Theory of Robust Control

Carsten Scherer
Mathematical Systems Theory
Department of Mathematics
University of Stuttgart
Germany

Contents

1 Introduction to Basic Concepts 6

1.1 Systems and Signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

1.2 Stability of LTI Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

1.3 Stable Inverses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

2 Robustness for SISO Systems 16

2.1 The Standard Tracking Configuration . . . . . . . . . . . . . . . . . . . . . 17

2.2 The Nyquist Stability Criterion . . . . . . . . . . . . . . . . . . . . . . . . 18

2.3 Classical Robustness Indicators . . . . . . . . . . . . . . . . . . . . . . . . 22

2.4 Plant Uncertainty . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

2.5 Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

2.6 Internal Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

3 Stabilizing Controllers for Interconnections 46

3.1 A Specific Tracking Interconnection . . . . . . . . . . . . . . . . . . . . . . 46

1
 3.2 The General Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

3.3 Stabilizing Controllers - State-Space Descriptions . . . . . . . . . . . . . . 49

3.4 Linear Fractional Transformations . . . . . . . . . . . . . . . . . . . . . . . 53

3.5 Stabilizing Controllers - Input-Output Descriptions . . . . . . . . . . . . . 54

3.6 Generalized Plants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58

3.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62

3.8 Back to the Tracking Interconnection . . . . . . . . . . . . . . . . . . . . . 62

4 Robust Stability Analysis 66

4.1 Uncertainties in the Tracking Configuration - Examples . . . . . . . . . . . 66

4.1.1 A Classical SISO Example . . . . . . . . . . . . . . . . . . . . . . . 66

4.1.2 A Modern MIMO Example . . . . . . . . . . . . . . . . . . . . . . 73

4.2 Types of Uncertainties of System Components . . . . . . . . . . . . . . . . 76

4.2.1 Parametric Uncertainties . . . . . . . . . . . . . . . . . . . . . . . . 76

4.2.2 Dynamic Uncertainties . . . . . . . . . . . . . . . . . . . . . . . . . 78

4.2.3 Mixed Uncertainties . . . . . . . . . . . . . . . . . . . . . . . . . . 81

4.2.4 Unstructured Uncertainties . . . . . . . . . . . . . . . . . . . . . . 82

4.2.5 Unstructured versus Structured Uncertainties . . . . . . . . . . . . 83

4.3 Summary on Uncertainty Modeling for Components . . . . . . . . . . . . . 85

4.4 Pulling out the Uncertainties . . . . . . . . . . . . . . . . . . . . . . . . . . 85

4.4.1 Pulling Uncertainties out of Subsystems . . . . . . . . . . . . . . . 86

4.4.2 Pulling Uncertainties out of Interconnections . . . . . . . . . . . . . 90

4.5 The General Paradigm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94

4.6 What is Robust Stability? . . . . . . . . . . . . . . . . . . . . . . . . . . . 97

4.7 Robust Stability Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . 98

4.7.1 Simplify Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . 98

2
 4.7.2 Reduction to Non-Singularity Test on Imaginary Axis . . . . . . . . 100

4.7.3 The Central Test for Robust Stability . . . . . . . . . . . . . . . . . 104

4.7.4 Construction of Destabilizing Perturbations . . . . . . . . . . . . . 105

4.8 Important Specific Robust Stability Tests . . . . . . . . . . . . . . . . . . . 106

4.8.1 M is Scalar . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107

4.8.2 The Small-Gain Theorem . . . . . . . . . . . . . . . . . . . . . . . 107

4.8.3 Full Block Uncertainties . . . . . . . . . . . . . . . . . . . . . . . . 109

4.9 The Structured Singular Value in a Unifying Framework . . . . . . . . . . 111

4.9.1 The Structured Singular Value . . . . . . . . . . . . . . . . . . . . . 113

4.9.2 SSV Applied to Testing Robust Stability . . . . . . . . . . . . . . . 115

4.9.3 Construction of Destabilizing Perturbations . . . . . . . . . . . . . 117

4.9.4 Example: Two Uncertainty Blocks in Tracking Configuration . . . . 118

4.9.5 SSV Applied to Testing the General Hypothesis . . . . . . . . . . . 122

4.10 A Brief Survey on the Structured Singular Value . . . . . . . . . . . . . . . 122

4.10.1 Continuity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124

4.10.2 Lower Bounds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124

4.10.3 Upper Bounds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125

4.11 When is the Upper Bound equal to the SSV? . . . . . . . . . . . . . . . . . 130

4.11.1 Example: Different Lower and Upper Bounds . . . . . . . . . . . . 130

5 Nominal Performance Specifications 136

5.1 An Alternative Interpretation of the H∞ Norm . . . . . . . . . . . . . . . . 136

5.2 The Tracking Interconnection . . . . . . . . . . . . . . . . . . . . . . . . . 138

5.2.1 Bound on Frequency Weighted System Gain . . . . . . . . . . . . . 138

5.2.2 Frequency Weighted Model Matching . . . . . . . . . . . . . . . . . 141

5.3 The General Paradigm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142

3
6 Robust Performance Analysis 144

6.1 Problem Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144

6.2 Testing Robust Performance . . . . . . . . . . . . . . . . . . . . . . . . . . 147

6.3 The Main Loop Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147

6.4 The Main Robust Stability and Robust Performance Test . . . . . . . . . . 150

6.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151

6.6 An Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152

7 Synthesis of H∞ Controllers 159

7.1 The Algebraic Riccati Equation and Inequality . . . . . . . . . . . . . . . . 159

7.2 Computation of H∞ Norms . . . . . . . . . . . . . . . . . . . . . . . . . . 162

7.3 The Bounded Real Lemma . . . . . . . . . . . . . . . . . . . . . . . . . . . 164

7.4 The H∞ -Control Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . 166

7.5 H∞ -Control for a Simplified Generalized Plant Description . . . . . . . . . 167

7.6 The State-Feedback Problem . . . . . . . . . . . . . . . . . . . . . . . . . . 168

7.6.1 Solution in Terms of Riccati Inequalities . . . . . . . . . . . . . . . 168

7.6.2 Solution in Terms of Riccati Equations . . . . . . . . . . . . . . . . 170

7.7 H∞ -Observer Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171

7.8 H∞ -Control by Output-Feedback . . . . . . . . . . . . . . . . . . . . . . . 172

7.8.1 Solution in Terms of Riccati Inequalities . . . . . . . . . . . . . . . 173

7.8.2 Solution in Terms of Riccati Equations . . . . . . . . . . . . . . . . 177

7.8.3 Solution in Terms of Indefinite Riccati Equations . . . . . . . . . . 179

7.9 What are the Weakest Hypotheses for the Riccati Solution? . . . . . . . . 183

7.10 Game-Theoretic Interpretation . . . . . . . . . . . . . . . . . . . . . . . . . 185

7.11 Interpretation and Parametrization of Controllers . . . . . . . . . . . . . . 196

7.11.1 Interpretation of the Output-Feedback Controller . . . . . . . . . . 206

7.11.2 Controller Parameterization . . . . . . . . . . . . . . . . . . . . . . 207

4
8 Robust Performance Synthesis 217

8.1 Problem Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217

8.2 The Scalings/Controller Iteration . . . . . . . . . . . . . . . . . . . . . . . 219

9 Youla Jabr Bongiorno Kucera Parametrizations 224

9.1 Algebraic Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 224

9.2 Stabilizable Plants and Controller Parametrization . . . . . . . . . . . . . 228

9.3 Double Bézout Identity for LTI Systems . . . . . . . . . . . . . . . . . . . 231

9.4 Youla Parametrization for Generalized Plants . . . . . . . . . . . . . . . . 234

10 A Brief Summary and Outlook 241

A Bisection 242

B Proof of Theorem 7.1 242

C Proof of Theorem 7.2 243

References 247

5
1 Introduction to Basic Concepts

1.1 Systems and Signals

In these notes we intend to develop the theory of robust control for linear time invariant
finite-dimensional systems that are briefly called LTI-systems. Recall that such systems
are described in the state-space as

ẋ = Ax + Bu, x(0) = x0
(1.1)
y = Cx + Du

with input u, output y, state x, and real matrices A, B, C, D of suitable size.

Here u, y, x are signals. Signals are functions of time t ∈ [0, ∞) that are piece-wise contin-
uous. (On finite intervals, such signals have only finitely many jumps as discontinuities.)
They can either take their values in R, or they can have k components such that they
take their values in Rk . To clearly identify e.g. x as a signal, we sometimes write x(.) to
stress this point. (Recall that x(.) denotes the signal as a whole, whereas x(t) denotes
the value of the signal at the time-instant t ≥ 0.)

Somewhat more precisely, (1.1) is defined by A ∈ Rn×n , B ∈ Rn×m , C ∈ Rk×n and

D ∈ Rk×n . Moreover, for an input trajectory u : [0, ∞) → Rm , there exists a unique state
response x : [0, ∞) → Rn of the system (1.1) which has the output y : [0, ∞) → Rk . It is
well-known that the output response y(.) to the input u(.) is given by
Z t
At
y(t) = Ce x0 + CeA(t−τ ) Bu(τ ) dτ + Du(t) for t ≥ 0.
0

We do not repeat the standard notions of controllability of the system or of the pair (A, B),
and of observability of the system or of (A, C). Nevertheless we recall the following very
basic facts:

• The Hautus test for controllability: (A, B) is controllable if and only if the
matrix  
A − λI B
has full row rank for all λ ∈ C.

• The Hautus test for observability: (A, C) is observable if and only if the matrix
 
A − λI
 
C

has full column rank for all λ ∈ C.

6
 • The system (1.1) or (A, B) is said to be stabilizable if there exists a feedback
matrix F such that A + BF has all its eigenvalues in the open left-half plane C< :=
{λ ∈ C | Re(λ) < 0}.
Recall the Hautus test for stabilizability: (A, B) is stabilizable if and only if
the matrix  
A − λI B
has full column rank for all λ ∈ C≥ , where C≥ denotes the closed right half-plane
C≥ := {λ ∈ C | Re(λ) ≥ 0}.

• The system (1.1) or (A, C) is said to be detectable if there exists an L such that
A + LC has all its eigenvalues in the open left-half plane C< .
Recall the Hautus test for detectability: (A, C) is detectable if and only if the
matrix  
A − λI
 
C
has full column rank for all λ ∈ C≥

The transfer matrix G(s) of the system (1.1) is defined as

G(s) = C(sI − A)−1 B + D

and is a matrix whose elements consist of real-rational and proper functions in s. Any
such function is a fraction of two polynomials in s with real coefficients; properness of
such a function means that the degree of the numerator is not larger than the degree of
the denominator.

Why does the transfer matrix pop up? Suppose the input signal u(.) has the Laplace-
transform Z ∞
û(s) = e−st u(t) dt.
0

Then the output y(.) of (1.1) also has a Laplace transform that can be calculated as

ŷ(s) = C(sI − A)−1 x0 + [C(sI − A)−1 B + D]û(s).

For x0 = 0 (which means that the state of the system starts with value zero at time 0),
the relation between the Laplace transform of the input and the output signals is hence
given by the transfer matrix as follows:

ŷ(s) = G(s)û(s).

The ‘complicated’ convolution integral in the time-domain is transformed into a the ‘sim-
pler’ multiplication operation in the frequency domain.

7
We have briefly addressed two different ways of representing a system: One representation
in the state-space defined with specific constant matrices A, B, C, D, and one in the
frequency domain defined via a real-rational proper transfer matrix G(s).

Remark 1.1 In this course we want to view a system as a device that processes signals;
hence a system is nothing but a mapping that maps the input signal u(.) into the output
signal y(.) (for a certain initial condition). One should distinguish the system (the map-
ping) from its representations, such as the one in the state-space via A, B, C, D, or that
in the frequency domain via G(s). System properties should be formulated in terms of
how signals are processed, and system representations are used to formulate algorithms
how certain system properties can be verified.

The fundamental relation between the state-space and frequency domain representation
is investigated in the so-called realization theory. Moving from the state-space to the
frequency domain just requires to calculate the transfer matrix G(s).

Conversely, suppose H(s) is an arbitrary matrix whose elements are real-rational proper
functions. Then there always exist real matrices AH , BH , CH , DH such that

H(s) = CH (sI − AH )−1 BH + DH

holds true. This representation of the transfer matrix is called a realization. Realizations
are not unique. Even more importantly, the size of the matrix AH can vary for various
realizations. However, there are realizations where AH is of minimal size, the so-called
minimal realization. There is a simple answer to the question of whether a realization is
minimal: This happens if and only if (AH , BH ) is controllable and (AH , CH ) is observable.

Task. Recapitulate how you can compute a minimal realization of an arbitrary real
rational proper transfer matrix H.

Pictorially, this discussion about the system representations in the time- and frequency-
domain and the interpretation as a mapping of signals (for a zero initial condition of the
state) can be depicted as follows:

8
 ẋ = Ax + Bu Realization
G(s)
y = Cx + Du

Z t Laplace Transform
y(t) = CeA(t−τ ) Bu(τ )dτ + Du(t) ŷ(s) = G(s)û(s)
0

We use the symbol  

A B
 
C D
both for the system mapping u → y as defined via the differential equation with initial
condition 0, and for the corresponding transfer matrix G.

1.2 Stability of LTI Systems

Recall that any matrix H(s) whose elements are real rational functions is stable if

• H(s) is proper (there is no pole at infinity) and

• H(s) has only poles in the open left-half plane C< (i.e. there is no pole in the closed
right half plane C≥ ).

Properness means that, for all entries of H, the degree of the numerator polynomial is
not larger than the degree of the denominator polynomial. H is called strictly proper
if the degree of the numerator polynomial is strictly smaller than the degree of the de-
nominator polynomial for all elements of H. Note that strict properness is equivalent to
lim|s|→∞ H(s) = 0.

For the set of real rational proper and stable matrices of dimension k × m we use the
special symbol
k×m
RH∞
and if the dimension is understood from the context we simply write RH∞ . Recall that
the three most important operations performed on stable transfer matrices do not lead
us out of this set: A scalar multiple of one stable transfer matrix as well as the sum and
the product of two stable transfer matrices (of compatible dimensions) are stable.

9
On the other hand, the state-space system (1.1) is said to be stable if A has all its
eigenvalues in the open left-half plane C< . We will denote the set of eigenvalues of A by
eig(A), the spectrum of A. Then stability of (1.1) is simply expressed as

eig(A) ⊂ C< .

We say as well that the matrix A is stable (or Hurwitz) if it has this property.

We recall the following relation between the stability of the system (1.1) and the stability
of the corresponding transfer matrix G(s) = C(sI − A)−1 B + D:

• If (1.1) (or A) is stable, then G(s) is stable.

• Conversely, if G(s) is stable, if (A, B) is stabilizable, and if (A, C) is detectable,

then (1.1) (or A) is stable.

Note that all these definitions are given in terms of properties of the representation.
Nevertheless, these concepts are closely related - at least for LTI systems - to the so-
called bounded-input bounded-output stability properties.

A vector valued signal u(.) is bounded if the maximal amplitude or peak

kuk∞ = sup ku(t)k is finite.

t≥0
p
Here ku(t)k just equals the Euclidean norm u(t)T u(t) of the vector u(t). The symbol
kuk∞ for the peak indicates that the peak is, in fact, a norm on the vector space of all
bounded signals; it is called the L∞ -norm.

The system (1.1) is said to be bounded-input bounded-output (BIBO) stable if it maps an

arbitrary bounded input u(.) into an output that is bounded as well. In short, kuk∞ < ∞
implies kyk∞ < ∞. It is an interesting fact that, for LTI systems, BIBO stability is
equivalent to the stability of the corresponding transfer matrix as defined earlier.

Theorem 1.2 The system (1.1) maps bounded inputs u(.) into bounded outputs y(.) if
and only if the corresponding transfer matrix C(sI − A)−1 B + D is stable.

 
A B
To summarize, for a stabilizable and detectable realization   of an LTI system,
C D
the following notions are equivalent: Stability of the system (1.1), stability of the cor-
responding transfer matrix C(sI − A)−1 B + D, and BIBO stability of the system (1.1)
viewed as an input output mapping.

Stability is a qualitative property. Another important issue is to quantify in how far

signals are amplified or quenched by a system. If we look at one input u(.) 6= 0, and if

10
we take kuk∞ and kyk∞ as a measure of size for the input and the output of the system
(1.1), the amplification for this specific input signal is nothing but

kyk∞
.
kuk∞

The worst possible amplification is obtained by finding the largest of these quotients if
varying u(.) over all bounded signals:

kyk∞
γpeak = sup . (1.2)
0<kuk∞ <∞ kuk∞

This is the so-called peak-to-peak gain of the system (1.1). Then it just follows from
the definition that
kyk∞ ≤ γpeak kuk∞
holds for all bounded input signals u(.): Hence γpeak quantifies how the amplitudes of the
bounded input signals are amplified or quenched by the system. Since γpeak is, in fact,
the smallest number such that this inequality is satisfied, there does exist an input signal
such that the peak amplification is actually arbitrarily close to γpeak . (The supremum in
(1.2) is not necessarily attained by some input signal. Hence we cannot say that γpeak is
attained, but we can come arbitrarily close.)

Besides the peak, we could as well work with the energy of a signal x(.), defined as
sZ
∞
kxk2 = kx(t)k2 dt,
0

to measure its size. Note that a signal with a large energy can have a small peak and
vice versa. (Think of examples!) Hence we are really talking about different physical
motivations if deciding for k.k∞ or for k.k2 as a measure of size.

Now the question arises when a system maps any signal of finite energy again into a signal
of finite energy; in short:

kuk2 < ∞ implies kyk2 < ∞.

It is somewhat surprising that, for the system (1.1), this property is again equivalent to
the stability of the corresponding transfer matrix C(sI −A)−1 B+D. Hence the qualitative
property of BIBO stability does not depend on whether one chooses the peak k.k∞ or the
energy k.k2 to characterize boundedness of a signal.

Remark 1.3 Note that this is a fundamental property of LTI system that is by no means
valid for other type of systems, even if they admit a state-space realization such as non-
linear system defined via differential equations.

11
Although the qualitative property of stability does not depend on the chosen measure
of size for the signals, the quantitative measure for the system amplification, the system
gain, is highly dependent on the chosen norm. The energy gain of (1.1) is analogously
defined as for the peak-to-peak gain defined by

kyk2
γenergy = sup .
0<kuk2 <∞ kuk2

Contrary to the peak-to-peak gain, one can nicely relate the energy gain of the system
(1.1) to the transfer matrix of the system. In fact, one can prove that γenergy is equal to
the largest value of
σmax (G(iω)) = kG(iω)k
f varying the frequency as ω ∈ R. Here σmax (A) denotes the maximum singular value of
the matrix A which equals the spectral norm of A. Let us hence introduce the abbreviation

kGk∞ := sup σmax (G(iω)) = sup kG(iω)k.

ω∈R ω∈R

As indicated by the symbol, this formula defines a norm on the vector space of all real-
k×m
rational proper and stable matrices RH∞ which is called the H∞ -norm.

We can conclude that the energy gain of the stable LTI system (1.1) is just equal to the
H∞ -norm of the corresponding transfer matrix:

γenergy = kGk∞ .

1.3 Stable Inverses

For any real-rational matrix G(s), we can compute the real rational function det(G(s)).
It is well-known that G(s) has a real-rational inverse if and only if det(G(s)) is not the
zero function (does not vanish identically). If G(s) is proper, it is easy to verify that it
has a proper inverse if and only if det(G(∞)) (which is well-defined since G(∞) is just a
real matrix) does not vanish.

The goal is to derive a similar condition for the proper and stable G(s) to have a proper
and stable inverse. Here is the desired characterization.

Lemma 1.4 The proper and stable matrix G(s) has a proper and stable inverse if and
only if the matrix G(∞) is non-singular, and the rational function det(G(s)) does not
have any zeros in the closed right-half plane.

12
Proof. Assume that G(s) has the proper and stable inverse H(s). From G(s)H(s) = I
we infer det(G(s)) det(H(s)) = 1 or
1
det(G(s)) = .
det(H(s))
Since H(s) is stable, det(H(s)) has no poles in C≥ ∪ {∞}. Therefore, the reciprocal
rational function and hence det(G(s)) does not have zeros in this set.

Conversely, let us assume that det(G(s)) has no zeros in C≥ ∪ {∞}. Then clearly
1
is proper and stable.
det(G(s))
Now recall that the inverse of G(s) is given by the formula
1
G(s)−1 = adj(G(s))
det(G(s))
where adj(G(s)) denotes the algebraic adjoint of G(s). These adjoints are computed by
taking products and sums/differences of the elements of G(s); since G(s) is stable, the
adjoint of G(s) is, therefore, a stable matrix. Then the explicit formula for G(s)−1 reveals
that this inverse must actually be stable as well.

Remark 1.5 It is important to apply this result to stable G(s) only. For example, the
proper unstable matrix  
s+1 1
s+2 s−1
G(s) =  
0 s+2
s+1

satisfies det(G(s)) = 1 for all s. Hence, its determinant has no zeros in the closed right-half
plane and at infinity; nevertheless, its inverse is not stable!

Let us now assume that the proper G has a realization

 
A B
G= .
C D

Recall that G has a proper inverse iff D = G(∞) is invertible. If D has an inverse, the
proper inverse of G admits the realization
 
−1 −1
A − BD C BD
G−1 =  .
−1 −1
−D C D

Why? A signal based arguments leads directly to the answer:

ẋ = Ax + Bw, z = Cx + Dw (1.3)

13
is equivalent to
ẋ = Ax + Bw, w = −D−1 Cx + D−1 z
and hence to

ẋ = (A − BD−1 C)x + BD−1 z, w = −D−1 Cx + D−1 z. (1.4)

This gives a test of whether the stable G has a proper and stable inverse directly in terms
of the matrices A, B, C, D of some realization.
 
A B
Lemma 1.6 Let G(s) be stable and let G(s) =   be a stabilizable and detectable
C D
realization. Then G(s) has a proper and stable inverse if and only if D is non-singular
and A − BD−1 C is stable.

Proof. Suppose G has the proper and stable inverse H. Then G(∞) = D is non-
singular. We can hence define the system (1.4); since the realization (1.3) is stabilizable
and detectable, one can easily verify (with the Hautus test) that the same is true for the
realization (1.4). We have argued above that (1.4) is a realization of H; since H is stable,
we can conclude that A − BD−1 C must be stable.

The converse is easier to see: If D is non-singular and A − BD−1 C is stable, (1.4) defines
the stable transfer matrix H. Since H is the inverse of G, as was seen above, we conclude
that G admits a proper and stable inverse.

Exercises

1) Suppose that G(s) is a real-rational proper matrix. Explain in general how you
can compute a state-space realization of G(s) with the command sysic of Matlab’s
Robust Control Toolbox, and discuss how to obtain a minimal realization.
(Matlab) Compute in this way a realization of
 
2
1/s 1/(s + 1) s/(s + 1)
G(s) =  .
2 3
(s − 3s + 5)/(2s + 4s + 1) 1/s 1/(s + 1)

2) Suppose that u(.) has finite energy, let y(.) be the output of (1.1), and denote the
Laplace transforms of both signals by û(.) and ŷ(.) respectively. Show that
Z ∞ Z ∞
∗
ŷ(iω) ŷ(iω) dω ≤ kGk∞ 2
û(iω)∗ û(iω) dω.
−∞ −∞

14
 Argue that this implies kyk2 ≤ kGk∞ kuk2 , and that this reveals that the energy
gain γenergy is not larger than kGk∞ .
Can you find a sequence uj (.) of signals with finite energy such that

kGuj k2
lim = kGk∞ ?
j→∞ kuj k2

3) Look at the system ẋ = ax + u, y = x. Compute the peak-to-gain of this system.

Determine a worst input u, i.e., an input for which the peak of the corresponding
output equals the peak-to-peak gain of the system.

4) For any discrete-time real-valued signal x = (x0 , x1 , . . .) let us define the peak as

kxk∞ := sup |xk |.

k≥0

Consider the SISO discrete-time system

xk+1 = Axk + Buk , yk = Cxk + Duk , x0 = 0, k ≥ 0

where all eigenvalues of A have absolute value smaller than 1 (discrete-time stabil-
ity). As in continuous-time, the peak-to-peak gain of this system is defined as

kyk∞
sup .
0<kuk∞ <∞ kuk∞

Derive a formula for the the peak-to-peak gain of the system!

Hint: If setting um := (u0 · · · um )T , y m := (y0 · · · ym )T , determine a matrix Mm
such that y m = Mm um . How are the peak-to-peak gain of the system and the norm
of Mm induced by the vector norm k.k∞ related?

5) Prove Theorem 1.2.

15
2 Robustness for SISO Systems

In this chapter robustness for single-input single-output (SISO) linear time invariant sys-
tems is considered. For this purpose we introduce, in Section 2.1, a standard tracking
configuration and characterize stability of the sensitivity and complementary sensitivity
transfer function.

We recall a version of the classical Nyquist criterion, which is a graphical test for checking
closed-loop stability by considering the Nyquist plot of the open-loop transfer function.

Furthermore we provide a mathematically precise definition of a multiplicative uncertainty

model and develop the related tests for checking robust stability and robust performance.

Finally, in Section 2.6, we address the notion of internal stabilization for a standard
feedback loop that consists of SISO components only.

To start, it is convenient to remind us about some facts about coprime polynomials.

Definition 2.1 Let p and q be real polynomials. A greatest common divisor of p

and q is a monic polynomial d = gcd(p, q) that divides p and q, such that every common
divisor of p and q also divides d. Two polynomials p and q are coprime if gcd(p, q) = 1.

We mainly use coprime polynomials because of property c) of the following theorem.

Theorem 2.2 For two real polynomials p and q, the following statements are equivalent.

a) p and q are coprime.

b) There exist real polynomials a and b such that bp + aq = 1.

c) p and q do not have common zeros.

Proof. b) follows from a) from the well-known lemma of Bézout.

b) ⇒ c): If q(s) = 0 for s ∈ C then b(s)p(s) = b(s)q(s) + a(s)q(s) = 1. This implies

p(s) 6= 0.

c) ⇒ a): Assume that p and q are not coprime. Then there exists a polynomial d with
deg(d) > 0 and gcd(p, q) = d. By the fundamental theorem of algebra there exists some
s ∈ C with d(s) = 0. Since d divides p and q, s is also a common zero of p and q.

Any real rational function G, and in particular any transfer function, can be expressed as
N
G= with real coprime polynomials N and D.
D

16
 r e u y
+ K G
−

Figure 1: The standard tracking configuration.

Since N and D have no common zeros, the zeros of G in C are the zeros of N and the
poles of G in C are the zeros of D, counted with the corresponding multiplicities. If not
emphasized explicitly, statements about poles or zeros should always be interpreted as
1 1
“counted with multiplicities”. Therefore s−1 and (s−1)2 do not have identical poles.

If the relative degree r(G) := deg(D) − deg(N ) of G is positive, then G is said to have
a zero at ∞ with multiplicity r(G). If G is not proper, which means deg(N ) > deg(D),
then G is said to have a pole at ∞ of multiplicity −r(G) = deg(N ) − deg(D).

2.1 The Standard Tracking Configuration

In this section, if nothing else is mentioned, we consider the feedback interconnection

y = Gu, u = K(r − y)

with the open-loop system G and the controller K. We assume that both G and K are
SISO transfer functions. This interconnection is the standard tracking configuration and
is depicted in Figure 1.

Roughly speaking, the standard goals in designing a controller K for the interconnection
can be expressed as follows:

• Stabilize the interconnection.

• The system output y should track r well, which means that the norm of the tracking
error e = r − y is small.
• The control action u should not be too large.

To analyze stability, let us define the loop transfer function or loop-gain

L := GK.

Then the interconnection equations can be written as

y = Le, e = r − y.

If the transfer function L is not identically equal to −1 we get

L 1
y = T r and e = Sr with T := and S := .
1+L 1+L

17
S is called the sensitivity (transfer function) and T the complementary sensitivity
(transfer function) of the feedback loop. Note that S + T = 1 always holds. It is also easy
to see that the poles of S and T are the zeros of 1 + L.

Theorem 2.3 The poles of S and the poles of T in C ∪ {∞} are exactly given by the
zeros of 1 + L in C ∪ {∞}.

N
Proof. L can be expressed as L = D
with real coprime polynomials N and D. Then
N D+N
1+L=1+ =
D D
and thus
D N
S= as well as T = .
D+N D+N
If D(λ) = 0, coprimeness of N and D implies N (λ) 6= 0 and therefore D(λ) + N (λ) 6= 0.
Hence D and D + N are coprime. This means that the poles of S in C are given by the
zeros of D + N in C which are the zeros of 1 + L in C.

Now let D(s) = dk sk + · · · + d0 and N (s) = nm sm + · · · + n0 with dk , nm 6= 0. Then S has

a pole at ∞ iff deg(D) > deg(D + N ) iff k = m and dk + nm = 0 iff L(∞) = −1 iff L + 1
has a zero at ∞.

The proof for T proceeds along the same arguments.

By Theorem 2.3, S is proper iff 1 + L(∞) 6= 0; similarly (and since stability includes
properness as a requirement), S is stable iff 1 + L(λ) 6= 0 for all λ ∈ C0 ∪ C+ ∪ {∞}.

2.2 The Nyquist Stability Criterion

We are interested in stability of (1 + L)−1 if L is a given loop transfer function; in

this section we proivde a classical graphical test for this property, the so-called Nyquist
stability criterion. In its essence, it is a graphical test for counting the number of zeros
of 1 + L that are encircled by a Nyquist contour of L.

For this purpose we consider the grey oriented curve Γ as depicted in Figure 2 (with
small semicircles of radius r > 0 avoiding poles of L on the imaginary axis and one large
semicircle of radius R > 0 in order to capture all poles of L in C= ∪ C> ); Let Ω denote
the open set that is encircled by Γ once in clockwise direction.

Definition 2.4 The curve Γ is called a Nyquist contour for the transfer function L if
it does not pass through any pole of L and if the set of poles of L in Ω is equal to the set
of poles of L in C0 ∪ C+ .

18
 Im

R
r

r Re
Ω
r
Γ

Figure 2: Nyquist contour Γ and encircled region Ω; crosses indicate poles of L in C= ∪C> .

If Γ does not pass through any pole of L, the remaining properties for Γ to qualify as a
Nyquist contour can also be expressed as follows:

• All poles of L in C= ∪ C> are encircled by Γ exactly once in clockwise direction.

(R is so large to ensure that no poles of L in C= ∪ C> are located outside Γ ∪ Ω).

• No pole of L in C< is encircled by Γ. (r is sufficiently small such that all poles of

L in C< are located outside Γ ∪ Ω.)

Since a transfer function L has only finitely many poles in C, we can always choose r > 0
small enough and R > 0 large enough in order to make sure that Γ is a Nyquist contour
for L. Then increasing R > 0 and/or decreasing r > 0 does not change this property.

The Nyquist plot of L is just defined as the image of a chosen Nyquist contour Γ under
L, with the orientation inherited from Γ:

L(Γ) = {L(λ) : λ ∈ Γ}.

Theorem 2.5 (Nyquist stability criterion). Let L be a proper transfer function and
Γ a Nyquist contour for L such that the Nyquist plot of L does not pass through −1 and
L(λ) 6= −1 for λ = ∞ as well as for all λ ∈ C= ∪C> with |λ| > R. Suppose that L has n0+
poles on the imaginary axis or in the open right-half plane (counted with multiplicities).
Then (1 + L)−1 is stable iff the Nyquist plot of L encircles −1 exactly n0+ times in the
counterclockwise direction.

Remark. We will see that the Nyquist plot of L always encircles −1 at most n0+ times in
the counterclockwise direction, irrespective of whether (1 + L)−1 is stable or not. Stability
of (1 + L)−1 is thus equivalent to the fact that the Nyquist plot of L encircles −1 at least
n0+ times in the counterclockwise direction.

19
Proof. As abbreviations let us denote by ZA (g), PA (g) the number of zeros, poles of
the function g in the set A ⊂ C. We prepare the proof with the following observations
involving the function f := L + 1:

• L and f have identical poles in C. Hence PΩ (L) = PΩ (f ).

• The curve f (Γ) is obtained from the Nyquist plot L(Γ) through a shift by 1 (to the
right) in the complex plane. Hence f (Γ) does not pass through zero, i.e., ZΓ (f ) = 0.

• By the assumption on L, f (λ) 6= 0 for λ = ∞ and λ ∈ {λ ∈ C= ∪ C> : |λ| ≥ R};

the latter set is the complement of Ω in C= ∪ C> . Therefore, (1 + L)−1 is stable iff

ZΩ (f ) = 0. (2.1)

• Since Γ is a Nyquist contour for L we have PΩ (L) = PC0 ∪C+ (L) = n0+ . Due to the
first bullet this implies PΩ (f ) = n0+ .

For the Nyquist plot of L, let N now denote the number of encirclements of −1 in the
counterclockwise direction. Then f (Γ) encircles 0 exactly −N times in clockwise direction.
By the principle of the argument we infer that −N = ZΩ (f ) − PΩ (f ) and thus

N = n0+ − ZΩ (f ). (2.2)

(This implies that N ≤ n0+ always holds as emphasized in the remark after the theorem.)

Proof of “if”. By hypothesis N = n0+ . Hence ZΩ (f ) = 0 and thus (1 + L)−1 is stable.

(This already follows from N ≥ n0+ which proves the statement in the remark.)

Proof of “only if”. If (1 + L)−1 is stable we have (2.1). Then (2.2) shows N = n0+ .

Remarks.

• The result can also be applied if L is an irrational meromorphic function (i.e. the
quotient of two functions that are analytic on C). As a typically example, it can be
used for loop transfer functions L(s) = e−sT H(s) with T > 0 and a transfer function
H, as emerging in feedback loops with a delay.

• For transfer functions L one could set R = ∞: The Nyquist contour then just
consists of the whole imaginary axis (still with the small semicircles to avoid poles
of L on the axis and leaving all poles of L in C< to the left); the Nyquist plot of L
then just consists of the image under L in union with L(∞); moreover, in the first
hypotheses of Theorem 2.5, no λ ∈ C= ∪ C> with |λ| > R exist and it only remains
to check L(∞) 6= −1.

20
 Nyquist plot of L (blue) and corresponding Nyquist contour (red)
6

−2

−4

−6
−10 −5 0 5

Figure 3: Nyquist plot of L for the satellite example. For the Nyquist contour we choose
r = 0.1 and R = 5.

Example 2.6 The transfer function of a flexible satellite is given by

0.036(s + 25.28)
G(s) = .
s2 (s2 + 0.0396s + 1)

A controller which renders S = (1 + GK)−1 stable is given by

7.9212(s + 0.1818)(s2 − 0.2244s + 0.8981)

K(s) = 2 .
(s + 3.899s + 4.745)(s2 + 1.039s + 3.395)

Recall the standard construction of stabilizing controllers based on the separation prin-
ciple. One takes a minimal realization (A, B, C, D) of G and chooses F and J such that
A + BF and A + JC are Hurwitz. Then (A + BF + JC + JDF, J, F, 0) is a realization
of a controller K which renders (1 + GK)−1 stable. Also recall that F and J can, for
example, be computed by solving an LQR and a Kalman filter problem.

In our case the loop gain L(s) equals

0.28516(s + 25.28)(s + 0.1818)(s2 − 0.2244s + 0.8981)

.
s2 (s2 + 3.899s + 4.745)(s2 + 0.0396s + 1.001)(s2 + 1.039s + 3.395)

L is strictly proper, i.e., L(∞) = 0 6= −1. Therefore (1 + L)−1 is proper. Moreover, L has
s = 0 with multiplicity two as the only pole in the closed right half-plane.

The Nyquist plot for R = 5 and r = 0.1 of L in Figure 3 encircles −1 two times in the
counterclockwise direction. The dashed lines correspond to negative frequencies. Hence
Theorem 2.5 implies that (1 + L)−1 has all its poles in C< .

21
Indeed, (1 + L(s))−1 equals

s2 (s2 + 3.899s + 4.745)(s2 + 0.0396s + 1.001)(s2 + 1.039s + 3.395)

.
(s2 + 1.322s + 0.606)(s2 + 1.77s + 0.9899)(s2 + 1.217s + 1.509)(s2 + 0.6677s + 1.3)

2.3 Classical Robustness Indicators

We now make a first step towards considering robustness of feedback systems. Suppose
that (1 + L)−1 is stable. If the loop transfer function changes from L to kL for k ∈ R,
we can ask ourselves whether the uncertain system (1 + kL)−1 remains stable for all
uncertainties k ∈ (k− , k+ ) for constants 0 ≤ k− < 1 < k+ ≤ ∞. Classically, the largest
1 < k+ ≤ ∞ and the smallest 0 ≤ k− < 1 such that (1+kL)−1 is stable for all k ∈ (k− , k+ )
are called the upper and lower gain-margins, respectively.

Similar the largest 0 < φ ≤ π such that (1 + cL)−1 is stable for all c with |c| = 1 and
arg(c) ∈ (−φ, φ) is called the phase-margin.

By the Nyquist stability criterion and corresponding remarks, these margins can be read
off from those points where the Nyquist plot of L crosses the negative real axis/the unit
circle. If changing L into kL for real k > 0 or into cL for complex |c| = 1, the effect onto
the Nyquist plot can be easily derived from the well-known relation

|cL(iω)| = |c||L(iω)| and arg(cL(iω)) = arg(c) + arg(L(iω)) for any c ∈ C,

with the usual interpretation of the equation on the right. Roughly speaking, if continu-
ously deforming L into cL, stability will be lost at that value of c for which the Nyquist
plot of cL happens to pass through −1 at some frequency ω ∈ R, since then (1 + cL)−1
has a pole at iω. For the gain-margins, we thus have to consider those frequencies for
which the Nyquist plot of L crosses the negative real axis, while for the phase-margin the
crossing points with the unit circle are relevant, as illustrated in in Figure 4.

Example 2.7 Recall the transfer functions from Example 2.6 and the Nyquist plot in
Figure 3, a zoomed-in version of which is given in Figure 5. The relevant crossing point on
the real axis is x1 ≈ −0.54, since the one at x2 ≈ −7.9 in Figure 3 moves to −∞ for r → 0
and is not relevant. Hence the gain-margins are equal to k− = 0 and k+ ≈ −1/(−0.54) ≈
1.85. Similar we can read off the crossing point of the Nyquist plot with the unit circle
which is closest to −1 in order to obtain the phase-margin 360 arctan(0.53/0.85)
2π
≈ 31.9 degrees.
With the command allmargin(L) in Matlab one extracts the complete information about
the points of crossings and the corresponding frequencies.

The relevance of phase-perturbations for loop transfer functions is motivated by consid-

22
 Im Im
−1
Re
−1 −1
−k− −k+
1
Re

Figure 4: Graphical illustration of gain-margins and phase-margin.

Nyquist plot of L (dashed for negative frequencies)

1

−0.53

−1
−0.85 −0.54 0 1

Figure 5: Nyquist plot with unit circle (dotted) and relevant crossing points.

23
 r e u y
+ K G
−
yd
dT

Figure 6: Standard tracking configuration with time-delay.

ering time-delays in feedback loops. A time-delay dT of T seconds takes u(·) into


 0 for 0 ≤ t < T,
y(t) =
 u(t − T ) for t≥T

if t is measured in seconds. Now recall that

Z ∞ Z ∞ Z ∞
−st −st
ŷ(s) = e y(t) dt = e u(t − T ) dt = e−s(τ +T ) u(τ ) dτ =
0 T 0
Z ∞
−sT
=e e−sτ u(τ ) dτ = e−sT û(s).
0

Hence the transfer function of a time-delay of T seconds is given by e−sT .

Even if such an element, which is not a real rational function, appears in a feedback-
loop, the Nyquist stability criterion stays valid! To model a delay in the measurement
of the system output by T seconds we consider the interconnection in Figure 6, which
shows that we have to now consider the loop transfer function L̃T (s) = e−sT L(s) with
L(s) = G(s)K(s) to analyze stability. If starting from some stable (1 + L)−1 , we can ask
when stability is lost if continuously increasing the delay time T .

This motivates to define the time-delay-margin Tm , which is the smallest T ≥ 0 for

which there exists a frequency ω ≥ 0 with

e−iωT L(iω) = −1;

if no such T exists we set Tm = ∞. Suppose that (1 + L)−1 is stable and L(∞) = 0.

An application of the Nyquist criterion allows to conclude that (1 + e−sT L(s))−1 has all
its poles in C< for all T ∈ [0, Tm ), which means that the loop with all these delays stays
stable.

Note that since |e−iωT | = 1 for all ω ∈ R and all T ≥ 0, there is a close relationship between
the phase and time-delay margin. Assuming that |L(iω)| = 1 only for finitely many
ω ∈ R ∪ {∞}, one can compute the time-delay-margin as follows. Compute ω1 , . . . , ωn
where the Nyquist plot of L crosses the unit circle. Then solve e−iωj Tj L(iωj ) = −1 for
Tj ∈ [0, 2π] and all 1 ≤ j ≤ n. Then the time-delay-margin equals inf 1≤j≤n Tj if n ≥ 1. If
the Nyquist plot of L does not cross the unit circle, the time-delay-margin equals ∞.

24
 Nyquist plot of exp(−sT)L(s) for various delay times
3

Imaginary Axis
1

−1

−2

−3
−3 −2 −1 0 1 2 3
Real Axis

Step response of error for various delay times

2

1
Amplitude

−1

−2
0 10 20 30 40 50 60 70 80 90 100
Time (seconds)

Figure 7: Responses of the system from Example 2.6 with time-delay T ∈ {0, 0.4, 0.9} in
the loop.

Example 2.8 The time-delay-margin for the transfer functions in Example 2.6 equals
2π 60.93◦
Tm ≈ 0.98 s ≈ .
360 1.085 rad
s

Mind the units of the frequency in this computation! Some responses of the system
affected by time-delay T ∈ {0, 0.4, 0.9} are given in Figure 7.

In practice, gain-variations and time-delays might occur simultaneously. It is hence no

longer sufficient to only consider real or complex unitary values of c if perturbing L to
cL. Common perturbations are modeled by complex numbers c, whose size is measured
in terms of their absolute value.

In view of the discussion so far, it is clear that the distance of the Nyquist plot of L from
the point −1 is the correct stability margin; if the transfer function L has no poles in C= ,
this distance (for a Nyquist contour with r = 0 and R = ∞) just equals

inf |L(iω) + 1|.

ω∈R

This distance is related to the sensitivity transfer function as follows.

Theorem 2.9 For real-rational L without poles in C= ∪ {∞} we have

 −1
1
inf |L(iω) + 1| = sup .
ω≥0 ω≥0 1 + L(iω)

25
 Nyquist plot of L
1

0.8

0.6

0.4

0.2

−0.2

−0.4

−0.6

−0.8

−1
−2.5 −2 −1.5 −1 −0.5 0

Figure 8: Nyquist plot of L in Example 2.10, unit circle and circle around −1 with radius
inf ω∈R |L(iω) + 1|.

This is a triviality. For a feedback-loop such that (1 + L)−1 is stable, it says that the
peak-value of the sensitivity transfer function

kSk∞

is inversely proportional to the distance of the Nyquist plot to −1. Roughly speaking,
the smaller kSk∞ is, the larger are the allowed common variations in gain and phase
(measured by |c| if replacing L with cL) without violating stability of the loop.

Example 2.10 Consider the Nyquist plot of

−0.46832(s + 3.3)(s − 2)(s + 0.55)(s2 + 0.8824s + 0.5882)

L(s) =
(s + 0.303)(s − 0.5)(s + 1.818)(s2 + 1.5s + 1.7)

which is depicted in Figure 8. One can see that the gain- and phase margins are much
larger than the distance of the Nyquist plot to −1. Hence the gain- and phase-margins
provide a wrong guideline for robust stability if considering commonly occurring variations
in gain and phase.

Up to this point we were rather sloppy in discussing robust stability and we did not
give complete proofs for our statements. The remaining parts of these notes serve to
develop a systematic theoretical basis for robustness questions that is also applicable to
multivariable systems.

26
2.4 Plant Uncertainty

The open-loop plant G is obtained by physical or experimental modeling. If the actual

physical system is described by H, we are typically confronted with a plant-model mis-
match, which just means G 6= H. If we design a controller for G, it should as well do a
good job for H, even if H deviates from G not only slightly but even significantly.

The general philosophy of robust control can be expressed as follows: Instead of designing
good controllers for just one model G of H, one rather considers a whole set of models H
with H ∈ H and designs a controller which does a good job for all elements in H.

A mathematical description of H is a so-called uncertainty model. Whether or not H ∈ H

is true, is a question of validating the uncertainty model and part of the field of system
identification.

Concretely, let us measure the deviation of G and H at some frequency ω with G(iω) 6= 0
by the relative error of their frequency responses:
|H(iω) − G(iω)| H(iω)
= −1 .
|G(iω)| G(iω)
In practice, this error is typically small at low frequency and large at high frequencies.
The variation of the size of this error over frequency is captured with a transfer function
W by requiring
H(iω)
− 1 < |W (iω)| if W (iω) 6= 0 and H(iω) = G(iω) otherwise.
G(iω)
Note that this holds iff there exists some ∆ω ∈ C with

H(iω) = G(iω)(1 + W (iω)∆ω ) and |∆ω | < 1.

Here G is the nominal system and W is a so-called uncertainty weight which captures
the size of the deviation of H(iω) from G(iω). This is said to be a multiplicative
uncertainty model.

The mathematically precise definition is based on a nominal model described by a

transfer function G 6= 0 and a weight described by a stable transfer function W 6= 0.

Definition 2.11 The multiplicative uncertainty model H related to a transfer function

G and a stable transfer function W is the set of all transfer functions H such that

1) H and G have the same number of poles in C> ,

2) H and G have identical poles on the imaginary axis C= and

3) |H(iω)/G(iω) − 1| < |W (iω)| for all ω ∈ R ∪ {∞} with W (iω) 6= 0.

27
 Bode Diagram
20

Magnitude (dB)
−20

−40

−60

−80

−100
−2 −1 0 1 2
10 10 10 10 10
Frequency (rad/s)

Figure 9: Graphical verification of 3) in Definition 2.11.

We emphasize that H/G does not have poles in C= ∪ {∞} due to 3). Also note that, in
practice, W typically is a high-pass filter.

Example 2.12 This continues Example 2.6 and 2.7. Suppose that some parameter in
our satellite model is not known exactly, and that is transfer function is actually given by
0.036(s + 25.28)
Gb (s) = for some b ∈ ( 21 b0 , 2b0 ) with b0 = 0.0396.
s2 (s2 + bs + 1)
Choose W (s) = 1.1s/(s + 0.5) to obtain
Gb (iω)
− 1 < |W (iω)| for all ω ∈ R with ω 6= 0,
Gb0 (iω)
as verified graphically through the plot in Figure 9. With Gb0 as the nominal model and
W as the uncertainty weight, define the uncertainty model H. Since all Gb as described
above have s = 0 as a double pole in C= ∪ C> we indeed infer Gb ∈ H.

It is emphasized that H contains many more models. In that sense we “cover” or “over-
bound” the actual uncertainty in the parameter b through an uncertainty model that also
comprises dynamic uncertainties. All models
 
0.036(s + 25.28) 1.1s
Gb0 (s)(1 + W (s)∆(s)) = 2 2 1+ ∆(s)
s (s + 0.0396s + 1) s + .5
for arbitrary ∆ ∈ RH∞ satisfying k∆k∞ < 1 belong to H as well:

Since ∆ cannot add poles in C> property 1) is still satisfied. Since 1 + W (0)∆(0) = 1,
we also observe that the double pole of Gb0 at s = 0 is never canceled; hence 2) holds.
The property 3) is satisfied because k∆k∞ < 1 implies |W (iω)∆(iω)| < |W (iω)| for
ω ∈ R ∪ {∞} with ω 6= 0.

For the designed controller K and models H ∈ H, samples of the Bode magnitude and
step response plots are depicted in Figure 10. We conclude that K does not robustly

28
 Sensitivity S (blue) and control sensitivity KS (green)

50

Magnitude (dB)
−50

−100

−3 −2 −1 0 1 2 3
10 10 10 10 10 10 10
Frequency (rad/s)

Error (blue) and control (green)

6

2
Amplitude

−2

−4

−6

−8
0 10 20 30 40 50 60 70 80 90 100
Time (seconds)

Figure 10: Samples of the Bode magnitude and step response.

stabilize H. This just means there exists some H ∈ H for which (1 + HK)−1 is not stable.

Theorem 2.13 Let H be given as in Definition 2.11. Suppose the transfer function
K does not cause any pole-zero cancellation in C= ∪ C> if forming GK and renders
(1 + GK)−1 proper and stable. Then

(1 + HK)−1 exists and is proper and stable for all H ∈ H (2.3)

if and only if
kW GK(1 + GK)−1 k∞ ≤ 1. (2.4)

The property (2.3) is the robust stability question with respect to the multiplicative
uncertainty model H under consideration.

The condition (2.4) provides an exact verifiable robust stability test: For the nominal
model G, verify whether all unstable poles of G and K appear in L = GK and whether
(1 + L)−1 is proper and stable. If this is the case, then just determine the complementary
sensitivity transfer function T = L(1 + L)−1 = GK(1 + GK)−1 and check kW T k∞ ≤ 1.

With L = GK we note that (2.4) reads as kW L(1 + L)−1 k∞ ≤ 1; if L has no poles in

C= then this condition has also a nice graphical interpretation. For this purpose observe

29
 Im
−1 Re

|W (iω)L(iω)|
L(iω)

Figure 11: Graphical interpretation of (2.4).

that (2.4) can be expressed as

|W (iω)L(iω)| ≤ | − 1 − L(iω)| for all ω ∈ R ∪ {∞}. (2.5)

For each frequency, this means that the distance of L(iω) to −1 is not smaller than
|W (iω)L(iω)|. We can equivalently say that the open disc Dω = {λ ∈ C| |λ − L(iω)| <
|W (iω)L(iω)|} does not contain −1. This is depicted in Figure 11

Proof. “⇐”: Choose any H ∈ H and define ∆ := W1 ( H

G
− 1). With 3) in Definition 2.11
we infer that
|∆(iω)| < 1 for all ω ∈ R ∪ {∞}. (2.6)
Hence ∆ is a transfer function without any poles in C0 . Now define

Lτ := (1 + τ ∆W )GK for τ ∈ [0, 1] .

Since 1 + ∆W = H/G we infer L0 = GK and L1 = HK. We know that (1 + L0 )−1 is

stable and we need to show stability of (1 + L1 )−1 . This will be done by applying the
Nyquist stability criterion and on the basis of the following key relation: By the very
definition we have Lτ + 1 = (L0 + 1) + τ ∆W L0 and thus

S(Lτ + 1) = 1 + τ ∆W T. (2.7)

for the stable transfer functions S = (1 + L0 )−1 and T = 1 − S = L0 (1 + L0 )−1 .

Let us choose R > 0, r > 0 defining a Nyquist contour Γ for G, H, K, W and ∆; so all
poles of these transfer functions in C= ∪ C> and none in C< are encircled by Γ. Note that
Γ is also a Nyquist contour for all Lτ with τ ∈ [0, 1]. We recall again that these properites
are not altered by increasing R > 0 and decreasing r > 0.

With (2.6) and kW T k∞ ≤ 1 we infer |∆(λ)W (λ)T (λ)| < 1 for all λ ∈ C= ∪ {∞}; if
applied for λ = ∞, we see that we can increase the given R > 0 in order to also ensure

|∆(λ)W (λ)T (λ)| < 1 for all λ ∈ C= ∪ C> with |λ| ≥ R; (2.8)

30
for reasons of continuity, we can decrease r > 0 to also guarantee |∆(λ)W (λ)T (λ)| < 1
for all points λ ∈ Γ on the indented imaginary axis; taken together this implies

|∆(λ)W (λ)T (λ)| < 1 for all λ ∈ Γ. (2.9)

After having fixed this Nyquist contour Γ for Lτ , we conclude from (2.7) and (2.9) that

Lτ (λ) 6= −1 for all λ ∈ Γ and all τ ∈ [0, 1]; (2.10)

otherwise we had 0 = |1+τ ∆(λ)W (λ)T (λ)| ≥ 1−τ |∆(λ)W (λ)T (λ)| > 0, a contradiction.
Let us now denote by N (Lτ ) the number of counterclockwise encirclements of −1 of Lτ (Γ).
Then (2.10) shows
N (L0 ) = N (L1 ). (2.11)
Indeed, let γ : [0, 1] → Γ be a (bijective piece-wise continuous differentiable) parametriza-
tion of Γ such that γ(t) moves in clockwise direction around Γ if t moves from zero to one.
It is then very easy to check that the map H : [0, 1] × [0, 1] 3 (τ, t) → Lτ (γ(t)) ∈ C
is continuous: Choose any (τ0 , t0 ) ∈ [0, 1] × [0, 1]; since Γ is compact and does not
pass through any pole of ∆, W and L0 = GK, there exists some constant b with
|∆(γ(t))W (γ(t))L0 (γ(t))| ≤ b for all t ∈ [0, 1]; for (τ, t) ∈ [0, 1] × [0, 1] we get

|Lτ (γ(t)) − Lτ0 (γ(t0 ))| ≤ |Lτ (γ(t)) − Lτ0 (γ(t))| + |Lτ0 (γ(t)) − Lτ0 (γ(t0 ))| ≤
≤ |τ − τ0 ||∆(γ(t))W (γ(t))L0 (γ(t))| + |Lτ0 (γ(t)) − Lτ0 (γ(t0 ))| ≤
≤ |τ − τ0 |b + |Lτ0 (γ(t)) − Lτ0 (γ(t0 ))| → 0 for (τ, t) → (τ0 , t0 ).

As a well-known fact in complex analysis, that already this fact allow us to conclude that
N (Lτ ) is constant for τ ∈ [0, 1], which in turn clearly leads to (2.11).

We are now ready to conclude the proof with the Nyquist stability criterion. First, since
Γ is a Nyquist contour for L0 , L1 , their Nyquist plots do not pass through −1. Second,
from (2.7) and (2.8), we infer Lτ (λ) 6= −1 for all λ ∈ C= ∪ C> with |λ| > R and τ = 0, 1.
If n0+ (g) denotes the number of poles of g in C0 ∪ C+ , Theorem 2.5 applied to L0 shows

N (L0 ) = n0+ (L0 ).

By the non-cancellation hypothesis and 1), 2) in Definition 2.11 we conclude

n0+ (L0 ) = n0+ (G) + n0+ (K) = n0+ (H) + n0+ (K) ≥ n0+ (L1 ). (2.12)

Combined with (2.11) we infer N (L1 ) ≥ n0+ (L1 ) which concludes the proof.

“⇒”: Set M := W T and suppose kM k∞ > 1. Then there exists some ω0 ∈ [0, ∞] with
|M (iω)| > 1. By continuity, we can make sure that iω0 is not a pole of G and K. If
1
defining δ0 := − M (iω 0)
we clearly have |δ0 | < 1. If δ0 is real set ∆(s) := δ0 and otherwise

31
construct a stable transfer function ∆ as in Lemma 2.14. We infer 1 + M (iω0 )∆(iω0 ) = 0.
With H := G(1 + W ∆) we obtain

1 + HK = (1 + GK)(1 + (1 + GK)−1 GKW ∆) = (1 + GK)(1 + M ∆).

