ISSN 0908-1208

Robust Control
Steen Tner-Clausen,

Palle Andersen

and Jakob Stoustrup

U96-4153

April 18, 2001

4th Edition

Department of Control Engineering, Institute of Ele troni Systems

Aalborg University, Fredrik Bajers Vej 7, DK-9220 Aalborg , Denmark

Page II of IV

Preamble
This note has been written for a basi ourse in robust and optimal ontrol at 9th term of the
System Constru tion Line, Institute of Ele troni Systems, Aalborg University.

Originally,

the note was intended for a ourse onsisting of six modules of four hours ea h. Currently, the
ourses at Aalborg University has been redu ed to ve modules, and hen e, the note ontains
material, that an not be in luded in the ourse. The note has been adapted to a level, whi h
an be expe ted of 9th term students at the System Constru tion Line.

The students are

expe ted to be a quainted with lassi al feedba k ontrol theory.

The purposes of the note is to provide an introdu tion to modern robust and optimal ontrol,
espe ially in

H1 and  theory.

In Chapter 1, a short introdu tion to the on ept of robust

ontrol is given; In Chapter 2 nominal and robust stability for single variable (SISO) systems
is des ribed. In Chapter 3 nominal and robust performan e for SISO systems is analyzed, and
the on epts

H1 and H2 optimal ontrol are introdu ed.

Chapter 4 gives an introdu tion

to the analysis of multi variable systems, and in Chapter 5, stability and performan e of
multi variable systems are studied. In Chapter 6, a solution to the

H1 ontrol problem is

presented. Finally, in Chapter 7, an introdu tion to the stru tured singular value
and ontroller design with

 is treated.

 is given,

Key Words
Robust optimal ontrol; robust stability; robust performan e;

 synthesis;

H1 optimal ontrol;  analysis;

Aliation of the Authors

The authors are with Grundfos A/S, and with the Department of Control Engineering, Institute of Ele troni Systems, Aalborg University, DK-9220 Aalborg , Denmark. The depart-

http://www. ontrol.au .dk.

{pa,js} ontrol.au .dk

ment has the following home page:

authors are:

Robust Control

The emails of the last two

Page III of IV

Contents
1 Robust Feedba k Control

1.1 Feedba k and Un ertainty . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1.2 Classi al Compensator Design Methods . . . . . . . . . . . . . . . . . . . . . . . . . .

1.3 Model Un ertainty. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1.4 Feedba k Systems with Model Un ertainties. . . . . . . . . . . . . . . . . . . . . . . .

1.5 Robust Control. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

2 Nominal and Robust Stability

2.1 A Model of the Pro es . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

2.2 Model Un ertainty . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

2.3 Nominal Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

11

2.4 Robust Stability. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

11

3 Nominal and Robust Performan e

13

3.1 Signal Norms. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

13

3.2 Norms for Systems and Transfer Fun tions. . . . . . . . . . . . . . . . . . . . . . . . .

15

3.3 Spe i ation of inputs. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

15

3.4 Requirements for Performan e. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

18

3.5

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

19

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

19

3.7 Robust Performan e. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

20

3.8

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

21

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

21

3.10 Loop Shaping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

23

3.6

3.9

H2 Optimal Control (LQ). .

H1 Optimal Control. . . .
H2 Robust Performan e .
H1 Robust Performan e.

4 An Introdu tion to Multivariable Systems

4.1 Poles and zeros of Multivariable Systems . . . . . . . . . . . . . . . . . . . . . . . . . .

26
26

4.1.1

Smith-M Millan Form of a Transfer Matrix . . . . . . . . . . . . . . . . . . . .

27

4.1.2

State Spa e Des riptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

30

4.2 Nominal Stability for Multivariable Systems . . . . . . . . . . . . . . . . . . . . . . . .

33

4.2.1

Internal Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

33

4.2.2

The Generalized Nyquist Theorem . . . . . . . . . . . . . . . . . . . . . . . . .

