (s) = cn sn + cn1 sn1 + cn2 sn2 + + c1 s + c0

D(s) = sn + pn1 sn1 + + p1 s + p0 .

Proof Taking the left causal Laplace transform of the differential equation, using Corollary E.8, gives
n
X
+
+
(u0 )(s) +
cj csj .
(y)(s) = N (s)L0
D(s)L0

This example demonstrates how tedious can be the matter of obtaining by hand the
solution to even a fairly simple initial value problem.
It is sometimes preferable to obtain the solution in the time-domain, particularly for
systems of high-order since the inverse Laplace transform will typically be difficult to obtain
in such cases. One way to do this is proposed in Section 3.6.3.

j=0

Since the initial conditions at t = 0 for y are all zero, we may express y = y0 + y1 where
Finish

3.6.2 Computing outputs for SISO linear systems in input/output form using the
left causal Laplace transform As mentioned in the preamble to this section, the right
causal Laplace transform is the more useful than its left brother for solving initial value
problems, by virtue of its encoding the initial conditions in a convenient manner. However,
it is possible to solve these same problems using the left causal Laplace transform. To do so,
since the initial conditions at t = 0 are all zero (we always work with causal inputs and
outputs), it turns out that the correct way to encode the initial conditions at t = 0+ is to
add delta-function inputs. While this is not necessarily the recommended way to solve such
equations, it does make precise the connection between the left and right causal transforms
in this case. Also, it allows us to better understand such things as the impulse response.
Let us begin by considering a proper SISO linear system (N, D) in input/output form.
For this system, let us consider inputs of the form
u(t) = 1(t)u0 (t) + c(t)
(1)
(deg(N )1)
u0 , u0 , . . . , u0

(3.14)

where u0 : R R has the property that

are continuous on [0, )
(deg(N ))
and that u0
is piecewise continuous on [0, ). Thus we allow inputs which satisfy
the derivative rule for the right causal Laplace transform (cf. Theorem E.7), and which
additionally allows a delta-function as input. This allows us to consider the impulse response
as part of the collection of problems considered. Note that by allowing a delta-function in
our input, we essentially mandate the use of the left causal Laplace transform to solve the
equation. Since we are interested in causal outputs, we ask that the output have initial
conditions
y(0) = 0, y (1) (0), . . . , y (n1) (0) = 0,
(3.15)
where, as usual, n = deg(D).
Let us first state the solution of the problem just formulated.

+
+
D(s)L0
(y0 )(s) = N (s)L0
(u0 )(s)

and
+
D(s)L0
(y1 )(s) =

n
X

cj csj .

j=0

3.6.3 Computing outputs for SISO linear systems in input/output form using the
causal impulse response To see how the Laplace transform method is connected with
solving equations in the time-domain, we make the observation, following from Theorem 3.22,
that the Laplace transform of the impulse response hN,D is the transfer function TN,D in
the case where (N, D) is strictly proper. Therefore, by Exercise EE.5, the inverse Laplace
+
transform of TN,D (s)L0+
(u)(s) is
Z

hN,D (t )u( ) d.

(3.16)

Let us address the question of which initial value problem of the form (3.12) has the expression (3.16) as its solution. That is, let us determine the proper initial conditions to obtain
the solution (3.16).
3.31 Proposition Let (N, D) be a strictly proper SISO linear system in input/output
form, let u : [0, ) R possess a right causal Laplace transform, and assume that
u, u(1) , . . . , u(deg(N )1) are continuous on [0, ) and that u(deg(N )) is piecewise continuous on
[0, ). Then the integral (3.16) is the solution to the initial value problem (3.12) provided
that the initial values y(0+), y (1) (0+), . . . , y (n1) (0+) are chosen as follows:
(i) let y(0+) = 0;

3.6 The matter of computing outputs

22/10/2004

(ii) recursively define y (k) (0+) =

k
X

103

cnj u(kj) (0+) pnj y (kj) (0+) , k = 1, . . . , n 1.

j=1

Proof From Proposition 3.27 it suffices to show that the initial conditions we have defined
are such that
n X
k1
X

pk sj y (kj1) (0+) ck sj u(kj1) (0+) = 0,

k = 1, . . . , n 1.

k=1 j=0

This will follow if the coefficient of each power of s in the preceding expression vanishes.
Starting with the coefficient of sn1 we see that y(0+) = 0. The coefficient of sn2 is then
determined to be
y (1) (0+) + pn1 y(0+) cn1 u(0+),
which gives y (1) (0+) = cn1 u(0+) pn1 y(0+). Proceeding in this way, we develop the
recursion relation as stated in the proposition.

Lets see how we may use this to obtain a solution to an initial value problem in the
time-domain using the impulse response hN,D .
3.32 Proposition Let (N, D) be a strictly proper SISO linear system in input/output
form, let u : [0, ) R possess a right causal Laplace transform, and assume that
u, u(1) , . . . , u(deg(N )1) are continuous on [0, ) and that u(deg(N )) is piecewise continuous
on [0, ). Let y0 , y1 , . . . , yn1 be the initial conditions as defined in Proposition 3.31, and
suppose that yh (t) solves the initial value problem

d
D dt
yh (t) = 0, y(0+) = y0 y0 , , y (1) (0+) = y1 y1 , . . . , y (n1) (0+) = yn1 yn1 .
Then the solution to the initial value problem (3.12) is given by
Z t
y(t) = yh (t) +
hN,D (t )u( ) d.

(3.17 )

Lets see how this works in a simple example.

3 Transfer functions (the s-domain)

22/10/2004

3.33 Example (Example 3.29 contd) We take (N (s), D(s)) = (1 s, s2 + 2s + 2) so the

differential equation is
y(t) + 2y(t)
+ 2y(t) = u(t) u(t).

As input we take u(t) = t and as initial conditions we take y(0+) = 1 and y(0+)

= 0.
We first obtain the homogeneous solution yh (t). We first determine the initial conditions,
meaning we need to determine the initial conditions y0 and y1 from Proposition 3.31. These
we readily determine to be y0 = 0 and y1 = c1 u(0+) p1 y0 = 0. Thus yh should solve the
initial value problem
yh (t) + 2y h (t) + 2yh (t) = 0,

yh (0+) = 1, y h (0+) = 0.

Recall how to solve this equation. One first determines the roots of the characteristic polynomial that in this case is s2 + 2s + 2. The roots we compute to be 1 i. This gives rise to
two linearly independent solutions to the homogeneous differential equation, and these can
be taken to be
y1 (t) = et cos t, y2 (t) = et sin t.
Any solution to the homogeneous equation is a sum of these two solutions, so we must have
yh (t) = C1 y1 (t) + C2 y2 (t) for appropriate constants C1 and C2 . To determine these constants
we use the initial conditions:
yh (0+) = C1 = 1
y h (0+) = C1 + C2 = 0,
from which we ascertain that C1 = C2 = 1 so that yh (t) = et (cos t + sin t).
We now need the impulse response that we can compute however we want (but see
Example 3.41 for a slick way to do this) to be
hN,D (t) = et (2 sin t cos t).

Proof That every solution of (3.12) can be expressed in the form of (3.17) follows from
Proposition 2.32. It thus suffices to show that the given solution satisfies the initial conditions. However, by the definition of yh and by Proposition 3.31 we have
Z 0
hN,D ( )u( ) d = y0 y0 = y0
y(0+) = yh (0+) +
0
Z t
d
(1)
y (1) (0+) = yh (0+) +
hN,D (t )u( ) d = y1 y1 + y1 = y1
dt t=0 0
..
.
Z t
dn1
(1)
hN,D (t )u( ) d = yn1 yn1 + yn1 = yn1 ,
y (n1) (0+) = yh (0+) + n1
dt
t=0 0
which are the desired initial conditions.

104

Now we may determine

Z t
Z t

hN,D (t )u( ) d =
e(t ) (2 sin(t ) cos(t )) d = 2t 1 + et cos t + 12 sin t
0

Therefore, the solution to the initial value problem

y(t) + 2y(t)
+ 2y(t) = t 1,

y(0+) = 1, y(0+)

is
y(t) = et

3
2

sin t + 2 cos t + 2t 1,

agreeing with what we determined using Laplace transforms in Example 3.29.

In practice, one does notat least I do notsolve simple differential equations this
way. Rather, I typically use the method of undetermined coefficients where the idea is
that after solving the homogeneous problem, the part of the solution that depends on the
right-hand side is made a general function of the same form as the right-hand side, but
with undetermined coefficients. One then resolves the coefficients by substitution into the
differential equation. This method can be found in your garden variety text on ordinary
differential equations, for example [Boyce and Diprima 2000]. A too quick overview is given
in Section B.1.

22/10/2004

3.6 The matter of computing outputs

105

3.6.4 Computing outputs for SISO linear systems Next we undertake the above
constructions for SISO linear systems = (A, b, c, D). Of course, we may simply compute
the transfer function T and proceed using the tools we developed above for systems in
input/output form. However, we wish to see how the additional structure for state-space
systems comes into play. Thus in this section we consider the initial value problem

x(t)
= Ax(t) + bu(t), x(0+) = x0
y(t) = ct x(t) + Du(t).

3 Transfer functions (the s-domain)

22/10/2004

Proof Note that y1 (t) = ct eAt x0 . Taking Laplace transforms and using Exercise EE.4 gives
ct adj(sI n A)x0
.
det(sI n A)

+
L0+
(y1 )(s) = ct (sI n A)1 x0 =

Thus we have
+
(y)1 (s) = ct adj(sI n A)x0 ,
D(s)L0+

(3.18)

Let us first make a somewhat trivial observation.

3.34 Lemma Let y(t) be the output for the SISO linear system = (A, b, ct , D) subject to
the input u(t) and the initial condition x(0+) = x0 : thus y(t) is defined by (3.18). Then
there exists unique functions y1 , y2 , y3 : [0, ) R satisfying:
(i) y1 (t) satisfies

x(t)
= Ax(t),
y1 (t) = ct x(t);

106

x(0+) = x0

(ii) y2 (t) satisfies

and since the right-hand side is a polynomial of degree at most n 1 (if deg(D) = n),
by (3.13)
this means that there are some initial conditions for which y1 (t) is a solution to

d
D dt
y1 (t) = 0. It remains to compute the initial conditions. However, this is a simple
computation, giving exactly the conditions of the lemma.

Now we look at a scalar differential equation for y2 (t).
3.36 Lemma Let = (A, b, ct , 01 ) be a SISO linear system, let u : [0, ) R possess a
right causal Laplace transform, and assume that u, u(1) , . . . , u(deg(N )1) are continuous on
[0, ) and that u(deg(N )) is piecewise continuous on [0, ). If y2 (t) is the output determined
by condition (ii) of Lemma 3.34, then
Z t
y2 (t) =
h (t )u( ) d,
0

x(t)
= Ax(t) + bu(t),
y1 (t) = ct x(t);

and furthermore y2 (t) solves the initial value problem

x(0+) = 0

d
dt

y2 (t) = N

(iii) y3 (t) satisfies

d
dt

u(t),

(k)
y2 (0+)

y3 (t) = Du(t);

(1)

y2 (0+) = 0, y2 (0+) = ct bu(0+), . . . ,

X
X
(n1)
ct Ai bu(j) (0+), . . . , y2
(0+) =
ct Ai bu(j) (0+),

i,j
i+j=k1

i,j
i+j=n2

(iv) y(t) = y1 (t) + y2 (t) + y3 (t).

Proof First note that y1 (t), y2 (t), and y3 (t) are indeed uniquely defined by the conditions
of (i), (ii), and (iii), respectively. It thus remains to show that (iv) is satisfied. However,
this is a straightforward matter of checking that y1 (t) + y2 (t) + y3 (t) as defined by the

conditions (i), (ii), and (iii) satisfies (3.18).
The idea here is that to obtain the output for (3.18) we first look at the case where D = 01
and u(t) = 0, obtaining the solution y1 (t). To this we add the output y2 (t), defined again
with D = 01 , but this time with the input as the given input u(t). Note that to obtain y1
we use the given initial condition x0 , but to obtain y2 (t) we use the zero initial condition.
Finally, to these we add y3 (t) = Du(t) to get the actual output.
Our objective in this section is to provide scalar initial value problems whose solutions are
y1 (t), y2 (t), and y3 (t). These are easily implemented numerically. We begin by determining
the scalar initial value problem for y1 (t).
3.35 Lemma Let = (A, b, ct , 01 ) be a SISO linear system and let y1 (t) be the output
determined by condition (i) of Lemma 3.34. Then y1 (t) solves the initial value problem
D

d
dt

y1 (t) = 0,

y1 (0+) = ct x0 ,

where (N, D) is the c.f.r. for T .

(1)
y1 (0+)

= ct Ax0 , . . . ,

(n1)
y1
(0+)

= ct An1 x0 ,

where (N, D) is the c.f.r. for T .

Rt
Proof That y2 (t) = 0 h (t )u( ) d is a simple consequence of the definition of y2 (t)
and Proposition 2.32. Whats more, taking the Laplace transform of y2 (t) we get
+
+
L0+
(y2 )(s) = T (s)L0+
(u)(s)

+
+
D(s)L0+
(y2 )(s) = N (s)L0+
(u)(s),

using Theorem 3.22, Exercise EE.5, and the fact that (N, D) is the c.f.r. of T . This
means
d
d
y2 (t) = N dt
u(t). To
that y2 (t) is indeed a solution of the differential equation D dt
determine the initial conditions for y2 (t), we simply differentiate the formula we have. Thus
we immediately get y2 (0+) = 0. For the first derivative we have
Z
d t
(1)
y2 (t) =
h (t )u( ) d
dt 0
Z t
(1)
= h (t)u(t) +
h (t )u( ) d.
0
(1)

Thus y (0+) = h (0+)u(0+). We may proceed, using mathematical induction if one wishes
to do it properly, to derive
Z t
X (i)
(k)
(k)
y2 (t) =
h (0+)u(j) (t) +
h (t )u( ) d.
i,j
i+j=k1

3.6 The matter of computing outputs

22/10/2004

107

Now we observe that a simple computation given h (t) = ct eAt b demonstrates that

3 Transfer functions (the s-domain)

(k)

ct x0 = 1,

ct Ax0 + ct bu(0+) = 0,

The above two lemmas give us the hard part of the output for (3.18)all that remains
is to add y3 (t) = Du(t). Let us therefore summarise how to determine the output for the
system (3.18) from a scalar differential equation.

and we compute the transfer function to be

3.37 Proposition Let = (A, b, ct , D) be a SISO linear system, let u : [0, ) R possess
a right causal Laplace transform, and assume that u, u(1) , . . . , u(deg(N )1) are continuous on
[0, ) and that u(deg(N )) is piecewise continuous on [0, ). If y(t) is the output defined
by (3.18) for the input u(t) and initial state x0 , then y(t) is the solution of the differential
equation

d
d
y(t) = N dt
u(t),
D dt

Therefore, the initial value problem given to us by Proposition 3.37 is

subject to the initial conditions

y(0+) = ct x0 + Du(0+),

..
.
ct Ai bu(j) (0+) + Du(n1) (0+).

i,j
i+j=n2
d
dt

modulo initial conditions. The initial conditions for y are derived as in Lemma 3.36, using
the fact that we now have
Z t
y(t) = eAt x0 +
h (t )u( ) d + Du(t),
0

= (A, b, ct , 01 ).
where

Now admittedly this is a lengthy result, but with any given example, it is simple enough
to apply. Let us justify this by applying the result to an example.
3.38 Example (Example 3.29 contd) We take = (A, b, ct , D) with

0
1
0
1
A=
, b=
, c=
, D = 01 .
2 2
1
1

y(t) + 2y(t)
+ 2y(t) = t 1,

y(0+) = 1, y(0+)

= 0.

Well now, if this isnt the same initial value problem encountered in Examples 3.29 and 3.33!
Of course, the initial state vector x0 was designed to accomplish this. In any case, we may
solve this initial value problem in the manner of either of Examples 3.29 and 3.33, and you
will recall that the answer is

3.39 Definition Let (N, D) be a SISO linear system in input/output form. The step response for (N, D) is the function 1N,D (t) whose Laplace transform is 1s TN,D (s). The ramp
response is the function RN,D (t) whose Laplace transform is s12 TN,D (s).

d
y(t) = N dt
u(t) follows since
+
+
t
1
L0+ (y)(s) = c (sI n A) b + D L0+ (u)(s),

Proof That y satisfies the differential equation D

1s
.
s2 + 2s + 2

3.6.5 Formulae for impulse, step, and ramp responses In this section we focus on
developing initial value problems for obtaining some of the basic outputs for SISO systems.
We provide this for both SISO linear systems, and those in input/output form. Although
the basic definitions are made in the Laplace transform domain, the essential goal of this
section is to provide initial value problems in the time-domain, and in so doing provide the
natural method for numerically obtaining the various responses. These responses can be
obtained by various control packages available, but it is nice to know what they are doing,
since they certainly are not performing inverse Laplace transforms!

i,j
i+j=k1

T (s) =

y(t) = et ( 23 sin t + 2 cos t) + 2t 1.

y (1) (0+) = ct Ax0 + ct bu(0+) + Du(1) (0+),

..
.
X
(k)
y (0+) = cAk x0 +
ct Ai bu(j) (0+) + Du(k) (0+),

y (n1) (0+) = cAn1 x0 +

22/10/2004

Thus this is simply the state-space version of the system we dealt with in Examples 3.29
and 3.33. As input we take again u(t) = t, and as initial state we select x0 = ( 45 , 15 ).
Following Proposition 3.37 we compute

h (0+) = ct Ak b,
and using this expression, the result follows.

108

Note that 1s is the Laplace transform of the unit step 1(t) and that s12 is the Laplace
transform of the unit slope ramp input u(t) = t1(t). Of course, we could define the response
to an input u(t) = tk 1(t) for k 2 by noting that the Laplace transform of such an input is
k!
. Indeed, it is a simple matter to produce the general formulas following what we do in
sk+1
this section, but we shall not pursue this level of generality here.
We now wish to produce the scalar initial value problems whose solutions are the impulse, step, and ramp responses Fortunately, the hard work has been done already, and we
essentially have but to state the answer.
3.40 Proposition Let (N, D) be a proper SISO linear system in input/output form with
D(s) = sn + pn1 sn1 + + p1 s + p0
N (s) = cn sn + cn1 sn1 + + c1 s + c0 ,
and let N,D = (A, b, ct , D) be the canonical minimal realisation of (N, D). The following
statements hold.

3.6 The matter of computing outputs

22/10/2004

109

110

3 Transfer functions (the s-domain)

22/10/2004

(i) If (N, D) is strictly proper then the impulse response hN,D is the solution to the initial
value problem

(1)
(n1)
d
D dt
hN,D (t) = 0, hN,D (0+) = ct b, hN,D (0+) = ct Ab, . . . , hN,D (0+) = ct An1 b.

To determine C1 and C2 we use the initial conditions. These give

(ii) The step response 1N,D (t) is the solution to the initial value problem

(1)
(n1)
d
1N,D (t) = c0 , 1N,D (0+) = D, 1N,D (0+) = ct b, . . . , 1N,D (0+) = ct An2 b.
D dt

Solving gives C1 = 1 and C2 = 2, and so we have

hN,D (0+) = C1 = 1
h N,D (0+) = C1 + C2 = 3.
hN,D (t) = et (2 sin t + cos t).
For the step response the initial value problem is
1N,D (t) + 21 N,D (t) + 21N,D (t) = 1, 1N,D (0+) = 0, 1 N,D (0+) = ct b = 1.

(iii) The ramp response RN,D (t) is the solution to the initial value problem

RN,D (t) = c1 + c0 t,

(1)

RN,D (0+) = 0, RN,D (0+) = D, . . . ,

(2)

(n1)

RN,D (0+) = ct b, . . . , RN,D (0+) = ct An3 b.

Proof Let us look first at the impulse response. Since the Laplace transform of hN,D is the
transfer function TN,D we have
+
D(s)L0+
(hN,D )(s) = N (s).

As we saw in the course of the proof of Proposition 3.27, the fact that the right-hand side
of this equation is a polynomial in s of degree at most
n 1 (if deg(D) = n) implies that

d
hN,D (t) = 0. The determination of the
hN,D is a solution of the differential equation D dt
initial conditions is a simple matter of differentiating hN,D (t) = ct eAt b the required number
of times, and evaluating at t = 0.
The last two statements follow from Lemma 3.36 choosing u(t) = 1(t) for the step
response, and u(t) = t1(t) for the ramp response.

The impulse response is generalised for proper systems in Exercise E3.1. One can, I
expect, see how the proposition gets generalised for inputs like u(t) = tk 1(t). Lets now
determine the impulse, step, and ramp response for an example so the reader can see how
this is done.
3.41 Example We take (N (s), D(s)) = (1 s, s2 + 2s + 2), the same example we have been
using throughout this section. We shall merely produce the initial value problems that give
the step and ramp response for this problem, and not bother with going through the details
of obtaining the solutions to these initial value problems. The canonical minimal realisation
is N,D = (A, b, ct , D) where

0
1
0
1
A=
, b=
, c=
, D = 01 .
2 2
1
1
We readily compute ct b = 1 and ct Ab = 3.
For the impulse response, we solve the initial value problem
N,D (t) + 2h N,D (t) + 2hN,D (t) = 0,
h

hN,D (0+) = 1,

h N,D (0+) = 3.

We obtain the solution by looking for a solution of the form est , and so determine that s
should be a root of s2 + 2s + 2 = 0. Thus s = 1 i, and so our homogeneous solution
will be a linear combination of the two linearly independent solutions y1 (t) = et cos t and
y2 (t) = et sin t. That is,
hN,D (t) = C1 et cos t + C2 et sin t.

In like manner, the initial value for the ramp response is

N,D (t) + 2R N,D (t) + 2RN,D (t) = t 1, RN,D (0+) = 0, R N,D (0+) = 0.
R
Doing the requisite calculations gives
1N,D (t) =

1
2

12 et (cos t + 3 sin t),

RN,D (t) =

t
2

1 + et cos t + 21 sin t .

You may wish to refer back toPSfrag

this
section on computing step responses later in the
PSfrag replacements
replacements
course of reading this book, since the step response is something we shall frequently write
ke(t)k
ke(t)k
down
e1 (t) without saying much about how we did
e1 (t)it.
e2 (t)
e3 (t)

e2 (t)
e3 (t)

3.42
y(t) Remark Those of you who are proponents
y(t) of the Laplace transform will wonder why
u(t)
u(t)
one
does not simply obtain the impulse response
or step response by obtaining the inverse
h (t)
hN,D
(t)
Laplace
transform of the transfer functionhN,D
or(t)the transfer function multiplied by 1s . While
1 (t)
this
not really a good alternative. Even numerically,
1N,D
(t) is theoretically possible, in practice it 1is
N,D (t)
(t)
(t)
determining
the
inverse
Laplace
transform
is
very difficult, even when it is possible.
(t)
(t)
fj (t)example. We take (N (s), D(s)) = (s2 + 3s +
fj (t)To indicate this, let us consider a concrete
Re response and step response were generated
Re
1,Im
s5 + 5s4 + 10s3 + 20s2 + 10s + 5). The impulse
Im
x1 of these were generated in two ways: (1) by
x1
numerically
and are shown in Figure 3.9. Both

x2
x1
x2
log
0.2
dB
deg 0.15
u = 0 ln
ln coth(|u| /2)
0.1
or 1
m
0.05

yos
0

tos 0
-0.05

,0
0

x2
x1
x2
log
dB
deg
u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

1 (t)

d
dt

h (t)

2.5

7.5

12.5

17.5

20,0
0

0.35
0.3
0.25
0.2
0.15
0.1
0.05
0

2.5

7.5

12.5

17.5

Figure 3.9 Impulse response hN,D (t) (left) and step response
1N,D (t) (right) for (N (s), D(s)) = (s2 +3s+1, s5 +5s4 +10s3 +
20s2 + 10s + 5)

computing numerically the inverse Laplace transform, and (2) by solving the ordinary differential equations of Proposition 3.40. In Table 3.1 can be seen a rough comparison of the
time taken to do the various calculations. Obviously, the differential equation methods are
far more efficient, particularly on the step response. Indeed, the inverse Laplace transform
methods will sometimes not work, because they rely on the capacity to factor a polynomial.

3.7 Summary

22/10/2004

111

Table 3.1 Comparison of using inverse Laplace transform and ordinary differential equations to obtain impulse and step response

Computation

Time using inverse Laplace transform

Time taken using ode

Impulse response
Step response

14s
5m15s

< 1s
1s

3.7 Summary
This chapter is very important. A thorough understanding of what is going on here is
essential and let us outline the salient facts you should assimilate before proceeding.
1. You should know basic things about polynomials, and in particular you should not hesitate at the mention of the word coprime.
2. You need to be familiar with the concept of a rational function. In particular, the words
canonical fractional representative (c.f.r.) will appear frequently later in this book.
3. You should be able to determine the partial fraction expansion of any rational function.
4. The definition of the Laplace transform is useful, and you ought to be able to apply it
when necessary. You should be aware of the abscissa of absolute convergence since it can
once in awhile come up and bite you.
5. The properties of the Laplace transform given in Proposition E.9 will see some use.
6. You should know that the inverse Laplace transform exists. You should recognise the
value of Proposition E.11 since it will form the basis of parts of our stability investigation.
7. You should be able to perform block diagram algebra with the greatest of ease. We will
introduce some powerful techniques in Section 6.1 that will make easier some aspects of
this kind of manipulation.
8. Given a SISO linear system = (A, b, ct , D) you should be able to write down its
transfer function T .
9. You need to understand some of the features of the transfer function T ; for example,
you should be able to ascertain from it whether a system is observable, controllable, and
whether it is minimum phase.
10. You really really need to know the difference between a SISO linear system and a SISO
linear system in input/output form. You should also know that there are relationships
between these two different kinds of objectsthus you should know how to determine
N,D for a given strictly proper SISO linear system (N, D) in input/output form.
11. The connection between the impulse response and the transfer function is very important.
12. You ought to be able to determine the impulse response for a transfer function (N, D)
in input/output form (use the partial fraction expansion).

112

3 Transfer functions (the s-domain)

22/10/2004

Exercises
The impulse response hN,D for a strictly proper SISO linear system (N, D)
in input/output

d
d
form has the property that every solution of the differential equation D dt
y(t) = N dt
u(t)
can be written as
Z t
hN,D (t )u( ) d
y(t) = yh (t) +
0

d
where yh (t) is a solution of D dt
yh (t) = 0 (see Proposition 3.32). In the next exercise, you
will extend this to proper systems.
E3.1 Let (N, D) be a proper, but not necessarily strictly proper, SISO linear system in
input/output form.
(a) Show that the transfer function TN,D for (N, D) can be written as
TN,D (s) = TN,
D
(s) + C
D)
in input/output
for a uniquely defined strictly proper SISO linear system (N,
D)
and C in terms of (N, D).
form, and constant C R. Explicitly determine (N,

d
d
(b) Show that every solution of D dt
y(t) = N dt
u(t) can be written as a linear

d y(t) = N
d u(t).
combination of u(t) and y(t) where y(t) is a solution of D
dt
dt

d
d
(c) Conclude that the solution of D dt y(t) = N dt u(t) can be written as
Z t
Z t
y(t) = yh (t) +
hN,
(t )u( ) d
D
(t )u( ) d + C
0

where (t) satisfies

(t t0 )f (t) dt = f (t0 )

for any integrable function f . Is (t) really a map from R to R?

E3.2 Determine the transfer function from r to y for the block diagram depicted in Figure E3.1.
R4 (s)

r(s)

R1 (s)

R2 (s)

R3 (s)

y(s)

R5 (s)

R6 (s)

Figure E3.1 A block diagram with three loops

At various time throughout the text, we will want to have some properties of real rational
functions as s becomes large. In the following simple exercise, you will show that this notion
does not depend on the sense in which s is allowed to go to infinity.
E3.3 Let (N, D) be a proper SISO linear system in input/output form.

Exercises for Chapter 3

113

(a) Show that the limit

P1 (s) = s2 + 1,

is real and independent of (, ]. Thus it makes sense to write

lim TN,D (s),

and this limit will be in R.

(b) Show that the limit in part (a) is zero if (N, D) is strictly proper.
(c) If (N, D) is not strictly proper (but still proper), give an expression for the limit
from part (a) in terms of the coefficients of D and N .
E3.4 For the SISO linear system = (A, b, ct , 01 ) with

0
1
A=
, b=
, c=

1
0
for R and > 0, determine T .
E3.5 For the SISO linear systems connected in series in Exercise E2.1, determine the transfer
function for the interconnected SISO linear system (i.e., for the A, b, c, and D
determined in that exercise).
E3.6 For the SISO linear systems connected in parallel in Exercise E2.2, determine the
transfer function for the interconnected SISO linear system (i.e., for the A, b, c, and
D determined in that exercise).
E3.7 For the SISO linear systems connected in the negative feedback configuration of Exercise E2.3, we will determine the transfer function for the interconnected SISO linear
system (i.e., for the A, b, c, and D determined in that exercise).
(a) Use Lemma A.2 to show that
c (sI n1 +n2

3 Transfer functions (the s-domain)

22/10/2004

E3.9 Consider the two polynomials that are not coprime:

lim TN,D (Rei )

114

1

sI n1 A1
b1 ct2
b1
A) b = 0t ct2
= ct2 U b1 ,
b2 ct1
sI n2 A2
0
1

where
U = (sI n2 A2 ) + b2 ct1 (sI n1 A1 )1 b1 ct2

b2 ct1 (sI n1 A1 )1 .

(b) Use your answer from part (a), along with Lemma A.3, to show that
1

T (s) = (1 + T2 (s)T1 (s)) T2 (s)T1 (s).

T1 (s)T2 (s)
.
1 + T1 (s)T2 (s)

Could you have deduced this otherwise?

(d) Use the computed transfer function to determine the characteristic polynomial
for the system. (Thus answer Exercise E2.3(b).)
E3.8 Let = (A, b, ct , D) be a SISO linear system with transfer function T . Show that
T is proper, and is strictly proper if and only if D = 01 .

P2 (s) = s3 + s2 + s + 1.

Answer the following questions:

(a) Construct two SISO linear system i = (Ai , bi , cti , 01 ), i = 1, 2, each with state
dimension 3, and with the following properties:
1. 1 is controllable but not observable;
2. 2 is observable but not controllable;
, i = 1, 2.
3. Ti (s) = PP12 (s)
(s)
Do not take any of Ai , bi , ci , i = 1, 2, to be zero.
(b) Is it possible to find a controllable and observable system 3 , with state dimension
3, for which T3 (s) = PP12 (s)
? If so find such a system, and if not explain why not.
(s)
(c) Is it possible to find an uncontrollable and unobservable system 4 , with state
(s)
dimension 3, for which T3 (s) = PP12 (s)
? If so find such a system, and if not explain
why not.
Hint: Use Theorem 2.41.
E3.10 Consider a SISO linear system (N, D) in input/output form, and suppose that N has
a real root z > 0 with multiplicity one. Thus, in particular, (N, D) is nonminimum
phase. Let u(t) = 1(t)ezt .
(a) Show that there exists a unique set of initial conditions y(0), y (1) (0), . . . , y (n1) (0)
d
d
for which the solution to the differential equation D( dt
)y(t) = N ( dt
)u(t), with
these initial conditions, is identically zero.
Hint: Think about what Proposition 3.27 says about the method of Laplace transforms for solving ordinary differential equations.
(b) Comment on this in light of the material in Section 2.3.3.
For the following three exercises we will consider open-circuit and short-circuit behaviour
of circuits. Let us make the definitions necessary here in a general setting by talking about
a SISO linear system = (A, b, ct , D). A pair of functions (u(t), y(t)) satisfying

x(t)
= Ax(t) + bu(t)
y(t) = ct x(t) + Du(t)
is open-circuit for if y(t) = 0 for all t and is short-circuit for if u(t) = 0 for all t.
E3.11 Consider the circuit of Exercise E1.7 with the output of Exercise E2.16.
(a) Compute the transfer function for the system.
(b) Determine all open-circuit pairs (u(t), y(t)).
(c) Determine all short-circuit pairs (u(t), y(t)).
(d) Comment on the open/short-circuit behaviour in terms of controllability, observability, and zero dynamics.
E3.12 Consider the circuit of Exercise E1.8 with the output of Exercise E2.17.
(a) Compute the transfer function for the system.
(b) Determine all open-circuit pairs (u(t), y(t)).
(c) Determine all short-circuit pairs (u(t), y(t)).
(d) Comment on the open/short-circuit behaviour in terms of controllability, observability, and zero dynamics.

Exercises for Chapter 3

115

E3.13 Consider the circuit of Exercise E1.9 with the output of Exercise E2.18.
(a) Compute the transfer function for the system.
(b) Determine all open-circuit pairs (u(t), y(t)).
(c) Determine all short-circuit pairs (u(t), y(t)).
(d) Comment on the open/short-circuit behaviour in terms of controllability, observability, and zero dynamics.
E3.14 For the coupled mass system of Exercises E1.4 and E2.19 (assume no damping), take
as input the case of = 0 described in Exercise E2.19thus the input is a force
u(t) applied to the leftmost mass. Determine an output that renders the system
unobservable.
E3.15 Using Theorem 3.15, determine the spectrum of the zero dynamics for the pendulum/cart system of Exercises E1.5 and E2.4 for each of the following linearisations:
(a) the equilibrium point (0, 0) with cart position as output;
(b) the equilibrium point (0, 0) with cart velocity as output;
(c) the equilibrium point (0, 0) with pendulum angle as output;
(d) the equilibrium point (0, 0) with pendulum angular velocity as output;
(e) the equilibrium point (0, ) with cart position as output;
(f) the equilibrium point (0, ) with cart velocity as output;
(g) the equilibrium point (0, ) with pendulum angle as output;
(h) the equilibrium point (0, ) with pendulum angular velocity as output.
E3.16 Consider the double pendulum of Exercises E1.6 and E2.5. In the following cases, use
Theorem 3.15 to determine the spectrum of the zero dynamics:
(a) the equilibrium point (0, 0, 0, 0) with the pendubot input;
(b) the equilibrium point (0, , 0, 0) with the pendubot input;
(c) the equilibrium point (, 0, 0, 0) with the pendubot input;
(d) the equilibrium point (, , 0, 0) with the pendubot input;
(e) the equilibrium point (0, 0, 0, 0) with the acrobot input;
(f) the equilibrium point (0, , 0, 0) with the acrobot input;
(g) the equilibrium point (, 0, 0, 0) with the acrobot input;
(h) the equilibrium point (, , 0, 0) with the acrobot input.
In each case, use the angle of the second link as output.
E3.17 Determine the spectrum of the zero dynamics for the linearised coupled tank system
of Exercises E1.11 and E2.6 for the following outputs:
(a) the height in tank 1;
(b) the height in tank 2;
(c) the difference of the heights in the tanks.
E3.18 Given the SISO linear system (N, D) in input/output form with
D(s) = s3 + 4s2 + s + 1,

N (s) = 3s2 + 1,

determine the canonical minimal realisation N,D . Is the triple (A, b, c) you found
complete? Was your first impulse to answer the previous question by doing calculations? Explain why these are not necessary.

116

3 Transfer functions (the s-domain)

22/10/2004

E3.19 State and prove a version of Theorem 3.20 that assigns to a SISO system (N, D) in
input/output form a SISO linear system = (A, b, ct , D) so that A and c are in
observer canonical form. Also produce the block diagram analogous to Figure 3.8 in
this case.
E3.20 Suppose you are handed a SISO linear system (N, D) in input/output form, and are
told that it comes from a SISO linear system = (A, b, ct , 01 )that is, you are told
that T = TN,D . Is it possible for you to tell whether (A, b) is controllable?
Hint: Consider Example 2.19 and Theorem 3.20.
E3.21 For a SISO linear system (N, D) in input/output form, show that the transfer function
TN,D has the property that TN,D (
s) = TN,D (s) for all s C. In particular, show that
s0 C is a zero or pole of TN,D if and only if s0 is a zero or pole, respectively.
E3.22 Verify Theorem 3.22 and Proposition 3.24 for Exercise E3.4 (recall that you had
computed the impulse response for this problem in Exercise E2.26).
E3.23 For the following SISO linear systems in input/output form, use Proposition 3.32 to
obtain the output corresponding to the given input and initial conditions.
(a) (N (s), D(s)) = (1, s + 3), u(t) = 1(t) (the unit step input), and y(0) = 1.
(b) (N (s), D(s)) = (1, s + 3), u(t) = 1(t)eat , a R, and y(0) = 0.
(c) (N (s), D(s)) = (s, s3 + s), u(t) = 1(t) cos t, and y(0) = 1, y(0)

= 0, and y(0) = 0.
(d) (N (s), D(s)) = (1, s2 ), u(t) = 1(t), and y(0) = 0 and y(0)

= 1.
E3.24 For the SISO linear systems in input/output form from Exercise E3.23 for which you
obtained the solution, apply Proposition 3.27 to obtain the same solution.
E3.25 For the following SISO systems in input/output form, use Proposition 3.40 to setup
the initial value problem for the step response, and use a computer package to plot
the step response.
(a) (N (s), D(s)) = (s + 1, s2 + s + 1).
(b) (N (s), D(s)) = (s2 + 2s + 1, s3 + 3s + 1).
(c) (N (s), D(s)) = (s 1, s4 + 15s3 + 20s2 + 10s + 2).
(d) (N (s), D(s)) = (s3 + 1, s5 + 9s4 + 20s3 + 40s2 + 50s + 25).
E3.26 Consider the differential equation
y(t) + 4y(t)
+ 8y(t) = 2u(t)
+ 3u(t),

(E3.1)

where u(t) is a specified function of time. If we define

(
1
, t [0, ]
u (t) =
0, otherwise,
and let y (t) be the solution to the differential equation (E3.1) when u = u , determine
lim0 y (t). (Note that in this problem although u is in U , u is not in U . Thus to
compute directly the solution to the differential equation (E3.1) with u = u is not
actually something you know how to do!)
E3.27 Consider the SISO linear system = (A, b, ct , D) given by

A= 0 , b= 1 , c= 1 , D= 0 ,
2

and let u(t) = 1(t)et .

Exercises for Chapter 3

117

(a) Show that if x(0) = 0 then the output for the input u is y(t) =
erf is the error function given by
Z t
2
2
e d.
erf(t) =
0

erf(it),
2i

where

(b) Laplace transform techniques are always limited by ones ability to compute the
inverse transform. Are there limitations in this example beyond the difficulty in
determining the inverse transform?

118

3 Transfer functions (the s-domain)

22/10/2004

120

This version: 22/10/2004

4 Frequency response (the frequency domain)

22/10/2004

by
= { R | i is a pole of T } .
The frequency response for is the function H : R\ C defined by H () = T (i).
Note that we do wish to think of the frequency response as a C-valued function, and not a
rational function, because we will want to graph it. Thus when we write T (i), we intend
to evaluate the transfer function at s = i. In order to do this, we suppose that all poles
and zeroes of T have been cancelled.
The following result gives a key interpretation of the frequency response.

Chapter 4
Frequency response (the frequency domain)
The final method we will describe for representing linear systems is the so-called frequency domain. In this domain we measure how the system responds in the steady-state
to sinusoidal inputs. This is often a good way to obtain information about how your system
will handle inputs of various types.
The frequency response that we study in this section contains a wealth of information,
often in somewhat subtle ways. This material, that builds on the transfer function discussed
in Chapter 3, is fundamental to what we do in this course.

Contents
4.1

The frequency response of SISO linear systems . . . . . . . . . . . . . . . . . . . . . . . . 119

4.2

The frequency response for systems in input/output form . . . . . . . . . . . . . . . . . . 122

4.3

Graphical representations of the frequency response . . . . . . . . . . . . . . . . . . . . . 124

4.4

4.5

4.6

4.3.1

The Bode plot . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124

4.3.2

A quick and dirty plotting method for Bode plots . . . . . . . . . . . . . . . . . . 129

4.3.3

The polar frequency response plot . . . . . . . . . . . . . . . . . . . . . . . . . . . 135

Properties of the frequency response . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135

4.4.1

Time-domain behaviour reflected in the frequency response . . . . . . . . . . . . . 135

4.4.2

Bodes Gain/Phase Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138

Uncertainly in system models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145

4.5.1

Structured and unstructured uncertainty . . . . . . . . . . . . . . . . . . . . . . . 146

4.5.2

Unstructured uncertainty models

. . . . . . . . . . . . . . . . . . . . . . . . . . . 147

Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150

4.1 Theorem Let = (A, b, ct , 01 ) be an complete SISO control system and let > 0. If
T has no poles on the imaginary axis that integrally divide ,1 then, given u(t) = u0 sin t
there is a unique periodic output yp (t) with period T = 2
satisfying (4.1) and it is given by

y(t) = u0 Re(H ()) sin t + u0 Im(H ()) cos t.

Proof We first look at the state behaviour of the system. Since (A, c) is complete, the
numerator and denominator polynomials of T are coprime. Thus the poles of T are exactly
the eigenvalues of A. If there are no such eigenvalues that integrally divide , this means that

there are no periodic solutions of period T for x(t)

= Ax(t). Therefore the linear equation
eAT u = u has only the trivial solution u = 0. This means that the matrix eAT I n is
invertible. We define
Z T
Z t
xp (t) = u0 eAt (eAT I n )1
eA b sin d + u0
eA(t ) b sin d,
(4.2)
0

and we claim that x(t) is a solution to the first of equations (4.1) and is periodic with period
2
. If T = 2
we first note that

Z T
eA b sin d +
xp (t + T ) = u0 eA(t+T ) (eAT I n )1
0
Z t+T
u0
eA(t+T ) b sin d
0
Z T
= eAt eAT (eAT I n )1
eA b sin d +
0
Z T
Z t+T
u0 eAt eAT
eA b sin d + u0
eA(t+T ) b sin d
0
T
Z T
= eAt (eAT (eAT I n )1 + eAT )
eA b sin d +
0
Z t
u0
eA(t ) b sin d.
0

4.1 The frequency response of SISO linear systems

Periodicity of xp (t) will follow then if we can show that

eAT (eAT I n )1 + eAT = (eAT I n )1 .

We first look at the state-space representation:

x(t)
= Ax(t) + bu(t)
y(t) = ct x(t) + Du(t).

But we compute
(4.1)

Let us first just come right out and define the frequency response, and then we can give its
interpretation. For a SISO linear control system = (A, b, c, D) we let R be defined

eAT (eAT I n )1 + eAT = (eAT + eAT (eAT I n ))(eAT I n )1

= (eAT I n )1 .
1

Thus there are no poles for T of the form i

where

4.1 The frequency response of SISO linear systems

22/10/2004

121

Thus xp (t) has period T . That xp (t) is a solution to (4.1) with the u(t) = u0 sin t follows
since xp (t) is of the form
Z t
eA(t ) b sin d
x(t) = eAt x0 + u0
0

provided we take
x0 = (eAT I n )1

eA b sin d.

For uniqueness, suppose that x(t) is a periodic solution of period T . Since it is a solution
it must satisfy
Z
t

eA(t ) b sin d

x(t) = eAt x0 + u0

for some x0 Rn . If x(t) has period T then we must have

Z T
Z t
eA(t ) b sin d +
eA(t ) b sin d = eAt eAT x0 + u0 eAt eAT
eAt x0 + u0
0
0
Z t+T
u0
eA(t+T ) b sin d
T
Z T
= eAt eAT x0 + u0 eAt eAT
eA(t ) b sin d +
0
Z t
u0
eA(t ) b sin d.

eA b sin d,

which means that x(t) = xp (t).

This shows that there is a unique periodic solution in state space. This clearly implies
a unique output periodic output yp (t) of period T . It remains to show that yp (t) has the
asserted form. We will start by giving a different representation of xp (t) than that given
in (4.2). We look for constant vectors x1 , x2 Rn with the property that
xp (t) = x1 sin t + x2 cos t.

4.2 Remarks 1. It turns out that any output from (4.1) with u(t) = u0 sin t can be written
as a sum of the periodic output yp (t) with a function yh (t) where yh (t) can be obtained
with zero input. This is, of course, reminiscent of the procedure in differential equations
where you find a homogeneous and particular solution.
2. If the eigenvalues of A all lie in the negative half-plane, then it is easy to see that
limt |yh (t)| = 0 and so after a long enough time, we will essentially be left with the
periodic solution yp (t). For this reason, one calls yp (t) the steady-state response and
yh (t) the transient response. Note that the steady-state response is uniquely defined
(under the hypotheses of Theorem 4.1), but that there is no unique transient responseit
depends upon the initial conditions for the state vector.
3. One can generalise this slightly to allow for imaginary eigenvalues i
of A for which

integrally divide , provided that b does not lie in the eigenspace of these eigenvalues.

4.3 Proposition Let (N, D) be a strictly proper SISO linear control system in input/output
form, and suppose that the poles of TN,D are in the negative half-plane. Then HN,D () =
N,D ().
h
Proof We have HN,D () = TN,D (i) and so
Z
hN,D (t)eit dt.
HN,D () =
0+

Substitution into (4.1) with u(t) = u0 sin t gives

=
=

22/10/2004

The matter of defining the frequency response for a SISO linear system in input/output
form is now obvious, I hope. Indeed, if (N, D) is a SISO linear system in input/output form,
then we define its frequency response by HN,D () = TN,D (i).
Let us see how one may recover the transfer function from the frequency response. Note
that it is not obvious that one should be able to do this. After all, the frequency response
function only gives us data on the imaginary axis. However, because the transfer function
is analytic, if we know its value on the imaginary axis (as is the case when we know the
frequency response), we may assert its value off the imaginary axis. To be perfectly precise
on these matters requires some effort, but we can sketch how things go.
The first thing we do is indicate a direct correspondence between the frequency response
and the impulse response. For this we refer to Section E.2 for a definition of the Fourier
transform. With the notion of the Fourier transform in hand, we establish the correspondence
between the frequency response and the impulse response as follows.

4 Frequency response (the frequency domain)

4.2 The frequency response for systems in input/output form

But this implies that we must have

x0 = eAT x0 + u0 eAT

122

By Exercise EE.2, min (hN,D ) < 0 since we are assuming all poles are in the negative halfplane. Therefore this integral exists. Furthermore, since hN,D (t) = 0 for t < 0 we have
Z
N,D ().
HN,D () =
hN,D (t)eit dt = h

x1 cos t x2 sin t = Ax1 sin t + Ax2 cos t + u0 b sin t

x1 Ax2 , Ax1 x2 = u0 b
i(x1 + ix2 ) A(x1 + ix2 ) = u0 b.

Since i is not an eigenvalue for A we have

This completes the proof.

x1 + ix2 = (iI n A)1 b

ct x1 = Re(H ()), ct x2 = Im(H ()).

The result follows since yp (t) = ct xp (t).

Now we recover the transfer function TN,D from the frequency response HN,D . In the
following result we are thinking of the transfer function not as a rational function, but as a
C-valued function.

22/10/2004

4.2 The frequency response for systems in input/output form

123

4.4 Proposition Let (N, D) be a strictly proper SISO linear control system in input/output
form, and suppose that the poles of TN,D are in the negative half-plane. Then, provided
Re(s) > min (hN,D ), we have
Z
HN,D ()
1
d.
TN,D (s) =
2 s i
Proof By Proposition 4.3 we know that hN,D is the inverse Fourier transform of HN,D :
Z
1
hN,D (t) =
HN,D ()eit d.
2

4 Frequency response (the frequency domain)

4.5 Remarks 1. I hope you can see the importance of the results in this section. What we
have done is establish the perfect correspondence between the three domains in which
we work: (1) the time-domain, (2) the s-plane, and (3) the frequency domain. In each
domain, one object captures the essence of the system behaviour: (1) the impulse response, (2) the transfer function, and (3) the frequency response. The relationships are
summarised in Figure 4.1. Note that anything you say about one of the three objects in

4.3 Graphical representations of the frequency response

One of the reasons why frequency response is a powerful tool is that it is possible to
succinctly understand its primary features by means of plotting functions or parameterised
curves. In this section we see how this is done. Some of this may seem a bit pointless at
present. However, as matters develop, and we get closer to design methodologies, the power
of these graphical representations will become clear. The first obvious application of these
ideas that we will encounter is the Nyquist criterion for stability in Chapter 7.

H () = |H ()| ei]H ()
where |H ()| is the absolute value of the complex number H () and ]H () is the
argument of the complex number H (). We take 180 < ]H () 180 . One then
constructs two plots, one of 20 log |H ()| as a function of log , and the other of ]H ()
as a function of log . (All logarithms we talk about here are base 10.) Together these
two plots comprise the Bode plot for the frequency response H . The units of the plot of
20 log |H ()| are decibels.2 One might think we are losing information here by plotting
the magnitude and phase for positive values of (which we are restricted to doing by using
log as the independent variable). However, as we shall see in Proposition 4.13, we do
not lose any information since the magnitude is symmetric about = 0, and the phase is
anti-symmetric about = 0.
Lets look at the Bode plots for our mass-spring-damper system. I used Mathematica
to generate all Bode plots in this book. We will also be touching on a method for roughly
determining Bode plots by hand.
4.6 Examples In all cases we have

impulse response o
(time domain)
eJJ

JJ JJJJ
JJ JJ
JJ JJF 1
JJ JJ
JJ JJ
JJ JJ
JJ JJ
F
JJ JJ
JJ
JJ
%

22/10/2004

4.3.1 The Bode plot What one normally does with the frequency response is plot it.
But one plots it in a very particular manner. First write H in polar form:

On the other hand, by Theorem 3.22 the transfer function TN,D is the Laplace transform of
hN,D so we have, for Re(s) > min (hN,D ).
Z
hN,D (t)est dt
TN,D (s) =
0+
Z Z
1
=
HN,D ()eit est dt d
2 0+
Z
HN,D ()
1
=
d.
2 s i
This completes the proof.

124

transfer function
(Laplace
domain)
r

rr 9
rr rr
rr rr
restrict to i rrrr rrrr
r
r
rr rr
rr rr
rr rrr analytic continuation
r
yrr rrr
rr

0
1
,
k
m
md

0
1
m

We take m = 1 and consider the various cases of d and k as employed in Example 2.33. Here
we can also consider the case when D 6= 01 .
1. We take d = 3 and k = 2.
(a) With c = (1, 0) and D = 01 we compute

frequency response
(frequency domain)

H () =
Figure 4.1 The connection between impulse response, transfer
function, and frequency response

Figure 4.1 must be somehow reflected in the others. We will see that this is true, and
will form the centre of discussion of much of the rest of the course.
2. Of course, the results in this section may be made to apply to SISO linear systems in the
form (4.1) provided that D = 01 and that the polynomials ct adj(sI n A)b and PA (s)
are coprime.

1
.
2 + 3i + 2

The corresponding Bode plot is the first plot in Figure 4.2.

(b) Next, with c = (0, 1) and D = 01 we compute
H () =

i
,
2 + 3i + 2

and the corresponding Bode plot is the second plot in Figure 4.2.
2

Decibels are so named after Alexander Graham Bell. The unit of bell was initially proposed, but when
it was found too coarse a unit, the decibel was proposed.

PSfrag replacements

-150

u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

0.5

1.5

100
50
0
-50
-100
-150
-1.5 -1 -0.5

log

,0
0

0.5

1.5

0
-20
-40
-60
-80
-1.5 -1 -0.5

log

0.5

1.5

150
100
50
0
-50
-100
-150

,0
0

-1.5 -1 -0.5

log

0.5

1.5

Figure 4.2 The displacement (top left), velocity (top right), and
acceleration (bottom) frequency response for the mass-spring
damper system when d = 3 and k = 2
k
(c) If we have c = ( m
, md ) and D = [1] we compute

H () =

u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

,0
0

-60
-80

150
100
50
0
-50
-100
-150

-40

PSfrag replacements

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)
h (t) 0
-1.5 -1 -0.5
hN,Dlog
(t)
1 (t)
1N,D (t)
(t)
(t)
fj (t)
Re
Im
x1
x2
x1
x2

u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

0.5

2. We take d = 2 and k = 1.

1
.
2 + 2i + 1

The corresponding Bode plot is the first plot in Figure 4.3.

1.5

0.5

1.5

150

u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

1
1.5
2

100
50
0
-50
-100
-150
-1.5 -1 -0.5

log

,0
0

-40
-60
-80

-1.5 -1 -0.5

log

0.5

1.5

150
100
50
0
-50
-100
-150

,0
0

-1.5 -1 -0.5

log

0.5

1.5

Figure 4.3 The displacement (top left), velocity (top right), and
acceleration (bottom) frequency response for the mass-spring
damper system when d = 2 and k = 1

(b) Next, with c = (0, 1) and D = 01 we compute

H () =

i
,
2 + 2i + 1

and the corresponding Bode plot is the second plot in Figure 4.3.
k
(c) If we have c = ( m
, md ) and D = [1] we compute

(a) With c = (1, 0) and D = 01 we compute

H () =

0.5

-20

2
,
2 + 3i + 2

The Bode plot for this frequency response function is the third plot in Figure 4.2.

22/10/2004

-100

0
-20

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)
4 Frequency response (the
h (t)frequency domain)
hN,D (t)
1 (t)
1N,D (t)
-10
(t)
(t)
-15
fj (t)
-20
Re
-25
Im
-30
x1
-35
x2
x1
-40
-1.5 -1 -0.5 0 0.5 1
1.5
2
-1.5 -1 -0.5 0
x2
log
log

deg

-50

0.5

150

u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

1
1.5
2

deg

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)
h (t) 0
-1.5 -1 -0.5
hN,Dlog
(t)
1 (t)
1N,D (t)
(t)
(t)
fj (t)
Re
Im
x1
x2
x1
x2

log

PSfrag replacements

deg

100

-1.5 -1 -0.5

PSfrag replacements

125

,0
0

150

deg

u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

the frequency response

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)
h126
(t)
hN,D (t)
1 (t)
1N,D (t)
(t)
(t)
fj (t)
Re
Im
x1
x2
x1
x2

deg

PSfrag replacements

t
t
ke(t)k
ke(t)k
e1 (t)
e1 (t)
e2 (t)
e2 (t)
e3 (t)
e3 (t)
y(t)
y(t)
u(t)
u(t)
4.3 Graphical representations
of
h22/10/2004
h (t)
(t)
hN,D (t)
hN,D (t)
1 (t)
1 (t)
-10
1N,D (t)
1N,D (t)
-10
(t)
(t)
-15
-20
(t)
(t)
-20
-30
fj (t)
fj (t)
-25
-40
Re
Re
-30
-50
Im
Im
-35
-60
x1
x1
-40
x2
x2
-70
x1
x1
-45
-80
-1.5 -1 -0.5 0
0.5
1
1.5
2
x2
x2
log

deg

PSfrag replacements

H () =

2
,
2 + 2i + 1

The Bode plot for this frequency response function is the third plot in Figure 4.3.
3. We take d = 2 and k = 10.

4.3 Graphical representations of the frequency response

-70
-1.5 -1 -0.5

0.5

log

100
50
0
-50
-100
-150

PSfrag replacements

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)
h (t) 0 0.5
-1.5 -1 -0.5
hN,Dlog
(t)
1 (t)
1N,D (t)
0
(t)
-20
(t)
fj (t)
-40
Re
-60
Im
x1
-80
x2
x1
-100
x2

u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

1
1.5
2

4. We
PSfrag replacements

0.5

,0
0

1.5

50
0
-50
-100
-150

log

,0
0

0.5

1.5

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)
h (t)
hN,D (t)
1 (t)
1N,D (t)
(t)
(t)
fj (t)
Re
Im
x1
x2
x1
x2

0.5

1.5

-1.5 -1 -0.5

0.5

1.5

log

150
100

PSfrag replacements

e1 (t)
e2 (t)

He()
3 (t) =

The corresponding Bode

40
20
0

2 + 1
y(t)
u(t)
(t) plot in Figure 4.5.
plot is thehfirst
hN,D (t)
1 (t)
1N,D (t)
60
(t)
(t)
40
fj (t)
20
Re
0
Im
x1
-20
x2
x1
-40
1
1.5
2
-1.5 -1 -0.5 0
x2
log

-20
-40
-60
-1.5 -1 -0.5

0.5

log

0.5

1.5

0.5

1.5

u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

-1.5 -1 -0.5

take d = 0 and k = 1.

t
(a) With c = (1, 0) and D = 01 we compute
ke(t)k

,0
0

150
100
50
0
-50
-100
-150

PSfrag replacements

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)
h (t) 0 0.5
-1.5 -1 -0.5
hN,Dlog
(t)
1 (t)
1N,D (t)
40
(t)
20
(t)
0
fj (t)
Re
-20
Im
-40
x1
x2
-60
x1
x2

150

u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

1
1.5
2

100
50
0
-50
-100
-150
-1.5 -1 -0.5

log

,0
0

u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

100

2
,
+ 2i + 10

The Bode plot for this frequency response function is the third plot in Figure 4.4.

150

-1.5 -1 -0.5

-60

150

H () =

-50

deg

-40

22/10/2004

k
(c) If we have c = ( m
, md ) and D = [1] we compute

deg

-30

4 Frequency response (the frequency domain)

50
0
-50
-100
-150

log

Figure 4.4 The displacement (top left), velocity (top right), and
acceleration (bottom) frequency response for the mass-spring
damper system when d = 2 and k = 10

(b) Next, with c = (0, 1) and D = 01 we compute

H () =

i
,
2 + 2i + 10

and the corresponding Bode plot is the second plot in Figure 4.4.

u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

,0
0

-1.5 -1 -0.5

0.5

1.5

-1.5 -1 -0.5

0.5

1.5

log

150
100
50

deg

-20

ke(t)k
e1 (t)
1
2 (t)
H () ee=
.
3 (t)
2
y(t) + 2i + 10
u(t)
(t) plot in Figure 4.4.
plot is thehfirst
hN,D (t)
1 (t)
1N,D (t)
-10
(t)
(t)
-20
fj (t)
-30
Re
-40
Im
x1
-50
x2
x1
-60
1
1.5
2
-1.5 -1 -0.5 0
x2
log

,0
0

The corresponding Bode

deg

u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

PSfrag replacements

(a) With c = (1, 0) and D = 01 we compute

deg

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)
h (t)
hN,D (t)
1 (t)
1N,D (t)
(t)
(t)
fj (t)
Re
Im
x1
x2
x1
x2

128

PSfrag replacements

127

deg

22/10/2004

0
-50
-100
-150

log

Figure 4.5 The displacement (top left), velocity (top right), and
acceleration (bottom) frequency response for the mass-spring
damper system when d = 0 and k = 1

22/10/2004

4.3 Graphical representations of the frequency response

129

i
,
2 + 1
and the corresponding Bode plot is the second plot in Figure 4.5.
k
(c) If we have c = ( m
, md ) and D = [1] we compute
H () =

and so
H() =

,
2 + 1
The Bode plot for this frequency response function is the third plot in Figure 4.5.
H () =

4.3.2 A quick and dirty plotting method for Bode plots It is possible, with varying
levels of difficulty, to plot Bode plots by hand. The first thing we do is rearrange the
frequency response in a particular way suitable to our purposes. The form desired is
k1
Y

(1 + ij1 )

j1 =1

H() =

(i)

k2
Y

1 + 2ij2 j ( j )2
2

j2 =1

k4
Y

(1 + ij4 )

j4 =1

k5
Y

(4.3)
2ij5 j
5

( j )2
5

j5 =1

where the s, s, and s are real, the s are all further between 1 and 1, and the s are
all positive. The frequency response for any stable, minimum phase system can always be
put in this form. For nonminimum phase systems, or for unstable systems, the variations
to what we describe here are straightforward. The form given reflects our transfer function
having
1. k1 real zeros at the points 1j , j1 = 1, . . . , k1 ,
1
q
2. k2 pairs of complex zeros at j2 (j2 1 j22 ), j2 = 1, . . . , k2 ,
3. k3 poles at the origin,
4. k4 real poles at the points 1j , j4 = 1, . . . , k4 , and
4
q
5. k5 pairs of complex poles at j5 (j5 1 j25 ), j5 = 1, . . . , k5 .
Although we exclude the possibility of having zeros or poles on the imaginary axis, one can
see how to handle such functions by allowing to become zero in one of the order two terms.
Let us see how to perform this in practice.
4.7 Example We consider the transfer function
T (s) =

1
s + 10
.
s(s2 + 4s + 8)

To put this in the desired form we write

1
10

s + 4s + 8 = 8 1 +
=8 1+
=8 1+

1 + 10s

1
s + 18 s2
2

2 14 8 s8 + ( s8 )2 )

2 12 s8 + ( s8 )2 .

4 Frequency response (the frequency domain)

22/10/2004

1
Thus we have a real zero at 10
, a pole at 0, and a pair of complex poles at 2 2i. Thus
we write
1 + 10s
1
T (s) =
80 s(1 + 2 12 s8 + ( s8 )2 )

(b) Next, with c = (0, 1) and D = 01 we compute

130

1 + i10
1
.
80 i(1 + 2i 12 8 ( 8 )2 )

I find it easier to work with transfer functions first to avoid imaginary numbers as long as
possible. You may do as you please, of course.

One can easily imagine that one of the big weaknesses of our computer-absentee plots is that
we have to find roots by hand. . .
Let us see what a Bode plot looks like for each of the basic elements. The idea is to see
what the magnitude and phase looks like for small and large , and to fill in the gaps in
between these asymptotes.
1. H() = K: The Bode plot here is simple. It takes the magnitude 20 log K for all values
of log . The phase is 0 for all log if K is positive, and 180 otherwise.
2. H() = 1 + i : For near zero the frequency response looks like 1 and so has the
value of 0dB. For large the frequency response looks like i , and so the log of the
magnitude will look like log . Thus the magnitude plot for large frequencies we have
|H()| 20 log dB. These asymptotes meet when 20 log = 0 or when = 1 . This
point is called the break frequency . Note that the slope of the frequency response for
large is independent of (since log = log + log ), but its break frequency does
depend on . The phase plot starts at 0 for small and for large , since the frequency
response is predominantly imaginary, becomes 90 . This Bode plot is shown in Figure 4.6
for = 1.
3. H() = 1 + 2i 0 ( 0 )2 : For small the magnitude is 1 or 0dB. For large the
frequency response looks like ( 0 )2 and so the magnitude looks like 40 log 0 . The two
asymptotes meet when 40 log 0 = 0 or when = 0 . One has to be a bit more careful
with what is happening around the frequency 0 . The behaviour here depends on the
value of , and various plots are shown in Figure 4.6 for 0 = 1. As decreases, the
undershoot increases. The phase starts out at 0 and goes to 180 as increases.
4. H() = (i)1 : The magnitude is 1 over the entire frequency range which gives |H()| =
20 log dB. The phase is 90 over the entire frequency range, and the simple Bode
plot is shown in Figure 4.7.
5. H() = (1 + i )1 : The analysis here is just like that for a real zero except for signs.
The Bode plot is shown in for = 1.
6. H() = (1 + 2i 0 ( 0 )2 )1 : The situation here is much like that for a complex zero
with sign reversal. The Bode plots are shown in Figure 4.8 for 0 = 1.
4.8 Remark We note that one often sees the language 20dB/decade. With what we have
done above for the typical elements in a frequency response function. A first-order element in
the numerator increases like 20 log for large frequencies. Thus as increases by a factor
of 10, the magnitude will increase by 20dB. This is where 20dB/decade comes from. If
the first-order element is in the denominator, then the magnitude decreases at 20dB/decade.
Second-order elements in the numerator increase at 40dB/decade, and second-order elements
in the denominator decrease at 40dB/decade. In this way, one can ascertain the relative

,0
0

ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)
(t)
4hFrequency
response (the frequency domain)
hN,D (t)
1 (t)
40
1N,D (t)
(t)
30
(t)
20
fj (t)
10
0
Re
Im
-10
x1
-20
x2
-30
x1
-1.5 -1 -0.5 0 0.5 1 1.5 2
x2
log

150
100
50
0
-50
-100
-150

u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

deg

u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0
-1.5 -1 -0.5 0 0.5 1 1.5 2
,0
log

150
100
50
0
-50
-100
-150
-1.5 -1 -0.5 0 0.5 1 1.5 2

log

,0
0

Figure 4.8 Bode plot for H() = (1 + 2i 2 )1 for =

0.2, 0.4, 0.6, 0.8

degree of the numerator and denominator of a frequency response function by looking at its
slope for large frequencies. Indeed, we have the following rule:
The slope of the magnitude Bode plot for HN,D at large frequencies is 20(deg(N )
deg(D))dB/decade.

To get a rough idea of how to sketch a Bode plot, the above arguments illustrate that the
asymptotes are the most essential feature. Thus we illustrate these asymptotes in Figure 4.19
(see the end of the chapter) for the essential Bode plots in the above list. From these one
can determine the character of most any Bode plot. The reason for this is that in (4.3) we
have ensured that any frequency response is a product of the factors we have individually
examined. Thus when we take logarithms as we do when generating a Bode plot, the graphs
simply add! And the same goes for phase plots. So by plotting each term individually by
the above rules, we end up with a pretty good rough approximation by adding the Bode
plots.
Let us illustrate how this is done in an example.
4.9 Example (Example 4.7 contd) We take the frequency response
H() =

-1.5 -1 -0.5 0 0.5 1 1.5 2

log

1
1 + i10
.
80 i 1 + 2i 12 8 ( 8 )2

Four essential elements will comprise the frequency response:

Figure 4.7 Bode plot for H() = (i)1 (left) and for H() =
(1 + i)1 (right)

22/10/2004

132

150
100
50
0
-50
-100
-150

40
30
20
10
0
-10
-20
-30

150

u = 0 ln
100
ln coth(|u| /2)
50
or 1
0
m

-50
yos -100

PSfrag replacements
tos 0 -150
-1.5 -1 -0.5 0 0.5 1 1.5 2
-1.5 -1 -0.5 0 0.5 1 1.5 2
,t0
log
log
ke(t)k
0
e1 (t)
e2 (t)
e3 (t)= 1 + i (left) and for H() =
Figure 4.6 Bode plot for H()
2
1 + 2i for = 0.2, y(t)
0.4, 0.6, 0.8 (right)
u(t)
h (t)
hN,D (t)
1 (t)
40
1N,D (t)
(t)
30
(t)
20
fj (t)
10
0
Re
Im
-10
x1
-20
x2
-30
x1
-1.5 -1 -0.5 0 0.5 1 1.5 2 x
-1.5 -1 -0.5 0 0.5 1 1.5 2
2
log
log

deg

u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

150
100
50
0
-50
-100
-150

deg

u = 0 ln
ln coth(|u| /2)
or 1
m

yos

Sfrag replacements
tos 0

,0 t
ke(t)k
0
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)
h (t)
hN,D (t)
1 (t)
1N,D (t)
(t)
(t)
fj (t)
Re
Im
x1
x2
x1
x2

deg

t
t
ke(t)k
ke(t)k
e1 (t)
e1 (t)
e2 (t)
e2 (t)
e3 (t)
e3 (t)
y(t)
y(t)
u(t)
u(t)
22/10/2004
4.3 Graphical representations
of the frequency response
131
h (t)
h (t)
hN,D (t)
hN,D (t)
1 (t)
1 (t)
40
40
1N,D (t)
1N,D (t)
(t)
(t)
30
30
(t)
(t)
20
20
fj (t)
fj (t)
10
10
0
0
Re
Re
Im
Im
-10
-10
x1
x1
-20
-20
x2
x2
-30
-30
x1
x1
-1.5 -1 -0.5 0 0.5 1 1.5 2 x
-1.5 -1 -0.5 0 0.5 1 1.5 2
x2
2
log
log

1. H1 () =

1
;
80

2. H2 () = 1 + i10;

,0
0

-3 -2 -1

log

Figure 4.9 Asymptotes for Example 4.7: H1 (left) and H2 (right)

2. H2 has a break frequency of =

in Figure 4.9.

1
10

or log = 1. The asymptotes for H2 are shown

3. H3 gives 20 log across the board. The asymptotes for H3 are shown in Figure 4.10.

4. H4 has a break frequency of = 8 or log 0.45. The asymptotes for H4 are shown
in Figure 4.10. Note that here we have = 12 , which is a largish value. Thus we do
not need to adjust the magnitude peak too much around the break frequency when we
use the asymptotes to approximate the actual Bode plot.
Now the phase angles.
1. H1 has phase exactly 0 for all frequencies.
2. For H2 , the phase is approximately 0 for log 10 < 1 or log < 2. For log 10 > 1
(or log > 0) the phase is approximately 90 . Between the frequencies log = 2
and log = 0 we interpolate linearly between the two asymptotic phase angles.
3. The phase for H3 is 90 for all frequencies.

4. For H4 , the phase is 0 for log 8 < 1 or log < log 8 1 0.55. For log 8 > 1,

(log > log 8 + 1 1.45), the phase is approximately 180 .

-100

150
100
50
0
-50
-100
-150

,0
e2 (t)
e3 (t)0
y(t)
u(t)
h (t)
hN,D (t)
1 (t)
1N,D (t)
50
(t)
(t)
0
fj (t)
Re
-50
Im
x1 -100
x2
x1
x2

u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

,0
0

u = 0 ln
ln coth(|u| /2)
or 1
m

PSfrag replacements
yos
t
ke(t)k
tos 0
-3 -2 -1 0
1
2
3
e41 (t)
log
e2,
(t)0

deg

-50

22/10/2004

150
100
50
0
-50
-100
-150
-3 -2 -1

log

e3 (t)
y(t)
Figure 4.10 Asymptotes for Example
u(t) 4.7: H3 (left) and H4 (right)
h (t)
hN,D (t)
1 (t)
1N,D (t)
50
(t)
(t)
0
fj (t)
Re
-50
Im
x1 -100
x2
x1
-3 -2 -1 0
1
2
3
4 x
-3 -2 -1 0
1
2
log
log
dB

log

150
100
50
0
-50
-100
-150

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)frequency domain)
4 Frequency response (the
h (t)
hN,D (t)
1 (t)
1N,D (t)
50
(t)
(t)
0
fj (t)
Re
-50
Im
x1 -100
x2
x1
-3 -2 -1 0
1
2
3
4 x
-3 -2 -1 0
2
log
log

150
100
50
0
-50
-100
-150
-3 -2 -1

log

u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0
2
3
4
,0
0

deg

-3 -2 -1

,0
0

u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0
2
3
4

u = 0 ln
ln coth(|u| /2)
or 1
m

PSfrag replacements y
os
t
ke(t)k
tos 0
e1 (t)

deg

150
100
50
0
-50
-100
-150

deg

1
( 8 )2 .

t
ke(t)k
e
(t)
1
H4 () = 1 + 2i 12 8
e2 (t)
e3 (t)
look first at the magnitudes.
y(t)
1
u(t) The asymptotes for H1 are shown in FigH1 will contribute 20 log 80
38.1dB.
h (t)
ure 4.9.
hN,D (t)
1 (t)
1N,D (t)
50
50
(t)
(t)
0
0
fj (t)
Re
-50
-50
Im
x
1
-100
-100
x2
x1
-3 -2 -1 0
1
2
3
4 x
-3 -2 -1 0
1
2
3
4
2
log
log

H3 () = (i)1 ;

deg

t
ke(t)k3.
e1 (t)
e2 (t)4.
e3 (t)
Lets
y(t)
u(t)1.
h (t)
hN,D (t)
1 (t)
1N,D (t)
(t)
(t)
fj (t)
Re
Im
x1
x2
x1
x2

u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

133

PSfrag replacements

deg

4.3 Graphical representations of the frequency response

22/10/2004

frag replacements

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)
134
h (t)
hN,D (t)
1 (t)
1N,D (t)
(t)
(t)
fj (t)
Re
Im
x1
x2
x1
x2

150
100
50
0
-50
-100
-150
-3 -2 -1

log

Figure 4.11 The sum of the asymptotes (left) and the actual Bode
plot (right) for Example 4.7

PSfrag replacements

135

x1
x2 1
x1
x2
log
0
dB
deg
PSfrag replacements
u = 0 ln
ln coth(|u| /2)t -1
ke(t)k
or
1
e1(t)
m
e2 (t)
-2
e3y(t)
os
y(t)
u(t)
tos
0
h (t)

,
hN,D
(t)0
0
1(t)
1N,D (t) 3
(t)
(t)
fj (t) 2
Im

The sum of the asymptotes are plotted in Figure 4.11 along with the actual Bode plot so you
can see how the Bode plot is essentially the sum of the individual Bode plots. You should
take care that you always account for , however. In our example, the value of in H4 is
quite large, so not much of an adjustment had to be made. If were small, wehave to add
a little bit of a peak in the magnitude around the break frequency log = log 8, and also
make the change in the phase a bit steeper.

4 Frequency response

4.10 Example (Example 4.12) Recall that we contrasted the two transfer functions
H1 () =

1 + i
,
2 + i + 1

H2 () =

1 i
.
2 + i + 1

We contrast the polar plots for these frequency responses in Figure 4.13. Note that, as
expected, the minimum phase system undergoes a smaller phase change if we follow it along
its parameterised polar curve. We shall see the potential dangers of this in Chapter 12.
Note that when making a polar plot, the thing one looses is frequency information. That
is, one can no longer read from the plot the frequency at which, say, the magnitude of the
frequency response is maximum. For this reason, it is not uncommon to place at intervals
along a polar plot the frequencies corresponding to various points.

4.4 Properties of the frequency response

It turns out that in the frequency response can be seen some of the behaviour we have
encountered in the time-domain and in the transfer function. We set out in this section to
scratch the surface behind interpreting the frequency response. However, this is almost an
art as much as a science, so plain experience counts for a lot here.
4.4.1 Time-domain behaviour reflected in the frequency response We have seen
in Section 3.2 we saw that some of the time-domain properties discussed in Section 2.3
were reflected in the transfer function. We anticipate being able to see these same features

,0
0

x1
x2 1
x1
x2
log
0
dB
deg
PSfrag replacements
u = 0 ln
ln coth(|u| /2)t -1
ke(t)k
or
1
e1(t)
m
e2 (t)
-2
e3y(t)
os
y(t)
u(t)
tos
0
h (t)

2
3,
hN,D
(t)0
0
1(t)
1N,D (t) 3
(t)
(t)
fj (t) 2

domain)

22/10/2004

-1

x1
x2 1
x1
x2
log
0
dB
deg
u = 0 ln
ln coth(|u| /2) -1
or 1
m
-2
yos

tos 0

-2

-1

-2

-1

x1
x2 1
x1
x2
log
0
dB
deg
u = 0 ln
ln coth(|u| /2) -1
or 1
m
-2
yos

tos 0

-2

4.3.3 The polar frequency response plot We will encounter in Chapter 12 another
representation of the frequency response H. The idea here is that rather than plotting
magnitude and phase as one does in a Bode plot, one plots the real and imaginary part of
the frequency response as a curve in the complex plane parameterised by (0, ). Doing
this yields the polar plot for the frequency response. One could do this, for example, by
taking the Bode plot, and for each point on the independent variable axis, put a point at
a distance |H()| from the origin in the direction ]H(). Indeed, given the Bode plot, one
can typically make a pretty good approximation of the polar plot by noting (1) the maxima
and minima of the magnitude response, and the phase at these maxima and minima, and
(2) the magnitude when the phase is 0, 90 , or 180 .
In Figure 4.12 are shown the polar plots for the basic frequency response functions.
Recall that in (4.3) we indicated that any frequency response will be a product of these
basic elements, and so one can determine the polar plot for a frequency response formed by
the product of such elements by performing complex multiplication that, you will recall, is
done in polar coordinates merely by multiplying radii, and adding angles.
For a lark, lets look at the minimum/nonminimum phase example in polar form.

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)
(the
h (t)frequency
hN,D (t)
1 (t)
1N,D (t) 3
(t)
(t)
fj (t) 2

4.4 Properties of the frequency response

22/10/2004

PSfrag replacements

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)
h136
(t)
hN,D (t)
1 (t)
1N,D (t)
(t)
(t)
fj (t)

-2

-1

3,0
0

Figure 4.12 Polar plots for H() = 1 + i (top left), H() =

1 + 2i 2 , = 0.2, 0.4, 0.6, 0.8 (top right), H() =
(1 + i)1 (bottom left), and H() = (1 + 2i 2 )1 ,
= 0.2, 0.4, 0.6, 0.8

reflected in the frequency response, and ergo in the Bode plot. In this section we explore
these expected relationships. We do this by looking at some examples.
4.11 Example The first example we look at is one where we have a pole/zero cancellation.
As per Theorem 3.5 this indicates a lack of observability in the system. It is most beneficial
to look at what happens when the pole and zero do not actually cancel, and compare it to
what happens when the pole and zero really do cancel. We take

0 1
0
1
A=
, b=
, c=
,
(4.4)
1
1
1
and we compute
H () =

1 + i
.
2 + i 1

-1

-0.5

0.5

1.5,0

-1

-0.5

0.5

1.5

,0
0

Figure 4.13 Polar plots for minimum phase (left) and nonminimum phase (right) systems

The Bode plots for three values of are shown in Figure 4.14. What are the essential features
here? Well, by choosing the values of 6= 0 to deviate significantly from zero, we can see
accentuated two essential points. Firstly, when 6= 0 the magnitude plot has two regions
where the magnitude drops off at different slopes. This is a consequence of there being two
different exponents in the characteristic polynomial. When = 0 the plot tails off at one
slope, indicating that the system is first-order and has only one characteristic exponent. This
is a consequence of the pole/zero cancellation that occurs when = 0. Note, however, that
we cannot look at one Bode plot and ascertain whether or not the system is observable.
There is also an effect that can be observed in the Bode plot for a system that is not
minimum phase. An example illustrates this well.

-20
-30
-40

150
100
50
0
-50
-100
-150

PSfrag replacements

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)
h (t) 0
-1.5 -1 -0.5
hN,Dlog
(t)
1 (t)
1N,D (t)
(t)
(t)
fj (t)
Re
Im
x1
x2
x1
x2

u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

,0
0

4.12 Example We consider two SISO linear systems, both with

0
1
0
A=
, b=
.
1 1
1

c2 =

1.5

150

u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

1
1.5
2

100
50
0
-50
-100
-150
-1.5 -1 -0.5

log

,0
0

0.5

1.5

0
-10
-20
-30
-40

-1.5 -1 -0.5

log

0.5

1.5

150
100
50
0
-50
-100
-150
-1.5 -1 -0.5

log

0.5

1.5

Figure 4.14 Bode plots for (4.4) for = 5, = 0, and = 5

1
.
1

0.5

(4.5)

The two output vectors we look at are

1
c1 =
,
1

22/10/2004

0
-10

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)
4 Frequency response (the
h (t)frequency domain)
hN,D (t)
1 (t)
0
1N,D (t)
(t)
(t)
-10
fj (t)
-20
Re
Im
x1
-30
x2
x1
-40
-1.5 -1 -0.5 0 0.5 1
1.5
2
-1.5 -1 -0.5 0
x2
log
log

deg

u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

PSfrag replacements

137

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)
h138
(t)
hN,D (t)
1 (t)
1N,D (t)
(t)
(t)
fj (t)
Re
Im
x1
x2
x1
x2

deg

,0
0

response

0.5

x1
x2
0
x1
x2
log
dB -0.5
deg
u = 0 ln
ln coth(|u| /2)
-1
or 1
m

yos -1.5

tos 0

0.5

x1
x2
0
x1
x2
log
dB -0.5
deg
u = 0 ln
ln coth(|u| /2)
-1
or 1
m

yos -1.5

tos 0

PSfrag replacements

4.4

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)
h (t)
hN,Dof
(t) the frequency
Properties
1 (t)
1N,D (t)
(t)
1
(t)
fj (t)

deg

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)
h (t)
hN,D
(t)
22/10/2004
1 (t)
1N,D (t)
(t)
1
(t)
fj (t)

(4.6)

Let us then denote 1 = (A, b, c1 , 01 ) and 2 = (A, b, c2 , 01 ). The two frequency response
functions are
1 + i
1 i
H1 () =
, H2 () =
.
2 + i + 1
2 + i + 1
The Bode plots are shown in Figure 4.15. What should one observe here? Note that the
magnitude plots are the same, and this can be verified by looking at the expressions for H1
and H2 . The differences occur in the phase plots. Note that the phase angle varies only
slightly for 1 across the frequency range, but it varies more radically for 2 . It is from this
behaviour that the term minimum phase is derived.

4.4.2 Bodes Gain/Phase Theorem In Bodes book [1945] one can find a few chapters
on some properties of the frequency response. In this section we begin this development, and
from it derive Bodes famous Gain/Phase Theorem. The material in this section relies on
some ideas from complex function theory
that we review in Appendix D. We start by examining some basic properties of frequency
response functions. Here we begin to see that the real and imaginary parts of a frequency
response function are not arbitrary functions.
4.13 Proposition Let (N, D) be a SISO linear system in input/output form with HN,D the
frequency response. The following statements hold:
(i) Re(HN,D ()) = Re(HN,D ());
(ii) Im(HN,D ()) = Im(HN,D ());
(iii) |HN,D ()| = |HN,D ()|;

PSfrag replacements

log

0.5

100
50
0
-50
-100
-150

,0
0

-1.5 -1 -0.5

log

0.5

u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

1
1.5
2

139

,0
0

0.5

1.5

22/10/2004

noting that SN,D is analytic on C \ ZPN,D . We recall that from the properties of the complex
logarithm we have

100
50

Re(SN,D (s)) = ln |TN,D (s)| ,

0
-50

Im(SN,D (s)) = ]TN,D (s)

for s C \ ZPN,D . Along similar lines we define ZPN,D R by

-100
-150
-1.5 -1 -0.5

log

0.5

1.5

ZPN,D = { R | i ZPN,D } .

Then we may define GN,D : R \ ZPN,D C by

GN,D () = SN,D (i) = ln(HN,D ()).

Re(GN,D ()) = ln |HN,D ()| ,

Proof We prove (i) and (ii) together. It is certainly true that the real parts of D(i)
and N (i) will involve terms that have even powers of and that the imaginary parts of
D(i) and N (i) will involve terms that have odd powers of . Therefore, if we denote by
D1 () and D2 () the real and imaginary parts of D(i) and N1 () and N2 () the real and
imaginary parts of N (i), we have
N2 () = N2 (),

D2 () = D2 (). (4.7)

We also have
N1 () + iN2 ()
D1 () + iD2 ()
N1 ()D1 () + N2 ()D2 ()
N2 ()D1 () N1 ()D2 ()
=
+i
,
D12 () + D22 ()
D12 () + D22 ()

HN,D () =

Im(GN,D ()) = ]HN,D ().

Note that these are almost the quantities one plots in the Bode plot. The phase is precisely
what is plotted in the Bode plot (against a logarithmic frequency scale), and Re(GN,D ())
is related to the magnitude plotted in the Bode plot by
Re(GN,D ()) = ln 10 log |HN,D ()| =

ln 10
|HN,D ()| dB.
20

Note that the relationship is a simple one as it only involves scaling by a constant factor.
The quantity Re(GN,D ()) is measured in the charming units of neppers.
Recall that (N, D) is stable if all roots of D lie in C and is minimum phase if all roots
of N lie in C . Here we will require the additional assumption that N have no roots on the
imaginary axis. Let us say that a SISO linear system (N, D) for which all roots of N lie in
C is strictly minimum phase.
4.14 Proposition Let (N, D) be a proper SISO linear system in input/output form that is
stable and strictly minimum phase, and let 0 > 0. We then have
Z
20 Re(GN,D ()) Re(GN,D (0 ))
Im(GN,D (0 )) =
d.
0
2 02

so that
N1 ()D1 () + N2 ()D2 ()
,
D12 () + D22 ()
N2 ()D1 () N1 ()D2 ()
Im(HN,D ()) =
.
D12 () + D22 ()
Re(HN,D ()) =

Proof Throughout the proof we denote by G1 () the real part of GN,D () and by G2 ()
the imaginary part of GN,D ().
Let U0 C be the open subset defined by

Using the relations (4.7) we see that

N1 ()D1 () + N2 ()D2 () = N1 ()D1 () + N2 ()D2 (),
N2 ()D1 () N1 ()D2 () = N2 ()D1 () N1 ()D2 (),
and from this the assertions (i) and (ii) obviously follow.
(iii) This is a consequence of (i) and (ii) and the definition of ||.
(iv) This follows from (i) and (ii) and the properties of arctan.

(4.8)

We can employ our previous use of the properties of the complex logarithm to assert that

(iv) ]HN,D () = ]HN,D () provided ]HN,D () (, ).

D1 () = D1 (),

4 Frequency response (the frequency domain)

SN,D (s) = ln(TN,D (s)),

Figure 4.15 Bode plots for the two systems of (4.5) and (4.6)H1
is on the left and H2 is on the right

N1 () = N1 (),

140

Now we turn our attention to the crux of the material in this sectiona look at how the
magnitude and phase of the frequency response are related. To relate these quantities, it is
necessary to represent the frequency response in the proper manner. To this end, for a SISO
linear system (N, D) in input/output form, we define ZPN,D C to be the set of zeros and
poles of TN,D . We may then define SN,D : C \ ZPN,D C by

150

deg

-1.5 -1 -0.5

150

deg

u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
Propertieshu(t)
of the frequency response
(t)
hN,D (t)
1 (t)
10
1N,D (t)
(t)
0
(t)
fj (t)
-10
Re
-20
Im
x1
-30
x2
x1
1 1.5 2
-1.5 -1 -0.5 0
x2
log
dB

4.4

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)
h22/10/2004
(t)
hN,D (t)
1 (t)
10
1N,D (t)
(t)
0
(t)
fj (t)
-10
Re
-20
Im
x1
-30
x2
x1
x2

U0 = { s C | Re(s) > 0, Im(s) 6= 0 } .

We now define a closed contour whose interior contains points in U0 . We do this in parts.
First, for R > 0 define a contour R in U0 by

R = {Rei |

2 }.

22/10/2004

4.4 Properties of the frequency response

141

Now for r > 0 define two contours

r,1 = {i0 + rei |

142

4 Frequency response (the frequency domain)

Indeed we have

2 },

r,2 = {i0 + rei |

F0 (Rei ) =

2 }.

0 ,3 = { i | < 0 0 r}
0 ,4 = { i | 0 + r 0 0 r}
0 ,5 = { i | + r 0 < } .

F0 (s) =

The closed contour we take is then

5
[

2i0 SN,D (Rei )

2i0 G1 (0 )

.
R2 e2i + 02
R2 e2i + 02

Since deg(N ) < deg(D) the first term on the right will behave like R2 as R and the
second
term will also behave like R2 as R . Since ds = iRei on R , the integrand in
R
F (s) ds will behave like R1 as R , and so our claim follows.
R 0
Now let us examine what happens as we let r 0. To evaluate the contributions of r,1
and r,2 we write

Finally define a contours r,j , j = 3, 4, 5, by

R,r = R

22/10/2004

SN,D (s) G1 (0 ) SN,D (s) G1 (0 )

.
s i0
s + i0

On r,1 we have s = i0 + rei so that on r,1 we have

r,j .
F0 (i0 + rei ) =

j=1

We show this contour in Figure 4.16. Now define a function F0 : U0 C by

Im
R

SN,D (i0 + rei ) G1 (0 ) SN,D (i0 + rei ) G1 (0 )

.
rei
2i0 + rei

We parameterise r,1 with the curve c : [ 2 , 2 ] C defined by c(t) = i0 + reit so that

c0 (t) = ireit . Thus, as r 0, F0 (s) ds behaves like
(SN,D (i0 ) G1 (0 ))ireit
= i(GN,D (0 ) G1 (0 )),
reit
using the parameterisation specified by c. Integrating gives
Z
F0 (s) ds = i(GN,D (0 ) G1 (0 )).
F0 (s) ds

r,5

r,4

r,1

r,1
Re

In similar fashion one obtains

Z
F0 (s) ds = i(GN,D (0 ) G1 (0 )).
r,2

r,2

Now we use Proposition 4.13 to assert that

GN,D (0 ) GN,D (0 ) = 2iG2 (0 ).

r,3

Therefore
Z
Z
F0 (s) ds +
r,1

Figure 4.16 The contour R,r used in the proof of Proposition 4.14

F0 (s) =

2i0 (SN,D (s) G1 (0 ))

.
s2 + 02

F0 (s) ds = i(GN,D (0 ) G1 (0 ) (GN,D (0 ) G1 (0 )))

r,2

= 2G2 (0 ).
Finally, we look at the integrals along the contours r,3 , r,4 , and r,5 as r 0 and for
fixed R > 0 . These contour integrals in the limit will yield a single integral along the a
portion of the imaginary axis:
Z
F0 (s) ds.
i[R,R]

Since (N, D) is stable and strictly minimum phase, SN,D is analytic on C+ , and so F0 is
analytic on U0 , and so we may apply Cauchys Integral Theorem to the integral of F0
around the closed contour R,r .
Let us evaluate that part of this contour integral corresponding to the contour R as R
get increasingly large. We claim that
Z
lim
F0 (s) ds = 0.
R

We can parameterise the contour in this integral by the curve c : [R, R] C defined by
c(t) = it. Thus c0 (t) = i, giving
Z
Z R
2i0 (SN,D (it) G1 (0 ))
F0 (s) ds =
i dt
02 t2
i[R,R]
R
Z R
20 (SN,D (it) G1 (0 ))
=
dt.
t2 02
R

143

Using Proposition 4.13 we can write this as

Z
Z R
40 (G1 (t) G1 (0 ))
F0 (s) ds =
dt.
t2 02
0
i[R,R]
Collecting this with our expression for the integrals along r,1 and r,2 , as well as noting
our claim that the integral along R vanishes as R , we have shown that
Z
Z
40 (G1 () G1 (0 ))
lim lim
F0 (s) ds =
d 2G2 (0 ).
r0 R
2 02
0
R,r
By Cauchys Integral Theorem, this integral should be zero, and our result now follows from
straightforward manipulation.

The following result gives an important property of stable strictly minimum phase systems.
4.15 Theorem (Bodes Gain/Phase Theorem) Let (N, D) be a proper, stable, strictly minimum phase SISO linear system in input/output form, and let GN,D () be as defined in (4.8).
0
0
For 0 > 0 define MN,D
: R R by MN,D
(u) = Re(GN,D (0 eU )). Then we have
]HN,D (0 ) =

0

dMN,D
ln coth u2 du.
du

Proof The theorem follows fairly easily from Proposition 4.14. By that result we have
Z
20 Re(GN,D ()) Re(GN,D (0 ))
d.
(4.9)
]HN,D (0 ) =
0
2 02
We make the change of variable = 0 eu , upon which the integral in (4.9) becomes
2

We note that

0
0
MN,D
(u) MN,D
(0)
1
du =
eu eu

0
0
MN,D
(u) MN,D
(0)
du.
sinh u

du
= ln coth u2 ,
sinh u
and we may use this formula, combined with integration by parts, to determine that
1

0
0
MN,D
(u) MN,D
(0)
du =
sinh u
Z
0

dMN,D
1
1

0
0
ln coth u2 (MN,D
(u) MN,D
(0)) ln coth u2 , (4.10)
0
du

and
1

0
0
MN,D
(u) MN,D
(0)
du =
sinh u
Z 0
0

dMN,D
1
1

0
0
ln coth u
+ (MN,D
(u) MN,D
(0)) ln coth u
. (4.11)
2
2
du

144

4 Frequency response (the frequency domain)

22/10/2004

Let us look at the first term in each of these integrals. At the limit u = 0 we may
compute
u3
2 u
+
coth u2 +
u 6 360
0
0
u
u
so that near u = 0, ln coth 2 ln 2 . Also, since MN,D (u) is analytic at u = 0, MN,D
(u)
0
MN,D (0) will behave linearly in u for u near zero. Therefore,
u
0
0
lim MN,D
(u) MN,D
(0)) ln coth u2 u ln .
u0
2
t
ke(t)k
Recalling that
lim
u
ln
u
=
0,
we
see
that
the
lower
limit
in
the above integrated exe1 (t) u0
pressions is zero.
e2 (t) At the other limits as u , ln coth u2 behaves like eu as u +
0
and like eu ase3 (t)
u . This, combined with the fact that MN,D
(u) behaves like ln u as
y(t)
u , implies the vanishing of the upper limits in the integrated expressions in both (4.10)
u(t)
and (4.11). hThus
we have shown that these integrated terms both vanish. From this the
(t)
result follows
easily.

hN,D (t)
1
(t)
It is not perhaps perfectly clear what is the import of this theorem, so let us examine it
1N,D (t)
for a moment.
In this discussion, let us fix 0 > 0. Bodes Gain/Phase Theorem is telling
(t)
us that the phase
(t) angle for the frequency response can be determined from the slope of the
Bode plot with
fj (t)decibels plotted against the logarithm of frequency. The contribution to
u
the phase of the
Reslope at some u = 0 ln is determined by the weighting factor ln coth 2
ImFigure 4.17. From the figure we see that the slopes near u = 0, or = 0 ,
which we plot in
x1
x2 6
x1
x2
log 5
dB
deg 4
PSfrag replacements

ln coth(|u| /2)

4.4 Properties of the frequency response

22/10/2004

or 1
m 2

yos
1
tos 0

,0
0

-1.5

-1

-0.5

0.5

u = 0 ln

1.5

Figure 4.17 The weighting factor in Bodes Gain/Phase Theorem

contribute most to the phase angle. But keep in mind that this only works for stable strictly
minimum phase systems. What it tells us is that for such systems, if one wishes to specify
a certain magnitude characteristic for the frequency response, the phase characteristic is
completely determined.
Lets see how this works in an example.
4.16 Example Suppose we have a Bode plot with y = |HN,D ()| dB versus x = log , and
that y = 20kx+b for k Z and b R. Thus the magnitude portion of the Bode plot is linear.

4.5 Uncertainly in system models

22/10/2004

145

146

To employ the Gain/Phase Theorem we should convert this to a relation in y = ln |HN,D ()|
versus x = ln 0 . The coordinates are then readily seen to be related by
x=

x
+ log 0 ,
ln 10

20
y.
ln 10

b ln 10
y = k
x + k ln 0 +
.
20

This will be investigated rather more systematically in Section 5.3.2, but for now the meaning
is rather pedestrian: it is the maximum value of the magnitude Bode plot.

In the terminology of Theorem 4.15 we thus have

4.5.1 Structured and unstructured uncertainty The reader may wish to recall our
general control theoretic terminology from Section 1.1. In particular, recall that a plant is
that part of a control system that is given to the control designer. What is given to the control
designer is a model, hopefully in something vaguely resembling one of the three forms we have
thus far discussed: a state-space model, a transfer function, of a frequency response function.
Of course, this model cannot be expected to be perfect. If one is uncertain about a plant
model, one should make an attempt to come up with a mathematical description of what
this means. There are many possible candidates, and they can essentially be dichotomised
as structured uncertainty and unstructured uncertainty . The idea with structured
uncertainty is that one has a specific type of plant model in mind, and parameters in that
plant model are regarded as uncertain. In this approach, one wishes to design a controller
that has desired properties for all possible values of the uncertain parameters.

b ln 10
= ku + k ln 0 +
20

so that duN,D = k. The Gain/Phase Theorem tells us that we may obtain the phase at any
frequency 0 as
Z

k
]HN,D (0 ) =
ln coth u2 du.

2

The integral is one that can be looked up (Mathematica evaluated it for me) to be 2 , so
that we have ]HN,D (0 ) = k
.
2
Lets see if this agrees with what we have seen before. We know, after all, a transfer
function whose magnitude Bode plot is 20k log . Indeed, one can check that choosing
the transfer function TN,D (s) = sk gives 20 log |HN,D ()| = 20k log , and so this transfer
function is of the type for which we are considering in the case when b = 0. For systems of
, which agrees with Bodes Gain Phase
this type we can readily determine the phase to be k
2
Theorem.

4.17 Example Suppose a mass is moving under a the influence of a control force u. The
precise value of the mass could be unknown, say m [m1 , m2 ]. In this case, a control design
should be thought of as being successful if it accomplishes stated objectives (the reader does
not know what these might be at this point in the book!) for all possible values of the mass
parameter m.

Despite the fact that it might be possible to explicitly apply Bodes Gain/Phase Theorem in some examples to directly compute the phase characteristic for a certain magnitude
characteristic, the value of the theorem lies not in the ability to do such computations, but
rather in its capacity to give us approximate information for stable and strictly minimum
phase systems. Indeed, it is sometimes useful to make the approximation
]HN,D (0 ) =

22/10/2004

design that have been developed up to this time rely on such a description. In the context
of this book, this culminates in Chapter 15 with a systematic design methodology keeping
robustness concerns foremost.
In this section it is helpful to introduce the H -norm for a rational function. Given
R R(s) we denote
kRk = sup{|R(i)|}.

Therefore the relation y = 20kx + b becomes

0
MN,D
(u)

4 Frequency response (the frequency domain)

0
dMN,D

,

du u=0 2

Typically, structured uncertainty can be expected to be handled on a case by case basis,

depending on the nature of the uncertainty. An approach to structured uncertainty is put
forward by Doyle [1982]. For unstructured uncertainty, the situation is typically different as
P in
one considers a set of plant transfer functions P that are close to a nominal plant R
some way. Again the objective is to design a controller that works for every plant in the set
of allowed plants.

(4.12)

and this approximation becomes better when one is in a region where the slope of the
magnitude characteristic on the Bode plot is large at 0 compared to the slope at other
frequencies.

4.5 Uncertainly in system models

As hinted at in Section 1.2 in the context of the simple DC servo motor example, robustness of a design to uncertainties in the model is a desirable feature. In recent years,
say the last twenty years, rigorous mathematical techniques have been developed to handle
model uncertainty, these going under the name of robust control. These matters will be
touched upon in this book, and in this section, we look at the first aspect of this: representing
uncertainty in system models. The reader will observe that this uncertainty representation
is done in the frequency domain. The reason for this is merely that the tools for controller

Uses of Gain/Phase
theorem

4.18 Example (Example 4.17 contd) Let us look at the mass problem above in a different
P (s) = 1/m
and define a set of
manner. Let us suppose that we choose a nominal plant R
s2
candidate plants

P
P = RP R(s) | RP R
(4.13)

for some > 0. In this case, we have clearly allowed a much larger class of uncertainty that
was allowed in Example 4.17. Indeed, not only is the mass no longer uncertain, even the
form of the transfer function is uncertain.

Thus we see that unstructured uncertainty generally forces us to consider a larger class
of plants, and so is a more stringent and, therefore, conservative manner for modelling
uncertainty. That is to say, by designing a controller that will work for all plants in P, we
are designing a controller for plants that are almost certainly not valid models for the plant

4.5 Uncertainly in system models

22/10/2004

147

22/10/2004

deg
0
u = 0 ln
ln coth(|u| /2)
-20
or 1
m
-40
yos
-60
tos 0

under consideration. Nevertheless, it turns out that this conservatism of design is made up
for by the admission of a consistent design methodology that goes along with unstructured
uncertainty. We shall now turn our attention to describing how unstructured uncertainty
may arise in examples.

148

(t)
fj (t)
Re
Im
x1
x2
x1 4 Frequency response (the frequency domain)
x2

4.5.2 Unstructured uncertainty models We shall consider four unstructured uncertainty models, although others are certainly possible. Of the four types of uncertainty we
present, only two will be treated in detail. A general account of uncertainty models is
the subject of Chapter 8 in [Dullerud and Paganini 1999]. Consistent with our keep our
treatment of robust control to a tolerable level of simplicity, we shall only look at rather
straightforward types of uncertainty.
The first type we consider is called multiplicative uncertainty . We start with a
P R(s) and let Wu R(s) be a proper rational function with no poles in
nominal plant R
P , Wu ) the set of rational functions RP R(s) with the properties
C+ . Denote by P (R

1. RP and RP have the same number of poles in C+ ,

P have the same imaginary axis poles, and
2. RP and R
R (i)

P

3.
1 |Wu (i)| for all R.
RP (i)
P , Wu ) if and only if
Another way to write this set of plants is to note that RP P (R
P
RP = (1 + Wu )R
where kk 1. Of course, is not arbitrary, even given kk 1. Indeed, this condition
only ensures condition 3 above. If further satisfies the first two conditions, it is said to be
allowable. In this representation, it is perhaps more clear where the term multiplicative
uncertainty comes up.
The following example is often used as one where multiplicative uncertainty is appropriate.
4.19 Example Recall from Exercise EE.6 that the transfer function for the time delay of a
function g by T is eT s g(s). Let us suppose that we have a plant transfer function
RP (s) = eT s R(s)
for some R(s) R(s). We wish to ensure that this plant is modelled by multiplicative
uncertainty. To do so, we note that the first two terms in the Taylor series for eT s are
1 T s, and thus we suppose that when T is small, a nominal plant of the form

In this case, the condition on Wu becomes

e0.1i

|Wu (i)|
s 1,
1 10

0.5

log

1.5

e i

10

Figure 4.18 1
s 1 (solid) and Wu =
10

1
100

2.5

(1 15 )

(dashed)

From Figure 4.18 (the solid curve) one can see that the magnitude of
e i

10

s 1
1 10
has the rough behaviour of tailing off at 40dB/decade at low frequency and having constant
magnitude at high frequency. Thus a model of the form
Wu (s) =

Ks2
s + 1

is a likely candidate, although other possibilities will work as well, as long as they capture
1
1
and K = 100
as
the essential features. Some fiddling with the Bode plot yields = 15
acceptable choices. Thus this choice of Wu will include the time delay plant in its set of
plants.

The next type of uncertainty we consider is called additive uncertainty . Again we
P and a proper Wu R(s) having no poles in C+ . Denote by
start with a nominal plant R
P , Wu ) the set of rational functions RP R(s) with the properties
P (R
P have the same number of poles in C+ ,
1. RP and R
P have the same imaginary axis poles, and
2. RP and R
P (i)| |Wu (i)| for all R.
3. |RP (i) R

P + Wu
RP = R

will do the job. Thus we are charged with finding a rational function Wu so that
R (i)

P

1 |Wu (i)| , R.

R(i)
1
.
10

-0.5

P , Wu ) will have the form

A plant in P+ (R

P (s) = R(s)
R
1 Ts

Let us suppose that T =

,0
0

where kk 1. As with multiplicative uncertainty, will be allowable if properties 1

and 2 above are met.
4.20 Example This example has a generic flavour. Suppose that we make measurements
of our plant to get magnitude information about its frequency response at a finite set of
frequencies {1 , . . . , k }. If the test is repeated at each frequency a number of times, we
P with the property that at the measured
might try to find a nominal plant transfer function R

22/10/2004

4.5 Uncertainly in system models

149

150

frequencies its magnitude is roughly the average of the measured magnitudes. One then can
determine Wu so that it covers the spread in the data at each of the measured frequencies.
Such a Wu should have the property, by definition, that
P (ij )| |Wu (ij )| ,
|RP (ij ) R

j = 1, . . . , k.

P , Wu ) and P+ (R
P , Wu ) are the
The above two classes of uncertainty models, P (R
two for which analysis will be carried out in this book. The main reason for this choice
is convenience of analysis. Fortunately, many interesting cases of plant uncertainty can be
modelled as multiplicative or additive uncertainty. However, for completeness, we shall give
two other types of uncertainty representations.
The first situation we consider where the set of plants are related to the nominal plant
as
P
R
, kk 1.
(4.14)
RP =
1 + Wu RP
This uncertainty representation can arise in practice.

4.6 Summary

4.21 Example Suppose that we have a plant of the form

1
s2 + as + 1

where all we know is that a [amin , amax ]. By taking

P (s) =
R

1
,
s2 + aavg s + 1

Wu (s) = 12 as,

where aavg = 12 (amax + amin ) and a = amax amin , then the set of plant is exactly as
in (4.14), if is a number between 1 and 1. If is allowed to be a rational function
satisfying kk 1 then we have embedded our actual set of plants inside the larger
uncertainty set described by (4.14). One can view this example of one where the structured
uncertainty is included as part of an unstructured uncertainty model.

The final type of uncertainty model we present allows plants that are related to the
nominal plant as
P
R
RP =
, kk 1.
1 + Wu
Let us see how this type of uncertainty can come up in an example.
finish

4.22 Example

P , Wu ) and P+ (R
P , Wu ) we made some assump4.23 Remarks 1. In our definitions of P (R
tions about the poles of the nominal plant and the poles of the plants in the uncertainty
set. The reason for these assumptions is not evident at this time, and indeed they can
be relaxed with the admission of additional complexity of the results of Section 7.3, and
those results that depend on these results. In practice, however, these assumptions are
not inconvenient to account for.

22/10/2004

2. All of our choices for uncertainty modelling share a common defect. They allow plants
that will almost definitely not be possible models for our actual plant. That is to say,
our sets of plants are very large. This has something of a drawback in our employment
of these uncertainty models for controller designthey will lead to a too conservative
design. However, this is mitigated by the existence of effective analysis tools to do robust
controller design for such uncertainty models.
3. When deciding on a rational function Wu with which to model plant uncertainty with
one of the above schemes, one typically will not want Wu to tend to zero as s . The
reason for this is that at higher frequencies is typically where model uncertainty is the
greatest. This becomes a factor when choosing starting point for Wu .

One could then hope that at the frequencies where data was not taken, the actual plant data
P , Wu ).
is in the data of the set P+ (R

RP (s) =

4 Frequency response (the frequency domain)

In this section we have introduced a nice piece of equipmentthe frequency response

functionand a pair of slick representations of itthe Bode plot and the polar plot. Here
are the pertinent things you should know from this chapter.
1. For a given SISO linear system , you should be able to compute and H .
2. The interpretation of the frequency response given in Theorem 4.1 is fundamental in
understanding just what it is that the frequency response means.
3. The complete equivalence of the impulse response, the transfer function, and the frequency response is important to realise. One should understand that something that one
observes in one of these must also be reflected somehow in the other two.
4. The Bode plot as a representation of the frequency response is extremely important.
Being able to look at it and understand things about the system in question is part of
the art of control engineering, although there is certainly a lot of science in this too.
5. You should be able to draw by hand the Bode plot for any system whose numerator
and denominator polynomials you can factor. Really the only subtle thing here is the
dependence on for the second-order components of the frequency response. If you think
of as damping, the interpretation here becomes straightforward since one expects larger
magnitudes for lower damping when the second-order term is in the denominator.
6. The polar plot as a representation of the frequency response will be useful to us in
Chapter 12. You should at least be able to sketch a polar plot given the corresponding
Bode plot.
7. You should be aware of why minimum phase system have the name they do, and be able
to identity these in obvious cases.
8. The developments of Section 4.4.2 are somehow essential, and at the same time somewhat
hard. If one is to engage in controller design using frequency response, clearly the fact that
there are essential restrictions on how the frequency response may behave is important.
9. It might be helpful on occasion to apply the approximation (4.12).

frag replacements

PSfrag replacements

,0
0

u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0
-1.5 -1 -0.5 0 0.5 1 1.5 2
,0
log

deg

ln coth(|u| /2)t
ke(t)k
or
1
e1(t)
m
e2 (t)

e3y(t)
os
y(t)
u(t)
tos
0
-1.5 -1 -0.5 0 0.5 1 1.5 h2 (t)

,0
log
hN,D(t)
0
1 (t)
1N,D (t)
(t)
(t)
fj (t)
Re
Im
x1
x2
x1
-1.5 -1 -0.5 0 0.5 1 1.5 2 x
2
log

150
100
50
0
-50
-100
-150
-1.5 -1 -0.5 0 0.5 1 1.5 2

log

40
30
20
10
0
-10
-20
-30

152

4 Frequency response (the frequency domain)

22/10/2004

Exercises
E4.1 Consider the vector initial value problem

x(t)
= Ax(t) + u0 sin tb,

x(0) = x0

where there are no eigenvalues for A of the form i

where
integrally divides . Let
xp (t) be the unique periodic solution constructed in the proof of Theorem 4.1, and
write the solution to the initial value problem as x(t) = xp (t) + xh (t). Determine an
expression for xh (t). Check that xh (t) is a solution to the homogeneous equation.
E4.2 Consider the SISO linear system = (A, b, ct , 01 ) with

0 0
0
1
A=
, b=
, c=
0 0
1
0
for 0 R and 0 > 0.
(a) Determine and compute H .
(b) Take u(t) = u0 sin t, and determine when the system satisfies the hypotheses of
Theorem 4.1, and determine the unique periodic output yp (t) guaranteed by that
theorem.
(c) Do periodic solutions exist when the hypotheses of Theorem 4.1 are not satisfied?
(d) Plot the Bode plot for H for various values of 0 0 and 0 > 0. Make sure
you get all cases where the Bode plot assumes a different character.
We have two essentially differing characterisations of the frequency response, one as the
way in which sinusoidal outputs appear under sinusoidal inputs (cf. Theorem 4.1) and one
involving Laplace and Fourier transforms (cf. Proposition 4.3). In the next exercise you will
explore the differing domains of validity for the two interpretations.

150
100
50
0
-50
-100
-150

PSfrag replacements
u = 0 ln

-1.5 -1 -0.5 0 0.5 1 1.5 2

log

deg

u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

150
100
50
0
-50
-100
-150

40
30
20
10
0
-10
-20
-30

deg

ln coth(|u| /2)t
ke(t)k
or
1
e1(t)
m
e2 (t)

e3y(t)
os
y(t)
u(t)
tos
0
h (t)

,0
hN,D(t)
0
1 (t)
1N,D (t)
(t)
(t)
fj (t)
Re
Im
x1
x2
x1
x2

deg

frag replacements
u = 0 ln

t
t
ke(t)k
ke(t)k
e1 (t)
e1 (t)
e2 (t)
e2 (t)
e22/10/2004
e4.6
151
3 (t)
3 (t)Summary
y(t)
y(t)
u(t)
u(t)
h (t)
h (t)
hN,D (t)
hN,D (t)
1 (t)
1 (t)
40
40
1N,D (t)
1N,D (t)
(t)
(t)
30
30
(t)
(t)
20
20
fj (t)
fj (t)
10
10
0
0
Re
Re
Im
Im
-10
-10
x1
x
1
-20
-20
x2
x2
-30
-30
x1
x1
-1.5 -1 -0.5 0 0.5 1 1.5 2 x
-1.5 -1 -0.5 0 0.5 1 1.5 2
x2
2
log
log

150
100
50
0
-50
-100
-150
-1.5 -1 -0.5 0 0.5 1 1.5 2

log

Figure 4.19 Asymptotes for magnitude and phase plots for

H() = 1 + i (top left), H() = 1 + 2i 2 (top right),
H() = (1+i)1 (bottom left), and H() = (1+2i 2 )1
(bottom right)

E4.3 Let = (A, b, ct , D) be complete. Show that the Fourier transform of h+

is defined if
and only if all eigenvalues of A lie in C . Does the characterisation of the frequency
response provided in Theorem 4.1 share this restriction?
E4.4 For the SISO linear system (N (s), D(s)) = (1, s + 1) in input/output form, verify
explicitly the following statements, stating hypotheses on arguments of the functions
where necessary.
(a) TN,D is the Laplace transform of hN,D .
(b) hN,D is the inverse Laplace transform of TN,D .
(c) HN,D is the Fourier transform of hN,D .
(d) hN,D is the inverse Fourier transform of HN,D .
Z
1
HN,D ()
(e) TN,D (s) =
d.
2 s i
(f) HN,D () = TN,D (i).
When discussing the impulse response, it was declared that to obtain the output for an
arbitrary input, one can use a convolution with the impulse response (plus a bit that depends
upon initial conditions). In the next exercise, you will come to an understanding of this in
terms of Fourier transforms.

Exercises for Chapter 4

153

154

E4.5 Suppose that f and g are functions of , and that they are the Fourier transforms
of functions f, g : (, ) R. You know from your course on transforms that the
inverse Fourier transform of the product fg is the convolution
Z
(f g)(t) =
f (t )g( ) d.

E4.7

Now consider a SISO linear system = (A, b, ct , 01 ) with causal impulse response
h+
. Recall that the output for zero state initial condition corresponding to the input
u : [0, ) R is
Z
t

h+
(t )u( ) d.

y(t) =
0

Make sense of the statement, The output y is equal to the convolution h+

u. Note
that you are essentially being asked to resolve the conflict in the limits of integration
between convolution in Fourier and Laplace transforms.
E4.6 We will consider in detail here the mass-spring-damper system whose Bode plots are
produced in Example 4.6. We begin by scaling the input u by the spring constant k
so that, upon division by m, the governing equations are

E4.8

E4.9

x + 20 x + 02 x = 02 u,
q
k
and = 2dkm . As output we shall take y = x. We allow d to be
where 0 = m
zero, but neither m nor k can be zero.
(a) Write this system as a SISO linear system = (A, b, ct , D) (that is, determine
A, b, c, and D), and determine the transfer function T .
(b) Determine for the various values of the parameters 0 and , and then determine the frequency response H .
(c) Show that
02
.
|H ()| = p 2
2
(0 )2 + 4 2 02 2
(d) How does |H ()| behave for 0 ? for 0 ?
d
d
(e) Show that d
|H ()| = 0 if and only if d
|H ()|2 = 0.
(f) Using the previous simplification, show that for < 12 there is a maximum for
H (). The maximum you determine should occur at the frequency
p
max = 0 1 2 2 ,

E4.10

4 Frequency response (the frequency domain)

22/10/2004

(i) Determine an expression for ]H (max ).

(j) Use your work above to give an accurate sketch of the Bode plot for in cases
when 12 , making sure to mark on your plot all the features you determined
in the previous parts of the question. What happens to the Bode plot as is
decreased? What would the Bode plot look like when = 0?
Construct Bode plots by hand for the following first-order SISO linear systems in
input/output form:
(a) (N (s), D(s)) = (1, s + 1);
(b) (N (s), D(s)) = (s, s + 1);
(c) (N (s), D(s)) = (s 1, s + 1);
(d) (N (s), D(s)) = (2, s + 1).
Construct Bode plots by hand for the following second-order SISO linear systems in
input/output form:
(a) (N (s), D(s)) = (1, s2 + 2s + 2);
(b) (N (s), D(s)) = (s, s2 + 2s + 2);
(c) (N (s), D(s)) = (s 1, s2 + 2s + 2);
(d) (N (s), D(s)) = (s + 2, s2 + 2s + 2).
Construct Bode plots by hand for the following third-order SISO linear systems in
input/output form:
(a) (N (s), D(s)) = (1, s3 + 3s2 + 4s + 2);
(b) (N (s), D(s)) = (s, s3 + 3s2 + 4s + 2);
(c) (N (s), D(s)) = (s2 4, s3 + 3s2 + 4s + 2);
(d) (N (s), D(s)) = (s2 + 1, s3 + 3s2 + 4s + 2).
For each of the SISO linear systems given below do the following:
1. calculate the eigenvalues of A;
2. calculate the transfer function;
3. sketch the Bode plots of the magnitude and phase of the frequency response;
4. by trial and error playing with the parameters, on your sketch, indicate the role
played by the parameters of the system (e.g., 0 in parts (a) and (b));
5. try to justify the name of the system by looking at the shapes of the Bode plots.
Here are the systems (assume all parameters are positive):
(a) low-pass filter :

A = 0 , b = 1 , c = 0 , D = 0 ;
(b) high-pass filter :

and should take the value

A = 0 ,
|H (max )| =

1
2

1 2

b= 1 ,

]H () = atan2(02 2 , 20 ),
where atan2 is the smart inverse tangent function that knows in which quadrant
you are.
(h) How does ]H () behave for 0 ? for 0 ?

D= 1 ;

(c) notch filter :

(g) Show that

c = 0 ,

0
1
,
2
0 0

(d) bandpass filter :

0
1
A=
,
2
0 0

0
,
1

0
,
0

D= 1 ;

D= 0 .

Exercises for Chapter 4

155

E4.11 Consider the coupled mass system of Exercises E1.4, E2.19, and E3.14. Assume no
damping and take the input from Exercise E2.19 in the case when = 0. In Exercise E3.14, you constructed an output vector c0 for which the pair (A, c0 ) was
unobservable.
(a) Construct a family of output vectors c by defining c = c0 + c1 for R and
c1 R4 . Make sure you choose c1 so that c is observable for 6= 0.
(b) Determine , and the frequency response H using the output vector c (allow
to be an arbitrary real number).
(c) Choose the parameters m = 1 and k = 1 and determine the Bode plot for values
of around zero. Do you notice anything different in the character of the Bode
plot when = 0?
E4.12 Consider the pendulum/cart system of Exercises E1.5, E2.4, and E3.15. For each of
the following linearisations:
(a) the equilibrium point (0, 0) with cart position as output;
(b) the equilibrium point (0, 0) with cart velocity as output;
(c) the equilibrium point (0, 0) with pendulum angle as output;
(d) the equilibrium point (0, 0) with pendulum angular velocity as output;
(e) the equilibrium point (0, ) with cart position as output;
(f) the equilibrium point (0, ) with cart velocity as output;
(g) the equilibrium point (0, ) with pendulum angle as output;
(h) the equilibrium point (0, ) with pendulum angular velocity as output,
do the following:
1. determine , and the frequency response H for the system;
2. for parameters M = 1 21 , m = 1, g = 9.81, and ` = 12 , produce the Bode plot for
the pendulum/cart system;
3. can you see reflected in your Bode plot the character of the spectrum of the zero
dynamics as you determined in Exercise E3.15?
E4.13 Consider the double pendulum system of Exercises E1.6, E2.5, E1.6 and E3.16. For
each of the following linearisations:
(a) the equilibrium point (0, 0, 0, 0) with the pendubot input;
(b) the equilibrium point (0, , 0, 0) with the pendubot input;
(c) the equilibrium point (, 0, 0, 0) with the pendubot input;
(d) the equilibrium point (, , 0, 0) with the pendubot input;
(e) the equilibrium point (0, 0, 0, 0) with the acrobot input;
(f) the equilibrium point (0, , 0, 0) with the acrobot input;
(g) the equilibrium point (, 0, 0, 0) with the acrobot input;
(h) the equilibrium point (, , 0, 0) with the acrobot input,
do the following:
1. determine , and the frequency response H for the system;
2. for parameters m1 = 1, m2 = 2, `1 = 12 , and `2 = 13 , produce the Bode plot for
the double pendulum;
3. can you see reflected in your Bode plot the character of the spectrum of the zero
dynamics as you determined in Exercise E3.16?
In each case, use the angle of the second link as output.

156

4 Frequency response (the frequency domain)

22/10/2004

E4.14 Consider the coupled tank system of Exercises E1.11, E2.6, and E3.17. For the linearisations in the following cases:
(a) the output is the level in tank 1;
(b) the output is the level in tank 2;
(c) the output is the difference in the levels,
do the following:
1. determine , and the frequency response H for the system;
1
1
2. for parameters = 13 , 1 = 1, A1 = 1, A2 = 12 , a1 = 10
, a2 = 20
, and g = 9.81,
produce the Bode plot for the tank system;
3. can you see reflected in your Bode plot the character of the spectrum of the zero
dynamics as you determined in Exercise E3.17?
E4.15 Suppose you are shown a Bode plot for a stable SISO linear system . Can you
tell from the character of the plot whether is controllable? observable? minimum
phase?
E4.16 Construct two transfer functions, one minimum phase and the other not, whose frequency response magnitudes are the same (you cannot use the one in the book). Make
Bode plots for each system, and make the relevant observations.
E4.17 Let (N, D) be a proper, stable SISO linear system of relative degree m. Characterise
the total phase shift,
]HN,D () ]HN,D (0),
in terms of the roots of N , assuming that N has no roots on iR.
In the next exercise, you will provide a rigorous justification for the term minimum phase.
E4.18 Let (N, D) be a SISO linear system in input/output form, and suppose that it is
minimum phase (thus N has no roots in C+ ). Let M (N, D) denote the collection of
D)
in input/output form for which
SISO linear systems (N,

R.
|HN,D ()| = HN,
D
() ,
Thus the magnitude Bode plots for all systems in M (N, D) are exactly the magnitude
Bode plot for (N, D).
(a) How are systems in M (N, D) related to (N, D)? This boils down, of course, to
identifying SISO linear systems that have a plot magnitude of 1 at all frequencies.
D)
M (N, D) denote 0 (N,
D)
= lim0 ]H () and (N,
D)
=
For (N,
N,D
lim ]HN,
D
(). Now let
D)
= (N,
D)
0 (N,
D).

(N,

E4.19
E4.20
E4.21
E4.22

With this notation prove the following statement.

D)
for any (N,
D)
M (N, D).
(b) (N, D) (N,
Construct polar plots corresponding to the Bode plots you made in Exercise E4.7.
Construct polar plots corresponding to the Bode plots you made in Exercise E4.8.
Construct polar plots corresponding to the Bode plots you made in Exercise E4.9.
For the SISO linear system of Exercise E4.2, plot the polar plots for various 0 0
and 0 > 0.

Exercises for Chapter 4

157

158

It is not uncommon to encounter a scheme for control design that relies on a plant being
stable, i.e., having all poles in C . This is in conflict with many plantsfor example, the
simple mass that have a pole at s = 0. In the next exercise, you will investigate a
commonly employed hack to get around plants having poles at s = 0.
E4.23 The transfer function
T (s) =

K
s + 1

is put forward as providing a stable approximation to 1s .

(a) Comment on how fiddling K and , particularly , render the approximation
better or worse.
(b) Explain in what sense this approximation is valid, and also how it is invalid.
E4.24 For the following SISO linear systems in input/output form, determine whether they
satisfy the hypotheses of Bodes Gain/Phase Theorem. If the system does satisfy
the hypotheses, verify explicitly that the theorem holds, and determine how good is
the approximation (4.12) at various frequencies. If the system does not satisfy the
hypotheses, determine explicitly whether the theorem does in fact hold.
(a) (N (s), D(s)) = (1, s).
(b) (N (s), D(s)) = (s, 1).
(c) (N (s), D(s)) = (1, s + 1).
(d) (N (s), D(s)) = (1, s 1).
(e) (N (s), D(s)) = (s + 1, 1).
(f) (N (s), D(s)) = (s 1, 1).
E4.25 Exercise on the approximate version of Bodes Gain/Phase Theorem.

complete

4 Frequency response (the frequency domain)

22/10/2004

This version: 22/10/2004

Part II
System analysis

162

This version: 22/10/2004

5 Stability of control systems

22/10/2004

5.1 Internal stability

Internal stability is a concept special to SISO linear systems i.e., those like

Chapter 5
Stability of control systems
We will be trying to stabilise unstable systems, or to make an already stable system even
more stable. Although the essential goals are the same for each class of system we have
encountered thus far (i.e., SISO linear systems and SISO linear systems in input/output
form), each has its separate issues. We first deal with these. In each case, we will see that
stability boils down to examining the roots of a polynomial. In Section 5.5 we give algebraic
criteria for determining when the roots of a polynomial all lie in the negative complex plane.

x(t)
= Ax(t) + bu(t)
y(t) = ct x(t) + Du(t).

(5.1)

Internal stability refers to the stability of the system without our doing anything with the
controls. We begin with the definitions.
5.1 Definition A SISO linear system = (A, b, ct , D) is
(i) internally stable if
lim sup kx(t)k <
t

for every solution x(t) of x(t)

= Ax(t);
(ii) internally asymptotically stable if
lim kx(t)k = 0

Contents
5.1

Internal stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162

5.2

Input/output stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165

5.3

5.4

5.5

5.6

5.2.1

BIBO stability of SISO linear systems . . . . . . . . . . . . . . . . . . . . . . . . . 165

5.2.2

BIBO stability of SISO linear systems in input/output form . . . . . . . . . . . . 168

Norm interpretations of BIBO stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170

5.3.1

Signal norms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170

5.3.2

Hardy spaces and transfer function norms

5.3.3

Stability interpretations of norms . . . . . . . . . . . . . . . . . . . . . . . . . . . 174

. . . . . . . . . . . . . . . . . . . . . . 172

Liapunov methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180

5.4.1

Background and terminology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180

5.4.2

Liapunov functions for linear systems . . . . . . . . . . . . . . . . . . . . . . . . . 182

Identifying polynomials with roots in C . . . . . . . . . . . . . . . . . . . . . . . . . . . 189

5.5.1

The Routh criterion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189

5.5.2

The Hurwitz criterion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193

5.5.3

The Hermite criterion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195

5.5.4

The Lienard-Chipart criterion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199

5.5.5

Kharitonovs test . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200

Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203

for every solution x(t) of x(t)

= Ax(t);
(iii) internally unstable if it is not internally stable.

Of course, internal stability has nothing to do with any part of other than the matrix
A. If one has a system that is subject to the problems we discussed in Section 2.3, then
one may want to hope the system at hand is one that is internally stable. Indeed, all the
bad behaviour we encountered there was a direct result of my intentionally choosing systems
that were not internally stableit served to better illustrate the problems that can arise.
Internal stability can almost be determined from the spectrum of A. The proof of the
following result, although simple, relies on the structure of the matrix exponential as we
discussed in Section B.2. We also employ the notation
C = {z C | Re(z) < 0} ,

C+ = { z C | Re(z) > 0} ,

C = { z C | Re(z) 0} , C+ = { z C | Re(z) 0} ,
iR = { z C | Re(z) = 0} .
With this we have the following result, recalling notation concerning eigenvalues and eigenvectors from Section A.5.
5.2 Theorem Consider a SISO linear system = (A, b, ct , D). The following statements
hold.
(i) is internally unstable if spec(A) C+ 6= .
(ii) is internally asymptotically stable if spec(A) C .
(iii) is internally stable if spec(A) C+ = and if mg () = ma () for spec(A)
(iR).
(iv) is internally unstable if mg () < ma () for spec(A) (iR).

5.1 Internal stability

22/10/2004

163

Proof (i) In this case there is an eigenvalue + i C+ and a corresponding eigenvector

u + iv which gives rise to real solutions
x1 (t) = et (cos tu sin tv),

x2 (t) = et (sin tu + cos tv).

Clearly these solutions are unbounded as t since > 0.

= Ax(t) will be a linear

(ii) If all eigenvalues lie in C , then any solution of x(t)
combination of n linearly independent vector functions of the form
tk et u or tk et (cos tu sin tv) or tk et (sin tu + cos tv)

(5.2)

for > 0. Note that all such functions tend in length to zero as t . Suppose that we
have a collection x1 , . . . , xn (t) of such vector functions. Then, for any solution x(t) we have,
for some constants c1 , . . . , cn ,
lim kx(t)k = lim kc1 x1 (t) + + cn xn (t)k

|c1 | lim kx1 (t)k + + |cn | lim kxn (t)k

= 0,
where we have used the triangle inequality, and the fact that the solutions x1 (t), . . . , xn (t)
all tend to zero as t .
(iii) If spec(A) C+ = and if further spec(A) C , then we are in case (ii), so
is internally asymptotically stable, and so internally stable. Thus we need only concern
ourselves with the case when we have eigenvalues on the imaginary axis. In this case,
provided all such eigenvalues have equal geometric and algebraic multiplicities, all solutions
will be linear combinations of functions like those in (5.2) or functions like
sin tu or

cos tu.

(5.3)

Let x1 (t), . . . , x` (t) be ` linearly independent functions of the form (5.2), and let
x`+1 (t), . . . , xn (t) be linearly independent functions of the form (5.3) so that x1 , . . . , xn

forms a set of linearly independent solutions for x(t)

= Ax(t). Thus we will have, for some
constants c1 , . . . , cn ,
lim sup kx(t)k = lim sup kc1 x1 (t) + + cn xn (t)k
t

|c1 | lim sup kx1 (t)k + + |c` | lim sup kx` (t)k +
t

|c`+1 | lim sup kx`+1 (t)k + + |cn | lim sup kxn (t)k
t

= |c`+1 | lim sup kx`+1 (t)k + + |cn | lim sup kxn (t)k .
t

Since each of the terms kx`+1 (t)k , . . . , kxn (t)k are bounded, their lim sups will exist, which
is what we wish to show.
(iv) If A has an eigenvalue = i on the imaginary axis for which mg () < ma () then

there will be solutions of x(t)

= Ax(t) that are linear combinations of vector functions of
the form tk sin tu or tk cos tv. Such functions are unbounded as t , and so is
internally unstable.

5.3 Remarks 1. A matrix A is Hurwitz if spec(A) C . Thus A is Hurwitz if and only
if = (A, b, ct , D) is internally asymptotically stable.

164

5 Stability of control systems

22/10/2004

2. We see that internal stability is almost completely determined by the eigenvalues of A.

Indeed, one says that is spectrally stable if A has no eigenvalues in C+ . It is only
in the case where there are repeated eigenvalues on the imaginary axis that one gets to
distinguish spectral stability from internal stability.
3. One does not generally want to restrict oneself to systems that are internally stable.
Indeed, one often wants to stabilise an unstable system with feedback. In Theorem 6.49
we shall see, in fact, that for controllable systems it is always possible to choose a
feedback vector that makes the closed-loop system internally stable.

The notion of internal stability is in principle an easy one to check, as we see from an
example.
5.4 Example We look at a SISO linear system = (A, b, ct , D) where

0
1
A=
.
b a
The form of b, c, and D does not concern us when talking about internal stability. The
eigenvalues of A are the roots of the characteristic polynomial s2 + as + b, and these are

a2 12 a2 4b.
The situation with the eigenvalue placement can be broken into cases.
1. a = 0 and b = 0: In this case there is a repeated zero eigenvalue. Thus we have spectral
stability, but we need to look at eigenvectors to determine internal stability. One readily
verifies that there is only one linearly independent eigenvector for the zero eigenvalue, so
the system is unstable.
2. a = 0 and b > 0: In this case the eigenvalues are purely imaginary. Since the roots are
also distinct, they will have equal algebraic and geometric multiplicity. Thus the system
is internally stable, but not internally asymptotically stable.
3. a = 0 and b < 0: In this case both roots are real, and one will be positive. Thus the
system is unstable.
4. a > 0 and b = 0: There will be one zero eigenvalue if b = 0. If a > 0 the other root will
be real and negative. In this case then, we have a root on the imaginary axis. Since it is
distinct, the system will be stable, but not asymptotically stable.
5. a > 0 and b > 0: One may readily ascertain (in Section 5.5 well see an easy way to do
this) that all eigenvalues are in C if a > 0 and b > 0. Thus when a and b are strictly
positive, the system is internally asymptotically stable.
6. a > 0 and b < 0: In this case both eigenvalues are real, one being positive and the other
negative. Thus the system is internally unstable.
7. a < 0 and b = 0: We have one zero eigenvalue. The other, however, will be real and
positive, and so the system is unstable.
8. a < 0 and b > 0: We play a little trick here. If s0 is a root of s2 + as + b with a, b < 0,
then s0 is clearly also a root of s2 as + b. From the previous case, we know that
s0 C , which means that s0 C+ . So in this case all eigenvalues are in C+ , and so
we have internal instability.
9. a < 0 and b < 0: In this case we are guaranteed that all eigenvalues are real, and
furthermore it is easy to see that one eigenvalue will be positive, and the other negative.
Thus the system will be internally unstable.

5.2 Input/output stability

22/10/2004

165

Note that one cannot really talk about internal stability for a SISO linear system (N, D)
in input/output form. After all, systems in input/output form do not have built into them a
notion of state, and internal stability has to do with states. In principle, one could define the
internal stability for a proper system as internal stability for N,D , but this is best handled
by talking directly about input/output stability which we now do.

5.2 Input/output stability

We shall primarily be interested in this course in input/output stability. That is, we want
nice inputs to produce nice outputs. In this section we demonstrate that this property is
intimately related with the properties of the impulse response, and therefore the properties
of the transfer function.
5.2.1 BIBO stability of SISO linear systems We begin by talking about input/output stability in the context of SISO linear systems. When we have understood
this, it is a simple matter to talk about SISO linear systems in input/output form.
5.5 Definition A SISO linear system = (A, b, ct , D) is bounded input, bounded output
stable (BIBO stable) if there exists a constant K > 0 so that the conditions (1) x(0) = 0

and (2) |u(t)| 1, t 0 imply that y(t) K where u(t), x(t), and y(t) satisfy (5.1).
Thus BIBO stability is our way of saying that a bounded input will produce a bounded
output. You can show that the formal definition means exactly this in Exercise E5.8.
The following result gives a concise condition for BIBO stability in terms of the impulse
response.
= (A, b, ct , 01 ).
5.6 Theorem Let = (A, b, ct , D) be a SISO linear system and define
Then is BIBO stable if and only if limt |h (t)| = 0.
Proof Suppose that limt |h (t)| =
6 0. Then, by Proposition 3.24, it must be the case
that either (1) h (t) blows up exponentially as t or that (2) h (t) is a sum of terms,
one of which is of the form sin t or cos t. For the first case we can take the bounded input
u(t) = 1(t). Using Proposition 2.32 and Proposition 3.24 we can then see that
Z
y(t) =
h (t ) d + Du(t).

166

5 Stability of control systems

22/10/2004

for some M > 0. Therefore, whenever u(t) 1 for t 0, we have

Z t

|y(t)| = h (t )u( ) d + Du(t)
0
Z t
|h (t )u( )| d + |D|

Z0 t

|h (t )| |u( )| d + |D|
Z0 t

|h (t )| d + |D|
0

M + |D| .
This means that is BIBO stable.

This result gives rise to two easy corollaries, the first following from Proposition 3.24, and
the second following from the fact that if the real part of all eigenvalues of A are negative
then limt |h (t)| = 0.
5.7 Corollary Let = (A, b, ct , D) be a SISO linear system and write
T (s) =
where (N, D) is the c.f.r.
negative half-plane.

N (s)
D(s)

Then is BIBO stable if and only if D has roots only in the

5.8 Corollary = (A, b, ct , D) is BIBO stable if spec(A) C .

The matter of testing for BIBO stability of a SISO linear system is straightforward, so
lets do it for a simple example.
5.9 Example (Example 5.4 contd) We continue with the case where

0
1
A=
,
b a

Since h (t) blows up exponentially, so too will y(t) if it is so defined. Thus the bounded
input u(t) = 1(t) produces an unbounded output. For case (2) we choose u(t) = sin t and
compute
Z t
Z t

sin (t ) sin d = 21 1 sin t t cos t ,
cos (t ) sin d = 21 t sin t.
0

Therefore, y(t) will be unbounded for the bounded input u(t) = sin t. We may then
conclude that is not BIBO stable.
Now suppose that limt |h (t)| = 0. By Proposition 3.24 this means that h (t) dies
off exponentially fast as t , and therefore we have a bound like
Z
|h (t )| d M
0

and we now add the information

0
,
1

1
,
0

D = 01 .

We compute
1
.
s2 + as + b
From Example 5.4 we know that we have BIBO stability if and only if a > 0 and b > 0.
Lets probe the issue a bit further by investigating what actually happens when we do
not have a, b > 0. The cases when is internally unstable are not altogether interesting
since the system is obviously not BIBO stable in these cases. So let us examine the cases
when we have no eigenvalues in C+ , but at least one eigenvalue on the imaginary axis.
T (s) =

22/10/2004

5.2 Input/output stability

167

1. a = 0 and b > 0: Here the eigenvalues are i b, and we compute

sin bt
h (t) = .
b

5 Stability of control systems

22/10/2004

We see that the state is behaving poorly, even though the output may be determined as
y(t) = 12 (e2t 1),
which is perfectly well-behaved. But we have seen this sort of thing before.

Thus the impulse response is bounded, but does not tend to zero as t . Theorem 5.6
predicts that there will be a bounded input
signal that produces an unbounded output
signal. In fact, if we choose u(t) = sin bt and zero initial condition, then one verifies
that the output is

Z t

sin( bt) t cos( bt)

.
y(t) =
ct eA(t ) b sin( b ) d =
2b
2 b
0
Thus a bounded input gives an unbounded output.
2. a > 0 and b = 0: The eigenvalues here are {0, a}. One may determine the impulse
response to be
1 eat
h (t) =
.
a
This impulse response is bounded, but again does not go to zero as t . Thus
there ought to be a bounded input that gives an unbounded output. We have a zero
eigenvalue, so this means we should choose a constant input. We take u(t) = 1 and zero
initial condition and determine the output as
Z t
1 eat
t
.
y(t) =
ct eA(t ) b d =
a
a2
0
Again, a bounded input provides an unbounded output.

168

As usual, when dealing with input/output issues for systems having states, one needs to
exercise caution for the very reasons explored in Section 2.3. This can be demonstrated with
an example.
5.10 Example Let us choose here a specific example (i.e., one without parameters) that will
illustrate problems that can arise with fixating on BIBO stability while ignoring other considerations. We consider the system = (A, b, ct , 01 ) with

0 1
0
1
A=
, b=
, c=
.
2 1
1
1
We determine that h (t) = e2t . From Theorem 5.6 we determine that is BIBO stable.
But is everything really okay? Well, no, because this system is actually not observable.
We compute

1 1
O(A, c) =
,
2 2
and since this matrix has rank 1 the system is not observable. How is this manifested in
the system behaviour? In exactly the way one would predict. Thus let us look at the state
behaviour for the system with a bounded input. We take u(t) = 1(t) as the unit step input,
and take the zero initial condition. The resulting state behaviour is defined by
1 t 1 t 1
Z t
e + e
x(t) =
eA(t ) b d = 3 1 t 6 1 2t 2 .
e 3e
0
3

Let us state a result that provides a situation where one can make a precise relationship
between internal and BIBO stability.
5.11 Proposition If a SISO linear system = (A, b, c0 , D) is controllable and observable,
then the following two statements are equivalent:
(i) is internally asymptotically stable;
(ii) is BIBO stable.
Proof When is controllable and observable, the poles of T are exactly the eigenvalues
of A.

When is not both controllable and observable, the situation is more delicate. The
diagram in Figure 5.1 provides a summary of the various types of stability, and which types

internally
asymptotically
stable GGG

+3 internally stable

GGGG
GGGG
GGGG
GGGG
GGGG
GGGG
GGGG
GGG
'

BIBO stable
Figure 5.1 Summary of various stability types for SISO linear systems

imply others. Note that there are not many arrows in this picture. Indeed, the only type of
stability which implies all others is internal asymptotic stability. This does not mean that
if a system is only internally stable or BIBO stable that it is not internally asymptotically
stable. It only means that one cannot generally infer internal asymptotic stability from
internal stability or BIBO stability. Whats more, when a system is internally stable but not
internally asymptotically stable, then one can make some negative implications, as shown in
Figure 5.2. Again, one should be careful when interpreting the absence of arrows from this
diagram. The best approach here is to understand that there are principles that underline
when one can infer one type of stability from another. If these principles are understood,
then matters are quite straightforward. A clearer resolution of the connection between BIBO
stability and internal stability is obtained in Section 10.1 when we introduce the concepts
of stabilisability and detectability. The complete version of Figures 5.1 and 5.2 is given
by Figure 10.1. Note that we have some water to put under the bridge to get there. . .
5.2.2 BIBO stability of SISO linear systems in input/output form It is now clear
how we may extend the above discussion of BIBO stability to systems in input/output form,
at least when they are proper.

5.2 Input/output stability

22/10/2004

internally stable but not

internally asymptotically
stable
MM

169

M MM
MM
M MM
M
if controllable and observable MMMM M
M MM
M MM

not internally
asymptotically
stable

22/10/2004

In this section, we offer interpretations of the stability characterisations of the previous

section in terms of various norms for transfer functions and for signals. The material in
this section will be familiar to those who have had a good course in signals and systems.
However, it is rare that the subject be treated in the manner we do here, although its value
for understanding control problems is now well established.

Figure 5.2 Negative implication when a system is internally stable,

but not internally asymptotically stable

5.12 Definition A proper SISO linear system (N, D) in input/output form is bounded input,
bounded output stable (BIBO stable) if the SISO linear system N,D is BIBO stable.
From Corollary 5.7 follows the next result giving necessary and sufficient conditions for
BIBO stability of strictly proper SISO systems in input/output form.
5.13 Proposition A proper SISO linear system (N, D) in input/output form is BIBO stable
if and only if TN,D has no poles in C+ .
The question then arises, What about SISO linear systems in input/output form that
are not proper? Well, such systems can readily be shown to be not BIBO stable, no matter
what the character of the denominator polynomial D. The following result shows why this
is the case.
5.14 Proposition If (N, D) is an improper SISO linear system in input/output form, then
there exists an input u satisfying the properties
(i) |u(t)| 1 for all t 0 and

d
d
(ii) if y satisfies D dt
y(t) = N dt
u(t), then for any M > 0 there exists t > 0 so that
|y(t)| > M .
Proof From Theorem C.6 we may write
N (s)
= R(s) + P (s)
D(s)
where R is a strictly proper rational function and P is a polynomial of degree at least 1.
Therefore, for any input u, the output y will be a sum y = y1 + y2 where
y2 (s) = P (s)
u(s).

5 Stability of control systems

5.3 Norm interpretations of BIBO stability

not BIBO stable

y1 (s) = R(s)
u(s),

170

(5.4)

If R has poles in C+ , then the result follows in usual manner of the proof of Theorem 5.6.
So we may as well suppose that R has no poles in C+ , so that the solution y1 is bounded.
We will show, then, that y2 is not bounded. Let us choose u(t) = sin(t2 ). Any derivative of
u will involve terms polynomial in t and such terms will not be bounded as t . But y2 ,
by (5.4), is a linear combination of derivatives of u, so the result follows.

5.3.1 Signal norms We begin our discussion by talking about ways to define the size
of a signal. The development in this section often is made in a more advanced setting where
the student is assumed to have some background in measure theory. However, it is possible
to get across the basic ideas in the absence of this machinery, and we try to do this here.
For p 1 and for a function f : (, ) R denote
Z
1/p
kf kp =
|f (t)|p dt

which we call the Lp -norm of y. Denote

Lp (, ) = f : (, ) R kf kp < .
Functions in Lp (, ) are said to be Lp -integrable. The case where p = is handled
separately by defining
kf k = sup{|f (t)| for almost every t}
0

as the L -norm of y. The L -norm is sometimes referred to as the sup norm. Here
almost every means except on a set T (, ) having the property that
Z
dt = 0.
T

We denote
L (, ) = { f : (, ) R | kf k < }
as the set of functions that we declare to be L -integrable. Note that we are dealing
here with functions defined on (, ), whereas with control systems, one most often has
functions that are defined to be zero for t < 0. This is still covered by what we do, and the
extra generality is convenient.
Most interesting to us will be the Lp spaces L2 (, ) and L (, ). The two sets
of functions certainly do not coincide, as the following collection of examples indicate.
5.15 Examples 1. The function cos t is in L (, ), but is in none of the spaces
Lp (, ) for 1 p < . In particular, it is not L2 -integrable.
1
2. The function f (t) = 1+t
is not L1 -integrable, although it is L2 -integrable; one computes
kf k2 = 1.
3. Define
(q
1
, t (0, 1]
t
f (t) =
0,
otherwise.
One then checks that kf k1 = 2, but that f is not L2 -integrable. Also, since limt1 f (t) =
, the function is not L -integrable.