Since 1 + GK has no pole at iω0 , we conclude that 1 + HK has a zero at iω0 . Hence,
(1 + HK)−1 does not exist, is not proper or is unstable. Moreover, for all ω ∈ R ∪ {∞}
with W (iω) 6= 0 we have

H(iω)
− 1 = |∆(iω)||W (iω)| ≤ |δ0 ||W (iω)| < |W (iω)|.
G(iω)

In this fashion we have constructed some H that is “almost” contained in H and which
destabilizes (1 + HK)−1 . As the only trouble, poles of G in C= ∪ C> might be canceled
in the product H = (1 + ∆W )G such that 1), 2) in Definition 2.11 are not valid.

To overcome this trouble let p1 , . . . , pm ∈ C= ∪C> be those poles of G at which cancelation

takes place; note that pj 6= ±iω0 . Since 1 + W ∆ vanishes at these points, W (pj ) 6= 0 for
j = 1, . . . , m. Lemma 2.15 shows that there exists a stable transfer function F satisfying
1
F (iω0 ) = 1, |F (pj )| < for j = 1, . . . , m and kF k∞ ≤ 1.
|W (pj )|

If we define H̃ = (1 + [F ∆]W )G, then all properties above persist to hold; in addition,
however, we infer

|1 + F (pj )∆(pj )W (pj )| ≥ 1 − |δ0 ||F (pj )||W (pj )| > 0 for j = 1, . . . , m.

Hence p1 , . . . , pm are not canceled any more in (1 + F ∆W )G and thus H̃ ∈ H.

In the proof of Theorem 2.13 the following two lemmata were used.

Lemma 2.14 Let ω0 > 0 and δ0 ∈ C. Choose

α − δ0
α = ±|δ0 |, β = iω0 .
α + δ0
Then the transfer function
s−β
∆(s) = α
s+β
is real-rational, proper and satisfies

∆(iω0 ) = δ0 as well as |∆(iω)| = |δ0 | for all ω ∈ R.

Either for α = |δ0 | or for α = −|δ0 | the transfer function ∆ is stable.

32
Proof. One can prove ∆0 = ∆(iω0 ) by direct calculations. Since |α| = |∆0 |, the vectors
that correspond to the complex numbers α + ∆0 and α − ∆0 are perpendicular. Hence
α−∆0
α+∆0
is purely imaginary. This implies that β is real. Therefore, the distances of iω to β
and to −β are identical such that |iω − β| = |iω + β| what implies |∆(iω)| = |α| = |∆0 |.
Moreover, a change of sign of α leads to the reciprocal of α−∆
α+∆0
0
which, again due to the
fact that this number is purely imaginary, changes the sign of the imaginary part. Hence
we can adjust the sign of α to render β non-negative. Then ∆ is stable. (Note that β
might vanish what causes no problem!)

Lemma 2.15 Let ω0 > 0, p1 , . . . , pm ∈ (C0 ∪ C+ )\{−iω0 , iω0 } located symmetrically with
respect to the real axis and positive real values α1 , . . . , αm be given. Then there exists a
stable transfer function F such that

F (iω) = 1, kF k∞ ≤ 1 and |F (pj )| < αj for j = 1, . . . , m.

Proof. For ξ > 0 and ω > 0 define the stable transfer function
2ξωs
Gξ (s) = .
s2 + 2ξωs + ω 2

Then observe that Gξ (iω) = 1, |Gξ (s)| < 1 for all s ∈ C0 ∪ C+ with s 6= ±iω and
2ξωs
s2 +2ξωs+ω 2
→ 0 for ξ → 0 for fixed s with s 6= ±iω (in particular for s = p1 , . . . , pm ).
Hence there exist ξ1 , . . . , ξm such that |Gξj (pj )| < αj for all 1 ≤ j ≤ m, and this persists to
hold if decreasing ξj > 0. Then F := Gξ̃ with ξ˜ = inf 1≤j≤m ξj has the desired properties.

NG NK
Remark 2.16 Let us represent G = DG
and K = DK
with coprime real numerator and
denominator polynomials.

• Then S and T can be expressed as

DG DK NG NK
S= and T = .
NG NK + DG DK NG NK + DG DK
Here NG NK + DG DK is the so-called characteristic polynomial of the feedback
loop. Recall that S (and/or T ) is stable iff NG NK + DG DK has only zeros in C< .

• The non-cancellation hypothesis means that NG , DK and DG , NK have no common

zeros in C= ∪ C> . The relation of “non-cancellation” and internal stabilization will
be clarified in Theorem 2.24.

33
 • Under the hypotheses of Theorem 2.13, the proof shows that (2.4) also guarantees
the absence of any pole-zero cancellations in C= ∪ C> for HK and any H ∈ H!

Example 2.17 Consider

1 1 1.1
G(s) = , H(s) = and W (s) = .
s−1 s−2 s+2
Then  
1 H(s) s+2
∆(s) = −1 =
W (s) G(s) 1.1(s − 2)
satisfies |∆(iω)| < 1 for all ω ∈ R ∪ {∞}. Hence there exists a uncertainty model H with
H ∈ H. The controller K ∈ R (a static gain) renders 1/(1 + GK) stable iff K + (s − 1)
has all its roots in C< iff K > 1. For K = 3 we get

3.3
kW T k∞ = = 0.825 ≤ 1.
(s + 2)2 ∞

Hence K is guaranteed to render 1/(1 + HK) stable by Theorem 2.13. Is that also true
for 1/(1 + Hτ K) with
Hτ = (1 + τ ∆W )G for τ ∈ (0, 1)?
Although one might be tempted to draw this conclusion, the answer is in general no. To see
this, observe that 1/(1+Hτ K) = (1+Lτ )−1 with Lτ = (1+τ ∆W )GK = (1−τ )GK +τ HK
defined as in the proof of Theorem 2.13. If τ 6= 0 and τ 6= 1 we can no longer guarantee
that n0+ (L0 ) ≥ n0+ (Lτ ) is true. It might very well happen that n0+ (L0 ) < n0+ (Lτ ) holds;
with the Nyquist contour and the notation in the proof of Theorem 2.13 we get

N (Lτ ) = N (L0 ) = n0+ (L0 ) < n0+ (Lτ ),

which actually implies instability of (1 + Lτ )−1 by Theorem 2.5.

Example 2.18 For the uncertainty model and transfer matrices in Example 2.12 and the
earlier designed controller we have kW T k∞ > 1.5 > 1, which confirms non-robustness as
seen by our simulations. If we reduce the size of the uncertainty by replacing the weight
through W̃ (s) = 0.8s/(s + 0.5), the simulations in Figure 12 seem to confirm robust
stability. However, this is misleading since kW̃ T k∞ ≈ 1.09 > 1 ensures that there exists
some system in the uncertainty model for which the sensitivity is not stable.

In general, the size of the uncertainty in our general uncertainty description can be scaled
with a factor r > 0 by replacing 3) in Definition 2.11 through

|H(iω)/G(iω) − 1| < r|W (iω)| for ω ∈ R ∪ {∞}, W (iω) 6= 0.

34
 Error (blue) and control (green)

From: In(1) To: [+sys(1)−r]

2

Amplitude
0

−1

−2
0 10 20 30 40 50 60 70 80 90 100
Time (seconds)

Figure 12: Sample plots for the rescaled uncertainty model Hr in Example 2.18.

For the corresponding uncertainty model Hr and under the hypothesis of Theorem 2.13
we then infer that

(1 + HK)−1 exists and is proper and stable for all H ∈ Hr (2.13)

if and only if kW GK(1 + GK)−1 k∞ ≤ 1r . This gives a formula for the corresponding
robust stability margin, the largest value of r for which (2.13) is valid.

Corollary 2.19 Under the hypotheses of Theorem 2.13, kW GK(1 + GK)−1 k−1
∞ is the
maximal r > 0 for which (2.13) is satisfied.

2.5 Performance

Consider again the interconnection depicted in Figure 1. Let S = (1 + L)−1 be stable

with L = GK and let (A, B, C, D) be a minimal realization of S. For a reference signal
with constant value r, the error response equals

e(t) = CeAt A−1 Br + (D − CA−1 B)r

and has the steady-state response (D − CA−1 B)r = S(0)r, since the transient response
CeAt A−1 Br converges to 0 for t → ∞ (just because S is stable, A is Hurwitz).

If r(t) = r0 eiωt is a complex sinusoidal signal with r0 ∈ C and ω > 0 we get

e(t) = CeAt (A − iωI)−1 Br0 + (C(iωI − A)−1 B + D)r0 eiωt .

Again, the transient response CeAt (A − iωI)−1 Br0 converges (exponentially fast) to 0 for
t → ∞ due to the stability of S and minimality of the realization. Hence the steady-state
response equals (C(iωI − A)−1 B + D)r0 eiωt = S(iω)r0 eiωt and is as well a sinusoidal signal
with the same frequency as r(.) and complex amplitude given by e0 = S(iω)r0 ; clearly
|S(iω)| is the amplification (or attenuation) factor for the amplitude of sinusoidal signals
of frequency ω in e = Sr. Hence |e(t)| ≈ |S(iω)||r0 | for sufficiently large t ≥ 0. Since

35
e is the tracking error for the configuration in Figure 1, sinusoidal reference signals of
frequency ω are tracked well if |S(iω)| is small. By
1 1 1
|S(iω)| = = 1 ,
|1 + G(iω)K(iω)| |K(iω)| | K(iω) + G(iω)|

this can be achieved with large values of |G(iω)| or of |K(iω)| if G(iω) 6= 0. Due to
u = KSr, the control action in the loop is related to |K(iω)S(iω)| = | K(iω) 1
+ G(iω)|−1 ;
this gain is large for large |K(iω)| and small |G(iω)|. Since |G(iω)| is often small for high
frequencies, large values of |K(iω)| imply
|G(iω)K(iω)| |G(iω)|
= 1 ≈ 1,
|1 + G(iω)K(iω)| | K(iω) + G(iω)|

which, roughly, means that the relative plant-model mismatch cannot go beyond 100% at
high frequencies in order not to endanger stability.

All this motivates that it is, in general, not a good design goal to try to reduce |S(iω)|
over all frequencies by control. Instead, in practice one chooses e.g. some frequency band
[ω1 , ω2 ] and tries to suppress |S(iω)| for ω ∈ [ω1 , ω2 ] only.

Mathematically, this is modeled with a performance weight Wp , a stable transfer func-

tion that allows to express the desired specification precisely as

kWp Sk∞ ≤ 1.

If this inequality holds, this simply means that

1
|S(iω)| ≤ for all ω ∈ R ∪ {∞}. (2.14)
|Wp (iω)|

If true, the tracking error is small “in the frequency range” [ω1 , ω2 ] if Wp had been taken
such that |Wp (iω)| is large at all these frequencies. In practice, Wp are band-pass filters.
To track constant references we choose ω1 = 0 and take Wp as a low-pass filter.

The inequality (2.14) has again a nice geometric interpretation. If iω is not a pole of L,
then (2.14) can be written as

|Wp (iω)| ≤ |1 + L(iω)| = | − 1 − L(iω)|.

This means that L(iω) should lie outside the open disk with center −1 and radius |Wp (iω)|
as depicted in Figure 13.

Example 2.20 This continues the Examples 2.6, 2.7 and 2.12. If we choose
√ !2 √ !2
s/ M + ωB s/ 5 + 0.1
Wp (s) = √ = √
s + ωB A s + 0.1 10−3

36
 Im
−1 Re
|Wp (iω)|

L(iω)

Figure 13: Geometric interpretation of (2.14).

Uncertain sensitivity S (blue) and inverse weight 1/Wp (red)
From: In(1) To: [+sys(1)−r]
30
20
10
Magnitude (dB)

0
−10
−20
−30
−40
−50
−60
−2 −1 0 1
10 10 10 10
Frequency (rad/s)

Figure 14: Samples of Bode magnitude plots.

one can check that kWp (1 + GK)−1 k ≤ 1. If H is given for the uncertainty weight W (s) =
0.7s/(s+0.5) we now ask whether the following robust performance specification holds:

kWp (1 + HK)−1 k∞ ≤ 1 for all H ∈ H.

It is easy to check kW T k∞ ≤ 1 which implies robust stability. However, the samples of

Bode magnitude plots in Figure 14 show that robust performance does not hold.

Theorem 2.21 Let Wp be a stable transfer function and let the hypothesis of Theorem
2.13 be satisfied. Then robust stability (2.3) and robust performance as characterized by

kWp (1 + HK)−1 k∞ ≤ 1 for all H ∈ H (2.15)

are satisfied if and only if

|Wp (iω)S(iω)| + |W (iω)T (iω)| ≤ 1 for all ω ∈ R ∪ {∞} (2.16)

holds with S = (1 + GK)−1 and T = GK(1 + GK)−1 .

37
Again, the essence is an easily verifiable characterization of robust stability and robust
performance in terms of the nominal system G, the controller K and the weights W and
Wp capturing the uncertainty and the performance specification. Note that (2.16) imply

kWp Sk∞ ≤ 1 and kW T k∞ ≤ 1,

i.e., nominal performance and robust stability. Also note that (2.16) cannot be expressed
as an H∞ -norm constraint; a sufficient condition for (2.16) can be expressed as
 
WS 1
 p  <√
WT 2
∞
.

Proof. “⇐”: (2.16) also implies |W (iω)T (iω)| ≤ 1 for all ω ∈ R ∪ {∞}. Since by
assumption (1 + GK)−1 and W are stable, this means kW T k∞ ≤ 1. Hence (1 + HK)−1
is stable for all H ∈ H by Theorem 2.13.

Suppose iω ∈ C0 is not a pole of G, K and Wp , W , 1/W . Let H ∈ H and define

∆ := W1 ( H
G
− 1). Due to 3) in Definition 2.11 we infer |∆(iω)| < 1. Also observe that
H = (∆W + 1)G. Moreover we conclude from (2.16) that

|Wp (iω)| + |W (iω)G(iω)K(iω)| ≤ |1 + G(iω)K(iω)|.

By using |∆(iω)| < 1 we obtain

|Wp (iω)| ≤ |1 + G(iω)K(iω)| − |W (iω)G(iω)K(iω)|

≤ |1 + G(iω)K(iω)| − |∆(iω)||W (iω)G(iω)K(iω)|
≤ ||1 + G(iω)K(iω)| − |∆(iω)W (iω)G(iω)K(iω)||
≤ ||1 + G(iω)K(iω) + ∆(iω)W (iω)G(iω)K(iω)||
≤ |1 + [1 + ∆(iω)W (iω)] G(iω)K(iω)|
= |1 + H(iω)K(iω)|

which shows |Wp (iω)(1 + H(iω)K(iω))−1 | ≤ 1. Due to stability of Wp and (1 + HK)−1

and continuity in ω, this inequality holds in fact for all ω ∈ R∪{∞}. Hence (2.15) follows.

“⇒”: If |Wp (iω̃)S(iω̃)|+|W (iω̃)T (iω̃)| > 1 for some ω̃ ∈ R∪{∞}, we can also find ω0 ∈ R
such that iω0 is not a pole of G and K and, still, |Wp (iω0 )S(iω0 )| + |W (iω0 )T (iω0 )| > 1.

If |W (iω0 )T (iω0 )| > 1, robust stability is not guaranteed by Theorem 2.13 and the proof
is finished. Otherwise we get |Wp (iω0 )S(iω0 )| > 1 − |W (iω0 )T (iω0 )| ≥ 0 and thus

|Wp (iω0 )| > |1 + L(iω0 )| − |W (iω0 )L(iω0 )| ≥ 0.

Lemma 2.22 for u = 1 + L(iω0 ) and v = W (iω0 )L(iω0 ) allows us to choose δ0 ∈ C with
|δ0 | < 1 and

|Wp (iω0 )| > |1 + L(iω0 ) + δ0 W (iω0 )L(iω0 )| = |1 + [1 + δ0 W (iω0 )]G(iω0 )K(iω0 )|. (2.17)

38
 Im
−1 Re
|Wp (iω)|

|W (iω)L(iω)|

L(iω)

Figure 15: Geometric interpretation of (2.16).

If constructing F ∆ ∈ RH∞ such that H := (1 + F ∆W )G ∈ H and F ∆(iω0 ) = δ0 as in

the proof of Theorem 2.13, we infer that iω0 is not a pole of H (since it was not a pole of
G) and |Wp (iω0 )| > |1 + H(iω0 )K(iω0 )| by (2.17). This shows kWp (1 + HK)−1 k∞ > 1.

Lemma 2.22 Let us be given α ∈ R, u, v ∈ C with α > |u| − |v| ≥ 0. Then there exist
δ0 ∈ C with |δ0 | < 1 such that α > |u + δ0 v|.

Proof. There exist φ, ψ ∈ [−π, π] with u = |u|eiφ and v = |v|eiψ . Define δ := ei(φ−ψ+π)
to obtain

|u + δv| = ||u|eiφ + |v|eiφ+iπ | = ||u|eiφ − |v|eiφ | = ||u| − |v|| = |u| − |v|.

Hence α > |u + δv|; since the inequality is strict, there exists some small  > 0 such that
α > |u + (1 − )δv| and we can choose δ0 := (1 − )δ.

If iω is not a pole of L, (2.16) can be written at this frequency as

|Wp (iω)| + |W (iω)L(iω)| ≤ | − 1 − L(iω)|.

This means that the distance of L(iω) to −1 is at least |Wp (iω)| + |W (iω)L(iω)|, i.e., the
two open disks related to the nominal performance specification and the robust stability
condition as shown in Figure 15 should not intersect.

2.6 Internal Stability

For transfer functions G and K let us now consider the feedback loop as depicted in
Figure 16. So far we have concentrated on controllers K which render the transfer function

39
 d
r e u y
+ K + G
−

Figure 16: Standard tracking configuration with disturbance d.

in e = (1 + GK)−1 r stable. If true, this does not always imply that the transfer functions
in u = K(1 + GK)−1 e or y = G(1 + GK)−1 d are stable. This motivates the following
definition.

Definition 2.23 K renders the feedback interconnection shwon in Figure 16 internally

stable if (1 + GK)−1 exists, is proper and if
1 K G
, ,
1 + GK 1 + GK 1 + GK
are all stable.

Let us emphasize the following two useful facts:

• Stability of (1 + GK)−1 implies stability of

GK(1 + GK)−1 = 1 − (1 + GK)−1 .

• If 1 + GK is not the zero transfer function we know that

 −1    
1 G 1 1 −G S −GS
  =  = .
−K 1 1 + GK K 1 KS S

The latter observation justifies the following alternative characterization of the conditions
in Definition 2.23: K renders the feedback interconnection internally stable iff
 −1
1 G
  exists and is stable.
−K 1

NG NK
Theorem 2.24 Suppose that G = D G
and K = D K
with coprime real numerator and
denominator polynomials. Then K renders the feedback interconnection internally stable
iff one of the following equivalent conditions holds:

1) There is no pole-zero cancellation in C= ∪ C> if forming GK and (1 + GK)−1 exists

and is stable.

40
  
1 G
2) The 2 × 2 transfer matrix   has a stable inverse.
−K 1

3) 1 + G(∞)K(∞) 6= 0 and the zeros of the characteristic polynomial NG NK + DG DK

of the loop are contained in C< .

Recall that “stable” includes the requirement of properness. We will not emphasize this
point in the sequel.

Proof. Define p := NG NK + DG DK . If 1 + GK 6= 0 then

 −1    
1 G S −GS D D −N D
  = = 1 G K G K
. (2.18)
−K 1 KS S p DG NK DG DK

1) ⇒ 3). Since S = (1 + GK)−1 is proper we get 1 + G(∞)K(∞) 6= 0. Suppose p(λ) = 0

for λ ∈ C0 ∪ C+ . Since S = DG DK /p is stable we infer DG (λ)DK (λ) = 0 and hence
NG (λ)NK (λ) = NG (λ)NK (λ) + DG (λ)DK (λ) = p(λ) = 0. Then either NG (λ) = 0 or
NK (λ) = 0.

If NG (λ) = 0 then DG (λ) 6= 0 since NG and DG are coprime. Hence DK (λ) = 0 is true.
This means that λ is a common zero of NG and DK that is cancelled when forming GK.
This is a contradiction to 1).

An analogous reasoning leads to a contradiction if NK (λ) = 0.

3) ⇒ 2). Since 1 + G(∞)K(∞) 6= 0, we have 1 + GK 6= 0 and the inverse in (2.18) exists

as a real rational function. Since S = (1 + GK)−1 , G and K are proper, the first equation
in (2.18) shows that the inverse is proper. Since p has only zeros in C< , the inverse is
stable as seen by the second equation in (2.18).

2) ⇒ 1). Since the inverse exists we infer 1 + GK 6= 0 as a function. Properness of the

inverse and (2.18) show that S = (1 + GK)−1 exists and is proper.

Let λ ∈ C= ∪ C> satisfy NG (λ) = 0 and DK (λ) = 0. On the one hand we clearly
have p(λ) = 0. On the other hand, we know DG (λ) 6= 0 (since NG , DG are coprime)
and NK (λ) 6= 0 (since NK , DK are coprime). Hence DG NK /p has a pole at λ, which
contradicts the stability of (2.18).

We infer that NG , DK have no common zeros in C= ∪ C> , and analogous arguments show
the same for DG , NK . We conclude that no C= ∪ C> pole-zero cancelation occurs in
GK.

41
With the concept of internal stabilization and in view of the hypothesis of Theorem 2.13
as well as the third bullet in Remark 2.16, we can rephrase Theorem 2.13 as follows: Sup-
pose K internally stabilizes the interconnection in Figure 16 with G. Then K internally
stabilizes this interconnection with any H ∈ H if and only if (2.4) holds true.

This sets the stage for the generalization to much more complicated general interconnec-
tions and structured uncertainties in the next chapters.

Exercises

1) Suppose you are given the system and controller

   
0 1 0
1 1
G =  −2 −3 1  and K =   with α = 1.
 
1 0
α −1 0

a) Compute GK and S = (1 + GK)−1 in the state-space by standard formulas

for the product and the inverse of realizations.
b) Is the resulting system-matrix Hurwitz?
c) Does K render S stable? Is GS stable? Is KS stable?
d) What is the trouble if α slightly deviates from 1?

2) For given transfer functions G and K let us consider the feedback interconnection
in Figure 16.
a) If GK is not equal to −1, show!
that the transfer matrix
! from the the reference
r e
and the input disturbance to the signal is given by
d u
 −1
1 G
  . (2.19)
−K 1

Express the inverse in terms of S, G and K.

b) For a given SISO system G, the SISO controller K is said to internally stabilize
the standard feedback configuration of Figure 16 if (2.19) is stable. Construct
an example such that K renders (2.19) stable except for the (1; 2)- entry.
c) Suppose G is stable. Then show that K internally stabilizes the tracking
configuration iff K renders K(1 + GK)−1 is stable.

3) Let L be a proper transfer function and Γ a Nyquist contour fo L with R = ∞ such

that the Nyquist plot of L does not pass through −1 and L(∞) 6= −1. Suppose

42
 that L has n0+ poles on the imaginary axis or in the open right-half plane (counted
with multiplicities). Show that (1 + L)−1 is stable if and only if the Nyquist plot of
L encircles −1 exactly n0+ times in the counterclockwise direction.

4) In 195x NASA launched a flexible satellite in order to spy on the Russians. Based
on the linear model G(s) = s2 0.036(s+25:28)
(s2 +0.00396s+1)
a team of control engineers designed
2 −0.2244s+0.8981)
and tested a linear controller K(s) = 7.9212(s+0:1818)(s
(s2 +3.899s+4.745)(s2 +1.039s+3.395)
that performed
sufficiently well for the job to be done. Now that the cold war is over, NASA
found another purpose for the satellite. Unfortunately, due to a recent defect in the
control hardware, the measurement got delayed by maximally T = 0.25 seconds. For
this purpose, a robustness analysis has been scheduled in order to verify whether
the system still performs sufficiently well, or whether a new controller has to be
designed and uploaded. Unfortunately, NASA ran out of money, due to jet another
fiscal cliff. For this reason NASA asked the University of Stuttgart (which is well-
known for their excellent students) whether the job could be done by a student free
of charge.
Consider the time-delayed feedback interconnection depicted in Figure 6.
a) Design a generic dynamical weight WT of order 1 that “tightly covers” dT − 1
with the delay operator dT and its transfer function e−sT and for arbitrary
T > 0. “Covering” is interpreted as in Example 2.12.
b) For the nominal model G and the weight WT , show that the delayed transfer
function H satisfies all properties in Definition 2.11.
c) Perform a robustness analysis based on Theorem 2.13 and answer the following
questions:
i) Does the system remain stable?
ii) What is the robust stability margin?
iii) What is the maximum tolerable delay-time T for which kWT (1 +
GK)−1 k∞ ≤ 1?
d) What would be your recommendation for NASA?
Hint: Although not proven rigorously, you can use the fact that (2.16) ⇒ (2.15)
in Theorem 2.13 stays valid for the delayed transfer function H. Use Matlab for
solving this exercise!

5) This continues Exercise 4) with H for the nominal system G and the uncertainty
weight WT .
a) Recall the first order weight WT from Exercise 4) and the second order perfor-
mance weight Wp from Example 2.20 and verify whether the controller achieves
robust performance for a delay-time of T = 0.25 seconds.

43
 b) What is the maximum tolerable time-delay Tmax for which the controller still
achieves robust performance?
c) For a time-delay of T = 0.25 seconds, what is the maximum bandwidth ωB in
the weight Wp for which the controller still achieves robust performance?
d) Let Wp again be as in Example 2.20 and choose a time-delay T > Tmax with
|Wp (iω)S(iω)| + |WT (iω)T (iω)| > 1 for some ω ∈ R. Determine some H ∈ H
such that kWp (1 + HK)−1 k∞ > 1.
Hint: Use Lemma 2.22.

6) For any transfer function ∆ define an uncertain system through

1
G∆ (s) = (1 + ∆(s)).
s
The nominal system is G0 (s) = 1/s and the controller is taken to be K(s) = 1. As
usual define L∆ = G∆ K and T := L0 (1 + L0 )−1 .
a) Show that T is stable and kT k∞ = 1.
b) If ∆ is a stable transfer function which satisfies ∆(0) = −1, its Taylor expansion
reads as ∆(s) = −1 + a1 s + a2 s2 + . . . around s = 0. If k∆k∞ ≤ 1 show that
a1 ≥ 0.
c) Show that (1 + L∆ )−1 is stable for all stable ∆ satisfying k∆k∞ ≤ 1.

The next few exercises are more mathematical orientated.

7) This continues Exercise 2) with transfer functions G and K.

a) Show that K internally stabilizes G iff (1 + GK)−1 is stable and no poles of G
and K in C0 ∪ C+ are canceled if forming GK.
b) Prove the Nyquist criterion for internal stability: Suppose m0+ is the number
of poles of G and of K in C0 ∪ C+ (including multiplicities). Then K inter-
nally stabilizes the tracking configuration iff the Nyquist plot of GK does not
pass through −1 and encircles −1 exactly m0+ times in the counterclockwise
direction.

8) Prove the following statements.

a) A continuous function f : C0 ∪ C+ → R is sub-harmonic if it satisfies, for all
z ∈ C+ and all r > 0 with r < Re(z), the inequality
Z 2π
1
f (z) ≤ f (z + reit )dt.
2π 0
If K ⊂ C0 ∪ C+ is a compact subset with boundary ∂K, show that

sup f (z) = sup f (z).

z∈∂K z∈K

44
 b) For any matrix M ∈ Ck×m show that kM k = maxkxk=1,yk=1 |x∗ M y|.
k×m
c) For G ∈ RH∞ show that z 7→ kG(z)k is subharmonic.
k×m
d) For G ∈ RH∞ prove the following “maximum modulus theorem”:

kGk∞ = sup kG(z)k.

z∈C0 ∪C+

9) Let L, W and ∆ be transfer functions with the following properties:

• W and T = L(1 + L)−1 are stable and satisfy kW T k∞ ≤ 1.
• ∆ has no poles in C0 and satisfies |∆(iω)| < 1 for all ω ∈ R ∪ {∞}.
Prove the following facts:
a) 1 + T W ∆ and (1 + T W ∆)−1 have no poles in C0 ∪ {∞}.
b) L and L∆ := L(1 + W ∆) have identical poles in C0 (including multiplicities).
c) Suppose that (1 + L∆ )−1 is stable. Then L and L∆ have the same number of
poles in C+ .
Hint: To prove the third statement one can use Rouché’s theorem.

10) With the transfer function L0 = N/D (where N , D are real coprime polynomials)
and T > 0 define the loop transfer function LT (s) = L0 (s)e−sT .
a) Show that the poles of (1 + LT )−1 in C0 ∪ C+ are given by the zeros of pT (s) :=
D(s) + N (s)e−sT in C0 ∪ C+ .
b) If |L0 (∞)| < 1 show that pT has only finitely many zeros in C0 ∪ C+ .
c) If |L0 (∞)| < 1 and pT has no zeros in C0 ∪ C+ show that (1 + LT )−1 is analytic
and bounded in C+ .
d) Given constants M > 0, R > 0, show that there exists T0 > 0 such that
M |1 − e−iωT | < 1 for all ω ∈ [−R, R], T ∈ [0, T0 ].
e) Suppose that |L0 (∞)| < 1 and that p0 has all its zeros in C− . Show that there
exists some T0 > 0 such that pT has only zeros in C− for T ∈ [0, T0 ). You can
assume that L0 has no poles in C0 .
Hint: Use the Nyquist criterion for LT .
f) If |L0 (∞)| > 1 show that for all T > 0 the polynomial pT has at least one zero
in C0 ∪ C+ .
Hint: Consider the zeros of the function f (s) = eT /s L0 (s−1 )−1 + 1 with an
essential singularity at 0 and apply Picard’s Great Theorem without proof.

45
3 Stabilizing Controllers for Interconnections

In the previous section we have discussed some robustness issues for SISO systems. In
practice, however, one encounters interconnections of multi input multi output (MIMO)
systems. In the most simplest case, typical components of such an interconnection are a
model of a considered physical plant and a to-be-designed feedback controller.

Before we deal with robustness for such interconnections of systems we consider the precise
definition of when a system is internally stabilized by a controller.

3.1 A Specific Tracking Interconnection

To be concrete, let us look at the typical one-degree of freedom control configuration in

Figure 17. Here G is the plant model, K is the to-be-designed controller, r is the reference
input signal, d is a disturbance at the plant’s output, n is measurement noise, e is the
tracking error, u is the control input, and y is the measured output.

It is important to note that we have explicitly specified those signals that are of interest
to us:

• Signals that affect the interconnection and cannot be influenced by the controller:
r, d, n.

• Signals with which we characterize whether the controller achieves the desired goal:
e should be kept as small as possible for all inputs r, d, n in a certain class.

• Signals via which the plant can be controlled: u.

• Signals that are available for control: y.

The interconnection does not only comprise the system components (G, K) and how
the signals that are processed by these components are related to each other, but it also
specifies those signals (e and r, d, n) that are related to the targeted task of the controller.

The corresponding open-loop interconnection is simply obtained by disconnecting the con-

troller as shown in Figure 18.

It is straightforward to arrive, without any computation, at the following input-output

description of the open-loop interconnection:
 
    d 
e I 0 −I G n
 =   .

y −I −I I −G  r

u

46
 -1
d
r y u e
+ K G + +

n
-1 +

Figure 17: Closed-loop interconnection

-1
d
r y u e
+ G + +

n
-1 +

Figure 18: Open-loop interconnection that corresponds to Figure 17

The input-output description of the closed-loop interconnection is then obtained by closing

the loop as
u = Ky.
A simple calculation reveals that
 
h   i   d
−1
e= I 0 −I + GK(I − (−G)K) −I −I I n
 
r

what can be simplified to

 
   d
e= −1 −1 −1 n.
(I + GK) −GK(I + GK) −(I + GK)  
r

As expected for this specific interconnection, we arrive at

 
  d 
e = S −T −S  n


47
with sensitivity S = (I + GK)−1 and complementary sensitivity T = GK(I + GK)−1 .

Let us now extract a general scheme from this specific example.

3.2 The General Framework

In an arbitrary closed-loop interconnection structure, let

• w denote the signal that affects the system and cannot be influence
 by the controller.
d
 
w is called generalized disturbance. (In our example, w =   n .)


• z denote the signal that allows to characterize whether a controller has certain
desired properties. z is called controlled variable. (In our example, z = e.)

• u denote the output signal of the controller, the so-called control input. (In our
example it’s just u.)

• y denote the signal that enters the controller, the so-called measurement output.
(In our example it’s just y.)

Any open-loop interconnection can then be generally described by (Figure 19)

      
z w P P w
  = P   =  11 12    (3.1)
y u P21 P22 u

where the system P comprises the subsystems that are involved in the interconnection
and the manner how these subsystems are connected with each other.

Even if we start with an interconnection of SISO systems, the resulting open-loop inter-
connection will generally be described by a MIMO system since one has to stack several
signals with only one component to vector valued signals.

In these whole notes we start from the fundamental hypothesis that P is an LTI system.
We denote the corresponding transfer matrix with the same symbol as
 
P11 (s) P12 (s)
P (s) =  .
P21 (s) P22 (s)

Let
ẋ = Ax + B1 w + B2 u
z = C1 x + D11 w + D12 u (3.2)
y = C2 x + D21 w + D22 u

48
 z w
P
y u

Figure 19: General open-loop interconnection

uK yK
K

Figure 20: Controller

denote a stabilizable and detectable state-space realization of P .

A controller (Figure 20) is any LTI system

yK = KuK . (3.3)
It can be described in the frequency domain by specifying its transfer matrix
K(s)
or via the stabilizable and detectable state-space realization

ẋK = AK xK + BK uK
(3.4)
y K = C K xK + D K u K .

The interconnection of the controller and the open-loop system as

uK = y and u = yK
leads to the closed-loop interconnection as depicted in Figure 21.

Remark 3.1 To have a minimal dimensions of the matrices and, hence, reduce the effort
for all subsequent computations, one should rather work with minimal (controllable and
observable) realizations for P and K. One can take these stronger hypothesis as the basis
for the discussion throughout these notes without the need for any modification.

3.3 Stabilizing Controllers - State-Space Descriptions

Let us now first compute a realization of the interconnection as

   
u y
 = K (3.5)
uK y

49
 z
P w

y u

Figure 21: General closed-loop interconnection

of the system (3.2) and the controller (3.4).

For that purpose it is advantageous to merge the descriptions of (3.2) and (3.4) as
    
ẋ A 0 B1 B2 0 x
    
 ẋ   0 A 0 0 BK    xK 
 
 K  K
    
 z  =  C1 0 D11 D12 0   w  .
    
    
 y K   0 CK 0 0 DK   u 
 
   
y C2 0 D21 D22 0 uK

To simplify the calculations notationally, let us introduce the abbreviation

 
A 0 B1 B2 0
   
A B1 B2  0 AK 0 0 BK 
 
   
 C 1 D 11 D 12  :=  C1
   0 D11 D12 0  . (3.6)
 
C 2 D 21 D 22  0
 CK 0 0 DK  
C2 0 D21 D22 0

The interconnection (3.5) leads to

 
  x  
yK   y
 K

=
 xK  + D 22
C 2 D 21 
 
y y
w

or  
  x
yK   
[I − D 22 ]  = C 2 D 21  xK  .
y
 
w

50
If I − D 22 is non-singular, we arrive at
 
  x
yK  
 = [I − D 22 ]−1

 C 2 D 21  x K

y
 
w

what finally leads to

   
ẋ    
  x

A B1 B2 
 [I − D 22 ]−1 C 2 D 21   xK
  
 ẋK  =  + 
C 1 D 11 D 12
   
z w

or
   
ẋ   x
  A + B 2 [I − D 22 ]−1 C 2 B 1 + B 2 [I − D 22 ]−1 D 21  
 ẋK  =    xK  . (3.7)
C 1 + D 12 [I − D 22 ]−1 C 2 D 1 + D 12 [I − D 22 ]−1 D 21
   
z w

This is an explicit formula for a state-space representation of the closed-loop interconnec-

tion.

On our way to derive this formula we assumed that I − D 22 is non-singular. This is a

condition to ensure that we could indeed close the loop; this is the reason why it is often
called a well-posedness condition for the interconnection.

Any controller should at least be chosen such that the interconnection is well-posed.
In addition, we require that the controller stabilizes the interconnection. This will just
amount to requiring that the matrix A + B 2 [I − D 22 ]−1 C 2 which defines the dynamics
of the interconnection is stable.

We arrive at the following fundamental definition of when the controller (3.4) stabilizes
the open-loop system (3.2).

Definition 3.2 The controller (3.4) stabilizes the system (3.2) if

 
I −DK
  is non-singular (3.8)
−D22 I

and if     −1  
A 0 B2 0 I −DK 0 CK
 +     (3.9)
0 AK 0 BK −D22 I C2 0
has all its eigenvalues in the open left-half plane C< .

51
Remark 3.3

• Verifying whether K stabilizes P is very simple: First check whether the realizations
of both P and K are stabilizable and detectable, then check (3.8), and finally verify
whether (3.9) is stable.

• If the realizations of both P and K are stabilizable and detectable, the realization of
the closed-loop interconnection (3.7) is in general not stabilizable and not detectable.

• Note that the definition only involves the matrices

   
A B2 AK BK
  and  .
C2 D22 CK DK
  
The matrices B1 and C1 only play a role in requiring that A, B1 B2 is stabi-
  
C1
lizable and A,   is detectable.
C2

• The same definition is in effect if the channel w → z is void and the system (3.2)
just reads as     
ẋ A B2 x
 =  .
y C2 D22 u

Note that the formulas (3.7) for the closed-loop interconnection simplify considerably if
either D22 or DK vanish. Let us look at the case when

D22 = 0.

Then (3.8) is always true. Due to

 −1  
I −DK I DK
  = ,
0 I 0 I

a straightforward calculation reveals that (3.7) now reads as

    
ẋ A + B2 DK C2 B2 CK B1 + B2 DK D21 x
    
 ẋK  =  C2 BK AK BK D21   xK .
    
z C1 + D12 DK C2 D12 CK D11 + D12 DK D21 w

Then the matrix (3.9) just equals

 
A + B2 DK C2 B2 CK
 .
C2 BK AK

52
Example 3.4 Consider again the classical configuration in section 3.1 with
 
−0.1 1 0 0
 
200 1  0 −20 1 0 
G(s) = =  (s).
 
10s + 1 (0.05s + 1) 2  0
 0 −20 64 

125 0 0 0

and a controller    
AK BK −3.3 1.3
K= = 
CK DK 1.7 0.3
with stabilizable and detectable realization. A stabilizable and detectable realization for
the open-loop interconnection P (depicted in Figure 18) is given by
 
−0.1 1 0 0 0 0 0
   
A B1 B2  0 −20 1 0 0 0 0 
 
   
 C1 D11 D12  =  0
   0 −20 0 0 0 64  .
 
C2 D21 D22  125 0
 0 1 0 −1 0  
−125 0 0 −1 −1 1 0

Since D22 = 0 clearly implies (3.8) and since the eigenvalues of (3.9) are contained in C− ,
we can conclude that K stabilizes P .

Next we want to derive a characterization of stabilizing controllers in terms of transfer

matrices. To do so, we first introduce linear fractional transformations.

3.4 Linear Fractional Transformations

Suppose P and K are given transfer matrices. Then the so-called lower linear fractional
transformation P ? K of P and K is defined as follows: Partition
 
P11 P12
P = 
P21 P22

such that P22 K is square, check whether the rational matrix I − P22 K has a rational
inverse, and set
P ? K := P11 + P12 K(I − P22 K)−1 P21 .
In the literature, the expression Fl (P, K) is often used instead of P ? K. Since a lower
linear fractional transformation is a particular form of a more general operation that
carries the name star-product, we prefer the symbol P ? K.

53
Similarly, the upper linear fractional transformation ∆ ? P of the rational matrices
∆ and P is defined as follows: Choose a partition
 
P11 P12
P = 
P21 P22

such that P11 ∆ is square, check whether the rational matrix I − P11 ∆ has a rational
inverse, and set
∆ ? P := P22 + P21 ∆(I − P11 ∆)−1 P12 .
One often finds the notation Fu (P, ∆) in the literature where one has to note that, un-
fortunately, the matrices ∆ and P appear in reverse order.

At this point, P ? K and ∆ ? P should be just viewed as abbreviations for the formulas
given above. The discussion to follow will reveal their system theoretic relevance.

3.5 Stabilizing Controllers - Input-Output Descriptions

Let us first see how to determine an input-ouput description of the closed-loop intercon-
nection as in Figure 21. For that purpose we only need to eliminate the signals u, y
in
z = P11 w + P12 u, y = P21 w + P22 u, u = Ky.
The last two relations lead to y = P21 w + P22 Ky or (I − P22 K)y = P21 w. If I − P22 K
does have a proper inverse, we obtain y = (I − P22 K)−1 P21 and, finally,

z = [P11 + P12 K(I − P22 K)−1 P21 ]w. (3.10)

This is a general formula how to obtain, from the input-ouput description P and from
that of the controller K, the corresponding input-output description of the closed-loop
interconnection. If we recall the definitions in Section 3.4, we observe that the closed-loop
input-output description by performing the lower linear fractional transformation of P
with K which has been denoted as P ? K:

z = (P ? K)w.

This is the mere reason why these fractional transformations play such an important role
in these notes and, in general, in robust control.

Note that (3.10) gives the transfer matrix that is defined by (3.7) and, conversely, (3.7)
is a state-space realization of (3.10).

Let us now observe that I − P22 K has a proper inverse if and only if I − P22 (∞)K(∞)
is non-singular. If we look back to the realizations (3.2) and (3.4), this just means that
I − D22 DK is non-singular, what is in turn equivalent to (3.8).

54
 z w
P + v1
y u

v2 v
+ K

Figure 22: Interconnection to test whether K stabilizies P .

Unfortunately, the relation for stability is not as straightforward. In general, if (3.7)

is stable, the transfer matrix defined through (3.10) is stable as well. However, the
realization (3.7) is not necessarily stabilizable or detectable. Therefore, even if (3.10)
defines a stable transfer matrix, the system (3.7) is not necessarily stable. For checking
whether K stabilizes P , it hence does not suffice to simply verify whether P11 + P12 K(I −
P22 K)−1 P21 defines a stable transfer matrix.

This indicates that we have to check more transfer matrices in the loop in Figure 21 than
just the one explicitly displayed by the channel w → z in order to guarantee that K
stabilizes P . It turns out that Figure 22 gives a suitable setup to define all the relevant
transfer matrices that have to be tested.

Theorem 3.5 K stabilizes P if and only if the interconnection as depicted in Figure 22

and defined through the relations
   
z w
  = P   , u = Kv + v1 , v = y + v2
y u

or, equivalently, by     
z P11 P12 0 w
    
v = 0 I −K  u  (3.11)
 1   
v2 −P21 −P22 I v
defines a proper transfer matrix
  
w z
   
 v1  →  u  (3.12)
   
v2 v

that is stable.

55
With this result we can test directly on the basis of the transfer matrices whether K
stabilizes P : One has to check whether the relations (3.11) define a proper and stable
transfer matrix (3.12).

Let us first clarify what this means exactly. Clearly, (3.11) can be rewritten as
        
v 0 I −K u   u
 1= w +     , z = P11 w + P12 0   . (3.13)
v2 −P21 −P22 I v v

These relations define a proper transfer matrix (3.12) if and only if

 
I −K
  has a proper inverse.
−P22 I

As well-known, this is true if and only if

 
I −K(∞)
  is non-singular.
−P22 (∞) I

Indeed, under this hypothesis, the first relation in (3.13) is equivalent to

   −1    −1  
u I −K 0 I −K v
 =   w +    1 .
v −P22 I P21 −P22 I v2

Hence (3.13) is nothing but

  −1    −1 
    I −K 0   I −K  
z  P11 + P12 0     P12 0    w
−P22 I P21 −P22 I
 
    
u= −1 −1
v .
          1 
 I −K 0 I −K 
v         v2
−P22 I P21 −P22 I

If we recall the formula

 −1  
−1 −1
I −K (I − KP22 ) K(I − P22 K)
  = ,
−P22 I (I − P22 K)−1 P22 (I − P22 K)−1

this can be rewritten to

    
z P11 + P12 K(I − P22 K)−1 P21 P12 (I − KP22 )−1 P12 K(I − P22 K)−1 w
    
u=
   K(I − P22 K)−1 P21 (I − KP22 )−1 K(I − P22 K)−1    v1  .
 

v (I − P22 K)−1 P21 (I − P22 K)−1 P22 (I − P22 K)−1 v2

(3.14)

We have arrived at a more explicit reformulation of Theorem 3.5.

56
Corollary 3.6 K stabilizes P if and only if I − P22 K has a proper inverse and all nine
transfer matrices in (3.14) are stable.

Remark 3.7 If the channel w → z is absent, the characterizing conditions in Theorem

3.5 or Corollary 3.6 read as follows: The transfer matrix
 
I −K
 
−P22 I

has a proper and stable inverse. Since K and P22 are, in general, not stable, it does not
suffice to simply verify whether the determinant of this matrix is stable; Lemma 1.4 does
not apply!

Proof of Theorem 3.5. We have already clarified that (3.11) defines a proper transfer
matrix (3.12) if and only if (3.8) is true. Let us hence assume the validity of (3.8).

Then we observe that (3.11) admits the state-space realization

    
ẋ A 0 B1 B2 0 x
    
 ẋ   0 A 0 0 BK    xK
 
 K  K 
    
 z  =  C1 0 D11 D12 0  . (3.15)
 w

   
    
 v1   0 −CK 0 I −D K
 u 
    
v2 −C2 0 −D21 −D22 I v

Using the abbreviation (3.6), this is nothing but

   
ẋ x
    
 ẋ  A B1 B2  xK
 
 K 
    
 z  =  C 1 D 11 D 12  .
 w

   
   
 v1 
  −C 2 −D 21 I − D 22
 u



v2 v

By (3.8), D̃ 22 := I − D 22 is non-singular. The same calculation as performed earlier

leads to a state-space realization of (3.12):
   
ẋ x
−1 −1 −1  
    
 ẋ  A + B D̃ C B + B D̃ D B D̃ x 
 K 2 22 2 1 2 22 21 2 22  K
 z  =  C 1 + D 12 D̃ −1 C 2 D 1 + D 12 D̃ −1 D 21 D 12 D̃ −1   w  .
    
   22 22 22    (3.16)
  −1 −1 −1  
 u  D̃ 22 C 2 D̃ 22 D 21 D̃ 22  v1 
   
v v2

57
Here is the crux of the proof: Since (3.2) and (3.4) are stabilizable/detectable realizations,
one can easily verify with the Hautus test that (3.15) has the same property. This implies
that the realization (3.16) is stabilizable and detectable as well.

Therefore we can conclude: The transfer matrix of (3.16) is stable if and only if the
−1
system (3.16) is stable if and only if A + B 2 D̃ 22 C 2 = A + B 2 (I − D 22 )−1 C 2 has all its
eigenvalues in C< . Since we have guaranteed the validity of (3.8), this property is (by
Definition 3.2) nothing but the fact that K stabilizes P .

3.6 Generalized Plants

Contrary to what one might expect, it is not possible to find a stabilizing controller K
for any P .

Example 3.8 Let us consider (3.1) with

 
1 1/s
P (s) =  .
1 1/(s + 1)

We claim that there is no K(s) that stabilizes P (s). Reason: Suppose we found a K(s)
that stabilizes P (s). Then the two transfer functions
1 1
P12 (s)(I − K(s)P22 (s))−1 =
s 1 − K(s)
s+1
K(s) 1 1
P12 (s)(I − K(s)P22 (s))−1 K(s) = s = 1 1
s − s+1 K(s) s K(s)
− s+1

are stable. But this cannot be true. To show that, we distinguish two cases:
1
• Suppose K(s) has no pole in 0. Then the denominator of K(s) is finite in s = 0
1− s+1
such that this function cannot have a zero in s = 0. This implies that the first of
the above two transfer functions has a pole in 0, i.e., it is unstable.
1 1
• Suppose K(s) does have a pole in 0. Then K(s)
vanishes in s = 0 such that 1 1
− s+1
K(s)
takes the value −1 in s = 0. Hence, the second of the above two transfer functions
has a pole in 0 and is, thus, unstable.

We arrive at the contradiction that at least one of the above two transfer functions is
always unstable. Roughly speaking, the pole s = 0 of the transfer function P12 (s) = 1s
cannot be stabilized via feeding y back to u since this is not a pole of P22 (s) as well.

58
Our theory will be based on the hypothesis that P does in fact admit a stabilizing con-
troller. For such open-loop interconnections we introduce a particular name.

Definition 3.9 If there exists at least one controller K that stabilizes the open-loop in-
terconnection P , we call P a generalized plant.

Fortunately, one can very easily check whether a given P is a generalized plant or not.
We first formulate a test for the state-space description of P .

Theorem 3.10 P with the stabilizable/detectable realization (3.2) is a generalized plant

if and only if (A, B2 ) is stabilizable and (A, C2 ) is detectable.

  
Since the realization (3.2) is stabilizable, we know that A, B1 B2 is stabilizable.
This does clearly not imply, in general, that the pair (A, B2 ) defining a system with fewer
inputs is stabilizable. A similar remark holds for detectability.

Let us now assume that (A, B2 ) is stabilizable and (A, C2 ) is detectable. Then we can
explicitly construct a controller that stabilizes P . In fact, stabilizability of (A, B2 ) and
detectability of (A, C2 ) imply that there exist F and L such that A + B2 F and A + LC2
are stable. Let us now take the controller K that is defined through

ẋK = (A + B2 F + LC2 + LD22 F )xK − Ly, u = F xK .

Note that this is nothing but the standard observer-based controller which one would
design for the system
ẋ = Ax + B2 u, y = C2 x + D22 u.
It is simple to check that K indeed stabilizes P . First, K is strictly proper (DK = 0)
such that (3.8) is obviously true. Second, let us look at
    −1  
A 0 B2 0 I 0 0 CK
 +   =
0 AK 0 BK −D22 I C2 0
     
A 0 B 0 I 0 0 F
= + 2   =
0 A + B2 F + LC2 + LD22 F 0 −L D22 I C2 0
    
A 0 B 0 0 F
= + 2  =
0 A + B2 F + LC2 + LD22 F 0 −L C2 D22 F
 
A B2 F
= .
−LC2 A + B2 F + LC2

59
We claim that this matrix is stable. This should be known from classical theory. However,
it can be verified by performing the similarity transformation (error dynamics!)
   −1
I 0 A B2 F I 0
   
I −I −LC2 A + B2 F + LC2 I −I

to arrive at  
A + B2 F B2 F
 
0 A + LC2
which is, obviously, stable since the diagonal blocks are.

This was the proof of the if-part in Theorem 3.10 with an explicit construction of a
stabilizing controller.

Proof of only if. To finish the proof, we have to show: If there exists a K that stabilizes
P , then (A, B2 ) is stabilizable and (A, C2 ) is detectable. If K stabilizes P , we know by
definition that
    −1  
A 0 B 0 I −DK 0 CK
A :=  + 2    
0 AK 0 BK −D22 I C2 0

is stable. This implies that (A, C2 ) is detectable. Let us prove this fact with the Hautus
test: Suppose Ax = λx, x 6= 0, and C2 x = 0. Then we observe that
     
x Ax x
A  =   = λ .
0 0 0
 
x
Hence   is an eigenvector of A with eigenvalue λ. Since A is stable, we infer Re(λ) <
0
0. This proves that (A, C2 ) is detectable.

Task. Show in a similar fashion that (A, B2 ) is stabilizable what finishes the proof.

Remark 3.11 If the channel w → z is absent, then (A, B2 ) and (A, C2 ) are obviously
stabilizable and detectable. (The matrices B1 , C2 , D11 , D12 , D21 in (3.2) are void.) Then
there always exists a controller K that stabilizes P .

This last remark reveals that we can always find a u = Ky that stabilizes y = P22 u. This
leads us to a input-output test of whether P is a generalized plant or not.

60
Theorem 3.12 Let u = Ky be any controller that stabilizes y = P22 u. Then P is a
generalized plant if and only if this controller K stabilizes the open-loop interconnection
P.

Again, this test is easy to perform: Find an (always existing) K that stabilizes P22 , and
verify that this K renders all the nine transfer matrices in (3.14) stable. If yes, P is a
generalized plant, if no, P is not.

Proof. Let K stabilize P22 .

If K also stabilizes P , we infer that there exists a stabilizing controller and, hence, P is
a generalized plant.

Conversely, let P be a generalized plant. We intend to show that K not only

 stabi-

v1
lizes P22 but even P . We proceed with state-space arguments. Recall that   =
v2
  
I −K u
    admits the state-space realization
−P22 I v
    
ẋ A 0 B2 0 x
    
 ẋ   0 A 0 BK x 
 K K  K
= . (3.17)
  
 
 v1   0 −CK I −DK  u 
    
v2 −C2 0 −D22 I v

Using the abbreviation (3.6), this is nothing but

   
ẋ  x
   
 ẋ 
 K  A B2 x 
 K
= .
 
 
 v1 
  −C 2 I − D 22  u


v2 v

By (3.8), I − D 22 is non-singular. Again, the same calculation as earlier leads to a

   −1  
u I −K v
state-space realization of   =    1  given by
v −P22 I v2
   
ẋ x
    
−1 −1
 ẋ
 K
 A + B 2 (I − D 22 ) C 2 B 2 (I − D 22 ) x
 K

= . (3.18)
   
 
−1 −1
 u


 (I − D 22 ) C 2 (I − D 22 )  v1



v v2

61
Since P is a generalized plant, (A, B2 ) is stabilizable and (A, C2 ) is detectable. Therefore,
the same is true for (3.17) and, similarly as in the proof of Theorem 3.5, also for (3.18).
Since K stabilizes P22 , the transfer matrix defined through (3.18) is stable. Since this
realization is stabilizable and detectable, we can conclude that A + B 2 (I − D 22 )−1 C 2 is
actually stable. Hence K stabilizes also P by definition.

3.7 Summary

For a specific control task, extract the open-loop interconnection (3.1).

Then test whether this open-loop interconnection defines a generalized plant by applying
either one of the following procedures:
  
• Find a state-space realization (3.2) of P for which A, B1 B2 is stabilizable (or
  
C1
even controllable) and A,   is detectable (or even observable), and check
C2
whether (A, B2 ) is stabilizable and (A, C2 ) is detectable. If yes, P is a generalized
plant, if no, P is not.
 
I −K
• Find any K such that   does have a proper and stable inverse. Then
−P22 I
verify whether this K renders all transfer matrices in (3.14) stable. If yes, P is a
generalized plant, if no, P is not.

If P turns out to be no generalized plant, the interconnection under consideration is not

suitable for the theory to be developed in these notes.

Suppose K stabilizes P . Then the closed-loop interconnection is described as

z = (P11 + P12 K(I − P22 K)−1 P21 )w = (P ? K)w

In the state-space, the closed-loop system admits the realization (3.7) with the abbrevia-
tions (3.6).

