Lecture Notes On Linear System Theory

Lecture Notes on Linear System Theory
John Lygeros and Federico Ramponi

Automatic Control Laboratory, ETH Zurich
CH-8092, Zurich, Switzerland
[email protected]
Department of Information Engineering, University of Brescia

Via Branze 38, 25123, Brescia, Italy
[email protected]
September 22, 2013

Contents
1 Introduction 1
1.1 Objectives of the course . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Proof methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Functions and maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2 Introduction to Algebra 11
2.1 Groups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2 Rings and fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.3 Linear spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.4 Subspaces and bases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.5 Linear maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.6 Linear maps generated by matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.7 Matrix representation of linear maps . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.8 Change of basis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3 Introduction to Analysis 33
3.1 Norms and continuity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.2 Infinite-dimensional normed spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.3 Completeness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.4 The space of square-summable functions . . . . . . . . . . . . . . . . . . . . . . . . . 42
3.5 Induced norms and matrix norms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
3.6 Ordinary differential equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.7 Existence and uniqueness of solutions . . . . . . . . . . . . . . . . . . . . . . . . . . 50
3.7.1 Background lemmas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.7.2 Proof of existence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.7.3 Proof of uniqueness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
4 Time varying linear systems: Solutions 58

4.1 Motivation: Linearization about a trajectory . . . . . . . . . . . . . . . . . . . . . . 58
i
4.2 Existence and structure of solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
4.3 State transition matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
5 Time invariant linear systems: Solutions and transfer functions 69

5.1 Time domain solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
5.2 Semi-simple matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
5.3 Jordan form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
5.4 Laplace transforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
6 Stability 83
6.1 Nonlinear systems: Basic definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
6.2 Linear time varying systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
6.3 Linear time invariant systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
6.4 Systems with inputs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
6.5 Lyapunov equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
7 Inner product spaces 99

7.1 Inner product . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
7.2 Orthogonal complement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
7.3 Adjoint of a linear map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
7.4 Finite rank lemma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
7.5 Application: Matrix pseudo-inverse . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
8 Controllability and observability 109

8.1 Nonlinear systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
8.2 Linear time varying systems: Controllability . . . . . . . . . . . . . . . . . . . . . . . 112
8.3 Linear time varying systems: Minimum energy control . . . . . . . . . . . . . . . . . 115
8.4 Linear time varying systems: Observability and duality . . . . . . . . . . . . . . . . 118
8.5 Linear time invariant systems: Observability . . . . . . . . . . . . . . . . . . . . . . . 122
8.6 Linear time invariant systems: Controllability . . . . . . . . . . . . . . . . . . . . . . 126
8.7 Kalman decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
9 State Feedback and Observer Design 130

9.1 Revision: Change of basis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
9.2 Linear state feedback for single input systems . . . . . . . . . . . . . . . . . . . . . . 132
9.3 Linear state observers for single output systems . . . . . . . . . . . . . . . . . . . . . 137
9.4 Output feedback and the separation principle . . . . . . . . . . . . . . . . . . . . . . 140
9.5 The multi-input, multi-output case . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
ii
A Class composition and outline 149
A.1 Class composition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
A.2 Class outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
B Notation (C&D A1) 151

B.1 Shorthands . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
B.2 Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
B.3 Logic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
C Basic linear algebra 153
iii
Chapter 1
Introduction
1.1 Objectives of the course

This course has two main objectives. The first (and more obvious) is for students to learn something
about linear systems. Most of the course will be devoted to linear time varying systems that evolve
in continuous time t R+ . These are dynamical systems whose evolution is defined through state
space equations of the form
x(t) = A(t)x(t) + B(t)u(t),

y(t) = C(t)x(t) + D(t)u(t),
where x(t) Rn denotes the system state, u(t) Rm denotes the system inputs, y(t) Rp denotes
the system outputs, A(t) Rnn , B(t) Rnm , C(t) Rpn , and D(t) Rpm are matrices of
appropriate dimensions, and where, as usual, x(t) = dx
dt (t) denotes the derivative of x(t) with respect
to time.
Time varying linear systems are useful in many application areas. They frequently arise as models
of mechanical or electrical systems whose parameters (for example, the stiffness of a spring or the
inductance of a coil) change in time. As we will see, time varying linear systems also arise when
one linearizes a non-linear system around a trajectory. This is very common in practice. Faced with
a nonlinear system one often uses the full nonlinear dynamics to design an optimal trajectory to
guide the system from its initial state to a desired final state. However, ensuring that the system
will actually track this trajectory in the presence of disturbances is not an easy task. One solution
is to linearize the nonlinear system (i.e. approximate it by a linear system) around the optimal
trajectory; the approximation is accurate as long as the nonlinear system does not drift too far away
from the optimal trajectory. The result of the linearization is a time varying linear system, which
can be controlled using the methods developed in this course. If the control design is done well, the
state of the nonlinear system will always stay close to the optimal trajectory, hence ensuring that
the linear approximation remains valid.
A special class of linear time varying systems are linear time invariant systems, usually referred to
by the acronym LTI. LTI systems are described by state equations of the form
x(t) = Ax(t) + Bu(t),

y(t) = Cx(t) + Du(t),
where the matrices A Rnn , B Rnm , C Rpn , and D Rpm are constant for all times
t R+ . LTI systems are somewhat easier to deal with and will be treated in the course as a special
case of the more general linear time varying systems.
1
c J. Lygeros & F. Ramponi, 2013
Lecture Notes on Linear System Theory, 2
The second and less obvious objective of the course is for students to experience something about
doing automatic control research, in particular developing mathematical proofs and formal logical
arguments. Linear systems are ideally suited for this task. There are two main reasons for this. The
first is that almost all the derivations given in the class can be carried out in complete detail, down to
the level of basic algebra. There are very few places where one has to invoke higher powers, such as
an obscure mathematical theorem whose proof is outside the scope of the course. One can generally
continue the calculations until he/she is convinced that the claim is true. The second reason is that
linear systems theory brings together two areas of mathematics, algebra and analysis. As we will
soon see, the state space, Rn , of the systems has both an algebraic structure (it is a vector space)
and a topological structure (it is a normed space). The algebraic structure allows us to perform
linear algebra operations, compute projections, eigenvalues, etc. The topological structure, on the
other hand, forms the basis of analysis, the definition of derivatives, etc. The main point of linear
systems theory is to exploit the algebraic structure to develop tractable algorithms that allow us
to answer analysis questions which appear intractable by themselves.
For example, consider the time invariant linear system
x(t) = Ax(t) + Bu(t) (1.1)
with x(t) Rn , u(t) Rm , A Rnn , and B Rnm . Given x0 Rn , T > 0 and a continuous
function u() : [0, T ] Rm (known as the input trajectory) one can show (Chapter 3) that there
exists a unique function x() : [0, T ] Rn such that
x(0) = x0 and x(t) = Ax(t) + Bu(t), for all t [0, T ]. (1.2)
This function is called the state trajectory (or simply the solution) of system (1.1) with initial
condition x0 under the input u(). As we will see in Chapter 3, u() does not even need to be
continuous for (1.2) to be true, provided one appropriately qualifies the statement for all t [0, T ].
System (1.1) is called controllable (Chapter 8) if and only if for all x0 Rn , for all x Rn , and for
all T > 0, there exists u() : [0, T ] Rm such that the solution of system (1.1) with initial condition
x0 under the input u() is such that x(T ) = x. Controllability is clearly an interesting property for
a system to have. If the system is controllable then we can guide it from any initial state to any
final state by selecting an appropriate input. If not, there may be some desirable parts of the state
space that we cannot reach from some initial states. Unfortunately, determining whether a system is
controllable directly from the definition is impossible. This would require calculating all trajectories
that start at all initial conditions. Except for trivial cases (like the linear system x(t) = u(t)) this
calculation is intractable, since the initial states, x0 , the times T of interest, and the possible input
trajectories u() : [0, T ] Rm are all infinite. Fortunately, linear algebra can be used to answer the
question without even computing a single solution (Chapter 8).

Theorem 1.1 System (1.1) is controllable if and only if the matrix B AB . . . An1 B Rnnm
has rank n.
The theorem shows how the seemingly intractable analysis question is the system (1.1) control-
lable? can be answered by a simple algebraic calculation of the rank of a matrix.
1.2 Proof methods

Most of the course will be devoted to proving theorems. The proof methods that we will encounter
are just a set of tools, grounded in mathematical logic and widely accepted in the mathematical
community, that let us say that a proposition is true, given that others are true. A Theorem is
indeed a logical statement that can be proven: This means that the truth of such statement can
be established by applying our proof methods to other statements that we already accept as true,
either because they have been proven before, or because we postulate so (for example the axioms
of logic), or because we assume so in a certain context (for example, when we say Let V be a vector
space . . . we mean Assume that the set V verifies the axioms of a vector space . . . ).
Theorems of minor importance, or theorems whose main point is to establish an intermediate step in
the proof of another theorem, will be called Lemmas, Facts, or Propositions; An immediate
consequence of a theorem that deserves to be highlighted separately is usually called a Corollary.
And a logical statement that we think may be true but cannot prove so is called a Conjecture.
The logical statements we will most be interested in typically take the form
pq
(p implies q). p is called the hypothesis and q the consequence.
Example (No smoke without fire) It is generally accepted that when there is smoke, there must
be some a fire somewhere. This knowledge can be encoded by the logical implication
If there is smoke then there is a fire

p q.
This is a statement of the form p q with p the statement there is smoke and q the statement
there is a fire.
Hypotheses and consequences may typically depend on one or more free variables, that is, objects
that in the formulation of hypotheses and consequences are left free to change.
Example (Greeks) Despite recent economic turbulence, it is generally accepted that Greek citizens
are also Europeans. This knowledge can be encoded by the logical implication
If X is a Greek then X is a European

p(X) q(X).
A sentence like X is a . . . is the verbal way of saying something belongs to a set; for example the
above statement can also be written as
X Greeks X Europeans,
where Greeks and Europeans are supposed to be sets; the assertion that this implication is true
for arbitrary X (X, X Greeks X Europeans) is equivalent to the set-theoretic statement of
inclusion:
Greeks Europeans.
You can visualize the implication and its set-theoretic interpretation in Figure 1.1.
There are several ways of proving that logical statements are true. The most obvious one is a direct
proof: Start from p and establish a finite sequence of intermediate implications, p1 , p2 , . . . , pn
leading to q
p p1 p2 . . . pn q.
We illustrate this proof technique using a statement about the natural numbers.
Definition 1.1 A natural number n N is called odd if and only if there exists k N such that
n = 2k + 1. It is called even if and only if there exists k N such that n = 2k.
One can indeed show that all natural numbers are either even or odd, and no natural number is
both even and odd (Problem 1.1).
Theorem 1.2 If n is odd then n2 is odd.

People
Europeans
Greeks
Figure 1.1: Set theoretic interpretation of logical implication.
Proof:
n is odd k N : n = 2k + 1
k N : n2 = (2k + 1)(2k + 1)
k N : n2 = 4k 2 + 4k + 1
k N : n2 = 2(2k 2 + 2k) + 1
l N : n2 = 2l + 1 (namely, l = 2k 2 + 2k N)
n2 is odd
This proof principle can also be exploited to perform proof by induction. Proof by induction concerns
propositions, pk , indexed by the natural numbers, k N, and statements of the form
k N, pk is true.
One often proves such statements by showing that p0 is true and then establishing and infinite
sequence of implications
p0 p1 p2 . . . .
Clearly proving these implications one by one is impractical. It suffices, however, to establish that
pk pk+1 for all k N, or in other words
[p0 (pk pk+1 , k N)] [pk , k N] .
We demonstrate this proof style with another statement about the natural numbers.
Definition 1.2 The factorial, n!, of a natural number n N is the natural number n! = n (n 1)
. . . 2 1. By convention, if n = 0 we set n! = 1.
Theorem 1.3 For all m, k N, (m + k)! m!k!.
Proof: It is easy to check that the statement holds for the special cases m = k = 0, m = 0 and
k = 1, and m = 1 and k = 0. For the case m = k = 1, (m + k)! = 2! 1!1! = m!k!.
Assume now that for some m, k N, (m + k)! m!k! (we call this the induction hypothesis) and
show that for m, k + 1, (m + k + 1)! m!(k + 1)! (also (m + 1)!k! for the case m + 1, k, by symmetry).
(m + k + 1)! = (m + k)!(m + k + 1)
m!k!(m + k + 1) (by the induction hypothesis)
m!k!(k + 1) (since m N)
= m!(k + 1)!
which completes the proof.
Even though there is no direct way to illustrate proof by induction using statements about Greeks
and other nationalities, one could in principle use a similar line of reasoning to prove a statement
like once a Greek always a Greek by arguing that children with at least one Greek parent are
themselves Greek.
Sometimes direct proof p q is difficult. In this case we try to find other statements that are
logically equivalent to p q and prove these instead. An example of such a statement is q p,
or in logic notation
(q p) (p q).
Example (Greeks (cont.)) The statement that all the Greeks are Europeans is also equivalent to
If X is not a European then X is not a Greek

q(X) p(X).
In turn, this is equivalent to the set theoretic statement
non-Europeans non-Greeks.
If we stipulate a priori that all the possible X we may consider in our discourse belong to some big
set (i.e., People), in fact this is also equivalent to
People \ Europeans = Europeansc Greeksc = People \ Greeks.
where \ denotes the difference of two sets and the superscript c denotes the set-complement with
respect to People.
Exercise 1.1 Visualize this set theoretic interpretation by a picture similar to Figure 1.1.
A proof where we show that p q is true by showing q p is true is known as a proof by

contraposition. We illustrate this proof technique by another statement about the natural numbers.
Theorem 1.4 If n2 is odd then n is odd.
Proof: Let p = n2 is odd, q = n is odd. Assume n is even (q) and show that n2 is even (p).
n is even k N : n = 2k
k N : n2 = 4k 2
l N : n = 2l (namely, l = 2k 2 N)
n2 is even.
Another common method that can be used to indirectly prove that p q is to suppose that p
is true, to suppose that q is false, and to apply other proof methods to derive a contradiction. A
contradiction is a proposition of the form r r (like There is smoke and there is no smoke, or
n is even and n is odd); all such statements are postulated to be false by virtue of their mere
structure, and irrespective of the proposition r. If, by assuming p is true and q is false we are able
to reach a false assertion, we must admit that if p is true the consequence q cannot be false, in other
words that p implies q. This method is known as proof by contradiction.
Example (Greeks and Chinese) Suppose the following implications: for all X,
X is a Greek X is a European
X is a Chinese X is an Asian
X is an Asian X is not a European
We show by contradiction that every Greek is not a Chinese, more formally
If X is a Greek then X is not a Chinese

p(X) q(X)
Indeed, suppose p(X) and the converse of q(X), that is, X is a Chinese. By direct deduction,
X is a Greek X is a Chinese

X is a European X is an Asian

X is a European X is not a European
Since the conclusion is a contradiction for all X, we must admit that p(X) q(X) is false or, which
is the same, that p(X) q(X). The set-theoretic interpretation is as follows: By postulate,
Europeans non-Europeans =
On the other hand, by deduction,
(Greeks Chinese) (Europeans non-Europeans)
It follows that Greeks Chinese is also equal to the empty set. Therefore (here is the point of the
above proof), Greeks non-Chinese.
Exercise 1.2 Visualize this set theoretic interpretation by a picture similar to Figure 1.1.
We will illustrate this fundamental proof technique with another statement, about rational numbers.
Definition 1.3 The real number x R is called rational if and only if there exist integers n, m Z
with m 6= 0 such that x = n/m.

Theorem 1.5 (Pythagoras) 2 is not rational.

Proof: (Euclid) Assume,
for the sake of contradiction,
that 2 is rational. Then there exist n, m Z
with m 6= 0 such that 2 = n/m. Since 2 > 0, without loss of generality we can take n, m N; if
they happen to be both negative multiply both by 1 and replace them by the resulting numbers.
Without loss of generality, we can further assume that m and n have no common divisor; if they do,
divide both by their common divisors until there are no common divisors left and replace m and n
by the resulting numbers. Now

n n2
2= 2= 2
m m
n = 2m2
2
n2 is even
n is even (Theorem 1.4 and Problem 1.1)
k N : n = 2k
k N : 2m2 = n2 = 4k 2
k N : m2 = 2k 2
m2 is even
m is even (Theorem 1.4 and Problem 1.1).
Therefore, n and m are both even and, according to Definition

??, 2 divides both. This contradicts
the fact that n and m have no common divisor. Therefore 2 cannot be rational.
Exercise 1.3 What is the statement p in Theorem 1.5? What is the statement q? What is the
statement r in the logical contradiction r r reached at the end of the proof?
Two statements are equivalent if one implies the other and vice versa,
(p q) is the same as (p q) (q p)
Usually showing that two statements are equivalent is done in two steps: Show that p q and then
show that q p. For example, consider the following statement about the natural numbers.
Theorem 1.6 n2 is odd if and only if n is odd.
Proof: n is odd implies that n2 is odd (by Theorem 1.2) and n2 is odd implies that n is odd (by
Theorem 1.4). Therefore the two statements are equivalent.
This is argument is related to the canonical way of proving that two sets are equal, by proving two
set inclusions A B and B A. To prove these inclusions one proves two implications:
X AXB
XBXA
or, in other words, X A X B.

Finally, let us close this brief discussion on proof techniques with a subtle caveat: If p is a false
statement then any implication of the form p q is true, irrespective of what q is.
Example (Maximal natural number) Here is a proof that there is no number larger than 1.
Theorem 1.7 Let N N be the largest natural number. Then N = 1.
Proof: Assume, for the sake of contradiction, that N > 1. Then N 2 is also a natural number and
N 2 > N . This contradicts the fact that N is the largest natural number. Therefore we must have
N = 1.
Obviously the theorem in this example is saying something quite silly. The problem, however,
is not that the proof is incorrect, but that the starting hypothesis let N be the largest natural
number is false, since there is no largest natural number.
g
X Y
f g f
Figure 1.2: Commutative diagram of function composition.
1.3 Functions and maps

A function
f :XY
maps the set X (known as the domain of f ) into the set Y (known as the co-domain of f ). This
means that for all x X there exists a unique y Y such that f (x) = y. The element f (x) Y is
known as the value of f at x. The set
{y Y | x X : f (x) = y} Y
is called the range of f (sometimes denoted by f (X)) and the set
{(x, y) X Y | y = f (x)} X Y
is called the graph of f .
Definition 1.4 A function f : X Y is called:
1. Injective (or one-to-one) if and only if f (x1 ) = f (x2 ) implies that x1 = x2 .

2. Surjective (or onto) if and only if for all y Y there exists x X such that y = f (x).
3. Bijective if and only if it is both injective and surjective, i.e. for all y Y there exists a unique
x X such that y = f (x).
Given two functions g : X Y and f : Y Z their composition is the function (f g) : X Z

defined by
(f g)(x) = f (g(x)).
Commutative diagrams help visualize function compositions (Figure 1.2).
Exercise 1.4 Show that composition is associative. In other words, for any three functions g : X
Y , f : Y Z and h : W X and for all w W , f (g h)(w) = (f g) h(w).
By virtue of this associativity property, we will simply use f gh : W Y to denote the composition
of three (or more) functions.
A special function that can always be defined on any set is the identity function, also called the
identity map, or simply the identity.
Definition 1.5 The identity map on X is the function 1X : X X defined by 1X (x) = x for all
x X.
Exercise 1.5 Show that the identity map is bijective.
Using the identity map one can also define various inverses of functions.
f f
g L f = 1X X Y X Y f g R = 1Y
gL gR
f
g f = 1X X Y f g = 1Y
g
Figure 1.3: Commutative diagram of function inverses.
Definition 1.6 Consider a function f : X Y .
1. The function gL : Y X is called a left inverse of f if and only if gL f = 1X .

2. The function gR : Y X is called a right inverse of f if and only if f gR = 1Y .
3. The function g : Y X is called an inverse of f if and only if it is both a left inverse and a
right inverse of f , i.e. (g f = 1X ) (f g = 1Y ).
f is called invertible if an inverse of f exists.
The commutative diagrams for the different types of inverses are shown in Figure 1.3.
Fact 1.1 If f is invertible then its inverse is unique.
Proof: Assume, for the sake of contradiction, that f is invertible but there exist two different
inverses, g1 : Y X and g2 : Y X. Since the inverses are different, there must exist y Y such
that g1 (y) 6= g2 (y). Let x1 = g1 (y) and x2 = g2 (y) and note that x1 6= x2 . Then
x1 = g1 (y) = 1X g1 (y) = (g2 f ) g1 (y) (g2 inverse of f )

= g2 (f g1 )(y) (composition is associative)
= g2 1Y (y) (g1 inverse of f )
= g2 (y) = x2 .
This contradicts the assumption that x1 6= x2 .
Problems for chapter 1

Problem 1.1 (Even and odd numbers) Show that every n N is either even or odd, but not
both.
Problem 1.2 (Inverses of functions) Consider two sets X and Y and a function f : X Y .
Show that:
1. f has a left inverse if and only if it is injective.

2. f has a right inverse if and only if it is surjective.
3. f is invertible if and only if it is bijective. In this case show also that all inverses (left-, right-
and two-sided) coincide.
4. If f is invertible then any two inverses (left-, right- or both) coincide.
Chapter 2
Introduction to Algebra
2.1 Groups
Definition 2.1 A group (G, ) is a set G equipped with a binary operation : G G G such
that:
1. The operation is associative: a, b, c G, a (b c) = (a b) c.

2. There exists an identity element: e G, a G, a e = e a = a.
3. Every element has an inverse element: a G, a1 G, a a1 = a1 a = e.
(G, ) is called commutative (or Abelian) if and only if in addition to 1-3 above
4. is commutative: a, b G, a b = b a.
Example (Common groups)

(R, +) is a commutative group. What is the identity element? What is the inverse?
(R, ) is not a group, since 0 has no inverse.
The set ({0, 1, 2}, +mod3) is a group. What is the identity element? What is the inverse? Is it
commutative? Recall that (0 + 1)mod3 = 1, (1 + 2)mod3 = 0, (2 + 2)mod3 = 1, etc.
The set of rotations of R2 (usually denoted by SO(2), or U(1) or S(1)) given by

cos() sin()
(, ] ,
sin() cos()
with the usual operation of matrix multiplication is a group1 . What is the identity? What is the
inverse?
Fact 2.1 For a group (G, ) the identity element, e, is unique. Moreover, for all a G the inverse
element, a1 , is unique.
1 For the time being, the reader is asked to excuse the use of matrices in the examples. Matrices will be formally
defined in the next section, but will be used in the meantime for informal illustrations.
11
Proof: To show the first statement, assume, for the sake of contradiction, that there exist two
identity elements e, e G with e 6= e . Then for all a G, e a = a e = a and e a = a e = a.
Then:
e = e e = e
which contradicts the assumption that e 6= e .
To show the second statement, assume, for the sake of contradiction, that there exists a G with
two inverse elements, say a1 and a2 with a1 6= a2 . Then
a1 = a1 e = a1 (a a2 ) = (a1 a) a2 = e a2 = a2 ,
which contradicts the assumption that a1 6= a2 .
2.2 Rings and fields

Definition 2.2 A ring (R, +, ) is a set R equipped with two binary operations, + : R R R
(called addition) and : R R R (called multiplication) such that:
1. Addition satisfies the following properties:

It is associative: a, b, c R, a + (b + c) = (a + b) + c.
It is commutative: a, b R, a + b = b + a.
There exists an identity element: 0 R, a R, a + 0 = a.
Every element has an inverse element: a R, (a) R, a + (a) = 0.
2. Multiplication satisfies the following properties:

It is associative: a, b, c R, a (b c) = (a b) c.
There exists an identity element: 1 R, a R, 1 a = a 1 = a.
3. Multiplication is distributive with respect to addition: a, b, c R, a (b + c) = a b + a c and
(b + c) a = b a + c a.
(R, +, ) is called commutative if in addition a, b R, a b = b a.
Example (Common rings)

(R, +, ) is a commutative ring.
(Rnn , +, ) with the usual operations of matrix addition and multiplication is a non-commutative
ring.
The set of rotations
cos() sin()
(, ] , +,
sin() cos()
with the same operations is not a ring, since it is not closed under addition.
(R[s], +, ), the set of polynomials of s with real coefficients, i.e. an sn + an1 sn1 + . . . + a0 for some
n N and a0 , . . . , an R is a commutative ring.
(R(s), +, ), the set of rational functions of s with real coefficients, i.e.
am sm + am1 sm1 + . . . + a0
bn sn + bn1 sn1 + . . . + b0
for some n, m N and a0 , . . . , am , b0 , . . . , bn R with bn 6= 0 is a commutative ring. We implicitly

assume here that the numerator and denominator polynomials are co-prime, that is they do not
have any common factors; if they do one can simply cancel these factors until the two polynomials
are co-prime. For example, it is easy to see that with such cancellations any rational function of the
form
0
bn sn + bn1 sn1 + . . . + b0
can be identified with the rational function 0/1, which is the identity element of addition for this
ring.
(Rp (s), +, ), the set of proper rational functions of s with real coefficients, i.e.
an sn + an1 sn1 + . . . + a0
bn sn + bn1 sn1 + . . . + b0
for some n N with a0 , . . . , an , b0 , . . . , bn R with bn 6= 0 is a commutative ring. Note that an = 0
is allowed, i.e. it is possible for the degree of the numerator polynomial to be less than or equal to
that of the denominator polynomial.
Exercise 2.1 Show that for every ring (R, +, ) the identity elements 0 and 1 are unique. Moreover,
for all a R the inverse element (a) is unique.
Fact 2.2 If (R, +, ) is a ring then:
1. For all a R, a 0 = 0 a = 0.
2. For all a, b R, (a) b = (a b) = a (b).
Proof: To show the first statement note that
a + 0 = a a (a + 0) = a a a a + a 0 = a a
(a a) + a a + a 0 = (a a) + a a
0 + a 0 = 0 a 0 = 0.
The second equation is similar. For the second statement note that
0 = 0 b = (a + (a)) b = a b + (a) b (a b) = (a) b.
The second equation is again similar.
Definition 2.3 A field (F, +, ) is a commutative ring that in addition satisfies
Multiplication inverse: a F with a 6= 0, a1 F , a a1 = 1.
Example (Common fields)

(R, +, ) is a field.
(Rnn , +, ) is not a field, since singular matrices have no inverse.
({A Rnn | Det(A) 6= 0}, +, ) is not a field, since it is not closed under addition.
The set of rotations
cos() sin()
(, ] , +.
sin() cos()
is not a field, it is not even a ring.
(R[s], +, ) is not a field, since the multiplicative inverse of a polynomial is not a polynomial but a
rational function.
(R(s), +, ) is a field.
(Rp (s), +, ) is not a field, since the multiplicative inverse of a proper rational function is not neces-
sarily proper.
Exercise 2.2 If (F, +, ) is a field then for all a R the multiplicative inverse element a1 is unique.
Exercise 2.3 Show that if (F, +, ) is a field and a 6= 0 then a b = a c a = c. Is the same true
for a ring? Illustrate using an example from R22 .
Given a field (F, +, ) and integers n, m N one can define matrices

a11 . . . a1m
.. F nm
A = ... ..
. .
an1 . . . anm
with a11 , . . . , anm F . One can then define the usual matrix operations in the usual way: ma-
trix multiplication, matrix addition, determinant, identity matrix, inverse matrix, adjoint matrix,
. . . . We assume that the reader is familiar with these operations from basic linear algebra, a brief
summary is provided in Appendix C.
Matrices can also be formed from the elements of a commutative ring (R, +, ).
Example (Transfer function matrices) As we will see in Chapter 5, the set of transfer functions
of time invariant linear systems with m inputs and p outputs are p m matrices whose elements are
proper rational functions,

x(t) = Ax(t) + Bu(t)
G(s) = C(sI A)1 B + D Rp (s)pm .
y(t) = Cx(t) + Du(t)
There are, however, subtle differences.
Fact 2.3 Assume that (R, +, ) is a commutative ring, A Rnn , and Det(A) 6= 0. It is not always
the case that A1 exists.
Roughly speaking, what goes wrong is that even though DetA 6= 0, the inverse (DetA)1 may not
exist in R. Then the inverse of the matrix defined as A1 = (Det(A))1 Adj(A) is not defined. The
example of the transfer function matrices given above illustrates the point: Assuming m = p, the
elements of the inverse of a transfer function matrix are not necessarily proper rational functions,
and hence the inverse matrix cannot be the transfer function of a linear time invariant system.
2.3 Linear spaces

We now come to the algebraic object of greatest interest for linear systems, namely linear (or vector)
spaces.
Definition 2.4 A linear space (V, F, , ) is a set V (of vectors) and a field (F, +, ) (of scalars)
equipped with two binary operations, : V V V (called vector addition) and : F V V
(called scalar multiplication) such that:
1. Vector addition satisfies the following properties:

It is associative: x, y, z V , x (y z) = (x y) z.
It is commutative: x, y V , x y = y x.
There exists an identity element: V , x V , x = x.
For every element there exists an inverse element: x V , (x) V , x (x) = .
2. Scalar multiplication satisfies the following properties:
It is associative: a, b F , x V , a (b x) = (a b) x.
Multiplication by the multiplication identity of F leaves elements unchanged: x V ,
1 x = x.
3. Scalar multiplication is distributive with respect to vector addition: a, b F , x, y V ,
(a + b) x = (a x) (b x) and a (x y) = (a x) (a y).
Exercise 2.4 Let (F, +, ) be a field. Show that (F, F, +, ) is a linear space.
As for groups, rings and fields the following fact is immediate.
Exercise 2.5 For every linear space (V, F, , ) the identity element is unique. Moreover, for all
x V there exists a unique inverse element x.
The following relations can also be established.
Fact 2.4 If (V, F, , ) is a linear space and 0 is the addition identity element of F then for all
x V , 0 x = . Moreover, for all a F , x V , (a) x = (a x) = a (x).
The proof is left as an exercise (Problem 2.3).