34

4.3 Frequen y Responses for Multivariable Systems . . . . . . . . . . . . . . . . . . . . . .

37

4.3.1

Ve tor Norms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

38

4.3.2

Indu ed Norms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

38

4.3.3

Singular Values . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

38

Robust Control

Page IV of IV

5 Robustness Analysis for Multivariable Systems

5.1 Nominal Performan e . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.1.1

Norms of Signals for Multivariable Systems . . . . . . . . . . . . . . . . . . . .

43
44
45

2  2 Blo k Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

49

5.2 Robust Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

50

5.1.2
5.2.1

The Small Gain Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

51

5.3 Robust Performan e . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

55

5.3.1

Spe i ations with Mixed Sensitivity Fun tions . . . . . . . . . . . . . . . . . .

58

5.3.2

The Signi an e of Zeros and Poles in the Right Half Plane . . . . . . . . . . .

60

6 Robust Design for Multivariable Systems

63

6.1 Loop Shaping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

63

6.2 Modeling Individual Channels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

64

6.3

66

H1 Control .

6.3.1
6.3.2

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

Remarks to the

H1 solution

The MatlabTM Toolboxes

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

68

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

72

7 Design of Robust Compensators based on  Theory

7.1  Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

77
77

7.1.1

Robust Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

77

7.1.2

Robust Performan e . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

83

7.1.3

Computing  . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

85

7.2  synthesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

88

7.2.1

Complex  Synthesis 

D-K

iteration . . . . . . . . . . . . . . . . . . . . . . .

Robust Control

89

Chapter 1

Robust Feedba k Control

Chapter 1 ontains a general presentation of lassi al feedba k ontrol, with an emphasis on
robustness to model un ertainty of feedba k systems. The need to develop design methods,
that expli itly an handle model un ertainty is demonstrated.

1.1 Feedba k and Un ertainty

Control of a dynami al pro ess by feedba k of a measured output is a well-known prin iple,
typi ally with the primary obje tive of keeping the output of the pro ess lose to a given
referen e.

d(s)
r(s)

e(s) - K (s)

Figure 1.1:

u(s)- G(s)

m(s)

-y(s)

Standard feedba k onguration.

In Figure 1.1, a standard feedba k onguration is shown, onsisting of a pro ess

ompensator

K (s).

G(s) and a

For systems with feedba k the following important properties an often

be established:






Systems that are unstable from nature, an be stabilized

The ee t of external disturban es an be attenuated
The time onstants or hara teristi frequen ies of the system an be shifted
Finally, often the above properties an be obtained even with an in omplete knowledge
of the properties of the pro ess itself

Page 2 of 92
Hen e, feedba k ontrolled systems have attra tive properties, that an be made robust with
some are.
The advantages of feedba k an be illustrated by onsidering the un ertain elements in a
pro ess.

There are two reasons why, an output from a pro ess is not ompletely known a

priori. First, the dynami al properties of the system are not entirely known, i.e., a model an
be onsidered only as an approximate des ription of a physi al pro ess.
disturban es an inuen e the pro ess. An output

Se ond, unknown

is not only a fun tion of the input

ontrolled by the ompensator, but also of a disturban e

u,

d, whi h often an not be measured.

The basi properties of feedba k ( losed loop ontrol) is exposed by a omparison with a
feedforward onguration (open loop ontrol). By open loop (OL) ontrol the result depend
ompletely on the a ura y by whi h the pro ess has been modeled, sin e the ompensator
determines the input based only on the model and the referen e. Deviations aused by the
disturban e

d and model un ertainties have a full impa t as a dis repan y between the a tual

and the expe ted output.

d(s)
r(s)

- K (s)
Figure 1.2:

u(s)- G(s)

m(s)

-y(s)

Feedforward onguration.

In ontrast to this, the feedba k ompensator ompares the a tual value of the output
with the referen e

r,

and determines the input

based on the error

e.

On top of the a

priori knowledge available on the system in terms of a mathemati al model, the feedba k
ompensator exploits the knowledge of the a tual behavior of the pro ess and the simultaneous

y. Hen e, it is possible to redu e

d as well as the ee t of the imperfe t modeling.

disturban es, whi h are impli itly given by the measurement

the ee t of the disturban e

Design of feedba k ompensators embarks from requirements on erning the stati and dynami al behavior of the ontrolled system. These requirements ould in lude the following,
[Lun89:

1. Stability: the operation point of the ontrolled systems must be stable.

r(t) and disturban es d(t) the

e(t) must tend to zero, as the time t tends to innity.