3.8 Back to the Tracking Interconnection

Let us come back to the specific tracking interconnection in Figure 17 for which we have
obtained    
P11 P12 I 0 −I G
P = = .
P21 P22 −I −I I −G

62
We claim that this is a generalized plant.

Input-output test: Let K stabilize P22 = −G. This means that that
   
(I − KP22 )−1 K(I − P22 K)−1 (I + KG)−1 K(I + GK)−1
 = 
(I − P22 K)−1 P22 (I − P22 K)−1 −(I + GK)−1 G (I + GK)−1
 
is well-defined and stable. Let us now look at (3.14). Since P21 (s) = −I −I I is
−1 −1
stable, the same is true of K(I − P22 K) P21 and (I − P22 K) P21 . Since P12 = G, we
infer
P12 (I − KP22 )−1 = G(I + KG)−1 = (I + GK)−1 G
and
P12 K(I − KP22 )−1 = KG(I + KG)−1 = I − (I + KG)−1
that are both stable. Hence it remains to check stability of

P ? K = P11 + P12 K(I − P22 K)−1 P21 :

−1
We  seen that P12 K(I − KP22 ) is stable. Since the same is true for P11 =
 have just
I 0 −I and P21 , we can indeed conclude that P ? K is stable. This reveals that all
nine transfer matrices in (3.14) are stable. By Theorem 3.12, P is a generalized plant.

State-space test: Let us assume that

G(s) = CG (sI − AG )−1 BG + DG

is a minimal realization. Then we observe that

   
CG   I 0 −I DG
P (s) =   (sI − AG )−1 0 0 0 BG +  
−CG −I −I I −DG

and hence P (s) admits a minimal realization with the matrix

   
A B1 B2 AG 0 0 0 BG
   
 C1
 D11 D12 
 =  CG I 0 −I DG  .

C2 D21 D22 −CG −I −I I −DG

Since (A, B2 ) = (AG , BG ) is controllable and (A, C2 ) = (AG , CG ) is observable, Theorem

3.10 implies that P is a generalized plant.

Note that all these tests are very simple, mainly due to the simplicity of the feedback
interconnection in Figure 17 under scrutiny. In practical circumstances one might en-
counter a much more complicated configuration where the tests have to be performed
numerically. We recommend Matlab’s Robust Control Toolbox command sysic to easily
build state-space realization even for complicated interconnections.

63
Exercises

1) Let P be a stable stable LTI system.

a) Show that K stabilizes P if and only if I − P22 K has a proper inverse and
K(I − P22 K)−1 is stable. (It suffices to check one instead of nine transfer
matrices.)
Is the same statement true if we replace K(I − P22 K)−1 with (I − P22 K)−1 ?
b) Show that the set of closed-loop transfer matrices P ? K where K varies over
all controllers that stabilize P is given by the set of all

P11 + P12 QP21

where Q is a free parameter in RH∞ . What is the relation between K and Q?

(This is the so-called Youla parameterization. Note that K enters in P ? K
in a non-linear fashion, whereas Q enters P11 + P12 QP21 in an affine fashion.
Hence the change of parameters K → Q leads to an affine dependence of the
closed-loop system on the so-called Youla-parameter. All this can be extended
to general systems P that are not necessarily stable.)

2) Which of the following transfer matrices P (s) define a generalized plant:

   
1/(s + 1) 1/(s + 2) 1/(s2 + s) 1/(s + 2)
 ,  
1/(s + 3) 1/s 1/(s + 3) 1/s
   
2
1/(s + 1) 1/(s + 2s) 1/(s + 1) 1/(s + 2)
 ,  ?
1/(s + 3) 1/s 1/(s + 3) 1/s2
   
z w
It is assumed that all signals in   = P   have one component.
y u

3) Suppose you have given the interconnection in Figure 23. We view Gj , j =

1, 2, 3, 4, 5, 6, 7 as possibly MIMO system components, and Kj , j = 1, 2, are possibly
MIMO controller blocks.

a) Compute the description P of the open-loop interconnection in terms of Gj .

Mind the fact that all components can have multiple inputs and outputs.
b) Find two examples with simple SISO components Gj such that the resulting
two open-loop interconnections P1 , P2 have the following properties:

P1 is no generalized plant. There exists a controller K that renders

S(P2 , K) stable but that does not stabilize P2 .

64
 G4

G2 + K1
−
−
w z
+ G1 G5 + G7
−
−
G3 + K2

G6

Figure 23: An interconnection

Is it possible to take P1 = P2 ?
c) (Matlab) Choose
1 s+1
G1 (s) = 1, G2 (s) = , G3 (s) = 2 , G4 (s) = 0,
s−1 s +1
1 s+2
G5 (s) = , G6 (s) = 1, G7 (s) = .
s (s + 3)(s − 2)

Show that P is a generalized plant. Design a controller K that stabilizes P .

Explain how you obtain K, and how you check whether K indeed stabilizes P .
Draw a Bode magnitude plot of the resulting closed loop system P ? K.

65
4 Robust Stability Analysis

All mathematical models of a physical system suffer from inaccuracies that result from
non-exact measurements or from the general inability to capture all phenomena that are
involved in the dynamics of the considered system. Even if it is possible to accurately
model a system, the resulting descriptions are often too complex to allow for a subsequent
analysis, not to speak of the design of a controller. Hence one rather chooses for a simple
model and takes a certain error between the simplified and the more complex model into
account.

Therefore, there is always a mismatch between the model and the system to be investi-
gated. A control engineer calls this mismatch uncertainty. Note that this is an abuse of
notation since neither the system nor the model are uncertain; it is rather our knowledge
about the actual physical system that we could call uncertain.

The main goal of robust control techniques is to take these uncertainties in a systematic
fashion into account when analyzing a control system or when designing a controller for
it.

In order to do so, one has to arrive at a mathematical description of the uncertainties.

Sometimes it is pretty obvious what to call an uncertainty (such as parameter variations
in a good physical model), but sometimes one just has to postulate a certain structure
of the uncertainty. Instead of being general, we shall first turn again to the specific
interconnection in Figure 17 and anticipate, on some examples, the general paradigm and
tools that are available in robust control.

4.1 Uncertainties in the Tracking Configuration - Examples

4.1.1 A Classical SISO Example

This example serves as preparation towards the general approach to deal with additive
uncertainties and as a review of Section 2.4.

Let us be concrete and assume that the model G(s) in Figure 17 is given as
200 1
G(s) = . (4.1)
10s + 1 (0.05s + 1)2
Suppose the controller is chosen as
0.1s + 1
K(s) = . (4.2)
(0.65s + 1)(0.03s + 1)

The code

66
s = zpk(’s’);
G = 200/(10*s + 1)/(0.05*s + 1)^2;
K = (0.1*s + 1)/(0.65*s + 1)/(0.03*s + 1);
systemnames=’G’;
inputvar=’[d;n;r;u]’;
outputvar=’[G+d-r;r-n-d-G]’;
input_to_G=’[u]’;
P = sysic;
S=lft(P,K);
[A,B,C,D]=ssdata(S);
eig(A)

actually computes realizations of the open-loop interconnection P , of the controller K,

and of the closed-loop interconnection P ? K denoted as (A, B, C, D). It turns out that
A is stable such that K stabilizes P .

Suppose that we know (for example from frequency domain experiments) that the fre-
quency response H(iω) of the actual stable plant H(s) does not coincide with that of the
model G(iω). Let us assume that we can even quantify this mismatch as

|H(iω) − G(iω)| < 1 for all ω ∈ R ∪ {∞}. (4.3)

Here is the fundamental question we would like to ask: If we replace G by H, does the
controller still stabilize the feedback interconnection?

If we knew H, we could just plug in H and test this property in the same way as we did
for G. Unfortunately, however, H could be any element of the set of all stable systems
H that satisfy (4.3). Hence, in principle, we would have to test infinitely many transfer
functions H what is not possible.

This motivates to look for alternative verifiable tests. Let us introduce the notation

∆(s) := H(s) − G(s)

for the plant-model mismatch. Then the actual plant is given as

H(s) = G(s) + ∆(s)

with some stable ∆(s) that satisfies

|∆(iω)| < 1 for all ω ∈ R ∪ {∞}. (4.4)

Therefore, our main question can be formulated as follows: Does the closed-loop inter-
connections as depicted in Figure 24 remain stable if ∆ is any stable transfer function
that satisfies (4.4)?

67
 -1
d
r y u e
+ K G+∆ + +

n
-1 +

Figure 24: Uncertain closed-loop interconnection

-1

z∆ w∆
∆
d
r y u e
+ K G + + +

n
-1 +

Figure 25: Rewritten uncertain closed-loop interconnection

We could also ask instead: Does there exists a stable ∆(s) with (4.4) that destabilizes the
closed-loop interconnection?

Roughly speaking, the answer is obtained by looking at the influence which the uncertainty
can exert on the interconnection. For that purpose we calculate the transfer function that
is ‘seen’ by ∆: Just rewrite the loop as in Figure 25 in which we have just introduced
notations for the input signal z∆ and the output signal w∆ of ∆. After this step we
disconnect ∆ to arrive at the interconnection in Figure 26. The transfer function seen by
∆ is nothing but the transfer function w∆ → z∆ .

For this specific interconnection, a straightforward calculation reveals that this transfer
function is given as
M = −(I + KG)−1 K.
As a fundamental result, we will reveal that the loop remains stable for a specific ∆ if
I − M ∆ does have a proper and stable inverse.

Let us motivate this result by putting the interconnection in Figure 25 into the the general

68
 -1

z∆ w∆
d
r y u e
+ K G + + +

n
-1 +

Figure 26: Closed-loop interconnection with disconnected uncertainty

z∆ w∆

z N w

Figure 27: Uncertain closed-loop interconnection

structure as in Figure 27 by setting

  z = e and collecting again all the signals d, n, r into
d
 
the vector-valued signal w = 
 n  as we did previously. Then Figure 26 corresponds to


r
Figure 28.

z∆ w∆
N
z w

Figure 28: Uncertain closed-loop interconnection

69
Mathematically, the system in Figure 28 with disconnected uncertainty is described as
         
z w N N w M N12 w
 ∆  = N  ∆  =  11 12   ∆  =   ∆ 
z w N21 N22 w N21 N22 w

where N is partitioned according to the (possibly vector valued) signals w∆ , w and z∆ ,

z. Then the transfer matrix seen by ∆ is nothing but N11 = M .

If we reconnect the uncertainty as

w∆ = ∆z∆ ,
we arrive at
z = [N22 + N21 ∆(I − M ∆)−1 N12 ]w.
This easily clarifies the above statement: Since the controller is stabilizing, all N11 = M ,
N12 , N21 , N22 are proper and stable. Only through the inverse (I − M ∆)−1 , improperness
or instability might occur in the loop. Therefore, if I − M ∆ does have a proper and stable
inverse, the loop remains stable.

Note that these arguments are not sound: We did not proved stability of the interconnec-
tion as defined in Definition 3.2. We will provide rigorous arguments in Section 4.7.

What have we achieved for our specific interconnection? We have seen that we need to
verify whether
I − M ∆ = I + (I + KG)−1 K∆
does have a proper stable inverse for all stable ∆ with (4.4). Let us apply the Nyquist
criterion: Since both M = −(I + KG)−1 K and ∆ are stable, this is true if the Nyquist
curve
ω → −M (iω)∆(iω) = (I + K(iω)G(iω))−1 K(iω)∆(iω)
does not encircle the point −1. This is certainly true if

|M (iω)∆(iω)| = |(I + K(iω)G(iω))−1 K(iω)∆(iω)| < 1 for all ω ∈ R ∪ {∞}. (4.5)

Due to (4.4), this is in turn implied by the condition

|M (iω)| = |(I + K(iω)G(iω))−1 K(iω)| ≤ 1 for all ω ∈ R ∪ {∞}. (4.6)

We conclude: If (4.6) is valid, the transfer function I − M ∆ = I + (I + KG)−1 K∆ does

have a proper and stable inverse for all stable ∆ with (4.4), and hence none of these
uncertainties can destabilize the loop.

To continue with the example, the code

70
s = zpk(’s’);
G = 200/(10*s + 1)/(0.05*s + 1)^2;
K = (0.1*s + 1)/(0.65*s + 1)/(0.03*s + 1);
systemnames = ’G’;
inputvar = ’[w;d;n;r;u]’;
outputvar = ’[u;w+G+d-r;r-w-n-d-G]’;
input_to_G = ’[u]’;
P = sysic;
N = lft(P, K);
[A, B, C, D] = ssdata(N);
eig(A)
M = N(1, 1);
om = logspace(-2, 2);
Mom = squeeze(freqresp(M, 1i*om));
loglog(om, abs(Mom));
grid on

determines the transfer matrix N , it picks out the left upper block M , the transfer function
seen by the uncertainty, and plots the magnitude of M over frequency; the result is shown
in Figure 29. Since the magnitude exceeds one at some frequencies, we see that we cannot
guarantee robust stability against all stable ∆ that satisfy (4.4).

Although this is a negative answer, the plot provides us with a lot of additional insight.

Let us first construct an uncertainty that destabilizes the loop. This is expected to happen
for a ∆ for which (I −M ∆)−1 has an unstable pole, i.e., for which I −M ∆ has an unstable
zero. Let us look specifically for a zero iω0 on the imaginary axis; then we need to have
M (iω0 )∆(iω0 ) = 1.
Let us pick ω0 such that |M (iω0 )| > 1. As the magnitude plot shows, such a frequency
indeed exists. Then the complex number
1
∆0 :=
M (iω0 )
indeed renders M (iω0 )∆0 = 1 satisfied. In our example, we chose ω0 = 5. If we calculate
∆0 and replace G by G + ∆0 , a state-space realization of the close-loop interconnection as
calculated earlier will have an eigenvalue 5i and is, hence, unstable. We have constructed
a complex number ∆0 that destabilizes the interconnection. Note, however, that complex
number are not in our uncertainty class that consisted of real rational proper transfer
functions only. Lemma 2.14 helps to find such destabilizing perturbation from ∆0 .

In fact, Lemma 2.14 says that we can construct a real-rational proper and stable ∆(s)
satisfying
∆(iω0 ) = ∆0 , |∆(iω)| = |∆0 | < 1 for all ω ∈ R ∪ {∞}.

71
 1
10

0
10

−1
10

−2
10

−3
10
−2 −1 0 1 2
10 10 10 10 10

Figure 29: Magnitude plot of M

In our case the construction leads to α = −0.8434 and β = −4.6257. As expected, the A
matrix of a realization of the closed-loop interconnection for G + ∆ turns out to have 5i
as an eigenvalue. We have hence found a stable destabilizing uncertainty whose frequency
response is smaller than 1.

To summarize, we have seen that the loop is not robustly stable against all the uncertain-
ties in the class we started out with. What can we conclude on the positive side? In fact,
Figure 29 shows that
|M (iω)| = |(I + K(iω)G(iω))−1 K(iω)| ≤ 4 for all ω ∈ R{∞}.
Therefore, (4.5) holds for all stable ∆ that satisfy
1
|∆(iω)| < for all ω ∈ R ∪ {∞}.
4
Hence, we can guarantee robust stability for all uncertainty in this smaller class.

In fact, the largest bound r for which we can still guarantee robust stability for any stable
∆ satisfying
|∆(iω)| < r for all ω ∈ R ∪ {∞}
is given by the reciprocal of the peak value of the magnitude plot:
!−1
r= sup |M (iω)| = kM k−1
∞.
ω∈R∪{∞}

72
We have discussed with this simple example how one can test robust stability by looking
at a magnitude plot of the transfer function ‘seen’ by ∆. If robust stability does not hold,
we have discussed how to construct a destabilizing perturbation.

4.1.2 A Modern MIMO Example

In the last section we have considered a very elementary example of a feedback intercon-
nection in which only one uncertainty occurs.

Let us hence look at a model that is described by the 2 × 2 transfer matrix

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

with minimal state-space realization

 
0 a 1 0
 
 −a 0 0 1 
G= .
 
 1 a 0 0
 
−a 1 0 0

Suppose that this is a model of a system in which certain tolerances for the actuators
have to be taken into account that are represented by parametric uncertainties. Let us
hence assume that the input matrix is rather given by
 
1 + δ1 0
 .
0 1 + δ2

Hence the actual system is

   
1 s − a2 a(s + 1) δ1 0
  I +  
s2 + a2 −a(s + 1) s − a2 0 δ2

or  
δ1 0
G(I + ∆) with ∆ =  
0 δ2
for some real numbers δ1 , δ2 with

|δ1 | < r, |δ2 | < r. (4.7)

Again, we are faced with a whole set of systems rather than with a single one. Uncertainty
now enters via the two real parameters δ1 , δ2 .

73
 -1
d
r y u e
+ K G(I + ∆) + +

n
-1 +

Figure 30: Uncertain closed-loop interconnection

−1

z∆ w∆
∆
d
r e
+ K + G + +
y u

n
−1 +

Figure 31: Rewritten uncertain closed-loop interconnection

Let us take the unity feedback controller

 
1 0
K= 
0 1

and consider again the interconnection in Figure 30.

As before we rewrite the interconnection in the obviously equivalent fashion as in Figure

31 and disconnect the uncertainty as in Figure 32.

The transfer matrix seen by ∆ is given as

 
1  −1 −a
M (s) = . (4.8)
s+1 a −1

As indicated earlier and as it will be developed in the general theory, for testing robust
stability we have to verify whether I − M ∆ has a proper and stable inverse for all ∆.

74
 −1

z∆ w∆
d
r e
+ K + G + +
y u

n
−1 +

Figure 32: Rewritten uncertain closed-loop interconnection

Recall that, by Lemma 1.4, this is true iff I − M (s)∆(s) is non-singular for s = ∞
(properness, no pole at infinity) and for all s in the closed right-half plane (stability,
no pole in closed right-half plane). Hence we have to check whether the determinant is
non-zero for all s ∈ C= ∪ C> ∪ {∞}. The determinant of
 
δ1 aδ2
1 + s+1 s+1
I − M (s)∆ =  
−aδ1 δ2
s+1
1 + s+1

is easily calculated to
1
s2 + (2 + δ1 + δ2 )s + (1 + δ1 + δ2 ) + (a2 + 1)δ1 δ2 .


(s + 1) 2

It does not have a zero at ∞. Moreover, its finite zeros are certainly confined to C< if
and only if
2 + δ1 + δ2 > 0, (1 + δ1 + δ2 ) + (a2 + 1)δ1 δ2 > 0.
For a = 10, Figure 33 depicts the region of parameters where this condition is not true.

Let us now first concentrate on one uncertainty at a time. For δ2 = 0, the stability
conditions holds for all δ1 in the big interval (−1, ∞). The same holds for δ1 = 0 and
δ2 ∈ (−1, ∞).

Let us now vary both parameters. If we try to find the largest r such that stability is
preserved for all parameters with (4.7), we have to inflate a box around zero until it hits
the region of instability as shown in Figure 33. (Why?) For the parameter a = 10, the
largest r turns out to be 0.1, and this value shrinks with increasing a.

In summary, the analysis with a single varying uncertainty (δ1 = 0 or δ2 = 0) just gives
a wrong picture of the robust stability region for common variations.

75
 Region of Instability for a=10
1

0.8

0.6

0.4

0.2

−0.2

−0.4

−0.6

−0.8

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

Figure 33: Dotted: Region of destabilizing parameters (δ1 , δ2 )

It is important to observe that we could very easily explicitly determine the region of
stability and instability for this specific problem. Since this is by no means possible in
general, we need to have a general tool that can be applied as well to arrive at similar
insights for more sophisticated structures. This is the goal in the theory to be developed
in the next sections.

4.2 Types of Uncertainties of System Components

Uncertainties that can be dealt with by the theory to be developed include parametric
and LTI dynamic uncertainties. Parametric uncertainties are related to variations of real
parameters (mass, spring constants, damping,...) in a system model, whereas LTI dynamic
uncertainty should capture unmodeled dynamics of a system.

4.2.1 Parametric Uncertainties

Consider the simple mechanical system depicted in Figure 34 modeled as

m1 ξ¨1 = u − f, m2 ξ¨2 = f, f = k(ξ1 − ξ2 ), y = ξ2

with nominal values m1 = m2 = k = 1.

76
 Position
Position
Measurement
ξ1
y = ξ2
Control
Force m1 m2
u Spring k

Figure 34: Simple mechanical system.

In the classical feedback interconnection (Figure 17) the controller

29.434(s + 0.2329)(s2 − 0.7687s + 2.917)

K(s) =
(s + 9.843)(s + 2.227)(s2 + 1.81s + 4.063)
is stabilizing if m1 , m2 , k are precisely known, which might be unrealistic.

Let us now suppose that m1 , m2 , k are only known within 20% of their nominal value,
which means that

m1 ∈ [0.8, 1.2], m2 ∈ [0.8, 1.2], k ∈ [0.8.1.2].

Each parameter triple defines another system. Hence, effectively, we consider a whole
family of systems. A fundamental question:

Is the controller robustly stabilizing: Does it internally stabilize all possible systems in
our family?

The Robust Control Toolbox offers a very simple approach in order to handle parametric
uncertainties. Instead of defining parameters as real numbers, just define them as un-
certain atom objects (command ureal) and build your interconnection in the same
fashion as before. The result is an uncertain system object. With usubs one can then
easily pick one particular system (or with usample a randomly chosen sample) out of the
whole family, and then test whether K stabilizes the resulting finitely many systems:

m1 = ureal(’m1’, 1, ’percent’, 20);

m2 = ureal(’m2’, 1, ’percent’, 20);
k = ureal(’k’, 1, ’percent’, 20);
s = zpk(’s’);
G1 = 1/(m1 * s^2);
G2 = 1/(m2 * s^2);
systemnames = ’G1 G2 k’;
inputvar = ’[u]’;
outputvar = ’[G2]’;

77
input_to_G1 = ’[u - k]’;
input_to_G2 = ’[k]’;
input_to_k = ’[G1 - G2]’;
G = sysic; % builds uncertain subsystem

systemnames = ’G’;
inputvar = ’[d; n; r; u]’;
outputvar = ’[G + d - r; r - n - d - G]’;
input_to_G = ’[u]’;
olic = sysic; % builds uncertain interconnection

% nominal stabilizing controller

K = 29.434 * (s + 0.2329) * (s^2 - 0.7687*s + 2.917) / (s + 9.843) / ...
(s + 2.227) / (s^2 + 1.81*s + 4.063);

clic = lft(olic, K);

cl0 = usubs(clic, ’m1’, 1, ’m2’, 1, ’k’, 1); % nominal system
clr = usample(clic, 10); % some samples
max(real(eig(cl0)));
plot(squeeze(eig(clr)),’*’);

Many familiar commands for LTI systems, like bode, are overlayed for uncertain systems.
Typically, the operation is performed for the nominal system and a specific number of
random samples (1 + 20 for bode). This provides a good intuition of sensitivity against
parameter variations.

In our example the closed loop poles for internal stability resulting from the above code
are depicted in Figure 36. However the samples of bode diagramms and step responses
of the transfer matrix from reference to error depicted in Figure 35 indicate that the
controller K does not internally stabilize all possible systems in our family.

4.2.2 Dynamic Uncertainties

One can estimate the frequency response of a real stable SISO plant by injecting sinusoidal
signals. If performing measurements at one frequency, one does usually not obtain just one
complex number that could be taken as an estimate for the plant’s response at frequency
ω, but, instead, it’s a whole set of complex numbers that is denoted by H(ω). Such an
experiment would lead us to the conclusion that any proper and stable H(s) that satisfies

H(iω) ∈ H(ω)

78
 Bode Diagram

From: r To: [+G+d−r]

100

50
Magnitude (dB)

−50
Reference to error
−100 Worst case
Nominal
−150
−2 −1 0 1 2
10 10 10 10 10
Frequency (rad/s)

Step Response

From: r To: [+G+d−r]

1.5

1
Amplitude

0.5

0
Step
−0.5 Worst case
Nominal
−1
0 10 20 30 40 50 60 70 80 90 100
Time (seconds)

Figure 35: Samples of bode diagramms and step responses of a system affected by para-
metric uncertainties

Closed loop poles for internal stability

2

−1

−2
−10 −9 −8 −7 −6 −5 −4 −3 −2 −1 0

Figure 36: Closed loop poles for internal stability of a system affected by parametric
uncertainties

79
is an appropriate model for the underlying plant. Since one can only perform a finite
number of measurements, H(ω) is usually only available at finitely many frequencies and
consists of finitely many points. Due to the lack of a nice description, this set is not
appropriate for the theory to be developed.

Hence we try to cover H(ω) with a set that admits a more appropriate description. This
means
H(ω) ⊂ G(iω) + W (iω)∆c for all ω ∈ R ∪ {∞},
where

• G(s) is a real rational proper transfer matrix

• ∆c is the open unit disk around 0: ∆c := {∆c ∈ C | |∆c | < 1}

• W (s) is a real rational weighting function.

At each frequency we have hence covered the unstructured set H(ω) with the disk
G(iω) + W (iω)∆c (4.9)
whose center is G(iω) and whose radius is |W (iω)|. In this description, G admits the
interpretation as a nominal system. The deviation from G(iω) is given by the circle
W (iω)∆c whose radius |W (iω)| varies with frequency. Hence, the weighting function W
captures how the size of the uncertainties depends upon the frequency; this allows to take
into account that models are, usually, not very accurate at high frequency; typically, W
is a high-pass filter.

Note that we proceeded similarly as in the parametric case: At frequency ω, we represent

the deviation by a nominal value G(iω) and by a W (iω)-weighted version of the open unit
disk.

The actual set of uncertainties is then defined as

∆ := {∆(s) ∈ RH∞ | ∆(iω) ∈ ∆c for all ω ∈ R ∪ {∞}}, (4.10)
the set of all proper and stable transfer functions that take their values along the imaginary
in the open unit disk. Note that this set is nothing but
{∆(s) ∈ RH∞ | k∆k∞ < 1} (4.11)
which is often called the open unit ball in RH∞ . (It is important to digest that (4.10)
and (4.11) are just the same!)

Finally, the uncertain system is described by

G∆ := G + W ∆ with ∆ ∈ ∆.
As for real uncertainties, we have obtained a whole set of systems that is now parameter-
ized by the uncertain dynamics ∆ in ∆.

80
Remark 4.1

1) The set of values ∆c must not necessarily be a circle for our results to apply. It
can be an arbitrary set that contains 0, such as a polytope. The required technical
hypothesis are discussed in Section 4.5. The deviation set

W (iω)∆c (4.12)

is then obtained by shrinking/stretching ∆c with factor |W (iω)|, and by rotating it

according to the phase of W (iω).

2) We could be even more general and simply allow for frequency dependent value sets
∆c (ω) that are not necessarily described as (4.12). Then we can more accurately
incorporate phase information about the uncertainty.

4.2.3 Mixed Uncertainties

Of course, in a certain system component, one might encounter both parametric and
dynamic uncertainties. As an example, suppose that the diagonal elements of
 
1 1
s+1 s+2
G(s) =   (4.13)
1 1
s+3 2s+1

are not affected by uncertainties, but the numerator 1 of the right upper element is affected
by perturbations such it actually equals
1 + W1 ∆1 1 W1
= + ∆1 where |∆1 | < 1
s+2 s+2 s+2
and the left lower element equals
1
(1 + W2 (s)∆2 (s)) where k∆2 k∞ < 1.
s+3
Here W1 is a constant weight, W2 (s) is a real rational weighting function, and ∆1 is a
parametric uncertainty in the unit interval (−1, 1), whereas ∆2 (s) is a (proper stable)
dynamic uncertainty that takes its values ∆(iω) on the imaginary axis in the open unit
disk {z ∈ C | |z| < 1}.

Hence, the uncertain system is described as

    
1 1 1
∆1 0 0 W1
G∆ (s) =  s+1 s+2  +   s+2  (4.14)
1 1 1
s+3 2s+1
0 ∆2 (s) s+3
W2 (s) 0

where ∆1 ∈ R is bounded as |∆1 | < 1, and ∆2 ∈ RH∞ is bounded as k∆2 k∞ < 1.

81
This amounts to
G∆ (s) = G(s) + ∆(s)W (s)
with a nominal system G(s), a matrix valued weighting W (s), and block-diagonally str-
cutured  
∆1 0
∆(s) =  .
0 ∆2 (s)

Note also that the diagonal blocks of ∆(s) have a different nature (∆1 is real, ∆2 is
dynamic) and they are bounded in size over frequency, where the bound for both is
rescaled to 1 by using weighting functions.

All these properties of ∆(s) (structure and bound on size) can be captured by simply
specifying a set of values that consists of complex matrices as follows:
 
∆1 0
∆c := {  | ∆1 ∈ R, |∆1 | < 1, ∆2 ∈ C, |∆2 | < 1}.
0 ∆2

The set of uncertainties ∆(s) is, again, just given by (4.10). We have demonstrated the
flexibility of the abstract setup (4.10) if we allow for subsets ∆c of matrices.

To conclude, we have brought the specific example back to the same general scheme: We
have parameterized the actual set of systems G∆ as G + ∆W where, necessarily, W has
to be chosen as a matrix, and the uncertainty ∆ ∈ ∆ turns out to admit a block-diagonal
structure.

Remark 4.2 As mentioned previously, we could considerably increase the flexibility in

uncertainty modeling if not only allowing to constrain the elements of the matrices in ∆c
by disks or real intervals; under the technical hypotheses as discussed in Section 4.5, all
the results to follow still remain valid.

4.2.4 Unstructured Uncertainties

Let us again consider the plant model G(s) in (4.13). Suppose this is an accurate model
at low frequencies, but it is known that the accuracy of all entries decreases at high
frequencies. With a (real rational) SISO high-pass filter W (s), the actual frequency
response is rather described as
G(iω) + W (iω)∆c
where  
∆11 ∆12
∆c =  
∆21 ∆22

82
is any complex matrix that is bounded as

k∆c k = σmax (∆c ) < 1.

Obviously, the size of the deviation W (iω)∆c at frequency ω is bounded as

kW (iω)∆c k = |W (iω)|k∆c k ≤ |W (iω)|.

(If |W (iω)| is not zero, we have a strict inequality.) Hence, the frequency dependence of
the size is captured by the dependence of |W (iω)| on ω. Note that this behavior is the
same for all the elements of the 2 × 2 matrix deviation W (iω)∆c .

We chose the maximal singular value to evaluate the size of the matrix W (iω)∆c since
this is an appropriate measure for the gain of the error at this frequency, and since the
theory to be developed will then turn out more satisfactory than for other choices.

Let us now subsume this specific situation in our general scenario: We take

∆c := {∆c ∈ C2×2 | k∆c k < 1}

and describe the uncertain system as

G + W∆

where ∆ ∈ ∆ as given in (4.10). Since ∆c is a set of full matrices without any specific
structural aspects, this type of uncertainty is called unstructured or a full-block uncer-
tainty.

4.2.5 Unstructured versus Structured Uncertainties

Continuing with the latter example, we might know more about the individual deviation
of each element of G(s). Suppose that, at frequency ω, the actual model is
 
W11 (iω)∆11 W12 (iω)∆12
G(iω) +   (4.15)
W21 (iω)∆21 W22 (iω)∆22

where the complex numbers ∆jk satisfy

|∆11 | < 1, |∆12 | < 1, |∆21 | < 1, |∆22 | < 1, (4.16)

and the real rational (usually high-pass) SISO transfer functions W11 (s), W12 (s), W21 (s),
W22 (s) capture the variation of size over frequency as in our SISO examples.

83
We could rewrite (4.15) as
 
  ∆11 0

W11 (iω) 0 W12 (iω) 0 ∆
21 0 
G(iω) +   
 
0 W21 (iω) 0 W22 (iω)  0 ∆12 


0 ∆22

or as
  
   ∆11 0 0 0

1 0

W11 (iω) 0 W12 (iω) 0  0 ∆21 0 0 1 0
 
G(iω) +   .

 
0 W21 (iω) 0 W22 (iω)  0 0 ∆12 0 
0 1

0 0 0 ∆22 0 1

Obviously, we have to live with structure in the uncertainty if we would like to use dif-
ferent weightings for the various elements in the uncertainty. However, the two displayed
structures differ: In the first case, we have two 2 × 1 blocks on the diagonal, whereas in
the second one we have four 1 × 1 blocks on the diagonal. What should we choose for?

As mentioned above, we would like to take as a measure of size of ∆c the largest singular
value; in fact, ∆c is supposed to be bounded as k∆c k < 1. Then we observe that
 
∆11 0
 
∆ 0 
 21
 <1


 0 ∆12 
 
0 ∆22

is equivalent to
   
∆ ∆12
 11  < 1,   < 1 or |∆11 |2 + |∆21 |2 < 1 and |∆12 |2 + |∆22 |2 < 1.
∆21 ∆22

This is not what we want. If we insist on (4.16), we have to work with the second structure
since  
∆11 0 0 0
 
 0 ∆ 0 0 
21
 <1
 

 0 0 ∆ 12 0 
 
0 0 0 ∆22
is equivalent to (4.16).

84
4.3 Summary on Uncertainty Modeling for Components

For MIMO models of system components, we can work with only rough descriptions of
modeling errors. Typically, the uncertain component is described as

G∆ = G + W1 ∆W2 (4.17)

with real rational weighting matrices W1 and W2 and full block or unstructured uncer-
tainties ∆ that belongs to ∆ as defined in (4.10) where

∆c := {∆c ∈ Cp×q | k∆c k < 1}.

If we choose for a more refined description of the uncertainties, the uncertainties in (4.17)
will admit a certain structure that is often block-diagonal. To be specific, the uncertainty
set ∆ will be given by (4.10) with
   


 δ1 0 


.
   





 . . 






   

 
  δ r

pj ×qj

∆c := ∆c =   | δj ∈ R, ∆j ∈ C , k∆c k < 1 .
 




 ∆1 




.
 





 . .







   

 
 0 ∆f 

Note that we have distinguished the real blocks δj that correspond to parametric uncer-
tainties from the complex blocks that are related to dynamic uncertainties. Note also
that
k∆c k < 1 just means |δj | < 1, k∆j k < 1.
The weighting matrices W1 and W2 capture the variation of the uncertainty with fre-
quency, and they determine how each of the blocks of the uncertainty appears in G∆ .

Finally, we have seen that there is a lot of flexibility in the choice of the structure. It is
mainly dictated by how refined one wishes to describe the individual uncertainties that
appear in the model.

4.4 Pulling out the Uncertainties

As we have seen in the example, a central ingredient in testing robust stability is to

calculate the transfer matrix ‘seen’ by ∆ in an open-loop interconnection. We would like
to explain how one can systematically perform these calculations.

85
 G∆
z w

Figure 37: Uncertain Component

z∆ w∆

G
z w

Figure 38: Uncertainty Pulled out of Component

4.4.1 Pulling Uncertainties out of Subsystems

First we observe that an interconnection is usually built from subsystems. These sub-
systems themselves might be subject to uncertainties. Hence we assume that they are
parameterized as G∆ with ∆ ∈ ∆ as shown in Figure 37.

In order to proceed one has to rewrite this system in the form as shown in Figure 38.

As an illustration, look at (4.14). This system can be rewritten as

z = (G22 + G21 ∆G12 )w

for
     
1 1 1
s+1 s+2
1 0 0 s+2
W1
G22 (s) =   , G21 (s) =   , G12 (s) =  .
1 1 1
s+3 2s+1
0 1 s+3
W2 (s) 0

If we define  
0 G12
G= ,
G21 G22

we observe that (4.14) is rewritten as

   
z∆ w∆
  = G  , w∆ = ∆z∆
z w

which are the algebraic relations that correspond to Figure 38.

86
The step of rewriting a system as in Figure 37 into the structure of Figure 38 is most
easily performed if just introducing extra signals that enter and leave the uncertainties.
We illustrate this technique with examples that are of prominent importance:

• Additive uncertainty:
    
z1 0 I w1
z = (G + ∆)w ⇐⇒  =   , w1 = ∆z1 .
z I G w

• Input multiplicative uncertainty.

    
z1 0 I w
z = G(I + ∆)w ⇐⇒   =    1  , w1 = ∆z1 .
z G G w

• Input-output multiplicative uncertainty.

z = (I + ∆1 )G(I + ∆2 )w

is equivalent to
    
z1 0 G G w1     
w1 ∆1 0 z
 1 .
    
 z2  =  0 0 I   w2  ,  =
w2 0 ∆2 z2
    
z I G G w

Let us show explicitly how to proceed in this example. Observe that z = (I +

∆1 )G(I + ∆2 )w can be written as

z = G(I + ∆2 )w + w1 , z1 = G(I + ∆2 )w, w1 = ∆1 z1 .

We have pulled out ∆1 . In a second step, we do the same with ∆2 . The above
relations are equivalent to

z = Gw + Gw2 + w1 , z2 = w, z1 = Gw + Gw2 , w1 = ∆1 z1 , w2 = ∆2 z2 .

The combination into matrix relations leads to the desired representation.

• Factor uncertainty. Let G2 have a proper inverse. Then

z = (G1 + ∆1 )(G2 + ∆2 )−1 w

is equivalent to
 
    w1    
z1 0 −G−1 G−1 w ∆
2 2   w2  ,  1  =  1  z1 .
 
 =
z I −G1 G−1 G1 G−1 w2 ∆2
 
2 2
w

87
 Again we perform the calculations for the purpose of illustration. We observe that
z = (G1 + ∆1 )(G2 + ∆2 )−1 w is nothing but

z = (G1 + ∆1 )ξ, (G2 + ∆2 )ξ = w

what can be rewritten as

z = G1 ξ + w1 , z1 = ξ, G2 ξ + w2 = w, w1 = ∆1 z1 , w2 = ∆2 z1 .

If we eliminate ξ via ξ = G−1

2 (w − w2 ), we arrive at

z = G1 G−1 −1 −1 −1
2 w − G1 G2 w2 + w1 , z1 = G2 w − G2 w2 , w1 = ∆1 z1 , w2 = ∆2 z1 .

Note that the manipulations are representatives of how to pull out the uncertainties, in
particular if they occur in the denominator as it happens in factor uncertainty. It is often
hard to pull out the uncertainties if just performing matrix manipulations. If we rather
use the input-output representations of systems including the signals, this technique is
often pretty straightforward to apply. As a general rule, blocks that occur in a rational
fashion can be pulled out. Finally, let us note that all these manipulations can also be
performed directly for a state-space description where the state and its derivative are as
well viewed as a signal.

Again, we include an example. Let

 
−1 + δ1 2
ẋ =  x
−1 −2 + δ2

denote a system with real uncertain parameters. This system can be written as

ẋ = Ax + Bw, z = Cx, w = ∆z

with        
−1 2 1 0 1 0 δ1 0
A= , B =  , C =  , ∆ =  .
−1 −2 0 1 0 1 0 δ2

Somewhat more general, suppose that

ẋ = A(δ)x + B(δ)w, z = C(δ)x + D(δ)w

where
 the matrices A(.), B(.), C(.), D(.) depend affinely on the parameter δ =
δ1 · · · δk . This just means that they can be represented as
     
k
A(δ) B(δ) A0 B0 X Aj Bj
 = + δj  .
C(δ) D(δ) C0 D0 j=1 Cj Dj

88
Let us factorize    
Aj Bj L1j  
 =  R1 R2 (4.18)
j j
Cj Dj L2j
 
L1j  
where   and Rj1 Rj2 have full column and row rank respectively. The original
L2j
system can be described as
    
ẋ A0 B0 L11 · · · L1k x
         
 z  C D
   0 0 L21 · · · L2k 
 w 
  w1 δ1 I 0 z1
 .   .  . 
 ..  =  ..   .. 
    
 z1  =  R1 R2 0 0   w1  ,
   1 2       
 .   . ...  . 
 ..   ..   ..  wk 0 δk I zk
    
1 2
zk Rk Rk 0 0 wk

wherethe sizes of the identity block in δj I is equal to the number of columns or rows of

L1  
 j  or R1 R2 respectively.
j j
L2j

Remark. We have chosen the factorization (4.18) such that the number of columns/rows
of the factors are minimal. This renders the size of the identity blocks in the uncertainty
minimal as well. One could clearly work with an arbitrary factorization; then, however,
the identity blocks will be larger and the representation is not as efficient as possible.

Again, we remark that we can represent any

    
ẋ A(δ) B(δ) x
 =  
z C(δ) D(δ) w
 
with elements that are rational functions (quotients of polynomials) of δ = δ1 · · · δk
without pole at δ = 0 by
    
ẋ A B1 B2 x
    
 z  =  C D D   w  , w∆ = ∆z∆
   1 11 12  
z∆ C2 D21 D22 w∆

where  
δ1 I 0

∆= ... 
.
 
0 δk I

89
 −1

w∆ z∆ d
r e
+ K G + +
y u

n
−1 +

Figure 39: Uncertain closed-loop interconnection

We observe that ∆ has a block-diagonal structure. Each block is given as

 
δj 0

δj I =  .. 
 . 

0 δj

and is said to be a real (δj ∈ R) block that is repeated if the dimension of the identity
matrix is at least two.

4.4.2 Pulling Uncertainties out of Interconnections

We have seen various possibilities how to represent (37) as (38) for components. Let us
now suppose this subsystem is part of a (possibly large) interconnection.

Again for the purpose of illustration, we come back to the tracking configuration as in
Figure 17 with a plant model G∆ . If we employ the representation in Figure 38, we arrive
at Figure 39.

How do we pull ∆ out of the interconnection? We simply disconnect ∆ and K to arrive

at Figure 40.

90
 −1

w∆ z∆ d
r e
+ G + +
y u

n
−1 +

Figure 40: Uncertain open-loop interconnection

It is then not difficult to obtain the corresponding algebraic relations as

 
w∆
    
z∆ G11 0 0 0 G12  d 
 
    
 e  =  G21 I 0 −I G22   n 
    
 
y G21 −I −I I −G22   r 

u
for the open-loop interconnection. The command sysic is very useful in automizing the
calculation of a state-space representation of this system.

After having determined this open-loop system, the uncertain uncontrolled system is
obtained by re-connecting the uncertainty as

w∆ = ∆z∆ .

Note that the uncertainty for the component is just coming back as uncertainty for the
interconnection. Hence the structure for the interconnection is simply inherited.

This is different if several components of the system are affected by uncertainties that are
to be pulled out. Then the various uncertainties for the components will appear on the
diagonal of an uncertainty block for the interconnection.

Let us again look at an illustrative example. Suppose we would like to connect a SISO
controller K to a real system. Since K has to be simulated (in a computer), the actu-
ally implemented controller will differ from K. Such a variation can be captured in an
uncertainty description: the implemented controller is

K(I + ∆K )

91
 -1

z2 w2 z1 w1
∆K ∆G
d
r y u e
+ + K G + + +

n
-1 +

Figure 41: Uncertain closed-loop interconnection

where ∆K is a proper and stable transfer matrix in some class that captures our knowledge
of the accuracy of the implemented controller, very similar to what we have been discussing
for a model of the plant.

Let us hence replace K by K(I + ∆K ) in the interconnection in Figure 25. Since this is
a multiplicative input uncertainty, we arrive at Figure 41.

If we disconnect K, ∆K , ∆G , the resulting open-loop interconnection is given as

 
    w1
z 0 0 0 0 0 I 
 1 

 z   −I 0 −I −I I G  
 w2 

 2   
    d 
 e  =  I 0 I 0 −I G   
,

 n 
   
  
 u   0 0 0 0 0 I  
    r  
y −I I −I −I I G  
u

and the perturbation enters as

     
w z ∆ 0
 1  = ∆ 1 , ∆ =  G .
w2 z2 0 ∆K

Since the signals z1 , z2 and the signals w1 , w2 are different, ∆ admits a block-diagonal
structure with ∆G and ∆K appearing on its diagonal.

Remark. Suppose that two uncertainties ∆1 , ∆2 enter an interconnection as in Figure

42. Then they can be pulled out as
   
w ∆
 1  =  1 z
w2 ∆2

92
 w1
∆1
z

w2
∆2

z1
∆1
w
+
z2
∆2

Figure 42: Special Configurations

or as  
  z1
w= ∆1 ∆2  
z2
respectively. Instead, however, it is also possible to pull them out as
    
w ∆ 0 z
 1= 1  1 
w2 0 ∆2 z2

by simply neglecting the fact that ∆1 and ∆2 are entered by the same signal, or that the
outputs sum up to one signal.

In summary, uncertainties might be structured or not at the component level. If pulling

them out of an interconnection, the resulting uncertainty for the interconnection is block-
diagonal, and the uncertainties of the components appear, possibly repeated, on the
diagonal.

If pulling uncertainties out of an interconnection, they will automatically have

a block-diagonal structure, even if the component uncertainties are not struc-
tured themselves.

93
4.5 The General Paradigm

We have seen how to describe a possibly complicated interconnection in the form

      
z∆ w∆ P11 P12 P13 w∆
      
 z  = P  w  =  P21 P22 P23   w  . (4.19)
      
y u P31 P32 P33 u

Here w∆ , z∆ are the signals that are introduced to pull out the uncertainties, w, z are
generalized disturbance and controlled variable, and u, y are control input and measured
output respectively.

The uncertainties will belong to a set ∆ that consists of proper and stable transfer ma-
trices. The perturbed uncontrolled interconnection is obtained by re-connecting the un-
certainty as
w∆ = ∆z∆ with ∆ ∈ ∆.
This leads to
   
z w
  = S(∆, P )   =
y u
     
P22 P23 P   w
=   +  21  ∆(I − P11 ∆)−1 P12 P13    . (4.20)
P32 P33 P31 u

If we connect the controller to the unperturbed open-loop interconnection as

y = Ku,

we obtain
   
z w
 ∆=P ?K ∆=
z w
     
P11 P12 P   w
=   +  13  K(I − P33 K)−1 P31 P32   ∆  v.
P21 P22 P23 w

The controlled and perturbed interconnection is obtained through

w∆ = ∆z∆ and u = Ky.

It does not matter in which order we reconnect ∆ or K. This reveals a nice property of
linear fractional transformations:

S(∆, P ? K) = S(S(∆, P ), K).

94
Hence the closed loop system admits the descriptions

z = S(∆, P ? K)w = S(S(∆, P ), K)w.

So far we were sloppy in not worrying about inverses that occur in calculating star products
or about any other technicalities. Let us now get rigorous and include the exact hypotheses
required in the general theory. All our technical results are subject to these assumption.
Hence they need to be verified before any of the presented results can be applied.

Hypothesis 4.3

• P is a generalized plant: there exists a controller u = Ky that stabilizes (4.19) in

the sense of Definition 3.2.

• The set of uncertainties is given as

∆ := {∆(s) ∈ RH∞ | ∆(iω) ∈ ∆c for all ω ∈ R ∪ {∞}} (4.21)

where ∆c is a value set of complex matrices (motivating the index c for complex)
that defines the structure and the size of the uncertainties. This set ∆c has to be
star-shaped with center 0:

∆c ∈ ∆c ⇒ τ ∆c ∈ ∆c for all τ ∈ [0, 1]. (4.22)

• The direct feedthrough P11 and ∆c are such that

I − P11 (∞)∆c is non-singular for all ∆c ∈ ∆c . (4.23)

Comments on the hypothesis

As a fundamental requirement of any controller, it should stabilize an interconnection.
Hence there should at least exist a stabilizing controller, and the tests developed in Section
3 can be applied to verify this fact.

In the second hypothesis, we define the considered class of uncertainties to be all proper
and stable transfer matrices that take their values along the imaginary axis in the set
∆c . We recall the specific examples we have seen earlier to illustrate this concept. It is
very important to obey that this value has to be star-shapedness with center 0: If ∆c is
contained in ∆c , then the whole line τ ∆c , τ ∈ [0, 1], that connects ∆c with 0 belongs to
∆c as well. Note that this implies 0 ∈ ∆c such that the zero transfer matrix is always
in the class ∆; this is consistent with ∆c to be viewed as a deviation from a nominal
value. For sets of complex numbers, Figure 43 shows an example and a counterexample
for star-shapedness.

95
 Figure 43: Star-shaped with center 0? Left: Yes. Right: No.

Remark. If ∆c is a set of matrices whose elements are just supposed to be contained in

real intervals or in circles around 0 (for individual elements) or in a unit ball of matrices
(for sub-blocks), the hypothesis (4.22) of star-shapedness is automatically satisfied.

The last property implies that, for any ∆ in our uncertainty set ∆, I−P11 (∞)∆(∞) is non-
singular such that I − P11 ∆ does have a proper inverse. This is required to guarantee that
S(∆, P ) can be calculated at all (existence of inverse), and that it defines a proper transfer
matrix (properness of inverse). At this stage, we don’t have a systematic technique to
test whether (4.23) holds true or not; this will be the topic of Section 4.9.5. However, if
P11 is strictly proper, it satisfies P11 (∞) = 0 and (4.23) is trivially satisfied.

Comments on weightings

• Note that we assume all weightings that are required to accurately describe the
uncertainty size or structure to be absorbed in P . These weightings do not need
to obey any specific technical properties; they neither need to be stable nor even
proper. The only requirement is to ask P being a generalized plant - this is the
decisive condition to apply our results. Of course, a wrongly chosen weighting
might preclude the existence of a stabilizing controller, and hence it might destroy
this property; therefore, an adjustment of the weightings might be a possibility to
enforce that P is a generalized plant.

• We could as well incorporate a weighting a posteriori in P . Suppose that we actually

intend to work with ∆b that is related with ∆ by

∆ = W1 ∆W
b 2

96
 for real rational weightings W1 and W2 . Then we simply replace P by Pb given as
 
W2 P11 W1 W2 P12 W2 P13
 
Pb = 
 P 21 W1 P 22 P 23


P31 W1 P32 P33

and proceed with Pb and ∆.

b (Note that this just amounts to pulling out ∆
b in
z = W1 ∆W
b 2 w.)

Comments on larger classes of uncertainties

• We could allow for value sets ∆c (ω) that depend on the frequency ω ∈ R ∪ {∞} in
order to define the uncertainty class ∆. Then we require ∆c (ω) to be star-shaped
with star center 0 for all ω ∈ R ∪ {∞}.

• We could even just stay with a general set of ∆ of real rational proper and stable
transfer matrices that does not admit a specific description at all. We would still
require that this set is star-shaped with center 0 (∆ ∈ ∆ implies τ ∆ ∈ ∆ for all
τ ∈ [0, 1]), and that I − P22 (∞)∆(∞) is non-singular for all ∆ ∈ ∆.

4.6 What is Robust Stability?

We have already seen when K achieves nominal stability: K should stabilize P in the
sense of Definition 3.2.

Robust Stabilization
We say that K robustly stabilizes S(∆, P ) against the uncertainties ∆ ∈ ∆ if K stabilizes
the system S(∆, P ) for any uncertainty ∆ taken out of the underlying class ∆.

Robust Stability Analysis Problem

For a given fixed controller K, test whether it robustly stabilizes S(∆, P ) against all
uncertainties in ∆.

Robust Stability Synthesis Problem

Find a controller K that robustly stabilizes S(∆, P ) against all uncertainties in ∆.

Although these definitions seem as tautologies, it is important to read them carefully: If

we have not specified a set of uncertainty, it does not make sense to talk of a robustly
stabilizing controller. Hence we explicitly included in the definition that robust stability
is related to a well-specified set of uncertainties. Clearly, whether or not a controller
robustly stabilizes an uncertain system, highly depends on the class of uncertainties that
is considered. These remarks are particularly important for controller design: If one has
found a controller that robustly stabilizes an uncertain system for a specific uncertainty

97
class, there is, in general, no guarantee whatsoever that such a controller is robustly
stabilizing for some other uncertainty class.1

4.7 Robust Stability Analysis

4.7.1 Simplify Structure

Starting from our general paradigm, we claim and prove that robust stability can be
decided on the basis of the transfer matrix that is seen by ∆. Let us hence introduce the
abbreviation    
N11 N12 M N12
P ?K =N = = 
N21 N22 N21 N22
where we highlight M ; this is the block referred to as the transfer matrix seen by the
uncertainty.

Theorem 4.4 If K stabilizes P , and if

I − M ∆ has a proper and stable inverse for all ∆ ∈ ∆,

then K robustly stabilizes S(∆, P ) against ∆.

Recall that I − M (s)∆(s) has a proper and stable inverse if and only if I − M (s)∆(s) is
non-singular for all s ∈ C= ∪ C> ∪ {∞} or, equivalently,

det(I − M (s)∆(s)) 6= 0 for all s ∈ C= ∪ C> ∪ {∞}.

Note that it is difficult to verify the latter property for all ∆ ∈ ∆. The crux of this result
is a structural simplification: Instead of investigating S(∆, P ) that depends possibly in
a highly involved fashion on ∆, we only need to investigate I − M ∆ which is just linear
in ∆. Hence, it is sufficient to look at this generic structure for all possible (potentially
complicated) interconnections.
1
These remarks are included since, recently, it has been claimed in a publication that robust control
design techniques are not useful, based on the following argument: The authors of this paper constructed
a controller that robustly stabilizes a system for some uncertainty class, and they tested this controller
against another class. It turned out that the robustness margins (to be defined later) for the alternative
uncertainty class are poor. The authors conclude that the technique they employed is not suited to
design robust controllers. This is extremely misleading since the actual conclusion should read as follows:
There is no reason to expect that a robustly stabilizing controller for one class of uncertainties is also
robustly stabilizing for another (possibly unrelated) class of uncertainties. Again, this is logical and seems
almost tautological, but we stress these points since the earlier mentioned severe confusions arose in the
literature.

98
Proof. Taking any ∆ ∈ ∆, we have to show that u = Ky stabilizes the system in (4.20)
in the sense of Definition 3.2. At this point we benefit from the fact that we don’t need
to go back to the original definition, but, instead, we can argue in terms of input-output
descriptions as formulated in Theorem 3.5. We hence have to show that
   
z w
  = S(∆, P )   , u = Kv + v1 , v = y + v2 (4.24)
y u
  

w z
   
 v1  →  u . Clearly, we can re-represent this
defines a proper and stable system    

v2 v
system as
   
z∆ w∆
   
 z  = P  w  , u = Kv + v1 , v = y + v2 , w∆ = ∆z∆ . (4.25)
   
y u

Recall that K stabilizes P . Hence the relations

   
z∆ w∆
   
 z  = P  w  , u = Kv + v1 , v = y + v2 ,
   
y u

define a stable LTI system

    
z∆ M N12 H13 H14 w∆
    
 z  N N22 H23 H24   w 
  21
= . (4.26)
  
 
 u   H31 H32 H33 H34 
  v1 
 
  
v H41 H42 H43 H44 v2
 
M N12
In addition to N =  , several other blocks appear in this representation whose
N21 N22
structure is not important; the only important fact is that they are all proper and stable.

If we reconnect w∆ = ∆z∆ in (4.26), we arrive at an alternative representation of (4.25)

or of (4.24) that reads as
       
z N22 H23 H24 N21    w 
−1
     
 u  =  H32 H33 H34  +  H31  ∆(I − M ∆) N12 H13 H14    v1  .
 
     
v H42 H43 H44 H41 v2

99
The essential point: Since both ∆ and (I − M ∆)−1 are proper and stable, and since, as
mentioned above, all the other blocks occurring in this formula are proper and stable as
well, this system defines a proper and stable transfer matrix as we had to prove.