Once we have found a few linear spaces we can always generate more by forming the so called
product spaces.
Definition 2.5 If (V, F, V , V ) and (W, F, W , W ) are linear spaces over the same field, the
product space (V W, F, , ) is the linear space comprising all pairs (v, w) V W with defined
by (v1 , w1 ) (v2 , w2 ) = (v1 V v2 , w1 W w2 ), and defined by a (v, w) = (a V v, a W w).
Exercise 2.6 Show that (V W, F, , ) is a linear space. What is the identity element for addition?
What is the inverse element?
Two types of linear spaces will play a central role in these notes. The first is constructed by taking
repeatedly the product of a field with itself.
Example (Finite product of a field) For any field, (F, +, ), consider the product space F n . Let
x = (x1 , . . . , xn ) F n , y = (y1 , . . . , yn ) F n and a F and define : F n F n F n by
x y = (x1 + y1 , . . . , xn + yn )
and : F F n F n by
a x = (a x1 , . . . , a xn ).
Note that both operations are well defined since a, x1 , . . . , xn , y1 , . . . , yn all take values in the same
field, F .
Exercise 2.7 Show that (F n , F, , ) is a linear space. What is the identity element ? What is
the inverse element x of x F n ?
The most important instance of this type of linear space in these notes will be (Rn , R, +, ) with the
usual addition and scalar multiplication for vectors. The state, input, and output spaces of linear
systems will be linear spaces of this type.
The second class of linear spaces that will play a key role in linear system theory are function spaces.
Example (Function spaces) Let (V, F, V , V ) be a linear space and D be any set. Let F (D, V )
denote the set of functions of the form f : D V . Consider f, g F (D, V ) and a F and define
: F (D, V ) F (D, V ) F (D, V ) by
(f g) : D V such that (f g)(d) = f (d) V g(d) d D
and : F F (D, V ) F (D, V ) by
(a f ) : D V such that (a f )(d) = a V f (d) d D
Note that both operations are well defined since a F , f (d), g(d) V and (V, F, V , V ) is a linear
space.
Exercise 2.8 Show that (F (D, V ), F, , ) is a linear space. What is the identity element? What
is the inverse element?
The most important instance of this type of linear space in these notes will be (F ([t0 , t1 ], Rn ), R, +, )
for real numbers t0 < t1 . The trajectories of the state, input, and output of the dynamical systems
we consider will take values in linear spaces of this type. The state, input and output trajectories
will differ in terms of their smoothness as functions of time. We will use the following notation to
distinguish the level of smoothness of the function in question:
C([t0 , t1 ], Rn ) will be the linear space of continuous functions f : [t0 , t1 ] Rn .

C 1 ([t0 , t1 ], Rn ) will be the linear space of differentiable functions f : [t0 , t1 ] Rn .
C k ([t0 , t1 ], Rn ) will be the linear space of k-times differentiable functions f : [t0 , t1 ] Rn .
C ([t0 , t1 ], Rn ) will be the linear space of infinitely differentiable functions f : [t0 , t1 ] Rn .
C ([t0 , t1 ], Rn ) will be the linear space of analytic functions f : [t0 , t1 ] Rn , i.e. functions
which are infinitely differentiable and whose Taylor series expansion converges for all t [t0 , t1 ].
Exercise 2.9 Show that all of these sets are linear spaces. You only need to check that they are
closed under addition and scalar multiplication. E.g. if f and g are differentiable, then so is f g.
Exercise 2.10 Show that for all k = 2, 3, . . .

C ([t0 , t1 ], Rn ) C ([t0 , t1 ], Rn ) C k ([t0 , t1 ], Rn ) C k1 ([t0 , t1 ], Rn )
C([t0 , t1 ], Rn ) (F ([t0 , t1 ], Rn ), R, +, ).
Note that all subsets are strict, so there must be functions that belong to one set but not the previous
one. Try to think of examples.
To simplify the notation, unless there is special reason to distinguish the operations and identity
element of a linear space from those of the field, from now on we will use the regular symbols +
and instead of and for the linear space operations of vector addition and scalar multiplication
respectively; in fact as for real numbers we will mostly ommit and simply write av instead of a v
for a F , v V . Likewise, unless explicitly needed we will also use 0 instead of to denote the
identity element of addition. Finally we will stop writing the operations explicitly when we define
the vector space and write (V, F ) or simply V whenever the field is clear from the context.
2.4 Subspaces and bases

Definition 2.6 Let (V, F ) be a linear space and W V . (W, F ) is a linear subspace of V if and only
if it is itself a linear space, i.e. for all w1 , w2 W and all a1 , a2 F , we have that a1 w1 +a2 w2 W .
Note that by definition a linear space and all its subspaces are linear spaces over the same field. The
equation provides a way of testing whether a given set W is a subspace or not: One needs to check
whether linear combinations of elements of W with coefficients in F are also elements of W .
Exercise 2.11 Show that if W is a subspace then for all n N, and ai F , wi W for i = 1, . . . , n
n
X
ai wi W.
i=1
Show further that V W . Hence show that W = V .
Example (Linear subspaces) In R2 , the set {(x1 , x2 ) R2 | x1 = 0} is subspace. So is the set

{(x1 , x2 ) R2 | x1 = x2 }. But the set {(x1 , x2 ) R2 | x2 = x1 + 1} is not a subspace and neither is
the set {(x1 , x2 ) R2 | (x1 = 0) (x2 = 0)}.
In R3 , all subspaces are:
1. R3
2. 2D planes through the origin.
3. 1D lines through the origin.
4. {0}.
For examples of subspace of function spaces consider (R[t], R) (polynomials of t R with real
coefficients). This is a linear subspace of C (R, R), which in turn is a linear subspace of C(R, R).
The set
{f : R R | t R, |f (t)| 1}
on the other hand is not a subspace of F (R, R).
Exercise 2.12 Show that {f : R R | t R, |f (t)| 1} is not a subspace. How about

{f : R R | M > 0, t R, |f (t)| M }
It is easy to see that the family of subspaces of a given a linear space is closed under finite addition
and intersection.
Exercise 2.13 Let {(Wi , F )}ni=1 be a finite family of subspaces of a linear space (V, F ). Show that
(ni=1 Wi , F ) is also a subspace. Is (ni=1 Wi , F ) a subspace?
Exercise 2.14 Let (W1 , F ) and (W2 , F ) be subspaces of (V, F ) and define
W1 + W2 = {w1 + w2 | w1 W1 , w2 W2 }.
Show that (W1 + W2 , F ) is a subspace of (V, F ).
A subset of a linear space will of course not be a subspace in general. Each subset of a linear space
does, however, generate a subspace in a natural way.
Definition 2.7 Let (V, F ) be a linear space and S V . The linear subspace of (V, F ) generated by S
is the smallest subspace of (V, F ) containing S.
Here, smallest is to be understood in the sense of set inclusion.
Exercise 2.15 What is the subspace generated by {(x1 , x2 ) R2 | (x1 = 0) (x2 = 0)}? What is
the subspace generated by {(x1 , x2 ) R2 | x2 = x1 + 1}? What is the subspace of R2 generated by
{(1, 2)}?
Definition 2.8 The span of S is the set defined by

( n )
X
Span(S) = ai vi n N, ai F, vi S, i = 1, . . . , n .
i=1
Fact 2.5 Let (V, F ) be a linear space and S V . The subspace generated by S coincides with
Span(S).
Proof: The fact that Span(S) is a subspace and contains S is easy to check. To show that is is
the smallest subspace containing S, consider another subspace, W , that contains S and an arbitrary
v Span(S); we will show that v W and hence Span(S) W . Since v Span(S) it can be
written as
Xn
v= ai vi
i=1
for some n N and ai F , vi S, i = 1, . . . , n. Since S W we must also have vi W ,

i = 1, . . . , n and hence v W (since W is a subspace).
The elements of Span(S) are known as linear combinations of elements of S. Notice that in general
the set S may contain an infinite number of elements; this was for example the case for the set
{(x1 , x2 ) R2 | (x1 = 0) (x2 = 0)} in Exercise 2.14. The span of S, however, is defined as the set
of all finite linear combinations of elements of S.
Definition 2.9 Let (V, F ) be a linear space. A set S V is called linearly independent if and only
if for all n N, vi S for i = 1, . . . , n
n
X
ai vi = 0 ai = 0, i = 1, . . . , n.
i=1
A set which is not linearly independent is called linearly dependent.
Note again that the set S may be infinite, but we only consider finite linear combinations to define
linear independence.
Exercise 2.16 Show that a set S V is linearly independent if and only if none of its elements
can be written as a finite linear combination of other elements in S. Show that every set S that
contains the identity element of vector addition is linearly dependent.
Example (Linearly independent set in Rn ) The following finite family of vectors are linearly
independent in (R3 , R):
{(1, 0, 0), (0, 1, 0), (0, 0, 1)}.
It is easy to show that any linear combination of these vectors is equal to (0, 0, 0) if and only if all
three coefficients are equal to 0.
Example (Linearly independent functions) Consider C([1, 1], R) and let fk (t) = tk for all
t [1, 1] and k N. Clearly fk (t) C([1, 1], R).
Fact 2.6 The vectors of the collection {fk (t)}

k=0 are linearly independent.
Proof: We need to show that for all n N

n
X
ai fi = ai = 0, i = 0, . . . , N.
i=0
Here C([1, 1], R) denotes the addition identity element for C([1, 1], R), i.e. the function
: [1, 1] R such that (t) = 0 for all t [1, 1].
One direction () is obvious. For the other () note that
n
X
f (t) = an fn (t) = a0 + a1 t + . . . + an tn = f (t) = 0 t [1, 1]
i=0
d dn
f (t) = f (t) = . . . = n f (t) = 0 t [1, 1]
dt dt
d dn
f (0) = f (0) = . . . = n f (0) = 0
dt dt
a0 = a1 = . . . = n!an = 0.
Hence a0 = a1 = . . . = an = 0 and {fk (t)}

k=0 are linearly independent.
Notice that in this case the collection {fk (t)}

k=0 contains an infinite number of elements.
At this stage one may be tempted to ask how many linearly independent vectors can be found in
a given vector space. This number is a fundamental property of the vector space, and is closely
related to the notion of a basis.
Definition 2.10 Let (V, F ) be a linear space. A set of vectors S V is a basis of (V, F ) if and
only if they are linearly independent and Span(S) = V .
One can then show the following.
Fact 2.7 Let (V, F ) be a linear space. If a basis of (V, F ) with a finite number of elements exists,
then every other basis of (V, F ) has the same number of elements.
Proof: Consider two bases of (V, F ), S and S and assume that S has a finite number of elements
{vi }ni=1 . Assume, for the sake of contradiction that S does not have the same number of elements.
Without loss of generality assume further that S has more elements than S; if not, interchange S
and S since in this case S must also have a finite number of elements. Take n + 1 elements from S ,
{vi }n+1 n
i=1 and recall that since S is a basis they must be linearly independent. Since {vi }i=1 generate
V we can write v1 as a linear combination
v1 = a1 v1 + . . . + an vn .
Note that at least one of the ai must be non-zero; otherwise v1 = 0 and the set {vi }n+1
i=1 cannot be
linearly independent (Exercise 2.15). Assume, without loss of generality that a1 6= 0 and write
1 a2 an
v1 = v1 v2 . . . vn
a1 a1 a1
(where we make use of the identities in Problem 2.3). Since {vi }ni=1 is a basis any element v V
can be expressed as a linear combination of v1 , v2 , . . . , vn and hence, by the above equation, as a
linear combination of v1 , v2 , . . . , vn . In particular, we can write
v2 = b1 v1 + b2 v2 + . . . + bn vn .
Note again that the b2 , . . . , bn cannot all be zero; otherwise v2 = b1 v1 and {vi }n+1
i=1 cannot be linearly
independent. Assume, without loss of generality, that b2 6= 0 and write
1 b1 bn
v2 = v v . . . vn .
b2 2 b2 1 b2
This shows that every vector is V can be written as a linear combination of v1 , v2 , v3 , . . . , vn .
Repeat for v3 , etc. until finally

vn+1 = c1 v1 + c2 v2 + . . . + cn vn .
This, however, contradicts the assumption that the set {vi }ni=1 is linearly independent.
The fact provides conditions under which the number of elements of a basis of a linear space in well
defined, and independent of the choice of the basis. A similar statement can be made for linear
spaces which do not have a basis with a finite number of elements. In this case there are again
families of vectors whose span covers almost all elements of the vector space. The proof, however,
is considerably more involved. It also relies on concepts covered in Chapters 3 and 7 that go beyond
the purely algebraic structure considered in this chapter.
Definition 2.11 Let (V, F ) be a linear space. If a basis of (V, F ) with a finite number of elements
exists, the number of elements of this basis is called the dimension of (V, F ) and (V, F ) is called
finite dimensional. If not, (V, F ) is called infinite dimensional.
Exercise 2.17 If (V, F ) has dimension n then any set of n+ 1 or more vectors is linearly dependent.
Example (Bases) The vectors {(1, 0, 0), (0, 1, 0), (0, 0, 1)} form a basis for the linear space R3 over
the field R. This is called the canonical basis of R3 and is usually denoted by {e1 , e2 , e3 }. However
other choices of basis are possible, for example {(1, 1, 0), (0, 1, 0), (0, 1, 1)}. In fact any three linearly
independent vectors will form a basis for R3 . This shows that R3 is finite dimensional and of
dimension 3. In the same way, the vectors {(1, 0, 0), (0, 1, 0), (0, 0, 1)} form the canonical basis for
the linear space C3 over the field C; therefore C3 over C has dimension 3. On the other hand, C3
is a linear space also over the field R; in this case it has dimension 6, the following being a basis:
{(1, 0, 0), (i, 0, 0), (0, 1, 0), (0, i, 0), (0, 0, 1), (0, 0, i)}.
Exercise 2.18 Find a basis for Rn over the field R. Hence, show that Rn has dimension n over R.
Show that Cn has dimension n over C and 2n over R.
The linear space F ([1, 1], R) is infinite dimensional. We have already shown that the collection
{tk | k N} F ([1, 1], R) is linearly independent. The collection contains an infinite number of
elements and may or may not span F ([1, 1], R). Therefore any basis of F ([1, 1], R) (which must
by definition span the set) must contain at least as many elements.
Let {b1 , b2 , . . . , bn } be a basis of a finite dimensional linear space (V, F ). By definition,

n
X
Span({b1 , . . . , bn }) = V therefore x V 1 , . . . , n F : x = i bi .
i=1
The vector = (1 , . . . , n ) F n is called the representation of x V with respect to the basis

{b1 , b2 , . . . , bn }.
Fact 2.8 The representation of a given x V with respect to a basis {b1 , . . . , bn } is unique.
The proof is left as an exercise (Problem ??).

Representations of the same vector with respect to different bases can of course be different.
Example (Representations) Let x = (x1 , x2 , x3 ) (R3 , R). The representation of x with respect
to the canonical basis is simply = (x1 , x2 , x3 ). The representation with respect to the basis
{(1, 1, 0), (0, 1, 0), (0, 1, 1)}, however, is = (x1 , x2 x1 x3 , x3 ) since

1 0 0 1 0 0
x = x1 0 + x2 1 + x3 0 = x1 1 + (x2 x1 x3 ) 1 + x3 1 .
0 0 1 0 0 1
Representations can also be defined for infinite dimensional spaces, but we will not get into the
details here. As an example, consider f (t) C ([1, 1], R). One can consider a representation of
f (t) with respect to the basis {tk | k N} defined through the Taylor series expansion. For example,
expansion about t = 0 gives
df 1 d2 f
f (t) = f (0) + (0)t + (0)t2 + . . . .
dt 2 dt2
2
which suggests that the representation of f (t) is = (f (0), df 1d f
dt (0), 2 dt2 (0), . . .). Making this state-
ment formal, however, is beyond the scope of these notes.
It turns out that all representations of an element of a linear space are related to one another:
Knowing one we can compute all others. To do this we need the concept of linear maps.
2.5 Linear maps

Definition 2.12 Let (U, F ) and (V, F ) be two linear spaces. The function A : U V is called
linear if and only if u1 , u2 U , a1 , a2 F
A(a1 u1 + a2 u2 ) = a1 A(u1 ) + a2 A(u2 ).
Note that both linear spaces have to be defined over the same field. For clarity we will sometimes
write A : (U, F ) (V, F ) if we need to specify the field over which the linear spaces are defined.
Example (Linear maps) Let (U, F ) = (Rn , R), (V, F ) = (Rm , R) and consider a matrix A Rmn .
Define
A : Rn Rm
u 7 A u.
It is easy to show that A is a linear map. Indeed:
A(a1 u1 + a2 u2 ) = A (a1 u1 + a2 u2 ) = a1 A u1 + a2 A u2 = a1 A(u1 ) + a2 A(u2 ).
Consider now f C([0, 1], R) and define the functions

A : (C([0, 1], R), R) (R, R)
Z 1
f 7 f (t)dt (integration)
0
A : (C([0, 1], R), R) (F ([0, 1], R), R)
Z t
f 7 g(t) = ea(t ) f ( )d (convolution with eat ).
0
Exercise 2.19 Show that the functions A and A are both linear.
It is easy to see that linear maps map the zero element of their domain to the zero element of their
co-domain.
Exercise 2.20 Show that if A : U V is linear then A(U ) = V .
Other elements of the domain may also be mapped to the zero element of the co-domain, however.
Definition 2.13 Let A : U V linear. The null space of A is the set
Null(A) = {u U | A(u) = V } U
and the range space of A is the set
Range(A) = {v V | u U : v = A(u)} V.
The word space in null space and range space is of course not accidental.
Fact 2.9 Show that Null(A) is a linear subspace of (U, F ) and Range(A) is a linear subspace of
(V, F ).
The proof is left as an exercise (Problem 2.5). It is easy to see that the properties of the null and
range spaces are closely related to the injectivity and surjectivity of the corresponding linear map,
and hence its invertibility (Problem ??).
Theorem 2.1 Let A : U V be a linear map and let b V .
1. A vector u U such that A(u) = b exists if and only if b Range(A). In particular, A is

surjective if and only if Range(A) = V .
2. If b Range(A) and for some u0 U we have that A(u0 ) = b then for all u U :
A(u) = b u = u0 + z with z Null(A).
3. If b Range(A) there exists a unique u U such that A(u) = b if and only if Null(A) =
{U }. In other words, A is injective if and only if Null(A) = {U }.
Proof: Part 1 follows by definition.

Part 2, ():
u = u0 + z A(u) = A(u0 + z) = A(u0 ) + A(z) = A(u0 ) = b.
Part 2, ():
A(u) = A(u0 ) = b A(u u0 ) = V z = (u u0 ) Null(A).
Part 3 follows from part 2.
Finally, we generalise the concept of an eigenvalue to more general linear maps of a (potentially
infinite dimensional) linear space.
Definition 2.14 Let (V, F ) be a linear space and consider a linear map A : V V . An element
F is called an eigenvalue of A if and only if there exists v V such that v 6= V and A(v) = v.
In this case, v is called an eigenvector of A for the eigenvalue .
Example (Eigenvalues) For maps between finite dimensional spaces defined by matrices, the in-
terpretation of eigenvalues is the familiar one from linear algebra. Since eigenvalues and eigenvectors
are in general complex numbers/vectors we consider matrices as linear maps between complex finite
dimensional spaces (even if the entries of the matrix itself are real). For example, consider the linear
space (C2 , C) and the linear map A : C2 C2 defined by the matrix

0 1
A=
1 0
through matrix multiplication; in other words, for all x C2 , A(x) = Ax.

Itis easy to see that the eigenvalues of A are 1 = j and 2 = j. Moreover, any vector of the
form
j j
c for any c C with c 6= 0 is an eigenvector of 1 and any vector of the form c is an
1 1
eigenvector of 2 .
Definition 2.14 also applies to infinite dimensional spaces, however. Consider, for example, the
linear space (C ([t0 , t1 ], R), R) of infinitely differentiable real valued functions of the interval [t0 , t1 ].
Consider also the function A : C ([t0 , t1 ], R) C ([t0 , t1 ], R) defined by differentiation, i.e. for all
f : [t0 , t1 ] R infinitely differentiable define
df
(A(f ))(t) = (t), t [t0 , t1 ].
dt
Exercise 2.21 Show that A is well defined, i.e. if f C ([t0 , t1 ], R) then A(f ) C ([t0 , t1 ], R).
Show further that A is linear.
One can see that in this case the linear map A has infinitely many eigenvalues. Indeed, any function
f : [t0 , t1 ] R of the form f (t) = et for R is an eigenvector with eigenvalue , since
d t
(A(f ))(t) = e = et = f (t), t [t0 , t1 ]
dt
which is equivalent to A(f ) = f .
Exercise 2.22 Let A : (V, F ) (V, F ) be a linear map and consider any F . Show that the set
{v V | A(v) = v} is a subspace of V .
2.6 Linear maps generated by matrices

Two of the examples in the previous section made use of matrices to define linear maps between
finite dimensional linear spaces. It is easy to see that this construction is general: Given a field F ,
every matrix A F mn defines a linear map A : (F n , F ) (F m , F ) by matrix multiplication, i.e.
for all x = (x1 , . . . , xn ) F n ,
Pn
j=1 a1j xj
..
A(x) = Ax = .
Pn .
j=1 amj xj
Naturally, for linear maps between finite dimensional spaces that are generated by matrices a close
relation between the properties of the linear map and those of the matrix exist.
Definition 2.15 Let F be a field, A F nm be a matrix and A : (F m , F ) (F n , F ) the linear

map defined by A(x) = Ax for all x F m . The rank of the matrix A (denoted by Rank(A)) is the
dimension of the range space, Range(A), of A. The nullity of A (denoted by Nullity(A)) is the
dimension of the null space, Null(A), of A.
The following facts can be used to link the nullity and rank of matrices. They will prove useful when
manipulating matrices later on.
Theorem 2.2 Let A F nm and B F mp .
1. Rank(A) + Nullity(A) = m.
2. 0 Rank(A) min{m, n}.
3. Rank(A) + Rank(B) m Rank(AB) min{Rank(A), Rank(B)}.
4. If P F mm and Q F nn are invertible then
Rank(A) = Rank(AP ) = Rank(QA) = Rank(QAP )

Nullity(A) = Nullity(AP ) = Nullity(QA) = Nullity(QAP ).

For square matrices there exists a close connection between the properties of A and the invertibility
of the matrix A.
Theorem 2.3 Let F be a field, A F nn be a matrix, and A : F n F n the linear map defined
by A(x) = Ax for all x F n . The following statements are equivalent:
1. A is invertible.
2. A is bijective.
3. A is injective.
4. A is surjective.
5. Rank(A) = n.
6. Nullity(A) = 0.
7. The columns aj = (a1j , . . . , anj ) F n form a linearly independent set {aj }nj=1 .
8. The rows ai = (ai1 , . . . , ain ) F n form a linearly independent set {ai }ni=1 .