2. Asymptoti ontrol: for a given lass of referen e inputs

error

3. Dynami al requirements: the ontrolled system must fulll a set of spe i ations, su h
as bounds on the step response, and requirements to some degree of de oupling between
the various signals, et .
4. Requirements on robustness: The properties that are spe ied for the ontrolled system
above, must be preserved under a given lass of variations in the pro ess dynami s.

In pro ess ontrol, the requirements of item 3 are often formulated as spe i ations for the
output

y(t) and for the ontrol signal u(t) by step or ramp shaped variations in referen e or
Robust Control

Page 3 of 92
disturban e. The spe i ations, e.g. an be bounds on overshoot
time

ts

Mp , rise time tr and settling

for a step response, see Figure 1.3.

The requirements an also be formulated in the frequen y domain as spe i ations for the
open loop or losed loop transfer fun tion, for example as onditions on a losed loop resonan e
peak

Mr

or bandwidth

f0 , see Figure 1.4.

y
+ 1%

Mp
1.0
0.9

-1%

tr

ts

time

Time domain design spe i ations: onditions on Mp , tr and ts for output y by a

step in the referen e signal r.

Figure 1.3:

dB

Closed loop
gain

Mr

-3

f0

Figure 1.4:

Frequen y domain design spe i ations: onditions on Mr and f0 .

1.2 Classi al Compensator Design Methods

A majority of the methods, that are applied to the design of ompensators, presume an exa t
model of the pro ess with an emphasis on items 1-3.

Robust Control

Page 4 of 92
Stability analysis in the frequen y plane was developed by Nyquist and Bode. Design of the
ompensators were arried out by methods based on 'trial and error'.
Design rules su h as Ziegler and Ni hols' are based on simple models, but they have su esfully
been applied to pro esses with a higher omplexity.
Root lo us analysis was developed by W.R. Evans, and aims at obtaining pres ribed dynami al properties by determining a ompensator, whi h provides a satisfa tory losed loop pole
pla ement.
Moreover, methods have been developed that provide expli it formulae for a ompensator
with a pres ribed pole pla ement.
Methods, that aim at satisfying performan e spe i ations formulated as onditions on the
integral of the square error, have been introdu ed by Newton, Gould & Kaiser and further
developed by Kalman et al.
Even though one of the obje tives of feedba k ontrol is to to redu e the ee ts of model
deviations, all these methods presume a perfe t model, and only indire tly a ounts for the
fa t, that a model an never be perfe t.
Sin e several of the methods have been applied su esfully for many years due to the inherent
robustness of feedba k ompensators, it an be questioned whether it is reasonable to in lude
model variations as an expli it design ondition. There are, however, many examples where
the introdu tion of feedba k not automati ally imply the robustness required to a tual model
variations. The following example originates from [Lun89.

Example 1.1 (Robust Stability)

A pro ess is des ribed by the model

G(s) =

1
1 + 1 s

whi h experimentally has been shown to provide approximately the same step response as the
real pro ess. The pro ess an, e.g., be ontrolled by a proportional ompensator with the gain
K . Theoreti ally, it should be possible without problems to sele t the gain K arbitrarily large.
It turns out in reality, however, that a more a urate model of the pro ess is given by

G^ (s) =

0 s

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

Now, it is apparent that the ontrol system will be ome unstable, if K is hosen too large. To
ensure stability, K must be hosen below the bound

K<

1 + 2
0

If for example 0 = 2 = 0:1
1 , K must satisfy K < 11 to ensure stability.

The example illustrate, that even if simple step response experiments lead to a model, that
approximately des ribe the pro ess, su h a model an not be used for ompensator design
without evaluating the region of validity of the model.