4.7.2 Reduction to Non-Singularity Test on Imaginary Axis

Recall that we need to verify whether I − M ∆ has a proper and stable inverse; for that
purpose one has to check whether the matrix I − M ∆ itself does not have zeros in the
closed right-half plane including infinity. Hence we need to check

det(I − M (s)∆(s)) 6= 0 for all s ∈ C= ∪ C> ∪ {∞}, ∆ ∈ ∆. (4.27)

This is complicated since we have to scan the full right-half plane, and we have to perform
the test for all dynamic uncertainties under consideration.

The following result shows that it suffices to test I − M (s)∆c for non-singularity only for
s = iω with ω ∈ R ∪ {∞}, and only for ∆c ∈ ∆c . Hence this reduces the original problem
to a pure problem in linear algebra, what might considerably simplify the test.

Let us first formulate the precise result.

Theorem 4.5 Suppose M is a proper and stable transfer matrix. If

det(I − M (iω)∆c ) 6= 0 for all ∆c ∈ ∆c , ω ∈ R ∪ {∞}, (4.28)

then
I − M ∆ has a proper and stable inverse for all ∆ ∈ ∆. (4.29)

Before we provide a formal proof, we would like to provide some intuition why the star-
shapedness hypothesis plays an important role in this result. Let us hence assume that
(4.28) is valid. Obviously, we can then conclude that

det(I − M (λ)∆(λ)) 6= 0 for all λ ∈ C= ∪ {∞}, ∆ ∈ ∆, (4.30)

since ∆(λ) is contained in ∆c if λ ∈ C= ∪ {∞} and ∆ ∈ ∆. Note that (4.27) and (4.30)
just differ by replacing C= ∪C> ∪{∞} with C= ∪{∞}. Due to C= ∪{∞} ⊂ C= ∪C> ∪{∞},
it is clear that (4.27) implies (4.30). However, we need the converse: We want to conclude
that (4.30) implies (4.27), and this is the non-trivial part of the story.

Why does this implication hold? We have illustrated the following discussion in Figure
44. The proof is by contradiction: Assume that (4.27) is not valid, and that (4.30) is true.
Then there exists a ∆ ∈ ∆ for which I − M ∆ has a zero s1 in C> (due to (4.27)) where

100
 Im(z)
crosses axis for some
τ0 at point s0 = iω0

s1
Re(z)

all zeros in one zero in

lhp for τ → 0 rhp for τ = 1

Figure 44: Movements of zeros

s1 is certainly not contained in C= ∪ {∞} (due to (4.30)). For this single ∆, we cannot
detect that it is destabilizing without scanning the right-half plane. However, apart from
∆, we can look at all the uncertainties τ ∆ obtained by varying τ ∈ [0, 1]. Since ∆c is
star-shaped, all these uncertainties are contained in the set ∆ as well. (Check that!) Let
us now see what happens to the zeros of

det(I − M (s)[τ ∆(s)]).

For τ = 1, this function has the zero s1 in C> . For τ close to zero, one can show that
all its zeros must be contained in C< . (The loop is stable for τ = 0 such that it remains
stable for τ close to zero since, then, the perturbation τ ∆ is small as well; we provide
a proof that avoids these sloppy reasonings.) Therefore, if we let τ decrease from 1 to
0, we can expect that the unstable zero s1 has to move from C> to C< . Since it moves
continuously, it must hit the imaginary axis on its way for some parameter τ0 .2 If this
zero curve hits the imaginary axis at iω0 , we can conclude that

det(I − M (iω0 )[τ0 ∆(iω0 )]) = 0.

Hence τ0 ∆(s) is an uncertainty that is still contained in ∆ (star-shapeness!) and for which
I − M [τ0 ∆] has a zero at iω0 . We have arrive at the contradiction that (4.30) cannot be
true either.
2
Contrary to what is often stated in the literature, this is not an elementary continuity argument!

101
It is interesting to summarize what these arguments reveal: If we find a ∆ such that
˜ for
(I − M ∆)−1 has a pole in the open right-half plane, then we can also find another ∆
which (I − M ∆) ˜ −1 has a pole on the imaginary axis.

Comments on larger classes of uncertainties

• If ∆c (ω) depends on frequency (and is star-shaped), we just have to check

det(I − M (iω)∆c ) 6= 0 for all ω ∈ R ∪ {∞}, ∆c ∈ ∆c (ω)

in order to conclude (4.29). The proof remains unchanged.

• If ∆ is a general set of real rational proper and stable transfer matrices without
specific description, we have to directly verify (4.30) in order to conclude (4.29).

Let us now provide the rigorous arguments to finish the proof of Theorem 4.5.

Proof. Recall that it remains to show (4.30) ⇒ (4.27) by contradiction. Suppose (4.30)
holds, but (4.27) is not valid. Hence there exists a s1 ∈ C> , s1 6∈ C= ∪{∞}, and a ∆1 ∈ ∆
such that
det(I − M (s1 )∆1 (s1 )) = 0. (4.31)
If we can show that there exists a s0 ∈ C= and a τ0 ∈ [0, 1] for which

det(I − M (s0 )[τ0 ∆1 (s0 )]) = 0, (4.32)

we arrive at a contradiction to (4.30) since τ0 ∆ ∈ ∆ and s0 ∈ C= .

To find s0 and τ0 , let us take the realization

 
A B
M ∆1 =  .
C D

We obviously have  
A B
I − M [τ ∆1 ] =  
−τ C I − τ D
and the well-known Schur formula leads to
det(I − τ D)
det(I − M (s)[τ ∆(s)]) = det(sI − A(τ )) (4.33)
det(sI − A)
if we abbreviate
A(τ ) = A + B(I − τ D)−1 τ C.
If we apply (4.30) for s = ∞ and ∆ = τ ∆1 , we infer that det(I − τ D) 6= 0 for all τ ∈ [0, 1].
In addition, A is stable such that det(s1 I −A) 6= 0. If we hence combine (4.33) and (4.31),
we arrive at
det(s1 I − A(1)) = 0 or s1 ∈ λ(A(1)).

102
Let us now exploit as a fundamental result the continuous dependence of eigenvalues of
matrices: Since A(τ ) depends continuously on τ ∈ [0, 1], there exists a continuous function
s(.) defined on [0, 1] and taking values in the complex plane such that

s(1) = s1 , det(s(τ )I − A(τ )) = 0 for all τ ∈ [0, 1].

(s(.) defines a continuous curve in the complex plane that starts in s1 and such that, for
each τ , s(τ ) is an eigenvalue of A(τ ).) Now we observe that A(0) = A is stable. Therefore,
s(0) must be contained in C< . We conclude: the continuous function Re(s(τ )) satisfies

Re(s(0)) < 0 and Re(s(1)) > 0.

Hence there must exist a τ0 ∈ (0, 1) with

Re(s(τ0 )) = 0.

Then s0 = s(τ0 ) and τ0 lead to (4.32) what is the desired contradiction.

For later purposes we can modify the proof to obtain the following result.

Lemma 4.6 Let P be stable with kP k∞ ≤ 1. Suppose K stabilizes P , that K has no

poles in C0 and satisfies kK(iω)k < 1 for all ω ∈ [0, ∞]. Then K is stable.

Proof. Choose ω ∈ [0, ∞] and τ ∈ [0, 1] arbitrary. Then we have

kτ P (iω)K(iω)k ≤ P (iω)k kK(iω)k < 1.

This implies
det(I − τ P (iω)K(iω)) 6= 0.
Choose again minimal realizations (Ap , Bp , Cp , Dp ) and (AK , BK , CK , DK ) of P and K
and abbreviate  
  Ap Bp CK Bp DK
A B  
PK =   :=  0 AK B K
.
C D
 
Cp Dp CK Dp DK
Define again A(τ ) := A + B(I − τ D)−1 τ C to obtain with the Schur formula

det(I − τ D)
det(I − P (s)[τ K(s)]) = det(sI − A(τ )).
det(sI − A)

103
Set now Si := (I − DK Dp )−1 and So := (I − Dp DK )−1 then observe that

A(1) = A + BSo C
   
Ap Bp CK BD  
=  +  p K  So Cp Dp CK
0 AK BK
   
Ap 0 Bp Si DK Cp Bp Si CK
= + 
0 AK BK So Cp BK So Dp CK
     
Ap 0 B 0 S Si DK 0 CK
= + p  i  
0 AK 0 BK So Dp So Cp 0
    −1  
Ap 0 Bp 0 I −DK 0 CK
= +    
0 AK 0 BK −Dp I Cp 0

is Hurwitz since K stabilizes

 P . With
 the same idea as in the proof of Theorem 4.5 we can
Ap Bp CK
conclude that A(0) =   is Hurwitz and hence AK is Hurwitz. This means
0 AK
that K is stable.

4.7.3 The Central Test for Robust Stability

We can easily combine Theorem 4.4 with Theorem 4.5 to arrive at the fundamental robust
stability analysis test for controlled interconnections.

Corollary 4.7 If K stabilizes P , and if

det(I − M (iω)∆c ) 6= 0 for all ∆c ∈ ∆c , ω ∈ R ∪ {∞}, (4.34)

then K robustly stabilizes S(∆, P ) against ∆.

Contrary to what is often claimed in the literature, the converse does in general not hold
in this result. Hence (4.34) might in general not be a tight condition. In practice and for
almost all relevant uncertainty classes, however, it often turns out to be tight. In order
to show that the condition is tight for a specific setup, one can simply proceed as follows:
if (4.34) is not true, try to construct a destabilizing perturbation, an uncertainty ∆ ∈ ∆
for which K does not stabilize S(∆, P ).

The construction of destabilizing perturbations is the topic of the next section.

104
4.7.4 Construction of Destabilizing Perturbations

As already pointed out, this section is related to the question in how far condition (4.34)
in Theorem 4.7 is also necessary for robust stability. We are not aware of definite answers
to this questions in our general setup, but we are aware of some false statements in the
literature! Nevertheless, we do not want to get into a technical discussion but we intend
to take a pragmatic route in order to construct destabilizing perturbations.

Let us assume that (4.34) is not valid. This means that we can find a complex matrix
∆0 ∈ ∆c and a frequency ω0 ∈ R ∪ {∞} for which

I − M (iω0 )∆0 is singular.

First step in constructing a destabilizing perturbation

Find a real rational proper and stable ∆(s) with

∆(iω0 ) = ∆0 and ∆(iω) ∈ ∆c for all ω ∈ R ∪ {∞}. (4.35)

This implies that ∆ is contained in our class ∆, and that I − M ∆ does not have a proper
and stable inverse (since I − M (iω0 )∆(iω0 ) = I − M (iω0 )∆0 is singular.)

Comments
Note that the construction of ∆ amounts to solving an interpolation problem: The func-
tion ∆(s) should be contained in our class ∆ and it should take the value ∆0 at the point
s = iω0 .

This problem has a trivial solution if ∆0 is a real matrix. Just set

∆(s) := ∆0 .

If ∆0 is complex, this choice is not suited since it is not contained in our perturbation
class of real rational transfer matrices. In this case we need to do some work. Note that
this was the whole purpose of Lemma ?? if ∆c is the open unit disk in the complex plane.
For more complicated sets (such as for block diagonal structures) we comment on the
solution of this problem later.

Second step in constructing a destabilizing perturbation

For the constructed ∆, check whether K stabilizes S(∆, P ) by any of our tests developed
earlier. If the answer is no, we have found a destabilizing perturbation. If the answer is
yes, the question of whether K is robustly stabilizing remains undecided.

Comments
In most practical cases, the answer will be no and the constructed ∆ is indeed destabiliz-
ing. However, one can find examples where this is not the case, and this point is largely
overlooked in the literature.

105
Let us provide (without proofs) conditions under which we can be sure that ∆ is destabi-
lizing:

• If ω0 = ∞ then ∆ is destabilizing.

• If ω0 is finite, if
 
A − iω0 I 0 B1 B2
 
0 ∆0 −I 0  has full row rank,


C1 −I D11 D12

and if  
A − iω0 I 0 B1
 
 0 ∆0 −I 
 has full column rank,
 


 C1 −I D11 

C2 0 D21
then ∆ is destabilizing.

• This latter test can be re-formulated in terms of P . If ω0 is finite, if iω0 is not a

pole of P (s), if
 
I − P11 (iω0 )∆0 P12 (iω) has full row rank,

and if  
I − ∆0 P11 (iω0 )
  has full column rank,
P21 (iω0 )
then ∆ is destabilizing.

Note that these conditions are very easy to check, and they will be true in most practical
cases. If they are not valid, the question of robust stability remains undecided. The failing
of the latter conditions might indicate that the process of pulling out the uncertainties
can be performed in a more efficient fashion by reducing the size of ∆.

4.8 Important Specific Robust Stability Tests

Testing (4.34) can be still pretty difficult in general, since the determinant is a complicated
function of the matrix elements, and since we still have to test an infinite number of
matrices for non-singularity.

On the other hand, very many robust stability test for LTI uncertainties have (4.34) at
their roots, and this condition can be specialized to simple tests in various interesting
settings. We can only touch upon the wealth of consequences.

106
4.8.1 M is Scalar

If it happens that M (iω) has dimension 1×1, then ∆c is simply a set of complex numbers.
In this case (4.34) just amounts to testing 1−M (iω)∆c 6= 0 for all ∆c ∈ ∆c . This amounts
to testing, frequency by frequency, whether the set

M (iω)∆c

contains 1 or not. If no, condition (4.34) holds true and we conclude robust stability.
If 1 is contained in this set for some frequency, (4.34) fails, and we might construct a
destabilizing uncertainty as in Section 4.7.4.

In many cases, ∆c is the open unit circle in the complex plane. If 1 ∈ M (iω0 )∆0 , Lemma
?? allows us to construct a proper real rational stable ∆(s) with ∆(iω0 ) = ∆0 . This is
the candidate for a destabilizing perturbation as discussed in Section 4.7.4.

4.8.2 The Small-Gain Theorem

Let us be specific and assume that the dimension of the uncertainty block is p × q. Then
we infer
∆c ⊂ Cp×q .
No matter whether or not ∆c consists of structure or unstructured matrices, it will cer-
tainly be a bounded set. Let us assume that we have found an r for which any

∆c ∈ ∆c satisfies k∆c k < r.

In order to check (4.34), we choose an arbitrary ω ∈ R ∪ {∞}, and any ∆c ∈ ∆c .

We infer that
det(I − M (iω)∆c ) 6= 0 (4.36)
is equivalent, by the definition of eigenvalues, to

1 6∈ λ(M (iω)∆c ). (4.37)

1 is certainly not an eigenvalue of M (iω)∆c if all these eigenvalues are in absolute value
smaller than 1. Hence (4.37) follows from

ρ(M (iω)∆c ) < 1. (4.38)

Since the spectral radius is smaller than the spectral norm of M (iω)∆c , (4.38) is implied
by
kM (iω)∆c k < 1. (4.39)

107
Finally, since the norm is sub-multiplicative, (4.39) follows from

kM (iω)kk∆c k < 1. (4.40)

At this point we exploit our knowledge that k∆c k < r to see that (4.40) is a consequence
of
1
kM (iω)k ≤ . (4.41)
r

We have seen that all the properties (4.37)-(4.41) are sufficient conditions for (4.34) and,
hence, for robust stability.

Why have we listed all these conditions in such a detail? They all appear - usually
separately - in the literature under the topic ‘small gain’. However, these conditions are
not often related to each other such that it might be very confusing what the right choice
is. The above chain of implications gives the answer: They all provide sufficient conditions
for (4.36) to hold.

Recall that we have to guarantee (4.36) for all ω ∈ R ∪ {∞}. In fact, this is implied if
(4.41) holds for all ω ∈ R ∪ {∞}, what is in turn easily expressed as kM k∞ ≤ 1r .

Theorem 4.8 If any ∆c ∈ ∆c satisfies k∆c k < r, and if

1
kM k∞ ≤ , (4.42)
r
then I − M ∆ has a proper and stable inverse for all ∆ ∈ ∆.

Again, one can combine Theorem 4.8 with Theorem 4.4 to see that (4.42) is a sufficient
condition for K to robustly stabilize S(∆, P ).

Corollary 4.9 If K stabilizes P , and if

1
kM k∞ ≤ ,
r
then K robustly stabilizes S(∆, P ) against ∆.

We stress again that (4.42) is only sufficient; it neglects any structure that might be
characterized through ∆c , and it only exploits that all the elements of this set are bounded
by r.

Remarks. Note that this result also holds for an arbitrary class ∆ of real rational proper
and stable matrices (no matter how they are defined) if they all satisfy the bound

k∆k∞ < r.

108
Moreover, we are not at all bound to the specific choice of k.k = σmax (.) as a measure
for the size of the underlying complex matrices. We could replace (also in the definition
of k.k∞ ) the maximal singular value by an arbitrary norm on matrices that is induced
by vector norms, and all our results remain valid. This would lead to another bunch of
small gain theorems that lead to different conditions. As specific examples, think of the
maximal row sum or maximal column sum which are both induced matrix norms.

4.8.3 Full Block Uncertainties

Let us suppose we have an interconnection in which only one subsystem is subject to

unstructured uncertainties. If this subsystem is SISO system, we can pull out the uncer-
tainty of the interconnection and we end up with an uncertainty block ∆ of dimension
1 × 1. If the subsystem is MIMO, the block will be matrix valued. Let us suppose that
the dimension of this block is p × q, and that is only restricted in size and bounded by r.

In our general scenario, this amounts to

∆c = {∆ ∈ Cp×q | k∆k < r}.

In other words, ∆ simply consists of all real rational proper and stable ∆(s) whose H∞
norm is smaller than r:
p×q
∆ := {∆ ∈ RH∞ | k∆k∞ < r}. (4.43)

1
Recall from Theorem 4.8: kM k∞ ≤ r
implies that I −M ∆ has a proper and stable inverse
for all ∆ ∈ ∆.

The whole purpose of this section is to demonstrate that, since ∆ consists of unstructured
uncertainties, the converse holds true as well: If I − M ∆ has a proper and stable inverse
for all ∆ ∈ ∆, then kM k∞ ≤ 1r .

1
Theorem 4.10 Let ∆ be defined by (4.43). Then kM k∞ ≤ r
holds true if and only if
I − M ∆ has a proper and stable inverse for all ∆ ∈ ∆.

We can put it in yet another form: In case that

1
kM k∞ > , (4.44)
r
we can construct - as shown in the proof - a real rational proper and stable ∆ with
k∆k∞ < r such that

I − M ∆ does not have a proper and stable inverse. (4.45)

109
This construction leads to a destabilizing perturbation for (I − M ∆)−1 , and it is a can-
didate for a destabilizing perturbation of the closed-loop interconnection as discussed in
Section 4.7.4.

Proof. This is what we have to do: If (4.44) holds true, there exists a ∆ ∈ ∆ with (4.45).

First step. Suppose that we have found ω0 ∈ R ∪ {∞} with kM (iω0 )k > 1r (which exists
by (4.44)). Recall that kM (iω0 )k2 is an eigenvalue of M (iω0 )M (iω0 )∗ . Hence there exists
an eigenvector u 6= 0 with

[M (iω0 )M (iω0 )∗ ]u = kM (iω0 )k2 u.

Let us define
1 ∗ u∗
v := M (iω0 ) u and ∆0 := v .
kM (iω0 )k2 kuk2
(Note that ∆0 has rank one; this is not important for our arguments.) We observe
kvk 1 1 1
k∆0 k ≤ ≤ 2
kM (iω0 )∗ uk ≤ <r
kuk kM (iω0 )k kuk kM (iω0 )k
and

[I − M (iω0 )∆0 ]u = u − M (iω0 )v =

1 ∗ kM (iω0 )k2
=u− M (iω0 )M (iω 0 ) u = u − u = 0.
kM (iω0 )k2 kM (iω0 )k2
We have constructed a complex matrix ∆0 that satisfies

k∆0 k < r and det(I − M (iω0 )∆0 ) = 0.

Second step. Once we are at this point, we have discussed in Section 4.7.4 that it suffices
to construct a real rational proper and stable ∆(s) satisfying

∆(iω0 ) = ∆0 and k∆k∞ < r.

Then this uncertainty renders (I − M ∆)−1 non-existent, non-proper, or unstable.

If ω0 = ∞ or ω0 = 0, M (iω) is real. Then u can be chosen real such that ∆0 is a real

matrix. Obviously, ∆(s) := ∆0 does the job.

Hence suppose ω0 ∈ (0, ∞). Let us now apply Lemma ?? to each of the components of
 
v1
u∗
   . 
. 
kuk2  . : There exist αj ≥ 0 and βj ≥ 0 with
= u1 · · · uq , v = 
vp

iω0 − αj iω0 − βj
uj = ±|uj | , vj = ±|vj | .
iω0 + αj iω0 + βj

110
Define the proper and stable
 
s−β1
±|v1 | s+β 1
..
   
s−αq
u(s) := ±|u1 | s−α
s+α1
1
· · · ±|uq | s+α q
, v(s) := 
 . .

s−βp
±|vp | s+βp

We claim that
∆(s) := v(s)u(s)
u ∗
does the job. It is proper, stable, and it clearly satisfies ∆(iω0 ) = v kuk2 = ∆0 . Finally,

observe that
q 2 q 2
2
X iω − αj
2
X u∗ 1
ku(iω)k = |uj | = |uj |2 = =
j=1
iω + αj j=1
kuk2 kuk2

and p p
2
2
X iω − βj
2
X
kv(iω)k = |vj | = |vj |2 = kvk2 .
j=1
iω + βj j=1

kvk
Hence k∆(iω)k ≤ ku(iω)kkv(iω)k = kuk
< r, what implies k∆k∞ < r.

4.9 The Structured Singular Value in a Unifying Framework

All our specific examples could be reduced to uncertainties whose values on the imaginary
axis admit the structure
 
p1 I 0
 .. 

 . 

 

 p n r I 

 

 δ1 I 

. .
∆c =  .  ∈ Cp×q (4.46)
 
 
δnc I
 
 
 

 ∆1


 .. 

 . 

0 ∆nf

and whose diagonal blocks satisfy

• pj ∈ R with |pj | < 1 for j = 1, . . . , nr ,

• δj ∈ C with |δj | < 1 for j = 1, . . . , nc ,

111
 • ∆j ∈ Cpj ×qj with k∆j k < 1 for j = 1, . . . , nf .

pj I is said to be a real repeated block, δj I is called a complex repeated block, and ∆j

is called a full (complex) block. The sizes of the identities can be different for different
blocks. Real full blocks usually do not occur and are, hence, not contained in the list.

Let us denote the set of all this complex matrices as ∆c . This set ∆c is very easy
to describe: one just needs to fix for each diagonal block its structure (real repeated,
complex repeated, complex full) and its dimension. If the dimension of pj I is rj , and the
dimension of δj I is cj , the µ-tools expect a description of this set in the following way:
 
−r1 0
 . .. 
 .. . 
 
 
 −rnr 0 
 
 
 c1 0 
 . .. 
 
blk =  .. . .
 
 cnc 0 
 
 
 p
 1 q1 

 . .. 
 .. . 
 
pnf qnf

Hence the row (−rj 0) indicates a real repeated block of dimension rj , whereas (cj 0)
corresponds to a complex repeated block of dimension cj , and (pj qj ) to a full block
dimension pj × qj .

Remark. In a practical example it might happen that the ordering of the blocks is
different from that in (4.46). Then the commands in the µ-Toolbox can still be applied
as long as the block structure matrix blk reflects the correct order and structure of the
blocks.

Remark. If ∆c takes the structure (4.46), the constraint on the size of the diagonal
blocks can be briefly expressed as k∆c k < 1. Moreover, the set r∆c consists of all
complex matrices ∆c that take the structure (4.46) and whose blocks are bounded in size
by r: k∆c k < r. Here r is just a scaling factor that will be relevant in introducing the
structured singular value.

The actual set of uncertainties ∆ is, once again, the set of all real rational proper and
stable ∆ whose frequency response takes it values in ∆c :

∆ := {∆(s) ∈ RH∞ | ∆(iω) ∈ ∆c for all ω ∈ R ∪ {∞}}.

Let us now apply the test (4.28) in Theorem 4.5. At a fixed frequency ω ∈ R ∪ {∞}, we

112
have to verify whether

I − M (iω)∆c is non-singular for all ∆c ∈ ∆c . (4.47)

This is a pure problem of linear algebra.

4.9.1 The Structured Singular Value

Let us restate the linear algebra problem we have encountered for clarity: Given the
complex matrix Mc ∈ Cq×p and the (relative open) set of complex matrices ∆c ⊂ Cp×q ,
decide whether
I − Mc ∆c is non-singular for all ∆c ∈ ∆c . (4.48)
The answer to this question is yes or no.

We modify the problem a bit. In fact, let us consider the scaled set r∆c in which we have
multiplied every element of ∆c with the factor r. This stretches or shrinks the set ∆c by
the factor r. Then we consider the following problem:

Determine the largest r such that I − Mc ∆c is non-singular for all ∆c in the

set r∆c . This largest value is denoted as r∗ .

In other words, calculate

r∗ = sup{r | det(I − Mc ∆c ) 6= 0 for all ∆c ∈ r∆c }. (4.49)

What happens here? Via the scaling factor r, we inflate or shrink the set r∆c . For small
r, I − Mc ∆c will be non-singular for any ∆c ∈ r∆c . If r grows larger, we might find some
∆c ∈ r∆c for which I − Mc ∆c will turn out singular. If no such r exists, we have r∗ = ∞.
Otherwise, r∗ is just the finite critical value for which we can assure non-singularity for
the set r∆c if r is smaller than r∗ . This is the reason why r∗ is called non-singularity
margin.

Remark. r∗ also equals the smallest r such that there exists a ∆c ∈ r∆c that renders
I −Mc ∆c singular. The above given definition seems more intuitive since we are interested
in non-singularity.

Definition 4.11 The structured singular value (SSV) of the matrix Mc with respect to
the set ∆c is defined as
1 1
µ∆c (Mc ) = = .
r∗ sup{r | det(I − Mc ∆c ) 6= 0 for all ∆c ∈ r∆c }

113
Remark. The non-singularity margin r∗ has been introduced by Michael Safonov,
whereas the structured singular value has been defined by John Doyle. Both concepts
are equivalent; the structured singular can be related in a nicer fashion to the ordinary
singular value what motivates its definition as the reciprocal of r∗ .

Let us now assume that we can compute the SSV. Then we can decide the original question
whether (4.48) is true or not as follows: We just have to check whether µ∆c (Mc ) ≤ 1. If
yes, then (4.48) is true, if no, then (4.48) is not true. This is the most important fact to
remember about the SSV.

Theorem 4.12 Let Mc be a complex matrix and ∆c an arbitrary (open) set of complex
matrices. Then

• µ∆c (Mc ) ≤ 1 implies that I − Mc ∆c is non-singular for all ∆c ∈ ∆c .

• µ∆c (Mc ) > 1 implies that there exists a ∆c ∈ ∆c for which I − Mc ∆c is singular.

Proof. Let us first assume that µ∆c (Mc ) ≤ 1. This implies that r∗ ≥ 1. Suppose that
there exists a ∆0 ∈ ∆c that renders I − Mc ∆0 singular. Since ∆c is relative open, ∆0 also
belongs to r∆c for some r < 1 that is close to 1. By the definition of r∗ , this implies that
r∗ must be smaller than r. Therefore, we conclude that r∗ < 1 what is a contradiction.

Suppose now that µ∆c (Mc ) > 1. This implies r∗ < 1. Suppose I − Mc ∆c is non-singular
for all ∆c ∈ r∆c for r = 1. This would imply (since r∗ was the largest among all r for
which this property holds) that r∗ ≥ r = 1, a contradiction.

It is important to note that the number µ∆c (Mc ) is depending both on the matrix Mc
and on the set ∆c what we explicitly indicate in our notation. For the computation of
the SSV, the µ-tools expect, as well, a complex matrix M and the block structure blk as
an input. In principle, one might wish to calculate the SSV exactly. Unfortunately, it
has been shown through examples that this is a very hard problem in a well-defined sense
introduced in computer science. Fortunately, one can calculate a lower bound and an
upper bound for the SSV pretty efficiently. In the µ-tools, this computation is performed
with the command mu(M, blk) which returns the row [upperbound lowerbound].

For the reader’s convenience we explicitly formulate the detailed conclusions that can be
drawn if having computed a lower and an upper bound of the SSV.

Theorem 4.13 Let Mc be a complex matrix and ∆c an arbitrary (open) set of complex
matrices. Then
1
• µ∆c (Mc ) ≤ γ1 implies that I − Mc ∆c is nonsingular for all ∆c ∈ γ1
∆c .

114
 1
• µ∆c (Mc ) > γ2 implies that there exists a ∆c ∈ γ2
∆c for which I − Mc ∆c is singular.

This is a straightforward consequence of the following simple fact:

α µ∆c (Mc ) = µ∆c (αMc ) = µα∆c (Mc ). (4.50)

A scalar scaling of the SSV with factor α is equivalent to scaling either the matrix Mc or
the set ∆c with the same factor.

Let us briefly look as well at the case of matrix valued scalings. Suppose U and V are
arbitrary complex matrices. Then we observe

det(I − Mc [U ∆c V ]) 6= 0 if and only if det(I − [V Mc U ]∆c ) 6= 0. (4.51)

This shows that

µU ∆c V (Mc ) = µ∆c (V Mc U ). (4.52)
Hence, if we intend to calculate the SSV with respect to the set

U ∆c V = {U ∆c V | ∆c ∈ ∆c },

we can do that by calculating the SSV of V Mc U with respect to the original set ∆c , and
this latter task can be accomplished with the µ-tools.

Before we proceed to a more extended discussion of the background on the SSV, let us
discuss its most important purpose, the application to robust stability analysis.

4.9.2 SSV Applied to Testing Robust Stability

For robust stability we had to check (4.47). If we recall Theorem 4.13, this condition
holds true if and only if the SSV of M (iω) calculated with respect to ∆c is smaller than
1. Since this has to be true for all frequencies, we immediately arrive at the following
fundamental result of these lecture notes.

Theorem 4.14 I − M ∆ has a proper and stable inverse for all ∆ ∈ ∆ if and only if

µ∆c (M (iω)) ≤ 1 for all ω ∈ R ∪ {∞}. (4.53)

We can again combine Theorem 4.8 with Theorem 4.4 to obtain the test for the general
interconnection.

Corollary 4.15 If K stabilizes P , and if

µ∆c (M (iω)) ≤ 1 for all ω ∈ R ∪ {∞},

then K robustly stabilizes S(∆, P ) against ∆.

115
Remark. In case that
∆c := {∆ ∈ Cp×q | k∆k < 1}
consists of full block matrices only (what corresponds to nr = 0, nc = 0, nf = 1), it
follows from the discussion in Section 4.8.3 that

µ∆c (M (iω)) = kM (iω)k.

Hence Theorem 4.14 and Corollary 4.15 specialize to Theorem 4.8 and Corollary 4.9 in
this particular case of full block uncertainties.

How do we apply the tests? We just calculate the number µ∆c (M (iω)) for each frequency
and check whether it does does not exceed one. In practice, one simply plots the function

ω → µ∆c (M (iω))

by calculating the right-hand side for finitely many frequencies ω. This allows to visually
check whether the curve stays below 1. If the answer is yes, we can conclude robust
stability as stated in Theorem 4.14 and Corollary 4.15 respectively. If the answer is no,
we reveal in the next section how to determine a destabilizing uncertainty.

In this ideal situation we assume that the SSV can be calculated exactly. As mentioned
above, however, only upper bounds can be computed efficiently. Still, with the upper
bound it is not difficult to guarantee robust stability. In fact, with a plot of the computed
upper bound of µ∆c M (iω) over the frequency ω, we can easily determine a number γ > 0
such that
µ∆c (M (iω)) ≤ γ for all ω ∈ R ∪ {∞} (4.54)
is satisfied. As before, we can conclude robust stability for the uncertainty set
1
{ ∆ | ∆ ∈ ∆} (4.55)
γ
which consists of all uncertainties that admit the same structure as those in ∆, but that
are rather bounded by γ1 instead of 1. This is an immediate consequence of Theorem 4.13.

We observe that the SSV-plot or a plot of the upper bound lets us decide the question
of how large we can let the structured uncertainties grow in order to still infer robust
stability.

If varying γ, the largest class γ1 ∆ is obtained with the smallest γ for which (4.54) is valid;
this best value is clearly given as

γ∗ = sup µ∆c (M (iω)).

ω∈R∪{∞}

Since γ1∗ is the largest possible inflating factor for the set of uncertainties, this number is
often called stability margin.

116
4.9.3 Construction of Destabilizing Perturbations

Let us suppose that we have found some frequency ω0 ∈ (0, ∞) for which

µ∆c (M (iω0 )) > γ0 .

Such a pair of frequency ω0 and value γ0 can be found by visually inspecting a plot of
the lower bound of the SSV over frequency as delivered by the µ-tools. Due to Theorem
4.13, there exists some ∆0 ∈ γ10 ∆c that renders I − M (iω0 )∆0 singular. Note that the
algorithm in the µ-tools to compute a lower bound of µ∆c (M (iω0 )) returns such a matrix
∆0 for the calculated bound γ0 .

Based on ω0 , γ0 and ∆0 , we intend to point out in this section how we can determine a
candidate for a dynamic destabilizing perturbation as discussed in Section 4.7.4.

Let us denote the blocks of ∆0 as

 
p1 I 0

 . ..


 
 

 pn r I 

δ10 I
 
 
 
..
∆0 =  .  ∈ Cp×q
 
 
δn0 c I
 
 
 

 ∆01 


 ... 

 
0 ∆0nf

with pj ∈ R, δj0 ∈ C, ∆0j ∈ Cpj ×qj .

According to Lemma ??, there exists proper and stable δj (s) with
1
δj (iω0 ) = δj0 and kδj k∞ ≤ |δj0 | < .
γ0

Since I − M (iω0 )∆0 is singular, there exists a complex kernel vector u 6= 0 with (I −
M (iω0 )∆0 )u = 0. Define v = ∆0 u. If we partition u and v according to ∆0 , we obtain
vj = ∆0j uj for those vector pieces that correspond to the full blocks. In the proof of
Theorem 4.10 we have shown how to construct a real rational proper and stable ∆j (s)
that satisfies
u∗j kvj k 1
∆j (iω0 ) = vj 2
and k∆j k∞ ≤ ≤ k∆0j k < .
kuj k kuj k γ0

(Provide additional arguments if it happens that uj = 0).

117
Let us then define the proper and stable dynamic perturbation
 
p1 I 0
 .. 

 . 

 

 pn r I 

 

 δ 1 (s)I 

∆(s) =  ...
.
 
 
δnc (s)I
 
 
 

 ∆1 (s) 


 ... 

 
0 ∆nf (s)

Since each of the diagonal blocks has an H∞ -norm that does not exceed γ0 , we infer
k∆k∞ < γ10 . Hence ∆ ∈ γ10 ∆. Moreover, by inspection one verifies that ∆(iω0 )u = v.
This implies that [I − M (iω0 )∆(iω0 )]u = u − M (iω0 )v = 0 such that I − M (s)∆(s) has
a zero at iω0 and, hence, its inverse is certainly not stable.

If ω0 = 0 or ω0 = ∞, this construction fails since δj0 and ∆0j cannot, in general, taken to
be real. We refer to the reference

A.L. Tits, M.K.H. Fan, On the small-µ theorem, Automatica 31 (1995) 1199-1201

for a construction of ∆ under these circumstances.

In summary, we have found an uncertainty ∆ that is destabilizing for (I − M ∆)−1 , and

that is a candidate for rendering the system S(∆, P ) controlled with K unstable.

4.9.4 Example: Two Uncertainty Blocks in Tracking Configuration

Let us come back to Figure 41 with

200 0.1s + 1
G(s) = and K(s) = 0.2 .
(10s + 1)(0.05s + 1)2 (0.65s + 1)(0.03s + 1)

In this little example we did not include any weightings for the uncertainties what is, of
course, unrealistic. Note that the uncertainties ∆1 and ∆2 have both dimension 1 × 1.
Pulling them out leads to the uncertainty
 
∆1 0
∆= 
0 ∆2

118
for the interconnection. Since both uncertainties are dynamic, we infer for this setup that
we have to choose
 
∆1 0
∆c := {  | ∆j ∈ C, |∆j | < 1, j = 1, 2}.
0 ∆2

For any complex matrix M, the command mu(M, [1 1; 1 1]) calculates the SSV of M with
respect to ∆c . The matrix [1 1; 1 1] just indicates that the structure consists of two blocks
(two rows) that are both of dimension 1 × 1.

The code

G=nd2sys( [200],conv([10 1],conv([0.05 1],[0.05 1])) );

K=nd2sys( [0.1 1], conv([0.65 1],[0.03 1]),0.2 );

systemnames=’G’;
inputvar=’[w1;w2;d;n;r;u]’;
outputvar=’[u;r-n-d-w1-G;G+w1+d-r;r+w2-n-d-w1-G]’;
input_to_G=’[u]’;
sysoutname=’P’;
cleanupsysic=’yes’;
sysic
N=starp(P,K);
M11=sel(N,1,1);
M22=sel(N,2,2);
M=sel(N,[1 2],[1 2]);
om=logspace(0,1);
clf;
vplot(’liv,m’,frsp(M11,om),’:’,frsp(M22,om),’:’,vnorm(frsp(M,om)),’--’);
hold on;grid on
Mmu=mu(frsp(M,om),[1 1;1 1]);
vplot(’liv,m’,Mmu,’-’);

computes the transfer matrix M seen by the uncertainty. Note that Mjj is the transfer
function seen by ∆j for j = 1, 2. We plot |M11 (iω)|, |M22 (iω)|, kM (iω)k, and µ∆c (M (iω))
over the frequency interval ω ∈ [0, 10], as shown in Figure 45. For good reasons to be
revealed in Section 4.10, the upper bound of the SSV coincides with the lower bound such
that, in this example, we have exactly calculated the SSV.

How do we have to interpret this plot? Since the SSV is not larger than 2, we conclude

119
 3

2.5

1.5

0.5

0
0 1
10 10

Figure 45: Magnitudes of M11 , M22 (dotted), norm of M (dashed), SSV of M (solid)

robust stability for all uncertainties that take their values in

 
1 ∆1 0
∆c = {  | ∆j ∈ C, |∆j | < 1 , j = 1, 2}.
2 0 ∆2 2

For this statement, we did not take into account that the SSV plot shows a variation
in frequency. To be specific, at the particular frequency ω, we can not only allow for
k∆c k < 0.5 but even for
1
k∆c k < r(ω) :=
µ∆c (M (iω))
and we can still conclude that I − M (iω)∆c is non-singular. Therefore, one can guarantee
robust stability for uncertainties that take their values at iω in the set
 
∆1 0
∆c (ω) := {  ∈ C2×2 | |∆1 | < r(ω), |∆2 | < r(ω)}.
0 ∆2

The SSV plot only leads to insights if re-scaling the whole matrix ∆c . How can we
explore robust stability for different bounds on the different uncertainty blocks? This
would correspond to uncertainties that take their values in
 
∆1 0
{  ∈ C2×2 | |∆1 | < r1 , |∆2 | < r2 }
0 ∆2

120
 20

18

16

14

12

10

0
0 1
10 10

Figure 46: SSV of RM with r1 = 1, r2 ∈ [0.1, 10].

for different r1 > 0, r2 > 0. The answer is simple: just employ weightings! Observe that
this set is nothing but  
r1 0
R∆c with R =  .
0 r2

In order to guarantee robust stability, we have to test µR∆c (M (iω)) ≤ 1 for all ω ∈
R ∪ {∞}, what amounts to verifying

µ∆c (RM (iω)) < 1 for all ω ∈ R ∪ {∞}

by the property (4.52). Again, we look at our example where we vary r2 in the interval
[0.1, 10] and fix r1 = 1. Figure 46 presents the SSV plots for these values.

Important task. Provide an interpretation of the plots!

Remark. For the last example, we could have directly re-scaled the two uncertainty
blocks in the interconnection, and then pulled out the normalized uncertainties. The
resulting test will lead, of course, to the same conclusions.

Similar statements can be made for the dashed curve and full block uncertainties; the

121
discussion is related to the set
 
∆11 ∆12
{∆c =   ∈ C2×2 | k∆c k < 1}.
∆21 ∆22

The dotted curves lead to robust stability results for

 
∆1 0
{  | ∆1 ∈ C, |∆1 | < 1}
0 0

or  
0 0
{  | ∆2 ∈ C, |∆2 | < 1}
0 ∆2
respectively.

Important task. Formulate the exact results and interpretations for the last three cases.

4.9.5 SSV Applied to Testing the General Hypothesis

Let us recall that we always have to verify the hypotheses 4.3 before we apply our results.
So far we were not able to check (4.23). For the specific set considered in this section,
this simply amounts to a SSV-test: (4.23) is true if and only if

µ∆c (P11 (∞)) ≤ 1.

Remark. We observe that the SSV is a tool that is by no means restricted to stability
tests in control. In fact, it is useful in any problem where one needs to check whether a
family of matrices is non-singular.

4.10 A Brief Survey on the Structured Singular Value

This section serves to provide some important properties of µ∆c (Mc ), and it should clarify
how it is possible to compute bounds on this quantity.

An important property of the SSV is its ‘monotonicity’ in the set ∆c : If two sets of
complex matrices ∆1 and ∆2 satisfy

∆1 ⊂ ∆2 ,

then we infer
µ∆1 (Mc ) ≤ µ∆2 (Mc ).

122
In short: The larger the set ∆c (with respect to inclusion), the larger µ∆c (Mc ).

Now it is simple to understand the basic idea of how to bound the SSV. For that purpose
let us introduce the specific sets
∆1 := {pI ∈ Rq×q | |p| < 1}
∆2 := {δI ∈ Cq×q | |δ| < 1}
∆3 := {∆c ∈ Cp×q | k∆c k < 1}
that correspond to one real repeated block (nr = 1, nc = 0, nf = 0), one complex repeated
block (nr = 0, nc = 1, nf = 0), or one full block (nr = 0, nc = 0, nf = 1). For these
specific structures one can easily compute the SSV explicitly:
µ∆1 (Mc ) = ρR (Mc )
µ∆2 (Mc ) = ρ(Mc )
µ∆3 (Mc ) = kMc k.
Here, ρR (M ) denotes the real spectral radius of Mc defined as
ρR (Mc ) = max{|λ| | λ is a real eigenvalue of Mc },
whereas ρ(M ) denotes the complex spectral radius of Mc that is given as
ρ(Mc ) = max{|λ| | λ is a complex eigenvalue of Mc }.

In general, we clearly have

∆1 ⊂ ∆c ⊂ ∆3
such that we immediately conclude
µ∆1 (Mc ) ≤ µ∆c (Mc ) ≤ µ∆3 (Mc ).
If there are no real blocks (nr = 0), we infer
∆2 ⊂ ∆c ⊂ ∆3
what implies
µ∆2 (Mc ) ≤ µ∆c (Mc ) ≤ µ∆3 (Mc ).
Together with the above given explicit formulas, we arrive at the following result.

Lemma 4.16 In general,

ρR (Mc ) ≤ µ∆c (Mc ) ≤ kMc k.
If nr = 0, then
ρ(Mc ) ≤ µ∆c (Mc ) ≤ kMc k.

Note that these bounds are pretty rough. The main goal in computational techniques is
to refine these bounds to get close to the actual value of the SSV.

123
4.10.1 Continuity

We have seen that the SSV of Mc with respect to the set {pI ∈ Rp×q | |p| < 1} is the real
spectral radius. This reveals that the SSV does, in general, not depend continuously on
Mc . Just look at the simple example
  
1 m  0 for m 6= 1
ρR   =
−m 1  1 for m = 0.

It shows that the value of the SSV can jump with only slight variations in m.

This is an important observation for practice. If the structure (4.46) comprises real blocks
(nr 6= 0), then
µ∆c (M (iω))
might have jumps if we vary ω. Even more dangerously, since we compute the SSV at
only a finite number of frequencies, we might miss a frequency where the SSV jumps to
very high levels. The plot could make us believe that the SSV is smaller than one and we
would conclude robust stability; in reality, the plot jumps above one at some frequency
which we have missed, and the conclusion was false.

The situation is more favorable if there are no real blocks nr = 0. Then µ∆c (Mc ) depends
continuously on Mc , and jumps do not occur.

Theorem 4.17 If nr = 0 such that no real blocks appear in (4.46), the function

Mc → µ∆c (Mc )

is continuous. In particular, If M is real-rational, proper and stable,

ω → µ∆c (M (iω))

defines a continuous function on R ∪ {∞}.

4.10.2 Lower Bounds

If one can compute some

1
∆0 ∈ ∆c that renders I − Mc ∆0 singular, (4.56)
γ
one can conclude that
γ ≤ µ∆c (Mc ).
The approach taken in the µ-tools is to maximize γ such that there exists a ∆0 as in (4.56).
There is no guarantee whether one can compute the global optimum for the resulting

124
maximization problem. Nevertheless, any step in increasing the value γ improves the
lower bound and is, hence, beneficial.

Note that the algorithm outputs a matrix ∆0 as in (4.56) for the best achievable lower
bound γ. Based on this matrix ∆0 , one can compute a destabilizing perturbation as
described in Section 4.9.3.

If the structure (4.46) only comprises real blocks (nc = 0, nf = 0), it often happens that
the algorithm fails and that the lower bound is actually just zero. In general, if real blocks
in the uncertainty structure do exist (nr 6= 0), the algorithm is less reliable if compared
to the case when these blocks do not appear (nr = 0). We will not go into the details of
these quite sophisticated algorithms.

4.10.3 Upper Bounds

If one can test that

1
for all ∆c ∈ ∆c the matrix I − Mc ∆c is non-singular, (4.57)
γ
one can conclude that
µ∆c (Mc ) ≤ γ.

We have already seen in Section (4.8.2) that kMc k ≤ γ is a sufficient condition for (4.57)
to hold.

How is it possible to refine this condition?

Simple Scalings
Let us assume that all the full blocks in (4.46) are square such that pj = qj . Suppose that
D is any non-singular matrix that satisfies
1
D∆c = ∆c D for all ∆c ∈ ∆c . (4.58)
γ
Then
kD−1 Mc Dk < γ (4.59)
implies that
I − [D−1 Mc D]∆c (4.60)
is non-singular for all ∆c ∈ γ1 ∆c . If we exploit D∆c = ∆c D, (4.60) is nothing but

I − D−1 [Mc ∆c ]D = D−1 [I − Mc ∆c ]D.

Therefore, not only (4.60) but even I − Mc ∆c itself is non-singular. This implies that
(4.57) is true such that γ is an upper bound for µ∆c (Mc ).

125
In order to find the smallest upper bound, we hence need to minimize

kD−1 Mc Dk

over the set of all matrices D that satisfy (4.58). Since D = I is in the class of all these
matrices, the minimal value is certainly better than kMc k, and we can indeed possibly
refine this rough upper bound through the introduction of the extra variables D. Since
the object of interest is a scaled version D−1 Mc D of Mc , these variables D are called
scalings. Let us summarize what we have found so far.

Lemma 4.18 We have

µ∆c (Mc ) ≤ inf kD−1 Mc Dk.

D satisfies (4.58) and is non−singular

In order to find the best upper bound, we have to solve the minimization problem on the
right. Both from a theoretical and a practical view-point, it has very favorable properties:
It is a convex optimization problem for which fast solvers are available. Convexity implies
that one can really find the global optimum.

In these notes we only intend to reveal that the fundamental reason for the favorable
properties can be attributed to the following fact: Finding a non-singular D with (4.58)
and (4.59) is a so-called Linear Matrix Inequality (LMI) problem. For such problems
very efficient algorithms have been developed in recent years.

As a first step it is very simple to see that (4.58) holds if and only if D admits the structure
 
D1 0
 .. 

 . 

 

 Dn r


 

 Dnr +1 

D= . . (4.61)
.
 

 
Dnr +nc
 
 
 

 d1 I 


 . ..


 
0 dnf I

with
Dj a non-singular complex matrix and dj a non-zero complex scalar
of the same size as the corresponding blocks in the partition (4.46). It is interesting to
observe that any repeated block in (4.46) corresponds to a full block in (4.61), and any
full block in (4.46) corresponds to a repeated block in (4.61).

126
In a second step, one transforms (4.59) into a linear matrix inequality: (4.59) is equivalent
to
[D−1 Mc D][D−1 Mc D]∗ < γ 2 I.
If we left-multiply with D and right-multiply with D∗ , we arrive at the equivalent in-
equality
Mc [DD∗ ]Mc∗ < γ 2 [DD∗ ].
Let us introduce the Hermitian matrix

Q := DD∗

such that the inequality reads as

Mc QMc∗ < γ 2 Q. (4.62)

Moreover, Q has the structure

 
Q1 0

 ... 

 
 

 Qnr 

 

 Qn r +1 

Q= .. (4.63)
.
 

 
Q
 
 nr +nc 
 

 q 1 I 


 . ..


 
0 qn f I

with

Qj a Hermitian positive definite matrix and qj a real positive scalar. (4.64)

Testing whether there exists a Q with the structure (4.63) that satisfies the matrix in-
equality (4.62) is an LMI problem.

Here we have held γ fixed. Typical LMI algorithms also allow to directly minimize γ in
order to find the best upper bound. Alternatively, one can resort to the standard bisection
algorithm as discussed in Section A.

A Larger Class of Scalings

Clearly, the larger the class of considered scalings, the more freedom is available to ap-
proach the actual value of the SSV. Hence a larger class of scalings might lead to the
computation of better upper bounds.

127
These arguments turn out to be valid, in particular, if the structure (4.46) comprises
real blocks. The fundamental idea to arrive at better upper bounds is formulated in the
following simple lemma.

Lemma 4.19 Suppose there exists a Hermitian P such that

 ∗  
∆ ∆
 c  P  c  ≥ 0 for all ∆c ∈ 1 ∆c (4.65)
I I γ

and that satisfies  ∗  

I I
  P  < 0. (4.66)
Mc Mc

Then (4.57) holds true and, hence, γ is an upper bound for µ∆c (Mc ).

Proof. The proof is extremely simple. Fix ∆c ∈ γ1 ∆c . We have to show that I − Mc ∆c

is non-singular. Let us assume the contrary: I − Mc ∆c is singular. Then there exists an
x 6= 0 with (I − Mc ∆c )x = 0. Define y := ∆c x such that x = Mc y. Then (4.65) leads to
 ∗    ∗  
∆c ∆ y y
0 ≤ x∗   P  c x =   P  .
I I x x

On the other hand, (4.66) implies

 ∗    ∗  
I I y y
0 > y∗   P y =   P  .
Mc Mc x x
 
y
(Since x 6= 0, the vector   is also non-zero.) This contradiction shows that I − Mc ∆c
x
cannot be singular.

Remark. In Lemma 4.19 the converse holds as well: If the SSV is smaller than γ, there
exists a Hermitian P with (4.65) and (4.66). In principle, based on this lemma, one could
compute the exact value of the SSV; the only crux is to parametrize the set of all scalings
P with (4.65) what cannot be achieved in an efficient manner.

Remark. In order to work with (4.58) in the previous section, we need to assume that
the full blocks ∆j are square. For (4.65) no such condition is required. The µ-tools can
handle non-square blocks as well.

128
For practical applications, we need to find a set of scalings that all fulfill (4.65). A very
straightforward choice that is implemented in the µ-tools is as follows: Let Pγ consist of
all matrices  
2
−γ Q S
P = 
S∗ Q
where Q has the structure (4.63) and (4.64), and S is given by
 
S1 0

 ... 

 
 

 Snr 

 

 0 

S= .. (4.67)
.
 

 
0
 
 
 

 0 


 . .. 

 
0 0

with
Sj a complex skew-Hermitian matrix: Sj∗ + Sj = 0.
 ∗  
∆c ∆c
For any P ∈ Pγ , the matrix   P  is block-diagonal. Let us now check for
I I
each diagonal block that it is positive semidefinite: We observe for

• real uncertainty blocks:

pj (−γ 2 Qj )pj + pj Sj + Sj∗ pj + Qj = Qj (−γ 2 p2j + 1) ≥ 0.

• complex repeated blocks:

δj∗ (−γ 2 Qj )δj + Qj = Qj (−γ 2 |δj |2 + 1) ≥ 0.

• complex full blocks:

∆∗j (−γ 2 qj I)∆j + qj I = qj (−γ 2 ∆∗j ∆j + I) ≥ 0.

We conclude that (4.65) holds for any P ∈ Pγ .

If we can find one P ∈ Pγ for which also condition (4.66) turns out to be true, we can
conclude that γ is an upper bound for the SSV. Again, testing the existence of P ∈ Pγ
that also satisfies (4.66) is a standard LMI problem.

129
The best upper bound is of course obtained as follows:

Minimize γ such that there exists a P ∈ Pγ that satisfies (4.66). (4.68)

Again, the best bound can be compute by bisection as described in Section A.

Remark. Only if real blocks do exist, the matrix S will be non-zero, and only in that
case we will benefit from the extension discussed in this section. We have described the
class of scalings that is employed in the µ-tools. However, Lemma 4.19 leaves room for
considerable improvements in calculating upper bounds for the SSV.

4.11 When is the Upper Bound equal to the SSV?

Theorem 4.20 If
2(nr + nc ) + nf ≤ 3,
then the SSV µ∆c (Mc ) is not only bounded by but actually coincides with the optimal value
of problem (4.68).

Note that this result is tight in the following sense: If 2(nr + nc ) + nf > 3, one can
construct examples for which there is a gap between the SSV and the best upper bound,
the optimal value of (4.68).

4.11.1 Example: Different Lower and Upper Bounds

Let  
1 s−2
2s+1
1 2s+4
 
M (s) =  s 1 .
 −1 s2 +s+1 (s+1)2 
3s −1
s+5 4s+1
1
Moreover, consider three different structures.

The first consists of two full complex blocks

  
 ∆ 0 
1
∆1 :=   | ∆1 ∈ C, |∆1 | < 1, ∆2 ∈ C2×2 , k∆2 k < 1 .
 0 ∆2 

We plot in Figure 47 kM (iω)k, µ∆1 (M (iω)), ρ(M (iω)) over frequency.

We conclude that I − M (iω0 )∆c is non-singular for all

1
• full blocks ∆c ∈ C3×3 with k∆c k < γ1

130
 3.5
complex spectral radius
SSV γ1
norm
3

2.5

γ2
2

γ3
1.5

1
10−2 10−1 100 ω0 = 10 102

Figure 47: Plots of kM (iω)k, µ∆1 (M (iω)) and ρ(M (iω)) over frequency.

1
• structured blocks ∆c ∈ γ2
∆1
 
δ 0 0
  1
• complex repeated blocks 
 0 δ 0  with |δ| <
 γ3
.
0 0 δ

In addition, there exists a ∆c that is

• a full block ∆c ∈ C3×3 with k∆c k < γ1 , γ < γ1

• a structured block ∆c ∈ γ1 ∆1 , γ < γ2

 
δ 0 0
 
• a complex repeated block  0 δ 0  with |δ| < 1 , γ < γ3
  γ

0 0 δ

that renders I − M (iω0 )∆c singular.