22/10/2004

5.3 Norm interpretations of BIBO stability

4. Define

171

(
ln t, t (0, 1]
f (t) =
0,
otherwise.

172

kf k22 =

1
T 2T

f 2 (t) dt

|f (t)| |f (t)| dt
Z
|f (t)| dt
kf k
=

= kf k kf k1 ,
as desired.

The relationships between the various Lp -spaces we shall care about and the pow operation are shown in Venn diagrammatic form in Figure 5.3.
pow

exists. For a power signal f we then define

pow(f ) =

22/10/2004

(iii) We have

Note that limt0+ f (t) = ; thus f is not L -integrable. Nonetheless, one checks that
if p is an integer, kf kp = (p!)1/p , so f is Lp -integrable for integers p [1, ). More
generally one has kf kp = (1 + p)1/p where the -function generalises the factorial to
non-integer values.

There is another measure of signal size we shall employ that differs somewhat from the
above measures in that it is not a norm. We let f : (, ) R be a function and say
that f is a power signal if the limit
Z T
1
f 2 (t) dt
lim
T 2T T

5 Stability of control systems

lim

1/2
f 2 (t) dt
,

L
1
which we call the average power of f . If we consider the function f (t) = (1+t)
2 we observe
that pow(f ) = 0 even though f is nonzero. Thus pow is certainly not a norm. Nevertheless,
it is a useful, and often used, measure of a signals size.
The following result gives some relationships between the various Lp -norms and the pow
operation.

5.16 Proposition The following statements hold:

(i) if f L2 (, ) then pow(f ) = 0;
(ii) if f L (, ) is a power signal then pow(f ) kf k ;
p
(iii) if f L1 (, ) L (, ) then kf k2 kf k kf k1 .
Proof (i) For T > 0 we have
Z

Figure 5.3 Venn diagram for relationships between Lp -spaces and

pow

f 2 (t) dt kf k22
Z T
1
1
f 2 (t) dt kf k22 .
2T T
T
T

The result follows since as T , the right-hand side goes to zero.

(ii) We compute
1
T 2T

f 2 (t) dt
Z T
1
kf k2 lim
dt
T 2T T

pow(f ) = lim

= kf k2 .

5.3.2 Hardy spaces and transfer function norms For a meromorphic complex-valued
function f we will be interested in measuring the size by f by evaluating its restriction to
the imaginary axis. To this end, given a meromorphic function f , we follow the analogue of
our time-domain norms and define, for p 1, the Hp -norm of f by
1/p
1 Z
|f (i)|p d
.
kf kp =
2

In like manner we define the H -norm of f by

kf k = sup |f (i)| .

While these definitions make sense for any meromorphic function f , we are interested in
particular such functions. In particular, we denote
RLp = {R R(s) | kRkp < }

22/10/2004

5.3 Norm interpretations of BIBO stability

173

for p [1, ) {}. Let us call the class of meromorphic functions f that are analytic in
C+ Hardy functions.1 We then have
H+
p = {f | f is a Hardy function with kf kp < },
for p [0, ) {}. We also have RH+
p = R(s) Hp as the Hardy functions with bounded
Hp -norm that are real rational. In actuality, we shall only be interested in the case when
p {1, 2, }, but the definition of the Hp -norm holds generally. One must be careful when
one sees the symbol kkp that one understands what the argument is. In one case we mean
it to measure the norm of a function of t defined on (, ), and in another case we use it
to define the norm of a complex function measured by looking at its values on the imaginary
axis.
Note that with the above notation, we have the following characterisation of BIBO stability.
5.17 Proposition A proper SISO linear system (N, D) in input/output form is BIBO stable
if and only if TN,D RH+
.

174

5 Stability of control systems

to be the function on the right

for some positive integer k and some C C. Let us define R
hand side of this approximation, and note that
Z 0 +
Z 0 +
2

2
C

R(i)
d =

d
k
(i(

))
0
0
0
Z 2
1
= |C|2
k d

= .
Thus the contribution to kRk2 of a pole on the imaginary axis will always be unbounded.
Conversely, if R is strictly proper with no poles on the imaginary axis, then one can find
a sufficiently large M > 0 and a sufficiently small > 0 so that
M

|R(i)| , R.

1 + i
One then computes
Z

The following result gives straightforward characterisations of the various rational function spaces we have been talking about.
5.18 Proposition The following statements hold:
(i) RL consists of those functions in R(s) that
(a) have no poles on iR and
(b) are proper;
(ii) RH+
consists of those functions in R(s) that
(a) have no poles in C+ and
(b) are proper;
(iii) RL2 consists of those functions in R(s) that
(a) have no poles on iR and
(b) are strictly proper.
(iv)

RH+
2

consists of those functions in R(s) that

(a) have no poles in C+ and

(b) are strictly proper.

R(s)
1

M 2
M
d = .
1 + i
2

This implies that kRk2

and so R RH+
2.
Clearly the above argument for RH+
2 also applies for RL2 .
M
2

After George Harold Hardy (1877-1947).

C
(s i0 )k

5.3.3 Stability interpretations of norms To characterise BIBO stability in terms of

these signal norms, we consider a SISO linear system (N, D) in input/output form. We wish
to flush out the input/output properties of a transfer function relative to the Lp signal norms
and the pow operation. For notational convenience, let us adopt the notation kkpow = pow
and let Lpow (, ) denote those functions f for which pow(f ) is defined. This is an abuse
of notation since pow is not a norm. However, the abuse is useful for making the following
definition.
5.19 Definition Let R R(s) be a proper rational function, and for u L2 (, ) let
yu : (, ) R be the function satisfying yu (s) = R(s)
u(s). For p1 , p2 [1, ) {}
{pow}, the Lp1 Lp2 -gain of R is defined by
kRkp1 p2 =

Proof Clearly we may prove the first and second, and then the third and fourth assertions
together.
(i) and (ii): This part of the proposition follows since a rational Hardy function is proper
if and only if lims |R(s)| < , and since |R(i)| is bounded for all R if and only if
R has no poles on iR. The same applies for RL .
(iii) and (iv) Clearly if R RH+
2 then lims |R(s)| = 0, meaning that R must be strictly
proper. We also need to show that R RH+
2 implies that R has no poles on iR. We shall
do this by showing that if R has poles on iR then R 6 RH+
2 . Indeed, if R has a pole at i0
then near i0 , R will essentially look like

22/10/2004

kyu kp2
.
uLp1 (,) kukp1
sup

u not zero

If (N, D) is SISO linear system in input/output form, then (N, D) is Lp1 Lp2 -stable if
kTN,D kp1 p2 < .

This definition of Lp1 Lp2 -stability is motivated by the following obvious result.
5.20 Proposition Let (N, D) be an Lp1 Lp2 stable SISO linear system in input/output
form and let u : (, ) R be an input with yu : (, ) R the function satisfying
yu (s) = TN,D (s)
u(s). If u Lp1 (, ) then
kyu kp2 kTN,D kp1 p2 kukp1 .
In particular, u Lp1 (, ) implies that yu Lp2 (, ).

22/10/2004

5.3 Norm interpretations of BIBO stability

175

Although our definitions have been made in the general context of Lp -spaces, we are
primarily interested in the cases where p1 , p2 {1, 2, }. In particular, we would like to
be able to relate the various gains for transfer functions to the Hardy space norms of the
previous section. The following result gives these characterisations. The proofs, as you will
see, is somewhat long and involved.
5.21 Theorem For a proper BIBO stable SISO linear system (N, D) in input/output form,

let C R be defined by TN,D (s) = TN,
D
(s) + C where (N, D) is strictly proper. Thus C = 0
if (N, D) is itself strictly proper. The following statements then hold:
(i) kTN,D k22 = kTN,D k ;
(iv) kTN,D k2 = ;
(ii) kTN,D k2 = kTN,D k2 ;
(v) kTN,D k khN,
D
k1 + |C|;
(iii) kTN,D k2pow = 0;
(vi) kTN,D kpow kTN,D k ;
(vii) kTN,D kpow2 = ;
(viii) kTN,D kpow = ;
(ix) kTN,D kpowpow = kTN,D k .
If (N, D) is strictly proper, then part (v) can be improved to kTN,D k = khN,D k1 .
Proof (i) By Parsevals Theorem we have kf k2 = kfk2 for any function f L2 (, ).
Therefore
yu k22
kyu k22 = k
Z
1
=
|
yu (i)|2 d
2
Z
1
|TN,D (i)|2 |
u(i)|2 d
=
2
Z
1
kTN,D k2
|
u(i)|2 d
2
= kTN,D k2 k
uk22
= kTN,D k2 kuk22 .
This shows that kTN,D k22 kTN,D k . We shall show that this is the least upper bound.
Let 0 R+ be a frequency at which kTN,D k is attained. First let us suppose that 0 is
finite. For > 0 define u to have the property
(p
/2, | 0 | < or | + 0 | <
u (i) =
0,
otherwise.
Then, by Parsevals Theorem, kuk2 = 1. We also compute

1
lim k
y u k =
|TN,D (i0 )|2 + |TN,D (i0 )|2
0
2
= |TN,D (i0 )|2
= kTN,D k2 .
If kTN,D k is not attained at a finite frequency, then we define u so that
(p

/2, 1 < or + 1 <
u (i) =
0,
otherwise.

176

5 Stability of control systems

22/10/2004

In this case we still have kuk2 = 1, but now we have

lim k
yu k = lim |TN,D (i)|2 = kTN,D k2 .
0

In either case we have shown that kTN,D k is a least upper bound for kTN,D k22 .
(ii) Here we employ the Cauchy-Schwartz inequality to determine
Z
hN,D (t )u( ) d
|yu (t)| =

Z
1/2 Z
1/2

h2N,D (t ) d
u2 ( ) d

= khN,D k2 kuk2
= kTN,D k2 kuk2 ,
where the last equality follows from Parsevals Theorem. Thus we have shown that
kTN,D k2 kTN,D k2 . This is also the least upper bound since if we take
u(t) =

hN,D (t)
,
kTN,D k2

we determine by Parsevals Theorem that kuk2 = 1 and from our above computations that
|y(0)| = kTN,D k2 which means that kyu k kTN,D k2 , as desired.
(iii) Since yu is L2 -integrable if u is L2 -integrable by part (i), this follows from
Proposition 5.16(i).
(iv) Let R+ have the property that TN,D (i) 6= 0. Take u(t) = sin t. By Theorem 4.1
we have
yu (t) = Re(T (i)) sin t + Im(T (i)) cos t + yh (t) + C
where limt yh (t) = 0. In this case we have kuk = 1 and kyu k2 = .
(v) We compute
Z

hN,
|y(t)| =
D
(t )u( ) d + Cu(t)

Z

=
hN,
D
( )u(t ) d + Cu(t)

Z

h ( )u(t ) d + |C| |u(t)|

N,D

Z

h ( ) d + |C|
kuk
N,D

= khN,
D
k1 + |C| kuk .
This shows that kTN,D k khN,
D
k1 + |C| as stated. To see that this is the least upper
bound when (N, D) is strictly proper (cf. the final statement in the theorem), fix t > 0 and
define u so that
(
+1, hN,D ( ) 0
u(t ) =
1, hN,D ( ) < 0.

5.3 Norm interpretations of BIBO stability

22/10/2004

177

Then we have kuk = 1 and

hN,D ( )u(t ) d

yu (t) =
Z

|hN,D ( )| d

5 Stability of control systems

22/10/2004

which gives (5.6) as desired. Using (5.5) combined with (5.6) we then have
Z
1
|TN,D (i)|2 (u)().
pow(yu )2 =
2

178

= khN,D k1 .
Thus kyu k khN,D k1 .
(vii) To carry out this part of the proof, we need a little diversion. For a power signal f
define
Z T
1
(f )(t) = lim
f ( )f (t + ) d
T 2T T
and note that (f )(0) = pow(f ). The limit in the definition of (f ) may not exist for all
, but it will exist for certain power signals. Let f be a nonzero power signal for which the
limit does exist. Denote by (f ) the Fourier transform of (f ):
Z
(f )(t)eit dt.
(f )() =

(5.7)

Provided that we choose u so that |TN,D (i)|2 (u)() is not identically zero, we see that
pow(yu )2 > 0 so that kyu k = .
(ix) By (5.7) we have pow(yu ) kTN,D k pow(u). Therefore kTN,D kpowpow kTN,D k .
To show that this is the least upper bound, let 0 R+ be
a frequency at which kTN,D k is
realised, and first suppose that 0 is finite. Now let u(t) = 2 sin 0 t. One readily computes
(u)(t) = cos 0 t, implying by (5.5) that pow(u) = 1. Also we clearly have

(u)() = ( 0 ) + ( + 0 ) ,
An application of (5.7) then gives

1
|TN,D (i0 )|2 + |TN,D (i0 )|2
2
= |TN,D (i0 )|2

pow(yu )2 =

= kTN,D k2 .

Therefore, since (f ) is the inverse Fourier transform of (f ) we have

Z
1
(f )() d.
pow(f )2 =
2

(5.5)

Now we claim that if yu is related to u by yu (s) = TN,D (s)

u(s) where u is a power signal for
which (u) exists, then we have
(yu )() = |TN,D (i)|2 (u)().

If kTN,D k is attained only in the limit as frequency goes to infinity, then the above argument
is readily modified to show that one can find a signal u so that pow(yu ) is arbitrarily close
to kTN,D k .
(vi) Let u L (, ) be a power signal. By Proposition 5.16(ii) we have pow(u)
kuk . It therefore follows that
kTN,D kpow =

(5.6)

kyu kpow
uL (,) kuk
sup

u not zero

Indeed note that

hN,D ()y(t)u(t + ) d,

yu (t)yu (t + ) =

so that

hN,D ( )(yu , u)(t ) d,

(yu )(t) =

1
T 2T

(f, g)(t) = lim

f ( )g(t + ) d.
T

In like manner we verify that

(yu , u)(t) =

h
N,D (t )(u)( ) d,

where h
N,D (t) = hN,D (t). Therefore we have (yu ) = hN,D hN,D (u), where signifies

convolution. One readily verifies that the Fourier transform of hN,D is the complex conjugate
of the Fourier transform of hN,D . Therefore

sup
uL (,)
uLpow (,)
u not zero

where

sup
uL (,)
uLpow (,)
u not zero

kyu kpow
.
kukpow

During the course of the proof of part (ix) we showed that there exists a power signal u with
pow(u) = 1 with the property that pow(yu ) = kTN,D k . Therefore, this part of the theorem
follows.
(viii) For k 1 define
(
k, t (k, k + k13 )
uk (t) =
0, otherwise,
and define an input u by
u(t) =

X
k=1

(yu )() = TN,D (i)TN,D (i)(u)(),

kyu kpow
kuk

uk (t).

5.3 Norm interpretations of BIBO stability

22/10/2004

For T 1 let k(T ) be the largest integer less than T .

Pk(T )1
1

Z T
Pk=1 k ,
k(T ) 1
2
u (t) dt =
,
k
Pk=1
T
k(T )1 1

1
+ t k(T
,
k=1
k
)

179

180

with a2j + b2j 6= 0, j = 1, . . . , k. In particular, since Re(p) 0, yu 6 L2 (, ).

(v) = (ii)

One computes
T k(T ) 1
T k(T )
t [k(T ) 1, k(T )].

1
T 0 2T

Since

Z N
1
u2 (t) dt
k 2N N
N
X
1
= lim
.
N
k
k=1

u2 (t) dt = lim

N
X
1
k=1

we have
1
T 0 2T

Z
<
1

u2 (t) dt = lim

ln N
= 0.
N

Thus u Lpow .

Note that our notion of BIBO stability exactly coincides with L L stability. The
following result summarises this along with our other notions of stability.
5.22 Theorem Let (N, D) be a proper SISO linear system in input/output form. The following statements are equivalent:
(i) (N, D) is BIBO stable;
(ii) TN,D RH+
;
(iii) (N, D) is L2 L2 -stable;
(iv) (N, D) is L L -stable;
(v) (N, D) is Lpow Lpow -stable.
Furthermore, if any of the preceding three conditions hold then
kTN,D k2,2 = kTN,D kpowpow = kTN,D k .
Proof We shall only prove those parts of the theorem not covered by Theorem 5.21, or
other previous results.
(iii) = (ii) We suppose that TN,D 6 RH+
so that D has a root in C+ . Let a < 0 have the
property that all roots of D lie to the right of { s C | Re(s) = a}. Thus if u(t) = eat 1(t)
then u L2 (, ). Let p C+ be a root for D of multiplicity k. Then
yu (s) = R1 (s) +

k
X
R2,j (s)
,
(s
p)k
j=1

where R1 , R2,1 , . . . , R2,k are rational functions analytic at p. Taking inverse Laplace transforms gives
k
X

yu (t) = y1 (t) +
tj eRe(p)t aj cos(Im(p)t) + bj sin(Im(p)t)
j=1

5.4 Liapunov methods

1
dt = ln N
t

lim

22/10/2004

Note that the theorem can be interpreted as saying that a system is BIBO stable if and
only if the energy/power of the output corresponding to a finite energy/power input is also
finite (here one thinks of the L2 -norm of a signal as a measure of its energy and of pow as its
power). At the moment, it seems like this additional characterisation of BIBO stability in
terms of L2 L2 and Lpow Lpow -stability is perhaps pointless. But the fact of the matter
is that this is far from the truth. As we shall see in Section 8.5, the use of the L2 -norm to
characterise stability has valuable implications for quantitative performance measures, and
their achievement through H methods. This is an approach given a thorough treatment
by Doyle, Francis, and Tannenbaum [1990] and Morris [2000].

Thus we have
lim

5 Stability of control systems

Finish

Liapunov methods for stability are particularly useful in the context of stability for
nonlinear differential equations and control systems. However, even for linear systems where
there are more concrete stability characterisations, Liapunov stability theory is useful as it
gives a collection of results that are useful, for example, in optimal control for such systems.
An application along these lines is the subject of Section 14.3.2. These techniques were
pioneered by Aleksandr Mikhailovich Liapunov (18571918); see [Liapunov 1893].
5.4.1 Background and terminology The Liapunov method for determining stability
has a general character that we will present briefly in order that we may understand the
linear results in a larger context. Let us consider a vector differential equation

x(t)
= f (x(t))

(5.8)

and an equilibrium point x0 for f ; thus f (x0 ) = 0. We wish to determine conditions for stability and asymptotic stability of the equilibrium point. First we should provide definitions
for these notions of stability as they are a little different from their linear counterparts.
5.23 Definition The differential equation (5.8) is:
(i) stable if there exists M > 0 so that for every 0 < M there exists < 0 so that if
kx(0) x0 k < then kx(t) x0 k < for every t > 0;
(ii) asymptotically stable if there exists M > 0 so that the inequality kx(0) x0 k < M
implies that limt kx(t) x0 k = 0.

Our definition differs from the definitions of stability for linear systems in that it is only
local. We do not require that all solutions be bounded in order that x0 be stable, only those
whose initial conditions are sufficiently close to x0 (and similarly for asymptotic stability).
The Liapunov idea for determining stability is to find a function V that has a local
minimum at x0 and whose time derivative along solutions of the differential equation (5.8)
is nonpositive. To be precise about this, let us make a definition.
5.24 Definition A function V : Rn R is a Liapunov function for the equilibrium point
x0 of (5.8) if
(i) V (x0 ) = 0,

Finish

22/10/2004

5.4 Liapunov methods

181

182

5 Stability of control systems

22/10/2004

(ii) V (x) 0, and

d
(iii) there exists M > 0 so that if kx(0) x0 k < M then dt
V (x(t)) 0.
If the nonstrict inequalities in parts (ii) and (iii) are strict, then V is a proper Liapunov
function.

Another characterisation of positive-definiteness involves the principal minors of M . The

following result is not entirely trivial, and a proof may be found in [Gantmacher 1959a].

We will not prove the following theorem as we shall prove it in the cases we care about in
the next section. Readers interested in the proof may refer to, for example, [Khalil 1996].

5.28 Theorem A symmetric n n matrix M is positive-definite if and only if all principal

minors of M are positive.

5.25 Theorem Consider the differential equation (5.8) with x0 an equilibrium point. The
following statements hold:
(i) x0 is stable if there is a Liapunov function for x0 ;
(ii) x0 is asymptotically stable if there is a proper Liapunov function for x0 .
Although we do not prove this theorem, it should nonetheless seem reasonable, particularly
d
V (x(t)) < 0 and since x0 is a strict
the second part. Indeed, since in this case we have dt
local minimum for V , it stands to reason that all solutions should be tending towards this
strict local minimum as t .
Of course, we are interested in linear differential equations of the form

x(t)
= Ax(t).
Our interest is in Liapunov functions of a special sort. We shall consider Liapunov functions
that are quadratic in x. To define such a function, let P Rnn be symmetric and let
V (x) = xt P x. We then compute

5.29 Proposition If M Rnn is positive-definite then there exists , > 0 so that for every
x Rn \ {0} we have
xt x < xt M x < xt x.
Proof Let T Rnn be a matrix for which D = T M T 1 is diagonal. Recall that T can
be chosen so that it is orthogonal, i.e., so that its rows and columns are orthonormal bases
for Rn . It follows that T 1 = T t . Let us also suppose that the diagonal elements d1 , . . . , dn
of D are ordered so that d1 d2 dn . Let us define = 21 d1 and = 2dn . Since for
x = (x1 , . . . , xn ) we have
n
X
xt Dx =
di x2i ,
i=1

it follows that
xt x < xt Dx < xt x

xt M x = xt T t DT x = (T x)t D(T x),

Note that the matrix Q = At P P A is itself symmetric. Now, to apply Theorem 5.25
we need to be able to characterise when the functions xt P x and xt Qx is nonnegative. This
we do with the following definition.

the result follows.

With this background and notation, we are ready to proceed with the results concerning
Liapunov functions for linear differential equations.
5.4.2 Liapunov functions for linear systems The reader will wish to recall from
Remark 2.18 our discussion of observability for MIMO systems, as we will put this to use in
this section. A Liapunov triple is a triple (A, P , Q) of n n real matrices with P and Q
symmetric and satisfying
At P + P A = Q.

The matter of determining when a matrix is positive-(semi)definite or negative(semi)definite is quite a simple matter in principle when one remembers that a symmetric
matrix is guaranteed to have real eigenvalues. With this in mind, we have the following
result whose proof is a simple exercise.
5.27 Proposition For M Rnn be symmetric the following statements hold:
(i) M is positive-definite if and only if spec(M ) C+ R;
(ii) M is negative-definite if and only if spec(M ) C R;
(iii) M is positive-semidefinite if and only if spec(M ) C+ R;

Along these lines, the following result from linear algebra will be helpful to us in the next
section.

for every x Rn \ {0}. Therefore, since

dV (x(t))

= x t (t)P x(t) + xt (t)P x(t)

dt
t
t
= x (t)(A P + P A)x(t).

5.26 Definition Let M Rnn be symmetric.

(i) M is positive-definite (written M > 0) if xt M x > 0 for all x Rn \ {0}.
(ii) M is negative-definite (written M < 0) if xt M x < 0 for all x Rn \ {0}.
(iii) M is positive-semidefinite (written M 0) if xt M x 0 for all x Rn .
(iv) M is negative-semidefinite (written M 0) if xt M x 0 for all x Rn .

(iv) M is negative-semidefinite if and only if spec(M ) C R.

We may now state our first result.

5.30 Theorem Let = (A, b, ct , D) be a SISO linear system and let (A, P , Q) be a Liapunov
triple. The following statements hold:
(i) if P is positive-definite and Q is positive-semidefinite, then is internally stable;
(ii) if P is positive-definite, Q is positive-semidefinite, and (A, Q) is observable, then
is internally asymptotically stable;
(iii) if P is not positive-semidefinite, Q is positive-semi-definite, and (A, Q) is observable,
then is internally unstable.

5.4 Liapunov methods

22/10/2004

183

Proof (i) As in Proposition 5.29, let , > 0 have the property that
t

x x < x P x < x x
for x Rn \ {0}. Let V (x) = xt P x. We compute
dV (x(t))
= xt (At P + P A)x = xt Qx,
dt
since (A, P , Q) is a Liapunov triple. As Q is positive-semidefinite, this implies that
Z t
dV (x(t))
V (x(t)) V (x(0)) =
dt 0
dt
0
for all t 0. Thus, for t 0,
xt (t)P x(t) xt (0)P x(0)
= xt (t)x(t) < xt (0)x(0)

= xt (t)x(t) < xt (0)x(0)

= kx(t)k <
kx(0)k .

Thus kx(t)k is bounded for all t 0, and for linear systems, this implies internal stability.
(ii) We suppose that P is positive-definite, Q is positive-semidefinite, (A, Q) is observable, and that is not internally asymptotically stable. By (i) we know is stable, so it
must be the case that A has at least one eigenvalue on the imaginary axis, and therefore
a nontrivial periodic solution x(t). From our characterisation of the matrix exponential in
Section B.2 we know that this periodic solution evolves in a two-dimensional subspace that
we shall denote by L. Whats more, every solution of x = Ax with initial condition in L is
periodic and remains in L. This implies that L is A-invariant. Indeed, if x L then
eAt x x
Ax = lim
L
t0
t
since x, eAt x L. We also claim that the subspace L is in ker(Q). To see this, suppose that
the solutions on L have period T . If V (x) = xt P x, then for any solution x(t) in L we have
Z T
Z T
dV (x(t))
0 = V (x(T )) V (x(0)) =
dt =
xt (t)Qx(t) dt.
dt
0
0
Since Q is positive-semidefinite this implies that xt (t)Qx(t) = 0. Thus L ker(Q), as
claimed. Thus, with our initial assumptions, we have shown the existence of an nontrivial
A-invariant subspace of ker(Q). This is a contradiction, however, since (A, Q) is observable.
It follows, therefore, that is internally asymptotically stable.
(iii) Since Q is positive-semidefinite and (A, Q) is observable, the argument from (ii)
shows that there are no nontrivial periodic solutions to x = Ax. Thus this part of the theorem will follow if we can show that is not internally asymptotically stable. By hypothesis,
Rn so that V (
tP x
< 0. Let x(t) be the solution of x = Ax with
there exists x
x) = x
. As in the proof of (i) we have V (x(t)) V (
x(0) = x
x) < 0 for all t 0 since Q is
positive-semidefinite. If we denote
r = inf { kxk | V (x) V (
x)} ,
then we have shown that kxk (t) > r for all t 0. This prohibits internal asymptotic
stability, and in this case, internal stability.

184

5 Stability of control systems

22/10/2004

5.31 Example (Example 5.4 contd) We again look at the 2 2 matrix

0
1
A=
,
b a
letting = (A, b, ct , D) for some b, c, and D. For this example, there are various cases
to consider, and we look at them separately in view of Theorem 5.30. In the following
discussion, the reader should compare the conclusions with those of Example 5.4.
1. a = 0 and b = 0: In this case, we know the system is internally unstable. However, it
turns out to be impossible to find a symmetric P and a positive-semidefinite Q so that
(A, P , Q) is a Liapunov triple, and so that (A, Q) is observable (cf. Exercise E5.16).
Thus we cannot use part (iii) of Theorem 5.30 to assert internal instability. We are off
to a bad start! But things start to look better.
2. a = 0 and b > 0: The matrices

b 0
0 0
P =
, Q=
0 1
0 0
have the property that (A, P , Q) are a Liapunov triple. Since P is positive-definite and
Q is positive-semidefinite, internal stability follows from part (i) of Theorem 5.30. Note
that (A, Q) is not observable, so internal asymptotic stability cannot be concluded from
part (ii).
3. a = 0 and b < 0: If we define

1 0 1
b 0
P =
, Q=
,
0 1
2 1 0
then one verifies that (A, P , Q) are a Liapunov triple. Since P is not positive-semidefinite
(its eigenvalues are { 21 }) and since Q is positive-definite and (A, Q) is observable (Q
is invertible), it follows from part (iii) of Theorem 5.30 that the system is internally
unstable.
4. a > 0 and b = 0: Here we take
2

a a
0 0
P =
, Q=
a 2
0 2a
and verify that (A, P , Q) is a Liapunov triple. The eigenvalues of P are { 12 (a2 + 2

a4 + 4)}. One may verify that a2 + 2 > a4 + 4, thus P is positive-definite. We also

compute

0
0
0 2a
,
O(A, Q) =
0
0
0 2a2
verifying that (A, Q) is not observable. Thus from part (i) of Theorem 5.30 we conclude
that is internally stable, but we cannot conclude internal asymptotic stability from (ii).
5. a > 0 and b > 0: Here we take

b 0
0 0
P =
, Q=
0 1
0 2a

22/10/2004

5.4 Liapunov methods

185

having the property that (A, P , Q) is a Liapunov triple. We compute

0
0
0
2a
,
O(A, Q) =
0
0
2ab 2a2
implying that (A, Q) is observable. Since P is positive-definite, we may conclude from
part (ii) of Theorem 5.30 that is internally asymptotically stable.
6. a > 0 and b < 0: Again we use

b 0
0 0
P =
, Q=
.
0 1
0 2a

186

5 Stability of control systems

22/10/2004

5.32 Theorem Let = (A, b, ct , D) be a SISO linear system with A Hurwitz. The following
statements hold:
(i) for any symmetric Q Rnn there exists a unique symmetric P Rnn so that
(A, P , Q) is a Liapunov triple;
(ii) if Q is positive-semidefinite with P the unique symmetric matrix for which (A, P , Q)
is a Liapunov triple, then P is positive-semidefinite;
(iii) if Q is positive-semidefinite with P the unique symmetric matrix for which (A, P , Q)
is a Liapunov triple, then P is positive-definite if and only if (A, Q) is observable.
Proof (i) We claim that if we define
Z
P =

eA t QeAt dt

(5.9)

Now, since P is not positive-semidefinite, from part (iii) of Theorem 5.30, we conclude
that is internally unstable.
7. a < 0 and b = 0: This case is much like case 1 in that the system is internally unstable,
but we cannot find a symmetric P and a positive-semidefinite Q so that (A, P , Q) is a
Liapunov triple, and so that (A, Q) is observable (again see Exercise E5.16).
8. a < 0 and b > 0: We note that if

0 0
b 0
P =
, Q=
,
0 1
0 2a

then (A, P , Q) is a Liapunov triple. First note that since A is Hurwitz, the integral does
indeed converge. We also have
Z t
Z t

At P + P A = At
eA t QeAt dt +
eA t QeAt dt A
0
Z 0
d At t At
e Qe
dt
=
dt
0

At t
At
= e Qe 0 = Q,
as desired. We now show that P as defined is the only symmetric matrix for which (A, P , Q)
also has the property that (A, P
, Q) is a Liapunov
is a Liapunov triple. Suppose that P
P . Then one sees that At + A = 0nn . If we let
triple, and let = P

then (A, P , Q) is a Liapunov triple. We also have

0
0
0 2a
.
O(A, Q) =
0
0
2
2ab 2a

(t) = eA t eAt ,
then

Thus (A, Q) is observable. Since P is not positive-definite and since Q is positivesemidefinite, from part (iii) of Theorem 5.30 we conclude that is internally unstable.
9. a < 0 and b < 0: Here we again take

0 0
b 0
P =
, Q=
.
0 1
0 2a
The same argument as in the previous case tells us that is internally unstable.

Note that in two of the nine cases in the preceding example, it was not possible to apply
Theorem 5.30 to conclude internal instability of a system. This points out something of a
weakness of the Liapunov approach, as compared to Theorem 5.2 which captures all possible
cases of internal stability and instability. Nevertheless, the Liapunov characterisation of
stability can be a useful one in practice. It is used by us in Chapters 14 and 15.
While Theorem 5.30 tells us how we certain Liapunov triples imply certain stability
properties, often one wishes for a converse to such results. Thus one starts with a system
= (A, b, ct , D) that is stable in some way, and one wishes to ascertain the character of
the corresponding Liapunov triples. While the utility of such an exercise is not immediately
obvious, it will come up in Section 14.3.2 when characterising solutions of an optimal control
problem.

d(t)
t
= eA t At + A eAt = 0nn .
dt
Therefore (t) is constant, and since (0) = , it follows that (t) = for all t. However,
since A is Hurwitz, it also follows that limt (t) = 0nn . Thus = 0nn , so that
= P.
P
(ii) If P is defined by (5.9) we have
Z
t

xt P x =
eAt x Q eAt x dt.
0

Therefore, if Q is positive-semidefinite, it follows that P is positive-semidefinite.

(iii) Here we employ a lemma.
1 Lemma If Q is positive-semidefinite then (A, Q) is observable if and only if the matrix P
defined by (5.9) is invertible.
Proof First suppose that (A, Q) is observable and let x ker(P ). Then
Z
t

eAt x Q eAt x dt = 0.
0

5.4 Liapunov methods

22/10/2004

187

188

5 Stability of control systems

22/10/2004

Since Q is positive-semidefinite, this implies that eAt x ker(Q) for all t. Differentiating
this inclusion with respect to t k times in succession gives Ak eAt x ker(Q) for any k > 0.
Evaluating at t = 0 shows that x is in the kernel of the matrix

Q
QA

O(A, Q) =
.
..

.
QAn1

If we let a = 2 and b = 1 then we compute

1+t
t
.
eAt = et
t 1 t

Since (A, Q) is observable, this implies that x = 0. Thus we have shown that ker(P ) = {0},
or equivalently that P is invertible.
Now suppose that P is invertible. Then the expression
Z
t

eAt x Q eAt x dt

Note that our choice of Q is positive-definite and that (A, Q) is observable. Therefore,
part (iii) of Theorem 5.32 implies that P is positive-definite. It may be verified that the
P s computed above are indeed positive-definite.
However, it is not necessary to make such hard work of this. After all, the equation

Again, a direct computation using (5.10) gives

3 1
P = 21 21 .
2

At P + P A = Q

is zero if and only if x = 0. Since Q is positive-semidefinite, this means that the expression
t

eAt x Q eAt x
is zero if and only if x = 0. Since eAt is invertible, this implies that Q must be positivedefinite, and in particular, invertible. In this case, (A, Q) is clearly observable.
H
With the lemma at hand, the remainder of the proof is straightforward. Indeed, from
part (ii) we know that P is positive-semidefinite. The lemma now says that P is positivedefinite if and only if (A, Q) is observable, as desired.

5.33 Example (Example 5.4 contd) We resume looking at the case where

0
1
A=
.
b a
Let us look at a few cases to flush out some aspects of Theorem 5.32.
1. a > 0 and b > 0: This is exactly the case when A is Hurwitz, so that part (i) of Theorem 5.32 implies that for any symmetric Q there is a unique symmetric P so that
(A, P , Q) is a Liapunov triple. As we saw in the proof of Theorem 5.32, one can determine P with the formula
Z
t
eA t QeAt dt.
(5.10)
P =
0

However, to do this in this example is a bit tedious since we would have to deal with
the various cases of a and b to cover all the various forms taken by eAt . For example,
suppose we take

1 0
Q=
0 1
and let a = 2 and b = 2. Then we have

cos t + sin t
sin t
et = et
2 sin t
cos t sin t
In this case one can directly apply (5.10) with some effort to get
5 1
P = 41 43 .
4

is nothing but a linear equation for P . That A is Hurwitz merely ensures a unique
solution for any symmetric Q. If we denote

p
p
P = 11 12
p12 p22
and continue to use
Q=

1 0
,
0 1

then we must solve the linear equations

p
p
0
1
1 0
0 b p11 p12
=
,
+ 11 12
p12 p22 b a
0 1
1 a p12 p22
subject to a, b > 0. One can then determine P for general (at least nonzero) a and b to
be
a2 +b+b2

1
2ab
2b
P =
1
b+1 .
2b

2ab

In this case, we are guaranteed that this is the unique P that does the job.
2. a 0 and b = 0: As we have seen, in this case there is not always a solution to the
equation
At P + P A = Q.
(5.11)
Indeed, when Q is positive-semidefinite and (A, Q) is observable, this equation is guaranteed to not have a solution (see Exercise E5.16). This demonstrates that when A is
not Hurwitz, part (i) of Theorem 5.32 can fail in the matter of existence.
3. a > 0 and b = 0: In this case we note that for any C R the matrix
2
a a
P0 = C
a 1
satisfies At P + P A = 022 . Thus if P is any solution to (5.11) then P + P 0 is also a
solution. If we take

0 0
Q=
,
0 2a

22/10/2004

5.5 Identifying polynomials with roots in C

189

then, as we saw in Theorem 5.30, if

2
a a
P =
,
a 2

190

From our discussion of Section 5.2 we see that it is very important that T have poles
only in the negative half-plane. However, checking that such a condition holds may not be
so easy. One way to do this is to establish conditions on the coefficients of the denominator
polynomial of T (after making pole/zero cancellations, of course). In this section, we present
three methods for doing exactly this. We also look at a test for the poles lying in C when
we only approximately know the coefficients of the polynomial. We shall generally say that
a polynomial all of whose roots lie in C is Hurwitz .
It is interesting to note that the method of Edward John Routh (18311907) was developed in response to a famous paper of James Clerk Maxwell2 (18311879) on the use of
governors to control a steam engine. This paper of Maxwell [1868] can be regarded as the
first paper in mathematical control theory.
5.5.1 The Routh criterion For the method of Routh, we construct an array involving
the coefficients of the polynomial in question. The array is constructed inductively, starting
with the first two rows. Thus suppose one has two collections a11 , a12 , . . . and a21 , a22 , . . . of
numbers. In practice, this is a finite collection, but let us suppose the length of each collection
to be indeterminate for convenience. Now construct a third row of numbers a31 , a32 , . . . by
defining
a11 a1,k+1a3k = a21 a1,k+1 a11 a2,k+1 . Thus a3k is minus the determinant of the matrix
a21 a2,k+1 . In practice, one writes this down as follows:
a11
a12
a1k
a21
a22
a2k
a21 a12 a11 a22 a21 a13 a11 a23 a21 a1,k+1 a11 a2,k+1
One may now proceed in this way, using the second and third row to construct a fourth row,
the third and fourth row to construct a fifth row, and so on. To see how to apply this to a
given polynomial P R[s]. Define two polynomials P+ , P R[s] as the even and odd part
of P . To be clear about this, if
P (s) = p0 + p1 s + p2 s2 + p3 s3 + + pn1 sn1 + pn sn ,
then
P+ (s) = p0 + p2 s + p4 s2 + . . . ,

P (s) = p1 + p3 s + p5 s2 + . . . .

Note then that P (s) = P+ (s2 ) + sP (s2 ). Let R(P ) be the array constructed as above with
the first two rows being comprised of the coefficients of P+ and P , respectively, starting
2

Maxwell, of course, is better known for his famous equations of electromagnetism.

22/10/2004

with the coefficients of lowest powers of s, and increasing to higher powers of s. Thus the
first three rows of R(P ) are
p0
p2
p2k
p1
p3
p2k+1
p1 p2 p0 p3 p1 p4 p0 p5 p1 p2k+2 p0 p2k+3
..
.. ..
.. ..
.
. .
. .

then (A, P , Q) is a Liapunov triple. What we have shown i that (A, P + P 0 , Q) is also
a Liapunov triple. Thus part (i) of Theorem 5.32 can fail in the matter of uniqueness
when A is not Hurwitz.

5.5 Identifying polynomials with roots in C

5 Stability of control systems

In making this construction, a zero is inserted whenever an operation is undefined. It is

readily determined that the first column of R(P ) has at most n + 1 nonzero components.
The Routh array is then the first column of the first n + 1 rows.
With this as setup, we may now state a criterion for determining whether a polynomial
is Hurwitz.
5.34 Theorem (Routh [1877]) A polynomial
P (s) = sn + pn1 sn1 + + p1 s + p0 R[s]
is Hurwitz if and only if all elements of the Routh array corresponding to R(P ) are positive.
Proof Let us construct a sequence of polynomials as follows. We let P0 = P+ and P1 = P
and let

P2 (s) = s1 P1 (0)P0 (s) P0 (0)P1 (s) .
Note that the constant coefficient of P1 (0)P0 (s) P0 (0)P1 (s) is zero, so this does indeed
define P2 as a polynomial. Now inductively define

Pk (s) = s1 Pk1 (0)Pk2 (s) Pk2 (0)Pk1 (s)
for k 3. With this notation, we have the following lemma that describes the statement of
the theorem.
1 Lemma The (k + 1)st row of R(P ) consists of the coefficients of Pk with the constant coefficient in the first column. Thus the hypothesis of the theorem is equivalent to the condition
that P0 (0), P1 (0), . . . , Pn (0) all be positive.
Proof We have P0 (0) = p0 , P1 (0) = p1 , and P2 (0) = p1 p2 p0 p3 , directly from the definitions. Thus the lemma holds for k = 0, 1, 2. Now suppose that the lemma holds for k 3.
Thus the kth and the (k + 1)st rows of R(P ) are the coefficients of the polynomials
Pk1 (s) = pk1,0 + pk1,1 s +
and
Pk = pk,0 + pk,1 s + ,
respectively. Using the definition of Pk+1 we see that Pk+1 (0) = pk,0 pk1,1 pk1,0 pk,1 .
However, this is exactly the term as it would appear in first column of the (k + 2)nd row of
R(P ).
H
Now note that P (s) = P0 (s2 ) + sP1 (s2 ) and define Q R[s] by Q(s) = P1 (s2 ) + sP2 (s2 ).
One may readily verify that deg(Q) n1. Indeed, in the proof of Theorem 5.36, a formula
for Q will be given. The following lemma is key to the proof. Let us suppose for the moment
that pn is not equal to 1.

22/10/2004

5.5 Identifying polynomials with roots in C

191

192

5 Stability of control systems

22/10/2004

the number of roots of Q within C must equal n 1 as deg(Q) = n 1. In particular,

P can have at most one root in C+ . This root, then, must be real, and let us denote it by
z0 > 0. Thus P (s) = P (s)(s z0 ) where P is Hurwitz. By Exercise E5.18 it follows that all
coefficients of P are positive. If we write

2 Lemma The following statements are equivalent:

(i) P is Hurwitz and pn > 0;
(ii) Q is Hurwitz, qn1 > 0, and P (0) > 0.
Proof We have already noted that P (s) = P0 (s2 ) + sP1 (s2 ). We may also compute

Q(s) = P1 (s2 ) + s1 P1 (0)P0 (s2 ) P0 (0)P1 (s2 ) .
(5.12)

P = pn1 sn1 + pn2 sn2 + + p1 s + p0 ,

then
P (s) = pn1 sn + (
pn2 z0 pn1 )sn1 + + (
p0 z0 p1 )s p0 z0 .

For [0, 1] define Q (s) = (1 )P (s) + Q(s), and compute

Q (s) = (1 ) + s1 P1 (0) P0 (s2 ) + (1 )s + s1 P0 (0) P1 (s2 ).

Thus the existence of a root z0 C+ contradicts the fact that P (0) > 0. Note that we have
also shown that pn > 0.
H

The polynomials P0 (s2 ) and P1 (s2 ) are even so that when evaluated on the imaginary axis
they are real. Now we claim that the roots of Q that lie on the imaginary axis are independent of , provided that P (0) > 0 and Q(0) > 0. First note that if P (0) > 0 and Q(0) > 0
then 0 is not a root of Q . Now if i0 is a nonzero imaginary root then we must have

(1 ) i01 P1 (0) P0 (02 ) + (1 )i0 + + i01 P0 (0) P1 (02 ) = 0.
Balancing real and imaginary parts of this equation gives
(1 )P0 (02 ) + P1 (02 ) = 0

01 P0 (0)P1 (02 ) P1 (0)P0 (02 ) + 0 (1 )P1 (02 ).

(5.13)

If we think of this as a homogeneous linear equation in P0 (02 ) and P1 (02 ) one determines
that the determinant of the coefficient matrix is

01 (1 )2 02 + ((1 )P0 (0) + P1 (0)) .
This expression is positive for [0, 1] since P (0), Q(0) > 0 implies that P0 (0), P1 (0) > 0.
To summarise, we have shown that, provided P (0) > 0 and Q(0) > 0, all imaginary axis
roots i0 of Q satisfy P0 (02 ) = 0 and P1 (02 ) = 0. In particular, the imaginary axis
roots of Q are independent of [0, 1] in this case.
(i) = (ii) For [0, 1] let
(
n,
[0, 1)
N () =
n 1, = 1.
Thus N () is the number of roots of Q . Now let
Z = { z,i | i {1, . . . , N ()}}
be the set of roots of Q . Since P is Hurwitz, Z0 C . Our previous computations then
show that Z iR = for [0, 1]. Now if Q = Q1 were to have a root in C+ this would
mean that for some value of one of the roots of Q would have to lie on the imaginary
axis, using the (nontrivial) fact that the roots of a polynomial are continuous functions of
its coefficients. This then shows that all roots of Q must lie in C . That P (0) > 0 is a
consequence of Exercise E5.18 and P being Hurwitz. One may check that qn1 = p1 pn
so that qn1 > 0 follows from Exercise E5.18 and pn > 0.
(ii) = (i) Let us adopt the notation N () and Z from the previous part of the proof.
Since Q is Hurwitz, Z1 C . Furthermore, since Z iR = , it follows that for [0, 1],

Now we proceed with the proof proper. First suppose that P is Hurwitz. By successive
applications of Lemma 2 it follows that the polynomials
Qk (s) = Pk (s2 ) + sPk+1 (s2 ),

k = 1, . . . , n,

are Hurwitz and that deg(Qk ) = n k, k = 1, . . . , n. Whats more, the coefficient of snk
is positive in Qk . Now, by Exercise E5.18 we have P0 (0) > 0 and P1 (0) > 0. Now suppose
that P0 (0), P1 (0), . . . , Pk (0) are all positive. Since Qk is Hurwitz with the coefficient of the
highest power of s being positive, from Exercise E5.18 it follows that the coefficient of s in
Qk should be positive. However, this coefficient is exactly Pk+1 (0). Thus we have shown
that Pk (0) > 0 for k = 0, 1, . . . , n. From Lemma 1 it follows that the elements of the Routh
array are positive.
Now suppose that one element of the Routh array is nonpositive, and that P is Hurwitz.
By Lemma 2 we may suppose that Pk0 (0) 0 for some k0 {2, 3, . . . , n}. Furthermore,
since P is Hurwitz, as above the polynomials Qk , k = 1, . . . , n, must also be Hurwitz,
with deg(Qk ) = n k where the coefficient of snk in Qk is positive. In particular, by
Exercise E5.18, all coefficients of Qk0 1 are positive. However, since Qk0 1 (s) = Pk0 1 (s2 ) +
sPk0 (s2 ) it follows that the coefficient of s in Qk0 1 is negative, and hence we arrive at a
contradiction, and the theorem follows.

The Routh criterion is simple to apply, and we illustrate it in the simple case of a degree
two polynomial.
5.35 Example Let us apply the criteria to the simplest nontrivial example possible: P (s) =
s2 + as + b. We compute the Routh table to be
b 1
R(P ) = a 0.
a 0

Thus the Routh array is b a a , and its entries are all positive if and only if a, b > 0.
Lets see how this compares
to what we know doing
the calculations by hand. The roots

of P are r1 = a2 + 12 a2 4b and r2 = a2 12 a2 4b. Let us consider the various cases.

1. If a2 4b < 0 then the roots are complex with nonzero imaginary part, and with real
part a. Thus the roots in this case lie in the negative half-plane if and only if a > 0.
2
We also have b > a4 and so b > 0 and hence ab > 0 as in the Routh criterion.
2
2. If a 4b = 0 then the roots are both a, and so lie in the negative half-plane if and
2
only if a > 0. In this case b = a4 and so b > 0. Thus ab > 0 as predicted.

22/10/2004

5.5 Identifying polynomials with roots in C

193

3. Finally we have the case when a2 4b > 0. We have two subcases.

(a) When a > 0 then we have negative half-plane roots if and only if a2 4b < a2 which
means that b > 0. Therefore we have negative half-plane roots if and only a > 0
and ab > 0.

(b) When a < 0 then we will never have all negative half-plane roots since a+ a2 4b
is always positive.
So we see that the Routh criterion provides a very simple encapsulation of the necessary
and sufficient conditions for all roots to lie in the negative half-plane, even for this simple
example.

194

5 Stability of control systems

and when n is odd we have

p1 pn1 p0 pn
pn
0
0
p1 pn3 p0 pn2 pn2 p1 pn1 p0 pn pn

..
..
..
..
H(Q) =
.
.
.
.

0
0
0
0
0
0
0
0

P (s) = s + pn1 s

1 0
0 p1

0 p0

..
T = ...
.

0 0

0 0
0 0

+ + p1 s + p0 R[s].

We denote the Hurwitz matrix by H(P ) Rnn and define

pn1
1
0
0

pn3 pn2 pn1

H(P ) = pn5 pn4 pn3 pn2

..
..
..
..
...
.
.
.
.
0
0
0
0

it by

0
0

0
.
..
.
p0

p1
p0

T = .
0

0
0

5.36 Theorem (Hurwitz [1895]) A polynomial

P (s) = sn + pn1 sn1 + + p1 s + p0 R[s]
is Hurwitz if and only if the n Hurwitz determinants 1 , . . . , n are positive.
Proof Let us begin by resuming with the notation from the proof of Theorem 5.34. In
particular, we recall the definition of Q(s) = P1 (s2 ) + sP2 (s2 ). We wish to compute H(Q) so
we need to compute Q in terms of the coefficients of P . A computation using the definition
of Q and P2 gives
Q(s) = p1 + (p1 p2 p0 p3 )s + p3 s2 + (p1 p4 p0 p5 )s3 + .

...

0
0
..
.

p1 p2 p0 p3

0
0
..
.

p1 p2 p0 p3

0
0

.. .
.

p3
p1

0
0
1
..
.

...

0
0

..
.

0 0

1 0
0 1

0
0
0
..
.

0 p1
0 p0
0
0

when n is even and by

Any terms in this matrix that are not defined are taken to be zero. Of course we also take
pn = 1. Now define H(P )k Rkk , k = 1, . . . , n, to be the matrix of elements H(P )ij ,
i, j = 1, . . . , k. Thus H(P )k is the matrix formed by taking the upper left k k block from
H(P ). Also define k = det H(P )k .
With this notation, the Hurwitz criterion is as follows.

One can then see that when n is even we have

pn1
p1 pn
0
0
pn3 p1 pn2 p0 pn1 pn1 p1 pn

..
..
..
H(Q) = ...
.
.
.

0
0
0
0
0
0
0
0

..
.

Now define T Rnn by

5.5.2 The Hurwitz criterion The method we discuss in this section is work of Adolf
Hurwitz (18591919). The key ingredient in the Hurwitz construction is a matrix formed
from the coefficients of a polynomial
n

22/10/2004

0
0

..
.

p3
p1

0
0
1
0
.. . .
..
.
.
.
0 p1
0 p0
0
0

0
0

..
.

0 0

1 0
0 1
0
0
..
.

when n is odd. One then verifies by direct calculation that

..
.

p4
H(P )T = H(Q)
.

p2
0 0 p0

(5.14)

1, . . . ,
n1 be the similar
We now let 1 , . . . , n be the determinants defined above and let
determinants corresponding to H(Q). A straightforward computation using (5.14) gives the

following relationships between the s and the s:

1 = p1
k+1 =

( b k c
k , k even
p1 2
,
d k2 e
k , k odd
p1

k = 1, . . . , n 1,

(5.15)

where bxc gives the greatest integer less than or equal to x and dxe gives the smallest integer
greater than or equal to x.
With this background notation, let us proceed with the proof, first supposing that P is
Hurwitz. In this case, by Exercise E5.18, it follows that p1 > 0 so that 1 > 0. By Lemma 2
1 > 0. A trivial induction
of Theorem 5.34 it also follows that Q is Hurwitz. Thus
argument on n = deg(P ) then shows that 2 , . . . , n > 0.
Now suppose that one of 1 , . . . , n is nonpositive and that P is Hurwitz. Since Q is
then Hurwitz by Lemma 2 of Theorem 5.34, we readily arrive at a contradiction, and this
completes the proof.

5.5 Identifying polynomials with roots in C

22/10/2004

195

The Hurwitz criterion is simple to apply, and we illustrate it in the simple case of a
degree two polynomial.
5.37 Example (Example 5.35 contd) Let us apply the criteria to our simple example of
P (s) = s2 + as + b. We then have

a 1
H(P ) =
0 b
We then compute 1 = a and 2 = ab. Thus 1 , 2 > 0 if and only if a, b > 0. This agrees
with our application of the Routh method to the same polynomial in Example 5.35.

5.5.3 The Hermite criterion We next look at a manner of determining whether a
polynomial is Hurwitz which makes contact with the Liapunov methods of Section 5.4. This
method is due to Charles Hermite (18221901) [see Hermite 1854]. Let us consider, as usual,
a monic polynomial of degree n:
n

P (s) = s + pn1 s

+ + p1 s + p0 .

Corresponding to such a polynomial, we construct its Hermite matrix as the n n matrix

P (P ) given by
Pi
k+i

k=1 (1) pnk+1 pnij+k , j i, i + j even

P (P )ij = P (P )ji ,
j < i, i + j even

0,
i + j odd.
As usual, in this formula we take pi = 0 for i < 0. One can get an idea of how this matrix
is formed by looking at its appearance for small values of n. For n = 2 we have

pp
0
P (P ) = 1 2
,
0 p0 p1
for n = 3 we have

p2 p3
0
p0 p3
0 ,
P (P ) = 0 p1 p2 p0 p3
p0 p3
0
p0 p1

and for n = 4 we have

p3 p4
0
p1 p4
0
0 p2 p3 p1 p4
0
p0 p3
.
P (P ) =
p1 p4
0
p1 p2 p0 p3
0
0
p0 p3
0
p0 p1
The following theorem gives necessary and sufficient conditions for P to be Hurwitz based
on its Hermite matrix. The slick proof using Liapunov methods comes from the paper of
Parks [1962].
5.38 Theorem A polynomial
P (s) = sn + pn1 sn1 + + p1 s + p0 R[s]
is Hurwitz if and only if P (P ) is positive-definite.

196

5 Stability of control systems

22/10/2004

Proof Let

pn1 pn2
1
0

1
A(P ) = 0
..
..
.
.
0
0

p1 p0

0
0

0
0
,
..
..
...
.
.

1
0

pn1
0

b(P ) = pn3 .
0

..
.

An unenjoyable computation gives

P (P )A(P ) + A(P )t P (P ) = b(P )b(P )t .
First suppose that P (P ) is positive-definite. By Theorem 5.30(i), since b(P )b(P )t is positivesemidefinite, A(P ) is Hurwitz. Conversely, if A(P ) is Hurwitz, then there is only one
symmetric P so that
P A(P ) + A(P )t P = b(P )b(P )t ,
this by Theorem 5.32(i). Since P (P ) satisfies this relation even when A(P ) is not Hurwitz,
it follows that P (P ) is positive-definite. The theorem now follows since the characteristic
polynomial of A(P ) is P .

Let us apply this theorem to our favourite example.
5.39 Example (Example 5.35 contd) We consider the polynomial P (s) = s2 + as + b which
has the Hermite matrix

a 0
P (P ) =
.
0 ab
Since this matrix is diagonal, it is positive-definite if and only if the diagonal entries are
zero. Thus we recover the by now well established condition that a, b > 0.

The Hermite criterion, Theorem 5.38, does indeed record necessary and sufficient conditions for a polynomial to be Hurwitz. However, it is more computationally demanding than
it needs to be, especially for large polynomials. Part of the problem is that the Hermite
matrix contains so many zero entries. To get conditions involving smaller matrices leads
to the so-called reduced Hermite criterion which we now discuss. Given a degree n
polynomial P with its Hermite matrix P (P ), we define matrices C(P ) and D(P ) as follows:
1. C(P ) is obtained by removing the even numbered rows and columns of P (P ) and
2. D(P ) is obtained by removing the odd numbered rows and columns of P (P ).
n+1
and D(P )
Thus, if n is even, C(P ) and D(P ) are n2 n2 , and if n is odd, C(P ) is n+1
2
2
n1
is n1

.
Let
us
record
a
few
of
these
matrices
for
small
values
of
n.
For
n = 2 we have
2
2

C(P ) = p1 p2 , D(P ) = p0 p1 ,
for n = 3 we have
C(P ) =

p2 p3 p0 p3
,
p0 p3 p0 p1

D(P ) = p1 p2 p0 p3 ,

and for n = 4 we have

C(P ) =

p3 p4
p1 p4
,
p1 p4 p1 p2 p0 p3

D(P ) =

p2 p3 p1 p4 p0 p3
.
p0 p3
p0 p1

Let us record a useful property of the matrices C(P ) and D(P ), noting that they are
symmetric.

5.5 Identifying polynomials with roots in C

22/10/2004

197

198

5 Stability of control systems

22/10/2004

5.40 Lemma P (P ) is positive-definite if and only if both C(P ) and D(P ) are positivedefinite.

Also note that C(P )1/2 is invertible, and we shall denote its inverse by C(P )1/2 . Note that
this inverse is also positive-definite. This then gives

Proof For x = (x1 , . . . , xn ) Rn , denote xodd = (x1 , x3 , . . .) and xeven = (x2 , x4 , . . .). A
simple computation then gives

C(P )1/2 A1 (P )C(P )1/2 = C(P )1/2 D(P )C(P )1/2 .

xt P (P )x = xtodd C(P )xodd + xteven D(P )xeven .

The matrix on the right is symmetric, so this shows that A1 (P ) is similar to a symmetric
matrix, allowing us to deduce that A1 (P ) has real eigenvalues. These eigenvalues are also
roots of the characteristic polynomial

(5.16)

Clearly, if C(P ) and D(P ) are both positive-definite, then so too is P (P ). Conversely,
suppose that one of C(P ) or D(P ), say C(P ), is not positive-definite. Thus there exists
x Rn so that xodd 6= 0 and xeven = 0, and for which
xtodd C(P )xodd 0.
From (5.16) it now follows that P (P ) is not positive-definite.

The Hermite criterion then tells us that P is Hurwitz if and only if both C(P ) and D(P )
are positive-definite. The remarkable fact is that we need only check one of these matrices
for definiteness, and this is recorded in the following theorem. Our proof follows that of
Anderson [1972].
5.41 Theorem A polynomial
P (s) = sn + pn1 sn1 + + p1 s + p0 R[s]
is Hurwitz if and only if any one of the following conditions holds:
(i) p2k > 0, k {0, 1, . . . , b n1
c} and C(P ) is positive-definite;
2
(ii) p2k > 0, k {0, 1, . . . , b n1
c} and D(P ) is positive-definite;
2
(iii) p0 > 0, p2k+1 > 0, k {0, 1, . . . , b n2
c} and C(P ) is positive-definite;
2
(iv) p0 > 0, p2k+1 > 0, k {0, 1, . . . , b n2
c} and D(P ) is positive-definite.
2
Proof First suppose that P is Hurwitz. Then all coefficients are positive (see Exercise E5.18)
and P (P ) is positive-definite by Theorem 5.38. This implies that C(P ) and D(P ) are
positive-definite by Lemma 5.40, and thus conditions (i)(iv) hold. For the converse assertion,
the cases when n is even or odd are best treated separately. This gives eight cases to look
at. As certain of them are quite similar in flavour, we only give details the first time an
argument is encountered.
Case 1: We assume (i) and that n is even. Denote
pn2

ppn2 ppn0
pn pn4
pn
1
0

0
0

0
0
A1 (P ) = 0
.
..

..
..
..
...
.

.
.
.
0
0

1
0
A calculation then gives C(P )A1 (P ) = D(P ). Since C(P ) is positive-definite, there exists
an orthogonal matrix R so that RC(P )Rt = , where is diagonal with strictly positive
diagonal entries. Let 1/2 denote the diagonal matrix whose diagonal entries are the square
roots of those of . Now denote C(P )1/2 = Rt 1/2 R, noting that C(P )1/2 C(P )1/2 = C(P ).

sn/2 +

(5.17)

p2
p0
pn2 n/21
s
+ + s + .
pn
pn
pn

Our assumption (i) ensures that is s is real and nonnegative, the value of the characteristic
polynomial is positive. From this we deduce that all eigenvalues of A1 (P ) are negative.
From (5.17) it now follows that D(P ) is positive-definite, and so P is Hurwitz by Lemma 5.40
and Theorem 5.38.
Case 2: We assume (ii) and that n is even. Consider the polynomial P 1 (s) = sn P ( 1s ).
Clearly the roots of P 1 are the reciprocals of those for P . Thus P 1 is Hurwitz if and only
if P is Hurwitz (see Exercise E5.20). Also, the coefficients for P 1 are obtained by reversing
those for P . Using this facts, one can see that C(P 1 ) is obtained from D(P ) by reversing
the rows and columns, and that D(P 1 ) is obtained from C(P ) by reversing the rows and
columns. One can then show that P 1 is Hurwitz just as in Case 1, and from this it follows
that P is Hurwitz.
Case 3: We assume (iii) and that n is odd. In this case we let
pn2

pn pn4
ppn1 0
pn
1
0

0 0

0 0
A2 (P ) = 0

..
..
..
..
...
.
.
.
.
0
0

1 0
and note that one can check to see that
C(P )A2 (P ) =

D(P ) 0
.
t
0
0

(5.18)

As in Case 1, we may define the square root, C(P )1/2 , of C(P ), and ascertain that

D(P ) 0
C(P )1/2 A2 (P )C(P )1/2 = C(P )1/2
C(P )1/2 .
t
0
0
Again, the conclusion is that A2 (P ) is similar to a symmetric matrix, and so must have real
eigenvalues. These eigenvalues are the roots of the characteristic polynomial
s(n+1)/2 +

pn2 (n+1)/21
p1
s
+ + s.
pn
pn

This polynomial clearly has a zero root. However, since (iii) holds, for positive real values of
s the characteristic polynomial takes on positive values, so the nonzero eigenvalues of A2 (P )
must be negative, and there are n+1
1 of these. From this and (5.18) it follows that the
2
matrix

D(P ) 0
t
0
0

22/10/2004

5.5 Identifying polynomials with roots in C

199

has one zero eigenvalue and n+1

1 positive real eigenvalues. Thus D(P ) must be positive2
definite, and P is then Hurwitz by Lemma 5.40 and Theorem 5.38.
Case 4: We assume (i) and that n is odd. As in Case 2, define P 1 (s) = sn P ( 1s ). In this
case one can ascertain that C(P 1 ) is obtained from C(P ) by reversing rows and columns,
and that D(P 1 ) is obtained from D(P ) by reversing rows and columns. The difference
from the situation in Case 2 arises because here we are taking n odd, while in Case 2 it was
even. In any event, one may now apply Case 3 to P 1 to show that P 1 is Hurwitz. Then
P is itself Hurwitz by Exercise E5.20.
Case 5: We assume (ii) and that n is odd. For > 0 define P R[s] by P (s) =
(s + )P (s). Thus the degree of P is now even. Indeed,
P (s) = pn sn+1 + (pn1 + pn )sn + + (p0 + p1 )s + p0 .
One may readily determine that
C(P ) = C(P ) + C
for some matrix C which is independent of . In like manner, one may show that

D(P ) + D 11 D 12
D(P ) =
2
D 12
p0
where D 11 and D 12 are independent of . Since D(P ) is positive-definite and a0 > 0, for
sufficiently small we must have D(P ) positive-definite. From the argument of Case 2 we
may infer that P is Hurwitz, from which it is obvious that P is also Hurwitz.
Case 6: We assume (iv) and that n is odd. We define P 1 (s) = sn P ( 1s ) so that C(P 1 )
is obtained from C(P ) by reversing rows and columns, and that D(P 1 ) is obtained from
D(P ) by reversing rows and columns. One can now use Case 5 to show that P 1 is Hurwitz,
and so P is also Hurwitz by Exercise E5.20.
Case 7: We assume (iii) and that n is even. As with Case 5, we define P (s) = (s+)P (s)
and in this case we compute

C(P ) + C 11 C 12
C(P ) =
C 12
p20

200

5 Stability of control systems

22/10/2004

delivering fewer determinantal inequalities to test. This results from their being a dependence
on some of the Hurwitz determinants. This is given thorough discussion by Gantmacher
[1959b]. Here we state the result, and give a proof due to Anderson [1972] that is more
elementary than that of Gantmacher.
5.42 Theorem A polynomial
P (s) = sn + pn1 sn1 + + p1 s + p0 R[s]
is Hurwitz if and only if any one of the following conditions holds:
(i) p2k > 0, k {0, 1, . . . , b n1
c} and 2k+1 > 0, k {0, 1, . . . , b n1
c};
2
2
n1
(ii) p2k > 0, k {0, 1, . . . , b 2 c} and 2k > 0, k {1, . . . , b n2 c};
c} and 2k+1 > 0, k {0, 1, . . . , b n1
c};
(iii) p0 > 0, p2k+1 > 0, k {0, 1, . . . , b n2
2
2
n2
(iv) p0 > 0, p2k+1 > 0, k {0, 1, . . . , b 2 c} and 2k > 0, k {1, . . . , b n2 c}.
Here 1 , . . . , n are the Hurwitz determinants.
Proof The theorem follows immediately from Theorems 5.28 and 5.41 and after one checks
that the principal minors of C(P ) are exactly the odd Hurwitz determinants 1 , 3 , . . ., and
that the principal minors of D(P ) are exactly the even Hurwitz determinants 2 , 4 , . . ..
This observation is made by a computation which we omit, and appears to be first been

noticed by Fujiwara [1915].
The advantage of the Lienard-Chipart test over the Hurwitz test is that one will generally
have fewer determinants to compute. Let us illustrate the criterion in the simplest case, when
n = 2.
5.43 Example (Example 5.35 contd) We consider the polynomial P (s) = s2 + as + b. Recall
that the Hurwitz determinants were computed in Example 5.37:
1 = a,

2 = ab.

Let us write down the four conditions of Theorem 5.42:

1. p0 = b > 0, 1 = a > 0;

and
D(P ) = D(P ) + D,
where C 11 , C 12 , and D are independent of . By our assumption (iii), for > 0 sufficiently
small we have C(P ) positive-definite. Thus, invoking the argument of Case 1, we may
deduce that D(P ) is also positive-definite. Therefore P is Hurwitz by Lemma 5.40 and
Theorem 5.36. Thus P is itself also Hurwitz.
Case 8: We assume (iv) and that n is even. Taking P 1 (s) = sn P ( 1s ) we see that C(P 1 )
is obtained from D(P ) by reversing the rows and columns, and that D(P 1 ) is obtained
from C(P ) by reversing the rows and columns. Now one may apply Case 7 to deduce that
P 1 , and therefore P , is Hurwitz.

5.5.4 The Li
enard-Chipart criterion Although less well-known than the criterion of
Routh and Hurwitz, the test we give due to Lienard and Chipart [1914]3 has the advantage of
3
Perhaps the relative obscurity of the test reflects that of its authors; I was unable to find a biographical
reference for either Lienard or Chipart. I do know that Lienard did work in differential equations, with the
Li
enard equation being a well-studied second-order linear differential equation.

2. p0 = b > 0, 2 = ab > 0;
3. p0 = b > 0, p1 = a > 0, 1 = a > 0;
4. p0 = b > 0, p1 = a > 0, 2 = ab > 0.
We see that all of these conditions are equivalent in this case, and imply that P is Hurwitz if
and only if a, b > 0, as expected. This example is really too simple to illustrate the potential
advantages of the Lienard-Chipart criterion, but we refer the reader to Exercise E5.22 to see
how the test can be put to good use.

5.5.5 Kharitonovs test It is sometimes the case that one does not know exactly
the coefficients for a given polynomial. In such instances, one may know bounds on the
coefficients. That is, for a polynomial
P (s) = pn sn + pn1 sn1 + + p1 s + p0 ,

(5.19)

5.5 Identifying polynomials with roots in C

22/10/2004

201

one may know that the coefficients satisfy inequalities of the form
pmin
i

pmax
,
i

202

5 Stability of control systems

22/10/2004

From this we deduce that for any allowable p we have

i = 0, 1, . . . , n.

(5.20)

In this case, the following remarkable theorem of Kharitonov [1978] gives a simple test for the
stability of the polynomial for all possible values for the coefficients. Since the publication of
Kharitonovs result, or more properly its discovery by the non-Russian speaking world, there
have been many simplifications of the proof [e.g., Chapellat and Bhattacharyya 1989, Dasgupta 1988, Mansour and Anderson 1993]. The proof we give essentially follows Minnichelli,
Anagnost, and Desoer [1989].

Re(Q1 (i)) = Re(Q2 (i)) Re(Pp (i)) Re(Q3 (i)) = Re(Q4 (i))
Im(Q1 (i)) = Im(Q4 (i)) Im(Pp (i)) Im(Q2 (i)) = Im(Q3 (i)).
This leads to the picture shown in Figure 5.4 for R(). The lemma follows immediately
Q2 (i)

Q3 (i)

Q1 (i)

Q4 (i)

5.44 Theorem Given a polynomial of the form (5.19) with the coefficients satisfying the inequalities (5.20), define four polynomials
max 2
Q1 (s) = pmin
+ pmin
s + pmax
s3 +
0
1 s + p2
3
3
Q2 (s) = pmin
+ pmax
s + pmax
s2 + pmin
0
1
2
3 s +
2
min 3
Q3 (s) = pmax
+ pmax
s + pmin
0
1
2 s + p3 s +

Q4 (s) =

pmax
0

pmin
1 s

2
pmin
2 s

pmax
s3
3

Figure 5.4 R()

+
from this.

Then P is Hurwitz for all

max
max
max
(p0 , p1 , . . . , pn ) [pmin
] [pmin
] [pmin
0 , p0
1 , p1
n , pn ]

if and only if the polynomials Q1 , Q2 , Q3 , and Q4 are Hurwitz.

Proof Let us first assume without loss of generality that pmin
> 0, j = 0, . . . , n. Indeed,
j
by Exercise E5.18, for a polynomial to be Hurwitz, its coefficients must have the same sign,
and we may as well suppose this sign to be positive. If
p = (p0 , p1 , . . . , pn )

min
[pmin
0 , p0 ]

min
[pmin
1 , p1 ]

Using the lemma, we now claim that if p is allowable, then Pp has no imaginary axis
roots. To do this, we record the following useful property of Hurwitz polynomials.
2 Lemma If P R[s] is monic and Hurwitz with deg(P ) 1, then ]P (i) is a continuous
and strictly increasing function of .
Proof Write
P (s) =

min
[pmin
n , pn ],

then let us say, for convenience, that p is allowable. For p allowable denote

j=1

where zj = j + ij with j < 0. Thus

Pp (s) = pn sn + pn1 sn1 + + p1 s + p0 .

It is clear that if all polynomials Pp are allowable then the polynomials Q1 , Q2 , Q3 , and
Q4 are Hurwitz. Thus suppose for the remainder of the proof that Q1 , Q2 , Q3 , and Q4 are
Hurwitz, and we shall deduce that Pp is also Hurwitz for every allowable p.
For R define
R() = { Pp (i) | p allowable} .
The following property of R() lies at the heart of our proof. It is first noticed by Dasgupta
[1988].
1 Lemma For each R, R() is a rectangle in C whose sides are parallel to the real and
imaginary axes, and whose corners are Q1 (i), Q2 (i), Q3 (i), and Q4 (i).
Proof We note that for R we have
4
Re(Q1 (i)) = Re(Q2 (i)) = pmin
pmax 2 + pmin
0
4 +

Re(Q3 (i)) = Re(Q4 (i)) = pmax

pmin 2 + pmax
4 +
0
4
4
Im(Q1 (i)) = Im(Q4 (i)) = pmin pmax 2 + pmin
4 +

Im(Q2 (i)) = Im(Q3 (i)) = pmax pmin 2 + pmax
4 + .
4

n
Y
(s zj )

]P (i) =

n
X
j=1

](i + |j | ij ) =

n
X

j
.
arctan
|j |
j=1

Since |j | > 0, each term in the sum is continuous and strictly increasing, and thus so too is
]P (i).
H
To show that 0 6 R() for R, first note that 0 6 R(0). Now, since the corners of
R() are continuous functions of , if 0 R() for some > 0, then it must be the case that
for some 0 [0, ] the point 0 C lies on the boundary of R(0 ). Suppose that 0 lies on
the lower boundary of the rectangle R(0 ). This means that Q1 (i0 ) < 0 and Q4 (i0 ) > 0
since the corners of R() cannot pass through 0. Since Q1 is Hurwitz, by Lemma 2 we must
have Q1 (i(0 + )) in the (, ) quadrant in C and Q4 (i(0 + )) in the (+, +) quadrant in
C for > 0 sufficiently small. However, since Im(Q1 (i)) = Im(Q4 (i)) for all R, this
cannot be. Therefore 0 cannot lie on the lower boundary of R(0 ) for any 0 > 0. Similar
arguments establish that 0 cannot lie on either of the other three boundaries either. This
then prohibits 0 from lying in R() for any > 0.
Now suppose that Pp0 is not Hurwitz for some allowable p0 . For [0, 1] each of the
polynomials
Q1 + (1 )Pp0
(5.21)

5.6 Summary

22/10/2004

203

is of the form Pp for some allowable p . Indeed, the equation (5.21) defines a straight line
from Q1 to Pp0 , and since the set of allowable ps is convex (it is a cube), this line remains
in the set of allowable polynomial coefficients. Now, since Q1 is Hurwitz and Pp0 is not, by
continuity of the roots of a polynomial with respect to the coefficients, we deduce that for
some [0, 1), the polynomial Pp must have an imaginary axis root. However, we showed
above that 0 6 R() for all R, denying the possibility of such imaginary axis roots.
Thus all polynomials Pp are Hurwitz for allowable p.

5.45 Remarks 1. Note the pattern of the coefficients in the polynomials Q1 , Q2 , Q3 , and
Q4 has the form (. . . , max, max, min, min, . . . ) This is charmingly referred to as the
Kharitonov melody .
2. One would anticipate that to check the stability for P one should look at all possible
extremes for the coefficients, giving 2n polynomials to check. That this can be reduced
to four polynomial checks is an unobvious simplification.
3. Anderson, Jury, and Mansour [1987] observe that for polynomials of degree 3, 4, or 5,
it suffices to check not four, but one, two, or three polynomials, respectively, as being
Hurwitz.
4. A proof of Kharitonovs theorem, using Liapunov methods (see Section 5.4), is given by
Mansour and Anderson [1993].

Let us apply the Kharitonov test in the simplest case when n = 2.
5.46 Example We consider
P (s) = s2 + as + b
with the coefficients satisfying
(a, b) [amin , amax ] [bmin , bmax ].
The polynomials required by Theorem 5.44 are
Q1 (s) = s2 + amin s + bmin
Q2 (s) = s2 + amax s + bmin
Q3 (s) = s2 + amax s + bmax
Q4 (s) = s2 + amin s + bmax .
We now apply the Routh/Hurwitz criterion to each of these polynomials. This indicates
that all coefficients of the four polynomials Q1 , Q2 , Q3 , and Q4 should be positive. This
reduces to requiring that
amin , amax , bmin , bmax > 0.
That is, amin , bmin > 0. In this simple case, we could have guessed the result ourselves since
the Routh/Hurwitz criterion are so simple to apply for degree two polynomials. Nonetheless,
the simple example illustrates how to apply Theorem 5.44.

5.6 Summary
The matter of stability is, of course, of essential importance. What we have done in this
chapter is quite simple, so let us outline the major facts.

204

5 Stability of control systems

22/10/2004

1. It should be understood that internal stability is a notion relevant only to SISO linear
systems. The difference between stability and asymptotic stability should be understood.
2. The conditions for internal stability are generally simple. The only subtleties occur when
there are repeated eigenvalues on the imaginary axis. All of this needs to be understood.
3. BIBO stability is really the stability type of most importance in this book. One should
understand when it happens. One should also know how, when it does not happen, to
produce an unbounded output with a bounded input.
4. Norm characterisations if BIBO stability provide additional insight, and offer a clarifying
language with which to organise BIBO stability. Furthermore, some of the technical
results concerning such matters will be useful in discussions of performance in Section 9.3
and of robustness in Chapter 15.
5. One should be able to apply the Hurwitz and Routh criteria freely.
6. The Liapunov method offer a different sort of characterisation of internal stability. One
should be able to apply the theorems presented.

Exercises for Chapter 5

205

Exercises
E5.1 Consider the SISO linear system = (A, b, c, D) of Example 5.4, i.e., with

0
1
A=
.
b a

E5.2

E5.3

E5.4
E5.5

E5.6

By explicitly computing a basis of solutions to x(t)

= Ax(t) in each case, verify by
direct calculation the conclusions of Example 5.4. If you wish, you may choose specific
values of the parameters a and b for each of the eight cases of Example 6.4. Make sure
that you cover sub-possibilities in each case that might arise from eigenvalues being
real or complex.
Determine the internal stability of the linearised pendulum/cart system of Exercise E1.5 for each of the following cases:
(a) the equilibrium point (0, 0);
(b) the equilibrium point (0, ).
For the double pendulum system of Exercise E1.6, determine the internal stability for
the linearised system about the following equilibria:
(a) the equilibrium point (0, 0, 0, 0);
(b) the equilibrium point (0, , 0, 0);
(c) the equilibrium point (, 0, 0, 0);
(d) the equilibrium point (, , 0, 0).
For the coupled tank system of Exercise E1.11 determine the internal stability of the
linearisation.
Determine the internal stability of the coupled mass system of Exercise E1.4, both
with and without damping. You may suppose that the mass and spring constant are
positive and that the damping factor is nonnegative.
Consider the SISO linear system = (A, b, ct , 01 ) defined by

0
1
A=
, b=
, c=

1
0
for R and > 0.
(a) For which values of the parameters and is spectrally stable?
(b) For which values of the parameters and is internally stable? Internally
asymptotically stable?
(c) For zero input, describe the qualitative behaviour of the states of the system
when the parameters and are chosen so that the system is internally stable
but not internally asymptotically stable.
(d) For zero input, describe the qualitative behaviour of the states of the system when
the parameters and are chosen so that the system is internally unstable.
(e) For which values of the parameters and is BIBO stable?
(f) When the system is not BIBO stable, determine a bounded input that produces
an unbounded output.

206

5 Stability of control systems

22/10/2004

E5.7 Consider the SISO linear system = (A, b, ct , 01 ) defined by

0
1
1 0
0 1
0
0

A=
0 0 , b = 0 , c = 0 ,
0 0
1
0

E5.8
E5.9

E5.10

E5.11

where R and > 0.

(a) For which values of the parameters and is spectrally stable?
(b) For which values of the parameters and is internally stable? Internally
asymptotically stable?
(c) For zero input, describe the qualitative behaviour of the states of the system
when the parameters and are chosen so that the system is internally stable
but not internally asymptotically stable.
(d) For zero input, describe the qualitative behaviour of the states of the system when
the parameters and are chosen so that the system is internally unstable.
(e) For which values of the parameters and is BIBO stable?
(f) When the system is not BIBO stable, determine a bounded input that produces
an unbounded output.
Show that (N, D) is BIBO stable
if and only if for every u L [0, ), the function
d
d
y satisfying D dt
y(t) = N dt
u(t) also lies in L [0, ).
Determine whether the pendulum/cart system of Exercises E1.5 and E2.4 is BIBO
stable in each of the following linearisations:
(a) the equilibrium point (0, 0) with cart position as output;
(b) the equilibrium point (0, 0) with cart velocity as output;
(c) the equilibrium point (0, 0) with pendulum angle as output;
(d) the equilibrium point (0, 0) with pendulum angular velocity as output;
(e) the equilibrium point (0, ) with cart position as output;
(f) the equilibrium point (0, ) with cart velocity as output;
(g) the equilibrium point (0, ) with pendulum angle as output;
(h) the equilibrium point (0, ) with pendulum angular velocity as output.
Determine whether the double pendulum system of Exercises E1.6 and E2.5 is BIBO
stable in each of the following cases:
(a) the equilibrium point (0, 0, 0, 0) with the pendubot input;
(b) the equilibrium point (0, , 0, 0) with the pendubot input;
(c) the equilibrium point (, 0, 0, 0) with the pendubot input;
(d) the equilibrium point (, , 0, 0) with the pendubot input;
(e) the equilibrium point (0, 0, 0, 0) with the acrobot input;
(f) the equilibrium point (0, , 0, 0) with the acrobot input;
(g) the equilibrium point (, 0, 0, 0) with the acrobot input;
(h) the equilibrium point (, , 0, 0) with the acrobot input.
In each case, use the angle of the second link as output.
Consider the coupled tank system of Exercises E1.11 and E2.6. Determine the BIBO
stability of the linearisations in the following cases:
(a) the output is the level in tank 1;

Exercises for Chapter 5

207

(b) the output is the level in tank 2;

(c) the output is the difference in the levels,
E5.12 Consider the coupled mass system of Exercise E1.4 and with inputs as described in
Exercise E2.19. For this problem, leave as an arbitrary parameter. We consider a
damping force of a very special form. We ask that the damping force on each mass be
given by d(x 1 + x 2 ). The mass and spring constant may be supposed positive, and
the damping constant is nonnegative.
(a) Represent the system as a SISO linear system = (A, b, c, D)note that in
Exercises E1.4 and E2.19, everything except the matrix A has already been determined.
(b) Determine for which values of , mass, spring constant, and damping constant
the system is BIBO stable.
(c) Are there any parameter values for which the system is BIBO stable, but for
which you might not be confident with the systems state behaviour? Explain
your answer.
E5.13 Let (N, D) be a SISO linear system in input/output form. In this exercise, if
u : [0, ) R is an input, yu will be the output defined so that yu (s) = TN,D (s)
u(s).
(a) For > 0 define an input u by
(
1
, t [0, ]
u (t) =
0, otherwise.

(b)

E5.14 Let
(a)
(b)

Determine
(i) lim0 kyu k2 ;
(ii) lim0 kyu k ;
(iii) lim0 pow(yu ).
If u(t) = sin(t) determine
(i) kyu k2 ;
(ii) kyu k ;
(iii) pow(yu ).
= (A, b, ct , D) be a SISO linear system.
Show that if A + At is negative-semidefinite then is internally stable.
Show that if A+At is negative-definite then is internally asymptotically stable.

E5.15 Let (A, P , Q) be a Liapunov triple for which P and Q are positive-definite. Show
that A is Hurwitz.
E5.16 Let

0 1
A=
0 a
for a 0. Show that if (A, P , Q) is a Liapunov triple for which Q is positivesemidefinite, then (A, Q) is not observable.
E5.17 Consider the polynomial P (s) = s3 + as2 + bs + c.
(a) Use the Routh criteria to determine conditions on the coefficients a, b, and c that
ensure that the polynomial P is Hurwitz.

208

5 Stability of control systems

22/10/2004

(b) Use the Hurwitz criteria to determine conditions on the coefficients a, b, and c
that ensure that the polynomial P is Hurwitz.
(c) Verify that the conditions on the coefficients from parts (a) and (b) are equivalent.
(d) Give an example of a polynomial of the form of P that is Hurwitz.
(e) Give an example of a polynomial of the form of P for which all coefficients are
strictly positive, but that is not Hurwitz.
E5.18 A useful necessary condition for a polynomial to have all roots in C is given by the
following theorem.
Theorem If the polynomial
P (s) = sn + pn1 sn1 + + p1 s + p0 R[s]
is Hurwitz, then the coefficients p0 , p1 , . . . , pn1 are all positive.
(a) Prove this theorem.
(b) Is the converse of the theorem true? If so, prove it, if not, give a counterexample.
The Routh/Hurwitz method gives a means of determining whether the roots of a polynomial
are stable, but gives no indication of how stable they are. In the following exercise, you
will examine conditions for a polynomial to be stable, and with some margin for error.
E5.19 Let P (s) = s2 + as + b, and for > 0 denote
R = { s C | Re(s) < } .
Thus R consists of those points lying a distance at least to the left of the imaginary
axis.
(a) Using the Routh criterion as a basis, derive necessary and sufficient conditions
for all roots of P to lie in R .

Hint: The polynomial P(s)

= P(s + ) must be Hurwitz.
(b) Again using the Routh criterion as a basis, state and prove necessary and sufficient
conditions for the roots of a general polynomial to lie in R .
Note that one can do this for any of the methods we have provided for characterising
Hurwitz polynomials.
E5.20 Consider a polynomial
P (s) = pn sn + pn1 sn1 + + p1 s + p0 R[s]
with p0 , pn 6= 0, and define P 1 R[s] by P 1 (s) = sn P ( 1s ).
(a) Show that the roots for P 1 are the reciprocals of the roots for P .
(b) Show that P is Hurwitz if and only if P 1 is Hurwitz.
E5.21 For the following two polynomials,
(a) P (s) = s3 + as2 + bs + c,
(b) P (s) = s4 + as3 + bs2 + cs + d,
do the following:
1. Using principal minors to test positive-definiteness, write the conditions of the
Hermite criterion, Theorem 5.38, for P to be Hurwitz.

Exercises for Chapter 5

209

2. Again using principal minors to test positive-definiteness, write the four conditions of the reduced Hermite criterion, Theorem 5.41, for P to be Hurwitz, and
ascertain which is the least restrictive.
E5.22 For the following two polynomials,
(a) P (s) = s3 + as2 + bs + c,
(b) P (s) = s4 + as3 + bs2 + cs + d,
write down the four conditions of the Lienard-Chipart criterion, Theorem 5.42, and
determine which is the least restrictive.
E5.23 Consider a general degree three polynomial
P (s) = s3 + as2 + bs + c,
where the coefficients satisfy
(a, b, c) [amin , amax ] [bmin , bmax ] [cmin , cmax ].

(E5.1)

Use Kharitonovs test, Theorem 5.44, to give conditions on the bounds for the intervals
for a, b, and c so that P is Hurwitz for all coefficients satisfying (E5.1).

210

5 Stability of control systems

22/10/2004

212

This version: 22/10/2004

Chapter 6
6.6

6 Interconnections and feedback

22/10/2004

6.5.1

Proportional control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263

6.5.2

Derivative control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263

6.5.3

Integral control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 264

6.5.4

Characteristics of the PID control law . . . . . . . . . . . . . . . . . . . . . . . . . 264

Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269

Interconnections and feedback

6.1 Signal flow graphs
We now begin to enter into the more design oriented parts of the course. The concept of
feedback is central to much of control, and we will be employing the analysis tools developed
in the time-domain, the s-domain, and the frequency domain to develop ways of evaluating
and designing feedback control systems. The value of feedback is at the same time obvious
and mysterious. It is clear that it ought to be employed, but it often has effects that are
subtle. Somehow the most basic feedback scheme is the PID controller that we discuss in
Section 6.5. Here we can get an idea of how feedback can effect a closed-loop system.
You may wish to take a look at the DC motor system we talked about in Section 1.2 in
order to see a very concrete display of the disadvantages of open-loop control, and how this
can be repaired to advantage with a closed-loop scheme.

Contents
6.1

6.2

6.3

6.4

6.5

Signal flow graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212

6.1.1

Definitions and examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212

6.1.2

Signal flow graphs and systems of equations . . . . . . . . . . . . . . . . . . . . . 215

6.1.3

Subgraphs, paths, and loops . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218

6.1.4

Cofactors and the determinants . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219

6.1.5

Masons Rule . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223

6.1.6

Sensitivity, return difference, and loop transmittance . . . . . . . . . . . . . . . . 226

Interconnected SISO linear systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229

6.2.1

Definitions and basic properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . 230

6.2.2

Well-posedness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233

6.2.3

Stability for interconnected systems . . . . . . . . . . . . . . . . . . . . . . . . . . 235

Feedback for input/output systems with a single feedback loop . . . . . . . . . . . . . . . 244

6.3.1

Open-loop versus closed-loop control . . . . . . . . . . . . . . . . . . . . . . . . . 244

6.3.2

Unity gain feedback loops . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247

6.3.3

Well-posedness and stability of single-loop interconnections . . . . . . . . . . . . . 248

Feedback for SISO linear systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 250

6.4.1

Static state feedback for SISO linear systems . . . . . . . . . . . . . . . . . . . . . 250

6.4.2

Static output feedback for SISO linear systems . . . . . . . . . . . . . . . . . . . . 255

6.4.3

Dynamic output feedback for SISO linear systems . . . . . . . . . . . . . . . . . . 258

The PID control law . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 261

We have seen in our discussion of block diagrams in Section 3.1 that one might be
interested in connecting a bunch of transfer functions like the one depicted in Figure 3.7. In
this section we indicate how one does this in a systematic and general way. This will enable
us to provide useful results on stability of such interconnected systems in Section 6.2.3. The
advantage of doing this systematically is that we can provide results that will hold for any
block diagram setup, not just a few standard ones.
The signal flow graph was first studied systematically in control by Mason[1953, 1956].
Many of the issues surrounding signal flow graphs are presented nicely in the paper of Lynch
[1961]. A general discussion of applications of graph theory may be found in the book of
Chen [1976]. Here the signal flow graph can be seen as a certain type of graph, and its
properties are revealed in this context. Our presentation will often follow that of Zadeh and
Desoer [1979].
6.1.1 Definitions and examples The notion of a signal flow graph can be seen as a
modification of the concept of a block diagram The idea is to introduce nodes to represent the
signals in a system, and then connect the nodes with branches, and assign to each branch
a rational function that performs the duty of a block in a block diagram. For example,
Figure 3.1 would simply appear as shown in Figure 6.1. Before we proceed to further
x1

/ x2

Figure 6.1 A simple signal flow graph

illustrations of how to construct signal flow graphs along the lines of what we did with block
diagrams, lets say just what we are talking about. Part of the point of signal flow graphs is
that we can do this in a precise way.
6.1 Definition Denote n = {1, 2, . . . , n} and let I n n.
(i) A signal flow graph with interconnections I is a pair (S, G) where S is a collection
{x1 , . . . , xn } of nodes or signals, and
G = {Gij R(s) | (i, j) I}
is a collection of rational functions that we call branches or gains. The branch Gij
originates from the node xj and terminates at the node xi .
(ii) A node xi is a sink if no branches originate from xi .
(iii) A node xi is a source if no branches terminate at xi .

6.1 Signal flow graphs

22/10/2004

213

214

6 Interconnections and feedback

x2 OO
OOO
oo7
OO1O
o
OOO
o
o
o
OO'
o
oo
x1 OO
7 x4
OOO
o
oo
OOO
ooo
o
O
o
oo 1
G31 OOO
O'
ooo
x3

(iv) A source xi is an input if only one branch originates from xi , and the gain of this
branch is 1.
(v) A sink xi is an output if only one branch terminates at xi and this branch has weight
1.

For example, for the simple signal flow graph of Figure 6.1, we have n = {1, 2}, I = {(1, 2)},
S = {x1 , x2 }, and G = {G21 = R}. The node x1 is a source and the node x2 is a sink. In
this case note that we do not have the branch from x2 to x1 . This is why we define I as we
dowe will almost never want all n2 possible interconnections, and those that we do want
are specified by I. Note that we assume that at most one branch can connect two given
nodes. Thus, if we have a situation like that in Figure 6.2, we will replace this with the

22/10/2004

G21 oooo

Figure 6.5 The signal flow graph for a parallel interconnection

/ x2 g

G32

/ x3

G43

/ x4

/ x5

1
G21

+3 x

Figure 6.6 The signal flow graph for a negative feedback loop

G021

Figure 6.2 Two branches connecting the same nodes

situation in Figure 6.3.

G21 +G021

is in Figure 6.6, and we have n = {1, 2, 3, 4, 5}, I = {(1, 2), (2, 3), (3, 4), (4, 2), (4, 5)},
S = {x1 , x2 , x3 , x4 , x5 }, and G = {G21 = 1, G32 , G43 , G24 = 1, G54 = 1}. Clearly, x1 is
an input and x5 is an output. From the graph we read the relationships x2 = x1 x4 ,
x3 = G32 x2 , x4 = G43 x3 , and x5 = x4 . This gives, in the usual manner,

/ x2

Figure 6.3 Branches added to give a single branch

If you think this obtuse and abstract, you are right. But if you work at it, you will
see why we make the definitions as we do. Perhaps a few more examples will make things
clearer.

x5 = x4 = G43 x3 = G43 G32 x2 = G43 G32 (x1 x4 )

G43 G32
x1 .
x4 = x5 =
1 + G43 G32

4. Suppose we wish to extract more information from the negative feedback loop of the
previous example. In Figure 6.7 we depict a situation where we have added an input
x6
1

6.2 Examples 1. The signal flow graph corresponding to the series block diagram of Figure 3.2 is shown in Figure 6.4. Note that here we have n = {1, 2, 3}, I = {(1, 2), (2, 3)},
x1

G21

/ x2

G32

/ x3

Figure 6.4 The signal flow graph for a series interconnection

S = {x1 , x2 , x3 }, and G = {G21 , G32 }. The node x1 is a source and x3 is a sink. There are
a possible n2 = 9 interconnections, and we only have two of them realised in this graph.
The transfer function from x1 to x2 is simply G21 ; that is, x2 = G21 x1 . We also read off
x3 = G32 x2 , and so x3 = G21 G32 x3 .
2. We now look at the representation of a parallel interconnection. The signal flow graph
is shown in Figure 6.5, and we have n = {1, 2, 3, 4}, I = {(1, 2), (1, 3), (2, 4), (3, 4)},
S = {x1 , x2 , x3 , x4 }, and G = {G21 , G31 , G42 = 1, G43 = 1}. One sees that x1 is a source
and x4 is a sink. From the graph we read the relations x2 = G21 x1 , x3 = G31 x1 , and
x4 = x2 + x3 . This gives x4 = (G21 + G31 )x1 .
3. The preceding two signal flow graphs are very simple in some sense. To obtain the transfer
function from the signal flow graph is a matter of looking at the graph and applying the
obvious rules. Let us now look at a feedback loop as a signal flow graph. The graph

/ x2 g

G32

/ x3

G43

/ x4

/ x5

Figure 6.7 Signal flow graph for negative feedback loop with extra
structure

signal x6 to the graph, and tapped an output signal x7 = x2 . Thinking about the control
situation, one might wish to think of x6 as a disturbance to the system, and of x7 as
being the error signal (this will be seen in a better context in Sections 6.3, 8.3, and 8.4).
In doing so we have added the input x6 to the existing input x1 and the output x7 to the
existing output x5 . Let us see how the two outputs get expressed as functions of the two
inputs. We have the relations x2 = x1 x4 , x3 = G32 x2 + x6 , x4 = G43 x3 , x5 = x4 , and

6.1 Signal flow graphs

22/10/2004

215

216

6 Interconnections and feedback

G11

x7 = x2 . We combine these to get

x5 = x4 = G43 x3 = G32 G43 x2 + G43 x6 = G32 G43 x1 G32 G43 x4 + G43 x6

G32 G43
G43
x5 = x4 =
x1 +
x6
1 + G32 G43
1 + G32 G43
x7 = x2 = x1 x4 = x1 G43 x3 = x1 G43 G32 x2 G43 x6
1
G43
x1
x6 .
x7 = x2 =
1 + G32 G43
1 + G32 G43

/ x1
G
G12

G21

/ x2
E
G22

We see that we essentially obtain a system of two linear equations that expresses the
input/output relations for the graph.

One can consider any node xj to effectively be an output by adding to the signal flow graph
a node xn+1 and a branch Gn+1,j with gain equal to 1. One can also add an input to any
node xj by adding a node xn+1 and a branch Gj,n+1 whose gain is 1.
6.1.2 Signal flow graphs and systems of equations A signal flow graph is also a
representation of a set of linear equations whose coefficients belong to any field. That is to
say, we consider a set of linear equations where the coefficients are anything that can be
added, multiplied, and divided. The particular field that is of interest to us is R(s). Whats
more, we consider a very particular type of linear equation; one of the form
(1 G11 )x1 G12 x2 G1n xn = u1
G21 x1 + (1 G22 )x2 G2n xn = u2
...
Gn1 x1 Gn2 x2 + (1 Gnn )xn = un ,

22/10/2004

(6.1)

where Gij R(s), i, j = 1, . . . , n. The reason for using this type of equation will become
clear shortly. However, we note that corresponding to the equations (6.1) is a natural signal
flow graph. The following construction indicates how this is determined.
6.3 From linear equation to signal flow graph Given a set of linear equations of the form (6.1),
perform the following steps:
(i) place a node for each variable x1 , . . . , xn ;
(ii) place a node for each input u1 , . . . , un ;
(iii) for each nonzero Gij , i, j = 1, . . . , n, draw a branch originating from node j and
terminating in node i, having gain Gij ;
(iv) for i = 1, . . . , n, draw a branch of gain 1 from ui to xi .
The result is a signal flow graph with nodes {x1 , . . . , xn , xn+1 = u1 , . . . , x2n = un } and gains
G = { Gij | i, j = 1, . . . , n} {G1,n+1 = 1, . . . , Gn,2n = 1}.

For example, in Figure 6.8 we show how this is done when n = 2. One can readily verify
that a balance at each node will yield the equations (6.1).
This establishes a graph for each set of equations of the form (6.1). It is also true that
one can go from a graph to a set of equations. To state how to do this, we provide the
following recipe.

Figure 6.8 A general 2 node signal flow graph

6.4 From signal flow graph to linear equation Given a signal flow graph (S, G), to any source
xi that is not an input, add a node xji = ui to S and a branch Giji to G. After doing this for
each such source, we arrive at a new signal flow graph (S0 , G0 ). For each node xi in (S0 , G0 )
that is not an input perform the following steps:
(i) let xj1 , . . . , xjk be the collection of nodes for which there are branches connecting them
with xi ;
(ii) form the product of xj1 , . . . , xjk with the respective branch gains;
(iii) set the sum of these products equal to xi .
The result is an equation defining the node balance at node xi . Some of the inputs may be
zero.

Thus we establish a 11 correspondence between signal flow graphs and linear equations
of the form (6.1) (with some inputs and gains possibly zero). Let us denote by GS,G the
matrix of rational functions that serves as the coefficient matrix in (6.1). Thus

1 G11 G12
G1n
G21 1 G22
G2n

GS,G = ..
(6.2)
.
..
..
.
.
.

.
.
.
Gn1
Gn2 1 Gnn
We call GS,G the structure matrix for (S, G). Note that GS,G is a matrix whose components
are rational functions. We denote the collection of n n matrices with rational function
components by R(s)nn . Again we note that for a given signal flow graph, of course, many
of the terms in this matrix might be zero. The objective of making the connection between
signal flow graphs and linear equations is that the matrix formulation puts at our disposal
all of the tools from linear algebra. Indeed, one could simply use the form (6.1) to study
signal flow graphs. However, this would sweep under the carpet the special structure of the
equations that results from their being derived as node equations of a signal flow graph. One
of the main objectives of this section is to deal with this aspect of the signal flow graph. But
for now, let us look at our signal flow graphs of Example 6.2 as systems of equations.
6.5 Examples (Example 6.2 contd) We shall simply write down the matrix G that appears
in (6.1). In each case, we shall consider, as prescribed by the above procedure, the system
with an input attached to each source that is not itself an input. Thus we ensure that the
matrix GS,G represents all states of the system.

6.1 Signal flow graphs

22/10/2004

217

u1
G21

/ x2

G32

/ x3

Figure 6.9 Series signal flow graph with input added

1. For the series interconnection we work with the signal flow graph of Figure 6.9, and we
determine

1
0
0

1
0 .
GS,G = G21
0
G32 1
2. For the parallel interconnection we work with the signal flow graph of Figure 6.10, and
u1

x2 OO
OOO
oo7
OO1O
o
OOO
o
o
o
OO'
o
oo
x1 OO
7 x4
OOO
o
oo
OOO
ooo
o
O
o
oo 1
G31 OOO
O'
ooo
x3
G21 oooo

Figure 6.10 Parallel signal flow graph with input added

we determine

GS,G

1
0
0
G21
1
0

=
G31
0
1
0
G42 G43

0
0
.
0
1

3. Finally, for the negative feedback interconnection we work with the signal flow graph of
Figure 6.11, and we determine
u1

22/10/2004

6.6 Definition Let (S, G) be a signal flow graph with interconnections I n n.

(i) A subgraph of (S, G) is a pair (S0 , G0 ) where S0 S and where G0 G satisfies
Gij G0 implies that xi , xj S0 . If S0 = {xi1 , . . . , xin0 }, then we say (S0 , G0 ) has
interconnections {i1 , . . . , in0 }. We also denote by I0 I the subset defined by G0 .
Thus (i0 , j 0 ) I0 if and only if Gi0 j 0 G0 .
(ii) A subgraph (S0 , G0 ) is connected if for any pair xi , xj S0 there exists nodes xi0 =
xi , xi1 , . . . , xik = xj S0 and gains G1 , . . . , Gk G0 so that G` {Gi` ,i`+1 , Gi`+1 ,i` },
` {0, . . . , k 1}.
(iii) A path in (S, G) is a sequence {xi1 , . . . , xik } of nodes with the property that (ij , ij+1 )
I for each j {1, . . . , k1}. That is to say, there is a branch connecting the jth element
in the sequence with the (j + 1)st element in the sequence.
(iv) A path {xi1 , . . . , xik } is simple if the set { (ij+1 , ij ) | j {1, . . . , k 1}} is distinct.
(v) A path {xi1 , . . . , xik } is directed if the vertices are distinct. We denote the set of all
directed paths in (S, G) by Path(S, G).
(vi) A forward path is a directed path from an input node of (S, G) to a node of (S, G).
The set of forward paths from an input xi to a node xj is denoted Pathji (S, G).
(vii) The number of branches in a directed path is the length of the directed path.
(viii) A loop is a simple path {xi1 , . . . , xik } with the property that xi1 = xik . The set of
loops in (S, G) we denote by Loop(S, G).
(ix) The product of the gains in a directed path is the gain of the directed path.
(x) The loop gain of a loop L is the product of the gains comprising the loop, and is
denoted GL .
(xi) A finite collection of loops is nontouching if they have no nodes or branch gains in
common.

Perhaps a few words of clarification about the less obvious parts of the definition are in
order.

6 Interconnections and feedback

6.1.3 Subgraphs, paths, and loops In the constructions of the next section, we need
to have a good understanding of paths through a signal flow graph. To do this, we make
some definitions.

218

G32

/ x2 g

/ x3

G43

/ x4

/ x5

Figure 6.11 Negative feedback signal flow graph with input added

GS,G

1
0
0
0
G21
1
0
G24

G32
1
0
=
0
0
0
G43
1
0
0
0
G54

0
0

0
.
0
1

Note that we may take some of the gains to be 1 when they appear that way in Example 6.2.

1. The notion of a connected subgraph has a simplicity that belies its formal definition.
It merely means that it is possible to go from any node to any other node, provided
one ignores the orientation of the branches.
2. Note that the nodes in a loop must be distinct. That a loop cannot follow a path that
goes through any node more than once.
3. It is easy to be misled into thinking that any directed path is a forward path. This is
not necessarily true since for a forward path originates at an input for (S, G).
Again, this looks pretty formidable, but is in fact quite simple. We can illustrate this
easily by looking at the examples we have given of signal flow graphs.
6.7 Examples 1. For the very simple signal flow graph of Figure 6.1 there is but one path of
length 1, and it is {x1 , x2 }. There are no loops.

6.1 Signal flow graphs

22/10/2004

219

{x2 , x3 },

and one path of length 2,

Again, there are no loops.
3. The parallel signal flow graph of Figure 6.5, forgetting for the moment that some of the
gains are predetermined to be 1, has the paths
{x2 , x4 },

{x1 , x3 },

It turns out that S,G is exactly the determinant of GS,G .

6.8 Proposition S,G = det GS,G .
Proof The proof is involved, but not difficult. We accomplish it with a series of lemmas.
First we note that the definition of the determinant gives an expression of the form
X
det GS,G = 1 +
G ,
(6.4)

{x3 , x4 },

of length 1, and paths

{x1 , x2 , x4 },

22/10/2004

(L1 ,...,Lk )
Loopk (G,S)

{x1 , x2 , x3 }.

{x1 , x2 },

6 Interconnections and feedback

most examples we shall see), Loopk (S, G) = . The determinant of a signal flow graph
(S, G) is defined to be

X (1)k X

(6.3)
S,G = 1 +
GL1 GLk .
k!
k1

2. For the series signal flow graph of Figure 6.4 we have two paths of length 1,
{x1 , x2 },

220

{x1 , x3 , x4 }

of length 2. There are no loops in this graph either.

4. The negative feedback graph of Figure 6.7 has an infinite number of paths. Let us list
the basic ones.

where G is a product of branch gains. We denote by (S , G ) the subgraph of (S, G)

comprised of the nodes and branches that are involved in a typical term G .
We explore some properties of the subgraph (SG , GG ).

(a) length 1:
{x1 , x2 },

{x2 , x3 },

{x3 , x4 },

{x4 , x5 },

{x3 , x6 },

1 Lemma For each node xi S , there is at most one branch in G terminating at xi , and
at most one branch in G originating from xi .

{x2 , x7 }.

Proof Let G = {G1 , . . . , Gk } be the gains that form G . By the definition of the determinant, for each row of GS,G there is at most one j {1, . . . , k} so that Gj is an element
of that row. A similar statement holds for each column of GS,G . However, from these statements exactly follows the lemma.
H

(b) length 2:
{x1 , x2 , x3 },

{x2 , x3 , x4 },

{x3 , x4 , x5 },

{x3 , x4 , x6 },

{x2 , x4 , x7 }.

{x2 , x3 , x5 , x5 },

{x3 , x4 , x5 , x6 }.

(d) length 4:
{x1 , x2 , x3 , x4 , x5 }.
Some of these paths may be concatenated to get paths of any length desired. The reason
for this is that there is a loop given by
{x2 , x3 , x4 }.

6.1.4 Cofactors and the determinants Next we wish to define, and state some properties of, some quantities associated with the matrix GS,G . These quantities we will put
to use in the next section in proving an important theorem in the subject of signal flow
graphs: Masons Rule. This is a rule for determining the transfer function between any input and any output of a signal flow graph. But this is getting ahead of ourselves. We have
a lot of work to do in the interim.
Let us proceed with the development. For k 1 we denote
Loopk (S, G) = {(Lj1 , . . . , Ljk ) (Loop(S, G))k | L1 , . . . , Lk are nontouching}.
That is, Loopk (S, G) consists of those k-tuples of loops, none of which touch the others. Note
that Loop1 (S, G) = Loop(S, G) and that for k sufficiently large (and not very large at all in

Proof Suppose that (S , G ) consists of a collection (S1 , G1 ), . . . , (S` , G` ) of connected

subgraphs. We shall show that each of these connected subgraphs is a loop. It then follows
from Lemma 1 that the loops will be nontouching. We proceed by contradiction, supposing
that one of the connected components, say (S1 , G1 ), is not a loop. From Lemma 1 it follows
that (S1 , G1 ) is a directed path between two nodes; suppose that these nodes are xi and xj .
Since no branches terminate at xi , there are no elements from row i of GS,G in G . Since no
branches originate from xj , there are no elements from column j of GS,G in G . It therefore
follows that the 1s in position (i, i) and (j, j) appear in the expression for G . However,
since S1 , G1 ) is a directed path between xi and xj it follows that there are some ki , kj n
H
so that Gki i and Gj,kj appear in G . This is in contradiction to Lemma 1.
3 Lemma If G is the product of the loop gains from k loops, then it has sign (1)k in the
expression for det GS,G .
Proof First note that the determinant of GS,G is unaffected by the numbering of the nodes.
Indeed, if one interchanges i and j in a given numbering scheme, then in GS,G , both the ith
and jth columns and the ith and jth rows are swapped. Since each of these operations gives
a change of sign in the determinant, the determinant itself is the same.
Suppose that (S , G ) is comprised of k nontouching loops of lengths `1 , . . . , `k . Denote
the nodes in these loops by
{x1 , . . . , x`1 , x1 }, {x`1 +1 , . . . , x`2 , x`+1 }, . . . , {x`k1 +1 , . . . , x`k , x`k1 +1 }.

6.1 Signal flow graphs

22/10/2004

221

According to the definition (A.1) of the determinant, to determine the sign of the contribution of G , we should determine the sign of the permutation

=

1
2

`1 1 `1 `1 + 1
`1
1 `1 + 2

`2 1
`2

`2
`1 + 1

`k1 + 1
`k1

`k 1
`k
`k
`k1 + 1

By `1 1 transpositions we may shift the 1 in the `1 st position to the 1st position. By `2 1

transpositions, the `1 + 1 in the (`1 + `2 )th position can be shifted to the (`1 + 1)st position.
Proceeding in this manner, we see that the sign of the permutation is given by

222

6 Interconnections and feedback

G34

G23

~
~

x2
oo7
o
ooo
ooo
x1 OO
OOO
OOO
O
G61 OOO
O'
x6 `
G21 oooo

G32

G76

/ x3

/ x7
`

G67

k
Y
sgn() =
(1)`j 1

(6.5)

22/10/2004

G43

G87

/ x4 O
OOO
OOOG54
OOO
OOO
'
x5
oo7
o
o
o
o
o
oo
ooo G85
/ x8 o

G78

Figure 6.12 A signal flow graph with multiple loops

j=1

Since each of the gains occurs with sign 1 in GS,G , they will contribute the sign
sgn()

2. Next we consider a new signal flow graph, namely the one depicted in Figure 6.12. This
graph has four loops:

k
Y
(1)`j

L1 = {x2 , x3 , x2 },

j=1

L2 = {x3 , x4 , x3 },

L3 = {x6 , x7 , x6 },

L4 = {x7 , x8 , x7 }.

We have Loop1 (S, G) = {L1 , L2 , L3 , L4 } and

to G . However, by (6.5), we have

Loop2 (S, G) = {(L1 , L3 ), (L1 , L4 ), (L2 , L3 ), (L2 , L4 ), (L3 , L1 ), (L4 , L1 ), (L3 , L1 ), (L4 , L2 )}.

sgn()

k
Y

An application of (6.3) gives

(1)`j = (1)k ,

S,G = 1 (GL1 + GL2 + GL3 + GL4 ) + (GL1 GL3 + GL1 GL4 + GL2 GL3 + GL2 GL4 ).

j=1

and so the lemma follows.

The three previous lemmas show that the terms G in the expression (6.4) for det GS,G
have the form of the terms in S,G . It remains to show that every term in G,G appears in
det GS,G . Thus suppose that (S1 , G1 ), . . . , (S` , G` ) is a collection of nontouching loops.
Since these loops are nontouching, it follows that in any given row or column, there can reside
at most one gain from the set G1 . . . G` . Therefore, an expansion of the determinant will
contain the product of the gains from the set G1 . . . G` . Furthermore, they must have

the same sign in det GS,G as in S,G by Lemma 3. This concludes the proof.
Let us see how to apply this in two examples, one of which we have seen, and one of
which is complicated enough to make use of the new terminology we have introduced.
6.9 Examples 1. Let us first look at the signal flow graph depicted in Figure 6.7. For clarity
it helps to ignore the fact that G24 = 1. This graph has a single loop
L = {x2 , x3 , x4 , x2 }.
This means that Loop1 (S, G) = {L} and Loopk (S, G) = for k 2. The gain of the loop
is simply the product G32 G43 G24 . Therefore, in this case, the determinant as per (6.3) is
S,G = 1 G32 G43 G24 .
If we now remember that G24 = 1 we get the term 1 + G32 G43 which appears in the
denominators in Example 6.24. This, of course, is no coincidence.

One may, of course, substitute into this expression the various gains given in terms of
the branch gains.

Now we turn to the cofactor of a path. For P Path(S, G) let GP denote the branches
of G with those comprising P removed, and those branches having a node in common with
P removed. We note that (S, GP ) is itself a signal flow graph. If (S, G) is connected, then
(S, GP ) may not be. For P Path(S, G) the cofactor of P is defined by CofP (S, G) = S,GP .
Let us illustrate this for the examples whose determinants we have computed.
6.10 Examples 1. For the signal flow graph depicted in Figure 6.7 let us consider various
paths. Again we do not pay attention to specific values assign to branch gains.
(a) P1 = {x1 , x2 , x3 , x4 , x5 }: If we remove this path, the graph has no loops and so
CofP1 (S, G) = 1.
(b) P2 = {x1 , x2 , x7 }: Removing this leaves intact the existing loop, and so the determinant of the graph remains unchanged. Therefore we have CofP2 (S, G) = S,G =
1 G32 G43 G24 .
(c) P3 = {x3 , x4 , x5 , x6 }: Removal of this path leaves a signal flow graph with no loops
so we must have CofP3 (S, G) = 1.
(d) P4 = {x2 , x3 , x4 , x6 , x7 }: Again, if P4 is removed, we are left with no loops and this
then gives CofP4 (S, G) = 1.
2. Next we look at the signal flow graph of Figure 6.12. We consider two paths.
(a) P1 = {x1 , x2 , x3 , x4 , x5 }: Removing this loop leaves two loops remaining: L3 and
L4 . The determinant of the resulting graph is, by (6.3),
CofP1 (S, G) = 1 (GL3 + GL4 ).

22/10/2004

6.1 Signal flow graphs

223

(b) P2 = {x1 , x6 , x7 , x8 , x5 }: This situation is rather like that for the path P1 and we
determine that
CofP2 (S, G) = 1 (GL1 + GL2 ).
Into these expressions for the cofactors, one may substitute the branch gains.

6.1.5 Masons Rule With the notion of cofactor and determinant clearly explicated,
we can state Masons Rule for finding the transfer function between an input to a signal
flow graph and any node in the graph. In this section we suppose that we are working with
a signal flow graph (S, G) with interconnections I n n. In order to simplify matters
and make extra hypotheses for everything we do, let us make a blanket assumption in this
section.
The following result gives an easy expression for the transfer function between the input
ui at xi and any other node. The following result can be found in the papers of Mason [1953,
1956]. It is not actually given a rigorous proof there, but one is given by Zadeh and Desoer
[1979]. We follow the latter proof.
6.11 Theorem (Masons Rule) Let (S, G) be a signal flow graph. For i, j {1, . . . , n} we
have
X
GP CofP (S, G)
xj =
ui .
S,G

224

xj =

det GS,G (x1 , j)

.
det GS,G

From Proposition 6.8 the theorem will follow if we can show that
X
det GS,G (x1 , j) =
GP CofP (S, G).
P Pathj1 (S,G)

Let (Sj , Gj ) be the subgraph of (S, G) corresponding to the matrix GS,G (x1 , j). One can
readily ascertain that one may arrive at (Sj , Gj ) from (S, G) by performing the following
operations:

2. the gain of a forward path P from x1 to xj ;

3. a product of loop gains for loops that share no nodes or branches with P .
Let us fix such a term corresponding to a given forward path P . Now let us renumber the
nodes so that we have

1
0
0
x1
Gk 1 1

0
0
2

..
..
..
...
.
.
.
.
0
`,n`

0
1
0
GS,G (x1 , j) = 0
(6.6)

G
0
jk
`1

G
0n`,`
R(n`)(n`) . Therefore we have
for some matrix G

det GS,G (x1 , j) = (1)`1 x1 (Gk2 1 )(Gk3 k2 ) (Gjk`1 ) det G

= x1 Gk2 1 Gk3 k2 Gjk det G.

Now note the form of the matrix in (6.6) shows that the signal flow graph corresponding
is obtained from (S, G) by removing all branches associated with the forward path P .
to G
From this it follows that all products in det GS,G (x1 , j) are of the form GP CofP (S, G) for
some forward path P .
It remains to show that any such expression will appear as a product in the expression
for det GS,G (x1 , j). This may be shown as follows. Let P be a forward path comprised of
gains Gk2 1 , Gk3 k2 , . . . , Gj,k`1 . The structure of (Sj , Gj ) implies that by adding the gain x1 ,
we have the gain for a loop in (Sj , Gj ). Now, as we saw in the proof of Proposition 6.8,
this term, multiplied by the loop gains of nontouching loops, is ensured to appear in the
determinant of (Sj , Gj ).

If a signal flow graph has multiple inputs u1 , . . . , uk , then one can apply Masons rule for
each input, and the resulting expression for a non-input node xj is then of the form

1. remove all branches originating from xj ;

xj =

By Proposition 6.8 we know that det GS,G (x1 , j) is comprised of a sum of products of loop
gains for nontouching loops. Thus in each of the products must appear a loop gain of the
form
f1 Gk2 1 Gk3 k2 Gjk`1 .
It follows that det GS,G (x1 , j) is a sum of products of terms falling into three types:

k
X

Tji ui ,

i=1

2. add a branch with gain x1 originating from xj and terminating at x1 .

Since the only nonzero term in the jth column is the term x1 appearing in the first row, an
expansion of the determinant about the jth column shows that det GS,G (x1 , j) has the form
Y
det GS,G (x1 , j) = x1
G .

22/10/2004

1. x1 ;

P Pathji (S,G)

Proof Without loss of generality, suppose that i = 1 and that node 1 is an input. Denote
x1 = (x1 , 0, . . . , 0) and let GS,G (x1 , j) be the matrix GS,G with the jth column replaced with
x1 . Cramers Rule then says that

6 Interconnections and feedback

where
Tji =

X
P Pathji (S,G)

GP CofP (S, G)
S,G

is the graph transmittance from the input ui to the node xj . Note that it is possible that
for a given i and j, Pathji (S, G) will be empty. In this case, we take the graph transmittance
to be zero.
The following result gives a useful interpretation of the set of all graph transmittances.
6.12 Corollary The matrix T S,G R(s)nn whose (i, j)th component is the graph transmittance Tij is the inverse of GS,G .

22/10/2004

6.1 Signal flow graphs

225

Proof As we saw in the proof of Theorem 6.11, the numerator term in the expression for
Tji is the determinant of the matrix obtained by replacing the jth column of GS,G by the
ith standard basis vector ei . By Cramers Rule we infer that Tji is the jth component of
the solution vector ti for the linear equation GS,G ti = ei . This means that if we define

T S,G = t1 tn ,
then we have
GS,G T S,G =

e1 en

= I n.

From the definition of inverse, the result follows.

Let us show how the above developments all come together to allow an easy determination
of the various transfer functions for the two examples we are considering.
6.13 Examples (Example 6.9 contd) In each example, we will number paths as we did in
Example 6.9.
1. For the signal flow graph of Figure 6.7 there are two sinks, x1 and x6 , and two sources,
x5 and x7 . To each of the sources, let us attach an input so that we are properly in the
setup of Theorem 6.11. Recalling our labelling of paths from Example 6.9, we have
Path51 (S, G) = {P1 },
Path56 (S, G) = {P3 },

Path71 (S, G) = {P2 },

Path76 (S, G) = {P4 }.

These are, of course, the paths whose cofactors we computed in Example 6.9. Now we
can compute, using Masons Rule, the coefficients in expressions of our two sinks x5 and
x7 involving the two sources x1 and x6 . We have
T51 =
T71 =
T56 =
T76 =

G21 G32 G43 G54

1 G32 G43 G24
G21 G72
1 G32 G43 G24
G36 G43 G54
1 G32 G43 G24
G36 G43 G24 G72
1 G32 G43 G24

and so we thus obtain the explicit expressions

x5 = T51 x1 + T56 x6 ,

x7 = T71 x1 + T76 x6 .

Lets see how this checks out when G21 = G54 = G72 = G36 = G24 = 1. In this case we
obtain
G32 G43
G43
1
G43
x1 +
x 6 , x7 =
x1
x6
x5 =
1 + G32 G43
1 + G32 G43
1 + G32 G43
1 + G32 G43
as we did when we performed the calculations by hand back in Example 6.2.
2. We shall compute the transfer function from x1 to x5 . We have Path51 (S, G) = {P1 , P2 }.
By Masons Rule, and using the determinant and cofactors we have already computed,
we have

G21 G32 G43 G54 1 (GL3 + GL4 ) + G61 G76 G87 G58 1 (GL1 + GL2 )
x5 =
x1 .
S,G

226

6 Interconnections and feedback

22/10/2004

As always, we may substitute into this the values for the branch gains to get an horrific
formula for the transfer function. But just imagine trying to do this by hand!

We have provided in this section a systematic way of deriving the transfer function
between various inputs and outputs in a signal flow graph. Whats more, we have identified
an important piece of structure in any such transfer function: the determinant of the graph.
We shall put this to use when studying stability of interconnected systems in Section 6.2.3.
6.1.6 Sensitivity, return difference, and loop transmittance Up to this point, the
discussion has been centred around the various transfer functions appearing in a signal flow
graph. Let us now look at other interesting objects, sensitivity, return difference, and loop
transmittance. These will turn out to be interesting in the special context of single-loop
systems in Section 6.3. The ideas we discuss in this section are presented also in the books
of Truxal [1955] and Horowitz [1963].
First we consider the notion of sensitivity. Let us first give a precise definition.
6.14 Definition Let (S, G) be a signal flow graph with interconnections I nn, let i, j n,
and let (`, k) I. Let Tji be the graph transmittance from node i to node j and let G`k be
the branch gain from node k to node `. The sensitivity of Tij to Gk` is
ji
S`k
=

(ln Tji )
,
(ln G`k )

where Tji is regarded as a function of the scalars Gsr , (s, r) I.

Let us try to gather some intuition concerning this definition. If f is a scalar function of
a scalar variable x then note that
d(ln f (x))
d(ln f (eln x ))
=
d(ln x)
d(ln x)
1 d(f (eln x ))
=
f (eln x ) d(ln x)
1 df (x) dx
=
f (x) dx d(ln x)
x df (x)
=
f (x) dx
f (x + x)/f (x)
= lim
.
x0
(x + x)/x
f (x))
The punchline is that d(ln
, evaluated at a particular x0 , gives the rate of f , normalised
d(ln x)
by f (x0 ), with respect to x, normalised by x0 . Thus one might say that

d(ln f (x))
d(% change in f )
=
.
d(ln x)
d(% change in x)
ji
In any event, S`k
measures the dependence of Tji on G`k in some sense.
ji
Let us now give a formula for S`k
in terms of graph determinants.

6.1 Signal flow graphs

22/10/2004

227

6.15 Proposition Let (S, G) be a signal flow graph with interconnections I nn, let i, j n,
and let (`, k) I. Let G(`,k) = G \ {G`k }. We then have
P
S,G(`,k)
P 0 Pathji (S,G(`,k) ) GP 0 CofP 0 (S, G(`,k) )
ji
P
S`k =

.
(6.7 )
S,G
P Pathji (S,G) GP CofP (S, G)
Proof In the proof we shall use the formula
We have Tji =

Fji
S,G

d(ln f (x))
d(ln x)

x df (x)
f (x) dx

derived above.

6 Interconnections and feedback

GP CofP (S, G).

G`k

/ x5

GP CofP (S, G) = G32 G43 .

P Pathji (S,G)

From this we see that

S,G
= (S,G )`k ,
G`k

0
1
1
+
=
.
1 + G32 G43 G32 G43
1 + G32 G43
We shall see this sensitivity function again in Section 6.3, and it will be an important object
when we come consider design issues throughout the remainder of the text.

51
S43
=

Now we consider the closely related concepts of return difference and loop transmittance.
First we look at loop transmittance. The loop transmittance is defined relative to a certain
branch in a signal flow graph. The idea is that we cut the branch, in doing so creating a new
signal flow graph with two new nodes, one a sink and one a source. The loop transmittance
is the transmittance in this new graph from the newly created source to the newly created
sink. Let us make this precise.

`k
S,G
(ln S,G )
(S,G )`k
S,G
=
=
,
(ln G`k )
S,G
S,G
`k . A moments thought tests the veracity of the formulae
thus defining Fji`k and
S,G
GP 0 CofP 0 (S, G(`,k) )

6.17 Definition Let (S, G) be a signal flow graph with interconnections I n n, and let
(j, i) I. Define a new signal flow graph (Sji , Gji ) with nodes {x1 , . . . , xn , xn+1 , xn+2 } and
with branches
Gij = (G \ {Gji ) {Gn+1,i = 1} {Gj,n+2 = Gji }.

P 0 Pathji (S,G(`,k) )

`k = S,G ,

S,G
(`,k)
giving the result.

/ x4

P 0 Path51 (S,G(4,3) )

Fji Fji`k
(ln Fji )
(Fji )`k
=
=
(ln G`k )
Fji
Fji

G43

S,G = 1 + G32 G43

X
GP 0 CofP 0 (S, G(4,3) ) = 0

where (Fji )`k and (S,G )`k are the terms in Fji and S,G that involve G`k . Therefore,

Fji`k =

/ x3

S,G(`,k) = 1

Now we note that each of the expressions Fji and S,G is a sum of terms, each of which is
either independent of G`k or depends linearly on G`k . Therefore we have
Fji
= (Fji )`k ,
G`k

G32

Figure 6.13 The signal flow graph for a negative feedback loop

(ln Tij )
(ln Fji )
(ln S,G )
=

.
(ln G`k )
(ln G`k )
(ln G`k )

G`k

/ x2 g

P Pathji (S,G)

Therefore

22/10/2004

6.16 Example We work with the single-loop signal flow graph of Figure 6.11, reproduced in
Figure 6.13. Let us determine the sensitivity of the transfer function T51 to the gain G43 . It
is easy to see that

where
Fji =

228

It is much easier to say in words what the symbols in the result mean. The signal flow
graph (S, G(`,k) ) is that obtained by removing the branch from node k to node `. Thus the
numerators,
X
S,G(`,k) and
GP 0 CofP 0 (S, G(`,k) ),
P 0 Pathji (S,G(`,k) )

in each of the terms on the right-hand side of (6.7) are simply the numerator and denominator
for the transfer function from node i to node j in the graph (S, G(`,k) ) as given by Masons
rule. Thus these can often be obtained pretty easily. Let us see how this works in an
example.

(i) The loop transmittance through Gji is the transmittance from xn+2 to xn+1 in the
graph (Sji , Gji ), and is denoted Lji .
(ii) The return difference through Gji is given by Rji = 1 Lji .

The return difference should be thought of as the difference between a unit signal transmitted
from node xn+2 and the signal that results at xn+1 .
As usual, we want to give a characterisation of the loop transmittance in terms of determinants and related notions.
6.18 Proposition Let (S, G) be a signal flow graph with interconnections I n n and let
(j, i) I. Let G(j,i) = G \ {Gji }. We then have
Rji =

S,G
,
S,G(j,i)

Lji = 1

S,G
.
S,G(j,i)

6.2 Interconnected SISO linear systems

22/10/2004

229

230

Proof If there is no loop containing the branch Gji then, by definition, Lji = 0 since no
forward path connects node xn+2 with node xn+1 in the graph (Sji , Gji ). This is consistent
with the proposition since in this case S,G = S,Gji . Thus the result holds if Gji is not part
of a loop. If it is part of loops L1 , . . . , Lk in (S, G), then these loops will be broken when the
graph (Sji , Gji ) is formed. Whats more, all elements of Pathn+2,n+1 (Sji , Gji ) can be arrived
at as follows:

6.2.1 Definitions and basic properties

interconnected SISO linear system.

2. Define a path P = {xn+2 , xj , . . . , xi , xn+1 }.

Thus the forward paths from xn+2 to xn+1 in (Sji , Gji ) are in 1 1-correspondence
with the loops containing Gji in (S, G), and which start with the branch Gji . If P
Pathn+2,n+1 (Sji , Gji ) then we denote by LP Loop(S, G) the loop associated to it. The
gain of P Pathn+2,n+1 (Sji , Gji ) is clearly exactly the gain of LP . Furthermore, the cofactor
of a path P Pathn+2,n+1 (Sji , Gji ) consists exactly of those terms in S,G not involving the
loop L Loop(S, G) giving rise to P . This fact can be employed to verify the equality
X
S,G = Sji ,Gji
GLP CofP (Sji , Gji ).

First of all, let us define what we mean by an

For example, the signal flow graphs of Figures 6.1, 6.4, 6.5, 6.6, and 6.12 are interconnected SISO linear systems, while that of Figure 6.7 is not.
We will want to determine the transfer function between any two signals in the signal
flow graph in order to discuss stability. Of course, Masons Rule makes this a comparatively
simple chore. Since our system does not necessarily have any inputs, in order to talk about
the transfer function from any signal to any other signal, we should introduce an input an
arbitrary node. We do this in the obvious way, as follows, supposing (S, G) to have n nodes.
For i {1, . . . , n} define the ith-appended system to be the signal flow graph (Si , Gi )
with Si = S {xn+1 = ui } and Gi = G {Gi,n+1 = 1}. Thus we simply tack on another
node with a branch of gain 1 going to node i. This renders the new node ui an input, and
the transfer function from node i to node j is then the graph transmittance Tji , this being
determined by Masons Rule. In particular, we define the transfer function of (S, G) by

P Pathn+2,n+1 (Sji ,Gji )

GP CofP (Sji , Gji ) = Sji ,Gji S,G .

P Pathn+2,n+1 (Sji ,Gji )

We also clearly have Sji ,Gji = S,G(j,i) since the loops in the graphs (Sji , Gji ) and (S, G(j,i) )
agree. Thus
P
S,G(j,i) S,G
P Pathn+2,n+1 (Sji ,Gji ) GP CofP (Sji , Gji )
=
,
Lji =
Sji ,Gji
S,G(j,i)
giving the result.

22/10/2004

6.20 Definition An interconnected SISO linear system is a connected signal flow graph
(S, G) with one source (the input) and one sink (the output). If the nodes for the system
are {x1 , . . . , xn },it is always assumed the source is x1 and the sink is xn . We assume that
the single source is an actual input (i.e., that there is one branch originating from x1 and
that the branch has gain 1). This can always be done without loss of generality by adding
the necessary node and branch if needed.
An interconnected SISO linear system (S, G) is proper (resp. strictly proper ) if all
gains in G are proper (resp. strictly proper).

1. Take a loop L {L1 , . . . , Lk }, ordered so that L = ({xi , xj , . . . }, {Gji , . . . }).

This then means exactly that

6 Interconnections and feedback

TS,G = Tn1 R(s).

If (NS,G , DS,G ) denotes the c.f.r. of TS,G , then we have reduced an interconnected SISO linear
system to a SISO linear system in input/output form.
Let us compute the transfer function for the examples we have been using.

This is, as always, easily exhibited in an example.

6.21 Examples We simply go through the steps and produce the transfer function after making the necessary simplifications. In each case, the gain Gij of a branch is represented by its
c.f.r. (Nij , Dij )).
1. For the series interconnection of Figure 6.4 we ascertain that the transfer function is

6.19 Example (Example 6.16 contd) We consider again the negative feedback loop of Figure 6.13. We had computed in Example 6.16 that
S,G(4,3) = 1,

S,G = 1 + G32 G43 ,

TS,G =

which gives
R43 = 1 + G32 G43 ,

L43 = G32 G43 .

We see how should interpret the loop transmittance in this example.

6.2 Interconnected SISO linear systems

The discussion of the previous section had a fairly general nature. Although this had the
advantage of allowing us to get a handle on the essential features of a signal flow graph, let us
bring things back into the realm of control systems. We do this in this section by introducing
the notion of an interconnected SISO linear system, and discussing the properties of such
things in a their general context. This treatment does not seem to appear in the current
control literature, oddly enough.

finish

N21 N32
.
D21 D32

2. For the parallel signal flow graph we have loops,

TS,G =

N21 N42 D31 D43 + N31 N43 D21 D42

.
D21 D42 D31 D43

3. Next we turn to the feedback loop of Figure 6.6. In Example 6.131 we computed the
transfer function from x1 to x5 to be
TS,G =

N21 N32 N43 N54

D21 D32 D43 D54
N32 N43 N24
1 D
32 D43 D24

6.2 Interconnected SISO linear systems

22/10/2004

231

Simplification gives
TS,G

In the case when G21 = G54 = G24 = 1 we get

22/10/2004

Let us perform these operations for the examples we have been toting around.
(6.8)

4. For the four loop signal flow graph of Figure 6.12 we determine the transfer function to
be
N61 N76 N87 D67 D78
.

TS,G =
D61 (D67 D76 (D78 D87 N78 N87 ) N67 N76 D78 D87 )
Of course, in simply writing the transfer function for an interconnected SISO linear
system we have eliminated a great deal of the structure of the signal flow graph. As when
one thinks of a SISO linear system = (A, b, ct , D) as being merely an input/output system
(see Section 2.3), one must take care if using only the transfer function to talk about an
interconnected SISO linear system. In order to see how this goes, we need to set up the
appropriate structure for an interconnected SISO linear system.
The first order of business is to set up equations of motion for a SISO linear system. We
give n sets of equations of motion, depending on where the input for the systems appears.
Of course, one can consider all inputs acting together by linearity. Note that the assumption
that x1 is an input, and the only source, ensures that the structure matrices for all of the
n appended systems are actually the same. Now we give a procedure for going from the
structure matrix GS,G of rational functions to a polynomial matrix AS,G and a polynomial
vector biS,G , corresponding to the ith-appended system (Si , Gi ).
6.22 Procedure for constructing (AS,G , biS,G ) from GS,G Given GS,G R(s)nn , do the following:
(i) let (Nij , Dij ) be the c.f.r. of each branch gain Gij G;
(ii) for each i {1, . . . , n}, let Gij1 , . . . , Gij` be the nonzero gains appearing in the ith row
of GS,G ;
(iii) multiply row i by the denominators Dij1 , . . . , Dij` ;
(iv) after doing this for each row, denote the resulting matrix by AS,G ;
(v) define biS,G to be the n-vector whose ith component is Dij1 . . . Dij` , the rest of the
components of biS,G being zero.

The equations of motion for the ith-appended system (Si , Gi ) are then the differential
equations

d
d
x(t) = biS,G dt
u(t),
(6.9)
AS,G dt
where x = (x1 , . . . , xn ) are the signals for (S, G) and where u = ui is the input at node i.
The procedure above systematises what one would do naturally in writing the equations of
motion obtain by doing a node balance. In this case, one would clear the denominators
so that all expressions would be polynomial, and so represent differential equations in the
time-domain. Let us introduce the notation R[s]nn for an n n matrix with components
in R[s]. Thus, for example, AS,G R[s]nn . Let us also define B S,G R(s) by

B S,G = b1S,G bnS,G ,
as a convenient way to catalogue the input vectors b1S,G , . . . , bnS,G .

6 Interconnections and feedback

6.23 Remark Clearly, the equations of motion for the ith appended system are exactly equivalent to the equations GS,G = ei , where ei is the ith standard basis vector. From this it
follows that
1
A1

S,G B S,G = GS,G .

N21 N32 N43 N54 D24

=
.
D32 D43 D21 D24 D54 N32 N43 N24 D21 D54
N32 N43
.
D32 D43 + N32 N43

232

6.24 Examples (Example 6.5 contd) We shall simply produce (AS,G , B S,G ) by applying Procedure 6.22.
1. We determine

1
0
0
1 0
0
0 , B S,G = 0 D21 0 .
AS,G = N21 D21
0
N32 D32
0 0 D32
2. We determine

1
0
0
0
N21
D21
0
0
,
AS,G =
N31
0
D31
0
0
N42 D43 N43 D43 D42 D43

B S,G

1 0
0
0
0 D21 0
0
.
=
0 0 D31
0
0 0
0 D42 D43

3. We determine

N21 D24

0
AS,G =

0
0

1
0

B S,G =
0
0
0

0
0
0
0
D21 D24
0
N24 D21 0

N32
D32
0
0

0
N43
D43
0
0
0
N54
D54

0
0
0
0
D21 D24 0
0
0

0
D32 0
0
.
0
0 D43 0
0
0
0 D54

4. Now let us also look at the four-loop example first introduced in Example 6.9 and shown
in Figure 6.12. We have not yet defined GS,G so let us do so:

1
0
0
0
0
0
0
0
G21
1
G23
0
0
0
0
0

G
1
G
0
0
0
0
32
34

0
G
1
0
0
0
0
43
.
GS,G =

0
0
0
G
1
0
0
G
54
58

G61

0
0
0
0
1
G
0
67

0
0
0
0
0 G76
1
G78
0
0
0
0
0
0
G87
1
We then have

6.2 Interconnected SISO linear systems

22/10/2004

0
0
0
0
0
0
0

N21 D23 D21 D23 N23 D21

0
0
0
0
0

0
N32 D34 D32 D34 N34 D32
0
0
0
0

0
0
N
D
0
0
0
0
43
43

=
0
0
0
N54 D58 D54 D58
0
0
N58 D54

N61 D67
0
0
0
0
D61 D67 N67 D61
0

0
0
0
0
0
N76 D78 D76 D78 N78 D76
0
0
0
0
0
0
N87
D87

1
0
0
0
0
0
0
0 D21 D23
0
0
0
0
0
0

0
0
D32 D34
0
0
0
0
0

0
0
0
D43
0
0
0
0
.
B S,G =
0
0
0
0 D54 D58
0
0
0

0
0
0
0
0
D61 D67
0
0

0
0
0
0
0
0
D76 D78
0
0
0
0
0
0
0
0
D87

AS,G

233

This procedure for coming up with (AS,G , B S,G ) is clearly simple enough, although perhaps
tedious, in any given example.

Now that we have the equations of motion (6.9) for an interconnected SISO linear system,
we can proceed to analyse these equations.
6.2.2 Well-posedness The next matter we deal with is a new one for us, the matter
of well-posedness. That there is something to talk about is best illustrated by an example.
6.25 Example We consider the signal flow graph of Figure 6.14 where we take RL (s) =

234

6 Interconnections and feedback

22/10/2004

Thus well-posedness is the requirement that n2 rational functions be proper. One would
like to derive simpler conditions for well-posedness. Starting down this road, the following
result gives an interpretation of well-posedness in terms of the determinant S,G .
6.27 Proposition A proper interconnected SISO linear system (S, G) is well-posed if and only
if lims S,G (s) 6= 0.
Proof Suppose that lims S,G (s) = L 6= 0. The graph transmittance Tji is given by
GP CofP (S, G)
.
S,G

Tji =

P Pathji (S,G)

Since (S, G) is proper, it follows that

lims GP (s)CofP (S, G)(s) is finite. Therefore,
X

lim Tji (s) =

P Pathji (S,G)

1
L

GP CofP (S, G)

proper

that

lims GP (s)CofP (S, G)(s)

lims S,G (s)

X
P Pathji (S,G)

lim GP (s)CofP (S, G)(s)

is finite. Thus Tji is proper.

Conversely, suppose that the matrix T S,G = G1
S,G of transmittances consists of proper
rational functions. Since the determinant of T S,G consists of sums of products of these proper
rational functions, it follows that det T S,G is itself a proper rational function. Therefore

1+ss2
.
s2 +s+1

lim det G1
S,G (s) 6= .

/ x2

/ x3

/ x4

From this we infer that

lim det GS,G (s) 6= 0,

so proving the result.

Figure 6.14 Unity gain feedback loop

Thus we have a perfectly well behaved collection of branch gains, and one can compute the
characteristic polynomial (see next section) to be PS,G (s) = 2(s + 2), which is Hurwitz. This
indicates that everything is pleasant. However, we compute the transfer function of the
system to be
1s
TS,G (s) =
.
2
This is problematic: an interconnection of proper branch gains has given rise to an improper
transfer function. Such cases are undesirable as improper transfer functions are certainly
not desirable (cf. Proposition 5.14).

Clearly one would like all graph transmittances to be proper rational functions, and this
leads to the following definition.

This gives the following sufficient condition for well-posedness, one that is satisfied in
many examples.
6.28 Corollary Let (S, G) be a proper interconnected SISO linear system. If each loop in (S, G)
contains a branch whose gain is strictly proper, then (S, G) is well-posed.
Proof Recall that the determinant is given by
S,G = 1 +

where G is a product of loop gains for nontouching loops. If (S, G) is proper, and each loop
contains a strictly proper branch gain, then we have
lim S,G (s) = 1,

6.26 Definition An interconnected SISO linear system (S, G) is well-posed if for each i
{1, . . . , n} the graph transmittance Tji R(s) is strictly proper for each j {1, . . . , n}.

implying well-posedness.

6.2 Interconnected SISO linear systems

22/10/2004

235

236

6 Interconnections and feedback

22/10/2004

This indicates that for many physical systems, whose branches will be comprised of
strictly proper rational functions, one can expect well-posedness to be typical. The following example reexamines our introductory example in light of our better understanding
of well-posedness.

By Proposition 5.13 we see that this input/output transfer function is BIBO stable, and so
on these grounds wed wash our hands of the stability question and walk away. In doing so,
wed be too hasty. To see why this is so, suppose that the system admits some noise n
as in
Figure 6.16. The transfer function between n
and y is then

6.29 Example (Example 6.25 contd) Let us still use the signal flow graph of Figure 6.14, but
now take
1 + 2 + as2
RL (s) = 2
,
s +s+1
where a R is unspecified. We compute

n
(s)

S,G (s) = 1 + RL (s) =

(1 + a)s2 + 2s + 2
.
s2 + s + 1

r(s)

s1
s+1

1
s1

y(s)

Figure 6.16 Trouble happening

We see that lims S,G (s) = 0 if and only if a = 1. Thus, even though this system does
not satisfy the sufficient conditions of well-posedness in Corollary 6.28, it is only in the very
special case when a = 1 that the system is not well-posed.

y(s)
1
=
n
(s)
s1

Well-posedness in a general context is discussed by Willems [1971], and for general (even
nonlinear) interconnected systems by Vidyasagar [1981]. When talking about well-posedness
for systems outside the rational function context we use here, one no longer has access to
simple notions like properness to characterise what it might mean for a system to be wellposed. Thus, for general systems, one uses a more basic idea connected with existence and
uniqueness of solutions. This is explored in Exercise E6.8.

which by Proposition 5.13 is not BIBO stable. So any slight perturbations in the signal as
it goes from the R1 block to the R2 block will potentially be dangerously magnified in the
output.

6.2.3 Stability for interconnected systems In Chapter 5 we looked at certain types

of stability: internal stability for SISO linear systems, and BIBO stability for input/output
systems. In this chapter, we are concerned with interconnections of systems in input/output
form, and the introduction of such interconnections gives rise to stability concerns that
simply do not arise when there are no interconnections. The difficulty here is that the
interconnections make possible some undesirable behaviour that is simply not captured by
the transfer function TS,G for the system.

6.31 Definition An interconnected SISO linear system (S, G) with nodes {x1 , . . . , xn } is
(i) internally stable if
lim sup kx(t)k <

A motivating example and definitions The following example makes this clear the difficulties one can encounter due to the introduction of even simple interconnections.