Robust Control

Page 5 of 92
The example further illustrate the main problem in the design of ompensators for systems
with model un ertainties: the desire of good performan e lead to a need for high gains. This
is in ontrast to robustness, sin e model deviations easily an lead to instability in systems
with high gains.
The fa t that pra ti al ompensators have been designed in the past, is due to experien ed
engineers who design ompensators with a ertain onservatism, by not taking the gains to
the theoreti al limit. For ompensators that are derived by minimizing a ost fun tion, some
robustness an be a hieved by in orporating ontrol signal in rements in the ost fun tion.
The development of methods that dire tly in orporate model un ertainties in ompensator
design, has a elerated within the past 20 years.

1.3 Model Un ertainty.

Normally, the rst step in a ompensator design would be the derivation of a model of the
pro ess to be ontrolled. Design of ompensators based on the model derived must take into
onsideration under whi h onditions the model is valid.

If ertainty is required that the

ompensator is going to work under all onditions, it is ne essary to augment the model of
the pro ess with a model, whi h expresses the possible deviations from the nominal model.
There are to reasons why, the output of the pro ess an not be predi ted exa tly by the derived
model of the preo ess. The pro ess an be inuen ed by disturban es, and the dynami s of
the model an deviate from that of the model, see Figure 1.5
Disturban es are external signals, that are independent of the pro ess inputs. The ee t of
disturban es an be aggregated at the output of the pro ess as an exogenous input

d, added

to the model output, as shown in Figure 1.5a.

Deviations between the dynami s of the pro ess and its model lead to dis repan ies that in
ontrast to disturban es are not aused by unknown disturban es, but highly depend on the

u. Hen e, model un ertainties an be represented

output w , see Figure 1.5b.

input

by an error model with input

and

Reje tion of disturban es is an integral part of ompensator design related to items 1-3. Thus,
in this ontext model un ertainties are emphasized. There are three main sour es for model
un ertainty:
1. In omplete knowledge on the pro ess; this type of un ertainty might be due to the fa t,
that the model has been derived from the laws of physi s, although the exa t parameters
of the pro ess an not be determined from the available knowledge on the pro ess. If
the model is determined experimentally, the a ura y of the model depends on whether
the pro ess has been ex ited by inputs, that are suited for determining a model, and to
what extent the pro ess has been inuen ed by disturban es during the experiment.
2. Model simpli ation: even though the original system might be known in great detail,
the model might have been redu ed in order to simplify the design task.
3. In omplete model stru ture: in general it is desirable to design ompensators based on
a linear model. Hen e, nonlinearities in a tuators or sensors must be omitted. Other
types of nonlinearities is aused by nonlinear dynami s of the pro ess itself. This type
of nonlinearities often result in parameters that depend on the operating point.

Robust Control

Page 6 of 92

d(s)

Unknown
Known

r(s)

e(s) - K (s)

u(s)- G(s)

m(s)

-y(s)

-y(s)

a)

z (s)

(s)

w(s)

Unknown
Known

r(s)

e(s) - K (s)

u(s)

- G(s)

m(s)

b)
Figure 1.5:

Disturban es vs. model un ertainty in the losed loop system. a) Closed loop system
exposed to disturban es. b) Closed loop system with model un ertainty.

4. Time varying parameters: also result in variation of the parameters of the model.
In any ir umstan e, the model will only be an approximate representation of the physi al
pro ess.
Systems with signi ant model un ertainties are simply refered to as

un ertain systems

irre-

spe tive of whether the un ertainties are aused by la k of information, model simpli ations,
nonlinearities, or time varying parameters.
In the literature, various on epts are asso iated with spe ial types of model un ertainties.

Parametri un ertainty

is the type of un ertainty that an be ompensated for by adjusting

the parameters of the model.

Stru tural un ertainty

is the type of un ertainty that relate to

an in omplete or in orre t model stru ture, e.g. by applying a linear model for pro ess that
exhibits nonlinear behavior, or by omitting dynami al elements in the pro ess.

Moreover,

model un ertainty is lassied subje t to whether only the overall level of un ertainty is
known, or whether also the hara ter of the un ertainty is known.

un ertainty

is distinguished from

unstru tured un ertainty.