As a second case, we take one repeated complex block and one full complex block:
  
 ∆ 0 
1
∆2 :=   | ∆1 ∈ C, |∆1 | < 1, δ2 ∈ C, |δ2 | < 1 .
 0 δ2 I2 

131
 1.7
γ1
1.65

1.6 γ2

1.55

1.5

1.45 γ3

1.4

1.35 upper bound

lower bound
1.3 −2
10 ω0 = 0.1 100 101 102

Figure 48: Plots of bounds on µ∆1 (M (iω)) and of ρ(M (iω)) over frequency.

As observed in Figure 48, the upper and lower bounds for the SSV are different. Still, the
complex spectral radius bounds the lower bound on the SSV from below.

We conclude that I − M (iω0 )∆c is non-singular for all structured ∆c in γ11 ∆2 . There
exists a structured ∆c in γ1 ∆2 , γ < γ2 , that renders I − M (iω0 )∆c singular.

As a last case, let us consider a structure with one real repeated block and one full complex
block:   
 ∆ 0 
1
∆3 :=   | ∆1 ∈ C, |∆1 | < 1, δ2 ∈ R, |δ2 | < 1 .
 0 δ2 I2 

Figure 49 shows that lower bound and upper bound of the SSV are further apart than in
the previous example, what reduces the quality of the information about the SSV. Since
the structure comprises a real block, the complex spectral radius is no lower bound on
the SSV (or its lower bound) over all frequencies, as expected.

The upper bound is smaller than γ1 for all frequencies. Hence I − M ∆ has proper and
stable inverse for all ∆ ∈ γ11 ∆.

There exists a frequency for which the lower bound is larger that γ2 : Hence there exists
a ∆ ∈ γ12 ∆ such that I − M ∆ does not have a proper and stable inverse.

We emphasize that the

132
 1.8
γ1
γ2
1.6

1.4

1.2

0.8
γ3 upper bound
lower bound
0.6
10−2 10−1 100 101 102

Figure 49: Plots of bounds on µ∆1 (M (iω)) and of ρ(M (iω)) over frequency.

Exercises

1) This is a continuation of Exercise 3 in Section 3.

(Matlab) Let the controller K be given as
 
−4.69 −1.28 −1.02 7.29 −0.937 −0.736 5.69 1.6
 
 −2.42 −3.29 −3.42 6.53 −1.17 −0.922 2.42 4.12 
 
 
 −1.48 0.418 −0.666 2.52 0 0 1.48 1.05 
 
K =  19 −0.491 1.34 −0.391 −6.16 −5.59 0.943 −0.334  .
 
 
 24.8
 2.35 3.73 −8.48 −7.58 −3.61 −3.45 −3.96 
 
 4.38
 4.02 3.2 −9.41 2.54 0 −4.38 −5.03 
20 −0.757 1.13 −1 −6.16 −5.59 0 0

Suppose the component G1 is actually given by

s−2
G1 + W1 ∆ with W1 (s) = .
s+2
For the above controller, determine the largest r such that the loop remains
stable for all ∆ with

∆ ∈ R, |∆| < r or ∆ ∈ RH∞ , k∆k∞ < r.

133
 Argue in terms of the Nyquist or Bode plot seen by the uncertainty ∆. Con-
struct for both cases destabilizing perturbations of smallest size, and check
that they lead, indeed, to an unstable controlled interconnection as expected.
a) (Matlab) Use the same data as in the previous exercise. Suppose now that the
pole of G2 is uncertain:
1
G2 (s) = , p ∈ R, |p − 1| < r.
s−p
What is the largest possible r such that K still stabilizes P for all possible real
poles p. What happens if the pole variation is allowed to be complex?
b) (Matlab) With W1 as above, let G1 be perturbed as

G1 + W1 ∆, ∆ ∈ RH∞ , k∆k∞ < r1

and let G2 be given as

1
G2 (s) = , p ∈ R, |p − 1| < r2 .
s−p
Find the largest r = r1 = r2 such that K still robustly stabilizes the system
in face of these uncertainties. Plot a trade-off curve: For each r1 in a suit-
able interval (which?), compute the largest r2 (r1 ) for which K is still robustly
stabilizing and plot the graph of r2 (r1 ); comment!

2) Suppose ∆c is the set of all matrices with k∆k < 1 that have the structure

∆ = diag(δ1 , . . . , δm ) ∈ Rm×m ,

and consider the rationally perturbed matrix A∆ = A + B∆(I − D∆)−1 C for real
matrices A, B, C, D of compatible size. Derive a µ-test for the following properties:
a) Both I − D∆ and A∆ are non-singular for all ∆ ∈ ∆c .
b) For all ∆ ∈ ∆c , I − D∆ is non-singular and A∆ has all its eigenvalues in the
open left-half plane.
c) For all ∆ ∈ ∆c , I − D∆ is non-singular and A∆ has all its eigenvalues in the
open unit disk {z ∈ C | |z| < 1}.

3) Let ∆c be the same as in Exercise 2. For real vectors a ∈ Rm and b ∈ Rm , give a

formula for µ∆c (abT ). (SSV of rank one matrices.)

134
4) a) For the data
   
12 −3 2 ∆11 ∆12 0
   
Mc =  −1 7 8  and ∆c = {∆c =  ∆21 ∆22 0  ∈ C3×3 | k∆c k < 1}
   
5 3 −1 0 0 ∆33

compute µ∆c (Mc ) with a Matlab m-file. You are allowed to use only the
functions max, eig and a for-loop; in particular, don’t use mu.
 
M1 M12
b) Let Mc =   and let ∆c be the set of all ∆c = diag(∆1 , ∆2 ) with
M21 M2
full square blocks ∆j satisfying k∆j k < 1. Give a formula for the value
 
M1 dM12
d∗ = min   .
d>0 1
d
M21 M 2
2

where kM k22 = trace(M ∗ M ). How does d∗ lead to an upper bound of µ∆c (M )?

(Matlab) Compare this bound with the exact value in the previous exercise.

135
5 Nominal Performance Specifications

In our general scenario (Figure 21), we have collected various signals into the generalized
disturbance w and the controlled variable z, and we assume that these signals are chosen to
characterize the performance properties to be achieved by the controller. So far, however,
we only included the requirement that K should render z = S(P, K)w stable.

For any stabilizing controller, one can of course just directly investigate the transfer
function elements of S(P, K) and decide whether they are satisfactory or not. Most often,
this just means that the Bode plots of these transfer functions should admit a desired
shape that is dictated by the interpretation of the underlying signals. In this context you
should remember the desired shapes for the sensitivity and the complementary sensitivity
transfer matrices in a standard tracking problem.

For the purpose of analysis, a direct inspection of the closed-loop transfer matrix is no
problem at all. However, if the interconnection is affected by uncertainties and if one would
like to verify robust performance, or if one wishes to design a controller, it is required to
translate the desired performance specifications into a weighted H∞ -norm criterion.

5.1 An Alternative Interpretation of the H∞ Norm

We have seen in Section 1.2 that the H∞ -norm of a proper and stable G is just the energy
gain of the corresponding LTI system.

Most often, reference or disturbance signals are persistent and can be assumed to be
sinusoids. Such a signal is given as

w(t) = w0 eiω0 t (5.1)

with w0 ∈ Cn and ω0 ∈ R. If ω0 = 0 and w0 ∈ Rn , this defines the step function w(t) = w0

of height w0 . (We note that this class of complex valued signals includes the set of all
real-valued sinusoids. We work with this enlarged class to simplify the notation in the
following arguments.) Let us choose as a measure of size for (5.1) the Euclidean norm of
its amplitude:
kw(.)kRMS := kw0 k.
As indicated, this defines indeed a norm on the vector space of all sinusoids.

Let w(.) as defined by (5.1) pass the LTI system defined by the proper and stable transfer
matrix G to obtain z(.). As well-known, limt→∞ [z(t) − G(iω0 )w0 eiω0 t ] = 0 such that the
steady-state response is
(Gw)s (t) = G(iω0 )w0 eiω0 t .

136
(The subscript means that we only consider the steady-state response of Gw.) We infer

k(Gw)s kRMS = kG(iω0 )w0 k

and hence, due to kG(iω0 )w0 k ≤ kG(iω)kkw0 k,

k(Gw)s kRMS
≤ kG(iω0 )k ≤ kGk∞ .
kwkRMS
Hence the gain of w → (Gw)s is bounded by kGk∞ . The gain actually turns out to be
equal to kGk∞ .

Theorem 5.1 Let G be proper and stable. Then

k(Gw)s kRMS
sup = kGk∞ .
w a sinusoid with kwkRMS >0 kwkRMS

The proof is instructive since it shows how to construct a signal that leads to the largest
amplification if passed through the system.

Proof. Pick the frequency ω0 ∈ R ∪ {∞} with kG(iω0 )k = kGk∞ .

Let us first assume that ω0 is finite. Then take w0 6= 0 with kG(iω0 )w0 k = kG(iω0 )kkw0 k.
(Direction of largest gain of G(iω0 ).) For the signal w(t) = w0 eiω0 t we infer
k(Gw)s kRMS kG(iω0 )w0 k kG(iω0 )kkw0 k
= = = kG(iω0 )k = kGk∞ .
kwkRMS kw0 k kw0 k
k(Gw)s kRMS
Hence the gain kwkRMS
for this signal is largest possible.

If ω0 is infinite, take any sequence ωj ∈ R with ωj → ∞, and construct at each ωj the

signal wj (.) as before. We infer
k(Gwj )s kRMS
= kG(iωj )k
kwj kRMS
and this converges to kG(iω0 )k = kGk∞ . Hence we cannot find a signal for which the
worst amplification is attained, but we can come arbitrarily close.

As a generalization, sums of sinusoids are given as

N
X
w(t) = wj eiωj t (5.2)
j=1

where N is the number of pair-wise different frequencies ωj ∈ R, and wj ∈ Cn is the

complex amplitude at the frequency ωj . As a measure of size for the signal (5.2) we
employ v
u N
uX
kwkRMS := t kwj k2 .
j=1

137
Again, this defines a norm on the vector space of all sums of sinusoids. For any w(.)
defined by (5.2), the steady-state response is
N
X
(Gw)s (t) = G(iωj )wj eiωj t
j=1

and has norm v

u N
uX
k(Gw)s kRMS =t kG(iωj )wj k2 .
j=1

Again by kG(iωj )wj k ≤ kG(iωj )kkwj k ≤ kGk∞ kwj k, we infer

k(Gw)s kRMS ≤ kGk∞ kwkRMS .

We arrive at the following generalization of the result given above.

Theorem 5.2 Let G be proper and stable. Then

k(Gw)s kRMS
sup = kGk∞ .
w a sum of sinusoids with kwkRMS >0 kwkRMS

Remark. We have separated the formulation of Theorem 5.1 from that of Theorem 5.2 in
order to stress that k(Gw)s kRMS
kwkRMS
can be rendered arbitrarily close to kGk∞ by using simple
sinusoids as in (5.1); we do not require sums of sinusoids to achieve this approximation.

5.2 The Tracking Interconnection

Let us come back to the interconnection in Figure 17, and let K stabilize the intercon-
nection.

5.2.1 Bound on Frequency Weighted System Gain

Often, performance specifications arise by specifying how signals have to be attenuated

in the interconnection.

Typically, the reference r and the disturbance d are most significant at low frequencies.
With real-rational proper and stable low-pass filters Wr , Wd , we hence assume that r, d
are given as
r = Wr r̃, d = Wd d˜

138
 ˜ are sinusoids or sums of sinusoids. Similarly, the measurement noise n
where r̃(.), d(.)
is most significant at high frequencies. With a real-rational proper and stable high-pass
filter Wn , we hence assume that n is given as

n = Wn ñ

where ñ(.) is a sum of sinusoids. Finally, the size of the unfiltered signals is assumed to
be bounded as   2
d˜
˜ 2 + kñk2
 
 ñ 
  = kr̃k2RMS + kdk RMS RMS ≤ 1.

r̃
RMS

Remark. In our signal-based approach all signals are assumed to enter the interconnec-
tion together. Hence it is reasonable to bound the stacked signal instead of working with
˜ RMS , kñkRMS . Recall, however, that the above inequality
individual bounds on kr̃kRMS , kdk
implies
kr̃k2RMS ≤ 1, kdk˜2 2
RMS ≤ 1, kñkRMS ≤ 1,

and that it is implied by

1 ˜2 1 1
kr̃k2RMS ≤ , kdk 2
RMS ≤ , kñkRMS ≤ .
3 3 3

The goal is to keep the norm of the steady-state error es small, no matter which of these
signals enters the interconnection. If we intend to achieve kes kRMS ≤ , we can as well
rewrite this condition with We := 1 as kẽs kRMS ≤ 1 for

ẽ = We e.

To proceed to the general framework, let us introduce

   
d d˜
   
z = e, z̃ = ẽ and w =   n  , w̃ =  ñ 
  

r r̃

as well as the weightings

 
Wd 0 0
 
Wz = We and Ww = 
 0 W n 0 .

0 0 Wr

In the general framework, the original closed-loop interconnection was described as

z = S(P, K)w.

139
 z̃ w̃
WZ WW
P

y u

Figure 50: Weighted closed-loop interconnection

Since the desired performance specification is formulated in terms of z̃ and w̃, we introduce
these signals with
z̃ = Wz z and w = Ww w̃
to get the weighted closed-loop interconnection (Figure 50)

z̃ = [Wz S(P, K)Ww ]w̃.

Recall that the desired performance specification was reduced to the fact that

kw̃kRMS ≤ 1 implies kz̃s kRMS ≤ 1.

By Theorem 5.2, this requirement is equivalent to

kWz S(P, K)Ww k∞ ≤ 1. (5.3)

We have arrived at those performance specifications that can be handled with the tech-
niques developed in these notes: Bounds on the weighted H∞ -norm of the performance
channels.

Let us recall that this specification is equivalent to

kWz (iω)S(P, K)(iω)Ww (iω)k ≤ 1 for all ω ∈ R ∪ {∞}.

This reveals the following two interpretations:

• Loop-shape interpretation. The shape of the frequency response ω →

S(P, K)(iω) is compatible with the requirement that the maximal singular value
of the weighted frequency response ω → Wz (iω)S(P, K)(iω)Ww (iω) does not ex-
ceed one. Roughly, this amounts to bounding all singular values of S(P, K)(iω)
from above with a bound that varies according the variations of the singular values
of Wz (iω) and Ww (iω). This rough interpretation is accurate if Wz (iω) and Ww (iω)
are just scalar valued since (5.3) then just amounts to
1
kS(P, K)(iω)k ≤ for all ω ∈ R ∪ {∞}.
|Wz (iω)Ww (iω)|

140
 It is important to note that one cannot easily impose a lower bound on the smallest
singular value of S(P, K)(iω)! Instead, desired minimal amplifications are enforced
by imposing upper bounds on the largest singular value of ‘complementary’ trans-
fer functions - for that purpose one should recall the interplay of sensitivity and
complementary sensitivity matrices.

• Disturbance attenuation interpretation. For all disturbances that are defined

through
N
X N
X
iωj t
w(t) = W (iωj )wj e with kwj k2 ≤ 1,
j=1 j=1

(5.3) implies that the steady-state response zs of z = S(P, K)w satisfies

N
X N
X
iωj t
zs (t) = zj e with kZ(iωj )zj k2 ≤ 1.
j=1 j=1

For pure sinusoids, any disturbance satisfying

w(t) = W (iω)weiωt with kwk2 ≤ 1

leads to a steady-state response

zs (t) = zeiωt with kZ(iω)zk2 ≤ 1.

This leads to a very clear interpretation of matrix valued weightings: Sinusoids of

frequency ω with an amplitude in the ellipsoid {Ww (iω)w | kwk ≤ 1} lead to a
steady-state sinusoidal response with amplitude in the ellipsoid {z | kWz (iω)zk ≤
1}. Hence Ww defines the ellipsoid which captures the a priori knowledge of the
amplitudes of the incoming disturbances and Wz defines the ellipsoids that captures
desired amplitudes of the controlled output. Through the use of matrix valued
weightings one can hence enforce spatial effects, such as quenching the output error
mainly in a certain direction.

Note that Wz S(P, K)Ww is nothing but S(P̃ , K) for

      
z̃ w̃ Wz P11 Ww Wz P12 w̃
  = P̃   =   .
y u P21 Ww P22 u

Instead of first closing the loop and then weighting the controlled system, one can as well
first weight the open-loop interconnection and then close the loop.

5.2.2 Frequency Weighted Model Matching

In design, a typical specification is to let one or several transfer matrices in an intercon-

nection come close to an ideal model. Let us suppose that the real-rational proper and

141
 z̃ w̃
P̃

y u

Figure 51: Weighted closed-loop interconnection

stable Wm is an ideal model. Moreover, the controller should render S(P, K) to match
this ideal model over certain frequency ranges. With suitable real-rational weightings W1
and W2 , this amounts to render kW1 (iω)[S(P, K)(iω) − Wm (iω)]W2 (iω)k ≤ γ satisfied for
all frequencies ω ∈ R ∪ {∞} where γ is small. By incorporating the desired bound γ into
the weightings (replace W1 by γ1 W1 ), we arrive at the performance specification
kW1 [S(P, K) − Wm ]W2 k∞ ≤ 1. (5.4)
To render this inequality satisfied, one tries ‘shape the closed-loop frequency response by
pushing it towards a desired model’. In this fashion, one can incorporate for each transfer
function of S(P, K) both amplitude and phase specifications. If there is no question about
which ideal model Wm to take, this is the method of choice.

Again, we observe that this performance specification can be rewritten as

kS(P̃ , K)k∞ ≤ 1
where P̃ is defined as  
W1 [P11 − Wm ]W2 W1 P12
P̃ =  .
P21 W2 P22

Remark. Note that the choice Wm = 0 of the ideal model leads back to imposing a
direct bound on the system gain as discussed before.

In summary, typical signal based performance specifications can be re-formulated as a

general frequency weighted model-matching requirement which leads to a bound on the
H∞ -norm of the matrix-weighted closed-loop transfer matrix.

5.3 The General Paradigm

Starting from       
z w P P w
  = P   =  11 12    , (5.5)
y u P21 P22 u

142
we have seen how to translate the two most important performance specifications on the
closed-loop system
S(P, K),
weighted gain-bounds and weighted model-matching, into the specification

kS(P̃ , K)k∞ ≤ 1 (5.6)

for the weighted open-loop interconnection

      
z̃ w̃ P̃ P̃ w̃
  = P̃   =  11 12    . (5.7)
y u P̃21 P22 u

So far, we have largely neglected any technical hypotheses on the weighting matrices that
are involved in building P̃ from P . In fact, any controller to be considered should stabilize
both P and P̃ . Hence we have to require that both interconnections define generalized
plants, and these are the only properties to be obeyed by any weighting matrices that are
incorporated in the interconnection.

Hypothesis 5.3 The open-loop interconnections (5.5) and (5.7) are generalized plants.

Note that P and P̃ have the same lower right block P22 . This is the reason why any
controller that stabilizes P also stabilizes P̃ , and vice versa.

Lemma 5.4 Let P and P̃ be generalized plants. A controller K stabilizes P if and only
if K stabilizes P̃ .

Proof. If K stabilizes P , then K stabilizes P22 . Since P̃ is a generalized plant and has
P22 as its right-lower block, K also stabilizes P̃ . The converse follows by interchanging
the role of P and P̃ .

Hence the class of stabilizing controller for P and for P̃ are identical.

From now on we assume that all performance weightings are already incorpo-
rated in P . Hence the performance specification is given by kS(P, K)k∞ ≤ 1.

Remark. In practical controller design, it is often important to keep P and P̃ separated.

Indeed, the controller will be designed on the basis of P̃ to obey kS(P̃ , K)k∞ ≤ 1, but
then it is often much more instructive to directly investigate the unweighted frequency
response ω → S(P, K)(iω) in order to judge whether the designed controller leads to the
desired closed-loop specifications.

143
6 Robust Performance Analysis

6.1 Problem Formulation

To test robust performance, we proceed as for robust stability: We identify the perfor-
mance signals, we pull out the uncertainties and introduce suitable weightings for the
uncertainties such that we arrive at the framework as described in Section 4.5. Moreover,
we incorporate in this framework the performance weightings as discussed in Section 5 to
reduce the desired performance specification to an H∞ norm bound on the performance
channel. In Figure 52 we have displayed the resulting open-loop interconnection, the
interconnection if closing the loop as u = Ky, and the controlled interconnection with
uncertainty.

We end up with the controlled uncertain system as described by

      
z∆ w∆ P11 P12 P13 w∆
      
 z  = P  w  =  P21 P22 P23   w  , u = Ky, w∆ = ∆z∆ , ∆ ∈ ∆.
      
y u P31 P32 P33 u

Let us now formulate the precise hypotheses on P , on the uncertainty class ∆, and on
the performance specification as follows.

Hypothesis 6.1

• P is a generalized plant.

• The set of uncertainties is given as

∆ := {∆ ∈ RH∞ | ∆(iω) ∈ ∆c for all ω ∈ R ∪ {∞}}

where ∆c is the set of all matrices ∆c structured as (4.46) and satisfying k∆c k < 1.

• The direct feed-through P11 and ∆c are such that

I − P11 (∞)∆c is non-singular for all ∆c ∈ ∆c .

• The performance of the system is as desired if the H∞ -norm of the channel w → z

is smaller than one.

We use the the brief notation

   
P22 P23 P21  
P∆ = S(∆, P ) =  +  ∆(I − P11 ∆)−1 P12 P13
P32 P33 P32

144
 z∆ w∆
z w
P
y u

z∆ w∆
z w
P

y u

z∆ w∆

z w
P

y u

Figure 52: Open-loop interconnection, controlled interconnection, uncertain controlled

interconnection.

145
for the perturbed open-loop interconnection. Then the unperturbed open-loop intercon-
nection is nothing but  
P22 P23
P0 = S(0, P ) =  .
P32 P33

Suppose that K stabilizes P . Then the perturbed and unperturbed controlled intercon-
nections are described by
z = S(P∆ , K)w and z = S(P0 , K)w
respectively. If K stabilizes P and if it leads to
kS(P0 , K)k∞ ≤ 1,
we say that K achieves nominal performance for P .

Accordingly, we can formulate the corresponding analysis and synthesis problems.

Nominal Performance Analysis Problem

For a given fixed controller K, test whether it achieves nominal performance for P .

Nominal Performance Synthesis Problem

Find a controller K that achieves nominal performance for P .

The analysis problem is very easy to solve: Check whether K stabilizes P , and plot
ω → kS(P0 , K)(iω)k in order to verify whether this value remains smaller than 1. Note
that H∞ -norm bounds of this sort can be verified much more efficiently on the basis of a
state-space test, as will be discussed in Section 7.2.

The synthesis problem amounts to finding a stabilizing controller K for P that renders
the H∞ -norm kS(P0 , K)k∞ smaller than 1. This is the celebrated H∞ -control problem
and will be discussed in Section 7.4.

The main subject of this section is robust performance analysis. If

K stabilizes P∆ = S(∆, P ) and kS(P∆ , K)k∞ ≤ 1 for all ∆ ∈ ∆,
we say that
K achieves robust performance for S(∆, P ) against ∆.

Let us again formulate the related analysis and synthesis problems explicitly.

Robust Performance Analysis Problem

For a given fixed controller K, test whether it achieves robust performance for S(∆, P )
against ∆.

Robust Performance Synthesis Problem

Find a controller K that achieves robust performance for S(∆, P ) against ∆.

146
6.2 Testing Robust Performance

Let us assume throughout that K stabilizes P what implies that N := S(P, K) is stable.
Introduce the partition
        
z w w M N12 w
 ∆  = S(P, K)  ∆  = N  ∆  =   ∆ .
z w w N21 N22 w

Then we infer

S(P∆ , K) = S(∆, N ) = N22 + N21 ∆(I − M ∆)−1 N12 .

Hence, K achieves robust performance if the robust stability condition

µ∆c (M (iω)) ≤ 1 for all ω ∈ R ∪ {∞}

or equivalently

det(I − M (iω)∆c ) 6= 0 for all ∆c ∈ ∆c , ω ∈ R ∪ {∞}

and the performance bound

kN22 + N21 ∆(I − M ∆)−1 N12 k ≤ 1 for all ∆ ∈ ∆

or equivalently

kN22 (iω) + N21 (iω)∆(iω)(I − M (iω)∆(iω))−1 N12 (iω)k ≤ 1

for all ∆ ∈ ∆, ω ∈ R ∪ {∞}

or equivalently

kN22 (iω) + N21 (iω)∆c (I − M (iω)∆c )−1 N12 (iω)k ≤ 1 for all ∆c ∈ ∆c , ω ∈ R ∪ {∞}

hold true.

Similarly as for robust stability, for a fixed frequency we arrive at a problem in linear
algebra which is treated in the next section.

6.3 The Main Loop Theorem

Here is the linear algebra problem that needs to be investigated: Given the set ∆c and
the complex matrix
 
Mc N12
Nc =   with N22 of size q2 × p2 ,
N21 N22

147
test whether the following two conditions hold:

det(I − Mc ∆c ) 6= 0 and kN22 + N21 ∆c (I − Mc ∆c )−1 N12 k ≤ 1 for all ∆c ∈ ∆c .

Here is the fundamental trick to solve this problem: The condition kN22 + N21 ∆c (I −
Mc ∆c )−1 N12 k = kS(∆c , Nc )k ≤ 1 is equivalent to

det(I − S(∆c , Nc )∆ b c ∈ Cp2 ×q2 , k∆

b c ) 6= 0 for all ∆ b c k < 1.

We just need to recall that the SSV of a complex matrix equals its norm if the uncertainty
structure just consists of one full block.

Let us hence define

∆ b c ∈ Cp2 ×q2 | k∆
b c = {∆ b c k < 1}.

We infer that, for all ∆c ∈ ∆c ,

det(I − Mc ∆c ) 6= 0 and kS(∆c , Nc )k ≤ 1

if and only if, for all ∆c ∈ ∆c and ∆

bc ∈ ∆
b c,

det(I − Mc ∆c ) 6= 0 and det(I − S(∆c , Nc )∆

b c ) 6= 0

if and only if, for all ∆c ∈ ∆c and ∆bc ∈ ∆

b c,
 
I − Mc ∆c −N12 ∆c
b
det   6= 0
−N21 ∆c I − N22 ∆
bc

if and only if, for all ∆c ∈ ∆c and ∆bc ∈ ∆

b c,
   
Mc N12 ∆ 0
det I −   c  6= 0.
N21 N22 0 ∆c
b

Note that we have used in these derivation the following simple consequence of the well-
known Schur formula for the determinant:

det(I − S(∆c , Nc )∆b c ) = det(I − [N22 + N21 ∆c (I − Mc ∆c )−1 N12 ]∆b c) =

 
b c )]) = det  I − Mc ∆c −N12 ∆c  .
b
b c ] − [N21 ∆c ](I − Mc ∆c )−1 [N12 ∆
= det([I − N22 ∆
−N21 ∆c I − N22 ∆
bc

This motivates to introduce the set of extended matrices

  
 ∆ 0 
c p2 ×q2
∆e :=   : ∆c ∈ ∆c , ∆c ∈ C
b , k∆c k < 1
b
 0 ∆ bc 

which consists of adjoining to the original structure one full block. We have proved the
following Main Loop Theorem.

148
Theorem 6.2 The two conditions

µ∆c (Mc ) ≤ 1 and kS(∆c , Nc )k ≤ 1 for all ∆c ∈ ∆c

are equivalent to
µ∆e (Nc ) ≤ 1.

This result reduces the desired condition to just another SSV-test on the matrix Nc with
respect to the extended structure ∆e .

Typically, a computation of µ∆e (Nc ) will lead to an inequality

µ∆e (Nc ) ≤ γ

with a bound γ > 0 different from one. The consequences that can then be drawn can be
easily obtained by re-scaling. In fact, this inequality leads to
1
µ∆e ( Nc ) ≤ 1.
γ
This is equivalent to
1
µ∆c ( Mc ) ≤ 1
γ
and
1 1 1 1
k N22 + N21 ∆c (I − Mc ∆c )−1 N12 k ≤ 1 for all ∆c ∈ ∆c .
γ γ γ γ
Both conditions are clearly nothing but

µ∆c (Mc ) ≤ γ

and
1 1
kN22 + N21 [ ∆c ](I − Mc [ ∆c ])−1 N12 k ≤ γ for all ∆c ∈ ∆c .
γ γ
We arrive at
1
det(I − Mc ∆c ) 6= 0 for all ∆c ∈ ∆c
γ
and
1
kN22 + N21 ∆c (I − Mc ∆c )−1 N12 k ≤ γ for all ∆c ∈ ∆c .
γ

Hence a general bound γ different from one leads to non-singularity conditions and a
performance bound γ for the class of complex matrices γ1 ∆c .

A more general scaling result that is proved analogously can be formulated as follows.

149
Lemma 6.3 The scaled SSV-inequality
  
γ1 I 0
µ∆e Nc   ≤ γ3
0 γ2 I

is equivalent to
γ1
det(I − Mc ∆c ) 6= 0 for all ∆c ∈ ∆c
γ3
and
γ3 γ1
kS(∆c , Nc )k ≤ for all ∆c ∈ ∆c .
γ2 γ3

This result allows to investigate the trade-off between the size of the uncertainty and the
worst possible norm kS(∆c , Nc )k by varying γ1 , γ2 and computing the SSV giving the
bound γ3 .

Note that we can as well draw conclusions of the following sort: If one wishes to guarantee

det(I − Mc ∆c ) 6= 0 and kS(∆c , Nc )k ≤ β for all ∆c ∈ α∆c

for some bounds α > 0, β > 0, one needs to perform the SSV-test
  
αI 0
µ∆e Nc   ≤ 1.
0 β1 I

6.4 The Main Robust Stability and Robust Performance Test

If we combine the findings of Section (6.2) with the main loop theorem, we obtain the
following result.

 
M N12
Theorem 6.4 Let N =   be a proper and stable transfer matrix. For all
N21 N22
∆ ∈ ∆,
(I − M ∆)−1 ∈ RH∞ and kS(∆, N )k∞ ≤ 1
if and only if
µ∆e (N (iω)) ≤ 1 for all ω ∈ R ∪ {∞}.

Combining all the insights we have gained so far leads us to the most fundamental result
in SSV-theory, the test of robust stability and robust performance against structured
uncertainties.

150
 ∆

z∆ w∆
 
M N12
 
N21 N22

z w

ˆ
∆

Figure 53: Equivalent Robust Stability Test

Corollary 6.5 If K stabilizes P , and if

µ∆e (N (iω)) ≤ 1 for all ω ∈ R ∪ {∞},

then K achieves robust performance for S(∆, P ) against all ∆ ∈ ∆.

At the outset, it looks more complicated to test robust performance if compared to robust
stability. However, the main loop theorem implies that the test of robust performance is
just another SSV test with respect to the extended block structure ∆e .

Accidentally (and with no really deep consequence), the SSV-test for robust performance
can be viewed as a robust stability test for the interconnection displayed in Figure 53.

6.5 Summary

Suppose that K stabilizes the generalized plant P and suppose that the controlled uncer-
tain system is described as
        
z∆ w∆ w∆ M N12 w∆
  = S(P, K)  =N =   , w∆ = ∆z∆
z w w N21 N22 w

with proper and stable ∆ satisfying

∆(iω) ∈ ∆c for all ω ∈ R ∪ {∞}.

Then the controller K achieves

151
 ∆

z∆ w∆
 
M N12
z w
 
N21 N22

Figure 54: Summary

• Robust stability if

µ∆c (M (iω)) ≤ 1 for all ω ∈ R ∪ {∞}.

• Nominal performance if

kN22 (iω)k ≤ 1 for all ω ∈ R ∪ {∞}.

• Robust performance if

µ∆e (N (iω)) ≤ 1 for all ω ∈ R ∪ {∞}.

In pictures, this can be summarized as follows. Robust stability is guaranteed by an

SSV-test on left-upper block M of N = S(P, K), nominal performance is guaranteed
by an SV-test on the right-lower block N22 of N = S(P, K), and robust performance is
guaranteed by an SSV-test on the whole N = S(P, K) with respect to the extended block
structure.

6.6 An Example

Suppose some controlled system is described with

 
1 s−2 s−0.1
1
 2s+1 2s+4 s+1

 −1 s 1
2
s +s+1 (s+1) 2 0.1 
N (s) =  .
 
3s −1 10

 s+5 4s+1 1 s+4 

1 0.1 s−1
s+2 s2 +s+1 s+1
1

Let ∆c be the set of ∆c with k∆c k < 1 and

 
∆1 0
∆c =   , ∆1 ∈ C2×2 , ∆2 ∈ C.
0 ∆2

152
 3.4

3.2

2.8

2.6 γ1

2.4 γ2

2.2

2
γ3
1.8 SSV RS
upper bound
1.6 lower bound
norm
1.4 −2 ω0
10 10−1 100 101 102

Figure 55: Norm of N , upper and lower bound on SSV of N , SSV of M .

The extended set ∆e consists of all ∆e with k∆e k < 1 and

 
∆1 0 0
 
∆e =  0 ∆ 2 0  , ∆1 ∈ C2×2 , ∆2 ∈ C, ∆
b ∈ C.
 
0 0 ∆ b

To test robust stability, we plot ω → µ∆c (M (iω)), to test nominal performance, we plot
ω → kN22 (iω)k, and the robust performance test requires to plot ω → µ∆e (N (iω)).

Let us first look at a frequency-by-frequency interpretation of the SSV plot of N with

respect to the extended structure (Figure 55).

With the upper bound, we infer µ∆e (N (iω0 )) ≤ γ1 what implies

1
kS(∆c , N (iω0 ))k ≤ γ1 for all ∆c ∈ ∆c .
γ1
At the frequency iω0 , one has a guaranteed performance level γ1 for the uncertainty set
1
γ1
∆c .

153
 γ1

2.7 γ2

2.6

2.5

upper bound
lower bound
2.4 −2
10 10−1 100 101 102

Figure 56: Upper and lower bound on SSV of N

With the lower bound, we infer µ∆e (N (iω0 )) > γ2 . This implies that
1
det(I − M (iω0 )∆c ) = 0 for some ∆c ∈ ∆c
γ2
or
1
kS(∆c , N (iω0 ))k > γ2 for some ∆c ∈ ∆c .
γ2
We can exploit the knowledge of the SSV curve for robust stability to exclude the first
property due to γ3 < γ2 . (Provide all arguments!) Hence, at this frequency we can violate
the performance bound γ2 by some matrix in the complex uncertainty set γ12 ∆c .

Let us now interpret the upper and lower bound plots of the SSV of N for all frequencies
(Figure 56).

Since the upper bound is not larger than 2.72 for all frequencies, we infer that
1
kS(∆, N )k∞ ≤ 2.72 for all ∆ ∈ ∆ ≈ 0.367∆
2.72
Since the lower bound is larger that 2.71 for some frequency, we infer that either
1
(I − M ∆)−1 is unstable for some ∆ ∈ ∆ ≈ 0.369∆
2.71
or that
1
kS(∆, N )k∞ > 2.71 for some ∆ ∈ ∆ ≈ 0.369∆.
2.71

154
 Maximal upper bound 1.91 Maximal lower bound 1.883
1.95

1.9

1.85

1.8

1.75

1.7

1.65

1.6 −2 −1 0 1 2
10 10 10 10 10

Figure 57: Upper and lower bound of SSV of scaled N .

The first property can be certainly excluded since Figure 55 reveals that µ∆c (M (iω)) ≤ 2.7
for all ω ∈ R ∪ {∞}.

Let us finally ask ourselves for which size of the uncertainties we can guarantee a robust
performance level of 2.

For that purpose let us plot (Figure 57) the SSV of

    
M N12 0.5I 0 0.5M N12
  = .
N21 N22 0 I 0.5N21 N22

Since the upper bound is not larger than 1.92 for all frequencies, we conclude
0.5
kS(∆, N )k∞ ≤ 1.92 for all ∆ ∈ ∆ ≈ 0.26∆.
1.92

155
Exercises

1) Look at a standard tracking configuration for a system G(I + ∆W ) with input

multiplicative uncertainty and a controller K that is described as

y = G(I + ∆W )u, k∆k∞ < 1, u = K(r − y).

The performance is as desired if the transfer matrix from references r to weighted

error V (r −y) has an H∞ -norm smaller than 1. Here G, K, ∆, V , W are LTI system
and the latter three are stable.

a) Set up the generalized plant.

 Show 
that the weighted closed-loop transfer
−M1 G M1
matrix has the structure   by computing M1 and M2 . Formulate
−M2 G M2
the µ-tests for robust stability, nominal performance and robust performance.
b) Now let all LTI systems G, K, ∆, V , W be SISO and define S = (I + KG)−1 ,
T = (I + KG)−1 KG. Show that K achieves robust performance iff

|V (iω)S(iω)| + |W (iω)T (iω)| ≤ 1 for all ω ∈ R ∪ {∞}. (6.1)

(Hint: This is a SSV-problem for rank one matrices!) If a controller achieves

robust stability and nominal performance, what can you conclude about robust
performance? How would you design robustly performing controllers by solving
an H∞ problem?
c) Let’s return to the MIMO case. Suppose that G is square and has a proper
inverse G−1 . Show that the SSV for the robust performance test is (at frequency
ω) bounded from above by

kG(iω)kkG(iω)−1 kkM1 (iω)G(iω)k + kM2 (iω)k.

If a controller achieves robust stability and nominal performance, what can you
now conclude for robust performance? Discuss the role of the plant condition
number kG(iω)kkG(iω)−1 k!
d) For any γ, construct
 complex matrices
 M1 , M2 and G such that kM1 Gk ≤ 1,
−M1 G M1
kM2 k ≤ 1, but µ( ) ≥ γ. Here, µ is computed with respect to
−M2 G M2
an uncertainty structure with two full blocks. What does this example show?
Hint: Construct the example such that kGkkG−1 k is large.

2) Consider the block diagram in Figure 58 where G and H are described by the
transfer functions
1 200
G(s) = and H(s) = .
(0.05s + 1)2 10s + 1

156
 d

∆ W 0.5

r e
+ K + G + H

−1

Figure 58: Tracking interconnection for exercise 2).

The magnitude of the uncertainty is not larger than 1% at low frequencies, it

does not exceed 100% at 30 rad/sec, and for larger frequencies it increases by
40 db per decade.

a) Design a weighting W that captures the specifications on the uncertainty, and

build the open-loop interconnection that corresponds to the block diagram.
b) Consider static gain controllers u = Ky with K in the interval [1, 3]. Which
controllers stabilize the interconnection? What do you observe if you increase
the gain of the controller for the tracking and disturbance suppression behavior
of the controlled system? What is the effect on robust stability with respect
to the given class of uncertainties?
c) With performance weightings of the form

s/b + 1
a∗ , a, b, c real,
s/c + 1

for the channels d → e and r → e, design a controller with the H∞ algorithm

to achieve the following performance specifications:
i. Disturbance attenuation of a factor 100 up to 0.1 rad/sec.
ii. Zero steady state response of tracking error (smaller than 10−4 ) for a ref-
erences step input.
√
iii. Bandwidth (frequency where magnitude plot first crosses 1/ 2 ≈ −3 dB
from below) from reference input to tracking error between 10 rad/sec and
20 rad/sec.
iv. Overshoot of tracking error is less than 7% in response to step reference.
Provide magnitude plots of all relevant transfer functions and discuss the re-
sults.

157
d) For the design you performed in 2c), what is the maximal size γ∗ of uncertainties
that do not destabilize the controlled system (stability margin). Compute a
destabilizing perturbation of size larger than γ∗ .
e) Extend the H∞ specification of 2c) by the uncertainty channel and perform
a new design with the same performance weightings. To what amount do
you need to give up the specifications to guarantee robust stability? If you
compare the two designs, what do you conclude about performing a nominal
design without taking robustness specifications into account?

158
7 Synthesis of H∞ Controllers

In this section we provide a self-contained and elementary route to solve the H∞ -problem.
We first describe how to bound or compute the H∞ -norm of stable transfer matrix in
terms of a suitable Hamiltonian matrix. Then we present a classical result, the so-called
Bounded Real Lemma, that characterizes (in terms of a state-space realization) when a
given transfer matrix has an H∞ -norm which is strictly smaller than a number γ. On
the basis of the Bounded Real Lemma, we will derive the celebrated solution of the H∞
control problem in terms of two algebraic Riccati equations and a coupling condition on
their solutions. We sacrifice generality to render most of the derivations as elementary as
possible.

7.1 The Algebraic Riccati Equation and Inequality

The basis for our approach to the H∞ problem is the algebraic Riccati equation or in-
equality. It occurs in proving the Bounded Real Lemma and it comes back in getting to
the Riccati solution of the H∞ problem.

Given symmetric matrices R ≥ 0 and Q, we consider the strict algebraic Riccati inequality

AT X + XA + XRX + Q < 0 (ARI)

and the corresponding algebraic Riccati equation

AT X + XA + XRX + Q = 0. (ARE)

Note that X is always assumed to be real symmetric or complex Hermitian. Moreover,

we allow for a general indefinite Q.

It will turn out that the solutions of the ARE with the property that A + RX has all
its eigenvalues in C< or in C> play a special role. Such a solutions are called stabilizing
or anti-stabilizing. If (A, R) is controllable, we can summarize the results in this section
as follows: The ARE or ARI have solutions if and only if a certain Hamiltonian matrix
defined through A, R, Q has no eigenvalues on the imaginary axis. If the ARI or the
ARE has a solution, there exists a unique stabilizing solution X− and a unique anti-
stabilizing solution X+ of the ARE, and all other solutions X of the ARE or ARI satisfy
X− ≤ X ≤ X+ . Here is the main result whose proof is given in the appendix.

Theorem 7.1 Suppose that Q is symmetric, that R is positive semi-definite, and that
(A, R) is controllable. Define the Hamiltonian matrix
 
A R
H :=  .
−Q −AT

159
Then the following statements are equivalent:

(a) H has no eigenvalues on the imaginary axis.

(b) AT X + XA + XRX + Q = 0 has a (unique) stabilizing solution X− .

(c) AT X + XA + XRX + Q = 0 has a (unique) anti-stabilizing solution X+ .

(d) AT X + XA + XRX + Q < 0 has a symmetric solution X.

If one and hence all of these conditions are satisfied, then

any solution X of the ARE or ARI satisfies X− ≤ X ≤ X+ .

We conclude that the stabilizing solution is the smallest among all solutions of the ARE
and the anti-stabilizing solution is the largest.

Remark. Note that H has a specific structure: The off-diagonal blocks are symmetric,
and the second block on the diagonal results from the first by reversing the sign and
transposing. Any such matrix is called a Hamiltonian matrix.

If (A, R) is only stabilizable, X+ does, in general, not exist. All other statements, however,
remain true. Here is the precise results that is proved in the appendix.

Theorem 7.2 Suppose that all hypothesis in Theorem 7.1 hold true but that (A, R) is
only stabilizable. Then the following statements are equivalent:

(a) H has no eigenvalues on the imaginary axis.

(b) AT X + XA + XRX + Q = 0 has a (unique) stabilizing solution X− .

(c) AT X + XA + XRX + Q < 0 has a symmetric solution X.

If one and hence all of these conditions are satisfied, then

any solution X of the ARE or ARI satisfies X− ≤ X.

The proof reveals that it is not difficult to construct a solution once one has verified that
H has no eigenvalues on the imaginary axis. We sketch the typical algorithm that is used
in software packages like Matlab.

Indeed, let H have no eigenvalues in C= . Then it has n eigenvalues in C< and n eigenvalues
in C> respectively. We can perform a Schur decomposition to obtain a unitary matrix T
with  
M11 M12
T ∗ HT =  
0 M22

160
where M11 of size n × n is stable and M22 of size n × n is anti-stable. Partition T into
four n × n blocks as  
U ∗
T = .
V ∗
The proof of Theorem 7.2 reveals that U is non-singular, and that the stabilizing solution
of the ARE is given by
X = V U −1 .
If the Schur decomposition is chosen such that M11 has all its eigenvalues in C> and M22
is stable, then the same procedure leads to the anti-stabilizing solution.

If Q is negative semi-definite, the eigenvalues of the Hamiltonian matrix on the imaginary

axis are just given by uncontrollable or unobservable modes. The exact statement reads
as follows.

Lemma 7.3 If R ≥ 0 and Q ≤ 0 then

 
A R
λ ∈ C= is an eigenvalue of H =  
T
−Q −A

if and only if

λ ∈ C= is an uncontrollable mode of (A, R) or an unobservable mode of (A, Q).

Proof. iω is an eigenvalue of H if and only if

   
A − iωI R A − iωI R
H − iωI =  = 
−Q −AT − iωI −Q −(A − iωI)∗

is singular.
 
If iω is an uncontrollable mode of (A, R), then A − iωI R does not have full row rank;
 
A − iωI
if it is an unobservable mode of (A, Q), then   does not have full column rank;
−Q
in both cases we infer that H − iωI is singular.

Conversely, suppose that H − iωI is singular. Then there exist x and y, not both zero,
with   
A − iωI R x
    = 0.
−Q −(A − iω)∗ y

161
This implies
(A − iωI)x + Ry = 0 and − Qx − (A − iωI)∗ y = 0. (7.1)
Left-multiply the first equation with y ∗ and the second equation with x∗ to get

y ∗ (A − iωI)x + y ∗ Ry = 0
−x∗ Qx − y ∗ (A − iωI)x = −x∗ Qx − x∗ (A − iωI)∗ y = 0.

This leads to
y ∗ Ry = x∗ Qx.
Since R ≥ 0 and Q ≤ 0, we infer Qx = 0 and Ry = 0. Then (7.1) implies (A − iωI)x = 0
and (A − iωI)∗ y = 0. If x 6= 0, iω is an unobservable mode of (A, Q), if y 6= 0, it is an
uncontrollable mode of (A, R).

Hence, if (A, R) is stabilizable, if Q ≤ 0, and if (A, Q) does not have unobservable modes
on the imaginary axis, the corresponding Hamiltonian matrix does not have eigenvalues
on the imaginary axis such that the underlying ARE has a stabilizing solution and the
ARI is solvable as well.

7.2 Computation of H∞ Norms

Consider the strictly proper transfer matrix M with realization

M (s) = C(sI − A)−1 B

where A is stable. Recall that the H∞ -norm of M is defined by

kM k∞ := sup kM (iω)k.
ω∈R

In general it is not advisable to compute the H∞ -norm of M by solving this optimization

problem. In this section we clarify how one can arrive at a more efficient computation of
this norm by looking, instead, at the following problem: Characterize in terms of A, B,
C whether the inequality
kM k∞ < 1 (7.2)
is true or not. Just by the definition of the H∞ norm, (7.2) is equivalent to

kM (iω)k < 1 for all ω ∈ R ∪ {∞}.

Since M is strictly proper, this inequality is always true for ω = ∞. Hence it remains to
consider
kM (iω)k < 1 for all ω ∈ R. (7.3)

162
If follows by continuity that this is true if and only if

det(M (iω)∗ M (iω) − I) 6= 0 for all ω ∈ R. (7.4)

Indeed, kM (iω)k < 1 implies that the largest eigenvalue of M (iω)∗ M (iω) is smaller than
1 such that det(M (iω)∗ M (iω) − I) 6= 0. Hence (7.3) implies (7.4). Conversely, suppose
(7.3) is not true. Then there exists a ω0 ∈ R for which kM (iω0 )k ≥ 1. Consider the real-
valued function ω → kM (iω)k which is continuous (as the norm of a rational function
without pole). Due to limω→∞ kM (iω)k = 0, there exists an ω1 > ω0 with kM (iω1 )k < 1.
By the intermediate value theorem, there exists some point in the interval ω∗ ∈ [ω0 , ω1 ]
with kM (iω∗ )k = 1. This implies det(M (iω∗ )∗ M (iω∗ ) − I) = 0 such that (7.4) is not true.

Since M is real rational we have M (iω)∗ = M (−iω)T . If we hence define

G(s) := M T (−s)M (s) − I,

(7.4) is the same as

det(G(iω)) 6= 0 for all ω ∈ R.
Since

M (−s)T = [C(−sI − A)−1 B]T = B T (−(sI + AT )−1 )C T = B T (sI − (−AT ))−1 (−C T ),

one easily obtains a state-space realization of G as

 
A 0 B
 
G= T T .
 −C C −A 0 
T
0 B −I

Let us now apply the Schur formula3 to this realization for s = iω. If we introduce the
abbreviation
     
T
A 0 B   A BB
H :=   −   (−I)−1 0 B T =  , (7.5)
T T T T
−C C −A 0 −C C −A

we arrive at
det(−I)
det(G(iω)) = det(iωI − H).
det(iωI − A) det(iωI + AT )

Now recall that A is stable such that both det(iωI − A) and det(iωI + AT ) do not vanish.
Hence
det(G(iω)) = 0 if and only if iω is an eigenvalue of H.
Hence (7.4) is equivalent to the fact that H does not have eigenvalues on the imaginary
axis. This leads to the following characterization of the H∞ -norm bound kM k∞ < 1.
det(D)
3
If D is nonsingular, det(C(sI − A)−1 B + D) = det(sI−A) det(sI − (A − BD−1 C))

163
  
A BB T
Theorem 7.4 kM k∞ < 1 if and only if   has no eigenvalues on the
−C T C −AT
imaginary axis.

To compute kM k∞ , we actually need to verify whether, for any given positive number γ,

kC(sI − A)−1 Bk∞ < γ (7.6)

is valid or not. Indeed, the inequality is the same as

   
1 −1 −1 1
k C (sI − A) Bk∞ < 1 or kC(sI − A) B k∞ < 1 (7.7)
γ γ

such that it suffices to re-scale either B or C by the factor γ1 to reduce the test to the one
with bound 1. We conclude: (7.6) holds if and only if
 
1 T
A γ2
BB
  has no eigenvalues on the imaginary axis
−C T C −AT

or, equivalently,
 
A BB T
  has no eigenvalues on the imaginary axis.
− γ12 C T C −AT

Why does this result help? It allows to check kM k∞ < γ, a test which involves computing
the norm at infinitely many frequencies, by just verifying whether a Hamiltonian matrix
that is defined through the data matrices A, B, C and the bound γ has an eigenvalue on
the imaginary axis or not. This allows to compute kM k∞ by bisection (Appendix A).

7.3 The Bounded Real Lemma

The characterization of kM k∞ < 1 in terms of the Hamiltonian matrix H is suitable

for computing the H∞ -norm of M , but it is not convenient to derive a solution of the
H∞ problem. For that purpose we aim at providing an alternative characterization of
kM k∞ < 1 in terms of the solvability of Riccati equations or inequalities. In view of our
preparations showing a relation of the solvability of the ARE or ARI with Hamiltonians,
this is not too surprising.

Theorem 7.5 Let M (s) = C(sI − A)−1 B with A being stable. Then kM k∞ < 1 holds if
and only if the ARI
AT X + XA + XBB T X + C T C < 0 (7.8)

164
has a solution. This is equivalent to the fact that the ARE

AT X + XA + XBB T X + C T C = 0 (7.9)

has a stabilizing solution.

We only need to observe that the stability of A implies that (A, BB T ) is stabilizable.
Then we can just combine Theorem 7.2 with Theorem 7.4 to obtain Theorem 7.5.

Since the H∞ -norms of C(sI − A)−1 B and of B T (sI − AT )−1 C T coincide, we can dualize
this result.

Theorem 7.6 Let M (s) = C(sI − A)−1 B with A being stable. Then kM k∞ < 1 holds if
and only if the ARI
AY + Y AT + BB T + Y C T CY < 0 (7.10)
has a solution. This is equivalent to the fact that the ARE

AY + Y AT + BB T + Y C T CY = 0 (7.11)

has a stabilizing solution.

Task. Provide the arguments why these statements are true.

Remarks.

• Recall how we reduced the bound (7.6) for some γ > 0 to (7.7). Hence (7.6) can be
characterized by performing the substitutions
1 1 T
BB T → BB T
or C T
C → C C
γ2 γ2
in all the four AREs or ARIs.

• If X satisfies (7.8), it must be non-singular: Suppose Xx = 0 with x 6= 0. Then

xT (7.8)x = xT BB T x < 0. This implies kB T xk < 0, a contradiction. If we note that
X −1 (7.9)X −1 implies

AX −1 + X −1 AT + BB T + X −1 C T CX −1 < 0,

we infer that Y = X −1 satisfies (7.10). Conversely, if Y solves (7.10), then Y −1

exists and satisfies (7.8).

165
7.4 The H∞ -Control Problem

Let us consider the generalized plant

 
    A B1 B2  
z w   w
  = P   =  C1 D11 D12    .
y u u
 
C2 D21 D22

Recall that y is the measured output available for control, u is the control input, z is the
controlled output, and w the disturbance input. We assume that P admits a stabilizing
controller such that

(A, B2 ) is stabilizable and (A, C2 ) is detectable.

As controllers we allow for any LTI system

 
AK BK
u = Ky =  y
CK DK

specified in the state-space through the parameter matrices AK , BK , CK , DK .

The goal in H∞ -control is to minimize the H∞ -norm of the transfer function w → z by

using stabilizing controllers. With the previous notation for the controlled closed-loop
system  
A B
z = S(P, K)w =   w,
C D
the intention is to minimize
 
A B
kS(P, K)k∞ =  
C D
∞

over all K that stabilizes P , i.e., which render A stable.

Similarly as for just determining the H∞ -norm of a transfer matrix, we rather consider the
so-called sub-optimal H∞ control problem: Given the number γ > 0, find a controller
K such that
K stabilizes P and achieves kS(P, K)k∞ < γ
or conclude that no such controller exists.

As usual, we can rescale to the bound 1 by introducing weightings. This amounts to

substituting P by either one of
     
1 1 1 1 1
P P P P P √ P
 γ 11 γ 12  ,  γ 11 12  ,  γ 11 γ 12 
1 √1 P21
P21 P22 P P22
γ 21 γ
P22

166
that read in the state-space as
     
A B1 B2 A 1
B B2 A √1 B1 B2
γ 1  γ 
   
 1C 1D 1D  ,  C 1 √1 C1 1 √1 D12
D ,
D12  D .
 
 γ 1 γ 11 γ 12   1 γ 11 γ γ 11 γ

 
1 1
C2 D21 D22 C2 D
γ 21
D22 C2 √ D21
γ
D22

Hence the problem is re-formulated as follows: Try to find a controller K such that
K stabilizes P and achieves kS(P, K)k∞ < 1. (7.12)
The conditions (7.12) read in the state-space as
λ(A) ⊂ C< and kC(sI − A)−1 B + Dk∞ < 1. (7.13)

Note that a K might or might not exist. Hence our goal is to provide verifiable conditions
formulated in terms of the generalized plant P for the existence of such a controller K. If
a controller is known to exist, we also need to devise an algorithm that allows to construct
a suitable controller K which renders the conditions in (7.12) or (7.13) satisfied.