6.30 Example We consider the simple block diagram configuration of Figure 6.15. Thus
s1
s+1

r(s)

1
s1

y(s)

Figure 6.15 Trouble waiting to happen

R1 (s) =

s1
s+1

and R2 (s) =

1
.
s1

As we saw in Section 3.1, y = R1 R2 r. This gives

y(s)
1
=
.
r(s)
s+1

We now address the difficulties of the above example with a notion of stability that makes
sense for interconnected systems. The following definition provides notions of stability that
are relevant for interconnected SISO linear systems.

t
d
AS,G dt

for every solution x(t) of

x(t) = 0;
(ii) internally asymptotically stable if
lim kx(t)k = 0

d
x(t) = 0;
for every solution x(t) of AS,G dt
(iii) internally unstable if it is not internally stable;
(iv) BIBO stable if there exists a constant K > 0 so that the conditions (1) x(0) = 0
and (2) |u(t)| 1, t 0 imply that xn (t) K where u(t) and x(t) satisfy (6.9) with
i = 1;
(v) interconnected bounded input, bounded output stable (IBIBO stable) if for
each i {1, . . . , n}, the graph transmittance Tji is BIBO stable for j {1, . . . , n}.
For interconnected systems, we have all the notions of stability for normal systems, plus
we have this new notion of IBIBO stability that deals with the input/output stability of the
interconnection. This new type of stability formalises the procedure of adding noise to each
signal, and tapping the output of each signal to see how the system reacts internally, apart
from just looking at how the given input and output nodes act. Thus, if noise added to a node
gets unstably magnified in the signal at some other node, our definition of IBIBO stability will

6.2 Interconnected SISO linear systems

22/10/2004

237

capture this. Note that internal stability for these systems does not follow in quite the same
way as for SISO linear systems since the equations for the two systems are fundamentally
d

x(t) = 0 versus x(t)

= Ax(t). Indeed, one can readily see that the latter
different: AS,G dt
is a special case of the former (see Exercise E6.3). A systematic investigation of equations
d
x(t) = 0 for an arbitrary matrix of polynomials P is carried out by
of the form P ( dt
Polderman and Willems [1998]. This can be thought of as a generalisation of the computation
of the matrix exponential. Since we shall not benefit from a full-blown treatment of such
systems, we do not give it, although some aspects of the theory certainly come up in our
characterisation of the internal stability of an interconnected SISO linear system.
Before we proceed to give results that we can use to test for the stability of interconnected
systems, let us see how the above definition of IBIBO stability covers our simple example.
6.32 Example (Example 6.30 contd) It is more convenient here to use the signal flow graph,
and we show it in Figure 6.17 with the nodes relabelled to make it easier to apply the
x1 (s)

s1
s+1

/ x2 (s)

1
s1

/ x3 (s)

Figure 6.17 The signal flow diagram corresponding to the block

diagram of Figure 6.15

definition of IBIBO stability. The appended systems are shown in Figure 6.18. We compute

238

6 Interconnections and feedback

22/10/2004

so its kernel is defined as usual. Following what we do for R-matrices, if C is a root

of PS,G , let us denote the algebraic multiplicity by ma () and the geometric multiplicity by
mg ().
6.33 Theorem Consider a proper interconnected SISO linear system (S, G). The following
statements hold.
(i) (S, G) is internally unstable if spec(PS,G ) C+ 6= .
(ii) (S, G) is internally asymptotically stable if spec(PS,G ) C .
(iii) (S, G) is internally stable if spec(PS,G ) C+ = and if mg () = ma () for
spec(PS,G ) (iR).
(iv) (S, G) is internally unstable if mg () < ma () for spec(PS,G ) (iR).
Proof Just as the proof of Theorem 5.2 follows easily once one understands the nature of
the matrix exponential, the present result follows easily once one understands the character
d
= 0. The following lemma records those aspects of this that are
of solutions of AS,G dt
useful for us. We simplify matters by supposing that we are working with complex signals.
If the signals are real, then the results follow by taking real and imaginary parts of what we
do here.
1 Lemma Let P C[s]nn and suppose that det P is not the zero polynomial. Let
{1 , . . . , ` } be the roots of det P with respective multiplicities m1 , . . . , m` . Every solution
d
of P dt
x(t) = 0 has the form

u1
1

s1
s+1

/ x2

1
s1

x(t) =
/ x3

` m
i 1
X
X

ij tj ei t

i=1 j=0
n

for some appropriate

ij C , i = 1, . . . , `, j = 0, . . . , mi 1. In particular, the set of
d
solutions of P dt
x(t) = 0 forms a R-vector space of dimension deg(det P ).

u2
1

s1
s+1

/ x2

1
s1

/ x3

Figure 6.18 The ith-appended systems for Figure 6.17 with i = 1

(top) and i = 2 (bottom)

the graph transmittances to be

T21 (s) =

s1
,
s+1

T31 (s) =

1
,
s+1

T32 (s) =

1
.
s1

Note that the transfer function T32 is BIBO unstable by Proposition 5.13. Therefore, the
interconnected system in Figure 6.17 is not IBIBO stable.

Conditions for internal stability Let us now produce some results concerning the various
types of stability. To get things rolling, we define the characteristic polynomial to be
PS,G = det AS,G . The algebraic multiplicity of a root of PS,G is the multiplicity of the
root of a polynomial in the usual sense (see Section C.1). The geometric multiplicity of a
root of PS,G is the dimension of the kernel of the matrix AS,G (). Note that AS,G () Rnn ,

Proof We prove the lemma by induction on n. It is obvious for n = 1 by well-known

properties of scalar differential equations as described in Section B.1. Now suppose the
lemma true for n {1, . . . , k 1} and let P R[s]kk . Note that neither the solutions of
d
x(t) = 0 nor the determinant det P are changed by the performing of
the equations P dt
elementary row operations on P . Thus we may suppose that P has been row reduced to
the form

P11 P12 P1k

P = ..

P
0

R[s](k1)(k1) . By the induction hypothesis every solution to Q d x(t) = 0
for some P
dt
has the form
` m
i 1
X
X
tj ei t ,
(t) =
x

ij
i=1 j=0

ij Cn , i = 1, . . . , `,
j = 0, . . . , mi 1. Thus {1 , . . . , } are the
for some appropriate
`

roots of det P with respective multiplicities m1 , . . . , m`. Substituting such a x(t) into the

6.2 Interconnected SISO linear systems

22/10/2004

first of the equations P

d
dt

239

x(t) = 0 gives a differential equation of the form

P11

d
dt

x1 (t) =

` m
i 1
X
X

ij tj ei t ,

(6.10)

i=1 j=0

j = 0, . . . , mi 1. Thus the constants ij will

for some appropriate ij C, i = 1, . . . , `,
s, the coefficients of P12 , . . . , P1k , and roots 1 , . . . , .
depend on the components of the
ij
`
To solve this equation (or at least come up with the form of a solution)
we recall that it will

d
be the sum of a homogeneous solution xih (t) satisfying P11 dt
x1h (t) = 0 and a particular
solution. Let us investigate the form of both of these components of the solution. The
homogeneous solution has the form
0

x1h (t) =

mi
` X
X

ij0 tj ei t ,

i=1 j=0

{01 , . . . , 0`0 }

where
are the roots of P11 with respective multiplicities m01 , . . . , m0`0 . If there
then the particular solution will have the form
are no common roots between P11 and det P
x1p (t) =

` m
i 1
X
X

ij00 tj ei t ,

i=1 j=0

do share
i.e., the same form as the right-hand side of (6.10). However, if P11 and det P
have a common
roots, we have to be a little more careful. Suppose that P11 and det P
. A moments reflection shows
root of multiplicity m0 for P11 and multiplicity m for det P
that the method of undetermined coefficients, as outlined in Procedure B.2, then produces
a particular solution corresponding to this common root of the form
0
m+m
X1

j00 tj et .

j=0

Collecting this all together yields the lemma by induction once we realise that det P =
.
P11 det P
H
n

(i) If spec(PS,G ) C+ then there exists a vector u C so that AS,G ()u = 0. If

x(t) = et u then we have

d
AS,G dt
x(t) = et AS,G ()u = 0.

d
Thus there is a solution of AS,G dt
x(t) = 0 that is unbounded.

d
(ii) From Lemma 1 we know that every solution of AS,G dt
x(t) = 0 is a vector function
with components being linear combinations of functions of the form
tj et

(6.11)

where spec(PS,G ). By hypothesis, it follows that every such function approaches 0 as
d
t , and so every solution of AS,G dt
x(t) = 0 approaches 0 as t . We refer to the
proof of Theorem 5.2 to see how this is done properly.

240

6 Interconnections and feedback

22/10/2004

d
(iii) By Lemma 1, every solution of AS,G dt
x(t) = 0 is a vector functions whose components are linear combinations of terms of the form (6.11) for Re() < 0 and terms of the
form
tj eit .
(6.12)
We must show that the condition that ma (i) = mg (i) implies that the only solutions
of the form (6.12) that are allowed occur with j = 0. Indeed, since ma (i) = mg (i), it
follows that there are ` , ma () linearly independent vectors, u1 , . . . , u` , in ker(AS,G (i)).
Therefore, this implies that the functions
ui eit , . . . , u` eit
are linearly independent solutions corresponding to the root i spec(PS,G ). By Lemma 1,
there are exactly ` such functions, so this implies that as long as ma (i) = mg (i), all
d
x(t) = 0 have the form (6.12) with j = 0. Now we
corresponding solutions of AS,G dt
proceed as in the proof of Theorem 5.2 and easily show that this implies internal stability.
(iv) We must show that the hypothesis that ma (i) > mg (i) implies that there is at
least one solution of the form (6.12) with j > 0. However, we argued in the proof of part (iii)
that the number of solutions of the form (6.12) with j = 0 is given exactly by mg (i).
Therefore, if ma (i) > mg (i) there must be at least one solution of the form (6.12) with
j > 0. From this, internal instability follows.

6.34 Remarks 1. Thus, just as with SISO linear systems, internal stability of interconnected
systems is a matter checked by computing roots of a polynomial, and possibly checking
the dimension of matrix kernels.

d
2. Note that Theorem 6.33 holds for arbitrary systems of the form P dt
x(t) = 0 where
P R[s]nn .

Lemma
1 of the proof of Theorem 6.33 is obviously important in the study of the system
d
AS,G dt
x(t) = 0, as it infers the character of the set of solutions, even if it does not give an
completely explicit formula for the solution as is accomplished by the matrix exponential.
In

d
particular, if deg PS,G = N then there are N linearly independent solutions to AS,G dt
x(t) =
0. Let us denote these solutions by
x1 (t), . . . , xN (t).
Linear independence implies that for each t the matrix

X(t) = x1 (t) xN (t)
has full rank in the sense that if there exists a function c : R RN so that X(t)c(t) = 0
for all t, then it follows that c(t) = 0 for all t.
Since the characteristic polynomial is clearly an interesting object, let us see how the
computation of the characteristic polynomial goes for the systems we are working with.
6.35 Examples As always, we bypass the grotesque calculations, and simply produce the
characteristic polynomial.
1. For the series interconnection of Figure 6.4 we ascertain that the characteristic polynomial
is simply PS,G = D21 D32 .

6.2 Interconnected SISO linear systems

22/10/2004

241

2. For the parallel signal flow graph we have

242

6 Interconnections and feedback

Proof This follows from the manner in which we arrived at AS,G from GS,G . Let us systematise this in a way that makes the proof easy. Let us order I by ordering n n as
follows:

PS,G = D21 D42 D31 D43 .

3. For the negative feedback loop of Figure 6.6 we compute

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

PS,G = D21 D54 (D32 D43 D54 N32 N43 N24 ).

In the case when G21 = G54 = G24 = 1 we get
PS,G = D32 D43 + N32 N43 .

(6.13)

Thus we order n n first ordering the rows, then ordering by column if the rows are equal.
This then gives a corresponding ordering of I nn, and let us denote the elements of I by
(i1 , j1 ), . . . , (i` , j` ) in order. Now, define A0 = GS,G and inductively define Ak , k {1, . . . , `},
by multiplying the ik th row of Ak1 by Dik ,jk . Thus, in particular AS,G = A` . Now note
that by the properties of the determinant,

4. For the four loop interconnection of Figure 6.12 we compute

det Ak = Dik ,jk det Ak1 .

PS,G = D21 D54 D58 D61 (D23 D32 (D34 D43 N34 N43 ) (D34 D43 N23 N32 ))
(D67 D76 (D78 D87 N78 N87 ) (D78 D87 N67 N76 ))

Thus we have

It is important to note that the characteristic polynomial for an interconnected SISO

linear system, is not the denominator of the transfer function, because in simplifying the
transfer function, there may be cancellations that may occur between the numerator and
the denominator. Let us illustrate this with a very concrete example.
6.36 Example We take the negative feedback loop of Figure 6.6 and we take as our gains the
following rational functions:
G21 (s) = 1,

G32 (s) =

s1
,
s+1

G43 (s) =

1
,
s1

G54 (s) = 1,

PS,G (s) = (s + 1)(s 1) + (s 1)1 = s2 + s 2.

If we write the transfer function as given by (6.8) we get
s1
1
=
.
s2 + 2 2
s+2

Note that PS,G (s) 6= s + 2!

Conditions for BIBO and IBIBO stability Next we wish to parlay our understanding of
the character of solutions to the equations of motion gained in the previous section into
conditions on IBIBO stability. The following preliminary result relates PS,G and S,G .
6.37 Proposition Let (S, G) be an interconnected SISO linear system with interconnections I,
and for (i, j) I, let (Nij , Dij ) be the c.f.r. for the gain Gij G. Then
Y
PS,G = S,G
Dij .
(i,j)I

det A1 = Di1 ,j1 det GS,G

det A2 = Di2 ,j2 det A1 = Di2 ,j2 Di1 ,j1 det GS,G
..
.
det A` = Di` ,j` det A`1 = det GS,G

`
Y

Dik ,jk .

k=1

The result now follows since PS,G = det AS,G and S,G = det GS,G .

Thus we see a strong relationship between the characteristic polynomial and the determinant. Since Theorem 6.33 tells us that the characteristic polynomial has much to do with
stability, we expect that the determinant will have something to do with stability. This
observation forms the basis of the Nyquist criterion of Chapter 7. In the following result,
we clearly state how the determinant relates to matters of stability. The following theorem
was stated for a simple feedback structure, but in the MIMO context, by Desoer and Chan
[1975]. In the SISO context we are employing, the theorem is stated, but strangely not
proved, by Wang, Lee, and He [1999].

G24 (s) = 1.

The characteristic polynomial, by (6.13), is

TS,G (s) =

22/10/2004

6.38 Theorem Let (S, G) be a proper, well-posed interconnected SISO linear system with interconnections I. The following statements are equivalent:
(i) (S, G) is internally asymptotically stable;
(ii) (S, G) is IBIBO stable;
(iii) the characteristic polynomial PS,G is Hurwitz;
(iv) the following three statements hold:
(a) S,G has all of its zeros in C ;
(b) there are no cancellations of poles and zeros in C+ in the formation of the individual loop gains;
(c) for any path P connecting the input of (S, G) to the output there are no cancellations of poles and zeros in C+ in the formation of the gain GP .
Furthermore, each of the above four statements implies the following statement:
(v) (S, G) is BIBO stable.

6.2 Interconnected SISO linear systems

22/10/2004

243

244

Proof Theorem 6.33 establishes the equivalence of (i) and (iii). Let us establish the equivalence of (ii) with parts (i) and (iii).
(iii) = (ii) We must show that all graph transmittances are BIBO stable transfer functions. Let T S,G be the matrix of graph transmittances as in Corollary 6.12. From Corollary 6.12 and Remark 6.23 we know that

6 Interconnections and feedback

22/10/2004

does not. This is also illustrated in this case by the conditions (iv a), (iv b) and (iv c) of
Theorem 6.38. Thus we compute
S,G = 1 +

1
s+2
=
s+1
s+1

and so condition (iv a) is satisfied. However, condition (iv b) is clearly not satisfied, and all
three conditions must be met for IBIBO stability.

1
T S,G = G1
S,G = AS,G B S,G .

In particular, it follows that for each (i, j) I we have

Tji =

6.3 Feedback for input/output systems with a single feedback loop

Qji
PS,G

Although in the previous two sections we introduced a systematic way to deal with very
general system interconnections, SISO control typically deals with the simple case where we
have an interconnection with one loop. In this section we concentrate on this setting, and
provide some details about such interconnections. We first look at the typical single loop
control problem with a plant transfer function that is to be modified by a controller transfer
function. In particular, we say what we mean by open-loop and closed-loop control. Then,
in Section 6.3.2 we simplify things and look at a generic single loop configuration, identifying
in it the features on which we will be concentrating for a large part of the remainder of these
notes.

for some Qji R[s]. From this it follows that if PS,G is Hurwitz then (S, G) is IBIBO stable.
(ii) = (i) From Corollary 6.12 and Remark 6.23 it follows that each of the graph
transmittances can be written as
Qji
Tji =
PS,G
for some Qij R[s]. We claim that there is at least one (i, j) I so that the polynomials
Qij and PS,G are coprime.
Now we use Proposition 6.37 to show that parts (iii) and (iv) are equivalent.
Finally, (v) follows from (ii), by definition of IBIBO stability.

6.39 Remark Note that BIBO stability of (S, G) obviously does not necessarily imply IBIBO
stability (cf. Example 6.30). This is analogous to SISO linear systems where internal stability
is not implied by BIBO stability. This is a property of possible pole/zero cancellations when
forming the transfer function TS,G . For SISO linear systems, we saw that this was related
to controllability and observability. This then raises the question of whether one can talk
intelligibly about controllability and observability for interconnected SISO linear systems.

One can, but we will not do so, referring instead to [Polderman and Willems 1998].

finish

6.3.1 Open-loop versus closed-loop control We shall mainly be interested in considering feedback as a means of designing a controller to accomplish a desired task. Thus
we start with a rational function RP R(s) that describes the plant (i.e., the system
about whose output we care) and we look to design a controller RC R(s) that stabilises
the system. The plant rational function should be thought of as unchangeable. One could
simply use an open-loop controller and design RC so that the open-loop transfer function RP RC has the desired properties. This corresponds to the situation of Figure 6.20.1
However, as we saw in Section 1.2, there are serious drawbacks to this methodology. To

If one determines the characteristic polynomial by computing the input/output transfer

function, then simplifying, one must not cancel factors in the numerator and denominator.
Let us recall why this is so.

r(s)

6.40 Example (Example 6.36 contd) We were looking at the negative feedback system depicted in Figure 6.19. Here the transfer function was determined to be
x1

/ x2 g

s1
s+1

/ x3

1
s1

/ x4

RC (s)

r(s)

RP (s)

RC (s)

RP (s)

y(s)

/ y(s)

Figure 6.20 Open-loop control configuration as a block diagram

(top) and a signal flow graph (bottom)

/ x5

Figure 6.19 Be careful computing the characteristic polynomial

TS,G (s) =

1
s+2

while the characteristic polynomial is PS,G (s) = s2 + s 2 = (s + 2)(s 1). Thus while
the denominator of the transfer function has all roots in C , the characteristic polynomial

get around these we design the controller to act not on the reference signal r, but on
the error signal e = r y. One may place the controller in other places in the block
diagram, and one may have other rational functions in the block diagram. However, for such
systems, the essential methodology is the same, and so let us concentrate on one type of
system interconnection for the moment for simplicity. The block diagram configuration for
the so-called closed-loop system we consider is depicted in Figure 6.21, and in Figure 6.22
we show some possible alternate feedback loops that we do not look at in detail.
1

We place a at a node in the signal flow graph that we do not care to name.

6.3 Feedback for input/output systems with a single feedback loop

22/10/2004

r(s)

RC (s)

RP (s)

245

RC (s)

RP (s)

6 Interconnections and feedback

22/10/2004

6.41 Input/output control design problem Given a rational function RP describing the plant,
find a rational function RC describing the controller, that make the closed-loop transfer
function behave in a suitable manner. In particular, one will typically wish for the poles of
the closed-loop transfer function to be in C .

y(s)

r(s)

246

/ y(s)

When doing this, the concerns that we will raise in Section 6.2.3 need to be taken into
account. Whats more, there are other concerns one needs to be aware of, and some of these
are addressed in Chapter 8. That is, one cannot look at the transfer function yr as being the
only indicator of system performance.
Let us look at an example of designing a closed-loop control law for a given plant.

Figure 6.21 Closed-loop control configuration as a block diagram

(top) and a signal flow graph (bottom)

r(s)

RC (s)

RP (s)

y(s)

6.42 Example We consider the plant transfer function RP (s) = 1s . Suppose we give the
system a step input: u(t) = 1(t). Then we have u(s) = 1s , and so the output in the Laplace
transform domain will be y(s) = s12 . From this we determine that y(t) = t. Thus the output
blows up as t .
Lets try to repair this with an open-loop controller. We seek a plant rational function
RC so that RC RP has all poles in C . If (NC , DC ) is the c.f.r. of RC , we have

RC (s)RP (s) =
RS (s)

r(s)

RR (s)

RC (s)

RP (s)

y(s)

Thus the partial fraction expansion of RC RP will always contain a term like as for some
a R unless we cancel the denominator of the plant transfer function with the numerator
of the controller transfer function. However, this is a bad idea. It essentially corresponds to
introducing unobservable dynamics into the system. So this leaves us with the term as in the
partial fraction expansion, and with a step response, the output will still blow up as t .
This motivates our trying a closed-loop scheme like that in Figure 6.21. We take RC (s) =
1
so that our closed-loop system is as depicted in Figure 6.23. The closed-loop transfer
s+1
r(s)

r(s)

RP (s)

y(s)

NC (s)
.
sDC (s)

1
s+1

1
s

y(s)

RC (s)

Figure 6.23 A closed-loop scheme for a simple plant

Figure 6.22 Some alternate feedback loops: transfer function in
feedback path (top); modiifed reference signal (middle); controller in feedback path (bottom)

function in this case is readily computed to be

1 1

The closed-loop transfer function from r to y is readily computed as

y
RP RC
=
.
r 1 + RP RC
The objective in the input/output scenario is described by the following problem statement.

RC (s)RP (s)
1
= s+1 1s 1 = 2
.
1 + RC (s)RP (s)
s +s+1
1 + s+1 s
The poles of the closed-loop transfer function are now 12 i
step response in the Laplace transform domain is
y(s) =

3
,
2

which are both in C . The

RC (s)RP (s)
1
1
u(s) = 2
.
1 + RC (s)RP (s)
s +s+1s

PSfrag replacements

1 (t)

ke(t)k
e1 (t)
e2 (t)
e3 (t)
22/10/2004
6.3 Feedback for input/output systems with a single feedback loop
247
y(t)
u(t)
The inverse Laplace
h (t) transform can be computed using partial fraction expansion. We have
hN,D (t)
1
1
1
s
1
= 2

,
s2 + s + 1 s
s s + s + 1 s2 + s + 1
1N,D (t)
(t)
from which wecan
(t) use our formulas of Section E.3 to ascertain that

fj (t)
y(t) = 1 et/2 cos 23 t + 13 sin 23 t .
Re
Im
x1
We plot this response
in Figure 6.24, Note that the closed-loop step response is now
x2
x1
x2
log 1.2
dB
1
deg
u = 0 ln 0.8
ln coth(|u| /2)
or 1 0.6
m
0.4
yos
0.2
tos 0

2
4
6
8
10
12
14
,0
t
0
Figure 6.24 Closed-loop step response for simple plant and controller

boundedsomething we were not able to legitimately accomplish with the open-loop controller. We shall develop ways of designing controllers for such systems later in the course.
6.3.2 Unity gain feedback loops In this section we focus on the feedback loop depicted in Figure 6.25. While in the previous section we had looked at the case where

248

6 Interconnections and feedback

control design. Although it is true that by restricting to such an interconnection we loose

some generality, it is not difficult to adapt what we say here to more general single-loop
configurations, perhaps with transfer functions in the feedback loop, or a transfer function
between the reference and the input to the loop.
The signals in the block diagram Figure 6.25 are the reference r(t), the output y(t),
the error e(t), the disturbance d(t), and the noise n(t). Throughout this chapter we will
encounter the various transfer functions associated with the block diagram Figure 6.25, so
let us record them here so that we may freely refer to them in the sequel:
y
RL
=
,
r 1 + RL
e
1
=
,
r 1 + RL

r(s)

d(s)

RL (s)

y
1
=
,

1
+
RL
d
1
e
=
,
d 1 + RL

y
RL
=
,
n

1 + RL
e
1
=
.
n

1 + RL

We see that there are essentially two transfer functions involved here:
TL =

RL
1 + RL

and SL =

1
.
1 + RL

These transfer functions are given the name of complementary sensitivity function,
and sensitivity function, respectively. Note that
TL + SL = 1.
Of course, the complementary sensitivity function is simply the closed-loop transfer function
from the input to the output. The sensitivity function is, in the parlance of Section 6.1.6,
the sensitivity of TL to RL . In Chapters 8 and 9 we will be seeing the importance of each of
these two transfer functions, and we will get a look at how they can interact in the design
process to make things somewhat subtle.
6.3.3 Well-posedness and stability of single-loop interconnections When making
a single-loop interconnection of the type in Figure 6.26, we have the notions of stability and
r(s)

e(s)

22/10/2004

RC (s)

RP (s)

y(s)

n
(s)

Figure 6.26 Block diagram single-loop feedback

Figure 6.25 A unity gain feedback loop

well-posedness given in Sections 6.2.3 and 6.2.2. Let us examine these ideas in our simple
single-loop context.
We first show how to easily characterise well-posedness for single-loop interconnections.

RL = RC RP , in this section our interests are more in the structure of the block diagram
than in the desire to design a controller rational function RC . In this case RL is often called
the open-loop transfer function for the interconnection, meaning that it is the transfer
function if the feedback connection is snipped. It is the relationship between the open-loop
transfer function and the closed-loop transfer function that lies at the heart of classical

6.43 Proposition Let RL = RC RP R(s) be proper and consider the interconnection of

Figure 6.26. Let L = (A, b, ct , D) be the canonical minimal realisation for RL . The
following statements are equivalent:
(i) the interconnection is well-posed;

22/10/2004

6.3 Feedback for input/output systems with a single feedback loop

249

250

6 Interconnections and feedback

22/10/2004

(i) RC S (RP );
(ii) the following two statements hold:

(ii) lims RL (s) 6= 1;

(iii) D 6= [1].
Proof (i) = (ii) Suppose that (ii) does not hold so that lims RL (s) = 1. Then we
have
RL (s)
lim TL (s) = lim
= ,
s
s 1 + RL (s)
and similarly lims SL (s) = . This implies that both TL and SL must be improper, so
the interconnection cannot be well-posed.
(ii) = (iii) Recall from the proof of Theorem 3.20 that since L is the canonical minimal
realisation of RL we have
RL (s) = TL (s) =

cn sn + cn1 sn1 + + c1 s + c0
sn + pn1 sn1 + + p1 s + p0

where D = [
cn ] and ci = ci +
cn pi , and where p0 , . . . , pn1 are the coefficients in the characteristic polynomial for A. One now computes lims TL (s) = cn . Thus lims RL (s) 6= 1
implies that D 6= [1], as desired.
(iii) = (i) From the previous step in the proof, if D = [d] then lims RL (s) = d.
Therefore,
d
RL (s)
=
lim TL (s) = lim
s
s 1 + RL (s)
1+d
is finite if d 6= 1. Also, if d 6= 1 then
lim SL (s) =

1
.
1+d

Thus indicates that if d 6= 1 then TL and SL are both proper.

(a) the characteristic polynomial DC DP + NC NP is Hurwitz;

(b) lims RC (s)RP (s) 6= 1;
(iii) the following three statements hold:
(a) RC and RP have no pole/zero cancellations in C+ ;
(b) 1 + RC RP has no zeros in C+ ;
(c) lims RC (s)RP (s) 6= 1.
Here (NC , DC ) and (NP , DP ) are the c.f.r.s of RC and RP , respectively.
6.46 Remark While Proposition 6.45 characterises the stabilising controllers, it does not answer the question concerning their existence, and if they exist, how many there are. We shall
deal with this in subsequent parts of the text, but since the questions are so fundamental,
let us now point to where the answers can be found.
1. For any strictly proper plant RP of order n, there exists a strictly proper controller
RC S (RP ) of order n (Theorems 10.27 and 13.2) and a proper controller RC S (RP )
of order n 1 (Theorem 13.2).
2. For any proper plant RP of order n, there exists a strictly proper controller RC S (RP )
of order n (Theorems 10.27 and 13.2).
3. For a proper plant RP , there is (essentially) a bijection from the set S (RP ) to RH+

(Theorem 10.37).

6.4 Feedback for SISO linear systems

The following obvious corollary indicates that well-posedness is not a consideration for
strictly proper loop gains.
6.44 Corollary If the loop gain RL in Figure 6.25 is strictly proper, then the interconnection
is well-posed.
Proof This follows from Proposition 6.43 since if RL is strictly proper than lims RL (s) =
0.

Next let us turn to stability of single-loop interconnections. It will be convenient to
introduce some notation. Given a plant RP , let us denote by S (RP ) the collection of IBIBO
stabilising controllers for which the closed-loop system is well-posed. Thus we consider the
interconnection shown Figure 6.26 and we take
S (RP ) = {RC R(s) the interconnection of Figure 6.26 is IBIBO stable}.

We look at our state-space model

x(t)
= Ax(t) + bu(t)
y(t) = ct x(t) + Du(t),

(6.14)

and ask, What should feedback mean for such a system? One can, of course, write the
transfer function T as a quotient of coprime polynomials, and then proceed like we did
above in the input/output case. However, it is not immediately clear what this means in
the framework of the SISO linear system = (A, b, ct , D). Indeed, it is not even clear what
types of transfer functions ought to be achievable via a SISO linear system because the
inputs for such a system appear in a specific way.
What needs to be undertaken is a description of feedback for SISO linear systems, independent of those in input/output form. The idea is that for the system (6.14) we should
take as feedback a linear function of the state x and the output y. We first consider the case
of pure state feedback, then allow the output to be fed back.

Of course, one can characterise the set of stabilising controllers fairly concretely using the
general machinery of Section 6.2.3. One readily sees that the following result follows directly
from Theorem 6.38 (the reader can provide a direct proof of this in Exercise E6.7).

6.4.1 Static state feedback for SISO linear systems State feedback should be of
the form u(t) = r(t) f t x(t) for some f Rn , where r is the reference signal. In block
diagram form, the situation is illustrated in Figure 6.27. One readily ascertains that the
closed-loop equations are

6.45 Proposition Let RP R(s) be proper. For RC R(s) the following statements are
equivalent:

x(t)
= (A bf t )x(t) + br(t)
y(t) = (ct Df t )x(t) + Dr(t).

6.4 Feedback for SISO linear systems

22/10/2004

251

(sI n A)1

6 Interconnections and feedback

22/10/2004

The following result says, at least, what we can do with the input/output relation. Note
that the result also says that for a controllable system, the eigenvalues of the closed-loop
system can be arbitrarily placed.

r(s)

252

y(s)

6.49 Theorem Let = (A, b, ct , D) be a SISO linear system with (A, b) controllable. A
rational function R R(s) is state compatible with if and only if there exists a monic
polynomial P R[s] of degree n so that
R(s) =

ct adj(sI n A)b + dPA (s)

,
P (s)

where D = [d], and where PA is the characteristic polynomial for A. In particular, if (A, b)
is controllable, then Ss () 6= .

f t

Proof The closed-loop transfer function for is

Tf (s) =

r(s)

(sI n A)1

(ct Df t )adj(sI n (A bf t ))b

+ D.
det(sI n (A bf t ))

4/ y(s)

Figure 6.27 The static state feedback configuration for the SISO
linear system (6.14) as a block diagram (top) and a signal flow
graph (bottom)

Motivated by this, we have the following definition.

6.47 Definition Let = (A, b, ct , D) be a SISO linear system. A state feedback vector is
a vector f Rn , and to a state feedback vector f we assign the closed-loop SISO linear
system f = (A bf t , b, ct Df t , D). The transfer function for f is called the closedloop transfer function. A rational function R R(s) is state compatible with is
there exists a state feedback vector f with the property that Tf = R.

The control problem here is a bit different than that of the input/output problem.
6.48 Static state feedback design problem Given the system = (A, b, ct , D), find a state
feedback vector f so that
(i) the closed-loop transfer function f has desirable properties and
(ii) the state variables are behaving in a nice fashion.
In particular, one will typically want the matrix A bf t to be Hurwitz.

By recognising that we have states, we are forced to confront the issue of their behaviour,
along with the behaviour of the input/output system. Following our notation for stabilising
controller transfer functions for input/output systems, given = (A, b, ct , D), let us denote
by Ss () the set of stabilising state feedback vectors. Thus

Ss () = f Rn | A bf t is Hurwitz .
The subscript s means state, as we shall shortly look at output feedback as well.

Since (A, b) is controllable, we may as well suppose that

0
1
0
0
0
0
0
1
0
0

0
0
0
1
0

A = ..
..
..
..
.. ,
...
.

.
.
.
.

0
0
0
0
1
p0 p1 p2 p3 pn1

0
0

0

b = .. ,
.

0
1

and let us write c = (c0 , c1 , . . . , cn1 ). If f = (f0 , f1 , . . . , fn1 ) a simple calculation gives

0 0 0
0
0 0 0
0

0 0 0
t
0
bf =

.. .. .. . .
..
. . .
.
.
f0 f1 f2 fn1
and so

0
0
0
..
.

1
0
0
..
.

0
1
0
..
.

0
0
1
..
.

..
.

t
A bf =

0
0
0
0

p0 f0 p1 f1 p2 f2 p3 f3

0
0
0
..
.
1
pn1 fn1

This shows that by choosing f appropriately, we may make the characteristic polynomial of
A bf t anything we like.
Now we note that
(ct Df t )adj(sI n A)b = (ct Df t )adj(sI n (A bf t ))b

PSfrag replacements

if c = (c0 , c1 , . . . , cn1 ) and D = [d]. Now we observe that if P (s) = det(sI n (A bf t ))

then we have
fn1 sn1 + + f1 s + f0 = P (s) PA (s).
Therefore
t

(c Df )adj(sI n (A bf ))b = c adj(sI n A)b d P (s) PA (s) .
The theorem now follows by straightforward simplification.

This result is important because it demonstrates that by choosing the appropriate static
state feedback for a controllable system , we may do as we please with the poles of the
closed-loop transfer function. And, as we have seen in Proposition 3.24 and Corollary 5.7,
the poles of the transfer function have a great deal of effect on the behaviour of the system.
Let us do this in an ad hoc way in an example.
6.50 Example We take

A=

0 1
,
1 0

0
.
1

Note that A is not Hurwitz as it has characteristic polynomial s2 + 1. Without any justification (for this, refer ahead to Example 10.15) we take as state feedback vector f = (3, 4).
We then have

0 1
0
0
1
3 4 =
A bf t =

.
1 0
1
4 4
Since (A bf t , b) is in controller canonical form, the characteristic polynomial can be read
from the bottom row: s2 + 4s + 4.
Lets look at the behaviour of the open-loop system. We compute

cos t sin t
eAt =
.
sin t cos t
If we provide the periodic input u(t) = 1(t) cos t and zero initial condition, the time-response
of the state of the system is

1
t sin t
2
x(t) = 1
(t cos t + sin t)
2
which we plot in Figure 6.28. Taking c = (1, 0), the corresponding output is
y(t) = 12 t sin t
which we plot in Figure 6.29.
For the closed-loop system we compute
2t

t
e + 2te2t
te2t
e(Abf )t =
4te2t
e2t 2te2t

ke(t)k
6 Interconnections
feedback
e1and
(t)
e2 (t)
e3 (t)
y(t)
u(t)
h (t)
hN,D (t)
1 (t)
1N,D (t)
(t)
(t)
fj (t)
Re
Im

22/10/2004

0.2
0.15
0.1
0.05

x1
x2
PSfrag replacements
log
dB
ke(t)k
deg
u = e0 1ln(t)

ln coth(|u|
e2/2)
(t)
1
ore3(t)
m

u(t)
yos
h (t)
tos(t)
0
hN,D
1 (t)

-2

-4

0
1N,D(t)
(t)
(t)
fj (t)
Re
Im
x1
x2
x1
10
x2
7.5
log
dB
5
deg
u = 0 ln
2.5
ln coth(|u| /2)
0
or 1
m -2.5

-5
yos
-7.5
tos 0

,0
0

-2

x1
x2
PSfrag replacements
log
dB
ke(t)k
deg
u = e0 1ln(t)

ln coth(|u|
e2/2)
(t)
1
ore3(t)
m

u(t)
yos
h (t)
tos(t)
0
hN,D
4
1 (t)

(cn1 dfn1 )sn1 + + (c1 df1 )s + (c0 df0 ).

254

0
-0.05
-0.1
-0.15

-0.15 -0.1 -0.05

0.05

0.1

0.15

0.2

17.5

0
1N,D(t)
(t)
(t)
j (t)
Figure 6.28 State response for fopen-loop
system (left) and closedRe
loop system (right) under static
Im state feedback
x1
x2
x1
x2
0.2
log
dB
deg 0.1
u = 0 ln
ln coth(|u| /2)
0
or 1
m
-0.1
yos

tos 0 -0.2

2.5
5
7.5
10 12.5 15
2.5
5
7.5
10 12.5 15 17.5 20,0
t
t
0

y(t)

for any vector f t . This is because both A and Abf t are in controller canonical form, which
means that the polynomials (ct Df t )adj(sI n A)b and (ct Df t )adj(sI n (A bf t ))b
are both given by

PSfrag replacements

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)
h (t)
hN,D (t)
1 (t)
1N,D (t)
(t)
(t)
fj (t)
Re
Im
x2

253

y(t)

6.4 Feedback for SISO linear systems

22/10/2004

Figure 6.29 Output response for open-loop system (left) and

closed-loop system (right) under static state feedback

from which we ascertain the state behaviour to be

3

3
4
25 e2t 25 te2t + 25
cos t + 25
cos t
x(t) =
4 2t
4 2t
4
3
25 e 5 te + 20 cos t 25 sin t
and the output to be
y(t) =

3
2
3
4
e2t te2t +
cos t +
cos t.
25
5
25
25

These are shown beside the open-loop response in Figures 6.28 and 6.29, respectively.
As expected, the addition of static state feedback has caused the system to behave in a
more suitable manner. In fact, for this example, it has taken an BIBO unstable system and
made it BIBO stable.

The matter of static state feedback is also attended to in Section 10.1.1where a better
understanding of when static state feedback can make a closed-loop system stableand

6.4 Feedback for SISO linear systems

22/10/2004

255

in Section 10.2.1where methods of constructing such feedback laws are discussed. We

also mention that those attracted to the signal flow graph technology for feedback might
be interested in looking at [Reinschke 1988] where a presentation of static state and static
output feedback (see the next section for the latter) appears in terms of graphs.
6.4.2 Static output feedback for SISO linear systems Now we consider feeding
back not the state, but the output itself. Thus we consider as feedback for the system the
quantity u(t) = r(t)F y(t) for F R, and where r is the reference signal. This is illustrated
diagrammatically in Figure 6.30.
F
x0

r(s)

(sI n A)1

y(s)

x0
1

22/10/2004

6.51 Definition Let = (A, b, ct , D) be a SISO linear system. An output feedback constant is a number F R with the property that F D 6= 1. To an output feedback number
FD
F
bct , 1 1+F
b, (1 +
F we assign the closed-loop SISO linear system F = (A 1+F
D
D
F D)1 ct , (1 + F D)1 D). The transfer function for F is called the closed-loop transfer
function. A rational function R R(s) is output compatible with is there exists an
output feedback number F with the property that TF = R.

6.52 Remarks Note that static output feedback is somewhat uninteresting for SISO systems.
This is because the feedback parameter is simply a scalar in this case. For MIMO systems,
the feedback is not via a scalar, but by a matrix, so things are more interesting in this case.
Nevertheless, as we shall see in Section 10.2.2, the static output feedback problem is difficult,
even for SISO systems.

6.53 Static output feedback design problem Given the system = (A, b, ct , D), find an output feedback constant F so that
(i) the closed-loop transfer function F has desirable properties and
(ii) the state variables are behaving in a nice fashion.
In particular, one typically want the matrix A F bct to be Hurwitz.

Following our earlier notation, given = (A, b, ct , D), we denote by So () the set of
stabilising output feedback constants. That is,

So () = F R | A F bct is Hurwitz .

6 Interconnections and feedback

Let us say what is the objective with this type of feedback.

r(s)

256

(sI n A)1

/ y(s)

Let us look at the form of rational functions compatible with a system under static
output feedback. This is analogous to Theorem 6.49, although we cannot make a statement
concerning the nature of the stabilising output feedback constants.

Figure 6.30 The static output feedback configuration for the SISO
linear system (6.14) as a block diagram (top) and a signal flow
graph (bottom)

Using the equations (6.14) we may determine the closed-loop equations. It turns out
that we require that F D 6= 1, a condition that is true, for instance, when D = 01 . This
condition is related to the well-posedness condition for input/output systems discussed in
Section 6.3.3. In any event, when the conputations are carried out we get

F
FD
bct x(t) 1
br(t)
1 + FD
1 + FD
1 t
1
y(t) = (1 + F D) c x(t) + (1 + F D) Dr(t),

x(t)
= A

for a reference signal r(t). These expressions simplify somewhat in the usual case when
D = 01 . With this in mind, we make the following definition which is the analogue of
Definition 6.47.

6.54 Theorem Let = (A, b, ct , D) be a SISO linear system with (A, c) observable and
D 6= 01 . A rational function R R(s) is output compatible with if and only if
t
FD
1 1+F
c adj(sI n A)b + DP (s)
D
R(s) =
,
P (s)
where P (s) = PA (s) +

F
ct adj(sI n
1+F D

Proof We without loss

0
1

A = 0
..
.

0
0

A)b and F R.

of generality assume that (A, c) are in observer canonical form:

0 0 0 p0
0 0 0 p1

1 0 0 p2

0 1 0 p3
, ct = 0 0 0 0 1 .
.. .. . . ..
..
. .
. .
.

0 0 0 pn2
0 0 1 pn1

6.4 Feedback for SISO linear systems

b0
b1
b2
..
.

0 0 0 bn1

so that

0
1

A
bct = 0
..
1 + FD
.

0
0

0
0
1
0
..
.

0
0
0
1
..
.

...

0
0
0
0
..
.

p0
p1
p2
p3

F
b
1+F D 0
F
b
1+F D 1
F
b
1+F D 2
F
b
1+F D 3

,0
0

F
bct
1+F D

is PA (s) +

given that A
function is

7.5

12.5

17.5

2.5

20,0

7.5

F
ct adj(sI n
1+F D

A F bct =
A)b since

12.5

17.5

0
1
.
2 2

This is the same closed-loop matrix as obtained in Example 6.50, so the closed-loop state
response will be the same as that shown in Figure 6.28. The closed-loop output response is
now computed to be
y(t) = (2t 1)e2t + cos t,
and this is shown in Figure 6.31. The salient fact, of course, is that where the open-loop
output was unbounded with the given input, the closed-loop output is now bounded.

FD t
F
FD
bct
1
b= 1
c adj(sI n A)b,
1 + FD
1 + FD
1 + FD

F
bct
1+F D

is in observer canonical form. Therefore, the closed-loop transfer

t
FD
c adj(sI n A)b
1 1+F
D
+ D,
TF (s) =
F
PA (s) + 1+F
ct adj(sI n A)b
D

and this is exactly as stated in the theorem.

2.5

Figure 6.31 Output response for open-loop system (left) and

closed-loop system (right) under static output feedback

ct adj(sI n A)b = bn1 sn1 + + b1 s + b0 .

We also have

ct sI n A

22/10/2004

F
0 0 0 pn2 1+F D bn2
F
0 0 1 pn1 1+F
b
D n1

Thus the characteristic polynomial of A

hN,D (t)
1 (t)
1N,D (t)
(t)
(t)
fj (t)
Re
Interconnections Im
and feedback
x1
x2
x1
1.5
x2
log
1
dB
deg
0.5
u = 0 ln
ln coth(|u| /2)
0
or 1
m
-0.5
yos
-1

tos 0

y(t)

Let us write b = (b0 , b1 , . . . , bn1 ). We then have

0 0 0
0 0 0

bct = 0 0 0
.. .. .. . .
. . .
.

257

y(t)

22/10/2004

hN,D (t)
1 (t)
1N,D (t)
(t)
(t)
fj (t)
Re
Im
258
x1
x2
x1
x2
40
log
dB
20
deg
u = 0 ln
ln coth(|u| /2)
0
or 1
m
-20
yos

tos 0 -40

Unsurprisingly, static output feedback for SISO systems does not give the same freedom
for pole placement as does static state feedback. Nevertheless, it is possible to have a positive
effect on a systems behaviour by employing static output feedback, as is indicated by the
following cooked example.
6.55 Example (Example 6.50 contd) We consider the SISO system = (A, b, ct , D) where

0 1
0
3
A=
, b=
, c=
, D = 01 .
1 0
1
4
Thus A and b are just as in Example 6.50. Therefore, if we give the open-loop system the
same input u(t) = 1(t) cos t with the same state initial condition x(0) = 0, the open-loop
state evolution will be the same as that in Figure 6.28. However, the output vector c we
now use differs from that of Example 6.50. The open-loop output is
2 sin t + 2t cos t + 32 t sin t
and shown in Figure 6.31. Let us use static output feedback with F = 1. We then compute

It is also not hard to come up with examples where static output feedback is not capable
of providing stable closed-loop behaviour (see Examples E6.17 and E6.18). The matter of
static output feedback is also attended to in Section 10.2.2, where methods of constructing
such feedback laws are discussed.
6.4.3 Dynamic output feedback for SISO linear systems Thus far, for SISO linear
systems, we have investigated only static feedback. This feedback is static because we have
not use any derivatives of the quantity being fed back. Now let us consider introducing
dynamics into the mix. The objective is not just to feedback the output, but also maybe
derivatives of the output, and to maybe have the fed back quantity also depend on the input.
Thus we start schematically with a block diagram as depicted in Figure 6.32. In the diagram,
P

r(t)

y(t)

controller

Figure 6.32 A proposed feedback loop for dynamic output feedback

P is an abbreviation for the block diagram for a SISO linear plant P = (AP , bP , ctP , D P )
(i.e., for a block diagram like Figure 3.6). Let us address the question of what lies within the

6.4 Feedback for SISO linear systems

22/10/2004

259

block labelled controller. For dynamic output feedback, in this block sits C , a controller
SISO system C = (AC , bC , ctC , D C ). Now note that the diagram of Figure 6.32 looks
schematically just like the bottom input/output feedback loop in Figure 6.22. Thus, in
being consistent with our discussion of Section 6.3, let us agree to consider a block diagram
schematic like Figure 6.33 to model dynamic output feedback. Therefore, dynamic output
u(t)
C

r(t)

y(t)

Figure 6.33 Schematic for dynamic output feedback used in text

feedback consists of connection two SISO linear systems. Note, however, that we have
come full circle back to the situation in Section 6.3 where we talked about feedback for
input/output systems. Indeed, one can view the designing of a controller rational function
RC as being equivalent to specifying its (say) canonical minimal realisation C .
Let us get a little more mathematical and write the interconnection of Figure 6.33 is
differential equation form. Let us denote the states for the plant by xP and the states for
the controller by xC . We let u(t) be the input to P which is also the output from C . We
then have
x P (t) = AP xP (t) + bP u(t)
x C (t) = AC xC (t) + bC (r(t) y(t))
y(t) = ctP xP (t) + D P u(t)
u(t) = ctC xC (t) + D C (r(t) y(t)).
To obtain the closed-loop system we use the last two equations to solve for y and u in
terms of the other variables. The equations to be solved are

ctP xP (t)
1 D P y(t)
.
= t
cC xC (t) + D C r(t)
DC
1
u(t)
We see that in order to solve this equation for y(t) and u(t) we must have 1 + D C D P 6= 0.
If this condition is satisfied, we say the interconnection is well-posed . This, it turns out,
is exactly the same as the definition of well-posed made in Section 6.3.3 (see Exercise E6.9).
After eliminating u, we are left with the input r(t), the output y(t), and the state (xP , xC ).
The following result says that the resulting equations are those for a SISO linear system, and
gives the form of the system. The proof is a direct calculation following the outline above.
6.56 Proposition Suppose that xP Rn and xC Rm . The closed-loop system for Figure 6.33 is a SISO linear system cl = (Acl , bcl , ctcl , D cl ) where

t

AP 0n,m
b 0m
1
DC
0n ctC
Acl =
+ (1 + D C D P )1 P
,
t
0m,n AC
0n bC D P
1
c 0m

P
b
0n
1
DC DC
0
bcl = (1 + D C D P )1 P
+ n ,
0m bC D P
1
0
bC

ctcl = (1 + D C D P )1 ctP D P ctC ,
D cl = (1 + D C D P )1 D C D P .

260

6 Interconnections and feedback

22/10/2004

Note that this is the content of Exercise E2.3, except that in that exercise no feedforward
terms were included.
Now that we have the closed-loop system on hand, we may state a problem one often
wishes to resolve by the use of dynamic output feedback.
6.57 Dynamic output feedback design problem Given a plant SISO linear system P , find a
controller SISO linear system C so that the closed-loop system is internally asymptotically
stable and well-posed.

As our final symbolic representation for stabilising controllers, given P =
(AP , bP , ctP , D P ), let us denote by S (P ) the set of SISO linear systems C =
(AC , bC , ctC , D C ) for which the closed-loop system cl is internally asymptotically stable.
The set S (P ) is closely related to the set of stabilising controller rational functions, S (RP ),
if RP = TP . In fact, the only essential difference is that all controllers C S (P ) will
give rise to proper controller transfer functions TC .
6.58 Remarks 1. Note that if the closed-loop system is internally asymptotically stable then
it is IBIBO stable. This may not appear obvious, but it actually is. Because all the
possible inputs and outputs in the system will simply be linear combinations of the
states xP and xC , it follows that if the states behave stably, then so too will all the
inputs and outputs.
2. As we mentioned above, the dynamic output feedback control problem, Problem 6.57,
and the input/output control problem, Problem 6.41, are very closely related. In the
input/output control problem, we were a little more vague, and asked for the closed-loop
transfer function to behave in a suitable manner. For the dynamic output feedback
control problem, we were a little more specific because we could be. Nevertheless, even
for the dynamic output feedback problem, one often simply wants internal asymptotic
stability as a matter of course, and additional requirements will be imposed additionally.
3. Related to the question we asked at the end of Section 6.3.1 is the question of whether
given a plant SISO linear system P , it is alway possible to come up with a controller SISO
linear system C so that the closed-loop system is internally asymptotically stable. This
question is answered in the affirmative in Section 10.2.3, at least under mild assumptions.
For example, if P is controllable and observable, this is possible. However, unlike the
analogous situation for input/output systems, it is possible for S (P ) to be empty.
Let us give an example of how dynamic output feedback can be used to do good. As
with all the controllers we have designed thus far, we simply give an ad hoc controller that
does the job. The matter of coming up with these in a systematic manner is something that
we get into in detail in subsequent chapters.
6.59 Example We look at the problem of a mass with no gravitational effects with the control
being a force applied to the mass. The differential equation is thus
m
x = u.
As output, let us use the position x(t) of the mass. Putting this into the form P =
(AP , bP , ctP , D P ) gives

1
0 1
0
, D P = 01 .
AP =
, bP = 1 , cP =
0
0 0
m

22/10/2004

6.5 The PID control law

261

262

6 Interconnections and feedback

Let us (magically) choose a controller SISO linear system C = (AC , bC , ctC , D C ) given by

2
2m
2 1
PSfrag
PSfrag replacements
AC =
, breplacements
, cC =
, D C = 01 .
C =
4 2
2
2m
ke(t)k
Aee1 (t)
tedious
2 (t)
e3 (t)
y(t)
u(t)
h (t)
hN,D (t)

ke(t)k

e1 (t)
calculation using Proposition 6.56
then gives
e2 (t)
e3 (t)

0 y(t)1 0
0
0 u(t)0 2
2
h (t)
.
Acl =
2
hN,D (t)
0 2 1
2
0 4 2
1N,D (t)

e(s)

u
(s)

q8 MMMM
MMM1
qqq
MMM
q
MMM
1
qqq
q
M&
q
T
s
I
/ Mq
8/
MMM
qqq
MMM
q
q
q
MMM
q
qqq1
TD s MMM
M& qqqq

1 qqq

e(s)

1 (t)

1
TI s

TD s

1N,D (t)
(t)
(t)
(t) checks that the eigenvalues of A are (t)
One
{1 + i, 1 + i, 1 i, 1 i}. Thus Acl is
cl
fj (t)
fj (t)
Hurwitz
as desired. This calculation is explained
in Example 10.30. The step response for
Re
Re
Im step response for the closed-loop system.
theIm
x1
x1plant is shown in Figure 6.34, alongside the
x2
x2
x1
x1 50
1.5
x2
x2
log
log
40
dB 1.25
dB
deg
deg
1
u = 0 ln
u = 0 ln 30
ln coth(|u| /2)
ln coth(|u| /2)
or 1 0.75
or 1 20
m
m
0.5

yos
yos 10
0.25

tos 0
tos 0
0
0

2
4
6
8
10
2
4
6
8
10

,0
,0
t
t
0
0

/u
(s)

Figure 6.35 The block diagram (top) and signal flow graph for a
PID controller

This can cause problems, and sometimes one considers the form

TD s
1
RC (s) = K 1 +
+
D s + 1 TI s

Figure 6.34 Step response for unit mass before (left) and after
(right) using dynamic output feedback

6.5 The PID control law

Minorsky, 1922

The so-called PID, for P roportional-I ntegral-Derivative, feedback is very popular, mainly
because it is simple and intuitive, and very often quite effective. The PID controller is
intended to apply to systems in input/output form, and so if one is dealing with a SISO linear
system , one needs to be aware that the PID controller only knows about the input/output
behaviour, and not the state behaviour. The idea is that one designs a controller that
provides an input to the plant based upon a sum of three terms, one of which is proportional
to the error, one of which is proportional to the time-derivative of the error, and the other
of which is proportional to the integral of the error with respect to time. In this section we
investigate each of these terms, and how they contribute to the controllers performance, and
why one should exercise some caution is choosing gains for the PID controller. Knowing what
differentiation and integration look like in the Laplace transform domain, we may represent
the controller transfer function for a PID control law as in Figure 6.35. The transfer function
for this controller is

1
.
RC (s) = K 1 + TD s +
TI s
The constant K is the gain, and TD and TI are the derivative time and reset time,
respectively. Note that the transfer function for the term TD s in the controller is not proper.

22/10/2004

for the PID controller, ensuring that all terms in the controller are proper. One can think
of the additional factor (D s + 1)1 as being a low-pass filter added to the second term.
In classical PID design, one makes D small compared to TD , thus minimising its effects.
However, more modern practise allows D to be set as a design parameter. In this chapter,
however, we shall always take D = 0, and deal with a straight, classical PID controller. In
the design of such controllers, one wishes to determine these constants to accomplish certain
objectives.
In investigating the PID controller types, we utilise a specific example. In general, the
behaviour of a given controller transfer function will depend to a very large extent upon the
nature of the plant being controlled. However, by looking at an example, we hope that we
can capture some of the essential features of each of P, I, and D. The example we use
will have a plant transfer function
RP (s) =

1
s2 + 3s + 2

The closed-loop transfer function is then

1
KRC (s) s2 +3s+2
1
1 + KRC (s) s2 +3s+2

(6.15)

and we consider how choosing a certain type of controller rational function RC and fiddling
with the gain K effects the closed-loop response.

1 (t)
1N,D (t)
(t)
(t)
fj (t)

263

PSfrag replacements

t
6.5.1 Proportional
control For proportional control we take RC (s) = 1. We compute
ke(t)k
the closed-loope1transfer
function to be
(t)
e2 (t)
K
e3 (t)
2 + 3s + 2 + K
s
y(t)
u(t)
and so the characteristic
polynomial is s2 + 3s + 2 + K. The roots of the characteristic
h (t)
polynomial are
the poles of the closed-loop transfer
function, and so these are of great

hN,D (t)
3
1
interest to us.
1We
(t) compute them to be 2 2 1 4K. In Figure 6.36 we plot the set of
1N,D (t)
1.5
(t)
(t)
fj (t)
1

,0
0

6 Interconnections and feedback

22/10/2004
x1
x2
The roots are always
real for K > 0, and the locus of roots is shown in Figure 6.37 as K
x1
x2
1
log
dB 0.75
deg
0.5
u = 0 ln
ln coth(|u| /2) 0.25
or 1
0
X
X
O
m
-0.25

yos
-0.5

-0.75
tos 0

-2.5
-2
-1.5
-1
-0.5
0
0.5
,0
Re
0
Figure 6.37 The locus of roots for (6.15) with derivative control

x1
x2
0.5
x1
x2
log
0
dB
deg
u = 0 ln
ln coth(|u| /2) -0.5
or 1
m

-1
yos

tos 0

264

6.5 The PID control law

22/10/2004

varies and TD is fixed to be 1. As we increase K, one root of the characteristic polynomial

gets closer and closer to zero. This suggests that for large derivative time, the response
will be slow. Derivative control also can suffer from being difficult to implement because
accurately measuring velocity is sometimes a difficult task. Furthermore, if used alone,
derivative control is incapable of determining steady-state error.
6.5.3 Integral control We finally look at integral control where we take RC (s) =
where the constant TI is called the reset time. The closed-loop transfer function is

-2.5

-2

-1.5

-1

-0.5

Figure 6.36 The locus of roots for (6.15) with proportional control

roots as K varies. Note that the roots are imaginary when K > 14 . In Figure 6.36, K = 14
thus corresponds to where the branches meet at 23 + i0.
Observe that by making K large we end up in a situation where the damping remains
the same as it was for K small (i.e., the value of the real part when the poles are complex
is always 23 ), but the frequency of the oscillatory component increases. This could be a
rather destructive effect in many systems, and indicates that the system response can suffer
when the proportional gain is too large. This is especially a problem for higher-order plants.
Proportional control can also suffer from giving a nonzero steady-state error.
6.5.2 Derivative control Here we take RC (s) = TD s for a constant TD called the
derivative time. This transfer function clearly corresponds to differentiating the error
signal. The closed-loop transfer function is
Ks
s2 + (3 + KTD )s + 2
which yields the roots of the characteristic polynomial as
p
1 + 6KTD + (KTD )2
3 KTD

.
2
2

1
TI s

K
.
TI s3 + 3TI s2 + 2TI s + K
The roots of this equation are too complicated to represent conveniently in closed form.
However, it is still possible to numerically plot the locus of roots, and we do this in Figure 6.38. We have fixed in this plot TI = 1 and are varying K. Note that as we increase the
gain K, the roots of the characteristic polynomial become oscillatory with decreasing real
part. Thus integral control by itself will tend to lead to poor behaviour. But it does have
the advantage of always giving zero steady-state error to certain common input types.
6.5.4 Characteristics of the PID control law We will discuss more in Chapter 8
the types of behaviour we are looking for in a good controller, but we summarise some of
the features of the various elements of a PID controller in Table 6.1. Often, while each
component of the PID controller may by itself not have completely desirable properties,
one can obtain satisfactory results by combining them appropriately. The table should be
regarded as providing some information on what to look for to better tune a PID controller.
Lets provide a simple example of a PID control application.
6.60 Example We consider a mass m falling under the influence of gravity as in Figure 6.39.
At time t = 0 the mass is at a height y = y0 and moving with velocity y(0)

= v0 . The mass
has attached to it a fan that is able to provide an upward force u. The differential equations
governing the system are
m
y (t) = mg + u.

e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)
h (t)
22/10/2004 hN,D (t)
1 (t)
1N,D (t)
(t)
(t)
fj (t)

6.5 The PID control law

265

266

6 Interconnections and feedback

22/10/2004

2
1.5

,0
0

-2.5

-2

-1.5

-1

PSfrag replacements

-0.5

Figure 6.39 Falling mass

0.5

Figure 6.38 The locus of roots for (6.15) with integral control

Table 6.1 Features of PID control

Type

Proportional

Derivative

Integral

Advantages

Disadvantages

1. Fast response

1. Potentially unstable for

large gains and higher-order
plants
2. Possible steady-state error for certain types of input

1. Good stability properties

1. Difficult to implement
2. Does not correct steadystate errors
3. Magnifies high-frequency
noise (see Exercise E6.20)

1. Corrects steady-state er- 1. Large gain leads to lightly

ror
damped oscillatory response

We take y(t) as our output. Note that this system is not quite of the type we have been
considering because of the presence of the gravitational force. For the moment, therefore,
lets suppose that g = 0, or equivalently that the motion of the mass is taking place in a
direction orthogonal to the direction of gravity. We will reintroduce the gravitational force
in Section 8.4. At time t = 0 we assume the mass to have state y(0) = 1 and y(0)

= 1, and it

ke(t)k
e1 (t)
e2 (t)
is our goal to make
e3 (t) it move to the height y = 0. Thus we take the reference signal r(t) = 0.
We shall investigate the effects of using proportional, derivative, and integral control.
u(t)
First we look
at making the applied force u proportional to the error. Thus the force we
(t) u = Ky. The differential equation is then
apply has thehform
hN,D (t)
1 (t)
m
y (t) + Ky(t) = 0, y(0) = 1, y(0)

=1
1N,D (t)
(t)
which has thesolution
q
q
q
(t)
m
t+ K
sin K
t.
y(t) = cos K
fj (t)
m
m
Re
In this case the Im
mass simply oscillates forever, only returning to its desired position periodx1 is plotted in Figure 6.40 for m = 1 and K = 28. Note that when we
ically. This output
x2
x1
1.5
x2
log
1
dB
deg 0.5
u = 0 ln
ln coth(|u| /2)
0
or 1
m
-0.5
yos
-1

tos 0

1
2
3
4
5
,0
t
0
y(t)

x1
x2
x1
0.5
x2
log
0
dB
deg
u = 0 ln -0.5
ln coth(|u| /2)
or 1
-1
m

yos -1.5

tos 0

Figure 6.40 Falling mass under proportional control

increase K this has the effect of decreasing the magnitude of the oscillations while increasing
their frequency.
Now we examine the situation when we have derivative control. Thus we take u =
KTD y.
The differential equation is
m
y (t) + KTD y(t)
= 0,
which has the solution
y(t) = 1 +

y(0) = 1, y(0)

m
1 e(KTD /m)t .
KTD

,0
0

6.5 The PID control law

267

1.4
1.2
1

22/10/2004

y(t)

0.8

1N,D (t)
(t)
(t)
fj (t)
Re
Im
x1
268
6 Interconnections and feedback
x2
x1
PSfrag replacements
1500
x2
log
ke(t)k
1000
e1dB
(t)
e2deg
(t)
500
u = 0eln

3 (t)
ln coth(|u| /2)
0
or u(t)
1
h(t)
m
-500

hN,D (t)
os
1y(t)
-1000
1N,D (t)
t(t)
os 0

,
(t)
1
2
3
4
0
t
fj(t)
0
Re
Im
Figure 6.42 Falling mass under integral control
x1
x2
x1
1.25
x2
log
1
dB
deg 0.75
u = 0 ln
0.5
ln coth(|u| /2)
or 1
0.25
m

0
yos
-0.25
tos 0

1
2
3
4
,0
t
0

0.6
0.4
0.2
1

Figure 6.41 Falling mass under derivative control

9
. Note here that the
This output is shown in Figure 6.41 for K = 28, m = 1, and TD = 28
mass ends up maintaining a steady height that is not the desired height. This is the problem
with derivative controlonce there is no velocity, the controller Rstops reacting.
t
We can also consider integral control. Here we take u = TKI 0 y( ) d . The differential
equation is thus
Z
K t
m
y (t) +
y( ) d = 0, y(0) = 1, y(0)

= 1.
(6.16)
TI 0

To integrate this equation it is most convenient to differentiate it once to get

... K
my + y(t) = 0,
TI

y(0) = 1, y(0)

= 1, y(0) = 0.

The initial condition y(0) = 0 comes to us from evaluating the equation (6.16) at t = 0. This
third-order equation may be explicitly solved, but the resulting expression is not particularly
worth recording. However, the equation may be numerically integrated and the resulting
7
9
, and TI = 10
. The behaviour
output is plotted in Figure 6.42 for K = 28, m = 1, TD = 28
here is oscillatory and unstable.
While none of the controllers individually performed in a reasonable manner, when we
combine them, we can get satisfactory performance. When we combine the three controllers,
the differential equation is
Z
K t
m
y + KTD y(t)
+ Ky(t) +
y( ) d = 0, y(0) = 1, y(0)

= 1.
TI 0
Again, to get rid of the integral sign, we differentiate to get
K
...
my + KTD y(t) + K y(t)
+ y(t) = 0,
TI

y(0) = 1, y(0)

= 1, y(0) = K(TD + 1).

This equation may in principle be solved explicitly, but we shall just numerically integrate
and show the results in Figure 6.43 for the same parameter values as were used in the plots
for the individual controllers. Note that the controller is behaving quite nicely, bringing the
mass to the desired height in a reasonable time.

y(t)

1N,D (t)
(t)
(t)
fj (t)
Re
Im
22/10/2004
x1
x2
x1
x2
log
dB
deg
u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

Figure 6.43 Falling mass under combined PID control

The parameters in the PID controller were not chosen completely by trial and error
9
hereI dont expect the number 28
often gets generated by trial and error. In Chapter 12
we will see how to do select controller parameters in a more systematic manner. That we
should expect to be able to do what we wish can be seen by the following result.
6.61 Proposition For the falling mass under PID control, the poles of the closed-loop system
can be chosen to duplicate the roots of any cubic polynomial except those of the form s3 +
as2 + bs + c where b 6= 0. If the closed-loop polynomial has the coefficient b = 0 then it must
also be the case that a = c = 0.
Proof The block diagram for the system with the PID feedback is shown in Figure 6.44.
The closed-loop transfer function is therefore

6.6 Summary

22/10/2004

r(s)

K + KTD s +

269

1
ms2

K
TI s

1
K + KTD s + TKI s
y(s)
ms2

=
r(s)
1 + ms1 2 K + KTD s + TKI s
s2 + T1D s + TD1TI
KTD
3
m s +K
T s2 + K
s+
m D
m

K
mTI

We wish to show that appropriate choices for K, TD , and TI can be made so that
s3 +

K
K
K
TD s2 + s +
= s3 + as2 + bs + c
m
m
mTI

for arbitrary a, b, c R. If b 6= 0 then

a
TD = ,
b

b
TI = ,
c

K = bm

accomplishes the task. If b = 0 case that K = 0 and so a and c must also be zero.

6 Interconnections and feedback

22/10/2004

5. Although we have not said much about controller design, one should understand the
issues surrounding the design problem for both SISO linear systems in input/output
form and for SISO linear systems.
6. Much of what we say as we go on will apply to the simple unity gain feedback loop. This
is simple, and one ought to be able to work with these fluently.
7. You should know what the PID control law is, and have some feel for how its proportional,
derivative, and integral components affect the performance of a system.

y(s)

Figure 6.44 Block diagram for falling mass with PID control

270

Thus we can place poles for the closed-loop transfer function pretty much anywhere we wish.
I used the proposition to select poles at {5, 2 2i}.
And with that we leave the wee falling mass to its own devices.

6.6 Summary
Of course, the notion of feedback is an important one in control. In this section, we have
discussed some of the issues surrounding feedback in a fairly general way. In later chapters,
we will be deciding how to deploy feedback in an effective manner. Let us do as we have
been doing, and list some of the more essential points raised in this chapter.
1. The first few sections of the chapter dealt with the setup surrounding interconnected
SISO linear systems. The generality here is somewhat extreme, but it serves to properly
illustrate the problems that can arise when interconnecting systems.
2. You might find it helpful to be able to switch freely from block diagrams to signal flow
graphs. Sometimes the latter are the more useful form, although the former are more
commonly used in practice.
3. One should know immediately how to compute the determinant and characteristic polynomial for simple block diagram configurations. One should also be able to use the
signal flow graph technology to determine the transfer function between various inputs
and outputs in a block diagram configuration.
4. One should be able to test a block diagram configuration for IBIBO stability.

Exercises for Chapter 6

271

272

Exercises
E6.1 Let (S, G) be a signal flow graph with GS,G its corresponding matrix. Show that the
gains in the ith column correspond to branches that originate from the ith node, and
that the gains in the jth row correspond to branches that terminate at the jth node.
E6.2 Consider the block diagram of Exercise E3.2.
(a) Draw the corresponding signal flow graph.
(b) Write the signal flow graph as (S, G) as we describe in the textthus you should
identify the nodes and how the nodes are connected.
(c) Write the structure matrix GS,G .
(d) Determine the pair (AS,G , B S,G ) as per Procedure 6.22.
(e) Write all simple paths through the signal flow graph which connect the input r
with the output y.
(f) Identify all loops in the graph by writing their gainsthat is, determine
Loop(S, G).
(g) For k 1 determine Loopk (S, G).
(h) Find S,G .
(i) Find TS,G .
(j) Write each of the rational functions R1 , . . . , R6 in Figure E3.1 as a numerator
polynomial over a denominator polynomial and then determine PS,G .
E6.3 Let = (A, b, ct , D) be a SISO linear system.

(a) Show that there exists P R[s]nn so that x(t)

= Ax(t) if and only if
d
x(t) = 0.
P dt
(b) Show that Theorem 5.2 then follows from Theorem 6.33. That is, show that the
hypotheses of Theorem 5.2 imply the hypotheses of Theorem 6.33.
E6.4 For the block diagram of Figure E6.1, determine the values of a, b, and c for which
r(s)

1
sb

1
(s + a)(s + c)

y(s)

s+a

6 Interconnections and feedback

Answer the following questions.

(a) Use Theorem 6.38 and the Routh/Hurwitz criteria to ascertain for which values
of a and b the interconnected system is IBIBO stable.
(b) For which values of a and b can IBIBO stability of the interconnected system
be inferred from looking only at the determinant of the graph without having to
resort to using the characteristic polynomial?
E6.6 In this exercise, you will investigate the matter of relating BIBO stability of individual
branch gains in a signal flow graph (S, G) to the IBIBO stability of the interconnection.

finish

(a) Is it possible for (S, G) to be IBIBO stable, and yet have branch gains that are
not BIBO stable? If it is not possible, explain why not. If it is possible, give an
example.
(b) Is it true that BIBO stability of all branch gains for (S, G) implies IBIBO stability of the interconnection? If it is true, prove it. If it is not true, give a
counterexample.
E6.7 Prove Proposition 6.45 directly, without reference to Theorem 6.38.
E6.8 Well-posedness and existence and uniqueness of solutions.
E6.9 For the feedback interconnection of Figure E6.2, let C = (A1 , b1 , ct1 , D 1 ) be the
r(s)

RC (s)

the interconnected system is IBIBO stable. For which values of a and b can stability
be inferred from the zeros of the determinant, without having to resort to looking at
the characteristic polynomial.
E6.5 For the block diagram of Exercise E3.2 (which you investigated as a signal flow graph
in Exercise E6.2), consider the following assignment of specific rational functions to
the block transfer functions:
1
1
R1 (s) = , R2 (s) = s 1, R3 (s) =
,
s
s+b
1
1
R4 (s) =
, R5 (2) =
, R6 (s) = s + a.
s+2
s+a

RP (s)

y(s)

Figure E6.2 Plant/controller feedback loop

canonical minimal realisation for RC and let P = (A2 , b2 , ct2 , D 2 ) be the canonical
minimal realisation for RP . Show that the interconnection is well-posed if and only if
D 1 D 2 6= [1].
E6.10 For the block diagram configuration of Figure E6.3, show that as K 0 the poles
r(s)

Figure E6.1 A block diagram for determining IBIBO stability

22/10/2004

RL (s)

y(s)

Figure E6.3 Unity gain feedback loop with variable gain

and zeros of the closed-loop transfer function approach those of RL .

E6.11 Let RP R(s) be a proper plant. For RC R(s) define RL = RC RP , as usual, and
let TL be the corresponding closed-loop transfer function. Suppose that RP has zeros
z1 , . . . , z` C+ {} and poles p1 , . . . , pk C+ (there may be other poles and zeros
of the plant, but we do not care about these). Prove the following result.

Exercises for Chapter 6

273

Proposition RC S (RP ) if and only if the following four statements hold:

(i) TL RH+
;
(ii) the zeros of 1 TL contain {p1 , . . . , pk }, including multiplicities;
(iii) the zeros of TL contain {z1 , . . . , z` }, including multiplicities;
(iv) lims RL (s) 6= 1.
E6.12 Let RP R(s) be a proper plant. For RC R(s) define RL = RC RP , as usual,
and let SL be the corresponding sensitivity function. Suppose that RP has zeros
z1 , . . . , z` C+ {} and poles p1 , . . . , pk C+ (there may be other poles and zeros
of the plant, but we do not care about these). Prove the following result.
Proposition RC S (RP ) if and only if the following four statements hold:
(i) SL RH+
;
(ii) the zeros of SL contain {p1 , . . . , pk }, including multiplicities;
(iii) the zeros of 1 SL contain {z1 , . . . , z` }, including multiplicities;
(iv) lims RL (s) 6= 1.
E6.13 Let RP R(s) be a BIBO stable plant. Show that there exists a controller RC R(s)
for which the interconnection of Figure E6.2 is IBIBO stable with closed-loop transfer
function TL if and only if TL , RTLP RH+
.
E6.14 Let RP R(s) be a proper plant transfer function, and let (S, G) be an interconnected
SISO linear system with the property that every forward path from the input to the
output passes through the plant. Show that IBIBO stability of (S, G) implies that
T
TS,G , RS,G
RH+
.
P
E6.15 In this exercise you will show that by feedback it is possible to move into C the
poles of a closed-loop transfer function, even when the poles of the plant are in C+ .
Consider the closed-loop system as depicted in Figure 6.21 with
RC (s) = 1,

RP (s) =

1
,
(s + 1)(s a)

with a > 0. Determine for which values of the gain K the closed-loop system has all
poles in C . Is the system IBIBO stable when all poles of the closed-loop system are
in C ?
In this exercise we explore the relationship between performing static state feedback for
SISO linear systems, and performing design for controller rational functions for input/output
systems.
E6.16 Consider a SISO linear system = (A, b, ct , D) with (A, b) controllable and (A, c)
observable. Let RP = T . For f = (f0 , f1 , . . . , fn1 ) Rn define the polynomial
F (s) = fn1 sn1 + + f1 s + f0 R[s].
Suppose that (A bf t , c) is observable.
(a) Show that there exists a controller rational function RC with the property that
the poles of the two transfer functions
Tf ,

RC RP
T =
1 + RC RP

274

6 Interconnections and feedback

22/10/2004

agree if and only if the polynomial NP divides the polynomial F over R[s]. In
particular, show that if NP is a constant polynomial, then it is always possible
to find a controller rational function RC with the property that the poles of Tf
and T agree.
Hint: Without loss of generality suppose that (A, b) is in controller canonical
form, and look at the proof of Proposition 10.13.
(b) Suppose that NP divides F over R[s] and by part (a) choose a controller rational
function RC with the property that the poles of Tf and T agree. What is the
difference of the numerators polynomials for Tf and T .
Thus the problem of placement of poles in feedback design for SISO linear systems
can sometimes be realised as feedback design for input/output systems. The following parts of the problem show that there are some important cases where controller
rational function design cannot be realised as design of a state feedback vector.
Let (N, D) be a strictly proper SISO linear system in input/output form with
= (A, b, ct , 01 ) the canonical minimal realisation. Suppose that N has no root at
s = 0. Let RP = TN,D .
(c) Is it possible, if RC is the controller rational function for a PID controller, to find
f Rn so that the poles of the transfer functions of part (a) agree?
E6.17 Consider the SISO linear system = (A, b, ct , 01 ) with

0 1
0
1
A=
, b=
, c=
.
0 0
1
0
For this system, answer the following.
(a) Show that there is no continuous function u(x1 ) with the property that for every
solution x(t) of the differential equation

x(t)
= Ax(t) + bu(x1 (t))

(E6.1)

satisfies limt kxk (t) = 0.

Hint: First prove that the function
1
V(x) = x22
2

u() d
0

is constant along solutions of the differential equation (E6.1).

(b) If f R2 is a state feedback vector for which A bf t is Hurwitz, what can be
said about the form of f from part (a).
E6.18 In this exercise we generalise Exercise E6.17 for linear feedback. We let =
(A, b, ct , 01 ) be a SISO linear system with tr(A) = 0 and ct b = 0.
(a) Show that there is no output feedback number F with the property that the
closed-loop system is internally asymptotically stable.
Hint: Show that if
PA (s) = sn + pn1 sn1 + + p1 s + p0
is the characteristic polynomial of A, then p0 = tr(A) (think of putting the matrix in complex Jordan canonical form and recall that trace is invariant under
similarity transformations).

Exercises for Chapter 6

275

(b) Something with Liapunov

E6.19 Let = (A, b, ct , 01 ) be a complete SISO linear system and suppose that z C
is a zero of T . Show that by static output feedback it is not possible to obtain a
closed-loop system with a pole at z.
One of the potential problems with derivative control is that differentiation can magnify
high frequency noise, as the following exercise points out.
E6.20 For a time signal
y(t) = As sin(s t) + An sin(n t + n ),
consisting of a sinusoidal signal (the first term) along with sinusoidal noise (the second
|As |
.
term), the signal-to-noise ratio is defined by S/N = |A
n|
(a) Show that for any such signal, the signal-to-noise ratio for y tends to zero as n
tends to infinity.
(b) Indicate in terms of Bode plots why differentiation is bad in terms of amplifying
high frequency noise.
E6.21 In this exercise we will investigate in detail the DC servo motor example that was
used in Section 1.2 to provide an illustration of some control concepts.
We begin by making sure we know how to put the model in a form we can deal
with. We model the system as a SISO linear system whose single state is the angular
velocity of the motor. The output is the angular velocity of the motor (i.e., the value
of the systems only state), and the input is the voltage to the motor.
(a) Determine = (A, b, ct , D). Your model should incorporate the time-constant
and motor gain kE as in Section 1.2, but do not include any effects from external
disturbances.
(b) Determine T .
We let the plant transfer function RP be T whose c.f.r. we write as (NP , DP ). For
the reasons we discussed in Section 1.2, an open-loop control scheme, while fine in an
idealised environment, lacks robustness. Thus we employ a closed-loop control scheme
like that depicted in Figure 6.21. For proportional control we use RC (s) = 1.
(c) Determine the closed-loop transfer function with gain K.
(d) Assuming that and kE are both positive, determine the range of gains K
for which the closed-loop system has all poles in C . That is, determine
Kmin and Kmax so that the closed-loop system has poles in C if and only if
K (Kmin , Kmax ).
Now we will see how the closed-loop systems frequency response represents its ability
to track sinusoidal inputs.
(e) Determine the frequency response for the closed-loop transfer function.
(f) Determine the output response to system when the reference signal is r(t) = cos t
for some > 0. Assuming that K (Kmin , Kmax ), what is the steady-state
response, yss (t).
(g) Show that limK yss (t) = cos t, and so as we boost the gain higher, we can in
principle exactly track a sinusoidal reference signal. Can you see this behaviour
reflected in the frequency response of the system?
E6.22 Consider the coupled masses of Exercise E1.4 (assume no friction). As input take the
situation in Exercise E2.19 with = 0. Thus the input is a force applied only to the
leftmost mass.

276

6 Interconnections and feedback

22/10/2004

We wish to investigate the effect of choosing an output on our ability to manipulate

the poles of the closed-loop transfer function. We first consider the case when the
output is the displacement of the rightmost mass.
(a) Determine the transfer function T for the system with this output.
(b) What are the poles for the transfer function? What does this imply about the
uncontrolled behaviour of the coupled mass system?
First we look at an open-loop controller as represented by Figure 6.20. We seek a
controller rational function RC whose c.f.r. we denote (NC , DC ). We also let RP = T ,
and denote by (NP , DP ) its c.f.r.
(c) Can you find a controller transfer function RC so that
1. DP and NC are coprime and
2. the open-loop transfer function has all poles in C ?
Why do we impose the condition 1?
Now we look for a closed-loop controller as represented by Figure 6.21. For simplicity,
we begin using a proportional control.
(d) For proportional control, suppose that RC (s) = 1, and derive the closed-loop
transfer function with gain K.
(e) Show that it is impossible to design a proportional control law for the system
with the properties
1. DP and NC are coprime and
2. the closed-loop transfer function has all poles in C ?
Hint: Show that for a polynomial s4 + as2 + b R[s], if s0 = 0 + i0 is a root,
then so are 0 i0 , 0 + i0 , and 0 i0 .
It turns out, in fact, that introducing proportional and/or derivative control into the
problem described to this point does not help. The difficulty is with our plant transfer
function. To change it around, we change what we measure.
Thus, for the remainder of the problem, suppose that the output is the velocity of
the leftmost mass (make sure you use the correct output).
(f) Determine the transfer function T for the system with this output.
(g) What are the poles for the transfer function?
The open-loop control problem here is the same as for the previous case where the
output was displacement of the rightmost mass. So now we look for a closed-loop
proportional controller for this transfer function.
(h) For proportional control, suppose that RC (s) = 1, and derive the closed-loop
transfer function with gain K.
(i) Choose m = 1 and k = 1, and show numerically that there exists K > 0 so that
the poles of the closed-loop transfer function all lie in C .
E6.23 Refer to Exercise E6.21. Set the time constant = 1 and the motor constant kE = 1.
Produce the Bode plots for plant transfer function, and for the closed-loop system
with the proportional controller with gains K = {1, 10, 100}. Describe the essential
differences in the Bode plots. How is your discovery of Exercise E6.21(g) reflected in
your Bode plots?
E6.24 Refer to Exercise E6.22, taking m = 1 and k = 1, and use the second output
(i.e., the velocity of the leftmost mass). Produce the Bode plots for plant transfer
function, and for the closed-loop system with the proportional controller with gains

Exercises for Chapter 6

277

K = {1, 10, 100}. Describe the essential differences in the Bode plots. In what way
does the open-loop transfer function differ from the rest?

E6.25 Consider the controller transfer function RC (s) = K 1 + TD s + T1I s .
(a) Can you find a SISO linear system = (A, b, ct , D) so that T = RC ? (Assume
that K, TD , and TI are finite and nonzero.)
(b) What does this tell you about the nature of the relationship between the Problems 6.41 and 6.57?
PID control is widely used in many industrial settings, due to its easily predictable behaviour,
at least when used with simple plants. In the next exercise you will see what one might
mean by simple.
E6.26 Consider the interconnection in Figure E6.4 with RP a proper plant. Suppose that if
r(s)

RC (s)

RP (s)

y(s)

Figure E6.4 Feedback loop for studying properties of PID control

RC = 1 then the interconnection is IBIBO stable. Show that there exists K0 > 1 and
TD,0 , TI,0 > 0 so that the controller

1
RC (s) = K 1 + TD s +
TI s
IBIBO stabilises the interconnection for all K [1, K0 ], TD [0, TD,0 ], and TI
[TI,0 , ).
Hint: Use the Nyquist criterion of Section 7.1 to show that the number of encirclements of 1+i0 does not change for a PID controller with the parameters satisfying
K [1, K0 ], TD [0, TD,0 ], and TI [TI,0 , ).
In the next exercise we will consider a difficult plant; one that is unstable and nonminimum
phase. For this plant you will see that any conventional strategies for designing a PID
controller, based on the intuitive ideas about PID control as discussed in Section 6.5, are
unlikely to meet with success.
E6.27 Consider, still using the interconnection of Figure E6.4, the plant
RP (s) =

1s
.
s(s 2)

Answer the following questions.

(a) Show that it is not possible to IBIBO stabilise the system using a PID controller
with positive parameters K, TD , and TI .
Hint: One can use one of the several polynomial stability tests of Section 5.5.
However, it turns out that the Routh test provides the simplest way of getting at
what we want here.

278

6 Interconnections and feedback

22/10/2004

Thus we must take at least one of the PID parameters to be negative. Let us consider
the simplest situation where we take K < 0, so that perhaps some of our intuition
about PID controllers persists.
(b) Show that if K < 0 then it is necessary for IBIBO stability that TI > 1.
(c) Show that if K < 0 and TI > 1 then it is necessary for IBIBO stability that
TD < 0.

Exercises for Chapter 6

279

280

This version: 22/10/2004

Chapter 7
Frequency domain methods for stability
In Chapter 5 we looked at various ways to test various notions of stability of SISO control
systems. Our stability discussion in that section ended with a discussion in Section 6.2.3 of
how interconnecting systems in block diagrams affects stability of the resulting system. The
criterion developed by Nyquist [1932] deals further with testing stability in such cases, and
we look at this in detail in this chapter. The methods in this chapter rely heavily on some
basic ideas in complex variable theory, and these are reviewed in Appendix D.

7 Frequency domain methods for stability

22/10/2004

limss0 f (s) does not exist, but there exists a k N so that the limit limss0 (s s0 )k f (s)
does exist. Recall that a meromorphic function on an open subset U C is a function
f : U C having the property that it is defined and analytic except at isolated points
(i.e., except at poles). Now we can state the Principle of the Argument, which relies on the
Residue Theorem stated in Appendix D.
7.1 Theorem (Principle of the Argument) Let U be a simply connected open subset of C and
let C be a contour in U . Suppose that f is a function which
(i) is meromorphic in U ,
(ii) has no poles or zeros on the contour C, and
(iii) has np poles and nz zeros in the interior of C, counting multiplicities of zeros and
poles.
Then
Z 0
f (s)
ds = 2i(nz np ),
C f (s)
provided integration is performed in a counterclockwise direction.
0

Contents
7.1

7.2

7.3

7.4

The Nyquist criterion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279

7.1.1

The Principle of the Argument . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279

7.1.2

The Nyquist criterion for single-loop interconnections . . . . . . . . . . . . . . . . 281

Proof Since f is meromorphic, ff is also meromorphic, and is analytic except at the poles
and zeroes of f . Let s0 be such a pole or zero, and suppose that it has multiplicity k.
Then there exists a meromorphic function f, analytic an s0 , with the property that f (s) =
(s s0 )k f(s). One then readily determines that
f 0 (s)
f0 (s)
k
=
+
f (s)
s s0
f(s)

The relationship between the Nyquist contour and the Bode plot . . . . . . . . . . . . . . 293
7.2.1

Capturing the essential features of the Nyquist contour from the Bode plot . . . . 293

7.2.2

Stability margins . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 294

Robust stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 303

7.3.1

Multiplicative uncertainty

7.3.2

Additive uncertainty

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 304

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 308

Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 311

for s in a neighbourhood of s0 . Now by the Residue Theorem the result follows since k is
positive if s0 is a zero and negative if s0 is a pole.

The use we will make of this theorem is in ascertaining the nature of the image of a closed
contour under an analytic function. Thus we let U C be an open set and c : [0, T ] U
a closed curve. The image of c we denote by C, and we let f be a function satisfying the
hypotheses of Theorem 7.1. Let us denote by c the curve defined by c(t) = f c(t). Since
f has no zeros on C, the curve c does not pass through the origin and so the function
F : [0, T ] C defined by F (t) = ln(
c(t)) is continuous. By the chain rule we have
f 0 (c(t)) 0
c (t),
f (c(t))

F 0 (t) =

7.1 The Nyquist criterion

t [0, T ].

The Nyquist criterion is a method for testing the closed-loop stability of a system based
on the frequency response of the open-loop transfer function.

Therefore

7.1.1 The Principle of the Argument In this section we review one of the essential
tools in dealing with closed-loop stability as we shall in this chapter: the so-called Principle
of the Argument. This is a result from the theory of complex analytic functions. That such
technology should be useful to us has been made clear in the developments of Section 4.4.2
concerning Bodes Gain/Phase Theorem. The Principle of the Argument has to do, as we
shall use it, with the image of closed contours under analytic functions. However, let us
first provide its form in complex analysis. Let U C be an open set, and let f : U C
be an analytic function. A pole for f is a point s0 U with the property that the limit

Using the definition of the logarithm we have

T
T
T

ln(f (c(t))) = ln |f (c(t))| + i]f (c(t)) .

Z
C

f 0 (s)
ds =
f (s)

Z
C

T
f 0 (c(t)) 0

c (t) dt = ln(f (c(t))) .
f (c(t))
0

Since c is closed, the first term on the right-hand side is zero. Using Theorem 7.1 we then
have
T

2(nz np ) = ]f (c(t)) .
(7.1)
0

In other words, we have the following.

PSfrag replacements

281

7.2 Proposition If C and f are as in Theorem 7.1, then the image of C under f encircles
the origin nz np times, with the convention that counterclockwise is positive.
Let us illustrate the principle with an example.

4
3
2

x1
x2 1
x1
x2
log
0
dB
deg
u = 0 ln -1
ln coth(|u| /2)
or 1 -2
m

yos -3

tos 0

x1
x2 1
x1
x2
log
0
dB
deg
u = 0 ln -1
ln coth(|u| /2)
or 1 -2
m

yos -3

tos 0

7.3 Example We take C to be the circle of radius 2 in C. This can be parameterised, for
example, by c : t
7 2eit , t [0, 2]. For f we take

7 Frequency

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)
domain methods
h (t)
hN,D (t)
1 (t)
4
1N,D (t)
(t)
(t) 3
fj (t)

1
,
(s + 1)2 + a2

a R.

Lets see what happens as we allow a to vary between 0 and 1.5. The curve c = f c is
defined by
1
, t [0, 2].
t 7 2it
(e + 1)2 + a2
Proposition 7.2 says that for a < 1 the image of C should encircle the origin in C two times
in the counterclockwise direction, and for a > 1 there should be no encirclements of the
origin. Of course, it is problematic to determine the image of a closed contour under a given
analytic function. Here we let the computer do the work for us, and the results are shown

in Figure 7.1. We see that the encirclements are as we expect.
7.1.2 The Nyquist criterion for single-loop interconnections Now we apply the
Principle of the Argument to determine the stability of a closed-loop transfer function. The
block diagram configuration we consider here is shown in Figure 7.2. The key observation is
that if the system is to be IBIBO stable then the poles of the closed-loop transfer function
T (s) =

RC (s)RP (s)
1 + RC (s)RP (s)

must all lie in C . The idea is that we examine the determinant 1 + RC RP to ascertain
when the poles of T are stable.
We denote the loop gain by RL = RC RP . Suppose that RL has poles on the imaginary
axis at i1 , . . . , ik where k > > 1 0. Let r > 0 have the property that
1
min {|i j |}.
r<
2 i,j{1,...,k}

(7.2)

,0
0

22/10/2004

PSfrag replacements

-3

-2

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)
h (t)
0
1
h-1
N,D (t)
Re
1 (t)
4
1N,D (t)
(t)
(t) 3
fj (t)

4,0

-3

-2

-1

x1
x2 1
x1
x2
log
0
dB
deg
u = 0 ln -1
ln coth(|u| /2)
or 1 -2
m

yos -3

tos 0

,0
0

-3

-2

-1

Figure 7.1 Images of closed contours for a = 0 (top left), a = 0.5

(top right), and a = 1.5 (bottom)

i6=j

r(s)

That is, r is smaller than half the distance separating the two closest poles on the imaginary
axis. Now we choose R so that
r
R > k + .
(7.3)
2
With r and R so chosen we may define a contour R,r which will be comprised of a collection
of components. For i = 1, . . . , k we define

r,k,+ = ii + rei 2 2 , r,k, = ii + rei 2 2 .
Now for i = 1, . . . , k 1 define
r,i,+ = {i | i + r < < i+1 r} ,

and also define

r,i,+ = { i | k + r < < R} ,

RC (s)

RP (s)

y(s)

Figure 7.2 A unity feedback loop

Finally we define
r,i, = { i | i r > > i+1 + r} ,

r,i, = { i | k r > > R} .

f (s) =

for stability

7.1 The Nyquist criterion

22/10/2004

PSfrag replacements

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)
h282
(t)
hN,D (t)
1 (t)
1N,D (t)
(t)
(t)
fj (t)

R = Rei

7.1 The Nyquist criterion

22/10/2004

283

The union of all these various contours we denote by R,r :

R,r =

k
[

r,i,+ r,i,

k
[

i=1

r,i,+
r,i,

[

R .

i=1

When RL has no poles on the imaginary axis, for R > 0 we write

[
R,0 = { i | R < < R} R .
Note that the orientation of the contour R,r is taken by convention to be positive in the
clockwise direction. This is counter to the complex variable convention, and we choose this
convention because, for reasons will soon see, we wish to move along the positive imaginary
axis from bottom to top. In any case, the idea is that we have a semicircular contour
extending into C+ , and we need to make provisions for any poles of RL which lie on the
imaginary axis. The situation is sketched in Figure 7.3. With this notion of a contour behind
Im

284

7 Frequency domain methods for stability

22/10/2004

(i) If RL has no poles on the imaginary axis then there exists M > 0 so that for any
M ). Furthermore
R > 0 the (R, 0)-Nyquist contour is contained in the disk D(0,
limR NR,0 is well-defined and we denote the limit by N,0 .
(ii) If RL is strictly proper, then for any r > 0 satisfying (7.2) and for any > 0 there
) for any R > R0 . Thus limR NR,r is
exists R0 > 0 so that NR,r \ NR0 ,r D(0,
well-defined and we denote the limit by N,r .
(iii) If RL is both strictly proper and has no poles on the imaginary axis, then the consequences of (ii) hold with r = 0, and we denote by N,0 the limit limR NR,0 .
Proof (i) Define R(s) = RL ( 1s ) for s 6= 0. Since RL is proper, the limit lims0 R(s) exists.
But, since RL is continuous, this is nothing more than the assertion we are trying to prove.
(ii) If RL is strictly proper then lims RL (s) = 0. Therefore, by continuity of RL we can
choose R0 sufficiently large that, for R > R0 , those points which lie in the (R, r)-Nyquist
). This is
contour but do not lie in the (R0 , r)-Nyquist contour reside in the disk D(0,
precisely what we have stated.

(iii) This is a simple consequence of (i) and (ii).
The punchline here is that the Nyquist contour is always bounded for proper loop gains,
provided that there are no poles on the imaginary axis. When there are poles on the imaginary axis, then the Nyquist contour will be unbounded, but for any fixed r > 0 sufficiently
small, we may still consider letting R .
Let us see how the character of the Nyquist contour relates to stability of the closed-loop
system depicted in Figure 7.2.

R
Re

Figure 7.3 The contour R,r

us, we can define what we will call the Nyquist contour.

7.4 Definition For the unity feedback loop of Figure 7.2, let RL be the rational function
RC RP , and let R and r satisfy the conditions (7.3) and (7.2). The (R, r)-Nyquist contour
is the contour RL (R,r ) C. We denote the (R, r)-Nyquist contour by NR,r .

When we are willing to live with the associated imprecision, we shall often simply say
Nyquist contour in place of (R, r)-Nyquist contour.
Let us first state some general properties of the (R, r)-Nyquist contour. At the same time
we introduce some useful notation. Since we are interested in using the Nyquist criterion
for determining IBIBO stability, we shall suppose RL to be proper, as in most cases we
encounter.
) =
7.5 Proposition Let RL be a proper rational function, and for > 0 let D(0,
{ s C | |s| } be the disk of radius centred at the origin in C. The following statements
hold.

7.6 Theorem (Nyquist Criterion) Let RC and RP be rational functions with RL = RC RP

proper. Let np be the number of poles of RL in C+ . First suppose that 1+RL has no zeros on
iR. Then the interconnected SISO linear system represented by the block diagram Figure 7.2
is IBIBO stable if and only if
(i) there are no cancellations of poles and zeros in C+ between RC and RP ;
(ii) lims RL (s) 6= 1;
(iii) there exists R0 , r0 > 0 satisfying (7.3) and (7.2) with the property that for every
R > R0 and r < r0 , the (R, r)-Nyquist contour encircles the point 1 + i0 in the
complex plane np times in the counterclockwise direction as the contour R,r is traversed
once in the clockwise direction.
Furthermore, if for any R and r satisfying (7.3) and (7.2) the (R, r)-Nyquist contour
passes through the point 1 + i0, and in particular if 1 + RL has zeros on iR, then the
closed-loop system is IBIBO unstable.
Proof We first note that the condition (ii) is simply the condition that the closed-loop
transfer function be proper. If the closed-loop transfer function is not proper, then the
resulting interconnection cannot be IBIBO stable.
By Theorem 6.38, the closed-loop system is IBIBO stable if and only if (1) all the closedloop transfer function is proper, (2) the zeros of the determinant 1 + RL are in C , and
(3) there are no cancellations of poles and zeros in C+ between RC and RP . Thus the first
statement in the theorem will follow if we can show that, when 1 + RL has no zeros on iR,
the condition (iii) is equivalent to the condition
(iv) all zeros of the determinant 1 + RL are in C .

22/10/2004

7.1 The Nyquist criterion

285

Since there are no poles or zeros of 1 + RL on R,r , provided that R > R0 and r < r0 , we can
apply Proposition 7.2 to the contour R,r and the function 1 + RL . The conclusion is that
the image of R,r under 1 + RL encircles the origin nz np times, with nz being the number
of zeros of 1 + RL in C+ and np being the number of poles of 1 + RL in C+ . Note that the
poles of 1 + RL are the same as the poles of RL , so np is the same as in the statement of the
theorem. The conclusion in this case is that nz = 0 if and only if the image of R,r under
1 + RL encircles the origin np times, with the opposite orientation of R,r . This, however, is
equivalent to the image of R,r encircling 1 + i0 np times, with the opposite orientation of
R,r .
Finally, if the (R, r)-Nyquist contour passes through the point 1 + i0, this means that
the contour 1 + RL (R,r ) passes through the origin. Thus this means that there is a point
s0 R,r which is a zero of 1 + RL . However, since all points on R,r are in C+ , the result
follows by Theorem 6.38.

Let us make a few observations before working out a few simple examples.
7.7 Remarks 1. Strictly proper rational functions always satisfy the condition (ii).
2. Of course, the matter of producing the Nyquist contour may not be entirely a straightforward one. What one can certainly do is produce it with a computer. As we will see,
the Nyquist contour provides a graphical representation of some important properties of
the closed-loop system.
3. By parts (i) and (ii) of Proposition 4.13 it suffices to plot the Nyquist contour only as we
traverse that half of R,r which sits in the positive imaginary plane, i.e., only for those
values of s along R,r which have positive imaginary part. This will be borne out in the
examples below.
4. When RL is proper and when there are no poles for RL on the imaginary axis (so we
can take r = 0), the (R, 0)-Nyquist contour is bounded as we take the limit R . If
we further ask that RL be strictly proper, that portion of the Nyquist contour which is
the image of R under RL will be mapped to the origin as R . Thus in this case
it suffices to determine the image of the imaginary axis under RL , along with the origin
in C. Given our remark 3, this essentially means that in this case we only determine
the polar plot for the loop gain RL . Thus we see the important relationship between the
Nyquist criterion and the Bode plot of the loop gain.
5. Heres one way to determine the number of times the (R, r)-Nyquist contour encircles
the point 1 + i0. From the point 1 + i0 draw a ray in any direction. Choose this
ray so that it is nowhere tangent to the (R, r)-Nyquist contour, and so that it does not
pass through points where the (R, r)-Nyquist contour intersects itself. The number of
times the (R, r)-Nyquist contour intersects this ray while moving in the counterclockwise
direction is the number of counterclockwise encirclements of 1 + i0. A crossing in the
clockwise direction is a negative counterclockwise crossing.

The Nyquist criterion can be readily demonstrated with a couple of examples. In each
of these examples we use Remark 7.7(5).
In the Nyquist plots below, the solid contour is the image of points in the positive
imaginary plane under RL , and the dashed contour is the image of the points in
the negative imaginary plane.

286

7 Frequency domain methods for stability

22/10/2004

1
7.8 Examples 1. We first take RC (s) = 1 and RP (s) = s+a
for a R. Note that conditions (i) and (ii) of Theorem 7.6 are satisfied for all a, so stability can be check by
verifying the condition (iii). We note that for a < 0 there is one pole of RL in C+ , and
otherwise there are no poles in C+ .
1
is strictly proper with no poles on the imaginary axis unless
The loop gain RL (s) = s+a
a = 0. So let us first consider the situation when a 6= 0. We need only consider
the image under RL of points on the imaginary axis. The corresponding points on the
Nyquist contour are given by

a
1

= 2
i 2
,
i + a
+ a2
+ a2

(, ).

1
This is a parametric representation of a circle of radius 2|a|
centred at
checked by verifying that
a
1 2 2
1

+
= 2.
2
2
+a
2a
2 + a2
4a

1
.
2a

This can be

The Nyquist contour is shown in Figure 7.4 for various nonzero a. From Figure 7.4 we
make the following observations:
(a) for a < 1 there are no encirclements of the point 1 + i0;
(b) for a = 1 the Nyquist contour passes through the point 1 + i0;
(c) for 1 < a < 0 the Nyquist contour encircles the point 1 + i0 one time in the
counterclockwise direction (to see this, one must observe the sign of the imaginary
part as runs from to +);
(d) for a > 0 there are no encirclements of the point 1 + i0.
Now let us look at the case where a = 0. In this case we have a pole for RL at s = 0, so
this must be taken into account. Choose r > 0. The image of { i | > r} is

i 0 < < 1r .
Now we need to look at the image of the contour r given by s = rei , [ 2 , 2 ]. One
readily sees that the image of r is
ei
,
r

[ , ]
2 2

which is a large semi-circle centred at the origin going from +i to in the clockwise
direction. This is shown in Figure 7.5.
This allows us to conclude the following:
(a) for a < 1 the system is IBIBO unstable since np = 1 and the number of counterclockwise encirclements is 1;
(b) for a = 1 the system is IBIBO unstable since the Nyquist contour passes through
the point 1 + i0;
(c) for 1 < a < 0 the system is IBIBO stable since np = 1 and there is one counterclockwise encirclement of the point 1 + i0;
(d) for a 0 the system is IBIBO stable since np = 0 and there are no encirclements
of the point 1 + i0.

PSfrag replacements
PSfrag replacements

PSfrag replacements

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
22/10/2004
u(t)
h (t)
hN,D (t)
1 (t)
1N,D (t)
2
(t)
(t)
fj (t) 1.5

7.1

-1.5

-1

-0.5

0.5

domain methods for stability

-1.5

-1

-0.5

0.5

1.5

,0
0

-10

-5

-1

-0.5

0.5

1 + RL (s) =

-1.5

-1

-0.5

0.5

1.5

-1.5

-1

-0.5

0.5

1.5

2,0
0

1
s

We can also check this directly by using Theorem 6.38. Since there are no unstable
pole/zero cancellations in the interconnected system, IBIBO stability is determined by
the zeros of the determinant, and the determinant is

x1
x2
x1
0.5
x2
log
0
dB
deg
u = 0 ln -0.5
ln coth(|u| /2)
or 1
-1
m

yos -1.5

tos 0

22/10/2004

Figure 7.5 The (, 0.1)-Nyquist contour for RL (s) =

-1.5

,0
0

5
x1
x2
x1
x2
log
0
dB
deg
u = 0 ln
ln coth(|u| /2)
-5
or 1
m

yos
-10
tos 0

x1
x2
x1
0.5
x2
log
0
dB
PSfrag replacements
deg
u = 0 ln t -0.5
ke(t)k
ln coth(|u|
/2)
1
1 (t)
ore
e2(t)
-1
m
e3 (t)

y(t)
yos -1.5
u(t)

htos
(t)
0
hN,D (t)

1
1.5
2
,0
1 (t)
0
1N,D (t)
2
(t)
(t)
fj (t) 1.5

288

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)
h7 (t)Frequency
hN,D (t)
1 (t)
1N,D (t)
(t)
(t) 10
fj (t)

x1
x2
x1
0.5
x2
log
0
dB
PSfrag replacements
deg
u = 0 ln t -0.5
ke(t)k
ln coth(|u|
/2)
1
1 (t)
ore
e2(t)
-1
m
e3 (t)

y(t)
yos -1.5
u(t)

htos
(t)
0
hN,D (t)

1
1.5
2
1,
(t)0
0
1N,D (t)
2
(t)
(t)
fj (t) 1.5

x1
x2
x1
0.5
x2
log
0
dB
deg
u = 0 ln -0.5
ln coth(|u| /2)
or 1
-1
m

yos -1.5

tos 0

287

x1
x2
x1
0.5
x2
log
0
dB
PSfrag replacements
deg
u = 0 ln t -0.5
ke(t)k
ln coth(|u|
/2)
1
1 (t)
ore
e2(t)
-1
m
e3 (t)

y(t)
yos -1.5
u(t)

htos
(t)
0
hN,D (t)

,0
1 (t)
0
1N,D (t)
2
(t)
(t)
fj (t) 1.5

criterion

x1
x2
x1
0.5
x2
log
0
dB
PSfrag replacements
deg
u = 0 ln t -0.5
ke(t)k
ln coth(|u|
/2)
1
1 (t)
ore
e2(t)
-1
m
e3 (t)

y(t)
yos -1.5
u(t)

htos
(t)
0
hN,D (t)

1,
(t)0
0
1N,D (t)
2
(t)
(t)
fj (t) 1.5

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
Theu(t)
Nyquist
h (t)
hN,D (t)
1 (t)
1N,D (t)
2
(t)
(t)
fj (t) 1.5

The zero of the determinant is a 1 which is in C exactly when a > 1, and this is
exactly the condition we derive above using the Nyquist criterion.
2. The previous example can be regarded as an implementation of a proportional controller.
1
Lets give an integral controller a try. Thus we take RC (s) = 1s and RP (s) = s+a
. Once
again, the conditions (i) and (ii) of Theorem 7.6 are satisfied, so we need only check
condition (iii).
In this example the loop gain RL is strictly proper, and there is a pole of RL at s = 0.
Thus we need to form a modified contour to take this into account. Let us expand the
loop gain into its real and imaginary parts when evaluated on the imaginary axis away
from the origin. We have
RL (i) =

-1.5

-1

-0.5

0.5

1
Figure 7.4 The (, 0)-Nyquist contour for RL (s) = s+a
, a = 2
(top left), a = 1 (top right), a = 12 (middle left), a = 21
(middle right), a = 1 (bottom left), and a = 2 (bottom right)

1.5

s+a+1
.
s+a

a
1
i
.
2 + a2
( 2 + a2 )

Let us examine the image of { i | > r} as r becomes increasingly small. For near
1
zero, the real part of the Nyquist contour is near 2 +a
2 , and as increases, it shrinks to
zero. Near = 0 the imaginary part is at sgn(a), and as increases, it goes to zero.
Also note that the imaginary part does not change sign. The image of { i | < r}
reflects this about the real axis. It only remains to examine the image of the contour r
around s = 0 given by s = rei where [ 2 , 2 ]. The image of this contour under RL
is
1
, [ 2 , 2 ].
rei (rei + a)

7.1 The Nyquist criterion

22/10/2004

289

For sufficiently small r we have

1
1
1
a
=
rei (rei + a)
rei 1 + ar ei

r
1 1
PSfrag replacements
= i 1 ei +
ret a
a
ke(t)ki
e
1
e1 (t)
=e2 (t) 2 + .
ar
a
e (t)

PSfrag replacements

1 (t)
1N,D (t)
(t)
(t)
fj (t)

,0
0

-10

-5

5
x1
x2
x1
x2
log
0
dB
deg
u = 0 ln
ln coth(|u| /2)
-5
or 1
m

yos
-10
tos 0

10 ,0
0

1 (t)
1N,D (t)
(t) 100
(t)
fj (t) 75
50

x1
x2
x1
25
x2
log
0
dB
deg
u = 0 ln -25
ln coth(|u| /2)
or 1
m -50

yos -75

tos 0

,0
0

-10

-5

-75

-50

-25

100

Figure 7.7 The (, 0.1)-Nyquist contour for RL (s) =

1
s2

We can now make the following conclusions regarding stability:

(a) when a < 0 the system is IBIBO unstable since np = 1 and there is one clockwise
encirclement of the point 1 + i0;
(b) when a = 0 the system is IBIBO unstable since the Nyquist contour passes through
the point 1 + i0;
(c) when a > 0 the system is IBIBO stable since np = 0 and there are no encirclements
of the point 1 + i0.

-5

22/10/2004

with the image of the contour r as r 0. On the imaginary axis we have RL (i) = 12 .
As goes from to 0 the Nyquist contour goes from 0 + i0 to + i0, and as
goes fromPSfrag
0+ to
+ the Nyquist contour goes from + i0 to 0 + i0. On the contour
replacements
r we have
t
ke(t)k
e2i
e1 (t)
RL (rei ) = 2 , [ 2 , 2 ].
e2 (t)
r
e3 (t)
As r 0 this describes
an infinite radius circle centred at the origin which starts at
y(t)
u(t)
+ i0 and goesharound
once in the clockwise direction. In particular, when a = 0 the
(t)
N,D (t)
Nyquist contour hpasses
through the point 1 + i0. The contour is shown in Figure 7.7.

5
x1
x2
x1
x2
log
0
dB
deg
PSfrag replacements
u = 0 ln t
ln coth(|u|
/2)
ke(t)k
-5
1
ore
1 (t)
e2(t)
m

e3 (t)
yos
y(t)
-10
u(t)
0
htos
(t)

h
(t)
N,D
5
10
,0
1(t)
0
1N,D (t)
(t)
(t) 10
fj (t)

5
x1
x2
x1
x2
log
0
dB
deg
u = 0 ln
ln coth(|u| /2)
-5
or 1
m

yos
-10
tos 0

-10

7 Frequency domain methods for stability

5
x1
x2
x1
x2
log
0
dB
deg
PSfrag replacements
u = 0 ln t
ln coth(|u|
/2)
ke(t)k
-5
1
ore
1 (t)
e2(t)
m

e3 (t)
yos
y(t)
-10
u(t)
0
htos
(t)

hN,D
(t)
,0
1(t)
0
1N,D (t)
(t)
(t) 10
fj (t)

y(t)

0, the contour r gets mapped

u(t) into a semi-circle of infinite radius, centred
h (t)
which goes clockwise from ha1N,D
isgn(a) to a12 isgn(a). In Figure 7.6 we
2 +
(t)

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)Thus, as r
h (t)
at 1 + i0,
hN,D (t) a2
1 (t)
1N,D (t)
(t)
(t) 10
fj (t)

290

We can also check IBIBO stability of the system via Theorem 6.38. Since there are no
unstable pole/zero cancellations, we can look at the zeros of the determinant which is
-10

-5

1
Figure 7.6 The (, 0.1)-Nyquist contours for RL (s) = s(s+a)
,a=
2 (top left), a = 1 (top right), a = 1 (bottom left), and
a = 2 (bottom right)

plot the Nyquist contours.

Now we look at the particular case when a = 0. In this case RL (s) = s12 , and so the
Nyquist contour is the image of the imaginary axis, except the origin, under RL , along

1 + RL (s) =

s2 + as + 1
.
s2 + as

By the Routh/Hurwitz criterion, the system is IBIBO stable exactly when a > 0, and
this is what we ascertained using the Nyquist criterion.
3. Our final example combines the above two examples, and implements a PI controller
1
where we take RC (s) = 1 + 1s and RP (s) = s+a
. Here the condition (ii) of Theorem 7.6
holds, but we should examine (i) just a bit carefully before moving on. We have RL (s) =
s+1 1
, and so there is a pole/zero cancellation here when a = 1. However, it is a stable
s s+a
pole/zero cancellation, so condition (i) still holds. We also ascertain that RL has one
pole in C+ when a < 0.

PSfrag replacements

1
rei + 1
a
rei 1 + ar ei

r i
rei + 1
1

e
+

=
arei
a
1 ei rei
1
= +
2 2 +
a
ar
a
a
ei a 1
=
+ 2 +
ar
a

RL (rei ) =

From this we conclude that for a 6= 0 the image as r 0 of r under RL is an infinite

radius semi-circle centered at a1
and going clockwise from a1
+ isgn(a) to a1

a2
a2
a2
isgn(a). In Figure 7.8 we show the Nyquist contours for various values of a 6= 0.
Now let us consider the image of the imaginary axis when a = 0. In this case we have
RL (i) =

1
1
i .
2

Thus the image of those points on the imaginary axis away from the origin describe a
parabola, sitting in C , passing through the origin, and symmetric about the real axis.
As concerns r when a = 0 we have
ei ei
RL (re ) = 1 +
r
r
i

which, as r 0, describes an infinite radius circle centred at the origin and going
clockwise from + i0 to + i0. This Nyquist contour is shown in Figure 7.9.
With the above computations and the corresponding Nyquist plots, we can make the
following conclusions concerning IBIBO stability of the closed-loop system.
(a) For a < 1 the system is IBIBO unstable since np = 1 and there is one clockwise
encirclement of 1 + i0.
(b) For a = 1 the system is IBIBO unstable since the Nyquist contour passes through
the point 1 + i0.
(c) For 1 < a < 0 the system is IBIBO stable since np = 1 and there is one counterclockwise encirclement of the point 1 + i0.
(d) For a 0 the system is IBIBO stable since np = 0 and there are no encirclements
of 1 + i0.

x1
2
x2
x1
x2
log
0
dB
PSfrag replacements
deg
t
u = ke(t)k
0 ln
ln coth(|u|
/2) -2
e1 (t)
1
ore
2 (t)
m
e3(t)
y(t) -4
yos
u(t)

h (t)
tos 0
hN,D
(t)

2
13,
(t)0
0
1N,D(t)
10
(t)
(t)
fj (t) 7.5

5
x1
x2
x1
2.5
x2
log
0
dB
deg
PSfrag replacements
u = 0 ln -2.5
ln coth(|u| /2)t
ke(t)k
1
ore(t)
1
-5
m
e2 (t)

e3y(t)
os
y(t) -7.5
tu(t)
0
hos
(t)

,
hN,D
(t)0
0
1(t)
1N,D (t)
(t)
4
(t)
fj (t)
2
x1
x2
x1
x2
log
0
dB
deg
u = 0 ln
ln coth(|u| /2)
-2
or 1
m

yos
-4
tos 0

,0
0

22/10/2004

x1
x2 1
x1
x2
log
0
dB
PSfrag replacements
deg
t
u = ke(t)k
0 ln
ln coth(|u|
/2) -1
e1 (t)
1
ore
2 (t)
m
e3(t)
y(t) -2
yos
u(t)

h (t)
tos 0
hN,D
(t)

1,
(t)0
0
1N,D(t)
10
(t)
(t)
fj (t) 7.5

for stability

-2

-1

5
x1
x2
x1
2.5
x2
log
0
dB
deg
PSfrag replacements
u = 0 ln -2.5
ln coth(|u| /2)t
ke(t)k
1
ore(t)
1
-5
m
e2 (t)

e3y(t)
os
y(t) -7.5
tu(t)
0
hos
(t)

5
7.5 hN,D
10,
(t)0
0
1(t)
1N,D (t)
(t)
4
(t)
fj (t)

-4

-2

Lets first consider the situation when a 6= 0. In this case, as goes from i to 0 , the
real part goes from 0 to a1
and the imaginary part goes from 0 to sgn(a). While the
a2
real part never changes sign, the imaginary part can change sign, so we must take care
about how it is behaving. As goes from 0+ to +, the resulting part of the Nyquist
contour is the mirror image about the real axis of what we have already computed. Now
let us look at the image of the contour r . When a 6= 0 we have

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
domain methods
u(t)
h (t)
hN,D (t)
1 (t)
1N,D (t) 6
(t)
(t)
fj (t)
4

-7.5

-5

-2.5

2.5

2
x1
x2
x1
x2
log
0
dB
deg
u = 0 ln
ln coth(|u| /2)
-2
or 1
m

yos
-4
tos 0

-7.5

-5

-2.5

2.5

7.5

In this example, RL is again strictly proper, and has a pole on the imaginary axis at the
origin. Thus to compute the Nyquist contour we determine the image of the imaginary
axis minus the origin, and tack on the image of the contour r = rei , [ 2 , 2 ]. We
first compute
a1
2 + a
RL (i) = 2
.
i
+ a2
( 2 + a2 )

7 Frequency

291

7.1 The Nyquist criterion

22/10/2004

PSfrag replacements

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
292
u(t)
h (t)
hN,D (t)
1 (t)
1N,D (t)
(t)
(t)
fj (t)

-4

-2

,0
0

-4

-2

s+1
,a=
Figure 7.8 The (, 0.2)-Nyquist contours for RL (s) = s(s+a)
1
2 (top left), a = 1 (top right), a = 2 (middle left), a = 12
(middle right), a = 1 (bottom left), and a = 2 (bottom right)

22/10/2004

7.2

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)
h (t)
The
relationship
hN,D (t)
1 (t)
1N,D (t) 30
(t)
(t)
fj (t)

between the Nyquist contour and the Bode plot

293

294

7 Frequency domain methods for stability

22/10/2004

s+1
7.9 Example (Example 7.83 contd) The loop gain, recall, is RL (s) = s(s+a)
. Let us write
this transfer function in the recommended form for making Bode plots. For a 6= 0 we have

RL (s) =

,0
0

11 1
(s + 1).
a s as + 1

Thus the frequency response for RL is a product of four terms:

x1
x2
10
x1
x2
log
0
dB
deg
u = 0 ln
ln coth(|u| /2) -10
or 1
m

yos -20

tos 0

1
H1 () = ,
a

-20

-10

Figure 7.9 The (, 0.2)-Nyquist contour for RL (s) =

s+1
s2

As always, we can evaluate IBIBO stability with Theorem 6.38. To do this we can
still simply look at the zeros of the determinant, because although there is a pole zero
cancellation when a = 1, it is a cancellation of stable factors so it does not hurt us. We
compute the determinant to be
1 + RL (s) =

s2 + (a + 1)s + 1
.
s2 + as

An application of the Routh/Hurwitz criterion suggests that we have IBIBO stability for
a > 1, just as we have demonstrated with the Nyquist criterion.

Note that the Nyquist criterion as we have shown is applicable only to interconnections
with a simple structure, namely a single loop. Bode [1945] discusses a version of the Nyquist
criterion for systems with multiple loops, and this is explored further by Zadeh and Desoer
[1979]. However, the development is too significant, and the outcome too modest (what are
obtained are sufficient conditions for IBIBO stability under restrictive hypotheses) to make
a presentation of these results worthwhile.

7.2 The relationship between the Nyquist contour and the Bode plot
The above examples, although simple, demonstrate that obtaining the Nyquist contour
can be problematic, at least by hand. This is especially well illustrated by the third of the
three examples where the capacity to change sign of the imaginary part of the restriction
of RL to the imaginary axis causes some difficulties which must be accounted for. A useful
observation here is that the Nyquist contour is in essence the polar plot for the loop gain,
taking care of the possibility of poles on the imaginary axis. The matter of constructing
a Bode plot is often an easier one than that of building the corresponding polar plot, so a
plausible approach for making a Nyquist contour is to first make a Bode plot, and convert
this to a polar plot as discussed in Section 4.3.3.
7.2.1 Capturing the essential features of the Nyquist contour from the Bode plot
Let us illustrate this with the third of our examples from the previous section.

i
H2 () = ,

H3 () =

1
,
1 + i a

H4 () = 1 + i.

Each of these is a simple enough function for the purpose of plotting frequency response,
and the effect of a is essentially captured in H3 .
Let us see if from Bode plot considerations we can infer when the imaginary part of the
transfer function changes sign as goes from 0+ to +. In Example 7.83 we determined
to 0
that as we vary in this way, the real part of the frequency response goes from a1
a2
and the imaginary part goes from sgn(a) to 0. The question is, For which values of a
does the imaginary part of the frequency response change sign as goes from 0+ to +?
Provided a1
< 0 this will happen if and only if the phase is at some finite frequency
.
a2
Lets take a look at the phase of the frequency response function.
1. First we take a < 0. Since ]H1 () = and ]H2 () = 2 , in order to have the total
phase equal , it must be the case that
}.
]H3 (
) + ]H4 (
) { 2 , 3
2
The phase of H4 varies from 0 to 2 , and the phase of H3 varies from 0 to 2 (because
a < 0!) Therefore we should aim for conditions on when ]H3 (
) + ]H4 (
) = 2 for
some finite frequency
. But one easily sees that for any a < 0 there will always be a
finite frequency
so that this condition is satisfied since

lim ]H3 () + ]H4 () = .

Thus as long as a < 0 there will always be a sign change in the imaginary part of the
frequency response as we vary from 0+ to +. The Bode plots for RL are shown in
Figure 7.10 for various a < 0.
2. Now we consider when a > 0. In this case we have ]H1 () = 0 and so we must seek

so that
]H3 (
) + ]H4 (
) { 2 , 3
}.
2
However, since a > 0 the phase of H3 will go from 0 to 2 . Therefore it will not be
possible for ]H3 () + ]H4 () to equal either 2 or 3
, and so we conclude that for
2
a > 0 the imaginary part of RL (i) will not change sign as we vary from 0+ to +.
The Bode plots for RL are shown in Figure 7.11 for various a > 0.

7.2.2 Stability margins The above example illustrates how the Bode plot can be
useful in determining certain aspects of the behaviour of the Nyquist contour. Indeed, if one
gives this a short moments thought, one reaches the realisation that one can benefit a great
deal by looking at the Bode plot for the loop gain. Let us provide the proper nomenclature
for organising such observations.

PSfrag replacements

-100
-150

-20
-40

-1.5 -1 -0.5

log

0.5

1.5

,0
0

20
0

-20
-40

22/10/2004

0.5

100
50
0
-50
-100
-150
-1.5 -1 -0.5

log

,0
0

0.5

1.5

u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

,0
0

150
100
50
0
-50
-100

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)
h (t) 0
-1.5 -1 -0.5
hN,Dlog
(t)
1 (t)
1N,D (t)
(t)
(t)
fj (t)
Re
Im
x1
x2
x1
x2

log

0.5

1.5

100
50
0
-50
-100
-150
-1.5 -1 -0.5

1.5

0.5

30
20
10
0
-10
-20
-30
-40

150

u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

1
1.5
2

100
50
0
-50
-100
-150
-1.5 -1 -0.5

log

,0
0

0.5

-150

PSfrag replacements

deg

u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

1
1.5
2

-1.5 -1 -0.5

150

-1.5 -1 -0.5

0.5

1.5

-1.5 -1 -0.5

0.5

1.5

log

150

deg

u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
7 Frequency domain methods
for stability
u(t)
h (t)
hN,D (t)
1 (t)
40
1N,D (t)
(t)
(t)
20
fj (t)
0
Re
Im
x1
-20
x2
x1
-40
-1.5 -1 -0.5 0
0.5
1
1.5
2
-1.5 -1 -0.5 0
x2
log
log
dB

PSfrag replacements

log

0.5

1.5

u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

150
100
50

deg

-50

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)
h (t) 0 0.5
-1.5 -1 -0.5
hN,Dlog
(t)
1 (t)
1N,D (t)
40
(t)
(t)
20
fj (t)
Re
0
Im
x1
-20
x2
x1
-40
x2

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
296
u(t)
h (t)
hN,D (t)
1 (t)
1N,D (t)
(t)
(t)
fj (t)
Re
Im
x1
x2
x1
x2

deg

PSfrag replacements

295

deg

Nyquist contour and the Bode plot

,0
0

100

deg

u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

150

PSfrag replacements

t
t
ke(t)k
ke(t)k
e1 (t)
e1 (t)
e2 (t)
e2 (t)
e3 (t)
e3 (t)
y(t)
y(t)
u(t)
u(t)
7.2 The relationship between
the
h22/10/2004
h (t)
(t)
hN,D (t)
hN,D (t)
1 (t)
1 (t)
1N,D (t)
1N,D (t)
30
(t)
(t)
20
(t)
(t)
10
fj (t)
fj (t)
0
Re
Re
-10
Im
Im
-20
x1
x1
x2
x2
-30
x1
x1
-40
-1.5 -1 -0.5 0
0.5
1
1.5
2
x2
x2
log

PSfrag replacements

0
-50
-100
-150

log

,0
0

s+1
Figure 7.10 Bode plots for RL (s) = s(s+a)
for a = 2 (top left),
a = 1 (top right), and a = 12 (bottom)

7.10 Definition Let RL R(s) be a proper rational function forming the loop gain for the
interconnection of Figure 7.12. Assume that there is no frequency > 0 for which RL (i) =
1 + i0.
(i) A phase crossover frequency , pc [0, ), for RL is a frequency for which
]RL (ipc ) = 180 .
Let pc,1 , . . . , pc,` be the phase crossover frequencies for RL , and assume these are ordered
so that
RL (ipc,1 ) < < RL (ipc,` ).
Also suppose that for some k {1, . . . , `} we have
RL (ipc,k ) < 1 < RL (ipc,k + 1 ).

s+1
Figure 7.11 Bode plots for RL (s) = s(s+a)
for a =
a = 1 (top right), and a = 2 (bottom)

r(s)

RL (s)

1
2

(top left),

y(s)

Figure 7.12 Unity gain feedback loop for stability margin discussion

1.5

22/10/2004

7.2 The relationship between the Nyquist contour and the Bode plot

297

298

7 Frequency domain methods for stability

(ii) The lower gain margin for RL defined by

22/10/2004
Im

1 + i0

Kmin = RL (ipc,k )1 (0, 1).

Re
1
Kmax

1
Kmin

min

1 < RL (ipc,1 ) < < RL (ipc,` ),

1
Re

max

then Kmin is undefined.

(iii) The upper gain margin for RL is defined by
PSfrag replacements
PSfrag replacements

If
RL (ipc,1 ) < < RL (ipc,` ) < 1,
then Kmax is undefined.
(iv) A gain crossover frequency , gc [0, ), for RL is a frequency for which
|RL (igc )| = 1.
Let gc,1 , . . . , gc,` be the gain crossover frequencies for RL , and assume these are ordered so
that
](RL (igc,1 )) < < ](RL (igc,` )).

x1 1
x2
x1
x2
log
0
dB
deg
u = 0 ln
ln coth(|u| /2) -1
or 1
m
-2
yos

tos 0

phase (right) margins from the

10
0
-10
-20
-30
-40
-50
-1.5 -1 -0.5

-1.5 -1 -0.5

log

Let us parse these definitions, as they are in actuality quite simple. First of all, we note
that it is possible to read the gain and phase margins off the Nyquist plot; this saves one
having to compute them directly using the definitions. Rather than try to state in a precise
way how to procure the margins from the Nyquist plot, let us simply illustrate the process in
Figure 7.13. The basic idea is that for the gain margins, one looks for the positive frequency
crossings of the negative real axis closest to 1 + i0 in each direction. The reciprocal of the
distances to the imaginary axis are the gain margins, as indicated in Figure 7.13. For the
phase margins, one looks for the positive frequency crossings of the unit circle closest to the
point 1 + i0. The angles to the negative real axis are then the phase margins, again as
indicated in Figure 7.13. We shall adopt the convention that when we simply say phase
margin, we refer to the smaller of the upper and lower phase margins.
The interpretations of gain crossover and phase crossover frequencies are clear. At a
gain crossover frequency, the magnitude on the Bode plot for RL will be 0dB. At a phase
crossover frequency, the graph will cross the upper or lower edge of the phase Bode plot.
Note that it is possible that for a given loop gain, some of the margins may not be defined.
Let us illustrate this with some examples.

,0
0

-1

u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

,0
0

0.5

1.5

150
100
50

deg

(vi) The upper phase margin for RL is defined by

(
](RL (igc,1 )) + 180 , ](RL (igc,1 )) 0
max =
undefined,
](RL (igc,1 )) > 0.

Figure 7.13 Getting gain

Nyquist plot

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
(left) and
u(t)
h (t)
hN,D (t)
1 (t)
1N,D (t)
(t)
(t)
fj (t)
Re
Im
x1
x2
x1
x2

(v) The lower phase margin for RL is defined by

(
180 ](RL (igc,` )), ](RL (igc,` )) 0
min =
undefined,
](RL (igc,` )) < 0.

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)
h (t)
hN,D (t)
1 (t)
1N,D (t)
(t)
(t)
fj (t)

Kmax = RL (ipc,k + 1 )1 (1, ).

0
-50
-100
-150

Figure 7.14 Nyquist and Bode plots for RL (s) =

log

0.5

1.5

10
s2 +4s+3

7.11 Examples 1. We consider the loop gain

RL (s) =

10
.
s2 + 4s + 3

In Figure 7.14 we note that there is one gain crossover frequency, and it is roughly at
gc = 2.1. The phase at the gain crossover frequency is about 110 , which gives the
upper phase margin as max 70 . The lower phase margin is not defined. Also, neither
of the gain margins are defined.
2. We take as loop gain
10
RL (s) = 2
.
s + 4s + 3
The Bode and Nyquist plots are shown in Figure 7.15. From the Bode plot we see
that there is one gain crossover frequency and it is approximately at gc = 2.1. The
phase at the gain crossover frequency it is about 70 . Thus the lower phase margin is

PSfrag replacements

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
22/10/2004
u(t)
h (t)
hN,D (t)
1 (t) 3
1N,D (t)
(t)
(t)
fj (t) 2

Nyquist contour and the Bode plot

,0
0

-4

-3

-2

-1

,0
0

300

-1.5 -1 -0.5

log

0.5

1.5

7 Frequency domain methods for stability

22/10/2004

Often when one reads texts on classical control design, one simply sees mentioned gain
margin and phase margin without reference to upper and lower. Typically, a result like
Proposition 7.12 is in the back of the minds of the authors, and it is being assumed that
the lower margins are undefined. In these cases, there is a more direct link between the
stability margins and actual stabilityat least when the loop gain is itself BIBO stableas
evidenced by Proposition 7.12. The following examples demonstrate this, and also show that
one cannot generally expect the converse of the statements in Proposition 7.12 to hold.

150

7.13 Examples 1. We consider

100

10
,
s2 + 4s + 3
which we looked at in Example 7.11. In this case, the hypotheses of part (ii) of Proposition 7.12 are satisfied, and indeed one can see from the Nyquist criterion that the system
is IBIBO stable.
2. Next we consider
10
RL (s) = 2
,
s + 4s + 3
which we also looked at in Example 7.11. Here, the lower phase margin is defined, so
the hypotheses of part (ii) of Proposition 7.12 are not satisfied. In this case, the Nyquist
criterion tells us that the system is indeed not BIBO stable.
3. The first of the preceding examples illustrates how one can use the conditions on gain
and phase margin of Proposition 7.12 to determine closed-loop stability when RL is
BIBO stable. The second example gives a system which violates the sufficient conditions
of Proposition 7.12, and is indeed not IBIBO stable. The question then arises, Is the
condition (ii) of Proposition 7.12 necessary for IBIBO stability? The answer is, No,
and we provide an example which illustrates this.
We take
3s 3
(1 s)2 (1 + 25
)
RL (s) =
s
s 3.
2
2(1 + s) (1 + 100 )(1 + 50
)
RL (s) =

deg

u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

299

10
0
-10
-20
-30
-40
-50

x1 1
x2
x1
x2
log
0
dB
deg
u = 0 ln
ln coth(|u| /2) -1
or 1
m
-2
yos

tos 0

7.2 The relationship

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)
between
the
h (t)
hN,D (t)
1 (t)
1N,D (t)
(t)
(t)
fj (t)
Re
Im
x1
x2
x1
x2

PSfrag replacements

0
-50
-100
-150
-1.5 -1 -0.5

log

0.5

1.5

10
Figure 7.15 Nyquist and Bode plots for RL (s) = s2 +4s+3

min 110 . We note that there the upper phase margin is undefined, as are the phase
crossover frequencies.

gain margin

3.
For cases when RL is BIBO stable, we can offer an interpretation of the gain and phase
margins in terms of closed-loop stability.
7.12 Proposition Let RL RH+
be a BIBO stable loop gain, and consider a unity gain feedback block diagram configuration like that of Figure 7.12. Either of the following statements
implies IBIBO stability of the closed-loop system:
(i) Kmin is undefined and either

This is a BIBO stable loop gain, and its Bode plot is shown in Figure 7.16. From the
Bode plot we can see that there is a gain crossover frequency gc satisfying something
like log gc = 0.8. The lower phase margin is defined at this frequency, and it is roughly
60 . Thus the lower phase margin is defined, and this loop gain is thus contrary to the
hypotheses of part (ii) of Proposition 7.12. Now let us examine the Nyquist contour for the
system. In Figure 7.17 we show the Nyquist contour, with the right plot showing a blowup
around the origin. This is a pretty messy Nyquist contour, but a careful accounting of
what is going on will reveal that there is actually a total of zero encirclements of 1 + i0.
Thus the closed-loop system with loop gain RL is IBIBO stable. This shows that the
converse of Proposition 7.12 is not generally true.

(a) Kmax > 1 or

(b) Kmax is undefined.
(ii) min is undefined and either
(a) max > 0.
(b) max is undefined.
Proof We use the Nyquist criterion for evaluating IBIBO stability. For a stable loop gain
RL , the closed-loop system is IBIBO stable if and only if there are no encirclements of
1 + i0. Note that the assumption that RL RH+
be BIBO stable implies that there are
no poles for RL on the imaginary axis, so the (, 0)-Nyquist contour is well-defined and
bounded.
(i) In this case, all crossings of the imaginary axis will occur in the interval (1, 0). This
precludes any encirclements of 1 + i0.
(ii) If (ii) holds, then all intersections at positive frequency of the (, 0)-Nyquist contour
with the unit circle in C will occur in the lower complex plane. Similarly, those crossings of
the unit disk at negative frequencies will take place in the upper complex plane. This clearly
precludes any encirclements of 1 + i0 which thus implies IBIBO stability.

gain margin

The above examples illustrate that one might wish to make the phase margins large and
positive, and make the gain margins large. However, this is only a rough rule of thumb. In
Exercise E7.10 the reader can work out an example where look at one stability margin while
ignoring the other can be dangerous. The following example from the book of Zhou [1996]
indicates that even when looking at both, they do not necessarily form a useful measure of
stability margin.

PSfrag replacements

contour and the Bode plot

301

22/10/2004

PSfrag replacements

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)
h (t)
hN,D (t)
1 (t)
1N,D (t)
(t)
(t)
fj (t)

deg

50
0
-50
-100
-150
-1

,0
0

log

,0
0

PSfrag replacements

t
Figure 7.16 Bode plot for BIBO
stable
transfer function with poske(t)k
itive phase margins
e1 (t)

2
1.5

x1
x2
x1
0.5
x2
log
0
dB
deg
u = 0 ln -0.5
ln coth(|u| /2)
or 1
-1
m

yos -1.5

tos 0
40
50 ,0

,0
0

-1.5

-1

-0.5

0.5

7.14 Example We take as our plant

2s
,
2s 1

and we consider two controllers, both of which may be verified to be stabilising:

RC,1 (s) = 1,

RC,2 (s) =

17 2
s + 33
s + 11
s + 32 s + 1
10
20 10
.
11
2
+ 1 20 s + 1 s + 32 s + 17
10

33
s
10

tells us. In
One can see

-2

,0
0

-1.5

-1

-0.5

0.5

-2

-1.5

-1

-0.5

0.5

2s
Figure 7.18 Nyquist plots for plant RP (s) = 2s1
with controller
RC,1 (top left), with controller RC,2 (top right), and both (bottom)

Figure 7.17 Nyquist contour for BIBO stable transfer function

with positive phase margins

RP (s) =

-1.5

1
x1
x2
x1
0.5
x2
log
dB
deg
0
u = 0 ln
ln coth(|u| /2)
or 1 -0.5
m

-1
yos

tos 0

x1
x2
x1
x2
0
log
dB
deg
u = 0 ln
ln coth(|u| /2)
or 1 -20
m

yos

tos 0 -40

-2

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)
h (t)
hN,D (t)
1 (t)
1N,D (t)
-1(t)-0.5
0
(t) Re
fj (t) 1.5

22/10/2004

e2 (t)
e3 (t)
y(t)
u(t)
h (t)
hN,D (t)
1 (t)
1N,D (t)
(t)
(t)
fj (t)

1
x1
x2
x1
0.5
x2
log
dB
deg
0
u = 0 ln
ln coth(|u| /2)
or 1 -0.5
m

-1
yos

tos 0

0.5
1

PSfrag replacements

100

,0
0

1
x1
x2
x1
0.5
x2
log
dB
deg
0
u = 0 ln
ln coth(|u| /2)
or 1 -0.5
m

-1
yos

tos 0

150

u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

Sfrag replacements

PSfrag replacements

t
t
ke(t)k
ke(t)k
e1 (t)
e1 (t)
e2 (t)
e2 (t)
e3 (t)
e3 (t)
y(t)
y(t)
302
7
Frequency
domain
methods
for stability
u(t)
u(t)
h (t)
h (t)
hN,D (t)
hN,D (t)
second controller is obviously carefully
contrived, but let us see what it
1This
1 (t)
(t)
1N,D
(t)
1N,Dthe
(t) loop gain R
7.18 are shown the Nyquist plots for
Figure
C,1 RP and RC,2 RP .
(t)
(t)
(t)
(t)
fj (t) 1.5
fj (t) 1.5

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)
7.2 The
h (t)relationship between the Nyquist
hN,D (t)
1 (t)
1N,D (t)
30
(t)
(t)
20
fj (t)
10
Re
Im
0
x1
-10
x2
x1
-20
-1
0
1
2
3
x2
log

1.5

that the gain and phase margins for the loop gain RC,2 RP are at least as good as those for
the loop gain RC,1 RP , but that the Nyquist contour for RC,2 RP passes closer to the critical
point 1 + i0. This suggests that gain and phase margin may not be perfect indicators of
stability margin.

With all of the above machinations about gain and phase margins out of the way, let us
give perhaps a simpler characterisation of what these notions are trying to capture. If the
objective is to stay far way from the point 1 + i0, then the following result tells us that
the sensitivity function is crucial in doing this.
7.15 Proposition inf |1 RL (i)| = kSL k1
.
>0

7.3 Robust stability

22/10/2004

303

Proof We compute

-1

-0.5

log

0.5

1.5

2,0

RP (s)

y(s)

Figure 7.20 Unity gain feedback loop for robust stability

Now we, as usual, consider the unity gain feedback loop of Figure 7.20. We wish to
design a controller RC which stabilises a whole set of plants. We devote this section to a
precise formulation of this problem for both uncertainty descriptions, and to providing useful
necessary and sufficient conditions for our problem to have a solution.
Let us be precise about this, as it is important that we know what we are saying.
P , Wu R(s) be a proper rational functions with kWu k < . A
7.17 Definition Let R

P , Wu ) (resp. P+ (R
P , Wu )) if
controller RC R(s) provides robust stability for P (R
P , W )u)
the feedback interconnection of Figure 7.20 is IBIBO stable for every RP P (R
P , Wu )).
(resp. for every RP P+ (R

Now we provide conditions for determining when a controller is robustly stabilising,
breaking our discussion into that for multiplicative, and that for additive uncertainty.

deg
u = 0 ln 6
ln coth(|u| /2)
or 1 4
m
2
yos
0
tos 0

RC (s)

22/10/2004

t
ke(t)k
ke(t)k
e1 (t)
e1 (t)
as
desired.

e2 (t)
e2 (t)
e3 (t)
e3 (t)
y(t) results says, simply, that the point on the
y(t)Nyquist contour which is closest to the point
The
u(t)
u(t)
+ i0 is a distance kSL k1
increase the stability margin, one may wish to
h1
h (t)
(t)
away. Thus, to
hN,D
(t) the sensitivity function small. This is
hN,D
(t)
make
a
reason for minimising the sensitivity function.
1 (t)
1 (t)
We
1N,D
(t) shall encounter others in Sections 8.5 1and
N,D (t)9.3.
(t)
(t)
(t)Let us illustrate Proposition 7.15 on thetwo
(t) loop gains of Example 7.14.
fj (t)
fj (t)
Re Example (Example 7.14 contd) We consider
Re
7.16
the plant transfer function RP and the two
Im
Im
x1
controller
transfer functions RC,1 and RC,2 ofx1Example 7.14. The magnitude Bode plots of
x2
x2
the
x1 sensitivity function for the two loop gains
x1 are shown in Figure 7.19. As expected, the
x2
x2

-1.5

7 Frequency domain methods for stability

r(s)

inf |1 RL (i)| = inf |1 + RL (i)|

>0
>0
1

1

= sup

>0 1 + RL (i)
PSfrag replacements
1
= kS
t L k ,

PSfrag replacements

,0
0

304

-1.5

-1

-0.5

log

0.5

1.5

Figure 7.19 Magnitude bode plot for the sensitivity function with
loop gain RC,1 RP (left) and RC,2 RP (right)

peak magnitude for the sensitivity with the loop gain RC,1 RP is lower than that for RC,2 RP ,
reflecting the fact that the Nyquist contour for the former is further from 1 + i0 than for
the latter.

7.3 Robust stability

We now return to the uncertainty representations of Section 4.5. Let us recall here the
P and a rational function Wu satisfying kWu k < ,
basic setup. Given a nominal plant R

P , Wu ) the set of plants satisfying

we denote by P (R
P ,
RP = (1 + Wu )R
P , Wu ) the set of plants satisfying
where kk 1. We also denote by P+ (R
P + Wu ,
RP = R
where kk 1. This is not quite a complete definition, since in Section 4.5 we additionally
P , Wu ) and P+ (R
P , Wu ) have the same
imposed the requirement that the plants in P (R
P .
number of unstable poles as does R

7.3.1 Multiplicative uncertainty The following important theorem gives simple conditions on when a controller is robustly stabilising. It was first stated by Doyle and Stein
[1981] with the first complete proof due to Chen and Desoer [1982].
P , Wu R(s) be a proper rational functions with kWu k < , and
7.18 Theorem Let R

P IBIBO stable in the interconnection of

suppose that RC R(s) renders the nominal plant R

P , Wu ) if and only if Wu TL < 1,
Figure 7.20. Then RC provides robust stability for P (R

where TL is the complementary sensitivity function, or closed-loop transfer function, for the
P .
loop gain RC R
P , Wu ) denote RL (RP ) = RC RP . By definition
Proof Let RC R(s). For RP P (R
P , Wu ) it follows that R
P and RP share the same imaginary axis poles for every
of P (R
P , Wu ). For r, R > 0 and for RP P (R
P , Wu ) let NR,r (RP ) be the (R, r)RP P (R
Nyquist contour for RL (RP ).
Now we make a simple computation:

Wu TL < 1

Wu (i)TL (i) < 1, R
W (i)R
L (i)
u

< 1, R

1 + RL (i)

L (i) < 1 + R
L (i) , R
Wu (i)R

L (i) < 1 R
L (i) , R.
Wu (i)R

This
gives a simple interpretation of the condition W2 TL < 1. We note that

L (i) is the distance from the point 1+i0 to the point R
L (i). Thus the condition
1 R

7.3 Robust stability

22/10/2004

305

L (i)
W2 TL < 1 is equivalent to the condition that the open disk of radius Wu (i)R

L (i) not contain the point 1 + i0. This is depicted in Figure 7.21. It is this
centred at R

306

7 Frequency domain methods for stability

22/10/2004

(i) ]G (i
) = and
(ii) the map 7 |G (i)| has a maximum at =
.
Proof Clearly if = 0 we may simply define G (s) = 1. Thus suppose that 6= 0. Consider
the rational function
02
T,0 (s) = 2
,
s + 20 s + 02
, 0 > 0. In Exercise E4.6 it was shown that for 12 < < 1 the function 7 |T (i)|
p
2
achieves a unique maximum
at this maxip at max = 0 1 2 . Furthermore, the phase
mum is given by atan2(, 1 2 2 ). Thus, as varies in the interval ( 12 , 1), the phase at
the frequency max varies between 4 and 0.
Now, given (, ], define
(

,
<0

= 5 2
10 , > 0.