Finally, for linear maps defined through matrices, one can give the usual interpretation to eigenvalues
and eigenvectors.
Theorem 2.4 Let F be a field, A F nn be a matrix, and A : F n F n be the linear map defined
by A(x) = Ax for all x F n . The following statements are equivalent:
1. C is an eigenvalue of A.
2. Det(I A) = 0.
3. There exists v Cn such that v 6= 0 and Av = v. Such a v is called a right eigenvector of A
for the eigenvalue .
4. There exists Cn such that 6= 0 and T A = T . Such an is called a left eigenvector of

A for the eigenvalue .
The proof is left as an exercise (Problem ??). The theorem shows (among other things) that the
eigenvalues of a linear map defined by a square matrix A F nn can be computed by finding the
roots of the characteristic polynomial of the matrix, defined by
A () = Det[I A] = sn + 1 sn1 + . . . + n1 s + n .
Notice that by definition the characteristic polynomial is a monic polynomial (the leading coefficient
is equal to 1) of degree n.
Definition 2.16 The spectrum of a matrix2 A F nn is the list of eigen values of A, denoted by
Spec[A] = {1 , . . . , n }.
Finally, combining the notion of eigenvalue with Theorem 2.3 leads to the following condition for
the invertibility of a matrix.
Theorem 2.5 A matrix A F nn is invertible if and only if none of its eigenvalues are equal to
the zero element 0 F .
2.7 Matrix representation of linear maps

In the previous section we saw that every matrix A F mn defines a linear map A : F n F m by
simple matrix multiplication. In fact it can be shown that the opposite is also true: Any linear map
between two finite dimensional linear spaces can be represented as a matrix by fixing bases for the
two spaces.
Consider a linear map
A : U V
between two finite dimensional linear spaces (U, F ) and (V, F ). Assume that the dimension of (U, F )
is n and the dimension of (V, F ) is m. Fix bases {uj }nj=1 for (U, F ) and {vi }m
i=1 for (V, F ). Let
yj = A(uj ) V for j = 1, . . . , n.
The vectors yj V can all be represented with respect to the basis {vi }m
i=1 of (V, F ). In other words
for all j = 1, . . . , n there exist unique aij F such that
m
X
yj = A(uj ) = aij vi .
i=1
The aij F can then be used to form a matrix

a11 . . . a1n
.. .. .. F mn .
A= . . .
am1 . . . amn
Since representations are unique (Fact 2.8), the linear map A and the bases {uj }nj=1 and {vi }m
i=1
uniquely define the matrix A F mn .
2 Sometimes a distinction is made between the spectrum of the matrix (list of eigenvalues with repeated eigenvalues
included) and the spectra list of a matrix (list of eigenvalues without repetitions). We will however not make this
distinction here.
Consider now an arbitrary x U . Again there exists unique representation F n of x with respect
to the basis {uj }nj=1 ,
Xn
x= j uj
j=1
Let us now see what happens to this representation if we apply the linear map A to x. Clearly
A(x) V , therefore there exists a unique representation F m of A(x) with respect to the basis
{vi }m
i=1 such that
Xm
A(x) = i vi .
i=1
What are the corresponding i F ? Recall that

Xn Xn n
X m
X m
X Xn
A(x) = A j uj = j A(uj ) = j aij vi = aij j vi .
j=1 j=1 j=1 i=1 i=1 j=1
By uniqueness of representation
n
X
i = aij j = A ,
j=1
where denotes the standard matrix multiplication. Therefore, when one looks at the representations
of vectors with respect to given bases, application of the linear map A to x U is equivalent to
multiplication of its representation (an element of F n ) with the matrix A F mn . To illustrate
this fact we will write things like
A
(U, F ) (V, F )
x 7 A(x)
mn
AF
{uj }nj=1 {vi }m
i=1
Fn 7 A F m
Theorem 2.6 The following relations between linear map operations and the corresponding matrix
representations hold:
1. Consider linear maps B : (U, F ) (V, F ) and A : (V, F ) (W, F ) where U , V and W
are finite dimensional linear spaces of dimensions n, m and p respectively. Then composition
C = A B : (U, F ) (W, F ) is also a linear map. Moreover, if we fix bases {uk }nk=1 , {vi }m
i=1
and {wj }pj=1 for the three spaces and
B A
(U, F ) (V, F ) (V, F ) (W, F )
BF mn and AF pm
{uk }nk=1 {vi }m
i=1 {vi }m
i=1 {wj }pj=1
then
C=AB
(U, F ) (W, F )
C=ABF pn
{uk }nk=1 {wj }pj=1
where denotes the standard matrix multiplication.
2. Consider an invertible linear map A : (V, F ) (V, F ) on an n-dimensional linear space V
and let A1 : (V, F ) (V, F ) denote its inverse. If A is the representation of A with respect
to a given basis of V , then A1 is the representation of A1 with respect to the same basis.
The proof is left as an exercise (Problem 2.8). Analogous statements can of course be made about
the representations of linear maps obtained by adding, or scaling other linear maps.
2.8 Change of basis

Given a linear map, A : (U, F ) (V, F ), selecting different bases for the linear spaces (U, F ) and
(V, F ) leads to different representations.
A
(U, F ) (V, F )
mn
AF
{uj }nj=1 {vi }m
i=1
AF mn
{uj }nj=1 {vi }m
i=1
In this section we investigate the relation between the two representations A and A.
Recall first that changing basis changes the representations of all vectors in the linear spaces. It is
therefore expected that the representation of a linear map will also change.
Example (Change of basis) Consider x = (x1 , x2 ) R2 . The representation of x with respect to

the canonical basis {u1 , u2 } = {(1, 0), (0, 1)} is simply = (x1 , x2 ). The representation with respect
to the basis {u1 , u2 } = {(1, 0), (1, 1)} is = (x1 x2 , x2 ) since

1 0 1 1
x = x1 + x2 = (x1 x2 ) + x2 .
0 1 0 1
To derive the relation between A and A, consider first the identity map
1U
(U, F ) (U, F )
x 7 1U (x) = x
IF nn
{uj }nj=1 {uj }nj=1
QF nn
{uj }nj=1 {uj }nj=1
I denotes the usual identity matrix in F nn

1 0 ... 0
0 1 ... 0

I= . . . F nn
.. .. . . ...

0 0 ... 1
where 0 and 1 are the addition and multiplication identity of F . The argument used to derive the
representation of a linear map as a matrix in Section 2.7 suggests that the elements of Q F nn
are simply the representations of 1U (uj ) = uj (i.e. the elements of the basis {uj }nj=1 ) with respect
to the basis {uj }nj=1 . Likewise
1
V
(V, F ) (V, F )
x 7 1V (x) = x
IF mm
{vi }m
i=1 {vi }m
i=1
mm
P F
{vi }m
i=1 {vi }m
i=1
Exercise 2.23 Show that the matrices Q F nn and P F mm are invertible. (Recall that 1U
and 1V are bijective functions.)
Example (Change of basis (cont.)) Consider the identity map

1 2
R2 R
R2
x 7 x

1 0 22
I= R
0 1
{(1, 0), (0, 1)} {(1, 0), (0, 1)}
(x1 , x2 ) 7 (x1 , x2 )
On the other hand,

1 2
R2 R
R2
x 7 x

1 1 22
Q= R
0 1
{(1, 0), (1, 1)} {(1, 0), (0, 1)}
(x1 , x2 ) 7 (x1 , x2 ) = (x1 + x2 , x2 )
and
1 2
R2 R
R2
x 7 x

1 1 R22
Q1 =
0 1
{(1, 0), (0, 1)} {(1, 0), (1, 1)}
(x1 , x2 ) 7 (x1 , x2 ) = (x1 x2 , x2 )
Now notice that A = 1V A 1U , therefore

1
U A V 1
(U, F ) (U, F ) (V, F ) (V, F )
nn
QF AF mn P F mm
{uj }nj=1 {uj }nj=1 {vi }m
i=1 {vi }m
i=1 .
By Theorem ?? the representation of the linear map
A = 1V A 1U : (U, F ) (V, F )
with respect to the bases {uj }nj=1 and {vi }m

i=1 is
A = P A Q F mn
where A is the representation of A with respect to the bases {uj }nj=1 and {vi }m
i=1 and denotes
ordinary matrix multiplication. Since P F mm and Q F nn are invertible it is also true that
A = P 1 A Q1 .
Matrices that are related to each other in this way will play a central role in subsequent calculations.
We therefore give them a special name.
Definition 2.17 Two matrices A F mn and A F mn are equivalent if and only if there exist
Q F nn , P F mm both invertible such that A = P A Q.
The discussion above leads immediately to the following conclusion.
Theorem 2.7 Two matrices are equivalent if and only if they are representations of the same linear
map.
Proof: The if part follows from the discussion above. For the only if part use the matrices to
define linear maps from F n to F m .
As a special case, consider the situation where (U, F ) = (V, F )

A
(U, F ) (U, F )
nn
AF
{uj }nj=1 {uj }nj=1
AF nn
{uj }nj=1 {uj }nj=1
As before
1
U A 1
U
(U, F ) (U, F ) (U, F ) (U, F )
nn
QF AF nn 1
Q F nn
{uj }nj=1 {uj }nj=1 {uj }nj=1 {uj }nj=1
therefore
A = Q1 A Q.
The last equation is usually referred to as the change of basis formula; the matrix Q is variously
referred to as a similarity transformation, a change of basis, a coordinate transformation, or a change
of coordinates. It can be seen that many of the matrix operations from linear algebra involve such
changes of basis: Elementary row/column operations, echelon forms, diagonalization, rotations, and
many others.
We shall see that many of the operations/properties of interest in linear system theory are unaffected
by change of basis in the state input or output space: Controllability, observability, stability, transfer
function, etc. This is theoretically expected, since these are properties of the system in question and
not the way we choose to represent it through a choice of basis. It is also very useful in practice:
Since we are effectively free when choosing the basis in which to represent our system, we will often
use specific bases that make the calculations easier.

Problem 2.1 (Groups)
1. Consider an arbitrary set S. Show that the set of bijective functions f : S S forms a group
under the operation of function composition. What does this group correspond to in the case
where S is finite?

2. Draw an equilateral triangle in R2 , say with vertices at (1, 0), ( 12 , 23 ) and ( 12 , 2 3 ). Con-
sider the set of symmetries of this triangle, i.e. the set of all functions f : R2 R2 that
map the triangle onto itself. Show that this set forms a group under the operation of function
composition. How many elements does this group contain? What is the identity element?
Repeat for the case of a square and a circle.
Problem 2.2 (Rings and Fields)
1. Let (R, +, ) be a ring and consider elements , R. Show that 0 = 0 = 0 and

() = () = ( ) where as usual 0 R denotes the addition identity and the
addition inverse of element R.
2. Let (R, +, ) be a ring. Show that there exists a unique element 0 R such that for all R,
+ 0 = 0 + = . Moreover, show that for all R there exists a unique () R such
that + () = () + = 0.
3. Let (F, +, ) be a field and consider , , F . Show that
6= 0, = = .
Is the same true for a ring? Justify your answer.
Problem 2.3 (Identity and inverse properties) Let (V, F, , ) be a linear space and 0 be the
addition identity element of F . Show that for all x V , 0 x = . Moreover, show that for all
a F , x V , (a) x = (a x) = a (x).
Problem 2.4 (Examples of linear spaces) Let (F, +, ) be a field.
1. Consider x = (1 , . . . , n ) F n and y = (1 , . . . , n ) F n and F . Define : F n F n

F n and : F F n F n by
x y = (1 + 1 , . . . , n + n )
x = ( 1 , . . . , n ).
Show that (F n , F, , ) is a linear space. What is the addition identity? What is the addition
inverse?
2. Let now (V, F, +, ) be a linear space and D an arbitrary set. Let F (D, V ) denote the set of all
functions f : D V . For f, g F (D, V ) and F define f g : D V and f : D V
by
(f g)(d) = f (d) + g(d)

( f )(d) = f (d)
for all d D. Show that (F (D, V ), F, , ) is a linear space. What is the addition identity?
What is the addition inverse.
Problem 2.5 (Subspaces) 1. Let U and V be linear spaces and let L(U, V ) denote the set of
linear functions A : U V . Show that L(U, V ) is a linear subspace of F (U, V ), the space
of all functions mapping U into V with the usual operations of function addition and scalar
multiplication.
2. Let U and V be linear spaces and A : U V be a linear function. Show that Range(A) is a
subspace of V and Null(A) is a subspace of U .
3. Let {Wi }ni=1 be a finite family of subspaces of V . Show that the the intersection and the direct
sum of these subspaces
n
\
Wi = {v V | i = 1, . . . , n, v Wi }
i=1
Mn
Wi = {v V | wi Wi , i = 1, . . . n, v = w1 + . . . wn }
i=1
are themselves subspaces of V .
Problem 2.6 (Basis and vector representation) Let V be a finite dimensional linear space.
1. Let W be a subspace of V . Show that W is also finite dimensional and its dimension can be
no greater than that of V .
2. Show that the representation of a given x V with respect to a basis {b1 , . . . , bn } is unique.
Problem 2.7 (Rank and nullity) Let (F, +, ) be a field and consider the linear maps A : (F n , F )
(F m , F ) and B : (F p , F ) (F n , F ) represented by matrices A F mn and B F pn respectively.
Show that:
1. 0 Rank(A) min{n, m} and Rank(A) + Nullity(A) = n.

2. Rank(A) + Rank(B) n Rank(BA) min{Rank(A), Rank(B)}.
(Hint: Let A : Range(B) F m be the restriction of A to Range(B). Then:

(a) Range(A B) = Range(A ) Range(A),
(b) Null(A ) Null(A).
To show part 2 apply the result from part 1 to A .)
Problem 2.8 (Invertible matrices) Let F be a field, A F nn be a matrix, and A : F n F n

the linear map defined by A(x) = Ax for all x F n . Show that the following statements are
equivalent:
1. A is invertible.
2. None of the eigenvalues of A is equal to zero.
3. A is bijective.
4. A in injective.
5. A is surjective.
6. Rank(A) = n.
7. Nullity(A) = 0.
8. The columns aj = (a1j , . . . , anj ) F n form a linearly independent set {aj }nj=1 .
9. The rows ai = (ai1 , . . . , ain ) F n form a linearly independent set {ai }ni=1 .
Problem 2.9 (Matrix representation properties)
1. Consider linear maps A : (V, F ) (W, F ) and B : (U, F ) (V, F ). Assume that U, V, W
have finite dimensions m, n, p respectively, and that A and B have representations A F pn
and B F nm with respect to given bases for the three spaces. Show that the composition
C = A B : (U, F ) (W, F ) has representation C = AB with respect to the same bases.
2. Consider a linear map A : (U, F ) (U, F ) where U has finite dimension n. Assume that A
has representation A F nn with respect to a given basis for U . Show that if A is invertible,
then A1 has representation A1 with respect to the same basis.
Problem 2.10 (Matrix representation examples)
1. Consider a linear map A : (U, F ) (U, F ) where U has finite dimension n. Assume there
exists a vector b U such that the collection {b, A(b), A A(b), . . . , An1 (b)} forms a basis for
U . Derive the representation of A and b with respect to this basis.
2. Consider a linear map A : (U, F ) (U, F ) where U has finite dimension n. Assume there exists
a basis bi , i = 1, . . . , n for U such that A(bn ) = bn and A(bi ) = bi + bi+1 , i = 1, . . . , n 1.
Derive the representation of A with respect to this basis.
3. Consider two matrices A, A Rnn related by a similarity transformation; i.e. there exists
Q Rnn invertible such that A = Q1 AQ. Show that Spec[A] = Spec[A].
Problem 2.11 (Matrix eigenvalues) Let F be a field, A F nn be a matrix, and A : F n F n

be the linear map defined by A(x) = Ax for all x F n . The following statements are equivalent:
1. C is an eigenvalue of A.
2. Det(I A) = 0.
3. There exists v Cn such that v 6= 0 and Av = v.
4. There exists Cn such that 6= 0 and T A = T .
Chapter 3
Introduction to Analysis
Consider a linear space (V, F ) and assume the F = R or F = C so that for a F the absolute value
(or modulus) |a| is well defined.
3.1 Norms and continuity

Definition 3.1 A norm on a linear space (V, F ) is a function k k : V R+ such that:
1. v1 , v2 V , kv1 + v2 k kv1 k + kv2 k (triangle inequality).

2. v V , a F , kavk = |a| kvk.
3. kvk = 0 v = V .
A linear space equipped with such a norm is called a normed linear space and is denoted by (V, F, kk).
A norm provides a notion of lenght for an element of a linear space. The norm can also be used
to define a notion of distance between elements of a linear space; one can think of kv1 v2 k as
the distance between two vectors v1 , v2 V .
Example (Normed spaces) In (Rn , R), the following are examples of norms:
n
X
kxk1 = |xi |, (1-norm)
i=1
v
u n
uX
kxk2 = t |xi |2 (Euclidean or 2-norm)
i=1
n
! p1
X
p
kxkp = |xi | for p 1, (p-norm)
i=1
kxk = max |xi | (infinity norm)
i=1,...,n
Exercise 3.1 Show that kxk1 , kxk2 , and kxk satisfy the axioms of a norm.
33
In fact the same definitions would also hold for the linear space (Cn , C). Different norms of course
lead to different notions of distance for the same linear space. One way to visualize these different
distance notions is through the so-called open balls.
Definition 3.2 If (V, F, kk) is a normed linear space, the (open) ball of radius r R+ centered at v V
is the set
B(v, r) = {v V | kv v k < r}.
It can be shown that B(v, r) is an open set in the sense of definition 3.3. The ball B(V , 1) is called
the unit ball of (V, F, k k).
Exercise 3.2 Draw the unit balls of the normed linear spaces (R2 , R, k k1 ), (R2 , R, k k2 ), and
(R2 , R, k k ).
Definition 3.3 A set S V is called bounded if S B(V , r) for some r.
For example, any ball B(v, r) is bounded, because B(v, r) B(V , kvk + r + 1).
Definition 3.4 Let (V, F, k k) be a normed space. A function v : N V is called a sequence in V .

The sequence converges to a point v V if and only if
> 0 N N m N, kv(m) vk < .
In this case v is called the limit of the sequence.
To denote sequences, we will mostly use the notation {vi } i=0 , where i N plays the role of the
i
index variable. To indicate that a sequence converges to some vector v V we write vi v
or limi vi = v. Note that by the definition of the norm, the statement the sequence {vi } i=0
converges to v V is equivalent to limi kvi vk = 0 in the usual sense in R.
Using the definition one can also define open and closed sets.
Definition 3.5 Let (V, F, k k) be a normed space. A set K V is called closed if and only if it
contains all its limit points; i.e. if and only if for all sequences {vi }
i=1 K if vi v V then
v K. K is called open if and only if its complement V \ K is closed.
Some useful properties of open and closed sets are stated in ??.
Consider now functions between two normed spaces
f : (U, F, k kU ) (V, F, k kV )
Definition 3.6 f is called continuous at u U if and only if

> 0 > 0 such that ku u kU < kf (u) f (u )kV < .
The function is called continuous on U if and only if it is continuous at every u U .
In other words, f is continuous at u if and only if

> 0 > 0 such that u B(u, ) f (u ) B(f (u), ).
or, with the usual notation, limu u f (u ) = f (u).
Continuity expresses the requirement that small changes in u U should lead to small changes in
f (u) V .
Exercise 3.3 Show that the set of continuous functions f : (U, F, k kU ) (V, F, k kV ) is a linear
subspace of F (U, V ).
Fact 3.1 The norm k k, as a function between the normed linear spaces (V, F, k k) and (R, R, | |),
is continuous on V .
Proof: The triangular inequality implies kxk = kx x0 + x0 k kx x0 k + kx0 k, and therefore

kxk kx0 k kx x0 k. Interchanging the role of x and x0 , we also get |kxk kx0 k| kx x0 k.
Now in the definition of continuity of k k at x0 :
> 0 > 0 kx x0 k < |kxk kx0 k| <
it sufficies to choose = to obtain |kxk kx0 k| kx x0 k < = .
Continuity of functions has an intimate relationship with the convergence of sequences and hence
with open and closed sets. This relationship is highlighted in Problem ??.
Note that in the definitions of convergence and continuity we have left open the choice of the norm.
One may wonder if, by choosing another norm, a a convergent sequence may cease to converge, or a
continuous function can cease to be continuous. To investigate this question we need the notion of
equivallent norms.
Definition 3.7 Consider a linear space (V, F ) with two norms, k ka and k kb . The two norms are
equivalent if and only if
mu ml > 0, v V ml kvka kvkb mu kvka .
Exercise 3.4 Show that in this case it is also true that ml , mu R+ such that v V , ml kvkb
kvka mu kvkb .
Example (Equivalent norms) Consider x Rn and the kxk1 and kxk norms defined above.
Then
n
X
kxk = max |xi | |xi | = kxk1
i=1,...,n
i=1
n
X n
X
kxk1 = |xi | max |xj | = n max |xi | = nkxk .
j=1,...,n j=1,...,n
i=1 i=1
Therefore kxk kxk1 nkxk and the two norms are equivalent (take ml = 1 and mu = n).
Fact 3.2 Let kka and kkb be two equivallent norms on a linera space (V, F ) with F = R or F = C.
A sequence {vi }i=0 V converges to some v V in (V, F, k ka ) if and only if it converges to v in
(V, F, k kb ).
Proof: Suppose that the sequence {vi } i=0 converges to v V with respect to the norm k ka , that
is for all > 0 there exists N N such that kvm vka < for all m N . Due to equivalence of the
norms, there exists mu > 0 such that ml kvka kvkb mu kvka for all v. Fix an arbitrary > 0.
Then there exists N N such that, for all m N , kvm vka < mu ; But then, with the same N ,
for all m N

kvm vkb mu kvm vka < mu = .
mu
Since function continuity, open and closed sets were all defined in terms of convergence of sequences,
this fact further implies that open/closed sets defined using one norm remain open/closed for any
other equivallent norm. Likewise, continuous functions defined with respect to a pair of norms
remain continuous with respect to any other pair of respectively equivalent norms.
The fact that kxk1 and kxk are equivalent norms on Rn is not a coincidence. A remarkable result
states, indeed, that any two norms on a finite-dimensional space are equivalent. To show this, we
will use the following fact, which is indeed a corollary of two fundamental theorems in real analysis.
Fact 3.3 (A corollary to the theorems of Heine-Borel and of Weierstrass). A continuous function
f : S R defined on a closed and bounded subset S of a finite-dimensional linear space V attains
a maximum M R and a minimum m R on S. In other words, there exist xM , xm S such that
M := sup f (x) = f (xM ) < +

xS
m := inf f (x) = f (xm ) >
xS
Proposition 3.1 Any two norms on a finite-dimensional space V are equivalent.
Proof: Assume that a certain norm k k has already been defined on V ; say that this is the norm
with respect to which the concepts of ball, convergence, and continuity have been defined. Let
kkA be another norm on V . The unit sphere S = {x V | kxk = 1} is bounded. Moreover, S is also
closed, as it is the inverse image of the closed set {1} under the continuous function k k : V R
(see Problem ?? and Fact ??). Therefore k kA , being continuous, admits a minimum m and a
maximum M on S. In particular M m > 0, because kxkA = 0 x = 0 kxk = 0 (it is not the
case for x S). Now, if x V and x 6= 0, x/kxk S. Hence,

x
kxkA = kxk mkxk
kxk A

x
kxkA = kxk M kxk
kxk A
The above inequalities are trivially true also when x = 0, and prove that k kA is equivalent to
k k. Since the choice of k kA was left open, it is immediate to see that any two norms, being both
equivalent to k k, are equivalent to each other.
3.2 Infinite-dimensional normed spaces

Consider now the linear space C([t0 , t1 ], Rn ) of all the continuous functions f : [t0 , t1 ] Rn . The
function k k : C([t0 , t1 ], Rn ) R+ defined by
kf k := sup |f (t)|
t[t0 ,t1 ]
is a norm on (C([t0 , t1 ], Rn ), R). Indeed, for all continuous functions f , kf k is a positive or null
quantity, and it is trivially equal to zero if and only if f (t) is the zero-function. Moreover,
kf k = sup |f (t)| = || sup |f (t)| = || kf k ,

t[t0 ,t1 ] t[t0 ,t1 ]
kf + gk = sup |f (t) + g(t)| sup |f (t)| + sup |g(t)| = kf k + kgk .

t[t0 ,t1 ] t[t0 ,t1 ] t[t0 ,t1 ]
Norms on function spaces can also be defined using integration. For example, for f C([t0 , t1 ], Rn )
we can define the following:
Z t1
kf k1 = kf (t)k2 dt
t
s0Z
t1
kf k2 = kf (t)k22 dt
t0
Z t1 p1
kf kp = kf (t)kp2 dt
t0
Here, the integral takes the role that the finite summation has in finite-dimensional spaces. Since
the 2-norm that appears in the integrands is a continuous function (Fact??), the integrands are also
continuous functions of the variable t, and all these quantities are well-defined. Moreover, all of them
are norms. Take for example the first one. Of course we have kf k1 0 for all f C([t0 , t1 ], Rn ); the
norm of the zero-function is of course equal to zero; on the other hand, given a continuous function
f , if f is nonzero at a point it must be nonzero in a whole subinterval of [t0 , t1 ], therefore kf k1 is
also nonzero. (Summing up, kf k1 = 0 if and only if f = 0.) Moreover,
Z t1 Z t1
kf k1 = |f (t)| dt = || |f (t)| dt = || kf k1
t0 t0
Z t1 Z t1
kf + gk1 = |f (t) + g(t)| dt (|f (t)| + |g(t)|) dt = kf k1 + kgk1
t0 t0
It is also easy to see that the use of the 2-norm in the integrands is arbitrary: Since norms in Rn
are all equivallent one could have used any other norm, leading to other norms on C([t0 , t1 ], Rn ).
Exercise 3.5 Show that the function norms obtained by replacing kf (t)k2 by another norm on Rn
in the integrands are equivalent.
Motivated by this fact and in analogy to finite-dimensional spaces, the reader may be tempted to
think that the all the norms introduced above are equivalent. Unfortunately, this is not the case.
Example (Non-equivalent norms) Consider the functions f C([0, 1], R) and the two norms
Z 1
kf k = sup |f (t)| and kf k1 = |f (t)|dt
t[0,1] 0
We will show that the two norms are not equivalent. Assume, for the sake of contradiction, that
there exist ml , mu R+ such that for all f C([0, 1], R) we have ml kf k1 kf k mu kf k1 . Since
Z 1 Z 1
kf k1 = |f (t)| dt kf k dt = kf k,
0 0
one can indeed choose ml = 1. On the other hand, it is impossible to find mu that satisfies the
other inequality for all f . Assume, for the sake of contradiction, that such an mu exists and consider
1
the family of functions fn (t) = tn C([0, 1], R) for n N. Clearly kfn k = 1 and kfn k1 = n+1 .
Therefore kfa k = (n + 1)kfa k1 . Selecting n + 1 > mu leads to a contradiction, showing that the
two norms are not equivalent.
3.3 Completeness
In our treatment of the solution of differential equations below we will be faced with the task of
establising the convergence of a series in a certain linear space. In principle, one should be able to
do so using Definition 3.2. In practice, however, this will not be possible as the limit is not known
beforehand. From the definition it is, however, apparent that if the sequence {vi } i=0 converges to
some v V the vi need to be getting closer and closer to v, and therefore closer and closer to each
other. Can we decide whether the sequence converges just by looking at kvi vj k?
Definition 3.8 A sequence {vi }

i=0 is called a Cauchy sequence if and only if
> 0 N N m N, kvm vN k < .
Exercise 3.6 Show that a sequence is Cauchy if and only if for all > 0 there exists N N such
that kvm vn k < for all m, n N .
Exercise 3.7 Show that every convergent sequence is Cauchy.
The converse however is not always true.
Definition 3.9 The normed space (V, F, k k) is called complete (or Banach) if and only if every
Cauchy sequence converges.
Fortunately many of the spaces we are interested in are complete.
Fact 3.4 The linear space (R, R, | k) is Banach.
This observation is in fact is the reason why the real number system was introduced in the first
place, and is so fundamental that in many expositions of analysis it is taken as an axiom. This fact
is also the main tool used to establishing completeness of many other spaces.
Theorem 3.1 Let F = R or F = C and assume that (V, F ) is finite-dimensional. Then (V, F, k k)
is a Banach space for any norm k k.
Proof: Suppose that F = R. Let {b1 , , bd } be a basis of V , so that every vector v V can be
Pd
represented in an unique way as v = k=1 xk bk for some xk R. On Rd let the function k k be
defined by d
X

kxk = k(x1 , , xd )k := xk bk

k=1
k k is a norm on R ; moreover, {vn }

d
n=0 is Cauchy or convergent in (V, F, k k) if and only if the
corresponding sequence of d-tuples {xn } d
n=0 is Cauchy or convergent in (R , R, k k ) (justify these
d
assertions). Thus, it suffices to show that (R , R, k k ) is complete.
Indeed, since Rd is finite-dimensional, kk is equivalent to kk , and Cauchy or convergent sequences
with respect to one norm are Cauchy or convergent also with respect to the other. Suppose therefore
that {xn }
n=0 is a Cauchy sequence. This implies that
> 0 N N m, n N, kxn xm k < . (3.1)
But then, we have the same property for each of the k-th components:
> 0 N N m, n N, |xnk xm
k | < .
In other words, the k-th components {xnk } n=0 form a Cauchy sequence in R. Since R is complete,
this sequence converges. Define x = (x1 , , xk , , xd ), where
xk := lim xnk
n
Finally, in 3.1, fix N and n, and let m :
> 0 N N n N, kxn xk < .
(we can take the limit within k k because the norm is continuous). Therefore, {xn } n=0 is conver-
gent, and the same happens for {vn } n=0 ; since the latter is arbitrary, (V, F ) is Banach.
To establish the same result for complex spaces (F = C) one can repeat the same proof after showing
that C itself is complete. This should now be easy, and is left as an exercise.
Corollary 3.1 Any finite-dimensional subspace W of a linear space (V, F, k k) is a closed subset
of V .
Proof: According to Definition 3.3, W is closed if for all sequences {wi }

i=1 W , if wi w V ,
then w W . Now if an arbitrary sequence {wi }
i=1 W is convergent in V , then it is Cauchy both
in V and in W . Since W is Banach, its limit point w belongs to W .
This fact will come in handy in Chapter 7 in the context of inner product spaces. The state, input
and output spaces of our linear systems will be of the form Rn , and will all be Banach spaces. One
may be tempted to think that this is more generally true. Unfortunately infinite-dimensional spaces
are less well-behaved.
Example (Non-Banach space) Consider the normed space (C([1, 1], R), R, k k1 ) of continuous
functions f : [1, 1] R with the 1norm
Z 1
kf k1 = |f (t)|dt.
1
For i = 1, 2, . . . consider the sequence of functions (Figure 3.1)