7.5 H∞ -Control for a Simplified Generalized Plant Description

The generalized plant reads as

ẋ = Ax + B1 w + B2 u
z = C1 x + D11 w + D12 u
y = C2 x + D21 w + D22 u.
The derivation of a solution of the H∞ -problem in its generality is most easily obtained
with LMI techniques. In these notes we only consider the so-called regular problem. This
amounts to the hypothesis that
D12 has full column rank and D21 has full row rank.
These assumptions basically imply that the full control signal u appears via D12 u in z,
and that the whole measured output signal y is corrupted via D21 w by noise.

In order to simplify both the derivation and the formulas, we confine the discussion to a
generalized plant with the following stronger properties:
   
T
    B1 T
0
D11 = 0, D22 = 0, D12 C1 D12 = 0 I ,   D21 =  . (7.14)
D21 I
Hence we assume that both P11 and P22 are strictly proper. Moreover, D12 does not only
have full column rank, but its columns are orthonormal, and they are orthogonal to the
columns of C1 . Similarly, the rows of D21 are orthonormal, and they are orthogonal to
the rows of B1 .

167
7.6 The State-Feedback Problem

Let us first concentrate on the specific control structure

u = Fx

what is often denoted as static state-feedback. The closed-loop system then reads as

ẋ = (A + B2 F )x + B1 w
z = (C1 + D12 F )x.

Hence the static-state feedback H∞ control problem is formulated as follows: Find an F

which renders the following two conditions satisfied:

λ(A + B2 F ) ⊂ C< and k(C1 + D12 F )(sI − A − B2 F )−1 B1 k∞ < 1. (7.15)

7.6.1 Solution in Terms of Riccati Inequalities

The gain F satisfies both conditions if and only if there exists a Y with

Y > 0, (A + B2 F )Y + Y (A + B2 F )T + B1 B1T + Y (C1 + D12 F )T (C1 + D12 F )Y < 0. (7.16)

Indeed, if F satisfies (7.15), we can apply Theorem 7.6 to infer that the ARI in (7.16) has
a symmetric solution Y . Since the inequality implies (A + B2 F )Y + Y (A + B2 F )T < 0,
and since A + B2 F is stable, we infer that Y is actually positive definite. Conversely, if
Y satisfies (7.16), then (A + B2 F )Y + Y (A + B2 F )T < 0 implies that A + B2 F is stable.
Again by Theorem 7.6 we arrive at (7.15).

In a next step, we eliminate F from (7.16). This will be possible on the basis of the
following lemma.

Lemma 7.7 For any F ,

(A + B2 F )Y + Y (A + B2 F )T + B1 B1T + Y (C1 + D12 F )T (C1 + D12 F )Y ≥

≥ AY + Y AT + B1 B1T − B2 B2T + Y C1T C1 Y.

Equality holds if and only if

F Y + B2 = 0.

Proof. Note that (7.14) implies

Y (C1 + D12 F )T (C1 + D12 F )Y = Y C1T C1 Y + Y F T F Y.

168
Moreover, we exploit (completion of the squares)

(B2 F Y ) + (B2 F Y )T + Y F T F Y = −B2 B2T + (F Y + B2T )T (F Y + B2T ).

Both equations imply

(A + B2 F )Y + Y (A + B2 F )T + B1 B1T + Y (C1 + D12 F )T (C1 + D12 F )Y =

= AY + Y AT + B1 B1T − B2 B2T + Y C1T C1 Y + (F Y + B2T )T (F Y + B2T ).

Since (F Y + B2T )T (F Y + B2T ) ≥ 0 and since (F Y + B2T )T (F Y + B2T ) = 0 if and only if

F Y + B2T = 0, the proof is finished.

Due to this lemma, any Y that satisfies (7.16) also satisfies

Y > 0, AY + Y AT + B1 B1T − B2 B2T + Y C1T C1 Y < 0. (7.17)

Note that the resulting ARI is independent of F ! Conversely, if Y is a matrix with (7.17),
it is non-singular such that we can set

F := −B2T Y −1 .

Since F Y + B2 vanishes, we can again apply Lemma 7.7 to see that (7.17) implies (7.16).
To conclude, there exists an F and a Y with (7.16) if and only if there exists a Y with
(7.17).

Let us summarize what we have found in the following Riccati inequality solution of the
state-feedback H∞ problem.

Theorem 7.8 The gain F solves the state-feedback H∞ -problem (7.15) if and only if
there exists a positive definite solution Y of the ARI

AY + Y AT + B1 B1T − B2 B2T + Y C1T C1 Y < 0. (7.18)

If Y > 0 is any solution of this ARI, then the gain

F := −B2T Y −1

renders A + B2 F stable and leads to k(C1 + D12 F )(sI − A − B2 F )−1 B1 k∞ < 1.

At this point it is unclear how we can test whether (7.18) has a positive definite solution.
One possibility is as follows: Observe that Y > 0 and (7.18) are equivalent to (Schur
complement)  
AY + Y AT + B1 B1T − B2 B2T Y C1T
Y > 0,   < 0.
C1 Y −I
These are two linear matrix inequalities and, hence, they can be readily solved by existing
software.

169
7.6.2 Solution in Terms of Riccati Equations

There is an alternative. We can reduce the solvability of (7.18) to a test for a certain
Hamiltonian matrix via Theorem 7.1. For that purpose we need an additional technical
hypothesis; we have to require that (AT , C1T C1 ) is controllable. By the Hautus test, this is
the same as (A, C1 ) being observable. (Why?) It is important to note that this assumption
is purely technical and makes it possible to provide a solution of the state-feedback H∞
problem by Riccati equations; it might happen that this property fails to hold such that
one has to rely on alternative techniques.

Theorem 7.9 Suppose that

(A, C1 ) is observable. (7.19)
Then there exists an F that solves the state-feedback H∞ -problem (7.15) if and only if the
Riccati equation
AY + Y AT + B1 B1T − B2 B2T + Y C1T C1 Y = 0 (7.20)
has an anti-stabilizing solution Y+ , and this anti-stabilizing solution is positive definite.
If Y+ > 0 denotes the anti-stabilizing solution of the ARE, the gain

F := −B2T Y+−1

leads to (7.15).

Proof. If there exists an F that satisfies (7.15), the ARI (7.18) has a solution Y > 0. By
Theorem 7.2, the ARE (7.20) has an anti-stabilizing solution Y+ that satisfies Y ≤ Y+ .
Hence Y+ exists and is positive definite.

Conversely, suppose Y+ exists and is positive definite. With F = −B2T Y+−1 we infer
F Y+ + B2 = 0 and hence

AY+ + Y+ AT + GGT − B2 B2T + Y+ C1T C1 Y+ + (F Y+ + B2T )T (F Y+ + B2T ) = 0.

By Lemma 7.7, we arrive at

(A + B2 F )Y+ + Y+ (A + B2 F )T + B1 B1T + Y+ (C1 + D12 F )T (C1 + D12 F )Y+ = 0. (7.21)

Let us first show that A + B2 F is stable. For that purpose let (A + B2 F )T x = λx with
x 6= 0. Now look at x∗ (7.21)x:

Re(λ)x∗ Y+ x + x∗ B1 B1T x + x∗ Y+ (C1 + D12 F )T (C1 + D12 F )Y+ x = 0.

Since Y+ ≥ 0, this implies Re(λ) ≥ 0. Let us now exclude Re(λ) = 0. In fact, this
condition implies (C1 + D12 F )Y+ x = 0. If we left-multiply D12 , we can exploit (7.14) to

170
 +
e z System
− w

Observer y
ẑ

Figure 59: H∞ -observer design.

infer F Y+ x = 0. This implies B2T x = 0 and hence AT x = λx. Since (A, B2 ) is stabilizable,
we obtain Re(λ) < 0, a contradiction. Therefore, the real part of λ must be negative,
what implies that A + B2 F is stable.

Since
A + B2 F + Y+ (C1 + D12 F )T (C1 + D12 F ) = A + Y+ C1T C1 ,
we infer that Y+ is the anti-stabilizing solution of (7.21). Hence the corresponding Hamil-
tonian matrix has no eigenvalues on the imaginary axis (Theorem 7.1) such that (7.15)
follows (Theorem 7.4).

7.7 H∞ -Observer Design

Suppose a given system is affected by a genralized disturbance w. Based on the measure-

ment output y, the goal is to estimate the output z of the system as closely as possible.

This should be achieved by an observer, a copy of the system (while neglecting the un-
known disturbance input) that processes the measured output y to generate the estimate
ẑ, which leads to the estimation error

e = z − ẑ.

In H∞ -observer design, the quality of the observer is measured in terms of the H∞ -norm
of the transfer matrix defined by w → e.

Let the system be described by

ẋ = Ax + B1 w, z = C1 x, y = C2 x + D21 w.

An observer is then defined with suitable gain L as

x̂˙ = Ax̂ + L(C2 x̂ − y), ẑ = C1 x̂.

For the error dynamics (with state ξ := x − x̂) as described by

ξ˙ = (A + LC2 )ξ + (B1 + LD21 )w, e = C1 ξ,

171
the goal is to achieve with some γ > 0:

eig(A + LC2 ) ⊂ C− and kC1 (sI − A − LC2 )−1 (B1 + LD21 )k∞ < γ (OD)

Any observer that guarantees (OD) leads to the following properties:

• For nonzero initial condition and the absence of the disturbances, the observer state
converges to the system state asymptotically:

lim (x(t) − x̂(t)) = 0

t→∞

• For zero initial conditions, the estimation error z − ẑ is endered small in worst-case
for all disturbances w, in the sense that the H∞ -norm of w → (z − ẑ) is smaller
than γ.

• The smallest achievable γ is called optimal H∞ -estimation level.

• Recall all our earlier discussions concerning the interpretation of this measure!
Morevoer, dynamic filters for capturing spectral properties of the disturbance are
again assumed to be absorbed into the system description at the input.

The following result is easily established by duality.

T T
Theorem 7.10 Let (A, C2 ) be detectable, B1 D21 = 0, D21 D21 = I and suppose that
(A, B1 ) is controllable. Then there exists an observer gain L which achieves (OD) if and
only if the anti-stabilizing solution X+ of the ARE

AT X + XA + XB1 B1T X − γ −2 C1T C1 − C2T C2 = 0

exists and is positive definite. If the anti-stabilizing solution X+ of the ARE exists and is
positife definite, L := −X+−1 C2T satisfies (OD).

For the proof just observe that (OD) is equivalent to

eig(AT + C2T LT ) ⊂ C− and k(B1T + D21

T T
L )(sI − AT − C2T LT )−1 C1T k∞ < γ.

Then apply Theorem 7.9 for (AT , C1T , C2T , B1T , D21
T
) substituting (A, B1 , B2 , C1 , D12 ).

7.8 H∞ -Control by Output-Feedback

Let us now come back to the H∞ -problem by output feedback control. This amounts to
finding the matrices AK , BK , CK , DK such that the conditions (7.13) are satisfied.

We proceed as in the state-feedback problem. We use the Bounded Real Lemma to rewrite
the H∞ norm bound into the solvability of a Riccati inequality, and we try to eliminate
the controller parameters to arrive at verifiable conditions.

172
7.8.1 Solution in Terms of Riccati Inequalities

For the derivation to follow we assume that the controller is strictly proper: DK = 0. The
proof for a controller with DK 6= 0 is only slightly more complicated and fits much better
into the LMI framework what will be discussed in the LMI course.

There exists a controller AK , BK , CK , DK = 0 that satisfies (7.13) if and only if there

exists some X with

X > 0, AT X + X A + X BB T X + C T C < 0. (7.22)

Indeed, (7.13) implies the existence of a symmetric solution of the ARI in (7.22) by
Theorem 7.5. Since A is stable, AT X + X A < 0 implies that X > 0. Conversely, (7.22)
implies AT X + X A < 0 for X > 0. Hence A is stable and we arrive, again by Theorem
7.5, at (7.13).

Let us now suppose that (7.22) is valid. Partition X and X −1 in the same fashion as A
to obtain    
X U Y V
X =  and X −1 =  .
UT X
b V T Yb
It is then obvious that
   
X U I 0
R=  and S =   satisfy SX = R.
I 0 Y V

We can assume without loss of generality that the size of AK is not smaller than n. Hence
the right upper block U of X has more columns than rows. This allows to assume that
U is of full row rank; if not true one just needs to slightly perturb this block without
violating the strict inequalities (7.22). Then we conclude that R has full row rank. Due
to SX = R, also S has full row rank.

Let us now left-multiply both inequalities in (7.22) with S and right-multiply with S T .
Since S has full row rank and if we exploit SX = R or X S T = RT , we arrive at

SRT > 0, SAT RT + RAS T + RBB T RT + SC T CS T < 0. (7.23)

173
The appearing blocks are most easily computed as follows:
 
X I
SRT =  
Y X + V UT Y
 
T T
XA + U BK C2 XAY + U BK C2 Y + XB2 CK V + U AK V
RAS T =  
A AY + B2 CK V T
 
XB1 + U BK D21
RB =  
B1
 
CS T = C1 C1 Y + D12 CK V T .

Let us now recall a relation between X, Y , U , V and define the new variables F and L
as in
Y X + V U T = I, L = X −1 U BK and F = CK V T Y −1 . (7.24)
Then the formulas simplify to
 
X I
SRT =  
I Y
 
T
X(A + LC2 ) X(A + LC 2 + B2 F )Y + U A K V
RAS T =  
A (A + B2 F )Y
 
X(B1 + LD21 )
RB =  
B1
 
CS T = C1 (C1 + D12 F )Y .

We conclude that (7.23) read as  

X I
 >0
I Y
and
 
T T T
(A + LC2 ) X + X(A + LC2 ) A + X(A + LC2 + B2 F )Y + U AK V
 +
A + [X(A + LC2 + B2 F )Y + U AK V T ]T (A + B2 F )Y + Y (A + B2 F )T
 
T T
X(B1 + LD21 )(B1 + LD21 ) X X(B1 + LD21 )B1
+ +
B1 (B1 + LD21 )T X B1 B1T
 
T T
C1 C1 C1 (C1 + D12 F )Y
+  < 0. (7.25)
T T
Y (C1 + D12 F ) C1 Y (C1 + D12 F ) (C1 + D12 F )Y

174
If we pick out the diagonal blocks of the last inequality, we conclude
 
X I
 >0 (7.26)
I Y
(A + LC2 )T X + X(A + LC2 ) + X(B1 + LD21 )(B1 + LD21 )T X + C1T C1 < 0 (7.27)
T
(A + B2 F )Y + Y (A + B2 F ) + B1 B1T T
+ Y (C1 + D12 F ) (C1 + D12 F )Y < 0. (7.28)

We can apply Lemma 7.7 to (7.28) and a dual version to (7.27). This implies that X and
Y satisfy (7.26) and the two algebraic Riccati inequalities

AT X + XA + XB1 B1T X + C1T C1 − C2T C2 < 0 (7.29)

AY + Y AT + B1 B1T − B2 B2T + Y C1T C1 Y < 0. (7.30)

We have shown: If there exists a controller that renders (7.13) satisfies, then there exist
symmetric matrices X and Y that satisfy the two algebraic Riccati inequalities (7.29)-
(7.30), and that these two matrices are coupled as (7.26). Note that these conditions
are, again, formulated only in terms of the matrices defining the generalized plant; the
controller parameters have been eliminated. It will turn out that one can reverse the
arguments: If X and Y satisfy (7.26),(7.29)-(7.30), one can construct a controller that
leads to (7.13).

Suppose X and Y satisfy (7.26),(7.29)-(7.30). Due to (7.26), X and Y are positive definite
and hence nonsingular. If we apply Lemma 7.7 to (7.30) and a dual version to (7.29), we
conclude that
L := −X −1 C2T and F := −B2T Y −1
lead to (7.26)-(7.28). Again due to (7.26), I − Y X is non-singular as well. Hence we can
find non-singular square matrices U and V satisfying V U T = I − Y X; take for example
U = I and V = I − Y X. If we set

BK = U −1 XL and CK = F Y V −T , (7.31)

we infer that the relations (7.24) are valid. If we now consider (7.25), we observe that
the only undefined block on the left-hand side is AK and this appears as U AK V T in the
off-diagonal position. We also observe that the diagonal blocks of this matrix are negative
definite. If we choose AK to render the off-diagonal block zero, we infer that (7.25) is
valid. This is clearly achieved for

AK = −U −1 [AT +X(A+LC2 +B2 F )Y +X(B1 +LD21 )B1T +C1T (C1 +D12 F )Y ]V −T . (7.32)

We arrive back to (7.23). Now S is square and non-singular. We can hence define X
through X := S −1 R. If we left-multiply both inequalities in (7.23) with S −1 and right-
multiply with S −T , we arrive at (7.22). This shows that the constructed controller leads
to (7.13) and the proof cycle is complete.

175
Theorem 7.11 There exist AK , BK , CK , DK that solve the output-feedback H∞ problem
(7.13) if and only if there exist X and Y that satisfy the two ARIs

AT X + XA + XB1 B1T X + C1T C1 − C2T C2 < 0, (7.33)

AY + Y AT + Y C1T C1 Y + B1 B1T − B2 B2T < 0, (7.34)

and the coupling condition  
X I
  > 0. (7.35)
I Y

Suppose that X and Y satisfy (7.33)-(7.35). Let U and V be square and non-singular
matrices with U V T = I − XY , and set L := −X −1 C2T , F := −B2T Y −1 . Then AK , BK ,
CK as defined in (7.32),(7.31) lead to (7.13).

Remarks. The necessity of the conditions (7.33)-(7.35) has been proved for a strictly
proper controller only. In fact, the conclusion does not require this hypothesis. On
the other hand, the controller as given in the construction is always strictly proper. In
general, if D11 = 0, the existence of a proper controller solving the H∞ -problem implies
the existence of a strictly proper controller that solves the problem. Note, however, that
one does in general not get a strictly proper controller that solves the H∞ -problem by
simply removing the direct feedthrough term from a non-proper controller that solves the
H∞ -problem!

Remark. If we recall (7.14), the formulas for the controller matrices can be simplified to

AK = −U −1 [AT + XAY + X(B1 B1T − B2 B2T ) + (C1T C1 − C2T C2 )Y ]V −T

BK = −U −1 C2T , (7.36)
CK = −B2T V −T .

For the particular choices U = X, V = X −1 − Y or U = Y −1 − X, V = Y , one arrives

at the specific formulas given in the literature. Note that all these constructions lead to
controllers that have the same order as the generalized plant: the dimension of Ac is equal
to the dimension of A.

We note again that (7.34)-(7.35) are equivalent to

 
T T T
A X + XA + C1 C1 − C2 C2 XB1
 <0
B1T X −I

and  
T
AY + Y A + B1 B1T − B2 B2T Y C1T
  < 0.
C1 Y −I

176
Hence testing the existence of solutions X and Y of (7.33)-(7.35) can be reduced to
verifying the solvability of a system of linear matrix inequalities. In fact, numerical
solvers provide solutions X, Y , and we have shown how to construct on the basis of these
matrices a controller that satisfies (7.13).

7.8.2 Solution in Terms of Riccati Equations

As an alternative, Theorem 7.1 allows to go back to Riccati equations if (A, B1 B1T ) and
(AT , C1T C1 ) are controllable as described in the following Riccati equation solution of the
H∞ problem.

Theorem 7.12 Suppose that

(A, B1 ) is controllable and (A, C1 ) is observable. (7.37)

There exist AK , BK , CK , DK that solves the output-feedback H∞ problem (7.13) if and

only if the Riccati equations

AT X + XA + XB1 B1T X + C1T C1 − C2T C2 = 0, (7.38)

AY + Y AT + Y C1T C1 Y + B1 B1T − B2 B2T = 0,

have anti-stabilizing solutions X+ , Y+ , and these solutions satisfy the coupling condition
 
X I
 +  > 0. (7.39)
I Y+

With any non-singular U and V satisfying XY + U V T = I, the formulas (7.36) define a

controller that satisfies (7.13).

Note that the conditions in this result are algebraically verifiable without relying on linear
matrix inequalities. One can test for the existence of anti-stabilizing solutions by verifying
whether the corresponding Hamiltonian matrix has no eigenvalues on the imaginary axis.
If the test is passed, one can compute these anti-stabilizing solutions. Then one simply
needs to check (7.39) what amounts to verifying whether the smallest eigenvalue of the
matrix on the left-hand side is positive.

Proof. Let AK , BK , CK , DK satisfy (7.13). By Theorem 7.11, there exist X, Y with

(7.33)-(7.35). If we apply Theorem 7.1 twice, we infer that the AREs (7.38)-(7.12) have
anti-stabilizing solutions X+ , Y+ with X ≤ X+ , Y ≤ Y+ . The last two inequalities clearly
imply (7.39).

177
Let us now suppose that X+ , Y+ exist and satisfy (7.39), and construct the controller
in the same fashion as described in Theorem 7.11. As easily seen, the matrix X we
constructed above will satisfy

X > 0, AT X + X A + X BB T X + C T C = 0. (7.40)

Let us prove that A is stable. For that purpose we assume that Ax = λx with x 6= 0. If
we left-multiply the ARE in (7.40) with x∗ and right-multiply with x, we arrive at

Re(λ)x∗ X x + x∗ X BB T X x + x∗ C T Cx = 0.

Since x∗ X x > 0, we infer Re(λ) ≤ 0. Let us show that Re(λ) = 0 cannot occur. In fact,
if Re(λ) = 0, we conclude
x∗ X B = 0 and Cx = 0.
Right-multiplying the ARE in (7.40) leads (with −λ = λ) to

x∗ X A = λx∗ X and, still, Ax = λx

b = X x, and partition x
Set x b and x according to A. Then we arrive at
   
  A B C
2 K
    B 1
yb∗ zb∗   = λ yb∗ zb∗ and yb∗ zb∗  =0
BK C2 AK BK D21

as well as
      
A B2 CK y y   y
    = λ   and C1 D12 CK   = 0.
BK C2 AK z z z

Right-multiplying with B1 and left-multiplying with C1T imply that zb∗ BK = 0 and CK z =
0. This reveals
yb∗ A = λb
y ∗ , yb∗ B1 = 0 and Ay = λy, C1 y = 0.
By controllability of (A, B1 ) and observability of (A, C1 ), we conclude yb = 0 and y = 0.
Finally,          
0 X U 0 0 Y V 0
  =  +    and   =  +  
T T
zb U ∗ z z V ∗ zb
lead to 0 = U z and 0 = V zb what implies z = 0 and zb = 0 since U and V are non-singular.
This is a contradiction to x 6= 0.

Since A is stable, the controller is stabilizing. Moreover, it is not difficult to verify that
the Riccati equation in (7.40) implies kC(sI − A)−1 Bk∞ ≤ 1. One can in fact show that
the strict inequality holds true; since this is cumbersome and not really relevant for our
considerations, we stop here.

178
7.8.3 Solution in Terms of Indefinite Riccati Equations

We observe that Y > 0 and

AY + Y AT + Y C1T C1 Y + B1 B1T − B2 B2T = 0, λ(A + Y C1T C1 ) ⊂ C> (7.41)

is equivalent to Y∞ > 0 and

AT Y∞ +Y∞ A+Y∞ (B1 B1T −B2 B2T )Y∞ +C1T C1 = 0, λ(A+(B1 B1T −B2 B2T )Y∞ ) ⊂ C< (7.42)

with
Y∞ = Y −1 .

Proof. Let us show that (7.41) implies (7.42). Left- and right-multiplying the ARE in
(7.41) with Y∞ = Y −1 leads to the ARE in (7.42). Moreover, Y > 0 implies Y∞ = Y −1 > 0.
Finally, the ARE in (7.41) is easily rewritten to

(A + Y C1T C1 )Y + Y (AT + Y −1 [B1 B1T − B2 B2T ]) = 0

what leads to
Y −1 (A + Y C1T C1 )Y = −(A + [B1 B1T − B2 B2T ]Y∞ )T .
Since A + Y C1T C1 is anti-stable, A + [B1 B1T − B2 B2T ]Y∞ must be stable.

The converse implication (7.42) ⇒ (7.41) follows by reversing the arguments.

Dually, we have X > 0 and

AT X + XA + XB1 B1T X + C1T C1 − C2T C2 = 0, λ(A + B1 B1T X) ⊂ C>

if and only if X∞ > 0 and

AX∞ + X∞ AT + X∞ (C1T C1 − C2T C2 )X∞ + B1 B1T = 0, λ(A + X∞ (C1T C1 − C2T C2 )) ⊂ C<

with
X∞ = X −1 .

Finally,  
X I
 >0
I Y
is equivalent to
X∞ > 0, Y∞ > 0, ρ(X∞ Y∞ ) < 1
with X∞ = X −1 , Y∞ = Y −1 .

179
Proof. The coupling condition is equivalent to X > 0, Y > 0, Y − X −1 > 0 (Schur) what
is nothing but X > 0, Y > 0, I − Y −1/2 X −1 Y −1/2 > 0 or, equivalently, X > 0, Y > 0,
ρ(Y −1/2 X −1 Y −1/2 ) < 1 what can be rewritten (since ρ(AB) = ρ(BA)) as X > 0, Y > 0,
ρ(X −1 Y −1 ) < 1.

Hence, all conditions in Theorem 7.12 can be rewritten in terms of so-called indefinite
algebraic Riccati equations. They are called indefinite since the matrices defining the
quadratic terms are, in general, not positive or negative semi-definite.

For the particular choice of U = Y −1 − X, V = Y , the formulas for the controller can be
rewritten in terms of X∞ = X −1 , Y∞ = Y −1 as follows:

AK = A − (I − X∞ Y∞ )−1 X∞ C2T C2 + (B1 B1T − B2 B2T )Y∞ ,

BK = (I − X∞ Y∞ )−1 X∞ C2T , CK = −B2T Y∞ .

Proof. If we left-multiply (7.41) by Y −1 , we obtain

AT + C1T C1 Y = −[Y −1 AY + Y −1 (B1 B1T − B2 B2T )].

Hence the formula for AK in (7.36) can be rewritten to

AK = −U −1 [(X − Y −1 )AY − C2T C2 Y + (X − Y −1 )(B1 B1T − B2 B2T )]V −T

what is nothing but

AK = A + (Y −1 − X)−1 C2T C2 + (B1 B1T − B2 B2T )Y −1 .

or
AK = A − (I − X −1 Y −1 )−1 X −1 C2T C2 + (B1 B1T − B2 B2T )Y −1 .
The formulas for BK and CK are obvious.

Why are the results in the literature usually formulated in terms of these indefinite AREs?
The simple reason is the possibility to relax the artificial and strong hypotheses that
(A, B1 ) and (A, C1 ) are controllable and observable to a condition on the non-existence
of uncontrollable or unobservable modes on the imaginary axis. The exact formulation
with a general bound γ and with an explicit controller formula is as follows.

Theorem 7.13 Suppose that

(A, B1 ), (A, C1 ) have no uncontrollable, unobservable modes in C= . (7.43)

180
Then there exist AK , BK , CK , DK that solves the output-feedback H∞ problem (7.13) if
and only if the unique X∞ and Y∞ with
1 T
AX∞ + X∞ AT + X∞ ( C C1 − C2T C2 )X∞ + B1 B1T = 0,
γ2 1
1
λ(A + X∞ ( 2 C1T C1 − C2T C2 )) ⊂ C<
γ
and
1
A T Y∞ + Y∞ A + Y∞ ( B1 B1T − B2 B2T )Y∞ + C1T C1 = 0,
γ2
1
λ(A + ( B1 B1T − B2 B2T )Y∞ ) ⊂ C<
γ2
exist, and they satisfy

X∞ ≥ 0, Y∞ ≥ 0, ρ(X∞ Y∞ ) < γ 2 .

If X∞ and Y∞ satisfy all these conditions, a controller with (7.13) is given by

   
  T 1 T T T
A B  A − ZX∞ C2 C2 + γ 2 B1 B1 − B2 B2 Y∞ ZX∞ C2 
 K K= 
CK DK
 
T
−B2 Y∞ 0

1
where Z = (I − X∞ Y∞ )−1 .
γ2

Remark 7.14

1) The constructed controller in Theorem 7.13 is also called central controller. This
terminology will be explained in Section 7.11.

2) X∞ and Y∞ are computed as for standard Riccati equations on the basis of the
Hamiltonian matrices
   
T 1 T T 1 T T
A C C − C2 C2
γ2 1 1
A B B − B2 B2
γ2 1 1
HX∞ =   and HY∞ =  
T T T
−B1 B1 −A −C1 C1 −A

as follows: Verify that HX∞ and HY∞ do not have

 eigenvalues
  onthe imaginary axis.
UX∞ U Y∞
Then compute (with Schur decompositions)  ,   and stable MX∞ ,
VX∞ V Y∞
MY∞ satisfying
       
UX∞ U U U
HX∞   =  X∞  MX∞ and HY∞  Y∞  =  Y∞  MY∞ .
VX∞ VX∞ VY∞ VY∞

181
 Verify that UX∞ and UY∞ are non-singular. Then

X∞ = VX∞ UX−1∞ and Y∞ = VY∞ UY−1

∞

exist and are the stabilizing solutions of the two indefinite AREs under considera-
tions.
After having verified that (the unique) X∞ and Y∞ exist, it remains to check whether
they are both positive semi-definite, and whether the spectral radius of X∞ Y∞ is
smaller than γ 2 .

2) Note that X∞ and Y∞ are, in general, not invertible. That’s why we insisted on
deriving formulas in which no inverse of X∞ or Y∞ occurs. If X∞ and Y∞ exist, one
can show:
X∞ has no kernel if and only if (A, B1 ) has no uncontrollable modes in the open
left-half plane, and Y∞ has no kernel if and only if (A, C1 ) has no unobservable
modes in the open left-half plane.

3) The optimal value

γ∗ = inf kS(P, K)k∞
K stabilizies P

is the smallest of all γ for which the stabilizing solutions X∞ , Y∞ of the indefinite
AREs exist and satisfy X∞ ≥ 0, Y∞ ≥ 0, ρ(X∞ Y∞ ) < γ 2 . The optimal value γ∗ can
hence be computed by bisection.

4) If γ ≤ γ∗ , it cannot be said a priori which of the conditions (existence of X∞ , Y∞ ,

positive semi-definiteness, coupling condition) fails.
For γ > γ∗ , but γ close to γ∗ , it often happens that
1
I− X∞ Y∞ is close to singular.
γ2
Hence computing the inverse of this matrix is ill-conditioned. This leads to an ill-
conditioned computation of the controller matrices as given in the theorem. There-
fore, it is advisable not to get too close to the optimum.
Note, however, that this is only due to the specific choice of the controller formulas.
Under our hypothesis, one can show that there always exists an optimal controller.
Hence there is a possibility, even γ = γ∗ , to compute an optimal controller a well-
conditioned fashion.

5) Let us consider the other extreme γ = ∞. Then one can always find X∞ ≥ 0 and
Y∞ ≥ 0 that satisfy

AX∞ + X∞ AT − X∞ C2T C2 X∞ + B1 B1T = 0, λ(A − X∞ C2T C2 ) ⊂ C<

182
 and

AT Y∞ + Y∞ A − Y∞ B2 B2T Y∞ + C1T C1 = 0, λ(A − B2 B2T Y∞ ) ⊂ C< .

(Why?) Moreover, the controller formulas read as

  
A − X∞ C2T C2 − B2 B2T Y∞ X∞ C2T

A B
 K K=  .

CK DK
−B2T Y∞ 0

Clearly, this controller stabilizes P . In addition, however, it has even the additional
property that it minimizes

kS(P, K)k2 among all controllers K which stabilize P .

Here, kM k2 is the so-called H2 -norm of the strictly proper and stable matrix M
which is defined via
Z ∞
1
2
kM k2 = trace(M (iω)∗ M (iω)) dω.
2π −∞

Hence the controller is a solution to the so-called H2 -control problem. Since the H2 -
norm can be seen to be identical to the criterion in LQG-control, we have recovered
the controller that solves the LQG problem as it is taught in an elementary course.
All this will be discussed in more detail and generality in the LMI course.

7.9 What are the Weakest Hypotheses for the Riccati Solution?

The command hinfys of the µ-tools to design an H∞ -controller requires the following
hypotheses for the system
 
  A B1 B2  
z   w
  =  C1 D11 D12   
y u
 
C2 D21 D22

describing the interconnection:

• (A, B2 ) is stabilizable, and (A, C2 ) is detectable.

• D21 has full row rank, and D12 has full column rank.
   
A − iωI B1 A − iωI B2
• For all ω ∈ R,   has full row rank, and   has full
C2 D21 C1 D12
column rank.

183
If not true, the second and the third hypotheses can be easily enforced as follows: With
some  > 0, solve the problem for the perturbed matrices
   
C 1 D 12
    D 11 0 0
  B B I 0
,  1  →  1
   
C1 D12 →  I 0  , D11 →  0 0 0  .
D21 D21 0 I
   
0 I 0 0 0

Let us denote the resulting generalized plant by

 
    A B1 I 0 B2  
z w   w
     C1 D11 0 0 D12 
 
z  w    w1 
 1  1 
  = P   =  I 0 0 0 0  .
 
 z2   w2   
   w 2

    0 0 0 0 I   
y u   u
C2 D21 0 I D22

The perturbation just amounts to introducing new disturbance signals w1 , w2 and new
controlled signals z1 , z2 in order to render all the hypotheses for  6= 0 satisfied.

Here are the precise conclusions that can be drawn for the relation of the H∞ problem
for the original interconnection P and for the perturbed interconnection P .

• K stabilizes P if and only if K stabilizes P . (Why?)

• For any K which stabilizes P and P , the gain-interpretation of the H∞ immediately

reveals that
kS(P, K)k∞ ≤ kS(P , K)k∞ .
Hence, if the controller K stabilizes P and achieves kS(P , K)k∞ < γ, then the very
same controller also stabilizes P and achieves kS(P, K)k∞ < γ. This property
does not depend on the size of  > 0!

• For any K stabilizing P and P , one has

kS(P , K)k∞ → kS(P, K)k∞ for  → 0.

Hence, if there exists a K stabilizing P and rendering kS(P, K)k∞ < γ satisfied,
the very same K stabilizes P and achieves kS(P , K)k∞ < γ for some sufficiently
small  > 0.

Note that there are other schemes to perturb the matrices in order to render the hypothe-
ses satisfied. In general, however, it might not be guaranteed that the first of these two
properties is true irrespective of the size of  > 0.

184
7.10 Game-Theoretic Interpretation

This more mathematically orientated chapter gives an extra interpretation of the H∞

control problem in terms of game theory. The main result will be that optimal H∞ control
is the optimal strategy for a player who plays control stragies and tries to minimize a cost
function. Optimal in the following sense: no matter what strategy its opponent plays,
its pay-off is bounded from above by an optimal value and a deviation from the optimal
strategy will leave its opponent the opportunity to drive the cost above the optimal value.

Let us now become more precise: We consider again the system

ẋ = Ax + B1 w + B2 u, x(0) = ξ
(SY)
z = C1 x + D12 u
as in the state-feedback case under the following assumptions:

Hypothesis 7.15 Suppose that

(A, B2 ) is stabilizable. (SF1)
   
T
D12 C1 D12 = 0 I . (SF2)
(A, C1 ) is observable. (SF3)
Moreover let Y∞ ≥ 0 satisfy as in Theorem 7.13
AT Y∞ + Y∞ A + Y∞ (γ −2 B1 B1T − B2 B2T )Y∞ + C1T C1 = 0, (7.44)
eig(A + (γ −2 B1 B1T − B2 B2T )Y∞ ) ⊂ C− (7.45)
with γ > 0.

Note that observability of (A, C1 ) implies Y∞ > 0. In linear quadratic game-theory, one
considers the control u and the generalized disturbance w to be manipulated by adversary
“players”. Here the u-player tries to minimize and the w-player tries to maximize the
quadratic cost (or pay-off)
Z ∞
kz(t)k2 − γ 2 kw(t)k2 dt. (CO)
0

Our intention is to touch upon a specific consequence from the existence of Y∞ - this is a
tip of a whole iceberg of game-theory.

The first result is the following fact.

Lemma 7.16 Suppose that Hypothesis 7.15 holds. If controlling (SY ) with ξ = 0 as
u = −B2T Y∞ x then Z ∞
kz(t)k2 − γ 2 kw(t)k2 dt ≤ 0 (7.46)
0
for all w ∈ L2 .

185
Proof. Since Y∞ > 0 we can define Y+ := Y∞−1 . Then Y+ is the anti-stabilizing solution
of the ARE
AY+ + Y+ A + γ −2 B1 B1T − B2 B2T + Y+ C1T C1 Y+ = 0
since
AT + C1T C1 Y+ = Y∞ (Y+ AT + Y+ C1T C1 Y+ )
= Y∞ (−A− γ −2 B1 B1T Y∞ + B2 B2T Y∞ )Y∞−1
= −Y∞ (A + (γ −2 B1 B1T − B2 B2T )Y∞ )Y∞−1 .
By Theorem 7.9 we can conclude that F = −B2T Y∞ satisfies

eig(A + B2 F ) ⊂ C− and k(C1 + D12 F )(sI − A − B2 F )−1 B1 k < γ.

Let G(s) be the transfermatrix of the System (SY) controlled by u = F x. Then G(s) =
(C1 + D12 F )(sI − A − B2 F )−1 B1 and by using Parceval’s identity
Z ∞ Z ∞
1
2
kf (t)k dt = kfˆ(iω)k2 dω for all f ∈ L2
0 2π −∞

we obtain Z ∞ Z ∞
2 1
kz(t)k dt = kẑ(iω)k2 dω
0 2π −∞
Z ∞
1
= kG(iω)ŵ(iω)k2 dω
2π −∞
Z ∞
2 1
≤ kGk∞ kŵ(iω)k2 dω
2π −∞
Z ∞
2
≤γ kw(t)k2 dt
0
for all w ∈ L2 .

From a game-theoretic point of view Lemma 7.16 means that if the u-player plays the
static-feedback strategy u = −B2T Y∞ x for the System (SY) with ξ = 0, the cost (CO)
is guaranteed to be bounded by 0 for any finite energy open-loop strategy the w-player
uses.

As a next step we want to generalize this result for arbitrary initial condition ξ.

Lemma 7.17 If Y∞ satisfies only (7.44) then along any trajectory of (SY) with (SF2)
holds: Z T
T T
x(T ) Y∞ x(T ) − ξ Y∞ ξ + kz(t)k2 − γ 2 kw(t)k2 dt
0
Z T Z T
= T 2
ku(t) + B2 Y∞ x(t)k − γ 2
kw(t) − γ −2 B1T Y∞ x(t)k2 dt.
0 0

186
Proof. (7.44) implies
d
x(t)T Y∞ x(t) = ẋ(t)T Y∞ x(t) + x(t)T Y∞ ẋ(t)
dt
= x(t)T AT Y∞ + Y∞ A x(t) + 2x(t)T Y∞ B1 w(t) + 2x(t)T Y∞ B2 u(t)
 

= − x(t)T C1T C1 x(t) − γ −2 x(t)T Y∞ B1 B1T Y∞ x(t) + x(t)T Y∞ B2 B2T Y∞ x(t)

+ 2x(t)T Y∞ B1 w(t) + 2x(t)T Y∞ B2 u(t)
= − x(t)T C1T C1 x(t) + γ 2 w(t)T w(t) − u(t)T u(t)
− γ 2 kw(t) − γ −2 B1T Y∞ x(t)k2 + ku(t) + B2T Y∞ x(t)k2
= − kz(t)k2 + γ 2 kw(t)k2 + ku(t) + B2T Y∞ x(t)k2
− γ 2 kw(t)γ −2 B1T Y∞ x(t)k2 .
Integration over [0, T ] proves the result.

Lemma 7.18 Suppose that Hypothesis 7.15 holds and suppose (SY) is controlled by u =
−B2T Y∞ x. Then we have for all w ∈ L2 :
Z ∞
kz(t)k2 − γ 2 kw(t)k2 dt ≤ ξ T Y∞ ξ.
0

Equality holds for the feedback strategy w = γ −2 B1T Y∞ x.

Proof. As in the proof of Lemma 7.16 it follows that eig(A − B2 B2T Y∞ ) ∈ C− . Now
choose w ∈ L2 arbitrary. Then any trajectory of the controlled system solves ẋ(t) =
(A − B2 B2T Y∞ )x(t) + B1 w(t) with w ∈ L2 . By using Young’s inequality for convolutions
we can conclude x ∈ L2 . Hence ẋ ∈ L2 and thus limT →∞ x(T ) = 0. Now Lemma 7.17
implies
Z ∞ Z ∞
T
ξ Y∞ ξ − 2 2 2
kz(t)k − γ kw(t)k dt = γ 2
kw(t) − γ −2 B1T Y∞ x(t)k2 dt ≥ 0.
0 0

Equality holds for w = γ −2 B1T Y∞ x. If w = γ −2 B1T Y∞ x any trajectory of the controlled

system solves ẋ(t) = (A − B2 B2T Y∞ + γ −2 B1 B1T Y∞ )x(t). Due to (7.45) we can conclude
x ∈ L2 and hence w ∈ L2 .

If the u-player plays the static state-feedback strategy u = −B2T Y∞ x, the incurred cost is
guaranteed to be bounded by ξ T Y∞ ξ, no matter which finite energy open-loop strategy
the w-player uses. The adversary w-player has the possibility to maximize its pay-off by
using the feedback strategy w = γ −2 B1T Y∞ x. For a particular ξ this translates into a finite
energy open-loop disturbance, which is the worst case disturbance from the perspective
of the u-player. The resulting cost equals ξ T Y∞ ξ.

Reversing the roles of w and u leads to an analogous insight.

187
Lemma 7.19 Suppose that Hypothesis 7.15 holds and let (SY) be affected by the worst-
case disturbance strategy w = γ −2 B1T Y∞ x. For any u ∈ L2 such that x ∈ L2 we have
Z ∞
T
ξ Y∞ ξ ≤ kz(t)k2 − γ 2 kw(t)k2 dt,
0

with equality for the feedback control u = −B2T Y∞ x.

If the w-player plays w = −γ 2 B1T Y∞ x, its guaranteed pay-off is ξ T Y∞ ξ for all finite energy
open-loop strategies of u that take care of stabilizing the state. The u-player can react
with u = −B2T Y∞ x to push its cost down to the lowest possible level ξ T Y∞ ξ. The order of
the decision-taking of the players is important in these conclusions. Note that the roles
of the players are not symmetric, since u has to take care of stabilization.

Let us hence only consider static-state feedback strategies for player u.

Theorem 7.20 Suppose that Hypothesis 7.15 holds. Then the following equation holds:
Z ∞
min −
sup kz(t)k2 − γ 2 kw(t)k2 dt = ξ T Y∞ ξ.
u=F x, eig(A+B2 F )⊂C w∈L2 0

If u plays the stabilizing strategy u = −F x we infer

Z ∞
sup kz(t)k2 − γ 2 kw(t)k2 dt ≥ ξ T Y∞ ξ,
w∈L2 0

with equality (as we already know) for F∞ = −B2T Y∞ . We conclude that it is rational
for the u-player to stick to u = F∞ x, since a deviation is guaranteed to be not beneficial
- it even does offer the opportunity for player w to drive the cost for u to values above
ξ T Y∞ ξ.

One can view this result as a time-domain interpretation of H∞ -control.

In order to proof this result, we need some preparations.

Lemma 7.21 Suppose that eig(A) ⊂ C+ . For every w ∈ L2 there exists a unique ξ, such
that the solution of ẋ = Ax + Bw, x(0) = ξ is square integrable.

Proof. Choose w ∈ L2 arbitrary. Then for any system trajectory we know that
Z t
−At
e x(t) = ξ + e−Aτ Bw(τ )dτ for t ≥ 0. (7.47)
0

Since −A is Hurwitz, the function τ 7→ e−Aτ Bw(τ ) is integrable on [0, ∞) by the Hölder
inequality.

188
If x ∈ L2 then ẋ ∈ L2 , since w ∈ L2 by assumption. Thus we obtain limt→∞ x(t) = 0.
R∞
From (7.47) we infer ξ = − 0 e−Aτ Bw(τ )dτ which proves uniqueness.
The solution for this initial condition is given by
Z ∞ Z t Z ∞
At −Aτ A(t−τ )
x(t) = −e e Bw(τ )dτ + e Bw(τ )dτ = − eA(t−τ ) Bw(τ )dτ.
0 0 t

Using again Young’s inequality for convolutions leads to x ∈ L2 and hence existence.

Lemma 7.22 Suppose that Hypothesis 7.15 holds and let Y be the stabilizing solution of
the standard ARE: AT Y +Y A−Y B2 B2T Y +C1T C1 = 0. Fix w ∈ L2 and choose the unique
response v ∈ L2 of v̇ = −(A − B2 B2T Y )T v − Y B1 w. If the control function u ∈ L2 for
(SY) assures x ∈ L2 then
Z ∞ Z ∞
2 T T
kz(t)k dt ≥ ξ Y ξ + 2v(0) ξ − v(t)T B2 B2T v(t) − 2v(t)T B1 w(t)dt. (7.48)
0 0

Equality holds for u(t, x) = −B2T Y x − B2T v(t).

The right hand side of the inequality only depends on ξ and on w(·), but not on the
control function u(·) or the resulting system trajectory of (SY).
This solves an LQ problem for a system driven by L2 -disturbances w(·). Therefore observe
that (7.48) for γ > 0 can be written as
T 
Z∞ Z∞
  
v(t) B2 B2T −B1 v(t)
kz(t)k2 − γ 2 kw(t)k2 dt ≥ ξ T Y ξ + 2v(0)T ξ −      dt.
w(t) −B1T γ 2 I w(t)
0 0

The optimal strategy requires full knowledge of w(·) and is non-causal. This just means
that at time t, the control action cannot be determined based on w|[0,t] only. This is true
since by the proof of Lemma 7.21 v(·) is given by
Z ∞
T T
v(t) = − e−(A−B2 B2 Y ) (t−τ ) Y B1 w(τ )dτ for t ≥ 0.
t

Hence knowledge of w(·) on [t, ∞) is required to calculate v(t) for a fixed t ≥ 0. Finally
note that under our assumptions there always exists a stabilitzing solution Y of the ARE:
AT Y + Y A − Y B2 B2T Y + C1T C1 = 0.

Proof. As in Lemma 7.17 the proof follows by completion-of-the-squares and by using

189
the fact that Y solves the given ARE.
d T
(x Y x + 2v T x) = ẋT Y x + xT Y ẋ + 2v̇ T x + 2v T ẋ
dt
= xT AT Y + Y A x + 2xT Y B1 w + 2xT Y B2 u
 

− 2v T (A − B2 B2T Y )x − 2wT B1T Y x + 2v T (Ax + B1 w + B2 u)

= − xT C1T C1 x + xT Y B2 B2T Y x + 2v T B2 B2T Y x + v T B2 B2T v
− v T B2 B2T v + 2uT B2T (Y x + v) + 2v T B1 w
= − kC1 xk2 − kuk2 + uT u + (Y x + v)T B2 B2T (Y x + v)
+ 2uT B2T (Y x + v) − v T B2 B2T v + 2v T B1 w
= − kzk2 + ku + B2T (Y x + v)k2 − v T B2 B2T v + 2v T B1 w.

Since x(T ) → 0, v(T ) → 0 for T → ∞ the statement follows by integration over [0, T ]
and taking the limit T → ∞.

Observe again that (7.48) can be written as

Z ∞
kz(t)k2 −γ 2 kw(t)k2 dt ≥ ξ T Y ξ + 2v(0)T
0
Z ∞
− v(t)T B2 B2T v(t) − 2v(t)T B1 w(t) + γ 2 w(t)T w(t)dt.
0

In the next step we intend to maximize the right-hand side over all pairs (w, v) as
in Lemma 7.22. For given w ∈ L2 by the proof of Lemma 7.21 we know that
R∞ T T
v(0) = 0 e(A−B2 B2 Y ) τ Y B1 w(τ )dτ . Thus the set of all v(0) for such pairs is equal
to the controllable subspace of (−(A − B2 B2T Y )T , −Y B1 ). Hence these pairs can as well
be parameterized by taking any v0 in the controllable subspace and choosing w ∈ L2
which assures v ∈ L2 for

v̇ = −(A − B2 B2T Y )T v − Y B1 w, v(0) = v0 . (AUX)

For notational simplicity let us assume that (AUX) is controllable, by Hautus this means
that
(A − B2 B2T Y, Y B1 B1T Y ) is observable. (OBS)
Let us first fix v0 . Then we have to solve the problem of minimizing
Z ∞
v(t)T B2 B2T v(t) − 2v(t)T B1 w(t) + γ 2 w(t)T w(t)dt
0
 T   
Z ∞ T
v(t) B B −B1 v(t)
=    2 2   dt
0 w(t) −B1T γ 2 I w(t)

over all w ∈ L2 with v ∈ L2 for solutions of (AUX).

190
This is actually a standard LQ problem with stability, but the cost function involves
cross-terms. This is the reasons why we provide a direct solution based on the following
ARE:

−(A − B2 B2T Y )Z − Z(A − B2 B2T Y )T + B2 B2T − γ −2 (I + ZY )B1 B1T (I + Y Z) = 0.

Lemma 7.23 Let Z be the anti-stabilizing solution of the ARE

(A − B2 B2T Y + γ −2 B1 B1T Y )Z + Z(A − B2 B2T Y + γ −2 B1 B1T Y )T

+γ −2 ZY B1 B1T Y Z + γ −2 B1 B1T − B2 B2T = 0.

If w ∈ L2 achieves v ∈ L2 for (AUX), the following inequality holds:

Z ∞
v(t)T B2 B2T v(t) − 2v(t)T B1 w(t) + γ 2 w(t)T w(t)dt ≥ v0T Zv0 .
0

Equality is achieved with w = γ −2 B1T (I + Y Z)v.

Proof. As before we use the fact that Z solves the ARE and complete the squares:
d T
v Zv = v̇ T Zv + v T Z v̇
dt
= v T [−(A − B2 B2T Y )Z − Z(A − B2 B2T Y )T ]v − 2v T ZY B1 w
= v T [−B2 B2T + γ −2 (I + ZY )B1 B1T (I + Y Z)]v − 2v T ZY B1 w
= − v T B2 B2T v + 2v T B1 w − γ 2 wT w − 2v T (I + ZY )B1 w
+ γ 2 wT w + γ −2 v T (I + ZY )B1 B1T (I + Y Z)v
= − v T B2 B2T v + 2v T B1 w − γ 2 wT w + γ 2 kw − γ −2 B2T (I + Y Z)vk2 .

Integrating over [0, T ] and taking the limit T → ∞ leads to

Z ∞
v(t)T B2 B2T v(t) − 2v(t)T B1 w(t)+γ 2 w(t)T w(t)dt
0
Z ∞
T
= v(0) Zv(0) + γ 2 kw − γ −2 B1T (I + Y Z)vk2 dt
0

since limT →∞ v(T ) = 0 by assumption. Hence we obtain the desired inequality.

Equality is enforced with w = γ −2 B1T (I + Y Z)v which results in

v̇ = −[(A − B2 B2T Y + γ −2 B1 B1T Y )T + γ −2 Y B1 B1T Y Z]v, v(0) = v0 .

Since Z is the anti-stabilizing solution of the given ARE we get eig(−(A − B2 B2T Y +
γ −2 B1 B1T Y )T −γ −2 Y B1 B1T Y Z) ⊂ C− and hence v ∈ L2 . This also implies w = γ −2 B1T (I +
Y Z)v ∈ L2 .

Since the following fact is very useful and needed in the following proofs, we formulate it
as a lemma.

191
Lemma 7.24 Let R = RT , Q1 = QT1 , Q2 = QT2 and let X1 and X2 be symmetric solutions
of
AT X1 + X1 A + X1 RX1 + Q1 = 0
AT X2 + X2 A + X2 RX2 + Q2 = 0.
Then ∆ := X2 − X1 solves the ARE

(A + RX1 )T ∆ + ∆(A + RX1 ) + ∆R∆ + Q2 − Q1 = 0

Proof. Direct calculation leads to

Q1 − Q2 = (AT X2 + X2 A + X2 RX2 ) − (AT X1 + X1 A + X1 RX1 )

= AT ∆ + ∆A + X2 RX2 − X1 RX1 + X1 RX2 − X1 RX2
= AT ∆ + ∆A + ∆RX2 + X1 R∆ + ∆RX1 − ∆RX1
= AT ∆ + ∆A + X1 R∆ + ∆RX1 + ∆R∆
= (A + RX1 )T ∆ + ∆(A + RX1 ) + ∆R∆.

The next Lemma shows a relation between the appearing ARE’s. The proof does not
need the assumptions (SF1) - (SF3).

Lemma 7.25 Let Y be as in Lemma 7.22 and let Y∞ ≥ 0 satisfy (7.44)-(7.45). If (OBS)
holds then Z in Lemma 7.23 exists, is positive definite and equals Z = (Y∞ − Y )−1 .

Proof. Set à := (A + γ −2 B1 B1T Y − B2 B2T Y ) and set ∆ := Y∞ − Y . Applying Lemma

7.24 to (7.44) and

AT Y + Y A + Y (γ −2 B1 B1T − B2 B2T )Y + C1T C1 − γ −2 Y B1 B1T Y = 0

leads to the equation

ÃT ∆ + ∆à + ∆(γ −2 B1 B1T − B2 B2T )∆ + γ −2 Y B1 B1T Y = 0.

Let x ∈ ker(∆). Then γ −2 kB1T Y xk2 = 0 and hence 0 = ∆Ãx = ∆(A − B2 B2T Y )x. This
means (A − B2 B2T Y ) ker(∆) ⊂ ker(∆) and ker(∆) ⊂ ker(Y B1 B1T Y ). By (OBS) we can
conclude ker(∆) = {0}.
Set Q := γ −2 Y∞ B1 B1T Y∞ + C1T C1 . Then observe that

0 = AT Y∞ + Y∞ A − Y∞ B2 B2T Y∞ + Q (7.49)

and
AT Y + Y A − Y B2 B2T Y + Q = γ −2 Y∞ B1 B1T Y∞ ≥ 0.

192
Choose x 6= 0 with (A − B2 B2T Y∞ )x = λx. Then we obtain by using the ARE for Y∞ :

0 = 2Re(λ)x∗ Y∞ x + x∗ Y∞ B2 BT
2 Y∞ x + γ
−2 ∗
x Y∞ B1 BT ∗ T
1 Y∞ x + x C1 C1 x

= 2Re(λ)x∗ Y∞ x + kBT 2
2 Y∞ xk + γ
−2
kBT 2
1 Y∞ xk + kC1 xk
2

Since Y∞ > 0 and kB2T Y∞ xk2 + γ −2 kB1T Y∞ xk2 + kC1 xk2 ≥ 0 we obtain Re(λ) ≤ 0. Hence
we obtain eig(A − B2 B2T Y∞ ) ⊂ C− ∪ C0 (i.e. Y∞ is the strong solution of the ARE (7.49)).
By Exercise 6 we can conclude Y ≤ Y∞ and hence ∆ ≥ 0.
Now observe that Z = ∆−1 > 0 satisfies the ARE in Lemma 7.23 and

− ∆−1 [Ã + γ −2 ZY B1 B1T Y ]T ∆ = Ã + (γ −2 B1 B1T − B2 B2T )∆ = A + (γ −2 B1 B1T − B2 B2T )Y∞ .