1 + i0

L (i)
Wu (i)R

Thus is guaranteed to live in the interval ( 4 , 0). Therefore, there exists ( 12 , 1) so

p
that = atan2(, 1 2 2 ). Now define 0 so that max =
. Now define
(
5
T,
, <0
0
G =
10
T,
, > 0.
0

L (i)
R

Figure 7.21 Interpretation of robust stability condition for multiplicative uncertainty

This function G , it is readily verified, does the job.

interpretation of the condition

W2 TL < 1 we shall employ in the proof.
P , Wu ) we
First suppose that Wu TL < 1. Note that for R and for RP P (R
have
P (i).
RL (RP )(i) = RC (i)RP (i) = (1 + (i)Wu (i))RC (i)R

L (i) with centre
Thus the point RL (RP )(i) lies in the closed disk of radius Wu (i)R
L (i). From Figure 7.21 we infer that the point of the the Nyquist contour NR,r (RP ) that
R
are the image of points on the imaginary axis will follow the points on the Nyquist contour
P ) while remaining on the same side of the point 1 + i0. Since Wu RH+ and
NR,r (R

L have the same poles on iR and the same number of

since is allowable, RL (RP ) and R
poles in C+ . Thus by choosing r0 < 0 sufficiently small and for R0 > 0 sufficiently large, the
P )
number of clockwise encirclements of 1 + i0 by NR,r (RP ) will equal the same by NR,r (R
for all R > R0 and r < r0 . From Theorem 7.6 we conclude IBIBO stability of the closed-loop
system with loop gain RL (RP ).
Now suppose that Wu TL 1. As depicted
in Figure

7.21, this implies the existence
L (i
L (i
of
0 so that the open disk of radius Wu (i
)R
) centred at R
) contains the
0 , r) be the closed disk of radius r centred at s0 . We claim
point 1 + i0. Denote by D(s

P , Wu ) to D R
L (i), Wu (i)R
L (i) defined by
that for each > 0, the map from P (R
P 7 RC (i)RP (i)
RP = (1 + Wu )R

(7.4)

is surjective. The following lemma helps us to establish this fact.

1 Lemma For any
0 and for any (, ] there exists a function G RH+
with
the properties

We now resume showing that the map defined by (7.4) is surjective.

It clearly suffices

L (i), Wu (i)R
L (i) lies
to show that the image of every point on the boundary of D R
in the image of the map, since all other points can then be obtained by scaling . For
(, ] let

L (i) + Wu (i)R
L (i) ei
s = R

L (i), Wu (i)R
L (i) . We wish to write
be a point on the boundary of D R
L (i) + (i)Wu (i)R
L (i)
s = RC (i)RP (i) = R
for an appropriate choice of allowable . Thus RH+
should necessarily satisfy

L (i) ei .

(i)Wu (i)RL (i) = Wu (i)R
L (i)) = . Therefore, define
It follows that |(i)| = 1 and that ](i) + ](Wu (i)R
L (i)).
= ](Wu (i)R
If G is as defined above, then defining
(s) =

G (s)
|G (i)|

does the job.

The remainder of the proof is now straightforward. Since the map defined by (7.4) is
P we
surjective, we conclude that there exists an allowable so that if RP = (1 + Wu )R
have
RL (RP )(i
) = RC (i
)RP (i
) = 1 + i0.
This implies that the Nyquist contour NR,r (RP ) for R sufficiently large passes through the
point 1 + i0. By Theorem 7.6 the closed-loop system is not IBIBO stable.

The proof of the above theorem is long-winded, but the idea is, in fact, very simple.
Indeed, the essential observation, repeated here outside the confines of the proof, is that
the condition kWu TL k < 1 is equivalent to the
in Figure 7.21, that,
condition, depicted

L (i) and centred at R
L (i), not
for each frequency , the open disk of radius Wu (i)R
contain the point 1 + j0.
One may also reverse engineer this problem as well. The idea here is as follows.
P and a controller RC which IBIBO stabilises
Suppose that we have our nominal plant R
the closed-loop system of Figure 7.20. At this point, we have not specified a set of plants
over which we want to stabilise. Now we ask, What is the maximum size of the allowable
P if the perturbed plant is to be stabilised by RC ? The following result
perturbation to R
makes precise this vague question, and provides its answer.
P , RC R(s) with R
P proper. Also, suppose that the interconnection
7.19 Proposition Let R
P . Let TL be the closed-loop transfer function
of Figure 7.20 is IBIBO stable with RP = R
L = RC R
P . The following statements hold:
for the loop gain R
1

(i) if Wu RH+
has the property that kWu k < TL then RC provides robust stability
P , Wu );
for P (R
1
(ii) for any TL there exists Wu RH+
satisfying kWu k = so that RC does not
P , Wu ).
robustly stabilise P (R
Proof (i) This follows directly from Theorem 7.18.

> 0 be a frequency
(ii) This part ofthe result
too follows from Theorem 7.18. Indeed, let
+
for which TL (i
) = TL . One
can
readily
define
W

RH
so that |Wu (i
)| =
u

kWu k = . It then follows that Wu TL 1, so we may apply Theorem 7.18.
Roughly, the proposition tells us that as long as we choose the uncertainty weight Wu so
1
that its H -norm is bounded by TL , then we are guaranteed that RC will provide robust
1
stability. If the H -norm of Wu exceeds TL , then it is possible, but not certain, that RC
will not robustly stabilise. In this latter case, we must check the condition of Theorem 7.18.
Let us illustrate the concept of robust stability for multiplicative uncertainty with a fairly
simple example.
7.20 Example We take as our nominal plant the model for a unit mass. Thus
P (s) = 1 .
R
s2
The PID control law given by
1
s
may be verified to stabilise the nominal plant; the Nyquist plot is shown in Figure 7.22. To
model the plant uncertainty, we choose
RC (s) = 1 + 2s +

Wu (s) =

as
,
s+1

308

domain methods for stability

22/10/2004

x1
x2
5
x1
x2
log
0
dB
deg
u = 0 ln
ln coth(|u| /2) -5
or 1
m

yos -10

tos 0

PSfrag replacements

t
t
ke(t)k
ke(t)k
e1 (t)
e1 (t)
e2 (t)
e2 (t)
,0
-10
-5 e3 (t) 0
5
10
15
e3 (t)
0
y(t)Re
y(t)
u(t)
u(t)
h (t)
h (t)
hN,D (t)
hN,D
(t) controller and nominal plant
Figure 7.22 Nyquist plot for
PID
1 (t)
1 (t)
1N,D (t)
1N,D (t)
(t)
(t)
(t)
(t)
fj (t)According to Theorem 7.18 we should fdetermine
for which values of a the inequality

j (t)

Re T
L we merely need to produce magnitude
L < 1 is satisfied. To determine WRe
W
T
u
u

Im
Im
x1 determine for which values of a the maximum
x1 plots for W T

Bode
u L for various values of a, and
x2
x2
L
x1 is shown the magnitude Bode plot for Wu T
x1
magnitude
does not exceed 0dB. In Figure 7.23
x2
x2
deg
u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

,0
0

0
-10
-20
-30
-40

-1.5

-1

-0.5

log

0.5

deg
u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

1.5

2,0

-10
-20

307

7.3 Robust stability

22/10/2004

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)
h (t)
7
Frequency
hN,D (t)
1 (t)
1N,D (t) 15
(t)
(t)
fj (t)

-30
-40
-1.5

-1

-0.5

log

0.5

1.5

Figure 7.23 The magnitude Bode plot for Wu TL when a = 1 (left)

and when a = 43 (right)

with a = 1. We see that we exceed 0dB by about 2.5dB. Thus we should reduce a by a
factor K having the property that 20 log K = 2.5 or K 1.33. Thus a 0.75. Thus let us
take a = 34 for which the magnitude Bode plot is produced in Figure 7.23. The magnitude
is bounded by 0dB, so we are safe, and all plants of the form
3

s(s) 1
RP (s) = 1 + 4
s + 1 s2

a > 0,

which has the desirable property of not tailing off to zero as s ; plant uncertainty will
generally increase with frequency. We shall determine for what values of a the controller RC
P , Wu ).
provides robust stability for P (R

will be stabilised by RC if is allowable. For example, the Nyquist plot for RP when
s1
= s+2
(which, it can be checked, is allowable), is shown in Figure 7.24.

7.3.2 Additive uncertainty
uncertainty.

Now let us state the analogue of Theorem 7.18 for additive

PSfrag replacements

22/10/2004

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)
h (t)
hN,D (t)
1 (t)
1N,D (t)
(t)
(t)
fj (t)

7.3 Robust stability

309

7 Frequency domain methods for stability

22/10/2004

x1
2
x2
x1
x2
log
0
dB
deg
u = 0 ln
ln coth(|u| /2) -2
or 1
m
-4
yos

tos 0

,0
0

310

1 + i0

|Wu (i)RC (i)|

-2

Figure 7.24 Nyquist plot for perturbed plant under multiplicative

uncertainty

P , Wu R(s) be a proper rational functions with kWu k < , and

7.21 Theorem Let R

P IBIBO stable in the interconnecsuppose that RC R(s) renders the nominal plant R
P , Wu ) if and only if
7.20. Then RC provides robust stability for P+ (R
tion
of Figure

P .
Wu RC SL < 1, where SL is the sensitivity function for the loop gain RC R

Proof We adopt the notation of the first paragraph of the proof of Theorem 7.18. The
following computation, mirroring the similar one in Theorem 7.18, is the key to the proof:

Wu RC SL < 1

Wu (i)RC (i)SL (i) < 1, R
W (i)R (i)
u

C

L (i) < 1, R
1+R

L (i) , R
|Wu (i)RC (i)| < 1 + R

L (i) , R.
|Wu (i)RC (i)| < 1 R

The punchline here is thus that the condition Wu RC SL < 1 is equivalent to the condition that the point 1 + i0 not be contained, for any R, in the open disk of radius
L (i). This is depicted in Figure 7.25.
|Wu (i)RC (i)| with centre R
The remainder of the proof now very much
follows that of Theorem 7.18, so we can
safely omit some details. First assume that Wu RC SL < 1. For R and for RP
P , Wu ) we have
P+ (R
L (i) + (i)RC (i)Wu (i).
RL (RP )(i) = RC (i)RP (i) = R
Thus the point RL (RP )(i) lies in the closed disk of radius |Wu (i)RC (i)| with centre
L (i). We may now simply repeat the argument of Theorem 7.18, now using Figure 7.25
R
rather than Figure 7.21, to conclude that the closed-loop system with loop gain RL (RP ) is
IBIBO stable.

L (i)
R

Figure 7.25 Interpretation of robust stability condition for additive uncertainty

Now suppose that kWu RC k 1. Thus there exists

0 so that |Wu (i
)RC (i
)|
1. Thus by Figure 7.25, it follows that 1 + i0 is contained in the open disk of
L (i
radius |Wu (i
), RC (i
)| with centre R
). As in the proof of Theorem 7.18, we
P , Wu ) to
may employ Lemma 1 of that proof to show that the map from P+ (R
L (i), |Wu (i)RC (i)| defined by
D R
P + Wu 7 RC (i)RP (i)
RP = R
is surjective for each > 0. In particular, it follows that there exists an allowable giving
P , Wu ) so that RL (RP )(i
RP P+ (R
) = 1 + i0. IBIBO instability now follows from

Theorem 7.6.
Again, the details in the proof are far more complicated than is the essential idea. This
essential idea is that for each R the open disk of radius |Wu (i)RC (i)| and centre
L (i) should not contain the point 1 + i0. This is what is depicted in Figure 7.25.
R
We also have the following, now hopefully obvious, analogue of Proposition 7.19.
P , RC R(s) with R
P proper. Also, suppose that the interconnection
7.22 Proposition Let R
P . Let SL be the sensitivity function for the loop
of Figure 7.20 is IBIBO stable with RP = R
L = RC R
P . The following statements hold:
gain R

1

(i) if Wu RH+
has the property that kWu k < RC SL then RC provides robust
P , Wu );
stability for P+ (R

1

(ii) for any RC SL there exists Wu RH+
satisfying kWu k = so that RC does
P , Wu ).
not robustly stabilise P+ (R
An example serves to illustrate the ideas for this section.

7.4 Summary

311

7.23 Example (Example 7.20 contd) We carry on look at the nominal plant transfer function
P (s) = 1
R
s2
PSfrag replacements

t
t
1
ke(t)k
ke(t)k
RC (s) =
e1 (t)
e11(t)+ 2s + .
s
e2 (t)
e2 (t)
e3 (t)
e3 (t)
y(t) that we may no longer use our Wu from
y(t)
Note
Example 7.20 as our plant uncertainty model.
u(t)
u(t)
since RC is improper, Wu is proper hbut
not strictly proper, and SL is proper but not
hIndeed,
(t)
(t)
hN,D (t)
hN,D
(t)
strictly
proper (one can readily compute that
this is so), it follows that Wu RC SL is improper.
1 (t)
1 (t)
Thus
1N,D (t)
1N,D
(t) we modify Wu to
(t) as
(t)
Wu = (t)
(t)
2
fj (s
(t) + 1)
fj (t)
Re
Re
toImmodel the plant uncertainty. Again, our Im
objective will be to determine the maximum
P , Wu ). Thus we should find the
x1 for P+ (R
x1 of a so that RC provides robust stability
value

x2
x2
maximum
value for a so that Wu RC SL x<1 1. In Figure 7.26 is shown the magnitude
x1
x2
x2
deg -20
u = 0 ln
ln coth(|u| /2) -40
or 1
-60
m

-80
yos
-100
tos 0 -120

,0
0

22/10/2004

x1
x2
5
x1
x2
log
0
dB
deg
u = 0 ln
ln coth(|u| /2) -5
or 1
m

yos -10

tos 0

,0
0

-5

Figure 7.27 Nyquist plot for perturbed plant under additive uncertainty

deg
-20
u = 0 ln
ln coth(|u| /2)
-40
or 1
m
-60

-80
yos
-100
tos 0
-120

domain methods for stability

which is stabilised by the PID controller

PSfrag replacements

312

22/10/2004

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)
h (t)
7
Frequency
hN,D (t)
1 (t)
1N,D (t) 15
(t)
(t)
fj (t)

-1.5

-1

-0.5

log

0.5

1.5

2,0

-1.5

-1

-0.5

log

0.5

1.5

Figure 7.26 The magnitude Bode plot for Wu RC SL when a = 1

(left) and when a = 12 (right)

Bode plot for Wu RC SL when a = 1. From this plot we see that we ought to reduce a by
a factor K having the property that 20 log K = 6 or K 2.00. Thus we take a = 12 , and
in Figure 7.26 we see that with this value of a we remain below the 0dB line. Thus we are
guaranteed that all plants of the form
P (s) +
RP (s) = R

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

are stabilised by RC provided that is allowable. The Nyquist plot for RP obtained by
taking = s1
is shown in Figure 7.27. This is a pretty sophisticated Nyquist plot, but
s+2
nonetheless, it is one for an IBIBO stable system.

7.4 Summary
In this chapter we have provided a graphical method for analysing the closed-loop stability
of a single loop feedback interconnection. You should understand the following things.
1. You need to understand what the Nyquist criterion, in the form of Theorem 7.6, is saying.

2. Drawing Nyquist plots can be a bit tricky. The essential idea is that one should focus
on the image of the various parts of the Nyquist contour R,r . In particular, one should
focus on the image of the imaginary axis for very large and very small frequencies. From
there one can try to understand of something more subtle is happening.
3. The gain and phase margins are sometimes useful measures of how close a system is to
being unstable.

Exercises for Chapter 7

313

Exercises
E7.1 For the following rational functions R and contours C, explicitly verify the Principle
of the Argument by determining the number of encirclements of the origin by the
image of C under R.

(a) R(s) = s12 and C = ei < .
i

(b) R(s) = s and C = e < .

(c) R(s) = s2s+1 and C = 2ei < .

E7.2 Let F (s) = ln s, and consider the contour C = ei < . Does the
Principle of the Argument hold for the image of C under F ? Why or why not?
E7.3 For the following loop gains,

, > 0,
(a) RL (s) = s2 (s+1)

(b) RL (s) = s(s1) , > 0, and

, > 0,
(c) RL (s) = (s+1)
s(s1)
do the following.
1. Determine the Nyquist contour for the following loop gains which depend on a
parameter satisfying the given conditions. Although the plots you produce may
be computer generated, you should make sure you provide analytical explanations
for the essential features of the plots as they vary with .
2. Draw the unity gain feedback block diagram which has RL as the loop gain.
3. For the three loop gains, use the Nyquist criterion to determine conditions on
for which the closed-loop system is IBIBO stable.
4. Determine IBIBO stability of the three closed-loop systems using the
Routh/Hurwitz criterion, and demonstrate that it agrees with your conclusions
using the Nyquist criterion.
In this next exercise you will investigate some simple ways of determining how to close a
Nyquist contour for loop gains with poles on the imaginary axis.

314

7 Frequency domain methods for stability

22/10/2004

(c) Determine a relationship between 0 , and k and a.

(d) Determine the relationship between
lim ]RL (i)

0,+

and 0 . This relationship will depend on k.

(e) Conclude that for a real rational loop gain, the closing of the Nyquist contour
is always in the clockwise direction and the closing arc subtends an angle of k.
E7.5 Let (N, D) be a proper SISO linear system in input/output form, and let n = deg(D)
and m = deg(N ).
(a) Determine lim ]TN,D (i) and lim ]TN,D (i).
(b) Comment on part (a) as it bears on the Nyquist plot for the system.
E7.6 Consider the SISO linear system (N (s), D(s)) = (1, s2 + 1) in input/output form.
(a) Sketch the Nyquist contour for the system, noting that the presence of imaginary
axis poles means that the (, r)-Nyquist contour is not bounded as r 0.
Consider the SISO linear system (N (s), D (s)) = (1, s2 + s + 1) which now has a
bounded Nyquist contour for > 0.
(b) Show, using the computer, that the Nyquist contour for (N , D ) approaches that
for the system of part (a) as 0.
E7.7 Let (N, D) be a SISO linear system in input/output form with deg(N ) = 0 and let

= max { Im(p) | p is a root of D} .

Answer the following questions.
(a) Show that |TD,N (i)| is a strictly decreasing function of for || >
.
(b) Comment on part (a) as it bears on the Nyquist plot for the system.
E7.8 Formulate and prove a condition for IBIBO stability for each of the two interconnected
systems in Figure E7.1, using the Principle of the Argument along the lines of the
r(s)

E7.4 Let RL R(s) be a proper loop gain with a pole at i0 of multiplicity k. For
concreteness, suppose that 0 0. If 0 6= 0, this implies that i0 is also a pole of
multiplicity k. Thus we write
(
1
0 = 0
k R(s),
RL (s) = s 1
R(s), 0 6= 0,
(s2 + 2 )k

RC (s)

RP (s)

y(s)

RS (s)

where i0 is neither a pole nor a zero for R.

(a) Show that
a
lim RL (i0 + rei ) = k eik + O(r1k ).
r0
r
where a is either purely real or purely imaginary. Give the expression for a.
(b) For 0 = 0 show that

lim ]RL (i) 2 , 0, 2 , .
0,

Denote 0 = lim0, ]RL (i).

r(s)

RI (s)

RC (s)

RP (s)

Figure E7.1 Alternate configurations for Nyquist criterion

Nyquist criterion of Theorem 7.6.

y(s)

Exercises for Chapter 7

315

E7.9 Let

In the next exercise you will investigate gain and phase margins for a simple plant and
controller. You will see that it is possible to have one stability margin be large while the
other is small. This exercise is taken from the text of [Zhou 1996].
E7.10 Take

as
s+b
, RC (s) =
,
as 1
bs + 1
where a > 1 and b > 0. Consider these in the standard unity gain feedback loop.
(a) Use the Routh/Hurwitz criterion to show that the closed-loop system is IBIBO
stable if b ( a1 , a).
Take b = 1.

2
(b) Show that Kmin = a1 , Kmax = a, min is undefined, and max = arcsin aa2 1
.
+1
(c) Comment on the nature of the stability margins in this case.
Fix a and take b ( a1 , a).
(d) Show that
1
lim Kmin = 2 , lim Kmax = a2 , lim max = 0.
ba
ba
ba
a
(e) Comment on the stability margins in the previous case.
(f) Show that
RP (s) =

b a

lim1 Kmax = 1,

b a

lim1 max = 2 arcsin

b a

a2 1
a2 + 1

7 Frequency domain methods for stability

r(s)

02
, RC (s) = 1.
2
s + 20 s
Plot the upper phase margin of the closed-loop system as a function of .
RP (s) =

lim1 Kmin = 1,

316

(g) Comment on the stability margins in the previous case.

In this chapter we have used the Nyquist criterion to assess IBIBO stability of input/output
feedback systems. It is also possible to use the Nyquist criterion to assess stability of static
state feedback, and the following exercise indicates how this is done.
E7.11 Let = (A, b, ct , D) be a SISO linear system.
(a) Let f Rn . Show that if is controllable then the closed-loop system f is
internally asymptotically stable if and only if
f t (sI n A)1 b
RH+
.
1 + f t (sI n A)1 b
Thus, we may consider closed-loop stability under static state feedback to be equivalent to IBIBO stability of the interconnection in Figure E7.2. The Nyquist criterion
can then be applied, as indicated in the following problem.
(b) Using part (a), prove the following result.
Proposition Let = (A, b, ct , 01 ) be a controllable SISO linear system and let np
be the number of eigenvalues of A in C+ . For a state feedback vector f Rn
define the loop gain Rf (s) = f t (sI n A)1 b. Assuming that (A, f ) is observable,
the following statements are equivalent:

f t (sI n A)1 b

22/10/2004

y(s)

Figure E7.2 A feedback interconnection for static state feedback

(i) the matrix A bf t is Hurwitz;

(ii) there exists r0 , R0 > 0 so that the image of R,r under Rf encircles the point
1 + i0 np times in the clockwise direction for every r < r0 and R > R0 .
To be concrete, consider the situation where

0 1
0
A=
, b=
.
1 1
1
For this system, answer the following questions.
(c) Show that if we take f = (0, 2) then A bf t is Hurwitz. Verify that the result
of part (a) of the problem holds.
(d) How many eigenvalues are there for A in C+ ?
(e) Plot the Nyquist contour for Figure E7.2 and verify that the Nyquist criterion of
part (b) holds.
E7.12 For the plant uncertainty set
RP =

P
R
P ,
1 + Wu R

kk 1,

use the idea demonstrated in the proofs of Theorems 7.18 and 7.21 to state and prove
a theorem providing conditions for robust stability.
E7.13 For the plant uncertainty set
RP =

P
R
,
1 + Wu

kk 1,

use the idea demonstrated in the proofs of Theorems 7.18 and 7.21 to state and prove
a theorem providing conditions for robust stability.

Exercises for Chapter 7

317

318

This version: 22/10/2004

Chapter 8
Performance of control systems
Before we move on to controller design for closed-loop systems, we should carefully investigate to what sort of specifications we should make our design. These appear in various
ways, depending on whether we are working with the time-domain, the s-domain, or the
frequency-domain. Also, the interplay of performance and stability can be somewhat subtle,
and there is often a tradeoff to be made in this regard. Our objective in this chapter is to
get the reader familiar with some of the issues that come up when discussing matters of
performance. Also, we wish to begin the process of familiarising the reader with the ways
in which various system properties can affect the performance measures. In many cases,
intuition is ones best friend when such matters come up. Therefore, we take a somewhat
intuitive approach in Section 8.2. The objective is to look at tweaking parameters in a couple
of simple transfer function to see how they affect the various performance measures. It is
hoped that this will give a little insight into how one might, say, glean some things about
the step response by looking at the Bode plot. The matter of tracking and steady-state error
can be dealt with in a more structured way. Disturbance rejection follows along similar lines.
The final matter touched upon in this chapter is the role of the sensitivity function in the
performance of a unity gain feedback loop. The minimisation of the sensitivity function is
often an objective of a successful control design, and in the last section of this chapter we
try to indicate why this might be the case.

8.6

8.7

8 Performance of control systems

22/10/2004

8.5.1

Basic properties of the sensitivity function . . . . . . . . . . . . . . . . . . . . . . 343

8.5.2

Quantitative performance measures . . . . . . . . . . . . . . . . . . . . . . . . . . 344

Frequency-domain performance specifications . . . . . . . . . . . . . . . . . . . . . . . . . 346

8.6.1

Natural frequency-domain specifications . . . . . . . . . . . . . . . . . . . . . . . . 346

8.6.2

Turning time-domain specifications into frequency-domain specifications

. . . . . 350

Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 350

8.1 Time-domain performance specifications

We begin with describing what is a common manner for specifying the desired behaviour
of a control system. The idea is essentially that the system is to be given a step input, and
the response is to have various specified properties. You will recall in Proposition 3.40, it
was shown how to determine the step response by solving an initial value problem. In this
chapter we will be producing many step responses, and in doing so we have used this initial
value problem method. In any case, when one gives a BIBO stable input/output system a
step input, it will, after some period of transient response, settle down to some steady-state
value. With this in mind, we make some definitions.
8.1 Definition Let (N, D) be a BIBO stable SISO linear system in input/output form, let
1N,D (t) be the step response, and suppose that limt 1N,D (t) = 1ss R. (N, D) is steppable provided that 1ss 6= 0. For a steppable system (N, D), let y(t) be the response to
the reference r(t) = 1(t)(1ss )1 so that limt y(t) = 1. We call y(t) the normalised step
response.
(i) The rise time is defined by

tr = sup y(t) t for all t [0, ] .

Contents
8.1

Time-domain performance specifications . . . . . . . . . . . . . . . . . . . . . . . . . . . . 318

8.2

Performance for some classes of transfer functions . . . . . . . . . . . . . . . . . . . . . . 320

8.3

8.2.1

Simple first-order systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 320

8.2.2

Simple second-order systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 321

8.2.3

The addition of zeros and more poles to second-order systems . . . . . . . . . . . 326

8.2.4

Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 327

Steady-state error . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 328

8.3.1

System type for SISO linear system in input/output form . . . . . . . . . . . . . . 329

8.3.2

System type for unity feedback closed-loop systems . . . . . . . . . . . . . . . . . 333

8.3.3

Error indices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 336

8.3.4

The internal model principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 336

8.4

Disturbance rejection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 336

8.5

The sensitivity function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 342

(ii) For (0, 1) the -settling time is defined by

ts, = inf |y(t) 1| < for all t [, ) .

(iii) The maximum overshoot is defined by yos = supt {y(t) 1} and the maximum
percentage overshoot is defined by Pos = yos 100%.
(iv) The maximum undershoot is defined by yus = supt {y(t)} and the maximum
percentage undershoot is defined by Pus = yus 100%.

For the most part, these definitions are self-explanatory once you parse the symbolism. A
possible exception is the definition of the rise time. The definition we provide is not the usual
one. The idea is that rise time measures how quickly the system reaches its steady-state
value. A more intuitive way to measure rise time would be to record the smallest time at
which the output reaches a certain percentage (say 90%) of its steady-state value. However,
the definition we provide still gives a measure of the time to reach the steady-state value,
and has the advantage of allowing us to state some useful results. In any event, a typical
step response is shown in Figure 8.1 with the relevant quantities labelled.
Notice that in Definition 8.1 we introduce the notion of a steppable input/output system.
The following result gives an easy test for when a system is steppable.

8.1 Time-domain performance specifications

22/10/2004

319

320

8 Performance of control systems

22/10/2004

8.2 Performance for some classes of transfer functions

yos

It is quite natural to specify performance criteria in the time-domain. However, in order to

perform design, one needs to see how system parameters affect the time-domain performance,
and how various performance measures get reflected in the other representations of control
systemsthe Laplace transform domain and the frequency-domain. To get a feel for this,
in this section we look at some concrete classes of transfer functions.

y(t)

ts,

8.2.1 Simple first-order systems The issues surrounding performance of control systems can be difficult to grasp, so our approach will be to begin at the beginning, describing
what happens with simple systems, then moving on to talk about things more general. In
this, the first of the three sections devoted to rather specific transfer functions, we will be
considering the situation when the transfer function

yus

Figure 8.1 Performance parameters in the time-domain

TL (s) =

8.2 Proposition A BIBO stable SISO linear system (N, D) in input/output form is steppable
if and only if lims0 TN,D (s) 6= 0. Furthermore, in this case limt 1N,D (t) = TN,D (0).
Proof By the Final Value Theorem, Proposition E.9(ii), the limiting value for the step
response is
lim 1N,D (t) = lim s1N,D (s) = lim TN,D (s).
s0

has a certain form, without giving consideration to the precise form of the loop gain RL .
The simplest case one can deal with is one where the system transfer function is first order.
If we normalise the transfer function so that it will have a steady-state response of 1 to a
unit step input, the transfer functions under consideration look like
T (s) =

Therefore limt 1N,D (t) 6= 0 if and only if lims0 1N,D (s) 6= 0, as specified.

So it is quite simple to identify systems which are not steppable. The idea here is quite
simple. When lims0 TN,D (s) = 0 this means that the input only appears in the differential
equation after being differentiated at least once. For a step input, this means that the righthand side of the differential equation forming the initial value problem is zero, and since
(N, D) is BIBO stable, the response must decay to zero. The following examples illustrate
both sides of the story.
8.3 Examples 1. We first take (N (s), D(s)) = (1, s2 +2s+2). By Proposition 3.40 the initial
value problem to solve for the step response is
1N,D (t) + 2 1 N,D (t) + 2 1N,D (t) = 1,

1
,
s + 1

and so essentially depend upon a single parameter . The step response of such a system is
depicted in Figure 8.2. The parameter is exactly the rise time for a first-order system, at

y(t)

RL (s)
1 + RL (s)

y(0) = 0, y(0)

= 0.

The step response here is computed to be

1N,D (t) =

1
2

12 et (cos t + sin t).

Note that limt 1N,D (t) = 12 , and so the system is steppable. Therefore, we can compute
the normalised step response, and it is
y(t) = 1 et (cos t + sin t).

t=
t

Figure 8.2 Step-response for typical first-order system

2. Next we take (N (s), D(s)) = (s, s + s2 + 2). Again by Proposition 3.40, the initial value
problem to be solved is
1N,D (t) + 2 1 N,D (t) + 2 1N,D (t) = 0,

y(0) = 0, y(0)

= 1.

The solution is 1N,D (t) = et sin t which decays to zero, so the system is not steppable.

This is consistent with Proposition 8.2 since lims0 TN,D (s) = 0.

least by our definition of rise time. Thus, by making smaller, we ensure the system will
more quickly reach its steady-state value.
Lets see how this is reflected in the behaviour of the poles of the transfer function, and in
the Bode plot for the transfer function. The behaviour of the poles in this simple first-order
case is straightforward. There is one pole at s = 1 . Thus making smaller moves this

8.2 Performance for some classes of transfer functions

22/10/2004

321

pole further from the origin, and further into the negative complex plane. Thus the issues in
designing a first-order transfer function are one: move the pole as far to the left as possible.
Of course, not many transfer functions that one will obtain are first-order.
The Bode plot for the transfer function T is very simple, of course: it has a single break
frequency at = 1 . A typical such plot is shown in Figure 8.3. Obviously, the greater

322

8 Performance of control systems

22/10/2004

which depends on the parameters and 0 . If we are interested in BIBO stable transfer
functions, the Routh/Hurwitz criterion, and Example 5.35 in particular, then we can without
loss of generality suppose that both and 0 are positive.
The nature of the closed form expression for the step response depends on the value of
. In any case, using Proposition 3.40(ii), we ascertain that the step response is given by

( 2 1)0 t

2
e(+ 1)0 t , > 1

1 + 2 1 1 e
1

y(t) = 1 e0 t (1 + 0 t),
=1

p
p

1 e0 t cos( 1 2 0 t) + sin( 1 2 0 t) ,
< 1.
1 2

This is plotted in Figure 8.4 for various values of . Note that for large there is no
= 0.2
= 0.4
= 0.7
1

y(t)

log

Figure 8.3 Bode plot for typical first-order system

values of break frequency correspond to quicker response times. Thus, for simple first-order
systems, the name of the game can be seen as pushing the break frequency as far as possible
to the right. This is, in fact, a general strategy, and we begin an exploration of this by
stating an obvious result.
i

8.4 Proposition The magnitude of T (s) at s =

Figure 8.4 Step response for typical second-order system for various values of

1 .
2

Proof This is a simple calculation since

whose magnitude we readily compute to be

overshoot, but that as decreases, the response develops an overshoot. Let us make some
precise statements about the character of the step response of a second-order system.

1
,
i+1

1 .
2

We also compute 20 log 12 3.01dB. Thus the magnitude of T ( i ) is approximately 3dB

which one can easily pick off from a Bode plot. We call 1 the bandwidth for the first-order
transfer function. Note that for first-order systems, larger bandwidth translates to better
performance, at least in terms of rise time.

8.5 Proposition Let y(t) be the step response for the SISO linear system (N, D) with transfer
function T,0 . The following statements hold:
(i) when 1 there is no overshoot;

2
(ii) when < 1 there is overshoot, and it has magnitude yos = e/ 1 and occurs at

time tos = 2 ;
1 0

(iii) if < 1 the rise time satisfies

8.2.2 Simple second-order systems First-order systems are not capable of sustaining
much rich behaviour, so lets see how second-order systems look. After again requiring a
transfer function which produces a steady-state of 1 to a unit step input, the typical secondorder transfer function we look at is
T,0 (s) =

02
,
s2 + 20 s + 02

(e0 tr

1 2) +

1 2 cos(0

1 2 tr ) + (0 tr + ) sin(0

1 2 tr ) =
p
tr 1 2 .

0 tr

Proof (i) and (ii) The proof here consists of differentiating the step response and setting it
to zero to obtain those times at which it is zero. When
p 1 the derivative is zero only when
t = 0. When < 1 the derivative is zero when 0 1 2 t = n, n Z+ . The smallest

22/10/2004

8.2 Performance for some classes of transfer functions

323

324

8 Performance of control systems

positive time will be the time of maximum overshoot, so we take n = 1, and from this the
stated formulae follow.
(iii) The rise time tr is the smallest positive time which satisfies
y(t
0) =

PSfrag replacements

y(t0 )

tos 0

t t0
t
ke(t)k
ke(t)k
e1 (t)
e1 (t)
since
e2 (t) slope of the line through the points (0, 0) and
e2 (t) the slope at the rise time should equal the
e3 (t)
e3 (t)
(t
has the form stated in the proposition.
0 , y(t0 )). After some manipulation this relation
y(t)
y(t)
u(t)
u(t)
hIn
(t)
h
(t)
Figure
8.5
we
plot
the
maximum
overshoot
and
the time at which it occurs as a function of

hN,D (t)
hN,D (t)
1 (t)
1 (t)
1N,D (t) 40
1N,D (t)
1
(t)
(t)
(t)
(t) 35
fj (t)
fj (t)
Re
Re 0.8
Im
Im 30
x1
x1
x2
x2 25
x1 0.6
x1
x2
x2
log 20
log
dB
dB
deg
deg
0.4
u = 0 ln 15
u = 0 ln
ln coth(|u| /2)
ln coth(|u| /2)
or 1 10
or 1
m
m
0.2

yos
5

tos 0

0.2
0.4
0.6
0.8
1
0.2
0.4
0.6
0.8
1,0
,0

0
0
yos

Figure 8.5 Maximum overshoot (left) and time of maximum overshoot (right) as functions of

= 0.1

PSfrag replacements

22/10/2004

= 0.5

10
5

= 0.9 and
0.25

9
50

0.5

0.75

1
0

1.25

1.5

1.75

Figure 8.6 Behaviour of rise time versus 0 for various < 1

dictates that we should have poles with nonzero imaginary part, so we consider the situation
when < 1. In this case, the poles are located as in Figure 8.7. Thus the poles lie on a circle

1 2 0

0
0

. Note that the overshoot decreases as increases, and that the time of overshoot increases
as increases.
The rise time tr is not easily understood via an analytical formula. Nevertheless, we may
numerically determine the rise time for various and 0 by solving the equation in part (iii) of
Proposition 8.5 and the results are shown in Figure 8.6. In that same figure we also plot the
curve tr = 59 0 , and we see that it is a very good approximationindeed it is indistinguishable
in our plot from the curve for = 0.9. Thus, for second-order transfer functions T,0 (s) with
< 1, it is fair to use tr 95 0 . Of course, for more general transfer functions, even more
general second-order transfer functions (say, with nontrivial numerators), this relationship
can no longer be expected to hold.
From the above discussion we see that overshoot is controlled by the damping factor
and rise time is essentially controlled by the natural frequency 0 . However, we note that the
time for maximum overshoot, tos , depends only upon , and indeed increases as increases.
Thus, there is a possible tradeoff to make when selecting if one chooses a small in the
-settling time specification. A commonly used rule of
thumb is that one should choose
= 12 which leads to yos = e 0.043 and tos 0 = 2 4.44.
Lets look now to see how the poles of T,p
0 depend upon the parameters and 0 . One
readily determines that the poles are 0 ( 1 2 ). As we have just seen, good response

Figure 8.7 Pole locations for T,0 when 0 < < 1

of radius 0 , and the angle they make with the imaginary axis is sin1 . Our rule of thumb
of taking = 12 then specifies that the poles lie on the rays emanating from the origin into
C at an angle of 45 from the imaginary axis.
Finally, for second-order systems, lets see how the parameters and 0 affect the Bode
plot for the system. In Exercise E4.6 the essential character of the frequency response for
T,0 is investigated, and let us just record the outcome of this.
8.6 Proposition If H,0 () = T,0 (i), then the following statements hold:

02
;
(i) |H,0 ()| = p 2
2
(0 )2 + 4 2 02 2
20
(ii) ]H,0 () = arctan 2
;
0 2
1
(iii) for < 2 , the function 7 |H,0 ()| has a maximum of
p
m = 0 1 2 2 ;
p
1 2 2
(iv) ]H,0 (m ) = arctan
.

1
2

1 2

325

8 Performance of control systems

22/10/2004

2.4
2.2
2
1.8

,0
0

8.2 Performance for some classes of transfer functions

22/10/2004

1N,D (t)
(t)
(t)
fj (t)
Re
Im
x1
326
x2
x1
x2
log
dB
deg
u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

at the frequency

In Figure 8.8 we label the typical points on the Bode plot for the transfer function T,0

1.6
1.4
1.2
1

0.1

0.2

0.3

0.4

0.5

0.6

0.7

Figure 8.9 Dependence of bandwidth on for second-order systems

|H,0 (m )|

detail

8.2.3 The addition of zeros and more poles to second-order systems In our general
buildup, the next thing we look at is the effect of adding to a second-order transfer function
either a zero, i.e., making a numerator which has a root, or an additional pole. The idea is
that we will investigate the effect that these have on the nature of the second-order response.
This is carried out by Jr. [1949].

Adding a zero To investigate what happens to the time signal when we add a zero to a
second-order system with a zero placed at 0 . If we normalise the transfer function to
that it has unit value at s = 0, we get

log

Figure 8.8 Typical Bode plot for second-order system with <

1
2

when < 12 . As we decrease the peak becomes larger, and shifted to the left. The phase,
as we decrease , tends to 90 at the peak frequency m .
Based on the discussion with first-order systems, for a second-order system with transfer
function T,0 , the bandwidth is that frequency ,0 > 0 for which |T,0 (i,0 )| = 12 . The
following result gives an explicit expression for bandwidth of second-order transfer functions.
Its proof is via direct calculation.
8.7 Proposition When <

1
2

the bandwidth for T,0 satisfies

p
,0
= 1 2 2 + 2(1 2 2 ) + 4 4 .
0
Thus we see that the bandwidth is directly proportional to the natural frequency 0 . The
dependence on is shown in Figure 8.9. Thus we duplicate our observation for first-order
systems that one should maximise the bandwidth to minimise the rise time. This is one of
the general themes in control synthesis, namely that, all other things being constant, one
should maximise bandwidth.

T (s) =

s + 0
02
.
0 s2 + 20 s + 02

For concreteness we take = 12 and 0 = 1. The step responses and magnitude Bode plots
are shown in Figure 8.10 for various values of . We note a couple of things.
8.8 Remarks 1. The addition of a zero increases the overshoot for < 3, and dramatically
so for < 1.
2. If the added zero is nonminimum phase, i.e., when < 0, the step response exhibits
undershoot. Thus nonminimum phase systems have this property of reacting in a manner
contrary to what we want, at least initially. This phenomenon will be explored further
in Section 9.1.
3. The magnitude Bode plot is the same for = 12 as it is for = 21 . Where the Bode
plots will differ is in the phase, as in the former case, the system is nonminimum phase.
4. When comparing the step response and the magnitude Bode plots for positive s, one
sees that the general tendency of larger bandwidths1 to produce shorter rise times is
preserved.

1
We have not yet defined bandwidth for general transfer functions, although it is clear how to do so. The
bandwidth, roughly, is the smallest frequency above which the magnitude of the frequency response remains
below 12 times its zero frequency value. This is made precise in Definition 8.25.

8.2 Performance for some classes of transfer functions

22/10/2004

327

328

8 Performance of control systems

3
1.2
PSF1Sigma(t)

PSF1Sigma(t)

= 10

= 0.5
=1
=2

2
1

=0
0
-1

1 =2
0.8
0.6

= 0.5

0.4

0.2

= 0.5
2

22/10/2004

8
PSFt

0
0

-20

-40

=2
-60
-80

-1

-0.5

0.5

-60
-80
-100

=0
-1.5

-40

= 21

-20

-120
1.5

log

Figure 8.10 The effect of an added zero for = 12 , 0, 12 , 1, 2

Adding a pole Now we look at the effect of an additional pole at 0 . The normalised
transfer function is
03
.
T (s) =
(s + 0 )(s2 + 20 s + 02 )

-1.5

-1

-0.5

0.5

1.5

log

Figure 8.11 The effect of an added pole for = 21 , 1, 2, 10

2. In terms of poles, larger bandwidth sometimes means closed-loop poles that are far
from the imaginary axis.
3. Large overshoot can arise when the Bode plot exhibits a large peak at some frequency.
This is readily seen for second-order systems, but can happen for other systems.

We once again fix = 21 and 0 = 1, and plot the step response for varying in Figure 8.11.
We make the following observations.

4. In terms of poles of the closed-loop transfer function, large overshoot can arise when
poles are close to the imaginary axis, as compared to their distance from the real axis.

8.9 Remarks
1. If a pole is added with < 3, the rise time will be dramatically increased.
This is also reflected in the bandwidth of the system increasing with .

5. Zeros of the closed-loop transfer function lying in C+ can lead to undershoot in the
step response, this having a deleterious effect on the systems performance.

2. The larger bandwidths in this case are accompanied by a more pronounced peak in
the Bode plot. As with second-order systems where this is a consequence of a smaller
damping factor, we see that there is more overshoot.

These rough guidelines can be useful in predicting the behaviour of a system based upon
the location of its poles, or on the shape of its frequency response. The former connection
forms the basis for root-locus design which is covered in Chapter 11. The frequency response
ideas we shall make much use of, as they form the basis for the design methodology of
Chapters 12 and 15. It is existence of the rigorous mathematical ideas for control design in
Chapter 15 that motivate the use of frequency response methods in design.

8.2.4 Summary This section has been something of a mixed bag of examples and
informal observations. We do not try to make it more than that at this point. Some of the
things covered here have a more general and rigorous treatment in Chapter 9. However,
it is worth summarising the gist of what we have said in an informal way. These are not
theorems. . .
1. Increased bandwidth can mean shorter rise times.

8.3 Steady-state error

An important consideration is that the difference between the reference signal and the
output should be as small a possible. When we studied the PID controller in Section 6.5

8.3 Steady-state error

22/10/2004

329

we noticed that with an integrator it was possible to make at least certain types of transfer
function have no steady-state error. Here we look at this in a slightly more systematic
manner. The first few subsections deal with descriptive matters.
8.3.1 System type for SISO linear system in input/output form For a SISO system
(N, D) in input/output form, a controlled output is a pair (r(t), y(t)) defined for t [0, )
with the property that

d
d
y(t) = N dt
r(t)
D dt
The error for a controlled output (r(t), y(t)) is e(t) = r(t) y(t), and the steady-state
error is
lim (r(t) y(t)),
t

and we allow the possibility that this limit may not exist. With this language, we have the
following definition of system type.
8.10 Definition Let k 0. A signal r U defined on [0, ) is of type k if r(t) = Ctk for
some C > 0. A BIBO stable SISO linear system (N, D) is of type k if limt (r(t) y(t))
exists and is nonzero for every controlled output (r(t), y(t)) with r(t) of type k.

In the definition of the type of a SISO linear system in input/output form, it is essential
that the system be BIBO stable, i.e., that the roots of D all lie in C . If this is not the case,
then one might expect the output to grow exponentially, and so error bounds like those in
the definition are not possible.
The following result attempts to flush out all the implications of system type.

330

8 Performance of control systems

22/10/2004

Now let (r(t), y(t)) be a controlled output with r(t) be an arbitrary signal of type k.
Note that we must have r(t)
r(t) for
yp (t) as a
= A
some A > 0. We may take yp (t) = A
d
d
y(t) = N dt
r(t) by linearity of the differential equation. This
particular solution to D dt
means that we must have
y(t) = c1 y1 (t) + + cn yn (t) + A
yp (t)
for some c1 , . . . , cn R. Thus we have

lim r(t) y(t) = lim A
r(t) c1 y1 (t) cn yn (t) A
yp (t)
t
t

= lim A r(t) yp (t) = AK.
t

This completes this part of the proof.

k
(ii) (iii) Let (r(t), y(t)) be a controlled output with r(t) = tk! , and suppose that
1
(1)
(n1)
y(0) = 0, y (0) = 0 . . . , y
(0) = 0. Note that r(s)
= sk+1 . Taking the Laplace transd
d
y(t) = N dt
r(t) gives D(s)
y (s) = N (s)
r(s). By
form of the differential equation D dt
Proposition E.9(ii) we have

lim r(t) y(t) = lim s r(s) y(s)
t
s0

= lim s r(s) TN,D (s)
r(s)
s0

= lim s
s0

1 TN,D (s)
sk+1

from which we ascertain that

8.11 Proposition Let (N, D) be a BIBO stable SISO linear system in input/output form. The
following statements are equivalent:
(i) (N, D) is of type k;
(ii) limt (r(t) y(t)) exists and is nonzero for some controlled output (r(t), y(t)) with
r(t) of type k.

1
(iii) lim k 1 TN,D (s) exists and is nonzero.
s0 s
The preceding three equivalent statements imply the following:
(iv) limt (r(t) y(t)) = 0 for every controlled output (r(t), y(t)) with r(t) of type ` with
` < k.
Proof (i) (ii) That (i) implies (ii) is clear. To show the converse, suppose that
limt (
r(t) y(t)) = K for some nonzero K and for some controlled output (
r(t), y(t))
with r(t) of type k. Supposethat deg(D) = n and let y1 (t), . . . , yn (t) be the n linearly
d
d
y(t) = 0. If yp (t) is a particular solution to D dt
y(t) = r(t)
independent solutions to D dt
we must have
y(t) = c1 y1 (t) + + cn yn (t) + yp (t)

(8.1)

From this the result clearly follows.

(iii) = (iv) Suppose that
1 TN,D (s)
=K
sk
for some nonzero constant K and let ` {0, 1, . . . , k 1}. Let (r(t), y(t)) be a controlled
output with r(t) a signal of type `. Since the roots of D are in C , we can without loss of
generality suppose that y(0) = 0, y (1) (0) = 0, . . . , y (n1) (0) = 0. We then have

1 TN,D (s)
lim r(t) y(t) = lim
s0
s`
k` 1 TN,D (s)
= lim s
s0
sk
= K lim sk` = 0.

This completes the proof.

Let us examine the consequences of this result by making a few observations.

for some c1 , . . . , cn R. By hypothesis we then have

lim r(t) c1 y1 (t) cn yn (t) yp (t) = lim r(t) yp (t) = K.

1 TN,D (s)
.
lim r(t) y(t) = lim
s0
sk

Here we have used the fact that the roots of D are in C so the solutions y1 (t), . . . , yn (t) all
decay to zero.

8.12 Remarks 1. Although we state the definition for systems in input/output form, it obviously applies to SISO linear systems and to interconnected SISO linear systems since
these give rise to systems in input/output form after simplification of their transfer functions.

respectively, we may ascertain using Proposition 3.40, that the step, ramp, and parabolic
responses are
y1 (t) =

1
2

+ 12 e2t et , y2 (t) = 41 (2t 3) 41 e2t + et ,

y3 (t) = 14 (2t2 6t + 7) + 41 e2t 2et ,

and the errors, ei (t) = ri (t) yi (t), i = 1, 2, 3, are plotted in Figure 8.12. Notice that
the step error response has a nonzero limit as t , but that the ramp and parabolic
responses grow without limit. This is what we expect from a type 0 system.
2. We take
2
TN,D (s) = 2
s + 3s + 2
which has type 1, using Proposition 8.11(iii). The step, ramp, and parabolic responses
are
y1 (t) = 1 + e2t 2et , y2 (t) = 21 (2t 3) 12 e2t + 2et ,
y3 (t) = 12 (2t2 6t + 7) + 21 e2t 4et ,
and the errors are shown in Figure 8.13. Since the system is type 1, the step response
gives zero steady-state error and the ramp response has constant steady-state error. The
parabolic input gives a linearly growing steady-state error.
3. The last example we look at is that with transfer function
TN,D (s) =

3s + 2
.
s2 + 3s + 2

,0
0

PSfrag replacements

ke(t)k
e2 (t)
e3 (t)
y(t)
u(t)
h (t)
hN,D (t)
1 (t)
1N,D (t)
(t)
(t)
fj (t)
Re
Im
x1
x2
x1
x2
log
1
dB
deg 0.8
u = 0 ln
ln coth(|u| /2)
0.6
or 1
m
0.4
yos
0.2
tos 0

,0
0

22/10/2004

e2 (t)

y(t)
u(t)
h (t)
hN,D (t)
1 (t)
1N,D (t)
(t)
(t)
fj (t)
Re
Im
2
4 x
1
x2 t
x1
x2
log
dB
deg
u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

systems

10,0

70
60
50
40

PSfrag replacements

ke(t)k
e1 (t)

e3 (t)
y(t)
u(t)
h
(t)

2
4
6
8
,0
hN,D (t)t
0
1 (t)
1N,D (t)
(t)
Figure 8.12 Step (top left), ramp
(t)(top right), and
fj (t)
tom) error responses for a system
of type 0
Re
PSfrag replacements
Im
x1
ke(t)k
x2
e1 (t)
x
1
2
e2 (t)
x2
log
1.75
y(t)
dB
u(t)
deg 1.5
h (t)
u
=

1.25
0 ln
hN,D (t)
ln coth(|u| /2)
1 (t)
1
1

1N,D (t)
m 0.75
(t)

(t)
0.5
yos
fj (t)
0.25
Re
tos 0

Im
2
2
4 x
6
8
10,0
1
0
x2 t
x1 30
x2
log 25
dB
deg 20
u = 0 ln
ln coth(|u| /2) 15
or 1
m
10
yos
5

tos 0

2
4
6
8
,0
t
0
10

parabolic (bot-

e2 (t)

8.13 Examples In each of these examples we look at a transfer function, decide what is its
type, and plot its response to inputs of various types.
1. We take
1
.
TN,D (s) = 2
s + 3s + 2
This transfer function is type 0, as may be determine by checking the limit of part (iii)
of Proposition 8.11. For a step reference, ramp reference, and parabolic reference,
(
1, t 0
r1 (t) =
0, otherwise
(
t, t 0
r2 (t) =
0, otherwise
(
t2 , t 0
r3 (t) =
0, otherwise,

ke(t)k
e1 (t)
e2 (t)

e3 (t)

Lets see how this plays out for some examples.

8
PSfrag replacements

y(t)
u(t)
h (t)
hN,D (t)
1 (t)
1N,D (t)
(t)
(t)
fj (t)
Performance of control
Re
Im
x1
x2 6
x1
x2
log 5
dB
deg 4
u = 0 ln
ln coth(|u| /2) 3
or 1
m
2
yos
1
tos 0

e3 (t)

2. The idea is that a system of type k can track up to a constant error a reference signal
which is a polynomial of degree k. Thus, for example, a system of type 0 can track a
step input up to a constant error. A system of type 1 can track a ramp input up to a
constant error, and can exactly track a step input for large time.

e1 (t)

331

e1 (t)

8.3 Steady-state error

22/10/2004

y(t)
u(t)
h (t)
hN,D (t)
1 (t)
1N,D (t)
(t)
(t)
fj (t)
332
Re
Im
x1
x2
x1
x2
log
1
dB
deg 0.8
u = 0 ln
ln coth(|u| /2)
0.6
or 1
m
0.4
yos
0.2
tos 0

Figure 8.13 Step (top left), ramp (top right), and parabolic (bottom) responses for a system of type 1

PSfrag replacements

ke(t)k
e1 (t)
e2 (t)
y(t)
u(t)
h (t)
hN,D (t)
1 (t)
1N,D (t)
(t)
(t)
fj (t)
Re
Im
x
2
4 1
x2
x1
x2
log
dB
deg
u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

error

333

334

8 Performance of control systems

22/10/2004

on the transfer function RL to determine the type of the closed-loop transfer function
TL (s) =

RL (s)
.
1 + RL (s)

These conditions are as follows.

8.14 Proposition Let RL R(s) be proper and define
2

,0
0

TL (s) =

1.4
1.2

RL (s)
.
1 + RL (s)

If (N, D) denotes the c.f.r. of TL , then (N, D) is of type k > 0 if and only if lims0 sk RL (s)
exists and is nonzero. (N, D) is of type 0 if and only if lims0 RL (s) exists and is not equal
to 1.

e3 (t)

,0
0

hN,D (t)
1 (t)
1N,D (t)
(t)
(t)
fj (t)
Re
Im
8.3 Steady-state
x1
x2
x1
x2
log 0.25
dB
deg
0.2
u = 0 ln
ln coth(|u| /2)
0.15
1
or
m
0.1

yos
0.05
tos 0

8
10
e2 (t)

e1 (t)

hN,D (t)
1 (t)
1N,D (t)
(t)
(t)
fj (t)
Re
Im
22/10/2004
x1
x2
x1
x2
1
log
dB
deg 0.8
u = 0 ln
0.6
ln coth(|u| /2)
0.4
1
or
m
0.2

0
yos
-0.2
tos 0

0.8
0.6
0.4

Proof We compute

0.2
2

,0
0

1 TL (s) =

1
.
1 + RL (s)

Thus, by Proposition 8.11(iii), (N, D) is of type k if and only if

Figure 8.14 Step (top left), ramp (top right), and parabolic (bottom) responses for a system of type 2

lim

This transfer function is type 2. The step, ramp, and parabolic responses are
y1 (t) = 1 2e2t + et ,

y2 (t) = t + e2t et ,

y3 (t) = t2 1 e2t + 2e2t .

The errors are plotted in Figure 8.14. Note that the step and ramp steady-state errors
shrink to zero, but the parabolic response has a constant error.

exists and is nonzero. For k > 0 we have

lim

1
1
=
,
sk (1 + RL (s))
lims0 sk RL (s)

and the result follows directly in this case. For k = 0 we have

lim

8.3.2 System type for unity feedback closed-loop systems To see how the steadystate error is reflected in a simple closed-loop setting, let us look at the situation depicted
originally in Figure 6.25, and reproduced in Figure 8.15. Thus we are not thinking here

1
sk (1 + RL (s))

1
1
=
.
sk (1 + RL (s))
1 + lims0 RL (s)

Thus, provided that RL (0) 6= 1 as hypothesised, the system is of type 0 if and only if
lims0 RL (s) exists.

The situation here, then, is quite simple. If RL (s) is proper, for some k 0 we can write

r(s)

RL (s)

y(s)

Figure 8.15 Unity gain feedback loop for investigating steady-state

error

so much about having a controller and a plant, but as combining these to get the transfer
function RL which is the loop gain in this case. In any event, we may directly give conditions

RL (s) =

NL (s)
sk DL (s)

with DL monic, DL and NL coprime, and DL (0) 6= 0. Thus we factor from the denominator
as many factors of s as we can. Each such factor is an integrator. The situation is depicted
in Figure 8.16. One can see, for example, why often the implementation of a PID control
law (with integration as part of the implementation) will give a type 1 closed-loop system.
To be precise, we can state the following.
8.15 Corollary For the unity gain feedback loop of Figure 8.15, the closed-loop system is of
type k > 0 if and only if there exists R R(s) with the properties
(i) R(0) 6= 0 and
(ii) RL (s) = s1k R(s).

8.3 Steady-state error

22/10/2004
1
s

r(s)

1
s

...

335

336

22/10/2004

(ii) The idea here is the same as that in the previous part of the result except that we
have
sk
1
e(s) = (1 TL (s))
r(s) = k
,
s + sk RL (s) s

y(s)

R(s)

8 Performance of control systems

with k 2, and so lims0 e(s) = 0.

Figure 8.16 A unity feedback loop with k integrators and R(0) 6= 0

The essential point is that the proposition will hold for any loop gain RL of type 1 (for
part (i)) or type 2 (for (ii)). An interesting consequence of the second part of the proposition
is the following.

Furthermore, if RL is the product of a plant transfer function RP with a controller

transfer function RC (s) = K(1 + TD s + T1I s ), then the closed-loop system will be of type 1
provided that lims0 RP (s) exists and is nonzero.

8.17 Corollary Let y(t) be the normalised step response for the closed-loop system depicted
in Figure 8.15. If the closed-loop transfer system is type k for k 2, then y(t) exhibits
overshoot.

Proof We need only prove the second part of the corollary as the first is a direct consequence
of Proposition 8.14. The closed-loop transfer function TL satisfies

Proof Since the error starts at e(0) = 1, in order that

Z
e(t) dt = 0,

1 TL (s) =

1
1 + K(1 + TD s +

,
1
)RP (s)
TI s

the error must at some time be negative. However, negative error means overshoot.

from which we determine that

1
1 TL (s)
=
s0
s
lims0 K(s + TD s2 +
lim

1
)RP (s)
TI

This issue of determining the behaviour of a closed-loop system which depends only
on properties of the loop gain is given further attention in Section 9.1. Here the effect of
unstable poles and nonminimum phase zeros for the plant is flushed out in a general setting.

and from this the result follows, since if lims0 RP (s) exists and is nonzero, then
lims0 s` RP (s) = 0 for ` > 0.

finish

Interestingly, there is more we can say about type k systems when k 1. The following
result gives more a detailed description of the steady-state error in these cases.
8.16 Proposition Let y(t) be the normalised step response for the closed-loop system depicted
in Figure 8.15. The following statements hold:
(i) if the closed-loop system is type 1 with lims0 sRL (s) = C with C a nonzero constant,
then
Z
e(t) dt = C1 ;
0

(ii) if the closed-loop system is type k with k 2 then

Z
e(t) dt = 0.
0

Proof (i) Since limt e(t) = 0, e(t) must be strictly proper. Therefore, by Proposition E.10, taking s0 = 0, we have
Z
e(t) dt = lim e(s).
0

Since
e(s) = (1 TL (s))
r(s) =
we have lims0 e(s) =

1
.
C

s
1
,
s + sRL (s) s

finish

8.3.3 Error indices

From standard texts (see Truxal, page 83).

8.3.4 The internal model principle To this point, the discussion has centred around
tracking type k signals, i.e., those that are powers of t. However, one often wishes to
track more exotic signals. The manner for doing so is suggested, upon reflection, by the
discussion to this point, and is loosely called the internal model principle.

8.4 Disturbance rejection

Our goal in this section is to do something rather similar to what we did in the last section,
except that we wish to look at the relationships between disturbances and the output. Since
a disturbance may enter a system in any of the signals, the appropriate setting here is that
of an interconnected SISO linear system (S, G). We will suppose that we are dealing with
such a system and that there are m nodes with the input being node 1 and the output being
node m.
We need a notion of output like that introduced in the previous section when talking
about transfer function types. We let i {2, . . . , m} and let (Si , Gi ) be the ith-appended
system. A j-disturbed output with input at node i is a pair (d(t), y(t)) defined on
[0, ) where y(t) is the response at node j for the input d(t) at node i for the interconnected
SISO linear system (Si , Gi ), where the input at node 1 is taken to be zero.
8.18 Definition Let (S, G) be an IBIBO stable interconnected SISO system as above and let
j {2, . . . , m}. The system is of disturbance type k at node j with input at node
i if limt y(t) exists and is nonzero for every j-disturbed output (d(t), y(t)) with input at
node i, where d(t) is of type k.

8.4 Disturbance rejection

22/10/2004

337

338

8 Performance of control systems

The idea here is very simple. We wish to allow a disturbance at any node which may
enter the system at any other node. Just where the disturbance enters a system and where
it is measured can drastically change the behaviour of the system. Typically, however, one
measures the disturbance at the output of the interconnected system. That is, typically
in the above discussion j is taken to be m. Before getting to an illustration by example,
let us provide a result which gives some obvious consequences of the notion of disturbance
type. The following result can be proved along the same lines as Proposition 8.11 and by
application of the Final Value Theorem.

the transfer function from d1 to y is

8.19 Proposition Let (S, G) be an IBIBO stable interconnected SISO linear system with input
in node 1 and output in node m. For j {2, . . . , m} the following statements are equivalent:
(i) (S, G) is of j-disturbance type k with input at node i;
(ii) limt y(t) exists and is nonzero for some j-disturbed output (d(t), y(t)) with input at
node i, where d(t) is of type k;
(iii) lims0 s1k Tji (s) exists and is nonzero.
The preceding three equivalent statements imply the following:
(iv) limt y(t) = 0 for every j-disturbed output (d(t), y(t)) with d(t) of type ` with ` < k.

and the transfer function from d3 to y is

Now let us consider a simple example.

RC (s)

// ff

RP (s)

// y(s)

Figure 8.17 A system ready to be disturbed

be considered, as shown in Figure 8.18. Note that one should strictly add another node to
d1 (s)

d3 (s)

1
RC (s)

1
RP (s)

// y(s)

Figure 8.18 Disturbances

the output node y, but this does not change anything, so it can be safely omitted. This will
generally be the case for an interconnected SISO linear system. In any case, the transfer
function for the signal flow graph is
TS,G (s) =

y(s)
RC (s)RP (s)
=
,
r(s)
1 + RC (s)(s)RP (s)

y(s)
RC (s)RP (s)
=
,
1 + RC (s)RP (s)
d1 (s)

the transfer function from d2 to y is

T2 (s) =

RP (s)
y(s)
=
,
1 + RC1 (s)RC2 (s)RP (s)
d2 (s)

T3 (s) =

y(s)
1
=
.

1
+
R
(s)R
C
P (s)
d3 (s)

Lets attach some concrete transfer functions at this point. Our transfer functions are
pretty artificial, but serve to illustrate the point. We take
1
RC (s) = ,
s
Check

8.20 Example Let us consider the interconnected SISO linear system whose signal flow graph
is shown in Figure 8.17. The system has three places where disturbances ought to naturally
r(s)

T1 (s) =

22/10/2004

RP (s) =

s
.
s+2

We compute the corresponding transfer functions to be

TS,G (s) =

s2
,
s2 + s + 2

T1 (s) =

s2
,
s2 + s + 2

T2 (s) =

s
,
s2 + s + 2

T3 (s) =

s+2
.
s2 + s + 2

A straightforward application of Proposition 8.11(iii) indicates that TS,G is of type 0. It is

also easy to see from Proposition 8.19(iii) that the system is of disturbance type 2 with
respect to the disturbance d1 , of disturbance type 1 with respect to the disturbance d2 , and
of disturbance type 0 with respect to d3 .

As we see in this example, the ability of a system to reject disturbances may differ
from the ability of a system to represent a reference signal. Furthermore, for disturbance
rejection, one typically wants the transfer function from the disturbance to the output to
have the property that the expected disturbance type gives zero steady-state error. Thus,
if step disturbances are what is expected, it is sufficient that the system be of disturbance
type 0.
Let us look at our falling mass example thinking of the gravitational force as a disturbance.
8.21 Example (Example 6.60 contd) The block diagram for the falling mass with the gravitational force is shown in Figure 8.19. We think of the input mg as a disturbance input
and so write d(t) = mg1(t). This input is then a step input if we think of holding the mass
then letting it go. In fact, in our analysis, lets agree to take y(0) = 0 and y(0)

= 0 and see
if the controller can return the mass to y = 0 after we let it fall under gravity. The transfer
function from the disturbance to the output is
Td (s) =

y(s)
ms2
=
.
2

ms + RC (s)
d(s)

Let us look at the form of this transfer function for the various controllers we employed in
Example 6.60.

8.4 Disturbance rejection

22/10/2004

339

RC (s)

1
ms2

y(s)

8 Performance of control systems

22/10/2004

1 (t)

r(s)

1N,D (t)
(t)
(t)
fj (t)
Re
Im
340
x1
x2
2
x1
x2
0
log
dB
deg -2
u = 0 ln
ln coth(|u| /2) -4
or 1
m
-6
yos
-8
tos 0

Figure 8.19 Block diagram for falling mass with gravitational force

1 (t)

ke(t)k
e1 (t)
e2 (t)
1. With proportional
control we have RC (s) = K and so
e3 (t)
ms2
y(t)
Td (s) =
.
u(t)
ms2 + K
h (t)
This transfer
function has poles on the imaginary axis, and so is not a candidate for having
hN,D (t)
its type defined. Nonetheless, we can compute the time response using Proposition 3.40
1N,D
given d(s)
=(t)
mg . We ascertain that the output to this disturbance is
(t) s
r
(t)
K
yd (t) = mg cos
t.
fj (t)
m
Re
One can see Im
that the system does not respond very nicely in this case to the step disturbance, and tox1further illustrate the point, we plot this step response in Figure 8.20 for
x2
x1
x2
10
log
dB
5
deg
u = 0 ln
ln coth(|u| /2)
0
or 1
m

-5
yos

tos 0 -10

1
2
3
4
5
,0
t
0
Figure 8.20 Response of falling mass to step disturbance with proportional control

m = 1, g = 9.81, and K = 28.

2. For derivative control we take RC (s) = KTD s which gives
ms
Td (s) =
.
ms + KTD
This disturbance transfer function is type 1 and so the steady-state disturbance to the
step input d(t) = mg is zero. Using Proposition 3.40 we determine that the response is
yd (t) = mgeKTD t/m

,0
0

Figure 8.21 Response of falling mass to step disturbance with

derivative control

PSfrag replacements

9
ke(t)k
and we plot
this response in Figure 8.21 for m = 1, g = 9.81, K = 28, and TD = 28
.
e
(t)
1
Note that this response decays to zero which is consistent with our observation that the
e2 (t)
error to a step
e3 (t) disturbance should decay to zero in the steady-state.
3. If we use any(t)
integral controller we take RC (s) = T1I s from which we compute
u(t)
h (t)
mTI s3
Td (s) =
.
hN,D (t)
mTI s3 + K

1N,Dthis
(t) transfer function for type, let us be concrete and choose m = 1, g = 9.81,
To examine
7
9

K = 28, T(t)
D = 28 , and TI = 10 . One determines (Mathematica !) that the roots of
(t)
3
mTI s + Kf are
then

j (t)
3
3
3
5 3 5i, 2 5.
Re
Im function has poles in C+ , the notion of type is not applicable. We plot
Since the transfer
x1
the response xto
the step gravitational disturbance in Figure 8.22, and we see that it is
2
x1 10000
x2
log
7500
dB 5000
deg
u = 0 ln
2500
ln coth(|u| /2)
0
1
or
m -2500

yos -5000
-7500
tos 0

1
2
3
4
5
,0
t
0
1 (t)

PSfrag replacements

Figure 8.22 Response of falling mass to step disturbance with integral control

badly behaved. Thus the given integral controller will magnify the disturbance.

1N,D (t)
(t)
(t)
fj (t)
Re
Im
x1
342
8 Performance of control systems
x2
x1
0.1
x2
log
0.05
dB
0
deg
u = 0 ln -0.05
ln coth(|u| /2)
or 1 -0.1
m -0.15

-0.2
yos
-0.25
tos 0

1
2
3
,0
t
0

PSfrag replacements

8.4 Disturbance rejection

341
ke(t)k
e1 (t)
4. Finally wee2combine
the three controllers into a nice PID control law: RC (s) = K +
(t)
The disturbance to output transfer function is then
KTD s + TKIes3.(t)
y(t)
u(t)
mTI s3
h (t)
Td (s) =
.
3
mTI s + KTD TI s2 + KTI s + K
hN,D (t)
9
7
For our 1parameters
m = 1, g = 9.81, K = 28, TD = 28
, and TI = 10
, we determine
N,D (t)
that the roots
(t)of the denominator polynomial are 5, 2 2i (recall that we had chosen
(t)
the PID parameters
in just this way). Since these roots are all in C , we may make the
(t) the disturbance type is 3. Therefore the steady-state error to a step
observationfj that
disturbance Re
should be zero. The response of this transfer function to the gravitational
Im
9
step disturbance
x1 is shown in Figure 8.23 for m = 1, g = 9.81, K = 28, TD = 28 , and
x2
x1
2
x2
log
dB 0
deg
u = 0 ln -2
ln coth(|u| /2)
or 1 -4
m
-6
yos
-8
tos 0

1
2
3
4
5
,0
t
0

1 (t)

The solution is plotted in Figure 8.24 for the chosen PID parameters. As should be the case,
the response is nice with the PID controller bringing the mass back to its initial height in
relatively short order.
This example is one which contains a lot of information, so you would be well served to
spend some time thinking about it.

8.5 The sensitivity function

In this section we introduce and study a transfer function which is, at least for certain
block diagram configurations, complementary to the transfer function. The block diagram
configuration we look at is the one we first looked at in Section 6.3.2, and the configuration
is reproduced in Figure 8.25. One may extend the discussion here to more general block

7
TI = 10
. This response decays to zero as it ought to.
Let us address the question of how to interpret our computations for the gravitational
disturbance to the falling mass. One needs to be careful not to misinterpret the disturbance
response for the system response. The input/output response of the system for the various
controllers was already investigated in Example 6.60. When examining the system including
the effects of the disturbance, one must take into account both the input/output response
and the response to the disturbance.
For example, with the PID controller we had chosen, our analysis of both these responses
suggests that the system with the chosen parameters should behave nicely with the gravitational force. To verify this, let us look at the situation where we hold the mass still at y = 0
and at time t = 0 let it go. Our PID controller is then charged with bringing the mass back
to y = 0. The initial value problem governing this situation is
Z
K t
y( ) d = mg, y(0) = 0, y(0)

= 0.
m
y + KTD y(t)
+ Ky(t) +
TI 0

e(s)

r(s)

d(s)
RL (s)

y(s)

n
(s)

Figure 8.25 Block diagram for investigating the sensitivity function

diagram configurations, but this unity gain feedback setup is one which can be handled nicely.
In this scenario, we are typically supposing that the loop gain RL is the product of a plant
transfer function RP and a controller transfer function RC . But unless we say otherwise, we
just take RL as a transfer function in its own right, and do not concern ourselves with from
where it comes.

To solve this equation we differentiate once to get the initial value problem
smooth this

y(0) = 0, y(0)

= 0, y(0) = g.

Figure 8.24 Response of falling mass with PID controller

Figure 8.23 Response of falling mass to step disturbance with PID

control

K
...
my + KTD y(t) + K y(t)
+ y(t) = 0,
TI

22/10/2004

y(t)

22/10/2004

The import of this simple result is that it holds for any loop gain RL . Thus the zeros and
poles for RL are immediately reflected in the closed-loop transfer function and the sensitivity
function. Let us introduce the following notation:
Z(SL ) = { s C | SL (s) = 0}
Z(TL ) = {s C | TL (s) = 0} .

(8.2)

This notation will be picked up again in Section 9.2.1.

Let us illustrate this with an example some of the simpler features of the sensitivity
function.
PSfrag replacements

PSfrag replacements

,0
0

150
100
50

deg

u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

0
-50
-100
-150
-1.5 -1 -0.5

log

0.5

u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

1
1.5
2
,0
0

150

22/10/2004

u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

,0
0

150
100
50
0
-50
-100
-150
-1.5 -1 -0.5

log

0.5

u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

1
1.5
2
,0
0

150
100
50
0
-50
-100
-150
-1.5 -1 -0.5

log

0.5

1.5

Figure 8.27 Bode plots for the closed-loop transfer function and
sensitivity function for the falling mass with integral controller
(left) and PID controller (right). In each case, the solid line
represents the closed-loop transfer function and the dashed line
the sensitivity function.

give the Bode plots for the transfer function (the solid line) and the sensitivity function
(the dashed line).
4. Finally, we look at a PID controller so that RC (s) = K + KTD s + T1I s , and we use the
same numerical values as above. As expected, in Figure 8.27 you will find the Bode plots
for the closed-loop transfer function and the sensitivity function.
In all cases, note that the gain attenuation at high frequencies leads to high sensitivities at
these frequencies, and all the associated disadvantages.

100
50

deg

t
8.23t Example (Example 6.60 contd) Let us
carry
on with our falling mass example. What
ke(t)k
ke(t)k
e1 (t)shall do is plot the frequency response of
e1 (t)
we
the closed-loop transfer function and the sene2 (t)
e2 (t)
sitivity
function on the same Bode plot. We
have RL (s) = RC (s) ms1 2 , and we shall use the
e3 (t)
e3 (t)
y(t) controller transfer functions of Exampley(t)
four
6.60. In all cases, we take m = 1.
u(t)
u(t)
h1.
h (t) 8.26 we give the Bode plots for the closed (t)We take RC (s) = K with K = 28. In Figure
hN,D (t)
hN,D (t)
1 (t)
1 (t)
30
10
1N,D (t)
1N,D (t)
(t)
(t)
5
20
(t)
(t)
0
10
fj (t)
fj (t)
-5
0
-10
Re
Re
Im
Im
-15
-10
x1
x1
-20
x2
x2
-20
-25
x1
x1
-1.5 -1 -0.5 0 0.5 1 1.5 2
-1.5 -1 -0.5 0 0.5 1 1.5 2
x2
x2
log
log

8 Performance of control systems

8.22 Proposition For the feedback interconnection of Figure 9.3, the following statements
hold:
(i) p C is a pole for RL if and only if SL (p) = 0 and TL (p) = 1;
(ii) z C is a zero for RL if and only if SL (z) = 1 and TL (z) = 0.

PSfrag replacements

344

t
t
ke(t)k
ke(t)k
e1 (t)
e1 (t)
e2 (t)Next we take RC (s) = KTD s with K = 28
e2 (t)and TD = 9 . In Figure 8.26 we give the Bode
2.
28
e3 (t)
e3 (t)
plots
for
the
closed-loop
transfer
function
and the sensitivity function.
y(t)
y(t)
u(t)
u(t)
1
7
Now we consider pure integral control with
h3.
h (t) RC (s) = TI s with TI = 10 . In Figure 8.26 we
(t)
hN,D (t)
hN,D (t)
1 (t)
1 (t)
10
10
1N,D (t)
1N,D (t)
(t)
(t)
5
5
(t)
(t)
0
0
fj (t)
fj (t)
-5
-5
-10
-10
Re
Re
Im
Im
-15
-15
x1
x1
-20
-20
x2
x
2
-25
-25
x1
x1
-1.5 -1 -0.5 0 0.5 1 1.5 2
-1.5 -1 -0.5 0 0.5 1 1.5 2
x2
x2
log
log

deg

8.5.1 Basic properties of the sensitivity function Let us first make some elementary
observations concerning the closed-loop transfer function and the sensitivity function. The
following result follows immediately from the definition of TL and SL .

PSfrag replacements

343

deg

8.5 The sensitivity function

22/10/2004

0
-50
-100
-150
-1.5 -1 -0.5

log

0.5

1.5

Figure 8.26 Bode plots for the closed-loop transfer function and
sensitivity function for the falling mass with proportional controller (left) and derivative controller (right). In each case, the
solid line represents the closed-loop transfer function and the
dashed line the sensitivity function.

loop transfer function (the solid line) and the sensitivity function (the dashed line).

8.5.2 Quantitative performance measures Our emphasis to this point has been essentially on descriptive measures of performance. However, it is very helpful to have on
hand quantitative measures of performance. Indeed, such performance measures form the
backbone of modern control theory, as from these precise performance characteristics spring
useful design methodologies. The reader will wish to recall from Section 5.3.1 the definitions
of the L2 and L -norms.
The quantitative measures we will provide are those on the error of a system to a step
input. Thus we work with the block diagram of Figure 6.25 which we reproduce in Figure 8.28. We shall assume that the closed-loop system is type 1 so that the steady-state
error to a step input is zero. In order to make meaningful comparisons, we deal with the
unit step response to the reference 1(t).
For design methodologies with a mathematical basis, the following two measures of per-

8.5 The sensitivity function

22/10/2004

345

346

8 Performance of control systems

22/10/2004

8.6 Frequency-domain performance specifications

r(s)

RL (s)

The time-domain performance specifications of Section 8.1 are traditionally what one
encounters as these are the easiest to achieve a feeling for. However, there are powerful
controller design methods which rely on specifying the performance requirements in the
frequency-domain, not the time-domain. In this section we address this in two ways. First
we look at some specifications that are actually most naturally given in the frequencydomain. These are followed by an explanation of how time-domain specifications can be
approximately turned into frequency-domain specifications.

y(s)

Figure 8.28 Block diagram for quantitative performance measures

8.6.1 Natural frequency-domain specifications There are a significant number of

performance specifications that are very easily presented in the frequency-domain. In this
section we list some of these. In a given design problem, one may wish to invoke only some
of these constraints and not others. Also, the exact way in which the specifications are best
presented can vary in detail, depending on the particular application. However, it can be
hoped that what we say here is a helpful guide in getting started in a given problem.
Before we begin, we remark that all of the specifications we give will be are what Helton
and Merrino [1998] call disk constraints. This means that the closed-loop transfer function
TL will be specified to satisfy a constraint of the form

formance are often the most useful:

Z
1/2
kek2 =
e2 (t) dt
0

kek = sup{|e(t)| for almost every t}.

The first is of course the L2 -norm of the error, and the second the L -norm of the error.
The error measure
Z
|e(t)| dt
kek1 =

|M (i) TL (i)| R(i),

[1 , 2 ]

is also used. For example, you will recall from Section 13.1 that the criterion used by Ziegler
and Nicols in their PID tuning was the minimisation of the L1 -norm. In practice, one will
sometimes wish to consider alternative performance measures. For example, the following
two, related to the L2 and L1 -norms, will serve to penalise errors which occur for large times,
but deemphasise large initial errors:
Z
1/2
kekt,2 =
te2 (t) dt
0
Z
t |e(t)| dt.
kekt,1 =

R+ defined?

for functions M : i[1 , 2 ] C M : i[1 , 2 ] R+ , and for 1 < 2 > 0. Thus at each
frequency in the interval [1 , 2 ], the value of TL at i must lie in the disk of radius R(i)
with centre M (i).
Stability margin lower bound Recall from Proposition 7.15 that the point on the Nyquist
contour closest to 1 + i0 is a distance kSL k away. Thus, to improve stability margins,
one should provide a lower bound for this distance. This means specifying an upper bound
on the H -norm of the sensitivity function. In terms of the transfer function this gives
|1 TL (i)| sm ,

> 0.

(8.3)

One can imagine choosing other weighting functions of time by which to multiply the error
to suit the particular nature of the problem with which one is dealing.
One of the reasons for the usefulness of the L2 and L norms is that they make it possible
to formulate statements in terms of transfer functions. For example, we have the following
result which was derived in the course of the proof of Theorem 5.22, if we recall that the
transfer function from the input to the error in Figure 8.28 is the sensitivity function.

It is possible that one may wish to enforce this constraint more or less at various frequencies.
For example, if one knows that a system will be operating in a certain frequency range, then
stability margins in this frequency range will be more important that in others. Thus one
may relax (8.3) by introducing a weighting function W and asking that

8.24 Proposition Consider the block diagram of Figure 8.28 and suppose that the closed-loop
system is type k for k 1. If e(t) is the error to an input u L2 [0, ), then

As a disk constraint this reads.

|W (i)(1 TL (i))| sm ,

|1 TL (i)|
kek2 kSL k kuk2 .
where SL is the sensitivity function of the closed-loop system.
This proposition tells us that if a design objective is to minimise the transfer of energy
from the input signal to the error signal, then one should aim to minimise the H -norm of
SL .

> 0,

sm
,
|W (i)|

> 0.

(8.4)

Tracking specifications Recall that the tracking error is minimised as in Proposition 8.24
by minimising kSL k . However, the kind of tracking error minimisation demanded by
Proposition 8.24 is extremely stringent. Indeed, it asks that for any sort of input, we
minimise the L2 -norm of the error. However, in practice one will wish to minimise the

22/10/2004

8.6 Frequency-domain performance specifications

347

348

|1 TL (i)| < Rtr ,

[1 , 2 ].

(8.5)

|(1 TL (i))
r(i)| tr ,

[1 , 2 ].

This is then made into the disk constraint

|1 TL (i)|

tr
,
|
r(i)|

[1 , 2 ].

bandwidth and
gain crossover
frequency

Thus N,D is, for typical systems, simply the frequency at which the magnitude part
of the Bode plot drops below, and stays below, the DC value minus 3dB. For most other
transfer functions, definition of bandwidth is still possible, but must be modified to suit
the characteristics of the system. Note that the bandwidth is not defined for systems with
imaginary axis poles. However, this is not an issue since such systems are not BIBO stable,
and we are typically interested in bandwidth for the closed-loop transfer function, where
BIBO stability is essential.
With is general notion of bandwidth, the typical bandwidth constraint one may pose
might take the form
|TL (i)| < bw , > bw .
(8.7)

(8.6)

Bandwidth constraints One of the more important features of a closed-loop transfer function is its bandwidth. As we saw in Section 8.2, larger bandwidth generally means quicker
response. However, one wishes to limit the bandwidth since it is often destructive to a systems physical components to have the response to high-frequency signals not be adequately
attenuated.
Let us take this opportunity to define bandwidth for fairly general systems. Motivated
by our observations for first and second-order transfer functions, we make the following
definition.
8.25 Definition Let (N, D) be a proper SISO linear system in input/output form and let

High-frequency roll-off In practise one does not want the closed-loop transfer function to
be proper, but strictly proper. This will ensure that the system will not be overly susceptible
to high-frequency disturbances in the reference. Furthermore, one will often want the closedloop transfer function to not only be proper, but to have a certain relative degree, so that it
falls off at a prescribed rate as . Note that
TL (i) =

|RC (i)|

as usual. When lim0 |HN,D ()| < , we consider the following five cases:
(i) (N, D) is strictly proper and steppable: the bandwidth of (N, D) is defined by
|HN,D ()|

N,D = inf
H (0) 12 for all >
;
| N,D |

(ii) (N, D) is not strictly proper and not steppable: the bandwidth of (N, D) is defined
by
|HN,D ()|

N,D = sup
H () 12 for all <
;
|
|
N,D

(iii) (N, D) is strictly proper, not steppable, and kTN,D k < : the lower cutoff frequency of (N, D) is defined by
|HN,D ()|

lower
= sup
T

N,D
12 for all <
k N,D k

upper
lower
and the bandwidth is given by N,D = N,D
N,D
;

RC (i)RP (i)
.
1 + RC (i)RP (i)

Therefore, assuming that RC RP is proper, for large frequencies TL (i) behaves like
RC (i)RP (i). To render the desired behaviour as a disk inequality, we first assume that
we insist on controllers that satisfy an inequality of the form

kTN,D k = sup{|HN,D ()|},

and the upper cutoff frequency of (N, D) is defined by

|HN,D ()|

upper
N,D
= inf
T
12 for all >
,
k N,D k

22/10/2004

(iv) (N, D) is strictly proper, not steppable, and kTN,D k = : the bandwidth of (N, D)
is not defined;
(v) (N, D) is not strictly proper and steppable: the bandwidth of (N, D) is not defined.
When lim0 |HN,D ()| = , bandwidth is undefined.

tracking error for inputs having a certain frequency response. To this end, we may specify
various sorts of specifications that are less restrictive than minimising kSL k .
A first case we consider is a constraint

This corresponds to tracking well signals whose frequency response is predominantly supported in the given range.
Another approach that may be taken occurs when one knows, or approximately knows,
the transfer function for the reference one wishes to track. That is, suppose that we wish
to track the reference r(t) whose Laplace transform is r(s). Generalising the analysis of
Section 8.3 in a straightforward way from type k reference signals to reference signals with
a general Laplace transform, we see that to track r well we should require

8 Performance of control systems

C
||nC

for some C > 0 and nC N. This is tantamount to making a relative degree constraint on
the controller. Since we are primarily concerned with high-frequency behaviour, the choice
of C is not critical. Assuming that such a constraint has been enforced, the high-frequency
approximation
TL (i) RC (i)RP (i)
gives rise to the inequality
|TL (i)|

C |RP (i)|
,
||nC

[1 , 2 ].

(8.8)

Controller output constraints The transfer function

R C SL =

RC
1 + RC RP

is sometimes called the closed-loop controller . It is the transfer function from the error
to the output from the plant. One would wish to minimise this transfer function in order to
ensure that the controller is not excessively aggressive. In practise, an excessively aggressive

8.6 Frequency-domain performance specifications

22/10/2004

349

350

controller might cause saturation, i.e., the controller may not physically be able to supply
the output needed. Saturation is a nonlinear effect, and should be avoided, unless its effects
are explicitly accounted for in the modelling.
In any event, the constraint we consider to limit the output of the closed-loop controller
is
|RC (i)(1 TL (i))| clc ,

|TL (i)| os ,

|TL (i)| clc |RP (i)| .

Now let us turn our attention to the reasonable ranges of frequency for the invocation of
such a constraint. Clearly the constraint will have the greatest effect for frequencies that are
zeros for RP that are on the imaginary axis. This we let iz1 , . . . , izk be the imaginary axis
zeros for RP , and take as our disk constraint
[z1 r1 , z1 + r1 ] [zk rk , zk + rk ],

Plant output constraints The idea here is quite similar to that for the closed-loop controller. To wit, the name closed-loop plant is often given to the transfer function
RP
.
1 + RC RP

finish

This is the transfer function from the output of the controller to the output of the plant.
That this should be kept from being too large is a reflection of the desire to not have the
plant overreact to the controller.
The details of the computation of the ensuing disk constraint go very much like that for
the closed-loop controller. We begin by considering the constraint
|RP (i)(1 TL (i))| clp ,
immediately giving rise to the disk constraint
|1 TL (i)|

clp
.
|RP (i)|

To determine when one should invoke this constraint, note that it most restrictive for frequencies near imaginary axis poles for RP . Thus we let p1 , . . . , pk be the imaginary axis
poles for RP and take the constraint
|1 TL (i)|

clp
,
|RP (i)|

for given r1 , . . . , rk > 0.

[p1 r1 , p1 + r1 ] [pk rk , pk + rk ],

(8.11)

(8.9)

for some r1 , . . . , rk > 0.

RP SL =

Summary We have given a list of typical frequency-domain constraints. In practise one

rarely enforce all of these simultaneously. Nevertheless, in a given application, any of the
constraints (8.3), (8.4), (8.5), (8.6), (8.8), (8.7), (8.9), (8.10), or (8.11) may prove useful.
Note that specifying the constants and various weighting functions will be a process of trial
and error in order to ensure that one specifies a problem that is solvable, and still has
satisfactory performance. In Chapter 9 we discuss in detail the idea that not all types of
frequency-domain performance specification should be expected to be achievable, particularly for unstable and/or nonminimum phase plants. The methods for obtaining controllers
satisfying the type of frequency-domain constraints we have discussed in this section may be
found in the book [Helton and Merrino 1998], where a software is demonstrated for doing
such design. This is given further discussion in Chapter 15 in terms of H methods.

TL
RC RP =
,
1 TL

this gives the disk constraint

|TL (i)| clc |RP (i)| ,

22/10/2004

Overshoot attenuation In Section 8.2.2 we saw that there is a relationship, at least for
second-order transfer functions, between a peak in the frequency response for the closed-loop
transfer function and a large overshoot. While no theorem to this effect is known to the
author, it may be desirable to enforce a constraint on kTL k of the type

where we leave the frequency unspecified for the moment. Using the relation
RC RP
TL =
1 + RC RP

8 Performance of control systems

(8.10)

8.6.2 Turning time-domain specifications into frequency-domain specifications It

is generally not possible to make a given time-domain performance specification and produce an exact, tractable frequency-domain specification. Nonetheless, one can make some
progress in producing reasonably effective and manageable frequency-domain specifications
from many types of time-domain specifications. The idea is to produce a transfer function having the desired time-domain behaviour, then using this as the basis for forming the
frequency-domain specifications.

8.7 Summary
When designing a controller the first step is typically to determine acceptable performance criterion. In this chapter, we have come up with a variety of performance measures.
Let us review what we have said.
1. Based upon a systems step response, we defined various performance features (rise time,
overshoot, etc.). These features should be understood at least in that one should be able
to compare two signals and determine which is the better with respect to certain of these
performance features.
2. Some of the character of system response are exhibited by a simple second-order transfer
function. In particular, the tradeoffs one has to make in controller design begin to show
up for such systems in that one cannot perfectly satisfy all performance measures.
3. For simple systems, often one can obtain an adequate understanding of the problems to
be encountered in system performance by using the observations seen when additional
poles and zeros are added to a second-order transfer function.
4. The effects of the existence of unstable poles and nonminimum phase zeros on the step
response should be understood.
5. System type is an easily understood concept. Particularly, one should be able to readily
determine the system type of a unity gain feedback loop with ease.

22/10/2004

8.7 Summary

351

6. For disturbance rejection, one should understand that a disturbance may affect a systems
dynamics in a variety of ways, and that the way these are quantified are via appended
systems, and systems types for these appended systems.
7. For design, the sensitivity function is an important consideration as concerns designing
controllers which are in some sense robust. The reader should be aware of some of the
tradeoffs which are necessitated by the need to have a good closed-loop response along
with desirable robustness and disturbance rejection characteristics.

352

8 Performance of control systems

22/10/2004

Exercises
E8.1 Consider the SISO linear system = (A, b, ct , D) with

0
1
A=
, b=
, c=
,

1
0

D = 01 ,

where 0 and > 0. For = 1 and {0, 0.1, 0.7, 1}, do the following.
(a) Determine the steady-state value of the output for a unit step input.
(b) For a unit step input and zero initial conditions, plot the output response (you
do not have to provide an analytical expression for it).
(c) From your plot, determine the rise time, tr .
(d) From your plot, determine the -settling time, ts, , for = 0.1.
(e) Determine the percentage overshoot Pos .
(f) Plot the location of the poles of the transfer function T in the complex plane.
(g) Produce the magnitude Bode plot for H (), and from it determine the bandwidth of the system.
E8.2 Denote by 1 the SISO linear system of Exercise E8.1. Add in series with the system
1 the first-order SISO system 2 = (A2 , b2 , ct2 , D 2 ) with

A2 = , b2 = 1 , c2 = 1 , D 2 = 01 .
That is, consider a SISO linear system having as its input the input to 2 and as its
output the output from 1 (see Exercise E2.1).
(a) Determine the transfer function for the interconnected system, and from this
ascertain the steady-state output arising from a unit step input.
Fix = 1 and = 1, and for {0, 0.1, 1, 5}, do the following.
(b) For a unit step input and zero initial conditions, plot the output response (you
do not have to provide an analytical expression for it).
(c) From your plot, determine the rise time, tr .
(d) From your plot, determine the -settling time, ts, , for = 0.1.
(e) From your plot, determine the percentage overshoot Pos .
(f) Plot the location of the poles of the transfer function in the complex plane.
(g) Produce the magnitude Bode plot for the interconnected system, and from it
determine the bandwidth of the system.
E8.3 Denote by 1 the SISO linear system of Exercise E8.1. Interconnect 1 with blocks
containing s and R as shown in the block diagram Figure E8.1. Thus the input
gets fed into the parallel block, whose output becomes the input into 1 .
(a) Determine the transfer function for the interconnected system, and from this
ascertain the steady-state output arising from a unit step input.
Fix = 1 and = 1, and for {1, 0, 1, 5}, do the following.
(b) For a unit step input and zero initial conditions, plot the output response (you
do not have to provide an analytical expression for it).
(c) From your plot, determine the rise time, tr .
(d) From your plot, determine the -settling time, ts, , for = 0.1.
(e) From your plot, determine the percentage overshoot Pos .

Exercises for Chapter 8

353

354

8 Performance of control systems

22/10/2004

Now suppose that you wish to consider the effect of a disturbance which enters the
system as in Figure E8.4.

s
u
(s)

y(s)

d(s)

r(s)

RC (s)

RP (s)

y(s)

Figure E8.1 Block diagram for system interconnection

(f) Plot the location of the poles and zeros of the transfer function in the complex
plane.
(g) Produce the magnitude Bode plot for the interconnected system, and from it
determine the bandwidth of the system.
(h) Produce the phase Bode plot for the interconnected system and make some comments on what you observe.
E8.4 Consider the closed-loop interconnection in Figure E8.2, and assume that it is IBIBO
r(s)

RC (s)

RP (s)

y(s)

Figure E8.2 A closed-loop system

stable. Suppose that the loop gain RC RP satisfies lims0 RP (s)RC (s) = K where
K 6= 1.
(a) Show that the closed-loop transfer function is type 0. If one desires zero steadystate error to a step input, will the closed-loop system be satisfactory?
C which has the property that, when put into the
(b) Design a rational function R
block diagram of Figure E8.3, the resulting closed-loop transfer function will be
RC (s)
r(s)

RP (s)

y(s)

C (s)
R

Figure E8.3 A modified closed-loop system

of type 1. For the closed-loop system you have just designed, what will be the
error of the system to a step input?

Figure E8.4 Adding a disturbance

(c) What is the system type with respect to this disturbance for the controller depicted in the block diagram Figure E8.2?
(d) What is the system type with respect to this disturbance for the modified controller depicted in the block diagram Figure E8.3 (i.e., that obtained by replacing
C and RC )?
the RC block with the parallel blocks containing your R
(e) Comment on the effectiveness of the modified controller as concerns rejection of
step disturbances in this case.
E8.5 Let RP R(s) be a strictly proper plant with RC S (RP ) a proper controller.
(a) Show that if the Nyquist plot for RL = RC RP lies in C+ then kSL k < 1. What
can you say about the performance of the resulting closed-loop system?
Now consider the two plants
RP,1 =

1
,
s+1

RP,2 =

s1
.
s2 + 2s + 1

(b) Produce the Bode plots for both plants with the controller RC (s) = 12 S (RP,1 )
S (RP,2 ). How do they differ?
(c) Produce the Nyquist plots for both loop gains RL,1 = RC RP,1 and RL,2 = RC RP,2 .
(d) Comment on the comparative performance of the closed-loop systems in light of
your work done in part (a).
(e) Produce the step response for both closed-loop systems and comment on their
comparative behaviour.
E8.6 In problems (a)(d) there are two SISO linear systems (N1 , D1 ) and (N2 , D2 ) in input/output form, but you are not told what they are. Instead, you are given a pair
of Bode plots, and a pair of step responses, one each for the pair of transfer functions
TN1 ,D1 and TN2 ,D2 . You are not told which Bode plot and which step response come
from the same transfer function.
In each case, do the following:
1. Indicate which Bode plot goes with which step response. correctly.
2. Indicate clearly (but not necessarily at great length) the features of the plots
that justify your choice in step 1.
(a)

-1

log

-2

-1

log

7.5

12.5

17.5

Step response 2

0
-50
-100
-150
-3

-2

-1

log

Bode plot 1

,0
0

1 (t)

-3

-2

-1

log

100

Step response 1

50
0

(d)

-50
-100
-150
-3

-2

-1

log

Bode plot 2

22/10/2004

1.2
1
0.8

1 (t)

deg

1 (t)

Bode plot 2

2.5

systems

0.6
0.4
0.2
0

2.5

7.5

12.5

17.5

Step response 2

0
-10
-20

-3

u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

2
3
4

1N,D (t)
1N,D (t)
(t)
(t)
(t)
(t)
fj (t)
fj (t)
Re
Re
356 Im
8 Performance of control
Im
x1
x1
x2
x2
x1
x1
x2
x
2
4
log
log
dB
dB
3
deg
deg
u = 0 ln
u = 0 ln
ln coth(|u| /2) 2
ln coth(|u| /2)
or 1
or 1
m
m 1

PSfrag replacements
PSfrag replacements
yos
yos
t
t 0
tos 0
tos 0
ke(t)k
ke(t)k

e1,
(t) -1
e
(t)
2.5
5
7.5
10 12.5 15 17.5 201,0
0
e2(t)
e2(t)
t
0
0
e3 (t)
e3 (t)
y(t)
y(t)
Step response 1
u(t)
u(t)
h (t)(c)
h (t)
hN,D (t)
hN,D (t)
1 (t)
1 (t)
1N,D (t)
1N,D (t)
0
(t)
(t)
-10
(t)
(t)
-20
fj (t)
fj (t)
Re
Re
-30
Im
Im
-40
x1
x1
x
x
-50
PSfrag replacements2
PSfrag replacements2
x1
x1
-3
-2
-1
0
1
2
3
4
x2
x2
ke(t)k
ke(t)k
log
e1 (t)
e1 (t)
e2 (t)
e2 (t)
150
e3 (t)
e3 (t)
u = 0 ln

u
=

0
y(t)
y(t)
100
ln coth(|u|u(t)
/2)
ln coth(|u|u(t)
/2)
50
1
1
orh(t)
orh(t)
0
m
m
hN,D(t)
hN,D(t)

-50
y
os
os
1N,Dy(t)
1
(t)
N,D
-100

(t)
(t)
-150
tos (t)
0
tos (t)
0

-3
-2
-1
0
1
2
3
4 fj (t)
fj,
(t)
,0
log
0
Re
Re
0
0
Im
Im
x1
x1
Bode plot 1
x2
x2
x1 1.2
x1
x2
x2
log
log
1
dB
dB
deg
deg
u = 0 ln 0.8
u = 0 ln
ln coth(|u| /2)
ln coth(|u| /2)
0.6
or 1
or 1
m
m
0.4

yos
yos
0.2

tos 0
tos 0
0

2
4
6
8
10,0
,0
t
0
0

-30
-40
-50
-3

-2

-1

log

150
100
50

deg

-2

150

deg

100

hN,D (t)

1 (t)

-3

deg

1 (t)

,0
0

150

deg

u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

355

hN,D (t)

PSfrag replacements

t
t
ke(t)k
ke(t)k
e1 (t)
e1 (t)
e2 (t)
e2 (t)
e3 (t)
e3 (t)
y(t)
y(t)
u(t)
u(t)
Exercises for Chapter
8
h (t)
h (t)
hN,D (t)
hN,D (t)
1 (t)
1 (t)
1N,D (t)
1N,D (t)
(t)
(t)
0
0
(t)
(t)
-20
-20
fj (t)
fj (t)
Re
Re
-40
-40
Im
Im
x1
x1
-60
-60
x
x
PSfrag replacements2
PSfrag replacements2
x1
x1
-3
-2
-1
0
1
2
3
4
x

ke(t)k2
ke(t)k2
log
e1 (t)
e1 (t)
e2 (t)
e2 (t)
150
150
e3 (t)
e3 (t)
u = 0 ln

u = 0 ln

y(t)
y(t)
100
100
ln coth(|u|u(t)
/2)
ln coth(|u|u(t)
/2)
50
50
1
1
orh(t)
orh(t)
0
0
m
m
hN,D(t)
hN,D(t)

-50
-50
os
os
1N,Dy(t)
1N,Dy(t)
-100
-100

(t)
(t)
-150
-150
tos (t)
0
tos (t)
0

-3
-2
-1
0
1
2
3
4 fj (t)
fj,
(t)

,
log

0
0
Re
Re
0
0
Im
Im
x1
x1
Bode plot 1
x2
x2
x1
x1
2
x2
x2
log
log
1.5
dB
dB
deg
deg 1.5
1
u = 0 ln
u = 0 ln
ln coth(|u| /2)
ln coth(|u| /2)
1
0.5
or 1
or 1
m
m

0
PSfrag replacements
PSfrag replacements
yos 0.5
yos
t
t
tos 0 -0.5
tos 0
ke(t)k
ke(t)k
0
e1,
(t)
e1,
(t)
2.5
5
7.5
10 12.5 15 17.5 20
0
0
e2(t)
e2(t)
t
0
0
e3 (t)
e3 (t)
y(t)
y(t)
Step response 1
u(t)
u(t)
h (t)(b)
h (t)
hN,D (t)
hN,D (t)
1 (t)
1 (t)
20
1N,D (t)
1N,D (t)
0
(t)
(t)
0
(t)
(t)
-20
fj (t)
fj (t)
-20
Re
Re
-40
Im
Im
-40
x1
x1
-60
x2
x2
-60
x1
x1
-3
-2
-1
0
1
2
3
4
x2
x2
log

1 (t)

PSfrag replacements

0
-50
-100
-150
-3

-2

-1

log

Bode plot 2
1.2
1
0.8
0.6
0.4
0.2
0

Step response 2

PSfrag replacements

100

deg

50
0
-50
-100
-150

1
0.8
0.6
0.4
0.2

,0
0

357

8 Performance of control systems

Step response 1

17.5

Step response 2

(sI n A)1

Figure 8.5 Block diagram for static state feedback

(sI n A)1

y(s)

Figure 8.6 Block diagram for static output feedback

the transfer function under static state feedback and static output feedback.

r(s)

22/10/2004

The material in this chapter has focused upon the unity gain feedback loop and its relation
to the solution of Problem 6.41 concerning design for input/output systems. In the next
two problems, you will investigate a few aspects of performance for Problem 6.48 where
static state feedback is considered, and for Problem 6.53 where static output feedback is
considered. Recall that the block diagram representation for static state feedback is as in
Figure 8.5, and that the block diagram representation for static output feedback is as in

r(s)

358

Figure 8.6. You may also wish to refer to Theorems 6.49 and 6.54 concerning the form of

deg

150

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)
Exercises for Chapter
8
h (t)
hN,D (t)
1 (t)
0
1N,D (t)
(t)
-25
(t)
-50
fj (t)
-75
Re
-100
Im
-125
x1
x2
-150
PSfrag replacements
x1
-175
-3 -2 -1
0
1
2
3
4 x
-3 -2 -1
0
1
2
log
log
ke(t)k
e1 (t)
e2 (t)
150
3 (t)
u = 0eln

100
ln coth(|u|y(t)
/2)
50
or u(t)
1
h(t)
0
m
hN,D (t)

-50
yos
-100
1N,D (t)

-150
t(t)
os 0

-3
-2
-1
0
1
2
3
4 (t)
-3
-2
-1
0
1
fj,
(t)
log
log
0
0
Re
Im
Bode plot 1
Bode plot 2
x1
x2
x1
0
x2
log
dB -10
deg
u = 0 ln
ln coth(|u| /2) -20
or 1
m
-30
yos
-40
tos 0

2.5
5
7.5
10 12.5 15 17.5 20,0
2.5
5
7.5
10 12.5
t
t
0
1 (t)

0
-25
-50
-75
-100
-125
-150
-175

1 (t)

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)
h (t)
hN,D (t)
1 (t)
1N,D (t)
(t)
(t)
fj (t)
Re
Im
x1
x2
PSfrag replacements
x1
x2
ke(t)k
e1 (t)
e2 (t)
3 (t)
u = 0eln

ln coth(|u|y(t)
/2)
u(t)
or 1
h(t)
m
hN,D (t)

yos
1N,D (t)

t(t)
os 0
(t)

fj,
(t)
0
0
Re
Im
x1
x2
x1
x2
log
dB
deg
u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

y(s)

E8.7 In this exercise we consider a controllable SISO linear system = (A, b, ct , 01 ) and a
state feedback vector f . You may suppose that (A, b) is in controller canonical form,
and that f Ss ().
(a) Compute the transfer function from the reference r(s) to the error r(s) y(s) for
the closed-loop system of Figure 8.5.
(b) Determine the system type for the closed-loop system of Figure 8.5. Note that
the system type will depend on the relationship between the state feedback vector
f and the system .
Now we will consider a concrete example of the above situation by taking

0 1
0
1
f
A=
, b=
, c=
, f= 0 .
f1
0 0
1
2
(c) What are the possible values for the system type for the closed-loop system in
this case?
(d) For what values of f0 and f1 does the system type achieve its maximum possible
value?
(e) Let f be a state feedback vector from part (d), i.e., so that the system type of the
closed-loop system is maximal. Plot the step response of the closed-loop system.
What is the steady-state error?
(f) Let f be a state feedback vector that is not of the type which answers
part (d), i.e., so that the system type of the closed-loop system is not maximal. Plot the step response of the closed-loop system. What is the steady-state
error?
E8.8 In this exercise we consider a controllable SISO linear system = (A, b, ct , 01 ) and an
output feedback constant F . You may suppose that (A, b) is in controller canonical
form.
(a) Compute the transfer function from the reference r(s) to the error r(s) y(s) for
the closed-loop system of Figure 8.6.

Exercises for Chapter 8

359

(b) Determine the system type for the closed-loop system of Figure 8.6. Note that
the system type will depend on the relationship between the output feedback
constant F and the system .
Now we will consider a concrete example of the above situation by taking

0 1
0
1
A=
, b=
, c=
.
0 0
1
1
(c) What are the possible values for the system type for the closed-loop system in
this case?
(d) For what values of F does the system type achieve its maximum possible value?
(e) Let F be an output feedback constant from part (d), i.e., so that the system type
of the closed-loop system is maximal. Plot the step response of the closed-loop
system. What is the steady-state error?
(f) Let F be a output feedback constant that is not of the type which answers
part (d), i.e., so that the system type of the closed-loop system is not maximal.
Plot the step response of the closed-loop system. What is the steady-state error?

360

8 Performance of control systems

22/10/2004

362

This version: 22/10/2004

9 Design limitations for feedback control

r(s)

Chapter 9

RL (s)

22/10/2004

y(s)

Design limitations for feedback control

Figure 9.1 Unity gain feedback loop for studying time-domain behaviour

In this chapter we shall follow up on some of the discussion of Chapter 8 with details
concerning performance limitations in both the frequency and time-domains. The objective
of this chapter is to clarify the way in which certain plant features can impinge upon the
attainability of certain performance specifications. That such matters can arise should be
clear from parts of our discussion in the preceding chapter. There we saw that even for
simple second-order systems there is a tradeoff to be made when simultaneously optimising,
for example, rise time and overshoot. In this chapter such matters will be brought into
clearer focus, and presented in a general context.
Many of the ideas we discuss here have been known for some time, but a very nice current
summary of results of the type we present may be found in the book of Seron, Braslavsky,
and Goodwin [1997]. The starting point for the discussion in this chapter might be thought
of as Bodes Gain/Phase Theorem stated in Section 4.4.2.

the loop gain RL itself affects the system performance. Note that it is possible to have RL not
be BIBO stable, but for the closed-loop system to be IBIBO stable. Thus, if RL = RC RP
for a controller transfer function RC and a plant transfer function RP , it is possible to
stabilise an unstable plant using feedback. However, we shall see in this section that unstable
plants, along with nonminimum phase plants, can impose on the system certain performance
limitations. Our treatment follows that of Seron, Braslavsky, and Goodwin [1997].
The integral constraints of the following result form the backbone for a part of what will
follow. The crucial thing to note here is that for any plant that has poles or zeros in C+ ,
there will be performance restrictions, regardless of what kind of controller one employs.

Contents

9.1 Theorem Suppose the unity gain feedback loop of Figure 9.1 is IBIBO stable and that the
closed-loop transfer function is steppable. Let y(t) be the normalised step response and let
e(t) = r(t) y(t) be the error. Let

= inf Re(s) > Re(p) when p is a pole of TL

9.1

Performance restrictions in the time-domain for general systems . . . . . . . . . . . . . . 361

9.2

Performance restrictions in the frequency domain for general systems . . . . . . . . . . . 368

9.3

9.4

9.2.1

Bode integral formulae . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 369

9.2.2

Bandwidth constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 374

9.2.3

The waterbed effect . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 375

9.2.4

Poisson integral formulae . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 377

The robust performance problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 381

9.3.1

Performance objectives in terms of sensitivity and transfer functions . . . . . . . . 382

9.3.2

Nominal and robust performance

. . . . . . . . . . . . . . . . . . . . . . . . . . . 385

Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 392

be the largest value of the real parts of the pole of TL . The following statements hold.
(i) If p C+ is a pole for RL then
Z
Z
1
ept e(t) dt = 0 and
ept y(t) dt =
.
TL (0)p
0
0
(ii) If z C+ is a zero for RL then
Z
ezt e(t) dt =
0

1
TL (0)z

Z
and

ezt y(t) dt = 0.

(iii) If RL has a pole at 0 then the following statements hold:

9.1 Performance restrictions in the time-domain for general systems

The objective when studying feedback control systems is to see how behaviour of certain
transfer functions affects the response. To this end, we concentrate on a certain simple
feedback configuration, namely the unity gain feedback loop of Figure 6.25 that we again
reproduce, this time in Figure 9.1. While in Section 8.3 we investigated the transfer function
TL of the closed-loop system of Figure 9.1, in this section we focus on how the character of

(a) if p C is a pole for RL with Re(p) > then

Z
ept e(t) dt = 0;
0

(b) if z C is a zero for RL with Re(z) > then

Z
1
ezt e(t) dt =
.
TL (0)z
0

22/10/2004

9.1 Performance restrictions in the time-domain for general systems

363

Proof Let us first determine that all stated integrals exist. Since the closed-loop system
is IBIBO stable, the normalised step response y(t) will be bounded, as will be the error.
Therefore, if Re(s) > 0, the integrals
Z
Z
est y(t) dt
est e(t) dt and
0

will exist. Now we claim that if lims0 RL (s) = then limt0 e(t) = 0. Indeed, by the
Final Value Theorem, Proposition E.9(ii), we have
lim e(t) = lim sSL (s)
r(s)

= lim

1
1
1 + RL (s) TL (0)

= 0,
where we have noted that for the normalised step response the reference is 1(t) TL1(0) . Now,
since limt e(t) = 0 when lims0 RL (s) = , in this case the integral
Z
est e(t) dt
0

will exist provided s is greater than the abscissa of absolute convergence for e(t). Since the
abscissa of absolute convergence is , it follows that the integrals of part (iii) exist.
(i) We have
Z
Z
ept e(t) dt = e(p) and
ept y(t) dt = y(p).
0

Therefore
Z

ept e(t) dt = SL (p)

r(p)

= lim

1
1
1 + RL (s) TL (0)s

= 0,
if Re(p) > 0. For the integral involving y(t), when Re(p) > 0 we have
Z
ept y(t) dt = TL (p)
r(p)
0

= lim

RL (s)
1
1 + RL (s) TL (0)s

1
=
.
TL (0)p
(ii) We compute
Z

ezt e(t) dt = SL (z)

r(z)

364

9 Design limitations for feedback control

if Re(z) > 0. In like manner we have

Z
ezt y(t) = TL (z)
r(z)
0

= lim
sz

1
1
1 + RL (s) TL (0)s

1
,
TL (0)z

RL (s)
1
1 + RL (s) TL (0)s

= 0,
again provided Re(z) > 0.
(iii) In this case, we can conclude, since the integrals have been shown to converge, that

the analysis of parts (i) and (ii) still holds.
It is not immediately obvious why these conclusions are useful or interesting. We shall
see shortly that they do lead to some results that are more obviously useful and interesting,
but lets make a few comments before we move on to further discussion.
9.2 Remarks 1. The primary importance of the result is that it will apply to any case where
a plant has unstable poles or nonminimum phase zeros. This immediately asserts that
in such cases there will be some restrictions on how the step response can behave.
2. Parts (i) and (ii) can be thought of as placing limits on how good the closed-loop response
can be in the presence of unstable poles or nonminimum phase zeros for the loop gain.
3. Part (iii a) ensures that even if the plant has no poles in C+ , provided that it has a pole
at s = 0 along with any other pole to the right of the largest closed-loop pole, there will
be overshoot.
4. Along similar lines, from part (iii b), if RL has a pole at the origin along with a zero,
even a minimum phase zero, that lies to the right of the largest pole of the closed-loop
system, then there will be undershoot.
5. We saw in Section 8.3 the ramifications of the assumption that lims0 RL (s) = . What
this means is that there is an integrator in the loop gain, and integrators give certain
properties with respect to rejection of disturbances, and the ability to track certain
reference signals.

Let us give a few examples that illustrate the remarks 3 and 4 above.
9.3 Examples 1. Let us illustrate part (iii a) of Theorem 9.1 with the loop gain RL (s) =
2(s+1)
. The closed-loop transfer function is
s(s1)
TL (s) =

2(s + 1)
,
s2 + s + 2

which is BIBO stable. The normalised step response is shown in Figure 9.2. The thing
to note, of course, is that the step response exhibits overshoot.
1s
2. Let us illustrate part (iii b) of Theorem 9.1 with the loop gain RL (s) = s(s+2)
. The
closed-loop transfer function is

= lim

22/10/2004

TL (s) =

1s
,
s2 + s + 1

which is BIBO stable. The normalised step response is shown in Figure 9.2, and as
expected there is undershoot in the response.

hN,D (t)

9.1 Performance

,0
0

1N,D (t)
(t)
(t)
fj (t)
Re
Im the time-domain
restrictions in
x1
x2
x1
1.2
x2
log
1
dB
deg 0.8
u = 0 ln
ln coth(|u| /2) 0.6
0.4
or 1
m
0.2

yos
0

tos 0 -0.2

8
10
2

for general systems

365

1 (t)

1N,D (t)
(t)
(t)
fj (t)
Re
Im
22/10/2004
x1
x2
x1
x2
1.5
log
dB
deg 1.25
u = 0 ln
1
ln coth(|u| /2)
or 1 0.75
m

0.5
yos
0.25
tos 0
0

,0
0

Figure 9.2 Normalised step responses for two loop gains with a
pole at zero: (1) on the right an unstable pole gives overshoot
and (2) on the left a nonminimum phase zero gives undershoot

As we saw in the previous remarks 3 and 4, an unstable pole in the loop gain immediately
implies overshoot in the step response and a nonminimum phase zero implies undershoot.
It turns out that we can be even more explicit about what is happening, and the following
result spells this out.
9.4 Proposition Consider the closed-loop system in Figure 9.1. Assuming the closed-loop
system is steppable and IBIBO stable, the following statements hold.
(i) If RL has a real pole at p C+ then there is an overshoot in the normalised step
response, and if tr is the rise time, then the maximum overshoot satisfies
(ptr TL (0))eptr + TL (0)
.
yos
ptr
If TL (0) = 1 then we can simplify this to
yos

ptr
.
2

Thus we have the inequality

yos ept
ptr + (eptr 1)TL (0)

p
p2 tr TL (0)
which, with simple manipulation, gives the first inequality of this part of the proposition.
For the second inequality one performs the Taylor expansion
eptt = 1 + ptr + 12 p2 t2r +
and a simple manipulation using TL (0) = 1 gives the second inequality.
(ii) For t ts, we have y(t) 1 . Therefore, using Theorem 9.1(ii), we have
Z
ezt y(t) dt = 0
0
Z ts,
Z
ezt y(t) dt =
ezt y(t) dt
=
Z
=

ts,

ts,
zt

Z
y(t) dt

ezt (1 ) dt

ts,

0
ts,

ezt y(t) dt

1
t

TL (0) tr

eptr
(eptr 1)ptr + TL (0)(1 eptr + ptr ) 1
+
1
2
pt
r
p tr TL (0)e
TL (0)
p
ptr + (eptr 1)TL (0)
=
.
p2 tr TL (0)

Proof (i) That there must be overshoot follows as stated in Remark 9.23. Our definition
of rise time implies that for t < tr we have y(t) < ttr . This means that

22/10/2004

Now we use this final inequality, along with the fact that yos y(t) 1 = ( TL1(0) 1) e(t)
to derive
Z
yos ept
= yos
ept dt
p
tr
Z
Z

ept e(t) dt +
ept TL1(0) 1 dt

t
tr
Z tr r
Z

pt
1
t
e

ept TL1(0) 1 dt
tr dt +
TL (0)

1
yus zts,
.
e
1

when t tr . Therefore, using Theorem 9.1(i),

Z
ept e(t) dt = 0
0
Z
Z tr
=
ept e(t) dt =
ept e(t) dt
tr
0
Z
Z tr
=
ept e(t) dt
ept TL1(0)

9 Design limitations for feedback control

(ii) If RL has a real zero at z C+ then there is an undershoot in the normalised step
response, and if ts, is the -settling time, then the undershoot satisfies

e(t)

366

(1 )ezts,
.
z

Now we use the definition of the undershoot and the previous inequality to compute
Z ts,
yus (1 ezts, )
= yus
ezt dt
z
0
Z ts,

ezt y(t) dt
0

(1 )ezts,

.
z
Thus we have demonstrated the inequality

t
tr

dt.

yus (1 ezts, )
(1 )ezts,

,
z
z
and from this the desired result follows easily.

22/10/2004

9.1 Performance restrictions in the time-domain for general systems

367

Let us check the predictions of the proposition on the previous examples where overshoot
and undershoot were exhibited.
9.5 Examples (Example 9.3 contd) 1. We resume looking at the loop gain RL (s) = 2(s+1)
s(s1)
where the normalised step response for the closed-loop system is shown on the left in
Figure 9.2. The closed-loop transfer function evaluated at s = 0 is TL (0) = 1. Therefore,
since we have a real pole at s = 1, the rise time and overshoot should together satisfy
yos t2r . Numerically we determine that yos 0.68. For this example, it turns out that
tr = 0 (the order of the transfer function is not high enough to generate an interesting
rise time with our definition. In any event, the inequality of part (i) of Proposition 9.4
is certainly satisfied.
1s
for which the closed-loop normalised step
2. Here we use the loop gain RL (s) = s(s+2)
response is shown on the right in Figure 9.2. Taking = 0.05, the undershoot and the
-settling time are computed as yus 0.28 and ts, 6.03. One computes
1
0.002,
ezts, 1
using z = 1. In this case, the inequality of part (ii) of Proposition 9.4 is clearly satisfied.

Note that for the examples, the estimates are very conservative. The reason that this is
not surprising is that the estimates hold for all systems, so one can expect that for a given
example they will not be that sharp.
When we have both an unstable pole and a nonminimum phase zero for the loop gain
RL , it is possible again to provide estimates for the overshoot and undershoot.
9.6 Proposition Consider the unity gain feedback loop of Figure 9.1, and suppose the closedloop system is IBIBO stable and steppable. Also suppose that the loop gain RL has a real
pole at p C+ and a real zero at z C+ . The following statements hold:
(i) if p < q then the overshoot satisfies
yos

q
.
TL (0)(p q)

Proof (i) From the formulas for concerning e(t) from parts (i) and (ii) of Theorem 9.1 we
have
Z
1
(ept ezt )(e(t)) dt =
.
TL (0)z
0
Since yos e(t) for all t > 0 this gives
Z
1
zp
yos
(ept ezt )(e(t)) dt =
.
TL (0)z
pz
0
From this the result follows.

9 Design limitations for feedback control

22/10/2004

(ii) Here we again use parts (i) and (ii) of Theorem 9.1, but now we use the formulas
concerning y(t). This gives
Z
1
(ezt ept )(y(t)) dt =
.
T
(0)p
L
0
Since yus y(t) for all t > 0 we have
Z
1
pz
(ezt ept )(y(t)) dt =
yus
,
TL (0)p
pz
0
giving the result.

An opportunity to employ this result is given in Exercise E9.1. The implications are
quite transparent, however. When one has an unstable real pole and a nonminimum phase
real zero, the undershoot or overshoot can be expected to be large if the zero and the pole
are widely separated.

9.2 Performance restrictions in the frequency domain for general

systems
In the previous section we gave some general results indicating the effects on the time response of unstable poles and nonminimum phase zeros of the plant. In this section we carry
this investigation into the frequency domain, following the excellent treatment of Seron,
Braslavsky, and Goodwin [1997]. The setup is the unity gain loop depicted in Figure 9.3.
Associated with this, of course, we have the closed-loop transfer function TL and the sensi-

r(s)

RL (s)

y(s)

p
;
TL (0)(q p)

(ii) if q < p then the undershoot satisfies

yus

368

Figure 9.3 The unity gain feedback loop for investigation of performance in the frequency domain

tivity function SL defined, as usual, by

TL (s) =

RL (s)
,
1 + RL (s)

SL (s) =

1
.
1 + RL (s)

These are obviously related, and in this section the aim is to explore fully the consequences
of this relationship, as well as explore the behaviour of these two rational functions as the
loop gain RL has poles and zeros in C+ . If one of the goals of system design is to reduce
error in the closed-loop system, then since the transfer function from input to error is the
sensitivity function, a goal might be to reduce the sensitivity function. However, it is simply
not possible to do this in any possible manner, and the zeros and poles of the loop gain RL
have a great deal to say about what is and is not possible.

22/10/2004

9.2 Performance restrictions in the frequency domain for general systems

369

370

9 Design limitations for feedback control

9.2.1 Bode integral formulae Throughout this section, we are dealing with the unity
gain feedback loop of Figure 9.3. The following result gives what are called the Bode
integral formulae. These results, for stable and minimum phase loop gains, are due to
Horowitz [1963]. The extension to unstable loop gains for the sensitivity function are due
to Freudenberg and Looze [1985], and for nonminimum phase loop gains for the transfer
function results are due to Middleton and Goodwin [1990].

1,j 2,j 3,j

9.7 Theorem Consider the feedback interconnection of Figure 9.3 and suppose that RL R(s)
is proper and has poles p1 , . . . , pk in C+ and zeros z1 , . . . , z` in C+ . If the interconnection
is IBIBO stable and if RL has no poles on the imaginary axis then the following statements
hold:
(i) if the closed-loop system is well-posed then
Z

22/10/2004

k
X
SL (i)
d = lim s(SL (s) SL ()) +
pj ;
ln

s
SL (i)
2
SL ()
j=1

(ii) if L(0) 6= 0 then

Z
0

`
X
TL (i) d
1
dTL (s)
1

ln
.
=
lim
+

2
s0
TL (0)
2 TL (0)
ds
z
j=1 j

Figure 9.4 The contour for the Bode integral formulae

For j {1, . . . , k}, define fj (s) by requiring that

Proof (i) For R > 0 sufficiently large and > 0 sufficiently small, we construct a contour
R, C+ comprised of 3k + 2 arcs as follows. We define iR by
= iR \ { [Im(p) , Im(p) + ] | p {p1 , . . . , pk }} .
Now for j {1, . . . , k} define 3 contours

1,j ,

2,j ,

and

3,j

1,j = { x + i(Im(pj ) ) C | x [0, Re(pj )]}

2,j = pj + rei C [
, ]
2 2
3,j = { x + i(Im(pj ) + ) C | x [0, Re(pj )]}
Now define

R = Rei C [
, ]
2 2

The contour R, is the union of these contours, and a depiction of it is shown in Fig(s)
ure 9.4. Since ln SSLL()
is analytic on and inside the contour R, for R sufficiently large and
sufficiently small, we have
Z
SL (s)
lim
ln
ds = 0.
0
SL ()
R R,
Let us analyse this integral bit by bit. Using the expression

SL (s)
SL (s)
+ i]SL (s)/SL (),
= ln
ln
SL ()
SL ()
we have

Z
lim

0
R

SL (s)
ds = 2i
SL ()

ln
0

SL (i)
d.
SL (i)

SL (s)
= ln(s pj ) + ln fj (s).
SL ()

,j be the contour obtained by closing ,j along the

Let ,j = 1,j 2,j 3,j , and let

imaginary axis. On and within ,j the function fj is analytic. Therefore, by Cauchys
Integral Theorem,
Z
Z
Z Im(pj )i
ln fj (s) ds =
ln fj (s) ds +
ln fj (i) d = 0

Combining this with

Z
ln
,j

SL (s)
ds =
SL ()

Im(pj )+i

Z
ln(s pj ) ds +

ln fj (s) ds
,j

gives
Z
ln
,j

SL (s)
ds =
SL ()

Im(pj )+i

ln(s pj ) ds +
,j

ln fj (i) d.
Im(pj )i

In the limit as goes to zero, the second integral on the right will vanish. Thus we shall
evaluate the first integral on the right. Since ln(s pj ) is analytic on ,j we use the fact
that the antiderivative of ln z is z ln z z to compute
Z
Im(pj )+i
ln(s pj ) ds = (s pj ) ln(s pj ) (s pj )
,j

Im(pj )i

= (Re(pj ) + i) ln(Re(pj ) + i) (Re(pj ) + i)

(Re(pj ) i) ln(Re(pj ) i) + (Re(pj ) i)

= iRe(pj ) ](Re(p) i) ](Re(p) + i) .

22/10/2004

9.2 Performance restrictions in the frequency domain for general systems

371

Taking the limit as 0 gives

372

9 Design limitations for feedback control

Now note that (9.1) implies that

lim

SL (s)
ds = 2iRe(pj ).
SL ()

c1 = lim s(RL (s) RL ())

s
1
1
= lim s

s
SL (s) SL ()

1
= 2
lim s SL () SL (s) .
SL () s

This then gives

lim
0

k Z
X
j=1

22/10/2004

= 2i

k
X

Re(pi ) = 2i

j=1

k
X

pj ,

j=1

This gives

with the last equality holding since poles of RL occur in complex conjugate pairs.
Now we look at the contour R . We have
lim ln

Z
ln

lim

SL (s)
SL (s)
ds = i Res ln
.
s=
SL ()
SL ()

The result now follows by summing the contours.

(ii) Here we define G(z) = RL (1/z) and

Since RL is proper we may write its Laurent expansion at s = as

RL (s) = RL () +

c1 c2
+ 2 + .
s
s

H(z) =
(9.1)

ln SL (s) = ln(1 + RL (s))

1
c2
1
c1
+
+ .
1 + RL () s 1 + RL () s2

SL (s)
L (s) ,
= ln 1 + R
SL ()

`
X
H(i)
1
d = lim z(H(z) H()) +
ln
.
H(i)
2 z
H()
z
j=1 j

Let us first show that

lim

Therefore, writing ln SL () = ln(1 + RL ()), we have

z(H(z) H())
1
dTL (s)
=
lim
.
H()
TL (0) s0 ds

First recall that

lim

where

L (s) = SL ()c1 + SL ()c2 + .

R
s
s2
Using the power series expansion for ln(1 + z),
ln(1 + z) = z +

z2
+ ,
2

for |z| < 1 we have

1
1
=
.
1 + 1/TL (1/z)
1 + G(z)

Thus G and H now play the part of RL and SL in part (i). Thus we have

Since SL (s) = (1 + RL (s))1 we have

= ln(1 + RL ()) ln 1 +

SL (s)
1
= i
lim s SL () SL (s) .
SL ()
SL () s

z(H(z) H())
= Res ln H(z),
z=
H()

following our calculations in the previous part of the proof. If the Taylor expansion of
ln TL (s) about s = 0 is
ln TL (s) = a0 + a1 s + a2 s2 + ,
then the Laurent expansion of ln H(z) about z = is
ln H(z) = a0 +

a1 a2
+ 2 + .
z
z

Therefore Resz= ln H(z) = a1 . However, we also have

2
SL (s)
L (s) + RL (s) +
ln
= R
SL ()
s
SL ()c1 SL ()c2
=

+ ,
s
s2

which is valid for s sufficiently large. This shows that

Res ln

SL (s)
= SL ()c1 .
SL ()

a1 = lim

Thus we have

Z
0

d ln TL (s)
1
dTL (s)
=
lim
.
ds
TL (0) s0 ds

H(i)
d = 1 lim dTL (s) .
ln
H(i)
2 TL (0) s0 ds

The result now follows by making the change of variable = 1 .

Although the theorem is of some importance, its consequences do require some explication. This will be done further in the next section, but here let us make some remarks that
can be easily deduced.

22/10/2004

9.2 Performance restrictions in the frequency domain for general systems

373

9.8 Remarks 1. Let us provide an interpretation for the first term on the right-hand side in
part (i). Using the definition of SL we have
lim

s(SL (s) SL ())

s(RL () RL(s))
= lim
s
SL ()
1 + RL (s)
1 N,D (0+)
,
=
1 + RL ()

where the last step follows from Proposition E.9(i). Here (N, D) is the c.f.r. of RL . On
this formula, let us make some remarks.
(a) If the relative degree of RL is greater than 0 then RL () = 0.
(b) If the relative degree of the plant is greater than 1 then 1 N,D (0+) = 0.
(c) From the previous two remarks, if the relative degree of RL is greater than 1 and if
the plant has no poles in C+ , then the formula
Z
ln |SL (i)| d = 0
0

holds, since SL (i) = 1 in these cases. This is the formula originally due to
Horowitz [1963], and is called the Horowitz area formula for the sensitivity
function. It tells us that the average of the area under the magnitude part of the
sensitivity Bode plot should be zero. Therefore, for any frequency intervals where
SL is less that 1, there are also frequency regions where SL will be greater than 1.
(d) If the relative degree of RL is greater than 1 but RL has poles in C+ , then the
formula from part (i) reads
Z
k
X
ln |SL (i)| d =
pj .
0

374

9 Design limitations for feedback control

22/10/2004

(a) If the scaled closed-loop system is type k for k > 1 then the steady-state error to
the ramp input will be zero.
(b) Based upon the previous remark, if the scaled closed-loop system is type k for k > 1
and if the loop gain RL has no zeros in C+ , then the formula

Z
TL (i) d

ln
=0
TL (0) 2
0
holds. Again, this is to be seen as an area integral for the magnitude Bode plot for
TL , but now it is weighted by 12 . Thus the magnitude smaller frequencies counts
for more in this integral constraint. This formula, like its sensitivity function counterpart, was first derived by Horowitz [1963], and is the Horowitz area formula
for the closed-loop transfer function.
(c) If the scaled closed-loop system is type k for k > 1 but the loop gain does have
zeros in C+ , then the formula

Z
`
TL (i) d X
1

ln
=
TL (0) 2
z
0
j=1 j
holds. Thus we see that as with poles for the sensitivity function integral, the zeros
for the loop gain shift the area up. Now we note that this effect gets worse as the
zeros approach the imaginary axis.
(d) Finally, suppose that the scaled closed-loop system is type 1 (so that the steady-state
error to the unit ramp is constant) and that RL has zeros in C+ . Then, provided
that the steady-state error to the unit ramp input is positive, we can compensate
for the effect of the zeros on the right hand side only by making the steady-state
error to the ramp larger.

j=1

Since the right-hand side is positive, this tells us that the poles in C+ have the effect
of shifting the area bias of the sensitivity function in the positive direction. That is
to say, with poles for RL in C+ , there will be a greater frequency range for which
SL will be greater than 1.
(e) Now let us consider the cases where the relative degree of RL is 0 or 1, and where
RL has poles in C+ . Here, if 1 N,D (0+) > 0 we have an opportunity to reduce the
detrimental effect of the positive contribution to the area integral from the poles.
2. Let us try to understand part (ii) in the same manner by understanding the first term
(s)
on the right-hand side. Define the scaled closed-loop transfer function TL (s) = TTLL(0)
.

Let e(t) be the error

signal
for
the
unit
ramp
input
of
the
transfer
function
T
L . Thus
(s)
s2 e(t) = 1 TTLL(0)
. We compute
TL (s)
1
1
dTL (s)
T (0)
lim
= lim L
s0
s0
TL (0)
ds
s
= lim s
e(s)
s0

= lim e(t),
t

where in the final step we have used Proposition E.9(ii). Let us make some remarks,
given this formula.

9.2.2 Bandwidth constraints The preceding discussion has alerted us to problems

we may encounter in trying to arbitrarily shape the Bode plots for the closed-loop transfer
function and the sensitivity function. Let us examine this matter further by seeing how
bandwidth constraints come into play.
For the unity gain closed-loop interconnection of Figure 9.3, if b is the bandwidth for
the closed-loop transfer function, then we must have

|RL (i)|
1
,
|1 + RL (i)|
2
1
|RL (i)|
,
2+1

> b
> b

This implies that bandwidth constraints for the closed-loop system translate into bandwidth
constraints for the loop gain, and vice versa. Typically, the bandwidth of the loop gain
will be limited by the plant, and the components available for the controller. The upshot is
that when performing a controller design, the designer will typically be confronted with an
inequality of the form
1+k
b
|RL (i)|
,
> b ,
(9.2)

where < 12 and where k N. The objective in this section is to see how these constraints
translate into constraints on the sensitivity function. This issue also came up in Section 8.6.1
in the discussion of high-frequency roll-off constraints.

22/10/2004

9.2 Performance restrictions in the frequency domain for general systems

375

376

9 Design limitations for feedback control

The following result gives a bound on the area under the tail of the sensitivity function
magnitude Bode plot.

Proof We first note without proof that if |s| < 12 then |ln(1 + s)| < 3|s|
. We then compute
2

Z
Z

|ln |SL (i)|| d
ln |SL (i)| d

b
Zb
|ln SL (i)| d

Zb
|ln(1 + RL (i)| d
=

ln |SL (i)|

9.9 Proposition For the closed-loop interconnection of Figure 9.3, if RL has relative degree
of greater than 1 and if RL satisfies a bandwidth constraint of the form (9.2), then

Z
b

ln |SL (i)| d
.

2k
b

22/10/2004

as desired.

1
d
1+k

3b1+k

2
b
=
,
2k

Let us see if we can explain the point of this.

9.10 Remark The idea is this. From Theorem 9.7(i) we know that since the relative degree
of RL is at least 2, the total area under the sensitivity function magnitude Bode plot will
be nonnegative. This means that if we have a region where we have made the sensitivity
function smaller than 1, there must be a region where the sensitivity function is larger
than 1. What one can hope for, however, is that one can smear this necessary increase in
magnitude of the sensitivity function over a large frequency range and so keep the maximum
magnitude of the sensitivity function under control. Proposition 9.9 tells us that if the loop
gain has bandwidth constraints, then the area contributed by the sensitivity function above
the bandwidth for the loop gain is limited. Therefore, the implication is necessarily that
if one wishes to significantly decrease the sensitivity function magnitude over a frequency
range, there must be a significant increase in the sensitivity function magnitude at frequencies
below the loop gain bandwidth. An attempt to illustrate this is given in Figure 9.5. In the
figure, the sensitivity function is made small in the frequency range [1 , 2 ], and the cost for
this is that there is a large peak in the sensitivity function magnitude below the bandwidth
since the contribution above the bandwidth is limited by Proposition 9.9.

Figure 9.5 The effect of loop gain bandwidth constraints on the

sensitivity function area distribiution

9.11 Theorem Consider the feedback interconnection of Figure 9.7, and assume that the
closed-loop system is IBIBO stable and well-posed. If RL has zero in C+ then for 0 <
1 < 2 there exists m > 0 so that
sup |SL (i)| kSL km
.

[1 ,2 ]

Proof Suppose that C+ is a zero of RL . Then SL (q) = 1 by Proposition 8.22. Let

D(0, 1) = { z C | |z| 1}
be the unit complex disk and define a mapping : C+ D(0, 1) by
s
.
(s) =
+s
This map is invertible and has inverse

9.2.3 The waterbed effect Let us first look at a phenomenon that indicates problems
that can be encountered in trying to minimise the sensitivity function over a frequency range.
This result is due to Francis and Zames [1984].

1 (z) =

z
.
1+z

The interval i[1 , 2 ] iR is mapped under to that portion of the boundary of D(0, 1)
given by

(i[1 , 2 ]) = ei [1 , 2 ] ,

22/10/2004

9.2 Performance restrictions in the frequency domain for general systems

where

ij
eij =
,
+ ij

377

[1 ,2 ]

Define = |2 1 | and let n >

be an integer. Define f : D(0, 1) C as

f (z) =

n1
Y

SL ze2ji/n .

j=0

Being the product of analytic functions in D(0, 1), f itself is analytic in D(0, 1). For each z
on the boundary of D(0, 1) there exists j {0, . . . , n 1} so that
ze2ji/n (i[1 , 2 ]).

1 = |f (0)|

sup f (ei )

n1

= sup SL (ei )
sup SL (ei )

[1 ,2 ]

sup |SL (i)| .

[1 ,2 ]

Taking m = n 1, the result follows.

9.12 Remark The idea here is simple. If, for a nonminimum phase plant, one wishes to reduce
the sensitivity function over a certain frequency range, then one can expect that the kSL k
will be increased in consequence.

9.2.4 Poisson integral formulae Now we look at other formulae that govern the
behaviour of the closed-loop transfer function and the sensitivity function. In order to state
the results in this section, we need to represent the closed-loop transfer function and the
sensitivity function in a particular manner. Recall from (8.2) the definitions Z(SL ) and
Z(TL ) of the sets of zeros for SL and TL in C+ . Suppose that Z(SL ) = {p1 , . . . , pk } and
Z(TL ) = {z1 , . . . , z` }, the notation reflecting the fact that zeros for SL are poles for RL and
zeros for TL are zeros for RL . Corresponding to these zeros are the Blaschke products for
SL and TL defined by
BSL (s) =

k
Y
pj s
,
p + s
j=1 j

BTL (s) =

`
Y
zj s
.
z + s
j=1 j

L by the equality
These functions all have unit magnitude. Now we define R
RL (s) =

L (s)BT (s)
R
L
.
BSL (s)

22/10/2004

SL (s) = SL (s)BSL (s),

TL (s) = TL (s)BTL (s),

and observe that SL and TL have no zeros in C+ .

With this notation, we state the Poisson integral formulae for the sensitivity function
and the closed-loop transfer function.
9.13 Theorem Consider the closed-loop interconnection of Figure 9.7. If the closed-loop system is IBIBO stable, the following equalities hold:
(i) if z = z + iz C+ is a zero of RL then
Z

z
ln |SL (i)| 2
d = ln BS1
(z) ;
L
z + ( z )2

(ii) if p = p + ip C+ is a pole of RL then

Z

p
ln |TL (i)| 2
d = ln BT1
(p) + p .
L
2
p + ( p )

Thus we have, using the Maximum Modulus Principle (Theorem D.9),

= kSL kn1

9 Design limitations for feedback control

In Section 14.2.2 we shall refer to this as an inner/outer factorisation of RL . The key fact
here is that the zeros and poles for RL have been soaked up into the Blaschke products so
L (s) has no zeros or poles in C+ . We similarly define SL and TL by
that R

j = 1, 2.

Define SL : D(0, 1) C by ( SL )(z) = SL (1 (z)). Thus SL is analytic in D(0, 1) and

SL (0) = 1. Furthermore, the following equalities hold:

sup SL (ei ) = sup |SL (i)| ,
sup SL (ei ) = kSL k .
[1 ,2 ]

378

(9.3)

Proof (i) This follows by applying Corollary D.8 to the function SL . To get the result, one
simply observes that ln SL is analytic and nonzero in C+ , that |SL (i)| = |SL (i)| for all
R, and that SL (z) = 1 from Proposition 8.22.
(ii) The idea here, like part (i), follows from Corollary D.8. In this case one observes
that ln TL is analytic and nonzero in C+ , that |TL (i)| = |TL (i)| for all R, and that

TL (p) = 1 from Proposition 8.22.
Let us now make some observations concerning the implications of the Poisson integral
formulae as it bears on controller design. Before we say anything formal, let us make some
easy observations.
9.14 Remarks 1. Like the Bode integral formulae, the Poisson integral formulae provide
limits on the behaviour of the magnitude portion of the Bode plots for the sensitivity
and complementary sensitivity functions.
2. Note that the weighting function
Ws () =

s
s2 + ( s )2

(9.4)

in each
of the integrands is positive, and that the Blaschke products satisfy BS1
(s) 1
L
1
and BTL (s) 1 for s C+ . Thus we see that the Poisson integral formulae do indeed
indicate that, for example, if there are zeros for the plant in C+ then the weighted integral
of the sensitivity function will be positive.
3. If a plant is both unstable and nonminimum phase, then one concludes that the weighted
area of sensitivity increase is greater than that for sensitivity decrease. One sees that this
effect is exacerbated if there is a near cancellation of an unstable pole and a nonminimum
phase zero.

22/10/2004

9.2 Performance restrictions in the frequency domain for general systems

379

380

9 Design limitations for feedback control

22/10/2004

Now let us make some more structured comments about the implications of the Poisson
integral formulae. Let us begin by examining carefully the weighting function (9.4) that
appears in the formulae.

This implies that

9.15 Lemma For s0 C+ and 1 > 0 define

Z
s0 (1 ) =

Exponentiating this inequality gives the result.