0 if t < 0
fi (t) = it if 0 t < 1/i (3.2)

1 if t 1/i.
It is easy to see that the sequence is Cauchy. Indeed, if we take N, m {1, 2, . . .} with m N ,
Z 1 Z 1/m Z 1/N
kfm fN k = |fm (t) fN (t)|dt = (mt N t)dt + (1 N t)dt
1 0 1/m
1/m 1/N
t2 t2 mN 1
= (m N ) + 1N = .
2 0 2 1/m 2mN 2N
Therefore, given any > 0 by selecting N > 1/(2) we can ensure that for all m N , kfm fN k < .
One can guess that the sequence {fi } i=1 converges to the function

0 if t < 0
f (t) =
1 if t 0.
Indeed Z 1/i
1
kfi f k1 = (1 it)dt = 0.
0 2i
The problem is that the limit function f F ([1, 1], R) is discontinuous and therefore f 6 C([1, 1], R).
Hence C([1, 1], R) is not a Banach space since the Cauchy sequence {fi }ni=1 does not converge to
an element of C([1, 1], R).
Mercifully, there are several infinite dimensional spaces that are Banach.
kfm fN k1
fm (t)
1
kfm fN k
fN (t)
1 1
1 m N 1 t
Figure 3.1: Examples of functions in non-Banach space.
Theorem 3.2 (C([t0 , t1 ], R, Rn ), k k ) is a Banach space.
Proof: We articulate the proof in three steps: First, given a sequence of continuous functions which
is Cauchy in the norm k k , we define a pointwise limit function; Second, we prove that the
sequence converges to this function; Third, we prove that this function is continuous.
Let therefore {fn } n=1 be a Cauchy sequence of functions defined and continuous on the closed
interval [t0 , t1 ]. For each t [t0 , t1 ], {fn (t)} n
n=1 is a Cauchy sequence of vectors in R (why?). Since
n
R is a Banach space (being finite-dimensional!), every such sequence has a limit. We define the
function f as follows:
f (t) := lim fn (t)
n
Now it happens that {fn }converges to f also with respect to the norm k k . Indeed, the fact
n=1
that {fn }
n=1 is Cauchy means that, for every > 0, there exists N N such that
n, m N, t [t0 , t1 ], kfn (t) fm (t)k kfn fm k <
In this equation, fix t and n, and let m :
n N, t [t0 , t1 ], kfn (t) f (t)k
Taking the supremum over t,

n N, kfn f k
hence kfn f k 0.
It remains to show that f is continuous. Let t [t0 , t1 ]. It holds:
kf (t) f (t)k = kf (t) fn (t) + fn (t) fn (t) + fn (t) f (t)k

kf (t) fn (t)k + kfn (t) fn (t)k + kfn (t) f (t)k
Now fix an > 0. Since {fn }n=1 converges to f with respect to k k , there exists n such that
kf fn k < /3, and therefore kf (t) fn (t)k < /3 for all t and in particular kfn (t) f (t)k < /3.
On the other hand, since each function of the sequence is continuous, and in particular so is fn ,
there exists > 0 such that kfn (t) fn (t)k < /3 whenever |t t| < . Thus, for all > 0 there
exists (n and) > 0 such that, if |t t| < , then
kf (t) f (t)k /3 + /3 + /3 = ,
and f is continuous.
Summarizing, given an arbitrary Cauchy sequence {fn } n
n=1 in C([t0 , t1 ], R, R ), we can construct a
n
function f C([t0 , t1 ], R, R ) such that kfn f k 0.
Exercise 3.8 Show that the sequence {fi }

i=1 defined in (??) is not Cauchy in (C([1, 1], R), R, k
k ).
3.4 The space of square-summable functions

We have seen that the space C([t0 , t1 ], Rn ) can be equipped with the norms
Z t1 p1
p
kf kp := |f (t)| dt
t0
In this course, we will need in particular the special case in which p = 2:

sZ
t1
kf k2 := |f (t)|2 dt
t0
Since every continuous function defined on the closed interval [t0 , t1 ] is bounded, its square-modulus
is also continuous, hence integrable (in the Riemann sense, that is, according to the definition of
integral given in calculus courses), so that k k2 is indeed a well-defined norm.
If we consider the space C(R+ , Rn ) of continuous functions over [0, +[ the discourse is different:
Even a continuous and bounded function does not need to be square-integrable (think of a nonzero
R +
constant function: 0 52 dt = +), in which case the norm is not well-defined. Nevertheless, the
space can be restricted. In principle, we could define
Z +
n 2
V = f C(R+ , R ) |f (t)| dt < + (3.3)
0
and it turns out that V would again be a well-behaved normed linear space.
Exercise 3.9 Show that the sum of square-integrable functions is also square-integrable (hence, V
is a subspace of C(R+ , Rn )).
Now, in the perspective of assuming C([t0 , t1 ], Rn ) and C(R+ , Rn ) as the environments for our
further developments, we could wishfully think of proving that C([t0 , t1 ], Rn ) and V are Banach.
Unfortunately, neither of them is complete (The previous example Non-Banach space shows this
fact for the space C([1, 1], R)).
If the only possible pathology was that of a single discontinuity of the limit function, or a finite
number of discontinuities, or even more, we could relax the continuity hypothesis and go for spaces
of piecewise-continuous functions or something like that. New issues would arise. Namely, if we allow
discontinuous functions to enter the picture, k k2 is not a well-defined norm anymore; Consider for
R +
example the function f that takes the value 1 at t = 0 and 0 everywhere else. Then 0 |f (t)|2 dt =
0, but f is not the zero function. This issue is easily solved by identifying those functions that are
equal except for a small set of points, such as finitely many points, or countably many points, or
anyway a set of points such that the integral of a function does not change if we alter the value of
the function on that set; f should therefore be identified with the zero function, because it differs
from it on only one point. Such functions are said to be equal almost everywhere.
But the situation is even worse. While the functions of interest are assumed to possess some sort of
continuity, there are always Cauchy sequences with respect to the norm k k2 that do not converge
to a function having that sort of continuity. Loosely speaking, what can happen is even that the
limit function can fail to be integrable in the sense of Riemann.
The canonical way in which this problem is solved is by extending the definition of integral. The
Lebesgue integral of a function is a technical construction that generalizes in many ways the notion of
integral that the reader remembers from calculus courses (the Riemann integral). When a function
f is integrable in the sense of Riemann (either over [t0 , t1 ] or over R+ ), the Lebesgue integral of
f always exists and coincides with its Riemann integral. But Lebesgue integration allows for a
much larger class of functions to be integrable, comprising functions way, way wilder than just
discontinuous at finitely many points. For example, the function d : [0, 1] R such that d(t) = 1
if t Q, and d(t) = 0 elsewhere is discontinuous at all points and is not integrable according to
R1
Riemann; but we mention, albeit without any justification, that the Lebesgue integral 0 d(t)dt
exists and is equal to 0. Lebesgue integration also provides more general convergence theorems, and
an easier way to reduce multiple integrals to iterated integrals of one variable (Fubinis theorem).
If the reader proceeds in the field of System Theory, he or she will have to seriously study Lebesgue
integration at some point. But since the long excursus into measure theory required by this study
leads far beyond the scope of this course, we will avoid this discussion altogether, leave to the reader
the bare statement that the Lebesgue integral is a more general notion of integral, and instead
rely, without providing a proof, on the following fundamental result of mathematical analysis:
Fact 3.5 Given a normed linear space (V, F, k k), there exist a complete normed linear space
(V , F, k k) and a norm-preserving linear function : V V (that is, k(x)k = kxk for all x V )
such that every point in V is a limit point of some sequence of points in (V ). V is called the
completion of V .
In other words, every normed space V can be mapped to a dense subspace of a complete space
V ; the map preserves the distances between points (it is an isometry), so that V and the dense
subspace can be identified to all practical purposes. To gain intuition, the reader can think at the
completion of V as the union of V and the set of all the missing limit points of the non-convergent
Cauchy sequences of V . For example, R is the completion of Q. On the other hand, if V is already
complete, then the completion of V is V itself (hence, the completion R is R). What is now of
greatest interest for us is the following
Fact 3.6 The completion of (C([t0 , t1 ], Rn ), kk2 ) is the space of Lebesgue square-integrable functions
Rt
over [t0 , t1 ], that is the set of all those functions f : [t0 , t1 ] Rn for which the integral t01 |f (t)|2 dt
exists (in the sense ofqLebesgue) and is finite; this space is a complete normed space with respect
R t1
to the norm kf k2 := 2
t0 |f (t)| dt, provided that functions which are equal almost everywhere
are identified. All the functions f whose square is integrable in the sense of Riemann belong to
this space, and their norm according to both the definitions of integral coincide. This completion is
denoted L2 ([t0 , t1 ], Rn ).
An analogous fact can be stated for the space (3.2), namely the completion of V is the Banach space
R +
L2 (R+ , Rn ) of those functions f : R+ Rn for which the Lebesgue integral 0 |f (t)|2 dt exists
qR
+
and is finite, equipped with the norm 0 |f (t)|2 dt.
L2 ([t0 , t1 ], Rn ) and L2 (R+ , Rn ) are now true Banach spaces. Let us restate it clearly: Every sequence
{fn }n=1 of Lebesgue square-integrable functions such that limm,n kfm fn k2 = 0 converges to
a function f which is also Lebesgue square-integrable; every continuous square-integrable function

that the reader can come up with, is also Lebesgue square-integrable; and finally, given a Lebesgue
square-integrable function f , there exist a continuous function fc arbitrarily close to f (kfc f k2 <
). Therefore, L2 ([t0 , t1 ], Rn ) and L2 (R+ , Rn ) are the complete versions of C([t0 , t1 ], Rn ) and
C(R+ , Rn ) at which we were aiming.
3.5 Induced norms and matrix norms

Consider now the space (Rmn , R).
Exercise 3.10 Show that (Rmn , R) is a linear space with the usual operations of matrix addition
and scalar multiplication. (Hint: One way to do this is to identify Rmn with Rnm .)
The following are examples of norms on (Rmn , R):

m X
X n
kAka = |aij | (cf. 1 norm in Rnm )
i=1 j=1
12
m X
X n
kAkF = a2ij (Frobenius norm, cf. 2 norm in Rnm )
i=1 j=1
kAk = max max |aij | (cf. norm in Rnm )

i=1,...,m j=1,...,n
More commonly used are the norms derived when one considers matrices as linear maps between
linear spaces. We start with a more general definition.
Definition 3.10 Consider the linear space of continuous functions f : U V between two normed
spaces (U, F, k kU ) and (V, F, k kV ). The induced norm of f is defined as
kf (u)kV
kf k = sup .
u6=U kukU
Fact 3.7 Consider two normed spaces (U, F, kkU ) and (V, F, kkV ) and a continuous linear function
A : U V . Then
kAk = sup kA(u)kV .
kukU =1
Notice that the induced norm depends not only on the function but also on the norms imposed on
the two spaces.
Example (Induced matrix norms) Consider A F mn and consider the linear map A : F n
F m defined by A(x) = A x for all x F n . By considering different norms on F m and F n different
induced norms for the linear map (and hence the matrix) can be defined:
kAxkp
kAkp = sup
xF n kxkp
Xm
kAk1 = max |aij | (maximum column sum)
j=1,...,n
i=1
1
kAk2 = max j (AT A) 2 (maximum singular value)
j=1,...,n
n
X
kAk = max |aij | (maximum row sum)
i=1,...,m
j=1
It turns out that the induced norm of linear maps is intimately related to their continuity.
Theorem 3.3 Consider two normed spaces (U, F, k kU ) and (V, F, k kV ) and a linear function
A : U V . The following statements are equivalent
1. A is continuous.
2. A is continuous at U .
3. The induced norm kAk is finite.
An easy corrolary is that linear functions between finite dimensional spaces are always continuous.
Corollary 3.2 All linear functions A : U V between two finite dimensional normed spaces
(U, F, k kU ) and (V, F, k kV ) are continuous.
Proof: Fix bases for the two spaces so that the linear function A is represented by multiplication
by a matrix A. Since both spaces are finite dimensional all norms are equivallent; for simplicity let
k kU and k kV both be the corresponding infinity norms.
Note that if A = 0 the function A is constant and hence continuous. Otherwise, consider arbitrary
an u U and note that kA(u) A(u )k kAk ku u k . Clearly kAk (the maximum row
sum) is a finite number. Therefore, for all > 0 there exists = /kAk such that ku u k <
implies that kA(u) A(u )k < . Hence the function is continuous.
Theorem 3.4 Consider linear functions A, A : (V, F, kkV ) (W, F, kkW ) and B : (U, F, kkU )
(V, F, k kV ) between normed linear spaces. Let k k denote the induced norms of the linear maps.
1. For all v V , kA(v)kW kAk kvkV .

2. For all a F , kaAk = |a| kAk.
3. kA + Ak kAk + kAk.
4. kAk = 0 A(v) = 0 for all v V (zero map).
5. kA Bk kAk kBk.
3.6 Ordinary differential equations

The main topic of these notes are dynamical systems of the form
x(t) = A(t)x(t) + B(t)u(t) (3.4)

y(t) = C(t)x(t) + D(t)u(t) (3.5)
where
t R, x(t) Rn , u(t) Rm , y(t) Rp

A(t) Rnn , B(t) Rnm , C(t) Rpn , D(t) Rpm
The difficult part conceptually is equation (3.3), a linear ordinary differential equation (ODE) with
time varying coefficients A(t) and B(t). Equation (3.3) is a special case of the (generally non-linear)
ODE
x(t) = f (x(t), u(t), t) (3.6)
with
t R, x(t) Rn , u(t) Rm
f : Rn Rm R Rn .
The only difference is that for linear systems the function f (x, u, t) is a linear function of x and u
for all t R.
In this section we are interested to find solutions (also known as trajectories, flows, . . . ) of
the ODE (3.5). In other words:
Given:
f : Rn Rm R Rn dynamics
n
(t0 , x0 ) R R initial condition
m
u() : R R input trajectory
Find:
x() : R Rn state trajectory
Such that:
x(t0 ) = x0
d
x(t) = f (x(t), u(t), t) t R
dt
While this definition is acceptable mathematically it tends to be too restrictive in practice. The
problem is that according to this definition x(t) should not only be continuous as a function of time,
but also differentiable; since the definition makes use of the derivative dx(t)/dt, it implicitly assumes
that the derivative is well defined. This will in general disallow input trajectories u(t) which are
discontinuous as a function of time. We could in principle only allow continuous inputs (in which
case the above definition of a solution should be OK) but unfortunately many interesting input
functions turn out to be discontinuous.
Example (Hasty driver) Consider a car moving on a road. Let y R denote the position of the
car with respect to a fixed point on the road, v R denote the velocity of the car and a R denote
its acceleration. We can then write a state space model for the car by defining

y
x= R2 , u = a R
v
and observing that

y(t) x2 (t)
x(t) = = .
v(t) u(t)
Defining
x2
f (x, u) =
u
the dynamics of our car can then be described by the (linear, time invariant) ODE
x(t) = f (x(t), u(t)).
Assume now that we would like to select the input u(t) to take the car from the initial state

0
x(0) =
0
to the terminal state

yF
x(T ) =
0
in the shortest time possible (T ) while respecting constraints on the speed and acceleration
v(t) [Vmax , Vmax ] and a(t) [amin , amax ] t [0, T ]
for some Vmax > 0, amin < 0 < amax . It turns out that the optimal solution for u(t) is discontinuous.
Assuming that yF is large enough it involves three phases:
1. Accelerate as much as possible (u(t) = amax ) until the maximum speed is reached (x2 (t) =
Vmax ).
2. Coast (u(t) = 0) until just before reaching yF .
3. Decelerate as much as possible (u(t) = amin ) and stop exactly at yF .
Unfortunately this optimal solution is not allowed by our definition of solution given above.
To make the definition of solution more relevant in practice we would like to allow discontinuous
u(t), albeit ones that are not too wild. Measurable functions provide the best notion of how
wild input functions are allowed to be and still give rise to reasonable solutions for differential
equations. Unfortunately the proper definition of measurable functions requires some exposure to
measure theory and is beyond the scope of these notes; the interested reader is referred to [?] for a
discussion. For our purposes the following, somewhat simpler definition will suffice.
Definition 3.11 A function u : R Rm is piecewise continuous if and only if it is continuous at

all t R except those in a set of discontinuity points D R that satisfy:
1. D left and right limits of u exist, i.e. limt + u(t) and limt u(t) exist and are finite.
Moreover, u( ) = limt + u(t).
2. t0 , t1 R with t0 < t1 , D [t0 , t1 ] contains a finite number of points.
We will use PC([t0 , t1 ], Rm ) to denote the linear space of piecewise continuous functions f : [t0 , t1 ]
Rm (and similarly for PC(R, Rm )).
Exercise 3.11 Show that (PC([t0 , t1 ], Rm ), R) is a linear space under the usual operations of func-
tion addition and scalar multiplication.
Definition 3.11 includes all continuous functions, the solution to the hasty driver example, square
waves, etc. Functions that are not included are things like 1/t or tan(t) (that go to infinity for some
t R), and wild things like

0
(t h1/2) (t 0)
1 1
u(t) = 1 t 2k+1 , 2k
h k = 1, 2, . . .

1 1 1
t ,
2(k+1) 2k+1
that have an infinite number of discontinuity points in the finite interval [0, 1/2].
Let us now return to our differential equation (3.5). Let us generalize the picture somewhat by
considering
x(t) = p(x(t), t) (3.7)
where we have obscured the presence of u(t) by defining p(x(t), t) = f (x(t), u(t), t). We will impose
the following assumption of p.
Assumption 3.1 The function p : Rn R Rn is piecewise continuous in its second argument,

i.e. there exists a set of discontinuity points D R such that for all x Rn
1. p(x, t) is continuous for all t R \ D.

2. For all D, limt + p(x, t) and limt p(x, t) exist and are finite and p(x, ) = limt + p(x, t).
3. t0 , t1 R with t0 < t1 , D [t0 , t1 ] contains a finite number of points.
Exercise 3.12 Show that if u : R Rm is piecewise continuous (according to Definition 3.11)

and f : Rn Rm R Rn is continuous then p(x, t) = f (x, u(t), t) satisfies the conditions of
Assumption 3.1.
We are now in a position to provide a formal definition of the solution of ODE (3.6).
Definition 3.12 Consider a function p : Rn R Rn satisfying the conditions of Assumption 3.1.

A continuous function : R Rn is called a solution of (3.6) passing through (t0 , x0 ) R Rn if
and only if
1. (t0 ) = x0 , and
d
2. t R \ D, is differentiable at t, and dt (t) = p((t), t).
The solution only needs to be continuous and differentiable almost everywhere, i.e. everywhere
except at the set of discontinuity points, D, of p. This definition allows us to include input trajectories
that are discontinuous as functions of time, but still reasonably well behaved.
Is this a good definition? For example, is it certain that there exists a function satisfying the
conditions of the definition? And if such a function does exist, is it unique, or can there be other
functions that also satisfy the same conditions? These are clearly important questions in control
theory where the differential equations usually represent models of physical systems (airplanes, cars,
circuits etc.). In this case the physical system being modeled is bound to do something as time
progresses. If the differential equation used to model the system does not have solutions or has
many, this is an indication that the model we developed is inadequate, since it cannot tell us what
this something is going to be.
The answer to these questions very much depends on the function p. Notice that in all subsequent
examples the function p(x, t) is independent of t; therefore all potential problems have noting to do
with the fact that p is piecewise continuous as a function of time.
Example (No solutions) Consider the one dimensional system

1 if x(t) 0
x(t) = sgn(x(t)) =
1 if x(t) < 0,
For the initial condition (t0 , x0 ) = (0, c) with c > 0, Definition 3.12 allows us to define the solution
(t) = c t for t c. There does not exist a function, however, that satisfies Definition 3.12 for
t > c.
Exercise 3.13 Verify this.
In particular, for the initial condition (t0 , x0 ) = (0, 0) the solution is undefined for all t > 0.
The problem here is that the function p(x, t) is discontinuous in x. The example suggests that to
be able to use Definition 3.12 effectively we need to exclude p(x, t) that are discontinuous in x.
Another alternative would be to relax the Definition 3.12 further to allow the so-called Filippov
solutions [?, ?], but this is beyond the scope of these notes.
Is the existence of a solution guaranteed if we restrict our attention to p(x, t) that are continuous in
x? Unfortunately not!
Example (Finite Escape Time) Consider the one dimensional system

x(t) = x(t)2
with (t0 , x0 ) = (0, c) for some c > 0. The function
c
x(t) =
1 tc
is a solution of this differential equation.
Notice that as t 1/c the solution escapes to infinity; no solution is defined for t 1/c. As in
the previous example, Definition 3.12 allows us to define a solution until a certain point in time,
but no further. There is a subtle difference between the two examples, however. In the latter case
the solution can always be extended for a little extra time; it is defined over the right-open interval
(, 1/c). In the former case, on the other hand, a dead end is reached in finite time; for initial
conditions (0, c) the solution is only defined over the closed interval (, c].
The problem here is that p(x, t) grows too fast as a function of x. To use Definition 3.12 we will
therefore need to exclude functions that grow too fast.
Even if these exclusions work and we manage to ensure that a solution according to Definition 3.12
exists, is this solution guaranteed to be unique, or can there be many functions (t) satisfying the
conditions of the definition for the same (t0 , x0 )? Unfortunately the answer is yes.
Example (Multiple Solutions) Consider the one dimensional system

1
x(t) = 2|x(t)| 2 sgn(x(t))
with (t0 , x0 ) = (0, 0). For all a 0 the function

(t a)2 ta
x(t) =
0 t<a
is a solution of the differential equation.
Notice that in this case the solution is not unique. In fact there are infinitely many solutions, one
for each a 0.
The problem here is that p(x, t) is too steep, since its slope goes to infinity as x tends to 0. To
use Definition 3.12 we will therefore need to also exclude functions that are too steep.
Functions that are discontinuous, grow too fast or are too steep can all be excluded at once by
the following definition.
Definition 3.13 The function p : Rn R Rn is globally Lipschitz in x if and only if there exists
a piecewise continuous function k : R R+ such that
x, x Rn , t R kp(x, t) p(x , t)k k(t)kx x k.
k(t) is called the Lipschitz constant of p at t R.
Example (Lipschitz functions) One can easily verify that linear functions are Lipschitz; we will
do so when we introduce linear systems. All differentiable functions with bounded derivatives are
Lipschitz. However, not all Lipschitz functions are differentiable. For example, the absolute value
function | | : R R is Lipschitz (with Lipschitz constant 1) but not differentiable at x = 0.
All Lipschitz functions arepcontinuous, but not all continuous functions are Lipschitz. For exam-
ple, the functions x2 and |x| from R to R are both continuous, but neither is Lipschitz. x2 is
locally Lipschitz, i.e. x0 R, > 0, kx0 : R R+ piecewise continuous such that x R:
kx x0 k < kp(x, t) p(x0 , t)k kx0 (t)kx x0 k

But there is no k(t) that will work for all x0 . x is not even locally Lipschitz, a finite number kx0 (t)
does not exist for x0 = 0.
In the following section we will show that global Lipschitz continuity is sufficient to ensure the
existence and uniqueness of solutions for ordinary differential equations. Global Lipshitz continuity
is indeed a tight sufficient condition for existence and uniqueness of solutions, albeit not a necessary
one. The examples outlined earlier in this section demonstrate that there exist differential equation
defined by non-Lipschitz functions that do not posess unique solutions. On the other hand, one can
also find differential equation defined by non-Lipschitz functions that do posess unique solutions for
all intial conditions; to establish this fact, however, more work is needed on a case by case basis. The
existence and uniqueness results discussed here can be further fine tuned (assuming for example local
Lipschitz continuity to derive local existence of solutions). In the Chapter 4, however, we will show
that linear differential equations, the main topic of these notes, always satisfy the global Lipschitz
assumption. We will therefore not pursue such refinements here, instead we refer the interested
reader to [?, ?, ?].
3.7 Existence and uniqueness of solutions

We are now in a position to state and prove a fundamental fact about the solutions of ordinary
differential equations.
Theorem 3.5 (Existence and uniqueness) Assume p : Rn R Rn is piecewise continuous

with respect to its second argument (with discontinuity set D R) and globally Lipschitz with
respect to its first argument. Then for all (t0 , x0 ) R Rn there exists a unique continuous function
: R Rn such that:
1. (t0 ) = x0 .
d
2. t R \ D, dt (t) = p((t), t).
The proof of this theorem is rather sophisticated. We will build it up in three steps:
1. Background lemmas.
2. Proof of existence (construction of a solution).
3. Proof of uniqueness.
3.7.1 Background lemmas
For a function f : [t0 , t1 ] Rn with f (t) = (f1 (t), . . . , fn (t)) define

Rt
1
Z t1 t0 f1 (t)dt
..
f (t)dt =
R .
.

t0 t1
t0 n f (t)dt
Fact 3.8 Let k k be any norm on Rn . Then for all t0 , t1 R,

Z t1 Z t1

f (t)dt kf (t)k dt .

t0 t0
Proof:(Sketch) Roughly speaking one can approximate the integral by a sum, use triangle inequality
on the sum and take a limit. Note that the absolute value on the right had side is necessary, since
the integral there will be negative if t1 < t0 .
Recall that m! = 1 2 . . . m denotes the factorial of a natural number m N.
Fact 3.9 The following hold:
1. m, k N, (m + k)! m! k!.
cm
2. c R, limm m! = 0.
Proof: Part 1 is easy to prove by induction (Theorem 1.2). For part 2, the case |c| < 1 is trivial,
since cm 0 while m! . For c > 1, take M > 2c integer. Then for m > M
cm cM cmM cM cmM cM 1
= mM
= mM 0.
m! M !(M + 1)(M + 2) . . . m M ! (2c) M! 2
The proof for c < 1 is similar.
Theorem 3.6 (Fundamental theorem of calculus) Let g : R R piecewise continuous with

Rt
discontinuity set D R. Then the function f (t) = 0 g( )d is continuous and for all t R \ D,
d
f (t) = g(t).
dt
Theorem 3.7 (Gronwall Lemma) Consider u(), k() : R R+ piecewise continuous, c1 0,

and t0 R+ . If for all t R+ Z t

u(t) c1 + k( )u( )d
t0
then for all t R Z t

u(t) c1 exp k( )d .
t0
Proof: Consider t > t0 (the proof for t < t0 is symmetric). Let

Z t
U (t) = c1 + k( )u( )d.
t0
Notice that u(t) U (t), since the absolute value is redundant as u() and k() are non-negative
and t t0 . By the fundamental theorem of calculus U is continuous and wherever k and u are
continuous
d
U (t) = k(t)u(t).
dt
Then
Rt Rt
k( )d k( )d
u(t) U (t) u(t)k(t)e t0 U (t)k(t)e t0
R
d t k( )d t k( )d
R
U (t) e t0 U (t)k(t)e t0 0
dt
R
d t k( )d d Rtt k( )d
U (t) e t0 + U (t) e 0 0
dt dt
d t k( )d
R
U (t)e t0 0
dt R t
k( )d
U (t)e t0
decreases as t increases
Rt R t0
k( )d
U (t)e t0
U (t0 )e t0 k( )d
t t0
Rt
k( )d
U (t)e t0
U (t0 ) = c1 t t0
Rt
k( )d
u(t) U (t) c1 e t0
which concludes the proof.
3.7.2 Proof of existence
We will now construct a solution for the differential equation
x(t) = p(x(t), t)
passing through (t0 , x0 ) R Rn using an iterative procedure. This second step of our existence-
uniqueness proof will itself involve three steps:
1. Construct a sequence of functions xm () : R Rn for m = 1, 2, . . ..