(7.50)
−2 T T −2 T
Since A + (γ B1 B1 − B2 B2 )Y∞ is stable, Ã + γ ZY B1 B1 Y is anti-stable.

We are now able to prove Theorem 7.20 under the additional assumption (OBS). If this is
not true, one argues by reducing the dynamics (AUX) to the controllable subspace which
is only notationally a bit more cumbersome.

Proof of Theorem 7.20. Take F with eig(A+BF ) ⊂ C− and control (SY) with u = F x.
Due to Young’s inequality for convolutions, for w ∈ L2 the controlled system responds
with x ∈ L2 and we can define
Z ∞
γwc (F ) := sup kz(t)k2 − γ 2 kw(t)k2 dt ≤ ∞.
w∈L2 0

For any w ∈ L2 and v ∈ L2 as in Lemma 7.22, we get

Z ∞
γwc (F ) ≥ kz(t)k2 − γ 2 kw(t)k2 dt
0
Z ∞
T T
≥ ξ Y ξ + 2v(0) ξ − v(t)T B2 B2T v(t) − 2v(t)T B1 w(t) + γ 2 w(t)T w(t)dt.
0

Let us now maximize the right-hand side over all such pairs (w, v). As remarked earlier,
we can instead maximize over w ∈ L2 for fixed v(0) = v0 in the controllable subspace of
(−(A − B2 B2T Y )T , −Y B1 ) such that v ∈ L2 .
By Lemma 7.23 and Lemma 7.25 we obtain

γwc (F ) ≥ ξ T Y ξ + 2v0T ξ − v0T Zv0 .

As a next step we maximize over v0 . Therefore observe that there exists v∗ with Zv∗ = ξ
and that Z has a positive definite square root. Hence

γwc (F ) ≥ ξ T Y ξ + 2v0T ξ − v0T Zv0

= ξ T Y ξ + 2v0T ξ − v0T Zv0 + ξ T Z −1 ξ − ξ T Z −1 ξ
1 1
= ξ T Y ξ + ξ T Z −1 ξ − kZ 2 v0 − Z − 2 ξk2 .

193
This shows that the maximum is achieved for v0 = Z −1 ξ = v∗ and the maximum value of
the right hand side is given by ξ T Y ξ + ξ T Z −1 ξ = ξ T Y∞ ξ by Lemma 7.25. This proves

min γwc (F ) ≥ ξ T Y∞ ξ.
u=F x, eig(A+B2 F )⊂C−

On the other hand, we know for F∞ = −B2T Y∞ by Lemma 7.18 that

min γwc (F ) ≤ γwc (F∞ ) ≤ ξ T Y∞ ξ.

u=F x, eig(A+B2 F )⊂C−

This proves Theorem 7.20.

Let us now consider the situation that w plays with open-loop strategies w ∈ L2 , and that
u responds with stabilizing open-loop strategies u ∈ L2 . The resulting pay-off is again
given by the value ξ T Y∞ ξ.

Theorem 7.26 Suppose that Hypothesis 7.15 holds. Then the following equation holds:
Z ∞
max min kz(t)k2 − γ 2 kw(t)k2 dt = ξ T Y∞ ξ.
w∈L2 u∈L2 such that x∈L2 0

If the disturbance w ∈ L2 acts on the system, the best choice for u leads to a cost not
larger than ξ T Y∞ ξ if u plays open-loop and takes care of stability. Moreover, the w-player
can push the cost for u up to the level ξ T Y∞ ξ by open-loop strategies w ∈ L2

Proof. For w ∈ L2 and u ∈ L2 such that x ∈ L2 define

Z ∞
γ(w, u) := kz(t)k2 − γkw(t)k2 dt.
0

Let w ∈ L2 be arbitrary. With v ∈ L2 satisfying v̇ = −(A − B2 B2T Y )T v − Y B1 w and for

the trajectory of ẋ = (A − B2 B2T Y )x + B1 w − B2 B2T v, x(0) = ξ we infer from the equality
in Lemma 7.22 and the inequality Lemma 7.23

γ(w, −B2T Y x − B2T v)

Z ∞
T T
= ξ Y ξ + 2v(0) ξ − v(t)T B2 B2T v(t) − 2v(t)T B1 w(t) + γ 2 w(t)T w(t)dt
0
≤ ξ Y ξ + 2v(0) ξ − v(0)T Zv(0)
T T

= ξ T Y ξ + ξ T Z −1 ξ − kZ 1/2 v(0) − Z −1/2 ξk2

≤ ξ T Y∞ ξ.

Hence we obtain
max min γ(w, u) ≤ ξ T Y∞ ξ.
w∈L2 u∈L2 s.t. x∈L2

194
On the other hand choose v(0) = v∗ such that v∗ = Z −1 ξ and choose w = γ −2 B1T (I +Y Z)v.
Then using the inequality in Lemma 7.22 and the equality in Lemma 7.23 leads to

max min γ(w, u)

w∈L2 u∈L2 s.t. x∈L2

≥ min γ(γ −2 B1T (I + Y Z)v, u)

u∈L2 s.t. x∈L2
Z ∞
T T
≥ ξ Y ξ + 2v∗ ξ − v(t)T B2 B2T v(t) − 2v(t)T B1 w(t) + γ 2 w(t)T w(t)dt
0
T
=ξ Yξ+ 2v∗T ξ + T
v∗ Zv∗ = ξ T Y∞ ξ.

This proves the claim.

The inner minimization of the u-player can be confined to stabilizing static state-feedback
strategies without changing the value. This should be surprising in view of the structure
of the optimal u-strategy for a fixed disturbance w ∈ L2 as described in Lemma 7.22!

Theorem 7.27 Suppose that Hypothesis 7.15 holds. Then the following equation holds:
Z ∞
T
ξ Y∞ ξ = min −
sup kz(t)k2 − γ 2 kw(t)k2 dt
u=F x, eig(A+B2 F )⊂C w∈L2 0
Z ∞
= sup inf −
kz(t)k2 − γ 2 kw(t)k2 dt.
w∈L2 u=F x, eig(A+B2 F )⊂C 0

In game-theory, this means that the game with dynamics (SY) and with w playing finite
energy open-loop strategies while u plays stabilizing static state-feedback strategies has
a value, and the value of the game equals ξ T Y∞ ξ.

Proof. The first equation is just Theorem 7.20. In order to shorten notation, define for
w ∈ L2 and u = F x with eig(A + B2 F ) ∈ C−
Z ∞
ζ(w, u) := kz(t)k2 − γkw(t)k2 dt.
0

For w ∈ L2 and u = F x with eig(A + B2 F ) ∈ C− arbitrary we trivially get

inf ζ(w, u) ≤ ζ(w, u).

u=F x, eig(A+B2 F )⊂C−

And hence we obtain for arbitrary u = F x such that eig(A + B2 F ) ⊂ C−

sup inf ζ(w, u) ≤ sup ζ(w, u).

w∈L2 u=F x, eig(A+B2 F )⊂C− w∈L2

With Theorem 7.20 this leads to

sup inf ζ(w, u) ≤ inf sup ζ(w, u) = ξ T Y∞ ξ.

w∈L2 u=F x, eig(A+B2 F )⊂C− u=F x, eig(A+B2 F )⊂C− w∈L2

195
Let w ∈ L2 and choose F with eig(A+B2 F ) ∈ C− . Then the solution of ẋ = (A+B2 F )x+
B1 w, x(0) = ξ is in L2 , due to Young’s inequality for convolutions. Hence u = F x ∈ L2
and Z ∞
inf kz(t)k2 − γ 2 kw(t)k2 dt ≤ inf −
ζ(w, u).
u∈L2 s.t. x∈L2 0 u=F x, eig(A+B2 F )⊂C

Taking the supremum over all w ∈ L2 implies with Theorem 7.26

Z ∞
T
ξ Y∞ ξ ≤ sup inf −
kz(t)k2 − γ 2 kw(t)k2 dt ≤ ξ T Y∞ ξ
w∈L2 u=F x, eig(A+B2 F )⊂C 0

and hence equality.

Note that the optimal strategies in Theorem 7.26 are generated by solving

v̇ = −(A − B2 B2T Y + γ −2 (I + ZY )B1 B1T Y )T v, v(0) = Z −1 ξ,

(7.51)
ẋ = (A − B2 B2T Y )x + (γ −2 B1 B1T (I + Y Z) − B2 B2T )v, x(0) = ξ.

and with u = −B2T Y x − B2T v as well as w = γ −2 B1T (I + Y Z)v.

Moreover we can exploit (7.50) wich reads more explicitly as

−Z[A + γ −2 (I + ZY )B1 B1T Y − B2 B2T Y ]T Z −1 = A + γ −2 B1 B1T Y∞ − B2 B2T Y∞ .

The coordinate change η = Zv in the dynamics of the optimal strategies (7.51) leads then
to
η̇ = (A + γ −2 B1 B1T Y∞ − B2 B2T Y∞ )η, η(0) = ξ
ẋ = Ax − B2 B2T Y (x − η) + (γ −2 B1 B1T Y∞ − B2 B2T Y∞ )η, x(0) = ξ.
Since ẋ − η̇ = (A − B2 B2T Y )(x − η) and x(0) − η(0) = 0, we infer x(t) = η(t) for all t ≥ 0.
Hence the worst-case disturbance and the optimal control are as well generated by static
state-feedback as

w = γ −2 B1T (I + Y Z)v = γ −2 B1T (Z −1 + Y )η = γ −2 B1T Y∞ x,

u = −B2T Y x − B2T v = −B2T Y x − B2T Z −1 η = −B2T Y∞ .

This fact will help in the next section to parameterize all controllers that solve the H∞ -
problem.

7.11 Interpretation and Parametrization of Controllers

In this section we want to give an interpretation of the controller constructed in Theorem

7.13 which solves the output-feedback H∞ -problem.
Moreover we will parameterize all controllers that solve the output-feedback H∞ -problem.
We will especially see that every such controllers equal Kpar ? Q for a fixed LTI system

196
Kpar and a stable parameter Q with kQk∞ < γ.
In order to do so, we need again some preparations.

The following lemma reveals conditions under which external and internal stability are
equivalent.

Lemma 7.28 Let P and K be a generalized plant and a controller with the usual realiza-
tios. Suppose that
 
A − λI B2
  has full column rank for all λ ∈ C0 ∪ C+ (7.52)
C1 D12

and  
A − λI B1
  has full row rank for all λ ∈ C0 ∪ C+ . (7.53)
C2 D21
Then K stabilizes P iff P ? K is stable.

+
If D12 has full column rank and D12 T
:= (D12 D12 )−1 D12
T
then (7.52) holds iff
+ +
(A − B2 D12 C1 , C1 − D12 D12 C1 ) is detectable.
+ T T −1 T
If D21 has full row rank and D21 := D21 (D21 D21 ) D12 then (7.53) holds iff
+ +
(A − B1 D21 C2 , B1 − B1 D21 D21 ) is stabilizable.

Also note that these properties are implied by controllability of (A, B1 ), observability of
(A, C1 ) and    
T
    B1 T
0
D12 C1 D12 = 0 I and   D21 =  .
D21 I

The last property was just (7.14). It is an exercise (Exercise 2) to prove Lemma 7.28 and
the properties above.

Definition 7.29 A transfer matrix G is said to be all-pass if it has no poles on the

imaginary axis and satisfies

G(iω)∗ G(iω) = I for all ω ∈ [0, ∞]. (7.54)

G is said to be inner if it is all-pass and stable.

197
Analogously to inner transfer matrices G ∈ RH∞ is said to be co-inner if

G(iω)G(iω)∗ = I for all ω ∈ [0, ∞].

Clearly a square transfer matrix G is inner if and only if G is co-inner. We will focus
on inner transfer matrices. The properties for co-inner transfer matrices can be obtained
by duality. Inner generalized plants P are very useful to us because of the properties in
Lemma 7.32, which is a characterization of stabilizing controllers K that achieve kP ?
Kk∞ < 1. They will appear when we modify the standard generalized plant with the
help of the “worst-case disturbance” −γ −2 B1T Y∞ x and the optimal control −B2T Y∞ x.

At first we present two characterizations of inner transfer matrices.

Lemma 7.30 G is inner if and only if kGuk2 = kuk2 for all u ∈ L2 .

Proof. Suppose G is inner then also G(iω)∗ G(iω) = I for all ω ∈ R. Let u ∈ L2 be
arbitrary. Since G is stable Gu ∈ L2 and we can use Parceval’s identity:
Z ∞
2
kGuk2 = kGu(t)k2 dt
0
Z ∞
1
= kG(iω)û(iω)k2 dω
2π −∞
Z ∞
1
= û(iω)∗ G(iω)∗ G(iω)û(iω)dω
2π −∞
Z ∞
1
= kû(iω)k2 dω = kuk22 .
2π −∞
Now suppose kGuk2 = kuk2 for all u ∈ L2 . Then G ∈ RH∞ and we get
kGuk2
kGk∞ = sup =1 (7.55)
0<kuk2 <∞ kuk2

We now show that

kG(iω)xk = kxk for all x and all ω ∈ R.
Due to (7.55) we only need to show “≥”. Assume there exist ω0 and x0 with kG(iω0 )x0 k <
kx0 k. Then the function u(t) = e−t (iω0 + 1)x0 ∈ L2 satisfies û(iω0 ) = x0 . We can now
use Parceval’s identity as above to obtain
Z ∞
0= kû(iω)k2 − kG(iω)û(iω)k2 dω
−∞

with kû(iω)k2 − kG(iω)û(iω)k2 ≥ 0 for all ω ∈ R and kû(iω0 )k2 − kG(iω0 )û(iω0 )k2 > 0.
Since G is stable and due to continuity, this is a contradiction.

The polarisation identity now implies

hy, G(iω)∗ G(iω)xi = hy, xi for all x, y and all ω ∈ R

198
and hence G(iω)∗ G(iω) = I for all ω ∈ [0, ∞].

Theorem 7.31 For G(s) = C(sI − A)−1 B + D with eig(A) ⊂ C− let X denote the
observability Gramian, the unique solution of the Lyapunov equation AT X +XA+C T C =
0. Then G is inner if
B T X + DT C = 0 and DT D = I.
If (A, B) is controllable, the statement holds with “iff ”.

Proof. Define G∗ (s) := G(−s)T to infer G(iω)∗ = G∗ (iω). Note that

 
T T
−A C
G∗ (s) has the realization  .
−B T DT

Therefore (7.54) holds iff the transfer matrix

 
   −AT C T C C T D
−AT C T A B
G∗ G = 
 
 = 0 A B 
−B T DT C D
 
−B T T T
D C D D

equals the identity matrix I. A coordinate change with X implies that G∗ G can also be
realized as
 
−AT AT X + XA + C T C XB + C T D
 = H ∗ + H + DT D
 
 0 A B
 
T T T T
−B B X +D C D D

with H(s) = (B T X + DT C)(sI − A)−1 B and H ∗ (s) := H(−s)T . If B T X + DT C = 0 and

DT D = I we get trivially G∗ (iω)G(iω) = I for all ω ∈ [0, ∞] and since G is stable by
assumption, G is inner.
If G is inner, we can conclude that H ∗ (iω) + H(iω) is constant for ω ∈ [0, ∞]. Since H
and H ∗ are strictly proper we obtain

0 = H ∗ (∞) + H(∞) = H ∗ (iω) + H(iω) = H(iω)∗ + H(iω)

for all ω ∈ R. Hence DT D = I.

As a next step we show that 0 = H(iω)∗ + H(iω) for all ω ∈ R implies H = 0. This
requires some arguments from complex analysis which can be read-off in any book on
complex analysis.

Choose arbitrary vectors u,v and set Ω := C\(eig(A) ∪ eig(−AT )). Then define the
analytic function f : Ω → C, f (z) = hu, H(z)vi and g : Ω → C, g(z) = hu, −H(−z)T vi.

199
By assumption f equals g on C0 and hence f equals g on Ω. Since f is analytic on C+
and g is analytic on C− , f has an analytic continuation f˜ on C. Strict properness of H
implies f (z) → 0 for |z| → ∞ and the same holds for the continuation f˜. Hence we can
conclude that f˜ is bounded on C and by Liouville f˜ is constant on C. Using again that
H is strict proper implies that f˜ vanishes on C. Hence hu, H(z)vi = 0 for all z ∈ C and
all u, v which means H = 0.

Now assume there exists x0 with (B T X + DT C)x0 6= 0. By controllability of (A, B) we

find a smooth u such that the solution of ẋ = Ax + Bu satisfies x(1) = x0 . We can make
sure that u ∈ L2 . Then y(1) = (B T X + DT C)x0 6= 0. On the other hand we get with
Parceval Z ∞
2
kyk2 = kHu(t)k2 dt
0
Z ∞
1
= kH(iω)û(iω)k2 dω = 0.
2π −∞
Since y is also continuous, this means y(t) = 0 for all t ∈ [0, ∞). This is a contradiction.

The proof also shows that G is inner if and only if G(−z)T G(z) = I for all z ∈ C.

If a generalized plant is inner and the transfer function from w to y has a stable inverse,
stabilizing controllers that achieve an H∞ -norm bound 1 are very special: they must be
contained in the open unit ball of RH∞ . The converse holds as well.

Lemma 7.32 Consider the inner generalized plant

      
z w P P w
  = P   =  11 12    with P21 −1
∈ RH∞ .
y u P21 P22 u

Then the transfer matrix K stabilizes P and achieves kP ? Kk∞ < 1 if and only if

K ∈ RH∞ and kKk∞ < 1.

This result is instrumental for parameterizing all suboptimal controllers.

Proof. “⇐”: Since P is inner, we can conclude that kP k∞ = 1 and hence

kP22 k∞ ≤ kP k∞ = 1.

We can apply Theorem 4.10 to obtain stability of (I − P22 K)−1 . Since K and P are also
stable K internally stabilizes P .
−1
Let kKk∞ ≤ γ with γ ∈ (0, 1) and define κ = kP21 (I − P22 K)k∞ . Now choose any
ω ∈ [0, ∞] and any complex vector x. Then z = (P ? K)(iω)x satisfies, with y =

200
(I − P22 (iω)K(iω))−1 P21 (iω)x the relations
    
z P (iω) P12 (iω) x
  =  11  . (7.56)
y P21 (iω) P22 (iω) K(iω)y

Since P (iω)∗ P (iω) = I, we infer

kzk2 + kyk2 = kxk2 + kK(iω)yk2 ≤ kxk2 + γ 2 kyk2 .

Since x = P21 (iω)−1 (I−P22 (iω)K(iω))y we have kxk2 ≤ κ2 kyk2 and thus, due to γ 2 −1 < 0
and κ > 0,
kzk2 ≤ [1 + (γ 2 − 1)κ−2 ]xk2 .
Since x, ω were arbitrary, we get k(P ? K)(iω)k < 1 for all ω and hence kP ? Kk∞ < 1.

“⇒”: Suppose supω∈[0,∞] kK(iω)k ≥ 1. Then we can choose ω ∈ [0, ∞] such that iω
is not a pole of K and some complex vector y 6= 0 with kK(iω)yk ≥ kyk. If defining
x := P21 (iω)−1 (I − P22 (iω)K(iω))y and z := (P ? K)(iω)x, the relation (7.56) holds
again. This now implies

kzk2 + kyk2 = kxk2 + kK(iω)yk2 ≥ kwk2 + kyk2

and thus k(P ? K)(iω)xk = kzk ≥ kxk. We infer k(P ? K)(iω)k ≥ 1, which in turn means
kP ? Kk∞ ≥ 1, a contradiction.

We conclude kK(iω)k < 1 for all ω ∈ [0, ∞]. In particular, K has no poles in C0 . Since
P is inner, it trivially satisfies kP k∞ ≤ 1. Since K stabilizes P we can conclude stability
of K by using Lemma 4.6.

We now re-describe the generalized plant P . Therefore let P be as usual

ẋ = A x + B1 w + B2 u
z = C1 x + D11 w + D12 u
y = C2 x + D21 w + D22 u
under the following hypothesis.

Hypothesis 7.33 Suppose that

(A, B2 ) is stabilizable and (A, C2 ) is detectable. (OF1)

   
T
    B1 T
0
D12 C1 D12 = 0 I and   D21 =  . (OF2)
D21 I
D11 = 0 and D22 = 0. (OF3)
(A, C1 ) is observable and (A, B1 ) is controllable. (OF4)

201
Moreover suppose that there exist Y∞ ≥ 0 with

AT Y∞ + Y∞ A + Y∞ (γ −2 B1 B1T − B2 B2T )Y∞ + C1T C1 = 0, (7.57)

−2 −
eig(A + (γ B1 B1T − B2 B2T )Y∞ ) ⊂C (7.58)

with γ > 0.

Lemma 7.17 motivates to introduce the new disturbance/control variable

d = w − γ −2 B1T Y∞ x and v = u − F∞ x, F∞ := −B2T Y∞ (7.59)

for the generalized plant P if ignoring y. Note that d is the deviations of the worst case
disturbance γ −2 B1T Y∞ x and that v is the deviation of the optimal control −B2T Y∞ x in
the sense of Lemma 7.18, Lemma 7.19 and Theorem 7.20. For x(0) = 0 and trajectories
of finite energy, Lemma 7.17 implies with T → ∞ that

kzk22 + γ 2 kdk22 = γ 2 kwk22 + kvk22 .

This means that the following system is inner:

    
ẋ A + B2 F∞ γ −1 B1 B2 x
    
G:   z  =  C1 + D12 F∞
  0   γw  .
D12    (7.60)
γd −γ −1 B1T Y∞ I 0 v

It is indeed possible to algebraically verify the following result.

Lemma 7.34 Suppose that Hypothesis 7.33 holds. Then G defined in (7.60) satisfies all
hypothesis of Lemma 7.32.

Proof. Let G denote the transfer matrix of (7.60) partioned accordingly. Since A +
B2 F∞ +γ −2 B1 B1T Y∞ is Hurwitz and (7.57) holds, we infer as already done eig(A+B2 F∞ ) ⊂
C− with Theorem 7.9 as in the proof of Lemma 7.16. Thus we get
 
−2 T −1
A + B2 F∞ + γ B1 B1 Y∞ γ B1
G ∈ RH∞ and G−1 21 =
  ∈ RH∞ .
γ −1 B1T Y∞ I

Due to (C1 + D12 F∞ )T (C1 + D12 F∞ ) = C1T C1 + Y∞ B2 B2T Y∞ , the ARE for Y∞ can be
written as
 T  
C1 + D12 F∞ C + D12 F∞
(A + B2 F∞ )T Y∞ + Y∞ (A + B2 F∞ ) +    1  = 0.
−1 T −1 T
−γ B1 Y∞ −γ B1 Y∞

202
 G
 
z A + B2 F∞ γ −1 B1 B2 γw
 
C +D F 0 D12 
 1 12 ∞ 
γd −γ −1 B1T Y∞ I 0

γ −1 v
P P̂
   
z A B1 B2 w d A + γ −2 B1 B1T Y∞ B1 B2
   
C 0 D 
 1 12 

 −F∞ 0 I 

C2 D21 0 C2 D21 0
y u y u

K K

Figure 60: Re-description of the generalized plant P .

Hence Y∞ is the observability Gramian corresponding to G. Moreover we trivially have

    
γ −1 B1T 0 I C1 + D12 F∞
  Y∞ +   =0
B2T T
D12 0 −γ −1 B1T Y∞
T
and D12 D12 = I. Hence G is inner by Theorem 7.31.

If we apply (7.59) to P while ignoring z, we optain

    
ẋ A + γ −2 B1 B1T Y∞ B1 B2 x
    
P̂ : 
v=
  −F∞ 0 I  d .
  (7.61)
y C2 D21 0 u

By the mere definitions of the systems (7.60)-(7.61), their interconnection (both uncon-
trollend and controlled) as depicted in Figure 60 just constitutes another description of
the original generalized plant P .

Similarly to P , G also satisfies the properties of Lemma 7.28. Then a direct application
of Lemma 7.32 with G leads to the following key relation.

Lemma 7.35 Suppose that Hypothesis 7.33 holds. Then the transfer matrix K stabilizes
P with kP ? Kk∞ < γ iff P̂ ? K is stable with kP̂ ? Kk∞ < γ.

203
Proof. Due to (OF2) and (OF4) P satisfies the hypothesis of Lemma 7.28. The same is
true for G. Since
+ +
(A + B2 F∞ − B2 D12 (C1 + D12 F∞ ), C1 + D12 F∞ − D12 D12 (C1 + D12 F∞ )) = (A, C1 )

is detectable and
(A + B2 F∞ + γ −2 B1 B1T Y∞ , 0)
is stabilizable.

Now suppose K stabilizes P with kP ? Kk∞ < γ. Then we get as depicted in Figure 60

(P ? K)γ −1 = G ? K̂ (7.62)

for K̂ := (P̂ ? K)γ −1 . Hence G ? K̂ ∈ RH∞ with kG ? K̂k∞ < 1. By Lemma 7.28, K̂
stabilizes G. Then Lemma 7.32 implies that K̂ is stable and kK̂k∞ < 1. Hence P̂ ? K is
stable with kP̂ ? Kk∞ < γ.

Conversely, P̂ ?K ∈ RH∞ and kP̂ ?Kk∞ < γ implies for K̂ := (P̂ ?K)γ −1 that K̂ ∈ RH∞
and kK̂k∞ < 1. By Lemma 7.32, K̂ stabilizes G with kG ? K̂k∞ < 1. By (7.62),
P ? K ∈ RH∞ with kP ? Kk∞ < γ. By Lemma 7.28, K also stabilizes P .

As a next step we express the conditions in Theorem 7.13 as follows.

Lemma 7.36 Suppose that (OF1) - (OF4) hold. Then there exist X∞ ≥ 0 and Y∞ ≥ 0
with

AX∞ + X∞ AT + X∞ (γ −2 C1T C1 − C2T C2 )X∞ + B1 B1T = 0

eig(A + X∞ (γ −2 C1T C1 − C2T C2 )) ⊂ C−
AT Y∞ + Y∞ A + Y∞ (γ −2 B1 B1T − B2 B2T )Y∞ + C1T C1 = 0
eig(A + Y∞ (γ −2 B1 B1T − B2 B2T )) ⊂ C−

and
ρ(X∞ Y∞ ) < γ 2
if and only if there exists Y∞ ≥ 0 and X̂∞ ≥ 0 with

AT Y∞ + Y∞ A + Y∞ (γ −2 B1 B1T − B2 B2T )Y∞ + C1T C1 = 0

eig(A + Y∞ (γ −2 B1 B1T − B2 B2T )) ⊂ C−

and, for  := A + γ −2 B1 B1T Y∞ and F∞ := −B2T Y∞ ,

ÂX̂∞ + X̂∞ ÂT + X̂∞ (γ −2 F∞

T
F∞ − C2T C2 )X̂∞ + B1 B1T = 0
eig(Â + X̂∞ (γ −2 F∞
T
F∞ − C2T C2 )) ⊂ C− .
−1
The transformation is achieved with X̂∞ = (X∞ − γ −2 Y∞ )−1 .

204
The two decoupled AREs with coupled solutions are thus rewritten into two coupled
AREs whose solutions are not coupled any more. More interestingly, the ARE for X̂∞
admits an important interpretation!

Proof. “⇒”: Due to (OF4) X∞ and Y∞ are invertible. We can write the given ARE’s as
−1 −1 −1 −1
AT X∞ + X∞ A + X∞ B1 B1T X∞ + γ −2 C1T C1 − C2T C2 = 0

and

AT (γ −2 Y∞ ) + (γ −2 Y∞ )A + (γ −2 Y∞ )B1 B1T (γ −2 Y∞ ) − γ −2 F∞
T
F∞ + γ −2 C1T C1 = 0.

−1
We can now apply Lemma 7.24 to ∆ = X∞ − γ −2 Y∞ and get

ÂT ∆ + ∆ + ∆B1 B1T ∆ + γ −2 F∞

T
F∞ − C2T C2 = 0 (7.63)

with  := A + γ −2 B1 B1T Y∞ . Since Y∞ is positive definite, it has a positive def-

1/2 1/2
inite square root. Hence we can use γ 2 > ρ(X∞ Y∞ ) = ρ(Y∞ X∞ Y∞ ) to obtain
−1/2 −1 −1/2
λmin (Y∞ X∞ Y∞ ) > γ −2 since X∞ is invertible. This implies
−1
∆ = X∞ − γ −2 Y∞ = Y∞1/2 (Y∞−1/2 X∞
−1 −1/2
Y∞ − γ −2 I)Y∞1/2 > 0.

Now observe that

−1
 + B1 B1T ∆ = A + B1 B1T X∞ = X∞ (−AT − (γ −2 C1T C1 − C2T C2 )X∞ )X∞
−1

is anti-stable. Hence we can conclude with X̂∞ := ∆−1 > 0 that

 + X̂∞ (γ −2 F∞
T
F∞ − C2T C2 ) = ∆−1 (−ÂT − ∆B1 B1T )∆

is Hurwitz. Together with (7.63) this completes “if”.

“⇐”: (OF4) implies again that Y∞ and X̂∞ are invertible since (Â, B1 ) is controllable.
Then X̃∞ := γ −2 Y∞ + X̂∞
−1
> 0 solves the ARE

AT X̃∞ + X̃∞ A + X̃∞ B1 B1T X̃∞ + γ −2 C1T C1 − C2T C2 = 0

−1
Now set X∞ := X̃∞ > 0 to see that X∞ solves the desired ARE. As before since eig(Â +
X̂∞ (γ F∞ F∞ − C2 C2 )) ⊂ C− we can conclude eig(A + X∞ (γ −2 C1T C1 − C2T C2 ) ⊂ C− .
−2 T T

Finally observe that

−1 −1
0 < X̂∞ = X∞ − γ −2 Y∞ = Y∞1/2 (Y∞−1/2 X∞
−1 −1/2
Y∞ − γ −2 I)Y∞1/2 .
1/2 −1/2 −1 −1/2
Since Y∞ > 0, this implies λmin (Y∞ X∞ Y∞ ) > γ −2 . Hence we can conclude γ 2 >
1/2 1/2
ρ(Y∞ X∞ Y∞ ) = ρ(X∞ Y∞ ).

205
7.11.1 Interpretation of the Output-Feedback Controller

Let us assume that there exists a K stabilizing P with kP ? Kk∞ < γ. By Theorem 7.13,
we can construct the central controller. We are now ready to provide an interpretation
of this controller and to prove that it really solves the H∞ -problem. By Lemma 7.35,
the latter requires to show that the controller in Theorem 7.13 renders P̂ ? K stable with
kP̂ ? Kk∞ < γ.

The optimal state-feedback controller for P̂ is u = F∞ x (with optimal value zero). If

only y is measured, one could try to reconstruct the required control action with an
H∞ -observer for P̂ with the structure

x̂˙ = Âx̂ + B2 u + L(C2 x̂ − y), û = F∞ x̂,

and then control P̂ with u = û. Theorem 7.10 and Lemma 7.36 motivate to choose the
observer gain L̂∞ = −X̂∞ C2T . This leads to the controller

x̂˙ = (Â + L̂∞ C2 )x̂ + B2 u − L̂∞ y, u = F∞ x̂. (7.64)

Due to X̂∞ = ZX∞ = (I − γ −2 X∞ Y∞ )−1 X∞ , this is exactly the controller in Theorem

7.13.

If we introduce the “worst-case disturbance” into the original generalized plant by w =

d + γ −2 B1T Y∞ x, we obtain
    
ẋ A + γ −2 B1 B1T Y∞ B1 B2 x
    
z= C1 0  d .
D12   
  
y C2 D21 0 u

This leads to the following interpretation of the central controller:

The output-feedback H∞ -controller results from replacing u in the H∞ -state-feedback
controller u = F∞ x with the estimate û, obtained by an H∞ -observer for the output F∞ x
of the generalized plant “in the presence of the worst-case disturbance γ −2 B1T Y∞ x”.

This is often called the separation principle in H∞ -theory. For γ = ∞, it reduces to the
classical separation principle for H2 -controllers, since then the “worst-case disturbance”
vanishes.

We still owe the proof that the constructed controller (7.64) based on the solutions of the
Riccati equations does indeed solve the H∞ -problem.

Lemma 7.37 Suppose that Hypothesis 7.33 holds and that there exists X∞ ≥ 0 with
AX∞ + X∞ AT + X∞ (γ −2 C1T C1 − C2T C2 )X∞ + B1 B1T = 0
eig(A + X∞ (γ −2 C1T C1 − C2T C2 )) ⊂ C−

206
and
ρ(X∞ Y∞ ) < γ 2 .
Moreover let  = A + γ −2 B1 B1T , X̂∞ = (X∞
−1
− γ −2 Y∞ )−1 , L̂∞ = −X̂∞ C2T and F∞ =
−B2T F∞ . Then the controller K defined by

x̂˙ = (Â + L̂∞ C2 )x̂ + B2 u − L̂∞ y, u = F∞ x̂.

stabilizes P and ensures kP ? Kk∞ < γ.

Proof. If interconnecting P̂ from (7.61) with this controller we obtain

    
ẋ Â B2 F∞ B1 x
 x̂˙  =  −L̂∞ C2 Â + L̂∞ C2 + B2 F∞ −L̂∞ D21   x̂  .
    
    
v −F∞ F∞ 0 d
The standard transformation into the error dynamics leads to
 
 + B2 F∞ B2 F∞ B1  
  Â + L̂∞ C2 −(B1 + L̂∞ D21 )
=
 0 Â + L̂∞ C2 −(B1 + L̂∞ D21 )   .
F∞ 0

0 F∞ 0

Now set X := X̂∞ . Then Lemma 7.36 implies

ÂT X + X Â + XB1 B1T X + γ −2 F∞

T
F∞ − C2T C2 = 0
eig(Â + B1 B1T X) ⊂ C+ .

The last statement follows as usual from

 + B1 B1T X = X −1 (ÂT + (γ −2 F∞
T
F∞ − C2T C2 )X̂∞ )X

since Â+ X̂∞ (γ −2 F∞

T
F∞ −C2T C2 ) is Hurwitz. Moreover observe that (Â, C2 ) is detectable,
T T
B1 D21 = 0 and D21 D21 = I. Theorem 7.10 implies

eig(Â + L̂∞ C2 ) ⊂ C− and kF∞ (sI − Â − L̂∞ C2 )−1 (B1 + L̂∞ D21 )k∞ < γ.

Now Lemma 7.35 finishes the proof.

7.11.2 Controller Parameterization

Let us go further and parameterize all suboptimal H∞ -controllers for P . The key technical
idea is to dualize and re-describe P̂ as done above for P . Let us first transpose P̂ to obtain
    
x̃˙ ÂT −F∞ T
C2T x̃
    
P̃ :    T 0 DT   w̃ 
 z̃  =  B1 21   
T
ỹ B2 I 0 ũ

207
Next, with X̂∞ we introduce the new variables

d˜ = w̃ + γ −2 F∞ X̂∞ x̃ and ṽ = ũ − L̂T∞ x̃, L̂∞ = −X̂∞ C2T

to arrive at the system

    
ẋ ÂT + C2T L̂T∞ −γ −1 F∞
T
C2T x̃
    
G̃ :  z̃  =  B T + DT L̂T 0 T 
D21 γ w̃ .
   1 21 ∞  
γd ˜ −1
γ F∞ X̂∞ I 0 ṽ

Due to the ARE for X̂∞ also G̃ satisfies the properties in Lemma 7.32. Moreover P̃ is the
feedback interconnection of G̃ with
    
x̃˙ ÂT − γ −2 F∞T
F∞ X̂∞ −F∞ T
C2T x̃
    
P̌ :  L̂T∞  ˜
 ṽ  =  0 I  d .
 

ỹ B2T + γ −2 F∞ X̂∞ I 0 d˜

Let us “transpose back” which results in

 
 − γ −2 F∞
T
F∞ −L̂∞ B2 + γ −2 X̂∞ F∞
T
 
Ptmp = 
 −F∞ 0 I .

C2 I 0

Just using the fact that the H∞ -norm of a transfer matrix stays invariant under transpo-
sition, we obtain the following result.

Lemma 7.38 Suppose that Hypothesis 7.33 holds and that there exists some X̂∞ ≥ 0
with

ÂX̂∞ + X̂∞ ÂT + X̂∞ (γ −2 F∞

T
F∞ − C2T C2 )X̂∞ + B1 B1T = 0
eig(Â + X̂∞ (γ −2 F∞
T
F∞ − C2T C2 )) ⊂ C− .

for  := A + γ −2 B1 B1T Y∞ and F∞ := −B2T Y∞ .

Then the LTI controller K renders P̂ ? K stable and achieves kP̂ ? Kk∞ < γ if and only
if Ptmp ? K ∈ RH∞ with kPtmp ? Kk∞ < γ.

Proof. First we need to verify some technical properties.

Since  + B2 F∞ is Hurwitz, the pair (Â, B2 ) is stabilizable. Set as before L̂∞ := −X̂∞ C2T .
Then (Â, C2 ) is detectable, since  + L̂∞ C2 is Hurwitz by Theorem 7.10. Hence P̂ is a
generalized plant.
+ T T −1 + +
Since B1 D21 = B1 D21 (D21 D21 ) = 0, (Â − B1 D21 C2 , B1 − B1 D21 D21 ) = (Â, B1 ) is
stabilizable by (OF4). Moreover (Â + B2 F∞ , 0) is clealy detectable. Hence P̂ satisfies all
hypothesis in Lemma 7.28.

208
The realization of P̃ is the transpose of that of P̂ . Hence P̃ is also a generalized plant
and its realization satisfies the hypothesis in Lemma 7.28.
T + T −1
Next set H := (D21 ) (B1T + D21
T T
L̂∞ ). Then H = (D21 D21 ) D21 (B1T + D21 T T
L̂∞ ) = L̂T∞ , by
(OF2). Moreover (ÂT + C2T L̂T∞ − C2T H, (B1T + D21 T T T
L̂∞ ) − D21 H) = (ÂT , B1T ) is detectable
and (ÂT +C2T L̂T∞ −γ −2 F∞
T
F∞ X̂∞ , 0) is stabilizable, since eig(Â+X̂∞ (γ −2 F∞
T
F∞ −C2T C2 )) ⊂
C− . Further since ÂT + C2T L̂T∞ is Hurwitz, G̃ is a generalized plant. Hence the realization
of G̃ satisfies the hypothesis in Lemma 7.28.

Next we show that G̃ satisfies the hypothesis of Lemma 7.32. We can use Theorem 7.31
to show that G̃ is inner since eig(ÂT + C2T L̂T∞ ) ⊂ C− .
First observe that
 
T T T −2 T −1 T
 + C2 L̂∞ − γ F∞ F∞ X̂∞ γ F∞
G̃−1
21 =
  ∈ RH∞
γ −1 F∞ X̂∞ I

since eig(Â + X̂∞ (γ −2 F∞

T
F∞ − C2T C2 )) ⊂ C− by assumption.
Second, due to (OF2), L̂∞ = −X̂∞ C2T and the ARE for X̂∞ we have
 T  
B1T + T T
D21 L̂∞ B1T + T T
D21 L̂∞
(ÂT + C2T L̂T∞ )T X̂∞ + X̂∞ (ÂT + C2T L̂∞ ) +    
γ −1 F∞ X̂∞ γ −1 F∞ X̂∞
=(ÂT + C2T L̂T∞ )T X̂∞ + X̂∞ (ÂT + C2T L̂∞ ) + B1 B1T + L̂∞ L̂T∞ + γ −2 X̂∞ F∞
T
F∞ X̂∞
=ÂX̂∞ + X̂∞ ÂT − 2X̂∞ C2T C2 X̂∞ + B1 B1T + L̂∞ L̂T∞ + γ −2 X̂∞ F∞
T
F∞ X̂∞
=0.

This means that X̂∞ is the observability Gramian for G̃.

Third, we trivially have
    
−γ −1 F∞ 0 I T T
B1T + D21 L̂∞
  X̂∞ +   =0
C2 D21 0 γ −1 F∞ X̂∞

and  T  
T T
0 D21 0 D21
    = I.
I 0 I 0

Hence G̃ is inner.

For transfer matrices P and K with proper (I − P22 K)−1 , observe that
T −1 T T
(P ? K)T = P11
T T
+ P21 (I − K T P22 ) K P12
T
= P11 T
+ P21 K T (I − P22
T
K T )−1 P12
T
= P T ? KT .

Therefore P ? K ∈ RH∞ with kP ? Kk∞ < 1 iff P T ? K T ∈ RH∞ with kP T ? K T k∞ < 1

209
This allows us to argue as follows: P̂ ? K ∈ RH∞ with kP̂ ? Kk∞ < γ iff P̃ ? K T ∈ RH∞
with kP̃ ? Kk∞ < γ iff (Lemma 7.28 for P̃ ) K T stabilizes P̃ with kP̃ ? K T k∞ < γ iff
(Lemma 7.35 for P̃ , G̃ and P̌ ) P̌ ? K T ∈ RH∞ with kP̌ ? Kk∞ < γ iff (transposing)
Ptmp ? K ∈ RH∞ with kPtmp ? Kk∞ < γ.

The system Ptmp has a very particular structure, which allows us to solve Q = Ptmp ? K
−1
for K as K = Q ? Ptmp and vice versa.

Together with Lemma 7.38, this has the following implication

K stabilizes P with kP ? Kk∞ < γ

⇔
Ptmp ? K ∈ RH∞ with kPtmp ? Kk∞ < γ
⇔
−1
K = Q ? Ptmp for some Q ∈ RH∞ with kQk∞ < γ.

Remark 7.39 All γ-suboptimal H∞ -controllers for P can be expressed as an LFT of a

−1
fixed system Ptmp with a parameter Q that is only constrained to be stable and bounded
in H∞ -norm by γ. This means that Q varies in the open RH∞ -ball of radius γ.

Theorem 7.40 Under the assumptions (OF1)- (OF4), let there exist an LTI controller
K which stabilizes P with kP ? Kk∞ < γ. With X∞ and Y∞ as in Theorem 7.13 define

Z∞ := (I − γ −2 X∞ Y∞ )−1 and F∞ := −B2T Y∞ , L∞ := −X∞ C2T

as well as
 
A + γ −2 B1 B1T Y∞ + B2 F∞ + Z∞ L∞ C2 −Z∞ L∞ Z∞ B2
 
Kpar := 
 F∞ 0 I .

−C2 I 0

Then K stabilizes P with kP ? Kk∞ < γ if and only if K = Kpar ? Q with the parameter
Q ∈ RH∞ satisfying kQk∞ < γ.

This explains the terminology for the central controller in Theorem 7.13:
 
−2 T
A + γ B1 B1 Y∞ + B2 F∞ + Z∞ L∞ C2 −Z∞ L∞
Kpar ? 0 =  .
F∞ 0

Proof. The system Ptmp has a very particular structure: It has a proper inverse since
−1
Ptmp (∞) is invertible and the (2, 2)-block of Ptmp and the (1, 1)-block of Ptmp are strictly

210
 −1
proper. Hence for LTI systems K and Q, the LFT’s Ptmp ? K and Q ? Ptmp are well-posed
and LTI.

Moreover the (1, 2) and the (2, 1) blocks of Ptmp are square and invertible. By computation
−1 −1
one verifies that Q = Ptmp ? K implies K = Q ? Ptmp . By symmetry, K = Q ? Ptmp implies
Q = Ptmp ? K.
−1
To prove the first implication, fix λ ∈ C such that it is not a pole of Ptmp , Ptmp ,
−1
(Ptmp,21 ) , K, Q and any complex vector y. Define u := K(λ)y and z := Q(λ)w
with w := Ptmp,21 (λ)−1 (I − Ptmp2 1 (λ)K(λ))y. Hence y = Ptmp,21 (λ)w + Ptmp,22 (λ)u and
z = Ptmp,11 (λ)w + Ptmp,12 (λ)u. These satisfy
       
z w w z
  = Ptmp (λ)   ⇔   = Ptmp (λ)−1   .
y u u y

Hence K(λ)y = u = [Q(λ) ? Ptmp (λ)−1 ]y, i.e. K(λ) = Q(λ) ? Ptmp (λ)−1 since y is free.
−1
Thus K = Q ? Ptmp almost everywhere.

It remains to verify the formula for Kpar . Note that B2 + γ −2 X̂∞ F∞

T
= (I + γ −2 X̂∞ Y∞ )B2
and
I + γ −2 X̂∞ Y∞ = X̂∞ (X̂∞−1
+ γ −2 Y∞ ) = X̂∞ X∞
−1
= Z∞ .
We hence clearly have
 
 + B2 F∞ + L̂∞ C Z∞ B2 −L̂∞
−1
 
Ptmp =
 −C2 0 I .

F∞ I 0
−1
Finally, the upper Q ? Ptmp is easily translated into the lower LFT as in
 
 + B2 F∞ + L̂∞ C2 −L̂∞ Z∞ B2
−1
 
Q ? Ptmp = F∞ 0 I   ? Q.
−C2 I 0

This completes the proof.

Exercises
1) Consider the standard ARE AT X +XA+XRX +Q = 0 for real matrices A, R = RT
and Q = QT and possibly non-Hermitian complex unknowns X.
If X is a stabilizing solution of the ARE, show that X = X T , X = X and that X
is unique.

211
2) Prove Lemma 7.28 and show that
+
a) If D12 has full column rank and D12 T
:= (D12 D12 )−1 D12
T
then (7.52) holds iff
+ +
(A − B2 D12 C1 , C1 − D12 D12 C1 ) is detectable.

+ T T −1 T
b) If D21 has full row rank and D21 := D21 (D21 D21 ) D12 then (7.53) holds iff
+ +
(A − B1 D21 C2 , B1 − B1 D21 D21 ) is stabilizable.

3) Let (A, B) ∈ Rn×(n+m) be stabilizable. With Q = QT , R = RT ≥ 0 define

 T   
x Q S x
q(x, u) :=     .
u ST R u

For ξ ∈ Rn let U (ξ) be the set of all pieveweise continuous (p.c.) control functions
u ∈ L2 [0, ∞)m such that the trajectory of ẋ = Ax + Bu, x(0) = ξ satisfies x ∈
L2 [0, ∞)n . Consider the so-called indefinite LQ-problem with stability:
Z ∞
v(ξ) = inf{ q(x(t), u(t))dt : ẋ(t) = Ax(t) + Bu(t), x(0) = ξ, u ∈ U (ξ)}.
0

Suppose that v(ξ) > −∞ for all ξ ∈ Rn .

a) Show that v(0) = 0.
b) Show that v satisfies the dissipation inequality
Z t2
v(x(t1 )) ≤ q(x(t), u(t))dt + v(x(t2 )) for any trajectory and 0 ≤ t1 ≤ t2 .
t1

c) For 0 ≤ t1 ≤ t2 show that v satisfies the optimality principle

Z t2 
v(x(t1 )) = inf q(x(t), u(t))dt + v(x(t2 )) .
u p.c. on [t1 ,t2 ] t1

d) Show that v satisfies the parallelogramm identity and that there exists c > 0
with |v(ξ)| ≤ ckξk2 for all ξ ∈ Rn . Argue (without details) why there hence
exists a symmetric matrix K with v(ξ) = ξ T Kξ.

4) Let the hypothesis of Exercise 3 hold.

a) Show that K satisfies the linear matrix inequality (LMI)
   
T
A K + KA KB Q S
 +  ≥ 0. (7.65)
T T
B K 0 S R

b) Show that X ≤ K if X = X T is any solution of the LMI.

212
5) In addition to the conditions in Exercise 3 assume that R > 0.
a) Show that the cost of the original problem and the one with the data
     
−1
Q̃ 0 I −SR Q S I 0
  :=    
−1/2 T −1 T −1/2
0 I 0 R S R −R S R

and  
I 0
(Ã, B̃) := (A, B)  
−R−1 S T R−1/2
are identical. How are close-to-optimal control functions related? Hence we
can assume w.l.o.g. that S = 0 and R = I, which is done from now on.
b) Show that X satisfies (7.65) if and only if it satisfies the ARI

AT X + XA − XBB T X + Q ≥ 0. (7.66)

c) For some C and T > 0 let P satisfy the Riccati differential equation Ṗ +AT P +
P A − P BB T P + C T C = 0 with P (T ) = 0 on [0, T ]. Show that
Z T
min kCx(t)k2 + ku(t)k2 dt = ξ T P (0)ξ.
ẋ=Ax+Bu, x(0)=ξ,u p.c. 0

d) If the minimum in c) is zero for all ξ ∈ Rn , show that C = 0.

e) Show that K even satisfies the ARE that corresponds to (7.66).
Hint: Choose C with C T C = AT K + KA − KBB T K + Q and exploit c), d)
and Exercise 3.

6) Consider the ARI AT X + XA − XBB T X + Q ≥ 0 and the corresponding ARE

for stabilizable (A, B) and Q = QT . All unknowns are real-symmetric. An ARE
solution Xs is called strong if it satisfies eig(A − BB T X) ⊂ C− ∪ C0 .
a) If Xs is a strong solution, show that

AT X + XA − XBB T X + Q ≥ 0 ⇒ X ≤ Xs . (7.67)

Hint: Use the difference trick; Prove and work with the fact that ker(∆) for
∆ = X − Xs is (A − BB T Xs )-invariant; continue to argue in coordinates such
that ∆ = diag(0, ∆2 ) with invertible ∆2 .
b) Conversely, if the ARE solution Xs satisfies (7.67), show that Xs is strong.
Hint: Use the difference trick and with a modal decomposition of A − BB T Xs
to construct a solution of the ARE that contradicts (7.67).
c) Show that Xs is unique.
d) If the ARI has a solution, show that the ARE has a strong solution.

213
 e) If Q ≥ 0 show that the ARE has a strong solution and that it is positive
semi-definite.

7) Consider ẋ = Ax+B1 w +B2 u, z = C1 x+D12 u. Suppose that (A, B2 ) is stabilizable,

that (A, C1 ) has no unobservable modes in C0 and that
   
T
D12 C1 D12 = 0 I .

Show that there exists some F with eig(A + B2 F ) ⊂ C− and k(C1 + D12 F )(sI −
A − B2 F )−1 B1 k∞ < γ if and only if there exists X ≥ 0 with

AT X + XA + X(γ −2 B1 B1T − B2 B2T )X + C1T C1 = 0,

eig(A + (γ −2 B1 B1T − B2 B2T )X) ⊂ C− .

For “only if” display the C− -unobservable modes of (A, C1 ) in a normal form.

8) Consider ẋ = Ax + B1 w + B2 u, z = C1 x + D12 u. Suppose that (A, B2 ) is stabilizable

and that D12 and
 
A − λI B2
  have full column rank for all λ ∈ C. (7.68)
C1 D12

Show that there exists some F with eig(A + B2 F ) ⊂ C− and k(C1 + D12 F )(sI −
A − B2 F )−1 B1 k∞ < γ if and only if the ARE

AT X +XA+γ −2 XB1 B1T X −(XB2 +C1T D12 )(D12

T
D12 )−1 (B2T X +D12
T
C1 )+C1T C1 = 0

has a solution X ≥ 0 with

eig(A + γ −2 B1 B1T X − B2 (D12

T
D12 )−1 (B2T X + D12
T
C1 )) ⊂ C− .

Can (7.68) be relaxed to requiring the rank condition on C0 only?

9) View Y+ (γ) in Theorem 7.9 as a function of γ > 0 on (γmin , ∞), where γmin is the
infimal γ > 0 such that the anti-stabilizing solution of the ARE exists. Let γopt be
the optimal value of the state-feedback H∞ -synthesis problem and let B1 6= 0.
a) Show that γmin = kC1 (sI + A + C1T C1 Y )−1 B1 k∞ > 0 with Y := Y+ (∞).
b) Show that Y+ (·) is non-decreasing and smooth on its domain of definition.
c) If Y+ (γmin ) > 0 show that the optimum is attained.
d) If Y+ (γmin ) 6> 0 show that Y+ (γopt ) ≥ 0 and Y+ (γopt ) 6> 0.
e) Under the latter hypothesis, any infimal sequence Fν for the optimal state-
feedback H∞ -problem satisfies kFν k → ∞ (i.e. is high-gain).
Hint: If not, assume (after taking a subsequence) that Fν converges to some
F∗ . Show that eig(A + B2 F∗ ) ⊂ C− ∪ C0 , that (sI − A − B2 F∗ )−1 B1 is stable
and that k(C + D12 F∗ )(sI − A − B2 F∗ )−1 B1 k∞ = γopt . Modify F∗ to stabilize
A + B2 F∗ without changing the norm.

214
10) Let P be a generalized plant with the usual state-space realization satisfy (OF1),
T
(OF3) with D12 , D21 of full column rank. Set Dc = (D12 T
D12 )−1 D12
T
and Dr =
T T −1
D21 (D21 D21 ) .
a) Show that there exists some K stabilizing P with kP ? Kk∞ < 1 if and only if
there exist symmetric X and Y with X > 0, Y − X −1 > 0 and

(A − B2 Dc C1 )Y + Y (A − B2 Dc C1 )T + Y C1T (I − D12 Dc )C1 Y + B1 B1T −

T
− B2 (D12 D12 )−1 B2T < 0

(A − B1 Dr C2 )T X + X(A − B1 Dr C2 ) + XB1T (I − Dr D21 )B1 X + C1 C1T −

T −1 T
− C2 (D21 D21 ) C2 < 0.

b) Let A be Hurwitz and Q = QT . For any α ∈ R show that AT X + XA + Q < 0

has a solution X = X T with X > αI.
−1
c) Let D21 be square and A − B1 D21 C2 be Hurwitz. Show that there exists some
K stabilizing P with kP ? Kk∞ < 1 if and only if there exists some Y > 0 with

(A − B2 Dc C1 )Y + Y (A − B2 Dc C1 )T + Y C1T (I − D12 Dc )C1 Y + B1 B1T −

T
− B2 (D12 D12 )−1 B2T < 0
   
Ag Bg Ah Bh
11) Let the realization of the SISO stable g =  , h =   be minimal
Cg 1 Ch 0
and suppose g has only zeros in C+ that are simple; they are denoted as z1 , . . . , zk .
For γ > 0 consider the following model-matching problem:

There exists f ∈ RH∞ with kh − gf k∞ < 1. (7.69)

Provide a solution along the following lines:

a) Set up a generalized plant to transform this problem to a standard H∞ -problem
and give a solution with one Lyapunov inequality.
b) If X satisfies (Ag − Bg Cg )X − XAh − Bg Ch = 0, show that (7.69) holds if and
only if the solution of the following Lyapunov equation is positive definite:

(Ag − Bg Cg )Y + Y (Ag − Bg Cg )T = Bg BgT − XBh BhT X T .

Hint: You can use X to transfer the “A”-matrix of the Lyapunov inequality
into block-diagonal form. Then one goes from the Lyapunov inequality to the
equation.
c) Show that w.l.o.g. it is possible to assume Ag − Bg Cg = diag(z1 , . . . , zk ) and
that Bg is the all-one vector.