1

Since BSL (z) 1, 1 < 1, and z (1 ) < , it follows that kSL k > 1. Thus, as with the
waterbed effect, the demand that the magnitude of the sensitivity function be made smaller
than 1 over a certain frequency range guarantees that its H -norm will be greater than 1.
In some sense, the estimate of Corollary 9.16 contains more information since it contains
explicitly in the estimate the location of the zero. This enables us to say, for example,
that if the phase lag contributed by the nonminimum phase zero is large over the specified
frequency range (that is, z (1 ) is near ), then the effect is to further increase the lower
bound on the H -norm of SL . To ameliorate this effect, one could design a controller so
that the loop gain magnitude is small for frequencies at or above the frequency where the
phase lag contributed by z is large.
A parallel process can be carried out for the complementary sensitivity function portion
of Theorem 9.13, i.e., for part (ii). In this case, a design objective is not to reduce TL over a
finite frequency range, but perhaps to ensure that for large frequencies the complementary
sensitivity function is not large. This is an appropriate objective where model uncertainties
are a consideration, and high frequency effects can be detrimental to closed-loop stability.
Thus here we ask that for some 2 > 0 and some positive 2 < 1 we have

Ws0 (1 ) d.

Then the following statements hold:

(i) if s0 is real then
s0 (1 ) = ]

s i
0
1
;
s0 + i1

(ii) if s0 is not real then

s0 (1 ) =

s i
1 s0 i1
0
1
]
+]
.
2
s0 + i1
s0 + i1

Proof Write s0 = 0 + i0 . The lemma follows from the fact that

Z 1
1 0
1 + 0
0
d = arctan
+ arctan
.
2
2
0
0
1 0 + ( 0 )

One should interpret this result as telling us that the length of the interval [1 , 1 ],
weighted by Ws0 , is related to the phase lag incurred by s0 C+ . Of course, in the Poisson
integral formulae, this phase lag arises from nonminimum phase zeros and/or unstable poles.
This precise description of the weighting function allows us to provide some estimates pertaining to reduction of the sensitivity function, along the same lines as given by the waterbed
affect, Theorem 9.11. Thus, for some 1 > 0 and for some positive 1 < 1 we wish for the
sensitivity function to satisfy
|SL (i)| 1 ,

[1 , 1 ].

9.16 Corollary Suppose that a proper loop gain RL has been factored as in (9.3) and that
the interconnection of Figure 9.3 is IBIBO stable. If the inequality (9.5) is satisfied and if
z C+ is a zero for RL then we have
z (1 )
1
1
z (1 )
B (z) z (1 )
SL
1

Proof Using Theorem 9.13 we compute

Z

=
ln BS1
(z)
ln |SL (i)| Wz () d
L

Z 1
Z 1
=
ln |SL (i)| Wz () d +
ln |SL (i)| Wz () d+
1
Z

ln |SL (i)| Wz () d.
1

|TL (i)| 2 ,

|| 2 .

(9.6)

In this case, calculations similar to those of Corollary 9.16 lead to the following result.
9.17 Corollary Suppose that a proper loop gain RL has been factored as in (9.3) and that
the interconnection of Figure 9.3 is IBIBO stable. If the inequality (9.6) is satisfied and if
p C+ is a pole for RL then we have

(9.5)

The following result tells what are the implications of such a demand on the sensitivity
function at other frequencies.

kSL k

z (1 ) ln 1 + ln kSL k ( z (1 )) ln BS1
(z) .
L

kTL k

p (2 )
1

p (2 )
B 1 (p) p (2 ) .
TL
2

The best way to interpret this result differs somewhat from how we interpret Corollary 9.16, since reduction of the H -norm of TL is not a design objective. However, the
closed-loop bandwidth is important, and Corollary 9.17 can be parlayed into an estimate as
follows.
9.18 Corollary Suppose that RL is proper and minimum phase, that the interconnection of
Figure 9.3 is IBIBO stable, and that RL has a real pole p C+ . If TL satisfies (9.6) for
2 = 12 then we have
p (b )
kTL k 2 p (b ) ,
where b is the bandwidth for the closed-loop system. If we further ask that the lower bound
on the H -norm of TL be bounded from above by Tmax , then we have

b p tan
.
ln T
max
2 + 2 ln 2

9.3 The robust performance problem

22/10/2004

381

Proof The first inequality follows directly from Corollary 9.17 after adding the simplifying hypotheses. The second inequality follows from straightforward manipulation of the
inequality
p (b )

2 p (b ) Tmax .

The upshot of the
corollary is that if one wishes to make the lower bound on the H -form
of TL smaller than 2, then the closed-loop bandwidth will exceed the location of the real,
unstable pole p.
Often in applications one wishes to impose the conditions (9.5) and (9.6) together, with
2 > 1 , and noting that (9.6) implies that |SL (i)| < 12 for || > 2 . The picture is
shown in Figure 9.6: one wishes to design the sensitivity function so that it remains below

382

9 Design limitations for feedback control

22/10/2004

difficulties to specify realistic performance objectives. The approach makes contact with the
topic of robust stability of Section 7.3, and enables us to state a control design problem, the
so-called robust performance problem. The resulting problem, as we shall see, takes the
form of a minimisation problem, and its solution is the subject of Chapter 15.
We shall continue in this section to use the unity gain feedback loop of Figure 6.25 that
we reproduce in Figure 9.7.
r(s)

RC (s)

RP (s)

y(s)

|SL (i)|

Figure 9.7 Unity gain feedback loop for robust performance problem
1

9.3.1 Performance objectives in terms of sensitivity and transfer functions If

you ever thought that it was possible to arbitrarily assign performance specifications to a
system, it is hoped that the content of Sections 9.1 and 9.2 have made you think differently.
But perhaps by now one is frightened into thinking that there is simply no way to specify
performance objectives that are achievable. This is not true, of course. But what is true is
that one needs to be somewhat cagey about how these specifications are stated.
Let us begin by reinforcing our results about the difficulty facing us by trying something
nave. According to Proposition 8.24, if we wish to minimise the energy of the error, we
should minimise the H -norm of SL . Thus, we could try to specify > 0 and then set out
to design a controller with the property that kSL k . This objective may be impossible
for many reasons, and let us list some of these.

Figure 9.6 Combined sensitivity and complementary sensitivity

function restrictions

the shaded area. Thus one will want to minimise the sensitivity function over a range of
frequencies, as well as attenuate the complementary sensitivity function for high frequencies.
Assuming that z C+ is a zero for RL , computations like those used to prove Corollary 9.16
then give
kSL k

1
1

z (1 )
z (2 )z (1 )

z (2 )

1
1 z (2 )
z (1 )
B z (2 )z (1 ) .
SL
1 + 2

(9.7)

The implications of this lower bound are examined in a simple case in Exercise E9.4.

9.3 The robust performance problem

So-called H design has received increasing attention in the recent control literature.
The basic methodology has its basis in the frequency response ideas outlined in Section 8.5.2
and Section 9.2. The idea is that, motivated by Proposition 8.24, one wishes to minimise
kSL k . However, as we saw in Section 9.2, this is not an entirely straightforward task.
Indeed, certain plant featuresunstable poles and/or nonminimum phase zeroscan make
this task attain a subtle nature. In this section we look at some ways of getting around these

9.19 Reasons why nave sensitivity minimisation fails

1. If RP is strictly proper, and we
wish to design a proper controller RC , then the loop gain RL = RC RP is strictly
proper. Thus TL is also strictly proper. Therefore, lim |TL (i)| = 0 from which it
follows (from the fact that SL + TL = 1) that lim |SL (i)| = 1. Thus, in such a
setting, we cannot possibly specify a H -norm bound for SL which is less than 1.
2. Based upon Proposition 8.22 we have two situations that provide limitations on possible
sensitivity reduction.
(a) If RP and RC have no poles in C+ and if either have a zero in C+ , then RL is
analytic in C+ and has a zero z C+ . From Proposition 8.22 we have SL (z) =
1. Thus it follows from the Maximum Modulus Principle (Theorem D.9) that
|SL (i)| 1 for some R.
(b) This is much like the previous situation, except we reflect things about the imaginary axis. Thus, if RP and RC have no poles in C and if either have a zero in
C , then RL is analytic in C and has a zero z C . Arguing as above, we see
that |SL (i)| 1 for some R.
3. If the relative degree of RL = RC RP exceeds 1 and if RL is strictly proper, then, from
Theorem 9.7 we know that kSL k > 1. Furthermore, this same result tells us that
this effect is exacerbated by RL having unstable poles.

22/10/2004

9.3 The robust performance problem

383

4. If RL is nonminimum phase and if we minimise |SL | over a certain range of frequencies,

then the result will be an increase in kSL k .
5. The previous point is reiterated in the various corollaries to the Poisson integral formulae given in Theorem 9.13. Here the location of nonminimum phase zeros is explicitly
seen in the estimate for the H -norm for the sensitivity function.

The above, apart from providing a neat summary of some of the material in Section 9.2,
indicates that the objective of minimising |SL (i)| over the entire frequency range is perhaps
not a good objective. Whats more, from a practical point of view, it is an excessively
stringent condition. Indeed, remember why we want to minimise the sensitivity function: to
reduce the L2 -norm of the error to a given input signal. However, in a given application,
one will often have some knowledge about the nature of the input signals one will have to
deal with. This knowledge may come in the form of, or possibly be converted into the form
of, frequency response information. Let us give some examples of how such a situation may
arise.
9.20 Examples 1. Suppose that we are interested in tracking sinusoids with frequencies in a
certain range with minimal error. Thus we take a class of nominal signals to be sinusoids
of the form
{ r(t) = a sin t | 0 a 1} .
Now we bias a subset of these inputs as follows:
Lref = { |Wp (i)a| sin t | 0 a 1} ,
where Wp R(s) is a function for which the magnitude of restriction to the imaginary
axis captures the behaviour we wish be giving more weight to certain frequencies. With
this set of inputs, the maximum error is
sup kek = sup |Wp (i)SL (i)|

rLref

= kWp SL k .
Note here that the measure of performance we use is the L -norm of the error. In the
following two examples, different measures will be used.
2. One can also think of specifying the character of reference signals by providing bounds
on the H2 -norm of the signal. Thus we could think of nominal signals as being those of
the form
{ rnom : [0, ) R | k
rnom k2 1} ,
and these are shaped to a set of possible reference signals

384

9 Design limitations for feedback control

22/10/2004

If we wish to minimise the L2 -norm of the error, then the maximum error is given by

sup kek2 = sup kWp SL rnom k2 rnom 1
rLref

= kWp SL k ,
where we have used part (i) of Theorem 5.21. Again, we have arrived at a specification
of the form of kWp SL k < .
3. The above argument can be carried out for nominal reference signals
{ rnom : [0, ) R | pow(
rnom ) 1} .
Similar arguments, now asking that the power spectrum of the error be minimised, lead in
the same way (using part (ix) of Theorem 5.21) to a condition of the form kWp SL k < .
4. Suppose that one knows from past experience that a controller will perform well if the
frequency response of the sensitivity function lies below that of a given transfer function.
Thus one would have a specification like
|SL (i)| |R(i)| ,

> 0,

(9.8)

where R R(s) comes from somewhere or other. In this case, if we define Wp (s) =
R(s)1 , this turns (9.8) into a condition of the form kWp TL k < 1.

Note that all specifications like kWp SL k < are by simple scaling transfered to a condition
of the form kWp SL k < 1. This is the usual form for these conditions to take, in practice.
The above examples present, in sort of general terms, possible natural ways in which
one can arrive at performance criterion of the form kWp SL k < 1, for some Wp R(s). In
making such specifications, one in only interested in the magnitude of Wp on the imaginary
axis. Therefore one may as well suppose that Wp has no poles or zeros in C+ .1 In this book,
we shall alway deal with specifications that come in the form kWp SL k < 1. Note also that
one might use the other transfer functions
RC RP
RC
RP
, RC S L =
, RP S L =
1 + RC RP
1 + RC RP
1 + RC RP
for performance specifications in the case when the loop gain RL is the product of RC with
RP . There is nothing in principle stopping us from using these other transfer functions; our
choice is motivated by a bald-faced demand for simplicity. Below we shall formulate criterion
that use our performance conditions of this section, and these results are complicated if one
uses the any of the other transfer functions in place of the sensitivity function.
There is a readily made graphical interpretation of the condition kWp SL k < 1, mirroring
the pictures in Figures 7.21 and 7.25. To see how this goes, note the following:
TL =

Lref = { r : [0, ) R | r = Wp rnom , k

rnom k2 1} ,

where Wp is a specified transfer function that shapes the energy spectrum of the signal
to a desired shape. Simple manipulation then shows that

Z

1
r(i) 2
Lref = r : [0, ) R

d 1 .
2 Wp (i)

kWp SL k < 1
|Wp (i)SL (i)| < 1, R
W (i)

p

< 1, R
1 + RL (i)
|Wp (i)| < |1 + RL (i)| , R
|Wp (i)| < |1 RL (i)| , R.

1
If Wp has, say, a zero z C+ , then we can write Wp = (s z)k W (s) where W (z) 6= 0. One can then
p (s) = (s + z)k W (s) has the same magnitude as Wp on the imaginary
easily verify that the new weight W
axis. Doing this for all poles and zeros, we see that we can produce a function with no poles or zeros in C+
that has the same magnitude on iR.

22/10/2004

9.3 The robust performance problem

385

Now note that |1 RL (i)| is the distance away from the point 1+i0 of the point RL (i)
on the Nyquist contour for RL . Thus the interpretation we make is that the Nyquist contour at frequency remain outside the closed disk centred at 1 + i0 of radius |Wp (i)|.
This is depicted in Figure 9.8. Note that we could just as well have described the con-

|Wp (i)|

1 + i0

RL (i)

386

9 Design limitations for feedback control

22/10/2004

P , Wu ) (resp. P+ (R
P , Wu )) and
(a) RC provides robust stability for P (R

(b) Wp SL < 1.
P , Wu ) (resp. P+ (R
P , Wu )) relative to
(ii) RC provides robust performance for P (R
Wp if
P , Wu ) (resp. P+ (R
P , Wu )) and
(a) RC provides robust stability for P (R

(b) kWp SL (RP )k < 1 for every RP P (RL , Wu ).

Thus by nominal performance we mean that the controller provides robust stability, and
meets the performance criterion for the nominal plant. However, robust performance requires
that we have robust stability and that the performance criterion is met for all plants in the
uncertainty set. Thus nominal performance consists of the two conditions

Wu TL < 1, Wp SL < 1

for multiplicative uncertainty, and

Wu RC SL < 1,

Wp SL < 1

for additive uncertainty. To interpret robust performance, note that for multiplicative perturbations we have
1
1 + RC RP
1
=
P
1 + RC (1 + )R

SL (RP ) =
Figure 9.8 Interpretation of weighted performance condition

dition kWp SL k < 1 just as we did in Figures 7.21 and 7.25 by saying that the circle
of radius |Wp (i)| and centre RL (i) does not contain the point 1 + i0. However, the
interpretation we have given has greater utility, at least as we shall use it. Also, note
that one can make similar interpretations using any of the other performance conditions
|Wp TL | , |Wp RC SL | , |Wp RP SL | < 1.
9.3.2 Nominal and robust performance Now that we have a method for producing frequency domain performance specifications that we can hope are manageable, let us
formulate some problems based upon combining this strategy with the uncertainty models
of Section 4.5 and the notions of robust stability of Section 7.3. The situation is roughly
this: we have a set of plants P and a performance weighting function Wp that gives a performance specification kWp SL k < 1. We wish to examine questions dealing with stabilising
all plants in P while also meeting the performance criterion.
The following definition makes precise the forms of stability possible in the framework
just described.
P R(s) be proper and suppose that RC R(s) renders IBIBO stable
9.21 Definition Let R
the interconnection of Figure 9.7. Let SL and TL denote the corresponding sensitivity and

closed-loop transfer functions. Let Wu RH+

and recall the definitions of P (RP , Wu )
P , Wu ) and for RP P (R
P , Wu ) or RP P+ (R
P , Wu ) denote by SL (RP ) and
and P+ (R
TL (RP ) the corresponding sensitivity and closed-loop transfer functions, respectively. Let
Wp R(s) have no poles or zeros in C+ .
P , Wu ) (resp. P+ (R
P , Wu )) relative
(i) RC provides nominal performance for P (R
to Wp if

=
=

1
P
1+RC R
P
RC R
+ 1+R

C RP

SL
,
1 + Wu TL

(9.9)

P , Wu ). A similarly style calculawhere is the allowable perturbation giving RP P (R

tion for additive perturbations gives
SL (RP ) =

SL
,
1 + Wu RC SL

(9.10)

P , Wu ).
where is the allowable perturbation giving RP P+ (R
To state a theorem on robust performance we need some notation. For R1 , R2 R(s) we
define a R-valued function of s by
s 7 |R1 (s)| + |R2 (s)| .
We denote this function by |R1 | + |R2 |, making a slight, but convenient, abuse of notation.
Although this function is not actually in the class of functions for which we defined the
H -norm, we may still denote

|R1 | + |R2 | = sup |R1 (i)| + |R2 (i)| .

This notation and the calculations of the preceding paragraph are useful in stating and
proving the following theorem.

9.3 The robust performance problem

22/10/2004

387

P R(s) be proper and suppose that RC R(s) renders IBIBO stable

9.22 Theorem Let R
the interconnection of Figure 9.7. Let SL and TL denote the corresponding sensitivity and

closed-loop transfer functions. Let Wu RH+

(i) R
C provides robust
performance for P (RP , Wu ) relative to Wp if and only if
|Wu TL | + |Wp SL | < 1;

P , Wu ) relative to Wp if and only if

(ii) R
performance for P+ (R
C provides robust
|Wu RC SL | + |Wp SL | < 1;

|Wu (i)TL (i)| + |Wp (i)SL (i)| < 1, R

From this it follows that

W S

Wp SL

p L

.
1 |Wu TL |
1 + Wu TL

From (9.11) it now follows that

Wp SL

< 1,
1 + Wu TL

388

9 Design limitations for feedback control

22/10/2004

is maximum. Let = ]Wu (i

)TL (i
). Using Lemma 1 from the proof of Theorem 7.18,
let G RH+
have
the
property
that ]G (i
) = and |G (i
)| = 1. Then = G is

allowable and the quantity (i

|Wu (i)RC (i)SL (i)| + |Wp (i)SL (i)| < 1, R

1 |Wu (i)RC (i)SL (i)| |1 + (i)Wu (i)RC (i)SL (i)|

for all R from which it follows that

Wp SL
Wp SL

1>

.
1 |Wu RC SL |
1 + Wu RC SL

and robust performance now follows from (9.9).

P , Wu ) relative to Wp .
Now suppose that RC provides robust performance for P (R
From (9.9) and Theorem 7.18 it follows that
W S

p L
kWu TL k < 1,
<1

1 + Wu TL

Robust performance now follows from (9.10).

P , Wu ) relative to Wp .
Now suppose that RC provides robust performance for P+ (R
By (9.10) and Theorem 7.21 this means that

Wp SL

kWu RC SL k < 1,
<1

1 + Wu RC SL

for all allowable . Let

0 be a frequency for which the function

for all allowable . Now let

0 be a frequency which maximises the function

|Wp (i)SL (i)|

1 |Wu (i)TL (i)|

|Wp (i)SL (i)|

.
1 |Wu (i)RC (i)SL (i)|

22/10/2004

9.3 The robust performance problem

389

390

9 Design limitations for feedback control

22/10/2004

Arguing as in part (i) we may find an allowable so that

1 |Wu (i
)RC (i)SL (i
)| = |1 + (i
)Wu (i
)RC (i)SL (i
)|.
|Wp (i)|

Again arguing as in (i) it follows that

1 + i0

This is an important theorem in SISO robust control theory, and it forms the basis
for many MIMO generalisations [see Dullerud and Paganini 1999]. It is useful because it
gives a single H -norm test for robust performance. For SISO systems this means that the
condition can be tested by producing Bode plots, and this is something that
is easily done.
For MIMO
systems, the matter of checking the conditions that generalise |Wu RC SL | +

|Wp SL | < 1 becomes a serious computational issue. In any event, the theorem allows
us to state a precisely formulated design problem from that simultaneously incorporates
stability, uncertainty, and performance.
9.23 Robust performance problem Given
P ,
(i) a nominal proper plant R
(ii) a function Wu RH+
,

P , Wu ) or P+ (R
P , Wu ), and
(iii) an uncertainty model P (R
(iv) a performance weight Wp R(s),
find a controller RC that
(v) stabilises the nominal system and

(vi) satisfies either |Wu TL | + |Wp SL | < 1 or |Wu RC SL | + |Wp SL | < 1, depending
on whether one is using multiplicative or additive uncertainty.

9.24 Remarks 1. The material leading up to the given statement in the robust performance
problem has its basis in the robust stability results of Doyle and Stein [1981] and Chen
and Desoer [1982], and seems to have its original statement in the book of Doyle, Francis,
and Tannenbaum [1990].
2. It is possible that the robust performance problem can have no solution. Indeed, it is
easy to come up with plant uncertainty models and performance weights that make the
problem unsolvable. An example of how to do this is the subject of Exercise E9.8.

3. A graphical
interpretation of the condition |Wu TL | + |Wp SL | < 1 (or |Wu RC SL | +

|Wp SL | < 1) is given in Figure 9.9. The picture says that for each frequency ,
the open disk
of radius |W

p (i)| centred at 1 + i0 should not intersect the open disk
L (i) centred at RL (i) (a similar statement holds, of course, for
of radius Wu (i)R
additive uncertainty).

In Chapter 15 we shall provide a way to find a solution to a slightly modified version
of the robust performance problem. In the stated form, it appears too difficult to admit a
simple solution. However, for now we can content ourselves with a couple of examples that
play with the problem in an ad hoc manner.
First we look at a case where we use multiplicative uncertainty.

L (i) or |Wu (i)RC (i)|

Wu (i)R

L (i)
R

Figure 9.9 Graphical interpretation of robust performance condition

9.25 Example (Example 7.20 contd) Recall that we have taken

P (s) = 1
R
s2
as our nominal plant and had used an uncertainty description of the form
Wu (s) =

as
,
s+1

a > 0.

In Example 7.20 we had concluded that provided that a < amax 34 , the controller
RC (s) = 1 + 2s +

1
s

P , Wu ). To make this into a robust performance problem, we need

robustly stabilises P (R
to provide a performance weight Wp . Let us suppose that we know that the reference signals
will have low energy below 10rad/ sec. Given our discussion in Example 9.20, one might say
that taking
Wp (s) =

1
2
s
10

1N,D (t)
(t)
(t)
fj (t)
Re
Im
x1
x2
22/10/2004
x1
x2

-20
-30

,0
0

391

-1.5

-1

-0.5

log

0.5

deg
u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

1.5

-20
-30

9.4 Summary
-1.5

2,0

-1

-0.5

log

0.5

1.5

in Figure 9.10, and one can see that it satisfies the robust performance constraint. Note
that not surprisingly the maximum allowable value for a is smaller than was the case in
Example 7.20 when we were merely looking to attain robust stability. The demand that our
controller also meet the performance specifications places upon it further restrictions. In the
current situation, if one wishes to allow greater variation in the set of plants contained in
the uncertainty description, one might look into backing off on the performance demands.
replacements
Next, we give an example alongPSfrag
similar
lines that uses additive uncertainty.

PSfrag replacements

t
t
ke(t)k
ke(t)k
e
e
1 (t)
1 (t)
9.26
Example (Example 7.23 contd) We proceed
along the lines of the previous example,
e2 (t)
e2 (t)
e3 (t)used the nominal plant and controller
e3 (t) carrying on from Example 7.23 where we
now
y(t)
y(t)
u(t)
u(t)
h (t)(s) = 1 + 2s + 1 .
h (t)
P (s) = 1 , R
R
C
hN,D (t)
s2 hN,D (t)
s
1 (t)
1 (t)
1
(t)
1N,D
(t)
N,D
To
model plant uncertainty we use
(t)
(t)
(t) as
(t)
fj (t)
fj (t)
Wu (s) =
,
Re(s + 1)2
Re
Im
Im
x1
x1
and
x2 we use the same performance weight Wpx2as in Example 9.25. With this data we give
L | + |Wxp1 SL | in Figure 9.11 for the case when a = 1.
x1 magnitude Bode plot for |Wu RC S
the
x2
x2
0

deg -20
u = 0 ln
ln coth(|u| /2) -40
or 1
-60
m

-80
yos
-100
tos 0
-120
dB

,0
0

-1.5

-1

-0.5

log

0.5

1.5

2,0
0

22/10/2004

-40

9 Design limitations for feedback control

The preceding example illustrate that the matter of checking a controller for robust
performance is largely a Bode plot issue. This is one thing that makes Theorem 9.22 a
valuable result in this day of the easily fabricated Bode plot.

-10

Figure 9.10 Bode plot of |Wu TL | + |Wp SL | for a = 1 (left) and for
a = 23 (right)

deg
u = 0 ln
-20
ln coth(|u| /2)
-40
or 1
m
-60

-80
yos
-100
tos 0

392

-10

-40

problem

deg
u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

9.3 The

1N,D (t)
(t)
(t)
fj (t)
Re
Im
x1
robustxx21 performance
x2

-1.5

-1

-0.5

log

0.5

1.5

Figure 9.11 Bode plot of |Wu RC SL | + |Wp SL | for a = 1 (left) and

for a = 14 (right)

Again, the easiest way to determine the upper bound on a is by trial and error. Doing so
gives amax 14 . In Figure 9.11 is shown the Bode plot of |Wu RC SL | + |Wp SL | for this value
of a, and one can see that it satisfies the bounds for robust performance.

When designing a controller, the first step is typically to determine acceptable performance criterion. In this chapter, we have come up with a variety of performance measures.
Let us review what we have said.
1.

Exercises for Chapter 9

393

Exercises
E9.1 Consider the pendulum/cart of Exercises E1.5 and E2.4. In this exercise, you will
use the cart position as the output, and you will take as the equilibrium position the
upright position for the pendulum.
(a) Compute the transfer function for the linearised system, and show that it has
both an unstable pole and a nonminimum phase zero.
(b) Choose parameter values for the system, then design a feedback vector f that
makes the closed-loop system IBIBO stable.
(c) For the closed-loop system, use Proposition 3.40 to compute the step response.
Verify that the bounds of Proposition 9.6 are satisfied.
E9.2 Poisson integral formulae for the pendulum on a cart.
Vidyasagar [1986] proposes a definition of undershoot different from the one we use, his definition being as follows. Let (N, D) be a BIBO stable, strictly proper, steppable SISO linear
(r)
system in input/output form and let r > 0 be the smallest integer for which 1N,D (0) 6= 0.
Note that limt 1N,D (t) = TN,D (0). The system (N, D) exhibits immediate undershoot
(r)
if 1N,D (0)TN,D (0) < 0. Thus, the system exhibits immediate undershoot when the initial
response heads off in a different direction that is attained by the steady-state response. In
the next exercise, you will explore this alternative definition of undershoot.
E9.3 Let (N, D) be a BIBO stable strictly proper, steppable SISO linear system in in(r)
put/output form and let r > 0 be the smallest integer for which 1N,D (0) 6= 0.
(a) Prove the following theorem of Vidyasagar [1986].
Theorem (N, D) exhibits immediate undershoot (in the sense of the above definition) if and only if N has an odd number of positive real roots.
Hint: Factor the numerator and denominator into irreducible factors, and use
the fact that
(r)
1N,D (0) = lim sr TN,D (s).
s

Now consider the system

(N (s), D(s)) = (s2 2s + 1, s3 + 2s2 + 2s + 2).
For this system, answer the following questions.
(b) Verify that this system is BIBO stable.
(c) According to the theorem of part (a), will the system exhibit immediate undershoot?
(d) Produce the step response for the system. Does it exhibit immediate undershoot
according to the definition of Vidyasagar? Would you say it exhibits undershoot?
E9.4 Let RL be a proper BIBO stable loop gain with the property that the standard unity
gain feedback loop (e.g., Figure 9.3) is IBIBO stable. Also suppose that RL has a single
real zero z C+ . We wish to have the sensitivity function fit under the shaded area
in Figure 9.6, and we take 1 = 34 2 . We also define 2 so that 2 is the closed-loop
bandwidth: 2 = 12 .
(a) For various values of 1 , plot the lower bound for kSL k given by (9.7) as a
function of zb .

394

9 Design limitations for feedback control

22/10/2004

(b) Would you expect better performance for the closed-loop system for larger of
smaller ratios zb ?
The following exercise should be done after you have designed a state feedback vector for
the pendulum/cart system. An arbitrary such state feedback vector is constructed in Exercise E10.11, and an optimal state feedback vector is determined in Exercise E14.7. For the
following exercise, you may use the parameter values, and the state feedback vector from
either of those exercises.
E9.5 Consider the pendulum/cart system of Exercises E1.5 and E2.4, and let f R4
be a stabilising state feedback vector, and consider the loop gain Rf as defined in
Exercise E7.11.
(a) For this loop gain, produce the magnitude Bode plot of the corresponding sensitivity and complementary sensitivity functions.
(b) Verify that the bounds of Corollaries 9.16 and 9.17 are satisfied, as well as that
of (9.7).
The next two exercises give conditions for robust performance for the uncertainty description
presented in Section 4.5, but that we have not discussed in detail in the text. The conditions
for robust stability for these descriptions you derived in Exercises E7.12 and E7.13. For these
plant uncertainty models, it turns out to be convenient to give the performance specifications
not on the sensitivity function SL , but on the closed-loop transfer function TL .
Thus the performance criterion for the following two exercises should take the
form kWp TL k < 1.
With this as backdrop, you may readily adapt the proof of Theorem 9.22 to prove the
conditions for robust performance in the next two exercises.
E9.6 For the plant uncertainty description of Exercise E7.12, and the performance criterion
kWp TL k < 1 (see above), show that a controller
RC stabilising the

nominal plant
P provides robust performance if and only if |Wu R
P SL | + |Wp TL | < 1.
R

E9.7 For the plant uncertainty description of Exercise E7.13, and the performance criterion
kWp TL k < 1 (see above), show that a controller
RC stabilising the nominal plant

P provides robust performance if and only if |Wu SL | + |Wp TL | < 1.
R

E9.8 Use the fact that SL + TL = 1 to show that in order for the robust performance
problem to have a solution, it must be the case that
min{Wp (i), Wu (i)} < 1,

Exercises for Chapter 9

395

This version: 22/10/2004

Part III
Controller design

398

This version: 22/10/2004

10 Stabilisation and state estimation

22/10/2004

10.1 Stabilisability and detectability

Chapter 10
Stabilisation and state estimation
While Chapter 5 dealt with various types of stability, and Chapter 6 provided a general
setting, with some specialisations in the later sections, for feedback, in this chapter we
combine feedback and stability to get stabilisation. The idea is quite simple: one wishes to
consider feedback that leaves a closed-loop system stable, or perhaps stabilises an unstable
system. In this chapter we also touch upon the matter of state estimation. The need for this
arises in practice where one can only measure outputs, and not the entire state. Therefore, if
one wishes to design feedback laws using the state of the system, it is necessary to reconstruct
the state from the output.
This is our first chapter concerned with controller design. As such, the design issue with
that we are concerned is merely stability. Design for performance is dealt with in later
chapters. An important outcome of this chapter is the parameterisation in Section 10.3 of
all stabilising dynamic output feedback controllers.

Contents
10.1 Stabilisability and detectability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 398
10.1.1 Stabilisability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 398
10.1.2 Detectablilty . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 400
10.1.3 Transfer function characterisations of stabilisability and detectability . . . . . . . 403
10.2 Methods for constructing stabilising control laws . . . . . . . . . . . . . . . . . . . . . . . 405
10.2.1 Stabilising static state feedback controllers . . . . . . . . . . . . . . . . . . . . . . 406
10.2.2 Stabilising static output feedback controllers . . . . . . . . . . . . . . . . . . . . . 408
10.2.3 Stabilising dynamic output feedback controllers . . . . . . . . . . . . . . . . . . . 416

In Chapters 2 and 3, we saw some interconnections between controllability, observability,

and pole/zero cancellations in the transfer function. At the time, we did not pay too much
attention to the nature of the poles and zero that were cancelled. In fact, the illustrative examples of Section 2.3 were cooked with the goal in mind of illustrating the possible disastrous
effects that lack of controllability and observability can have. This need not always be the
case. Indeed, it is possible that the a system can be both uncontrollable and unobservable,
yet be a system that is tolerable. In this section, we provide the language that expresses the
form of this tolerability.
10.1.1 Stabilisability In Theorem 6.49 we saw that if a system = (A, b, ct , D) is
controllable, then it is possible by state feedback to render the closed-loop system internally
asymptotically stable, and so make the system BIBO stable. The controllability hypothesis
is not necessary, and this is captured by the notion of stabilisability. To wit, the system is
stabilisable if Ss () 6= . That is, is stabilisable if there exists f Rn so that A bf t
is Hurwitz. Note that stabilisability depends only on (A, b), so we may sometimes say that
(A, b) is stabilisable rather than saying is stabilisable. The following result describes
stabilisability.
10.1 Proposition Let = (A, b, ct , D) be a SISO linear system and let T Rnn be invertible with the property that

b1
A11 A12
, Tb =
(10.1 )
T AT 1 =
0n`
0n`,` A22
where (A11 , b1 ) is in controller canonical form (cf. Theorem 2.39). The following are equivalent:
(i) is stabilisable;
(ii) A22 is Hurwitz;
(iii) the matrix

sI n A b
has rank n for all s C+ .
Proof (i) = (ii) Let f Ss (). Then T (A bf t )T 1 is Hurwitz. Now write

10.3 Parameterisation of stabilising dynamic output feedback controllers . . . . . . . . . . . . 420

T t f = (f 1 , f 2 ) R` Rn` .

10.3.1 More facts about RH+

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 421
10.3.2 The Youla parameterisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 423
10.4 Strongly stabilising controllers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 429
10.5 State estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 429
10.5.1 Observers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 429
10.5.2 Luenberger observers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 431
10.5.3 Static state feedback, Luenberger observers, and dynamic output feedback . . . . 433
10.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 438

We then have
T (A bf t )T 1 =

A11 b1 f t1 A12 b1 f t2
.
0n`,`
A22

This matrix is Hurwitz if and only if A11 b1 f t1 and A22 are Hurwitz, and that A22 is
Hurwitz is our assertion.
1

(ii) = (iii)
Let us define
A and b to be the expressions for T AT and T b in (10.1).n

The matrix sI n A b has rank n exactly when there exists no nonzero vector x R
with the property that

= xt (sI n A)
xt b = 0tn 0
b
xt sI n A

22/10/2004

10.1 Stabilisability and detectability

399

So suppose that x has the property that this equation does hold for some s0 C+ . Let us
write x = (x1 , x2 ) R` Rn` . Thus we have

A12
b1
= xt1 xt2 s0 I ` A11
b
xt s 0 I n A
0n`,`
s0 I nk` A22 0n`

= xt1 (s0 I ` A11 ) xt1 A12 + s0 xt2 (I nk` A22 ) xt1 b1 ,

400

10 Stabilisation and state estimation

2. Now we consider an example that is not controllable but is stabilisable. We define

= (A, b, ct , 01 ) by

0 1
1
0
A=
, b=
, c=
.
1 0
1
1
We compute
C(A, b) =

so that

xt1 (s0 I ` A11 ) = 0t` , xt1 b1 = 0,

(10.2)
xt1 A12 + s0 xt2 (I nk` A22 ) = 0tn`

Since (A11 , b1 ) is controllable, by Exercise E2.13 the matrix s0 I ` A11 b1 is full rank so
that the first two of equations (10.2) implies that x1 = 0` . The third of equations (10.2) then
says that x2 is a vector in the eigenspace of At22 with eigenvalue s0 C+ . However, since

b
A22 is Hurwitz this implies that x2 = 0n` . This shows that the matrix s0 I ` A
has rank n for s0 C+ . Now we note that

T 1
= T s0 I ` A b
b
.
s0 I ` A
1

and s0 I ` A b agree and the result follows.
b
Therefore, the ranks of s0 I ` A

(iii) = (i) This is Exercise E10.2.
The following corollary is obvious from the implication (i) = (ii) of the above result.
10.2 Corollary A SISO linear system = (A, b, ct , D) is stabilisable if it is controllable.
Let us explore these results with some examples.
10.3 Examples Note that if a system is controllable, then it is stabilisable. Therefore, interesting things concerning stabilisability will happen for uncontrollable systems.
1. Let us first consider a system that is not controllable and not stabilisable. We let =
(A, b, ct , 01 ) be defined by

0 1
1
0
A=
, b=
, c=
.
1 0
1
1
We compute

C(A, b) =

1 1
1 1

that has rank 1, so the system is indeed uncontrollable. To put the system into the
proper form to test for stabilisability, we use the change of basis matrix T defined by

1 1
T 1 =
.
1 1
Note that the first column of T 1 is the input vector b, and the other column is a vector
not collinear with b. We then compute

1 0
1
T AT 1 =
, Tb =
.
0 1
0
These are in the form of (10.1). Note that A22 = [1] which is not Hurwitz. Thus the
system is not stabilisable.

22/10/2004

1 1
1 1

so the system is uncontrollable. Now we define T by

1 1
T 1 =
.
1 1
Again note that b forms the first column of T . We also compute

1 0
1
T AT 1 =
, Tb =
,
0 1
0
which is in the form of (10.1). Now we have A22 = [1] which is Hurwitz, so the system
is stabilisable.

10.1.2 Detectablilty The notion of detectability is dual to stabilisability in the same
manner that observability is dual to controllability. But let us be precise. A SISO linear
system = (A, b, ct , D) is detectable if there exists a vector ` Rn with the property
that the matrix A `ct is Hurwitz. The following result is analogous to Proposition 10.1.
10.4 Proposition Let = (A, b, ct , D) be a SISO linear system and let T Rnn be invertible with the property that

A11 0nk,k
(10.3 )
T AT 1 =
, ct T 1 = ct1 0tnk
A21 A22
where (A11 , c1 ) are in observer canonical form (cf. Theorem 2.40). The following statements
are equivalent:
(i) is detectable;
(ii) A22 is Hurwitz;
(iii) the matrix

sI n A
t
c
has rank n for all s C+ .
Furthermore, the condition
(iv) there exists an output feedback constant F R so that the closed-loop system F is
internally asymptotically stable,
implies the above three conditions.
Proof (i) = (ii) Let ` Rn have the property that A `ct is Hurwitz. Writing T ` =
(`1 , `2 ) Rk Rnk we have

A11 `1 ct1 0k,nk
T (A `ct )T 1 =
.
t
A21 `2 c1 A22

22/10/2004

10.1 Stabilisability and detectability

401

This matrix is Hurwitz if and only if the matrices A11 `1 ct1 and A22 are Hurwitz. Therefore,
if A `ct is Hurwitz then A22 is Hurwitz as claimed.
and c
be the matrix and vector in (10.3). The matrix
(ii) = (iii) Let A

sI n A
t

c
has rank n if the only vector x Rn for which

0
sI n A
x= n
t
0

402

10 Stabilisation and state estimation

22/10/2004

10.6 Remark The content of the Propositions 10.1 and 10.4 is easily described in words. A
system is stabilisable if the uncontrollable dynamics, represented by the matrix A22 in (10.1),
are themselves asymptotically stable. Thus, even though the system may not be controllable,
this does not hurt you as far as your ability to render the system stable by static state
feedback. Similarly, a system is detectable if the unobservable dynamics, represented by
the matrix A11 in (10.3), are asymptotically stable. Therefore an unobservable system may
be made stable under static output feedback if it is detectable. The consequences of these

observations are explored in Section 10.2.
Let us explore these detectability results via examples.

is the zero vector. So let x be a vector for which the above equation is satisfied. Then,
writing x = (x1 , x2 ) Rk Rnk and letting s0 C+ , we have

s0 I k A11
0k,nk

x
s0 I n A

A21
s0 I nk A22 1
x=
t
x2

c
0tnk
ct1

(s0 I k A11 )x1

= (s0 I kn A22 )x2 A21 x1 .
ct1 x1

10.7 Examples Note that if a system is observable, then it is detectable. Therefore, interesting things concerning interesting things for detectability will happen for unobservable
systems.
1. Let us first consider a system that is not observable and not detectable. We let =
(A, b, ct , 01 ) be defined by

0 1
0
1
A=
, b=
, c=
.
1 0
1
1
We compute

The right-hand side is zero if and only if

(s0 I kn A22 )x2 A21 x1 = 0nk ,
(s0 I k A11 )x1 = 0k , ct1 x1 = 0.

O(A, c) =
(10.4)

Since (A11 , c1 ) is observable, the last two of equations (10.4) imply that x1 = 0k (see
Exercise E2.14). Now the first of equations (10.4) imply that x2 is in the eigenspace of A22
for the eigenvalue s0 C+ . However, since A22 is Hurwitz, this implies that x2 = 0nk .
Thus we have shown that if A22 is Hurwitz then the matrix

sI n A
t

c
has full rank. Now we note that

T sI n A
sI n A
=
T
1
ct
t
c

The following corollary follows from the implication (i) = (ii).

10.5 Corollary A SISO linear system = (A, b, ct , D) is detectable if it is observable.

which has rank 1, so the system is indeed unobservable. To put the system into the
proper form to test for detectability, we use the change of basis matrix T defined by

1 1
Tt =
.
1 1
Note that the first column of T t is the vector c itself, whereas the second column is a
vector in ker(ct ). We then compute

1 0
T AT 1 =
, ct T 1 = 1 0 .
0 1
These are in the form of (10.3). Note that A22 = [1] which is not Hurwitz. Thus the
system is not detectable.
2. Now we consider an example that is not observable but is detectable. We define =
(A, b, ct , 01 ) by

0 1
0
1
A=
, b=
, c=
.
1 0
1
1

so that if A11 is Hurwitz, it also follows that the matrix

sI n A
t
c
has full rank, as claimed.
(iii) = (i) This is Exercise E10.3.
(iv) = (i) This follows since ` = F b has the property that A `ct is Hurwitz.

1 1
1 1

We compute
O(A, c) =

1 1
1 1

so the system is unobservable. Now we define T by

1 1
Tt =
.
1 1

10.1 Stabilisability and detectability

22/10/2004

403

Again note that c forms the second column of T and that the first column is in ker(ct ).
We compute

1 0
T AT 1 =
, ct T 1 = 1 0 ,
0 1
which is in the form of (10.3). Now we have A22 = [1] which is Hurwitz, so the system
is detectable.

10.1.3 Transfer function characterisations of stabilisability and detectability The
results of the previous two sections were concerned with state-space characterisations of
stabilisability and detectability. In this section we look into how these may be manifested
in the transfer function. This treatment follows closely that of Section 3.3.
First let us look at the detectability result.
10.8 Proposition Let = (A, b, ct , 01 ) be a SISO linear system and define polynomials
P1 (s) = ct adj(sI n A)b,

P2 (s) = det(sI n A).

If (A, b) is controllable then is detectable if and only if the GCD of P1 and P2 has no
roots in C+ .
Proof We may as well assume that P1 and P2 are not coprime since the result follows from
Theorem 3.5 and Corollary 10.5 otherwise. Thus we may as well suppose that (A, c) are not
observable, and that

A11 A12
A=
, ct = 0tk ct2
0k,nk A22
where (A22 , c2 ) is observable. Therefore, if we write b = (b1 , b2 ) Rk Rnk then we
compute
ct (sI n A)b = ct2 (sI nk A22 )b.
Therefore

ct adj(sI n A22 )b2

ct adj(sI n A)b
= 2
.
det(sI n A)
det(sI nk A22 )

But we also have

det(sI n A) = det(sI k A11 ) det(sI nk A22 ),
from which we conclude that
ct adj(sI n A)b = det(sI k A11 )ct2 adj(sI n A22 )b2 .
Since (A22 , c2 ) is observable, the GCD of P1 and P2 must be exactly det(sI k A11 ). By
Proposition 10.4, the roots of the GCD are in C if and only if is detectable.

The consequences of the above result are readily observed in the detectability example
we have already introduced.
10.9 Examples (Example 10.7 contd)

0
A=
1

1. Here we had

1
0
1
, b=
, c=
.
0
1
1

404

10 Stabilisation and state estimation

22/10/2004

so that we compute
ct (sI 2 A)b = 1 s,

det(sI 2 A) = s2 1

The GCD of these polynomials is s 1 which has the root 1 C+ . Thus, as we have
seen, the system is not detectable.
2. Here we take

0 1
0
1
A=
, b=
, c=
1 0
1
1
and so compute
ct (sI 2 A)b = s + 1,

det(sI 2 A) = s2 1.

The GCD of these polynomials is s + 1 which has the single root 1 C , so that the
system is detectable.

Now let us give the analogous result for stabilisability.
10.10 Proposition Let = (A, b, ct , 01 ) be a SISO linear system and define polynomials
P1 (s) = ct adj(sI n A)b,

P2 (s) = det(sI n A).

If (A, c) is observable then is stabilisable if and only if the GCD of P1 and P2 has no
roots in C+ .
Proof The idea is very much like the proof of Proposition 10.8. We assume that (A, b) is
not controllable and that

A11 A12
b1
A=
, b=
0n`,` A22
0n`
with (A11 , b1 ) controllable. By arguments like those in the proof of Proposition 10.8 we
show that the GCD of P1 and P2 is det(sI n` A22 ). By Proposition 10.1 the roots of the
GCD are in C if and only if is stabilisable.

Again, we may use our existing stabilisability example to illustrate the consequences of
this result.
10.11 Examples (Example 10.3 contd) 1. First we take

0 1
1
0
A=
, b=
, c=
,
1 0
1
1
which is observable as it is in observer canonical form. We compute
ct adj(sI 2 A)b = 1 s,

det(sI 2 A) = s2 1.

The GCD of these polynomials is s 1 whose root is 1 C+ , leading us to the correct

conclusion that the system is not stabilisable.

10.2 Methods for constructing stabilising control laws

22/10/2004

2. Next we take
A=

0 1
,
1 0

1
,
1

405

0
.
1

406

10 Stabilisation and state estimation

22/10/2004

are often present other more demanding criterion given in the form of specific performance
criterion (see Chapter 8), or a demand for robustness of the controller to uncertainties in
the plant model (see Chapter 15).

We compute
ct adj(sI 2 A)b = s + 1,

det(sI 2 A) = s2 1,

so the GCD of these polynomials is s + 1 whose roots are in C . Thus we conclude that
the system is stabilisable.

Finally, we state a result that characterises stabilisability and detectability in terms of
cancellation of poles and zeros in the transfer function. This result is rather analogous to
Corollary 3.13.

10.2.1 Stabilising static state feedback controllers There is a method for systematically determining the state feedback vector f that will produce the desired poles for the
closed-loop transfer function. The formula is called Ackermanns formula [Ackermann
1972].
10.13 Proposition Let (A, b, ct , D) be a controllable SISO linear system and suppose that the
characteristic polynomial for A is
PA (s) = sn + pn1 sn1 + + p1 s + p0 .

10.12 Corollary Let = (A, b, ct , D) be a SISO linear system and define polynomials
P1 (s) = ct adj(sI n A)b,

P2 (s) = det(sI n A).

The following statements are equivalent:

(i) is stabilisable and detectable;
(ii) the GCD of P1 and P2 has no roots in C+ .

has the property that the characteristic polynomial of the matrix A bf t is P .

We comment that without additional information, one cannot decide whether a system
is not stabilisable or not detectable by simply look at the numerator and denominator of
the transfer function. This is made clear in Exercise E10.6. Also, note that now we can
complete the implications indicated in Figures 5.1 and Figure 5.2. The result is shown in
Figure 10.1.

internally
asymptotically
stable _g GG

+3 internally stable

GGGGG
GGGG
GGGG
GGGG
GGGG
GGGG
GGGG
G

Let P R[s] be monic and degree n. The state feedback vector f defined by

f t = 0 0 1 (C(A, b))1 P (A).

BIBO stable,
stabilisable, and
detectable
Figure 10.1 Precise relationship between internal and BIBO stability

Proof For the proof of this result, it is convenient to employ a different canonical form than
the controller canonical form of Theorem 2.37. From Exercise E2.32 we recall that if we
define

1 pn1
0
0
0
0
1
pn1 0
0

..
..
.. ,
. . ..
T 1 = ...

. .
.
.
.

0
0
0
1 pn1
0
0
0
0
1
1
then the matrix T = T 1
has the property that
1 (C(A, b))

pn1 pn2 p1 p0
1
0

0
0

0
1

0
0
T AT =
,
..
..
..
..
..
.
.
.
.
.
0
0

1
0

1
0

T b = 0 .
..
.
0

One readily verifies that the vector

f = (n1 pn1 , . . . , 1 p1 , 0 p0 )
has the property that the matrix T AT 1 T bf t has as characteristic polynomial
sn + n1 sn1 + + 1 s + 0 .

10.2 Methods for constructing stabilising control laws

In this section, under certain assumptions, we provide explicit formulas for constructing
stabilising controllers of various types. The formulas we give are most interesting in that
they show that it is in principle possible to explicitly design stabilising controllers of various
types. However, as a means for designing controllers, it should be emphasised that the
techniques we give here may not in and of themselves be that useful as they only address
one aspect of controller design; the necessity that the closed-loop system be stable. There

The actual state feedback vector will then be defined by f t = f t T .

= T AT 1 , and note that by Cayley-Hamilton we have
We denote A
n + pn1 A
n1 + + p1 A
+ p0 I n = 0n .
A
Suppose that
P (s) = sn + n1 sn1 + + 1 s + 0

22/10/2004

10.2 Methods for constructing stabilising control laws

407

so that

408

10 Stabilisation and state estimation

22/10/2004

To apply Proposition 10.13 we compute

n1 + + 1 A
+ 0 I n ,
=A
n + n1 A
P (A)
C(A, b) =

and subtracting from this the previous expression gives

= (n1 pn1 )A
n1 + + (1 p1 )A
+ (0 p0 )I n .
P (A)

(10.5)

One readily determines that if {e1 , . . . en } is the standard basis for R , then
etn Ak = etnk ,

In this case we compute f to be

f t = 0 1 C(A, b)1 P (A) = 3 4 .
This was the state feedback vector presented out of the blue in Example 6.50, and now we
see where it came from.

k = 1, . . . , n 1.

Therefore, using (10.5), we have

= (n1 pn1 )et + + (1 p1 )et + (0 p0 )et .
etn P (A)
1
n1
n
It remains to transform f back into the original coorBut this shows that f t = et1 P (A).
dinates to get the state feedback vector f . This we do by recalling that f t = f t T with
1
T = T 1
1 C(A, b) . Thus we arrive at
T = et P (T AT 1 )T = et T P (A)
f t = etn P (A)
n
n
where we have used the fact that P (T M T 1 ) = T P (M )T 1 for T , M Rnn with T
t
invertible. Now we see from Exercise E2.31 that etn T 1
1 = en , and so this proves that
1
t
1
f t = etn T 1
1 (C(A, b)) P (A) = en (C(A, b)) P (A),

as desired.

0 1
.
1 0

10.14 Remarks 1. The static state feedback of Proposition 10.13 gives an explicit controller
that stabilises the closed-loop system, provided that the system is controllable. There are,
of course, other stabilising state feedbacks. In fact, there are algorithms for computing
static state feedback vectors that place the poles in desirable places. Often an ad hoc
guess for pole locations will not work well in practice, as, for example, model inaccuracies
may have adverse effects on the behaviour of the actual system with a controller designed
as in Proposition 10.13. A popular method for designing stabilising static state feedback
vectors is the linear quadratic regulator (LQR), where the state feedback is defined
to have an optimal property. This is discussed in many books, a recent example of
which is [Morris 2000]. In Chapter 14 we look at this for SISO systems using polynomial
machinery.
2. In the proof of Proposition 10.13 we have made use of the canonical form developed in
Exercise E2.32. The matter of choosing which canonical form to employ is often a matter
of convenience.

Let us illustrate this technology with an example.
10.15 Example (Example 6.50 contd) We recall that in this example we had

0 1
0
A=
, b=
.
1 0
1
Suppose we wish the closed-loop characteristic polynomial to be s2 + 4s + 4 which has the
repeated root 2. Thus we compute

3 4
P (A) = A2 + 4A + 4I 2 =
.
4 3

In Proposition 10.13 the assumption of controllability is made. However, only the assumption of stabilisability need be made for stabilisation under static state feedback (pretty
much by the very definition of stabilisability). The details of how to construct a stabilising
static state feedback for a system that is stabilisable but not controllable is the subject of
Exercise E10.7. Let us here state the result.
10.16 Proposition If = (A, b, ct , D) is stabilisable, it is possible to explicitly construct a
feedback vector f with the property that the closed-loop system f is internally asymptotically
stable, i.e., so that f Ss ().
For this result, we also have the following corollary, that is the analogous result for
detectability.
10.17 Corollary If = (A, b, ct , D) is detectable, it is possible to explicitly construct a vector
` Rn with the property that the matrix A `ct is Hurwitz.
Proof Since is detectable, it follows from part (iii) of Proposition 10.1 and part (i) of
= (At , c, bt , D) is stabilisable. Therefore, by ProposiProposition 10.4 that the system
tion 10.16, there exists a vector f Rn so that At cf t is Hurwitz. Therefore, A f ct is
also Hurwitz, and the result follows by taking ` = f .

Note that the result merely tells us to construct a feedback vector as in Proposition 10.13
(or Exercise E10.7 if (A, c) is not observable) using At and c in place of A and b.
10.2.2 Stabilising static output feedback controllers Interestingly, the problem of
stabilisation by static output feedback is significantly more difficult than the problem of
stabilisation by static state feedback (see [Syrmos, Abdallah, Dorato, and Grigoriadis 1997]
for a fairly recent survey, and [Geromel, de Souza, and Skelton 1998] for convexity results).
For example, a system can be stabilised by static state feedback if and only if it is stabilisable,
and stabilisability is a condition one can computationally get a handle on. However, Blondel
and Tsitsiklis [1997] essentially show that the problem of determining whether a system is
stabilisable by static output feedback is NP-hard.1 Thus there is no easily computable check
1
Let us recall, for our own amusement, roughly the idea behind this NP-business. The class of P problems
are those for which there is a solution algorithm whose computational complexity satisfies a polynomial
bound. The class of problems NP are those with the property that every solution can be verified as actually
being a solution with an algorithm whose computational complexity satisfies a polynomial bound. A famous
open problem is, Does P=NP? This problem was posed by Cook [1971] and listed by Smale [1998] as
one of the most important problems in mathematics. A problem is NP-hard if every problem in NP can
be reduced to it. An NP-hard problem may not be in NP, and all known NP-hard problems are not
solvable by an algorithm whose complexity satisfies a polynomial bound. A problem is NP-complete if it
is NP-hard, and further is in NP. These are, in essence, the hardest of the problems in NP. All known
NP-complete problems are not solvable by an algorithm whose complexity satisfies a polynomial bound.

22/10/2004

10.2 Methods for constructing stabilising control laws

409

to see when it is even possible for a system to be stabilisable under static output feedback,
never mind an algorithm to compute a static output feedback that actually stabilises.
We shall therefore have to be satisfied with a discussion of static output feedback that
is not as complete as the corresponding discussion surrounding static state feedback. To
motivate what we do say, recall from Theorem 6.54 that one may express the form of the
closed-loop transfer functions available via static output feedback. We wish to determine
conditions to test whether the poles of the closed-loop transfer function are those of a given
polynomial.
Our first result gives a state space test, and is extracted from [van der Woude 1988]. For
a SISO linear system = (A, b, c, D), for k {1, . . . , n} define a n k matrix by

C k (A, b) = b Ab Ak1 b
and define a k n matrix by

O k (A, c) =

ct
ct A
..
.
ct Ak1

Thus C k (A, b) is comprised of the first k columns of C(A, b) and O k (A, c) is comprised of
the first k rows of O(A, c). With this notation, we have the following result.
10.18 Theorem Let = (A, b, c, D) be a complete SISO linear system, and let P R[s] be
a degree n monic polynomial. The following statements are equivalent:
(i) it is possible to choose an output feedback constant F so that the poles of the closed-loop
system F are the roots of P ;

(ii) for some k {1, . . . , n 1} we have P (A) ker(O k (A, c)) image(C nk (A, b)).

(iii) for all k {1, . . . , n 1} we have P (A) ker(O k (A, c)) image(C nk (A, b)).
Proof First recall from Theorem 6.54 that the poles of the closed-loop transfer function F
are the roots of
det(sI n (A F bct )).
First let us show that (i) is equivalent to (ii) with k = 1. First suppose that (i) holds.
Note that if x ker(ct ) then
(A F bct )x = Ax.
Therefore,
(A F bct )2 x = A2 x F bct Ax.
Thus (A F bct )2 x = A2 x + y 2 where y 2 span {b}. An easy induction now shows that
for j {2, . . . , n} we have
(A F bct )j x = Aj x + y j ,

where y j span b, Ab, . . . , Aj2 b . By the Cayley-Hamilton theorem we have
P (A F bct ) = 0n .
Combined with our above calculations this shows that for x ker(ct ) we have
0 = P (A F bct )x = P (A)x + y

410

10 Stabilisation and state estimation

22/10/2004

where y span b, Ab, . . . , An2 b . Thus we conclude that

P (A) ker(O 1 (A, c)) image(C n1 (A, b)),
and, therefore, that (ii) holds with k = 1.
Now suppose that (ii) holds with k = 1. We first claim that

image(C n1 (A, b)) = ker etn (C(A, b))1 ,

(10.6)

if en is the nth standard basis vector for Rn . Indeed, if x image(C n1 (A, b)) if and only
if the components of x in the basis {b, Ab, . . . , An1 b} are of the form (x1 , . . . , xn1 , 0).
However, these components are exactly (C(A, b))1 x:

x1
..
.

= (C(A, b))1 x.
xn1
0
The relation (10.6) now follows immediately, and by our current hypothesis, so does the
relation

P (A) ker(ct ) ker etn (C(A, b))1

ker(ct ) ker etn (C(A, b))1 P (A)
F ct = etn (C(A, b))1 P (A),
for some F R. Now, since (A, b) is controllable, by Proposition 10.13, if we define f Rn
by
f t = et (C(A, b))1 P (A),
then det(sI n (A bf t )) = P (s). This completes the proof of our assertion that (i) is
equivalent to (ii) with k = 1.
Next let us show that (ii) holds with k = 1 if and only if it holds with k = n 1. The
equivalence of (i) and (ii) with k = 1 tells us that there exists F R so that
P (s) = det(sI n (At cbt ))

(10.7)

if and only if

P (At ) ker(O 1 (At , b)) image(C n1 (At , c)).
Now we note that O 1 (At , b) = C 1 (A, b)t and C n1 (At , c) = O n1 (A, c)t . Thus there exists
an F so that (10.7) is satisfied if and only if

P (At ) ker(C 1 (A, b)t ) image(O n1 (A, c)t )
Let us make use of a lemma.
1 Lemma Let T Rnn , M Rnk , and L R(nk)n . Then the following statements are
equivalent:
(i) T t (ker(M t )) image(Lt );
(ii) T (ker(L)) image(M ).

22/10/2004

10.2 Methods for constructing stabilising control laws

411

Proof Assume that (i) holds. Statement (i) asserts that T t maps ker(M t ) to image(Lt ).
We claim that this implies that T maps the orthogonal complement of image(Lt ) to the orthogonal complement of ker(M t ). Indeed, if x is in the orthogonal complement to image(Lt )
and if y ker(M t ) then we have
hT x, yi = hx, T yi .
Since (i) holds, it follows that T y image(Lt ), showing that hT x, yi = 0 if x is in the
orthogonal complement to image(Lt ) and if y ker(M t ). That is, T maps the orthogonal
complement of image(Lt ) to the orthogonal complement of ker(M t ). However, the orthogonal complement of image(Lt ) is exactly ker(L) and the orthogonal complement of ker(M t )
is exactly image(M ). Thus we have shown that
T (ker(L)) image(M ),
thus (ii) holds. Clearly this proof is symmetric, so the converse holds trivially.

Applying the lemma to the case when T = P (A), u = C 1 (A, b), and L = O n1 (A, c) we
complete the proof of the fact that (ii) with k = 1 implies (ii) with k = n 1. The converse
implication is a matter of reversing the above computations.
To this point we have shown that the following three statements are equivalent:
1. it is possible to choose an output feedback constant F so that the poles of the closedloop system F are the roots of P ;

2. P (A) ker(O 1 (A, c)) image(C n1 (A, b));

3. P (A) ker(O n1 (A, c)) image(C 1 (A, b)).
We complete the proof by showing that if (ii) holds with k = ` then it also holds with
k = ` 1.
Without loss of generality, suppose that (A, c) are in observer canonical form so that

0 0 0 0 p0
b0
0
1 0 0 0 p1

b1
0

0 1 0 0 p2

b2
0

A = 0 0 1 0 p3 , b = .. , c = .. .
.. .. .. . . ..

.
..
.
. . .

. .
.

0
bn2
0 0 0 0 pn2
bn1
1
0 0 0 1 pn1

412

10 Stabilisation and state estimation

22/10/2004

We then compute

P (A) ker(O `1 (A, c)) = span y, Ay, . . . , An` y

= span y, Ay, . . . , An`1 y + A span y, Ay, . . . , An`1 y

= P (A) ker(O ` (A, c)) + AP (A) ker(O ` (A, c))

image(C n` (A, b)) + A image(C n` (A, b))
= image(C n`+1 (A, b)).
This completes the proof.

The theorem gives an insight into the difficultly in determining which closed-loop poles
are available to us. Unlike its static state feedback counterpart Proposition 10.13, the
conditions given by Theorem 10.18 for determining whether the closed-loop poles are the
roots of a polynomial P involve the polynomial P itself . This is what makes the problem a
computationally difficult one. Let us illustrate Theorem 10.18 on an example.
10.19 Example (Example 6.55 contd) We take

0 1
0
A=
, b=
,
1 0
1

3
,
4

D = 01 ,

noting that A and b are as in Example 10.15. For F R we determine that

det(sI 2 (A F bct )) = s2 + 4F s + 1 + 3F.
Thus, if we let PF (s) = s2 + 4F s + 1 + 3F , we see that the only places we may assign poles
under static output feedback are at the roots of PF for some F . Let us see how this checks
in with Theorem 10.18. A simple computation gives

3F 4F
PF (A) =
.
4F 3F
Because n = 2, the only cases that are applicable for statements (ii) and (iii) of Theorem 10.18
are when k = 1. We then have O 1 (A, c) = ct and C 1 (A, b) = b. Therefore
ker(O 1 (A, c)) = span {(4, 3)} ,

image(C 1 (A, b)) = span {(0, 1)} .

We compute

PF (A)

A direct computation than shows that

4
0
=
.
3
25F

and A`1 e1 = e` for ` {1, . . . , n 1}. If we let y = P (A)e1 it therefore follows that

P (A) ker(O ` (A, c)) = span y, Ay, . . . , An`1 y .

Therefore PF (A) ker(O 1 (A, c)) image(C 1 (A, b)), just as stated in Theorem 10.18.
Conversely, let P (s) = s2 + as + b and compute

b1
a
P (A) =
,
a b 1

Now assume that

and

ker(O ` (A, c)) = span {e1 , . . . , en` }

P (A) ker(O ` (A, c)) image(C n` (A, b)).

P (A)

4
4b 3a 4
=
.
3
3 4a 3b

10.2 Methods for constructing stabilising control laws

22/10/2004

Thus we have

413

if and only if 4b 3a 4 = 0. The linear equation

a
3 4
=4
b
does not have a unique solution, of course. One may check that (a, b) = (0, 1) is a solution.
All solutions are then of the form
x ker((3, 4)t ) = span {(4, 3)} .

That is to say, if the poles of the closed-loop system are to roots of P (s) = s2 + as + b, then
we must have (a, b) = (0, 1) + F (4, 3). Thus P (s) = s2 + 4F s + (1 + 3F ), just as we noticed
initially.

Note that Theorem 10.18 does not tell us when we may choose an output feedback
constant F with the property that the closed-loop system F is stable. The following
result of Kucera and de Souza [1995] gives a characterisation of stabilisability in terms of
the Liapunov ideas of Section 5.4 and the Riccati equations ideas that will come up in
Section 14.3.2.
10.20 Theorem For a SISO linear system = (A, b, ct , 01 ), the following two statements
are equivalent:
(i) So () 6= ;
(ii) the following two conditions hold:
(a) is stabilisable and detectable;
(b) there exists F R and g Rn so that

22/10/2004

= A F bct , b, C, 01 ) where
= (A
This means that the two output system
t
c
C=
F ct
is detectable since (A F bct ) LC is Hurwitz if

L = ` b .
From the MIMO version of Exercise E10.4, the result now follows.

10.21 Remark In Section 14.3.2 we will see that the equation (10.9) comes up naturally in
an optimisation problem. This equation is called an algebraic Riccati equation, and
there are well-developed numerical schemes for obtaining solutions; it is a nice equation
numerically. However, in the statement of Theorem 10.20 we see that not only must the
algebraic Riccati equation be satisfied, but the subsidiary condition (10.8) must also be met.
This, it turns out, takes the numerical problem out of the nice category in which the
algebraic Riccati equations sits. Again, with static output feedback, things are never easy.
Now let us adopt a different, more constructive approach. While the approach is more
constructive, it is not a sharp construction, as we shall see. First, the following result gives
a crude necessary condition for an output feedback constant to stabilise.
10.22 Proposition Let , P , and Q be as in Theorem 10.23 and write

F ct + bt P = g t ,

At P + P A P bbt P = cct gg t .

Define
(10.9 )

Proof (i) = (ii) Let F R have the property that A F bct is Hurwitz. Then clearly
(A, b) is stabilisable and (A, c) is detectable. Since A F bct is Hurwitz, by part (ii) of
Theorem 5.32 there exists a unique positive-semidefinite matrix P that satisfies
t

P (s) = sn + pn1 sn1 + + p1 s + p0

Q(s) = qm sm + + q1 s + q0 .

(10.8 )

where P is the unique positive-semidefinite solution of the equation

t t

10 Stabilisation and state estimation

Let ` R have the property that A `ct is Hurwitz. We then note that

ct
A `ct = (A F bct ) + ` b
t .
Fc

4b 3a 4
image(C 1 (A, b))
3 4a 3b

(0, 1) + x,

414

(A F bc ) P + P (A F bc ) = (1 + F )cc .

pj
qj > 0, j = 0, . . . , m
q
jpj

= max qj qj < 0, j = 0, . . . , m .

Flower = min
Fupper

If qj 0 for all j {0, 1, . . . , m} then take Flower = and if qj 0 for all j

{0, 1, . . . , m} then take Fupper = . If the closed-loop system F is internally asymptotically
stable for some F R, then it must be the case that F (Flower , Fupper ).

At P + P A P bbt P = cct (bt P F ct )t (bt P F ct ).

Proof By Theorem 6.54, the poles of the closed-loop transfer function F are the roots of
the polynomial
PF (s) = P (s) + F Q(s).

This part of the proof now follows by taking g = P bP F c.

(ii) = (i) Let F and g be as in the statement of the theorem. One then directly verifies
using the properties of F and g that

If PF is Hurwitz then all coefficients of PF be positive (see Exercise E5.18). In particular,

we should have
pj + F qj > 0, j = 1, . . . , m.

Straightforward manipulation of this expression gives

(A F bct )t P + P (A F bct ) = (1 + F 2 )cct .

This relation will be satisfied in F (Flower , Fupper ).

22/10/2004

10.2 Methods for constructing stabilising control laws

415

Thus we can now restrict our search for feasible output feedback constants to those
in the interval (Flower , Fupper ). However, we still do not know when an output feedback
constant F does stabilise. Let us address this by giving a result of Chen [1993]. To state the
result, we need the notion of a generalised eigenvalue. Given matrices M 1 , M 2 Rnn ,
a generalised eigenvalue is a number C that satisfies

416

10 Stabilisation and state estimation

10.25 Example ((10.19) contd) We resume our example where

0 1
0
3
A=
, b=
, c=
, D = 01 .
1 0
1
4
We compute
P (s) = s2 + 1,

det(M 1 M 2 ) = 0.
We denote by (M 1 , M 2 ) the set of generalised eigenvalues. We also recall from Section 5.5.2
the n n Hurwitz matrix H(P ) that we may associate to a polynomial P R[s] of degree
n. Note that if Q R[s] has degree less than n, then we may still define an n n matrix
H(Q) by thinking of Q as being degree n with the coefficients of the higher order terms
being zero.
10.23 Theorem Let = (A, b, ct , 01 ) be a controllable SISO linear system. Denote
P (s) = det(sI n A),

22/10/2004

Q(s) = 4s + 3,

which gives
Flower = 0,

Fupper = .

One may also determine that

H(P ) =

0 1
,
0 1

H(Q) =

4 0
.
0 3

We then compute
det(H(P ) + H(Q)) = 12s2 + 4s.

Q(s) = c adj(sI n A)b

Thus

with H(P ) and H(Q) the corresponding Hurwitz matrices, supposing both to be n n. Let

(H(P ), H(Q)) R = 13 , 0 ,

(H(P ), H(Q)) R = {1 , . . . , k }

so 1 = 13 and 2 = 0. We note that (1 , 2 ) 6 (Flower , Fupper ), so from Theorem 10.23

we may only conclude that there are no stabilising output feedback constants in (1 , 2 ).
However, note that in this example, any F (Flower , Fupper ) is actually stabilising. Thus
while Theorem 10.23 provides a way to perhaps obtain some stabilising output feedback
constants, it does not provide all of them. This further points out the genuine difficulty of
developing a satisfactory theory for stabilisation using static output feedback, even in the
SISO context.

with 1 k . The following statements hold:

(i) for i {1, . . . , k}, the closed-loop system i is not internally asymptotically stable;
(ii) for j {1, . . . , k 1}, the closed-loop system F is internally asymptotically stable
for all F (j , j+1 ) if and only if F is internally asymptotically stable for some
F (j , j+1 ).
We do not present the proof of this theorem, as the elements needed to get the proof
underway would take us too far afield. However, let us indicate how one might use the
theorem.
10.24 Method for generating stabilising output feedback constants Suppose you are given =
(A, b, ct , D).
1. Compute
P (s) = det(sI n A), Q(s) = ct adj(sI n A)b.
2.
3.
4.
5.
6.

Compute Flower and Fupper as in Proposition 10.22.

Compute H(P ) and H(Q) as in Theorem 10.23.
Compute and order {1 , . . . , k } = (H(P ), H(Q)) R.
If (j , j+1 ) 6 (Flower , Fupper ), then any F (j , j+1 ) is not stabilising.
If (j , j+1 ) (Flower , Fupper ), then choose F (j , j+1 ) and check whether PF (s) =
P (s) + F Q(s) is Hurwitz (use, for example, the Routh/Hurwitz criterion).
7. If PF is Hurwitz, then PF is Hurwitz for any F (j , j+1 ).

Let us try this out on a simple example.

10.26 Remark While the above discussion suggests that obtaining a fully-developed theory
for stabilisation by static output feedback may be troublesome, in practice, things are not
as grim as they have been painted out to be, at least for SISO systems. Indeed, the matter
of finding stabilising output feedback constants is exactly the problem of finding constants
F so that the polynomial
PF (s) = P (s) + F Q(s)
is Hurwitz. A standard way to do this is using root-locus methods developed by Evans [1948,
1950], and presented here in Chapter 11. It is also possible to use the Nyquist criterion to
obtain suitable values for F . Note, however, that both of these solution methods, while
certainly usable in practice, are graphical, and do not involve concrete formulas, as do the
corresponding formulas for static state feedback in Section 10.2.1 and for dynamic output
feedback in Section 10.2.3. Thus ones belief in such solutions methods is exactly as deep as
ones trust in graphical methods.

10.2.3 Stabilising dynamic output feedback controllers In this section we will show
that it is always possible to construct a dynamic output feedback controller that renders
the resulting closed-loop system internally asymptotically stable, provided that the plant
is stabilisable and detectable. This is clearly interesting. First of all, we should certainly
expect that we will have to make the assumption of stabilisability and detectability. If these
assumptions are not made, then it is not hard to believe that there will be no way to make the

22/10/2004

10.2 Methods for constructing stabilising control laws

417

plant internally asymptotically stable under feedback since the plant has internal unstable
dynamics that are neither controlled by the input nor observed by the output.
First we recall that if P = (AP , bP , ctP , D P ) is stabilisable and detectable there exists
two vectors f , ` Rn with the property that the matrices AP bP f t and AP `ctP are
Hurwitz. With this as basis, we state the following result.
10.27 Theorem Let P = (AP , bP , ctP , D P ) be a SISO linear plant. Then the following
statements are equivalent:
(i) P is stabilisable and detectable;
(ii) there exists a SISO linear controller C = (AC , bC , ctC , D C ) with the property that the
closed-loop system is internally asymptotically stable.
Furthermore, if either of these equivalent conditions is satisfied and if f , ` Rn have the
property that the matrices AP bP f t and AP `ctP are Hurwitz, then the SISO linear
controller C defined by
AC = AP `ctP bP f t + D P `f t ,
bC = `, ctC = f t , D = 01
has the property that the closed-loop system is internally asymptotically stable.
Proof (i) = (ii) Using Proposition 6.56 we compute the closed-loop system matrix Acl to
be

AP
bP f t
.
Acl =
`ctP AP `ctP bP f t
Now define T R2n2n by

I I
T = n n =
0n I n

T 1

I I n
= n
.
0n I n

One readily computes that

T Acl T 1 =

AP `ctP
0n
.
A bf t
`ctP

In particular,
spec(Acl ) = spec(AP `ctP ) spec(A bf t ),
which says that the eigenvalues of Acl are in C , as desired, provided we choose f and `
appropriately. This also proves the second assertion of the theorem.
(ii) = (i) If P is neither stabilisable nor detectable, it is also neither controllable nor
observable. Therefore, by Theorem 2.41 we may suppose that AP , bP , and cP have the form

A11 A12 A13 A14
b1
0j
c2
0kj A22 0j` A24
b2

,
AP =
, bP = , cP =
0`j 0`k A33 A34
0`
0`
0mj 0mk 0m` A44
0m
c4
for suitable j, k, `, and m. The assumption that P is neither stabilisable nor detectable is
equivalent to saying that the matrix A33 is not Hurwitz. First let us suppose that A33 has

418

10 Stabilisation and state estimation

22/10/2004

a real eigenvalue C+ with v an eigenvector. Consider a controller SISO linear system

C = (AC , bC , ctC , D C ), giving rise to the closed-loop equations
x P (t) = AP xP (t) + bP u(t)
x C (t) = AC xC (t) bC y(t)
y(t) = ctP xP (t) + D P u(t)
u(t) = ctC xC (t) D C y(t).
If we choose initial conditions for xP and xC as

0j
0k

xP (0) =
v , xC (0) = 0,
0m
the resulting solution to the closed-loop equations will simply be

0j

t 0k
xP (t) = e , xC (t) = 0.
v
0m
In particular, the closed-loop system is not internally asymptotically stable. If the eigenvalue

in C+ is not real, obviously a similar argument can be constructed.
10.28 Remark The stabilising controller C constructed in the theorem has the same order, i.e., the same number of state variables as the plant P . It can be expected that
frequently one can do much better than this and design a significantly simpler controller.
In Section 10.3 we parameterise (almost) all stabilising controllers which includes the one of

Theorem 10.27 as a special case.
This gives the following interesting corollary that relates to the feedback interconnection
of Figure 10.2. Note that this answer the question raised at the end of Section 6.3.1.
r(s)

RC (s)

RP (s)

y(s)

Figure 10.2 Unity gain feedback loop

10.29 Corollary Let RP R(s) which we assume to be represented by its c.f.r. (NP , DP ).
Then it is possible to compute a controller rational function RC with the property that the
closed-loop interconnection of Figure 10.2 is IBIBO stable.
Proof This follows immediately from Theorem 10.27 since the canonical minimal realisation
of RP is controllable and observable, and so stabilisable and detectable.

22/10/2004

10.2 Methods for constructing stabilising control laws

419

Let us see how this works in an example. We return to the example of Section 6.4.3,
except now we do so with a methodology in mind.
10.30 Example (Example 6.59) In this example we had

1
0 1
0
,
AP =
, bP = 1 , cP =
0
0 0
m

D P = 01 .
t

AP `ctP

10 Stabilisation and state estimation

22/10/2004

TP (s) = ms1 2PSfrag

. This
is not, of course, a BIBO stable transfer function. However, if we use
replacements
t
the controller rational
function
ke(t)k
e1 (t)
4m(2s + 1)
e2 (t)
,
RC (s) = TC (s) = 2
e3 (t)
s + 4s + 8
y(t)
then we are guaranteedu(t)
by Corollary 10.29 that the closed-loop configuration of Figure 10.2
h (t)
is IBIBO stable. In hFigure
10.3 is shown the Nyquist plot for RL = RC RP . Note that the
N,D (t)
1 (t)
1N,D (t)
(t)
(t)
fj (t)

2
x1
x2
x1
x2
log
0
dB
deg
u = 0 ln
ln coth(|u| /2)
-2
or 1
m

yos
-4
tos 0

To apply Theorem 10.27 we need to design vectors f and ` so that AP bP f and

are Hurwitz. To do this, it is required that (AP , bP ) be stabilisable and that (AP , cP ) be
detectable. But we compute

0 1
1 0
C(A, b) =
, O(A, c) =
.
1 0
0 1

420

Thus P is controllable (and hence stabilisable) and observable (and hence detectable). Thus
we may use Proposition 10.13 to construct f and Corollary 10.17 to construct `. In each
case, we ask that the characteristic polynomial of the closed-loop matrix be (s) = s2 + 2s2
which has roots 1 i. Thus we define

f t = 0 0 1 C(A, b)1 (A) = 2m 2m

`t = 0 0 1 O(At , c)1 (At ) = 2 2 .

,0
0

-4

-2

Here, following Proposition 10.13, we have used

(A) = A2 + 2A + 2I 2 ,

(At ) = (At )2 + 2At + 2I 2 .

Using Theorem 10.27 we then ascertain that

2 1
AC = AP `ctP bP f t =
4 2

2
bC = ` =
, ctC = f t = 2m 2m .
2
Let us check that the closed-loop system, as defined by Proposition 6.56, is indeed internally
asymptotically stable. A straightforward application of Proposition 6.56 gives

0 1 0
0
0 0 2
2

Acl =
2 0 2 1 ,
2 0 4 2

0
0

t

bcl =
2 , ccl = 1 0 0 0 , D cl = 01 .
2
One may check that the eigenvalues of the Acl are {1 i} where each root is has algebraic
multiplicity 2. As predicted by the proof of Theorem 10.27, these are the eigenvalues of
AP bP f t and AP `ctP .
Now let us look at this from the point of view of Corollary 10.29. Instead of thinking
of the plant as a SISO linear system P , let us think of it as a rational function RP (s) =

Figure 10.3 Nyquist plot for dynamic output feedback problem

Nyquist criterion is indeed satisfied. However, one could certainly make the point that the
gain and phase margins could use improvement. This points out the general drawback of
the purely algorithmic approach to controller design that is common to all of the algorithms
we present. One should not rely on the algorithm to produce a satisfactory controller out of
the box. The control designer will always be able to improve an initial design by employing
the lessons only experience can teach.

10.3 Parameterisation of stabilising dynamic output feedback

controllers
The above discussion of construction stabilising controllers leads one to a consideration
of whether it is possible to describe all stabilising controllers. The answer is that it is,
and it is best executed in the rational function framework. We look at the block diagram
configuration of Figure 10.4. We think of the plant transfer function RP as being proper and
fixed. The objective is to find all controller transfer functions RC so that the interconnection
is IBIBO stable. This will happen when the transfer functions between all inputs and outputs
are in RH+
. The four relevant transfer functions are
1
RC
T1 =
, T2 =
,
1 + RC RP
1 + RC RP
(10.10)
RP
RC RP
T3 =
, T4 =
.
1 + RC RP
1 + RC RP
Let us first provide a useful properties rational functions in RH+
.

22/10/2004

10.3 Parameterisation of stabilising dynamic output feedback controllers

r(s)

RC (s)

RP (s)

421

422

10 Stabilisation and state estimation

22/10/2004

+
is the product of a function invertible in RH+
, and a function in RH all of whose zeros lie
in C+ .
The following result introduces the notion of the coprime factorisation. This will play
for us an essential role in determining useful representations for stabilising controllers.

y(s)

10.33 Theorem Rational functions R1 , R2 RH+

are coprime if and only if there exists
1 , 2 RH+
so that
1 R1 + 2 R2 = 1.
(10.12 )

Figure 10.4 The block diagram configuration for the investigation

of stabilising controller parameterisation

We call (1 , 2 ) a coprime factorisation of (R1 , R2 ).

10.3.1 More facts about RH+
It is useful to investigate in more detail some algebraic

properties of RH+
. These will be useful in this section, and again in Chapter 15. Many of
our results in this section may be found in [Fuhrmann 1996].
For a rational function R R(s), the c.f.r. is a representation of R by the quotient of
coprime polynomials where the denominator polynomial is monic. Let us look at another
way to represent rational functions. We shall say that R1 , R2 RH+
are coprime if they
have no common zeros in C+ and if at least one of them is not strictly proper.
10.31 Definition A coprime fractional representative of R R(s) is a pair (RN , RD )
with the following properties:
(i) RN , RD RH+
;
(ii) RD and RN are coprime in RH+
;
N
.

(iii) R = R
RD

Proof First suppose that 1 , 2 RH+

exist as stated. Clearly, if z C+ is a zero of, say,
R1 is cannot also be a zero of R2 as this would contradict (10.12). Whats more, if both R1
and R2 are strictly proper then we have

lim 1 (s)R1 (s) + 2 (s)R2 (s) = 0,
s

again in contradiction with (10.12).

Now suppose that R1 , R2 RH+
are coprime and suppose that R1 is not strictly proper.
Let (N1 , D1 ) and (N2 , D2 ) be the c.f.r.s for R1 and R2 . Denote = (s+1), let `j = deg(Nj,+ ),
and let kj be the relative degree of (Nj , Dj ), j = 1, 2. Thus k1 = 0. Write
j Nj,+ ,
Rj = R
`j +kj

j , j = 1, 2, is invertible in RH+ . Suppose that (

where R
1 , 2 ) are a coprime factorisation

N2,+
of ( N`11,+
+k1 , `2 +k2 ). Then

The following simple result indicates that any rational function has a coprime fractional
representative.
10.32 Proposition If R R(s) then R has a coprime fractional representative.

RN (s) =

N (s)
,
(s + 1)k

RD (s) =

D(s)
.
(s + 1)k

Note that unlike the c.f.r., there is no unique coprime fractional representative. However,
it will be useful for us to come up with a particular coprime fractional representative. Given
a polynomial P R[s] we may factor it as P (s) = P (s)P+ (s) where all roots of P lie in
C and all roots of P+ lie on C+ . This factorisation is unique except in a trivial way; the
coefficient of the highest power of s may be distributed between P and P+ in an arbitrary
way. Now let (N, D) be the c.f.r. for R R(s). We then have
R(s) =

N (s)
N (s)N+ (s)
N (s)(s + 1)`+k N+ (s)
=
=
D(s)
D(s)
D(s)
(s + 1)`+k
(s)(s+1)`+k

(10.11)

N
RH+

D(s)
D(s)
+
+
+
1
and that N (s)(s+1)`+k RH . Generally, if Q RH and Q RH , then we say that Q
+
is invertible in RH+
. The formula (10.11) then says that any rational function R RH

where ` = deg(N+ ) and k is the relative degree of (N, D). Note that

N1,+
N2,+
+ 2 `2 +k2 = 1
`1

11 R
1 N1,+ + 2 R
21 R
2 N2,+ = 1.
1 R
`1
`2 +k2
1

Proof Let (N, D) be the c.f.r. of R and let k = max{deg(D), deg(N )}. Then (RN , RD ) is a
coprime fractional representative where

j = 1, 2,

1 and R
2 are invertible in RH+
11 , 2 R
21 ) is a coprime factoriSince R
1 R
this shows that (
sation of (R1 , R2 ). Thus we can assume, without loss of generality, that R1 and R2 are of
the form
P2
P1
R1 = `1 , R2 = `2 +k ,

where the coprime polynomials P1 and P2 have all roots in C+ , deg(Pj ) = `j , j = 1, 2, and
`2 `1 . By Lemma C.4 we may find polynomials Q1 , Q2 R[s] so that
Q1 P1 + Q2 P2 = `1 +`2 +k ,
and with deg(Q2 ) < deg(P1 ). Thus we have
Q1 P1
Q2 P2
+
= 1.
`2 +k `1 `1 `2 +k

(10.13)

Q2
P2
Since R2 RH+
, `2 +k is proper. Since deg(Q2 ) < deg(P1 ), `1 is strictly proper. Therefore,
P1
the second term in (10.13) is strictly proper. Since `1 is not strictly proper, it then follows
1
that `Q2 +k
is also not strictly proper since we must have
Q P
Q2 P2
1
1
lim
+ `1 `2 +k = 1
s `2 +k `1

22/10/2004

10.3 Parameterisation of stabilising dynamic output feedback controllers

and
lim

2
`1

P2
`2 +k

423

424

10 Stabilisation and state estimation

10.35 Lemma Let RP R(s) be proper with coprime fractional representative (P1 , P2 ). If
(1 , 2 ) and (
1 , 2 ) are coprime factorisations for (P1 , P2 ), then the map

= 0.

1 + ()P
1 + P2
2
7

2 P1
2 ()P
1

Therefore, if we take
Q1
Q2
1 = `2 +k , 2 = `1 ,

we see that the conditions of the theorem are satisfied.

The matter of determining rational functions 1 and 2 in the lemma is not necessarily
straightforward. However, in the next section we shall demonstrate a way to do this, at
least if one can find a single stabilising controller. The following corollary, derived directly
from the computations of the proof of Theorem 10.33, declares the existence of a particular
coprime factorisation that will be helpful in the course of the proof of Theorem 10.37.
10.34 Corollary If R1 , R2 RH+
are coprime with R1 not strictly proper, then there exists
a coprime factorisation (1 , 2 ) of (R1 , R2 ) having the property that 2 is strictly proper.
Proof As in the initial part of the proof of Theorem 10.33, let us write
1 P1 ,
R1 = R
`1

2 P2 ,
R2 = R
`2 +k

1 and R
2 are invertible in RH+ and where `j = deg(Pj ), j = 1, 2. In the proof of
where R

Theorem 10.33 were constructed 1 and 2 (in the proof of Theorem 10.33, these are the
final 1 and 2 ) with the property that
1

P1
P2
+ 2 `2 +k = 1
`1

1 j , j = 1, 2, then (1 , 2 ) is a coprime
and 2 is strictly proper. Now we note that if j = R
j
21 RH+ , and so is proper.
factorisation of (R1 , R2 ) and that 2 is strictly proper since R

10.3.2 The Youla parameterisation Now we can use Theorem 10.33 to ensure a
means of parameterising stabilising controllers by a single function in RH+
. Before we
begin, let us establish some notation that we will use to provide an important preliminary
results. We let RP R(s) be proper with coprime fractional representative (P1 , P2 ), and
let (1 , 2 ) be a coprime factorisation for (P1 , P2 ). We call RH+
admissible for the
coprime fractional representative (P1 , P2 ) and the coprime factorisation (1 , 2 ) if
2
1. 6=
, and
P1

2. lim 2 (s) (s)P1 (s) 6= 0.
s

Now we define
Spr (P1 , P2 , 1 , 2 ) =

22/10/2004

1 + P2

admissible
.
2 P1

At this point, this set depends on the choice of coprime factorisation (1 , 2 ). The following
lemma indicates that the set is, in fact, independent of this factorisation.

from Spr (P1 , P2 , 1 , 2 ) to Spr (P1 , P2 , 1 , 2 ) is a bijection if

= + 1 2 2 1 .
()

Proof First note that ()

RH+
so the map is well-defined. To see that the map is a
bijection, it suffices to check that the map
2
2
1 + P
1 + ()P
7

1
2 P1
2 ()P
is its inverse provided that
= + 1 2 2 1 .
()
This is a straightforward, if slightly tedious, computation.

The upshot of the lemma is that the set Spr (P1 , P2 , 1 , 2 ) is independent of 1 and 2 .
Thus let us denote it by Spr (P1 , P2 ). Now let us verify that this set is in fact only dependent
on RP , and not on the coprime fractional representative of RP .
10.36 Lemma If RP R(s) is proper, and if (P1 , P2 ) and (P1 , P2 ) are coprime fractional
representatives of RP , then Spr (P1 , P2 ) = Spr (P1 , P2 ).
Proof Since we have
RP =

P1
P1
=
P2
P2

it follows that Pj = U Pj for an invertible U RH+

. If (1 , 2 ) is a coprime factorisation of
(P1 , P2 ) it then follows that (U 1 1 , U 1 2 ) is a coprime factorisation of (P1 , P2 ). We then
have

(
)
1 + P2

Spr (P1 , P2 ) =
admissible
2 P1

(
)
P2
U 1 1 + U

=
admissible
P1
U 1 2 U

(
)
2 P2
1 + U

=
admissible
2 P1
2 U

1 + P2
=

admissible
2 P1
= Spr (P1 , P2 ),
as desired.

Now we are permitted to denote Spr (P1 , P2 ) simply by Spr (RP ) since it indeed only
depends on the plant transfer function. With this notation we state the following famous
result due to [Youla, Jabr, and Bongiorno 1976], stating that Spr (RP ) is exactly the set of
proper stabilising controllers. Thus, in particular, Spr (RP ) S (RP ).

22/10/2004

10.3 Parameterisation of stabilising dynamic output feedback controllers

425

10.37 Theorem Consider the block diagram configuration of Figure 10.4 and suppose that RP
is proper. For the plant rational function RP , let (P1 , P2 ) be a coprime fractional representative with (1 , 2 ) a coprime factorisation of (P1 , P2 ):

426

10 Stabilisation and state estimation

Let D = C1 P1 + C2 P2 . We claim that D and

1 , 2 RH+
have the property that

1
D

22/10/2004

+
are in RH+
. Clearly D RH . Also, if

1 C1 + 2 C2 = 1,
1 P1 + 2 P2 = 1.

(10.14 )
(such functions exist by Theorem 10.33), then we have

Then there is a one-to-one correspondence between the set of proper rational functions RC
R(s) that render the interconnection IBIBO stable and the set

1 + P2
Spr (RP ) =

admissible
.
(10.15 )
2 P1
Furthermore, if RP is strictly proper, every RH+
is admissible.
Proof Let us first translate the conditions (10.10) into conditions on coprime fractional
representatives for RP and RC . Let (P1 , P2 ) be a coprime fractional representative of RP as
in the statement of the theorem. Also denote a coprime fractional representative of RC as
(C1 , C2 ). Then the four transfer functions of (10.10) are computed to be
C2 P2
,
T1 =
C1 P1 + C2 P2
C2 P1
T3 =
,
C1 P1 + C2 P2

C1 P2
T2 =
,
C1 P1 + C2 P2
C1 P1
T4 =
.
C1 P1 + C2 P2

By the assumption that the transfer functions T1 , T2 , T3 , and T4 are all in RH+
, it follows
that D1 RH+
. Thus D is proper and not strictly proper so that we may define a new
coprime fractional representative for RC by ( CD1 , CD2 ) so that
C1 P1 + C2 P2 = 1.
We therefore have

(10.16)
=

We are thus charged with showing that these four functions are in RH+
.
Now let RH+
and let RC be the corresponding rational function defined by (10.15).
Let
C1 = 1 + P2 , C2 = 2 P1 .
We claim that (C1 , C2 ) is a coprime fractional representative of RC . This requires us to show
that C1 , C2 RH+
, that the functions have no common zeros in C+ , and that at least one
+
of them is not strictly proper. That C1 , C2 RH+
follows since , 1 , 2 , P1 , P2 RH . A
direct computation, using (10.14), shows that
C1 P1 + C2 P2 = 1.

(1 C1 + 2 C2 )(1 P1 + 2 P2 )
1
=
D
D
= 1 1 T4 + 1 2 T2 + 2 1 T3 + 2 2 T1 .

(10.17)

From this it follows that C1 and C2 have no common zeros. Finally, we shall show that
RC is proper. By Lemma 10.35 we may freely choose the coprime factorisation (1 , 2 ). By
Corollary 10.34 we choose (1 , 2 ) so that 1 is strictly proper. Since

lim 1 (s)P1 (s) + 2 (s)P2 (s) = 1,
s

it follows that 2 is not strictly proper. Therefore, C2 = 2 P1 is also not strictly proper,
provided that is admissible.
Now consider the case when RP is proper (i.e., the final assertion of the theorem). Note
that if RP is strictly proper, then so is P1 . Condition 1 of the definition of admissibility
then follows since if = P21 , then would necessarily be improper. Similarly, if P1 is strictly
proper, it follows that lims C2 (s) = 2 (s) 6= 0. Thus condition 2 for admissibility holds.
Now suppose that RC R(s) stabilises the closed-loop system so that the four transfer
functions (10.10) are in RH+
. Let (C1 , C2 ) be a coprime fractional representative of RC .

1 2 P 1
1
=
C1 C2 P2
1

P
C2 2
1
1 =
,
P2
C1 1 1

if we take = 2 C1 1 C2 RH+
. It therefore follows that
C1 = 1 + P2 ,

C2 = 2 P1 ,

as desired.

10.38 Remarks 1. This is clearly an interesting result as it allows us the opportunity to write
down all proper stabilising controllers in terms of a single parameter RH+
. Note that
there is a correspondence between proper controllers and those obtained in the dynamic
output feedback setting. Thus the previous result might be thought of as capturing all
the stabilising dynamics output feedback controllers.
2. Some authors say that all stabilising controllers are obtained by the Youla parameterisation. This is not quite correct.
3. In the case that RP is not strictly proper, one should check that all admissible s give
loop gains RL = RC RP that are well-posed in the unity gain feedback configuration of
Figure 10.4. This is done in Exercise E10.15.

Things are problematic at the moment because we are required to determine 1 , 2 RH+

with the property that (10.14) holds. This may not be trivial. However, let us indicate a
possible method for determining 1 and 2 . First let us show that the parameter RH+

uniquely determines the controller.

10.39 Proposition Let RP be a strictly proper rational function with (P1 , P2 ) a coprime fractional representative. Let 1 , 2 RH+
have the property that
1 P1 + 2 P2 = 1.

10.3 Parameterisation of stabilising dynamic output feedback controllers

22/10/2004

427

428

1 + P2

7
2 P1
from RH+
into the set of stabilising controllers for the block diagram configuration of Figure 10.4 is injective.
Proof Let be the indicated map from RH+
into the set of stabilising controllers, and
suppose that (1 ) = (2 ). That is,
1 + 2 P2
1 + 1 P 2
=
.
2 1 P1
2 2 P 1
This implies that
1 2 P1 P2 + 1 P2 2 2 P1 1 + 1 2 = 1 2 P1 P2 1 P1 1 + 2 P2 2 + 1 2
1 (1 P1 + 2 P2 ) = 2 (1 P1 + 2 P2 )
1 = 2 ,

as desired.

Let us carry this out for an example.

10.42 Example (Example 6.59 contd) We return to the example where RP =
form the above steps.
1. A coprime fractional representative of RP is given by (P1 , P2 ) where

Let us now see how to employ a given stabilising controller to determine 1 and 2 .
10.40 Proposition Let RP be a strictly proper rational function with (P1 , P2 ) a coprime fractional representative. If RC is a stabilising controller for the block diagram configuration of
Figure 10.4, define 1 and 2 by
RC
1 =
,
P2 + P 1 R C

22/10/2004

2. Construct the canonical minimal realisation P = (AP , bP , ctP , D P ) for RP .

3. Using Proposition 10.13, construct f Rn so that AP bP f t is Hurwitz.
4. Using Corollary 10.17 construct ` Rn so that AP `ctP is Hurwitz. Note that this
amounts to performing the construction of Proposition 10.13 with A = AtP and b = c.
5. Using Theorem 10.27 define a stabilising controller SISO linear system C =
(AC , bC , ctC , D C ).
6. Define the stabilising controller rational function RC = TC .
7. Determine 1 , 2 RH+
using Proposition 10.40.
8. The set of all stabilising controllers is now given by

1 + P2
admissible .

Spr (RP ) =

2 P1

Then the map

=
=

10 Stabilisation and state estimation

1
2 =
.
P2 + P 1 R C

Then the following statements hold:

(i) 1 , 2 RH+
;
(ii) 1 P1 + 2 P2 = 1.

P1 (s) =
check

2. As in Example 6.59, we have

0 1
AP =
,
0 0

1/m
,
(s + 1)2

bP =

0
1
m

P2 (s) =

s2
.
(s + 1)2

1
,
0

cP =

D P = 01 .

3. In Example 10.30 we computed

Proof (i) Let (C1 , C2 ) be a coprime fractional representative for RC . Then

1 =
=

4. In Example 10.30 we also computed

C1
C2

P2 +

C1
P1 C
2

C1
.
C1 P1 + C2 P2

As in the proof of Theorem 10.37, it follows then that if D = C1 P1 +C2 P2 then D, D1 RH+
.
+
+
Since C1 RH+
, it follows that 1 RH . A similar computation gives 2 RH .

(ii) This is a direct computation.

2m
,
2m

5. Using Example 10.30 we have

2 1
2
AC =
bC =
,
4 2
2

2
.
2

ctC = 2m 2m ,

D C = 01 .

6. In Example 10.30 we computed

This result allows us to compute 1 and 2 given a coprime fractional representative for a
plant transfer function. This allows us to produce the following algorithm for parameterising
the set of stabilising controllers.

RC (s) =

4m(2s + 1)
.
s2 + 4s + 8

7. From Proposition 10.40 we calculate, after some simplification,

10.41 Algorithm for parameterisation of stabilising controllers Given a proper plant transfer
function RP perform the following steps.
1. Determine a coprime fractional representative (P1 , P2 ) for RP using Proposition 10.32.

1 =

(s + 1)2 (s2 + 4s + 8)
,
(s2 + 2s + 2)2

2 =

4m(s + 1)2 (2s + 1)

.
(s2 + 2s + 2)2

1
,
ms2

and per-

22/10/2004

10.4 Strongly stabilising controllers

429

430

8. Finally, after simplification, we see that the set of stabilising controllers is given by
(
)

m 4m(s + 1)4 (2s + 1) + (s)s2 ((s2 + 2s + 2)2 )
+
Spr (RP ) =

RH
.

m(s + 1)4 (s2 + 4s + 8) (s)((s2 + 2s + 2)2 )

u(t)

e(t)

observer

finish

(t)
x

Figure 10.5 A schematic for an observer using the error

10.43 Proposition Let = (A, b, ct , D) be an observable SISO linear system satisfying (10.18). There exists oo (s), oi (s) R[s]n1 with the property that

d
d
x(t) = oo dt
y(t) + oi dt
u(t).

With these functions we compute the set of stabilising controllers to be

)
(

m (s)s2 m(s2 + 1)
+

RH
,

+ m(s2 + 1)

10.4 Strongly stabilising controllers

In the previous section we expressed all controllers that stabilise a given plant. Of course,
one will typically not want to allow any form for the controller. For example, one might
wish for the controller transfer function to itself be stable. This is not always possible, and
in this section we explore this matter.

10.5 State estimation

One of the problems with using static or dynamic state feedback is that the assumption
that one knows the value of the state is typically over-optimistic. Indeed, in practice, it is
often the case that one can at best only know the value of the output at any given time.
Therefore, what one would like to be able to do is infer the value of the state from the
knowledge of the output. A moments reflection should suggest that this is possible for
observable systems. A further moments reflection should lead one to allow the possibility
for detectable systems. These speculations are correct, and lead to the theory of observers
that we introduce in this section.
10.5.1 Observers Let us give a rough idea of what we mean by an observer before
we get to formal definitions. Suppose that we have a SISO linear system = (A, b, ct , D)
evolving with its usual differential equations:
x = Ax(t) + bu(t)
y(t) = ct x(t) + Du(t).

x(t)

y(t)

2 (s) = m.

which is somewhat more pleasing than our expression derived using our rules.

plant

22/10/2004

This is a somewhat complicated expression. It can be simplified by using simpler expressions for 1 and 2 . In computing 1 and 2 above, we have merely applied our rule
verbatim. Indeed, simpler functions that also satisfy 1 P1 + 2 P2 = 1 are
1 (s) = 1,

10 Stabilisation and state estimation

(10.18)

An observer should take as input the original input u(t), along with the measured output
y(t). Using these inputs, the observer constructs an estimate for the state, and we denote
(t). In Figure 10.5 we schematically show how an observer works. Our
this estimate by x
first result shows that an observer exists, although we will not use the given observer in
practice.

Proof In (10.18), differentiate y(t) n 1 times successively with respect to t, and use the

equation for x(t)

to get

y(t)
u(t)
ct
D
0
0 0
y (1) (t) ct A
ct b
(1)

D
0 0

u (t)
t
y (2) (t) ct A2
ct Ab
u(2) (t)
c
b

0
0

=
x(t) +

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

.
.
.
.
.
.
.
t n2
t n3
t
t n1
(n1)
(n1)
c A b c A b c b D
cA
y
(t)
u
(t).
Since is observable, O(A, c) is invertible, and so we have

y(t)
D
0
ct b
y (1) (t)
D

(2)

t
ct b
x(t) = O(A, c)1 y (t) O(A, c)1 c Ab

..
..
..

.
.
.
n2
t
t
c A b c An3 b
y (n1) (t)

u(t)
0
(1)

0 u (t)

(2)

0
u (t)

..
..

.
.
t
c b D
u(n1) (t),

..
.

which proves the proposition if we take

D
0
1
ct b
s
D

2
t
ct b
oo = O(A, c)1 s , oi = O(A, c)1 c Ab

..
..
..
.

.
.
n2
sn1
ct A b ct An3 b

0
0
0
..
.

0
1

0 s

2
0
s .
.. ..
. .
sn1
ct b D

...

0
0
0
..
.

While the above result does indeed prove the existence of an observer which exactly
reproduces the state given the output and the input, it suffers from repeatedly differentiating
the measured output, and in practice this produces undesirable noise. To circumvent these
problems, in the next section we introduce an observer that does not exactly measure the
(t)
state. Indeed, it is an asymptotic observer , meaning that the error e(t) = x(t) x
satisfies limt e(t) = 0.

10.5 State estimation

22/10/2004

431

432

y(t)

i(t)

u(t)

PA (s) = sn + pn1 sn1 + + p1 s + p0 .

internal
model

Let P R[s] be monic and degree n. The observer gain vector ` defined by

0

1
` = P (A)(O(A, c))
0
1

(t)
x

Figure 10.6 The schematic for a Luenberger observer

10.5.2 Luenberger observers The error schematic in Figure 10.5 has the feature that
it is driven using the error e(t). The asymptotic observer we construct in this section instead
uses the so-called innovation defined as i(t) = y(t) y(t) where y(t) is the estimated output
(t) + Du(t). The schematic for the sort of observer is shown in Figure 10.6. Note
y(t) = ct x
that the inputs to the observer are the actual measured output y(t) and the actual input u(t),
(t). The vector ` Rn we call the observer
and that the output is the estimated state x
gain vector . There is a gap in the schematic, and that is the internal model. We use as
the internal model the original system model, but now with the inputs as specified in the
schematic. Thus we take the estimated state to satisfy the equation
(t) = A
x
x(t) + bu(t) + `i(t)
(t) + Du(t)
y(t) = ct x
i(t) = y(t) y(t).

(10.19)

(t).
From this equation, the following lemma gives the error e(t) = x(t) x
(t), y(t), and i(t) sat10.44 Lemma If = (A, b, ct , D) is a SISO linear system and if x
(t) x(t) satisfies the differential equation
isfy (10.19), then e(t) = x

e(t)
= (A `ct )e(t).
Proof Subtracting

x(t)
= Ax(t) + bu(t),

22/10/2004

10.45 Proposition Let (A, b, ct , D) be an observable SISO linear system and suppose that the
characteristic polynomial for A is

ct
D

10 Stabilisation and state estimation

(t) = A
x
x(t) + bu(t) + `i(t),

has the property that the characteristic polynomial of the matrix A `ct is P .
Proof Note that observability of (A, c) is equivalent to controllability of (At , c). Therefore,
by Proposition 10.13 we know that if

`t = 0 0 1 (C(At , c))1 P (At ),
(10.20)
internal model
principle?

then the matrix At c`t has characteristic polynomial P . The result now follows by taking
the transpose of equation (10.20), and noting that the characteristic polynomial of At c`t
is equal to that of A `ct .

10.46 Remarks 1. As expected, the construction of the observer gain vector is accomplished
along the same lines as the static state feedback vector as in Proposition 10.13. Indeed,
the observer gain vector is obtained by using the formula of Proposition 10.13 with At in
place of A, and with c in place of b. This is another example of the duality between
controllability and observability.
2. The eigenvalues of A `ct are called the observer poles for the given Luenberger
observer.
3. The notion of an observer is lurking in the proof of Theorem 10.27. This is flushed out
in Section 10.5.3.
4. We are, of course, interested in choosing the observer gain vector ` so that A `ct is
Hurwitz. This can be done if is observable, or more generally, detectable. To this end,
let us denote by D() those observer gain vectors for which A `ct is Hurwitz.

Let us illustrate this with an example.

and using the second and third of equations (10.19) we get

10.47 Example We consider the SISO linear system = (A, b, ct , D) with

0 1
0
A=
, c=
.
1 0
1

e(t)
= Ae(t) `i(t)
= Ae(t) `(y(t) y(t))
(t))
= Ae(t) `ct (x(t) x
= Ae(t) `ct e(t),
as desired.

The lemma now tells us that we can make our Luenberger observer an asymptotic observer
provided we choose ` so that A `ct is Hurwitz. This is very much like the Ackermann
pole placement problem, and indeed can be proved along similar lines, giving the following
result.

Suppose that we wish a closed-loop characteristic polynomial of P (s) = s2 + 4s + 4. We

compute

1 2
0 1
P (A) = A2 + 2A + 2I 2 =
, O(A, c) =
.
2 1
1 0
Then we have
` = P (A(O(A, c))1

0
1
=
,
1
2

10.5 State estimation

22/10/2004

433

giving
A `ct =
PSfrag replacements

434

10 Stabilisation and state estimation

a somewhat optimistic means of designing a controller. However, if one uses an observer to

estimate the state, one can use the estimated state for static state feedback. The schematic
is depicted in Figure 10.8. If the observer is a Luenberger observer satisfying (10.19) and if

0 2
,
1 2

PSfrag replacements

t
which has the desired characteristic polynomial.
ke(t)k
(t)
Now let us see how the observer does atee1observing.
In Figure 10.7 we show the results
(t)

t
ke(t)k
e1 (t)
e2PSfrag
(t)
replacements
e3 (t)
y(t)
u(t)
h (t)
e1 (t)
hN,D (t)
2
1 (t)
e2 (t)
1N,D (t)
e3 (t)
(t)
(t)
y(t)
fj (t)
1
u(t)
Re
Im
h (t)

e3 (t)
y(t)
u(t)
h (t)
hN,D (t)
1 (t)
1N,D (t)
(t)
(t)
fj (t)
Re
Im
x1
x2

x2
log
dB
deg
u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

,0
0

observer
(t)
x
ft

Figure 10.8 Static state feedback using the estimated state

-1

the plant is the usual SISO state representation satisfying (10.18), then one may verify that
the equations governing the interconnection of Figure 10.8 are

x(t)
A
bf t
x(t)
b
=
+
r(t)
(t)
`ct A bf t `ct x
b
(t)
x
(10.21)

x(t)
y(t) = ct Df t
+ Dr(t).
(t)
x

-2
-3

-2

-1

,0
0

The next result records the characteristic polynomial and the closed-loop transfer function
for these equations.

10.48 Theorem Let = (A, b, ct , D) be a SISO linear system with f Rn a static state
feedback vector and ` Rn an observer gain vector for a Luenberger observer (10.19).
Suppose that the observer is combined with state feedback as in Figure 10.8, giving the closedloop equations (10.21). Also consider the interconnection of Figure 10.9 where

x(0) = (1, 1),

RP (s)

y(t)

of a simulation of equations (10.18) and (10.19) with

D = 01 ,

RC (s)

r(t)

Figure 10.7 The state (top left), the estimated state (top right),
and the norm of the error for a Luenberger observer

b = (0, 1),

y(t)

-1

plant

ke(t)k

,0
0

u(t)

r(t)

hN,D (t)
1 (t)
1N,D (t)
(t)
-1
(t)
fj (t)
Re
-2
Im
x1 -2
-3
x2
x1
x2 1.4
log
1.2
dB
deg
1
u = 0 ln
ln coth(|u| /2) 0.8
or 1
m 0.6

yos 0.4
0.2
tos 0

x1
x2
log
dB
deg
u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

22/10/2004

RO (s)

(0) = 0,
x

Note that the error is decaying to zero exponentially as expected.

u(t) = 1(t).

10.5.3 Static state feedback, Luenberger observers, and dynamic output feedback
In this section we bring together the ideas of static state feedback, Luenberger observers,
and dynamic output feedback. It is by no means obvious that these should all be tied
together, but indeed they are. To make the connection we make the obvious observation
that if one does not possess accurate knowledge of the state, then static state feedback seems

Figure 10.9 A dynamic output feedback loop giving the closedloop characteristic polynomial of static state feedback using a
Luenberger observer to estimate the state

10.5 State estimation

22/10/2004

435

det(sI n A + `ct )
det(sI n A + bf t + `(ct Df t ))
RP (s) = ct (sI n A)1 b + D
f t adj(sI n A)`
.
det(sI n A + `ct )

Then the following statements hold:

(i) the closed-loop characteristic polynomial for (10.21) is the product of the characteristic
polynomials of A bf t and A `ct .
(ii) the closed-loop system for (10.21) (i.e., the transfer function from r to y) is the same
transfer function for the interconnection of Figure 10.9, and whats more, both transfer
functions are exactly the transfer function for (A bf , b, ct Df t , 01 ).
Proof (i) The A-matrix for the observer/feedback system (10.21) is

A
bf t
Acl =
t
t
t .
`c A bf `c

Now, the equation governing the observer states is

(t) = (A `ct )
x
x(t) + (b D`)u(t) + `y(t).
Taking left causal Laplace transforms gives

(s) = (sI n A + `ct )1 (b D`)
x
u(s) + `
y (s) .
Let us define
T 1 (s) = (sI n A + `ct )1 (b D`),

T 1

I n I n
0 In

I I
= n n .
0 In

A `ct
0
=
t .
t
`c
A bf

Thus we see that the characteristic polynomial of T Acl T 1 , and therefore the characteristic
polynomial of Acl , is indeed the product of the characteristic polynomials of A `ct and
A bf t , as claimed.
(ii) Let us use the new coordinates corresponding to the change of basis matrix T defined
in (10.22). Thus we take

t
0
cl = A t`c
A
t
`c
A bf

b
0
cl = T
b
=
b
b
t

t
cl = c Df t T 1 = ct ct Df t
c
cl = D,
D
cl , c
cl , b
cl ). We have
= (A
tcl , D
and determine the transfer function for

t 1
0
cl )1 = (sI n A + `c )
(sI 2n A
t 1 ,

(sI n A + bf )
where denotes a term whose exact form is not relevant. One then readily computes
cl = (ct Df t )(sI n A + bf t )1 b.
cl )1 b
tcl (sI 2n A
T (s) = c

T 2 (s) = (sI n A + `ct )1 `

so that
(10.22)

It is a simple computation to show that

T Acl T 1

22/10/2004

NC (s) = det(sI n A + `ct ),

DC (s) = det(sI n A + bf t + `(ct Df t ))
t
NP (s) = D det(sI n A) + c adj(sI n A)b, DP (s) = det(sI n A)
NO (s) = f t adj(sI n A)`,
DO (s) = det(sI n A + `ct ).

Let us define an invertible 2n 2n matrix

T =

10 Stabilisation and state estimation

From this it follows that the transfer function of the closed-loop system (10.21) is as claimed.
Now we determine this same transfer function in a different way, proceeding directly from
the closed-loop equations (10.21). First let us define

RC (s) =

RO (s) =

436

(s) = T 1 (s)
x
u(s) + T 2 (s)
y (s).
(t), we have
Given that u(t) = r(t) f t x
u(s) = r(s) f t T 1 (s)
u(s) f t T 1 (s)
y (s)
=

u(s) =

1
f t T 2 (s)
y(s) +
r(s).
1 + f t T 1 (s)
1 + f t T 1 (s)

From the proof of Lemma A.3 we have that

1 + f t T 1 (s) = 1 + f t (sI n A + `ct )1 (b D`)

= det 1 + f t (sI n A + `ct )1 (b D`)
= det(I n + (sI n A + `ct )1 (b D`)f t )
= det (sI n A + `ct )1 (sI n A + `(ct Df t ) + bf t
det(sI n A + `(ct Df t ) + bf t )
det(sI n A + `ct )
DC (s)
=
.
NC (s)
=

We also clearly have

f t T 2 (s) = f t (sI n A + `ct )1 `
f t adj(I n A + `ct )`
det(I n A + `ct )
f t adj(I n A)`
=
det(I n A + `ct )
NO (s)
=
,
DO (s)
=

10.5 State estimation

22/10/2004

437

where we have used Lemma A.5. Thus we have

u(s) =

(10.23)

NP (s)
u(s),
DP (s)

(10.24)

Also noting that

we may eliminate r(s) from equations (10.23) and (10.24) to get

T (s) =

y(s)
NC (s)NP (s)
=
,
r(s)
DC (s)DP (s) + NP (s)NO (s)

10 Stabilisation and state estimation

22/10/2004

Let us illustrate Theorem 10.48 via an example.

NO (s) NC (s)
NC (s)
y(s) +
r(s).
DO (s) DC (s)
DC (s)

y(s) =

438

(10.25)

if we use the fact that DO = NC .

Let us now turn to the transfer function of the interconnection of Figure 10.9. The
transfer function may be computed in the usual manner using Masons Rule:
RC (s)RP (s)
DO (s)NC (s)NP (s)
y(s)
=
=
.
r(s)
1 + RC (s)RP (s)RO (s)
DC (s)DP (s)DO (s) + NC (s)NP (s)NO (s)
Since DO = NC this may be simplified to
y(s)
NC (s)NP (s)
=
.
r(s)
DC (s)DP (s) + NP (s)NO (s)
Comparing this to (10.25), we see that indeed the transfer function of Figure 10.9 is the

same as that of the closed-loop system (10.21).
10.49 Remarks 1. One of the consequences of part (i) of the theorem is that one can separately choose the observer poles and the controller poles. This is a phenomenon that
goes under the name of the separation principle.
2. The change of coordinates represented by the matrix defined in (10.22) can easily be seen
) to (e, x). It is not surprising, then, that
as making a change of coordinates from (x, x
in these coordinates we should see that the characteristic polynomial gets represented as
a product of two characteristic polynomials. Indeed, we have designed the observer so
that the error should be governed by the matrix A `ct , and we have designed the state
feedback so that the state should be governed by the matrix A bf t .
3. Note that for the interconnection of Figure 10.9 there must be massive cancellation in
the transfer function since the system nominally has 3n states, but the denominator
of the transfer function has degree n. Indeed, one can see that in the loop gain there
is directly a cancellation of a factor det(sI n A + `ct ) from the numerator of RC
and the denominator of RO . What is not so obvious at first glance is that there is an
additional cancellation of another factor det(sI n A + `ct ) that happens when forming
the closed-loop transfer function. Note that these cancellations are all stable provided
one chooses ` so that A `ct is Hurwitz. Thus they need not be disastrous. That there
should be this cancellation in the closed-loop equations (10.21) is not surprising since the
control does not affect the error. Thus the closed-loop system with the observer is not
controllable (also see Exercise E10.23). One should also be careful that the characteristic
polynomial for the closed-loop system (10.21) is different from that for the interconnection
Figure 10.9.

10.50 Example (Example 10.47 contd) Recall that we had

0 1
0
A=
, c=
.
1 0
1
Suppose that we want both observer poles and controller poles to be roots of the polynomial
P (s) = s2 + 4s + 4. In Example 10.47 we found the required observer gain vector to be
` = (1, 2). To find the static state feedback vector we employ Ackermanns formula from
Proposition 10.13. The controllability matrix is

0 1
C(A, b) =
.
1 0
Thus we compute

f t = 0 1 (C(A, b))1 P (A) = 1 2 .
This gives the closed-loop system matrix

0 1 0 0
1 0 1 2

Acl =
0 1 0 2
0 2 2 4
for the interconnections of Figure 10.8. For the interconnection of Figure 10.8 we determine
that the closed-loop input and output vectors are

0
0
1

, ccl = 1 .
bcl =
0
0
1
0
(t) for the initial conditions
In Figure 10.10 we show the state x(t) and the estimated state x
(0) = (0, 0), and for u(t) = 1(t). Note that the quantities approach the
x(0) = (1, 1) and x
same value as t , as should be the case as the observer is an asymptotic observer. In

Figure 10.11 we show the output for the interconnection of Figure 10.8.

10.6 Summary
1. Stabilisability and detectability extend the notions of controllability and observability.
The extensions are along the lines of asking that the state behaviour that you cannot
control or observe is stable.
2. Ackermanns formula is available as a means to place the poles for a SISO linear system
in a desired location.

PSfrag replacements

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)
h (t)
hN,D (t)
1 (t)
1N,D (t)
(t)
(t)
fj (t)
Re
Im

0.5

-0.5

PSfrag replacements

ke(t)k
e1 (t)
e2 (t)
e3 (t)

439

440

10 Stabilisation and state estimation

22/10/2004

show that A is Hurwitz.

Hint: Show that (E10.1) implies that

1.5

P = eA t PeAt +

eA cct eA d.

0
1

0.5

1.5

e2 (t)
e3 (t)
y(t)
u(t)
h (t)
hN,D (t)
1 (t)
1N,D (t)
(t)
(t)
fj (t)
Re
Im
x1
x2

log
dB
deg
u = 0 ln
ln coth(|u| /2)
or 1
m

yos

1
1.5tos 0

u(t)
h (t)
hN,D (t)
1 (t) 0
-1
-0.5
0.5
1N,D (t) x1
,0
(t)
0
(t)
fj (t)
Figure 10.10
Re The state (left) and the
Im
the closed-loop
system
x1
x2
x1
x2
1.5
log
dB 1.25
deg
1
u = 0 ln
ln coth(|u| /2) 0.75
or 1
0.5
m

0.25
yos

0
tos 0
-0.25
1
2
3
4
,0
t
0

-0.5

-1

-0.5

0.5

1.5

estimated state (right) for

In the next exercise, we will introduce the notion of a linear matrix inequality (LMI ).
Such a relation is, in general, a matrix equation, invariant under transposition (i.e., if one
takes the matrix transpose of the equation, it remains the same), for an unknown matrix.
Since the equation is invariant under transposition, it makes sense to demand that the
unknown matrix render the equation positive or negative-definite or semidefinite. In recent
years, LMIs have become increasingly important in control theory. A survey is [El Ghaoui
and Niculescu 2000]. The reason for the importance of LMIs is one can often determine
their solvability using convexity methods. These are often numerically tractable. This
idea forms the backbone, for example, of the approach to robust control taken by Dullerud
and Paganini [1999].
E10.5 Consider a SISO linear system = (A, b, ct , D).
(a) Use the Liapunov methods of Section 5.4 to show that is stabilisable if and
only if there exists f Rn so that the linear matrix inequality
(A bf t )P + P (At f bt ) < 0

y(t)

,0
0

x1
x2
log
dB
deg
u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

t
ke(t)k
Exercises for
10
e1Chapter
(t)

has a solution P > 0.

(b) Use your result from part (a) to prove the following theorem.
5

Theorem For a stabilisable SISO linear system = (A, b, ct , D), a state feedback vector f Rn stabilises the closed-loop system if and only if there exists
P > 0 and g Rn so that
(i) f = P 1 g and
(ii) so that the LMI

P

At
A b
P
g
<0
+
bt
gt

Figure 10.11 The output for the interconnection of Figure 10.8

Exercises
E10.1 Show that (A, v) is stabilisable if and only if (At , v) is detectable.
E10.2 Show that = (A, b, ct , D) is stabilisable if and only if the matrix

sI n A b

is satisfied.
The point of the exercise is that the LMI gives a way of parameterising all stabilising
state feedback vectors.
E10.6 Consider the two polynomials that have a common factor in C+ :

has rank n for all s C+ .

E10.3 Show that = (A, b, ct , D) is detectable if and only if the matrix

sI n A
t
c

P1 (s) = s2 1,

has rank n for all s C+ .

E10.4 If (A, c) is detectable and if P Rnn is positive-semidefinite and satisfies
At P + P A = cct ,

(E10.1)

P2 (s) = s3 + s2 s 1.

Construct three SISO linear systems i = (Ai , bi , cti , 01 ), i = 1, 2, 3, with the following properties:
1. Ti = PP12 , i = 1, 2, 3;
2. 1 is stabilisable but not detectable;
3. 2 is detectable but not stabilisable;
4. 3 is neither stabilisable nor detectable.

Exercises for Chapter 10

441

E10.7 Let = (A, b, ct , D) be a SISO linear system.

(a) If is stabilisable show how, using Proposition 10.13, to construct a stabilising
state feedback vector f in the case when is not controllable.
(b) Show that if is not stabilisable then it is not possible to construct a stabilising
state feedback vector.
E10.8 Let (A, b) Rnn Rn be controllable and let P R[s] be a monic polynomial of
degree n. Show that there exists a unique static state feedback vector f Rn with
the property that P (s) = det(sI n (A bf t )).
Hint: Refer to Exercise E2.38.
E10.9 Suppose that (A, b) Rnn Rn is controllable, and let P R[s] be a monic
polynomial of degree n 1.
(a) Show that there exists a static state feedback vector f Rn with the following
properties:
1. A bf t possesses an invariant (n 1)-dimensional subspace V Rn with
the property that for each basis {v 1 , . . . , v n1 } for V , {v 1 , . . . , v n1 , b} is a
basis for Rn ;
2. The characteristic polynomial of (A bf t )|V is P .
(b) For V as in part (a), define a linear map AV : V V by AV (v) = prV A(v),
where prV : Rn V is defined by

442

E10.11

E10.12

prV (a1 v 1 + + an1 v n1 + an b) = a1 v 1 + + an1 v n1 ,

for {v 1 , . . . , v n1 } a basis for V .
E10.10 Consider the pendulum/cart system of Exercises E1.5. In this problem, we shall
change the input for the system from a force applied to the cart, to a force applied
to the mass on the end of the pendulum that is tangential to the pendulum motion;
see Figure E10.1. For the following cases,

E10.13

E10.14

Figure E10.1 The pendulum/cart with an alternate input

(a)
(b)
(c)
(d)
(e)

the
the
the
the
the

equilibrium
equilibrium
equilibrium
equilibrium
equilibrium

point
point
point
point
point

(0, 0) with cart position as output;

(0, 0) with cart velocity as output;
(0, 0) with pendulum angle as output;
(0, 0) with pendulum angular velocity as output;
(0, ) with cart position as output;

10 Stabilisation and state estimation

22/10/2004

(f) the equilibrium point (0, ) with cart velocity as output;

(g) the equilibrium point (0, ) with pendulum angle as output;
(h) the equilibrium point (0, ) with pendulum angular velocity as output,
do the following:
1. obtain the linearisation = (A, b, ct , D) of the system;
2. obtain the transfer function for ;
3. determine whether the system is observable and/or controllable;
4. determine whether the system is detectable and/or stabilisable.
Consider the pendulum/cart system of Exercises E1.5 and E2.4. Construct a state
feedback vector f that makes the linearisation of pendulum/cart system stable about
the equilibrium point with the pendulum pointing up. Verify that the closed-loop
system has the eigenvalues you asked for.
Consider the double pendulum system of Exercises E1.6 and E2.5. For the following
equilibrium points and input configurations, construct a state feedback vector f that
makes the linearisation of double pendulum about that equilibrium point stable:
(a) the equilibrium point (0, , 0, 0) with the pendubot input;
(b) the equilibrium point (, 0, 0, 0) with the pendubot input;
(c) the equilibrium point (, , 0, 0) with the pendubot input;
(d) the equilibrium point (0, , 0, 0) with the acrobot input;
(e) the equilibrium point (, 0, 0, 0) with the acrobot input;
(f) the equilibrium point (, , 0, 0) with the acrobot input.
In each case, take the output to be the angle of the second link. Verify that the
closed-loop system has the eigenvalues you asked for.
Consider the coupled tank system of Exercises E1.11, E2.6. For the following equilibrium points and input configurations, construct a state feedback vector f that
makes the linearisation of double pendulum about that equilibrium point stable:
(a) the output is the level in tank 1;
(b) the output is the level in tank 2;
(c) the output is the difference in the levels.
In each case, take the output to be the angle of the second link. Verify that the
closed-loop system has the eigenvalues you asked for.
Let = (A, b, ct , 01 ) with

0 1 0 0 0
0 0 1 0 0

0 0 0 1 0

A = .. .. .. .. . . .. .
. . . .
. .

0 0 0 0 1
0 0 0 0 0
Answer the following questions.
(a) Show that is stabilisable by static output feedback only if all components of
c are nonzero and of the same sign.

Exercises for Chapter 10

443

444

E10.20 Consider the plant transfer function

RP (s) =

In the next exercise you will demonstrate what is a useful feature of

Recall that
a ring is a set R with the operations of addition and multiplication, and these operations
should satisfy the natural properties associated with addition and multiplication (see [Lang
1984]). A ring is commutative when the operation of multiplication is commutative (as for
example, with the integers). A commutative ring is an integral domain when the product
of nonzero elements is nonzero. An ideal in a ring R is a subset I for which
1. 0 I,
2. if x, y I then x + y I, and
3. if x I then ax I and xa I for all a R.
An ideal I is principal if it has the form
I = { xa | a R} .
The ideal above is said to be generated by x. An integral domain R is a principal ideal
domain when every ideal I is principal.
E10.19 Answer the following questions.
(a) Show that RH+
is an integral domain.
(b) Show that RH+
is a principal ideal domain.
(c) Show that H+
is not a principal ideal domain.

s
.
s+1

For this plant, do the following.

s
(a) Show that (P1 (s), P2 (s)) = ( s+1
, 1) is a coprime fractional representative for RP .
(b) Show that
s + 1 (s + 1)(s + 4)
(1 (s), 2 (s)) =
,
(s + 2)2
(s + 2)2

for an admissible then the unity gain feedback loop with RL = RC RP is

well-posed.
E10.16 Let RP RH+
be a stable plant transfer function. Show that the set of controllers
RC for which the closed-loop system in Figure 10.4 is given by

RH+
.
1 RP

RH+
.

22/10/2004

Normally, only condition 1 is given in the definition of admissibility for a function RH+

to parameterise Spr (RP ). This is a genuine omission of hypotheses. In the next exercise,
you will show that, in general, the condition 2 also must be included.

(b) If A is as given, and if all components of c are nonzero and of the same sign, is
it true that can be stabilised using static output feedback? If so, prove it. If
not, find a counterexample.
E10.15 (a) Show that the closed-loop transfer function for any one of the stabilising controllers from Theorem 10.37 is an affine function of . That is to say, show that
the closed-loop transfer function has the form R1 + R2 for rational functions
R1 and R2 .
(b) Show that in the expression R1 + R2 from part (a) that R1 RH+
, and is
strictly proper if and only if RP is strictly proper.
(c) Conclude that if
1 + P2
RC =
2 P1

E10.17 Consider the plant transfer function RP (s) = s12 , and consider the interconnection
of Figure 10.4. Show that there is a rational function RC (s), not of the form given
in Theorem 10.37, for which the closed-loop system is IBIBO stable.
Hint: Consider PD control.
E10.18 Exercise on restricted domain for poles.

10 Stabilisation and state estimation

is a coprime factorisation of (P1 , P2 ).

s
.
s+1

Is admissible?
(d) For the function RH+
from part (c), show that the controller
RC =

Finish

1 + P2
2 P1

is improper. Thus for the conclusions of Theorem 10.37 to hold, in particular, for
the controller parameterisation to be one for proper controller transfer functions,
the condition 2 for admissibility is necessary.
(e) Show that despite our conclusions of part (d), the unity gain interconnection
with loop gain RL = RC RP , using RC from part 2, is IBIBO stable.
E10.21 Consider the plant
s
RP (s) =
.
(s + 1)(s 1)
For this plant, do the following.
(a) Find a coprime fractional representative (P1 , P2 ) for RP (choose the obvious one,
if you want to make life easy).
(b) Show that
2(s + 2)
RC (s) =
s 12
stabilises RP .
(c) Use the stabilising controller from the previous step to construct a coprime
factorisation (1 , 2 ) for (P1 , P2 ).
(d) Write the expression for the set of proper stabilising controllers for RP depending
on the parameter RH+
.
(e) For two different values of , produce the Nyquist plots for the loop gain RC RP
and comment on the properties of the controller you have chosen.
E10.22 Let = (A, b, ct , 01 ) be a SISO linear system.

Exercises for Chapter 10

445

(a) Show that for a Luenberger observer with observer gain vector ` the state and
the estimated state together satisfy the vector differential equation

x(t)
A
0
x(t)
b
=
+
u(t).
(t)
`ct A `ct x
b
(t)
x
(b) What is the differential equation governing the behaviour of x(t) and the error
(t).
e(t) = x(t) x
E10.23 Verify that the system (10.21) is not controllable.
Hint: Consider the system in the basis defined by the change of basis matrix
of (10.21).

446

10 Stabilisation and state estimation

22/10/2004

448

This version: 22/10/2004

11 Ad hoc methods I: The root-locus method

22/10/2004

11.1 The root-locus problem, and its r

ole in control
The aim in this section is to provide some situations in control where a certain type of
problem involving a certain type of polynomial arises. Once this has been nailed down, we
can talk in generality about this problem, and some of its broad properties. We reserve for
Section 11.2 a more or less complete discussion of the properties of such polynomials.

Chapter 11
Ad hoc methods I: The root-locus method
The root-locus method we study in this section was put forward in the papers of
Evans [1948, 1950]. The study is of roots of polynomials when the coefficients depend
linearly on a parameter. In control systems, the parameter is typically the gain of a feedback loop, and our interest is in choosing the gain so that the closed-loop system is IBIBO
stable. As we saw in Section 10.2.2, with static output feedback of SISO systems, there
naturally arises a control problem where one has a polynomial with coefficients linear in a
parameter, and stabilisation requires choosing this parameter so that the polynomial is Hurwitz. In Section 11.1.1 below, we discuss some control problems where this scenario arises.
The manner of studying such problems in this chapter is to study how the roots move in
the complex plane as functions of the parameter. That is to say, we look at the locus of all
roots of the polynomial as the parameter varies, hence the name root-locus.
In many introductory texts, one will find a laying out of a design method using rootlocus methods. We do not devote significant effort to this for the reason that, according
to Horowitz [1963], It appears, therefore, that the root locus approach to the sensitivity
problem is justified only in systems where there are few dominant poles and zeros. These
systems are quite well understood in any case (see Section 13.2.3).

Contents
11.1 The root-locus problem, and its r
ole in control . . . . . . . . . . . . . . . . . . . . . . . . 448
11.1.1 A collection of problems in control . . . . . . . . . . . . . . . . . . . . . . . . . . . 448
11.1.2 Definitions and general properties . . . . . . . . . . . . . . . . . . . . . . . . . . . 449

11.1.1 A collection of problems in control Our first task is to indicate that there
are a collection of control problems which can be reduced to a problem of a certain type.
11.1 Examples 1. To get things rolling, let us consider the unity gain feedback loop of Figure 11.1 where the loop gain is known up to a multiplicative constant K, and let us
r(s)

RL (s)

y(s)

Figure 11.1 Root-locus from a negative feedback configuration

suppose that we ask that K be nonnegative. The transfer function from r(s) to y(s) is,
as usual,
y(s)
KRL (s)
=
.
r(s)
1 + KRL (s)
If RL has the c.f.r. (NL , DL ), then the characteristic polynomial for the interconnection
is DL + KNL . If RL is strictly proper, as is quite often the case, then we have deg(NL ) <
deg(DL ). The design objective is to determine whether there is a constant K0 0 so
that the characteristic polynomial DL + K0 NL is Hurwitz. One may even want to choose
K so that certain performance objectives are met. This is discussed in Section 11.3.
2. As we saw in Section 6.4.2, the closed-loop characteristic polynomial for static output
feedback of a SISO linear system = (A, b, ct , 01 ) is given by
PA (s) + F ct adj(sI n A)b

11.2 Properties of the root-locus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 451

11.2.1 A rigorous discussion of the root-locus

. . . . . . . . . . . . . . . . . . . . . . . . 451

11.2.2 The graphical method of Evans . . . . . . . . . . . . . . . . . . . . . . . . . . . . 460

11.3 Design based on the root-locus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 461
11.3.1 Location of closed-loop poles using root-locus

. . . . . . . . . . . . . . . . . . . . 461

11.3.2 Root sensitivity in root-locus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 462

11.4 The relationship between the root-locus and the Nyquist contour . . . . . . . . . . . . . . 463
11.4.1 The symmetry between gain and frequency . . . . . . . . . . . . . . . . . . . . . . 463
11.4.2 The characteristic gain and the characteristic frequency functions . . . . . . . . . 465

where F is the output feedback constant. Since ct adj(sI n A)b is a polynomial of degree
at most n 1, this will have the form of P1 + F P2 where deg(P2 ) < deg(P2 ).
3. It is possible that the variable constant K may not appear in as simple a manner as
indicated in Figure 11.1. However, in some such cases, one can still reduce the characteristic polynomial to one of the desired form. Suppose that we have a plant with
RP (s) = ms1 2 and we wish to stabilise it in a unity gain feedback loop with a PID controller RC (s) = K(1 + TD s + T1I s). Let us suppose that, for whatever reason, we are
interested only in changing the reset time TI , and not the gain K. Furthermore, we
restrict our interest to TI > 0. Thus we are not immediately in the situation illustrated
in Figure 11.1. Nevertheless, we proceed. The closed-loop characteristic polynomial is
s3 + KTD s2 + Ks +

K
.
TI

11.1 The root-locus problem, and its role in control

449

Now note that we may write this closed-loop characteristic polynomial as P1 + P2 if we

take
1
P1 = s3 + KTD s2 + Ks, P2 = K, = .
TI
Now, as TI runs from 0 to , runs from to 0. Thus, even though we are not
immediately in the form of Figure 11.1, we can write the characteristic polynomial is the
form of a sum of two polynomials, one with a positive coefficient. Note that this may
not always be possible, but sometimes it is.

In each of the preceding three examples, we arrive at a characteristic polynomial that is
of the form P1 + KP2 where deg(P2 ) < deg(P1 ). It is such polynomials that are of interest
to us in this chapter.
11.1.2 Definitions and general properties
section, we make the following definition.

Based on the discussion of the preceding

11.2 Definition Let N, D R[s] have the following properties:

(i) D is monic;
(ii) either N or N is monic;
(iii) deg(N ) < deg(D).
A (N, D)-polynomial family is a family of polynomials of the form P(N, D) =
{D + KN | K 0}. For a fixed K 0 we shall denote PK = D + KN R[s]. The
root-locus of an (N, D)-polynomial family is the subset
RL(P(N, D)) = { z C | PK (z) = 0 for some K 0}
of the complex plane.1 A (N, D)-polynomial family is Hurwitz if there exists K 0 so
that PK is Hurwitz.

11.3 Remark Note that we do ask for D to be monic and either N or N to be monic when
defining a (N, D)-polynomial family. That D should be taken as monic seems natural. As
for N , clearly one can make either N or N monic by simply redefining K if necessary.
Thus we loose no generality with these assumptions.

Note that the problem of determining whether a (N, D)-polynomial family is Hurwitz is exactly equivalent to the problem of determining whether a SISO linear system
= (A, b, ct , D) admits a stabilising static output feedback controller (see Exercise E11.2).
Therefore, to determine whether a (N, D)-polynomial family is Hurwitz is as difficult as
proving the existence of stabilising static output feedback. This problem, as was noted at
the beginning of Section 10.2.2 is NP hard. Thus we cannot expect to solve this problem
is an easily computable manner. However, we shall provide a rough description of the rootlocus for a (N, D)-polynomial family which suggests that you might be able to numerically
ascertain whether such a family is Hurwitz.
Note that the root-locus consists of collections of roots of polynomials that depend continuously, in this case linearly, on a parameter K. Let us record some properties of the roots
of such parameterised polynomials.
1

Note that S denotes the closure of the set S, meaning the set and all of its limit points.

450

11 Ad hoc methods I: The root-locus method

22/10/2004

11.4 Lemma Let P (s, K) = sn + pn1 (K)sn1 + + p1 (K)s + p0 (K) be a polynomial

with coefficients differentiable functions of the parameter K R. For K R, denote
by {z1 (K), . . . , zn (K)} the roots of P (s, K).
(i) The functions R 3 K 7 zi (K) C, i = 1, . . . , n, may be chosen to be continuous.
(ii) If the roots zi (K), i = 1, . . . , n, are chosen to be continuous functions, then, if zi (K0 )
is a root of multiplicity one for P (s, K0 ) then there exists > 0 so that zi |[K0 , K0 +]
so that
(a) fi (K0 ) = zi (K0 ) and
(b) P (zi (K), K) = 0.
More succinctly, near nondegenerate roots of P (s, K0 ), the roots are differentiable
functions of the parameter K.
11.5 Remark With the notation of the lemma, we may write the roots of P (s, K) as
{z1 (K), . . . , zn (K)}. We shall do this frequently, and we always make the assumption that
PSfrag replacements
this is done in such a way that the functions K 7 zi (K), i = 1, . . . , n, are continuous.
t
We do notke(t)k
prove this lemma, although part (ii) is easily proved using the implicit function
e1 (t)
theorem (see Exercise
E11.3). Let us illustrate the lemma with a simple example.
e2 (t)
e3 (t)
y(t) us take P (s, K) = s2 + K. For K 0 the roots are {z1 (K) =
11.6 Example Let

K, z2 (K) h=u(t)
K} and for K > 0 the roots are {z1 (K) = i K, z2 (K) = i K}. In
(t)
Figure 11.2 hwe
plot the locus of roots for this polynomial as K runs from 4 to 4. Note that
N,D (t)
1 (t)
1N,D (t) 2
(t)
(t)
fj (t)
1

x1
x2
x1
x2
log
0
dB
deg
u = 0 ln
ln coth(|u| /2)
or 1 -1
m

yos

tos 0 -2

22/10/2004

,0
0

-2

-1

Figure 11.2 Locus of roots for P (s, K) = s2 + K

if K varies slightly, the location of the roots also change only slightly. Whats more, as long
as K 6= 0 the roots are distinct, and the locus of roots near such values of K are smooth

curves in the complex plane (lines in this case). However, for the repeated root when K = 0,
the character of the locus of roots is not smooth near this repeated root. Indeed, at the origin
where this repeated root lies, the locus of roots has an intersection. This will typically be
the case, and constitutes one of the more challenging aspects of making root-locus plots.

11.2 Properties of the root-locus

In this section we do two things. First we provide a list of provable properties of the
root-locus of a (N, D)-polynomial family. These are presented with proofs as these do not
seem to be part of the standard discussion of the root-locus method. The presentation in
this first part of the section produces for us a good understanding of what is known and
unknown about the nature of the root-locus. The next thing we do is put this understanding
together to develop a methodology for graphically producing the root-locus for a (N, D)polynomial family. It is this graphical method which forms the bulk of the presentation on
the root-locus method in most standard texts.
11.2.1 A rigorous discussion of the root-locus We follow here the paper of Krall
[1961], starting with a (N, D)-polynomial family P(N, D). The essential idea is that when
K = 0 the roots of PK are obviously those of D. Suppose that deg(D) = n so that P0 has n
roots, if one counts multiplicities. Now, as K increases, Lemma 11.4 suggests that these n
roots should move about continuously in the complex plane. Now suppose that m = deg(N ).
What will turn out to happen is that m of the n roots of PK will start at the roots for D = P0
and end up at the roots of N as K . One must then account for the remaining n m
roots, which, it turns out, shoot off to infinity in predictable ways.
Lets get down to it. For concreteness let us write

(11.1)

The following result gives a simple characterisation of the centre of gravity in terms of the
roots of N and D.
n

11 Ad hoc methods I: The root-locus method

22/10/2004

Now, through the centre of gravity we construct n m rays in the complex plane. We
denote these rays by 1 , . . . , nm PSfrag
and replacements
define them by
t
t
ke(t)k
ke(t)k
(

(2j1)i/(nm)
e1 (t)
e
(t)
r 0 , N monic
1
CG(P(N, D)) + re e (t)
e2 (t)
2

=
(11.2)
j
e3 (t)
e3 (t)
r0 ,
CG(P(N, D)) + re2ji/(nm)
-N monic.
y(t)
y(t)

PSfrag replacements

u(t)
hWe
(t) call
hN,D (t)
1 (t)
1N,D (t) 3
(t)
(t)
fj (t) 2
x1
x2 1
x1
x2
log
0
dB
deg
u = 0 ln
ln coth(|u| /2) -1
or 1
m
-2
yos

tos 0

u(t)

h (t) D). In Figure 11.3 we show the asymptotes

these rays the asymptotes for P(N,
hN,D (t)
1 (t)
1N,D (t)
(t)
(t)
fj (t)

,0
0

x1
x2 1
x1
x2
log
0
dB
deg
u = 0 ln
ln coth(|u| /2) -1
or 1
m
-2
yos

tos 0

-2

-1

3,0

-2

-1

Figure 11.3 Asymptotes if CG(P(N, D)) = 1 + i0 and n m = 3,

and when N is monic (left) and N is monic (right)

when the centre of gravity is at 1 + i0 and when n m = 3. Note that generally the
asymptotes will depend on whether N or N is monic.
With the notion of asymptotes in place, we may state the following result which indicates
in what manner some of the roots of PK go to infinity.

D(s) = sn + pn1 sn1 + + p1 s + p0

N (s) = sm + qm1 sm1 + + q1 s + q0 .
We define the centre of gravity of P(N, D) to be
(
qm1
, N monic
pn1nm
CG(P(N, D)) =
pn1 +qm1
nm , N monic.

452

451

11.2 Properties of the root-locus

22/10/2004

X
1 X
zj
k , where z1 , . . . , zn are the roots of D
n m j=1
k=1
are the roots for N .