2. Show that for all t1 t0 t2 this sequence is a Cauchy sequence in the Banach space
C([t1 , t2 ], Rn ), k k ). Therefore the sequence converges to a limit () C([t1 , t2 ], Rn ).
3. Show that the limit () is a solution to the differential equation.
Step 2.1: We construct a sequence of functions xm () : R Rn for m = 1, 2, . . . by the so called

Picard iteration:
1. x0 (t) = x0 t R
Rt
2. xm+1 (t) = x0 + t0 p(xm ( ), )d, t R.
The generated sequence of functions is known at the Picard Iteration. Notice that all the functions
xm () generated in this way are continuous by construction.
Consider any t1 , t2 R such that t1 t0 t2 . Let
k = sup k(t) and T = t2 t1 .

t[t1 ,t2 ]
Notice that under the conditions of the theorem both k and T are non-negative and finite. Let k k
be the infinity norm on Rn . Then for all t [t1 , t2 ]
Z t Z t

kxm+1 (t) xm (t)k = x
0 + p(x m ( ), )d x0 p(xm1 ( ), )d

0 t 0 t
Z t

= [p(xm ( ), ) p(xm1 ( ), )]d
t0
Z t

kp(xm ( ), ) p(xm1 ( ), )k d (Fact 3.6)
t
Z 0t

k( ) kxm ( ) xm1 ( )k d (p is Lipschitz in x)
t0
Z t

k kxm ( ) xm1 ( )k d
t0
For m = 0 Z t Z
t2

kx1 (t) x0 (t)k kp(x0 , )k d kp(x0 , )k d = M,
t0 t1
for some non-negative, finite number M (which of course depends on t1 and t2 ). For m = 1
Z t

kx2 (t) x1 (t)k k kx1 ( ) x0 ( )k d
t0
Z t

k M d = kM |t t0 |.
t0
For m = 2,
Z t

kx3 (t) x2 (t)k k kx2 ( ) x1 ( )k d
t0
Z t 2
k (t t0 )2
k
kM | t0 |d = M .
t0 2
In general, for all t [t1 , t2 ],
[k|t t0 |]m [kT ]m

kxm+1 (t) xm (t)k M M .
m! m!
Step 2.2: We show that the sequence xm () is a Cauchy sequence in the Banach space (C([t1 , t2 ], Rn ), k
k ). Note that xm () is continuous by construction by the fundamental theorem of calculus. We
therefore need to show that xm () has bounded infinity norm for all m and that
> 0 N N m N, kxm () xN ()k .

We start with the second statement. Take m N 0 integers and consider

kxm () xN ()k = kxm () xm1 () + xm1 () . . . + xN +1 () xN ()k

mN 1
X
= [xi+N +1 () xi+N ()]

i=0
mN
X1
kxi+N +1 () xi+N ()k
i=0
mN
X1 (kT )i+N
M (Step 2.1)
i=0
(i + N )!
mN
X1 (kT )N (kT )i
M (Fact 3.7)
i=0
N! i!
mN
X1
(kT )N (kT )i
=M
N! i=0
i!

(kT )N X (kT )i
M
N ! i=0 i!
(kT )N kT
=M e .
N!
Note that, in particular, if we take N = 0 this implies that
kxm () x0 ()k M ekT .

This, in turn, by the triangle inequality and the fact that x0 (t) = x0 for all t [t1 , t2 ] implies that
kxm ()k kx0 k + M ekT < , which establishes that xm () has bounded infinity norm hence
xm () (C([t1 , t2 ], Rn ), k k ) as required. Moreover, by Fact 3.7
(kT )N
lim = 0,
N N!
therefore,
> 0 N N m N, kxm () xN ()k
and the sequence {xm ()} n
m=0 is Cauchy in the Banach space (C([t1 , t2 ], R ), kk). Hence it converges
n
in the infinity norm to some continuous function () : [t1 , t2 ] R with k()k < . Notice that
convergence in the infinity norm implies that
m m
kxm () ()k 0 sup kxm (t) (t)k 0
t[t1 ,t2 ]
m
kxm (t) (t)k 0 t [t1 , t2 ]
m
xm (t) (t) t [t1 , t2 ]
Step 2.3: Finally, we show that the limit function () : [t1 , t2 ] Rn solves the differential equation.
We need to verify that:
1. (t0 ) = x0 ; and
d
2. t [t1 , t2 ] \ D, dt (t) = p((t), t).
For the first part, notice that by construction x0 (t0 ) = x0 and for all m > 0 since
Z t0
xm+1 (t0 ) = x0 + p(xm ( ), )d = x0 .
t0
Therefore xm (t0 ) = x0 for all m N and limm xm (t0 ) = (t0 ) = x0 .

For the second part, for t [t1 , t2 ] consider
Z t Z t

[p(xm ( ), ) p(( ), )]d kp(xm ( ), ) p(( ), )k d (Fact 3.6)

t0 t0
Z t

k( )kxm ( ) ( )kd
(p is Lipschitz)
tZ0
t

k sup kxm (t) (t)kd
t0 t[t1 ,t2 ]
Z t

kkxm () ()k d
t0
= kkxm () ()k |t t0 |
kkxm () ()k T.
m
Since, by Step 2.2, kxm () ()k 0,
Z t Z t
m
p(xm ( ), )d( ) p(( ), )d.
t0 t0
Therefore Rt
xm+1 (t) = x0 + t0 p(xm ( ), )d

Rt
(t) = x0 + t0
p(( ), )d
By the fundamental theorem of calculus is continuous and
d
(t) = p((t), t) t [t1 , t2 ] \ D.
dt
Therefore, our iteration converges to a solution of the differential equation for all t [t1 , t2 ]. Since
[t1 , t2 ] are arbitrary, we can see that our iteration converges to a solution of the differential equation
for all t R, by selecting t1 small enough and t2 large enough to ensure that t [t1 , t2 ].
3.7.3 Proof of uniqueness
Can there be more solutions besides the that we constructed in Section 3.7.2? It turns out that
this is not the case. Assume, for the sake of contradiction, that there are two different solutions
(), () : R Rn . In other words
1. (t0 ) = (t0 ) = x0 ; and

d d
2. R \ D, d ( ) = p(( ), ) and d ( ) = p(( ), ).
Therefore
d
(( ) ( )) = p(( ), ) p(( ), ), R \ D
d
Z t
(t) (t) = [p(( ), ) p(( ), )]d, t R
t0
Z t Z t

k(t) (t)k
kp(( ), ) p(( ), )kd k( )k( ) ( )kd
t0 t0
Z t

k(t) (t)k c1 + k( )k( ) ( )kd c1 0.
t0
Letting u(t) = k(t) (t)k and applying the Gronwall lemma leads to
Rt
k( )d
k(t) (t)k c1 e t0
, c1 0.
Letting c1 0 leads to
k(t) (t)k = 0 (t) = (t) t R
which contradicts the fact that () 6= ().
This concludes the proof of existence and uniqueness of solutions of ordinary differential equations.
From now on we can talk about the solution of the differential equation, as long as the conditions
of the theorem are satisfied. It turns out that the solution has several other nice properties, for
example it varies continuously as a function of the initial condition and parameters of the function
p.
Unfortunately, the solution cannot usually be computed explicitly as a function of time. In this case
we have to rely on simulation algorithms to approximate the solution using a computer. The nice
properties of the solution (its guarantees of existence, uniqueness and continuity) come very handy
in this case, since they allow one to design algorithms to numerically approximate the solutions and
rigorously evaluate their convergence properties. Unfortunately for more general classes of systems,
such as hybrid systems, one cannot rely on such properties and the task of simulation becomes much
more challenging.
One exception to the general rule that solutions cannot be explicitly computed is the case of linear
systems, where the extra structure afforded by the linearity of the function p allows us to study the
solution in greater detail. We pursue this direction in Chapter 4.

Problem 3.1 (Norms) Let F be either R or C. Show that the following are well-defined norms
for the linear spaces (F n , F ), (F mn , F ) and C([t0 , t1 ], F n ), F , respectively:
1. ||x|| = maxi |xi |, where x = (x1 , . . . , xn ) F n ;

P
2. ||A|| = maxi j |ai,j |, where A = [ai,j ] F mn ;
3. ||f || = maxt[t0 ,t1 ] ||f (t)|| , where f C([t0 , t1 ], F n ).
For x F n , show in addition P

that the norms ||x|| , ||x||1 and ||x||2 are equivalent (Hint: you may
n n
assume Schwarzs inequality: i=1 (|xi | |yi |) ||x||2 ||y||2 , x, y F ).
Problem 3.2 (Ball of a given norm) Let B(a; r), with a V and r R+ , denote a ball of a
normed vector space (V, F, || ||), and let be the zero vector. Show that:
1. v1 , v2 B(a; r) v1 + (1 )v2 B(a; r), [0, 1] (B(a; r) is convex);

2. v B(; r) v B(; r) (B(; r) is balanced);
3. v V r (0, +) such that v B(; r).
Problem 3.3 Let (V, F, k k) be a normed space.
1. The sets V and are both open and closed.
2. If K1 , K2 V are open sets then K1 K2 is an open set.

3. If K1 , K2 V are closed sets then K1 K2 is a closed set.

4. Let {Ki V | i I} be a collection of open sets, where I is an arbitrary (finite, or infinite)
index set. Then iI Ki is an open set.
5. Let {Ki V | i I} be a collection of closed sets, where I is an arbitrary (finite, or infinite)
index set. Then iI Ki is a closed set.
(Hint: Show 1, 3 and 5, then show that 3 implies 2 and 5 implies 4.)
Problem 3.4 (Continuity) Let f : (U, F, k kU ) (V, F, k kV ) be a function between two normed
spaces. The following statements are equivalent:
1. f is continuous.
2. For all sequences {ui }
i=1 U
lim ui = u lim f (ui ) = f (u).

i i
3. For all K V open, the set
f 1 (K) = {u U | f (u) K}
is open.
4. For all K V closed, the set f 1 (K) is closed.
Problem 3.5 (Equivalent Norms) Let (V, F ) be a linear space. Let ||||a and ||||b be equivalent
norms on (V, F ). Let v V , X V and let {vi }iN be a sequence of elements of V . Show that:
1. {vi }iN is Cauchy w.r.t. || ||a {vi }iN is Cauchy w.r.t. || ||b ;
i i
2. vi v w.r.t. || ||a vi v w.r.t. || ||b ;
3. X is dense in V w.r.t. || ||a X is dense in V w.r.t. || ||b .
Problem 3.6 (Induced norms) Let (U, F, || ||U ) and (V, F, || ||V ) be normed spaces, and let U
be the zero vector of U .
1. Show that the induced norm

||F (u)||V
||F || = sup
uU: u6=U ||u||U
is a well-defined norm for the space of all operators F : U V (you may assume that the
space of operators is a linear space over F ).
2. Let A : U V be a linear operator. For the induced norm ||A||, show that
||A|| = sup ||A(u)||V .

u: ||u||U =1
Problem 3.7 (ODE solution properties) Consider p : Rn R Rn Lipschitz continuous in its

first argument and piecewise continuous in its second. For t, t0 R, x0 Rn , let s(t, t0 , x0 ) denote
the unique solution of the differential equation

s(t, to , x0 ) = p(s(t, t0 , x0 ), t), with s(t0 , t0 , x0 ) = x0 .
t
Consider arbitrary t, t1 , t0 R. Show that:
1. For all x0 Rn , s(t, t1 , s(t1 , t0 , x0 )) = s(t, to , x0 ).

2. For all x0 Rn , s(t0 , t, s(t, t0 , x0 )) = x0 .
3. The function s(t, t0 , ) : Rn Rn is continuous.
Problem 3.8 (Population dynamics)
Problem 3.9 (Discontinuous dynamics)

Chapter 4
Time varying linear systems:

Solutions
We now return to the system
x(t) = A(t)x(t) + B(t)u(t) (4.1)

y(t) = C(t)x(t) + D(t)u(t) (4.2)
where
t R, x(t) Rn , u(t) Rm , y(t) Rp
and
A() : R Rnn , B() : R Rnm ,

C() : R Rpn , D() : R Rpm
which will concern us from now on.
4.1 Motivation: Linearization about a trajectory

Why should one worry about time varying linear systems? In addition to the fact that many physical
systems are naturally modeled of the form (4.1)(4.2), time varying linear systems naturally arise
when one linearizes non-linear systems about a trajectory.
Consider a non-linear system modeled by an ODE
x(t) = f (x(t), u(t))
(4.3)
y(t) = h(x(t), u(t))
with
t R, x(t) Rn , u(t) U Rm , f : Rn U Rn .
To ensure that the solution to (4.3) is well defined assume that f is globally Lipschitz in its first
argument and continuous in its second argument and restrict attention to input trajectories u() :
R U which are piecewise continuous.
Assume that the system starts at a given initial state x0 Rn at time t = 0 and we would like to
drive it to a given terminal state xF Rn at some given time T > 0. Moreover, we would like to
accomplish this with the minimum amount of effort. More formally, given a function
l : Rn U R (known as the running cost or Lagrangian)
58
we would like to solve the following optimization problem:

RT
minimize 0
l(x(t), u(t))dt
over u() : [0, T ] U piecewise continuous
subject to x(0) = x0
x(T ) = xF
x() is the solution of x(t) = f (x(t), u(t)).
Essentially where l(x, u) encodes the effort of applying input u at state x.

The most immediate way to solve this optimization problem is to apply a theorem known as the
maximum principle [?], which leads to an optimal input trajectory u() : [0, T ] U and the
corresponding optimal state trajectory x () : [0, T ] Rn such that x (0) = x0 , x (T ) = xF and
x (t) = f (x (t), u (t)) for almost all t [0, T ].
Notice that the optimal input trajectory is open loop. If we were to apply it to the real system
the resulting state trajectory, x(t), will most likely be far from the expected optimal state trajectory
x (t). The reason is that the differential equation model (4.3) of the system is bound to include
modeling approximation, ignore minor dynamics, disturbance inputs, etc. In fact one can show
that the resulting trajectory x(t) may diverge exponentially from the optimal trajectory x (t) as a
function of t, even for very small errors in the dynamics f (x, u).
To solve this problem and force the real system to track the optimal trajectory we need to introduce
feedback. Several ways of doing this exist. The most direct is to solve the optimization problem
explicitly over feedback functions, i.e. determine a function g : Rn R U such that u(t) = g(x(t), t)
is the optimal input if the system finds itself at state x(t) at time t. This can be done using the
theory of dynamic programming [?]. The main advantage is that the resulting input will explicitly
be in feedback form. The disadvantage is that the computation needed to do this is rather wasteful;
in particular it requires one to solve the problem for all initial conditions rather than just the single
(known) initial condition x0 . In most cases the computational complexity associated with doing this
is prohibitive.
A sub-optimal but much more tractable approach is to linearize the non-linear system about the
optimal trajectory (x (), u ()). We consider perturbations about the optimal trajectory
x(t) = x (t) + x(t) and u(t) = u (t) + u(t).
We assume that x(t) Rn and u(t) Rm are small (i.e. we start close to the optimal trajectory).
The plan is to use u(t) to ensure that x(t) remains small (i.e. we stay close to the optimal
trajectory). Note that

x(t) = x (t) + x(t) x(t) = x (t) + x(t)
= f (x (t), u (t)) + x(t).
Assume that
f1 (x, u)
..
f (x, u) = .
fn (x, u)
where the functions fi : Rn U R for i = 1, . . . , n are differentiable. For t [0, T ] define
f1 f1
x1 (x (t), u (t)) . . . xn (x (t), u (t))
f .. .. ..
(x (t), u (t)) = . . . = A(t) R
nn
x fn fn

x1 (x (t), u (t)) . . . xn (x (t), u (t))
f1 f1
u1 (x (t), u (t)) . . . um (x (t), u (t))
f .. .. ..
(x (t), u (t)) = . . . = B(t) R
nm
.
u fn f

u1 (x (t), u (t)) . . . um (x (t), u (t))
n
Then
x(t) = f (x(t), u(t)) = f (x (t) + x(t), u (t) + u(t))

f f
= f (x (t), u (t)) + (x (t), u (t))x(t) + (x (t), u (t))u(t) + h.o.t.
x u
If x(t) and u(t) are small then the higher order terms are small compared to the terms linear in
x(t) and u(t) and the evolution of x(t) is approximately described by the linear time varying
system

x(t) = A(t)x(t) + B(t)u(t).
We can now use the theory developed in these notes to ensure that x(t) remains small and hence
the nonlinear system tracks the optimal trajectory x (t) closely.
4.2 Existence and structure of solutions

More formally, let (X, R), (U, R), and (Y, R) be finite dimensional linear spaces, of dimensions n, m,
and p respectively1 . Consider families of linear functions,
A(t) : X X B(t) : U X
C(t) : X Y D(t) : U Y
p
parametrized by a real number t R. Fix bases, {ei }ni=1 for (X, R), {fi }m
i=1 for (U, R), and {gi }i=1
for (Y, R). Let A(t), B(t), C(t), and D(t) denote respectively the representation of the linear maps
A(t), B(t), C(t), and D(t) with respect to those bases,
A(t) B(t)
(X, R) (X, R) (U, R) (X, R)
nn
A(t)R B(t)Rnm
{ei }ni=1 {ei }ni=1 {fi }m
i=1 {ei }ni=1
C(t) D(t)
(X, R) (Y, R) (U, R) (Y, R)
C(t)Rpn B(t)Rnm
{ei }ni=1 {gi }pi=1 {fi }m
i=1 {gi }pi=1 .
For the rest of these notes we will be interested in dynamical systems of the form
x(t) = A(t)x(t) + B(t)u(t) (4.4)

y(t) = C(t)x(t) + D(t)u(t) (4.5)
where x(t) Rn , u(t) Rm , and y(t) Rp denote the representations of elements of (X, R), (U, R),
p
and (Y, R) with respect to the bases {ei }ni=1 , {fi }m
i=1 , and {gi }i=1 respectively.
To ensure that the system (4.4)-(4.5) is well-posed we will from now on impose following assumption:
Assumption 4.1 A(), B(), C(), D() and u() are piecewise continuous.
Fact 4.1 Given u() : R Rm and (t0 , x0 ) R Rn there exists a unique solution x() : R Rn
and y() : R Rp for the system (4.4)-(4.5).
1 The complex numbers, C, can also be used as the field, at the expense of some additional complication in
dimension counting. For simplicity we will think of linear spaces as defined over the field of real numbers, unless
otherwise specified (e.g. for eigenvalue calculations).
Proof:(Sketch) Define p : Rn R Rn by
p(x, t) = A(t)x + B(t)u(t).
Exercise 4.1 Show that under Assumption 4.1 p satisfies the conditions of Theorem 3.7. (Take D
to be the union of the discontinuity sets of A(), B(), and u()).
The existence and uniqueness of x() : R Rn then follows by Theorem 3.7. Defining y() : R Rp
by
y(t) = C(t)x(t) + D(t)u(t) t R
completes the proof.
The unique solution of (4.1)-(4.2) defines two functions
x(t) = s(t, t0 , x0 , u) state transition map

y(t) = (t, t0 , x0 , u) output response map
mapping the input trajectory u() : R Rm and initial condition (t0 , x0 ) R Rn to the state and
output at time t R respectively.
Theorem 4.1 Let Dx be the union of the discontinuity sets of A(), B() and u() and Dy the union
of the discontinuity sets of C(), D() and u().
1. For all (t0 , x0 ) R Rn , u() PC(R, Rm )

x() = s(, t0 , x0 , u) : R Rn is continuous and differentiable for all t R \ Dx .
y() = (, t0 , x0 , u) : R Rp is piecewise continuous with discontinuity set Dy .
2. For all t, t0 R, u() PC(R, Rm ), x() = s(t, t0 , , u) : Rn Rn and (t, t0 , , u) : Rn Rp
are continuous.
3. For all t, t0 R, x01 , x02 Rn , u1 (), u2 () PC(R, Rm ), a1 , a2 R
s(t, t0 , a1 x01 + a2 x02 , a1 u1 + a2 u2 ) = a1 s(t, t0 , x01 , u1 ) + a2 s(t, t0 , x02 , u2 )

(t, t0 , a1 x01 + a2 x02 , a1 u1 + a2 u2 ) = a1 (t, t0 , x01 , u1 ) + a2 (t, t0 , x02 , u2 ).
4. For all t, t0 R, x0 Rn , u PC(R, Rm ),
s(t, t0 , x0 , u) = s(t, t0 , x0 , PC ) + s(t, t0 , Rn , u)

(t, t0 , x0 , u) = (t, t0 , x0 , PC ) + (t, t0 , Rn , u)
where Rn = (0, . . . , 0) denotes the zero element of Rn and PC () : R Rm denotes the zero
function PC (t) = (0, . . . , 0) for all t R.
Proof: Part 1 follows from the definition of the solution. Part 4 follows from Part 3, by setting
u1 = U , u2 = u, x01 = x0 , x02 = Rn , and a1 = a2 = 1. Part 2 follows from Part 4, by noting
that s(t, t0 , , u) = s(t, t0 , , U ) + s(t, t0 , Rn , u) and s(t, t0 , , U ) : Rn Rn is a linear function
between finite dimensional linear spaces (and hence continuous by Corollary 3.2); the argument for
is similar. So we only need to establish Part 3.
Let x1 (t) = s(t, t0 , x01 , u1 ), x2 (t) = s(t, t0 , x02 , u2 ), x(t) = s(t, t0 , a1 x01 + a2 x02 , a1 u1 + a2 u2 ), and
(t) = a1 x1 (t) + a2 x2 (t). We would like to show that x(t) = (t) for all t R. By definition
x(t0 ) = a1 x01 + a2 x02 = a1 x1 (t0 ) + a2 x2 (t0 ) = (t0 ).

Moreover, if we let u(t) = a1 u1 (t) + a2 u(t) then for all t R \ D
x(t) = A(t)x(t) + B(t)u(t)

(t) = a1 x1 (t) + a2 x2 (t)
= a1 (A(t)x1 (t) + B(t)u1 (t)) + a2 (A(t)x2 (t) + B(t)u2 (t))
= A(t)(a1 x1 (t) + a2 x2 (t)) + B(t)(a1 u1 (t) + a2 u2 (t))
= A(t)(t) + B(t)u(t).
Therefore x(t) = (t) since the solution to the linear ODE is unique. Linearity of follows from the
fact that C(t)x + D(t)u is linear in x and u.
For simplicity, from now on we will just use 0 to denote zero vectors both for the finite dimensional
space Rn and for the infinite dimensional space PC (). The interpretation should be clear from the
context.
4.3 State transition matrix

By Part 4 of Theorem 4.1 the solution of the system can be partitioned into two distinct components:
s(t, t0 , x0 , u) = s(t, t0 , x0 , 0) + s(t, t0 , 0, u)

state transition = zero input transition + zero state transition
(t, t0 , x0 , u) = (t, t0 , x0 , 0) + (t, t0 , 0, u)

output response = zero input response + zero state response.
Moreover, by Part 3 of Theorem 4.1, the zero input components s(t, t0 , x0 , U ) and (t, t0 , x0 , U )
are linear in x0 Rn . Therefore, in the basis used for the representation of A(), s(t, t0 , x0 , U ) has
a matrix representation (which will in general depend on t and t0 ). This representation is called the
state transition matrix and is denoted by (t, t0 ); note that the representation will depend on t and
t0 . In other words,
s(t,t0 ,,U )
(X, R) (X, R)
nn
(t,t0 )R
{ei }ni=1 {ei }ni=1
and, assuming s(t, t0 , x0 , 0) refers to the representation of the solution with respect to the basis
{ei }ni=1 ,
s(t, t0 , x0 , 0) = (t, t0 )x0 .
Exercise 4.2 Show that the representation of (t, t0 , , 0) : X Y with respect to the bases {ei }ni=1
and {gi }pi=1 is given by C(t)(t, t0 ); in other words
(t, t0 , x0 , 0) = C(t)(t, t0 )x0 .
Therefore the state transition matrix (t, t0 ) completely characterizes the zero input state transi-
tion and output response. We will soon see that, together with the input trajectory u(), it also
characterizes the complete state transition and output response.
Theorem 4.2 (t, t0 ) has the following properties:

1. (, t0 ) : R Rnn is the unique solution of the linear matrix ordinary differential equation

(t, t0 ) = A(t)(t, t0 ) with (t0 , t0 ) = I.
t
Hence it is continuous for all t R and differentiable everywhere except at the discontinuity
points of A(t).
2. For all t, t0 , t1 R, (t, t0 ) = (t, t1 )(t1 , t0 ).
1
3. For all t1 , t0 R, (t1 , t0 ) is invertible and its inverse is [(t1 , t0 )] = (t0 , t1 ).
Proof: Part 1. Recall that s(t, t0 , x0 , U ) = (t, t0 )x0 is the solution to the linear differential
equation
x(t) = A(t)x(t) with x(t0 ) = x0 .
For i = 1, . . . , n consider the solution xi () = s(, t0 , xi (t0 ), U ) to the linear differential equation
starting at the representation xi (t0 ) = (0, . . . , 1, . . . , 0) Rn of the basis vector ei

0 0
.. ..
. .

xi (t) = A(t)xi (t) with xi (t0 ) =
1 = xi (t) = (t, t0 ) 1 .