215
d) With the standard unit vector ej show that eTj XBh = h(zj ). Derive a formula
for eTj Y ek .
e) Show that (7.69) holds iff the matrix with entries

1 h(zj )h(zk )
−
zj + zk zj + zk

is positive definite.

216
8 Robust Performance Synthesis

In the last Section we have dealt with designing a controller that achieves nominal per-
formance. The related problem of minimizing the H∞ -norm of the controlled system has
been the subject of intensive research in the 1980’s which culminated in the very elegant
solution of this problem in terms of Riccati equations.

In view of our analysis results in Section 6, the design of controllers that achieve robust
stability and robust performance amounts to minimizing the SSV (with respect to a
specific structure) of the controlled system over all frequencies. Although this problem
has received considerable attention in the literature, the related optimization problem
has, until today, not found any satisfactory algorithmic solution.

Instead, a rather heuristic method has been suggested how to attack the robust per-
formance design problem which carries the name D/K-iteration or scalings/controller-
iteration. Although it cannot be theoretically justified that this technique does lead to
locally or even globally optimal controllers, it has turned out pretty successful in some
practical applications. This is reason enough to describe in this section the pretty simple
ideas behind this approach.

8.1 Problem Formulation

We assume that we have built the same set-up us for testing robust performance in Section
6: After fixing the performance signals, pulling out the uncertainties, and including all
required uncertainty and performance weightings, one arrives at the controlled uncertain
system as described by
      
z∆ w∆ P11 P12 P13 w∆
      
 z  = P  w  =  P21 P22 P23   w  , u = Ky, w∆ = ∆z∆ , ∆ ∈ ∆
      
y u P31 P32 P33 u

under the Hypotheses 6.1. Let us again use the notation P∆ := S(∆, P ). The goal in this
section is to design a controller K which stabilizes P∆ and which leads to

kS(P∆ , K)k ≤ 1

for all uncertainties ∆ ∈ ∆.

In order to formulate again the robust performance analysis test, let us recall the ex-
tended block-structure
  
 ∆ 0 
c
∆e :=  ˆ c ∈ Cp2 ×q2 , k∆
 : ∆c ∈ ∆c , ∆ ˆ ck < 1
 0 ∆ ˆc 

217
where p2 /q2 is the number of components of w/z respectively.

Then K achieves robust performance if it stabilizes the nominal system P and if it renders
the inequality
µ∆e (S(P, K)(iω)) ≤ 1 for all ω ∈ R ∪ {∞} (8.1)
satisfied.

Finding a K which achieves (8.1) cannot be done directly since not even the SSV itself
can be computed directly. The main idea is to achieve (8.1) by guaranteeing that a
computable upper bound on the SSV is smaller than one for all frequencies. Let us recall
that the set of scalings D that corresponds to ∆c for computing an upper bound is given
by all  
D1 0

 ... 

 
 

 Dnr 

 

 Dn r +1 

D= ..
.
 

 
D
 
 nr +nc 
 

 d 1 I 


 . ..


 
0 dnf I
where Dj are Hermitian positive definite matrices and dj are real positive numbers. The
class of scalings that corresponds to ∆e is then defined as
  
 D 0 
De :=  >0|D∈D .
 0 I 

Remarks. With choosing this class of scalings, we recall that we ignore the fact that the
uncertainty structure comprises real blocks. Moreover, the scaling block that corresponds
to the full block included for the performance channel in the extension is set (without loss
of generality) equal to the identity matrix.

With these class of scalings we have

µ∆e (S(P, K)(iω)) ≤ inf kD−1 S(P, K)(iω)Dk.

D∈De

We conclude that any stabilizing controller which achieves

inf kD−1 S(P, K)(iω)Dk ≤ 1 for all ω ∈ R ∪ {∞} (8.2)

D∈De

218
also guarantees the desired inequality (8.1). Hence instead of designing a controller that
reduces the SSV directly, we design one that minimizes the upper bound of the SSV which
is obtained with the class of scalings De : SSV design is, actually, upper bound design!

Let us slightly re-formulate (8.2) equivalently as follows: There exists a frequency depen-
dent scaling D(ω) ∈ D e such that

kD(ω)−1 S(P, K)(iω)D(ω)k < 1 for all ω ∈ R ∪ {∞}.

This leads us to the precise formulation of the problem that we intend to solve.

Robust performance synthesis problem. Minimize

sup kD(ω)−1 S(P, K)(iω)D(ω)k (8.3)

ω∈R∪{∞}

over all controllers K that stabilize P , and over all frequency dependent scalings D(ω)
with values in the set De .

If the minimal value that can be achieved is smaller than one, we are done: We guarantee
(8.2) and hence also (8.1).

If the minimal value is larger than one, the procedure fails. Since we only consider
the upper bound, it might still be possible to push the SSV below one by a suitable
controller choice. Hence we cannot draw a definitive conclusion in this case. In practice,
one concludes that robust performance cannot be achieved, and one tries to adjust the
weightings in order to still be able to push the upper bound on the SSV below one.

8.2 The Scalings/Controller Iteration

Unfortunately, it is still not possible to minimize (8.3) over the controller K and the
frequency dependent scalings D(ω) together. Therefore, it has been suggested to iterate
the following two steps: 1) Fix the scaling function D(ω) and minimize (8.3) over all
stabilizing controllers. 2) Fix the stabilizing controller K and minimize (8.3) over all
scaling functions D(ω).

This procedure is called D/K-iteration and we use the more appropriate terminology
scalings/controller iteration. It does not guarantee that we really reach a local or even a
global minimum of (8.3). Nevertheless, in each step of this procedure the value of (8.3)
is reduced. If it can be rendered smaller than one, we can stop since the desired goal is
achieved. Instead, one could proceed until one cannot reduce the value of (8.3) by any of
the two steps.

Let us now turn to the details of this iteration.

219
First Step. Set
D1 (ω) = I
and minimize
sup kD1 (ω)−1 S(P, K)(iω)D1 (ω)k = kS(P, K)k∞
ω∈R∪{∞}

over all K that stabilize P . This is nothing but a standard H∞ problem! Let the optimal
value be smaller than γ1 , and let the controller K1 achieve this bound.

After Step k we have found a scaling function Dk (ω) and a controller Kk that stabilizes
P and which renders

sup kDk (ω)−1 S(P, Kk )(iω)Dk (ω)k < γk (8.4)

ω∈R∪{∞}

for some bound γk satisfied.

Scalings optimization to determine Dk+1 (ω). Given Kk , perform an SSV robust

performance analysis test. This amounts to calculating at each frequency ω ∈ R ∪ {∞}
the upper bound
inf kD−1 S(P, Kk )(iω)Dk (8.5)
D∈D e

on the SSV. Typical algorithms also return an (almost) optimal scaling Dk+1 (ω). This
step leads to a scaling function Dk+1 (ω) such that

sup kDk+1 (ω)−1 S(P, Kk )(iω)Dk+1 (ω)k < γ̂k

ω∈R∪{∞}

for some new bound γ̂k .

Controller optimization to determine Kk+1 . We cannot optimize (8.3) over K for

an arbitrary scaling function D(ω) = Dk+1 (ω). This is the reason why one first has to fit
this scaling function by a real rational D̂(s) that is proper and stable, that has a proper
and stable inverse, and that is chosen close to D(ω) in the following sense: For a (small)
error bound  > 0, it satisfies

kDk+1 (ω) − D̂(iω)k ≤  for all ω ∈ R ∪ {∞}.

With D̂, one then solves the H∞ -control problem

inf kD̂−1 S(P, K)D̂k∞

K stabilizes P

to find an almost optimal controller Kk+1 . This step leads to a Kk+1 such that

sup kDk+1 (ω)−1 S(P, Kk )(iω)Dk+1 (ω)k < γk+1 (8.6)

ω∈R∪{∞}

holds for some new bound γk+1 .

220
We have arrived at (8.4) for k → k + 1 and can iterate.

Let us now analyze the improvements that can be gained during this iteration. During
the scalings iteration, we are guaranteed that the new bound γˆk can be chosen with
γ̂k < γk .
However, it might happened that the value of (8.5) cannot be made significantly smaller
than γk at some frequency. Then the new bound γ̂k is close to γk and the algorithm is
stopped. In the other case, γ̂k is significantly smaller than γk and the algorithm proceeds.

During the controller iteration, one has to perform an approximation of the scaling func-
tion Dk+1 (ω) by D̂(iω) uniformly over all frequencies with a real rational D̂(s). If the
approximation error is small, we infer that
−1
sup kD̂−1 (iω)S(P, K)(iω)D̂(iω)k ≈ sup kDk+1 (ω)S(P, K)(iω)Dk+1 (ω)k
ω∈R∪{∞} ω∈R∪{∞}

for both K = Kk and K = Kk+1 . For a sufficiently good approximation, we can hence
infer from (8.6) that
sup kD̂−1 (iω)S(P, Kk )(iω)D̂(iω)k < γ̂k .
ω∈R∪{∞}

Since Kk+1 is obtained by solving an H∞ -optimization problem, it can be chosen with

kD̂−1 S(P, Kk+1 )D̂k∞ ≤ kD̂−1 S(P, Kk )D̂k∞ .
This implies
sup kD̂−1 (iω)S(P, Kk+1 )(iω)D̂(iω)k < γ̂k .
ω∈R∪{∞}

Again, if the approximation of the scalings is good enough, this leads to

−1
sup kDk+1 (ω)S(P, Kk+1 )(iω)Dk+1 (ω)k < γ̂k .
ω∈R∪{∞}

Hence the bound γk+1 can be taken smaller than γ̂k . Therefore, we can conclude
γk+1 < γk .
Note that this inequality requires a good approximation of the scalings. In practice it
can very well happen that the new bound γk+1 cannot be made smaller than the previous
bound γk ! In any case, if γk+1 can be rendered significantly smaller than γk , the algorithm
proceeds, and in the other cases it stops.

Under ideal circumstances, we observe that the sequence of bounds γk is monotonically

decreasing and, since bounded from below by zero, hence convergent. This is the only
convergence conclusion that can be drawn. There are no general implications about
the convergence of Kk or of the scaling functions Dk (ω), and no conclusions can be drawn
about optimality of the limiting value of γk .

Remarks.

221
 • The µ-tools support the fitting of D(ω) with rational transfer matrices D̂(iω). This
is done with GUI support on an element by element basis of the function D(ω),
where the user has control over the McMillan degree of the rational fitting function.
In addition, automatic fitting routines are available as well.

• Since D̂ and D̂−1 are both proper and stable, minimizing kD̂−1 S(P, K)D̂k∞ over K
amounts to solving a standard weighted H∞ -problem. The McMillan degree of an
(almost) optimal controller is given by

2 ∗ McMillan degree of D̂ + McMillan degree of P.

Keeping the order of K small requires to keep the order of D̂ small. Note that this
might be only possible at the expense of an large approximation error during the
fitting of the scalings. Again, no general rules can be given here and it remains to
the user to find a good trade-off between these two aspects.

• If the order of the controller is too large, one should perform a reduction along the
lines as described in the chapters 7 and 9 in [ZDG] and chapter 11 in [SP].

To learn about the practice of applying the scalings/controller we recommend to run the
himat-demo around the design of a pitch axis controller for the simple model of an airplane
which is included in the µ-Toolbox. This examples comprises a detailed description of
how to apply the commands for the controller/scalings iteration. Moreover, it compares
the resulting µ-controller with a simple loop-shaping controller and reveals the benefits
of a robust design.

Exercise

1) Consider the following simplified model of a distillation column

 
1 87.8 −86.4
G(s) =  
75s + 1 108.2 −109.6

(with time expressed in minutes) in the tracking configuration of Figure 61. The
uncertainty weighting is given as
 
s+0.1
0
W (s) =  0.5s+1 
s+0.1
0 0.5s+1

and the real rational proper and stable ∆ with k∆k∞ < 1 is structured as
 
∆1 (s) 0
∆(s) =  .
0 ∆2 (s)

222
 W ∆

r e u
+ K + G

−1

Figure 61: Tracking interconnection of simplified distillation column.

a) Provide an interpretation of this uncertainty structure, and discuss how the

uncertainty varies with frequency.
b) Choose the decoupling controllers

β
Kβ (s) = G(s)−1 , β ∈ {0.1, 0.3, 0.5, 0.7}
s
Discuss in terms of step responses in how far these controller lead to a good
closed-loop response! Now choose the performance weightings
s+β
Wp (s) = α I, α ∈ [0, 1], β ∈ {0.1, 0.3, 0.5, 0.7}
s + 10−6
with a small perturbation to render the pure intergrator stable. Test robust
stability, nominal performance, and robust performance of Kβ for α = 1 and
discuss.
Reveal that Kβ does not lead to good robust performance in the time-domain
by determining a simple real uncertainty which leads to bad step responses in
r → e.
c) Perform a two step scalings/controller iteration to design a controller that en-
forces robust peformance, possibly by varying α to render the SSV for robust
performance of the controlled system close to one. Discuss the resulting con-
troller in the frequency-domain and the time-domain, and compare it with the
decoupling controllers Kβ . Show that this controller exhibits a much better
time-response for the ‘bad’ disturbance constructed in the previous exercise.

Remark. You should make a sensible selection of all the plots that you might
collect during performing this exercise in order to nicely display in a non-repetitive
fashion all the relevant aspects of your design!

223
9 Youla Jabr Bongiorno Kucera Parametrizations

Our main goal in this section is to parameterize all LTI controller K that stabilize a
given LTI generalized plant P . Theorem 9.16 will show that under some hypothesis K
stabilizes P iff there exist some stable Q with K = J ? Q. A state-space realization of
J is given in Hypothesis 9.15. To develop this result, we will start from a more general
setting. More precisely we will replace the set of stable systems RH∞ by an integral
domain R with unit 1 and replace the set of systems R(s), the real rational functions, by
F , the field of fractions corresponding to R. We introduce in Section 9.1 the algebraic
framework for such general systems and develop algebraic stabilization theory in Section
9.2. Afterwards the Youla parametrization (Theorem 9.13) allows us to parametrize all
stabilizing controllers for general systems and the double Bézout identity for LTI systems
in Section 9.3 allows us to give explicit state space realizations of the systems in Theorem
9.13 in the LTI case. This will finally lead to Theorem 9.16.

9.1 Algebraic Framework

Let R be an integral domain with unit 1. An integral domain is a nonzero commutative

ring in which the product of any two nonzero elements is nonzero. Moreover let F be the
corresponding field of fractions, which is the smallest field in which R can be embedded.
Sometimes F is denoted by Quot(R) or Frac(R). We denote the set of matrices with
entries in R by M (R) or Rm×n , if we want to specify the dimension of the matrices.
Analogously we denote the set of matrices with entries in F by M (F ) or F m×n . The
identity matrix is denoted as I. If dimensions are not specified and if multiplying or
adding matrices, dimensions are assumed to be compatible. In the sequel, elements in
M (F ) are called systems and those in M (R) are called stable systems.

Example 9.1

• With R = RH∞ and F = R(s), the real-rational functions, the results in this section
can be spezialized to what we are familiar with.

• R = H∞ , the Banach algebra of bounded holomorphic functions in C> . This e.g.

allows to handle systems with multiple delays.

• R equals the set of real-rational functions with all its poles in S ⊂ C ∪ {∞}. If
S = {∞}, then R is the set of all real polynomials which are not constant.

• R equals the set of n/d with n ∈ R[z1 , . . . , zm ] and with d ∈ R[z1 , . . . , zm ] \ {0}
having all zeros in Cn \ Dn where Dm = {(z1 , . . . , zm ) ∈ Cm : |zi | ≤ 1}. This allows
to handle stable multi-dimensional discrete-time systems.

224
In order to develop the algebraic stabilization theory we need to make use of right and
left coprime factorizations of systems.

Definition 9.2 G ∈ M (F ) has a right factorization (over R) if there exist M, N ∈

M (R) such that M is square and invertible in M (F ) with G = N M −1 . The factorization
is coprime if there exist X, Y ∈ M (R) with XM + Y N = I.

Any G ∈ F m×n defines a linear map G : F n → F m . If G has a right factorization observe

that y = Gu iff y = N ξ and u = M ξ for some ξ ∈ F n . Hence we infer
    
 y  N
  : y = Gu, u ∈ F m = im   .
 u  M

This is a so-called image representation (of the graph) of the system G.

The advantage of stability of the factors emerges if considering “bounded input pairs”
(y, u) ∈ Rm × Rn with y = Gu only: all pairs (N ξ, M ξ) for ξ ∈ Rn clearly have this
property. Coprimeness implies that we do indeed reach all such pairs in this fashion.

Lemma 9.3 If G = N M −1 is a right coprime factorization (rcf ) then

     
 y   N 
  : u ∈ Rn , y = Gu ∈ Rm =   ξ : ξ ∈ Rm .
 u   M 

Proof. “⊃”: If y = N ξ and u = M ξ for ξ ∈ Rn , we infer y ∈ Rm , u ∈ Rn and ξ = M −1 u

implying y = N M −1 u.

“⊂”: If y = Gu = N M −1 u for y ∈ Rm , u ∈ Rn , let ξ := M −1 u to get

   
y N
  =   ξ.
u M
 
N  
Co-primeness just means that   has a left-inverse L = X Y ∈ M (R). We infer
M
   
y N
Rn 3 L   =   ξ = ξ
u M

and thus ξ ∈ Rn . (Hence ξ must be “bounded”).

It is clear how to define the dual notion of Definition 9.2.

225
Definition 9.4 G ∈ M (F ) has a left factorization (over R) if there exist M, N ∈
M (R) such that M is square
 and invertible in M (F ) with G = M −1 N . The factorization
is coprime if M N has a right-inverse in M (R).

Left-factorizations give rise to so-called kernel representations:

  
 y   
  : y = Gu, u ∈ F m = ker M −N .
 u 

Example 9.5 Let G be a transfer matrix with realization ẋ = Ax + Bu, y = Cx + Du.

1) If F renders A + BF Hurwitz then the pre-compensation u = F x + ξ leads to a

right factorization: Indeed we obtain

ẋ = (A + BF )x + Bξ
y = (C + DF )x + Dξ
u = F x + Iξ.

This motivates to define

 
  A + BF B
N  
  =  C + DF D  ∈ RH∞ . (9.1)
M
 
F I
 
A B
Since I is invertible we get M −1 =   and hence
−F I
   
A + BF −BF B A BF B  
A B
N M −1 = 
   
 =  0 A + BF 0  =
0 B A     = G.
C D

C + DF −DF D C −DF D
 
I 0
The second equation follows by transforming the system with T =  .
−I I

2) 
The factorization
 in 1) is however in general not coprime. To see this let G =
0 1
  = 1. Clearly F = −1 renders A − BF Hurwitz and
0 1
   
0 1 −1 1
M −1 =   = s + 1 and N =  = s .
1 1 s −1 1 s+1

226
 Now suppose that the factorization is coprime. Choose u = 1 ∈ RH∞ to get
s
y = Gu = 1 ∈ RH∞ . By Lemma 9.3 there exists ξ ∈ RH∞ with 1 = u = N ξ = s+1 .
s+1
Hence ξ = s 6∈ RH∞ .

3) If A is Hurwitz the factorization in 1) is coprime since X = M −1 ∈ RH∞ and

Y = 0 ∈ RH∞ implies XM + Y N = I.

4) If L renders A + LC Hurwitz, we get analogously to 1) a left factorization as

 
  A + LC L B + LD
M̃ Ñ =  .
C I D

If in addition F renders A + BF Hurwitz, we will see in Section 9.3 that the above
left factorization and the right factorization (9.1) are coprime.

Definition 9.6 Q ∈ M (R) is called unimodular or a unit if Q−1 exists and satisfies
Q−1 ∈ M (R).

The next result shows that rcf’s only differ by unimodular right factors.

Lemma 9.7 If N1 M1−1 = N2 M2−1 are two right coprime factorizations then there exists
a unimodular Q with M1 = M2 Q and N1 = N2 Q.

Proof. If N1 M1−1 = N2 M2−1 are two coprime factorizations, we infer

       
M M M M
 1  Q =  2  and  2  Q−1 =  1 
N1 N2 N2 N1

for Q := M1−1 M2 . Left-multiplying with the stable left-inverses L1 and L2 corresponding

to the two factorizations implies
   
M2 M
Q = L1   ∈ M (R) and Q−1 = L2  1  ∈ M (R).
N2 N1

Of course, dually, left coprime factorizations just differ by unimodular left factors.

In RH∞ , exactly the proper and stable transfer matrices with a proper and stable inverse
are the units. Such transfer matrices are often called outer or minimum-phase.

In this context you should keep in mind the topological consequences of the small-gain
theorem. For example, if M ∈ RH∞ and kM k∞ < 1 then I − M is a unit.

227
Lemma 9.8 Let A, B ∈ M (F ) and let AB be a unit.

• If A ∈ M (R) and B is invertible then B −1 is stable.

• If B ∈ M (R) and A is invertible then A−1 is stable

Proof. If AB = Q is a unit then Q−1 A ∈ M (R) is a right inverse of B and since B is

invertible we get Q−1 A = B −1 . This shows B −1 ∈ M (R). The second statement follows
analogously.

9.2 Stabilizable Plants and Controller Parametrization

We now define stabilization of a system G by a controller K in (almost) complete analogy

to Section 3.

Definition 9.9 The system G ∈ M (F ) is stabilized by K ∈ M (F ) if

 −1
I −K
  ∈ M (R).
−G I

Precisely it is meant that the inverse exists in M (F ) and is stable.

Let us assume that G admits the rcf G = N M −1 . Then trivially

 −1   −1
I −K M 0 M −K
  =   . (9.2)
−G I 0 I −N I

Let us first see that stabilizing controllers are characterized by stability of the inverse on
the right.

Lemma 9.10 If G = N M −1 is a rcf then K stabilizes G iff

 −1
M −K
  ∈ M (R). (9.3)
−N I

Proof. “If” is trivial by (9.2) since M ∈ M (R) . For “only if”, XM + Y N = I implies
     
X Y M 0 0 −Y M −K
  +   = I.
0 I 0 I 0 0 −N I

Right-multiplying (9.3) shows by (9.2) that the inverse is stable.

228
Theorem 9.11 G ∈ M (F ) with the rcf N M −1 is stabilizable iff there exist X̃, Ỹ ∈ M (R)
such that X̃M − Ỹ N = I and X̃ is invertible.

Proof. To show “only if”, let K stabilize G and define

   −1
X̃ Ỹ M −K
  :=   ∈ M (R). (9.4)
Ñ M̃ −N I

This implies     
X̃ Ỹ M −K I 0
  = . (9.5)
0 I −N I −N I

Hence X̃M − Ỹ N = I and X̃ is invertible, since the r.h.s. is invertible. We also infer
K = X̃ −1 Ỹ and that this is a lcf.

For “if”, let X̃, Ỹ be as described. Then K := X̃ −1 Ỹ implies (9.5). Since the left factor
!
M −K
on the left is stable and the r.h.s. is a unit, Lemma 9.8 implies stability of .
−N I
By Lemma 9.10 this implies that K stabilizes G.

The proof shows that stabilizing controllers admit a lcf X̃ −1 Ỹ with factors as in Theorem
9.11. Moreover if X̃, Ỹ are taken as in Theorem 9.11 then K = X̃ −1 Ỹ stabilizes G. Let’s
see what happens if G admits both a rcf and a lcf.

Lemma 9.12 Let G have both the rcf N M −1 and the lcf M̃ −1 Ñ and X̃, Ỹ ∈ M (R) satisfy
X̃M − Ỹ N = I. Then
      
X̃ Ỹ M I X̃ Ỹ
   =   and   is a unit.
Ñ M̃ −N 0 Ñ M̃

Proof. The equation is a consequence of M̃ −1 Ñ − N M −1 = 0. If we take stable X, Y

with M̃ X − Ñ Y = I, we obtain
    
X̃ Ỹ M −Y I Ỹ X − X̃Y
  = .
Ñ M̃ −N X 0 I

All these matrices are stable and the r.h.s. is obviously even a unit. Since the r.h.s is
invertible the same holds for both matrices on the l.h.s. By Lemma 9.8, all matrices are
units and the claim follows.

Note that since G has the rcf N M −1 there clearly always exist X̃ Ỹ ∈ M (R) with
X̃M − Ỹ N = I.

229
Theorem 9.13 (Youla parametrization) Let G = N M −1 = M̃ −1 Ñ be a rcf and lcf of G
and choose X̃, Ỹ ∈ M (R) with X̃M − Ỹ N = I. Then K stabilizes G iff

K = (X̃ + QÑ )−1 (Ỹ + QM̃ ) (9.6)

for some Youla parameter Q ∈ M (R) such that X̃ + QÑ is invertible.

For a controller described as in (9.6), Q enters affinely in

 −1    
I −K M X̃ M Ỹ M  
  =  +   Q Ñ M̃ . (9.7)
−G I N X̃ N Ỹ +I N

Note that (9.6) is a lcf of K. If K0 stabilizes G = N M −1 , we have seen in Theorem

9.11 that it admits a lcf K0 = X̃ −1 Ỹ with stable X̃, Ỹ such that X̃M − Ỹ N = I and
X̃ is invertible. Therefore, the parametrization in Theorem 9.13 can be based on any
stabilizing controller K0 for G, and K0 is then reobtained back from (9.6) with Q = 0.
!
I Q
Proof. “ ⇐ ”: left-multiply the equation in Lemma 9.12 with ∈ M (R) to get
0 I
   
X̃ + QÑ Ỹ + QM̃ M I
   =  .
Ñ M̃ −N 0

Hence (9.6) is a lcf of K. Moreover we trivially infer

    
X̃ + QÑ Ỹ + QM̃ M −K I 0
  = .
0 I −N I −N I

By Lemma 9.8 and 9.10 we conclude that K stabilizes G. Right-multiplying the inverse
of the right-factor and using (9.2), the last equation clearly implies
    −1
M 0 X̃ + QÑ Ỹ + QM̃ I −K
  =  .
N I 0 I −G I

This is the claimed expression for the inverse.

“⇒”: Let K̃ be any stabilizing controller. Lemma 9.12 implies

    
X̃ Ỹ M −K̃ I S
  = 
Ñ M̃ −N I 0 T

for some S and T . Moreover, since both matrices on the left-hand side are invertible
the right-hand side is invertible. By left-multiplying the inverse of the r.h.s., we get with

230
Q := −ST −1 that
   −1
X̃ + QÑ Ỹ + QM̃ M −K̃
 =  ∈ M (R). (9.8)
−1 −1
T Ñ T M̃ −N I

This implies that X̃ + QÑ and Ỹ + QM̃ are stable. As in the proof of Theorem 9.11, we
also infer that X̃ + QÑ is invertible and that we have K̃ = (X̃ + QÑ )−1 (Ỹ + QM̃ ). (9.8)
even shows that this is a lcf.

Stability of Q itself follows from Lemma 9.12 and

    
I Q X̃ Ỹ X̃ + QÑ Ỹ + QM̃
  =  ∈ M (R).
0 I Ñ M̃ Ñ M̃

9.3 Double Bézout Identity for LTI Systems

If G ∈ R(s) is proper, the ingredients to parametrize all stabilizing controllers can be

obtained by state-space computations.

For this purpose we choose a stabilizable and detectable realization

 
A B
G= .
C D

If F , L are taken with eig(A + BF ) ⊂ C< , eig(A + LC) ⊂ C< then

 
(A + BF ) + L(C + DF ) −L
K= 
F 0

is an observer-based stabilizing controller for G. It is now not difficult to construct

matrices in RH∞ which satisfy the double Bézout identity
  
X̃ −Ỹ M Y
  =I (9.9)
−Ñ M̃ N X

and G = N M −1 = M̃ −1 Ñ (as well as K = X̃ −1 Ỹ = Y X −1 ).

231
Introduce the abbreviations AF := A + BF , CF = C + DF as well as AL := A + LC,
BL := B + LD. Then
 −1
 −1  A 0 B 0

I −K  0 (A + BF ) + L(C + DF ) 0 −L 
  =   =

−G I  0
 −F I 0 
−C 0 −D I
   
A BF B 0 AF BF B 0
   
 −LC A + BF + LC −LD −L   0 A −B −L 
L L
= = =
   
 0 F I 0    F F I 0 

 
C DF D I CF DF D I
  
AF B 0 AL −BL −L   
   M 0 X̃ Ỹ
= F I 0  F
 I 0  =:
  .
N I 0 I
CF D I 0 0 I
In complete analogy but with another coordinate change:
 
 −1  A BF B 0

I −K  −LC A + BF + LC −LD −L 
  = 
=

−G I  0
 F I 0 

C DF D I
  
AL 0 BL L AF −LC −LD −L
  
 −LC A −LD −L   0 A BL L 
F L
= =
  

 0 F I 0  F 0 I 0 
  
C CF D I CF C D I
  
AF 0 −L AL BL L   
   I Y I 0
= F I 0   0 I 0  =:   .
0 X Ñ M̃
  
CF 0 I C D I

Note that, by their respective definitions, M , X̃ and M̃ , X have proper inverses.

Hence        
M   0 0 Y   I 0
  X̃ Ỹ + =  Ñ M̃ + 
N 0 I X 0 0
which implies      
M   Y   I 0
  X̃ Ỹ −  Ñ M̃ = 
N X 0 −I

232
and thus   
M Y X̃ −Ỹ
   = I.
N X −Ñ M̃
We can also get to the equivalent version
  
M −Y X̃ Ỹ
   = I.
−N X Ñ M̃

Note again that, by the definition of M , N , Ñ , M̃ , X, Y , X̃ and Ỹ both matrices on the

l.h.s. have proper inverses. Hence (9.9) holds and in particular X̃M − Ỹ N = I. These
(or those with reversed order) are called double Bézout identities.

From    
I −K M 0 X̃ Ỹ
   =I
−G I N I 0 I

we infer (−GM + N )X̃ = 0 and thus

G = N M −1 .

In a similar fashion one can extract K = X̃ −1 Ỹ . The double Bézout identity then implies
Ñ M − M̃ N = 0 and thus
G = M̃ −1 Ñ .
Hence the assumptions from Theorem 9.13 are satisfied and every controller KQ = (X̃ +
QÑ )−1 (Ỹ + QM̃ ) with Q ∈ M (R) such that X̃ + QÑ is invertible, stabilizes G. Moreover
we regain the controller K by choosing Q = 0.

Observe that the realizations in the Bézout identity can be compactly expressed as
   
  AF B −L   AL −BL L
M Y   X̃ −Ỹ  
  =  F I 0 ,  = F I 0 . (9.10)
N X −Ñ M̃
   
CF D I C −D I
Moreover, we can just read off that
 
     AF BF B 0
 
M X̃ M Ỹ M 0  0 A −B −L 
X̃ Ỹ
L L
 =  =
 .

(9.11)
N X̃ N Ỹ +I N I 0 I 
 F F I 0 

CF DF D I
We obtain explicit state-space formulas for all matrices in Theorem 9.13.

Note that these relations were obtained with hardly any computations, in contrast to
what you typically see in the literature. Variants (e.g. the extension to time-varying
state-space systems) are also easy to obtain.

233
9.4 Youla Parametrization for Generalized Plants

If P ∈ M (F ) we consider again the generalized plant

   
z w
 =P 
y u
and define the notion of stabilizing controllers as in Lecture 3.

Definition 9.14 K ∈ M (F ) stabilizes P ∈ M (F ) if I − P22 K is invertible and if

   
P11 P12 0 P12  
 0 I 0  +  I  K(I − P22 K)−1 P21 P22 I ∈ M (R).
   
   
P21 P22 I P22

To shorten the exposition, we assume that there exists a stabilizing controller for P , and
that K stabilizes P iff K stabilizes P22 . The Springer-book by Francis (1987) develops
algebraic tests for F = R(s) and R = RH∞ (as we did in state-space) for these properties.

We can thus parametrize all stabilizing controllers for P , by using the parametrization
for G := P22 in Theorem 9.13. With the notations from there we get

P ? K = P11 + P12 K(I − P22 K)−1 P21 =

 −1  
  I −K 0
= P11 + P12 I 0     P21 =
−P22 I I
= P11 + P12 [M Ỹ + M QM̃ ]P21
and hence obtain affine dependence of the controlled system on Q.

Our main goal is to obtain this parametrization for LTI generalized plants directly on the
basis of state-space descriptions.

Hypothesis 9.15 Let P be a transfer matrix with

 
A B1 B2
 
P =  1 11 12  , (A, B2 ) stabilizable , (A, C2 ) detectable.
C D D 

C2 D21 D22
Choose F , L that render A + B2 F and A + LC2 Hurwitz and define
 
   A + B2 F −B2 F B1 B2

T1 T2  0 A + LC2 B1 + LD21 0 
 =  

T3 0  C1 + D12 F −D12 F D 11 D 12

 
0 C2 D21 0

234
(in which the lower right transfer matrix indeed vanishes) as well as
 
A + B2 F + LC2 + LD22 F L −(B2 + LD22 )
 
J =  −F 0 I .

C2 + D22 F I −D22

Theorem 9.16 Under Hypotheses 9.15, the LTI controller K stabilizes P iff there exists
some Q ∈ RH∞ with det(I + D22 Q(∞)) 6= 0 such that

K = J ? Q.

The set of all closed-loop transfer matrices P ? K that are achievable by stabilizing con-
trollers is given as

{T1 + T2 QT3 : Q ∈ RH∞ , det(I + D22 Q(∞)) 6= 0}.

In many (not all!) situations we can assume w.l.o.g. that D22 = 0 by pushing this direct
feedthrough term to the controller. Then the set of stabilized controlled system admits
the beautiful affine parametrization

T1 + T2 QT3 with free Q ∈ RH∞ .

Remark. The parametrization of all sub-optimal H∞ in Theorem 7.40 is different: The

parameter does not enter the controlled system affinely!

Proof. We can apply Theorem 9.13 with the matrices (9.10) and (9.11) taken for
G = P22 . Note that (9.6) describes stabilizing controllers in M (R(s)), while we target at
a parametrization of all stabilizing controllers that are LTI.

If Q ∈ RH∞ and det(X̃(∞) + Q(∞)Ñ (∞)) = det(I + Q(∞)D22 ) 6= 0, then K in (9.6) is

obviously proper.

Conversely, if K is LTI and stabilizes P22 , we can express it as (9.6) with Q ∈ RH∞ such
that X̃ + QÑ has a rational inverse. Then we get, as in the proof of Theorem 9.13,
    
X̃ + QÑ Ỹ + QM̃ M −K I 0
  = ,
0 I −N I −N I

which is equivalent to
 −1   
X̃ + QÑ Ỹ + QM̃ M −K I 0
  =  .
0 I −N I N I

235
Hence the inverse of X̃ + QÑ is actually proper, which in turn implies 0 6= det(X̃(∞) +
Q(∞)Ñ (∞)) = det(I + Q(∞)D22 ).

The rest of the proof are straightforward computations.

If K is as in (9.6) then u = Ky iff (X̃ +QÑ )u = (Ỹ +QM̃ )y iff X̃u− Ỹ y = Q(−Ñ u+ M̃ y)
iff     
ŷ M̃ −Ñ y
 =    , û = Qŷ
û −Ỹ X̃ u

iff (use (9.10))

 
  A + LC2 L −B − LD22  
ŷ   y
 = C 2 I −D 22
   , û = Qŷ
û u
 
F 0 I

iff  
  A + LC2 + BF + LD22 F L −B − LD22  
ŷ   y
 = C2 + D22 F I −D22    , û = Qŷ.
u û
 
−F 0 I
Permuting the outputs shows u = (J ? Q)y and thus K = J ? Q.

The interconnection of     
ẋ A B1 B2 x
    
 z  = C D D w 
   1 11 12   
y C2 D21 D22 u
and     
ξ˙ A+B2 F +LC2 +LD22 F L −(B2 + LD22 ) ξ
    
u =  −F 0 I y
    
ŷ C2 + D22 F I −D22 û
can be obtained with the feedback u = ũ from
    
ẋ A 0 B1 0 B2 x
    
 ξ˙   LC A+B F +LC +LD F LD −(B +LD ) LD  ξ 
   2 2 2 22 21 2 22 22  
    
 z = C 0 D 0 D12  w  .
 
   1 11
    
 ŷ   C2
   C2 +D22 F D21 −D22 D22  û 
 
ũ 0 −F 0 I 0 u

236
Without tedious computations, the lower LFT clearly equals
 
   A −B 2 F B1 B2
 
z  LC 2 A + B2 F + LC 2 LD 21 −B  w
2
 =   
ŷ  C1
 −D 12 F D 11 D  û
12 

C2 C2 D21 0

which leads with a state-coordinate change to

 
   A + B2 F −B 2 F B 1 B2
 
z  0 A + LC2 B1 + LD21 0  w 
 =   .
ŷ  C1 + D12 F −D12 F D 11 D 
12  û

0 C2 D21 0

With T1 , T2 , T3 as defined above this is

    
z T T w
  =  1 2  .
ŷ T3 0 û

This indeed implies P ? J ? Q = P ? K = T1 + T2 QT3 as claimed.

Remarks

Note that the representations for T1 and T2 simplify to

   
A + B2 F B2 A + LC2 B1 + LD21
T2 =   and T3 =  .
C1 + D12 F D12 C2 D21

Also T1 can be expressed in the following ways:

   
A+B2 F −B2 F B1 A+B2 F −LC2 −LD21
   

 0 A+LC2 B1 +LD21  =
 
 0 A+LC 2 B1 +LD 21
=

C1 +D12 F −D12 F D11 C1 +D12 F C1 D11
   
A + LC2 B1 + LD21 A + B2 F B1
= T2  + =
−F 0 C1 + D21 F D11
   
A + B2 F −L A + LC2 B1 + LD21
=  T3 +  .
C1 + D12 F 0 C1 D11

Applications

There are very many ways to exploit the Youla parametrization:

237
 • Convex optimization for optimal feedback controller synthesis.

• Switching or interpolating controllers (also on-line for time-varying transitions) with

guarantees for stability and performance.

• Designing infinite dimensional controllers for finite-dimensional plants, such as delay

system and synthesis of Smith predictors.

• Reduction of control to functional analytic approximation problems for infinite di-

mensional plants or more general classes of systems.

• Solution of math problems by control techniques (Nehari, Nevanlinna-Pick,

Carathéodory-Fejér, ...)

Of course, employing Youla-techniques is often not straightforward.

Exercises

1) Consider the standard tracking configuration given in the figure below with a to-
be-controlled stable SISO system g and a to-be-designed controller k

r e u y
+ k g
−

a) Compute the transfer functions Tur and Tyr from r to u and from r to y.
b) The set V of all strictly proper rational functions without pole on the imaginary
axis is a vector space. Argue that the 2-norm of the frequency response of some
strictly proper f defined as
s Z
∞
1
kf k2 := |f (iω)|2 dω
2π −∞

comes from an inner product on V .

c) Show that the subspace of all strictly proper and stable rational functions,
denoted as RH2 , is orthogonal to the subspace of all strictly proper and anti
stable rational functions, denoted as RH2⊥ and argue that any strictly proper
rational function f without pole on the imaginary axis can be uniquely decom-
posed into the sum of fs ∈ RH2 and fu ∈ RH2⊥ as f = fs + fu .
d) Consider now the problem given in the figure below, which is called a model
matching problem: Given stable g, m (m strictly proper), design a strictly
proper controller k which internally stabilizes the closed loop system and min-
imizes the 2-norm of the transfer function Tzr from r to z, which is called the
matching error.

238
 m

y −
e u g
+ k +
r − z

Show that this problem can be reduced to

inf kgq − mk2 . (9.12)

q∈RH2

e) Give the optimal solution of the reduced and original problem, if g −1 is sta-
ble and proper, and argue that there exists almost optimal solutions q , k
dependinf on some positive  if g −1 is stable but not propber.
f) Suppose that g is stable and has no zeros on the imaginary axis and at infinity.
It is a fact that there exists a factorization g = gi go , where go and go−1 are
proper and stable and gi has magnitude 1 on the whole imaginary axis (such
s−α0
as e.g. s+α 0
for α0 > 0). Use this fact to reduce the problem (9.12) to the
equivalent problem
inf kgi x − mk2 . (9.13)
x∈RH2

g) Determine the transfer function r such that the optimization problem (9.13)
can be transformed to the equivalent problem

inf kx − rk2 .
x∈RH2

h) Use again the fact thet r can be decomposed into the sum of a strictly proper
stable rs and a strictly proper anti stable ru to design the optimal controller
kopt and to compute inf x∈RH2 kx − rk2 .

1) Suppose we have the structured system

      
y1 u1 P11 0 u
 =P =  1 
y2 u2 P21 P22 u2

with P22 being strictly proper. The aim of this exercise is that the set of all (lower)
triangular stabilizing controllers K can be parameterized as
 
Q11 0
K = Y11 + Y12 Q(I − Y22 Q)−1 Y21 with Q =   ∈ RH∞
Q21 Q22

for some (lower) triangular Y11 , Y12 , Y21 and Y22 .

239
a) Suppose P is stable and let K stabilize P . Show that Q = K(I − P K)−1 is a
Youla parameter for K. Moreover show that Q is triangular iff K is triangular.
b) Now choose a minimal realization
 
  A B1 B2
P11 0  
P =  =  C1 D11 0  .
P21 P22
 
C2 D21 D22

Show that due to C1 (sI − A)−1 B2 = 0 the controllable subspace of (A, B2 ) is

contained in the unobservable subspace of (A, C1 ). Show that we can hence
find a state coordinate change such that
 
  A11 0 0
A B2  
  =  A21 A22 B22  .
C1 0
 
C11 0 0

c) By b) we can realize P as
   
A11 0 B11 0 AK K
0 B11 0
   11 
A A B B   AK AK K K 
 21 22 21 22   21 22 B21 B22 
P =  and define K =  .
 C11 0 D11 0   CK K
0 D11 0 
   11 
K K K K
C21 C22 D21 D22 C21 C22 D21 D22

Show that K stabilizes P iff (A11 , B11 ), (A22 , B22 ) are stabilizable and
(A11 , C11 ), (A22 , C22 ) are detectable.
d) By c) we can find for P with minimal realization as in b) triangular F and L
such that A + BF and A + LC are Hurwitz. Define
 
  A + BF L −B
Y11 Y12  
Y =  =  −F 0 I .
Y21 Y22
 
C I 0

Show that all stabilizing controllers are given by K = Y ? Q with Q ∈ RH∞ .

e) Show that if Q ∈ RH∞ is triangular then K = Y ? Q is a triangular stabilizing
controller for P and show that if K is a triangular stabilizing controller for
P then Q = (I + ZY22 )−1 Z with Z = Y12−1 (K − Y11 )Y21−1 is the corresponding
Youla parameter and is triangular.

240
10 A Brief Summary and Outlook

We have seen that the generalized plant framework is very general and versatile to
capture a multitude of interconnection structures. This concept extends to much larger
classes of uncertainties, and it can be also generalized to some extent to non-linear systems
where it looses, however, its generality.

The structured singular value offered us computable tests for robust stability and
robust performance against structured linear time-invariant uncertainties. It is possible
to include parametric uncertainties which are, however, computationally more delicate.

Nominal controller synthesis, formulated as reducing the H∞ -norm of the perfor-

mance channel, found a very satisfactory and complete solution in the celebrated Riccati
approach to H∞ -control.

Robust controller synthesis is, from a theoretical point of view and despite intensive
efforts, still in its infancy. The scaling/controller iteration and variations thereof form - to
date - the only possibility to design robustly peforming controllers. It offers no guarantee
of optimality, but it has proven useful in practice.

These notes were solely concerned with LTI controller analysis/synthesis, with LTI un-
certainties, and with H∞ -norm bounds as performance specifications. Linear Matrix
Inequalities (LMIs) techniques offer the possibilities for extensions in the following direc-
tions:

• There is a much greater flexibility in choosing the performance specification, such as

taking into account H2 -norm constraints (stochastic noise reduction), positive real
conditions and amplitude constraints.

• The LMI framework allows the extension to multi-objective design in which the
controller is built to guarantee various performance specs on different channels.

• It is possible to include time-varying parametric, non-linear static and non-linear

dynamic uncertainties in the robustness tests. The uncertainties will be described in
this framework by Integral Quadratic Constraints (IQCs) which are generalizations
of the scalings techniques developed in these notes.

• Finally, LMI techniques allow to perform a systematic design of gain-scheduling

controller to attack certain class of non-linear control problems.

241
A Bisection

At several places we encountered the problem to compute a critical value γcritical , such
as in computing the upper bound for the SSV or the optimal value in the H∞ problem.
However, the algorithms we have developed only allowed to test whether a given number γ
satisfies γcritical < γ or not. How can we compute γcritical just by exploiting the possibility
to perform such a test?

The most simple technique is bisection. Fix a level of accuracy  > 0.

• Start with an interval [a1 , b1 ] such that a1 ≤ γcritical ≤ b1 .

• Suppose one has constructed [aj , bj ] with aj ≤ γcritical ≤ bj .

Then one tests whether
aj + b j
.
γcritical <
2
We assume that this test can be performed and leads to either yes or no as an
answer.
aj +bj
If the answer is yes, set [aj+1 , bj+1 ] = [aj , 2
].
aj +bj
If the answer is no, set [aj+1 , bj+1 ] = [ 2
, bj ].

• If bj+1 − aj+1 >  then proceed with the second step for j replaced by j + 1.
If bj+1 − aj+1 ≤  then stop with aj+1 ≤ γcritical ≤ aj+1 + .

Since the length of [aj+1 , bj+1 ] is just half the length of [aj , bj ], there clearly exists an index
for which the length of the interval is smaller than . Hence the algorithm always stops.
After the algorithm has stopped, we have calculated γcritical up to the absolute accuracy
.

B Proof of Theorem 7.1

Note that

AT X + XA + XRX + Q = (−A)T (−X) + (−X)(−A) + (−X)R(−X) + Q

and
A + RX = −((−A) + R(−X)).
Hence we can apply Theorem 7.2 to the ARE/ARI

(−A)T Y + Y (−A) + Y RY + Q = 0, (−A)T Y + Y (−A) + Y RY + Q < 0

if (−A, R) is stabilizable. This leads to the following result.

242
Corollary B.1 Suppose that all hypothesis in Theorem 7.2 hold true but that (−A, R) is
only stabilizable. Then the following statements are equivalent:

(a) H has no eigenvalues on the imaginary axis.

(b) AT X + XA + XRX + Q = 0 has a (unique) anti-stabilizing solution X+ .

(c) AT X + XA + XRX + Q < 0 has a symmetric solution X.

If one and hence all of these conditions are satisfied, then

any solution X of the ARE or ARI satisfies X ≤ X+ .

Combining this corollary with Theorem 7.2 implies Theorem 7.1.

C Proof of Theorem 7.2

Let us first show the inequality X− ≤ X if X− is the stabilizing solution of the ARE
(existence assumed) and X is any other solution of the ARE or ARI.

The key is the easily proved identity

(AT Z + ZA + ZRZ + Q) − (AT Y + Y A + Y RY + Q) =

= (A + RY )T (Z − Y ) + (Z − Y )(A + RY ) + (Z − Y )R(Z − Y ). (C.1)

If we set Z = X and Y = X− , and if we exploit R ≥ 0, we obtain

0 ≥ AT X + XA + XRX + Q ≥ (A + RX− )T (X − X− ) + (X − X− )(A + RX− ).

Since A + RX− is stable, we infer that X − X− ≥ 0.

Minimality of X− implies that there is at most one stabilizing solution. In fact, if X1 and
X2 are two stabilizing solutions of the ARE, we can infer X1 ≤ X2 (since X2 is smallest)
and X2 ≤ X1 (since X1 is smallest); this leads to X1 = X2 .

Before we can start the proof, we have to establish an important property of the Hamil-
tonian matrix H. First we observe that the particular structure of H can be expressed as
follows. Defining  
0 −I
J :=  ,
I 0
the matrix JH is symmetric! It is a pretty immediate consequence of this fact that
the eigenvalues of any Hamiltonian matrix are located symmetrically with respect to the
imaginary axis in the complex plane.

243
Lemma C.1 Suppose H is a matrix such that JH is symmetric. If H has k eigenvalues
in C< then it has also k eigenvalues in C> (counted with their algebraic multiplicity) and
vice versa.

Proof. By JH = (JH)T = H T J T = −H T J we infer JHJ −1 = −H T . Hence H and

−H T are similar and, thus, their characteristic polynomials are identical: det(H − sI) =
det(−H T − sI). Since H has even size, we infer det(−H T − sI) = det(H T + sI) =
det(H + sI) = det(H − (−s)I). The resulting equation det(H − sI) = det(H − (−s)I) is
all what we need: If λ ∈ C< is a zero of det(H − sI) (of multiplicity k), −λ is a zero (of
multiplicity k) of the same polynomial.

Now we can start the stepwise proof of Theorem 7.2.

Proof of (a) ⇒ (b). Since H has no eigenvalues in C= , it has n eigenvalues in C< and in
C> respectively (Lemma C.1). Then there exists a unitary matrix T with
 
M M12
T ∗ HT =  
0 M22

where M of size n × n is stable and M22 of size n × n is antistable. Let us denote the first
n columns of T by Z to infer
HZ = ZM.
 
U
Now partition Z =   with two square blocks U and V of size n. The difficult step
V
is now to prove that U is invertible. Then it is not so hard to see that X− := V U −1 is
indeed a real Hermitian stabilizing solution of the ARE.

We proceed in several steps. We start by showing

V ∗ U = U ∗ V.

Indeed, HZ = ZM implies Z ∗ JHZ = Z ∗ JZM . Since the left-hand side is symmetric,

so is the right-hand side. This implies (Z ∗ JZ)M = M ∗ (Z ∗ J ∗ Z) = −M ∗ (Z ∗ JZ) by
J ∗ = −J. Since M is stable, we infer from M ∗ (Z ∗ JZ) + (Z ∗ JZ)M = 0 that Z ∗ JZ = 04
what is indeed nothing but V ∗ U = U ∗ V . Next we show that

U x = 0 ⇒ RV x = 0 ⇒ U M x = 0.

Indeed, U x = 0 and the first row of HZ = ZM imply U M x = (AU + RV )x = RV x and

thus x∗ V ∗ U M x = x∗ V ∗ RV x. Since x∗ V ∗ U = x∗ U ∗ V = 0, the left-hand side and hence
4
Recall that AX − XB = 0 has no nonzero solution iff A and B have no eigenvalues in common.

244
the right-hand side vanish. By R ≥ 0, we conclude RV x = 0. From U M x = RV x we
infer U M x = 0. Now we can establish that

U is invertible.

It suffices to prove ker(U ) = {0}. Let us assume that ker(U ) is nontrivial. We have
just shown x ∈ ker(U ) ⇒ U x = 0 ⇒ U M x = 0 ⇒ M x ∈ ker(U ). Hence ker(U ) is a
nonzero M -invariant subspace. Therefore, there exists an eigenvector of M in ker(U ),
i.e., an x 6= 0 with M x = λx and U x = 0. Now the second row of HZ = ZM yields
(−QU − AT V )x = V M x and thus AT V x = −λV x. Since U x = 0, we have RV x = 0
(second step) or RT V x = 0. Since (A, R) is stabilizable and Re(−λ) > 0, we infer V x = 0.
Since U x = 0, this implies Zx = 0 and hence x = 0 because Z has full column rank.
However, this contradicts the choice of x as a nonzero vector.

Since U is nonsingular we can certainly define

X− := V U −1 .

X− is Hermitian since V ∗ U = U ∗ V implies U −∗ V ∗ = V U −1 and hence (V U −1 )∗ = V U −1 .

X− is a stabilizing solution of the ARE. Indeed, from HZ = ZM we infer HZU −1 =
ZM U −1 = ZU −1 (U M U −1 ). This leads to
   
I I
H =  (U M U −1 ).
X− X−

The first row of this identity shows A + RX− = U M U −1 such that A + RX− is stable.
The second row reads as −Q − AT X− = X− (A + RX− ) what is nothing but the fact that
X− satisfies the ARE.

So far, X− might be complex. Since the data matrices are real, we have

AT X + XA + XRX + Q = AT X̄ + X̄A + X̄RX̄ + Q, A + RX = A + RX̄.

Consequently, with X− , also its complex conjugate X̄− is a stabilizing solution of the
ARE. Since we have already shown that there is at most one such solution, we infer
X = X̄− such that X− must be necessarily real.

Proof of (b) ⇒ (a). Just evaluating both sides gives

    
I 0 I 0 A + RX− R
H =  .
X− I X− I 0 −(A + RX− )T

Hence  
A + RX− R
H is similar to  . (C.2)
0 −(A + RX− )T

245
Therefore, any eigenvalue of H is an eigenvalue of A + RX− or of −(A + RX− )T such
that H cannot have eigenvalues in C= .

Proof of (a) ⇒ (c). We perturb Q to Q + I and denote the corresponding Hamiltonian

as H . Since the eigenvalues of a matrix depend continuously on its coefficients, H has
no eigenvalues in C= for all small  > 0. By (b), there exists a symmetric X with
AT X + XA + XRX + Q + I = 0 what implies (c).

Proof of (c) ⇒ (b): Suppose Y satisfies the strict ARI and define P := AT Y + Y A +
Y RY + Q < 0. If we can establish that

(A + RY )T ∆ + ∆(A + RY ) + ∆R∆ + P = 0 (C.3)

has a stabilizing solution ∆, we are done: Due to (C.1), X = Y + ∆ satisfies the ARE
and renders A + RX = (A + RY ) + R∆ stable. The existence of a stabilizing solution of
(C.3) is assured as follows: Clearly, (A + RY, R) is stabilizable and (A, P ) is observable
(due to P < 0). Due to R ≥ 0 and P < 0, Lemma 7.3 is applicable and we can conclude
that the Hamiltonian matrix corresponding to (C.3) does not have any eigenvalues on the
imaginary axis. Hence we can apply the already proved implication (a) ⇒ (b) to infer
that (C.3) has a stabilizing solution.

246
References

[BDGPS] G.J. Balas, J.C. Doyle, K. Glover, A. Packard, R. Smith, µ-Analysis and
Synthesis Toolbox, The MahtWorks Inc. and MYSYNC Inc. (1995).

[DFT] J. Doyle, B. Francis, A. Tannenbaum, Feedback Control Theory, Macmillan Pub-

lishing Company (1990).

[DGKF] J. Doyle, K. Glover, P. Khargonekar, B. Francis, State-space solutions to stan-

dard H∞ and H2 control problems, IEEE Trans. Automat. Control 34 (1989)
831-847.

[DP] G.E. Dullerud. F. Paganini, A Course in Robust Control, Springer-Verlag, Berling,

(1999).

[DGKF] B. Francis, A course in H∞ control theory, Springer-Verlag, Berlin, (1987).

[G] M. Green, D. Limebeer, Linear Robust Control, Dover (2012).

[SGC] C.W. Scherer, P. Gahinet, M. Chilali, Multi-objective control by LMI optimiza-

tion, IEEE Trans. Automat. Control 42 (1997) 896-911.

[SP] S. Skogestad, I. Postlethwaite, Multivariable Feedback Control, Analysis and De-

sign, John Wiley & Sons, Chichester (1996).

[ZDG] K. Zhou, J.C. Doyle, K. Glover, Robust and Optimal Control, Prentice-Hall,
London (1996).

247