11.8 Proposition Let P(N, D) be a (N, D)-polynomial family and denote the roots of PK =
D + KN by {z1 (K), . . . , zn (K)}. Then there exists distinct j1 , . . . , jnm {1, . . . , n} so that

(i) if N is monic, limK zjk (K) CG(P(N, D)) K 1/(nm) e(2k1)i/(nm) = 0, and

(ii) if N is monic, limK zj (K) CG(P(N, D)) K 1/(nm) e2ki/(nm) = 0,
k

for each k = 1, . . . , n m.
Proof The bulk of the proof is contained in the following result.

11.7 Lemma CG(P(N, D)) =

and 1 , . . . , m

Proof For a monic polynomial

1 Lemma The proposition holds when CG(P(N, D)) = 0 + i0.

Proof In this case we can write

PK (s) = sn +asn1 +pn2 sn2 + +p1 s+p0 Kei sm +asm1 +qm2 sm2 + +q1 s+q0 ,

P (s) = s + an1 s

+ + a1 s + a0 ,

one may easily show that the sum of the roots of P is equal to an1 (see Exercise EC.3).
Thus the sum of the roots for D is pn1 , and the sum of the roots for N is qm1 is N is
monic and qn1 is N is monic.

where K 0 and {0, }. The main point is that since CG(P(N, D)) = 0+i0, the next
to highest coefficients of D and N must be equal if N is monic. Now fix k {1, . . . , n m}
and make the substitution
s = weik ,

k =

2k +
, w C,
nm

11.2 Properties of the root-locus

22/10/2004

453

so that
ik

PK (we ) = w + aw

n1 ik

n2
X

aj w

K w + aw

m1 ik

m2
X

bj w ,

j=0

454

11 Ad hoc methods I: The root-locus method

when {1, e2i/(nm) , e2i/(nm) , . . . }. Therefore, by the inequality (11.3), nm 1 is

greater than when . By the triangle inequality we also have

a

a

+ eik || eik

for some coefficients a0 , . . . , an2 , b0 , . . . , bm2 whose exact character are not of much interest
to us. Now define
M=

n2
X

|aj | (3 |a| + 1)

n2j

4 j
3

j=0

and
= K 1/(nm) ,

m2
X

Therefore, for we have

|bj | (3 |a| + 1)

m2j

j=0

4 j

|f ()| > (1 )m1 31 1 >

|g()|

The key observation to make is that M is independent of K. Let

(nm)/2 o
3m M
.
i/(2(nm))
|1 e
|
2

n

K0 = max (3 |a| + 1)nm , (3m+1 M )(nm)/2 , 1

M
=
2

1 m
.
3

Therefore, for we have |f ()| > |g()|. From Rouches theorem, Theorem D.6, we may
conclude that the number of zeros for f + g within is the same as the number of zeros f ,
namely 1, provided that K > K0 . This shows that there is a zero 0 inside so that

If K > K0 then we have

1/(nm)

1 m
.
3

Thus there exists K0 (redefine this if necessary) so that for all K > K0 , f has only the zero
at 1 + i0 within . By definition of M , for we have

3m M
3m M
=
= 2/(nm) .
2

= K 1/(nm) > (3 |a| + 1)nm

22/10/2004

n eik PK (0 eik ) = 0

> 3 |a| + 1,

PK (z0 ) = 0,

where

and

z0 = 0 eik = K 1/(nm) 0 eik .

3m M
3m M
1
= 2/(nm) <
2/(nm) = ,
K
3
m+1
(nm)/2
(3
M)

The lemma now follows since limK 0 = 1, since limK = 0.

With the lemma in hand, it is now comparatively straightforward to prove the proposition. Let us proceed in the case when N is monic so that

and
m

3 M
<
K 2/(nm)

3 M
1
i/(2(nm))
.
(nm)/2 2/(nm) = 1 e
2
1
i/(2(nm))
1e
|
|
2
3m M

Now make the substitution =

(11.3)

C so that, with our previous substitution for s, we have

n2
m2

X
X
a
a
aj jn j m + eik m1 +
bj jm j
n eik PK (eik ) = n + eik n1 +

j=0
j=0

CG(P(N, D)) =

Now define

a
f () =
+ eik ( nm 1)

n2
m2

X
1 X
g() = 2
aj jn+2 j
bj jm+2 j .
j=0
j=0

Now let be the circle of radius centred at 1+i0, with as previously defined. Since < 31 ,
the real part of k C is positive on for k = 1, . . . , n m. Note that nm 1 vanishes

pnq qn1
.
nm

Making the substitution s = w + CG(P(N, D)) gives

PK (w + CG(P(N, D))) = wn +

n2
m2
a

X
1 X
= m1 eik ( nm 1) + 2
aj jn+2 j
bj jm+2 j .
j=0
j=0

nqn1 mpn1 n1 X
w
+
aj w j +
nm
j=0
m2

nqn1 mpn1 m1 X
K wm +
w
+
bj w j ,
nm
j=0

for some coefficients a0 , an2 , b0 , . . . , bm2 whose exact form is not of particular interest to us.
Now, by Lemma 1, there is a collection of roots {j1 (K), . . . , jnm (K)} to this polynomial
that satisfy

lim jk (K) K 1/(nm) e2ki/(nm) = 0.
K

These clearly give rise to the roots {zj1 (K), . . . , zjnm (K)} as given in the statement of the
proposition.

Let us see how this works in a simple example.

check

1 (t)
1N,D (t)
(t)
(t)
fj (t)

11.2 Properties of the root-locus

455

11.9 Example Let us take (N (s), D(s)) = (1, s2 ). Note that CG(P(N, D)) = 0 + i0.
First let us consider the case where N = 1 so that PK (s) = s2 + K. Since n m = 2 we
11.8(i),in the
expect there to be 2 roots that shoot off to infinity.
According
to Proposition

limit as K these roots should behave like Kei/2 = i K and Ke3i/2 = i K as

K gets large. Not only do the roots behave like this as K gets large, they are exactly given
by these expressions for all K!
that
When N = 1 we have PK (s) = s2 K.
In this case,Proposition
11.8(ii) predicts

as K the two roots of PK behave like Kei = K and Ke2i = + K. Again,

these happen to be the exact values of the roots. Clearly we do not expect this to generally
be the case.

Now let us consider what happens to those remaining m roots of PK as K varies.
11.10 Proposition Let P(N, D) be a (N, D) polynomial family with {z1 (K), . . . , zn (K)} the
roots of PK for K 0. Also denote {1 , . . . , m } as the roots for N . There exists distinct
j1 , . . . , jm {1, . . . , n} so that for each k = 1, . . . , m, the curve K 7 zjk (K) is a continuous
curve starting from zjk (0) and satisfying limK zjk (K) = k .
Proof By part (i) of Lemma 11.4 we know that the curve K 7 zj (K) is continuous for
every j = 1, . . . , n. Now fix k {1, . . . , m} and suppose that the multiplicity of the root k
is `. Choose > 0 and let be the circle of radius centred at k . Take sufficiently small
that there are no roots of N within but zk . The number of zeros of PK within is given
by the Principle of the Argument to be
Z
1
PK0 (s)
ds.
2i PK (s)
Using PK (s) = D(s) + KN (s) we compute
Z
Z
Z
PK0 (s)
1
N 0 (s)
1
1 D0 (s)N (s) D(s)N 0 (s)
1
ds =
ds +
ds.
2i PK (s)
2i N (s)
2i K ( K1 D(s) + N (s))N (s)
Since is bounded, both polynomials N and D are bounded on . Therefore, we may
choose K sufficiently large that K1 |D(s)| < |N (s)| for s . Therefore we see that the
number of zeros of PK within is given by ` + where may be made as small as we please
by increasing K. Thus we conclude that for K sufficiently large, PK has ` zeros in . As
this holds for all > 0, we conclude that ` zeros of PK limit to k as K . As this is
true for all roots k for N , this proves the proposition.

Let us illustrate this proposition in another simple example.
11.11 Example We take (N (s), D(s)) = (s + 1, s2 ) so that PK (s) = s2 + Ks + K. Note that
CG(P(N, D)) = 1 + i0. Since n m = 1 and N is monic, by part (i) of Proposition 11.8
we expect to see one of the roots approaching 1 + Kei = 1 K as K . The other root
should start at one of the roots for D and end up in the root = 1 for N . We compute
the roots to be
q
K2 ( K2 )2 K,
and the root-locus is shown in Figure 11.4. Note that one of the roots does indeed start
out at 0 + i0, a root for D, and moves continuously toward 1 + i0, a root for N . In the
root-locus of Figure 11.4, the path taken by this root is not unique. It can start at 0 + i0

456

11 Ad hoc methods I: The root-locus method

x1
x2
1
x1
x2
log
0.5
dB
deg
u = 0 ln
ln coth(|u| /2)
0
O
or 1
m
-0.5
yos

tos 0
-1

-3
-2.5
-2
-1.5
-1
-0.5
,0
Re
0
Im

22/10/2004

Figure 11.4 Root locus for (N (s), D(s)) = (s + 1, s2 )

and go down along the semi-circle, then turn right toward 1 + i0, or it can go up along the
other semi-circle, and then again turn right toward 1 + i0. Note that Proposition 11.10
does not preclude this ambiguity. Also, we see in Figure 11.4 that one of the roots is going
off to as predicted.
Let us see if we can verify the limiting behaviour analytically. For K large, both roots
of PK are real. Let us write
q
q
z1 (K) = K2 ( K2 )2 K, z2 (K) = K2 + ( K2 )2 K
Note that
z1 (K) (1 K) = 1 +

K
2

= 1+

K
2

q
( K2 )2 K
q
( K2 ) 1 K4

= 1 + K2 ( K2 ) 1
= K1 + ,

1 4
2K

18 ( K4 )2 +

so that limK (z1 (K) (1 K)) = 0 as desired. Similarly we have

q
z2 (K) (1) = 1 K2 + ( K2 )2 K
q
= 1 K2 + ( K2 ) 1 K4
= 1 K2 + ( K2 ) 1
= K1 + ,

1 4
2K

18 ( K4 )2 +

so that limK (z2 (K) (1))

= 0, as predicted. In both of these computations, we have

used the Taylor series for 1 x about x = 0.
Now we wish to examine the way that the root-locus passes through certain points on
the root-locus. Let z0 RL(P(N, D)) so that z0 is a root of PK0 for some K0 0. If
the multiplicity of z0 is one, then the root-locus passes differentiably through z0 . Denote by

11.2 Properties of the root-locus

22/10/2004

457

zi (K) that root for PK for which zi (K0 ) = z0 . We define the arrival angle of RL(P(N, D))
at z0 to be the angle a (z0 ) (, ] for which

458

11 Ad hoc methods I: The root-locus method

(a) if N is monic, take
n

a,k (j ) =

zi (K) z0
= a eia (z0 )
lim
KK0 K K0
for some suitable a > 0. This make precise the intuitive notion of arrival angle. Similarly,
the departure angle of RL(P(N, D)) at z0 is the angle d (z0 ) (, ] for which

1 X
](j z )
` =1

m
X

KK0

zi (K) z0
= d eid (z0 )
K K0

lim

KK0

zi (K) z0
= a,k eia,k (z0 ) ,
K K0

for some suitable a,1 , . . . , a,` > 0.

d,1 (z0 ), . . . , d,` (z0 ), by
lim

KK0

1 X
](j z )
` =1

j=1

Similarly, we may define ` departure angles,

D(zjk (K)) Kei N (zjk (K)) = 0

zi (K) z0
= d,k eid,k (z0 ) ,
K K0

k = 1, . . . , `,

(a) if N is monic, take

1 X
](zj )
` =1

n
X

Kei N (zjk (K))

= 1.
D(zjk (K))

m
X

](s j )

j=1

n
X

](s zj ) = 2k

j=1

where k Z. Therefore,
`](s zj ) =

m
X

n
X

](s )

](s z ) + 2k,

k Z.

=1
excluding zj

In the limit as K 0, the result follows, noting the relation between and N or N being
monic.
(ii) The argument is exactly as in part (i), except we let K .

Let us examine this again in an example.

](zj z ) (2k 1) ,

k = 1, . . . , `;

=1
excluding zj

(b) if N is monic, take

1 X
](zj )
` =1

Taking complex logarithms of both sides, using (11.4), we obtain

11.12 Proposition Let P(N, D) be an (N, D)-polynomial family and let {z1 , . . . , zn } be the
roots of D and {1 , . . . , m } be the roots of N .
(i) If zj has multiplicity `, then d (zj ) is the following collection of ` angles:

k = 1, . . . , `.

=1
excluding j

Proof Let us for convenience write PK = D Kei N and assume that N is monic. Thus,
we recover the monic cases for N and N by taking = and = 0, respectively. We then
write
n
m
Y
Y
D(s) =
(s zj ), N (s) =
(s j ).
(11.4)

d,k (zj ) =

](j ) + 2k ,

(i) For j {1, . . . , n}, let ` be the multiplicity of zj . For k {1, . . . , `} let zjk (K) be a
root of PK which approaches zj as K 0. Note that

for some suitable d,1 , . . . , d,` > 0. For any multiplicity of the root z0 , let us denote by
a (z0 ) the collection of all arrival angles, and by d (z0 ) the collection of all departure
angles.
We are now ready to state the character of some of the arrival and departure angles,
namely the departure angles for the roots of D and the arrival angles for the roots of N . In
practice, these are the most helpful to know in terms of being able to produce the root-locus
for an (N, D)-polynomial family.

d,k (zj ) =

m
X

j=1

k = 1, . . . , `,

k = 1, . . . , `;

(b) if N is monic, take

d,k (j ) =

for some suitable d > 0. Since the root-locus is differentiable at z0 , one can easily see that
we must have d (z0 ) {a (z0 ) + , a (z0 ) }. When z0 has multiplicity ` > 1, things are
not so transparent since there will be more than one arrival and departure angle. In this
case we have ` roots zj1 (K), . . . , zj` (K) which can pass through z0 when K0 . We may define
` arrival angles, a,1 (z0 ), . . . , a,` (z0 ), by

](j ) + (2k 1) ,

=1
excluding zj

lim

22/10/2004

n
X

](zj z ) 2k ,

11.13 Example (Example 11.11 contd) We again take (N (s), D(s)) = (s+1, s2 ). In this case
we have a root z1 = 0 for D of multiplicity 2 and a root 1 = 1 for N of multiplicity 1.
Since N is monic, Proposition 11.12 predicts that
d (z1 ) = { 2 , 3
},
2

k = 1, . . . , `.

=1
excluding zj

(ii) If j has multiplicity `, then a (j ) is the following collection of ` angles:

This is indeed consistent with Figure 11.4.

a (1 ) = {}.

The next result tells us that we should expect to see certain parts of the real axis within
the root-locus.

11.2 Properties of the root-locus

22/10/2004

459

11.14 Proposition Let P(N, D) be an (N, D)-polynomial family, and denote the roots of D
by {z1 , . . . , zn } and the roots of N by {1 , . . . , m }.
(i) Suppose that N is monic and let x0 R C. Then x0 RL(P(N, D)) if and only if
the set
{ j | Re(zj ) > x0 } { j | Re(j ) > x0 }
has odd cardinality.2
(ii) Suppose that N is monic and let x0 R C. Then x0 RL(P(N, D)) if and only
if the set
{ j | Re(zj ) > x0 } { j | Re(j ) > x0 }
has even cardinality.
Proof (i) First note that since the roots of PK come in complex conjugate pairs, roots with
nonzero imaginary part will always contribute an even number to the cardinality of the sets
{ j | Re(zj ) > x0 } ,

{ j | Re(j ) > x0 } .

Thus it suffices to consider only the sets

{ j | zj real and zj > x0 } ,

{ j | j real and j > x0 }

for odd cardinality. Let x0 R RL(P(N, D)). Then

D(x0 ) + KN (x0 ) = 1

KN (x0 )
= 1.
D(x0 )

Taking complex logarithms, and using the fact that D and N can be written as in (11.4),
gives
m
n
X
X
](x0 j )
](x0 zj ) = (2k 1)
(11.5)
j=1

j=1

for k Z. Clearly, the sums can be taken as being over real roots since the terms corresponding to a complex root for N or D and its conjugate will cancel from (11.5). If the
cardinality of the set
{ j | Re(zj ) > x0 } { j | Re(j ) > x0 }
is odd, then we will have
m
X
j=1

](x0 j )

n
X

](x0 zj ) = (2r 1)

j=1

for some r Z. Thus this part of the proposition follows.

(ii) The proof here goes just as in part (i), except that we write D(x0 ) KN (x0 ) = 1
and assume N monic.

As always, let us check the conclusions of the proposition on an example.
11.15 Example (Example 11.11 contd) We again take (N (s), D(s)) = (s + 1, s2 ). The roots
of D are {z1 = 0, z2 = 0} and the roots of N are {1 = 1}. Thus, if x0 R C, there are
an odd number of zeros for both D and N to the right of x0 is and only if x0 < 1. Then
Proposition 11.14 tells us that
RL(P(N, D)) R = { x + i0 | x < 1} .
This is indeed consistent with Figure 11.4.
2

The cardinality of a set S is simply the number of points in S.

460

11 Ad hoc methods I: The root-locus method

11.2.2 The graphical method of Evans Now we may present the graphical technique
typically presented in classical texts for producing the root-locus. This rather ingenious
technique was that developed by Evans [1948, 1950]. For us, this is simply a matter of
applying the results of the preceding section, and we shall only enumerate the steps one
typically takes in such a construction.
11.16 Steps for making a plot of the root-locus Given:
an
(N, D)-polynomial
family
P(N, D).
1. Compute the roots {z1 , . . . , zn } for D and place an X at the location of each root in C.
2. Compute the roots {1 , . . . , m } for N and place an O at the location of each root in C.
3. Compute the centre of gravity using (11.1) or Lemma 11.7.
4. Draw the n m asymptotes using (11.2).
5. Use Proposition 11.14 to determine RL(P(N, D)) R.
6. Use Proposition 11.12 to determine how the root-locus departs from the roots of D.
7. Use Proposition 11.12 to determine how the root-locus arrives at the roots of N .
8. If you are lucky, you can give a reasonable guess as to how the root locus behaves. For
filling in the gaps in a root-locus plot, a useful property of the root-locus is that it is
invariant under complex conjugation.

The last step is in some sense the most crucial. It is possible that one can do the steps
preceding it, and still get the root-locus wrong. Some experience is typically involved in
knowing how a typical root-locus diagram looks, and then extrapolating this to a given
example. Thankfully, computers typically do a good job with producing root-locus plots.
These can run into problems when there are repeated roots of PK , however. Thus one should
check that a computer produced root-locus has the essential properties, mainly the correct
number of branches.
Let us go through the steps outlined above in an example to see how well it works.
11.17 Example We take (N (s), D(s) = (s + 1, s4 + 6s3 + 14s2 + 16s + 8).
1. The roots for D are {z1 = 2, z2 = 2, z3 = 1 i, z4 = 1 + i}.
2. The roots for N are {1 = 1}.
3. CG(P(N, D)) = 53 .
4. The asymptotes are given by

i
1 = 53 + re 3 r 0

2 = 53 + rei r 0
5

5i
3 = 3 + re 3 r 0 .
We show these asymptotes in Figure 11.5.
5. The point on the real-axis which lie on the root-locus are those points x < 2.
6. The departure angles from the roots of D are
d (z1 ) = d (z2 ) = {0, P i},

22/10/2004

d (z3 ) = { 2 },

These departure angles are shown in Figure 11.6

d (z4 ) = { 3
}.
2

ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)
22/10/2004 h (t)
hN,D (t)
1 (t)
1N,D (t)
(t)
(t)
fj (t)

PSfrag replacements

11.3 Design based on the root-locus

461

11 Ad hoc methods
3

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)
I: The
h (t)root-locus
hN,D (t)
1 (t)
1N,D (t) 3
(t)
(t)
fj (t) 2

method

22/10/2004

,0
0

-3

-2

-1

Figure 11.5 The asymptotes for (N (s), D(s) = (s + 1, s4 + 6s3 +

14s2 + 16s + 8).

7. The arrival angles from the roots of N are

a (1 ) = {}.
8. We play connect the dots, hoping that we will arrive at a decent approximation of
the actual root-locus. The actual root locus, along with the skeleton produced by our
procedure, is shown in Figure 11.7. Note that without the knowledge of which roots of
D go where as K , it is in actuality difficult to know the details of the character of
the root locus. Nevertheless, in simple examples one can often figure out the character
of the root locus.

11.3 Design based on the root-locus

The root-locus method can be thought of as a design technique. As such, it is most
useful for plants that can be stabilised using simple controllers. For such plants, the rootlocus method allows one to evaluate a one-parameter family of controllers by investigating
the root-locus plot as the parameter varies. This assumes that one can put the system into
a form where as the parameter varies we produce an (N, D)-polynomial family. We argued
in Section 11.1.1 that this can sometimes be done, although it will by no means always be
the case.
11.3.1 Location of closed-loop poles using root-locus For systems that have simple
enough behaviour, one can attempt to determine the quality of the performance of the system
based on the location of the transfer function poles in the complex plane. For example, as we

x1
x2 1
x1
x2
log
0
dB
deg
u = 0 ln
ln coth(|u| /2) -1
or 1
m
-2
yos

tos 0

x1
x2 1
x1
x2
log
0
dB
deg
u = 0 ln
ln coth(|u| /2) -1
or 1
m
-2
yos

tos 0

x1
x2 1
x1
x2
log
0
dB
deg
u = 0 ln
ln coth(|u| /2) -1
or 1
m
-2
yos

tos 0

,0
0

PSfrag replacements

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)
h462
(t)
hN,D (t)
1 (t)
1N,D (t)
(t)
(t)
fj (t)

-3

-2

-1

2,0

-3

-2

-1

Figure 11.6 Departure angles from roots of D(s) = s4 +6s3 +14s2 +

16s + 8 (left) and arrival angle from the roots of N (s) = s + 1
(right).

saw in Section 8.2, first and second-order transfer functions have performance attributes that
are easily related to pole locations. Sometimes in practice one is able to design a controller
that makes a system behave similarly to one of these two simple transfer functions, so one
can use them as a basis for control design. With this as flimsy justification, we deal with
interconnections like that depicted in Figure 11.8, and make the following assumption.
In this section we will assume that the closed-loop interconnection of Figure 11.8
is designed so that for some values of K, the closed-loop transfer function is
stable, and has a pair of complex conjugate poles whose real part is larger than
that of all other poles. We call these complex conjugate poles the dominant
poles.
The idea is that we pretend the dominant poles allow us to think of the closed-loop system
as being second-order, and we base our choice of K on this consideration. In practice, one
may wish to find a second-order transfer function that well approximates the system by, say,
matching Bode plots as best one can.
Our approach is to carefully observe the relationship between system performance and
pole location for second-order transfer functions. Thus we let
T,0 (s) =

02
.
s2 + 20 s + 02

For this transfer function, let us list some of the more important performance measures,
some approximately, in terms of the parameters and 0 .
11.18 Performance measures for second-order transfer functions Consider the transfer function
T,0 .
1.
11.3.2 Root sensitivity in root-locus

See Bishop and Dorf.

ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)
(t) The relationship between the root-locus and the Nyquist contour
22/10/2004 h11.4
hN,D (t)
1 (t)
1N,D (t) 3
(t)
(t)
fj (t) 2
x1
x2 1
x1
x2
log
0
dB
deg
u = 0 ln
ln coth(|u| /2) -1
or 1
m
-2
yos

tos 0

463

464

11 Ad hoc methods I: The root-locus method

22/10/2004

r(s)

(sI n A)1

y(s)

k 1
X

(sI n A)1
X

r(s)
-3

,0
0

-2

-1

(kI 1 D)1

y(s)

Figure 11.7 The skeleton and the actual root-locus

(N (s), D(s) = (s + 1, s4 + 6s3 + 14s2 + 16s + 8).

Figure 11.9 Equivalent block diagrams for studying the relationship between the root-locus and the Nyquist contour

for

Let us write
r(s)

RL (s)

y(s)

K(s) = ct (sI n A)1 b + D R(s)11

S(k) = ct (kI 1 D)1 b + A R(k)nn .

The following lemma tells us how we should interpret these matrices of rational functions.

Figure 11.8 Interconnection for investigating closed-loop pole locations

11.19 Lemma The following statements hold:

(i) K(s) is the open-loop transfer function from r to y for the uppermost block diagram
in Figure 11.9, i.e., the transfer function if the feedback loop is snipped;
(ii) S(k) is the A-matrix for the closed-loop system in Figure 11.9, i.e., if the closed-loop
c
b,
then A
= S(k).
= (A,
t , D),
transfer function has canonical minimal realisation

11.4 The relationship between the root-locus and the Nyquist contour
It turns out that there are some rather unobvious connections between the root-locus
and the Nyquist contour. This is explained in a MIMO setting in [MacFarlane 1982], and
we only look at the SISO case here.
11.4.1 The symmetry between gain and frequency We begin with a proper rational
function R with c.f.r. (N, D) and canonical minimal realisation = (A, b, ct , D). We next
place R, represented by its canonical minimal realisation, into a positive feedback loop with
feedback gain k 1 as shown in the upper block diagram in Figure 11.9. There are a few things
to note here: (1) there is no negative sign where the feedback enters the summer with the
reference r, (2) except for the sign, the block diagram is the same as that for static output
feedback as studied in Section 6.4.2, and (3) the top block diagram in Figure 11.9 is equivalent
to the bottom block diagram in the same figure. The second of these block diagrams has
the advantage of making apparent a symmetry that exists between the parameter s and the
parameter k. It is this symmetry which we study here.

cite info

Proof The first assertion is clear. The second follows from the same computations that
gave the closed-loop system under static output feedback in Section 6.4.2.

Note that the closed-loop system is internally stable if and only if spec(S(k)) C .
We wish to understand a similar relationship between closed-loop stability and K(s). The
following lemma gives us the essence of this relationship
11.20 Lemma If s C \ spec(A) and k C \ spec(D), then the following statements are
equivalent:
(i) det(sI n S(k)) = 0;
(ii) det(kI 1 K(s)) = 0.

22/10/2004

11.4 The relationship between the root-locus and the Nyquist contour

465

Proof First note that if s and k are as hypothesised, then S(k) and K(s) are well-defined.
Since is assumed controllable and observable, the poles of the closed-loop transfer function for the system in Figure 11.9 are exactly the eigenvalues of S(k), as a consequence of
Lemma 11.19(ii). However, the closed-loop transfer function is
K(s)
k 1 K(s)
=
,
1 k 1 K(s)
kI 1 K(s)
giving the lemma.

From this we have as a consequence the test for closed-loop stability in terms of K(s).
11.21 Corollary spec(S(k)) C if and only if spec(K(s)) C .
Proof The closed-loop system of Figure 11.9 is BIBO stable if and only if spec(S(k)) C
if and only if the poles of the closed-loop transfer function lie in C . The result then follows

from Lemma 11.20.

466

11 Ad hoc methods I: The root-locus method

Exercises
E11.1 Let P(N, D) = { D + KN | K 0} be a (N, D)-polynomial family.
(a) Show that there exists RL R[s] so that the closed-loop characteristic polynomial for the interconnection of Figure 11.1 is D + KN . Explicitly give RL in
terms of N and D.
(b) Provide RL for the (N, D)-polynomial family of Example 11.13.
In Section 11.1.1 we saw that the problem of static output feedback leads naturally to
a (N, D)-polynomial family. In the next exercise, you will show that the converse also
happens, i.e., that a (N, D)-polynomial family leads naturally to a static output feedback
problem.
E11.2 Let P(N, D) = { D + KN | K 0} be a (N, D)-polynomial family. Show that
there exists a SISO linear system = (A, b, ct , D) so that the closed-loop characteristic polynomial for the static output feedback interconnection of Figure E11.1 is

What we have done to this point how the gain k C and the frequency s C have a
symmetric relationship in determining the closed-loop stability of the system in Figure 11.9.
Now let us see how this investigation bears fruit.
11.4.2 The characteristic gain and the characteristic frequency functions
we saw in the preceding discussion was that the two meromorphic functions

What

K
x0
r(s)

Fk (s) = det(sI n S(k)) R(s)

Fs (k) = det(kI 1 K(s)) R(k).
With this as motivation we have the following result giving us the precise manner in which
the gain and frequency are related. First we make two important definitions. To do so, the
reader will wish to recall our discussion in Section D.5 on algebraic functions and Riemann
surfaces.

22/10/2004

(sI n A)1

y(s)

Figure E11.1 Static output feedback and (N, D)-polynomial families

exactly PK .
E11.3 Prove part (ii) of Lemma 11.4 using the implicit function theorem.
The development of the root-locus properties for an (N, D)-polynomial family assumed that
deg(N ) < deg(D). It is also possible to proceed when confronted with the case when
deg(N ) > deg(D).
E11.4 Let N, D R[s] and assume that D is monic, that either N or N is monic, and
that deg(N ) > deg(D). Define

P(N,
D) = {D + KN | K 0} ,
PK = D + KN , and let

P(N,
RL(
D)) = { z C | PK (z) = 0 for some K 0}.
D
R[s] satisfying the conditions of Definition 11.2 so
Show that there exists N,

D))
= RL(
P(N,
that RL(P(N,
D)).

Exercises for Chapter 11

E11.5 Let = (A, b, ct , 01 ) where

0 1 0

A = 0 0 1 ,
0 0 0

0
b = 0 ,
1

467

1
.
1 1

Show using root-locus method that it is not possible to stabilise the system using
static output feedback.

468

11 Ad hoc methods I: The root-locus method

22/10/2004

470 12 Ad hoc methods II: Simple frequency response methods for controller design 22/10/2004

This version: 22/10/2004

Chapter 12
Ad hoc methods II: Simple frequency response
methods for controller design

characteristics. The term compensation is used for the process of designing a controller
rational function, reflecting the fact that one is compensating for deficiencies in the innate
properties of the plant. In Section 6.5 we investigated the PID controller, and discussed some
of its properties. In this section we introduce another class of controller transfer functions
which are useful in shaping the frequency response. We also discuss PID compensation in a
manner slightly different from our previous discussion.
12.1.1 Lead and lag compensation The first type of compensation we look at deals
essentially with the manipulation of the phase of the closed-loop system.
12.1 Definition A lead/lag compensator is a rational function of the form

In Chapter 7 we introduced a frequency domain technique, the Nyquist criterion, for

studying closed-loop stability. The stability margins introduced in Section 7.2 suggest that
the frequency domain techniques may be useful in control design, as they offer tangible
objectives for the properties of the closed-loop system. The design ideas we discuss in this
chapter are essentially ad hoc, but are frequently useful in situations where controller design
is not essentially difficultat least as concerns obtaining closed-loop stabilitybut where
one would like guidance in improving the performance of a controller. The methods we use
in this chapter fall under the broad heading of loop shaping , as the objective is to shape
the Nyquist plot to have desired properties. In this chapter the emphasis is on choosing a
priori a form for the controller, then using the design methodology to improve the closedloop response. In Chapter 15, the design method does not assume a form for the controller,
but rather the form of the controller is decided upon by the objectives of the problem.

Contents
12.1 Compensation in the frequency domain . . . . . . . . . . . . . . . . . . . . . . . . . . . . 469
12.1.1 Lead and lag compensation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 470
12.1.2 PID compensation in the frequency domain . . . . . . . . . . . . . . . . . . . . . . 473
12.2 Design using controllers of predetermined form . . . . . . . . . . . . . . . . . . . . . . . . 475
12.2.1 Using the Nyquist plot to choose a gain . . . . . . . . . . . . . . . . . . . . . . . . 476
12.2.2 A design methodology using lead and integrator compensation . . . . . . . . . . . 477
12.2.3 A design methodology using PID compensation . . . . . . . . . . . . . . . . . . . 483
12.2.4 A discussion of design methodologies . . . . . . . . . . . . . . . . . . . . . . . . . 486
12.3 Design with open controller form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 488
12.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 488

12.1 Compensation in the frequency domain

The discussion of the previous few sections has focused on analysis of closed-loop systems
using frequency domain techniques. In this discussion, certain key contributors to stability
were identified, principally gain and phase margin. Faced with a plant transfer function
RP , one will want to design a controller rational function RC which has certain desirable

RC (s) =

K(1 + s)
1 + s

for K, , R with 6 {0, 1} and K, 6= 0. It is a lead compensator when || > 1 and

a lag compensator when || < 1.

can be put into the form of a lead/lag
Note that any rational function of the form A s+z
s+p
compensator by defining
1
p
A
= , = , K= .
p
z

The form for the lead/lag compensator as we define it is convenient for analysing the nature
of the compensator, as we shall see.
The following result gives the essential properties of a lead/lag compensator.
s
12.2 Proposition If RC (s) = 1+
is a lead/lag compensator then, with HC () = RC (i),
1+ s
the following statements hold:
(i) lim0 ]HC () = 0;
(ii) (a) lim ]HC () = 0 if > 0;
(b) lim ]HC () = 180 if < 0 and < 0;
(c) lim ]HC () = 180 if < 0 and > 0;
(iii) lim0 |HC ()| = 0dB;
(iv) lim |HC ()| = 20 log dB;

(v) for > 0, the function 7 |]HC ()| achieves its maximum at m = (| | )1 ,
1
and this maximum value is m = arcsin +1 . Furthermore,

(a) if > 1, ]HC () has a maximum at m and the maximum value is between 0
and 90 , and
(b) if < 1, ]HC () has a minimum at m and the minimum value is between 90
and 0 ;

(vi) |HC (m )| = .
Proof (i) This is evident since lim0 HC () = 1.
(ii) In this case, the assertion follows since lim HC () =
signs of the real and imaginary parts gives the desired result.
(iii) This follows since lim0 HC () = 1.
(iv) This is evident since lim HC () = .

Keeping track of the

12.1 Compensation in the frequency domain

22/10/2004

471

(v) We compute
Re(HC ()) =

1 + 2 2
,
1 + 22

Im(HC ()) =

( 1)
,
1 + 22

so that

]HC () = ] ( 1) + i(1 + 2 2 ) .

In differentiating this with respect to , we need to take care of the various branches of
arctan. Let us first consider > 0 and > 1 so that
( 1)
]HC () = arctan
.
1 + 2 2

((1 + 2 2 2 2 ) 1)
d]HC ()
=
,
d
(1 + 2 2 )(1 + 2 2 2 )

1
which has a zero at = 2 . If > 0 and < 1, let us define =
]HC () = arctan

1 + 1 2 2

2.5

7.5

d]HC ()
((1 + ) )
=
,
d
(1 + 2 2 )( 2 + 2 2 )

.
2

which has a zero at =

The computations are the same for < 0, except that the
phase angle has the opposite sign. One may also check that the second derivative is not zero
1
, and so the function must have a maximum or minimum at this frequency. A
at = |
|
straightforward substitution gives
1
.
2

If > 1 this angle is positive, and if < 1 it is negative. That the value is bounded in
magnitude by 90 is a consequence of the maximum argument of 1 + i being 90 for
> 0 and the minimum argument of (1 + i )1 being 90 for > 0. A similar statement
holds for < 0. To get the final form for m given in the statement of the proposition, we
x
recall that if tan = x then sin = 1+x
2 . Taking x = tan m gives the result.
(vi) This is a simple matter of substituting the expression for m into HC ().

12.3 Remarks 1. One sees from part (v) of Proposition 12.2 why the names lead and lag
compensator are employed. For a lead compensator with , > 0 the phase angle
has a frequency window in which the phase angle is increased, and similarly for a lag
compensator with , > 0 there is a frequency window in which the phase is decreased.
2. A plot of the maximum phase lead m for > 1 is shown in Figure 12.1. When < 1
we can define = 1 and then see that
m = arcsin

1
.
+1

17.5

so that
Therefore, to determine the maximum phase lag when < 1, one can reads off the phase
angle from Figure 12.2 at 1 , and changes the sign of the resulting angle.
Note that the equation sin m = 1
can be solved for given m , and the result is
+1
=

12.5

or 1

Figure 12.1 Maximum phase shift versus for lead compensator

and versus 1 for lag compensator

Differentiating this with respect to gives

tan m =

yos

tos 0

,0
0

Differentiating this with respect to gives

(1 1 )

1N,D (t)
(t)
(t)
fj (t)
Re
Im
472 12 Ad hoc methods
II: Simple frequency response methods for controller design 22/10/2004
x1
x2
x1
x2 80
log
dB
deg 60
u = 0 ln
ln coth(|u| /2)

1 + sin m
.
1 sin m

(12.1)

3. Although in Proposition 12.2 we have stated the characterisation of lead/lag compensators in fairly general terms, typically one deals with the case when , > 0.
4. If one plots the Bode plot for a lead/lag compensator by hand using the methods of
1
.
Section 4.3.2, there are two break frequencies, one at 1 = 1 and another at 2 =
Depending on whether the compensator is lead or lag, one or the other of these frequencies
will be the smaller. In either case, the geometric mean of these two break frequencies is, by

definition, 1 2 which is equal to m . On a Bode plot for the lead/lag compensator we

therefore have log m = 12 (log 1 + log 2 ), showing that on the Bode plot, the maximum
phase shift occurs exactly between the two break frequencies.
5. The behaviour of the magnitude portion of the frequency
response at m is also nice on

a Bode plot. By (vi) the magnitude at m is 20 log = 10 log , and so is half of the
limiting value of the magnitude as .

Let us look at some simple examples of lead/lag compensators. We illustrate the various
things which can happen when choosing parameters in the compensator. But do keep in
mind that one normally uses a lead/lag compensator with > 0, and either 0 < < 1 (lag
compensator) or > 1 (lead compensator).
12.4 Examples 1. We look at a lead compensator with transfer function
RC (s) =

1 + 10s
,
1+s

meaning that = 1 and = 10. Based on Proposition 12.2 and Remark 12.34, it is
actually easy to guess what the Bode plot for this transfer function will look like. The

12.1 Compensation in the frequency domain

22/10/2004
PSfrag replacementstwo

473

1
break frequencies are 1 =PSfrag
1 adreplacements
2 = 10
. Thus the maximum phase lead will occur

,0
0

150
100
50

deg

u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

0
-50
-100
-150
-1.5 -1 -0.5

log

0.5

u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

1
1.5
2

150
50
0
-50

TD
1
RC (s) = K 1 +
+
+ TD s ,
TI
TI s
and this expression is genuinely different from the PID compensator we investigated in
Section 6.5. However, the alteration is not one of any substance (the difference is merely
a constant factor K TTDI ), and it turns out that the form (12.2) is well-suited to frequency
response methods.
The following result gives some of the essential features of the transfer function RC in
(12.2).
12.6 Proposition Let RC be a PID compensator of the form (12.2) and define HC () =
RC (i) for > 0. The following statements hold:
(i) lim0 |HC ()| = TKI ;

I
(v) the function 7 |HC ()| has a minimum of K TDT+T
at m =
I
(vi) ]HC (m ) = 0;

-100
-150
-1.5 -1 -0.5

0.5

log

,0
0

If we expand this form of the PID compensator we get

(ii) lim 1 |HC ()| = KTD ;

(iii) lim0 ]HC () = 90 ;
(iv) lim ]HC () = 90 ;

100

deg

1
1
1
t
t
magnitude
part of the Bode plot will start at
ke(t)kat log m = 2 (log 1 + log 10 ) = 2 . The
ke(t)k
e1 (t)0dB and turn up at about log = 1 until
e1 (t) it reaches log = 0 where it will level off at
e2 (t)
e2 (t)
e3 (t)20dB. Given that = 10, from Figure 12.1
e3 (t) we determine that the maximum phase lead
y(t)
y(t)

is
about
55
(the
calculation
gives

54.9 ). The actual Bode plot can be found in

m
u(t)
u(t)
h (t)Figure 12.2. One should note that the maximum
h (t)
phase shift does indeed occur between
hN,D (t)
hN,D (t)
1 (t)
1 (t)
20
0
1N,D (t)
1N,D (t)
(t)
(t)
(t)
(t)
15
-5
fj (t)
fj (t)
10
-10
Re
Re
Im
Im
x1
x1
5
-15
x2
x2
x1
x

1
0
-20
-1.5 -1 -0.5 0 0.5 1
1.5
2
-1.5 -1 -0.5 0 0.5 1
1.5
2
x2
x2
log
log

474 12 Ad hoc methods II: Simple frequency response methods for controller design 22/10/2004

1.5

TD TI

Proof The first four assertions are readily ascertained from the expression

Figure 12.2 Bode plots for a lead compensator RC (s) =

(left) and lag compensator RC (s) =

1
1+ 10
s
1+s

1+10s
1+s

TD
1
+ TD s
+
RC (s) = K 1 +
TI
TI s

(right)

for PID compensator. The final two assertions are easily derived from the decomposition
the two break frequencies, for example.
2. Here we take the lag compensator
RC (s) =

HC () =
1
1 + 10
s
,
1+s

of HC () into its real and imaginary parts, and then differentiating the magnitude to find
extrema.

1
so that = 1 and = 10
. Here on determines that the two break frequencies are 1 = 1
and 2 = 10. Thus log m = 12 (log 1 + log 10) = 12 . The magnitude portion of the Bode
plot will start at 0dB and turn down at log = 0 until it reaches log = 1 where it levels
off at 20dB. The maximum phase lag can be determined from Figure 12.1. Since < 1
we need to work with 1 which is 10 in this case. Thus, from the graph, we determine that
the maximum phase lag is about 55 . The actual Bode plot is shown in Figure 12.2.

12.1.2 PID compensation in the frequency domain We now look at how a PID
controller looks in the frequency domain. First we need to modify just a little the type of
PID controller which we introduced in Section 6.5.
12.5 Definition A PID compensator is a rational function of the form
RC (s) =
for K, TI , TD > 0.

K
(1 + TD s) s +
s

K
TD TI 2 1
(TD + TI ) + i
TI

1
TI

(12.2)

12.7 Remarks 1. The frequency m at which the magnitude of the frequency response
achieves its minimum is the geometric mean of the frequencies T1D and T1I . Thus on
a Bode plot, this minimum will occur at the midpoint of these two frequencies.
2. If one redefines a normalised frequency
= TI and a normalised derivative time TD =
TD
,
the
frequency
response
for
a
PID
compensator, as a function of the normalised
TI
frequency, is given by

TD
2 1
HC () = K TD + 1) + i
.

This identifies the normalised derivative time and the gain K as the essential parameters
in describing the behaviour of the frequency response of a PID compensator. Of these,
of course the dependence on the gain is straightforward, so it is really the normalised
derivative time TD = TTDI which is most relevant.

12.2 Design using controllers of predetermined form

475

476 12 Ad hoc methods II: Simple frequency response methods for controller design 22/10/2004

We have a good idea of what the Bode plot will look like for a PID compensator based
on the structure outline in Proposition 12.6. Lets look at two typical examples to see how
it looks in practice.

Let us begin, however, with a simple design situation where one wants to determine the
effects of adjusting a gain.

22/10/2004

12.8 Example In each of these examples, we take K = 1 since the effect of changing K is
merely reflected in a shift of the magnitude
Bode plot by log K.
PSfrag replacements
PSfrag replacements
tWe take TD = 1 and TI = 10. In this case, tProposition 12.6 tells us that the magnitude
1.
ke(t)k
ke(t)k 1
e1 (t)Bode plot will have its minimum at m
e1 (t)
= 10 or log m = 12 . The value of the
e2 (t)
e2 (t)
TD +TI
11
m , the phase will be zero. Below this
= 10 . At the frequency
e3 (t)minimum will be
e3 (t)
TI
y(t)
y(t)
u(t)frequency the phase is negative and above
u(t)it the phase is positive. Putting this together
h (t)gives the frequency response shown in Figure
h (t) 12.3.

,0
0

40
30
20
10
0

-2

-1

log

150
100
50
0
-50
-100
-150
-2

-1

log

u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

2
3
,0
0

r(s)

RC (s)

RP (s)

y(s)

Figure 12.4 Closed-loop system

80
70
60

we have chosen RC (perhaps it is a PID controller) and we wish to tune the gain K. To
determine whether a given gain K produces an IBIBO stable closed-loop system, we could,
for example, plot the Nyquist contour for the loop gain RL = KRC RP . However, if we have
to do this for very many gains, it might get a bit tedious. The following result relieves us of
this tedium.

hN,D (t)
1 (t)
1N,D (t)
(t)
(t)
fj (t)
Re
Im
x1
x2
x1
3
x2

50
40
30
20

-2

-1

log

150

12.9 Proposition For the closed-loop interconnection of Figure 12.4, define RL = RC RP , and
let R, r > 0 be selected so that the (R, r)-Nyquist contour is defined. Then, for K 6= 0, the
number of encirclements of 1 + i0 by the (R, r)-Nyquist contour for KRL is equal to the
number of encirclements of K1 + i0 by the (R, r)-Nyquist contour for RL .

100
50

deg

u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

deg

hN,D (t)
1 (t)
1N,D (t)
(t)
(t)
fj (t)
Re
Im
x1
x2
x1
x2

12.2.1 Using the Nyquist plot to choose a gain We work with the unity gain
feedback setup of Figure 6.21 which we reproduce here in Figure 12.4. In this scenario,

0
-50
-100
-150
-2

-1

log

Figure 12.3 The Bode plots for a PID compensator with K = 1,

TD = 1, and TI = 10 (left) and with K = 1, TD = 10, and
TI = 1 (right)

1
2. Now we take TD = 10 and TI = 1. Again we have m = 10 . The magnitude Bode
1
I
plot has its minimum at m = 10 and the value of the minimum is TDT+T
= 11. At
I

the frequency m the phase is 0 , and it is negative below this frequency and positive
above it. The Bode plot is shown in Figure 12.3.

12.2 Design using controllers of predetermined form

In this section we indicate how the procedures of the preceding sections can be helpful
in controller design. The basic idea of the frequency response methods we discuss here are
that one designs controllers which do various things to the systems frequency response, and
so alter both the Bode plot for the loop gain as well as the Nyquist plot. The technique of
using the frequency response to design a controller is called loop shaping since, as we shall
see, the idea is to shape the Nyquist plot to have a shape we like.
In Sections 12.2.2 and 12.2.3 we discuss ways to design controller rational functions which
shape the Nyquist contour, or equivalently the Bode plot, to have desired characteristics.

Proof A point s C is on the (R, r)-Nyquist contour for KRL if and only if K1 s is on the
(R, r)-Nyquist contour for RL . That is, the (R, r)-Nyquist contour for KRL is the (R, r)Nyquist contour for RL scaled by a factor of K. From this the result follows. If K < 0
the result still holds as the (R, r)-Nyquist contour for KRL is reflected about the imaginary
axis.

This result allows one to simply plot the Nyquist contour for RC RP , and then determine
stability for the closed-loop system with gain K merely by counting the encirclements of
K1 + i0. The following example illustrates this.
1
12.10 Example Let us look at the open-loop unstable system with RP (s) = s1
, and we wish
to stabilise this using a proportional controller, i.e., RC (s) = 1. Let us use the Nyquist
criterion to determine for which values of the gain the closed-loop system is stable. The
1
Nyquist contour for RL (s) = RC (s) = RP (s) = s1
is shown in Figure 12.5. Since the loop
gain has 1 pole in C+ , by Proposition 12.9, for IBIBO stability of the closed-loop system we
must have one counterclockwise encirclement of K1 + i0, provided K 6= 0. From Figure 12.5
we see that this will happen if and only if K > 1.
Of course, we can determine this directly also. The closed-loop transfer function is

TL (s) =

K
KRL (s)
=
.
1 + KRL (s)
s+K 1

The Routh/Hurwitz criterion (if you wish to hit this with an oversized hammer) says that
we have closed-loop IBIBO stability if and only if K > 1.

using controllers of predetermined form

477

-1

,0
0

478 12 Ad hoc methods II: Simple frequency response methods for controller design 22/10/2004
(ii) Design a phase lead controller rational function RC,1 so that the closed-loop system
will meet the phase and gain margin requirements.
(iii) Use phase lag compensation or integral control to boost the low frequency gain of the
controller transfer function. These terms will assist in the tracking of step inputs, and
the rejection of disturbances. Take care here not to degrade the stability of the system.
(iv) Should the plant not have sufficient attenuation at high frequencies, add a term which
rolls off at high frequencies to alleviate the systems susceptibility to high frequency
noise. Most plants will have sufficient attenuation at high frequencies, however.

22/10/2004

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)
h (t)
hN,D (t)12.2 Design
1 (t)
1N,D (t)
(t)
(t)
0.4
fj (t)
Re
Im
x1
0.2
x2
x1
x2
log
0
dB
deg
u = 0 ln
ln coth(|u| /2)
-0.2
or 1
m

yos
-0.4
tos 0

We can illustrate this design methodology with an example. Although this example is
illustrative, it is simply not possible to come up with a design methodology, or an example,
which will work in all cases.
-0.8

-0.6

-0.4

-0.2

Figure 12.5 Stabilising an open-loop unstable system using the

Nyquist criterion

12.12 Example We consider the problem of controlling a mass in space in the absence of
gravity. We again suppose that the motive force is applied through a propeller, as in Example 6.60. The governing differential equations are
m
y (t) = f (t) + d(t),

12.2.2 A design methodology using lead and integrator compensation In the previous section we assumed that we had in hand the nature of the controller we would be
using, and that the only matter to account for was the gain. In doing this, we glossed over
how one can choose the controller rational function RC . There are, in actuality, many ways
in which one can use PID control elements, in conjunction with lead/lag compensators, to
make a frequency response look like what we need it to look like. We shall explore just a
few design methodologies, and these can be seen as representative of how one can approach
the problem of controller design.
We look at a methodology which first employs a lead compensator to obtain a desired
phase margin for stability, and then combines this with an integrator for low frequency
performance and disturbance rejection. The rough idea is described in the following algorithm.
12.11 A design methodology Consider the closed-loop system of Figure 12.6. To design a
r(s)

RC (s)

RP (s)

y(s)

Figure 12.6 Closed-loop system for loop shaping

controller transfer function RC , proceed as follows.

(i) Select a gain crossover frequency for your system. This choice of frequency is based
upon requirements for transient response since larger gain crossover frequencies are
related to shorter rise times. One cannot expect to specify an arbitrary gain crossover
frequency, however. The bandwidth of a system will be limited by the bandwidth of
the system components.

where f is the input force from the propeller, and d(t) is a disturbance force on the system.
In the Laplace transform domain, the block diagram is as depicted in Figure 12.7. In this

d(s)

r(s)

RC (s)

f(s)

1
ms2

y(s)

Figure 12.7 Block diagram for controller design for mass

example we fix m = 1. Let us suppose that we have determined that a respectable value
for the gain crossover frequency is gc (in stating this, we are implicitly assuming that we
will have just one gain crossover frequency). For suitable stability we require that the phase
margin for the system be at least 40 . We also ask that the system have zero steady-state
error to a step disturbance.
Now we can go about ascertaining how to employ our design methodology to obtain the
specifications. First we look at the Bode plot for the plant transfer function RP (s) = ms1 2 .
This is shown in Figure 12.8. At the gain crossover frequency gc = 10rad/ sec (i.e., log gc =
1), the phase of the plant transfer function is 180 . Indeed, the phase of the plant transfer
function is 180 for all frequencies. To get the desired phase margin, we employ a lead
compensator. Indeed, we should choose a lead compensator which gives us the phase margin
we desire, and hopefully with some to spare. We need a minimum phase lead of 40 , and
40
from (12.1) we can see that this means that 1+sin
4.60. Let us be really conservative
1sin 40
and choose = 20. Now, we not only need for the phase shift to be at least 40 , we need it to
exceed this value at a specific frequency. But
Proposition 12.2 tells us how to manage this:
we need to specify m . We have m = (| | )1 , and solving this for with = 20 and

PSfrag replacements

form

479

100
50

h (t)
hN,D (t)
1 (t)
1N,D (t)
(t)
(t)
fj (t)

0
-50
-100
-150
-1.5 -1 -0.5

0.5

log

,0
0

1.5

x1
x2
20
x1
x2
log
0
dB
deg
u = 0 ln
ln coth(|u| /2) -20
or 1
m

yos -40

tos 0

80
60

Figure 12.9 we give the Nyquist and Bode

,0
0

-20

-10

,0
0

-40

-20

,0
60

-1.5 -1 -0.5

log

0.5

1.5

150
100
50
0
-50
-100
-150

log

0.5

1.5

Figure 12.10 The (100, 0.25)-Nyquist contour and the Bode plot
for mass and lead compensator with increased gain

60
40
20
0
-20
-40

-1.5 -1 -0.5

log

0.5

1.5

150
100
50

deg

u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

,0
0

10
x1
x2
x1
x2
log
0
dB
deg
u = 0 ln
ln coth(|u| /2)
-10
or 1
m

yos
-20
tos 0

where we are for the

u(t)
h (t)
moment.
In
hN,D (t)
1 (t)
1N,D (t)
(t)
(t)
fj (t)
Re
Im
x1
x2
x1
x2

y(t)
Lets
u(t) evaluate
h (t)
hN,D (t)
1 (t)
1N,D (t)
(t)
(t) 20
fj (t)

u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

20
0

deg

m = 10 gives = 2015 0.022. PSfrag

We have
been sufficiently conservative with our choice of
replacements
1
. With this specification,
that we can afford to make a nice number,t so lets take = 50
PSfrag replacements
ke(t)k
t lead compensator is
the
e
(t)
1
ke(t)k
e2 (t) 1 + 2s
e1 (t)
RC,1 (s)e3=(t)
e2 (t)
1 .
y(t) 1 + 50 s
e3 (t)

-20

Figure 12.8 Plant Bode plot for mass system

hN,D (t)
1 (t)
1N,D (t)
(t)
(t)
fj (t)
Re
Im
x1
x2
x1
x2

150

deg

u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

480 12 Ad hoc methods II: Simple frequency response methods for controller design 22/10/2004
and so the system is IBIBO stable. From the Bode plot, we see that the phase is about 105
(about 104.2 to be more precise) at gc , which gives a phase margin of about 75 , well
in excess of what we need. There is a problem at the moment, however, in that the desired
gain crossover frequency is, in fact, not a gain crossover frequency since the transfer function
has less magnitude than we desirePSfrag
at replacements
gc . We can correct this by boosting the gain of the
t
PSfrag replacements
lead compensator. The matter of just how
much to boost the gain is easily resolved. At
ke(t)k
t
e
(t)
1
ke(t)k
see that the magnitude is about 15dB (about
= gc , from the Bode plot of Figure 12.9 we
e2 (t)
e1 (t)
e3 (t)
14.1dB
to be more precise). So we need to
adjust K so that 20 log K = 15 or K 5.62.
e2 (t)
y(t)
e3 (t)
1
.
The
Nyquist
and
Bode
plots
for the gain boosted led compensator are
Let
us
take
K
=
5
u(t)
y(t)
2
h (t)
u(t)
provided
in Figure 12.10. Also note from the
Bode plot that the magnitude is falling off at

22/10/2004

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)
h (t)12.2 Design using controllers of predetermined
hN,D (t)
1 (t)
1N,D (t)
75
(t)
50
(t)
25
fj (t)
0
Re
Im
-25
x1
-50
x2
x1
-75
-1.5 -1 -0.5 0
0.5
1
1.5
2
x2
log

0
-50
-100

40dB/decade, which means that the system is type 2 by Proposition 8.14. Therefore this
system meets our objective to track step inputs. In Figure 12.11 we plot the response of the
closed-loop system to a unit step input which we obtained by using Proposition 3.40. Note
that the error decays to zero as predicted.
Lets now look at how the system handles step disturbances since we have asked that these
be handled gracefully (specifically, we want no error to step disturbances). In Figure 12.12
we display the response of the system to a unit step disturbance (no input). Clearly this is
not satisfactory, given our design specifications. To see what is going on here, we determine
the transfer function from the disturbance to the output to be

-150
-1.5 -1 -0.5

log

0.5

1.5

Td (s) =

RP (s)
.
1 + RC,1 (s)RP (s)

We compute
Figure 12.9 The (100, 0.25)-Nyquist contour and the Bode plot for
mass and lead compensator

plots for RC1 RP . From the Nyquist plot we see that there are no encirclements of 1 + i0,

lim Td (s) =

lims0

1
RP (s)

1
1
=
= 1,
lims0 RC,1 (s)
+ lims0 RC,1 (s)

and so the system has disturbance type 0 with respect to this disturbance. This is not
satisfactory for the purpose of eliminating error resulting from a step input, so we need to

PSfrag replacements

repair this in some way. As we have seen in the past, a good way to do this is to add an
integrator to the controller transfer function. Lets see how this works. We work with the
new controller rational function
RC,2 (s) =

KI
1 + 2s
+
1 .
s
1 + 10
s

First note that lims0 RC,2 (s) = and so this immediately means that the system type with
respect to the disturbance is at least 1 (and is in fact exactly 1). We look to ascertain how
changing the integrator gain KI affects system stability. In Figure 12.13 the Bode plots are
shown for various values of KI . We can see that, as expected since an integrator will decrease
the phase, the phase margin becomes worse as KI increases. Whats more, one can check the
Nyquist criterion to show that the system is in fact unstable for K sufficiently large. Thus
1
we choose a not too large integrator gain of KI = 10
. Lets see how this choice of controller
performs. First, in Figure 12.14 we display the step response (no disturbance) of the system
with the integrator added to the controller. Note that the overshoot and the settling time

100
50
0
-50
-100
-150

0.5

log

PSfrag replacements

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)
h (t) 0
-1.5 -1 -0.5
hN,Dlog
(t)
1 (t)
1N,D (t)
(t)
(t)
fj (t)
Re
Im
x1
x2
x1
x2

0.5

150
125
100
75
50
25
0
-25

150

u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

1
1.5
2

100
50
0
-50
-100
-150
-1.5 -1 -0.5

log

,0
0

0.5

1.5

,0
0

150

-1.5 -1 -0.5

deg

u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)
frequency hresponse
methods for controller design 22/10/2004
(t)
hN,D (t)
1 (t)
1N,D (t)
125
(t)
100
(t)
fj (t)
75
Re
50
Im
25
x1
0
x2
x1
-25
1
1.5
2
-1.5 -1 -0.5 0
0.5
1
1.5
2
x2
log
dB

hoc methods II: Simple

u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

,0
0

-1.5 -1 -0.5

0.5

1.5

-1.5 -1 -0.5

0.5

1.5

log

150
100
50

deg

1 (t)

Figure 12.12 Step disturbance response of mass with lead compensator

PSfrag replacements

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)
12 Ad
h482
(t)
hN,D (t)
1 (t)
100
1N,D (t)
(t)
80
(t)
60
fj (t)
40
Re
20
Im
x1
0
x2
-20
x1
x2

deg

481

1 (t)

1N,D (t)
(t)
(t)
fj (t)
Re
Im
22/10/2004
12.2 Design using controllers of predetermined form
x1
x2
PSfrag replacements
x1
1
x2
ke(t)k
log
e1dB
(t) 0.8
e2deg
(t)
u = 0eln
0.6
3 (t)
ln coth(|u|y(t)
/2)
or u(t)
1
0.4
h(t)
m

hN,D (t)
yos 0.2

1N,D (t)
t(t)
os 0
0

,
(t)
2
4
6
8
10
12
14
0
fj(t)
t
0
Re
Im
x1 Figure 12.11 Step response of mass with lead compensator
x2
x1
x2 0.175
log
dB 0.15
deg 0.125
u = 0 ln
ln coth(|u| /2)
0.1
or 1
0.075
m

0.05
yos
0.025
tos 0
0

2
4
6
8
10
12
14
,0
t
0

0
-50
-100
-150

log

Figure 12.13 Bode plots for mass with lead and integrator com1
(top left), KI = 10 (top right), and
pensation: KI = 10
KI = 100 (bottom)

have increased from the situation when we simply employed the lead compensator, but may
be considered acceptable. If they are not, then one should iterate the design methodology
until satisfactory performance is achieved. Also in Figure 12.14 we display the response of
the system to a step disturbance (no input). Note that this response ends up at zero as we
have ensured by designing the controller as we have done. The decay of the effect of the
error is quite slow, however. One may wish to improve this by increasing the integrator gain.
Now that we have added the integrator to reject disturbances, we need to ensure that
the other performance specifications are still met. A look at Figure 12.13 shows that our
gain crossover frequency has decreased. We may wish to repair this by further boosting the
gain on the lead compensator.

This example, even though simple, shows some of the tradeoffs which must take place in

,0
0

PSfrag replacements

12.2 Design using controllers of predetermined form

483

1
0.8

0.6

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
484 12 Ad hoc y(t)
methods II: Simple frequency response methods for controller design 22/10/2004
u(t)
hfor
(t) this transfer function is reproduced in Figure 12.15. We suppose that we
The Bode plot
hN,D (t)
1 (t)
1N,D (t)
75
(t)
50
(t)
25
fj (t)
0
Re
Im
-25
x1
-50
x2
x1
-75
-1.5 -1 -0.5 0
0.5
1
1.5
2
x2
log

0.4
0.2
0

0.15

u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

0.125
0.1
0.075
0.05
0.025
0

100

,0
0

150
100
50

deg

1 (t)

1N,D (t)
(t)
(t)
PSfrag replacements
fj (t)
Re
ke(t)k
Im
22/10/2004
e1 (t)
x1
e2 (t)
x2
e3 (t)
x1
y(t)
x2
u(t)
log

hdB
(t)
hN,Ddeg
(t)
u = 0 ln
ln coth(|u|
1N,D/2)
(t)
or (t)
1
(t)
m

fj (t)
yos
Re

Im
tos
x01
x02
,
1
x
0
x2
log
dB
deg
u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

0
-50
-100
-150
-1.5 -1 -0.5

log

0.5

1.5

Figure 12.15 Plant Bode plot for mass system

Figure 12.14 Step response of mass (top) and response to step
disturbance (bottom) with lead and integrator compensation

designing a controller to meet sometimes conflicting performance requirements.

12.2.3 A design methodology using PID compensation Lets see in this section
how one can employ a PID controller in a systematic way in the frequency domain to obtain
a desired response. The rough idea of PID design is outlined as follows.
12.13 A design methodology Consider the closed-loop system of Figure 12.6. To design a
controller transfer function RC , proceed as follows.
(i) Select a gain crossover frequency consistent with the demands for quick response and
the limitations of the system components.
(ii) Commit to TI being quite a bit larger than TD .
(iii) Adjust TD so that the phase margin requirements are met.
(iv) Adjust K so that the gain crossover frequency is as desired.

Let us illustrate this in an example.
12.14 Example (Example 12.12 contd) We again take the plant transfer function
RP (s) =

1
.
ms2

are given the criterion of having a phase margin of at least 65 at as high a gain crossover
frequency as possible.
Let us decide to go with TTDI = 51 . In order to achieve the prescribed phase margin,
we must use the PID compensator to produce a phase for the loop gain RC RP which is
somewhere negative and greater than 115 . Note that given the relative degree of the
plant and the PID controller, the phase for large frequencies will approach 90 . The

1
contribution to the phase made by the controller changes sign at m = TD TI . Given
TD
1
1
1
our choice of TI = 5 , this means that m = 5 TD . Thus choosing a large derivative time
will decrease the frequency at which the phase becomes negative. In Figure 12.16 we show
the situation for TD = 15 and TD = 2. For the smaller value of TD , we see that the phase
margin requirements are met at a quite high frequency (around 20rad/ sec). However, a
peek at the Nyquist contour for TD = 51 shows that there are 2 clockwise encirclements of
1 + i0. Since RC RP has no poles in C+ , this means the system is not IBIBO stable for the
given PID parameters. When the derivative time is TD = 2, the phase requirement is met
at roughly 2rad/ sec, but now the Nyquist contour is predicting IBIBO stability. Thus we
see that we should expect there to be a tradeoff between stability and performance in this
design methodology. Let us fix TD = 1 (and hence TI = 5), for which we plot the Bode plot
and Nyquist contour in Figure 12.17.
We now choose the gain K in order to make the gain crossover frequency large. From
Figure 12.17 we see that when the phase is 115 the frequency is roughly given by log =
0.5, so we take gc = 100.5 3.16. At this frequency the magnitude of the frequency response
is about 10dB. Thus we need to choose K so that 20 log K = 10, or K 3.16. Lets make

PSfrag replacements

-100
-150
-1.5 -1 -0.5

log

0.5

50
0
-50
-100
-150
-1.5 -1 -0.5

log

0.5

1.5

x1
10
x2
x1
x2
0
log
dB
deg
u = 0 ln -10
ln coth(|u| /2)
or 1
m -20

yos

tos 0 -30

-50

-25

100,0

-30

-20

-10

Figure 12.16 Effect of varying PID derivative time for RP = s12 :

TD = 15 (left) and TD = 2 (right). In both cases K = 1 and
TI = 5TD , and Nyquist plots are on top and Bode plots are on
the bottom

3 14
(1 + TD s) s +
s

u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0
PSfrag replacements

,0t
ke(t)k
0
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)
h (t)
hN,D (t)
1 (t)
1N,D (t)
(t)
(t)
fj (t)
Re
Im
x1
x2
x1
x2

u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

,0
0

methods for controller design 22/10/2004

log

0.5

100
50
0
-50
-100
-150
-1.5 -1 -0.5

log

0.5

x1
x2
x1
1
1.5
2
x2
log
dB
deg
u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0
1
1.5
2

PSfrag replacements

-1.5 -1 -0.5

150

100
80
60
40
20
0
-20

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)
frequency response
h (t)
hN,D (t)
1 (t)
1N,D (t)
30
(t)
(t)
fj (t)

-10

-20

-30
-20

,0
0 t

-10

ke(t)k
e1 (t)
e2 (t)
e3 (t)and Bode plot
Figure 12.17 Nyquist plot (left)
y(t)
with PID controller and K u(t)
=
1, TD = 1, and
h (t)
hN,D (t)
1 (t)
1N,D (t)
100
(t)
(t)
fj (t)

(right) for mass

TI = 5

-1.5 -1 -0.5

log

0.5

150
100
50
0
-50
-100
-150
-1.5 -1 -0.5

log

0.5

x1
x2
x
1
1.5
2
x2
log
dB
deg
u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0
1.5
2

,0
0

-50

-100
-50

100

Figure 12.18 Nyquist plot (left) and Bode plot (right) for mass
with PID controller and K = 3 14 , TD = 1, and TI = 5

this K = 3 14 so that our final controller is given by

RC (s) =

100
80
60
40
20
0
-20
-40

,0
0

1.5

100

deg

0
-50

x1
25
x2
x1
x2
0
log
dB
deg
u = 0 ln -25
ln coth(|u| /2)
or 1
m -50

yos
-75
tos 0

u = ln
ln coth(|u| /2)
or
1t
ke(t)k
m
e1(t)

e2 (t)
os
e3y(t)

y(t)
tos
0
u(t)

1
1.5
2h (t)
,0
hN,D(t)
0
1 (t)
1N,D (t)
(t)
(t)
fj (t)

150

0
PSfrag replacements

Ad hoc methods II: Simple

deg

100

deg

ln coth(|u| /2)
t
or
1
ke(t)k
m
e1(t)

e2 (t)
e3y(t)
os

y(t)
tos
0
u(t)

h (t)
0
hN,D,
(t)
0
1 (t)
1N,D (t)
(t)
(t)
fj (t)

PSfrag replacements

150

= 0 ln
PSfrag u
replacements

485

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)
12
h486
(t)
hN,D (t)
1 (t)
1N,D (t)
(t)
(t)
fj (t)
Re
Im
x1
x2
x1
x2

PSfrag replacements

t
t
ke(t)k
ke(t)k
e1 (t)
e1 (t)
e2 (t)
e2 (t)
e3 (t)
e3 (t)
y(t)
y(t)
u(t)
u(t)
12.2 Design using controllers
of predetermined form
h22/10/2004
h (t)
(t)
hN,D (t)
hN,D (t)
1 (t)
1 (t)
100
1N,D (t)
1N,D (t)
100
(t)
(t)
80
(t)
(t)
75
60
fj (t)
fj (t)
50
40
Re
Re
25
20
Im
Im
0
x1
x1
0
-25
x2
x2
-20
x1
x1
-50
-1.5 -1 -0.5 0
0.5 1
1.5
2
-1.5 -1 -0.5 0
0.5
x2
x2
log
log

deg

PSfrag replacements

1
TI

The Bode and Nyquist plots for the corresponding loop gain are shown in Figure 12.18.
From the Nyquist plot we see that the system is IBIBO stable, and we also see that out
phase margin requirements are satisfied.
In Figure 12.19 is the response of the system to a unit step input. We may also compute
the effect of a disturbance entering between the controller and the plant. The corresponding

transfer function is
Td (s) =

RP (s)
s
= 3
9
1 + RC (s)RP (s)
s + 3 41 s2 + 3 10
s+

and the response to a unit step is shown in Figure 12.20.

13
20

12.2.4 A discussion of design methodologies The above examples give an idea of

how one can use ideas of frequency response in designing controller rational functions. In-

12.2 Design using controllers of predetermined form

In the previous section, the emphasis was on tuning controllers of a specified type. This
can be a difficult exercise for plants which are unstable and/or nonminimum phase (see
Exercise E12.6). The difficulty is that be fixing the form on the controller, one also limits
what can be done to the sensitivity function and the closed-loop transfer function. It may
be possible that the plant will not allow stabilisation by a controller of a certain form.

0.8
0.6
0.4

12.4 Summary

0.2
0

Figure 12.19 Step response of mass with PID controller

0.2
0.15
0.1
0.05
0

488 12 Ad hoc methods II: Simple frequency response methods for controller design 22/10/2004

12.3 Design with open controller form

1 (t)

,0
0

487

1.2

1 (t)

1N,D (t)
(t)
(t)
fj (t)
Re
Im
22/10/2004
x1
x2
x1
PSfrag replacements
x2
log
ke(t)k
e1dB
(t)
e2deg
(t)
u = 0eln

3 (t)
ln coth(|u|y(t)
/2)
or u(t)
1
h(t)
m

hN,D (t)
yos

1N,D (t)
t(t)
os 0

,
(t)
0
fj(t)
0
Re
Im
x1
x2
x1
x2
log
dB
deg
u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

Figure 12.20 Unit step disturbance of mass with PID controller

deed, it is interesting to compare the examples since they use the same plant with different
control design strategies. For example, the one thing we notice straight away is the better
disturbance response for the PID compensator. This should come as no surprise since the
reset time for the integrator is significantly larger for the PID compensator.
We should emphasise that the process is rarely as straightforward as the examples suggest,
and that in practice one will usually have to iterate the design process to account for the
various tradeoffs which exist between stability and performance. Also, the methodologies we
discuss above can only be expected to have a reasonable chance at success when the plant
transfer function RP has no poles or zeros in C+ . If RP does have poles in C+ , then one
must design the controller RC so that the point 1 + i0 is encircled. If RP has zeros in C+ ,
then it typically turns out to be more difficult to design a stabilising controller that meets
goals for stability margins. In such cases, mundane considerations of gain and phase margin
become less satisfactory measures of a good design unless weighed against other factors.

In this section, we have been able to assimilate our knowledge gained to this point to
generate some design methods for controller rational functions. One should note that our
methods are not guaranteed to work, but for simple applications, they provide a starting
point for serious controller design. Let us review some of the basic ideas.
1. The basis for the frequency response design methodology is the Nyquist criterion. This
can be a somewhat subtle notion, so it would be best to understand it. A good way to
do this is to study the proof of the Nyquist criterion since it is quite simple, given the
Principle of the Argument.
2. We have presented two design methodologies using frequency response: one for lead
compensation and the other for PID compensation. One should make sure to understand
the idea behind these design methodologies.

Exercises for Chapter 12

489

490 12 Ad hoc methods II: Simple frequency response methods for controller design 22/10/2004

Exercises
C

The next three exercises have to do with designing circuits to implement various controllers.
Although we look at only a three specific controller transfer functions, it is possible, in
principle, to design a circuit using only passive resistors, capacitors, and inductors, to realise
any transfer function in RH+
. A means of doing this was first pointed out in the famous
paper of Bott and Duffin [1949].

R
R
2

E12.1 Consider the circuit of Figure E12.1.

+
V2

Figure E12.3 Circuit for notch filter

C
+

Figure E12.1 Circuit for lead compensation

+
V1

(a) Determine the differential equation governing the output voltage V2 given the
input voltage V1 .
(b) Convert this differential equation to a transfer function, and show that the
resulting transfer function is that of a notch filter.
(c) Determine expressions for K, , and in the standard form for a lag compensator in terms of Rand C.
2
E12.4 Consider the plant transfer function RP (s) = s2 (s+2)
. Design a PID controller using
frequency domain methods which produces an IBIBO stable system with a phase
margin of at least 65 at as large a gain crossover frequency as possible. Check the
stability of your design using the Nyquist criterion, and produce the step response,
and the response to a step disturbance which enters the loop between the controller
and the plant.

s
E12.5 Consider a controller transfer function RC (s) = K 1+
+ T1I s where K, , , and
1+ s
TI are all finite and nonzero. Determine a SISO linear system = (A, b, ct , D) so
that T = RC .
E12.6 In this problem you will design a controller for the unstable, nonminimum phase
plant shown in Figure E12.4.
r(s)

RC (s)

2s
s2 + s 3

y(s)

Figure E12.2 Circuit for lag compensation

Figure E12.4 A closed-loop system with an unstable, nonminimum

phase plant

(a) Why is the plant unstable? nonminimum phase?

(b) Produce the Nyquist and Bode plots for the plant transfer function. Is the
closed-loop system stable with RC (s) = 1?
(c) Is it possible to design a proportional controller RC (s) = K which renders the
closed-loop system IBIBO stable?
(d) We will design a stabilising controller by manipulating the Nyquist contour to
look like the cartoon in Figure E12.5. Sketch the Bode plot for such a Nyquist

Exercises for Chapter 12

491

492 12 Ad hoc methods II: Simple frequency response methods for controller design 22/10/2004
and the lead/lag controller transfer function
RC (s) =

K(1 + s)
.
1 + s

Answer the following two questions.

(a) Show that it is not possible to design a lead/lag controller RC for which RC
S (RP ).
(b) Show that for any proper, second-order plant transfer function RP R(s) there
exists a lead/lag controller RC for which RC S (RP ).
E12.9 Consider the plant transfer function
RP (s) =
Figure E12.5 The desired Nyquist contour

and the PID controller transfer function

1
.
RC (s) = K 1 + TD s +
TI s

contour, being particularly concerned with the phase of the Bode plot. Assume that the loop gain RL = RC RP is strictly proper and that its maximum
magnitude occurs at = 0.
(e) Design a phase lead controller RC with the property that the Bode plot for
RL = RC RP has the Bode plot phase which looks qualitatively like that you
drew in part (d). For the moment, design RC to that RC (0) = 1. Plot the
Nyquist contour for RL to verify that it has the shape shown in Figure E12.5.
(f) With the lead compensator you have designed, is your closed-loop system IBIBO
stable? By adding a gain K to the lead compensator from part (e), it can be
ensured that the system will be IBIBO stable for some values of K. Using your
Nyquist contour from part (e), for which values of K will the closed-loop system
be IBIBO stable? Pick one such value of K and produce the Nyquist contour to
verify that the controller you have designed does indeed render the closed-loop
system IBIBO stable.
Congratulations, you have just designed a controller for an unstable, nonminimum
phase plant, albeit a contrived one. Let us see how good this controller is.
(g) Comment on the gain and phase margins for your system.
(h) Produce the step response for the closed-loop system, and comment on its performance.
E12.7 In this exercise you will be given a plant transfer function for which it is not possible
to achieve arbitrary design objectives. The transfer function is
P (s) s z
RP (s) = R
sp
for s, p R positive. Thus, we know that our plant has an unstable real pole p, and
a nonminimum phase real zero z. All other poles and zeros are assumed to be stable
and minimum phase, respectively. Let RC be a stable, minimum phase controller
transfer function and define RL = RC RP .
(a) Define a
E12.8 Consider the plant transfer function
1
RP (s) = 3
,
s s2 + s 1

1
,
s4 s3 + s2 s + 1

Answer the following two questions.

(a) Show that it is not possible to design a PID controller RC for which RC
S (RP ).
(b) Show that for any proper, third-order plant transfer function RP R(s) there
exists a PID controller RC for which RC S (RP ).
E12.10 Consider the coupled tank system of Exercises E1.11, E2.6, and E3.17. Take as
1
1
system parameters = 13 , 1 = 1, A1 = 1, A2 = 21 , a1 = 10
, a2 = 20
, and g = 9.81.
As output, take the level in tank 2.
(a) Design a PID controller for the system that achieves a phase margin of at least
75 with as large a gain crossover frequency as possible.
We now wish to simulate the behaviour of the system with the controller that you
have designed. The states of the system are nominally h1 and h2 . However, since the
controller involves an integration, an additional state will need to be defined. With
this in mind, answer the following question.
(b) Suppose that a reference output h2,ref has been specified, and that the control
u is specified by the PID controller you designed in part (a). Develop the
nonlinear state differential equations for the system, starting with the equations
you linearised in Exercise E2.6. As indicated in the preamble to this part of the
problem, you will have three state equations.
(c) With the controller that you have designed, simulate the differential equations
from part (b) with initial conditions equal to the equilibrium initial conditions,
and subject to a reference step input of size 41 .
(d) Plot the height in both tanks, and comment on the behaviour of your controller.

Finish this

Exercises for Chapter 12

493

494

This version: 22/10/2004

13 Advanced synthesis, including PID synthesis

22/10/2004

the integral of the absolute value of the error due to a unit step input. Thus one minimises
Z
|e(t)| dt,

Chapter 13
Advanced synthesis, including PID synthesis
We saw in Chapters 11 and 12 that one can use ad hoc methods to choose PID parameters
that can serve as acceptable starting points for final designs of such controllers. These
classical methods, while valuable in terms of providing some insight into the process of
control design, can often be surpassed in effectiveness by more modern methods. In this
chapter we survey some of these, noting that they rely on some of the more sophisticated
ideas in the text to this point. This explains why they may not form a part of the typical
introductory text dealing with PID control.

Contents

where e(t) is the difference between the step response and the desired response to a step
input. In particular, it is assumed that this integral is finite. Since the methods in this
section are ad hoc, they should not be thought of as being guarantees, but rather as a good
starting point for beginning a final tuning of the parameters. A methodology for this is the

subject of [Hang, Astrom, and Ho 1990].

13.1.1 First method We shall work with systems in input/output form. Thus let RP
be a plant transfer function with c.f.r. (N, D), and let 1N,D (t) be the step response for the
system. We assume that (N, D) is BIBO stable. Define a parameter R+ by

= sup 1 N,D (t) .
t0

Thus is the maximum

slope of the step response. Let t R be the smallest time

satisfying 1 N,D (t) = . Thus t is the time at which the slope of the step response reaches
its maximum value. Typically this time is unique. We then define R by

13.1 Ziegler-Nichols tuning for PID controllers . . . . . . . . . . . . . . . . . . . . . . . . . . . 493

= t

13.1.1 First method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 494

13.1.2 Second method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 494
13.1.3 An application of Ziegler-Nicols tuning . . . . . . . . . . . . . . . . . . . . . . . . 496

The meaning of is as shown in Figure 13.1. With the parameters and at hand, we can
specify the parameters in a P, PI, or PID control law of the form

13.2 Synthesis using pole placement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 500

RC (s) = K

(P)
1
RC (s) = K 1 +
(PI)
TI s

1
(PID).
RC (s) = K 1 + TD s +
TI s

13.2.1 Pole placement using polynomials . . . . . . . . . . . . . . . . . . . . . . . . . . . 500

13.2.2 Enforcing design considerations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 507

13.2.3 Achievable poles using PID control . . . . . . . . . . . . . . . . . . . . . . . . . . 513
13.3 Two controller configurations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 516
13.3.1 Implementable transfer functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 517

1N,D (t )
.

(13.1)

In Table 13.1 we tabulate choices of the parameter values for various types of controllers.

13.3.2 Implementations that meet design considerations . . . . . . . . . . . . . . . . . . . 521

13.4 Synthesis using controller parameterisation . . . . . . . . . . . . . . . . . . . . . . . . . . 521
13.4.1 Properties of the Youla parameterisation . . . . . . . . . . . . . . . . . . . . . . . 521
13.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 522

13.1 Ziegler-Nichols tuning for PID controllers

The ideas we discuss here are the result of an empirical investigation by Ziegler and Nicols
[1942]. We give two methods for specifying PID parameters. The first will be applicable quite
often, especially for BIBO stable plants, whereas the second makes some assumptions about
the nature of the system. In each case, the criterion for optimisation was the minimisation of

Table 13.1 Controller parameters for the first Ziegler-Nicols tuning method

Controller type

Controller parameters

PID

9
10
, TI = 3
10
6
,
T
=
2
,
I
5

TD =

13.1.2 Second method We resume the setting above with a plant RP with c.f.r.
(N, D). In this method, we make the following assumption.

13.1 Ziegler-Nichols tuning for PID controllers

22/10/2004

495

496

13 Advanced synthesis, including PID synthesis

22/10/2004

Note that there are some simple cases in which the Ziegler-Nicols criterion will not apply
(see Exercise E13.1). However, for cases where the method does apply, it can be a useful
starting point. It also has the advantage that it can be applied to an experimentally obtained
step
response.
PSfrag replacements

1N,D (t)

t
1

13.1 Assumption The closed-loop transfer function with proportional control,

T =

KRP
,
1 + KRP

is BIBO stable for K very near zero. Furthermore, if the gain K is increased from K = 0,
there exists a critical gain Ku where exactly one pair of the poles of the transfer function
crosses the imaginary axis, with the remaining poles in C .

Under the conditions of the assumption, at the gain Ku the step response will exhibit an
oscillatory behaviour for sufficiently large times. If the poles on the imaginary axis are at
iu , the period of this oscillation will be Tu = 2u . With this information, the criterion for
choosing the parameters in the controller (13.1) are as given in Table 13.2.
Table 13.2 Controller parameters for the second Ziegler-Nicols
tuning method

Controller type Controller parameters

PID

Ku
2
9Ku
,
20
6Ku
,
10

TI = 56 Tu
TI =

Tu
,
2

TD =

Tu
8

1N,D (t)

Figure 13.1 The definitions of and for Ziegler-Nicols PID tuning

13.1.3 ke(t)k
An application of Ziegler-Nicols tuning Let us apply the Ziegler-Nicols tuning methods eto1 (t)
an example. Suppose that we have a rotor spinning on a shaft supported by
(t)
bearings. Thee2angular
position of the rotor will satisfy a differential equation of the form
e3 (t)
y(t)
J + d + k = u,
u(t)
h (t)
where J ishthe
inertia of the rotor, d accounts for the viscous friction in the bearings, k is
N,D (t)
the shaft spring
1 (t)constant, and u is the torque applied to the shaft. This then gives a plant
transfer function
1
(t)
RP (s) = 2
.
(t)
Js + ds + k
fj (t)
1
Let us take J = 1, d = 10
, and k = 2.
Re
Let us lookIm
at the first Ziegler-Nicols method. The step response for the plant is shown
in Figure 13.2. xOne
may compute and graphically. However, in this case it is possible to
1
x2
x1
1
x2
log
0.8
dB
deg
u = 0 ln 0.6
ln coth(|u| /2)
or 1
m 0.4

yos 0.2

tos 0

2.5
5
7.5
10 12.5 15 17.5 20
,0
t
0
Figure 13.2 Step response for rotor on shaft

compute these numerically since the step response is a known function on t. To find t one
determines where 1N,D (t ) = 0, where (N, D) is the c.f.r. for RP . We compute t 1.08 and
then easily compute 0.70 and 0.40. The values for the P, PI, or PID parameters
are shown in Table 13.3. The three corresponding step responses for the closed-loop transfer
function, normalised so that they have the same steady state value, are shown in Figure 13.3.
Notice that the closed-loop performance is actually rather abysmal. Furthermore, it is quite
evident that what is needed in more derivative time. If we arbitrarily set TD = 1 in the
PID controller, the resulting step response is shown in Figure 13.4. Note that the response
is now more settled. This exercise points out that the Ziegler-Nicols tuning method does
not produce guaranteed effective control laws. Indeed, the system we have utilised is quite
benign, and it still needed some adjustment to work well.

13.1 Ziegler-Nichols
PSfrag replacements tuning for PID controllers

PI
PSfrag replacements

PID

ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)
h (t)
hN,D (t)

parameters

(t)
(t)
K 3.45,
TI
fj (t)
Re
K 4.60,
TI
Im
x1
x2
x1 1.4
x2
log 1.2
dB
1
deg
u = 0 ln
ln coth(|u| /2) 0.8
or 1 0.6
m
0.4
yos
0.2
tos 0
0

8.55
= 0.78, TD 0.19

,0
0

1N,D (t)
(t)
(t)
fj (t)
Re
Im
x1
2.5
5
7.5 x 10 12.5 15 17.5 20,0
2
0
x1 t
1.5
x2
log
dB 1.25
deg
1
u = 0 ln
ln coth(|u| /2)
or 1 0.75
m
0.5

yos
0.25
tos 0
0

2.5
5
7.5
10
,0
t
0

0.8
0.6
0.4
0.2
0

2.5

7.5

12.5

17.5

PSfrag replacements

t
ke(t)k

e1 (t)
Figure 13.4 Normalised
step response for rotor example using first
e2 (t)
e3 (t) tuning method and derivative time adjusted to
Ziegler-Nicols
y(t)
TD = 1 u(t)

2.5

7.5

12.5

17.5

h (t)
hN,D (t)
1 (t)
1N,D (t)
(t)
(t)
fj (t)

PSfrag replacements

17.5

Figure 13.3 Normalised step responses for rotor example using

first Ziegler-Nicols tuning method: P (top left), PI (top right),
and PID (bottom)

Let us now move onto the second of the Ziegler-Nicols methods. We cannot use the
rotor example, because it does not satisfy Assumption 13.1. So let us come up with a plant
transfer function that does work. An example is
RP (s) =

12.5

22/10/2004

1
.
s3 + 3s2 + 4s + 1

In Figure 13.5 is a plot of the behaviour of the poles of the closed-loop system as a function
of K with RC (s) = K. As we can see, the roots behave as specified in Assumption 13.1, so
we can proceed with that design methodology. The method asks that we find the critical
gain Ku for which the roots cross the imaginary axis. One can do this by trial and error,
looking at the step response. For example, in Figure 13.6 we plot the step response for two
values of K. As we can see, for the plot on the left K < Ku and for the plot on the right
K > Ku . One can imagine iteratively finding something quite close to Ku by looking at such
plots. However, I found Ku by numerically determining when the real part of the poles for

ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)
h (t)
hN,D (t)
1N,D (t)
(t)
(t)
fj (t)
Re
Im
x1
x2
x1
x2
1.5
log
dB 1.25
deg
u = 0 ln
1
ln coth(|u| /2)
or 1 0.75
m
0.5

yos
0.25
tos 0
0

,0
0

x1
x2
x1
x2
0 PSfrag replacements
X
log
dB
deg
ke(t)k
u = 0 ln
e1 (t)
ln coth(|u| /2)
e2 (t)
or 1 -1
e3 (t)
X
m
y(t)

u(t)
yos
h (t)

tos 0 -2
hN,D (t)

,0
-2.5 -2 -1.5 -1 -0.5
Re
0
1N,D (t)

(t)
(t)
fj (t)
Figure 13.5 Behaviour of poles for plant
Re
RC (s) = K as K varies
Im
x1
x2
x1
2.5
x2
log
2
dB
deg 1.5
u = 0 ln
ln coth(|u| /2)
1
or 1
m
0.5

yos
0

tos 0 -0.5

2.5
5
7.5
10 12.5 15 17.5 20,0
t
0

0.5

RP (s) =

1
s3 +3s2 +4s+1

and

1 (t)

1N,D
(t)
K
3.83

13 Advanced synthesis, including PID synthesis

1 (t)

,0
0

Controller type

u(t)
h (t)
hN,D (t)
Controller

1 (t)

1N,D (t)
(t)
(t)
fj (t)
Re
Im
x1
x2
x1
x2
log
dB 1.5
deg
u = 0 ln
ln coth(|u| /2)
1
or 1
m

yos 0.5

tos 0
0

ke(t)k
e1 (t)

Table 13.3 Controller parameters

e2 (t) for the rotor example first
e3 (t)
Ziegler-Nicols tuning method
y(t)

1 (t)

ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)
h (t)
hN,D (t)

497

22/10/2004
PSfrag replacements

1N,D (t)
(t)
(t)
fj (t)
Re
Im
498
x1
x2
x1
x2
log
dB
deg
u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

2.5

7.5

12.5

Figure 13.6 The step response of the plant RP (s) = s3 +3s21+4s+1

and RC (s) = K for K = 10 (left) and K = 12 (right)

17.5

13.1 Ziegler-Nichols tuning for PID controllers

22/10/2004

499

1
KRP (s)
= 3
1 + KRP (s)
s + 3s2 + 4s + 1 + K

are zero. The answer is approximately Ku 11.0. With this value of K the imaginary part
of the poles is then 2.0. Thus we have Tu = 2u 3.14. The corresponding values for the
PID parameters are displayed in Table 13.4, and the normalised closed-loop step responses

ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)
h (t)
hN,D (t)

1 (t)

1N,D (t)
(t)
(t)
fj (t)
Re
Im
arex1 shown
x2
x1
x2 1.4
log
dB 1.2
deg
1
u = 0 ln
ln coth(|u| /2) 0.8
or 1
m 0.6
0.4
yos
0.2
tos 0
0

e1 (t)
e2 (t)
e3 (t)
Controller
y(t)
u(t)
K h5.5
(t)
hN,D (t)

Controller type
P

K 4.95, TI 2.62

PID

K (t)
6.6, TI = 1.57, TD 0.39

in Figure ke(t)k
13.7. The
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)
h (t)
hN,D (t)

1N,D (t)
(t)
(t)
fj (t)
Re
Im
x
2.5
5
7.5 110 12.5
x2 t
x1
x2 1.4
log 1.2
dB
deg
1
u = 0 ln
ln coth(|u| /2) 0.8
or 1 0.6
m
0.4
yos
0.2
tos 0
0

,0
0

13.2.1 Pole placement using polynomials In this section we engage in a rather general discussion of the closed-loop poles using purely polynomial methodology. We consider
the standard unity gain feedback loop of Figure 13.8. Our objective is to characterise some
RC (s)

RP (s)

y(s)

PSfrag replacements

22/10/2004

In this section we use a form of pole placement to indicate how to select PID parameters
based upon the location of poles. Clearly one cannot choose a PID controller to place the
poles anywhere for a general controller. However, in this section we see exactly how well we
can do.

r(s)

parameters

1N,D (t)

(t)
fj (t)
Re
Im
PID responsex1is respectable,
x2
x1
x2
1.5
log
dB 1.25
deg
u = 0 ln
1
ln coth(|u| /2)
or 1 0.75
m
0.5

yos
0.25
tos 0
0

2.5
5

Figure 13.8 Feedback loop for polynomial pole placement

but might benefit from more

17.5

20,0

7.5

12.5

17.5

of the possible closed-loop characteristic polynomials for the system. The main result is the
following.
13.2 Theorem Consider the interconnection of Figure 13.8 and let (NP , DP ) be the c.f.r. for
RP , supposing deg(DP ) = n. The following statements hold:
(i) if deg(NP ) n 1 and if P R[s] is monic and degree 2n 1, then there exists a
proper RC R(s) with c.f.r. (NC , DC ) so that the closed-loop characteristic polynomial
of the interconnection, DC DP + NP NC , is exactly P ;
(ii) if deg(NP ) n and if P R[s] is monic and degree 2n, then there exists a strictly
proper RC R(s) with c.f.r. (NC , DC ) so that the closed-loop characteristic polynomial
of the interconnection, DC DP + NP NC , is exactly P .
Proof The following result contains the essential part of the proof.
1 Lemma (Sylvesters theorem) For polynomials

1 (t)

,0
0

replacements
Table 13.4 ControllerPSfrag
parameters
for example using second
Ziegler-Nicols tuning method
ke(t)k

1 (t)

PSfrag replacements

13 Advanced synthesis, including PID synthesis

13.2 Synthesis using pole placement

the closed-loop transfer function

T (s) =

500

P (s) = pn sn + pn1 sn1 + + p1 s + p0

Q(s) = qn sn + qn1 sn1 + + q1 s + q0 ,
2.5

7.5

12.5

17.5

Figure 13.7 Normalised step responses for example using second

Ziegler-Nicols tuning method: P (top left), PI (top right), and
PID (bottom)

derivative time given its largish overshoot.

with p2n + qn2 6= 0, define their eliminant as the 2n 2n

pn
0
0
qn
pn1 pn
0
qn1
.
.. . .
..
..
.
.
.
.
.
.

p2 pn
q1
p
M (P, Q) = 1
p0 p1 pn1 q0

p0 pn2
0
0
.
.. . .
..
..
..
.
.
.
.
0
0 p0
0

matrix

0
0
qn
0

. . . ..

q2 qn
.
q1 qn1

q0 qn2

..
..

. .
0

Then P and Q are coprime if and only if det M (P, Q) 6= 0.

13.2 Synthesis using pole placement

22/10/2004

502

501

Proof Note that if P and Q are not coprime then there exists z C so that
P (s) = (s z)(
pn1 s
+ + p1 s + p0 )
Q(s) = (s z)(
qn1 sn1 + + q1 s + q0 ).

P (s) = s2n1 + p2n2 s2n2 + + p1 s + p0 .

This implies that

Now define a vector in R2n by

bn1
..
.

p2n1
b1
..

= M (DP , NP )1 . .
a
n1
p1
.
..
p0

(
qn1 sn1 + + q1 s + q0 )P (s) (
pn1 sn1 + + p1 s + p0 )Q(s) = 0.
We now balance the coefficients of powers of s in this expression:
s2n1 : pn qn1 qn pn1 = 0
s2n2 : pn1 qn1 + pn qn2 qn1 pn1 qn pn2 = 0
..
.
s1 : p0 q1 + p1 q0 q0 p1 q1 p0 = 0
s0 : p0 q0 q0 q0 = 0.

One verifies by direct computation that the resulting equation

bn1
..
.

p2n1
b1

b
M (DP , NP ) 0 = .
an1 p1
.
..
p0

a1
a0

(
qn1 sn1 + + q1 s + q0 )P (s) = (
pn1 sn1 + + p1 s + p0 )Q(s)
qn1 sn1 + + q1 s + q0
Q(s)
=
.
P (s)
pn1 sn1 + + p1 s + p0

Since either P or Q has degree n, it must be the case that P and Q have a common factor. H

(13.2)

One readily ascertains that this is exactly equivalent to

qn1
..
.

q
M (P, Q) 0 = 0.

p
n1
.
..

Now reversing the argument for the preceding part of the proof shows that

22/10/2004

Now we proceed with the proof.

(i) Since NP and DP are coprime, their eliminant M (DP , NP ) is invertible by Lemma 1.
Now let P R[s] be monic and of degree 2n 1 and write

This then implies that det M (P, Q) = 0 since not all of the coefficients
qn1 , . . . , q0 , pn1 , . . . , p0 can vanish.
Now suppose that det M (P, Q) = 0. This implies that there is a nonzero vector x R2n
so that M (P, Q)x = 0. Let us write

qn1
..
.

q0
x=
.
pn1

.
..

13 Advanced synthesis, including PID synthesis

is exactly the result of equating DC DP + NC NP = P , provided that we define

RC (s) =

finish

an1 sn1 + + a1 s + a0
.
sn1 + bn2 sn2 + + b1 s + b0

To complete the proof, we must also show that the numerator and denominator in this
expression for RC are coprime.
(ii) In this case we take
P (s) = s2n + p2n1 s2n1 + + p1 s + p0 ,
and define a vector in R2n+1 by

bn
bn1
.

.
p2n
.

p2n1
b1

.
(DP , NP )1
b0 = M
.. .

an1
p1
.
..
p0

a1
a0

(13.3)

13.2 Synthesis using pole placement

22/10/2004

(DP , NP ) is defined by
Here the matrix M

0t
p2n
(DP , NP ) =
M
,
m(DP ) M (DP , NP )

503

13 Advanced synthesis, including PID synthesis

s2
.
s+4
RC (s)RP (s)
(s 1)(s 2)
T (s) =
= 3
.
1 + RC (s)RP (s)
s + 3s2 + 4s + 2
RC (s) =

(13.4)

Note that RC is indeed a proper, but not a strictly proper, controller.

To achieve a proper controller we use part (ii) of Theorem 13.2. To do this, we must
specify a closed-loop characteristic polynomial of degree 4. Let us go with

an1 sn1 + + a1 s + a0
,
+ bn1 sn1 + + b1 s + b0

then this controller satisfies the conclusions of this part of the proposition.

P (s) = s4 + 4s3 + 7s2 + 6s + 2.

13.3 Remarks 1. This result is analogous to Theorem 10.27 in that it provides an explicit
formula, in this case either (13.2) or (13.3), for a stabilising controller for the feedback
loop of Figure 13.8. In fact, in each case we can achieve a prescribed characteristic
polynomial of a certain type.
2. In Theorem 10.27 the characteristic polynomial had to be of degree 2n and had to be
writable as a product of two polynomials (this latter restriction is only a restriction when
n is odd). However, in Theorem 13.2 we go this one better because the characteristic
polynomial had degree one less, 2n 1, at least in cases when RP is strictly proper.
This means we have a controller whose denominator has one degree less than that of
Theorem 10.27. This can be advantageous.
3. One of the things we loose in Theorem 13.2 is the separation principle interpretation
available for Theorem 10.27 (cf. Theorem 10.48).

An example illustrates how to explicitly apply Theorem 13.2.

(DP , NP ) in the equation (13.3). Using the description

Next we must determine the matrix M
of this provided in the proof we get

1 0 0 0
0
1 1 0 0
0

(DP , NP ) = 0 0 1 1 0 .
M

1 1 0 1 1
0 0 1 0
1
An application of (13.3) gives

1
1
4 4

(DP , NP ) 7 = 5 .
M

6 1
2
3
This then gives the strictly proper controller

13.4 Example Let us consider the unstable, nonminimum phase plant

RP (s) =

1s
.
s2 + 1

We first wish to design a proper controller that produces the closed-loop characteristic polynomial
P (s) = s3 + 3s2 + 4s + 2.
We determine the eliminant M (DP , NP ) to be

1
0
M (DP , NP ) =
1
0

0 0
0
1 1 0
.
0 1 1
1 0
1

22/10/2004

This then gives the controller

and where m(DP ) R2n is a vector containing the coefficients of DP , with the coefficient
(DP , NP ) is
of sn1 in the first entry, and with the last n entries being zero. Clearly M
invertible since M (DP , NP ) is invertible. Now one checks that if
RC (s) =

504

RC (s) =

s 3
.
s2 + 4s + 5

The corresponding closed-loop transfer function is

T (s) =

RC (s)RP (s)
(s + 3)(s 1)
= 4
.
1 + RC (s)RP (s)
s + 4s3 + 7s2 + 6s + 2

Note that the order of the two controllers we have designed is as predicted by Theorem 13.2,
and are comparable to the order of the plant.
The closed-loop characteristic polynomials are designed to be stable. In Figure 13.9 we
show the Nyquist plots for both systems. One can get some idea of the stability margins of
the closed-loop system from these.

Let us discuss this a little further by looking into a couple of related results. The first is
that if we wish to increase the denominator degree of the controller, we may.

An application of (13.2) gives

1
1
4
1 3

M (DP , NP ) =
.
4
1
2
2

13.5 Corollary Consider the interconnection of Figure 13.8 and let (NP , DP ) be the c.f.r. for
RP , supposing deg(DP ) = n deg(NP ). If P R[s] is monic, of degree k 2n 1, and if
the coefficient of s2n1 in P is nonzero, then there exists RC R(s) with c.f.r. (NC , DC )
so that the closed-loop characteristic polynomial of the interconnection, DC DP + NP NC , is
exactly P .

PSfrag replacements

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)
h22/10/2004
(t)
hN,D (t)
1 (t)
4
1N,D (t)
(t)
(t) 3
fj (t)

13.2

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)
Synthesis
h (t)using
hN,D (t)
1 (t)
4
1N,D (t)
(t)
(t) 3
fj (t)

,0
0

505

-2

-1

5,0

22/10/2004

5s3 16s2 8s 20
s3 + 24

achieves the desired characteristic polynomial. However, this is not the only controller that
accomplishes this task. Indeed, one may verify that any controller of the form
5s3 16s2 8s 20 + a1 (s3 + s2 + s + 1) + a2 (s2 + 1)
s3 + 24 + a1 (s2 1) + a2 (s 1)
will achieve the same characteristic polynomial. Thus we see explicitly how the freedom of
the high degree characteristic polynomial plays out in the controller. This may be helpful
in practice to fine tune the controller do have certain properties.

-2

-1

It is not surprising that we should be able to achieve a characteristic polynomial of degree

higher than that of Theorem 13.2. It is also the case that we can generally expect to do not
better. The following result gives this in precise terms.

Figure 13.9 Nyquist plots for polynomial pole placement for

RP (s) = s21+1 for a proper controller (left) and a strictly proper
controller (right)

13.8 Proposition For the interconnection of Figure 13.8, let RP be proper with c.f.r. (NP , DP )
and deg(DP ) = n. If k < 2n 1 then there exists a monic polynomial P of degree k so that
there is no proper controller RC with the property that deg(DC ) = k n, where (NC , DC )
is the c.f.r. for RC .

Proof Define
DC,1 (s) = sk + bk1 sk1 + + bn sn
by asking that the polynomial DC,1 DP P have degree 2n 1. Thus DC,1 is obtained
by equating the coefficients of sk , . . . , s2n in the polynomials DC,1 DP and P . Now define
1
(P DC,1 DP ). By construction of DC,1 , P is monic and has degree 2n 1. Thus,
P = p2n1
by Theorem 13.2, there exists

RC =

gives the corollary.

NC
DC

13.6 Remark The controller RC in the corollary will not be unique as it will be, for example,

in Theorem 13.2.
Let us now see that the above result works in an example.
13.7 Example (Example 13.4 contd) Let us see what happens when we choose a closed-loop
characteristic polynomial whose degree is too high. Let us suppose that we wish to achieve
the closed-loop characteristic polynomial
P (s) = s5 + 5s4 + 12s3 + 16s2 + 12s + 4.

k < 2n 1
k > 2n + 1
k + 1 > 2k 2n + 2.

This means that the linear equation for determining the coefficients of RC has more equations
than unknowns. Since a linear map from a vector space into one of larger dimension is
incapable of being surjective, the proposition follows.

C . Taking
with the property that DC,2 DP + NC,2 NP = P , if (NC,2 , DC,2 ) is the c.f.r. for R
NC = p2n1 NC,2 ,

Proof The most general proper controller RC for which deg(DC ) = kn will have 2(nk+1)
undetermined coefficients. The equation DC DP +NC NP = P gives a linear equation in these
coefficients by balancing powers of s. This linear equation is one with k + 1 equations in
2(k n + 1). However, we have

=
=

a
n1 s
+ + a
1 s + a
0
sn1 + bn2 sn2 + + b1 s + b0

DC = DC,1 + p2n1 DC,2 ,

13 Advanced synthesis, including PID synthesis

RC (s) =

C (s) =
R

506

We should look for a controller of order 5 2 = 3. One may verify that the controller

x1
x2 1
x1
x2
log
0
dB
deg
u = 0 ln -1
ln coth(|u| /2)
or 1 -2
m

yos -3

tos 0

x1
x2 1
x1
x2
log
0
dB
deg
u = 0 ln -1
ln coth(|u| /2)
or 1 -2
m

yos -3

tos 0

pole placement

This means that we cannot expect to stabilise a general plant except with a controller
whose denominator has comparable degree. This is easily illustrated with an example
13.9 Example (Example 13.4 contd) Let us continue with our example where
RP (s) =

1s
.
s2 + 1

We wish to show that we cannot find a proper controller RC with the property that the
closed-loop characteristic polynomial for the interconnection of Figure 13.8 is an arbitrary
Hurwitz polynomial of degree 2. Clearly a proper controller of degree 1 or greater will lead
to a closed-loop characteristic polynomial of degree 3 or greater. Thus we may only use a
constant controller: RC (s) = K. In this case the closed-loop characteristic polynomial is
readily determined to be
P (s) = s2 Ks + 1 + K.

e2 (t)
e3 (t)
y(t)
u(t)
h (t)
22/10/2004 hN,D (t)
1 (t)
1N,D (t)
(t)
(t)
fj (t)

13.2 Synthesis using pole placement

507

508

13 Advanced synthesis, including PID synthesis

22/10/2004

(d) the closed-loop characteristic polynomial of the interconnection, DC DP + NP NC ,

is exactly P .
Proof (i) Let us write
RP (s) =

0.5
x1
x2
x1
x2
log
0
dB
deg
u = 0 ln
ln coth(|u| /2)
or 1 -0.5
m

yos

tos 0
-1

,0
0

cn1 sn1 + + c1 s + c0
+ dn1 sn1 + + d1 s + d0

and
P (s) = s2n+k1 + p2n+k2 s2n+k2 + + p1 s + p0 .
The idea is the same as in the proof of Theorem 13.2 in that the issue is matching coefficients.
The content lies in ascertaining the form of the coefficient matrix. In this case we define a
(2n + k) k matrix Ak (NP ) by

-1

-0.5

X
0

0.5

0
0
..
..
.
.

0
0

0
c
Ak (NP ) = n1
cn2 cn1
.
..
..
.

0
0
0
0

Figure 13.10 Locus of roots for s2 Ks + 1 + K where K (1, 0)

Note that this polynomial is Hurwitz only for K (1, 0). Thus the class of closed-loop
characteristic polynomials of degree 2 that we may achieve is limited. Indeed, in Figure 13.10
we show the locus of roots for those values of K for which the closed-loop system is IBIBO
stable. The placement of poles is clearly restricted.

13.2.2 Enforcing design considerations Sometimes one wants to design a controller
with certain properties. One of the most often encountered of these is that RC have a pole at
s = 0 so as to ensure asymptotic tracking of step inputs and rejection of step disturbances.
With this in mind, we have the following result.
13.10 Proposition Consider the interconnection of Figure 13.8 and suppose that (NP , DP ) if
the c.f.r. for RP . For k N the following statements hold:
(i) if deg(NP ) n 1 and if P R[s] is monic and degree 2n + k 1, then there exists
a controller RC R(s) with c.f.r. (NC , DC ) with the following properties:
(a)
(b)
(c)
(d)

RC is proper;
RC has a pole of order at least k at s = 0;
deg(DC ) = n + k 1;
the closed-loop characteristic polynomial of the interconnection, DC DP + NP NC ,
is exactly P ;

(ii) if deg(NP ) n and if P R[s] is monic and degree 2n + k, then there exists a
controller RC R(s) with c.f.r. (NC , DC ) with the following properties:
(a) RC is strictly proper;
(b) RC has a pole of order at least k at s = 0;
(c) deg(DC ) = n + k;

...

0
..
.

0 0

0 0
.
0 0
.. ..
. .

c0 c1
0 c0
0
..
.

The way to assemble Ak (NP ) in practice is to put the components of the polynomial NC
into the last column of Ak (NP ), starting with c0 in the last row and working backwards.
The next column to the left is made by shifting the last column up one row and placing a
zero in the last row. One proceeds in this way until all k columns have been formed. Now
define

M (DP , NP )
M k (DP , NP ) =
Ak (NP ) R(2n+k)(2n+k) .
0
One now defines a vector in R2n+k by

bn+k1

..
p2n+k1
.

p2n+k2
bk

= M k (DP , NP )1 .. .
an+k1
.
.
..
p0
a0
One now checks that if
RC (s) =

an+k1 sn+k1 + + a1 s + a0
,
bn+k1 sn+k1 + + bk sk

then RC has the properties stated in the proposition.

(ii) Here we write
RP (s) =

cn sn + cn1 sn1 + + c1 s + c0
sn + dn1 sn1 + + d1 s + d0

13.2 Synthesis using pole placement

22/10/2004

509

and

510

13 Advanced synthesis, including PID synthesis

22/10/2004

We then compute

P (s) = s2n+k + p2n+k1 s2n+k1 + + p1 s + p0 .

k (NP ) R(2n+k+1)k by
The next step is to define A

0
0 0 0
.. . .
. .
..
. .. ..
.

0
0

cn
0 0 0
k (NP ) =
A

.
c
c

0
0

n1 n

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

0
0 c0 c1
0

PSfrag replacements

t
ke(t)k
e1 (t) we take
Thus
e2 (t)
e3 (t)
y(t)
u(t)
hThe
(t) Nyquist plot
hN,D (t)
1 (t)
1N,D (t) 6
(t)
(t)
fj (t)
4

0 c0

Then we let
k (DP , NP ) =
M

(DP , NP )
M
0

k (NP )
A

R(2n+k2)(2n+k2) ,

x1
2
x2
x1
x2
log
0
dB
deg
PSfrag replacements
u = 0 ln
ln coth(|u| /2)t -2
ke(t)k
or
1
e1(t)
m
e2 (t) -4
e3y(t)
os
y(t)
tu(t)
os 0
h (t)
,0
hN,D
(t)
0
1 (t)
1N,D (t) 20
(t)
(t)
fj (t)

The rational function

RC (s) =

(DP , NP ) is as defined in (13.4). Then we define a vector in R2n+k+1 by

where M

bn+k

.
.
p2n+k
.

p2n+k1
bk

k (DP , NP )1

=M
.. .
an+k1
.
.
..
p0
a0
an+k1 sn+k1 + + a1 s + a0
bn+k sn+k + + bk sk

has the desired properties.

for the loop

ke(t)k
e1 (t)
e25s
(t)2 s +
RC (s) =e3 (t) 2
y(t)s + 9s
u(t)
(t)
gain RC RP ishshown
in
hN,D (t)
1 (t)
1N,D (t) 6
(t)
(t)
fj (t)
4

-4

-2

13.11 Example (Example 13.4 contd) We take

RP (s) =

1s
,
s2 + 1

and design a controller with a pole at s = 0 that produces a desired characteristic polynomial.
First let us achieve a pole of degree k = 1 at the origin. To derive a proper controller, P
must have degree 2n + k 1 = 4. Let us take
P (s) = s4 + 4s3 + 7s2 + 6s + 2.
The matrix M 1 (DP , NP ) is then constructed

1
0

M 1 (DP , NP ) =
1
0
0

as in the proof of Proposition 13.10 as

0 0
0
0
1 1 0
0

0 1 1 0
.
1 0
1 1
0 0
0
1

Figure 13.11.

-4

-2

-4

-2

x1
x2
x1
x2
0
log
dB
deg
u = 0 ln
ln coth(|u| /2)
or 1 -10
m

yos

tos 0 -20

,0
0

x1
2
x2
x1
x2
log
0
dB
deg
u = 0 ln
ln coth(|u| /2) -2
or 1
m
-4
yos

tos 0

Let us apply this in an example, as the proof, as was the proof of Theorem 13.2, is
constructive.

x1
2
x2
x1
x2
log
0
dB
deg
u = 0 ln
PSfrag replacements
ln coth(|u| /2) -2
t
or
1
ke(t)k
m
e1(t)
-4
e2 (t)
os
e3y(t)
y(t)
tos
0
u(t)

h6,
(t)
4
0
hN,D(t)
0
1 (t)
1N,D (t) 6
(t)
(t)
fj (t)
4

1
1
4 9

M 1 (DP , NP )1
7 = 5 .
6 1
PSfrag replacements
2
t 2

-10

20 ,0
0

1s
and controllers
s2 +1
8s2 3s+4
(top right),
s3 +5s2 +19s

Figure 13.11 Nyquist plot for plant RP (s) =

RC (s) =
RC (s) =

5s2 s+2
(top
s2 +9s
19s3 +8s2 +16s+4
s3 +24s2

left), RC (s) =

(bottom left and right)

If we demand that RC be proper, then we use part (ii) of Proposition 13.10. We must

22/10/2004

13.2 Synthesis using pole placement

511

now specify a polynomial of degree 2n + k = 5; let us take

P (s) = s5 + 5s4 + 12s3 + 16s2 + 12s + 4.
We compute

1
0

1 (DP , NP ) = 1
M
0

0
0

0
1
0
1
0
0

0 0
0
0
0 0
0
0

1 1 0
0
,
0 1 1 0

1 0
1 1
0 0
0
1

giving

1
1
5 5

1 (DP , NP )1 12 = 19 .
M
16 8

12 3
4
4
Thus we have

8s2 3s + 4
.
s3 + 5s2 + 19s
One may see the Nyquist plot for the controller in Figure 13.11.
We may also design a controller giving a pole of degree k = 2 at s = 0. This requires the
specification of a closed-loop characteristic polynomial of degree 2n + k 1 = 5; let us take
RC (s) =

P (s) = s5 + 5s4 + 12s3 + 16s2 + 12s + 4.

again. We compute

1
0

1
M 2 (DP , NP ) =
0

0
0
and

0 0
0
0
0
1 1 0
0
0

0 1 1 0
0

1 0
1 1 0

0 0
0
1 1
0 0
0
0
1

1
1
5 24

12 19

M 2 (DP , NP )1
16 = 8 .

12 16
4
4

Thus we take

19s3 + 8s2 + 16s + 4

.
s3 + 24s2
The Nyquist plot for this plant/controller is shown in Figure 13.11. This is not strictly
proper, and one may of course apply part (ii) of Proposition 13.10 to get a strictly proper
controller. However, it is hopefully clear how to do this at this point.

RC (s) =

512

13 Advanced synthesis, including PID synthesis

22/10/2004

13.12 Remark It is interesting to look at the Nyquist plots for the controllers designed in
Examples 13.4 and 13.11. First of all, the Nyquist plots get increasingly more complicated
as we increase the controller order. Ones ability to design controllers of this complexity with
ad hoc methods is quite limited. On the other hand, one can see that the Nyquist plot for
the third controller of Example 13.11 (the bottom Nyquist plots in Figure 13.11) has very
bad gain and phase margins. One would be suspicious of such a controller, even though the
theory tells us that it produces a desirable closed-loop characteristic polynomial. This points
out the importance of using all the tools at ones disposal when designing a controller.
One can also ask that more general polynomials appear in the denominator of RC . Indeed,
the proof of Proposition 13.10 can easily, if tediously, be adapted to prove the following
result.
13.13 Corollary Consider the interconnection of Figure 13.8 and suppose that (NP , DP ) if the
c.f.r. for RP . For F R[s] a monic polynomial of degree k, the following statements hold:
(i) if deg(NP ) n 1 and if P R[s] is monic and degree 2n + k 1, then there exists
a controller RC R(s) with c.f.r. (NC , DC ) with the following properties:
(a)
(b)
(c)
(d)

RC is proper;
DC has F as a factor;
deg(DC ) = n + k 1;
the closed-loop characteristic polynomial of the interconnection, DC DP + NP NC ,
is exactly P ;

(ii) if deg(NP ) n and if P R[s] is monic and degree 2n + k, then there exists a
controller RC R(s) with c.f.r. (NC , DC ) with the following properties:
(a)
(b)
(c)
(d)

RC is strictly proper;
DC has F as a factor;
deg(DC ) = n + k;
the closed-loop characteristic polynomial of the interconnection, DC DP + NP NC ,
is exactly P .

Just when one might wish to do this is a matter of circumstance. The following example
gives one such instance.
13.14 Example Suppose we are given a plant RP with c.f.r. (NP , DP ) where deg(DP ) = n.
A reasonable design specification is that the closed-loop system be able to track well signals
of a certain period . This would seem to demand that the transfer function from the input
to the error should be zero at s = i. For the interconnection of Figure 13.8 the transfer
function from the input to the error is the sensitivity function for the loop,
SL (s) =

1
.
1 + RC (s)RP (s)

By ensuring that RC has a pole at s = i, we can ensure that SL (i) = 0. Thus we should
seek a controller of the form
NC (s)
RC (s) =
.
C (s)
(s2 + 2 )D
Corollary 13.13 indicates that we can find a proper such controller provided that the closedloop characteristic polynomial is specified to be of degree 2n + k 1. If the controller is

22/10/2004

13.2 Synthesis using pole placement

513

514

13 Advanced synthesis, including PID synthesis

to be strictly proper, then we must allow the closed-loop characteristic polynomial to have
degree 2n + k. While we do not produce the explicit formula for the coefficients in NC and
C , one can easily produce such a formula by balancing polynomial coefficients, just as is
D
done in Theorem 13.2 and Proposition 13.10.

If one defines

One may also wish to specify that the numerator of RC have roots at some specified
locations. The following result tells us when this can be done, and the degree of the closedloop polynomial necessary to guarantee the required behaviour. We omit the details of the
proof, as these go much like the proofs of Theorem 13.2 and Proposition 13.10, except that
there are more complications.

then R = RC , where RC is as in (13.5).

RC is proper;
NC has F as a factor;
deg(DC ) = n + k 1;
the closed-loop characteristic polynomial of the interconnection, DC DP + NP NC ,
is exactly P ;

(ii) if deg(NP ) n and if P R[s] is monic and degree 2n + k, then there exists a
controller RC R(s) with c.f.r. (NC , DC ) with the following properties:
(a)
(b)
(c)
(d)

RC is strictly proper;
NC has F as a factor;
deg(DC ) = n + k;
the closed-loop characteristic polynomial of the interconnection, DC DP + NP NC ,
is exactly P .

Note that if the GCD of NC and DP is F , then F is guaranteed to appear as a factor in

the closed-loop characteristic polynomial.
13.2.3 Achievable poles using PID control We begin by noting that in this section,
as in the rest of this chapter, we use a PID controller that renders proper the derivative term
in the controller. Thus we take a controller transfer function of the form

1
TD s
RC (s) = K 1 +
+
.
(13.5)
D s + 1 TI s
The advantage of this from our point of view in this section is contained in the following
result.
13.16 Lemma Let R R[s] be any rational function of the form
R(s) =

a2 s 2 + a1 s + a0
.
s2 + b 1 s

TD =

a0 a1 b1 + a2 b21
,
b1 (a1 b1 a0 )

TI =

a1 b 1 a0
,
a0 b 1

D =

1
,
b1

Proof We compute
RC (s) =

K(1 +

TD 2
)s
D

+ K( 1D +
s2

1
)s
TI

K
D TI

1
s
D

The lemma follows by setting

13.15 Proposition Consider the interconnection of Figure 13.8 and suppose that (NP , DP ) if
the c.f.r. for RP . For F R[s] a monic polynomial of degree k, the following statements
hold:
(i) if deg(NP ) n 1 and if P R[s] is monic and degree 2n + k 1, then there exists
a controller RC R(s) with c.f.r. (NC , DC ) with the following properties:
(a)
(b)
(c)
(d)

a1 b 1 a0
,
b21

22/10/2004

a2 = K(1 +

TD
),
D

a1 = K( 1D +

1
),
TI

a0 =

K
,
D TI

b1 =

1
,
D

and solving for a0 , a1 , a2 , and b1 .

na
ve PID
controllers
with positive
parameters

It is appropriate to employ a PID controller for first and second-order plants. To design
PID controllers using the machinery of Sections 13.2.1 and 13.2.2, we use Proposition 13.10
with k = 1 to take the integrator into account. Let us see explicitly how to do this for
general first and second-order plants.
13.17 Proposition Consider the three strictly proper plant transfer functions
R (s) =

1
,
s + 1

R,0 , (s) =

02 ( s + 1)
,
+ 20 s + 02

R1 ,2 , (s) =

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

and the two polynomials

P1 (s) = s3 + as2 + bs + c,

P2 (s) = s4 + as3 + bs2 + cs + d.

The following statements hold:

(i) if RP = R and if
RC (s) =

(a 1)s2 + b s + c s( s + 1)
s2 + s

for R, then the closed-loop polynomial of the interconnection of Figure 13.8 is P1 ;

(ii) if RP = R,0 , and if
a2 s 2 + a1 s + a0
s2 + b 1 s
b c + d 2 02 + a 02 2a0 2 03 + 402 2
a2 =
02 (1 + 2 02 2 0 )
c
d
a b + c 2 d 3 + 02 20
a1 = 2 2
0
0
1 + 2 02 2 0
d
a0 = 2
0
a b + c 2 d 3 + 02 20
b1 =
1 + 2 02 2 0
RC (s) =

then the closed-loop polynomial of the interconnection of Figure 13.8 is P2 ;

then the closed-loop polynomial of the interconnection of Figure 13.8 is P2 .

13.18 Remarks 1. Using Lemma 13.16 one can turn the controllers of Proposition 13.17 into
PID controllers of the form (13.5).
2. Note that the three plants R , R,0 , and R1 ,2 , cover all possible strictly proper first
and second-order plants.
3. The essential point is not so much the formulae themselves as their existence. That is to
say, the main point is that for a first-order plant, a PID controller can be explicitly found
that produces a desired third-order characteristic polynomial, and that for a secondorder plant, a PID controller can be explicitly found that produces a desired fourth-order
characteristic polynomial.
4. In practice, one would not use the formulae of Proposition 13.17, but would simply
2 +a s+a
1
0
design a controller of the type RC (s) = a2 ss(s+b
by enforcing a pole at s = 0 as in
1)
Proposition 13.10.
5. Note that for first-order plants, there is some freedom in the design of an appropriate PID
controller (characterised by the presence of the parameter in part (i) of the proposition).
However, the PID controller of part (ii) is uniquely specified by the desired characteristic
polynomial.

Let us illustrate a PID design using Proposition 13.17.
13.19 Example (Example 12.14 contd) So that we may contrast our design with that of the
ad hoc PID design of Section 12.2.3, we take RP = s12 . Let us design a controller with poles
at {5, 5, 2 2i}. Thus we require the characteristic polynomial
P = s4 + 14s3 + 73s2 + 180s + 200.
By contrast, the closed-loop characteristic polynomial of Example 12.14 is
s3 +

13 2
s
4

39
s
10

t
t
ke(t)k
Converting
this to the PID form of (13.5) ke(t)k
using
Lemma 13.16 gives
e1 (t)
e1 (t)
e2 (t)
e2 (t)
580
2997
29
1
e3 (t)
3 (t),
K=
, TD = ey(t)
TI = , D = .
y(t)
49
8120
35
14
u(t)
u(t)
hIn
(t)
h (t)
Figure 13.12 we show the Nyquist plot for
the loop gain RC RP . Note
hN,D (t)
hN,D (t)
1 (t)
1 (t)
1N,D (t)
1N,D (t) 3
60
(t)
(t)
(t)
(t)
fj (t)
fj (t) 2
40
x1
x2
x1
x2
log
dB
deg
u = 0 ln
ln coth(|u| /2)
or 1
m

PSfrag replacements
yos

ke(t)k
teos1 (t)
0

e2 (t)
e3,
(t)0
0
y(t)
u(t)
h (t)
hN,D (t)

22/10/2004

PSfrag replacements

that this controller

x1
x2 1
x1
x2
log
0
dB
deg
u = 0 ln
ln coth(|u| /2) -1
or 1
m
PSfrag replacements
-2
yos
ke(t)k
teos1 (t)
0

e2 (t)
-2
-1
0
1
0
20
40
60
e3,
(t)0
Re
Re
0
y(t)
u(t)
(t)
13.12 Nyquist plot forhplant
RP (s) = s12 and controller
hN,D (t)
2
Im

(a2 /b2 )s2 + (a1 /b2 )s + (a0 /b2 )

s2 + (b1 /b2 )s
a2 = 22 d 2 12 22 + 1 2 (1 + a2 ) + 12 (1 + a2 b22 ) + (1 + 2 a1 2 + c12 22 )
a1 = 1 2 + a1 2 c12 22 d 2 1 2 (1 + 2 ) + (1 b1 2 + d12 22 + c1 2 (1 + 2 ))
a0 = (d( 1 )1 ( 2 )2 )
b2 = ( 1 )( 2 )
b1 = 1 + 2 a1 2 c 2 1 2 + d 3 1 2 + (1 + b1 2 ),

RC (s) =

13 Advanced synthesis, including PID synthesis

-20

-40

-60
-40

-20

Figure
RC (s) =

73s +180s+200
s(s+14)

1N,D (t)
1N,D (t)
(t)
(t)
(t)
(t)
fj (t) off the bat presented us with respectable
fj (t)gain and phase margins. If one wished, this
has
Re
Re further refinements to the stability margins.
controller
could be used as a starting point for
Im
Im
x1 the response to a step disturbance between
1
13.13 is shown the step response and
InxFigure
x2
x2
x1
x1
1
x2
x2 1.4
log 1.2
log
0.8
dB
dB
deg 0.6
deg
1
u = 0 ln
u = 0 ln
0.4
ln coth(|u| /2)
ln coth(|u| /2) 0.8
or 1 0.6
or 1
0.2
m
m
0.4

0
yos
yos
0.2
-0.2
tos 0 -0.4
tos 0
0

2
4
6
8
10
2
4
6
8
10,0
,0
t
t
0
0
1 (t)

(iii) if RP = R1 ,2 , and if

516

PSfrag replacements

515

1 (t)

13.2 Synthesis using pole placement

22/10/2004

Figure 13.13 Step response (top) and response to step disturbance

(bottom)

13
,
20

which has roots of approximately {0.197, 1.53 0.984i}. The lower degree of the characteristic polynomial is Example 12.14 is a consequence of the derivative term not being
proper as in (13.5). In any event, an application of Proposition 13.17, or more conveniently
of Proposition 13.10, gives
73s2 + 180s + 200
RC (s) =
.
s(s + 14)

which enters the system at the output. The response time is quite good, although the
overshoot is a bit large, and could be improved with a larger derivative time, perhaps.

13.3 Two controller configurations

In the preceding section we considered the achieving of a specified closed-loop characteristic polynomial using a unity gain feedback configuration as in Figure 13.8. Next we

13.3 Two controller configurations

22/10/2004

517

turn our attention to a richer specification of the closed-loop transfer function to allow the
determination of not only the closed-loop characteristic polynomial, but also additional features of the closed-loop transfer function. To do this we must consider a more complicated
interconnection, and we consider the interconnection of Figure 13.14. There are now two
r(s)

Rff (s)

RP (s)

518

13 Advanced synthesis, including PID synthesis

Proof (i) = (ii) If R is implementable, the interconnection of Figure 13.14 is IBIBO

stable. In particular, the closed-loop transfer function is IBIBO stable, and since this is
by hypothesis R, we have R RH+
to the
. The transfer function from the plant input u
reference r should also be BIBO stable. This means that
y/
r
R
u
=
=
RH+
.
r y/
u
RP

y(s)

Thus (ii) holds.

(ii) = (iii) As R =
be proper. Since

Rfb (s)

Figure 13.14 A two controller feedback loop

controllers we may specify, and we call Rfb the feedback controller and Rff the feedforward controller . It certainly makes sense that having two controllers makes it possible
to do more than was possible in Section 13.2. The objective of this section is to quantify
how much can be done with the richer configuration, and to detail exactly how to do what
is possible.
13.3.1
after.

Implementable transfer functions

First let us be clear about what we are

13.20 Definition Let RP R(s) be a proper plant. R R(s) is implementable for RP if

there exists proper feedback and feedforward controllers Rfb , Rff R(s) so that
(i) R is the closed-loop transfer function for the interconnection of Figure 13.14,
(ii) the interconnection is IBIBO stable, and
(iii) the only zeros of R in C+ are those of RP , including multiplicities.
We denote by I (RP ) the collection of implementable closed-loop transfer functions for the
plant RP .

The next result tells us that it is possible to achieve an implementable transfer function
with an interconnection of the form Figure 13.14.
13.21 Theorem Let RP be a proper plant with c.f.r. (NP , DP ), and let R R(s) have c.f.r.
(N, D). The following conditions are equivalent.
(i) R is implementable;
(ii) R, RRP RH+
and the strictly nonminimum phase zeros of R and RP agree, including
multiplicities;
(iii) the following three conditions hold:
(a) D is Hurwitz;
(b) deg(D) deg(N ) deg(DP ) deg(NP );
(c) the roots of N in C+ are exactly the roots of N in C+ , including multiplicities.

22/10/2004

N
D

obviously D is Hurwitz so (iii a) holds. Since

R
RP

RH+
,

R
RP

must

N DP
R
=
,
(13.6)
RP
DNP
this implies that deg(DNP ) deg(N DP ). Since the degree of the product of polynomials is
the sum of the degrees, it follows that
deg(D) + deg(NP ) deg(N ) + deg(DP ),
from which (iii b) follows. If R, RRP RH+ + it follows from (13.6) that all roots of NP in
C+ must also be roots of N , and vice versa. Thus (iii c) holds.
(iii) = (i) Suppose that R has c.f.r. (N, D) satisfying the conditions of (iii). Let F be
and NP = F N
P . Then define P1 = DN
P . Note
the GCD of N and NP and write N = F N
that by (iii a) and (iii c), P1 is Hurwitz. Now let P2 be an arbitrary Hurwitz polynomial
having the property that deg(P1 P2 ) = 2n 1, where n = deg(DP ). By Corollary 13.5 we
may find polynomials Dff and Nfb so that
Dff DP + Nfb NP = P1 P2 .

(13.7)

P2 . We now claim that if we take Rff =

We also take Nff = Dfb = N
and Rfb =
then
the interconnection of Figure 13.14 is IBIBO stable. The relevant transfer functions that
must be checked as belonging to RH+
are
Nff
Dff

RP Rff Rfb
,
1 + RP Rff Rfb
RP Rfb
T3 =
,
1 + RP Rff Rfb
RP
T5 =
,
1 + RP Rff Rfb
T1 =

Nfb
,
Dfb

RP Rff
1 + RP Rff Rfb
Rff Rfb
T4 =
1 + RP Rff Rfb
Rff
T6 =
1 + RP Rff Rfb
Rfb
T7 =
.
1 + RP Rff Rfb
T2 =

NEED T3 and T7
To see that all of the transfer functions are have no poles in C+ first note that they can
all be written as rational functions with denominator
P2 (Dff DP + Nfb NP ) = N
P1 P 2 .
Dff Dfb DP + Nff Nfb NP = N
2

(13.8)

is Hurwitz by (iii c). Thus

P2 is Hurwitz by design, P1 is Hurwitz by (iii a) and (iii c), and N
all transfer functions T1 and T7 are analytic in C+ .
Let us now give the numerator for each of the transfer functions T1 through T7 when put
over the denominator (13.8), and use this to ascertain that these transfer functions are all

13.3 Two controller configurations

22/10/2004

519

520

P2 ), which we now obtain.

proper. In doing this, it will be helpful to have a bound on deg(N
We suppose that deg(P1 ) = k so that deg(P2 ) = 2n k 1. Now,

13 Advanced synthesis, including PID synthesis

P2 this gives
6. T6 : The numerator is DP Dfb Nff . Since Nff = N
T6 =

P )
deg(P1 ) = deg(D) + deg(N
= deg(D) + deg(NP ) deg(F )
deg(DP ) + deg(N ) deg(F )
),
= n + deg(N

7. T7 : The numerator is DP Dff Nfb , giving

T7 =

Finally we check that the closed-loop transfer function is R. Indeed, the closed-loop
transfer function is T2 and we have

P2 this gives
1. T1 : The numerator is NP Nff Nfb . Since Nff = N
NP Nfb
.
P1 P 2

NP Dfb
.
P1 P 2

13.22 Remarks 1. The conditions R, RRP RH+

are actually the weakest possible for a
class of interconnections more general than that of Figure 13.14. Indeed, if (S, G) is
any interconnected SISO linear system with the property that every forward path from
the reference r to the output y passes through the plant, then one readily sees that the
transfer function from the reference r to the input u to the plant is exactly RRP (cf. the
proof that (i) implies (ii) in Theorem 13.21 below). Thus we have the following statement:

(13.10)

Again, deg(P1 P2 ) = 2n 1 and deg(NP ) n. In (13.9) we showed that deg(Dfb )

n 1, thus T2 is proper.
3. T3 : The numerator of T3 is NP Nfb Dff , giving
T3 =

If (S, G) is an interconnected SISO linear system with the property that every
forward path from the input to the output passes through the plant, then the
interconnection is IBIBO stable with transfer function R R(s) only if R, RRP
RH+
.

NP Nfb Dff
P1 P22
N

We have deg(NP ) n and deg(P1 P2 ) = 2n 1. Since Dff and Nfb satisfy (13.7),
deg(Dff ), deg(Nfb ) n 1.
P2 this gives
4. T4 : The numerator is DP Nff Nfb . Since Nff = N
T4 =

DP Nfb
.
P1 P 2

We have deg(DP ) = n and deg(P1 P2 ) = 2n 1. Since Nfb satisfies (13.7), deg(Nfb )

n 1. This shows that T4 is proper.
P2 this gives
5. T5 : The numerator is NP Dff Dfb . Since Dfb = N
T5 =

NP Dff
.
P1 P 2

We have deg(DP ) = n and deg(P1 P2 ) = 2n 1. Since Dff satisfies (13.7), deg(Dff )

n 1. This shows that T5 is proper.

NP
NP P2
N
N
N
NP Dfb
=
=
=
= T2 ,

D
P1 P2
P 1 P2
D NP

using (13.10).
Thus we have shown that the interconnection of Figure 13.14 is IBIBO stable with transfer
function R. This means that R is implementable.

We have deg(P1 P2 ) = 2n 1 and deg(NP ) n. Since Nfb is obtained from (13.7),

deg(Tfb ) n 1. From this we deduce that T1 is proper.
P2 this gives
2. T2 : The numerator of T2 is NP Nff Dfb . Since Nff = N
T2 =

DP Dff Nfb
P1 P22
N

We have deg(DP ) = n and deg(P1 P2 ) = 2n 1. Since Dff and Nfb satisfy (13.7) we
have deg(Dff ), deg(Nfb ) n 1.

(13.9)

Now we proceed with our calculations.

T1 =

DP Dfb
.
P 1 P2

We have deg(DP ) = n and deg(P1 P2 ) = 2n 1. By (13.9) deg(Dfb ) n 1, showing

that T6 is proper.

using (iii b). Therefore

)
deg(P2 ) = 2n 1 deg(P1 ) 2n 1 n deg(N

deg(N P2 ) n 1.

22/10/2004

finish

(This is Exercise E6.14.) This indicates that the conditions R, RRP RH+
are the weakest one can impose on a closed-loop transfer function if it is to be realisable by some
reasonable interconnection. The additional hypothesis that R have no nonminimum
phase zeros other than those of RP is reasonable: the nonminimum phase zeros of RP
must appear in R, and we would not want any more of these than necessary, given the
discussions of Chapters 8 and 9 concerning the effects of nonminimum phase zeros.
2. As with our results of Section 13.2, Theorem 13.21 is constructive.
3. Note that the interconnection of Figure 13.14 is the same as that of Figure 10.9. Indeed,
there is a relationship between the controllers constructed in Theorem 13.21 and the
combination of the observer combined with static state feedback in Theorem 10.48. The
procedure of Theorem 13.21 is a bit more flexible in that the order of the closed-loop
characteristic polynomial is not necessarily 2n.
Let us first illustrate via an example that implementability is indeed different that stabilisation of the closed-loop system.

13.4 Synthesis using controller parameterisation

22/10/2004

13.23 Example Let RP (s) = ss2

2 1 . One can verify that R(s) =
Indeed, clearly R RH+
and we also have

2s
s2 +2s+2

521
is implementable.

R(s) =

1. 6=

RC (s)RP (s)
,
1 + RC (s)RP (s)

R
1 s2
=
.
(R 1)RP
s(s + 3)

Note that although R is implementable, the interconnection of Figure 13.8 is not IBIBO
stable. Thus we see that implementability is different from an IBIBO interconnection of the
form Figure 13.8.

13.3.2 Implementations that meet design considerations As we saw in Section 13.2.2, one often wants to enforce more than IBIBO stability on ones feedback loop.
Let us see how these considerations can be enforced in the two controller configuration of
Figure 13.14.

13.4 Synthesis using controller parameterisation

In Section 10.3 we described the set of proper controllers that stabilise a proper plant
RP , and the parameterisation came to us in terms of a free function in RH+
. In this section
we turn to using this parameterisation to provide a useful guide to controller design. We
will be concerned with the standard plant/controller unity feedback loop, and we reproduce
this in Figure 13.15 for easy reference.

2
, and
P1

We wish to determine how the choice of the parameter affects the properties of the closedloop system. For this purpose, let us suppose that we decide that we wish to place all
poles of the closed-loop system in some region Cdes in the complex plane. Note that we
can always do this, since by, for example, Theorem 10.27 or Theorem 13.2, we may construct controllers that achieve any closed-loop characteristic polynomial, provided it has
sufficiently high degree (twice the order of the plant for a strictly proper controller, and one
less than this for a proper controller). Thus we are permitted to talk about the set of all
proper controllers for which the closed-loop poles lie in Cdes , and let us denote this set of
controllers by Spr (RP , Cdes ). The following result describes these controllers using the Youla
parameterisation.
13.24 Proposition Let RP be a proper plant and Cdes C be the set of desirable closed-loop
pole locations. Also let (P1 , P2 ) be a coprime fractional representative for RP with (1 , 2 )
a coprime factorisation of (P1 , P2 ). Then

1 + P2
Spr (RP , Cdes ) =

admissible
and

has
all
poles
in
C
.
des
2 P1
Proof

The result is perhaps surprisingly obvious, at least in statement, and clearly provides
a useful tool for controller design.

13.5 Summary
1. Ziegler-Nichols tuning is available for doing PID control design.

r(s)

RC (s)

RP (s)

22/10/2004

we can solve this for RC to get

RC (s) =

13 Advanced synthesis, including PID synthesis

2. lim 2 (s) (s)P1 (s) 6= 0.

R(s)
1 s2
= 2
RH+
.
RP (s)
s + 2s + 2
Since we have

522

y(s)

Figure 13.15 Interconnection for studying controller parameterisation

13.4.1 Properties of the Youla parameterisation We first recall the Youla parameterisation of Theorem 10.37. Given a proper plant RP , the set Spr (RP ) of proper controllers
that render the interconnection of Figure 13.15 IBIBO stable is given by

1 + P2
Spr (RP ) =
admissible .

2 P1
Recall that (P1 , P2 ) is a coprime fractional representative of RP and (1 , 2 ) is a coprime
factorisation of (P1 , P2 ). Also recall that the collection of admissible functions are defined
by

Exercises for Chapter 13

523

524

13 Advanced synthesis, including PID synthesis

Exercises

r(s)

Rff (s)

First, the first method.

(a) Produce the step response numerically using Proposition 3.40.
(b) Graphically (or with the computer, if you can) determine the values of and ,
and so determine the parameters K, TI , and TD for the three cases of a P, PI,
and PID controller.
(c) Using Proposition 3.40, determine the closed-loop step response.
(d) Comment on the performance of the three controllers.
Now the second method.
(e) Verify that Assumption 13.1 is satisfied for the plant transfer function RP .
Hint: A good way to do this might be to use a computer package to solve the
equation obtained by setting the real part of the poles to zero.
(f) Determine by trial and error (or with the computer, if you can) the value Ku
of the proportional gain that makes two poles of the transfer function negative,
and the period of the oscillatory part of the step response. Use these numbers
to give the corresponding Ziegler-Nicols values for the parameters K, TI , and
TD for the three cases of a P, PI, and PID controller.
(g) Using Proposition 3.40, determine the closed-loop step response.
(h) Comment on the performance of the three controllers.
E13.3 Exercises on controlling state examples using pole placement on output.
E13.4 Let RP be a proper plant and let R I (RP ). Suppose that the feedback and
feedforward controllers giving R as the closed-loop transfer function in Figure 13.14
are Rfb and Rff , respectively. Let (N, D), (Nfb , Dfb ), and (Nff , Dff ) be the c.f.r.s for
R, Rfb , and Rff , respectively.
(a) Show that all interconnections of Figure E13.1 have closed-loop transfer function
R. Note that the recipe of Theorem 13.21 gives Nff = Dfb , so let us assume that
this is the case.
(b) Show that the interconnection (a) of Figure E13.1 will not generally be IBIBO
stable.
(c) Show that the interconnection (b) of Figure E13.1 will not generally be IBIBO
stable.

y(s)

E13.1 In this exercise, you will show that the Ziegler-Nicols tuning method cannot be
applied, even in some quite simple cases.
a
(a) Show that for any first-order plant, say RP = s+b
, both of the Ziegler-Nicols
methods will not yield PID parameter values.
a
(b) Show that there exists a second-order plant, say of the form RP (s) = s2 +bs+c
, so
that Assumption 13.1 will not be satisfied.
E13.2 In this exercise, you will apply the Ziegler-Nicols tuning method to design PID
controllers for the plant transfer function
1
RP (s) = 3
.
s + s2 + 2s + 1

RP (s)

22/10/2004

Nfb (s)
Dff (s)

(a)

r(s)

Nff (s)
Nfb (s)

Nfb (s)
Dff (s)

RP (s)

y(s)

Nfb (s)
Dff (s)

(b)

Figure E13.1 Alternate two controller configuration implementations

526

This version: 22/10/2004

14 An introduction to H2 optimal control

22/10/2004

14.4.1 LQR and pole placement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 550

14.4.2 Frequency domain interpretations . . . . . . . . . . . . . . . . . . . . . . . . . . . 550
14.4.3 H2 model matching and LQG . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 550

Chapter 14

14.5 Stability margins for optimal feedback . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 550

14.5.1 Stability margins for LQR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 551

An introduction to H2 optimal control

14.5.2 Stability margins for LQG . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 553

14.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 553

The topic of optimal control is a classical one in control, and in some circles, mainly
mathematical ones, optimal control is synonymous with control. That we have covered
as much material as we have without mention of optimal control is hopefully ample demonstration that optimal control is a subdiscipline, albeit an important one, of control. One
of the useful features of optimal control is that is guides one is designing controllers in a
systematic way. Much of the control design we have done so far has been somewhat ad hoc.
What we shall cover in this chapter is a very bare introduction to the subject of optimal
control, or more correctly, linear quadratic regulator theory. This subject is dealt with
thoroughly in a number of texts. Classical texts are [Brockett 1970] and [Bryson and Ho
1975]. A recent mathematical treatment is contained in the book [Sontag 1998]. Design issues
are treated in [Goodwin, Graebe, and Salgado 2001]. Our approach closely follows that of
Brockett. We also provide a treatment of optimal estimator design. Here a standard text
is [Bryson and Ho 1975]. More recent treatments include [Goodwin, Graebe, and Salgado
2001] and [Davis 2002]. We have given this chapter a somewhat pretentious title involving
H2 . The reason for this is the frequency domain interpretation of our optimal control
problem in Section 14.4.2.