. .
.. ..
0 0
Note that xi (t) is equal to the ith column of (t, t0 ). Putting the columns next to each other

(t, t0 ) = x1 (t) x2 (t) . . . xn (t)
shows that

d 1 d 2 d n
(t, t0 ) = x (t) x (t) . . . x (t)
t dt dt dt

= A(t)x1 (t) A(t)x2 (t) . . . A(t)xn (t)

= A(t) x1 (t) x2 (t) . . . xn (t) = A(t)(t, t0 ).
Moreover,

1 0 ... 0
0 1 ... 0
(t0 , t0 ) = x1 (t0 ) x2 (t0 ) . . . xn (t0 ) = .. .. . . .. = I R
nn
.
. . . .
0 0 ... 1
Part 1 follows.
Part 2. Consider arbitrary t0 , t1 R and let L(t) = (t, t0 ) and R(t) = (t, t1 )(t1 , t0 ). We would
like to show that L(t) = R(t) for all t R. Note that (by Part 1)
L(t1 ) = (t1 , t0 )
R(t1 ) = (t1 , t1 )(t1 , t0 ) = I (t1 , t0 ) = (t1 , t0 ).
Therefore L(t1 ) = R(t1 ). Moreover, (also by Part 1)

d d
L(t) = (t, t0 ) = A(t)(t, t0 ) = A(t)L(t)
dt dt
d d d
R(t) = [(t, t1 )(t1 , t0 )] = [(t, t1 )](t1 , t0 ) = A(t)(t, t1 )(t1 , t0 ) = A(t)R(t)
dt dt dt
Therefore L(t) = R(t) by existence and uniqueness of solutions of linear differential equations.
Part 3. First we show that (t, t0 ) is nonsingular for all t, t0 R. Assume, for the sake of contra-
diction, that it is not, i.e. there exists t, t0 R such that (t, t0 ) is singular. Then the columns
of (t, t0 ) are linearly dependent (Theorem 2.3) and there exists x0 Rn with x0 6= 0 such that
(t, t0 )x0 = 0.
Let x( ) = (, t0 )x0 . Notice that x(t) = (t, t0 )x0 = 0 and
d d
x( ) = (, t0 )x0 = A( )(, t0 )x0 = A( )x( ).
d d
Therefore x( ) is the unique solution to the differential equation:
d
x( ) = A( )x( ) with x(t) = 0. (4.6)
d
The function x( ) = 0 for all R clearly satisfies (4.6); therefore it is the unique solution to (4.6).
Let now = t0 . Then
0 = x(t0 ) = (t0 , t0 )x0 = I x0 = x0
which contradicts the fact that x0 6= 0. Therefore (t, t0 ) cannot be non-singular.
To determine its inverse, recall that for all t, t0 , t1 R,
(t, t1 )(t1 , t0 ) = (t, t0 )
and let t = t0 . Then
(t0 , t1 )(t1 , t0 ) = (t0 , t0 ) = I [(t1 , t0 )]1 = (t0 , t1 ).
In addition to these, the state transition matrix also has several other interesting properties, some
of which can be found in the exercises. We can now show that the state transition matrix (t, t0 )
completely characterizes the solution of linear time varying differential equations.
Theorem 4.3 For all t, t0 R, x0 Rn , u()) PC(R, Rm ),

Rt
s(t, t0 , x0 , u) = (t, t0 )x0 + t0
(t, )B( )u( )d
state transition = zero input transition + zero state transition
Rt
(t, t0 , x0 , u) = C(t)(t, t0 )x0 + C(t) t0 (t, )B( )u( )d + D(t)u(t)
output response = zero input response + zero state response.
Proof: Several methods exist for proving this fact. The simplest is to invoke the rule of Leibniz for
differentiating integrals.
"Z # Z
b(t) b(t)
d d d
f (t, )d = f (t, )d + f (t, b(t)) b(t) f (t, a(t)) a(t).
dt a(t) a(t) t dt dt
(for the sake of comparison, notice that the fundamental theorem of calculus is the special case
a(t) = t0 , b(t) = t and f (t, ) = f ( ) independent of t.)
We start by showing that
Z t
s(t, t0 , x0 , u) = (t, t0 )x0 + (t, )B( )u( )d
t0
Rt
If we let L(t) = s(t, t0 , x0 , u) and R(t) = (t, t0 )x0 + t0 (t, )B( )u( )d we would like to show
that for all t R, L(t) = R(t). Notice that by definition L(t0 ) = s(t0 , t0 , x0 , u) = x0 and
d d
L(t) = s(t, t0 , x0 , u) = A(t)s(t, t0 , x0 , u) + B(t)u(t) = A(t)L(t) + B(t)u(t).
dt dt
We will show that R(t) satisfies the same differential equation with the same initial condition; the
claim then follows by the existence and uniqueness theorem.
Note first that
Z t0
R(t0 ) = (t0 , t0 )x0 + (t, )B( )u( )d = I x0 + 0 = x0 = L(t0 ).
t0
Moreover, by the Leibniz rule

Z t
d d
R(t) = (t, t0 )x0 + (t, )B( )u( )d
dt dt t0
Z t
d d
= (t, t0 ) x0 + (t, )B( )u( )d
dt dt t0
Z t
d d
= A(t)(t, t0 )x0 + (t, )B( )u( )d + (t, t)B(t)u(t) t (t0 , t0 )B(t0 )u(t0 ) t0
t0 t dt dt
Z t
= A(t)(t, t0 )x0 + A(t)(t, )B( )u( )d + I B(t)u(t)
t0
Z t
= A(t) (t, t0 )x0 + (t, )B( )u( )d + B(t)u(t)
t0
= A(t)R(t) + B(t)u(t).
Therefore R(t) and L(t) satisfy the same linear differential equation for the same initial condition,
hence they are equal for all t by uniqueness of solutions.
To obtain the formula for (t, t0 , x0 , u) simply substitute the formula for s(t, t0 , x0 , u) into y(t) =
C(t)x(t) + D(t)u(t).
Theorem 4.3 shows that

s(t, t0 , x0 , U ) = (t, t0 )x0 zero input transition
Z t
s(t, t0 , 0, u) = (t, )B( )u( )d zero state transition
t0
(t, t0 , x0 , U ) = C(t)(t, t0 )x0 zero input response
Z t
s(t, t0 , 0, u) = C(t) (t, )B( )u( )d + D(t)u(t) zero state response
t0
Let us now analyze the zero state transition and response in greater detail. By Theorem 4.1, the zero
state transition and the zero state response are both linear functions s(t, t0 , 0, ) : PC(R, Rm ) Rn
and (t, t0 , 0, ) : PC(R, Rm ) Rp respectively.
s(t,t0 ,0,)
(PC(R, Rm ), R) (Rn , R)
Z t
u() 7 (t, )B( )u( )d
t0
and
(t,t0 ,0,)
(PC(R, Rm ), R) (Rp , R)
Z t
u() 7 C(t) (t, )B( )u( )d + D(t)u(t).
t0
Fix the basis {fj }m m

j=1 for (R , R) used in the representation of B(t) R
nm
. Fix t0 and
m
consider a family of functions (,) () PC(R, R ) parametrized by > 0 and defined by

0 if t <
1
(,) (t) = if t < +

0 if t + .
For j = 1, . . . , m, consider the zero state transition2 s(t, t0 , 0, (,) (t)fj ) under input (,) (t)fj .
Since s(t0 , t0 , 0, (,) (t)fj ) = 0 and the input is zero until t = ,
s(t, t0 , 0, (,)(t)fj ) = 0 t < .
Exercise 4.3 Show this by invoking the existence-uniqueness theorem.
For t + and assuming is small

Z t
s(t, t0 , 0, (,) (t)fj ) = (t, )B( )(,) ( )fj d
t0
Z +
1
= (t, )B( ) fj d

Z
1 +
= (t, + )( + , )B( )fj d

Z
(t, + ) +
= ( + , )B( )fj d

(t, + )
[( + , )B()fj ]

0
(t, )(, )B()fj = (t, )B()fj
Therefore
lim s(t, t0 , 0, (,)(t)fj ) = (t, )B()fj .
0
Formally, if we pass the limit inside the function s and define
(t) = lim (,) (t)

0
we obtain
s(t, t0 , 0, (t)fj ) = (t, )B()fj Rm .
The statement is formal since to pass the limit inside the function we first need to ensure that the
function is continuous.
Exercise 4.4 We already know that the function s(t, t0 , 0, ) : PC(R, Rm ) Rn is linear. What
more do we need to check to make sure that it is continuous?
Moreover, strictly speaking (t) is not an acceptable input function, since it is equal to infinity
at t = and hence not piecewise continuous. Indeed, (t) is not a real valued function at all,
it just serves as a mathematical abstraction for an input pulse of arbitrarily small length. This
mathematical abstraction is known as the impulse function or the Dirac pulse. Even though in
practice the response of a real system to such an impulsive input cannot be observed, by applying
2 Strictly speaking, we should use the Rm representation of the basis vector f U instead of f itself. The reader
i i
is asked to excuse this slight abuse of the notation.
as input piecewise continuous functions (,) (t) for small enough one can approximate the state
transition s(t, t0 , 0, (t)fj ) arbitrarily closely. Notice also that for t
s(t, t0 , 0, (t)fj ) = s(t, , B()fj , U )
i.e. the zero state transition due to the impulse (t)fj is also a zero input transition starting with
state B()fj at time .
Repeating the process for all {fj }m
j=1 leads to m vectors (t, )B()fj for j = 1, . . . , m. Order-
ing these vectors according to their index j and putting them one next to the other leads to the
impulse transition matrix, K(t, ) Rnm , defined by

(t, )B() if t
K(t, ) =
0 if t < .
The (i, j) element of K(t, ) contains the trajectory of state xi when the impulse function (t) is
applied to input uj . Note that, even though these elements cannot be measured in practice, the
impulse transition matrix is still a well defined, matrix valued function for all t, R.
Substituting s(t, 0, 0, (t)fj ) into the equation y(t) = C(t)x(t) + D(t)u(t) leads to
(t, t0 , 0, (t)fj ) = C(t)(t, )B()fj + D(t)fj (t) Rm .
and the impulse response matrix, H(t, ) Rpm , defined by

C(t)(t, )B() + D(t) (t) if t
H(t, ) =
0 if t < .
Note that, unlike K(t, ), H(t, ) is in general not a well defined matrix valued function since it in
general contains an impulse in its definition (unless of course D(t) = 0 for all t R).

Problem 4.1 (Invariant Subspaces) Let L : V V be a linear map on an n-dimensional vector
space V over the field F . Recall that a subspace M V is called L-invariant if Lx M for every
x M . Suppose that V is a direct sum of two subspaces M1 and M2 , i.e., M1 M2 = {0}, and
M1 + M2 = V . If both M1 and M2 are L-invariant, show that there exists a matrix representation
A F nn of the form
A11 0
A=
0 A22
with Dim(A11 ) = Dim(M1 ) and Dim(A22 ) = Dim(M2 ). (Recall that the sum of subspaces M and
N of a vector space X is the set of all vectors of the form m + n where m M and n N . A
vector space X is the direct sum of two subspaces M and N if every vector x X has a unique
prepresentation of the form x = m + n where m M and n N ; we write X = M N .)
Problem 4.2 (Eigenvalues and Invariant Subspaces) Let A be a real-valued n n matrix.

Suppose that +i is a complex eigenvalue of A and x+iy is the corresponding complex eigenvector,
where , R and x, y Rn . Show that x iy is also an eigenvector with eigenvalue i. Let
V be the 2-dimensional subspace spanned by x and y, i.e., V is the set of linear combinations with
real-valued coefficients of the real-valued vectors x and y. Show that V is an invariant subspace of
A, namely, if z V then we have Az V .
Problem 4.3 (ODEs)

1. Let f : Rn Rn be Lipschitz with Lipschitz constant K [0, +). For t R, let x(t) be
the solution of x(t) = f (x(t)), with x(0) = x0 . Let x Rn be such that f (x) = 0. Show that
||x(t) x|| eKt ||x0 x||, t R+ . (Here || || is the Euclidean norm on (Rn , R) for which K
is defined. Hint: use the Gronwall Lemma).
2. Let A : R+ Rnn , B : R+ Rnm and u : R+ Rm be piecewise continuous functions.
Show that, for any x0 Rn , the linear ODE
(
x(t) = A(t)x(t) + B(t)u(t), t R+ ,
x(0) = x0
has a unique solution x() : R+ Rn . (Hint: you may assume that if A(t) is piecewise
continuous then so is its induced norm).
Problem 4.4 (Linear ODEs) Let A() : R Rnn be piecewise continuous. Consider the fol-
lowing linear ODE:
x(t) = A(t)x(t), (4.7)
and let (t, t0 ) be the state transition matrix.

1. Show that t (t0 , t) = (t0 , t)A(t).
(Hint: differentiate the identity (t0 , t)(t, t0 ) = I.)
2. Let X0 be a convex set (i.e. x, x X0 x + (1 )x X0 , [0, 1]). Let s(t, t0 , x0 ) be
the solution of (1) associated with the initial condition x(t0 ) = x0 . Show that the set
X(t) = {s(t, t0 , x0 ) : x0 X0 }
is convex for all t R.

3. A function X() : R Rnn is said to be a fundamental matrix for the matrix differential
equation X(t) = A(t)X(t) if it is a solution and it is nonsingular for all t R. Show that,
given any fundamental matrix X(), it holds that (t, t0 ) = X(t) X(t0 )1 .
Problem 4.5 (Change of basis)

Chapter 5
Time invariant linear systems:

Solutions and transfer functions
Let us now turn to the special case where all the matrices involved in the system dynamics are
constant, in other words
x(t) = Ax(t) + Bu(t) (5.1)

y(t) = Cx(t) + Du(t) (5.2)
with t R+ , x(t) Rn , u(t) Rm , y(t) Rp and A Rnn , B Rnm , C Rpn and D Rpm
are constant matrices.
5.1 Time domain solution

Define the exponential of the matrix A Rnn by
A2 t2 Ak tk
eAt = I + At + + ...+ + . . . Rnn . (5.3)
2! k!
Theorem 5.1 For all t, t0 R+ , (t, t0 ) = eA(tt0 ) .
Proof: Exercise. Show that eA(tt0 ) satisfies the conditions of Part 1 of Theorem 4.2 by taking the
derivative of the expansion in (5.3) with respect to t. The result follows by uniqueness of solutions
of ordinary differential equations.
The above theorem together with the different properties of (t, t0 ) established in Chapter 4 imme-
diately lead to the following corollary.
Corollary 5.1 The state transition matrix, solution, impulse transition, and impulse response of a
time invariant linear system satisfy the following properties:
1
1. For all t, t1 , t0 R, eAt1 eAt2 = eA(t1 +t2 ) and eAt = eAt .
2. For all t, t0 R, (t, t0 ) = (t t0 , 0).
69
3. For all t, t0 R, x0 Rn , u() PC(R, Rm ),

Z t
s(t, t0 , x0 , u) = eA(tt0 ) x0 + eA(t ) B( )u( )d
t0
Z t
A(tt0 )
(t, t0 , x0 , u) = Ce x0 + C eA(t ) B( )u( )d + Du(t).
t0
4. For all t, R the

eA(t) B if t
K(t, ) = K(t , 0) =
0 if t < .

CeA(t) B + D0 (t ) if t
H(t, ) = H(t , 0) =
0 if t < .
From the above it becomes clear that for linear time invariant systems the solution is independent
of the initial time t0 ; all that matters is how much time has elapsed since then, i.e. t t0 . Without
loss of generality we will therefore take t0 = 0 and write
Z t
x(t) = s(t, 0, x0 , u) = eAt x0 + eA(t ) B( )u( )d
0
Z t
At
y(t) = (t, 0, x0 , u) = Ce x0 + C eA(t )B( )u( )d + Du(t)
0
At
e B if t 0
K(t) = K(t, 0) =
0 if t < 0.
At
Ce B + D0 (t) if t 0
H(t) = H(t, 0) =
0 if t < 0.
Notice that in this case the integral that appears in the state transition and output response is
simply the convolution of the input u() with the impulse transition and impulse response matrices
respectively,
x(t) = eAt x0 + (K u)(t)

y(t) = CeAt x0 + (H u)(t).
5.2 Semi-simple matrices

For the time invariant case, the state transition matrix eAt can be computed explicitly. In some cases
this can be done directly from the infinite series; this is the case for example for nilpotent matrices,
i.e. matrices for which there exists N N such that AN = 0. More generally, one can use Laplace
transforms or eigenvectors to do this. We concentrate primarily on the latter method; Laplace
transforms will be briefly discussed in Section 5.4. We start with the simpler case of the so-called
semi-simple matrices.
Definition 5.1 A matrix A Rnn is semi-simple if and only if its right eigenvectors {vi }ni=1 Cn
are linearly independent in the linear space (Cn , C).
Theorem 5.2 Matrix A Rnn is semi-simple if and only if there exists a nonsingular matrix
T Cnn and a diagonal matrix Cnn such that A = T 1 T .
Proof: () Take
T = [v1 v2 . . . vn ]1 Cnn
(which is invertible since {vi }ni=1 are linearly independent). Then
AT 1 = [Av1 Av2 . . . Avn ] = [1 v1 2 v2 . . . n vn ] = T 1
where i C are the corresponding eigenvalues. Multiplying on the right by T leads to
A = T 1 T.
() Assume that there exists matrices T Cnn nonsingular and Cnn diagonal such that
A = T 1 T AT 1 = T 1 .
Let T 1 = [w1 . . . wn ] where wi Cn denoted the ith column of T 1 and

1 . . . 0

= ... . . . ...
0 . . . n
with i C. Then
Awi = i wi
and therefore wi is a right eigenvector of A with eigenvalue i . Since T 1 is invertible its columns
(and eigenvectors of A) {wi }ni=1 are linearly independent.
If A is semi-simple, then its eigenvectors are linearly independent and can be used as a basis. Let
us see what the representation of the matrix A and the state transition matrix eAt with respect to
this basis are.
A
(Cn , C) (Cn , C)
ARnn
{ei }ni=1 {ei }ni=1 (basis leading to representation of A by A)
A=T AT 1 =
{vi }ni=1 {vi }ni=1 (eigenvector basis)
Recall that if x = (x1 , . . . , xn ) Rn is the representation of x with respect to the basis {ei }ni=1 its
representation with respect to the complex basis {vi }ni=1 will be the complex vector x = T x Cn .
The above formula simply states that
x = T x = T Ax = T AT 1 x = x.
Therefore, if A is semi-simple, its representation with respect to the basis of its eigenvectors is the
diagonal matrix of its eigenvalues. Notice that even though A is a real matrix, its representation
is in general complex, since the basis {vi }ni=1 is also complex.
What about the state transition matrix?
Fact 5.1 If A is semi-simple

e1 t ... 0

eAt = T 1 et T = T 1 ... ..
.
..
. T (5.4)
0 ... en t
Proof: Exercise. Show that Ak = T 1 k T and substitute into the expansion (5.3). Or differentiate
(??) and invoke existence-uniqueness.
In other words:
s(t,0,,U )
(Rn , R) (Rn , R)
x0 7 x(t)
eAt Rnn
{ei }ni=1 {ei }ni=1 (basis leading to representation of A by A)
t At 1
e =T e T
{vi }ni=1 {vi }ni=1 (eigenvector basis).
Once again, note that the matrices T , T 1 and et will general be complex. Fact 5.1, however,
shows that when taking their product the imaginary parts will miraculously all cancel and we will
be left with a real matrix.
Fact 5.1 shows that if a matrix is semi-simple the calculation of the matrix exponential is rather
straightforward. It would therefore be desirable to establish conditions under which a matrix is
semisimple.
Definition 5.2 A matrix A Rnn is simple if and only if its eigenvalues are distinct, i.e. i 6= j
for all i 6= j.
Theorem 5.3 All simple matrices are semi-simple.
Proof: Assume, for the sake of contradiction, that A Rnn is simple, but not semi-simple. Then
i 6= j for all i 6= j but {vi }ni=1 are linearly dependent in (Cn , C). Hence, there exist a1 , . . . , an C
not all zero, such that
Xn
ai vi = 0.
i=1
Without loss of generality, assume that a1 6= 0 and multiply the above identity by (A 2 I)(A
3 I) . . . (A n I) on the left. Then
n
X
a1 (A 2 I)(A 3 I) . . . (A n I)v1 + ai (A 2 I)(A 3 I) . . . (A n I)vi = 0.
i=2
Concentrating on the first product and unraveling it from the right leads to
a1 (A 2 I)(A 3 I) . . . (A n I)v1 =a1 (A 2 I)(A 3 I) . . . (Av1 n v1 )

=a1 (A 2 I)(A 3 I) . . . (1 v1 n v1 )
=a1 (A 2 I)(A 3 I) . . . v1 (1 n )
=a1 (A 2 I)(A 3 I) . . . (Av1 n1 v1 )(1 n )
etc
which leads to
n
X
a1 (1 2 )(1 3 ) . . . (1 n )v1 + ai (i 2 )(i 3 ) . . . (i n )vi = 0.
i=2
Each term of the sum on the right will contain a term of the form (i i ) = 0. Hence the sum on
the right is zero, leading to
a1 (1 2 )(1 3 ) . . . (1 n )v1 = 0.
But, v1 6= 0 (since it is an eigenvector) and a1 6= 0 (by the assumption that {vi }ni=1 are linearly
dependent) and (1 i ) 6= 0 for i = 2, . . . , n (since the eigenvalues are distinct). This leads to a
contradiction.
In addition to simple matrices, other matrices are simple, for example diagonal matrices and orthog-
onal matrices. In fact one can show that semi-simple matrices are dense in Rnn , in the sense
that
A Rnn , > 0, A Rnn semi-simple : kA A k < .
(Here k k denotes any matrix norm, anyway they are all equivalent). The reason is that for a matrix
to be non semi-simple the matrix of its eigenvectors T Cnn must be singular, i.e. we must have
Det[T ] = 0. But this is a fragile condition, as arbitrary small perturbations of the matrix A will in
general lead to Det[T ] 6= 0. In fact it is not hard to convince ourselves that almost all matrices
in Rnn are semi-simple, since the condition Det[T ] = 0 imposes a single constraint on the n2
dimensional space of complex matrices1 .
In summary, almost all matrices are semi-simple, though not all.
Example (Non semi-simple matrices) The matrix

1 0
A1 = 0 1
0 0
is not semi-simple. Its eigenvalues are 1 = 2 = 3 = (hence it is not simple) but there is only
one eigenvector

1 0 x1 x1 x1 + x2 = x1 1
0 1 x2 = x2 x2 + x3 = x2 v1 = v2 = v3 = 0

0 0 x3 x3 x3 = x3 0
Note that A1 has the same eigenvalues as the matrices

1 0 0 0
A2 = 0 0 and A3 = 0 0
0 0 0 0
A2 is also not semi-simple, it has only two linearly independent eigenvectors v1 = v2 = (1, 0, 0) and
v3 = (0, 0, 1). A3 is semi-simple, with the canonical basis as eigenvectors. Notice that none of the
three matrices is simple.
5.3 Jordan form

Can non semi-simple matrices be written in a form that simplifies the computation of their state
transition matrix eAt ? This is possible through a change of basis that involves their eigenvectors.
The problem is that since in this case there are not enough linearly independent eigenvectors, the
basis needs to be completed with additional vectors. It turns out that there is a special way of doing
this using the so called generalized eigenvectors, so that the representation of the matrix A in the
resulting basis has a particularly simple form, the so-called Jordan canonical form.
To start with, notice that for the eigenvector vi Cn corresponding to eigenvalue i C of A Rnn
Avi = i vi (A i I)vi = 0 vi Null[A i I].
Recall that Null[A i I] is a subspace of Cn .

1 As an aside, these statements can also be made for simple and invertible matrices.
Definition 5.3 The algebraic mutiplicity of an eigenvalue C of a matrix A Rnn is the

number of times appears in the spectrum Spec[A]. The geometric multiplicity of is the dimension
of Null[A I].
Example (Non semi-simple matrices (cont.)) For the matrices considered above
Null[A1 I] = Span{(1, 0, 0)} Dim[Null[A1 I]] = 1

Null[A2 I] = Span{(1, 0, 0), (0, 0, 1)} Dim[Null[A1 I]] = 2
Null[A3 I] = Span{(1, 0, 0), (0, 1, 0), (0, 0, 1)} Dim[Null[A1 I]] = 3.
Notice that in all three cases the algebraic multiplicity of the eigenvalue is 3. The problem with
A1 and A2 is that the geometric multiplicity of is smaller than its algebraic multiplicity. This
implies that there are not enough eigenvectors associated with eigenvalue (in other words linearly
independent vectors in the null space) to form a basis.
The above example suggests that the reason matrices fail to be semi-simple is a discrepancy between
their algebraic and geometric multiplicities. To alleviate this problem we need to complete the basis
with additional vectors. For this purpose we consider the so called generalized eigenvectors.
Definition 5.4 A Jordan chain of length N at eigenvalue C is a family of vectors {v j }j=0

Cn such that
1. v 0 = 0,
2. {v j }j=1 are linearly independent, and
3. [A I]v j = v j1 for j = 1, 2, . . . , .
A Jordan chain {v j }j=0 is called maximal if it cannot be extended, i.e. there does not exist v Cn
linearly independent from {v j }j=1 such that [A I]v = v . The elements of all the maximal Jordan
chains at are the called generalized eigenvectors of .
Fact 5.2 Let {v j }j=0 Cn be a Jordan chain of length at eigenvalue C of the matrix
A Rnn :
1. v 1 is an eigenvector of A with eigenvalue .

2. v j Null[(A I)j ] for j = 0, 1, . . . , .
3. Null[(A I)j ] Null[(A I)j+1 ].
Proof: Part 1: By definition [A I]v 1 = v 0 = 0, hence Av 1 = v 1 .

Part 2: [A I]j v j = [A I]j1 v j1 = . . . = [A I]1 v 1 = v 0 = 0.
Part 3: Let v Null[(A I)j ], i.e. [A I]j v = 0. Then [A I]j+1 v = 0 and hence v
Null[(A I)j+1 ].
Example (Non semi-simple matrices (cont.)) For the matrices considered above, A1 has one
maximal Jordan chain of length = 3 at , with
v 1 = (1, 0, 0), v 2 = (0, 1, 0), v 3 = (0, 0, 1).
A2 has two maximal Jordan chains at , one of length 1 = 2 and the other of length 2 = 1, with
v11 = (1, 0, 0), v12 = (0, 1, 0), v21 = (0, 0, 1).

Finally, A3 has three maximal Jordan chains at each of length 1 = 2 = 3 = 1, with

v11 = (1, 0, 0), v21 = (0, 1, 0), v31 = (0, 0, 1).
Notice that in all three cases the generalized eigenvectors are the same, but they are partitioned
differently into chains. Note also that in all three cases the generalized eigenvectors taken all together
form a linearly independent collection.
Consider now a change of basis

1
T = v11 . . . v11 v21 . . . v22 . . .
comprising the generalized eigenvectors as the columns of the matrix T 1 . It can be shown that the
matrix of generalized eigenvectors is indeed invertible.
Theorem 5.4 Assume that the matrix A Rnn has k linearly independent eigenvectors v1 , . . . , vk
Cn with corresponding maximal Jordan chains {vij }j=0
i
Cn , i = 1, . . . , k. Then
Pk
1. The matrix v11 . . . v11 . . . vk1 . . . vkk . . . Cnn is invertible. In particular, i=1 i = n.
2. A can be written in the Jordan canonical form A = T 1 JT where J Cnn is block-diagonal

i 1 0 . . . 0 0
J1 0 . . . 0 0 i 1 . . . 0
0 J2 . . . 0 0
.. Ci i , i = 1, . . . , k
J = . . . . Cnn , Ji = ... .. .. . .
. .
.
. .. .
.. .. . . ..
0 0 0 . . . i 1
0 0 . . . Jk
0 0 0 ... 0 i
and i C is the eigenvalue corrseponding to the Jordan chain {vij }j=0

i
. The matrices Ji are
known as the Jordan blocks of the matrix A.
The proof of this fact is rather tedious and will be omitted, see [?]. Notice that there may be multiple
Jordan chains for the same eigenvalue , in fact their number will be the same as the number of
linearly independent eigenvectors associated with . If k = n (equivalently, all Jordan chains have
length 1) then the matrix is semi-simple, T 1 is the matrix of eigenvectors of A, and J = .
Example (Non semi-simple matrices (cont.)) In the above example, the three matrices A1 ,
A2 and A3 are already in Jordan canonical form. A1 comprises one Jordan block of size 3, A2 two
Jordan blocks of sizes 2 and 1 and A3 three Jordan blocks, each of size 1.
How does this help with the computation of eAt ?
Theorem 5.5 eAt = T 1 eJt T where

J1 t
t2 ei t ti 1 ei t
e 0 ... 0 ei t tei t 2! ... (i 1)!
0 J2 t ti 2 ei t
e . . . 0 0 ei t te i t
...
eJt = . .. .. .. and eJi t =
.. .. ..
(i 2)!
..
, i = 1, . . . , k.

.. . . . ..
.
. . . .
0 0 . . . e Jk t 0 0 0 ... ei t
Proof: Exercise. Show that Aj = T 1 J j T , then that

j Jt
J1 0 . . . 0 e 1 0 ... 0
0 Jj . . . 0 0 e J2 t ... 0
2
Jj = . .. . .
Jt
.. , and hence that e = .. .. .. .. .
.. . . . . . . .
0 0 . . . Jk j 0 0 . . . e Jk t
Finally show that

t2 ei t ti 1 ei t
ei t tei t 2! ... (i 1)!
ti 2 ei t
0 ei t te i t
...
e Ji t =
.. .. ..
(i 2)!
..