To this end

stru tured

Model un ertainty an be des ribed by dening a set of possible models. In this set or family
of models, ea h member an represent the original pro ess, but it is unknown whi h spe i
member that a tually does.

Dependent on the hara ter of the model un ertainty, the set

of possible models appear in dierent ways. By mapping all possible models in the Nyquist
plane, a bounded region in the plane is asso iated with ea h frequen y rather than an isolated
point.

Robust Control

Page 7 of 92
Norm bounded model deviations were suggested by Doyle in onne tion with ompensator
design. The absolute value of the model deviation was limited to a ertain frequen y dependent
quantity. This orresponds to ir ular bounds in the Nyquist plane.

1.4 Feedba k Systems with Model Un ertainties.

Model un ertainty an not in general be treated as disturban es, although this is often done
in pra ti e, espe ially if the un ertainty is relatively small. Exogenous disturban es an not
inuen e the stability of a linear losed loop system. However, as it was seen from Example 1.1,
a model deviation an inuen e stability. In the example, the proportional gain had to be
within a ertain upper bound, depending on the parameters of the 'real' model.
In the sequel, a series of on epts are introdu ed whi h are signi ant for un ertain feedba k
systems:

Sensitivity is a measure of how a ertain property (e.g.

the losed loop transfer fun tion)

depend on dierential deviations of a parameter or on the entire ve tor of parameters.

Sensitivity analysis deal with small deviations in the neighborhood of a nominal value,
and does not refer to the admissible size of a deviation.

Robustness

des ribe the apability of a system to exhibit satisfa tory properties for all

models in a given family. Hen e, robustness refer to an a eptable region for the desired
properties and for given bounds of model deviations.

Neither sensitivity nor robustness indi ate the likelihood of a ertain model deviation
to arise in a ertain time interval.

Reilability

is a measure for the apability of a system

to fun tion satisfa torily within a ertain time range. In reliability analysis, statisti al
methods are applied. Thus, quantitative measures for reliability are probabilities, that a
system fun tions satisfa torily at a ertain time instant, or within a ertain time range.

Adaptivity

is the apability of a ontrol system to adjust to hanges in the pro ess.

Some time after the hanges in the pro ess have o ured, an adaptive system will have
adjustet itself, and will again fun tion satisfa torily.
In feedba k ontrol, sensitivity, robustness, and adaptivity play ru ial roles. Designing robust ontrollers, involves looking for linear ompensators that ensure satisfa tory
performan e for all possible model variations. As an alternative, adaptive ontrol methods try to adjust the ompensator to the model that provides the best t at any time
instan e.
Reliability analysis is mainly used in onne tion with implementation of ompensators and
other pro ess ontrol equipment, and mainly to answer questions like:

How large is the

probability that the pro ess ontrol equipment performs satisfa torily?

1.5 Robust Control.

These le ture notes addresses the analysis of robustness of ontrol systems and the design of
robust ontrol systems. This has been the subje t of intense resear h during the past years.
In this ontext, the following resear h results are ru ial:

Robust Control

Page 8 of 92

In 1976, Youla et al. [YJB76a, [YJB76b demonstrated, that it is possible to parameterize all ompensators that stabilize a given system. This greatly simplies the design
of a stabilizing ontroller.

In 1979, Doyle and Stein [DS79 pointed out that the good phase and amplitude margins
of the LQ ontroller, easily are ruined by an observer, and also suggested a method to
re over these appealing properties.

In 1981, Zames, [Zam81, suggested to apply the

H1 norm rather than the traditonal

H2 norm in order to evaluate the performan e of a ompensator.

In 1981, Doyle and Stein, [DS81, showed that model un ertainties an be des ribed
well as norm bounded deviations. Along with the

H1 norm, norm bounded deviations

onstitute an ex ellent tool to des ribe the robustness of a ompensator.

In 1982,

Doyle [Doy82 suggested to generalize this to stru tured un ertainties by introdu ing
the stru tured singular value,

.

In 1989, Doyle et al. [DGKF89 derived expli it design formulae for ompensator design
that minimized the

H1 norm based on state spa e models.

Robust Control