14.1 Problems in optimal control and optimal state estimation

We begin by clearly stating the problems we consider in this chapter. The problems come
in two flavours, optimal feedback and optimal estimation. While the problem formulations
have a distinct flavour, we shall see that their solutions are intimately connected.
14.1.1 Optimal feedback We shall start with a SISO linear system = (A, b, ct , 01 )
which we assume to be in controller canonical form, and which we suppose to be observable.
As was discussed in Section 2.5.1, following Theorem 2.37, this means that we are essentially
studying the system
x(n) (t) + pn1 x(n1) + . . . p1 x(1) (t) + p0 x(t) = u(t)
y(t) = cn1 x(n1) (t) + cn2 x(n2) (t) + + c1 x(1) (t) + c0 x(t),

(14.1)

where x is the first component of the state vector x,

PA () = n + pn1 n1 + + p1 + p0

Contents
14.1 Problems in optimal control and optimal state estimation . . . . . . . . . . . . . . . . . . 526
14.1.1 Optimal feedback . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 526
14.1.2 Optimal state estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 528

is the characteristic polynomial for A, and c = (c0 , c1 , . . . , cn1 ). Recall that U denotes
the set of admissible inputs. For u U defined on an interval I I and for x0 =
(x00 , x01 , . . . , x0,n1 ) R, the solution for (14.1) corresponding to u and x0 is the unique
map xu,x0 : I R that satisfies the first of equations (14.1) and for which

14.2 Tools for H2 optimisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 530

x(0) = x00 , x(1) (0) = x01 , , . . . , x(n1) (0) = x0,n1 .

14.2.1 An additive factorisation for rational functions . . . . . . . . . . . . . . . . . . . . 531

14.2.2 The inner-outer factorisation of a rational function

. . . . . . . . . . . . . . . . . 531

14.2.3 Spectral factorisation for polynomials . . . . . . . . . . . . . . . . . . . . . . . . . 532

14.2.4 Spectral factorisation for rational functions . . . . . . . . . . . . . . . . . . . . . . 535
14.2.5 A class of path independent integrals . . . . . . . . . . . . . . . . . . . . . . . . . 536
14.2.6 H2 model matching . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 539
14.3 Solutions of optimal control and state estimation problems . . . . . . . . . . . . . . . . . 541

The corresponding output, defined by the second of equations (14.1), we denote by yu,x0 : I
R. We fix x0 Rn and define Ux0 to be the set of admissible inputs u U for which
1. u(t) is defined on either
(a) [0, T ] for some T > 0 or
(b) [0, ),
and

14.3.1 Optimal control results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 541

2. if yu,x0 is the output corresponding to u and x0 , then

14.3.2 Relationship with the Riccati equation . . . . . . . . . . . . . . . . . . . . . . . . 544

(a) yu,x0 (T ) = 0 if u is defined on [0, T ] or

(b) limt yu,x0 (t) = 0 if u is defined on [0, ).

14.3.3 Optimal state estimation results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 548

14.4 The linear quadratic Gaussian controller . . . . . . . . . . . . . . . . . . . . . . . . . . . 550

22/10/2004

14.1 Problems in optimal control and optimal state estimation

527

Thus Ux0 are those controls that, in possibly infinite time, drive the output from y = y(0)
to y = 0. For u Ux0 define the cost function
Z

2
(t) + u2 (t) dt.
yu,x
Jx0 (u) =
0
0

Thus we assign cost on the basis of equally penalising large inputs and large outputs, but
we do not penalise the state, except as it is related to the output.
With this in hand we can state the following optimal control problem.

528

14 An introduction to H2 optimal control

22/10/2004

let us do this for the sake of context. We let = (A, B, C, 0) be a MIMO linear system
satisfying the equations

x(t)
= Ax(t) + Bu(t)
y(t) = Cx(t).
The natural formulation of a cost for the system is given directly in terms of states and
outputs, rather than inputs and outputs. Thus we take as cost
Z

J=
xt (t)Qx(t) + ut (t)Ru(t) dt,
0

14.1 Optimal control problem in terms of output Seek a control u Ux0 that has the property
that Jx0 (u) Jx0 (
u) for any u Ux0 . Such a control u is called an optimal control law .

14.2 Remark Note that the output going to zero does not imply that the state also goes to
zero, unless we impose additional assumptions of (A, c). For example, if (A, c) is detectable,
then by driving the output to zero as t , we can also ensure that the state is driven to
zero as t .

Now let us formulate another optimal control problem that is clearly related to Problem 14.1, but is not obviously the same. In Section 14.3 we shall see that the two problems
are, in actuality, the same. To state this problem, we proceed as above, except that we seek
our control law in the form of state feedback. Thus we now work with the state equations

x(t)
= Ax(t) + bu(t),
and we no longer make the assumption that (A, b) is in controller canonical form, but we
do ask that be complete. We recall that Ss () Rn is the subset of vectors f for which
A bf t is Hurwitz. If x0 Rn , we let xf ,x0 be the solution to the initial value problem

x(t)
= (A bf t )x(t),

x(0) = x0 .

Thus xf ,x0 is the solution for the state equations under state feedback via f with initial
condition x0 . We define the cost function
Z

Jc,x0 (f ) =
xtf ,x0 (t)(cct + f f t )xf ,x0 (t) dt.

where Q R
and R R
are symmetric, and typically positive-definite. The positivedefiniteness of R is required for the problem to be nice, but that Q can be taken to be only
positive-semidefinite. The solution to this problem will be discussed, but not proved, in

Section 14.3.2.
14.1.2 Optimal state estimation Next we consider a problem that is dual to that
considered in the preceding section. This is the problem of constructing a Luenberger observer that is optimal in some sense. The problem we state in this section is what is known
as the deterministic version of the problem. There is a stochastic version of the problem
statement that we do not consider. However, the stochastic version is very common, and is
indeed where the problem originated in the work of Kalman [1960] and Kalman and Bucy
[1960]. The stochastic problem can be found, for example, in the recent text of Davis [2002].
Our deterministic approach follows roughly that of Goodwin, Graebe, and Salgado [2001].
The version of the problem we formulate for solution in Section 14.3.3 is not quite the typical formulation. We have modified the typical formulation to be compatible with the SISO
version of the optimal control problem of the preceding section.
Our development benefits from an appropriate view of what a state estimator does.
We let = (A, b, ct , 01 ) be a complete SISO linear system, and recall that the equations
, the output y, and the estimated output y are
governing the state x, the estimated x
(t) = A
x
x(t) + bu(t) + `(y(t) y(t))
(t)
y(t) = ct x

x(t)
= Ax(t) + bu(t)
y(t) = ct x(t).

Note that this is the same as the cost function Jx0 (u) if we take u(t) = f t xf ,x0 (t) and
yu,x0 (t) = ct xf ,x0 (t). Thus there is clearly a strong relationship between the cost functions
for the two optimal control problems we are considering.
In any case, for c Rn and x0 Rn , we then define the following optimal control
problem.
14.3 Optimal control problem in terms of state Seek f Ss () so that Jc,x0 (f ) Jc,x0 (f )
for every f Ss (). Such a feedback vector f is called an optimal state feedback
vector .

14.4 Remark Our formulation of the two optimal control problems is adapted to our SISO
setting. More commonly, and more naturally, the problem is cast in a MIMO setting, and

and output estimate y can be rewritten as

The equations for the state estimate x
(t) = (A `ct )
x
x(t) + bu(t) + `y(t)
(t).
y(t) = ct x

(14.2)

The form of this equation is essential to our point of view, as it casts the relationship between
the output y and the estimated output y as the input/output relationship for the SISO linear
system ` = (A `ct , `, ct , 01 ). The objective is to choose ` in such a way that the state
tends to zero as t . If the system is observable this is equivalent to
error e = x x
asking that the output error also i = y y has the property that limt i(t) = 0. Thus we
wish to include as part of the cost of a observer gain vector ` a measure of the output
error. We do this in a rather particular way, based on the following result.

22/10/2004

14.1 Problems in optimal control and optimal state estimation

529

14.5 Proposition Let = (A, b, ct , 01 ) be a complete SISO linear system. If the output error
corresponding to the initial value problem
(t) = A
(0) = e0
x
x(t) + `(y(t) y(t)),
x
(t)
y(t) = ct x

x(t)
= Ax(t) + b(t),
x(0) = 0
y(t) = ct x(t)
satisfies limt i(t) = 0 then the output error has this same property for any input, and for
any initial condition for the state and estimated state.
Proof Since the system is observable, by Lemma 10.44 the state error estimate tends to
zero as t for all inputs and all state and estimated state initial conditions if and only
if ` D(). Thus we only need to show that by taking u(t) = 0 and y(t) = h (t) in (14.2),
the output error i will tend to zero for all error initial conditions if only if ` D(). The
output error for the equations given in the statement of the proposition are

(t)
e(t)
= x(t)
x
(t) + `h (t)
= Ax(t) A
x(t) + `ct x
t
= (A `c )e(t),
for t > 0, where we have used the fact that x(t) = h (t) = ct eAt b. From this the proposition
immediately follows.

What this result allows us to do is specialise to the case when we consider the particular
output that is the impulse response. That is, one portion of the cost of the state estimator
will be the penalise the difference between the input to the system (14.2) and its output,
when the input is the impulse response for . We also wish to include in our cost a penalty
for the size of the estimator. This penalty we take to be the L2 -norm of the impulse
response for the system ` . This gives the total cost for the observer gain vector ` to be
Z
Z t
Z
2
J(`) =
h (t)
h` (t )h ( ) d dt +
h` (t)2 dt.
(14.3)
0

With the above as motivation, we pose the following problem.

14.6 Optimal estimator problem Seek ` Rn so that J(`) J(`) for every ` D(). Such
an observer gain vector ` is called an optimal observer gain vector .

14.7 Remark The cost function for the optimal control problems of the preceding section
seem somehow more natural than the cost (14.3) for the state estimator `. This can be
explained by the fact that we are formulating a problem for an estimator in terms of what
is really a filter. We do not explain here the distinction between these concepts, as these
are best explained in the context of stochastic control. For details we refer to [Kamen and
Su 1999]. We merely notice that when one penalises a filter as if it were an estimator,
one deals with quantities like impulse response, rather than with more tangible things like
inputs and outputs that we see in the optimal control problems. One can also formulate
more palatable versions cost for an estimator that make more intuitive sense than (14.3).
However, the result will not then be the famous Kalman-Bucy filter that we come up with
in Section 14.3.3. For a description of natural estimator costs, we refer to the deterministic

530

14 An introduction to H2 optimal control

22/10/2004

Kalman filter described by Sontag [1998]. We also mention that the problem looks more
natural, not in the time-domain, but in the frequency domain, where it relates to the model

matching problem that we will discuss in Section 14.2.6.
14.8 Remark Our deterministic development of the cost function for an optimal state estimator differs enough from that one normally sees that it is worth quickly presenting the
normal formulation for completeness. There are two essential restrictions we made in our
above derivation that need not be made in the general context. These are (1) the use of
output rather than state error and (2) the use of the impulse response for as the input
to the estimator (14.2). The most natural setting for the general problem, as was the case
for the optimal control setting of Remark 14.4, in MIMO. Thus we consider a MIMO linear
system = (A, B, C, 0). The state can then be estimated, just as in the SISO case, via a
Luenberger observer defined by an observer gain vector L Rmr . To define the analogue
of the cost (14.3) in the MIMO setting we need some notation. Let S R be symmetric
and positive-definite. The S-Frobenius norm of a matrix M C is given by
1/2
kM kS = tr(M SM )
.
With this notation, the cost associated to an observer gain vector L is given by
Z
Z
At t

t (At C t Lt )t
e eAt t C t Lt e(At C t Lt )t dt +
L e
dt,
J(L) =

(14.4)

for symmetric, positive-definite matrices Rnn and Rrr . While we wont discuss
this in detail, one can easily see that (14.4) is the natural analogue of (14.3) after one removes
the restrictions we made to render the problem compatible with our SISO setting. We shall

present, but not prove, the solution of this problem in Section 14.3.3.

14.2 Tools for H2 optimisation

Before we can solve the optimal feedback and the optimal state estimation problems
posed in the preceding section, we must collect some tools for the job. The first few sections
deal with factorisation of rational functions. Factorisation plays an important role in control
synthesis for linear systems. While in the SISO context of this book these topics are rather
elementary, things are not so straightforward in the MIMO setting; indeed, significant effort
has been expended in this direction. The recent paper of Oara and Varga [2000] indicates
tens of papers have been dedicated to this since the original work of Youla [1961]. An
approach for controller synthesis based upon factorisation is given by Vidyasagar [1987].
In this book, factorisation will appear in the current chapter in our discussion of optimal
control, as well as in Chapter 15 when we talk about synthesising controllers that solve
the robust performance problem. Indeed, only the results of Section 14.2.3 will be used in
this chapter. The remaining result will have to wait until Chapter 15 for their application.
However, it is perhaps in the best interests of organisation to have all the factorisation results
in one place. In Section 14.2.5 we consider a class of path independent integrals. These will
allow us to simplify the optimisation problem, and make its solution obvious. Finally, in
Section 14.2.6 we describe H2 model matching. The subject of model matching will be
revisited in Chapter 15 in the H context.
For the most part, this background is peripheral to applying the results of Sections 14.3.
Thus the reader looking for the shortest route to the point can skip ahead to Section 14.3
after understanding how to compute the spectral factorisation of a polynomial.

14.2 Tools for H2 optimisation

22/10/2004

531

14.2.1 An additive factorisation for rational functions Our first factorisation is

a simple one for rational functions. The reader will wish to recall some of our notation
+
for system norms from Section 5.3.2. In that section, RH+
2 and RH denoted the strictly
proper and proper, respectively, rational functions with no poles in C+ . Also recall that
RL denotes the proper rational functions with no poles on the imaginary axis. Given a
rational function R R(s) let us denote R R(s) as the rational function R (s) = R(s).
With this notation, let us additionally define

+

RH
2 = R R(s) | R RH2

+

RH
= R R(s) | R RH .

Thus RH
2 and RH denote the strictly proper and proper, respectively, rational functions
with no poles in C . Now we make a decomposition using this notation.

14.9 Proposition If R RL then there exists unique rational functions R1 , R2 R(s) so

that
(i) R = R1 + R2 ,
(ii) R1 RH
2 , and
(iii) R2 RH+
.
Proof For existence, suppose that we have produced the partial fraction expansion of R.
Since R has no poles on iR, by Theorem C.6 it follows that the partial fraction expansion
will have the form
1 + R
2 + C
R=R

+

1 and R2 = R
2 +C gives existence
where R1 RH2 , R2 RH2 , and C R. Defining R1 = R

1 RH
of the stated factorisation. For uniqueness, suppose that R = R1 + R2 for R
2 and
2 RH+
1 = R
1
R
.
Then,
uniqueness
of
the
partial
fraction
expansion
guarantees
that R

2 = R
2 + C.
and R

Note that the proof of the result is constructive; it merely asks that we produce the
partial fraction expansion.

532

14 An introduction to H2 optimal control

22/10/2004

Thus the last 12 (n `0 ) zeros are the complex conjugates of the 12 (n `0 ) preceding zeros.
For the remainder of the proof, let us denote m = 12 (n `0 ). If we then define
Rin =

k
Y
(s zj )`j
j=1

(s + zj )`j

Rout =

R
,
Rin

(14.5)

then clearly Rin and Rout satisfy (i), (ii), and (iii). To see that these are unique, suppose that
in and R
out are two inner and outer functions having the properties that R = R
in R
out and
R
1
in (0) = 1. Since R
out
in .
R
is analytic in C+ , if z C+ is a zero for R it must be a zero for R
Thus
k
Y
in =
R
(s zj )`j T (s)
j=1

in is inner we have
for some T R(s). Because R
1=

k
k
Y
Y
(s zj )`j T (s) (1)`j (s + zj )`j T (s)
j=1

j=1

= T (s)T (s)(1)`0

k
k
Y
Y
(s zj )`j
(s + zj )`j
j=1

j=1

which gives
1
Qk

T (s)T (s) = (1)`0 Qk

`j
j=1 (s zj )

j=1 (s

+ zj )`j

From this we conclude that either

T (s) = Qk

j=1 (s

)`j

or T (s) = Qk

j=1 (s

+ zj )`j

`0
in RH+

The facts that R

ensures that
and Rin (0) = (1)

T (s) = Qk

14.2.2 The inner-outer factorisation of a rational function A rational function

1
R R(s) is inner if R RH+
is
and if RR = 1, and is outer if R RH and if R
analytic in C+ . Note that outer rational functions are exactly those that are proper, BIBO
stable, and minimum phase. The following simple result tells the story about inner and
outer rational functions as they concern us.

and from this the follows uniqueness.

14.10 Proposition If R RH+

then there exists unique Rin , Rout R(s) so that
(i) R = Rin Rout ,
(ii) Rin is inner and Rout is outer, and
(iii) Rin (0) = (1)`0 , where `0 0 is the multiplicity of the root s = 0 for R.

14.2.3 Spectral factorisation for polynomials Spectral factorisation, while still quite
simple in principle, requires a little more attention. Let us first look at the factorisation of
a polynomial, as this is easily carried out.

Proof Let z0 , . . . , zk C+ be the collection of distinct zeros of R in C+ with each having

multiplicity `j , j = 0, . . . , k. Suppose that the zeros are ordered so that
(
0,
j=0
zj =
zj+ 1 (n+`0 ) = zj+`0 , j = 1, . . . , 21 (n `0 ).

14.11 Definition A polynomial P R[s] admits a spectral factorisation if there exists

a polynomial Q R[s] with the properties
(i) P = QQ and
(ii) all roots of Q lie in C .
If P satisfies these two conditions, then P admits a spectral factorisation by Q.

j=1 (s

+ zj )`j

Note that the proof is constructive; indeed, the inner and outer factor whose existence is
declared are given explicitly in (14.5).

14.2 Tools for H2 optimisation

22/10/2004

533

Let us now classify those polynomials that admit a spectral factorisation. A polynomial
P R[s] is even if P (s) = P (s). Thus an even polynomial will have nonzero coefficients
only for even powers of s, and in consequence the polynomial will be R-valued on iR.
14.12 Proposition Let P R[s] have a zero at s = 0 of multiplicity k0 0. Then P admits
a spectral factorisation if and only if
(i) P is even and

(ii) if k20 is odd then P (i) 0 for R, and if k20 is even then P (i) 0 for R.
Furthermore, if P admits a spectral factorisation by Q then Q is uniquely defined by requiring that the coefficient of the highest power of s in Q be positive.
Proof First suppose that P satisfies (i) and (ii). If z is a root for P , then so will z be a
root since P is even. Since P is real, z and hence
z are also roots. Thus, if we factor P it
will have the factored form
P (s) = A0 s2k0

k1
Y

(a2j1 s2 )

k2
Y

(s2 + b2j2 )

j2 =1

j1 =1

k3
Y

(14.6)

k1
k3
Y
Y
p

A0 s k 0
(s + aj1 )
(s + j3 )2 + j23
j3 =1

satisfies the conditions of the definition of a spectral factorisation.

Now suppose that P admits a spectral factorisation by Q R[s]. Since Q must have all
roots in C , we may write
k1
Y
j1 =1

22/10/2004

Proof Suppose that

P (s) = pn sn + + p1 s + p0 .

Since P P is even, if z is a root, so is z. Thus if z is a root, so is z, and therefore

z.
Arguing as in the proof of Proposition 14.12 this gives
P (s)P (s) = p2n s2k0

k1
Y

(a2j1 s2 )

j1 =1

ks
Y

(s2 j22 )2 + 2j22 (s2 + j22 ) + j42 ,

j2 =1

where 2k0 + 2k1 + 2k2 + 4k3 = 2 deg(P ). Here aj1 > 0, j1 = 1, . . . , k1 , j2 0, j2 = 1, . . . , k2 ,

and j2 > 0, j2 = 1, . . . , k3 . This form for P P is a consequence of all imaginary axis roots
necessarily having even multiplicity. It is now evident that
Q(s) = |pn | sk0

k1
Y

(s + aj1 )

k2
Y

(s + j2 )2 + j22

j2 =1

is a spectral factor. The final assertion of the corollary is clear.

j3 =1

j1 =1

Q(s) = B0 sk0

14 An introduction to H2 optimal control

j1 =1

(s2 j23 )2 + 2j23 (s2 + j23 ) + j43 .

where 2k0 + 2k1 + 2k2 + 4k3 = 2n = deg(P ). Here aj1 > 0, j1 = 1, . . . , k1 , bj2 > 0,
j2 = 1, . . . , k2 , j3 0, j3 = 1, . . . , k3 , and j3 > 0, j3 = 1, . . . , k3 . One may check that
condition (ii) implies that A0 > 0. By (ii), P must not change sign on the imaginary axis.
This implies that all nonzero imaginary roots must have even multiplicity, since otherwise,
terms like (s2 + b2j2 ) will change sign on the imaginary axis. Therefore we may suppose that
k2 = 0 as the purely imaginary roots are then captured by the third product in (14.6). Now
we can see that taking
Q(s) =

534

(s + aj1 )

k2
Y
j2 =1

(s2 + b2j2 )

k3
Y

Let us extract the essence of Proposition 14.12.

14.14 Remark The proof of Proposition 14.12 is constructive. To determine the spectral
factorisation, one proceeds as follows.
1. Let P be an even polynomial for which P (i) does not change sign for R. Let A0
be the coefficient of the highest power of s. Divide P by |A0 | so that the resulting polynomial is monic, or 1 times a monic polynomial. Rename the resulting polynomial
P .
2. Compute the roots for P that will then fall into one of the following categories:
(a) k0 roots at s = 0;
(b) k1 roots s = aj1 for aj1 real and positive (and the associated k1 roots s = aj1 );
(c) k2 roots s = j2 + ij2 for j2 and j2 real and nonnegative (and the associated
3k1 roots s = j2 ij2 , s = j2 + ij2 , and s = j2 ij2 ).
3. A left half-plane spectral factor for P will then be

((s + j3 )2 + j23 ),

j3 =1

where aj1 > 0, j1 = 1, . . . , k1 , bj2 > 0, j2 = 1, . . . , k2 , j3 , j3 > 0, j3 = 1, . . . , k3 . Computing

QQ gives exactly the expression (14.6) with A0 = B02 , thus showing that P satisfies (i)
and (ii).
Uniqueness up to sign of the spectral factorisation follows directly from the above computations.

We denote the polynomial specified by the theorem by [P ]+ which we call the left halfplane spectral factor for P . The polynomial [P ] = Q we call the right half-plane
spectral factor of P .
The following result gives a common example of when a spectral factorisation arises.
14.13 Corollary If P R[s] then P P admits a spectral factorisation. Furthermore, if P
has no roots on the imaginary axis, then the left half-plane spectral factor will be Hurwitz.

Q(s)
=

k1
Y
j1 =1

(s + aj1 )

k2
Y

(s + j2 )2 + j22 .

j2 =1

4. A left half-plane spectral factor for P is then Q =

|A0 |Q.

Lets compute the spectral factors for a few even polynomials.

14.15 Examples 1. Let us look at the easy case where P (s) = s2 + 1. This polynomial
is certainly
even. However, note that it changes sign on the imaginary axis. Indeed,

P ( 2i) = 1 and P (0) = 1. This demonstrates why nonzero roots along the imaginary
axis should appear with even multiplicity if a spectral factorisation is to be admitted.
2. The previous example can be fixed by considering instead P (s) = s4 + 2s2 + 1 =
(s2 + 1)2 . This polynomial has roots {i, i, i, i}. The left half-plane spectral factor is
then [P (s)]+ = s2 + 1, and this is also the right half-plane spectral factor in this case.

22/10/2004

14.2 Tools for H2 optimisation

535

3. Let P (s) = s2 + 1. Clearly P is even and is nonnegative on the imaginary axis. The
roots of P are {1, 1}. Thus the left half-plane spectral factor is [P (s)]+ = s + 1.
4. Let P (s) =s4 s2 +
1. Obviously
P is even and nonnegative on iR. This polynomial

has roots 23 + i 12 , 23 i 12 , 23 + i 21 , 23 i 21 . Thus we have
[P (s)]+ = (s +

3 2
)
2

1
4

536

14.16 Definition A rational function R R(s) admits a spectral factorisation if there

exists a rational function R(s) with the properties
(i) R = ,
(ii) all zeros and poles of lie in C .
If satisfies these two conditions, then R admits a spectral factorisation by .

n
X

Mij (i) (t)(j) (t) dt = F (a), (b), (1) (a), (1) (b), . . . , (n1) (a), (n1) (b) , (14.7)

a i,j=0

where Mij , i, j = 0, . . . , n, are the components of the matrix M . Thus the matrix M is
integrable when the integral in (14.7) depends only on the values of and its derivatives at
the endpoints, and not on the values of on the interval (a, b). As an example, consider the
two symmetric 2 2 matrices

0 1
1 0
M1 =
, M2 =
.
1 0
0 0
The first isintegrable on [a, b] since

Let us now classify those rational functions that admit a spectral factorisation.
14.17 Proposition Let R R(s) be a rational function with (N, D) its c.f.r. R admits a
spectral factorisation if and only if both N and D admits a spectral factorisation.
Proof Suppose that N and D admit a spectral factorisation by [N ]+ and [D]+ , respectively.
Then we have
[N ]+ [N ]
R=
.
[D]+ [D]

2
X

N N
.
D D

Since N and D are coprime, so are N N and D D . Since D D is also monic, it follows
that the c.f.r. of R is (N N , D D ). Thus the c.f.r.s of R admits spectral factorisations.
Following our notation for the spectral factor for polynomials, let us denote by [R]+ the
rational function guaranteed by Proposition 14.17. As with polynomial spectral factorisation, there is a common type of rational function for which one wishes to obtain a spectral
factorisation.
14.18 Corollary If R R(s) then RR admits a spectral factorisation. Furthermore, if R
RL then [R]+ RH+
.
Proof Follows directly from Proposition 14.17 and Corollary 14.13.

M1,ij (i) (t)(j) (t) dt =

a i,j=0

dt = 2 (b) 2 (b).
2(t)(t)

M 2 is not integrable. Indeed we have

]
Clearly [N
is a spectral factorisation of R.
[D]+
Conversely, suppose that R admits a spectral factorisation by and let (N , D ) be the
c.f.r. of . Note that all roots of N and D lie in C . We then have

R = =

22/10/2004

14.2.5 A class of path independent integrals In the proof of Theorem 14.25, it will
be convenient to have at hand a condition that tells us when the integral of a quadratic
function of t and its derivatives only depends on the data evaluated at the endpoints. To
this end, we say that a symmetric matrix M R(n+1)(n+1) is integrable if for any interval
[a, b] R there exists a map F : R2n R so that for any n times continuously differentiable
function : [a, b] R, we have
Z

14.2.4 Spectral factorisation for rational functions One can readily extend the notion of spectral factorisation to rational functions. The notion for such a factorisation is
defined as follows.

14 An introduction to H2 optimal control

2
X

M2,ij (i) (t)(j) (t) dt =

2 (t) dt,

a i,j=0

and this integral will depend not only on the value of at the endpoints of the interval, but
3
on its value between the endpoints. For example, the two functions 1 (t) = t and 2 (t) = t3
have the same value at the endpoints of the interval [1, 1], and their first derivatives also
have the same value at the endpoints, but
Z 1
Z 1
2
2
22 (t) dt = .
21 (t) dt = ,
3
63
1
1
The following result gives necessary and sufficient conditions on the coefficients Mij ,
i, j = 1, . . . , n, for a symmetric matrix M to be integrable.
14.19 Proposition A symmetric matrix M R(n+1)(n+1) is integrable if and only if the
polynomial
n
X

P (s) =
Mij si (s)j + (s)i sj
i,j=0

is the zero polynomial.

Proof First let us consider the terms
Z
a

(i) (t)(j) (t) dt

14.2 Tools for H2 optimisation

22/10/2004

537

when i + j is odd. Suppose without loss of generality that i > j. If i = j + 1 then the
integral
Z b
(j) (t) t=b
(j+1) (t)(j) (t) dt =

2 t=a
a
If i > j + 1 then a single integration by parts yields
Z b
t=b Z b

(i1) (t)(j+1) (t) dt.
(i) (t)(j) (t) dt = (i1) (t)(j) (t)
t=a

One can carry on in this manner 21 (i j 1) times until one arrives at an expression that
is a sum of evaluations at the endpoints, and an integral of the form
Z b
(k+1) (t)(k) (t) dt

538

14 An introduction to H2 optimal control

22/10/2004

14.20 Corollary A symmetric matrix M R(n+1)(n+1) is integrable if and only if there exists
a symmetric matrix P Rnn so that
Z

n
X

Mij (i) (t)(j) (t) dt = xt (b)P x(b) xt (a)P x(a),

a i,j=0

where x(t) = (t), (1) (t), . . . , (n1) (t) .
Proof From Proposition 14.19 we see that if M is integrable, then the only contributions
to the integral come from terms of the form
Z b
Mij (i) (t)(j) (t) dt,
a

where i + j is odd. As we saw in the proof of Proposition 14.19, the resulting integrated
expressions are sums of pairwise products of the form
t=b

(k) (t)(`) (t)

which is then evaluated to

(k) (t) t=b

.
2
t=a
Thus all terms where i + j is odd lead to expressions that are only endpoint dependent.
Now we look at the terms of the form
Z b
(i) (t)(j) (t) dt

t=a

where k, ` {0, 1, . . . , n 1}. This shows that there is a matrix P Rnn so that

Z
1
(i
2

j) times to get an expression that

when i + j is even. We integrate by parts as above
is a sum of terms that are evaluations at the endpoints, and an integral of the form
Z
b ((i+j)/2) 2
1
(1)i + (1)j

(t) dt.
2
a
Thus we will have
Z bX
n
n Z

2
1X b
Mij (i) (t)(j) (t) dt = I0 +
Mij (1)i + (1)j ((i+j)/2) (t) dt,
2
a i,j=0
i,j=0 a

(14.8)

where I0 is a sum of terms that are evaluations at the endpoints. In order for the expression (14.8) to involve evaluations at the endpoints for every function , it must be the case
that
n
X

2
Mij (1)i + (1)j ((i+j)/2) (t) = 0

n
X

Mij (i) (t)(j) (t) dt = xt (b)P x(b) xt (a)P x(a).

a i,j=0

That P may be taken to be symmetric is a consequence of the fact that the expressions
xt (b)P x(b) and xt (a)P x(a) depend only on the symmetric part of P (see Section A.6).
It is in the formulation of the corollary that the notion of path independence in the title
of this section is perhaps best understood, as here we can interpret the integral (14.7) as a
path integral in Rn+1 where the coordinate axes are the values of and its first n derivatives.
The following result is key, and combines our discussion in this section with the spectral
factorisation of the previous section.
14.21 Proposition If P, Q R[s] are coprime then the polynomial P P + QQ admits a
spectral factorisation. Let F be the left half-plane spectral factor for this polynomial. Then
for any n times continuously differentiable function : [a, b] R, the integral

i,j=0

Z b

for all t [a, b]. Now we compute directly that

n
X

n
X

Mij (1)i + (1)j si+j =
Mij si (s)j + (s)i sj .

i,j=0

d
dt

2
(t) + Q

d
dt

2
(t) F

d
dt

2
(t)
dt

(14.9)

is expressible in terms of the value of the function and its derivatives at the endpoints of
the interval [a, b].

Proof For the first assertion, we refer to Exercise E14.2. Let us suppose that n = deg(P )
deg(Q) so that deg(F ) = deg(P ). Let us also write

i,j=0

From this the result follows.

Note that the equation (14.9) tells us that M is integrable if and only if for k = 0, . . . , 2n
we have
X
Mij = 0.
i+j=2k

The result yields the following corollary that makes contact with the Riccati equation
method of Section 14.3.2.

P (s) = pn sn + + p1 s + p0
Q(s) = qn sn + + q1 s + q0
F (s) = fn sn + + f1 s + f0 .

14.2 Tools for H2 optimisation

22/10/2004

539

According to Proposition 14.19, we should show that the coefficients of the even powers of
s in the polynomial P 2 + Q2 F 2 vanish. By definition of F we have
P P + QQ = F F ,

14 An introduction to H2 optimal control

22/10/2004

14.23 Lemma If u is the inverse Laplace transform of RH+

2 then
Z

1
JT1 ,T2 () =
y(t)2 + u (t)2 dt,
2 0
where y satisfies

or
n
X

540

i=0
n
X

p i si

n
X

(1)j pj sj +

j=0

n
X

qi s i

n
X

i=0

(1)i pi pj s2k +

k=0 i+j=2k

n
X
X

(1)j qj sj =

n
X

j=0

(1)i qi qj s2k =

k=0 i+j=2k

fi si

n
X

i=0

(1)j fj sj

j=0

(1)i fi fj s2k .

In particular, it follows that for k = 0, . . . , n we have

X
X
(pi pj + qi qj )
fi fj = 0.

A1 0
,
0 A2

0
,
b2

x(0) = x0 ,

ct = ct1 ct2 ,

x0 =

b1
.
0

Proof By Parsevals theorem we immediately have

Z
Z
1
|(i)|2 d =
|u (t)|2 dt,
2 0

since RH+
2 . We also have

i+j=2k

However, these are exactly the coefficients of the even powers of s in the polynomial P 2 +
Q2 F 2 , and thus our result follows.

14.2.6 H2 model matching In this section we state a so-called model matching
problem that on the surface has nothing to do with the optimal control problems of Section 14.1. However, we shall directly use the solution to this model matching problem to
solve Problem 14.6, and in Section 14.4.2 we shall see that the famous LQG control scheme
is the solution of a certain model matching problem.
The problem in this section is precisely stated as follows.
2
+
14.22 H2 model matching problem Given T1 , T2 RH+
2 , find RH2 so that kT1 T2 k2 +

kk22 is minimised.

T1 (s) = ct1 (sI n1 A1 )1 b1 ,

T2 (s)(s) = ct2 (sI n2 A2 )1 b2 (s),

giving
L 1 (T1 (s))(t) = ct1 eA1 t b,

L 1 (T2 (s)(s))(t) =

c22 eA2 (t ) b2 u ( ) d.

In other words, the inverse Laplace transform of T1 (s) (s)T2 (s) is exactly y, if y is
specified as in the statement of the lemma. The result then follows by Parsevals theorem
since T1 (s) (s)T2 (s) RH+

2.
The punchline is that we now know how to minimise JT1 ,T2 () by finding u as per the
methods of Section 14.3. The following result contains the upshot of translating our problem
here to the previous framework.
,
14.24 Proposition Let j = (Aj , bj , ctj , 01 ) be the canonical minimal realisation of Tj RH+

2
j {1, 2}, let = (A, b, ct , 01 ) be as defined in Lemma 14.23, and let f = f t1 f t2 be
1 = (A1 , b1 , f t , 01 ) and
2 = (A2
the solution to Problem 14.3 for . If we denote
1
b2 f 2 , b2 , f 22 , 01 ), then the solution to the model matching problem Problem 14.22 is

The norm kk2 is the H2 -norm defined in Section 5.3.2:

Z
kf k22 =
|f (i)|2 d.

The idea is that given T1 and T2 , we find so that the cost

Z
Z
JT1 ,T2 () =
|T1 (i) (i)T2 (i)|2 d +

with
A=

n
X
X
k=0 i+j=2k

i+j=2k

x(t)
= Ax(t) + bu (t)
y(t) = ct x(t),

= (1 + T 2 )T 1 .

|(i)|2 d

is minimised. The problem may be thought of as follows. Think of T1 as given. We wish to

see how close we can get to T1 with a multiple of T2 , with closeness being measured by the
H2 -norm. The cost should be thought of as being the cost of the difference of T1 from T2 ,
along with a penalty for the size of the matching parameter .
To solve the H2 model matching problem we turn the problem into a problem much
like that posed in Section 14.1. To do this, we let j = (Aj , bj , ctj , 01 ), j {1, 2}, be
the canonical minimal realisations of T1 and T2 . The following lemma gives a time-domain
characterisation of the model matching cost JT1 ,T2 ().

Proof By Theorem 14.25, Corollary 14.26, and Lemma 14.23 we know that u (t) = f x(t)
where x satisfies the initial value problem

x(t)
= (A bf t )x(t),
Then we have
A bf t =

x(0) = x0 .

A1
0
t
t .
b2 f 1 A2 b2 f 2

Therefore
(sI n (A bf t )1

(sI n1 A1 )1
0
=
t 1
t
t 1 ,
1
(sI n2 (A2 b2 f 2 )) b2 f 1 (sI n1 A1 )
(sI n2 (A2 b2 f 2 ))

22/10/2004

14.3 Solutions of optimal control and state estimation problems

541

542

14 An introduction to H2 optimal control

22/10/2004

Now define F = [DD + N N ]+ and write

giving
(sI n (A bf t )1 x0 =

(sI n1 A1 )1 b1
.
(A2 b2 f t2 ))1 b2 f t1 (sI n1 A1 )1 b1

(sI n2

Z
Jx0 (u) =
0

d
dt

2
xu,x0 (t) + D

This finally gives

14.3 Solutions of optimal control and state estimation problems

With the developments of the previous section under our belts, we are ready to state our
main results. The essential results are those for the optimal control problems, Problems 14.1
and 14.3, and these are given in Section 14.3.1. By going through the model matching problem of Section 14.2.6, the optimal state estimation problem, Problem 14.6, is seen to follow
from the optimal control results. The way of approaching the optimal control problem we
undertake is not the usual route, and is not easily adaptable to MIMO control problems.
In this latter case, the typical development uses the so-called Riccati equation. In Section 14.3.2 we indicate how the connection between our approach and the Riccati equation
approach is made. Here the developments of Section 5.4, that seemed a bit unmotivated at
first glance, become useful.
14.3.1 Optimal control results

The main optimal control result is the following.

14.25 Theorem Let = (A, b, ct , 01 ) be an observable SISO linear system in controller

canonical form with (N, D) the c.f.r. for T . Let F R[s] be defined by
F = [DD + N N ]+ + D.
For x0 Rn , a control u Ux0 solves Problem 14.1 if and only if it satisfies

d
u(t) = F dt
xu,x0 (t),
and is defined on [0, ).
Proof First note that observability of implies that the polynomial DD + N N is even
and positive on the imaginary axis (see Exercise E14.2), so the left half-plane spectral factor
[DD + N N ]+ does actually exist. The system equations are given by (14.1), which we
write as

d
d
D dt
xu,x0 (t) = u(t), N dt
xu,x0 (t) = y(t).
Let u Ux0 . If u is defined on [0, T ] then we may extend u to [0, ) by asking that u(t) = 0
for t > T . Note that if we do this, the value of Jx0 (u) is unchanged. Therefore, let us
suppose that u is defined on [0, ). Now let us write
Z

2
Jx0 (u) =
yu,x
(t) + u2 (t) dt
0
0
Z

2

2
d
d
=
N dt
xu,x0 (t) + D dt
xu,x0 (t)
dt.
0

2
xu,x0 (t) F

d
dt

2
xu,x0 (t)
dt+
Z b

2
d
F dt
xu,x0 (t) dt.
a

(s) = f t1 (sI n1 A1 )1 b1 + f t2 (sI n2 (A2 b2 f t2 ))1 b2 f t1 (sI n1 A1 )1 b1 ,

as given in the statement of the result.

d
dt

By Proposition 14.21 the first integral depends only on the value of xu,x0 (t) and its derivatives
at its initial point and terminal point. Thus it cannot be changed by changing the control.
This means the best we can hope to achieve will be by choosing u so that

d
xu,x0 (t) = 0.
F dt
Note that if xu,x0 does satisfy this condition, then it and its derivatives do indeed go to zero
d
since the zeros of F are in C (see Exercise E14.2). Clearly, since D dt
xu,x0 (t) = u, this
can be done if and only if u(t) satisfies

d
d
u(t) = F dt
xu,x0 (t) + D dt
xu,x0 (t),
as specified in the theorem statement.

Note that, interestingly, the control law is independent of the initial condition x0 for the
state. Also, motivated by Corollary 14.20, let P Rnn be the symmetric matrix with the
property that
Z b

2

2

2
d
d
d
N dt
(t) + D dt
(t) F dt
(t)
dt = x(b)P x(b) x(a)P x(a),
a

where x(t) = ((t), (1) (t), . . . , (n1) (t)). We then see that the cost of the optimal control
with initial state condition x0 is
Jx0 (u) = xt0 P x0 .
Let us compare Theorem 14.25 to something we are already familiar with, namely the
method of finding a state feedback vector f to stabilise a system.
14.26 Corollary Let = (A, b, ct , 01 ) be a controllable and detectable SISO linear system, and
let T Rnn be an invertible matrix with the property that (T AT 1 , T b) is in controller
canonical form. If f = (f0 , f1 , . . . , fn1 ) Rn is defined by requiring that
fn1 sn1 + + f1 s + f0 = [D(s)D(s) + N (s)N (s)]+ D(s),
where (N, D) is the c.f.r. for T , then f = T t f is a solution of Problem 14.3.
Proof Let us denote by
c
b,
= (A,
t , 01 ) = (T AT 1 , T b, T t ct , 01 )

is in controller canonical form. We make the obser b)

the system in coordinates where (A,
0 = T x0
vation that the cost function is independent of state coordinates. That is to say, if x
then Jc,x0 (f) = Jc,x0 (f ), where f = T t f . Thus it suffices to prove the corollary for the
system in controller canonical form.

543

Thus we proceed with the proof under this assumption. Theorem 14.25 gives the form of
the control law u in terms of xu,x0 that must be satisfied for optimality. It remains to show
that the given u is actually in state feedback form. However, since the system is in controller
canonical form we have xu,x0 (t) = (x(t), x(1) (t), . . . , x(n1) (t)), and so it does indeed follow
that

d
d
d
d +
d
[D dt
D dt
+ N dt
N dt
] D dt
xu,x0 (t) = f t xu,x0 (t),
with f as defined in the statement of the corollary. The corollary now follows since in this
case we have Jc,x0 (f ) = Jx0 (u).

14.27 Remarks 1. The corollary gives, perhaps, the easiest way of seeing what is going on
with Theorem 14.25 in that it indicates that the control law that solves Problem 14.1
is actually a state feedback control law. This is not obvious from the statement of the
problem, but is a consequence of Theorem 14.25.
2. The assumption of controllability in the corollary may be weakened to stabilisability.
In this case, to construct the optimal state feedback vector, one would put the system
into the canonical form of Theorem 2.39, with (A1 , b1 ) in controller canonical form.
One then constructs f t = [ f t1 0t ], where f 1 is defined by applying the corollary to
1 = (A1 , b1 , ct1 , 01 ).

Let us look at an example of an application of Theorem 14.25, or more properly, of
Corollary 14.26.
14.28 Example (Example 6.50 contd) The system we look at in this example had

0 1
0
A=
, b=
.
1 0
1
In order to specify an optimal control problem, we also need an output vector c to specify
an output cost in Problem 14.3. In actuality, one may wish to modify the output vector
to obtain suitable controller performance. Let us look at this a little here by choosing two
output vectors

1
0
c1 =
, c2 =
.
0
1
To simplify things, note that (A, b) is in controller canonical form. Thus Problems 14.1
and 14.3 are related in a trivial manner. If 1 = (A, b, ct1 , 01 ) and 2 = (A, b, ct2 , 01 ), then
one readily computes
1
s
T1 (s) = 2
, T2 (s) = 2
.
s +1
s +1
Let us denote (N1 (s), D1 (s)) = (1, s2 + 1) and (N2 (s), D2 (s)) = (s, s2 + 1). We then compute
D1 (s)D1 (s) + N1 (s)N1 (s) = s4 + 2s2 + 2
D2 (s)D2 (s) + N2 (s)N2 (s) = s4 + s2 + 1.
Following the recipe of Remark 14.14, one determines that

4
[D1 (s)D1 (s) + N1 (s)N1 (s)]+ = s2 + 2 2 sin 8 s + 2
[D2 (s)D2 (s) + N2 (s)N2 (s)]+ = s2 + s + 1.

544

14 An introduction to H2 optimal control

22/10/2004

Then one follows the recipe of Corollary 14.26 and computes

4
[D1 (s)D1 (s) + N1 (s)N1 (s)]+ D1 (s) = 2 2 sin 8 s + 2 1
[D2 (s)D2 (s) + N2 (s)N2 (s)]+ D2 (s) = s.
Thus the two optimal state feedback vectors are

21
f1 =
4
,
2 2 sin 8

f2 =

0
.
1

The eigenvalues of the closed-loop system matrices

A bf t1 ,

A bf t2

PSfrag replacements
aret 4 2 sin 8 i 1 sin2 8
{0.46t i1.10} and { 12 i 23 }. In Figure 14.1
ke(t)k
ke(t)k
e1 (t) plotted the trajectories for the closed-loop
e1 (t) system in the (x1 , x2 )-plane for the initial
are

PSfrag replacements

e2 (t)
e3 (t)
y(t)
u(t)
h (t)
hN,D (t)
1 (t)
1N,D (t)
(t)
(t)
fj (t)
Re
Im

x1
x2
log
dB
deg
u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

,0
0

e2 (t)
e3 (t)
y(t)
u(t)
h (t)
hN,D (t)
1 (t)
1N,D (t)
(t)
(t)
fj (t)
Re
Im

1.5

0.5

-0.5

-1

-0.5

0.5

x1
x2
log
dB
deg
u = 0 ln
ln coth(|u| /2)
or 1
m

yos

1
1.5tos 0

1.5

0.5

14.3 Solutions of optimal control and state estimation problems

22/10/2004

-0.5

-1

-0.5

0.5

1.5

,0
0

Figure 14.1 Optimal trajectories for output vector c1 (left) and c2

(right)

condition (1, 1). One can see a slight difference in that in the optimal control law for c2
the x2 -component of the solution is tending to zero somewhat more sharply. In practice,
one uses ideas such as this to refine a controller based upon the principles we have outlined
here.

14.3.2 Relationship with the Riccati equation In classical linear quadratic regulator
theory, one does not normally deal with polynomials in deriving the optimal control law.
Normally, one solves a quadratic matrix equation called the algebraic Riccati equation.1
In this section, that can be regarded as optional, we make this link explicit by proving the
following theorem.
1
After Jacopo Francesco Riccati (16761754) who made original contributions to the theory of differential
equations. Jacopo also had a son, Vincenzo Riccati (17071775), who was a mathematician of some note.

22/10/2004

14.3 Solutions of optimal control and state estimation problems

545

14.29 Theorem If = (A, b, c, 01 ) is a controllable SISO linear system, then there exists at
least one symmetric solution P Rnn of the equation
At P + P A P bbt P = cct ,

546

14 An introduction to H2 optimal control

2. let Q1 R[s] be defined by

bt P 1 (sI n A)1 b =

(14.10 )

and, furthermore, exactly one of these solutions is positive-definite. The equation (14.10) is
called the algebraic Riccati equation.
Proof Let
c
b,
= (A,
t , 01 ) = (T AT 1 , T b, T t ct , 01 )

is in controller canonical form. Note that if

b)
be the system in state coordinates where (A,
is a solution for the equation
P

4. substitute the above definitions into the equation () and the resulting expression turns
out to be
P F
P F
+
= 0;

DD
DD
5. as D and F are monic, this reduces to
P
P
+
= 0;
F F

which is nonnegative for s = i. Let us denote this rational function by R. Defining

F = [RDD ]+ we see that
F (s)F (s)
D(s)D(s)

(here we have used the fact that A and At have the same eigenvalues). Note that F is monic
and of degree n. Because A is in controller canonical form, a simple computation along the
lines of that in the proof of Theorem 6.49 shows that if f Rn is defined by
f t (sI n A)1 b =

F (s) D(s)
D(s)

P
F

and

P
F

have no common poles

(bt P 1 f t )(sI n A)1 b = 0;

(14.15)

8. this implies that the components of the vector bt P 1 f t are zero, or that P 1 b = f ,
as desired.
Thus we have shown that (14.10) has a solution, and the solution P 1 we found was
positive-definite. Let us show that this is the only positive-definite solution. Let P 2 be
positive-definite and suppose that
At P 2 + P 2 A P 2 bbt P 2 = cct .

(14.16)

Now argue as follows:

(14.11)

then det(sI n (A bf t )) = F (s). By the spectral factorisation theorem, A bf t is thus

Hurwitz. By Theorem 5.32(i) there exists a unique positive-definite solution to the equation
(At f bt )P + P (A bf t ) = cct f f t .

6. since F is analytic in C+ , the rational functions

so (14.14) holds if and only if P = P = 0;

(14.14)

7. by the definition of P , P (s) = 0 implies that

Tt (s)T (s) + 1 = bt (sI n A)cct (sI n A)1 b + 1,

Tt (s)T (s) + 1 =

Q1 (s)
;
D(s)

3. let P = Q1 F + D;

b
t P
=
+P
A
P
b
tP
t ,
cc
A
T 1 is a solution to (14.10). Therefore, without loss of generality we assume
then P = T P
that (A, b) is in controller canonical form. Let T (s) = ct (sI n A)1 b be the transfer
function for with (N, D) its c.f.r. Then we have

22/10/2004

1. multiply (14.16) by 1, add and subtract P 2 s, and multiply on the left and right by
bt (sI n A)1 and (sI n A)1 b, respectively, to get an expression ();
2. Define Q2 R[s] by requiring that
bt P 2 (sI n A)1 b =

(14.12)

Let us denote this solution by P 1 . A straightforward computation shows that the previous
equation is equivalent to

Q2 (s)
;
D(s)

3. following the arguments in the existence part of the proof, show that P 2 b = f ;
4. this implies that (14.16) is equivalent to

At P 1 + P 1 A P 1 bbt P 1 = cct (P 1 b f )(P 1 b f )t .

(14.13)

Thus the existence part of the theorem will follow if we can show that P 1 b = f . Let us
show this.
The argument is outlined as follows:
1. multiply (14.13) by 1, add and subtract P 1 s, and multiply on the left and right by
bt (sI n A)1 and (sI n A)1 b, respectively, to get an equation ();

(At f bt )P 2 + P 2 (A bf t ) = cct f f t ;
5. by Theorem 5.32(i), P 2 = P 1 .
This completes the proof.

During the course of the proof of the theorem, we arrived at the relationship between
the solution to the algebraic Riccati equation and the optimal state feedback vector f . Let
us record this.

22/10/2004

14.3 Solutions of optimal control and state estimation problems

547

14.30 Corollary Consider a controllable SISO linear system = (A, b, ct , 01 ). If f is the

optimal state feedback vector of Corollary 14.26 and P is the unique positive-definite solution
to the algebraic Riccati equation (14.10), then f = P b.
Proof By Corollary 14.26 the vector f = (f0 , f1 , . . . , fn1 ) is defined by
fn1 sn1 + + f1 s + f0 = [D(s)D(s) + N (s)N (s)]+ D(s).
However, this is exactly the relation (14.11) in the proof of Theorem 14.29, since in the
statement of the theorem we had assumed (A, b) to be in controller canonical form. During
the course of the same proof, the relation f = P b was also shown to be true when (A, b)
is in controller canonical form. The result now follows from the transformation property for
control systems under state similarity stated in Proposition 2.36.

14.31 Example (Example 14.28 contd) We resume with the situation where

0 1
0
A=
, b=
.
1 0
1
and where we used the two output vectors

1
c1 =
,
0

c2 =

The two feedback vectors were computed to be

21
f1 =
4
,
2 2 sin 8

0
.
1

f2 =

respectively. Let us denote by P 1 and P 2 the two corresponding positive-definite matrices

guaranteed by Theorem 14.29. Referring to the equation (14.12) in the proof of Theorem 14.29, we see that the matrices P 1 and P 2 satisfy
j = 1, 2.

Now we refer to the proof of part (i) of Theorem 5.32 to see that
Z
t
eAj t Qj eAj t dt, j = 1, 2,
Pj =
0

where
Aj = A bf tj ,

Qj = cj ctj + f j f tj ,

14 An introduction to H2 optimal control

22/10/2004

2. Note that the matrix on the right-hand side of the algebraic Riccati equation is the matrix
that determines the penalty given to states (as opposed to control) in the cost function of
Problems 14.1 and 14.3. One can easily imagine using more general symmetric matrices
to define this cost, thus looking at cost functions of the form
Z

xtu,x0 (t)Qxu,x0 (t) + Ru2 dt.
Jx0 (u) =
0
nn

where Q R
is symmetric and positive-semidefinite, and R > 0. This can also be
seen to be generalisable to multiple inputs by making the cost associated to the input
be of the form ut (t)Ru(t) for an m m symmetric matrix Q. Making the natural
extrapolation gives the analogue of equation (14.10) to be
At P + P A P BR1 B t P = Q.
This is indeed the form of the algebraic Riccati equation that gets used in MIMO generalisations of our Theorem 14.25.
3. The optimal feedback vector f determined in this section is often referred to as the
linear quadratic regulator (LQR).

14.3.3 Optimal state estimation results With the optimal control results of the
preceding two sections, and with the (now known) solution of the model matching problem
of Section 14.2.6, we can prove the following result which characterises the solution to the
optimal state estimation problem.

0
,
1

(At f j bt )P j + P j (A bf tj ) = cj ctj f j f tj ,

548

14.33 Theorem For = (A, b, ct , D) a complete SISO linear system, the following statements regarding ` Rn are equivalent:
(i) ` is a solution for Problem 14.6;
(ii) if (N, D) is the c.f.r. of T , if T Rnn is the invertible matrix with the property that
(T AT 1 , T t c) is in observer canonical form, and if ` = (`0 , `1 , . . . , `n1 ) is defined
by requiring that
`n1 sn1 + + `1 s + `0 = [D(s)D(s) + N (s)N (s)]+ D(s),
then ` = T 1 `;
(iii) ` = P c where P is the unique positive-definite solution to the algebraic Riccati equation

j = 1, 2.

One may do the integration (I used Mathematica , of course) to obtain

2 4 8 sin 8 2 1
1 0
P1 =
,
P
=
.
2
0 1
2 1 2 4 2 sin 8
In each case we readily verify that f j = P j b, j = 1, 2, just as predicted by Corollary 14.30.
14.32 Remarks 1. The algebraic Riccati equation must generally be solved numerically. Indeed, note that Theorem 14.25 provides essentially the same information as the algebraic
Riccati equation (as made precise in Corollary 14.30). In the former case, we must find
the left half-plane spectral factor of an even polynomial, and this involves finding the
roots of this polynomial. This itself is something that typically must be done numerically.

AP + P At P cct P = bbt .
=
Proof First let us demonstrate the equivalence of (ii) and (iii). Let us define
(At , c, bt , D), and let T Rnn be the invertible matrix which puts (At , c) into controller
D)
is the
canonical form. From Corollaries 14.26 and 14.30 we infer the following. If (N,
c.f.r. for T then if ` is defined by
D(s)

(s)N
(s)]+ D(s),

`n1 sn1 + + `1 s + `0 = [D(s)

+N
we have ` = T t `. Now we note that (T At T 1 , T c) is in controller canonical form if and
only if (T t AT t , T c) is in observer canonical form. From this we infer that T = T t . The
D)
= (N, D).
equivalence of parts (ii) and (iii) now follow from the fact that (N,

22/10/2004

14.3 Solutions of optimal control and state estimation problems

549

We now show the equivalence of parts (i) and (iii). First we note that by Parsevals
Theorem we have
Z
Z

1
T (i) T (i)T (i)2 d + 1
T (i)2 d,
J(`) =
`
2
2 `

550

(1 + T 2 (s))T 1 (s) = (1 ct (sI n A + `ct )1 `)ct (sI n A)1 `

= ct (sI n A)1 (I n `ct (sI n A + `ct )1 )`
= ct (sI n A + `ct `ct )(sI n A + `ct )`
= ct (sI n A + `ct )`,
as desired.

T (s) = ct (sI n A)1 b = bt (sI n At )1 ct

T` (s) = ct (sI n A + `ct )1 ` = `t (sI n At + c`t )1 c.
We shall think of these transfer functions as being in this form for the moment. Now,
thinking of T` as the unknown, we wish to choose this unknown to minimise J(`). This,
however, is exactly the model matching problem posed in Section 14.2.6. The solution
to the model matching problem is given, if we know the solution to the optimal control
problem Problem 14.3. But since we now know this, we do indeed know the solution to
the model matching problem, and let us translate the solution into our current notation.
1 =
The transfer function T` minimising J(`) is given by T` = (1 + T 2 )T 1 where
t
t
t
t
t
t
t
t t
t P
,

(A , c, `1 , 01 ) and 2 = (A c`2 , c, `2 , 01 ), and where ` = [ `1 `1 ] is given by ` = b

is the unique positive-definite solution to the algebraic Riccati equation
where P

with

t

= A 0t ,
A
0 A

= 0 ,
b
c

22/10/2004

thus giving

where ` = (A `ct , `, ct , 01 ). Since the transfer functions are scalar, they may be transposed without changing anything. That is to say, we have

b
t P
tP
+P
A
P
b
=
t ,
A
cc

14 An introduction to H2 optimal control

14.4 The linear quadratic Gaussian controller

In Section 10.5.3 we showed how one could combine state estimation via a Luenberger
observer with static state feedback to obtain a controller that worked by using the estimated
states in the state feedback law.
14.4.1 LQR and pole placement We know that if the state feedback vector f is
chosen according to Corollary 14.26, then the matrix A bf t will be Hurwitz. However, we
have said nothing about the exact nature of the eigenvalues of this matrix. In this section
we address this issue.

(14.17)
14.4.2 Frequency domain interpretations Our above presentations for the linear
quadratic regulator and the optimal state estimator are presented in the time-domain. While
we present their solutions in terms of spectral factorisation of polynomials, we also see that
these solutions are obtainable using the algebraic Riccati equation, and this is how these
problems are typically solved in the MIMO case. However, it is also possible to give frequency
response formulations and solutions to these problems, and in doing so we make a connection
to the H2 model matching problem of Section 14.2.6.

t = bt bt .
c

Writing

= P 11 P 12
P
P 12 P 22
and expanding (14.17), we arrive at the following four equations:
AP 11 + P 11 At P 12 cct P 12
AP 12 + P 12 At P 12 cct P 22
AP 12 + P 12 At P 22 cct P 12
AP 22 + P 22 At P 22 cct P 22

= bbt
= bbt
= bbt
= bbt .

(14.18)

Since P is symmetric and positive-definite, so too is P 22 (since xt P xt > 0 for all x of the
form xt = [ 0 xt2 ]). The last of equations (14.18) then uniquely prescribes P 22 as being the
matrix prescribed in part (iii) of the theorem. Adding the last two of equations (14.18) gives
A(P 12 + P 22 ) + (P 12 + P 22 )At P 22 cct (P 12 + P 22 ) = 0.
This gives P 12 + P 22 = 0 by . We then have

t t
P 11 P 22

`1 `2 = 0 ct
= ct P 22 ct P 22 ,
P 22 P 22
and we take ` = `2 . Now we obtain
T 1 (s) = `t (sI n At )1 c = ct (sI n A)1 `
T 2 (s) = `t (sI n At + c`t )1 c = ct (sI n A + `ct )1 `,

14.4.3 H2 model matching and LQG In this section we show that LQG control may
be posed as an H2 model matching problem. This will provide us with a natural segue to the
next chapter where we discuss a more difficult model matching problem, that of H model
matching. By representing the somewhat easily understood LQG control in the context of
model matching, we hope to motivate the less easily understood material in the next chapter.

14.5 Stability margins for optimal feedback

In this section we wish to investigate some properties of our optimal feedback law. We
shall work in the setting of Corollary 14.26. It is our desire to talk about the gain and phase
margin for the closed-loop system in this case. However, it is not quite clear that it makes
sense to do this as gain and phase margins are defined in the context of unity gain feedback
loops, not in the context of static state feedback. Therefore, the first thing we shall is recall
from Exercise E7.11 the connection between static state feedback and unity gain feedback
loop concepts. In particular, recall that A bf t is Hurwitz if and only if
f t (sI n A)1 b
RH+
.
1 + f t (sI n A)1 b

14.5 Stability margins for optimal feedback

22/10/2004

551

This result tells us that closed-loop stability of the closed-loop system f is equivalent to
IBIBO stability of a unity gain feedback loop with loop gain RL (s) = f t (sI n A)1 b.
Note that this enables us to employ our machinery for these interconnections, thinking of
f (sI n A)1 b as being the loop gain.
14.5.1 Stability margins for LQR In particular, as is made clear in Exercise E7.11,
we may employ the Nyquist criterion to determine closed-loop stability under static state
feedback. Recall that rather than the poles of the loop gain in C+ , one uses the eigenvalues
of A in C+ to compare with the encirclements of 1 + i0. Let us illustrate the Nyquist
criterion on an unstable system.
14.34 Example We take
A=

0 1
,
1 2

0
,
1

1
.
1

u(t)
h (t)

t
1
In Figure 14.2 is shown
hN,D (t)the Nyquist plot for the loop gain Rf (s) = f (sI 2 A) b. One
2

14.35 Theorem Let = (A, b, ct , 01 ) be a controllable SISO linear system and let f Rn
be the optimal state feedback vector of Corollary 14.26. We have

1 f (iI n A)1 b 1
for every R. In other words, the point Rf (i) is at least distance 1 away from the point
1 + i0.
Proof We assume without loss of generality that (A, b) is in controller canonical form. We
let P be the unique positive-definite matrix guaranteed by Theorem 14.29. Thus
(14.19)

Now perform the following computations:

1. to (14.19), add and subtract iP ;
2. multiply the resulting equation on the left by bt (iI n At )1 and on the right by
(iI n A)1 b;
3. use the identity f = P b from Corollary 14.30.
The result of these computations is

1 + f t (iI n A)1 b2 = 1 + bt (iI n At )1 cct (iI n A)1 b.
Since the matrix cct is positive-semidefinite, we have
bt (iI n At )1 cct (iI n A)1 b 0.

This interesting result tells us that the Nyquist contour remains outside the circle of
radius 1 in the complex plane with centre 1 + i0. Thus, it remains well clear of the critical
point.

14.36 Example (Example 14.28 contd) We resume looking at the system with

0 1
0
A=
, b=
,
1 0
1

,0
0

22/10/2004

The discussion to this point has largely been with respect to general feedback vectors.
However, notice that the Nyquist plot of Example 14.34, done for the optimal state feedback
vector of Corollary 14.26, has an interesting feature: the Nyquist contour remains well clear
of the critical point 1 + i0. The following result tells us that we can generally expect this
to happen when we use the optimal state feedback vector.

The result follows.

1.5

x1
x2
x1
0.5
x2
log
0
dB
deg
u = 0 ln -0.5
ln coth(|u| /2)
or 1
-1
m

yos -1.5

tos 0

14 An introduction to H2 optimal control

At P + P A P bbt P + cct = 0nn .

We ascertain that A has a repeated eigenvalue of +1, thus np = 2. Exercise E7.11(b)

indicates that any stabilising state feedback vector f will have the property that the Nyquist
plot for the loop gain Rf (s) = f t (sI s A)1 b will encircle the origin twice in the clockwise
direction. Let us test this, not for just any stabilising state feedback vector, but for the
optimal one. Without going through the details as we have done this for one example
PSfrag replacements
already, the optimal
feedback state vector is
t
ke(t)k

#
"
e1 (t)
1 +q 2
q
e2 (t)

.
f
=
e3 (t)
7
241 + 72 + 241 2
y(t)
2
1 (t)
1N,D (t)
(t)
(t)
fj (t)

552

-2.5

-2

-1.5

-1

-0.5

0.5

and where wed considered the two output vectors

1
0
c1 =
, c2 =
,
0
1

leading to the optimal state feedback vectors

21
f1 =
,
2 4 2 sin 8

Figure 14.2 Nyquist plot for optimal state feedback with two unstable eigenvalues

can see that it encircles the origin twice in the clockwise direction, as predicted.

f2 =

0
,
1

respectively. In Figure 14.3 we give the Nyquist plots for both loop gains Rf 1 and Rf 2 . As
predicted by Theorem 14.35, both Nyquist contours remain outside the circle of radius 1
centred at 1 + i0.

PSfrag replacements

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)
h22/10/2004
(t)
hN,D (t)
1 (t)
1N,D (t) 6
(t)
(t)
fj (t)
4

,0
0

Summary

553

-4

-2

14 An introduction to H2 optimal control

22/10/2004

Exercises

E14.1 Show that an inner function has a magnitude Bode plot that is a constant 0dB.
E14.2 Let P, Q R[s] have no common roots on the imaginary axis.
(a) Show that the polynomial P P + QQ is even and positive when evaluated on
iR.
(b) Conclude that P P + QQ admits a spectral factorisation.
(c) Show that [P P + QQ ]+ is Hurwitz (i.e., show that [P P + QQ ]+ has no roots
on the imaginary axis).
E14.3 Show that if polynomials R1 , R2 R(s) admit a spectral factorisation, then so does
R1 R2 .
E14.4 Show that if R RH+
, then its outer factor is also a spectral factor.
E14.5 In Theorem 14.25 the optimal control law drives the output to zero, but it is not
said what happens to the state. In this exercise, you will redress this problem.
(a) Show that not only does limt y(t) = 0, but that limt y (k) (t) = 0 for all
k > 0.
(b) Now use the fact that in the statement of Theorem 14.25, is said to be observable to show that limt x(t) = 0.

x1
2
x2
x1
x2
log
0
dB
deg
u = 0 ln
ln coth(|u| /2) -2
or 1
m
-4
yos

tos 0

554

x1
2
x2
x1
x2
log
0
dB
deg
u = 0 ln
ln coth(|u| /2) -2
or 1
m
-4
yos

tos 0

t
ke(t)k
e1 (t)
e2 (t)
e3 (t)
y(t)
u(t)
14.6
h (t)
hN,D (t)
1 (t)
1N,D (t)
(t)
(t)
fj (t)

6,0
0

-4

-2

Figure 14.3 Nyquist plots for the optimal state feedback vectors
of Example 14.36

14.5.2 Stability margins for LQG Results of the nature of Theorem 14.35 demand
full knowledge of the state. Doyle [1978] shows that as soon as one tries to estimate the state
from the output using a Kalman filter, the stability margin of Theorem 14.35 disappears.
Although state estimation, and in particular Kalman filtering, is something we do not cover
in this book, the reader should be aware of these lurking dangers.

14.6 Summary

In Exercise E10.5 you provided a characterisation of a stabilising state feedback vector using
a linear matrix inequality (LMI). In the following exercise, you will further characterise the
optimal state feedback vector using LMIs. The characterisation uses the algebraic Riccati
equation of Theorem 14.29.
E14.6
E14.7 For the pendulum on a cart of Exercises E1.5 and E2.4, choose parameter values for
the mass of the cart, the mass of the pendulum, the gravitational constant, and the
length of the pendulum arm to be M = 1 12 , m = 1, g = 9.81, and ` = 12 . For each of
the following linearisations:
(a) the equilibrium point (0, ) with cart position as output;
(b) the equilibrium point (0, ) with cart velocity as output;
(c) the equilibrium point (0, ) with pendulum angle as output;
(d) the equilibrium point (0, ) with pendulum angular velocity as output,
do the following:
1. construct the optimal state feedback vector of Problem 14.3;
2. compute the closed-loop eigenvalues;
3. plot a few trajectories of the full system with the control determined by the
state feedback vector of part (1).
E14.8 For the double pendulum system of Exercises E1.6 and E2.5, choose parameter values
for the first link mass, the second link mass, the first link length, and the second
link length to be m1 = 1, m2 = 2, `1 = 21 , and `2 = 31 . For each of the following
linearisations:
(a) the equilibrium point (0, , 0, 0) with the pendubot input;
(b) the equilibrium point (, 0, 0, 0) with the pendubot input;
(c) the equilibrium point (, , 0, 0) with the pendubot input;
(d) the equilibrium point (0, , 0, 0) with the acrobot input;

Exercises for Chapter 14

555

(e) the equilibrium point (, 0, 0, 0) with the acrobot input;

(f) the equilibrium point (, , 0, 0) with the acrobot input,
do the following:
1. construct the optimal state feedback vector of Problem 14.3;
2. compute the closed-loop eigenvalues;
3. plot a few trajectories of the full system with the control determined by the
state feedback vector of part (1).
E14.9 Consider the coupled tank system of Exercises E1.11, E2.6, and E3.17. Choose the
1
1
parameters = 31 , 1 = 1, A1 = 1, A2 = 21 , a1 = 10
, a2 = 20
, and g = 9.81. For the
linearisations in the following cases:
(a) the output is the level in tank 1;
(b) the output is the level in tank 2;
(c) the output is the difference in the levels,
do the following:
1. construct the optimal state feedback vector of Problem 14.3;
2. compute the closed-loop eigenvalues;
3. plot a few trajectories of the full system with the control determined by the
state feedback vector of part (1).
E14.10 Let = (A, b, ct , 01 ) be controllable and suppose that A is Hurwitz. Let f be the
optimal state feedback vector of Corollary 14.26 with Rf (s) = f t (sAn A)1 b the
corresponding loop gain. Show that the gain margin for the resulting Nyquist plot
is infinite, and that the phase margin exceeds 60 .

556

14 An introduction to H2 optimal control

22/10/2004

558

This version: 22/10/2004

Chapter 15

15 An introduction to H control theory

22/10/2004

Readers wishing only to be able to apply the methods in this chapter should go through
the chapter with the above skeleton as a guide. However, those wishing to see the
details will be happy to see very little omitted. It should also be emphasised that
Mathematica and Maple packages for solving these problems may be found at the URL
http://mast.queensu.ca/~math332/.1

An introduction to H control theory

Contents
15.1 Reduction of robust performance problem to model matching problem . . . . . . . . . . . 558

The task of designing a controller which accomplishes a specified task is a challenging

chore. When one wishes to add robustness to the mix, things become even more challenging. By robust design, what is meant is that one should design a controller which works not
only for a given model, but for plants which are close to that model. In this way, one can
have some degree of certainty that even if the model is imperfect, the controller will behave in
a satisfactory manner. The development of systematic design procedures for robust control
can be seen to have been initiated with the important paper of Francis and Zames [1984].
Since this time, there have been many developments and generalisations. The understanding
of so-called H methods has progressed to the point that a somewhat elementary treatment
is possible. We shall essentially follow the approach of Doyle, Francis, and Tannenbaum
[1990]. For a recent account of MIMO developments, see [Dullerud and Paganini 1999]. The
book by Francis [1987] is also a useful reference.
Although all of the material in this chapter can be followed by any student who has
come to grips with the more basic material in this book, much of what we do here is a
significant diversion from control theory, per se, and is really a development of the necessary
mathematical tools. Since a complete understanding of the tools is not necessary in order
to apply them, it is perhaps worthwhile to outline the bare bones guide to getting through
this chapter. This might be as follows.
1. One should first make sure that one understands the problem being solved: the robust performance problem. The first thing done is to modify this problem so that it is tractable;
this is the content of Problem 15.2.
2. The modified robust performance problem is first converted to a model matching problem,
which is stated in generality in Problem 15.3. The content of this conversion is contained
in Algorithm 15.5.
3. The model matching problem is solved in this book in two ways. The first method involves Nevanlinna-Pick interpolation, and to apply this method, follow the steps outlined
in Algorithm 15.18. It is entirely possible that you will have to make some modifications
to the algorithm to ensure that you arrive at a proper controller. The necessary machinations are discussed following the algorithm.
4. The second method for solving the model matching problem involves approximation of
unstable rational functions by stable ones via Neharis Theorem. To apply this method,
follow the steps in Algorithm 15.29. As with Nevanlinna-Pick interpolation, one should
be prepared to make some modifications to the algorithm to ensure that things work
out. The manner in which to carry out these modifications is discussed following the
algorithm.
5. Other problems that can be solved in this manner are discussed in Section 15.5.

15.1.1 A modified robust performance problem . . . . . . . . . . . . . . . . . . . . . . . . 558

15.1.2 Algorithm for reduction to model matching problem . . . . . . . . . . . . . . . . . 560
15.1.3 Proof that reduction procedure works . . . . . . . . . . . . . . . . . . . . . . . . . 562
15.2 Optimal model matching I. Nevanlinna-Pick theory . . . . . . . . . . . . . . . . . . . . . 566
15.2.1 Picks theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 566
15.2.2 An inductive algorithm for solving the interpolation problem . . . . . . . . . . . . 568
15.2.3 Relationship to the model matching problem . . . . . . . . . . . . . . . . . . . . . 570
15.3 Optimal model matching II. Neharis Theorem . . . . . . . . . . . . . . . . . . . . . . . . 570
15.3.1 Hankel operators in the frequency-domain . . . . . . . . . . . . . . . . . . . . . . 571
15.3.2 Hankel operators in the time-domain . . . . . . . . . . . . . . . . . . . . . . . . . 573
15.3.3 Hankel singular values and Schmidt pairs . . . . . . . . . . . . . . . . . . . . . . . 576
15.3.4 Neharis Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 578
15.3.5 Relationship to the model matching problem . . . . . . . . . . . . . . . . . . . . . 580
15.4 A robust performance example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 580
15.5 Other problems involving H methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . 580

15.1 Reduction of robust performance problem to model matching

problem
For SISO systems, the problem of designing for robust performance turns out to be reducible in a certain sense to two problems investigated by mathematicians coming under the
broad heading of complex function approximation, or, more descriptively, model matching.
In this section we shall discuss how this reduction takes place, as in itself it is not an entirely
obvious step.
15.1.1 A modified robust performance problem First recall the robust performance
P , a function Wu RH+ given an
problem, Problem 9.23. Given a proper nominal plant R

P , Wu ) or P+ (R
P , Wu ), and a performance weight Wp R(s), we seek
uncertainty set P (R
a controller RC that stabilises the nominal plant and satisfies either

|Wu TL | + |Wp SL | < 1 or |Wu RC SL | + |Wp SL | < 1,

These are not currently implemented. Hopefully they will be in the near future.

22/10/2004

15.1 Reduction of robust performance problem to model matching problem

559

depending on whether one is using multiplicative or additive uncertainty. The robust performance problem, it turns out, is quite difficult. For instance, useful necessary and sufficient
conditions for there to exist a solution to the problem are not known. Thus our first step
in this section is to come up with a simpler problem that is easier to solve. The simpler
problem is based upon the following result.
15.1 Lemma Let R1 , R2 R(s) and denote by |R1 | + |R2 | the R-valued function s 7
(|R1 (s)| + |R2 (s)|) and by |R1 |2 + |R2 |2 the R-valued function s 7 (|R1 (s)|2 + |R2 (s)|2 ).
If

|R1 |2 + |R2 |2 < 1

2
then

|R1 | + |R2 | < 1.

Proof Define

S1 = (x, y) R2 x, y > 0, |x + y| < 1 ,

S2 = (x, y) R2 x, y > 0, x2 + y 2 < 21 .
The result will follow if we can show that S1 S2 . However, we note that S1 = [0, 1] [0, 1]
and S2 is the circle of radius 12 centred at the origin. Clearly S1 S2 (see Figure 15.1).

560

15 An introduction to H control theory

22/10/2004

3. RC is proper: We make this assumption mostly as a matter of convention. If we arrive

at a controller that is improper but solves the problem, then it is often possible to
modify the controller so that it is proper and still solves the problem.
Thus our problem becomes.
15.2 Modified robust performance problem Given
P ,
(i) a nominal proper plant R
(ii) a function Wu RH+
,

P , Wu ) or P+ (R
P , Wu ), and
(iii) an uncertainty model P (R
(iv) a performance weight Wp RH+
,

so that Wu and Wp have no common imaginary axis zeros, find a proper controller RC that
(v) stabilises the nominal system and

(vi) satisfies either |Wu TL |2 +|Wp SL |2 < 1 or |Wu RC SL |2 +|Wp SL |2 < 1, depending
on whether one is using multiplicative or additive uncertainty.

P , Wu , and Wp it will not
As should be clear from Figure 15.1, it is possible that for a given R
be possible to solve the modified robust performance problem even though a solution may
exist to the robust performance problem. Thus we are sacrificing something in so modifying
the problem, but what we gain is a simplified problem that can be solved.
15.1.2 Algorithm for reduction to model matching problem The objective of this
section is to convert the modified robust performance problem into the model matching
problem. We shall concentrate in this section on multiplicative uncertainty, with the reader
filling in the details for additive uncertainty in Exercise E15.3.
First let us state the model matching problem.

+
15.3 A model matching problem Let T1 , T2 RH+
. Find RH (s) so that kT1 T2 k
is minimised.

Figure 15.1 Interpretation of modified robust performance condition

This leads to a modification of the robust performance problem, and we state this formally
since it is this problem to which we devote the majority of our effort in this chapter. Note
that we make a few additional assumptions in the statement of the problem that are not
present in the statement of Problem 9.23. Namely, we assume now the following.
1. Wp RH+
: Thus we add the assumption that Wp be proper since, without loss of
generality as we are only interested on the value of Wp on iR, we can suppose that all
poles of Wp lie in C .
2. Wu and Wp have no common imaginary axis zeros: This is an assumption that, if not
satisfied, can be satisfied with minor tweaking of Wu and Wp .

The model matching problem may not have a solution. In fact, it will often be the case in
applications that it does not have a solution. However, as we shall see as we get into our
development, even when the problem has no solution, it can be used as a guide to solve the
problem that is actually of interest to us, namely the modified robust performance problem,
Problem 15.2. Some issues concerning existence of solutions to the model matching problem
are the topic of Exercise E15.2
15.4 Remark Note that since T1 , T2 RH+
, if is to be a solution of the model matching
problem, then it can have no imaginary axis poles. Also, since the model matching problem
only cares about the value of on the imaginary axis, we can without loss of generality (by
multiplying a solution to the model matching problem by an inner function that cancels
all poles in in C+ ) suppose that has no poles in C+ .

Let us outline the steps in performing the reduction of Problem 15.2 to Problem 15.3.
After we have said how to perform the reduction, we will actually prove that everything
works. The reader will wish to recall the notion of spectral factorisation for rational functions
(Proposition 14.17) and the notion of a coprime factorisation for a pair of rational functions
(Theorem 10.33).

22/10/2004

15.1 Reduction of robust performance problem to model matching problem

561

P ,
15.5 Algorithm for obtaining model matching problem for multiplicative uncertainty Given R
Wu , and Wp as in Problem 15.2.
1. Define
Wp Wp Wu Wu
U3 =
.
Wp Wp + Wu Wu
2. If kU3 k 12 , then Problem 15.2 has no solution.
P .
3. Let (P1 , P2 ) be a coprime fractional representative for R
4. Let (1 , 2 ) be a coprime factorisation for P1 and P2 :
1 P1 + 2 P2 = 1.

562

15 An introduction to H control theory

As we are working with multiplicative uncertainty (see Exercise E15.3 for additive uncertainty), the problem we start out with, of course, is to find a proper RC R(s) that
satisfies

|Wu TL |2 + |Wp SL |2 < 1

2
P , an uncertainty model Wu RH+ , and a performance
for a given nominal proper plant R

weight Wp RH+
. We choose a coprime fractional representative (P1 , P2 ) for RP and an
associated coprime factorisation (1 , 2 ). By Theorem 10.37, any proper stabilising controller
is then of the form
1 + P2
RC =
.
2 P1
A simple computation then gives

5. Define

TL = P1 P2 + 1 P1 ,
R1 = Wp 2 P2 ,
R 2 = W p P1 P 2 ,

S1 = Wu 1 P1 ,
S2 = Wu P1 P2 .

R 1 = W p 2 P 2 ,
R 2 = W p P1 P2 ,

R1 R2 + S1 S2
V,
Q

9. Define U4 = [ 12 U3 ]+ .
10. Define
T1 =

U1
,
U4

T2 =

|Wu TL |2 + |Wp SL |2 = |R1 R2 |2 + |S1 S2 |2 .

Up to this point, everything is simple enough. Now we claim that there exists functions
U1 , U2 RH+
and U3 R(s), defined in terms of R1 , R2 , S1 , and S2 , and having the
property that

|R1 R2 |2 + |S1 S2 |2 = |U1 U2 |2 + U3 .
(15.1)

U2 = QV.

U2
.
U4

11. Let be a solution to Problem 15.3, and by Remark 15.4 suppose that it has no poles
in C+ .
12. If kT1 T2 k 1 then Problem 15.2 has no solution.
13. The controller
1 + P2
RC =
,
2 P1
is a solution to Problem 15.2.

S 1 = W u 1 P 1 ,
S 2 = W u P1 P 2 ,

we then obtain

R1 R2 + S1 S2
V
Q

U1 =

SL = (P1 P2 2 P2 ).

Defining

6. Define Q = [R2 R2 + S2 S2 ]+ .
7. Let V be an inner function with the property that

has no poles in C+ .
8. Define

22/10/2004

The above procedure provides a way to produce a controller satisfying the modified
robust performance problem, provided one can find in Step 11. That is to say, we have
reduced the finding of a solution to the modified robust performance problem to that of
solving the model matching problem. It remains to show that all constructions made in
Algorithm 15.5 are sensible, and that all claims made are true. In the next section we will
do this formally. However, before we get into all the details, it is helpful to give a glimpse
into how Algorithm 15.5 comes about.

That these functions exist, and are as stated in Steps 1 and 8 of Algorithm 15.5, will be
proved in the subsequent section. Finally, with U4 as defined in Step 9, in the next section
we shall show that

|U1 U2 |2 + U3 < 1

kU41 U1 U41 U2 k < 1.

2
With this rough justification behind us, let us turn to formal proofs of the validity of
Algorithm 15.5. Readers not interested in this sort of detail can actually skip to Section 15.4.
P ,
15.1.3 Proof that reduction procedure works Throughout this section, we let R
Wp , and Wp are as stated in Problem 15.2.
In Step 6 of Algorithm 15.5, we are asked to compute the spectral factorisation of R2 R2 +
S2 S2 . Let us verify that this spectral factorisation exists.
15.6 Lemma R2 R2 + S2 S2 admits a spectral factorisation.
Proof We have
R2 R2 + S2 S2 = P1 P1 P2 P2 (Wp Wp + Wu Wu ).
Since P1 , P2

RH+

we may find an inner-outer factorisation

P1 = P1,in P1,out ,

P2 = P2,in P2,out

22/10/2004

15.1 Reduction of robust performance problem to model matching problem

563

by Proposition 14.10. Therefore

since P1,in is inner. Since P1,out is outer, it follows that P1,out is a left spectral factor for
P1 P1 . Similarly P2 P2 admits a spectral factorisation by the outer factor P2,out of P2 . Thus
P1 P1 and P2 P2 admit a spectral factorisation, and so too then does P1 P1 P2 P2 . Now let
(Np , Dp ) and (Nu , Du ) be the c.f.r.s of Wp and Wu . We then have

R1 R1 + S1 S1 U1 U1 = R1 R1 + S1 S1

Du ,

Our next result declares that U1 , U2 , and U3 are as they should be, meaning that they
satisfy the relation (15.1).
15.7 Lemma If U1 , U2 , and U3 as defined in Steps 8 and 1 satisfy

|R1 R2 |2 + |S1 S2 |2 = |U1 U2 |2 + U3 .

Proof It is sufficient that the relation

R1 R2 R1 R2 + S1 S2 S1 S2 = U1 U2 U1 U2 + U3 (15.2)
holds for all and for s = i. Doing the manipulation shows that if the three relations

R2 R2 + S2 S2 = U2 U2
R1 R2 + S1 S2 = U1 U2
(15.3)
R1 R1 + S1 S1 = U1 U1 + U3 .
hold for s = i, then (15.2) will indeed hold for all . We let Q = [R2 R2 + S2 S2 ]+ . If V is
an inner function with the property that

1
2

then Problem 15.2 has no solution.

U3 = R1 R1 + S1 S1

(R1 R2 + S1 S2 )(R1 R2 + S1 S2 )
R2 R2 + S2 S2

(this follows from (15.4)). Now work backwards through the proof of Lemma 15.7 to see
that

U3 = R1 R2 R1 R2 + S1 S2 S1 S2 U1 U2 U1 U2
for any admissible RH+
. Therefore, if kU3 k

then

However, a simple working through of the definitions of R1 , R2 , S1 , and S2 shows that this
implies that for any stabilising controller RC we must have

|Wu TL |2 + |Wp SL |2 1 ,

2
as desired.

Next we show that with T1 and T2 as defined in Step 10, the modified robust performance
problem is indeed equivalent to the model matching problem.
15.9 Lemma With T1 and T2 as defined in Step 10 we have

|Wu TL |2 + |Wp SL |2 < 1

2
where

2 + P1
.
1 P2

Proof As outlined at the end Section 15.1.2, the condition

|Wu TL |2 + |Wp SL |2 < 1

Thus U2 satisfies the first of equations (15.3). Also,

R1 R2 + S1 S2
V
Q

1
2

|R1 R2 |2 + |S1 S2 |2 1 .

U2 U2 = QV Q V = QQ = R2 R2 + S2 S2 .

U1 =

(15.4)

Proof One may verify by direct computation that

RC =
has no poles in C+ , then if U2 = QV we have

(R1 R2 + S1 S2 )(R1 R2 + S1 S2 )
,
R2 R2 + S2 S2

using our solution for U1 and the fact that V is inner. A straightforward substitution of the
definitions of R1 , R2 , S1 , and S2 now gives U3 as in Step 1.

15.8 Lemma If kU3 k

have no imaginary
and
Since Wp , Wu
by hypothesis, Dp and Du , and therefore
axis roots. Also by assumption, Np and Nu have no common roots on iR. One can then
show (along the lines of Exercise E14.2) that this infers that Np Np Du Du + Nu Nu Dp Dp has
constant sign on iR. Since it is clearly even, one may infer from Proposition 14.12 that
Np Np Du Du + Nu Nu Dp Dp admits a spectral factorisation. Therefore, by Proposition 14.17,
so too does Wp Wp +Wu Wu . Finally, by Exercise E14.3 we conclude that R2 R2 +S2 S2 admits
a spectral factorisation.

R1 R2 + S1 S2
V
Q

22/10/2004

Our next lemma verifies Step 2 of Algorithm 15.5.

Np Np Du Du + Nu Nu Dp Dp
.
Dp Dp Du Du

RH+

15 An introduction to H control theory

clearly satisfies the second of equations (15.3). To verify the last of equations (15.3), we may
directly compute

P1 P1 = P1,in P1,out P1,in

P1,out
= P1,out P1,out
,

Wp Wp + Wu Wu =

564

is equivalent to

|U1 U2 |2 + U3 < 1 .

kT1 T2 k ,

22/10/2004

15.1 Reduction of robust performance problem to model matching problem

565

566

Also note that by definition, U3 (i) 0 for all R, and that U3 = U3 , the latter fact
implying that U3 is even. Therefore, provided that kU3 k < 12 , 21 U3 admits a spectral
factorisation. Now we compute

|U1 U2 |2 + U3 < 1

2

|U1 (i) (i)U2 (i)|2 + U3 (i) < 1 , R
2
1
2
|U1 (i) (i)U2 (i)| + U3 (i) < , R
2
1
2
|U1 (i) (i)U2 (i)| < U3 (i), R
2
|U1 (i) (i)U2 (i)|2 < [ 21 U3 (i)]+ [ 12 U3 (i)],
|U1 (i) (i)U2 (i)|2 < |U4 (i)|2 , R
1

U4 (i)U1 (i) (i)U41 (i)U2 (i)2 < 1,

finish

It is necessary that T1 , T2
in order to fit them into the model matching problem.
The next lemma ensures that this follows from the constructions we have made.
15.10 Lemma T1 , T2 RH+
.
1
2

Proof First note that kU3 k < by the time we have gotten to defining T1 and T2 . Therefore
U4 = [ 21 U3 ]+ is strictly proper, and so is invertible in RH+
. The lemma will then follow
if we can show that U1 , U2 RH+
.
First let us show that U2 RH+
. By definition, U2 is the left half-plane spectral factor
of
P1 P1 P2 P2 (Wp Wp + Wu Wu ).
As such, it is the product of the two quantities
[Wp Wp + Wu Wu ]+ .

+
Since each of P1 P2 RH+
, [P1 P1 P2 P2 ] RH . Since

we have
[Wp Wp + Wu Wu ]+ =

Np Np Du Du + Nu Nu Dp Dp
,
Dp Dp Du Du
[Np Np Du Du + Nu Nu Dp Dp ]+
.
Dp Du

+
+

+
Since Wp , Wu RH+
, it follows that [Wp Wp + Wu Wu ] RH . Thus U2 RH .
Now let us show that U1 RH+
.
A
computation
shows
that

U1 =

2 P2 Wp Wp 1 P1 Wu Wu have poles on iR. Thus, our claim will follow if we can show that
P1 P2
has no poles on iR. This is true since the imaginary axis zeros of P1 P2 and
[P1 P2 P2 P2 ]

[P1 P2 P2 P2 ] agree in location and multiplicity. Thus we have shown that U1 is analytic in
C+ . That U1 RH+

will now follow if we can show that U1 is proper.

The first solution we shall give to the model matching problem comes from a seemingly
unrelated interpolation problem. To state the problem, we need a little notation. We let
+,C
RH+,C
be the collection of proper functions in C(s) with no poles in C+ . Thus RH is just
+
like RH , except that now we allow the functions to have complex coefficients. Note that
kk still makes sense for functions in RH+,C
. Now the interpolation problem is as follows.

RH+

Wp Wp + Wu Wu =

22/10/2004

15.2 Optimal model matching I. Nevanlinna-Pick theory

By definition of T1 and T2 , the lemma follows.

[P1 P1 P2 P2 ]+ ,

15 An introduction to H control theory

2 P2 Wp Wp 1 P1 Wu Wu
P1 P2
V.

[P1 P2 P2 P2 ]
[Wp Wp + Wu Wu ]

The inner function V is designed so that this function has no poles in C+ . We also claim
that U1 has no poles on iR. Since Wp , Wu RH+
and since they have no common imaginary
axis zeros, it follows that [Wp Wp + Wu Wu ] has no zeros on iR. Clearly, neither P1 P2 nor

15.11 Nevanlinna-Pick interpolation problem Let {a1 , . . . , ak } C+ and let {b1 , . . . , bk } C

collections of distinct points. A Pick pair is then a pair (aj , bj ), i {1, . . . , k}. Suppose
that if (ak , bk ) is a Pick pair, then so is (
aj , bj ), j = 1, . . . , k.
Find R RH+
so that
(i) kRk 1 and
(ii) R(aj ) = bj , j = 1, . . . , n.

This problem was originally solved by Pick [1916], and was solved independently by
Nevanlinna [1919]. Nevanlinna also gave an algorithm for finding a solution to the interpolation problem [Nevanlinna 1929]. In this section, we will state and prove Picks necessary and
sufficient condition for the solution of the interpolation problem, and also give an algorithm
for determining a solution.
15.12 Remark We should say that the problem solved by Pick is somewhat different than
the one we state here. The difference occurs in three ways.
1. Pick actually allowed kRk = 1. However, our purposes will require that the H -norm
of R be strictly less than 1.
2. Pick was not interested in making the restriction that if a every Pick pair have the
property that its complex conjugate also be a Pick pair.
3. Pick was interested in the case where the points a1 , . . . , an lie in the open disk D(0, 1)
of radius 1 centred at 0. This is not a genuine distinction, however, as the map s 7 1s
1+s
bijectively maps C+ onto D(0, 1), and so translates the domain of concern for Pick to
our domain.
4. Finally, Pick allowed interpolating functions to be general meromorphic functions,
bounded and analytic in C+ . For obvious reasons, our interest is in the subset of such
functions that are in R(s), i.e., functions in RH+

.
15.2.1 Picks theorem Picks conditions for the existence of a solution to Problem 15.11 are quite simple. The proof that these conditions are necessary and sufficient
is not entirely straightforward. In this section we only prove necessity, as sufficiency will
follow from our algorithm for solving the Nevanlinna-Pick interpolation problem in the ensuing section. Our necessity proof follows [Doyle, Francis, and Tannenbaum 1990]. For the
t
statement of Picks theorem, recall that M Cnn is Hermitian if M = M = M .

15.2 Optimal model matching I. Nevanlinna-Pick theory

22/10/2004

567

568

A Hermitian matrix is readily verified to have real eigenvalues. Therefore, the notions of
definiteness presented in Section 5.4.1 may be applied to Hermitian matrices.

15 An introduction to H control theory

We now compute
Z

e(aj +a` )t dt =

15.13 Theorem (Picks Theorem) Problem 15.11 has a solution if and only if the Pick matrix , the complex k k symmetric matrix M with components
1 bj b`
,
Mj` =
aj + a
`

k
X

j, ` = 1, . . . , k,

j,`=1

k
X

where

aj t

cj e .

j=1

k
X
j=1

k
X

hR ( )eaj (t ) d

cj eaj t

hR ( )eaj dt

cj eaj t R(aj )

j=1

k
X

c1

x = ... .
cn

This can be considered as an input to the transfer function R, with the output computed
by separately computing the real and imaginary parts. Moreover, if hR denotes the inverse
Laplace transform for R, the complex output will be
Z
hR ( )u(t ) d
y(t) =
cj

1 bj b`
c` 0,
aj + a
`

x M x 0,

j=1

which we recognise as being equivalent to the expression

is positive-semidefinite.

k
X

1
,
aj + a
`

thus giving

Proof of necessity Suppose that Problem 15.11 has a solution R. For c1 , . . . , ck C, not all
zero, consider the complex input u : (, 0] C given by
u(t) =

22/10/2004

cj bj eaj t ,

Thus we have shown that the Pick matrix is positive-definite.

Check

15.2.2 An inductive algorithm for solving the interpolation problem In this section
we provide a simple algorithm for solving the Nevanlinna-Pick interpolation problem in the
situation when the Pick matrix is positive-definite. In doing so, we also complete the proof
of Theorem 15.13. The algorithm we present in this section follows [Marshall 1975].
Before we state the algorithm, we need to introduce some notation. First note that by
the Maximum Modulus Principle, it is necessary that for the Nevanlinna-Pick interpolation
problem to have a solution, each of the numbers b1 , . . . , bk satisfy |bj | 1. Thus we may
assume this to be the case when we seek a solution to the problem. For b C satisfying
|b| < 1, define the Blaschke function Bb R(s) associated with b by

,
bR

1 bs
2
Bb (s) =
2
s + 2Re(b)s + |b|

, otherwise.

1 + 2Re(b)s + |b|2 s2

j=1

where we have used the definition of the Laplace transform. By part (i) of Theorem 5.21,
and since kRk 1, it follows that
Z 0
Z 0
|y(t)|2 dt
|u(t)|2 dt.

Substituting the current definitions of u and y gives

Z 0 X
Z 0 X
k
k
2
2

cj bj eaj t dt
cj eaj t dt

j=1

k
X

j,`=1

Z
=

k
X

j,`=1

j=1

cj eaj t c` ea` t dt

k
X

cj bj eaj t c`b` ea` t dt

j,`=1

cj c` (1 bj b` )e(aj +a` )t dt 0.

Some easily verified relevant properties of Blaschke functions are the subject of Exercise E15.4. Also, for a C define a function Aa R(s) by

,
aR

s+a

Aa (s) = s2 2Re(a)s + |a|2

, otherwise.
2
s + 2Re(a)s + |a|2
Again, we refer to Exercise E15.4 for some of the easily proven properties of such functions.
Let us begin by solving the Nevanlinna-Pick interpolation problem when k = 1.
15.14 Lemma Let a1 C+ and let b1 C have the property that |b1 | < 1. The associated Nevanlinna-Pick interpolation problem has an infinite number of solutions if it has one
solution, and the set of all solutions is given by

Re(R) R(s) = Bb1 R1 (s)Aa1 (s) , R1 C(s) has no poles in C+ , and kR1 k < 1 .

22/10/2004

15.2 Optimal model matching I. Nevanlinna-Pick theory

569

Proof First note that the Nevanlinna-Pick interpolation problem does indeed have a solution, namely the trivial solution R0 (s) = b1 .
Now let R1 C(s) have no poles in C+ and suppose that kR1 k < 1. If

R(s) = Bb1 R1 (s)Aa1 (s)
then R is the composition of the functions

570

15 An introduction to H control theory

22/10/2004

The lemmas gives the form of all solutions to the Nevanlinna-Pick interpolation problem
in the cases when k = 1 and k = 2 with the points being complex conjugates of one another.
It turns out that with this case, one can construct solutions to the general problem. To
do this, one makes the clever observation (this was Nevanlinnas contribution) that one can
reduce a k point interpolation problem to a k 1 or k 2 point interpolation problem by
properly defining the new k 1 or k 2 points. We say how this is done in a definition.

s 7 R1 (s)Aa1 (s)
s 7 Mb1 (s).

15.16 Definition For k > 1, let

The first of these functions in analytic in C+ since both R1 and Aa1 are. Also, by Exer 1).
cise E15.4 and since kR1 k < 1, the first of these functions maps C+ onto the disk D(0,
1) and maps it onto itThe second of these functions, by Exercise E15.4, is analytic in D(0,
self. Thus we can conclude that R C(s) as defined has no poles in C+ and that kRk < 1.
Whats more, we claim that R(a1 ) = b1 if R1 (a1 ) = b1 . Indeed

R(a1 ) = Bb1 R1 (a1 )Aa1 (a1 ) = Bb1 (0) = b1 ,

be as in Problem 15.11. Define

{a1 , . . . , ak }, {b1 , . . . , bk } C

using the definitions of Aa1 and Bb1 . It only remains to show that Re(R) solves Problem 15.11. Clearly, since b1 must be real, it follows that Re(R(a1 )) = b1 . Furthermore, since
kRe(R)k < kRk , it follows that kRe(R)k < 1. Finally, Re(R) can have no poles in C+
since R has no poles in C+ .
Now suppose that R R(s) solves Problem 15.11. Define R1 C(s) by

k =

(
k 1, bk R
k 1, otherwise.

The Nevanlinna reduction of the numbers {a1 , . . . , ak } and {b1 , . . . , bk } is the collection
of numbers
{
a1 , . . . , a
k }, {b1 , . . . , bk } C
defined by

With this in mind, we state the algorithm that forms the main result of this section.
15.17 Algorithm for solving the Nevanlinna-Pick interpolation problem Given points

Bb (R(s))
R1 (s) = 1
.
Aa1 (s)

{a1 , . . . , ak }, {b1 , . . . , bk } C

The function in the numerator is analytic in C+ and has a zero at s = a1 . Therefore, since
the only zero of Aa1 is at zero, R1 is analytic in C+ . Furthermore, the H -norm of the
numerator is strictly bounded by 1, and since the H -norm of the denominator equals 1, we
conclude that kR1 k < 1. This concludes the proof of the lemma.

as in Problem 15.11, additionally assume that |bj | < 1, j = 1, . . . , k, and that the Pick
matrix is positive-definite.
1.

As stated in the proof of the lemma, if |b1 | < 1, then the one point interpolation problem
always has the trivial solution R0 (s) = b1 . Let us also do this in the case when k = 2 and
we have a2 = a
1 6= a1 and b2 = b2 6= b2 .

15.2.3 Relationship to the model matching problem The above discussion of the
Nevanlinna-Pick interpolation problem is not obviously related to the model matching problem, Problem 15.3.

15.15 Lemma Let {a1 , a2 = a

1 } C+ and let {b1 , b2 = b1 } C have the property that |b1 | <
1. Also suppose that a1 6= a2 and b1 6= b2 . The associated Nevanlinna-Pick interpolation
problem has an infinite number of solutions if it has one solution, and the set of all solutions
is given by

Re(R) R(s) = Bb1 R1 (s)Aa1 (s) , R1 C(s) has no poles in C+ , and kR1 k < 1 .

15.18 Model matching by Nevanlinna-Pick interpolation Given T1 , T2 RH+

Proof Problem 15.11 has the solution

Rs (s) =

15.3 Optimal model matching II. Neharis Theorem

In this section we present another method for obtaining a solution, or an approximate
solution, to the model matching problem, Problem 15.3. The strategy in this section involves
significantly more development than does the Nevanlinna-Pick procedure from Section 15.2.
However, the algorithm produced in actually easier to apply than is that using NevanlinnaPick theory. Unfortunately, the methods in this section suffer from on occasion producing
a controller that is improper, and one must devise hacks to get around this, just as with
Nevanlinna-Pick theory. Our presentation in this section follows that of Francis [1987].

22/10/2004

15.3 Optimal model matching II. Neharis Theorem

571

15.3.1 Hankel operators in the frequency-domain The key tool for the methods
of this section is something new for us: a Hankel operator of a certain type. To initiate
this discussion, let us note that as in Proposition 14.9, but restriction to functions in RL2 ,
+
we have a decomposition RL2 = RH
2 RH2 . That is to say, any strictly proper rational
function with no poles on iR has a unique expression as a sum of a strictly proper rational
function with no poles in C+ and a strictly proper rational function with no poles in C .
This is no surprise as this decomposition is simply obtained by partial fraction expansion.
The essential idea of this section puts this mundane idea to good use. Let us denote by

+ : RL2 RH+
2 and : RL2 RH2 the projections.
Let us list some operators that are readily verified to have the stated properties.
1. The Laurent operator with symbol R: Given R RL and Q RL2 , one readily
sees that RQ RL2 . Thus, given R RL we have a map R : RL2 RL2 defined by
R (Q) = RQ. This is the Laurent operator with symbol R.
2. The Toeplitz operator with symbol R: Clearly, if R RL and if Q RH+
2 , then
+
+
RQ RL2 . Therefore, + (RQ) RH+
2 . Thus, for R RL , the map R : RH2 RH2
defined by R (Q) = + (RQ) is well-defined, and is called the Toeplitz operator with
symbol R.
3. The Hankel operator with symbol R: Here again, if R RL and if Q RH
2,

then RQ RL2 . Now we map this rational function into RH+

2 using the projection .

Thus, for R RL , we define a map R : RH+

2 RH2 by R (Q) = (RQ). This is
the Hankel operator with symbol R.

572

15 An introduction to H control theory

+
for R1 , R2 RL2 . In like fashion, the Toeplitz operator has an adjoint R : RH+
2 RH2
defined by
hR (R1 ), R2 i2 = hR1 , R (R2 )i2 , R1 , R2 RH+
2,
+
and the Hankel operator has an adjoint R : RH
2 RH2 defined by

hR (R1 ), R2 i2 = hR1 , R (R2 )i2 ,

15.20 Proposition For R RL2 the following statements hold:

(i) R = R ;
(ii) R = R ;
(iii) R (Q) = + (R (Q)).
Proof (i) We compute
hR (R1 ), R2 i2 =
=
=
=
=

R (Q) = R (Q) + R (Q).

The Laurent, Toeplitz, and Hankel operators are linear. Thus it makes sense to ask
+
questions about the nature of their spectrum. However, the spaces RL2 , RH
2 , and RH2
are infinite-dimensional, so these issues are not immediately approachable as they are in
finite-dimensions. The good news, however, is that these operators are essentially finitedimensional. The easiest way to make sense of this is with state-space techniques, and this
is done in the next section.
It also turns out that the Laurent, Toeplitz, and Hankel operators are defined on spaces
with an inner product. Indeed, on RL2 we may define an inner product by
Z
1
R1 (i)R2 (i) d.
(15.5)
hR1 , R2 i2 =
2
Note that this is an inner product on a real vector space. This inner product may clearly be
+
applied to any functions in RL2 , including those in the subspaces RH
2 and RH2 . Indeed,

+
RH2 and RH2 are orthogonal with respect to this inner product (see Exercise E15.5). One
may define the adjoint of any of our operators with respect to this inner product. The
adjoint of the Laurent operator with symbol R is the map R : RL2 RL2 defined by the
relation
hR (R1 ), R2 i2 = hR1 , R (R2 )i2

R1 RH+
2 , R2 RH2 .

The following result gives explicit formulae for the adjoints.

15.19 Remarks 1. Note that the Toeplitz and Hankel operators together specify the value of
the Laurent operator when applied to functions in RH+
2 . That is to say, if R RL2 then

2. It is more common to see the Laurent, Toeplitz, and Hankel operators defined for general
analytic functions rather than just rational functions. However, since our interest is
entirely in the rational case, it is to this is that we restrict our interest.

22/10/2004

Z
1
R (R1 )(i)R2 (i) d
2
Z
1
R(i)R1 (i)R2 (i) d
2
Z
1
R1 (i)R(i)R2 (i) d
2
Z
1
R1 (i)R (i)R2 (i) d
2
hR1 , R (R2 )i2 .

This then gives

= R as desired.
(ii) For R1 , R2 RH+
2 we compute
hR (R1 ), R2 i2 =
=
=
=

hR (R1 ), R2 i2
hR1 , R (R2 )i2
hR1 , R (R2 )i2
hR1 , R (R2 )i2 ,

and this part of the proposition follows.

(iii) For R1 RH+

2 and R2 RH2 we compute
hR (R1 ), R2 i2 = hR (R1 ), R2 i2
= hR1 , R (R2 )i2
= hR1 , R (R2 )i2 ,
and from this the result follows.

In the next section, we will come up with concrete realisations of the Hankel operator
and its adjoint using time-domain methods.

15.3 Optimal model matching II. Neharis Theorem

22/10/2004

573

15.3.2 Hankel operators in the time-domain The above operators defined in the rational function domain are simple enough, but they have interesting and nontrivial counter 2 (, ) those functions
parts in the time-domain. To simplify matters, let us denote by L
of time that, when Laplace transformed, give functions in RL2 . As we saw in Section E.3,
this consists exactly of sums of products of polynomial functions, trigonometric functions,
2 (, 0] the subset of L
2 (, )
and exponential functions of time. Let us denote by L
2 [0, ) the subset of L
2 (, )
consisting of functions that are bounded for t < 0, and by L
consisting of functions that are bounded for t > 0. Note that
2 (, ) = L
2 (, 0] L
2 [0, ).
L

f1 (t)f2 (t) dt

(15.6)

RL2

/L
2 [0, )
L

R (u) L
2 (, 0].
15.22 Lemma
Proof For t 0 we have

/ RL+
2

eA bu( ) d.

Since A has all eigenvalues in C+ , A has all eigenvalues in C , so the integral converges.
R (u) L
2 (, 0], as
Also, for the same reason, eAt is bounded for t 0. This shows that
claimed.

Now let us show that this is indeed the same as the frequency domain Hankel operator.
+
15.23 Proposition Let R RL and write R = R1 + R2 with R1 RH
2 and R2 RH .

Let r1 L2 (, 0] be the inverse Laplace transform of R1 and define R as in (15.7). Then

the diagram

2 [0, )
L

15.21 Proposition The Laplace transform is a bijection

2 (, ) to RL2 ,
(i) from L
2 (, 0] to RH , and
(ii) from L
2
2 [0, ) to RH+
(iii) from L
2.
Furthermore, the diagrams
+

R (u)(t) = 0 for t > 0. We claim that

R is the time-domain version of the Hankel
We take
operator R . Let us first prove that its takes its values in the right space.

2 (, ), then obviously L
2 (, 0] and L
2 [0, ) are orthogonal. We hope that it will
on L
be clear from context what we mean when we use the symbol h, i2 in two different ways,
one for the frequency-domain, and the other for the time-domain.
The following result summarises the previous discussion.

2 (, )
L

2 [0, ) and for t 0, let us

which is the anticausal impulse response for R1 . Now, for u L
define
Z
R (u)(t) =
r1 (t )u( ) d.
(15.7)

R (u)(t) = ct eAt

If we employ the inner product

hf1 , f2 i2 =

22/10/2004

2 (, 0] is the inverse
the impulse response for 1 . Note that (A) C+ . Thus if r1 L
Laplace transform of R1 we have
(
ct eAt b, t 0
r1 (t) =
0,
t > 0,

2 (, ) L
2 [0, )
L

2 (, 0].
L2 (, ) L

15 An introduction to H control theory

2 (, ) can be uniquely decomposed into a sum of two funcThat is, every function in L
tions, one that is bounded for t < 0 and one that is bounded for t > 0. It is clear that this
+
decomposition corresponds exactly to the decomposition RL2 = RH
2 RH2 that uses the
partial fraction expansion. Let us also define projections
+:

574

RL+
2

/L
2 (, 0]
L

/ RL
2

commutes.

2 (, )
L

RL2

/L
2 (, 0]
L

/ RL
2

commute.
We now turn our attention to describing how the operators of Section 15.3.1 appear in
the time-domain, given the correspondence of Proposition 15.21. Our interest is particularly
in the Hankel operator. Given Proposition 15.21 we expect the analogue of the frequency
2 [0, ) to L
2 (, 0], given R RL . To do this, given
domain Hankel operator to map L
+
R RL , let us write R = R1 + R2 for R1 RH
2 and R2 RH as in Proposition 14.9.
We then let 1 = (A, b, ct , 01 ) be the complete SISO linear system in controller canonical
form with the property that T1 = R1 . Therefore, the inverse Laplace transform for R1 is

2 [0, ) and denote y =

R (u) L(,

Proof Let u L
0]. Denote as usual the Laplace
transforms of u and y by u and y. We then have
R (
u) = ((R1 + R2 )
u).

Note that R2 u RH+

u) = (R1 u). Now let us compute the
2 . Therefore, ((R1 + R2 )

t , 01 ) be a complete SISO linear system

inverse Laplace transform of R1 u. Let = (A, b, c
has all eigenvalues in C . Now we compute
defined so that T = u. Thus A
Z
L 1 (R1 u)(t) =
r1 (t )u( ) d

Z
= ct eAt
1( t)eA bu( ) d

Z
= ct eAt
1( t)eA bu( ) d,
0

15.3 Optimal model matching II. Neharis Theorem

22/10/2004

575

(L 1 (R1 u)) is nonzero only for t < 0 so that

since for < 0, u( ) = 0. Now note that
we can write
Z
(L 1 (R1 u)))(t) = ct eAt
(
eA bu( ) d.
0

(L 1 (R1 u)) =
R (
Thus
u). By Proposition 15.21 this means that

R , as claimed.
or, equivalently, that R L = L

Up to this point, the value of the time-domain formulation of a Hankel operator is

not at all clear. The simple act of multiplication and partial fraction expansion in the
frequency-domain becomes a little more abstract in the time-domain. However, the value
of the time-domain formulation is in its presenting a concrete representation of the Hankel
operator and its adjoint. To come up with this representation, we introduce some machinery
harking back to our observability and controllability discussion in Sections 2.3.1 and 2.3.2.
In particular, we begin to dig into the proof of Theorem 2.21. We resume with the situation
+
where R = R1 + R2 RL with R1 RH
2 and R2 RH . As above, we let 1 =
t
(A, b, c , 01 ) be the canonical minimal realisation of R1 . With this notation, we define a
2 [0, ) Rn by
map CR : L
Z
eA bu( ) d,
CR (u) =
0

2 (, 0]
and we call this the controllability operator . Similarly, we define OR : Rn L
by
(
ct eAt x, t 0
(OR (x))(t) =
0,
t < 0,
and we call this the observability operator . Note that the adjoint of the controllability
2 [0, ) satisfying
operator will be the map CR : Rn L
2 [0, ), x Rn ,
uL

2 (, 0] Rn satisfying
and the adjoint of the observability operator will be the map OR : L
hOR (x), yi2 = hx, OR (y)i ,

(iii) OR (y) =

15 An introduction to H control theory

22/10/2004

eA t cy(t) dt;
( R
t At (t )
cy(t) dt, 0
(y))( ) = 0 b e
(iv) (
R
0,
> 0;

(v) CR is injective;

R (u),
L 1 ( (R1 u)) = L 1 (R (
u)) =

hCR (u), xi = hu, CR (x)i2 ,

576

Proof (i) It suffices to show that the left diagram commutes, since if it does, the right
diagram will also commute by the definition of the adjoint. However, the left diagram may
R.
be easily seen to commute by virtue of the very definitions of CR , OR , and
(ii) This follows from the definition of CR and the inner product in equation (15.6).
(iii) This follows from the definition of OR and the inner product in equation (15.6).
(iv) This follows from the right diagram in part (i), along with parts (ii) and (ii).
t
(v) It suffices to show that CR (x) = 0 if and only if x = 0. If (CR (x))( ) = bt eA x = 0
for all then successive differentiation with respect to and evaluation at = 0 gives
bt x = 0,

bt At x = 0, . . . , (1)n bt (At )n1 x = 0.

Since (A, b) is controllable, this implies that x = 0.

Thus the above result gives a simple way of relating a Hankel operator and its adjoint
to operators with either a domain or a range that is finite-dimensional. In the next section,
we shall put this to good use.
15.3.3 Hankel singular values and Schmidt pairs Recall that a singular value for
a linear map A between inner product spaces is, by definition, an eigenvalue of A A. One
readily verifies that singular values are real and nonnegative. Our objective in this section
R in the
is to find the singular values of a Hankel operator R using our representation
time-domain.
As in the previous section, for R RL we write R = R1 + R2 with R1 RH+
2 and
t
R2 RH
. We let 1 = (A, b, c , 01 ) be the complete SISO linear system in controller
canonical form so that T1 = R1 . We next introduce the controllability Gramian
Z
t
eAt bbt eA t dt,
CR =
0

2 (, 0], x Rn .
yL

2 (, 0] and L
2 [0, ).
In each case, the inner product of equation (15.6) is being used on L
The following result summarises the value of introducing this notation.

and the observability Gramian

Z
OR =

eA t cct eAt dt.

15.24 Proposition With the above notation, the following statements hold:
(i) the diagrams
R J
JJ
v:
JJOR
vv
JJ
vv
v
JJ
v
v
v
$
/L
2 [0, )
2 (, 0]
L

commute;
(
t
bt eA x, t 0

(ii) (CR (x))( ) =

0,
t > 0;

R H
t:
HH
tt
HHCR
HH
tt
t
HH
t
t
$
t
/L
2 [0, )
2 (, 0]
L

These are each elements of Rnn . We have previously encountered the controllability
Gramian in the proof of Theorem 2.21, and the observability Gramian may be used in a
similar manner. However, here we are interested in their relationship with the Hankel operator R . The following result gives this relationship, as well as providing a characterisation
of C R and O R in terms of the Liapunov ideas of Section 5.4.

15.25 Proposition With the above notation, the following statements hold:
(i) (At , C R , bbt ) is a Liapunov triple;
(ii) (A, O R , cct ) is a Liapunov triple;

22/10/2004

15.3 Optimal model matching II. Neharis Theorem

577

578

(iii) CR CR = C R ;
(iv) OR OR = O R ;
(v) O R is invertible.

15 An introduction to H control theory

22/10/2004

If u = CR O R x then we claim that u 6= 0. Indeed, from part (v) of Proposition 15.24, CR is

injective, and from part (v) of Proposition 15.25, O R is injective. Thus u = 0 if and only if

x = 0. Thus we see that 2 is an eigenvalue for

R R with eigenvalue u.
Thus we have a characterisation of all nonzero singular values of the Hankel operator as
eigenvalues of an n n matrix. This is something of a coup. We shall suppose that the
nonzero singular values are arranged in descending order 1 2 k , so that 1
denotes the largest of the singular values. We shall call 1 , . . . , k the Hankel singular
values for the Hankel operator R .
Now we wish to talk about the size of a Hankel operator R . Since R is a linear map

between two inner product spacesfrom RH+

2 to RH2 we may simply define its norm in
the same manner in which we defined the induced signal norms in Definition 5.19. Thus we
define
kR (Q)k2
kR k = sup
.
kQk2
QRH+

Proof (i) Since A is Hurwitz, by part (i) of Theorem 5.32 there is a unique symmetric
matrix P so that (At , P , bbt ) is a Liapunov triple. Whats more, the proof of Theorem 5.32
gives P explicitly as
Z

eAt bbt eA t dt.

P =
0

Now one sees trivially that (A, P , b, bt ) is also a Liapunov triple. This part of the
proposition now follows because C R = P .
(ii) The proof here is exactly as for part (i).
(iii) This follows from the characterisations of CR and CR given in Proposition 15.24.
(iv) This follows from the characterisations of OR and OR given in Proposition 15.24.
(v) Since O R is square, injectivity is equivalent to invertibility. Suppose that O R is not
invertible. Then, since O R is positive-semidefinite, there exists x Rn so that xt O R x = 0,
or so that
Z
t
xt eA t cct eAt x dt.

Q not zero

This is called the Hankel norm of the Hankel operator R . The following result follows
easily from Theorem 15.26 if one knows just a little more operator theory than is really
within the confines of this course. However, it is an essential result for us.

This means that ct eAt x = 0 for all t [0, ). Differentiating successively with respect to
t at t = 0 gives
ct x = 0, ct Ax = 0, . . . , (1)n1 ct An1 x = 0.
This implies that (A, c) is not observable. It therefore follows that O R is indeed injective.
This, then, is interesting as it affords us the possibility of characterising the singular
values of the Hankel operator in terms of the eigenvalues of an n n matrix. This is
summarised in the following result.
15.26 Theorem The nonzero eigenvalues of the following three operators,
(i) R R ,

R , and
(ii)
(iii) C R O R ,
agree.

Proof That the eigenvalues for R R and

R R agree is a simple consequence of Proposition 15.23: the Laplace transform or its inverse will deliver eigenvalues and eigenvectors for

either of R R or
R R given eigenvalues and eigenvectors for the other.

Now let 2 > 0 be an eigenvalue for

R R with eigenvector u L2 (, 0]. By part (i)
of Proposition 15.24 this means that
CR OR OR CR (u) = 2 u
CR CR OR OR CR (u) = 2 CR (u).

If x = CR (u) then x 6= 0 since otherwise it would follow that = 0. This shows that is
an eigenvalue of C R O R with eigenvector x.
Now suppose that 2 6= 0 is an eigenvalue for C R O R with eigenvector x. Thus
2

C R OR x = 2 x
CR O R C R O R = 2 CR O R x.

Proof We can prove this using matrix norms.

Next, let us look a little more closely at eigenvectors induced by singular values. Thus

we let 2 be a nonzero singular value for
R R with eigenvector u1 L2 [0, ). Now define
2 (, 0] by u2 = 1
R (u1 ). Then one readily computes
u2 L

R (u1 ) = u2

R (u2 ) = u1 .

When an operator and its adjoint possess the same eigenvalue in this manner, the resulting
eigenvectors (u1 , u2 ) are called a -Schmidt pair for the operator. Of course, if Rj is the
Laplace transform of uj , j = 1, 2, then we have

15.27 Corollary If 1 is the largest Hankel singular value, then kR k = 1 .

finish

R (R1 ) = R2
R (R2 ) = R1 ,

so that (R1 , R2 ) RH+

2 RH2 are a -Schmidt pair for R . The matter of finding Schmidt
pairs for Hankel operators is a simple enough proposition as one may use Theorem 15.26.
Indeed, suppose that 2 > 0 is an eigenvalue for C R O R with eigenvector x. Then a R is readily verified to be given by (u1 , u2 ) where
Schmidt pair for

1
O (O R x)
R
u2 = OR (x).
u1 =

15.3.4 Neharis Theorem In this section we state and prove a famous theorem of Nehari [1957]. This theorem is just one in a sweeping research effort in Hankel norm approximation, with key contributions being made in a sequence of papers by Adamjan, Arov, and

22/10/2004

15.3 Optimal model matching II. Neharis Theorem

579

Krein [1968a, 1968b, 1971]. Our interest in this section is in a special version of this rather
general work, as we are only interested in rational functions, whereas Nehari was interested
in general H functions.
15.28 Theorem Let R0 RH
, let 1 > 0 be the largest Hankel singular value for R0 , and

let (R1 , R2 ) RH+

2 RH2 be a 1 -Schmidt pair. Then
inf

RRH+

kR0 Rk = 1 ,

Proof First let us show that 1 is a lower bound for kR0 Rk . For any R RH+
2 we
compute, using part (i) of Theorem 5.21,
sup
QRH+
2

k(R0 R)Qk2
kQk2

sup
QRH+
2
Q not zero

15.3.5 Relationship to the model matching problem The previous buildup has been
significant, and it is perhaps not transparent how Hankel operators and Neharis Theorem
relate in any way to the model matching problem. The relationship is, in fact, quite simple,
and in this section we give a simple algorithm for obtaining a solution to the model matching
problem using the tools of this section. However, as with Nevanlinna-Pick theory, there is a
drawback in that on occasion a hack will have to be employed. Nonetheless, the process is
systematic enough.
Let us come right out and state the algorithm.

15.4 A robust performance example

It turns out that the robust performance problem is only one of a number of problems
falling under the umbrella of H control. In this section we briefly indicate some other
problems whose solution can be reduced to a model matching problem, and thus whose
solution can be obtained by the methods in this chapter.

k (R0 R)Qk2
kQk2
k (R0 )Qk2
kQk2

= kR k .
Now let (R1 , R2 ) be a 1 -Schmidt pair and write Q = (R0 R)R1 for R RH+
. Since
+
+

R1 RH+
2 , R0 (R1 ) RH2 . Since RR1 RH2 , (Q) = (R0 R1 ) = R0 (R1 ). Therefore,
we compute
0 kQ R0 (R1 )k22
= kQk22 + hR0 (R1 ), R0 (R1 )i2 2 hQ, R0 (R1 )i2

= kQk22 + hR0 (R1 ), R0 (R1 )i2 2 (Q), R0 (R1 ) 2
= kQk22 hR0 (R1 ), R0 (R1 )i2

= kQk22 R1 , R0 R0 (R1 ) 2
= kQk22 12 hR1 , R1 i2
= kQk22 12 kR1 k22
kR0 Rk2 kR1 k22 12 kR1 k22

= kR0 Rk2 12 kR1 k22
0.
This shows that Q = R0 (R1 ), or, equivalently,
(R0 R)R1 = R0 (R1 ) = 1 R2 ,
as claimed.

22/10/2004

15.5 Other problems involving H methods

Q not zero

15 An introduction to H control theory

15.29 Model matching by Hankel norm approximation Given T1 , T2 RH+

and if R RH+
satisfies R1 (R0 R) = 1 R2 then kR0 Rk = 1 .

kR0 Rk =

580

Exercises for Chapter 15

581

582

Exercises
E15.1 Exercise in the -norm giving a Banach algebra.
E15.2 Exercise on existence of solutions to the model matching problem.
E15.3 Verify that the following algorithm for reducing the modified robust performance
problem for additive uncertainty actually works.
15.30 Algorithm for obtaining model matching problem for additive uncertainty Given
P , Wu , and Wp as in Problem 15.2.
R
1. Define
Wp Wp Wu Wu
U3 =
.
Wp Wp + Wu Wu
2. If kU3 k 21 , then Problem 15.2 has no solution.
P .
3. Let (P1 , P2 ) be a coprime fractional representative for R
4. Let (1 , 2 ) be a coprime factorisation for P1 and P2 :
1 P1 + 2 P2 = 1.
5. Define
R 1 = W p 2 P 2 ,
R 2 = W p P1 P 2 ,

S1 = Wu 1 P1 ,
S2 = Wu P1 P2 .

6. Define Q = [R2 R2 + S2 S2 ]+ .
7. Let V be an inner function with the property that
R1 R2 + S1 S2
V
Q
has no poles in C+ .
8. Define
U1 =

R1 R2 + S1 S2
V,
Q

U2 = QV.

9. Define U4 = [ 12 U3 ]+ .
10. Define

U1
U2
, T2 =
.
U4
U4
11. Let be a solution to Problem 15.3.
12. If kT1 T2 k 1 then Problem 15.2 has no solution.
13. The controller
1 + P2
RC =
,
2 P1
is a solution to Problem 15.2.
T1 =

E15.4 Mobius functions

+
E15.5 Show that RH
2 and RH2 are orthogonal with respect to the inner product on RL2
defined in equation (15.5).

Finish

15 An introduction to H control theory

22/10/2004

This version: 22/10/2004

Part IV
Background material

586

This version: 22/10/2004

A Linear algebra

22/10/2004

and scalar multiplication must satisfy the rules

5. a(b v) = (ab)v (associativity );

Appendix A
Linear algebra
Formulation of the time-domain setting for linear systems requires fluency with linear
algebra. Those of you taking this course are expected to be very familiar with essentials of
linear algebra. In this appendix we will review some of those essentials, mainly to introduce
the notation we use. The presentation is distinguished by a chain of sometimes not obvious
statements made in sequence. That is, nothing is proved in this appendix.

Contents
A.1 Vector spaces and subspaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 585
A.2 Linear independence and bases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 586
A.3 Matrices and linear maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 587
A.3.1 Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 587
A.3.2 Some useful matrix lemmas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 589
A.3.3 Linear maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 591
A.4 Change of basis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 592

6. 1 v = v;
7. a(v1 + v2 ) = a v1 + a v2 (distributivity );
8. (a1 + a2 )v = a1 v + a2 v (distributivity again).
One can also consider vector spaces over C, and on occasion we will do so, but not frequently.
Thus we shall simply call a vector space over R a vector space when no confusion will arise
from our doing so.
The vector space of primary importance to us is the collection Rn of n-tuples of real
numbers. Thus an element of Rn is written (x1 , . . . , xn ). We will write this also as a column
vector:

x1
..
. .
xn
Vector addition and scalar multiplication in Rn are done component-wise:
(x1 , . . . , xn ) + (y1 , . . . , yn ) = (x1 + y1 , . . . , xn + yn ),

a(x1 , . . . , xn ) = (ax1 , . . . , axn ).

We shall sometimes wish to work with Cn , the set of n-tuples of complex numbers. This can
be thought of as a vector space over either R or C with the component-wise operations. We
denote by x the vector with components (x1 , . . . , xn ). Thus vectors are generally denoted
with lowercase bold letters. We simply denote the zero vector by 0.
A subset U of a vector space V is a subspace if u1 + u2 U for all u1 , u2 U and if
a u U for all a R and all u U . Note that a subspace of a vector space is itself a vector
space.

A.5 Eigenvalues and eigenvectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 593

A.6 Inner products . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 594

A.1 Vector spaces and subspaces

We will work with the real numbers, denoted R, and with the complex numbers,
denoted C. We
follow the convention of mathematicians rather than of electrical engineers
and denote i = 1. If we are in a situation where we wish to refer to either R or C, we
will write F. R+ denotes the set of positive real numbers.
A vector space over R is a set V with two operations: (1) vector addition, denoted
v1 + v2 V for v1 , v2 V , and (2) scalar multiplication, denoted a v V for a R and
v V . Vector addition must satisfy the rules
1. v1 + v2 = v2 + v1 (commutativity );
2. v1 + (v2 + v3 ) = (v1 + v2 ) + v3 (associativity );
3. there exists a unique vector 0 V with the property that v + 0 = v for every v V
(zero vector ), and
4. for every v V there exists a unique vector v V such that v + (v) = 0 (negative
vector ),

A.2 Linear independence and bases

Let V be a vector space. A collection of vectors {v1 , . . . , vn } V is linearly independent if the equality
c1 v1 + + cn vn = 0
holds only for c1 = = cn = 0. A collection of vectors {v1 , . . . , vn } spans V if for every
vector v V there exists constants c1 , . . . , cn R so that
v = c1 v1 + + cn vn .
The subset {v1 , . . . , vn } is a basis for V if it is linearly independent and spans V . Note
that the number of basis vectors for a vector space V is a constant independent of the choice
of basis. This constant is the dimension of V , denoted dim(V ). It is possible that a vector
space will not have a finite basis. In such cases it is said to be infinite-dimensional . We
may talk in particular about bases for Rn . The standard basis for Rn is given by the n
vectors {e1 = (1, 0, . . . , 0), e2 = (0, 1, . . . , 0), . . . , en = (0, 0, . . . , 1)}.
Given a basis {v1 , . . . , vn } for V and a vector v V , there is a unique collection of
constants c1 , . . . , cn R so that
v = c1 v1 + + cn vn .

22/10/2004

A.3 Matrices and linear maps

587

These are the components of v in the given basis.

For vectors v1 , . . . , vk V we define the span of these vectors to be the subspace
span {v1 , . . . , vk } = { c1 v1 + + ck vk | c1 , . . . , ck R} .
Note that this does not require that the vectors be linearly independent. However, if the
vectors {v1 , . . . , vk } are linearly independent, then they form a basis for span {v1 , . . . , vk }.

A.3 Matrices and linear maps

There are close relationships between matrices and linear maps. However, let us break
up our discussion to cover some matrix-specific topics.
A.3.1 Matrices The set of n m matrices we
matrix with components

a11 a12 a13

a21 a22 a23

..
..
..
...
.
.
.
an1 an2 an3

denote by Rnm . We denote by A the

a1m
a2m

.. .
.
anm

Thus matrices are generally denoted with uppercase bold letters. At times we will think of
vectors as n 1 matrices. We denote the n m matrix of zeros by 0n,m . If A Rnm then
we define the transpose of A, denoted At , to be the matrix whose rows are the columns
of A. Thus

a11 a12 an1

a11 a12 a1m
a21 a22 a2m
a12 a22 an2

=
At = ..
A = ..
..
..
..
.. .
..
..
.

.
.
.
.
.
.
.
a1m a2m anm
an1 an2 anm
If A Rnp and B Rpm then we may multiply these to get AB Rnm . The (i, j)th
element of AB is
p
X
aik bkj .
k=1

Of special interest are the n n matrices, i.e., the square matrices. I n denotes the
n n identity matrix , i.e., the matrix whose entries are all zero, except for 1s on the
diagonal, and 0n denotes the n n matrix of zeros. The trace of A Rnn , denoted tr(A),
as the sum of the diagonal elements of A:
tr(A) =

n
X

aii .

i=1

Also useful is the determinant of A Rnn , denoted det A. Let us recall the inductive
definition. The determinant of a 1 1 matrix [a] is simply a. The determinant of a 2 2
matrix is defined by

a
a
det 11 12 = a11 a22 a12 a21 .
a21 a22
Now we indicate how to compute the determinant of an n n matrix provided one knows
how to compute the determinant of an (n 1) (n 1) matrix. One does this as follows.

588

A Linear algebra

22/10/2004

For a fixed i {1, . . . , n}, let a1 , . . . , an be the components of the ith row of A Rnn .
Also let A(, ) be the (n 1) (n 1) matrix obtained by deleting the ith row and the jth
column of A. With this notation we define
det A = (1)i+j aj det A(, ).
Thus, to compute the determinant of an n n matrix, one must compute the determinant
of n matrices of size (n 1) (n 1).
There is another definition of the determinant that we will use in Section 6.1. Let Sn
be the collection of permutations of (1, . . . , n). We denote an element Sn by indicating
what it does to each element in the sequence (1, . . . , n) like so:

1
2

n
=
.
(1) (2) (n)
A transposition is a permutation that consists of the swapping of two elements of (1, . . . , n).
A permutation is odd (resp. even) if it is the composition of an odd (resp. even) number
of transpositions. We define sgn : Sn {1, 1} by
(
1,
is even
sgn() =
1, is odd.
With this notation, it can be shown that
X
sgn()a(1)1 a(2)2 a(n)n .
det A =

(A.1)

If and only if det A 6= 0 there exists A1 Rnn so that AA1 = A1 A = I n . In this

case we say A is invertible of nonsingular . If A is not invertible, it is singular .
We shall require the cofactor matrix of an n n matrix A. If i, j {1, . . . , n}, define
the (i, j)th cofactor to be
Cij = (1)i+j det A(, ).
The matrix Cof(A) is then the n n matrix whose (i, j)th element is Cij . One then verifies
that
A(Cof(A))t = (det A)I n .
The matrix (Cof(A))t we denote adj(A), and we note that if A is invertible, then
A1 =

1
adj(A).
det A

(A.2)

We call adj(A) the adjugate of A.

Let us next consider the linear equation Ax = b which we wish to solve for a given
b Rn and A Rnn . We will only look at the easy case where A is invertible. If for
i {1, . . . , n}, A(b, i) denotes the matrix A but with the ith column replaced with b, then
Cramers Rule states that the ith component of the solution vector x is given by
det A(b, i)
.
det A
More general than a cofactor of A is the notion of a minor. A kth-order minor of A
is the determinant of a k k matrix obtained from A by removing any collection of n k
rows and n k columns. The principal minors of A are the n determinants of the upper
left k k blocks of A for k {1, . . . , n}.
xi =

A.3 Matrices and linear maps

22/10/2004

589

590

A Linear algebra

22/10/2004

A.3.2 Some useful matrix lemmas The following results will be useful to us, and
the proofs are simple enough to give here.

A.3 Lemma If A Rnm and B Rmn then I n AB is nonsingular provided that

I m BA is nonsingular, and furthermore the relation

A.1 Lemma If A is an invertible k k matrix, B is a k (n k) matrix, C is a (n k) k

matrix, and D is an (n k) (n k) matrix, then

A B
det
= det A det(D CA1 B).
C D

A(I m BA)1 = (I n AB)1 A,

Proof We observe that

Ik
0k,(nk) A B
A
B
=
.
1
1
C D
CA
I nk
0(nk),k D CA B
The determinant of the leftmost matrix is 1, and since the determinant of an upper block
diagonal matrix is the determinant of the diagonal blocks, the determinant of the rightmost
matrix is det A det(D CA1 B). From this the result follows.

A.2 Lemma If A is an invertible k k matrix, B is a k (n k) matrix, C is a (n k) k
matrix, and D is an (n k) (n k) matrix, then

1
X Y
A B
=
,
U V
C D
where
X = A1 + A1 B(D CA1 B)1 CA1
Y = A1 B((D CA1 B)1
1

U = (D CA B) CA

holds in this case.

Proof We first show that det(I n AB) = det(I m BA). First we note that the matrix

In A
B Im
is transformed to the matrix

B Im
In A

by k column switches, for the same k as was used to make the row switches. Note that
each row and column switch changes the determinant by a factor of 1. Thus we have, also
employing Lemma A.1,

I
A
I
B
det(I m BA) = det n
= det m
= det(I n AB),
B Im
A In
as desired. This shows that I n AB is nonsingular if and only if I m BA is nonsingular.
Now we make a simple computation,

=
=

provided that the inverse exists.

Multiply the first k rows by A1 to get

I k A1 B A1
0
.
C
D
0 I nk
Now subtract from the second n k rows the first k rows multiplied by C to get

Ik
A1 B
A1
0
.
1
1
0 D CA B CA
I nk
1

Multiply the second n k rows by (D CA B)1 times the first k rows to get

I k A1 B
A1
0
1
1
1
1
1 .
0 I nk (D CA B) CA
(D CA B)
Finally, subtract from the second n k rows A1 B times the first k rows to yield the result.

by k row switches, for some suitable k. Now this last matrix can be transformed to

Im B
A In

V = (D CA1 B)1 ,

Proof We perform row operations on

A B Ik
0
.
C D 0 I nk

A ABA = A ABA
A(I m BA) = (I n AB)A
(I n AB)1 A = A(I m BA)1 ,

as desired.

A.4 Lemma Let A Rnn , B Rnm , and C Rmn . If A and A + BC are invertible
then I m + CA1 B is invertible and
(A + BC)1 = A1 A1 B(I m + CA1 B)1 CA1 .
Proof We compute

A1 A1 B(I m + CA1 B)1 CA1 (A + BC)
= I n + A1 BC A1 B(I m + CA1 B)1 C A1 B(I m + CA1 B)1 CA1 BC

= I n + A1 B I m (I m + CA1 B)1 (I m + CA1 B)1 CA1 B C

= I n + A1 B I m (I m + CA1 B)1 (I m + CA1 B C

= I n + A1 B I m I m C
= I n.
This gives the lemma by uniqueness of the inverse.

A.3 Matrices and linear maps

22/10/2004

591

A.5 Lemma Let M Rnn and let u, v Rn . If M and M + uv t are invertible we have
adj(M + uv t )u = (adjM )u
v t adj(M + uv t ) = v t (adjM ).
Proof Let us show the first equality, and the second follows in much the same manner.
Using Lemma A.4 we compute

M 1 =

A Linear algebra

22/10/2004

for some uniquely defined a1i , . . . , ani R. The nm numbers a`i , ` = 1, . . . , n, i = 1, . . . , m

are called the components of L relative to the given bases.
If L : U V is a linear map, we define subspaces
ker(L) = { u U | L(u) = 0} U
image(L) = { L(u) | u U } V
which we call the kernel and image of L, respectively. You may know these as nullspace
and range. The Rank-Nullity Theorem states that dim(ker(L)) + dim(image(L)) =
dim(U ). A linear map L : U V is injective if the equality L(u1 ) = L(u2 ) implies that
u1 = u2 . L is surjective if for each v V there exists u U so that L(u) = v. If L is
both injective and surjective then it is bijective or invertible. If L is invertible then there
exists a linear map L1 : V U so that L L1 = idV and L1 L = idU , where id denotes
the identity map.
If A Rnm and we write A using its columns as

A = a1 am .

adj(M + uv t )
u
det(M + uv t )
M 1 uv t M 1 u
= M 1 u
1 + vtM u
M 1 u
=
.
1 + vtM u

(M + uv t )1 u =

We also have

592

adjM
det M

Thus each of the vectors a, . . . , am are in Rn . We define the columnspace of A to be the

subspace of Rn given by span {a, . . . , am }. We observe that the columnspace and the image
of A, thought of as a linear map, coincide.
If L : V V is a linear transformation, a subspace U V is L-invariant if L(u) U
for every u U . Suppose that V is finite-dimensional and that U is an invariant subspace
for L and let {v1 , . . . , vn } be a basis for Rn with the property that {v1 , . . . , vk } is a basis for
U . Since U is L-invariant we must have

and, by Lemma A.3.

1 + v t M 1 u = det(1 + v t M 1 u)
= det(I n + M 1 uv t )
det(M + uv t )
=
.
det M
This then gives
M 1 u
(adjM )u
det M
(adjM )u
=
=
.
1 + vtM u
det M det(M + uv t )
det(M + uv t )
Thus
(M + uv t )1 u =

(adjM )u
,
det(M + uv t )

from which we deduce that adj(M + uv t )u = (adjM )u, as desired.

A.3.3 Linear maps Let U and V be vector spaces. A linear map from U to V is a
map L : U V satisfying L(u1 + u2 ) = L(u1 ) + L(u2 ) for all u1 , u2 U and L(a u) = a L(u)
for all a R and u U . If U = V then L is a linear transformation. A special type of
linear map, one from the vector space Rm to the vector space Rn , is defined by an n m
matrix A. If x Rm then A(x) Rn is defined by its ith component being
n
X

Aij xj .

j=1

Typically we will write Ax for A(x).

Let L : U V be a linear map between finite-dimensional vector spaces, and let
{u1 , . . . , um } be a basis for U and {v1 , . . . , vn } be a basis for V . For i {1, . . . , m} we
may write
L(ui ) = a1i v1 + + ani vn

L(v1 ) = a11 v1 + + ak1 vk

..
.
L(vk ) = a1k v1 + + akk vk
L(vk+1 ) = a1,k+1 v1 + + ak,k+1 vk + ak+1,k+1 vk+1 + + an,k+1 vn
..
.
L(vn ) = a1n v1 + + akn vk + ak+1,n vk+1 + + ann vn .
Thus the matrix representation for L in this basis has the form

A11 A12
.
0nk,k A22

(A.3)

A.4 Change of basis

One on occasion wishes to ascertain how the components of vectors and linear maps
change when one changes basis. We restrict our consideration to linear transformations. So
let V be a vector space with {v1 , . . . , vn } and v1 , . . . , vn } two bases for V . Since these are
both bases, there exists an invertible n n matrix T so that
vi =

n
X
j=1

tji vj ,

i = 1, . . . , n.

A.5 Eigenvalues and eigenvectors

22/10/2004

593

T is called the change of basis matrix . For v V , let c1 , . . . , cn and c1 , . . . , cn be the

components of v in the respective bases {v1 , . . . , vn } and v1 , . . . , vn }. One readily determines
= T c. Also, let L : V V be a linear transformation and let aij , i, j = 1, . . . , n,
that c
and a
ij , i, j = 1, . . . , n, be the components of L in the respective bases {v1 , . . . , vn } and
=
v1 , . . . , vn }. A simple computation using the definition of components shows that A
T AT 1 . We shall most frequently encounter the change of basis in the scenario when
V = Rn , {v1 = e1 , . . . , vn = en }, and {
v1 = x1 , . . . , vn = xn }. Thus we are changing from
the standard basis for Rn to some nonstandard basis. In this case, the change of basis matrix
is simply defined by

T 1 = x1 xn .
The transformation that sends A Rnn to the matrix T AT 1 is called a similarity
transformation. One readily verifies both trace and determinant are unchanged under
similarity transformations. That is, for any invertible T Rnn we have
tr(T AT

) = tr(A),

det(T AT

) = det(A).

594

A Linear algebra

22/10/2004

then
An + pn1 An1 + + p1 A + p0 I n = 0n .
Youll also recall that spec(A) = spec(At ), but that the eigenvectors of A and At will
generally differ.
A linear transformation L : V V on a finite-dimensional vector space is diagonalisable if it possesses a basis of eigenvectors. One can easily check that if L is diagonalisable
then there exists a basis {v1 , . . . , vn } for V in which the matrix of components for L is diagonal. Note that a matrix with complex eigenvalues can never be diagonalised if V is a
R-vector space.

A.6 Inner products

An inner product on a R-vector space V assigns to a pair of vectors v1 , v2 V a
number hv1 , v2 i R, and the assignment satisfies
1. hv1 , v2 i = hv2 , v1 i (symmetry ),

For this reason, the notion of trace and determinant can be applied to any linear transformation by simply defining them in an arbitrary basis.

2. ha1 v1 + a2 v2 , vi = a1 hv1 , vi + a2 hv2 , vi (bilinearity ), and

3. hv, vi = 0 if and only if v = 0 (positive-definiteness).