..
. . . . .
0 0 0 ... ei t
by differentiating with respect to t, showing that the result is equal to Ji eJi t and invoking uniqueness
(or in a number of other ways).
So the computation of the matrix exponential becomes easy once again. Notice that if k = n then we
are back to the semi-simple case. As in the semi-simple case, all matrices involved in the product will
in general be complex. However, the theorem ensures that when taking the product the imaginary
parts will cancel and the result will be a real matrix.
Example (Non semi-simple matrices (cont.)) In the above example,

2 t t
et tet t2 et e tet 0 e 0 0
e A1 t = 0 et tet , eA2 t = 0 et 0 , e A3 t = 0 et 0 ,
0 0 et 0 0 et 0 0 et
Notice that in all cases eAt consists of linear combinations of elements of the form
ei t , tei t , . . . , ti 1 ei t
for i Spec[A] and i the length of the longest Jordan chain at i . In other words
X
eAt = (t)et (5.5)
Spec[A]
where for each Spec[A], (t) C[t]nn is a matrix of polynomials of t with complex coefficients
and degree at most equal to the length of the longest Jordan chain at . In particular, if A is semi-
simple all Jordan chains have length equal to 1 and the matrix exponential reduces to
X
eAt = et
Spec[A]
where Cnn are constant complex matrices. Notice again that even though in general both
the eigenvalues and the coefficients of the corresponding polynomials (t) are complex numbers,
because the eigenvalues appear in complex conjugate pairs the imaginary parts for the sum cancel
out and the result is the real matrix eAt Rnn .
5.4 Laplace transforms

To establish a connection to more conventional control notation, we recall the definition of the
Laplace transform of a signal f () : R+ Rnm :
Z
F (s) = L{f (t)} = f (t)est dt Cnm
0
where the integral is interpreted element by element and we assume that it is well defined (for a
careful discussion of this point see, for example, [?]). The Laplace transform L{f (t)} transforms the
real matrix valued function f (t) Rnm of the real number t R+ to the complex matrix valued
function F (s) Cnm of the complex number s C. The inverse Laplace transform L1 {F (s)}
performs the inverse operation; it can also be expressed as an integral, even though in the calculations
considered here one mostly encounters functions F (s) that are Laplace recognizable transforms of
known functions f (t); in particular the functions F (s) will typically be proper rational functions of
s whose inverse Laplace transfrom can be computed by partial fraction expansion.
Fact 5.3 The Laplace transform (assuming that it is defined for all functions concerned) has the
following properties:
1. It is linear, i.e. for all A1 , A2 Rpn and all f1 () : R+ Rnm , f2 () : R+ Rnm
L{A1 f1 (t) + A2 f2 (t)} = A1 L{f1 (t)} + A2 L{f2 (t)} = A1 F1 (s) + A2 F2 (s)

d
2. L dt f (t) = sF (s) f (0).
3. L {(f g)(t)} = F (s)G(s) where (f g)() : R+ Rpm denotes the convolution of f () :
R+ Rpn and g() : R+ Rnm defined by
Z t
(f g)(t) = f (t )g( )d.
0
Proof: Exercise, just use the definition and elementary calculus.

Fact 5.4 For all A Rnn and t R+ , L eAt = (sI A)1 .
Proof: Recall that

d At d At
e = AeAt L e = L AeAt
dt dt

sL eAt eA0 = AL eAt

sL eAt I = AL eAt

(sI A)L eAt = I.
The claim follows by multiplying on the left by (sI A); notice that the matrix is invertible for all
s C, except the eigenvalues of A.
Let us look somewhat more closely to the structure of the Laplace transform, (sI A)1 , of the
state transition matrix. This is an n n matrix of strictly proper rational functions of s, which as
we saw in Chapter 2 form a sub-ring of (Rp (s), +, ). To see this recall that by definition
Adj[sI A]
(sI A)1 =
Det[sI A]
The denominator is simply the characteristic polynomial,
A (s) = sn + 1 sn1 + . . . + n R[s]
of the matrix A, where i R for i = 1, . . . , n.

In the numerator is the adjoint of the matrix (sI A). Recall that the (i, j) element of this
n n matrix is equal to the determinant (together with the corresponding sign (1)i (1)j ) of the
sub-matrix of (sI A) formed by removing its j th row and the ith column. Since sI A has all
terms in s on the diagonal when we eliminate one row and one column of sI A we eliminate at
least one term containing s. Therefore the resulting sub-determinants will be polynomials of order
at most n 1 in s, in other words
Adj[sI A] R[s]nn and (sI A)1 Rp (s)nn .

Given this structure, let us write (sI A)1 more explicitly as

M (s)
(sI A)1 = (5.6)
A (s)
where M (s) = M0 sn1 + . . . + Mn2 s + Mn1 with Mi Rnn for i = 0, . . . , n 1.
Theorem 5.6 The matrices Mi satisfy
M0 = I
Mi = Mi1 A + i I for i = 1, . . . , n 1
Mn1 A + n I = 0.
Proof: Post multiplying equation (5.5) by (sI A)A (s) leads to

A (s)I = M (s)(sI A) (sn + 1 sn1 + . . . + n )I = M0 sn1 + . . . + Mn2 s + Mn1 (sI A).
Since the last identity must hold for all s C the coefficients of the two polynomials on the left and
on the right must be equal. Equating the coefficients for sn leads to the formula for M0 , for sn1
leads to the formula for M1 , etc.
The theorem provides an easily implementable algorithm for computing the Laplace transform of
the state transition matrix without having to invert any matrices. The only thing that is needed is
the computation of the characteristic polynomial of A and some matrix multiplications. In addition,
the following useful fact about square matrices can be deduced as a corollary.
Theorem 5.7 (Cayley-Hamilton) Every square matrix A Rnn satisfies its characteristic poly-
nomial, i.e.
A (A) = An + 1 An1 + . . . + n I = 0 Rnn .
Proof: By the last equation of Theorem 5.6, Mn1 A + n I = 0. From the next to last equation,
Mn1 = Mn2 A + n1 I; substituting this into the last equation leads to
Mn2 A2 + n1 A + n I = 0.
Substituting Mn2 from the third to last equation, etc. leads to the claim.
Corollary 5.2 Let A Rnn be a square matrix. For any k N, Ak can be written as a linear
combination of {I, A, A2 , . . . , An1 }.
The proof is left as an exercise. Stated another way, the corollary shows that all powers of
an n-dimensional square matrix live in a low-dimansional subspace of the n2 dimensional linear
space of square matrices; this subspace has dimension at most n and is spanned by the matrices
{I, A, A2 , . . . , An1 }.
The Laplace transform can also be used to compute the response of the system. Notice that taking
Laplace transforms of both sides of the differential equation governing the evolution of x(t) leads to
L
x(t) = Ax(t) + Bu(t) = sX(s) x0 = AX(s) + BU (s)
leading to
X(s) = (sI A)1 x0 + (sI A)1 BU (s). (5.7)
How does this relate to the solution of the differential equation that we have already derived? We
have shown that Z t
At
x(t) = e x0 + eAt Bu( )d = eAt x0 + (K u)(t)
0
where (K u)(t) denotes the convolution of the impulse state transition
K(t) = eAt B
with the input u(t). Taking Laplace transforms we obtain
X(s) = L{x(t)} =L{eAt x0 + (K u)(t)}

=L{eAt }x0 + L{(K u)(t)}
=(sI A)1 x0 + L{eAt B}L{u(t)}
=(sI A)1 x0 + (sI A)1 BU (s)
which coincides with equation (5.6), as expected.

Equation (5.6) provides an indirect, purely algebraic way of computing the solution of the differential
equation, without having to compute a matrix exponential or a convolution integral. One can form
the Laplace transform of the solution, X(s), by inverting (sI A) (using, for example, the matrix
multiplication of Theorem 5.6) and substituting into equation (5.6). From there the solution x(t)
can be computed by taking an inverse Laplace transform. Since (sI A)1 Rp (s)nn is a matrix of
strictly proper rational functions, X(s) Rp (s)n will be a vector of strictly proper rational functions,
with the characteristic polynomial in the denominator. The inverse Laplace transform can therefore
be computed by partial fraction expansions, at least for many reasonable input functions (constants,
sines and cosines, exponentials, ramps, polynomials, and combinations thereof).
Taking the Laplace transform of the output equation leads to
L
y(t) = Cx(t) + Du(t) = Y (s) = CX(s) + DU (s).
By (5.6)
Y (s) = C(sI A)1 x0 + (sI A)1 BU (s) + DU (s)
which for x0 = 0 (zero state response) reduces to
Y (s) = C(sI A)1 BU (s) + DU (s) = G(s)U (s). (5.8)
Definition 5.5 The function G() : C Cpm defined by
G(s) = C(sI A)1 B + D (5.9)
is called the transfer function of the system.
Comparing equation (5.7) with the zero state response that we computed earlier
Z t
y(t) = C eA(t ) Bu( )d + Du(t) = (H u)(t)
0
it is clear that the transfer function is the Laplace transform of the impulse response H(t) of the
system
G(s) = L{H(t)} = L{CeAt B + D0 (t)} = C(sI A)1 B + D.
Substituting equation (5.5) into (5.8) we obtain

M (s) CM (s)B + DA (s)
G(s) = C B+D = . (5.10)
A (s) A (s)
Since K(s) is a matrix of polynomials of degree at most n 1 and A (s) is a polynomial of degree
n we see that G(s) Rp (s)pm is a matrix of proper rational functions. If moreover D = 0 then the
rational functions are strictly proper.
Definition 5.6 The poles of the system are the values of s C are the roots of the denominator
polynomial of G(s).
From equation (5.9) it becomes apparent that all poles of the system are eigenvalues (i.e. are
contained in the spectrum) of the matrix A. Note, however, that not all eigenvalues of A are
necessarily poles of the system, since there may be cancellations of common factors in the numerator
and denominator when forming the fraction (5.9). It turns out that such cancellations are related
to the controllability and observability properties of the system. We will return to this point in
Chapter 8, after introducing these notions.

p
Problem 5.1 (Change of basis) Let {ui }m n m
i=1 , {xi }i=1 , {yi }i=1 be bases of the linear spaces (R , R),
n p m
(R , R) and (R , R), respectively. Let u() : R+ R be piecewise continuous. For t R+ , consider
the linear time-invariant system:
(
x(t) = Ax(t) + Bu(t),
(5.11)
y(t) = Cx(t) + Du(t),
p
with x(t0 ) = x0 , where all matrix representations are given w.r.t. {ui }m n
i=1 , {xi }i=1 , {yi }i=1 . Now let
xi }ni=1 be another basis of (Rn , R) and let T Rnn represent the change of basis from {xi }ni=1 to
{e
xi }ni=1 .
{e
1. Derive the representation of the system w.r.t. bases {ui }m xi }ni=1 , {yi }pi=1 .
i=1 , {e
2. Compute the transition map (t,e t0 ) and the impulse response matrix H(t,
e ) w.r.t. the new
representation. How do they compare with the corresponding quantities (t, t0 ) and H(t, )
in the original representation?
Problem 5.2 (Time-invariant Systems) Consider the linear time-invariant system of Problem 5.1.
1. Show that k

X A(t t0 )
(t, t0 ) = exp A(t t0 ) = .
k!
k=0
2. Given two matrices A1 , A2 Rnn show that, if A1 A2 = A2 A1 , then A2 exp(A1 t) = exp(A1 t)A2
and exp (A1 + A2 )t = exp(A1 t) exp(A2 t). Also show that these properties may not hold if
A1 A2 6= A2 A1 .
3. Show that the impulse response matrix satisfies H(t, ) = H(t , 0) (i.e. it depends only on
the difference t ).
Problem 5.3 (Discretization of Continuous-time Systems)
1. Consider the time-varying linear system
x(t) = A(t)x(t) + B(t)u(t),

()
y(t) = C(t)x(t) + D(t)u(t),
with initial condition x(t0 ) = x0 Rn , t t0 . Consider a set of time instants tk , with

k = 0, 1, 2, . . . , such that tk < tk+1 for all k. Let u(t) be constant between subsequent time
instants: u(t) = uk k N, t [tk , tk+1 ). Let xk+1 = x(tk+1 ) and yk+1 = y(tk+1 ) be the
state and the output of system () sampled at times tk . Show that there exist matrices Ak ,
Bk , Ck and Dk such that
xk+1 = Ak xk + Bk uk ,
()
yk = Ck xk + Dk uk .

2. Now assume that () is time-invariant, i.e. A(t), B(t), C(t), D(t) = (A, B, C, D), t t0 ,
and that there exists a fixed T > 0, tk+1 tk = T , k. Provide simplified expressions for Ak ,
Bk , Ck and Dk and show that they are independent of k.
Problem 5.4 (Realization) Consider the following n-th order scalar differential equation with
constant coefficients:
y (n) (t) + a1 y (n1) (t) + . . . + an1 y (1) (t) + an y(t) = u(t), t R+ , (5.12)
where y (i) (t) denotes the i-th derivative of y at t, {ai } R and u() : R+ R is a piecewise
continuous input. Show that (5.11) can be put in the form (5.10) for an appropriate definition of
the state x(t) Rn and of matrices A, B, C, D.
Problem 5.5 (Jordan Blocks) Let , 1 , 2 F , with F = R or F = C. Compute exp(A t) for

the following definitions of matrix A:

1
.. ..
0 1 . .
1. A = 1 ; 2. A = ; 3. A = F nn ,
0 2 0 1

where the elements not shown are zeroes. (Hint: in 3, consider the decomposition A = I + N , and
Pk k!
make use of (I + N )k = i=0 i!(ki)! (I)i N ki .)
Problem 5.6 (Jordan blocks and matrix exponential) For i = 1, . . . , m, let i C ni ni . De-
fine n = n1 + . . . + nm and the block diagonal matrix

1
2

= diag(1 , 2 , . . . , m ) = .. Rnn . ()
.
m

1. Show that exp() = diag exp(1 ), exp(2 ), . . . , exp(m ) .
2. Compute exp(i ) for the following definitions of i : (assume each entry is real)

1
1
.. ..
(a) i = .. ; (b) = . . ;
. i
1
ni

(c) i = R22 ,

where the elements not shown are zeroes. [Hint: in (b), consider the decomposition = I +N ,
P
and make use of (I + N )k = ki=0 i!(ki)!
k!
(I)i N ki .]
3. Assume that v = x + iy and v = x iy, with x, y Rn , are complex eigenvectors of a matrix

A Rnn , i.e. Av = v and Av = v for some = + i with , R. Let (), with
1 C 22 , be the Jordan decomposition of A Rnn corresponding to a basis of the form
{v, v , v3 , v4 , . . . vn }.
(a) Write the expression of 1 .

e 1 , 2 , . . . , m )
(b) Find a new basis and the corresponding change of basis T such that T T 1 = diag(
e
with 1 R 22 e
(real). What is the expression of 1 ?
Problem 5.7 (Modal Analysis) For t R+ , consider the ODE x = Ax, x(0) = x0 . Compute
exp(At) for the cases listed below. In each case provide a rough plot of the parametric curves x2 (t)
vs. x1 (t) for some initial conditions x0 .

0
1. A= , for < 0 and > 0.
0

2. A= , for < 0 and > 0.

2 1
3. A= 1 , for 1 < 0 < 2 , 1 , 2 < 0 and 1 , 2 > 0.
0 2

1 1
4. A= , for < 0, > 0 and = 0.
1 +1
Problem 5.8 (Matrix powers) Consider A Rnn , C Rpn and B Rnm . Show that
1. For any k N, Ak can be written as a linear combination of {I, A, A2 , . . . , An1 }.

2. CAk B = 0 for all k N if and only if CAk B = 0 for k = 0, 1, . . . , n 1.
Problem 5.9 (Nilpotent matrices) Show that a matrix A Rnn is nilpotent (i.e. there exists
k N such that Ak = 0) if and only if all of its eigenvalues are equal to zero. Show further that in
this case k n.
Problem 5.10 (Transfer function) Let A Rnn , B Rnm , C Rpn and D Rpm . Let
u() : R+ U be piecewise continuous. For t R+ , consider the linear time-invariant system:
(
x(t) = Ax(t) + Bu(t),
y(t) = Cx(t) + Du(t).
1. Consider a change of basis x = T x. Compute the transfer function with respect to the new
basis and compare it to the transfer function in the original basis.
2. Assume now that A is semisimple. Use your answer in part 1 to provide a simple formula for
G(s) in terms of the eigenvalues of A.
Chapter 6
Stability
Consider again the time varying linear system
x(t) = A(t)x(t) + B(t)u(t)

y(t) = C(t)x(t) + D(t)u(t).
Stability addresses the question of what happens to the solutions of this system as time, t, increases.
Do they remain bounded, will they get progressively smaller, or will they diverge to infinity. Stability
deals first and foremost with the properties of the differential equation. We will therefore ignore the
output equation for the time being. We will also start with the zero input case (u = U , or u(t) = 0
for all t R), i.e. by considering the solutions of
x(t) = A(t)x(t). (6.1)
We will then return to inputs and outputs in Section 6.4. As we did for the definition of the solutions
of the differential equation, we will start by considering general nonlinear systems (Section 6.1, then
specialize to linear time varying systems (Section 6.2), then specialize further to linear time invariant
systems (Section 6.3).
6.1 Nonlinear systems: Basic definitions

Consider again a general nonlinear, time varying system defined by a differential equation
x(t) = p(x(t), t) (6.2)
for t R, x(t) Rn and p : Rn R Rn . To ensure existence and uniqueness of solutions assume

that p is Lipschitz continuous in its first argument and piecewise continuous in its second. Let
s(t, t0 , x0 ) denote the unique solution of (6.2) at time t R passing through x0 Rn at time t0 R.
Though in general the computation of the solution function s : R Rn R Rn is impossible, for
some x0 the solution may become particularly easy.
Definition 6.1 A state x Rn is called an equilibrium of system (6.2) if and only if p(x, t) = 0 for
all t R.
The following fact is an immediate consequence of this observation.
Fact 6.1 If x Rn is an equilibrium of system (6.2) then s(t, t0 , x) = x for all t, t0 R.
83
Proof: Note that for the proposed solution s(t0 , t0 , x) = x and
d d
s(t, t0 , x) = x = 0 = p(x, t) = p(s(t, t0 , x), t).
dt dt
The conclusion follows by existence and uniqueness of solutions.
The fact shows that a solution which passes through an equilibrium, x, at some point in time is
doomed to stay on the equilibrium for all times. We call this constant solution the equilibrium
solution defined by the equilibrium x.
What if a solution passes close to the equilibrium, but not exactly through it? Clearly such a solution
will no longer be identically equal to the equilibrium, but will it move away from the equilibrium,
or will it remain close to it, even converge to it, and if so at what rate?
To answer these questions we first need to fix a norm on Rn to be able to measure distances. Any
norm will do since they are all equivalent, for simplicity we will use the Euclidean norm throughout.
Equiped with this norm, we can now formalize the above questions in the following definition.
Definition 6.2 Let x Rn be an equilibrium of system (6.2). This equilibrium is called:
1. Stable if and only if for all t0 R, and all > 0, there exists > 0 such that
kx0 xk < ks(t, t0 , x0 ) xk < , t t0 .
2. Unstable if and only if it is not stable.

3. Uniformly stable of and only if for all > 0 there exists > 0 such that for all t0 R
kx0 xk < ks(t, t0 , x0 ) xk < , t t0 .
4. Locally asymptotically stable if and only if it is stable and for all t0 R there exists M > 0
such that
kx0 xk M lim ks(t, t0 , x0 ) xk = 0.
t
5. Globally asymptotically stable if and only if it is stable and for all (t0 , x0 ) R Rn
lim ks(t, t0 , x0 ) xk = 0.
t
6. Locally exponentially stable if and only if for all t0 R there exist , m, M > 0 such that for
all all t t0
kx0 xk M ks(t, t0 , x0 ) xk mkx0 xke(tt0 ) .
7. Globally exponentially stable if and only if for all t0 R there exist , m > 0 such that for all
x0 Rn and all t t0
ks(t, t0 , x0 ) xk mkx0 xke(tt0 ) .
Special care is needed in the above definition: The order of the quantifiers is very important. Note
for example that the definition of stability implicitly allows to depend on t0 and ; one sometimes
writes (t0 , ) to highlight this dependence. On the other hand, in the definition of uniform stability
can depend on but not on t0 , i.e. the same must work for all t0 ; one sometimes uses the
notation () to highlight this fact. Likewise, the definition of global exponential stability requires
and m to be independent of x0 , i.e. the same and m must work for all x0 Rn ; a variant of
this definition where m and are allowed to depend on x0 is sometimes referred to as semi-global
exponential stability.
The definition distinguishes stability concepts along three axes. The most fundamental distinction
deals with the convergence of nearby solutions to the equilibrium. The equilibrium is called unstable
if we cannot keep solutions close to it by starting sufficiently close, stable if we can keep solutions
as close as we want by starting them sufficiently close, asymptotically stable if in addition nearby
solutions converge to the equilibrium, and exponentially stable if they do so at an exponential
rate. The second distinction deals with how these properties depend on the starting time, t0 : For
uniform stability the starting time is irrelevant, the property holds the same way irrespective of
when we look at the system. The third distinction deals with how these properties depend on the
starting state, x0 : Local implies that the property holds provided we start close enough to the
equilibrium, whereas global requires that the property holds irrespective of where we start. Note
that this distinction is irrelevant for stability and uniform stability, since the conditions listed in
the definition are required to hold provided we start close enough. One can also combinatorially
mix these qualities to define other variants of stability notions: Uniform local asymptotic stability
(where the equilibrium is uniformly stable and the convergence rate is independent of the starting
time), uniform global exponential stability, etc. We will not pursue these variants of the definitions
here, since most of them turn out to be irrelevant when dealing with linear systems.
It is easy to see that the notions of stability introduced in Definition 6.2 are progressively stronger.
Fact 6.2 Consider an equilibrium of system 6.2. Then the following statements are true:
If the equilibrium is then it is also

uniformly stable stable
locally asymptotically stable stable
globally asymptotically stable locally asymptotically stable
locally exponentially stable locally asymptotically stable
globally exponentally stable globally asymptotically stable
globally exponentally stable locally exponentially stable
Proof: Most of the statements are obvious from the definition. Asymptotic stability requires
stability, global asymptotic stability implies that the conditions of local asymptotic stability hold
for any M > 0, etc. The only part that requires any work is showing that local/global exponential
stability implies local/global asymptotic stability.
Consider a globally1 exponentially stable equilibrium x, i.e. assume that for all t0 there exist
, m > 0 such that for all x0 , ks(t, t0 , x0 ) xk mkx0 xke(tt0 ) for all t t0 . For t0 R and
> 0 take = /m. Then for all x0 Rn such that kx0 xk < and all t t0
ks(t, t0 , x0 ) xk mkx0 xke(tt0 ) < me(tt0 ) = e(tt0 ) .
Hence the equilibrium is stable. Moreover, since by the properties of the norm ks(t, t0 , x0 ) xk 0,
0 lim ks(t, t0 , x0 ) x0 k lim mkx0 xke(tt0 ) = 0.
t t
Hence limt ks(t, t0 , x0 ) xk = 0 and the equilibrium is asymptotically stable.
It is also easy to see that the stability notions of Definition 6.2 are strictly stronger one from the
other; in other words the converse implications in the table of Fact 6.2 are in general not true. We
show this through a series of counter-examples.
Example (Stable, non-uniformly stable equilibrium) For x(t) R consider the linear, time
varying system
2t
x(t) = x(t) (6.3)
1 + t2
1 The argument for local exponential stability is effectively the same.

Exercise 6.1 Show that the system has a unique equilibrium at x = 0. Show further that
1 + t20
s(t, t0 , x0 ) = x0
1 + t2
by differentiating and invoking existence and uniqueness of solutions.
Typical trajectories of the system for x0 = 1 and different values of t0 are shown in Figure ??.
It is easy to see that x = 0 is a stable equilibrium. Indeed, given t0 R and > 0 let = /(1 + t20).
Then for all x0 Rn such that kx0 k < ,

1 + t20

ks(t, t0 , x0 )k = x0 < .
1+t 2 1 + t2
However, the equilibrium is not uniformly stable: For a given we cannot find a that works for all
t0 R. To see this, notice that for t0 0, ks(t, t0 , x0 )k reaches a maximum of (1 + t20 )kx0 k at t = 0.
Hence to ensure that ks(t, t0 , x0 )k < we need to ensure that (1 + t20 )kx0 k < which is impossible
to do by restricting
p x0 alone; for any 0 < < and kx0 k < we can make ks(0, t0 , x0 )k > by
taking t0 < / 1.
Example (Stable, non asymptotically stable equilibrium) For x(t) R2 consider the linear,
time invariant system

0 x01
x(t) = x(t) with x(0) = x0 = .
0 x02
Since the system is linear time invariant we can take t0 = 0 without loss of generality.
Exercise 6.2 Show that the system has a unique equilibrium x = 0. Show further that

cos(t) sin(t)
(t, 0) =
sin(t) cos(t)
by differentiating and invoking Theorem 4.2.
Typical trajectories of the system for different values of x0 are shown in Figure ??. Clearly

x01 cos(t) x02 sin(t)
x(t) = (t, 0)x0 = .
x01 sin(t) + x02 cos(t)
Using the 2-norm leads to

kx(t)k2 = kx0 k2 .
Therefore the system is uniformly stable (take = ) but not asymptotically stable (since in general
limt kx(t)k = kx0 k 6= 0).
Exercise 6.3 Show that linear time invariant system x(t) = 0 with x(t) R also has a stable, but
not asymptotically stable equilibrium x = 0. Does this system have any other equilibria? Are they
stable? Asymptotically stable?
Example (Asymptotically stable, non exponentially stable equilibrium) Let us return to

the system of equation (6.3). Recall that x = 0 is a stable equilibrium. Moreover,
1
lim ks(t, t0 , x0 )k = kx0 k(1 + t20 ) lim =0
t t 1 + t2
for all t0 , x0 . Hence the equilibrium is globally asymptotically stable. It is, however not exponentially
stable (not even locally). Assume, for the sake of contradiction, that for all t0 R there exist
, m, M > 0 such that for all x0 R with kx0 k M and all t t0 ,
ks(t, t0 , x0 )k mkx0 ke(tt0 )
In particular, for x0 = M and t0 = 0 this would imply that for all t 0

et
ks(t, 0, 1)k mM et < mM.
1 + t2
et
Since 1+t2 as t this is a contradiction.
Notice that so far all the counter-examples have involved linear systems (sometimes time varying).
As we will see in the next section, for linear systems local and global asymptotic stability are
equivalent concepts. To distinguish between them we therefore need to resort to non-linear systems.
Example (Locally, non-globally asymptotically stable equilibrium) Consider a two dimen-

sional state vector x = (x1 , x2 ) R2 whose evolution is goverened by the following differential
equations
x1 (t) = x2 (t)
x2 (t) = x1 (t) x1 (t)3 x2 (t).
It is easy to see that this system has three equilibria, at (0, 0), (1, 0) and (1, 0) respectively.
Exercise 6.4 Verify that these are indeed the only equilibria of this system. Is the system linear or
nonlinear? What is the function p in x(t) = p(x(t), t)? Is it globally Lipshitz in its first argument?
Clearly none of the three equilibria can be globally asymptotically stable: It is impossible for all
trajectories to converge to a particular equilibrium, since those starting at another equilibrium will
stay put. To study whether some of these equilibria are stable or locally asymptotically stable one
can compute the linerization of the system about each of the equilibria and study the stability of
the resulting linear system using the methods presented later in this chapter. It is reasonable to
assume that nearby the equilibrium, where the terms neglected in the linearization are small, the
behavior of the nonlinear system will be similar to that of its linearization. Hence if the linearization
is asymptotically stable one whould expect the equilibrium to be locally asymptotically stable for
the nonlinear system. This argument can in fact be formalized, leading to the so-called Lyapunov
Indirect Method for checking local asymptotic stability, or instability of nonlinear systems. The
interested reader is referred to [?] or [?] for more information on this topic.
For this system, linearization suggests that one would expect the equilibria (1, 0) and (1, 0) to be
locally asymptotically stable and the equilibrium (0, 0) to be unstable (Problem ??). This can be
visually confirmed by simulating the system for various initial conditions and plotting the trajectories
x(t). The most informative way of doing this is to generate a parametric plot of x1 (t) against x2 (t)
parametrized by t. This so-called phase plane plot for this system is shown in Figure ??.
Before restricting our attention to linear system we point out two more general facts about the
stability concepts introduced in Definition 6.2. The first is an intimate relation between stability
and continuity. To expose this link we need to think of the function mapping initial conditions to
state trajectories from the initial time t0 onwards. Since state trajectories are continuous functions
of time, for each t0 R one can think of this function as a map between the state space Rn and the
space of continuous functions C([t0 , ), Rn )
s(, t0 , ) : Rn C([t0 , ), Rn )
x0 7 {s(, t0 , x0 ) : [t0 , ) Rn }.
The strange notation is meant to alert the reader to the fact that we consider s(, t0 , ) for fixed t0
as a function mapping a vector (denoted by the placeholder ) to a function of time (denoted by
the placeholder , left over after x0 is substituted for ).
Recall that for the stability definitions we have equipped Rn with a norm k k. We now equip
C([t0 , ), Rn ) with the corresponding infinity norm
ks(, t0 , x0 )kt0 , = sup ks(t, t0 , x0 )k, (6.4)

tt0
where we include t0 in the notation to make the dependence on initial time explicit. Notice that the
first norm in Equation (6.4) is a norm on the infinite dimensional function space (i.e., s(, t0 , x0 )
C([t0 , ), Rn ) is though of as a function of time), whereas the second norm is a norm on the finite
dimensional state space (i.e., s(t, t0 , x0 ) Rn is the value of this function for the specific time
t [t0 , )).
Fact 6.3 An equilibrium, x, of system (6.2) is stable if and only if for all t0 R the function
s(, t0 , ) mapping the normed space (Rn , k k) into the normed space (C([t0 , ), Rn ), k kt0 , ) is
continuous at x.
Proof: The statement is effectively a tautology. Fix t0 R and recall that, according to Defini-
tion 3.6, s(, t0 , ) is continuous at x if and only for all > 0 there exists > 0 such that
kx0 xk < ks(, t0 , x0 ) s(, t0 , x)kt0 , < .
By Equation (6.4) this is equivalent to
kx0 xk < ks(t, t0 , x0 ) s(t, t0 , x)k < t t0 ,
which, recalling that s(t, t0 , x) = x for all t t0 is in turn equivallent to
kx0 xk < ks(t, t0 , x0 ) xk < t t0 ,
which is precisely the definition of stability.
A similar relation between uniform stability and uniform continuity (where the above is indepen-
dent of t0 ) can also be derived in the same way.
The second general fact relates to the possible rate of convergence. The strongest notion of stability
in Definition 6.2, namely exponential stability, requires that solutions converge to the equilibrium
exponentially (i.e. rather quickly) in time. Could they converge even faster? Could we, for example,
introduce another meaningful stability definition that requires solutions to coverge with a rate of
2
et for some > 0? And if not, can we at least increase in the exponential convergence? The
following fact reveals that Lipschitz continuity imposes a fundamental limit on how fast convergence
can be.
Fact 6.4 Let x be an equilibrium of system (6.2) and assume that there exists k > 0 such that for
all x, x Rn , kp(x, t) p(x , t)k kkx x k. Then for all t0 R and all t t0
kx0 xkek(tt0 ) ks(t, t0 , x0 ) xk kx0 xkek(tt0 )
Proof: If x0 = x the claim is trivially true, we therefore restrict attention to the case x0 6= x. Note
that in this case we must have s(t, t0 , x0 ) 6= x for all t; if s(t, t0 , x0 ) = x for some t then s(t, t0 , x0 )
must be the equilibrium solution and s(t, t0 , x0 ) = x for all t which, setting t = t0 contradicts the
fact that x0 6= x.
Recall that for simplicity we are using the Euclidean norm. Hence ks(t, t0 , x0 ) xk2 = (s(t, t0 , x0 )
x)T (s(t, t0 , x0 ) x) and

d
ks(t, t0 , x0 ) xk2 = d s(t, t0 , x0 )T (s(t, t0 , x0 ) x) + (s(t, t0 , x0 ) x)T d s(t, t0 , x0 )
dt dt dt