A.5 Eigenvalues and eigenvectors

For a linear transformation L : V V , if we have a pair (, v) R (V \ {0}) which
satisfies L(v) = v then we say 0 is an eigenvalue for L with eigenvector v. The
collection of all eigenvectors for , along with the zero vector, forms the eigenspace for .
The eigenspace is easily seen to be a subspace. When V is finite-dimensional, the matter of
computing eigenvalues and eigenvectors is most easily done in a basis {v1 , . . . , vn } where L
is represented by its component matrix A Rnn . Thus the problem is reduced to finding
(, x) R (Rn \ {0}) satisfying Ax = x. This implies that A I n must be singular.
Thus eigenvalues must be roots of the characteristic polynomial

On a C-vector space, we do things a little differently. If V is a C-vector space, a Hermitian

inner product is an assignment (v1 , v2 ) 7 hv1 , v2 i C with the assignment now satisfying
1. hv1 , v2 i = hv2 , v1 i (symmetry ),
2. ha1 v1 + a2 v2 , vi = a
1 hv1 , vi + a
2 hv2 , vi (bilinearity ), and
3. hv, vi = 0 if and only if v = 0 (positive-definiteness).
Here z denotes the complex conjugate of z C. On Rn there is the standard inner
product defined by
n
X
hx, yi =
xi yi ,
i=1

PA () = det(I n A).
Such roots may well be complex, and recall that a complex root of a real polynomial always
occurs along with its complex conjugate. We denote by
spec(A) = { C | is an eigenvalue for A} ,
which we call the spectrum of A. If is an eigenvalue of A, the algebraic multiplicity
of 0 , denoted ma (0 ), is the largest integer k for which we can write PA () = ( 0 )k Q()
for some polynomial Q() satisfying Q(0 ) 6= 0. The geometric multiplicity of 0 is
the maximum number of linearly independent eigenvectors possessed by 0 . Recall that
mg (0 ) ma (0 ). If we have a real matrix A, we will often talk about multiplicities
thinking of A as a complex matrix, since A may very well have complex eigenvalues. Note
that the above discussion of finding eigenvalues for L using its matrix of components in a
basis is obviously a basis independent operation.
Also recall the Cayley-Hamilton Theorem which says that a matrix satisfies its own
characteristic polynomial. That is to say, if
PA () = n + pn1 n1 + + p1 + p0

i.e., the dot product. For Cn the standard Hermitian inner product is defined by
hx, yi =

n
X

xi yi .

i=1

Given R-inner product spaces U and V and a linear map L : U V , we define the
adjoint of L as the linear map L : V U satisfying
hL (v), uiU = hv, L(u)iV ,

u U, v V.

If A Rnm is thought of as a linear map between the R-vector spaces Rm and Rn with their
standard inner products, then the adjoint of A is simply the standard transpose: A = At .
If A Cnn is thought of as a linear map between the R-vector spaces Cm and Cn with their
standard Hermitian inner products, then the adjoint of A is the standard transpose with
t . Given a linear map L : U V
each element of the matrix additional conjugated: A = A
between inner product spaces (or Hermitian inner product spaces), a singular value for L
is an eigenvalue of L L : V V .
A linear transformation L : V V of a R-inner product space (resp. a C-Hermitian
inner product space) is symmetric (resp. Hermitian) if L = L . One easily deduces

22/10/2004

A.6 Inner products

595

that the eigenvalues of a symmetric or Hermitian linear transformation are always real. If
V = Rn so that L is an n n matrix, then symmetry of L is exactly the condition that L
be a symmetric matrix. In this case, there exists an orthogonal matrix R so that the matrix
RLRt is diagonal.
A linear transformation L : V V of a R-inner product space (resp. a C-Hermitian inner
product space) is skew-symmetric (resp. skew-Hermitian) if L = L . One readily
checks that any transformation L on a R-inner product space is a sum of its symmetric
part 21 (L + L ) and its skew-symmetric part 12 (L L ).
On an inner product space or a Hermitian inner product space V we also
p have defined
a norm which assigns to a vector v V a real number defined by kvk = hv, vi. A norm
may be verified to satisfy the triangle inequality :
kv1 + v2 k kv1 k + kv2 k .

596

A Linear algebra

22/10/2004

Exercises
EA.1 Let V = span {(1, 2, 0, 0), (0, 1, 0, 1)} R4 . Construct a 4 4 matrix that leaves the
subspace V invariant.

Exercises for Chapter A

597

598

This version: 22/10/2004

B Ordinary differential equations

22/10/2004

B.1 Construction of solutions corresponding to a root Let C be a root of the characteristic

equation of multiplicity k.
1. If R then k linearly independent solutions are

Appendix B

y1 (t) = et , y2 (t) = tet , . . . , yk (t) = tk1 et .

2. If 6 R write = + i. Then 2k linearly independent solutions are

Ordinary differential equations

y1 (t) = et sin t, y2 (t) = et cos t, y3 (t) = tet sin t, y4 (t) = tet cos t, . . . ,
y2k1 (t) = tk1 et sin t, y2k (t) = tk1 et cos t.
In this appendix we provide a very quick review of differential equations at the level
prerequisite to use this book. The objective, as in the other appendices, is not to provide a
complete overview, but to summarise the main points. The text itself covers various aspects
of differential equations theory beyond what we say here; indeed, it is possible to think
of control theory as a sub-discipline of differential equations, although I do not like to do
so. We deal with scalar equations in Section B.1, and with systems of equations, using the
matrix exponential, in Section B.2.

Applying the construction to all roots of the characteristic equation yields n linearly independent solutions y1 (t), . . . , yn (t) to the homogeneous equation. Furthermore, every solution of
the homogeneous equation is a linear combination of these n linearly independent solutions.
Thus we take yh to be a general linear combination of the n linearly independent solutions,
with at the moment unspecified coefficients:
yh (t) = c1 y1 (t) + + cn yn (t).

Contents
B.1 Scalar ordinary differential equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 597
B.2 Systems of ordinary differential equations . . . . . . . . . . . . . . . . . . . . . . . . . . . 600

B.1 Scalar ordinary differential equations

Let us consider a differential equation of the form
y (n) (t) + pn1 y (n1) (t) + + p1 y (1) (t) + p0 y(t) = u(t).
(i)

These are the solutions corresponding to the root .

(B.1)

Next one determines what is called a particular solution, which we shall denote yp .
A particular solution is, by definition, any solution to the differential equation (B.1). Of
course, one cannot expect to arrive at such a solution in a useful way for arbitrary righthand side functions u. What we shall describe here is an essentially ad hoc procedure,
useful for certain types of functions u. This procedure typically goes under the name of the
method of undetermined coefficients. The idea is that one uses the fact that certain
types of functions of timepolynomial functions, trigonometric functions, and exponential
functionshave their form unchanged by the act of differentiation. Thus if one takes yp to
be of this form, then
yp(n) (t) + pn1 yp(n1) (t) + + p1 yp(1) (t) + p0 yp (t)

(B.2)

d y(t)
.
dti

and u(t) is a known function of t. In the text, we will establish various

Here y (t) =
methods for solving such equations for general functions u. Students are expected to be able
to solve simple examples of such equations with alacrity. Let us recall how this is typically
done.
First one obtains a solution, called a homogeneous solution and denoted here by yh ,
to the equation
(n)

(n1)

yh (t) + pn1 yh

(1)

(t) + + p1 yh (t) + p0 yh (t) = 0.

To do this, one seeks solutions of the form et for C. Substitution into the homogeneous
equations then gives the polynomial
n + pn1 n1 + + p1 + p0 = 0.
This is the characteristic equation and it will have n roots, counting multiplicities. Thus
it will have roots 1 , . . . , ` , with respective multiplicities k1 , . . . , k` , with k1 + + k` = n.
Corresponding to a root a , one constructs ka linearly independent solutions if the root is
real, and 2ka linearly independent solutions if the root is complex (a complex root and its
conjugate yield the same 2ka solutions). These solutions are constructed as follows.

will also have the same form. Therefore, one hopes to be able to determine yp by comparing
u to the expression (B.2). This hope is founded, and leads to the following procedure.
B.2 Method of undetermined coefficients Suppose that u(t) = t` et (a1 sin t + a2 cos t). Let
k be the smallest integer with the property that tk ytest (t) is a not solution to the homogeneous
equation, where ytest is any function of t of the form
ytest (t) = P (t)et (1 sin t + 2 cos t),
where 1 and 2 are arbitrary constants, and P is an arbitrary nonzero polynomial of degree
at most `. Often we will have k = 0. Then seek a particular solution of the form
yp (t) = tk Q(t)et (b1 sin t + b2 cos t),
where Q is a polynomial of degree `. Substitute into (B.1) to determine b1 , b2 , and the
coefficients of Q. If u is a linear combination,
u(t) = 1 u1 (t) + + m um (t),

B.1 Scalar ordinary differential equations

22/10/2004

599

of terms of the form t` et (a1 sin t + a2 cos t), then the above procedure may be applied
separately to each term in the linear combination separately, yielding separate particular
solutions yp,1 , . . . , yp,m . Then the particular solution is
yp (t) = 1 yp,1 (t) + + m yp,m (t).

y(t) = yh (t) + yp (t) = c1 y1 (t) + + cn yn (t) + yp (t).

The constants c1 , . . . , cn can be determined if we have an initial value problem where the
initial values y(0), y (1) (0), . . . , y (n1) (0) are specified.
It is worth demonstrating the procedure to this point on a simple example.
B.3 Example We consider the differential equation
...
y(t) + 3
y (t) + 3y + y(t) = tet .
First let us find a solution to the homogeneous equation
...
yh (t) + 3
yh (t) + 3y h + yh (t) = 0.
The characteristic equation is 3 + 32 + 3 + 1 = 0, which has the single root = 1 of
multiplicity 3. Our recipe for the homogeneous solution then gives
yh (t) = c1 et + c2 tet + c3 t2 et .
Now let us obtain a particular solution. By the above procedure, we take
ytest = (1 t + 0 )et .
We calculate that ytest (t) and tytest identically solve the differential equation for any 1
and 0 . Also, we compute that t2 ytest (t) satisfies the homogeneous provided that 1 = 0.
In exasperation, we compute that the only way that t3 ytest (t) will satisfy the homogeneous
equation if and only if 1 = 0 = 0. Thus we have k = 3, using the notation of our procedure
for obtaining particular solutions. Therefore, we seek a particular solution of the form
yp (t) = t3 (q1 t + q0 )et .
Substitution into the ode gives

Thus the particular solution is

4 t

yp (t) =

te
.
24

The general solution is then

t4 et
.
24
To determine the unknown constants c1 , c2 , and c3 , we use the initial conditions. One may
determine that
y(t) = yh (t) + yp (t) = c1 et + c2 tet + c3 t2 et +

c1 = y(0),

c2 = y(0) + y(0),

+ y(0)).
c3 = 21 (y(0) + 2y(0)

22/10/2004

B.2 Systems of ordinary differential equations

Next we look at first-order systems of ordinary differential equations:

x(t)
= Ax(t),
n

(B.3)

where x R and A R . The possibility of an additional term on the right-hand side

of the form f (t) is interesting, but this is handled in the text. The solution to (B.3) is given
to us by the matrix exponential . For any n n matrix A define
eA = I n + A +

Ak
A2
+ +
+
2!
k!

One may verify that this series converges for every A, and it is a straightforward calculation to verify that x(t) = eAt x0 is the solution to the differential equation (B.3) with
the initial condition x(0) = x0 . Let us present a way to compute the matrix exponential.
Corresponding to the linear system (B.3) will be n linearly independent vector solutions
x1 (t), . . . , xn (t). If A is diagonalisable, it is relatively easy to compute these solutions, but
if A is not diagonalisable, there is a bit more work involved. Nonetheless, there is a recipe
for doing this, and we present it here. Note that this recipe will always work, and it can be
applied, even when A is diagonalisablein this case it just simplifies.
First one computes the eigenvalues for A. There will be n of these in total, counting
algebraic multiplicities and complex conjugate pairs. One treats each eigenvalue separately.
For a real eigenvalue 0 with algebraic multiplicity k = ma (0 ), one must compute k linearly
independent solutions. For a complex eigenvalue 0 with algebraic multiplicity ` = ma (0 ),
0 is also necessarily an eigenvalue
one must compute 2` linearly independent solutions, since
with algebraic multiplicity `.
We first look at how to deal with real eigenvalues. Let 0 be one such object with
algebraic multiplicity k. It is a fact that the matrix (A 0 I n )k will have rank n k, and
so will have a kernel of dimension k by the Rank-Nullity Theorem. Let u1 , . . . , uk be a basis
for ker((A 0 I n )k ). We call each of these vectors a generalised eigenvector . If the
geometric multiplicity of 0 is also k, then the generalised eigenvectors will simply be the
usual eigenvectors. If mg (0 ) < ma (0 ) then a generalised eigenvector may or may not be an
eigenvector. Corresponding to each generalised eigenvector ui , i = 1, . . . , k, we will define a
solution to (B.3) by
xi (t) = e0 t exp((A 0 I n )t)ui .
(B.4)
Note that because ui is a generalised eigenvector, the infinite series exp((A 0 I n )t)ui will
have only a finite number of termsat most k in fact. Indeed we have

tk1
t2
(A 0 I n )k1 ui ,
exp((A 0 I n )t)ui = I n + t(A 0 I n ) + (A 0 I n )2 + +
2!
(k 1)!

(6q0 + 24q1 t)et = tet ,

1
.
24

B Ordinary differential equations

Once one has a particular solution, the general solution is then simply the sum of the
homogeneous and particular solution:

from which we ascertain that q0 = 0 and q1 =

600

since the remaining terms in the series will be zero. In any case, it turns out that the k vector
functions x1 (t), . . . , xk (t) so constructed will be linearly independent solutions of (B.3). This
tells us how to manage the real case.
Lets see how this goes in a few examples.
B.4 Examples 1. We take a simple case where

7 4
A=
.
6 3

22/10/2004

B.2 Systems of ordinary differential equations

601

The characteristic polynomial for A is PA () = 2 + 4 + 3 = ( + 1)( + 3). Thus

the eigenvalues for A are 1 = 1 and 2 = 3. Since the eigenvalues are distinct, the
algebraic and geometric multiplicities will be equal, and the generalised eigenvectors will
simply be eigenvectors. An eigenvector for 1 = 1 is u1 = (2, 3) and an eigenvector for
2 = 3 is u2 = (1, 1). Our recipe then gives two linearly independent solutions as

1
2
.
, x2 (t) = e3t
x1 (t) = et
1
3
2. A more interesting case is the following:

2 1
0
A = 0 2 0 .
0
0 1

B Ordinary differential equations

and the generalised eigenvectors span the kernel of this matrix, and so we may take
u1 = (1, 0, 0) and u2 = (0, 1, 0) as generalised eigenvectors. Applying the formula (B.4)
gives

1
0 1 0 1
x1 (t) = e2t 0 + te2t 0 0 0 0
0
0 0 1 0
2t
e
= 0
0

22/10/2004

Now let us look at the complex case. Thus let 0 be a complex eigenvalue (with nonzero
0 will also be an eigenvalue of
imaginary part) of algebraic multiplicity `. This means that
algebraic multiplicity ` since A, and hence PA (), is real. Thus we need to find 2` linearly
independent solutions. We do this by following the exact same idea as in the real case,
except that we think of A as being a complex matrix for the moment. In this case it is still
true that the matrix (A 0 I n )` will have an `-dimensional kernel, and we can take vectors
u1 , . . . , u` as a basis for this kernel. Note, however, that since (A 0 I n )` is complex, these
vectors will also be complex. But the procedure is otherwise identical to the real case. One
then constructs ` complex vector functions
z j (t) = e0 t exp((A 0 I n )t)uj .

Since the matrix is upper triangular, the eigenvalues are the diagonal elements: 1 = 2
and 2 = 1. The algebraic multiplicity of 1 is 2. However, we readily see that
dim(ker(A 1 I 3 )) = 1 and so the geometric multiplicity is 1. So we need to compute
generalised eigenvectors in this case. We have

0 0 0
2

(A 1 I 3 ) = 0 0 0
0 0 1

(B.5)

Each such complex vector function will be a sum of its real and imaginary parts: z j (t) =
xj (t) + iy j (t). It turns out that the 2` real vector functions x1 (t), . . . , x` (t), y 1 (t), . . . , y ` (t)
are linearly independent solutions to (B.3).
Lets see how this works in some examples.
B.5 Examples 1. An example with complex roots is

1 1
0
A = 1 1 0 .
0
0 2
The characteristic polynomial is PA () = r3 + 4r2 + 6r + 4. One ascertains that the
1 = 1 i, 3 = 2. Lets deal with the
eigenvalues are then 1 = 1 + i, 2 =
complex root first. We have

i 1
0
0
A 1 I 3 = 1 i
0
0 1 i
from which we glean that an eigenvector is u1 = (i, 1, 0). Using (B.5) the complex
solution is then

i
(1+i)t
1 .
z 1 (t) = e
0
Using Eulers formula, ei = cos + i sin , we have

i cos t + sin t
sin t
cos t
z 1 (t) = et cos t + i sin t = et cos t + iet sin t
0
0
0

and

0
0 1 0 0
2t
x2 (t) = e 1 + te 0 0 0 1
0
0 0 1 0
2t
te
= e2t .
0
2t

thus giving

Finally we determine that u3 = (0, 0, 1) is an eigenvector corresponding to 2 = 1, and

so this gives the solution

0
x3 (t) = 0 .
et
Thus we arrive at our three linearly independent solutions.

602

sin t
cos t
x1 (t) = et cos t , y 1 (t) = et sin t .
0
0
Corresponding to the real eigenvalue 3 we readily determine that

0
2t
0
x2 = e
1
is a corresponding solution. This gives three linearly independent real solutions x1 (t),
y 1 (t), and x2 (t).

22/10/2004

B.2 Systems of ordinary differential equations

603

2. Let us look at a complex root with nontrivial multiplicity. The smallest matrix to possess
such a feature will be one which is 4 4, and about the simplest example for which the
geometric multiplicity is less than the algebraic multiplicity is

0 1 1 0
1 0 0 1

A=
0 0 0 1 .
0 0 1 0

604

We now determine the two linearly independent real solutions corresponding to u1 . We

have
z 1 (t) = eit (u1 + t(A iI 4 )u1 )

0
i

it 0
it 1
= e + te
i
0
1
0

0
0
0
1 !
0
0
1
0

= (cos t + i sin t) + i + t + it
0
1
0
0
1
0
0
0

t sin t
t cos t
t cos t
t sin t

=
sin t + i cos t
cos t
sin t

t sin t
t cos t
t cos t
t sin t

= x1 (t) =
sin t , y 1 (t) = cos t .
sin t
cos t

22/10/2004

For u2 we have
z 2 (t) = eit (u2 + t(A iI 4 )u2 )

i
1

= eit
0
0

0
1 !
1
0

= (cos t + i sin t)
0 + i 0
0
0

sin t
cos t
cos t
sin t

=
0 + i 0
0
0

sin t
cos t
cos t
sin t

= x2 (t) =
0 , y2 = 0 .
0
0

The eigenvalues are 1 = i and 2 = i, both with algebraic multiplicity 2. One readily
determines that the kernel of AiI 4 is one-dimensional, and so the geometric multiplicity
of these eigenvalues is just 1. Thus we need to compute complex generalised eigenvectors.
I have used Mathematica for the computations below. We compute

1 i i 1

i 1 1 i

(A iI 4 )2 = 2
0
0 1 i
0
0
i 1
and one checks that u1 = (0, 0, i, 1) and u2 = (i, 1, 0, 0) are two linearly independent
generalised eigenvectors. We compute

i
0
1
0

(A iI 4 )u1 = , (A iI 4 )u2 =
.
0
0
0
0

B Ordinary differential equations

Thus we have the four real linearly independent solutions x1 (t), x2 (t), y 1 (t), and y 2 (t).
We still havent gotten to the matrix exponential yet, but all the hard work is done.
Using the above methodology we may in principle compute for any n n matrix A, n
linearly independent solutions x1 , . . . , xn (t).1 If we assemble the resulting solutions into the
columns of a matrix X(t):

X(t) = x1 (t) xn (t) ,
the resulting matrix is an example of a fundamental matrix. Generally, a fundamental
matrix is any nn matrix function of t whose columns form n linearly independent solutions
to (B.3). What we have done above is give a recipe for computing a fundamental matrix
(there are an infinite number of these). The following result connects the construction of a
fundamental matrix with the matrix exponential.
B.6 Theorem Given any fundamental matrix X(t) we have eAt = X(t)X 1 (0).
Thus, once we have a fundamental matrix, the computation of the matrix exponential is
just algebra, although computing inverses of matrices of any size is a task best left to the
computer.
Lets work this out for our four examples.
B.7 Examples 1. If
A=

7 4
.
6 3

1
Note that the solutions x1 , . . . , xn are those obtained from both real and complex eigenvalues. Therefore,
the solutions denoted above as y i (t) for complex eigenvalues will be included in the n linearly independent
solutions, except now I am calling everything xj (t).

B.2 Systems of ordinary differential equations

22/10/2004

605

It is then an easy calculation to arrive at

3t

3e 2et 2e3t + 2et
eAt = X(t)X 1 (0) =
3t
t
3t
t
3e 3e
2e + 3e

2 1
0
A = 0 2 0 .
0
0 1

we had determined a fundamental matrix to be

e
te2t 0
e2t 0 .
X(t) = 0
0
0
et
It so happens that in this example we lucked out and eAt = X(t) since X(0) = I 3 .
3. For

1 1
0
A = 1 1 0 .
0
0 2
we had determined the fundamental matrix

t
e sin t et cos t 0
0 .
X(t) = et cos t et sin t
0
0
e2t
A straightforward computation yields

et cos t et sin t 0
= X(t)X 1 (0) = et sin t et cos t 0 .
0
0
e2t

eAt

4. Finally, we look at the 4 4 example we worked out:

0 1 1 0
1 0 0 1

A=
0 0 0 1 .
0 0 1 0
The fundamental matrix we obtained was

t sin t t cos t sin t cos t

t cos t t sin t cos t sin t
.
X(t) =
sin t cos t
0
0
cos t
sin t
0
0
A little manipulation gives

eAt

cos t sin t t cos t t sin t

sin t cos t t sin t t cos t
.
=
0
0
cos t
sin t
0
0
sin t cos t

B Ordinary differential equations

22/10/2004

Now you know how to compute the matrix exponential. However, you can get Maple to
do this. I use Mathematica for such chores. But you are expected to know how in principle
to determine the matrix exponential. Most importantly, you should know precisely what it
is.

then we have determined a fundamental matrix to be

t

2e
e3t
X(t) =
.
3e2t e3t

2. When

606

Exercises for Chapter B

607

Exercises
EB.1 Solve the initial value problem
x(t)
+ x(t) = A1(t),

x(0) = 0,

where > 0, and 1(t) is the unit step function:

(
1, t 0
1(t) =
0, otherwise.
Draw a graph of the solution.
EB.2 A mass moving in a gravitational field is governed by the differential equation m
y (t) =
mg. Solve this differential equation with initial conditions y(0) = y0 and y(0)

= v0 .
EB.3 Obtain the general solution to the differential equation
x(t) + 20 x(t)
+ 02 x(t) = A cos t
for , 0 , and A positive constants. You will have to deal with three cases depending
on the value of .
EB.4 Compute by hand eAt for the following matrices:

2 0
1

(a) A = 1 2 0 ;
0
0 3

(b) A =
for R and > 0.

EB.5 Use a computer package

2 3
1
3 2 0

(a) A =
0
0 2
0
0 3

0 1 1 0
0 0 0 2

(b) A =
1 0 0 0
0 1 0 0
0 0 0 0

to determine eAt for the following matrices:

0
1
;
3
2

0
0

0
.
1
0

608

B Ordinary differential equations

22/10/2004

610

This version: 22/10/2004

C Polynomials and rational functions

22/10/2004

2. If
P1 () = 2 + 2 + 1,

P2 () = + 1,

then the LCM of P1 and P2 is Q() = + 2 + 1.

Appendix C

If P F[] is given by
P () = ak k + ak1 k1 + + a1 + a0 ,

Polynomials and rational functions

a root for P is a number F with the property that

ak k + ak1 k1 + + a1 + a0 = 0.
We shall be frequently be encountering and manipulating polynomials, so it will be helpful
to have on hand some basic facts concerning such objects. That polynomials might be useful
to us can be seen by noting that we have already defined the characteristic polynomial which
will be very important to us in these notes.

Contents
C.1 Polynomials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 609
C.2 Rational functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 611

C.1 Polynomials
We denote by F[] the set of polynomials in indeterminant and with coefficients in F,
where F is either R or C. Thus a typical element of F[] looks like
P () = ak k + ak1 k1 + + a1 + a0

(C.1)

where a0 , . . . , ak F with ak 6= 0. We call k the degree of P which we denote by deg(P ).

You should not think of as being an element of F, but rather as just being a placeholder.
If we wish to plug in values from F into P we shall generally say when we do this. Of course,
you can add and multiply polynomials in just the ways with which you are familiar.
A polynomial P of the form (C.1) is monic if ak = 1; that is, a monic polynomial is
one where the coefficient of the highest power of the indeterminant is +1. For example, the
characteristic polynomial is always a monic polynomial.1 Given two polynomials P1 , P2
F[] their least common multiple (LCM ) is the unique monic polynomial Q F[] of
least degree with the property that Q = P1 R1 and Q = P2 R2 for some R1 , R2 F[].
C.1 Examples In each of these examples, the polynomials may be thought of as in either R[]
or C[].
1. If
P1 () = + 2, P2 () = 2 3,
then the LCM of P1 and P2 is Q() = 2 12 23 .
1

This is why we defined PA () by det(I n A) instead of by det(A I n ).

We denote the set of roots of P by spec(P ), mirroring our notation for eigenvalues for
matrices. If is a root of P then there exists Q F[] so that P () = ( )Q(). If is a
root of P and if m is the unique integer with the property that P () = ( )m Q() where
is not a root of Q, then m is the multiplicty of the root. If F = C, a root C is (1) in
the positive half-plane if Re() > 0, (2) in the negative half-plane if Re() < 0, and
(3) on the imaginary axis if Re() = 0. We denote the positive half-plane by C+ and the
negative half-plane by C . We shall often also denote the imaginary axis by iR (meaning
{i | R}). By C+ we mean the positive half-plane along with the imaginary axis, and
similarly be C we mean the negative half-plane along with the imaginary axis.
The greatest common divisor (GCD) of two polynomials P1 , P2 F[] is the unique
monic polynomial T of greatest degree so that P1 = T Q1 and P2 = T Q2 for some Q1 , Q2
F[].
C.2 Examples In each of these examples, the polynomials may be thought of as in either R[]
or C[].
1. If
P1 () = 2 + 1, P2 () = 3
then the GCD of P1 and P2 is T () = 1.
2. If
P1 () = 3 2 + 6 + 3,

P2 () = 2 2 6 8,

then the GCD of P1 and P2 is T () = + 1.

If the GCD of polynomials P1 and P2 is T () = 1, then P1 and P2 are said to be coprime.

A polynomial P F[] with deg(P ) > 0 is irreducible if there are no polynomials P1 , P2
F[], both with degree less than P , with the property that P = P1 P2 . Thus an irreducible
polynomial cannot be factored.
C.3 Examples 1. By the Fundamental Theorem of Algebra, the only irreducible monic polynomials in C[] are of the form P () = + a for some a C.
2. As we know, there are polynomials of degree two in R[] which are irreducible. For
example P () = 2 + 1 is irreducible in R[], but not in C[]. The irreducible monic
polynomials in R[] are of the form
(a) P () = + a for some a R or
(b) P () = 2 + b + c where b, c R satisfy b2 4c < 0.
One can easily show that an irreducible monic polynomial of degree 2 in R[] must have
the form ( )2 + 2 for R and > 0 (see Exercise EC.1).

C.2 Rational functions

22/10/2004

611

The following result is related to the Euclidean algorithm for F[] which, you will
recall, states that for any polynomials P, Q F[] there exists polynomials F, R F[] with
the properties

612

C Polynomials and rational functions

D() = 0. As a polynomial, the only polynomial which is zero is the zero polynomial. If
P F[] is a non-zero polynomial, then the two rational functions
R1 =

1. P = F Q + R and
2. deg(D) < deg(Q).
That is to say, a polynomial can be written as a product of a given polynomial and a
remainder. With this in mind, we can prove the following result.
C.4 Lemma Let P1 , P2 , F F[] be polynomials with P1 and P2 coprime. Then there exists
Q1 , Q2 F[s] so that
(i) deg(Q1 ) < deg(P2 ) and
(ii) Q1 P1 + Q2 P2 = F .
1, Q
2 R[s] so that
Proof Since P1 and P2 are coprime it follows that there exists Q

1 F )P1 + (Q
2 F )P2 = F,
(Q
from which it follows that for any G R[s] we have
1 F P2 G)P1 + (Q
2 F + P1 G)P2 = F.
(Q
The Euclidean algorithm asserts that there exists a G, R F[] so that
1 F = GP2 + R
Q
and deg(R) < deg(P2 ). That is to say, there exists G F[] so that
1 F P2 G) < deg(P2 ).
deg(Q
1 F P2 G and Q2 = Q
2 F +P1 G gives the result.
Choosing this G and then defining Q1 = Q

C.2 Rational functions

We also wish to talk about objects which are quotients of polynomials. A rational
function over F with indeterminate is a quotient of two elements of F[]. Thus we
write a rational function over F as
R() =
Thus
R() =

N ()
,
D()

N, D F[].

ak k + ak1 k1 + + a1 + a0
b` ` + b`1 `1 + + b1 + b0

where a0 , . . . , ak , b0 , . . . , b` F, and not all the bi s are zero. We denote the set of rational
functions over F with indeterminate by F(). One should take care not to unduly concern
oneself about things like whether the rational function blows up for certain values of where

PN
,
PD

R2 =

N
D

will in fact represent the same rational function. This is exactly the same thing we do when
we say that 13 and 26 are the same rational number. Note that for R F() there are unique
coprime monic polynomials N, D F[] so that
R=a

N
D

(C.2)

for some a F.
C.5 Example Consider the rational function

1 P1 + Q
2 P2 = 1.
Q
Therefore we have

22/10/2004

R() =

2 2 2 12
.
3 2 15 + 18

We can write this as

R() =

2
3

+2
2

which is the unique representation of the form (C.2).

Given a rational function R F() with a F and N, D F[] defined by (C.2), we call
(aN, D) the canonical fractional representative of R. We will be frequently in need of
this simple concept, so shall abbreviate if c.f.r.. A rational function R with c.f.r. (N, D) is
proper if deg(N ) deg(D), and strictly proper if deg(N ) < deg(D). A rational function
which is not proper is improper .
Let R F() which we write in its unique representation (C.2) for some coprime monic
polynomials N, D F[]. A zero of R is defined to be a root of N and a pole of R is defined
to be a root of D. Note that in this way we get around the problem of the function R not
being defined at poles. Two rational functions R1 , R2 F() are coprime if they have no
common zeros.
The final thing we do is provide a discussion of the so-called partial fraction expansion.
Recall that the idea here is to take a rational function and expand it as a sum of rational
functions whose denominators are powers of an irreducible polynomial. Thus, for example
3

1
1
19 1
1 1
3 3 + 2
= +5
+
+
.
5 2 + 3 + 9
1
( 3)2
4 3 4 +1

It is hard to come across an accurate description of how this is done, so let us provide one
here.
C.6 Theorem Let R F() and suppose that
R=

N
D

where N, D F[] are coprime, and take D to be monic.

There exists

C.2 Rational functions

22/10/2004

613

(i) m irreducible monic polynomials D1 , . . . , Dm F[],

(ii) positive integers j1 , . . . , jm ,
(iii) j1 + + jm polynomials N1,1 (x), . . . , N1,j1 , . . . , Nm,1 , . . . , Nm,jm F[], and
(iv) a polynomial Q F[] of degree deg(N )deg(D) (take Q = 0 if deg(N )deg(D) < 0),
with the properties
(v) D1 , . . . , Dm are coprime (i.e., distinct),
(vi) deg(Ni,k ) < deg(Di ) for i = 1, . . . , m and k = 1, . . . , ji ,
(vii) Ni,k and Di are coprime for i = 1, . . . , m and k = 1, . . . , ji , and
ji
m X
X
Ni,k
(viii) R =
+ Q.
(Di )k
i=1 k=1
Furthermore the objects described in (i)(iv) are the unique such objects with the properties (v)(viii).
The expression in part (viii) is called the partial fraction expansion of R. The proof of
this is straightforward, but requires some buildup, and we refer to [Lang 1984, Theorems 5.2
and 5.3].
It will turn out that we are primarily interested in the case when deg(N ) deg(D), and
in this case Q will be a constant, possibly zero, when deg(N ) = deg(D), and zero when
deg(N ) < deg(D). For a rational function R C() which satisfies our degree condition,
Theorem C.6 tells us that we may write
R() =

jm
m X
X
i=1 k=1

i,k
+
( i )k

ji
r X
X
i=1 k=1

(C.4)

for real numbers i , i = 1, . . . , r, i,k , i = 1, . . . , r, k = 1, . . . , ji , ai,k , bi,k , i = 1, . . . , m,

k = 1, . . . , `i , i , i , i = 1, . . . , `, and b.
This sort of leaves open how we compute the constants in the partial fraction expansion
for R = N/D. We shall say here how to do it when R C(). In this case 1 , . . . , m are
the poles of R, and ji is the multiplicity of the pole i . That is, D() = ( i )ji Q() where
Q and ( i ) are coprime. It turns out that
i,ji

1
dkji
=
( )k R()
kj
i
(k ji )! d
=i

R() = 1
2. We take
R() =

1,1 = ( 1)

+1
R() =
+2

5 + 4
.
2 + 2

and
2,1 = ( + 2)

5 + 4
9

= =3

( 1)( + 2) =1 3

5 + 4
6

= 2.
=

( 1)( + 2) =2 3

Thus the partial fraction expansion is

R() =
3. We take

2
3
+
.
1 +2

3 2 + 5 + 2
.
3 3 2 + 3

The roots of the denominator polynomial are 1 = 3, 2 = i, and 3 = i. Since we have

complex roots, there will be different partial fraction expansions, depending on whether
we are thinking of R C() or R R(). Let us take the complex case first. Using (C.5)
we have
3 2 + 5 + 2
1,1 = ( 3)
= 1

( + 3)( i)( + i) =3

3 2 + 5 + 2
2,1 = ( i)
= 1 + 2i
( + 3)( i)( + i) =i
3 2 + 5 + 2
= 1 2i .
3,1 = ( + i)

( + 3)( i)( + i) =i
Thus the partial fraction expansion over C is
1 2i
1 + 2i
1

.
3
i
+i

The partial fraction expansion over R turns out to be

As usual, this is self-explanatory in examples.

C.7 Examples 1. First let us show that one cannot dispense with the constant term if the
numerator and denominator polynomials have the same degree. If we take

1
.
+2

The first thing to do is factor the denominator: 2 + 2 = ( 1)( + 2). Thus in the
parlance of (C.3) we have 1 = 1 and 2 = 2. These roots have multiplicity 1 and so
j1 = j2 = 1. By (C.5) we then have

R() =
(C.5)

22/10/2004

then its partial fraction expansion is

(C.3)

i
XX
i,k
ai,k + bi,k
+
k + b
k
( i )
(
i )2 + i2
i=1 k=1

C Polynomials and rational functions

R() =

for some uniquely defined complex numbers , i , i = 1, . . . , m, and i,k , i = 1, . . . , m,

k = 1, . . . , ji . For R R(), things are a bit more complicated. We may write
R() =

614

R() =

1
2 + 1

.
3 2 + 1

Thus, employing the symbols in (C.4) we have 1 = 3, 1 = 0, and 1 = 1. The easiest

way to determine this is to compute the complex partial fraction expansion, and then
recombine the complex conjugate pairs over a common denominator.

C.2 Rational functions

22/10/2004

4. We take
R() =

615

616

EC.1 Let P () R[] be monic, irreducible, and degree two. Show that there exists R
and > 0 so that P () = ( )2 + 2 .
EC.2 Note that R[] is naturally a subset of C[]. If P () R[], denote by P () the same
polynomial, except thought of as being in C[]. Show that polynomials P1 (), P2 ()
R[] are coprime if and only if P1 (), P2 () are coprime.
EC.3 For
P () = n + pn1 n1 + + p1 + p0 F[], F {R, C},

2 2 + 1
1 d2

( + 1)3 3
=2

2
2
2 d
+ 3 + 3 + 1 =1

2 2 + 1
d

( + 1)3 3
= 4
1,2 =

d
+ 3 2 + 3 + 1 =1

2
2 + 1

1,3 = ( + 1)3 3
= 3.

+ 3 2 + 3 + 1 =1

1,1 =

Thus the partial fraction expansion is

2
4
3

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

22/10/2004

Exercises

2 2 + 1
.
3
+ 3 2 + 3 + 1

The root of the denominator polynomial is 1 which has multiplicity 3. We use (C.5) to
get

R() =

C Polynomials and rational functions

Heavyside coverup?

show that sum of the roots of P , counting multiplicities, is equal to pn1 .

EC.4 Determine the c.f.r. of the following rational functions:
+1
(a)
;
3 2 + 6
2
3 + 6 + 9
(b) 3
;
+ 5 2 + 7 + 3
2 + 2
(c)
;
( + 1)3
2 2 + 4
(d)
.
3 3 + 9 2 + 3 + 9
EC.5 Determine the partial fraction expansion of the following rational functions (for functions with complex poles, determine both the real and complex partial fraction expansions):
2 1
(a)
;
+2
3 + 4
(b) 2
;
+ 3 + 2
2 2 + 1
;
(c)
2 3 + 18 2 + 48 + 32
2
+2
(d)
.
( 2 + 1)2
EC.6 Let R C(). Show that R R() if and only if R(s) = R(
s) for every s C which
is not a pole for R.

Exercises for Chapter C

617

618

This version: 22/10/2004

D Complex variable theory

22/10/2004

Appendix D

]z ] 2 , ]

]z ]0, 2 ]

Complex variable theory

We shall require some basic facts about functions of a complex variable. We assume the
reader to have some background in such matters. Certainly we anticipate that the reader is
fully functional in manipulating complex numbers. For more details on what we say here,
we refer to [Lang 1993]. An excellent introduction to those topics in complex variable theory
that are useful in control may be found, complete with all details, in Appendix A of [Seron,
Braslavsky, and Goodwin 1997].

Contents

]z ] 2 , ]

]z ]0, 2 ]

Figure D.1 The values for the argument of a complex number

D.1 The complex plane and its subsets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 617

D.2 Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 618

The closed disk of radius r centered at z is denoted:

r) = { z 0 C | |z 0 z| r} .
D(z,

D.3 Integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 620

D.4 Applications of Cauchys Integral Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . 621
D.5 Algebraic functions and Riemann surfaces . . . . . . . . . . . . . . . . . . . . . . . . . . . 623

D.1 The complex plane and its subsets

The complex plane C is the set of ordered pairs (x, y) of complex
numbers. We will

follow the usual convention

of writing (x, y) = x + iy where i = 1. Note that we do not

use the symbol j for 1. Only electrical engineers, and those under their influence, use
this crazy notation. Complex numbers of the form x + i0 for x R are real and complex
numbers of the form z = 0 + iy for y R are called imaginary . For z = x + iy we denote
Re(z) = x and Im(z) = y as the real part and imaginary part, respectively, of z. We
will assume the reader knows how to add and multiply complex numbers:
(x1 + iy1 ) + (x2 + iy2 ) = (x1 + x2 ) + i(y1 + y2 )
(x1 + iy1 )(x2 + iy2 ) = (x1 x2 y1 y2 ) + i(x2 y1 + x1 y2 ).
The
p complex conjugate of z = x + iy is the complex number z = x iy. By |x + iy| =
x2 + y 2 we denote the modulus of z. If z is real, then |z| is the usual absolute value.
By ]x + iy = atan2(x, y) we denote the argument of z. Here atan2 : R2 (, ] is the
intelligent arctangent that knows which quadrant one is in. This is illustrated in Figure D.1.
Let us make a few definitions concerning the nature of subsets of C. First we denote by
D(z, r) the open disk of radius r centered at z:
D(z, r) = { z 0 C | |z 0 z| < r} .

Now we have the following.

D.1 Definition Let S C.
(i) S is open if for any z S there exists an > 0 so that D(z, ) S.
(ii) S is closed if its complement, C \ S, is open.
(iii) A boundary point for S is a point s C so that for every > 0 there exists
s1 , s2 D(s, ) so that s1 S and s2 6 S.
(iv) S C is connected any two points in S can be connected by a polygonal path
consisting of a finite number of line segments.
(v) D is simply connected if every closed curve in D can be continuously contracted to
a point.1 .
(vi) S is a domain if it is open and connected.
(vii) S is a region if it is a domain together with a possibly empty subset of its boundary.
(viii) A region is closed if it contains all of its boundary points.
(ix) A region S is bounded if there exists R > 0 so that S D(0, R).

D.2 Functions
Let D C be a domain. A function f : D C is continuous at z0 if for every > 0
there exists > 0 so that |z z0 | < implies that |f (z) f (z0 )| < . If f is continuous at
1
This definition refers ahead to Section D.3 for the notion of a closed curve. We do not often use the idea
of simple connectivity, but it intuitively means no holes.

D.2 Functions

22/10/2004

619

every point in D then f is simply continuous. The function f is differentiable at z0 if

the limit
f (z) f (z0 )
lim
zz0
z z0
exists and is independent of the manner in which the limit is taken.2 The limit when
it so exists is the derivative and denoted f 0 (z0 ). If we write z = x + iy and f (z) =
u(x, y)+iv(x, y) for R-valued functions u and v, then it may be shown that f is differentiable
at z0 = x0 + iy0 if and only if (1) u and v are differentiable at (x0 , y0 ) and (2) the CauchyRiemann equations are satisfied at z0 :
v
u
(x0 , y0 ) =
(x0 , y0 ),
x
y

620

D Complex variable theory

Analyticity of f guarantees pointwise convergence of the Taylor series in a closed disk of

positive radius. If z0 is an isolated singularity for f then the Taylor series is not a promising
approach to representing the function. However, one can instead use the Laurent series
given by

X
X
bj
.
f (z) =
aj (z z0 )j +
(z

z0 )j
j=0
j=1
The matter of expressing the coefficients in terms of f obviously cannot be done by evaluations of f and its derivatives at z0 . However, there are formulas for the coefficients involving
contour integrals. So. . .

u
v
(x0 , y0 ) = (x0 , y0 ).
y
x

A function f : D C on a domain D is analytic at z0 D if there exists > 0 so that

f is differentiable at every point in D(z0 , ). If R D is a region, we say f is analytic in R
if it is analytic at each point in R. Note that this may necessitate differentiability of f at
points outside R.
Analytic functions may fail to be defined at isolated points. Let us be systematic about
characterising such points.

22/10/2004

D.3 Integration
Much of what interests in complex variable theory centres around integration. In this
section we give a rapid overview of the essential facts.
A curve in C is a continuous map c : [a, b] C. A closed curve in C is a curve
c : [a, b] C for which c(a) = c(b). Thus a closed curve forms a loop with no intersections
(see Figure D.2). A curve c defined on [a, b] is simple if the restriction of c to (a, b) is
Im

D.2 Definition Let f : D C be analytic.

(i) A point z0 D is an isolated singularity for f if there exists > 0 so that f is
defined and analytic on D(z0 , r) \ {z0 } but is not defined on D(z0 , r).
(ii) An isolated singularity z0 for f is removable if there exists an r > 0 and an analytic
function g : D(z0 , r) C so that g(z) = f (z) for z 6= z0 .
(iii) An isolated singularity z0 for f is a pole if

c(a) = c(b)
c
c
Re

(a) limzz0 |f (z)| = and

(b) there exists k > 0 so that the function g defined by g(z) = (z z0 )k f (z) is analytic
at z0 . The smallest k Z for which this is true is called the order of the pole.
(iv) An isolated singularity z0 for f is essential if it is neither a pole nor a removable
singularity.
(v) A function f : D C is meromorphic if it analytic except possibly at a finite set of
poles.

Another important topic in the theory of complex functions is that of series expansions.
Let D be a domain. If f : D C is analytic at z0 D then one can show that all derivatives
of f exist at z0 . The Taylor series for f at z0 is then the series
f (z) =

aj (z z0 )j .

j=0

where the coefficients are defined by

aj =
2

f (j) (z0 )
.
j!

Thus for any sequence {zk } converging to z0 , the sequence

converge to the same complex number.

f (zk )f (z0 )
zk z0

should converge, and should

[
a

]
b

Figure D.2 A closed curve in C

injective. Thus for each t1 , t2 (a, b) the points c(t1 ) and c(t2 ) are distinct. Sometimes
a simple closed curve is called a Jordan curve. The Jordan Curve Theorem then states
that a simple closed curve separates C into two domains, the interior and the exterior. This
also allows us to make sense of the orientation of a simple closed curve. We shall speak of
simple closed curves as having clockwise orientation or counterclockwise orientation. Let
us agree not to give these precise notation as the meaning will be obvious in any application
we encounter.
Sometimes we will wish for a curve to have more smoothness, and so speak of a differentiable curve as one where the functions u, v : [a, b] R defined by c(t) = u(t) + iv(t)
are differentiable. For short, we shall call a differentiable curve an arc. In such cases we
denote
du
dv
c0 (t) =
+i .
dt
dt

22/10/2004

D.4 Applications of Cauchys Integral Theorem

621

622

D Complex variable theory

The length of a differentiable curve c : [a, b] C is given by

|c0 (t)| dt.

A contour is a curve that is a concatenation of a finite collection of disjoint differentiable

curves.
If c : [a, b] C is a curve then we define
Z

Z
c(t) dt =

Z
u(t) dt + i

22/10/2004

C1
z0

v(t) dt,
a

where u and v are defined by c(t) = u(t)+iv(t). Now we let D be a domain in C, c : [a, b] D
be an arc, and f : D C be a continuous function. We define
Z
Z b
f (c(t))c0 (t) dt.
(D.1)
f (z) dz =

One may verify that this integral does not in fact depend on the parameterisation of c, and
so really only depends onR the shape
of the image of c in U C. We shall typically denote
R
C = image(c) and write C = c . If c is a contour, then one may similarly define the integral
by defining it over each of the finite arcs comprising c. If F : D C is differentiable with
continuous derivative f , then one verifies that
Z
f (z) dz = F (c(b)) F (c(a)),
c

for a contour c : [a, b] C.

The following theorem lies at the heart of much of complex analysis, and will be useful
for us here.
D.3 Theorem (Cauchys Integral Theorem) Let D C be a simply connected domain, suppose that f : D C is analytic on the closure of D, and let C be a simple closed contour
contained in D. Then
Z
f (z) dz = 0.
C

Figure D.3 Contours for definition of Laurent series coefficients

is the Laurent series for f at z0 then we have

Z
1
f (z)
cj =
dz, j = 0, 1, . . .
2i C0 (z z0 )j+1
Z
1
f (z)
dz, j = 1, 2, . . .
cj =
2i C1 (z z0 )j+1
The residue of an analytic function f at an isolated singularity z0 is the coefficient c1
in the Laurent series for f at z0 . We denote the residue by
Z
1
Res f (z) =
f (z) dz,
z=pj
2i C
where C is some sufficiently small circular contour centred at z0 . The Residue Theorem is
also important for us.

D.4 Applications of Cauchys Integral Theorem

Cauchys Integral Theorem forms the basis for much that is special in the theory of
complex variables. We shall give a few of the applications that are of interest to us in this
book.
Let us begin by providing formulas for the coefficients in the Laurent expansion in terms
of contour integrals. The following result does the job.
D.4 Proposition Let f : D C be analytic and let z0 D be an isolated singularity for f .
Let C0 and C1 be circular contours centred at z0 with C1 smaller than C0 (see Figure D.3).
If

X
f (z) =
cj (z z0 )j
j=

D.5 Theorem (Residue Theorem) Let D C be a domain with C a simple, clockwiseoriented, closed contour in D. Let f : D C be meromorphic in the interior of C and
analytic on C. Denote the poles of f in the interior of C by p1 , . . . , pk . Then
Z
f (s) ds = 2i
C

k
X
j=1

Res f (s).

s=pj

D.6 Theorem
Another useful result is the Poisson Integral Formula.

22/10/2004

D.5 Algebraic functions and Riemann surfaces

623

624

D.7 Theorem (Poisson Integral Formula) Let D C be a domain containing the positive
complex plane C+ and let f : D C be analytic in C+ . Additionally, we will suppose that
if for R > 0 we define m(R) > 0 by
sup |f (Rei )|,

m(R) =

D Complex variable theory

Let us not formally define a what we mean by a Riemann surface, but proceed by example.
We consider first the degree 1 case where f satisfies the equation
a1 (z)f (z) + a0 (z) = 0

(D.2 )

f (z) =

[ 2 , 2 ]

lim

m(R)
= 0.
R

If z0 = x0 + iy0 C+ then we have

f (z0 ) =

f (iy)

f (z) = f0 (z) + f1 (z) + + fn (z)

x0
dy.
x20 + (x x0 )2

where f0 C[z] and where

The Poisson Integral Formula has the following useful corollary, stated by Freudenberg
and Looze [1985].

fj (z) =

R(z, w) =

an (z)f (z)n + + a1 (z)f (z) + a0 (z) = 0

where a0 , a1 , . . . , an C[z]. If an is not the zero polynomial then n is the degree of the
algebraic function f . Upon reflection, one sees that there is a problem with making this
definition precise since an algebraic function will not have as many as n possible solutions
for each z C. Thus an algebraic function is multi-valued. Since this is not an entirely clear
notion, one should attempt to come up with a framework in which an algebraic function can
be defined in a precise manner. The way in which this is done is by asking that an algebraic
function take its values not in C, but in what is called a Riemann surface. A classical
introductory reference is [Springer 1981].

ak (z)wk + + a1 (z)w + a0 (z)

b` (z)w` + + b1 (z)w + b0 (z)

It is not clear why we should be interested in this, but let us look at this in the degree 1
case. In this case, since f (z) is itself a rational function, R(, f ()) is a rational function
in , so has a partial fraction expansion. Using this, one may then explicitly evaluate the
integral (D.3) as being the sum of rational functions and logarithmic functions.

D.9 Theorem If f : D C is an analytic function on a domain D, then |f | has no maximum

on D unless f is constant.

In our discussion in Section 11.4 we shall need some not quite elementary concepts from
the theory of complex variables. An algebraic function is a function f of a complex
variable z satisfying an equation of the form

j = 1, . . . , k.

for a0 , a1 , . . . , ak , b0 , b1 , . . . , b` C[z]. In the study of Riemann surfaces it is useful to examine

integrals of the type
Z z
R(, f ()) d.
(D.3)

Finally, we state a sort of stray result, but one that is standard in complex variable
theory, the Maximum Modulus Principle.

D.5 Algebraic functions and Riemann surfaces

cmj ,j
c1,j
+ +
,
z pj
(z pj )mj

Now let R be a rational function in z and w; thus

D.8 Corollary Suppose that D is a domain containing C and that f : D C is analytic and
nonzero in C, with the possible exception of zeros on the imaginary axis. Also, assume that
ln f satisfies the equality (D.2). Then for each z0 = x0 + iy0 C+ we have
Z
x0
1
ln |f (iy)| 2
dy.
ln |f (z0 )| =

x0 + (x x0 )2

From this result it follows that if f is analytic in a closed bounded region R, then the
maximum value taken by |f | must occur on the boundary of R.

a0 (z)
C(z).
a1 (z)

Thus degree 1 algebraic functions are simply rational functions. Let us examine some of
the properties of such functions that may be helpful in our examination of higher-order
Riemann surfaces. Suppose that f has poles at p1 , . . . , pk C with respective multiplicities
m1 , . . . , mk . Write f using a complex partial fraction expansion:

then f has the property that

22/10/2004

analytic
continuation

Exercises for Chapter D

Exercises
ED.1 Graphical calculation of residues from Truxal (page 27)

625

626

D Complex variable theory

22/10/2004

626

This version: 22/10/2004

E Fourier and Laplace transforms

22/10/2004

PSfrag replacements

Fourier and Laplace transforms

For the most part, our use of transforms will be rather pedestrian. However, some
of the technical material, especially in Chapter 15, requires that we actually know a little
more than is often classically covered. Thus this appendix is a broader, although not terribly
detailed, treatment of Fourier and Laplace transforms than may be required of students only
engaging in the more straightforward parts of the book. Some uses of the Fourier or Laplace
transforms benefit from a cursory knowledge of distributions. We begin our discussion with
a presentation of distributions along these lines.

Contents
E.1 Delta-functions and distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 627
E.1.1 Test functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 627
E.1.2 Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 629
E.2 The Fourier transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 630
E.3 The Laplace transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 632
E.3.1 Laplace transforms of various flavours . . . . . . . . . . . . . . . . . . . . . . . . . 632
E.3.2 Properties of the Laplace transform . . . . . . . . . . . . . . . . . . . . . . . . . . 635
E.3.3 Some useful Laplace transforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . 638

E.1 Delta-functions and distributions

When rigour is not a concern, a delta-function at 0 is a function (t) with the property
that
Z
Z
(t) dt = 1,
f (t)(t) dt = f (0)

for any function f of unspecified character. It is actually not difficult to show that the
existence of such a function is an impossibility. However, there is a way to rectify this in
such a way that all the improper manoeuvres typically done with delta-functions are legal.
The idea on making this precise is due to Schwartz [1966].
E.1.1 Test functions
is the closure of the set

If f : R R is a function, the support of f , denoted supp(f ),

{ x R | f (x) 6= 0} .

A test function on R is a function : R R with the properties that

Re
One may verifyIm
that this function is indeed infinitely differentiable, and it clearly has compact
support. The function
is plotted in Figure E.1.

x1
x2
x1
x2 0.35
log
dB 0.3
deg 0.25
u = 0 ln
ln coth(|u| /2) 0.2
or 1
m 0.15

0.1
yos
0.05
tos 0
0

,0
-0.75 -0.5 -0.25 0
0.25 0.5 0.75
1
t
0
(t)

Appendix E

1. is infinitely differentiable and

ke(t)k
2. supp()
is bounded.
e1 (t)
The second condition
normally is stated as having compact support. We see that a test
e2 (t)
e3 (t) except on some interval of finite length. Note that this precludes from
function vanishes
being analytic.y(t)
This makes it somewhat non-obvious how to define test functions. However,
u(t)
here is a common
h (t) example of one such.
hN,D (t)
E.1 Example 1Let
(t) > 0 and define
1N,D (t)
(
2
(t)
exp( 2t2 ), |t| <
(t) =
0,
|t| .
f (t)

Figure E.1 The test function when = 0.5

Note that the set of test functions forms a vector space since the sum of two test functions
is also a test function, and any scalar multiple of a test function is also a test function. This
is then an infinite-dimensional vector space, and we denote it by T . Let us define the notion
of convergence in this vector space T . A sequence of test functions {j }jN converges to
zero if
1. there exists an interval I of finite length so that supp(j ) I for all j N and
(k)

2. for each k N0 = {0} N, the sequence of functions {j }jN converges uniformly to

the zero function.
We then say that a sequence of test functions {j }jN converges to a test function if
the sequence {j }jN converges to zero. A linear map L : T R is continuous if
the sequence {L(j )}jN of numbers converges for every convergent sequence {j }jN of test
functions. If I R is a closed interval of finite length, then TI denotes the subspace of T
consisting of those test functions for which supp() I.

22/10/2004

E.1 Delta-functions and distributions

627

E.1.2 Distributions Finally, with the above terminology in place, a distribution is

a linear map L : T R having the property that the restriction of L to TI is continuous
for each closed interval I of finite length.
Let us consider some examples of distributions.
E.2 Example 1. Let f : R R have the property that f is integrable over any interval of
finite length. Then associated to f is the distribution Lf defined by
Z
f (t)(t) dt.
Lf () =

Thus integrable functions may themselves be thought of as distributions. That is to say,

distributions generalise functions.
2. Let us indicate that we may think of the delta-function as a distribution. Consider the
linear map : T R defined by L() = (0). If I R is a closed interval of finite
length which does not contain 0 then the restriction of to TI is obviously continuous: it
is identically zero. Now let I be a closed finite-length interval containing 0. Let {j }jN
be a sequence of test functions converging to a test function and so that j TI for
j N. By definition of convergence in T , the sequence {j (0)}jN converges to (0).
This shows that the sequence {(j )}jN converges, so showing that is continuous on
TI as desired. Thus is indeed a distribution as we have defined it.
Note that the notion of a distribution suggests that we write () for the value of
applied to a test function . However, custom dictates that we write
Z
() =
(t)(t) dt = (0).

3. If L is a distribution, then one may verify that the linear mapping L : T R defined
by

L()
= L()
is itself a distribution. This is the derivative of L, indicating that one can always
differentiate a distribution. If f is a continuously differentiable function, then one may

verify that L f = Lf . Also, the derivative of the delta-function is defined by ()

= (0).
4. Let us combine 2 and 3 to show that (t) is the derivative of the unit step function 1(t).
By definition of the derivative we have, for every test function ,
Z
Z

=
dt =
= (t) = (0),

1()
= 1()
1(t)(t)
(t)
0

as desired.

It turns out that many of the manipulations valid for functions are valid for distributions.
As we saw in Example 3 above, one can differentiate an arbitrary distribution in straightforward manner. Since distributions are generalisations of locally integrable functions, this
by implication means that it is possible to define the derivative, in the distributional sense,
of functions that are not differentiable! One says that a distribution has order k if k is the
smallest integer for which L = f (k+1) for an integrable function f . Thus the order measures
how far away a distribution is from a being function. One may show that every distribution
has a finite order. For example, since is the derivative of the locally integrable function

628

E Fourier and Laplace transforms

22/10/2004

t 7 1(t), has order 0. Likewise, is order 1. If L is a distribution of order k and if f is at

least k-times continuously differentiable, then it is possible to define the product of L with
f to be the distribution f L given by (f L)() = L(f ).
E.3 Example To multiply by a function and get a distribution, f has to be at least continuous since has order 0, In this case, if f is continuous, then (f )() = f (0)(0) for
T.

If one can define the product f L for a function f and a distribution L, then the result can
be differentiated in the distributional sense. One can easily show that the derivative of this
d
provided that the product fL
(f L) = fL + f L,
product satisfies the usual product rule: dt
makes sense as a distribution.
E.4 Example Let f : R R be k-times continuously differentiable and define g(t) = 1(t)f (t).
This function may be differentiated n times in a distributional sense, and the derivatives are
computed using the product rule:
g(t) = 1(t)f (t)
g (1) (t) = f (0)(t) + 1(t)f (1) (t)
g (2) (t) = f (0) (1) (t) + f(0)(t) + 1(t)f (2) (t)
..
.
g (n) (t) =

n
X

f (nj) (0) (j1) (t) + 1(t)f (n) (t).

j=1

This formula is very useful in Section 3.6.2 when discussing the solution of differential equations using the left causal Laplace transform.

As we have seen, a distribution is generally not representable as a function. However,
what is true is that any distribution is a limit of a sequence of functions. This necessitates
saying what we mean by convergence of distributions. A sequence {Lj }jN of distributions
converges to a distribution L if for every T , the sequence of numbers {Lj ()}jN converges to L(). What is then true is that every distribution is the limit of a sequence {fj }jN
of infinitely differentiable functions, with these functions being regarded as distributions.
E.5 Example One can show that = limj fj where
n2 t2
j
fj (t) = exp
.
2
2
In Figure E.2 we plot fj for a couple of values of j, and we can see the anticipated behaviour
of concentration of the function near 0.

E.2 The Fourier transform

Before we begin talking about transforms, we need a few basic notions of integration. Some of this is discussed in more generality and detail in Section 5.3. A function

1N,D (t)
(t)
(t)
Re
Im
x1
x2
x1
x2
log
dB
deg
u = 0 ln
ln coth(|u| /2)
or 1
m

yos

tos 0

E.2 The Fourier transform

22/10/2004

630

E Fourier and Laplace transforms

lim

22/10/2004

f: (, ) C so that
Z

fj (t)

,0
0

629

-4

-2

Figure E.2 A sequence of functions converging to

f : (, ) C1 is L2 -integrable if
Z

1/2
|f (t)|2 dt
< .

In this case, we denote by kf k2 the quantity on the left-hand side of the inequality. In like
manner, a function f : (, ) C is L1 -integrable if
Z
|f (t)| dt < .

f() fT () dt = 0.

Thus the functions fT converge in mean to the function f which we call the L2 -Fourier
transform of the L2 -integrable function f . Note that the L2 -Fourier transform has the
interesting property of mapping L2 -integrable functions to L2 -integrable functions. The
inverse of the L2 -Fourier transform must therefore exist. Indeed, given an L2 -integrable
function f: R C, its inverse Fourier transform is given by the L2 -integrable function
f : (, ) C defined by
Z
1
f (t) =
f()eit d.
2
It will on occasion be convenient to use the notation F for the map that sends an L1 integrable function to its Fourier transform. Thus F (f ) = f. We likewise denote the
inverse for the L2 -Fourier transform by F 1 .
For L2 -integrable functions f and g, one can readily verify Parsevals Theorem which
states that
Z
Z
1
f()
g () d.
f(t)g(t) dt =
2

In particular, it follows that kf k2 = 12 kfk2 . Also important to us is the notion of convolution. Given L1 -integrable functions f and g, we define a new L1 -integrable function,
denoted f g and called the convolution of f with g, defined by
Z
(f g)(t) =
f (t )g( ) d.

In this case, we denote by kf k1 the quantity on the left-hand side of the inequality. Finally,
a function f : (, ) C is L -integrable if |f (t)| < for almost every t R. In this
case, we denote by kf k the least upper bound of |f (t)|, t R.
With these notions of integrability at hand, we may define the Fourier transform of
an L1 -integrable function f : (, ) C as the function f: (, ) C given by given
by
Z
f() =
f (t)eit dt.

Convolution in the time-domain is equivalent to multiplication in the frequency domain:

(f g)= fg.
We shall encounter convolution in various contexts using both Fourier and Laplace transforms. It is important to realise just what kind of convolution with which one is dealing!

E.3 The Laplace transform

Note that f() kf k1 , so that f is L -integrable. If we wish to emphasise that we are
transforming an L1 -integrable function, we shall state that we are using the L1 -Fourier
transform. It is also possible to take the Fourier transform of an L2 -integrable function
f : (, ) C, and this is done as follows. For T > 0 define fT : (, ) C by
(
f (t), |t| T
fT (t) =
0,
otherwise.
One may verify that fT is L1 -integrable for any finite T , so that its L1 -Fourier transform, fT , exists. Whats more, it can be shown that there exists an L2 -integrable function
1
We deal in this appendix always with C-valued functions of t. In the majority of instances in the text,
the functions are R-valued. However, for the general presentation here, it is convenient to consider C-valued
functions.

The Laplace transform is related to the Fourier transform in a sort of simple way. However, since the relationship can get a little complicated, we try to be clear about the notion
of the Laplace transform. Also, we will consider Laplace transforms of various sorts of functions, so we must take care to distinguish these with proper notation. We remark that this
is not normally done in control texts, with the result that there are sometimes contradictions present that apparently go unnoticed. A good introductory discussion of the Laplace
transform, minus any detailed discussion of distributions, can be found in [Schiff 1999]. The
distributional theory is carried out in detail by Zemanian [1987].
E.3.1 Laplace transforms of various flavours Let us first establish some notation.
Let f : (, ) C be an L1 -integrable function. If Re(s) = 0 then the integral
Z
f (t)est dt
(E.1)

22/10/2004

E.3 The Laplace transform

631

exists; it is, of course, simply the Fourier transform. By continuity of the integral, if Re(s)
is sufficiently close to zero, the integral (E.1) exists. To define the Laplace transform, we
wish to ascertain a subset of C so that if s lies in this set, the integral (E.1) exists. In doing
so, we do not necessarily require that f itself be L1 -integrable.
Let f have the property that the function f (t)ect is L1 -integrable for some c R. That
is, for some c R we have
Z

|f (t)| ect dt < .

(E.2)

For such functions, call the real number

632

E Fourier and Laplace transforms

22/10/2004

transforms. The left causal Laplace transform for a simple distribution f as given
+
(f ) : Rc (f0 ) C defined by
in (E.4) is the map L0
Z
Z
+
f (t)est ds.
L0
(f )(s) = lim
f (t)est ds =
0

Note that the left causal transform includes the effect of the delta-functions. Indeed, if f
is k-times continuously differentiable we can simply evaluate the contributions of the deltafunctions and see that
k
X
+
L0
(f )(s) = L (f0 ) +
c j sj .

min (f ) = inf { c R | the inequality (E.2) is satisfied}

j=0

the minimum abscissa of absolute convergence. Similarly we define

max (f ) = sup { c R | the inequality (E.2) is satisfied} .
We call max (f ) the maximum abscissa of absolute convergence. We then define

In contrast to this, the right causal Laplace transform does not include the contri+
butions of the delta-functions. It is defined to be the map L0+
(f ) : Rc (f0 ) C defined
by
Z
Z
+
f (t)est ds.
L0+
(f )(s) = lim
f (t)est ds =
0

+
In this case, since the delta-functions do not contribute, we simply have L0+
(f )(s) = L (f0 ).
Thus for genuine functions f (i.e., that do not involve distributions), we have

Rc (f ) = { s C | min (f ) < Re(s) < max (f )}

which we call the region of absolute convergence. We define the two-sided Laplace
transform of f as L (f ) : Rc (f ) C defined by
Z
L (f )(s) =
f (t)est dt.
(E.3)

For functions f : (, ) C that have the property that f (t)ect is L2 -integrable, the
inverse Laplace transform exists, and is defined by
Z +i
1
L (f )(s)est ,
L 1 (L (f )) =
2i i

+
+
L (f ) = L0
(f ) = L0+
(f ),

provided that f (t) = 0 for t < 0. Note that for functions f vanishing for negative times we
always have max (f ) = .
We will also have occasion to consider functions of time that are zero for positive times.
As above, let f be a simple distribution as given by (E.4), but now assume that f0 (t) = 0
for t > 0. We shall of course assume that the two-sided Laplace transform of f0 still exists.

The left anticausal Laplace transform of f is the map L0

(f ) : Rc (f0 ) C defined
by
Z
0

L0
(f )(s) =

f (t)est ds.

where is any number satisfying min (f ) < < max (f ).

Most commonly we will be dealing with the Laplace transform of functions that are zero
for negative times. We will also want to be able to take Laplace transforms of distributions
involving delta-functions and derivatives of delta-functions at 0. In these cases, it matters
when performing an integral over [0, ) whether we take the lower limit as 0+ or 0. To
be precise, suppose that f0 : (, ) C possesses a two-sided Laplace transform and has
the property that f0 (t) = 0 for t < 0. Now consider the distribution
f (t) = f0 (t) +

k
X

(j)

cj (t).

(E.4)

Note that the left anticausal transform does not include the effect of the delta-functions.

Thus we may write L0

(f )(s) = L (f0 ). Proceeding in the natural manner the right

anticausal Laplace transform is defined to be the map L0+

(f ) : Rc (f0 ) C defined by

L0+
(f )(s) =

f (t)est ds.

This transform does include the effects of the delta-functions. If f is k-times continuously
differentiable we can simply evaluate the contributions of the delta-functions and see that

j=0

Thus f is a distribution that is a sum of a function which vanishes for negative times and a
finite sum of delta-functions and derivatives of delta-functions. One can consider taking the
Laplace transform of more general distributions, but we shall not need this level of generality,
and its use necessitates a significant diversion [see Zemanian 1987]. A distribution of the
form (E.4), and for which f0 vanishes for negative times and possesses a two-sided Laplace
transform, will be called simple. For simple distributions we define two kinds of Laplace

L0+
(f )(s) = L (f0 ) +

k
X

c j sj .

j=0

Thus for genuine functions f (i.e., that do not involve distributions), we have

L (f ) = L0
(f ) = L0+
(f ),

provided that f (t) = 0 for t > 0. For anticausal functions f we always have min (f ) = .

22/10/2004

E.3 The Laplace transform

633

E.6 Notation When there is no possibility of confusion, we shall denote the Laplace transform
of a function f by f. If f is a simple distribution that is not a function, we shall always
specify which transform we use. However, if f is a function with no distributional component,
the various transforms are all equal, and there is no potential for confusion, unless one wants

to consider the Laplace transform of the derivative, cf. Theorem E.7.

634

E Fourier and Laplace transforms

where 1 is to be understood in the distribution sense, i.e., 1(t)

= (t). We now compute
+
L0
(f )(s) = lim

0+
R

f(t)est dt

= lim

0+
R

(t)f(t) + 1(t)f(t) est dt

E.3.2 Properties of the Laplace transform An important property of the Laplace

transform is how it acts with respect to the derivative. This is a place where, as we see, the
left and right transforms differ.

= f(0) + lim

E.7 Theorem Let f : R R be causal and continuous on [0, ) and suppose that f is
piecewise continuous on [0, ). If the Laplace transform of f exists then

= f (0) + lim

+
+
+
(f )(s)
(f )(s) = sL0
(f )(s) = sL0+
L0

for s Rc (f ).
Proof We use integration by parts to compute
Z R
+
L0+
(f )(s) = lim
f(t)est dt
0+
R

Z
R

= lim f (t)est + lim
0+
R

0+
R

f (t)sest dt.

Now, if s Rc (f ) then it must be the case that limR f (R)esR = 0. Taking also the
limit as 0+ we see that

= f (0) +

1(t)f(t)est dt

f (t)est dt

0+
R
+
(f )(s)
sL0

+
(f )(s).
f (0) = sL0

Here we use the fact that since f does not involve a delta-function at t = 0, we have
+
+
L0
(f )(s) = L0+
(f )(s).

We see that the right causal Laplace transform has the capacity to involve the value
of f at 0+. For this reason, it is useful to use this transform when solving initial value
problems, if that is your preferred way to do such things. However, for general control
theoretic discussions, the left causal Laplace transform is often the preferred tool since it
provides a simple relationship between the Laplace transform of a function and its derivative.
In any case, both shall appear at certain times in the text.
A repeated application of Theorem E.7 gives the following result.
E.8 Corollary Let f : R R be causal and suppose that f, f (1) , . . . , f (n1) are continuous on
[0, ) and that f (n) is piecewise continuous on [0, ). Then
+
+
L0+
(f )(s) = sn L0+
(f )(s)

+
+
L0+
(f )(s) = sL0+
(f )(s) f (0+)

as stated.
For the second part of the theorem, define f: R R by
(
f (x), x 0
f(x) =
f (0), x < 0.

0+
R

+
+
(f )(s) f (0+)
(f )(s) = sL0+
L0+

and

22/10/2004

n1
X

sj y (nj1) (0+)

j=0

and
+
+
L0
(f )(s) = sn L0
(f )(s)

if s Rc (f ).

Thus f is continuous on R and f is piecewise continuous on R. We then have f (t) = 1(t)f(t)

where 1(t) is the unit step function, and so we obtain

The convolution also has the same useful properties for Laplace transforms as for Fourier
transforms. To be clear, suppose that the Laplace transforms of f, g : (, ) C exist
and that Rc (f ) Rc (g) 6= . If the convolution f g is defined, then its Laplace transform
is defined, and we further have

f(t) + 1(t)f(t),
f(t) = 1(t)

L (f g) = L (f )L (g);
thus the Laplace transform of the convolution is the product of the Laplace transforms, and
it is defined on the region Rc (f ) Rc (g) C. In the text, we shall consider the cases when
f and g are both zero for either positive or negative times. For example, if f is a causal
function then
Z
(f g)(t) =
f ( )g(t ) d
0

E.3 The Laplace transform

22/10/2004

635

636

(f g)(t) =

f (t )g( ) d.
0

Similar statements hold for anticausal functions.

The following result is often helpful.

exists and is equal to

E.9 Proposition Let f : R R be a causal function, continuous on [0, ), and suppose that
f is piecewise continuous on [0, R). If the Laplace transform of f exists, then the following
statements hold:
+
(i) f (0+) = lims sL0+
(f )(s) (s real);
+
(ii) lim f (t) = lim sL0+ (f )(s) (s real), provided that the limit on the left exists.
t

22/10/2004

Proof Note that min (f ) , and so if Re(s0 ) > , then s0 is in the domain of definition
of the transform. Therefore the integral
Z
es0 t f (t) dt

and if g is a causal function then

E Fourier and Laplace transforms

+
L0+
(f )(s0 ).

E.3.3 Some useful Laplace transforms You will recall some standard Laplace transforms which we collect in Table E.1. Here 1(t) is the unit step function defined by

Proof (i) By Theorem E.7 we have

Z
+
f(t)est dt = sL0+
(f )(s) f (0+).

f (t)

f(s)

min (f )

f (t) + g(t)
af (t)

f(s) + g(s)
af(s)

unknown
min (f )

f(t)

1(t)

If we take the limits as s we may switch the limit and the integral since the integral
converges absolutely when s 0. This gives

Rb (t)

f (0+) + sf(s) unknown

1
0
s
b
0
s2

f (t)
eat
sin t
cos t

f(s)
1
sa

s2 + 2
s
s2 + 2

min (f )
a
0
0

+
0 = lim sL0+
(f )(s) f (0+)
s

Table E.1 Some common Laplace transforms (an unknown

means that is it not generally determinable in terms of min (f ))

from which our first assertion follows.

(ii) By Theorem E.7 we have
Z
+
f(t)est dt = sL0+
(f )(s) f (0+).

(
1, t 0
1(t) =
0, otherwise

We take the limit as s 0 of both sides, and move the limit inside the integral since the
integral is absolutely convergent in a neighbourhood of s = 0. This gives
Z
+
f(t) dt = lim sL0+
f )(s) f (0+)
0+

and Rb (t) is the ramp function defined by

(
bt, t 0
Rb (t) =
0, otherwise.

+
lim f (t) f (0+) = lim sL0+
(f )(s) f (0+)

from which the result follows, under the proviso that limt f (t) exists.

ref

We will also need some not so common Laplace transforms in order to make a conclusive
stability analysis using the impulse response. You can look up that

The second assertion of the proposition is often called the Final Value Theorem, and it
does require the hypothesis that limt f (t) exist.
The following result is one of a similar nature, but involves the integral of the function of
time in terms of its Laplace transform. These results will be interesting for us in Section 8.3
when we discuss various aspects of controller performance.
+
(f )(s) is a strictly
E.10 Proposition Let f (t) have the property that its Laplace transform L0+
proper rational function with the property that there exists > 0 so that if s is a pole of
+
L0+
(f )(s) then Re(s) . Then, for any s0 with Re(s0 ) > we have
Z
+
es0 t f (t) dt = lim L0+
(f )(s).
0

ss0

f (t) = tk eat sin t

!

X
k!
k+1
m
2m+1
k2m

f (s) =
(1)

(s a)
k+1
2m + 1
2
(s a) + 2
02mk

and

f (t) = tk eat cos t

!

X
k!
k+1
f(s) =
(1)m
2m (s a)k2m+1

2m
2 + 2 k+1
(s

a)
02mk+1

E.3 The Laplace transform

22/10/2004

637

where, you recall, for integers k and ` with k ` we have

k!
k
=
.
`
`!(k `)!
These are pretty ugly expressions. Lets observe some general features. The Laplace transforms f(s) are polynomials where the degree of the numerator is strictly less than that of
the denominator. As special cases of these formulas we have
f (t) = tk eat

f(s) =

f (t) = eat sin t

f(s) =

f (t) = eat cos t

f(s) =

f (t) = t sin t

f(s) =

f (t) = t cos t

f(s) =

k!
(s a)k+1

(s a)2 + 2
sa
(s a)2 + 2
2s
(s2 + 2 )2
s2 2
.
(s2 + 2 )2

We shant fiddle much with the definition of the inverse Laplace transform, but it will
be helpful to know a couple of them which cannot be obviously gleaned from Table E.1. We
have
1
f(s) =
k
(s )2 + 2

k
et X 2k i 1
di
= f (t) = k1 2k
(2t)i1 i cos t
k1
4
dt
i=1

and
s
k
(s )2 + 2

k

et X 2k i 1 (2t)i1 di
f (t) = k1 2k
cos t + sin t
i
k1
4
(i

1)!
dt
i=1

k1

X
2k i 2 (2t)i1 di
sin t
2
i
k1
(i 1)! dt
i=1

f(s) =

The import of these ridiculous expressions is contained in the following, now self-evident,
statement.
E.11 Proposition The functions
1
(s )2 + 2

k ,

s
(s )2 + 2

k ,

, R, k N,

are in one-to-one correspondence with the Laplace transforms of functions which are linear
combinations of
t` et sin t and t` et cos t,

, R, k N, ` = 0, . . . , k 1.

(E.5 )

638

E Fourier and Laplace transforms

22/10/2004

Note that we have in this section established (without proof) a perfect correspondence
between functions which are rational in the Laplace transform variable s, and functions
of time of the form (E.5). That is to say, a strictly proper rational function in s is the
Laplace transform of a linear combination of functions like (E.5) and the Laplace transform
of a function like (E.5) is a rational function in s. This fact will be useful in our later
investigations of the behaviour of SISO linear systems.

Exercises for Chapter E

639

Exercises
EE.1 What are the abscissa of absolute convergence, min (f ) and max , for the function
f (t) = 1(t)eat ? How are the values of min (f ) and max (f ) reflected in the properties
of the Laplace transform for f ?
EE.2 Show that min (f ) = and max (f ) = for functions f of the form f (t) =
1(t)t` et sin t or f (t) = 1(t)t` et sin t.
EE.3 Let f (t) = 1(t) sin t. Compute lims0 sf(s) and determine whether limt f (t) =
lims0 sf(s). Explain your conclusion in terms of Proposition E.9(ii).
EE.4 Show that the Laplace transform of eAt is (sI n A)1 . What do you think are the
abscissa of absolute convergence?
EE.5 Let f and g be functions whose causal right Laplace transforms exist and whose
domains have a nonempty intersection. Prove that the inverse Laplace transform of
f(s)
g (s) is either of the expressions
Z t
Z t
f (t )g( ) d,
g(t )f ( ) d.
0

EE.6 Fix T > 0 and a function g whose causal right Laplace transform exists and is g.
Show that the causal right Laplace transform of the function f (t) = g(t T ) is
f(s) = eT s g(s). The function f is called the time delay of g by T .
EE.7 Using the adjugate (see Section A.3.1), determine the inverse of the matrix sI 3 A
where s R and

1 0 1
A = 0 2 1 .
1 1 1
You may suppose that s is such that det(sI 3 A) 6= 0.

640

E Fourier and Laplace transforms

22/10/2004

644

This version: 22/10/2004

Bibliography

Philadelphia, ISBN 0-03-060289-0, major revision of 1970 version, and more thorough
than the 1999 Oxford University Press version.

Bibliography

Chen, J. [1993] Static output feedback stabilization for SISO systems and related problems:
Solutions via generalised eigenvalues, Control Theory and Advanced Technology, 10(4),
22332244.

Ackermann, J. E. [1972] Der entwurf linearer regelungsysteme im zustandsraum, Regelungstech. Prozess-Datenvararb, 7, 297300.

Chen, M. J. and Desoer, C. A. [1982] Necessary and sufficient condition for robust stability
of linear distributed feedback systems, International Journal of Control, 35(2), 255267.

Adamjan, V. M., Arov, D. Z., and Krein, M. G. [1968a] Infinite Hankel matrices and generalized problems of Caratheodory-Fejer and F. Riesz, Functional Analysis and its Applications, 2(1), 119.

Chen, W. K. [1976] Applied Graph Theory, second edition, number 13 in North-Holland

Series in Applied Mathematics and Mechanics, North-Holland, Amsterdam/New York.

[1968b] Infinite Hankel matrices and generalized problems of Caratheodory-Fejer and I.

Schur, Functional Analysis and its Applications, 2(4), 117.
[1971] Analytic properties of Schmidt pairs for a Hankel operator and the generalized
Schur-Takagi problem, Mathematics of the USSR-Sbornik, 15, 3173.
Anderson, B. D. O. [1972] The reduced Hermite criterion with application to proof of the
Lienard-Chipart criterion, Institute of Electrical and Electronics Engineers. Transactions
on Automatic Control, AC-17(5), 669672.
Anderson, B. D. O., Jury, E. I., and Mansour, M. [1987] On robust Hurwitz polynomials,
Institute of Electrical and Electronics Engineers. Transactions on Automatic Control,
32(10), 909913.

Cook, S. A. [1971] The complexity of theorem-proving techniques, in Conference Record of

Third Annual ACM Symposium on Theory of Computing, pages 151158.
Dasgupta, S. [1988] Kharitonovs theorem revisited, Systems & Control Letters, 11(5), 381
384.
Davis, J. H. [2002] Foundations of Deterministic and Stochastic Control, Systems & Control:
Foundations & Applications, Birkhauser, Boston/Basel/Stuttgart, ISBN 0-8176-4104-9.
Desoer, C. A. and Chan, W. S. [1975] The feedback interconnection of lumped linear timeinvariant systems, Journal of the Franklin Institute, 300(5/6), 335351.
Dorf, R. C. and Bishop, R. H. [2001] Modern Control Systems, 9th edition, Prentice-Hall,
Englewood Cliffs, NJ, ISBN 0-13-030660-6.

Basile, G. and Marro, G. [1968] Luoghi caratteristici dello spazio degli stati relativi al controllo dei sistemi lineari, LElettrotecnica, 55(12), 17.

Doyle, J. C. [1978] Guaranteed margins for LQG regulators, Institute of Electrical and Electronics Engineers. Transactions on Automatic Control, AC-23(4), 756757.

Blondel, V. and Tsitsiklis, J. M. [1997] NP-hardness of some linear control design problems,
SIAM Journal on Control and Optimization, 35(6), 21182127.

[1982] Analysis of feedback systems with structured uncertainty, Institution of Electrical

Engineers. Proceedings. D. Control Theory and Applications, 129(6), 242250.

Bode, H. W. [1945] Network Analysis and Feedback Amplifier Design, Van Nostrand Reinhold
Co., London, ISBN 0-882-75242-1.

Doyle, J. C., Francis, B. A., and Tannenbaum, A. R. [1990] Feedback Control Theory, Macmillian, New York, ISBN 0-02-330011-6.

Bott, R. and Duffin, R. J. [1949] Impedance synthesis without use of transformers, Journal
of Applied Physics, 20, 816.

Doyle, J. C. and Stein, G. [1981] Multivariable feedback design: Concepts for a classical modern synthesis, Institute of Electrical and Electronics Engineers. Transactions on Automatic
Control, AC-26(1), 416.

Boyce, W. E. and Diprima, R. C. [2000] Elementary Differential Equations, 7th edition,

John Wiley and Sons, New York, New York, ISBN 0-471-31998-8.
Brockett, R. W. [1970] Finite Dimensional Linear Systems, John Wiley and Sons, New York,
New York.

Dullerud, G. E. and Paganini, F. [1999] A Course in Robust Control Theory, number 36

in Texts in Applied Mathematics, Springer-Verlag, New York-Heidelberg-Berlin, ISBN
0-387-98945-5.

Bryson, A. E. and Ho, Y. C. [1975] Applied Optimal Control, John Wiley and Sons, New
York, New York, ISBN 0-470-11481-9, revised printing.

El Ghaoui, L. and Niculescu, S.-I., (editors) [2000] Advances in Linear Matrix Inequality
Methods in Control, Advances in Design and Control, Society for Industrial and Applied
Mathematics, Philadelphia, Pennsylvania, ISBN 0-89871-438-9.

Cannon, Jr., R. H. [1967] Dynamics of Physical Systems, McGraw-Hill, New York, ISBN
007-009754-2.

Evans, W. R. [1948] Graphical analysis of control systems, Transactions of the American

Institute of Electrical Engineers, 67, 547551.

Chapellat, H. and Bhattacharyya, S. P. [1989] An alternative proof of Kharitonovs theorem,

Institute of Electrical and Electronics Engineers. Transactions on Automatic Control,
34(4), 448450.

[1950] Control systems synthesis by root locus method, Transactions of the American
Institute of Electrical Engineers, 69, 6669.

Chen, C.-T. [1984] Linear System Theory and Design, HRW Series in Electrical and Computer Engineering, Holt, Rinehart and Winston, New York, Chicago, San Francisco,

Francis, B. A. [1987] A Course in H Control Theory, number 88 in Lecture Notes in

Control and Information Sciences, Springer-Verlag, New York-Heidelberg-Berlin, ISBN
0-387-17069-3.

Bibliography

645

646

Bibliography

Francis, B. A. and Zames, G. [1984] On H -optimal sensitivity theory for SISO feedback
systems, Institute of Electrical and Electronics Engineers. Transactions on Automatic
Control, 29(1), 916.

Kalman, R. E. [1960] A new approach to linear filtering and prediction theory, American
Society of Mechanical Engineers. Transactions of the ASME. Series D. Journal of Basic
Engineering, 82, 3545.

Franklin, G. F., Powell, J. D., and Emani-Naeini, A. [1994] Feedback Control of Dynamics
Systems, third edition, Addison Wesley, Reading, MA, ISBN 0-201-52747-2.

Kalman, R. E. and Bucy, R. S. [1960] New results in linear filtering and prediction theory,
American Society of Mechanical Engineers. Transactions of the ASME. Series D. Journal
of Basic Engineering, 83, 95108.

Freudenberg, J. S. and Looze, D. P. [1985] Right half plane poles and zeros and design tradeoffs in feedback systems, Institute of Electrical and Electronics Engineers. Transactions on
Automatic Control, 30(6), 555565.
Fuhrmann, P. A. [1996] A Polynomial Approach to Linear Algebra, Universitext, SpringerVerlag, New York-Heidelberg-Berlin, ISBN 0-387-94643-8.

Fujiwara, M. [1915] Uber

die Wurzeln der algebraischen Gleichungen, The Tohoku Mathematical Journal. Second Series, 8, 7885.
Gantmacher, F. R. [1959a] The Theory of Matrices, volume 1, Chelsea, New York, ISBN
0-8218-1376-5, translated from the German by K. A. Hirsch.
[1959b] The Theory of Matrices, volume 2, Chelsea, New York, ISBN 0-8218-2664-6,
translated from the German by K. A. Hirsch.
Geromel, J. C., de Souza, C. C., and Skelton, R. E. [1998] Static output feedback controllers:
Stability and convexity, Institute of Electrical and Electronics Engineers. Transactions on
Automatic Control, 43(1), 120125.
Goodwin, G. C., Graebe, S. F., and Salgado, M. E. [2001] Control System Design, PrenticeHall, Englewood Cliffs, NJ, ISBN 0-13-958653-9.

Hang, C. C., Astrom, K. J., and Ho, W. K. [1990] Refinements of the Ziegler-Nicols tuning formula, Technical Report CI-90-1, National University of Singapore, Department of
Electrical Engineering.
Hautus, M. L. J. [1969] Controllability and observability conditions of linear autonomous
systems, Koninklijke Nederlandse Akademie van Wetenschappen. Proceedings. Series A.
Mathematical Sciences, 31, 443448.
Helton, J. W. and Merrino, O. [1998] Classical Control Using H Methods, Theory, Optimization, and Design, Society for Industrial and Applied Mathematics, Philadelphia,
Pennsylvania, ISBN 0-89871-419-2.
Hermite, C. [1854] Sur le nombre des racines dune equation alg`ebrique comprise entre des
limites donnees, Journal f
ur die Reine und Angewandte Mathematik, 52, 3951.
Horowitz, I. M. [1963] Synthesis of Feedback Systems, Academic Press, New York.

Hurwitz, A. [1895] Uber

di Bedingungen unter welchen eine Gleichung nur Wurzeln mit
negativen reellen Teilen besitz, Mathematische Annalen, 46(273-284).
Jr., J. R. M. [1949] The effect of pole and zero location on the transient response of linear
dynamics systems, Institution of Electrical Engineers. Proceedings. D. Control Theory and
Applications, 37, 516529.
Kailath, T. [1980] Linear Systems, Information and System Sciences Series, Prentice-Hall,
Englewood Cliffs, NJ, ISBN 0-13-536961-4.

Kalman, R. E., Ho, Y.-C., and Narendra, K. S. [1963] Controllability of linear dynamical
systems, Contributions to Differential Equations, 1, 189213.
Kamen, E. W. and Su, J. K. [1999] Introduction to Optimal Estimation, Advanced Textbooks
in Control and Signal Processing, Springer-Verlag, New York-Heidelberg-Berlin, ISBN 185233-133-X.
Khalil, H. K. [1996] Nonlinear Systems, second edition, Prentice-Hall, Englewood Cliffs, NJ,
ISBN 0-13-228024-8.
Kharitonov, V. L. [1978] Asymptotic stability of an equilibrium position of a family of systems
of linear differential equations, Differentsialnye Uravneniya, 14, 20862088.
Krall, A. M. [1961] An extension and proof of the root-locus method, Journal of the Society
for Industrial and Applied Mathematics, 9(4), 644653.
Kucera, V. and de Souza, C. E. [1995] A necessary and sufficient condition for output feedback stabilizability, Automatica. The Journal of IFAC. The International Federation of
Automatic Control, 31(9), 13571359.
Lang, S. [1984] Algebra, second edition, Addison Wesley, Reading, MA, ISBN 0-201-05487-6.
[1993] Complex Analysis, third edition, number 103 in Texts in Applied Mathematics,
Springer-Verlag, New York-Heidelberg-Berlin, ISBN 0-387-97886-0.
Liapunov, A. M. [1893] A special case of the problem of stability of motion, Matematicheski
Sbornik. Novaya Seriya, 17, 252333.
Lienard, A. and Chipart, M. [1914] Sur la signe de la partie reelle des racines dune equation
algebrique, Journal de Mathematiques Pures et Appliquees. Neuvi`eme Ser, 10(6), 291346.
Lynch, W. A. [1961] Linear control systemsa signal-flow-graph viewpoint, in Adaptive Control Systems, chapter 2, pages 2349, McGraw-Hill, New York.
MacFarlane, A. G. J. [1982] Complex variable methods in feedback systems analysis and
design, in Design of Modern Control Systems, D. J. Bell, P. A. Cook, and N. Munro,
editors, chapter 2, pages 1845, number 20 in IEE Control Engineering Series, Peter
Peregrinus, Ltd., Stevanage, ISBN 0-906048-74-5.
Mansour, M. and Anderson, B. D. O. [1993] Kharitonovs theorem and the second method of
Lyapunov, Systems & Control Letters, 20(3), 3947.
Marshall, D. E. [1975] An elementary proof of the Pick-Nevanlinna interpolation theorem,
The Michigan Mathematical Journal, 21, 219223.
Mason, S. J. [1953] Feedback theory: Some properties of signal flow graphs, Proc. IRE, 41,
11441156.
[1956] Feedback theory: Further properties of signal flow graphs, Proc. IRE, 44, 920926.

Bibliography

647

Maxwell, J. C. [1868] On governors, Proceedings of the Royal Society. London. Series A.

Mathematical and Physical Sciences, 16, 270283.
Middleton, R. H. and Goodwin, G. C. [1990] Digital Control and Estimation. A Unified
Approach, Prentice-Hall, Englewood Cliffs, NJ, ISBN 0-132-11665-0.
Minnichelli, R. J., Anagnost, J. J., and Desoer, C. A. [1989] An elementary proof of
Kharitonovs stability theorem with extensions, Institute of Electrical and Electronics Engineers. Transactions on Automatic Control, 34(9), 995998.
Morris, K. A. [2000] Introduction to Feedback Control, Harcourt Brace & Company, New
York, ISBN 0-12-507660-6.
Nehari, Z. [1957] On bounded bilinear forms, Annals of Mathematics. Second Series, 65(1),
153162.

Nevanlinna, R. [1919] Uber

beschrankte Funktionen, die in gegebenen Punkten
vorgeschriebene Werte annehmen, Annales Academiae Scientiarum Fennicae. Series A I.
Mathematica Dissertationes, 13(1).

[1929] Uber
beschr
ankte analytisch Funktionen, Annales Academiae Scientiarum Fennicae.
Series A I. Mathematica Dissertationes, 32(7).
Nyquist, H. [1932] Regeneration theory, Bell Laboratories Technical Journal, 11, 126147.

648

Bibliography

Sontag, E. D. [1998] Mathematical Control Theory: Deterministic Finite Dimensional Systems, second edition, number 6 in Texts in Applied Mathematics, Springer-Verlag, New
York-Heidelberg-Berlin, ISBN 0-387-98489-5.
Springer, G. [1981] Introduction to Riemann Surfaces, second edition, Chelsea, New York,
ISBN 0-8284-0313-9.
Syrmos, V. L., Abdallah, C. T., Dorato, P., and Grigoriadis, K. M. [1997] Static output
feedbacka survey, Automatica. The Journal of IFAC. The International Federation of
Automatic Control, 33(2), 125137.
Truxal, J. G. [1955] Automatic Feedback Control System Synthesis, McGraw-Hill Electrical
and Electronic Engineering Series, McGraw-Hill, New York.
van der Woude, J. [1988] A note on pole placement by static output feedback for single-input
systems, Systems & Control Letters, 11(4), 285287.
Vidyasagar, M. [1981] Input-Output Analysis of Large-Scale Interconnected Systems, number 29 in Lecture Notes in Control and Information Sciences, Springer-Verlag, New YorkHeidelberg-Berlin, ISBN 0-387-10501-8.
[1986] On undershoot and nonminimum phase zeros, Institute of Electrical and Electronics
Engineers. Transactions on Automatic Control, 31(5), 440.

Oara, C. and Varga, A. [2000] Computation of general inner-outer and spectral factorizations,
Institute of Electrical and Electronics Engineers. Transactions on Automatic Control,
45(12), 23072325.

[1987] Control System Synthesis. A Factorization Approach, number 7 in MIT Press Series
in Signal Processing, Optimization, and Control, MIT Press, Cambridge, MA, ISBN 0262-22027-X.

Parks, P. C. [1962] A new proof of the Routh-Hurwitz stability criterion using the second
method of Liapunov, Proceedings of the Cambridge Philosophical Society, 58, 694702.

Pick, G. [1916] Uber

die Beschrankunger analytischer Funktionen, welche durch vorgegebebe
Funkionswerte bewirkt werden, Mathematische Annalen, 77, 723.

Wang, Q.-G., Lee, T.-H., and He, J.-B. [1999] Internal stability of interconnected systems,
Institute of Electrical and Electronics Engineers. Transactions on Automatic Control,
44(3), 593596.

Polderman, J. W. and Willems, J. C. [1998] Introduction to Mathematical Systems Theory, number 26 in Texts in Applied Mathematics, Springer-Verlag, New York-HeidelbergBerlin, ISBN 0-387-98266-3.
Reinschke, K. J. [1988] Multivariable Control: A Graph-Theoretic Approach, number 108
in Lecture Notes in Control and Information Sciences, Springer-Verlag, New YorkHeidelberg-Berlin, ISBN 3-540-18899-1.
Routh, E. J. [1877] A treatise on the stability of a given state of motion, Cambrige University,
Adams Prize Essay.
Schiff, J. L. [1999] The Laplace Transform, Undergraduate Texts in Mathematics, SpringerVerlag, New York-Heidelberg-Berlin, ISBN 0-387-98698-7.
Schwartz, L. [1966] Theorie des distributions, Publications de lInstitut de Mathematique de
lUniversite de Strasbourg, Hermann, Paris.
Seron, M. M., Braslavsky, J. H., and Goodwin, G. C. [1997] Fundamental Limitations in
Filtering and Control, Communications and Control Engineering Series, Springer-Verlag,
New York-Heidelberg-Berlin, ISBN 3-540-76126-8.
Smale, S. [1998] Mathematical problems for the next century, The Mathematical Intelligencer,
20(2), 715.

Willems, J. C. [1971] Analysis of Feedback Systems, MIT Press, Cambridge, MA, ISBN
0-26-223046-1.
Youla, D. C. [1961] On the spectral and inner-outer factorizations of rational matrices, IRE
Transactions on Information Theory, IT-7, 172189.
Youla, D. C., Jabr, H. A., and Bongiorno, J. J. [1976] Modern Wiener-Hopf design of optimal controllers. I. The single-input-output case, Institute of Electrical and Electronics
Engineers. Transactions on Automatic Control, AC-21(1), 313.
Zadeh, L. A. and Desoer, C. A. [1979] Linear System Theory, Robert E. Krieger Publishing Company, Huntington, New York, ISBN 0-88275-809-8, reprint of 1963 McGraw-Hill
original edition.
Zemanian, A. H. [1987] Distribution Theory and Transform Analysis. An Introduction to
Generalized Functions, with Applications, second edition, Dover Publications, Inc., New
York, original edition, McGraw-Hill, New York, 1965.
Zhou, K. [1996] Robust and Optimal Control, Prentice-Hall, Englewood Cliffs, NJ, ISBN
0-13-456567-3, with John C. Doyle and Keith Glover.
Ziegler, J. G. and Nicols, N. B. [1942] Optimum settings for automatic controllers, ASME
Transactions, 64(8), 759.

650

This version: 22/10/2004

Symbol Index
Symbol

: Description and page(s) when applicable

(i)

f 0 (z)
L
S\A
f g
z
f
kk

,
||
0n
0m,n
1(t)
1N,D
A
A(, )

:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:

adj(A)
b
ct
C(A, b)
c.f.r.
C
C+
C+
C
C
C[]
C()
D
D(z, r)
D(z, r)
(t)
det A
dim(V )

:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:

the ith derivative with respect to time, 597

derivative of f at z, 619
adjoint of the linear map L, 594
set theoretic complement of A in S, 618
convolution of f and g, 632
complex conjugate of z, 617
Fourier transform of f , 631
norm, 595
derivative with respect to time
empty set
equal by definition, i
modulus of a complex number, or absolute value of a real number, 617
n n matrix of zeros, 587
m n matrix of zeros, 587
the unit step function, 607
step response of SISO linear system (N, D) in input/output form, 108
state dynamics matrix for a SISO linear system, 24
(n 1) (n 1) matrix obtained by deleting the ith row and the jth
column of A Rnn , 588
adjugate of the matrix A, 588
input vector for a SISO linear system, 24
output vector for a SISO linear system, 24
controllability matrix for (A, b), 38
canonical fractional representative, 612
complex numbers, 617
complex numbers with positive real part, 610
complex numbers with nonnegative real part, 610
complex numbers with negative real part, 610
complex numbers with nonpositive real part, 610
polynomials with complex coefficients in indeterminant , 609
rational functions with complex coefficients in indeterminant , 611
feedforward term for a SISO linear system, 24
open disk of radius r centred at z, 618
closed disk of radius r centred at z, 618
the delta-function at 0, 629
determinant of the matrix A, 587
dimension of the vector space V , 586

Symbol Index

Df
e(t)
ei
F
F[s]nn
F
F 1
h+
N,D

:
:
:
:
:
:
:
:

h
N,D

:
:

hN,D

:
:

h
H
I
id
Im
image
In
iR
I
ker
lim sup
L (f )
+
L0+
(f )
+
L0
(f )

L0+
(f )

L0
(f )
L 1
n

:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:

(N, D)
N
O(A, c)
F(s)nn
Re
R
R+
R[]

:
:
:
:
:
:
:
:

the Jacobian matrix for a map f : Rn Rm , 12

the error for an observer, 430
ith standard basis vector for Rn , 586
general field, standing typically for either R or C, 585
set of n n matrices with components in F[s], 231
Fourier transform, 632
inverse Fourier transform, 632
causal impulse response for the SISO linear system (N, D) in input/output form, 94
causal impulse response for the SISO linear system , 49
anticausal impulse response for the SISO linear system (N, D) in input/output form, 94
anticausal impulse response for the SISO linear system , 53
causal impulse response for the SISO linear system (N, D) in input/output form, 94
causal impulse response for the SISO linear system , 49
frequency response for SISO linear system , 120
a generic interval, 26
identity map, 592
imaginary part of a complex number, 617
image of a linear map or matrix, 592
n n identity matrix, 587
imaginary axis, 610
set of intervals, 26
kernel of a linear map or matrix, 592
limit of the least upper bounds, 96
two-sided Laplace transform of f , 633
the right causal Laplace transform, 634
the left causal Laplace transform, 634
the right anticausal Laplace transform, 634
the left anticausal Laplace transform, 634
two-sided inverse Laplace transform, 633
state dimension of a generic SISO linear system = (A, b, ct , D), or the
degree of the denominator polynomial in a generic SISO linear system
(N, D) in input/output form, 24
generic SISO linear system in input/output form, 91
matrix whose eigenvalues are the spectrum for the zero dynamics, 46
observability matrix for (A, c), 34
set of n n matrices with components in F(s), 216
real part of a complex number, 617
real numbers, 585
positive real numbers, 585
polynomials with real coefficients in indeterminant , 609

Symbol Index
R()
RN,D
Rc (f )

:
:
:
:

max (f )
min (f )
N,D
Sn
span {S}
spec(A)
spec(P )
T (A, b)

:
:
:
:
:
:
:
:

At
TN,D

:
:

tr(A)
T
T
u
U

x
y
yh
yp
Z
Zk

:
:
:
:
:
:
:
:
:
:
:
:

651

rational functions with real coefficients in indeterminant , 611

ramp response of SISO linear system (N, D) in input/output form, 108
region of absolute convergence for f , 633
SISO linear system of the form (A, b, ct , D) (might on occasion be
MIMO), 24
maximum abscissa of absolute convergence of f , 633
minimum abscissa of absolute convergence of f , 633
canonical minimal realisation of input/output system (N, D), 93
set of permutations of (1, . . . , n), 588
smallest subspace containing S, 587
spectrum of A Rnn , 593
set of roots of a polynomial P , 610
largest subspace of Rn on which every curve can be followed by =
(A, b, c, D), 42
transpose of the matrix A, 587
transfer function for SISO linear system (N, D) in input/output form,
91
trace of the matrix A, 587
the vector space of test functions, 628
transfer function of a SISO linear system , 79
input to a SISO linear system, 24
set of admissible input, 26
set of imaginary axis poles for T , 119
state vector for a SISO linear system, 24
output of a SISO linear system, 24
homogeneous solution to a scalar ordinary differential equation, 597
particular solution to a scalar ordinary differential equation, 598
set of integers, i
sequence of subspaces defined for computing the spectrum of the zero
dynamics, 45
subspace used in computation of zero dynamics, 45

652

Symbol Index

654

This version: 22/10/2004

Index
abscissa of absolute convergence, 122, 363,
641
maximum, 633
minimum, 633
adjoint, 594
adjugate, 79, 588
admissible, 26, 49, 53, 93, 94, 116, 423, 425,
426, 443, 444, 526, 564
allowable, 147, 148
amplifier, 15
analogue computer, 93
analytic, 619
anticausal, 26, 53, 94, 98, 567, 574
appended system, 230
arc, 620
argument, 617
asymptote, 130, 150
asymptotic observer, 430
bandwidth, 347, 374, 375, 477, 483
and gain crossover frequency, 348, 477,
483
and performance, 321, 325
and rise time, 321, 325327
for first-order systems, 321
for general systems, 326, 347
for second-order systems, 325
basis, 586
bijective, 592
Blaschke product, 377
block diagram, 3, 76, 92, 93, 112, 116, 244,
279, 352, 353, 420
and signal flow graphs, 212, 271
for PID control, 261, 268
for state-space models, 79, 250, 255, 259,
357, 418
negative feedback, 64, 77, 213
parallel, 63, 76, 213
series, 63, 76, 213
unity gain, 247, 281, 333, 342, 344, 361,
368, 382
Bode integral formulae, 369, 375, 382
and performance, 372
Bode plot, 124, 124, 132, 135, 138, 152,
154156, 276, 308, 348, 352, 353,
389392, 554

and minimum phase, 137, 156

and Nyquist plot, 135, 285, 293, 294,
475, 490
and the H -norm, 146
by hand, 129
for closed-loop transfer function, 374
for first-order systems, 321
for lead/lag compensation, 470
for PID compensator, 473
for second-order systems, 153, 324
for sensitivity function, 373375
magnitude, 124, 132, 133, 135, 146, 148,
154, 156, 308, 311, 321, 348, 352,
353, 373375, 377, 383, 390, 391,
480, 554
phase, 124, 135, 137, 144, 154, 294, 306,
325, 353, 470, 491
Bodes Gain/Phase Theorem, 138, 143, 157,
279, 361
boundary point, 618
bounded, 618
branch, 212
tree, 10
break frequency, 130, 321, 472
c.f.r., see canonical fractional representative
cancellation, 81, 8387, 93, 120, 136, 189,
241243, 246, 250, 276, 284, 288,
290, 398, 405, 560
canonical form, 54, 59
controllability, 70
controllable, 407
controller, 55, 56, 59, 70, 93, 253, 358,
398, 406, 526, 527, 541, 542, 545,
552, 573
first Luenberger-Brunovsky, 70, 71
Jordan, 73, 274
observability, 71
observer, 57, 58, 59, 93, 116, 256, 400,
404, 411
second Luenberger-Brunovsky, 56, 58
canonical fractional representative, 612
canonical minimal realisation, 93, 108, 115,
248, 259, 272, 274, 418, 428, 575
capacitor, 10
Cauchys Integral Theorem, 621

Cauchy-Riemann equations, 619

causal, 26, 48, 94, 95, 153, 635
Cayley-Hamilton Theorem, 36, 37, 40, 41,
56, 61, 593
change of basis
matrix, 593
change of coordinates, 54, 5759, 67, 81, 407,
542, 545, 592
for linear maps, 593
for vectors, 593
characteristic equation, 597
characteristic polynomial, 55, 6365, 78, 82,
137, 249, 252, 257, 271, 272, 274,
526, 593
for interconnected systems, 237, 240242
Chipart, 199
closed circuit, 114
closed set, 618
closed-loop, 6, 211, 244, 251, 253, 255, 256,
258, 259, 279, 333, 397, 417, 425
poles, 246, 252, 253, 268, 273, 275, 276,
281, 364, 409, 497, 499
response, 247, 258, 261, 364, 367, 393,
469, 480, 496, 499, 523, 544
cofactor
of a matrix, 588
of a signal flow graph, 219, 222, 223
columnspace, 592
complete, 41, 43, 92, 115, 120, 275, 409, 573,
574, 576
complex conjugate, 617
complex numbers, 585
components
of a linear map, 592
of a vector, 587
connected, 218, 618
continuous, 618, 628
contour, 140, 279, 281, 286, 313, 369, 621
controllability, 38, 38, 41, 42, 44, 45, 48, 55,
57, 5961, 6668, 70, 73, 92, 93, 97,
114116, 156, 164, 168, 252, 260,
273, 358, 398400, 403, 404, 406,
408, 410, 415, 417, 441, 442, 542,
545, 547, 552, 555, 576
and coprime polynomials, 81, 8387
for interconnected systems, 243
for MIMO systems, 43
Gramian, 576
matrix, 38, 41
operator, 575
controllability canonical form, 70

Index
controlled output, 329
controller, 1, 6, 7, 146, 150, 244, 246, 249,
250, 273, 304, 307, 315, 333, 338,
342, 362, 374, 382, 384, 386, 389,
391, 392, 394, 397, 416418, 420,
443, 469, 475, 477, 483, 486, 488,
557, 558, 561, 581
optimal, 525, 543, 544, 550
proper, 260, 354, 425, 426, 444, 557, 560,
562, 570
stabilising, 405, 418, 420, 422, 423, 426,
427, 443, 444, 487, 562, 564
strongly stabilising, 429
controller canonical form, 55, 56, 59, 70, 93,
253, 358, 398, 406, 526, 527, 541,
542, 545, 552, 573
convergence
of test functions, 628
convolution, 27, 49, 95, 102, 103, 107, 112,
152, 165, 177, 632, 641
for Laplace transforms, 636
convolution for Fourier transforms, 632
coprime, 81, 8387, 91, 93, 114, 123, 136,
250, 276, 334, 403, 535, 538, 610,
611, 612, 612, 616
factorisation, 422, 423425, 444, 521,
560562, 581
fractional representative, 421, 423427,
444, 521, 561, 562, 581
rational functions, 421, 421, 422, 423
coupled masses, 15, 67, 115, 155, 205, 207,
275
coupled tanks, 18
Cramers Rule, 223, 588
curve, 620
closed, 620
differentiable, 620
Jordan, 620
length, 621
simple, 620
cutoff frequency
lower, 347
upper, 347
dB/decade, 130
DC servo motor, 2, 11, 15, 211, 275, 276
decibel, 124
degree, 609
delta-function, 629
derivative, 619
of a distribution, 629

Index
derivative control, 263, 264, 266, 276, 339,
443
derivative time, 261, 263, 474, 499
design problem
for dynamic output feedback, 260
for input/output systems, 246
for static output feedback, 256
for static state feedback, 251
detectable, 429
determinant
of a matrix, 193, 587
of a signal flow graph, 219, 220, 223,
242, 281, 284
diagonalisable, 73, 594, 600
differentiable, 619
dimension, 586
directed path, 218
discrete, 26
disk
closed, 618
open, 617
distribution, 629
simple, 633
disturbance, 1, 3, 6, 214, 248, 336338, 340,
354, 364, 477, 478, 480, 482, 485, 490
disturbance type, 336339, 341, 480, 481
disturbed output, 336
domain, 618
double pendulum, 16, 65, 68, 69, 115, 155,
205, 206, 442, 554
dynamic output feedback, 258, 260, 416, 420,
426
eigenspace, 82, 122, 399, 401, 593
eigenvalue, 47, 73, 85, 89, 120122, 152, 154,
162, 164, 166, 168, 181, 183, 205,
315, 399, 401, 418, 545, 551, 593, 600
algebraic multiplicity, 162, 419, 593, 600,
602
closed-loop, 252, 417, 442, 544, 554, 555
generalised, 415
geometric multiplicity, 162, 593, 600
eigenvector, 42, 82, 162, 593
generalised, 600
electric circuit, 10, 17, 67, 114, 115
for lag compensation, 489
for lead compensation, 489
for notch filter, 489
equations of motion for an interconnected
SISO linear system, 231

655
equilibrium point, 8, 12, 12, 16, 17, 31, 65,
66, 180
error, 24, 6, 214, 244, 248, 261, 329, 331,
333, 336, 340, 344, 345, 353, 358,
362, 363, 368, 382384, 430, 431, 494
essentially bounded, 26
even, 588
feedback, 2, 6, 113, 164, 211, 244, 273, 362
dynamic output, 258, 260, 416, 420, 426
for state-space models, 250, 255, 258,
272274, 357, 398, 405, 417, 441,
442, 527, 542, 546, 554
negative, 15, 64, 77, 213, 214, 219, 230,
241, 243, 244
static output, 255, 255, 256, 275, 357,
408, 416, 447
static state, 250, 251, 251, 273, 274, 315,
357, 398, 402, 406, 416, 440442,
527, 542, 546, 550
unity gain, 78, 247, 283, 299, 304, 313,
333, 334, 342, 344, 357, 361, 362,
367369, 382, 418, 443, 476, 500
feedforward, 79, 96
final value problem, 53
Final Value Theorem, 637
first Luenberger-Brunovsky canonical form,
70, 71
Fourier transform, 631
L1 , 631
L2 , 632
inverse, 632
frequency response, 119, 135, 138, 148, 207,
279, 383, 469
and impulse response, 122
and periodic solutions, 120, 275
and the transfer function, 120, 122
for first-order systems, 321
for input/output systems, 122
for lead/lag compensator, 470, 477
for PID compensator, 473, 483
for second-order systems, 153, 324
for SISO linear systems, 120
graphical representations, 124, 130, 135,
138
rearranging for Bode plots, 129
gain, 4, 15, 275, 344, 447, 474, 476, 495
L2 L2 , 567
Lp1 Lp2 , 174
in a signal flow graph, 212, 218, 272
gain crossover frequency, 297, 297, 477, 483

656
and bandwidth, 348, 477, 483
gain margin, 299, 300, 420, 469, 478, 487,
550, 555
limitation of, 300
lower, 297, 297
upper, 297, 297
general solution, 599
graph transmittance, 224, 230, 236
Hardy
function, 173
space, 172, 175
Hermite, 195
criterion, 195, 197
reduced, 196, 197
matrix, 195
Hermite matrix, 196
Hermitian, 594
Hermitian inner product, 594
homogeneous solution, 104, 109, 122, 152,
597
Horowitz area formula, 373, 374
Hurwitz, 163, 186, 189, 190, 193, 202, 207,
208, 261, 274, 316, 398, 400, 408,
413, 415417, 428, 440, 447, 449,
527, 533, 545, 554, 555, 577
criterion, 193, 199, 203, 208, 272, 290,
293, 313, 322, 476
matrix, 193, 415
identity matrix, 587
image, 592
imaginary part, 617
improper, 91, 169, 233, 311, 444, 560, 570,
612
impulse response, 48, 49, 53, 69, 94, 103,
108110, 116, 165
and convolution, 49, 53, 102, 103, 107,
112, 152
and frequency response, 122
and the transfer function, 95, 96, 122
anticausal, 53, 574
as a solution to an initial value problem,
108
as limit of realisable outputs, 51, 207
causal, 49
for input/output systems, 94
inductor, 10
infinite-dimensional, 586
initial value problem, 27, 99, 102, 105, 108,
116, 152, 527, 599
injective, 592

Index
inner product, 594
innovation, 431
input, 1, 2, 24, 2528, 33, 3537, 39, 44, 49,
61, 79, 90, 105, 107, 152, 153, 226,
230, 242, 248, 259, 260, 345, 368,
383, 526, 548, 567
bounded, 165, 167, 174, 180, 205, 206,
258
for a signal flow graph, 213
integrable, 536, 536, 538
L1 , 631
L2 , 631
L , 170, 631
Lp , 170
integral control, 264, 267, 288, 340, 478
interconnected SISO linear system, 229, 230,
235, 236, 242
interconnection, 212, 212, 229, 230, 233, 235,
236, 242, 248, 259, 271, 272, 279, 281
negative feedback, 213
parallel, 213
series, 213
interval, 25
invariant subspace, 36, 41, 45, 59, 87, 183,
592, 596
invertible, 588, 592
Jacobian, 12
kernel, 592
Kharitonovs test, 201, 209
Kirchhoffs laws, 10, 11
Laplace transform, 76, 78, 79, 92, 95, 98, 99,
101103, 106, 107, 110, 123, 152,
153, 260, 320, 573, 577, 578, 632,
633, 641
and differential equations, 98, 99, 101,
102, 108
and the matrix exponential, 106, 641
anticausal
left, 634
right, 634
causal, 78, 92, 95, 98, 99, 101103, 106,
107
left, 78, 92, 95, 98, 99, 101, 634
right, 98, 99, 101103, 106, 107, 634
inverse, 99, 100, 102, 108, 110, 179, 567,
573, 574, 577, 633, 641
two-sided, 633
Laurent series, 620
least common multiple, 609

Index
length, 218
Liapunov, 180, 545
function, 180, 181, 182, 195, 203
proper, 181
triple, 182, 186, 207, 576
Lienard, 199
Lienard-Chipart criterion, 200, 209
linear
map, 591
transformation, 591
linearisation, 8, 12, 12, 16, 17, 31, 31, 65, 66
linearly independent, 586, 600, 602, 604
link, 10
loop, 218, 219
loop gain, 218, 281, 293, 320, 333, 336, 342,
343, 362, 382, 384, 444, 475, 478,
480, 481, 484, 491, 551, 555
bandwidth constraints, 374, 375
poles of, 273, 362, 364, 365, 367, 369, 378
proper, 233, 248, 283285, 313, 334, 369,
371
stable, 299, 300, 369
strictly proper, 249, 284, 285, 354
unstable, 362, 364, 365, 367369, 382
zeros of, 273, 362, 364, 365, 367, 369, 378
Masons Rule, 223, 230
mass-spring-damper system, 7, 24, 28, 49,
78, 79, 124, 153, 260, 264, 307, 310,
338, 343, 478, 483
matrix exponential, 27, 36, 52, 95, 98, 106,
107, 162, 183, 600, 641
Maximum Modulus Principle, 377, 382, 568,
623
measurable, 26
meromorphic, 172, 280, 566, 619
method of undetermined coefficients, 104,
598
MIMO, 242
MIMO systems, 24, 37, 43, 54, 75, 182, 256,
389, 530, 541, 548, 557
minimum phase, 47, 90, 91, 93, 129, 135,
137, 140, 143, 156, 326, 336, 362,
364, 365, 367369, 377, 381, 383,
393, 488, 490, 491, 531
minor, 588
principal, 182, 588
modulus, 617
monic, 609
multiplicity, 179, 280, 313, 531, 533, 534,
566, 597, 610

657
algebraic, 162, 237, 419, 593, 600, 602
geometric, 162, 237, 593, 600
negative-definite, 181, 207, 440
negative-semidefinite, 181, 207, 440
neppers, 140
node, 212
noise, 236, 248, 478
nominal performance, 385, 385
nonlinear, 12, 16, 17, 31, 65, 66, 180
nonminimum phase, 47, 90, 91, 93, 129, 135,
137, 140, 143, 156, 326, 336, 362,
364, 365, 367369, 377, 380, 381,
383, 393, 488, 490, 491, 531
nonsingular, 588
nontouching, 218
norm, 170, 595
2 2, 179
H , 345
H2 , 383, 531
H , 146, 172, 175, 179, 302, 376,
381384, 389, 394, 531, 566
Hp , 172, 173
L1 , 345
L2 , 345
L , 345
L1 , 171
L2 , 171, 180, 207, 383, 384
L , 170, 171, 207, 383, 494
Lp , 170
Lpow , 171, 174, 207
pow pow, 179
p1 p2 , 174, 175
and stability, 174, 175, 179
for signal in time-domain, 170
for transfer function, 172
Hankel, 578, 578
sup, 170
Nyquist contour, 283, 283, 284, 285, 302,
313, 314, 385, 476, 490
and the Bode plot, 135, 285, 293, 294,
475, 490
and the polar plot, 293
determining the number of
encirclements, 285
for state-space models, 552, 555
imaginary axis poles, 313, 314
plotting of, 285
Nyquist criterion, 279, 281, 284, 299, 304,
313, 314, 416, 420, 444, 476, 488,
490, 551

658

Index
for state-space models, 315, 551

observability, 33, 34, 36, 41, 44, 45, 48, 57,

5961, 6668, 70, 73, 93, 97, 114,
115, 136, 137, 155, 156, 167, 168,
182, 186, 207, 246, 260, 273, 398,
400, 401, 403, 404, 408, 417, 429,
430, 432, 442, 554
and coprime polynomials, 81, 83, 8587,
93
for interconnected systems, 243
for MIMO systems, 37, 182
Gramian, 576
matrix, 34, 36
operator, 575
observability canonical form, 71
observer, 429
asymptotic, 430
gain vector, 431
observer canonical form, 5759, 93, 116, 256,
400, 404, 411
observer gain vector, 431
odd, 588
open circuit, 114
open set, 618
open-loop, 1, 3, 4, 6, 15, 211, 244, 246, 247,
253, 254, 257, 258, 275277, 476
optimal state feedback, 546, 554
stability margins, 550
optimal state feedback vector, 527
order
of a distribution, 629
order of a pole, 619
orientation, 620
originates, 212
output, 1, 24, 33, 36, 37, 79, 90, 122, 152,
180, 226, 230, 242, 244, 248, 255,
258260, 276, 336, 526, 527, 543, 553
bounded, 165, 258
computation of, 98, 105107
for a signal flow graph, 213
periodic, 120, 122, 152
unbounded, 165, 167, 205, 206, 258
output compatible, 256, 256
output feedback constant, 256, 400, 409, 414,
415
overshoot, 318, 328, 336, 361, 481, 499
and peaks in Bode plot, 322, 327, 328
and poles at 0, 364
and rise time, 365, 367
and unstable poles, 365, 367

and zeros, 326

for second-order systems, 322
Parsevals Theorem, 632
partial fraction expansion, 97, 100, 111, 531,
571, 573, 575, 613, 616
particular solution, 122, 329, 598
path, 218
PD control, 443
pendulum, 8, 12, 32
pendulum on cart, 16, 65, 68, 69, 115, 155,
205, 206, 393, 441, 442, 554
performance, 317, 361
and bandwidth, 321, 325
and nonminimum phase zeros, 362, 364,
365, 367, 369, 373, 375, 378
and the sensitivity function, 342, 344,
345, 368, 369, 373375, 377, 378,
381384
and the transfer function, 320, 321, 361,
364, 369, 374, 377, 378, 382
and unstable poles, 362, 364, 365, 367,
369, 373, 378
for first-order systems, 320
for second-order systems, 321
for state-space models, 357
frequency-domain criteria, 320, 368, 369,
373375, 378, 382, 478, 483
time-domain criteria, 318, 335, 336, 361,
362, 364, 365, 367
with additional poles, 327
with zeros, 326
periodic solution, 120, 122, 152, 183, 275
phase crossover frequency, 295, 297
phase margin, 297, 299, 300, 315, 420, 477,
478, 483, 484, 487, 490, 550, 555
limitation of, 300
lower, 297, 297
upper, 297, 297
PI control, 290, 494, 496, 523
PID control, 91, 261, 264, 274, 277, 307, 310,
328, 334, 341, 344, 473, 483, 490, 493
Ziegler-Nicols tuning, 345, 493, 496, 523
piecewise continuous, 26, 49, 94
plant, 1, 2, 7, 146, 149, 157, 244, 246, 249,
250, 258261, 273, 303, 315, 333,
335, 342, 361, 362, 374, 385, 416,
418, 420, 425, 427, 429, 469, 478,
494, 523
minimum phase, 373, 487
nominal, 146149, 303, 304, 307, 309,

Index
386, 389, 394, 558, 560, 562
nonminimum phase, 272, 273, 354, 362,
364, 368, 377, 378, 380, 381, 490, 491
stable, 443, 487, 493
unstable, 272, 273, 336, 362, 364, 368,
378, 380, 381, 488, 490, 491
Poisson Integral Formula, 622
Poisson integral formulae, 377, 393
polar plot, 135, 135, 156
and Nyquist plot, 285, 293
pole, 116, 120, 129, 140, 149, 168, 253,
273276, 279281, 327, 343, 368,
377, 398, 405, 409, 412, 414, 443,
612, 616, 619
for first-order systems, 320
for second-order systems, 323
imaginary axis, 120, 147, 148, 173,
283285, 304, 305, 313, 314, 339,
362, 369, 495, 531, 560, 571
placement, 257, 406, 407
stable, 122, 123, 157, 189, 246, 531, 535,
559, 566, 569, 571
unstable, 147, 148, 169, 173, 242, 284,
303, 336, 340, 362, 364, 365,
367369, 373, 374, 378, 381, 382,
384, 385, 387, 393, 487, 491, 560, 571
poles
closed-loop, 328
Popov-Belevitch-Hautus, 67, 398, 400, 439
positive-definite, 39, 181, 182, 186, 207, 545,
547, 567, 570, 594
positive-semidefinite, 39, 181, 182, 186, 207,
413, 439, 577
power, 180
average, 171
signal, 171, 177
spectrum, 384
Principle of the Argument, 279, 280, 284,
313, 314, 455, 488
proper, 91, 92, 94, 99, 102, 108, 109, 112,
140, 143, 147, 148, 165, 168,
173175, 179, 230, 233, 272, 273,
277, 283285, 295, 304, 307,
309311, 313, 334, 347, 369, 371,
382, 385, 387, 389, 420, 422427,
443, 444, 491, 531, 557560, 562,
566, 571, 612
proportional control, 263, 264, 266, 275, 276,
285, 339, 443, 476, 490, 495
proportional gain, 261, 523

659
ramp response, 108, 109, 331
as a solution to an initial value problem,
108
Rank-Nullity Theorem, 592
rational function, 76, 80, 112, 146, 147, 173,
212, 421, 422, 531, 535, 560, 571,
573, 579, 611, 616
real numbers, 585
real part, 617
reference trajectory, 1, 35, 244, 248, 250,
255, 318, 328, 331, 338, 364, 383
region, 618
region of absolute convergence, 633
relative degree, 91, 132, 373, 375, 382, 421,
422, 484
reset time, 261, 264
residue, 622
Residue Theorem, 622
resistor, 10
rise time, 318, 361
alternate definition, 318
and additional poles, 327
and bandwidth, 321, 325327, 477
and overshoot, 365, 367
for first-order systems, 320
for second-order systems, 322
robust performance, 381, 385, 386, 386, 389,
394, 580
for multiplicative uncertainty, 387
graphical interpretation, 389
modified problem, 559, 560, 581
problem, 389, 389, 530, 558, 560, 564
robust stability, 303, 304, 316, 382, 385, 386,
389
for additive uncertainty, 304, 309, 310,
387
graphical interpretation, 309
for multiplicative uncertainty, 304, 307
graphical interpretation, 305
root, 189, 610
Routh, 189
array, 190
criterion, 190, 192, 199, 203, 207, 272,
290, 293, 313, 322, 476
saturation, 349
second Luenberger-Brunovsky canonical
form, 56, 58
sensitivity function, 248, 273, 309, 310, 342,
343, 345, 368, 384, 385, 394, 488

660
and performance, 342, 344, 345, 368,
369, 373375, 377, 378, 381384
and stability margin, 302, 303
complementary, 248, 304
minimisation, 303, 317, 345, 375, 381,
382
sensor, 2, 244
settling time, 318, 482
and undershoot, 365, 367
for second-order systems, 323
signal, 212
signal flow graph, 212, 212, 219, 223, 229,
271
and block diagrams, 212, 271
negative feedback, 213
parallel, 213
series, 213
signal-to-noise ratio, 275
simple distribution, 633
singular, 588
singular value, 594
singularity
essential, 619
isolated, 619
removable, 619
sink, 212
SISO linear system, 24
SISO linear system in input/output form, 91
SISO nonlinear system, 31, 65, 66
skew-Hermitian, 595
skew-symmetric, 595
skew-symmetric part, 595
source, 212
span, 587
spans, 586
spectrum, 115, 155, 156, 162, 163, 166, 181,
417, 571, 574, 593
for zero dynamics, 88
stability, 180, 181
L2 L2 , 179
Lpow Lpow , 179
Lp1 Lp2 , 174
asymptotic, 180, 181
BIBO, 165, 165, 166, 168, 169, 173, 179,
180, 205207, 272
for interconnected systems, 236
closed-loop, 254, 258, 260, 273, 274, 279,
284, 299, 313, 315, 362, 393, 398,
400, 405, 408, 415, 416, 418, 419,
440, 443, 469, 476, 490, 551
IBIBO, 235237, 242, 248, 271273, 281,

Index
284, 299, 304, 313, 314, 336, 337,
353, 362, 393, 418, 420, 425, 443,
444, 447, 476
internal, 162, 162, 163, 165, 182, 205207
for interconnected systems, 236, 238
internal asymptotic, 162, 162, 163, 168,
182, 205207, 260, 274, 398, 408,
416, 417
for interconnected systems, 236, 238
Liapunov methods, 180182, 184, 186,
275, 413, 440
relationship between internal and BIBO,
168, 405
spectral, 164, 205, 206
transfer function, 281
stable
controller, 429
loop gain, 299, 300, 369
plant, 443, 487, 493
transfer function, 91, 122, 129, 140,
143145, 156, 157, 175, 179, 318,
319, 329, 393, 420, 493, 494, 531
zero dynamics, 47
standard basis, 586
standard Hermitian inner product, 594
standard inner product, 594
state, 24, 27, 28, 33, 3537, 40, 43, 49, 54,
57, 61, 6365, 67, 81, 84, 98, 105,
153, 167, 205, 206, 250, 253, 259,
418, 526, 527, 542, 545, 548, 554
controller, 259
lack of in input/output systems, 90, 92,
93, 165, 167, 261
periodic, 120, 152
plant, 259
solution, 27
state compatible, 251, 252
state feedback vector, 251, 358, 542
static output feedback, 255, 255, 256, 275,
357, 408, 416, 447
static state feedback, 250, 251, 251, 273, 274,
315, 357, 398, 402, 406, 412, 416,
440442, 527, 542, 546, 550
steady-state error, 263, 264, 329, 331, 333,
335, 338, 341, 344, 353, 358, 359,
374, 478, 480
steady-state response, 122, 318
step function, 49, 53, 108, 607
step input, 318, 320, 321, 331, 478
step response, 108, 109, 110, 318, 319, 331,
344, 354, 358, 359, 364, 393, 481,

Index
490, 523
as a solution to an initial value problem,
108
first-order, 320
normalised, 318, 335, 336, 362
second-order, 322
with additional poles, 327
with zeros, 326
steppable, 318, 318, 347
strictly minimum phase, 140, 143
strictly proper, 91, 94, 103, 109, 112, 113,
122, 123, 175, 230, 274, 284, 285,
311, 335, 382, 393, 421423, 425,
443, 531, 565, 571, 612
structure matrix, 216, 271
subgraph, 218
subspace, 586
summer, 77
support, 627
surjective, 592
symmetric, 594
symmetric part, 595
Taylor series, 619
terminates, 212
test function, 627
time delay, 147, 641
touching, 219
trace, 587
transfer function, 75, 90, 92, 105, 112,
114116, 154, 156, 230, 273, 352,
358, 393, 409, 442, 489, 490, 545
and changes of coordinate, 81
and controllability, 81, 83, 87, 398
and detectability, 403, 405
and frequency response, 120, 122
and impulse response, 95, 96, 112, 165
and observability, 83, 85, 87, 398
and performance, 361, 364, 369, 374,
377, 378, 382
and stabilisability, 403405
and zero dynamics, 87, 90
closed-loop, 245, 246, 247, 251, 256, 260,
272, 273, 281, 304, 307, 328,
334336, 343, 348, 361, 364, 367,
368, 373, 374, 377, 378, 385, 394,
406, 409, 414, 443, 488, 495
first-order, 320
for input/output systems, 90, 91
for interconnected systems, 223, 235,
240, 243

661
for negative feedback interconnection,
113
for parallel interconnection, 113
for series interconnection, 113
for SISO linear systems, 78, 79, 80
improper, 233, 284
invertibility, 88
norms for, 172
open-loop, 244, 247, 279
proper, 102, 113, 284
second-order, 153, 321
stable, 91, 122, 129, 140, 143145, 156,
157, 175, 179, 318, 319, 329, 393,
420, 493, 494, 531
strictly proper, 102, 113
unstable, 393
transient response, 122, 318, 477
transposition, 588
tree, 10
triangle inequality, 595
type, 329, 331
2 or greater, 335, 336, 374
0, 331, 353
1, 331, 335, 345, 353, 374
and poles at 0, 334
for state-space models, 358, 359
for unity gain feedback loops, 333, 334
of a system, 329
of an input, 329
uncertainty, 2, 4, 145, 149, 150, 303, 304,
307, 316, 385, 386, 389, 391, 394,
406, 550
additive, 148, 149, 303, 308, 310, 385,
386, 389, 391, 558, 560, 562, 581
multiplicative, 147, 147, 149, 303305,
307, 385, 386, 389, 558, 560562
structured, 146
unstructured, 146, 147
undershoot, 318
and nonminimum phase zeros, 326, 328,
365, 367
and poles at 0, 364
and settling time, 365, 367
immediate, 393
unstable, 43, 129, 162, 182, 205, 206, 254,
284, 417, 476, 551, 557
loop gain, 362, 364, 365, 367369, 382
plant, 336, 362, 364, 368, 381, 488, 490,
491
transfer function, 393

662

Index
zero dynamics, 90

vector space, 585

waterbed effect, 375
well-posed, 233, 233, 248, 255, 259, 272, 284,
369, 376, 426, 443
zero, 116, 129, 140, 271, 275, 280, 284, 313,
326, 336, 343, 368, 374, 377, 398,
405, 421, 531, 612
dynamics, 43, 46, 47, 68, 69, 87, 88, 90,
114, 115, 155, 156

imaginary axis, 533, 559, 560, 565

initial condition, 48, 49, 165, 236
input, 35, 122, 205
minimum phase, 364, 535, 542
nonminimum phase, 242, 284, 326, 362,
364, 365, 367369, 374, 376, 378,
381, 382, 384, 385, 387, 393, 422,
425, 487, 491
stable, 242
Ziegler-Nicols tuning, 345, 493, 496, 523

Index

663

Full download Practical Methods for Optimal Control and Estimation Using Nonlinear Programming Second Edition Advances in Design and Control John T. Betts pdf docx
100% (16)
Full download Practical Methods for Optimal Control and Estimation Using Nonlinear Programming Second Edition Advances in Design and Control John T. Betts pdf docx
60 pages
HyBase6100 Service Manual V1.0 en
78% (23)
HyBase6100 Service Manual V1.0 en
88 pages
Linear Optimal Control Systems
No ratings yet
Linear Optimal Control Systems
22 pages
EE5103/ME5403: Computer Control Systems: Associate Professor
No ratings yet
EE5103/ME5403: Computer Control Systems: Associate Professor
64 pages
WinPlot Major Features Summary
No ratings yet
WinPlot Major Features Summary
29 pages
DPOCexam2008midterm Solution
No ratings yet
DPOCexam2008midterm Solution
12 pages
Seminar Report On Amoled Technology
88% (8)
Seminar Report On Amoled Technology
22 pages
Chem 1220 Assignment 4
No ratings yet
Chem 1220 Assignment 4
5 pages
A Mathematical Approach To Control System PDF
No ratings yet
A Mathematical Approach To Control System PDF
666 pages
Dynamics of Linear Systems: September 27, 2017
No ratings yet
Dynamics of Linear Systems: September 27, 2017
101 pages
Review - 3 - Load Forecasting PDF
No ratings yet
Review - 3 - Load Forecasting PDF
25 pages
Notes On Linearisation (H.K.Khalil)
No ratings yet
Notes On Linearisation (H.K.Khalil)
11 pages
Lab 4 QUBE-Servo Bump Test Modeling Workbook (Student)
No ratings yet
Lab 4 QUBE-Servo Bump Test Modeling Workbook (Student)
6 pages
Razumikhin and Krasovskii Stability Theorems For Time-Varying Time-Delay Systems
No ratings yet
Razumikhin and Krasovskii Stability Theorems For Time-Varying Time-Delay Systems
11 pages
Design of A Linear State Feedback Controller
No ratings yet
Design of A Linear State Feedback Controller
27 pages
Nonlinear and Adaptive Control: Zhengtao Ding
No ratings yet
Nonlinear and Adaptive Control: Zhengtao Ding
116 pages
Lec Note - Robust Control - 1501
No ratings yet
Lec Note - Robust Control - 1501
83 pages
Balanced Truncation
No ratings yet
Balanced Truncation
15 pages
Material: Glad & Ljung Ch. 12.2 Khalil Ch. 4.1-4.3 Lecture Notes
No ratings yet
Material: Glad & Ljung Ch. 12.2 Khalil Ch. 4.1-4.3 Lecture Notes
42 pages
LTI
No ratings yet
LTI
67 pages
Advance Control System
100% (1)
Advance Control System
3 pages
LQG/LQR Controller Design: Undergraduate Lecture Notes On
No ratings yet
LQG/LQR Controller Design: Undergraduate Lecture Notes On
37 pages
Robust Control
No ratings yet
Robust Control
2 pages
Switched Systems
100% (2)
Switched Systems
185 pages
Robust Control Notes
No ratings yet
Robust Control Notes
415 pages
Session1 SC16B047
No ratings yet
Session1 SC16B047
7 pages
Design Modelingand Controlofa Reaction Wheel Balanced Inverted Pendulum
No ratings yet
Design Modelingand Controlofa Reaction Wheel Balanced Inverted Pendulum
10 pages
Nonlinear Tracking Differentiator For Velocity Estimation From Shaft Encoder
No ratings yet
Nonlinear Tracking Differentiator For Velocity Estimation From Shaft Encoder
12 pages
Dynamic Programming Value Iteration
100% (1)
Dynamic Programming Value Iteration
36 pages
Use of Kalman Filters in Time and Frequency Analysis
No ratings yet
Use of Kalman Filters in Time and Frequency Analysis
78 pages
Ece380 Notes PDF
0% (1)
Ece380 Notes PDF
171 pages
AN APPLICATION OF OPTiMAL CONTROL THEORY TO AN ANTI - TANK WEAPON SYSTEM
No ratings yet
AN APPLICATION OF OPTiMAL CONTROL THEORY TO AN ANTI - TANK WEAPON SYSTEM
174 pages
Extended State Observer Based Controller
50% (2)
Extended State Observer Based Controller
5 pages
Problem Statement: Using MATLAB For STK Automation
No ratings yet
Problem Statement: Using MATLAB For STK Automation
10 pages
EACT633 - Optimal Control and It's Applications Worksheet For Chapter 4, 5 & 6
No ratings yet
EACT633 - Optimal Control and It's Applications Worksheet For Chapter 4, 5 & 6
3 pages
LQR Feedforward
No ratings yet
LQR Feedforward
20 pages
PID Controller: Control Engineering and Signal Processing Michala Srnová 2009/2010
No ratings yet
PID Controller: Control Engineering and Signal Processing Michala Srnová 2009/2010
12 pages
Stability: EE-601 Linear System Theory
No ratings yet
Stability: EE-601 Linear System Theory
26 pages
Control and Instrumentation Lab (AV 435)
No ratings yet
Control and Instrumentation Lab (AV 435)
3 pages
Algorithm For Fixed-Range Optimal Trajectories
No ratings yet
Algorithm For Fixed-Range Optimal Trajectories
86 pages
CH 4
No ratings yet
CH 4
113 pages
Nonlinear Control of The Reaction Wheel Pendulum
No ratings yet
Nonlinear Control of The Reaction Wheel Pendulum
7 pages
EE5103 Part1 Lecture 1
No ratings yet
EE5103 Part1 Lecture 1
66 pages
M.gopal Control Systems.
No ratings yet
M.gopal Control Systems.
206 pages
Robust Pole Placement Using Linear Quadratic Regulator Weight Selection Algorithm
No ratings yet
Robust Pole Placement Using Linear Quadratic Regulator Weight Selection Algorithm
5 pages
EE5103/ME5403 Lecture Three Analysis of Discrete-Time Systems
No ratings yet
EE5103/ME5403 Lecture Three Analysis of Discrete-Time Systems
49 pages
Practical Guide Optimal Control Theory PDF
No ratings yet
Practical Guide Optimal Control Theory PDF
39 pages
Adaptive UnitI
No ratings yet
Adaptive UnitI
38 pages
State Space A To NPGMV Control
No ratings yet
State Space A To NPGMV Control
43 pages
EE5103/ME5403 Lecture Two Computer Control and Discrete-Time Systems
No ratings yet
EE5103/ME5403 Lecture Two Computer Control and Discrete-Time Systems
49 pages
ECE1647 NonlinearControl
100% (1)
ECE1647 NonlinearControl
138 pages
Stability of Nonlinear Systems
No ratings yet
Stability of Nonlinear Systems
18 pages
Lyapunov Tutorial
100% (2)
Lyapunov Tutorial
104 pages
Sliding Mode Control Tutorial PDF
No ratings yet
Sliding Mode Control Tutorial PDF
7 pages
Complete Quadratic Lyapunov Functionals Using Bessel LegendreInequality
No ratings yet
Complete Quadratic Lyapunov Functionals Using Bessel LegendreInequality
6 pages
The Kalman Filter: State-Space Derivation For Mass-Spring-Damper System
No ratings yet
The Kalman Filter: State-Space Derivation For Mass-Spring-Damper System
10 pages
Control Book
No ratings yet
Control Book
422 pages
EE5103/ME5403 Lecture Four Pole-Placement Problem State-Space Approach
No ratings yet
EE5103/ME5403 Lecture Four Pole-Placement Problem State-Space Approach
46 pages
23 DeepLearning PDF
No ratings yet
23 DeepLearning PDF
74 pages
2004 - A Mathematical Approach To Classical Control - Autor Lewis
No ratings yet
2004 - A Mathematical Approach To Classical Control - Autor Lewis
678 pages
Fundamentals of Switching Theory
100% (1)
Fundamentals of Switching Theory
245 pages
Two-Dimensional Calculus
From Everand
Two-Dimensional Calculus
Robert Osserman
5/5 (1)
Analysis in Euclidean Space
From Everand
Analysis in Euclidean Space
Kenneth Hoffman
No ratings yet
Questions Bank EE5I SAP 22524 021120
No ratings yet
Questions Bank EE5I SAP 22524 021120
3 pages
Oculo Orbital Trauma: MDCT Findings
No ratings yet
Oculo Orbital Trauma: MDCT Findings
32 pages
GLG The Expert Interview Blueprint
No ratings yet
GLG The Expert Interview Blueprint
17 pages
Download EU Foreign Policy and Post Soviet Conflicts Stealth Intervention 1st Edition Nicu Popescu ebook All Chapters PDF
100% (5)
Download EU Foreign Policy and Post Soviet Conflicts Stealth Intervention 1st Edition Nicu Popescu ebook All Chapters PDF
81 pages
Hyper Hype 1
No ratings yet
Hyper Hype 1
8 pages
Using Homophones Correctly: 200 Administration Bldg. 651-603-6233
No ratings yet
Using Homophones Correctly: 200 Administration Bldg. 651-603-6233
3 pages
Teena FDHG
No ratings yet
Teena FDHG
2 pages
Vihasifine Chem PVT LTD, (Vihasi.m@gmail - Com)
No ratings yet
Vihasifine Chem PVT LTD, (Vihasi.m@gmail - Com)
647 pages
Flint Kote
No ratings yet
Flint Kote
3 pages
AbberantInfluence Eldritch CO
No ratings yet
AbberantInfluence Eldritch CO
17 pages
Case Study 1
No ratings yet
Case Study 1
8 pages
Borders Maths Coursework
100% (2)
Borders Maths Coursework
5 pages
Group Presentation and Individual Assessment Rubric
No ratings yet
Group Presentation and Individual Assessment Rubric
2 pages
Research
No ratings yet
Research
42 pages
Experiement 11 .MSK Home Automation WEBIOPi
No ratings yet
Experiement 11 .MSK Home Automation WEBIOPi
21 pages
Dissolution of A Partnership Firm
50% (2)
Dissolution of A Partnership Firm
5 pages
Jurnal Empowerment 1
No ratings yet
Jurnal Empowerment 1
10 pages
WORK MEASUREMENT , TIME STUDY
No ratings yet
WORK MEASUREMENT , TIME STUDY
2 pages
Maharlika
No ratings yet
Maharlika
4 pages
Jde Peets To Hold Extraordinary General Meeting of SH
No ratings yet
Jde Peets To Hold Extraordinary General Meeting of SH
2 pages
1.talking Facts: Simple Present
No ratings yet
1.talking Facts: Simple Present
4 pages
Sensor Principle MH-C MH-C2 Electrical Data: Vert-X 48E - 24V / 4 - 20ma Applications
No ratings yet
Sensor Principle MH-C MH-C2 Electrical Data: Vert-X 48E - 24V / 4 - 20ma Applications
5 pages
PPMP20010 Assessment 1 Part B (Project Controls Report) - Briefing Slides
No ratings yet
PPMP20010 Assessment 1 Part B (Project Controls Report) - Briefing Slides
15 pages
Problem Set For ISI CMI Entrance 18 6 2024
No ratings yet
Problem Set For ISI CMI Entrance 18 6 2024
34 pages
Lucky File
No ratings yet
Lucky File
63 pages
Volvo_ECR35D_ECR40D_Specs
No ratings yet
Volvo_ECR35D_ECR40D_Specs
16 pages
Mepco Online Billl
No ratings yet
Mepco Online Billl
1 page