= p(s(t, t0 , x0 ), t)T (s(t, t0 , x0 ) x) + (s(t, t0 , x0 ) x)T p(s(t, t0 , x0 ), t)

p(s(t, t0 , x0 ), t)T (s(t, t0 , x0 ) x) + (s(t, t0 , x0 ) x)T p(s(t, t0 , x0 ), t)
kp(s(t, t0 , x0 ), t)T k ks(t, t0 , x0 ) xk + k(s(t, t0 , x0 ) x)T k kp(s(t, t0 , x0 ), t)k
= 2ks(t, t0 , x0 ) xk kp(s(t, t0 , x0 ), t)k
= 2ks(t, t0 , x0 ) xk kp(s(t, t0 , x0 ), t) p(x, t)k
2kks(t, t0 , x0 ) xk ks(t, t0 , x0 ) xk.
On the other hand,

d
ks(t, t0 , x0 ) xk2 = 2ks(t, t0 , x0 ) xk d ks(t, t0 , x0 ) xk .
dt dt
Since s(t, t0 , x0 ) 6= x, combining the two equations we must have

d
ks(t, t0 , x0 ) xk kks(t, t0 , x0 ) xk
dt
or in other words
d
kks(t, t0 , x0 ) xk ks(t, t0 , x0 ) xk kks(t, t0 , x0 ) xk.
dt
Applying the Gronwall Lemma (Theorem 3.9) to the right inequality leads to
ks(t, t0 , x0 ) xk kx0 xkek(tt0 ) .
From the right inequality (adapting the steps of the proof of the Gronwall Lemma) we have
d d d
ks(t, t0 , x0 ) xkek(tt0 ) = (ks(t, t0 , x0 ) xk) ek(tt0 ) + ks(t, t0 , x0 ) xk ek(tt0 )
dt dt dt
kks(t, t0 , x0 ) xkek(tt0 ) + ks(t, t0 , x0 ) xkkek(tt0 ) = 0.
Hence for all t t0 ,
ks(t, t0 , x0 ) xkek(tt0 ) ks(t0 , t0 , x0 ) xkek(t0 t0 ) = kx0 xk,
which leads to
ks(t, t0 , x0 ) xk kx0 xkek(tt0 ) .
The fact reveals that convergence to an equilibrium can be at most exponential and the rate can
be at most equal to the Lipschitz constant of p. On the bright side, the fact also shows that if
an equilibrium is unstable divergence cannot be any faster than exponential. Even though a fixed
Lipshitz constant is assumed to simplify the proof it is easy to see that the claim still holds if the
Lipschitz constant is time varying but bounded from above and below; one simply needs to replace
k by its lower bound in the left inequality and its upper bound in the right inequality. The lower
bound on the Lipschitz constant also provides a bound for the rate of exponential convergence.
6.2 Linear time varying systems

We note that Definition 6.2 is very general and works also for nonlinear systems. Since for linear
systems we know something more about the solution of the system it turns out that the conditions of
the definition are somewhat redundant in this case. In particular, since s(t, t0 , x0 , U ) = (t, t0 )x0 ,
the solution is linear with respect to the initial state and all the stability definitions indirectly refer
to the properties of the state transition matrix (t, t0 ).
First note that for all t0 R+ if x0 = 0 then
s(t, t0 , 0, 0) = (t, t0 )x0 = 0 Rn t R+
is the solution of (6.1). Another way to think of this observation is that if x(t0 ) = 0 for some
t0 R+ , then
x(t0 ) = A(t)x(t0 ) = 0
therefore the solution of the differential equation does not move from 0. Either way, a zero input
solution that passes through the state x(t0 ) = 0 at some time t0 R+ will be identically equal to
zero for all times.
Exercise 6.5 Can there be other x0 6= 0 such that s(t, t0 , x0 , 0) = x0 for all t R+ ?
Theorem 6.1 The equilibrium x = 0 of (6.1) is:
1. Stable if and only if for all t0 R, there exists M > 0 such that
ks(t, t0 , x0 , U )k M kx0 k t t0 , x0 Rn .
2. Locally asymptotically stable if and only if limt (t, 0) = 0 Rnn .
Proof: Part 1: By Fact ??, the equilibrium x = 0 is stable if for all t0 R the function
s(, t0 , , 0) : (Rn , k k2 ) (C([t0 , ), Rn ), k kt0 , )

x0 7 s(, t0 , x0 , 0) = (, t0 )x0
is continuous at x0 = 0. Recall that for all t0 R+ this function is linear in x0 (by Theorem 4.1).
Therefore (by Theorem 3.5) the function is continuous at x0 = 0 if and only if it is continuous at all
x0 Rn , if and only if its induced norm is finite. In other words, the equilibrium solution is stable
if and only if for all t0 R+ there exists M > 0 such that
ks(, t0 , x0 , U )kt0 ,
sup M ks(, t0 , x0 , U )kt0 , M kx0 k2 x0 Rn
x0 6=0 kx0 k2
ks(t, t0 , x0 , U )k2 M kx0 k2 t t0 , x0 Rn .
Part 2: Assume first that limt (t, 0) = 0 and show that the equilibrium solution is asymptotically
stable. For all (x0 , t0 ) Rn R+
ks(t, t0 , x0 , 0)k2 = k(t, t0 )x0 k2 = k(t, 0)(0, t0 )x0 k2 k(t, 0)k k(0, t0 )x0 k2
(where we use the induced 2-norm for the matrix (t, 0)). Therefore, for all x0 Rn , t0 R,
lim ks(t, t0 , x0 , U )k = 0.
t
To establish that the equilibrium is asymptotically stable it therefore suffices to show that it is stable.
Recall that the function s(, t0 , x0 , U ) = (, t0 )x0 is continuous (by Theorem 4.1). Therefore, since
limt ks(t, t0 , x0 , 0)k = 0, we must have that s(, t0 , x0 , 0) is bounded, i.e. there exists M > 0 such
that for all x0 Rn
s(t, t0 , x0 , U ) M kx0 k t t0 .
Hence the equilibrium solution is stable by Part 1.
Conversely, assume that the equilibrium solution is locally asymptotically stable. Then there exists
M > 0 such that for all t R and all x0 Rn with kx0 k M ,
lim s(t, t0 , x0 , 0) = lim (t, t0 )x0 = lim ((t, 0)(0, t0 )x0 ) = lim ((t, 0)) (0, t0 )x0 = 0 (6.5)
t t t t
Consider the canonical basis {ei }ni=1 n

for R and take t0 = 0. Letting x0 = ei in (6.5) shows that the
ith column of (t, 0) tends to zero. Repeating for i = 1, . . . , n establishes the claim.
From this theorem one can further conclude that when it comes to linear systems (even linear time
varying systems) local stability notions are equivallent to global ones.
Fact 6.5 The equilibrium solution x = 0 of (6.1) is
1. Globally asymptotically stable if and only if it is locally asymptotically stable.

2. Globally exponentially stable if and only if it is locally exponentially stable.
6.3 Linear time invariant systems

For linear time invariant systems the stability discussion simplifies even further. For one thing, it is
easy to see that in this case there is no difference between uniform and non-uniform stability notions.
Fact 6.6 The equilibrium x = 0 of the linear time invariant system x(t) = Ax(t) is uniformly stable
if and only if it is stable.

We have seen that for linear systems there is no distinction between local and global asymp-
totic/exponential stability (Fact ??). Asymptotic stability, however, is still a weaker requirement
than exponential stability, even for time varying linear systems. For time invariant linear systems it
turns out that asymptotic stability and exponential stability are also equivallent. Moreover, one can
determine whether a time invariant linear system is exponentially stable through a simple algebraic
calculation.
Theorem 6.2 For linear time invariant systems the following statements are equivalent:
1. The equilibrium x = 0 is asymptotically stable.

2. The equilibrium x = 0 is exponentially stable.
3. For all Spec[A], Re[] < 0.
We start the proof by establishing the following lemma.
Lemma 6.1 For all > 0 there exists m > 0 such that for all t R+
keAt k me(+)t
where k k denotes any induced norm on Rnn and = max{Re[] | Spec[A]}.
Proof: Recall that the existence of the Jordan canonical form implies that
X
eAt = (t)et
Spec[A]
where (t) C[t]nn are n n matrices of polynomials in t with complex coefficients. Consider
(for simplicity) the infinity induced norm k k for Cnn . Then

X X

keAt k = (t)e t
k (t)k et .
Spec[A] Spec[A]

For Spec[A] let = + j with , R and note that

t (+j)t t jt
e = e = e e = et |cos(t) + j sin(t)| = et .
Therefore, since = Re[] ,

X
keAt k k (t)k et .
Spec[A]
Recall that k (t)k is the maximum among the rows of (t) of the sum of the magnitudes of the
elements in the row. Since all entries are polynomial, then there exists a polynomial p (t) R[t]
such that p (t) k (t)k for all t R+ . If we define the polynomial p(t) R[t] by
X
p(t) = p (t),
Spec[A]
then
keAt k p(t)et . (6.6)
t
Since p(t) is a polynomial in t R+ , for any > 0 the function p(t)e is continuous and
limt p(t)et = 0. Therefore p(t)et is bounded for t R+ , i.e. there exists m > 0 such
that p(t)et m for all t R+ . Substituting this into equation (6.6) leads to
> 0 m > 0, keAt k me(+)t .
Proof: (Of Theorem 6.2) We have already seen that 2 1 (Fact 6.2).
3 2: If all eigenvalues have negative real part then
= max{Re[] | Spec[A]} < 0.
Consider (0, ) and set = ( + ) > 0. By Lemma 6.1 there exists m > 0 such that
keAt k met .
Therefore for all (x0 , t0 ) Rm R+ and all t t0

ks(t, t0 , x0 , 0)k = eA(tt0 ) x0 eA(tt0 ) kx0 k mkx0 ke(tt0 ) .
Hence the equilibrium solution is exponentially stable.

1 3: By contraposition. Assume there exists Spec[A] such that Re[] 0 and let v Cn
denote the corresponding eigenvector. Without loss of generality take t0 = 0 and note that
s(t, 0, v, 0) = et v Cn .
This follows by existence and uniqueness of solutions, since e0 v = v and

d t
e v = et v = et Av = Aet v.
dt
For clarity of exposition we distinguish two cases.
First case: is real. Then
ks(t, 0, v, 0)k = et kvk.
Since 0, ks(t, 0, e1 , 0)k is either constant (if = 0), or tends to infinity as t tends to infinity (if
> 0). In either case the equilibrium solution cannot be asymptotically stable.
Second case: is complex. Let = + j C with , R, 0 and 6= 0. Let also
v = v1 + jv2 Cn for v1 , v2 Rn . Recall that since A Rnn eigenvalues come in complex
conjugate pairs, so = j is also an eigenvalue of A with eigenvector v = v1 jv2 . Since v is
an eigenvector by definition v 6= 0 and either v1 6= 0 or v2 6= 0. Together with the fact that 6= 0
this implies that v2 6= 0; otherwise A(v1 + jv2 ) = ( + j)(v1 + jv2 ) implies that Av1 = ( + j)v1
which cannot be satisfied by any non-zero real vector v1 . Therefore, without loss of generality we
can assume that kv2 k = 1. Since
j(v1 + jv2 ) + j(v1 jv2 ) jv + jv
v2 = =
2 2
by the linearity of the solution
s(t, 0, 2v2 , 0) = js(t, 0, v1 + jv2 , 0) + js(t, 0, v1 jv2 , 0)
= je(+j)t (v1 + jv2 ) + je(j)t (v1 jv2 )

= et jejt (v1 + jv2 ) + jejt (v1 jv2 )
= et [j(cos(t) + j sin(t))(v1 + jv2 ) + j(cos(t) j sin(t))(v1 jv2 )]
= 2et [v2 cos(t) + v1 sin(t)] .
Hence
ks(t, 0, 2v2 , 0)k = 2et kv2 cos(t) + v1 sin(t)k .
Note that
kv2 cos(t) + v1 sin(t)k = kv2 k = 1
whenever t = k/ for k N. Since 0 the sequence
n o

{ks(k/, 0, 2v2 , 0)k}k=0 = 2ek/
k=0
is bounded away from zero; it is either constant (if = 0) or diverges to infinity (if > 0). Hence
ks(t, 0, 2v1 , 0)k cannot converge to 0 as t tends to infinity and the equilibrium solution cannot be
asymptotically stable.
We have shown that 1 3 2 1. Hence the three statements are equivalent.
Theorem 6.2 does not hold is the system is time varying: We have already seen an example of
a system that is asymptotically but not exponentially stable, so the three statements cannot be
equivalent. How about 3 1? In other words if the eigenvalues of the matrix A(t) Rnn have
negative real parts for all t R is the equilibrium solution of x(t) = A(t)x(t) asymptotically stable?
This would provide an easy test of stability for time varying systems. Unfortunately, however, this
statement is not true.
Example (Unstable time varying system) For a (1, 2) consider the matrix

1 + a cos2 (t) 1 a cos(t) sin(t)
A(t) =
1 + a cos(t) sin(t) 1 + a sin2 (t)
Exercise 6.6 Show that the eigenvalues 1 , 2 of A(t) have Re[i ] = 2

2 for i = 1, 2. Moreover,
(a1)t
e cos(t) et sin(t)
(t, 0) = (a1)t .
e sin(t) et cos(t)
Therefore, if we take x0 = (1, 0) then

(a1)t
e cos(t)
s(t, 0, x0 , U ) = ks(t, 0, x0 , U )k2 = e(a1)t .
e(a1)t sin(t)
Since a (1, 2), Re[i ] < 0 for all t R+ but nonetheless ks(t, 0, x0 , U )k2 and the equilibrium
solution is not stable.
As the example suggests, determining the stability of time varying linear systems is in general
difficult. Some results exist (for example, if A(t) has negative eigenvalues, is bounded as a function
of time and varies slowly enough) but unfortunately, unlike time invariant linear systems, there
are no simple, general purpose methods that can be used to investigate the stability of time varying
linear systems.
An argument similar to that of Theorem 6.2 can be used to derive stability conditions for time
invariant systems in terms of the eigenvalues of the matrix A.
Theorem 6.3 The equilibrium solution of a linear time invariant systems is stable if and only if
the following two conditions are met:
1. For all Spec[A], Re[] 0.

2. For all Spec[A] such that Re[] = 0, all Jordan blocks of have dimension 1.
6.4 Systems with inputs

What if the inputs are non-zero? Based on our discussion of the structure of the solutions s(t, t0 , x0 , u)
and (t, t0 , x0 , u) it should come as no surprise that the properties of the zero input solution
s(t, t0 , x0 , 0) to a large extend also determine the properties of the solution under non-zero inputs.
Recall that for simplicity we consider the Euclidean norm for Rn and for f () : R Rn we defined
kf ()kt0 , = sup kf (t)k.
tt0
Similar definitions can be given for the matrix valued functions A(t), B(t), etc. using the corre-
sponding induced norms. Recall also that for existence and uniqueness of solutions we assume that
A(t), B(t), etc. are piecewise continuous functions of time.
Theorem 6.4 Assume that
1. The equilibrium solution is exponentially stable, i.e. there exists m, > 0 such that for all
(x0 , t0 ) Rn R+ and all t R+ , ks(t, t0 , x0 , U )k mkx0 ke(tt0 ) .
2. For all t0 R, kA()kt0 , , kB()kt0 , , kC()kt0 , , kD()kt0 , are bounded.
Then for all (x0 , t0 ) Rn R and all u() : [t0 , ) Rm with ku()kt0 , bounded,
m
ks(, t0 , x0 , u)kt0 , mkx0 ket0 + kB()kt0 , ku()kt0 ,
hm i
k(, t0 , x0 , u)kt0 , mkC()kt0 , kx0 ket0 + kC()kt0 , kB()kt0 , + kD()kt0 , ku()kt0 , .

If in addition limt u(t) = 0 then limt s(t, t0 , x0 , u) = 0 and limt (t, t0 , x0 , u) = 0.
One can see that also in the case of systems with inputs there is an intimate relation between stability
and continuity of the solution functions in an appropriate function space. To see this consider the
functions
s(, t0 , x0 , ) : PC([t0 , ), Rm ) C([t0 , ), Rn )

{u() : [t0 , ) Rm } 7 {s(, t0 , x0 , u) : [t0 , ) Rn }
(, t0 , x0 , ) : PC([t0 , ), Rm ) PC([t0 , ), Rp )
{u() : [t0 , ) Rm } 7 {(, t0 , x0 , u) : [t0 , ) Rp }
The strange notation is again meant to alert the reader to the fact that we consider s(, t0 , x0 , ) for
fixed t0 , x0 as a function mapping a piecewise continuous function of time (denoted by the placeholder
) to a continuous function of time (denoted by the placeholder , left over after u is substituted for
). Then Theorem ?? directly implies the following.
Corollary 6.1 If the linear time varying system is exponentially stable then the functions s(, t0 , x0 , )
and (, t0 , x0 , ) are continuous.

The properties guaranteed by Theorem ?? are known as the bounded input bounded state (BIBS)
and the bounded input bounded output (BIBO) properties. We just note that exponential stability
is in general necessary, since asymptotic stability is not enough.
Example (Unstable with input, asymptotically stable without) For x(t), u(t) R and
t R+ consider the system
1
x(t) = x + u.
1+t
Exercise 6.7 Show that (t, t0 ) = (1 + t0 )/(1 + t).
Let x(0) = 0 and apply the constant input u(t) = 1 for all t R+ . Then
Z t Z t
1+ t + t2 /2
s(t, 0, 0, 1) = (t, )u( )d = d =
0 0 1+t 1+t
even though the equilibrium solution is asymptotically stable and the input is bounded since
ku()k = 1 < .
6.5 Lyapunov equation

In addition to computing eigenvalues, there is a second algebraic test that allows us to determine
the asymptotic stability of linear time invariant systems, by solving the linear equation
AT P + P A = Q
(known as the Lyapunov equation). To formally state this fact recall that a matrix P Rnn is
called symmetric if and only if P T = P . A symmetric matrix, P = P T Rnn is called positive
definite if and only if for all x Rn with x 6= 0, xT P x > 0; we then write P = P T > 0.
Theorem 6.5 The following statements are equivalent:
1. The equilibrium solution of x(t) = Ax(t) is asymptotically stable.

2. For all Q = QT > 0 the equation AT P + P A = Q admits a unique solution P = P T > 0.
The proof of this fact is deferred to the next section. For the time being we simply show what can
go wrong with the Lyapunov equation if the system is not asymptotically stable.
Example (Lyapunov equation for scalar systems) Let x(t) R and consider the linear system
x(t) = ax(t) for a R; recall that the solution is simply x(t) = eat x(0). For this system the
Lyapunov equation becomes
aT p + pa = 2pa = q
for p, q R. The matrix q is automatically symmetric and it is positive definite if and only if
q > 0. If a < 0 (i.e. the equilibrium solution is asymptotically stable) the Lyapunov equation has
the unique positive definite solution p = q/(2a) > 0. If a = 0 (i.e. the equilibrium solution is stable
but not asymptotically stable) the Lyapunov equation does not have a solution. Finally, if a > 0 (i.e.
the equilibrium solution is unstable) the Lyapunov equation has a unique solution p = q/(2a) < 0
but no positive definite solution.

Problem 6.1 (Stability) For t R+ , consider the ODE x(t) = A(t)x(t), where, for a R,

1 + a cos2 (t) 1 a cos(t) sin(t)
A(t) =
1 a cos(t) sin(t) 1 + a sin2 (t),
1. Show that
e(a1)t cos(t) et sin(t)
(t, 0) =
e(a1)t sin(t) et cos(t).
2. Deduce how the stability properties of the system change with the value of a.
3. Find the values of a for which the eigenvalues of A(t) have negative real part for all t 0.
Compare the result with that of 2.
Problem 6.2 (Misc.)
1. Let A Rnm , B Rmp and || || be any induced norm on the corresponding matrix spaces.
Show that ||A B|| ||A|| ||B||.
2. Let p(t) be any polynomial in t. Show that > 0 m > 0 : |p(t)| met , t 0.
Problem 6.3 (Laplace Transform) Consider the time-invariant system:

(
x(t) = Ax(t) + Bu(t),
y(t) = Cx(t) + Du(t).
1. Let x(0) = 0. Make use of the Laplace transform to derive an input-output (ARMA) model
of the system of the following form: for ai R and Bi Rpm , i = 0, 1, . . . , n,
a0 y (n) (t) + a1 y (n1) (t) + . . . + an y(t) = B0 u(n) (t) + B1 u(n1) (t) + . . . + Bn u(t),
where y (i) (t) (resp. u(i) (t)) denotes the i-th derivative of y (resp. of u) at t. You may assume
that y (i) (0) = 0 and u(i) (0) = 0 for i = 0, . . . n.

A11 0 0
2. Let A = ,B= and C = C1 C2 , where, for 1 n2 n, A22 Rn2 n2 ,
A21 A22 B2
B2 Rn2 m , C2 Rpn2 and all other matrix blocks have consistent dimensions. Compute
the transfer function of the system and use the result to build a state-space representation of
order (i.e. size of the state evolution matrix) n2 having the same transfer function.

M11 0
(Hint: verify that the inverse of an invertible block triangular matrix M = , with
M21 M22
M11 0
M11 and M22 square, is also block triangular, i.e. M 1 = ; in particular, M22

M21 M22
must be equal to... .)
Problem 6.4 (Stability)
d

1. Consider a system x(t) = f (x(t)). Suppose that we have dt x(t)T P x(t) x(t)T Qx(t),
where P and Q are symmetric positive definite matrices (here the time-derivative is taken
along solutions of the system). Prove that under this condition the system is exponentially
stable, in the sense that its solutions satisfy kx(t)k cet kx(0)k for some c, > 0. Note
that this statement is true whether the system is linear or not. Further, if f (x) = Ax, then
show that the above condition is equivalent to AT P + P A Q.

2. Consider the system x(t) = Ax(t) + Bu(t) such that eAt cet for some c, > 0.
(a) Prove that if u is bounded over all time (in the sense that sup0t< ku(t)k M for some
M ), then x is also bounded, for any initial condition.
(b) Now restrict attention to the zero initial condition (x0 = 0). We can view the system above
as a linear operator from the normed linear space of bounded functions u : [0, [ Rm
with norm kuk := sup0t< ku(t)k, to the normed linear space of functions x : [0, [
Rn with the norm kxk := sup0t< kx(t)k. What can you say about the induced norm
of this operator, using the calculations you made in (a)? What can you say about its
continuity?
Problem 6.5 (Discrete-time Systems) Consider a discrete-time linear system x(k+1) = Ax(k)+
Bu(k), k = 0, 1, . . ..
1. Write the formula for the solution x(k) at time k starting from some initial state x(0) at time
0.
2. Assume A is semi-simple. Under what conditions on the eigenvalues of A is the discrete-time
system x(k + 1) = Ax(k) (no controls) asymptotically stable? Stable? Justify your answers.
(Stability definitions are the same as for continuous-time systems, just replace t by k.)
3. Lyapunovs second method for discrete-time system x(k + 1) = f (x(k)) involves the difference
V (x) := V (f (x)) V (x) instead of the derivative V (x); with this substitution, the statement
is the same as in the continuous-time case. Derive the counterpart of the Lyapunov equation
for the LTI discrete-time system x(k + 1) = Ax(k).
Problem 6.6 (Linearization example) Consider a two dimensional state vector x = (x1 , x2 )
R2 whose evolution is goverened by the following differential equations
x1 (t) = x2 (t)
x2 (t) = x1 (t) x1 (t)3 x2 (t).
1. Compute all equilibria of the system.

2. Compute the linearization of the system about its equilibria. The calculation is the same as
the one in Section 4.1, replacing the optimal trajectory, x (t), with the equilibrium solution
x (t) = x for all t, where x is each of the equilibria of the system.
3. Compute the eigenvalues of the matrices of the resulting linearizations and speculate about
the stability of the equilibria of the nonlinear system by invoking Theorem 6.2. Simulate the
system to confirm your intuition.
Problem 6.7 (When local is global) The equilibrium solution x = 0 of the linear time varying
system x(t) = A(t)x(t) is
1. Globally asymptotically stable if and only if it is locally asymptotically stable.

2. Globally exponentially stable if and only if it is locally exponentially stable.
Is it possible for the linear time varying system to have more than one equilibria? If yes, is it possible
for any of them to be locally asymptotically stable in such a case?
Problem 6.8 (When all is uniform) Show that the equilibrium x = 0 of the linear time invariant
system x(t) = Ax(t) is uniformly stable if and only if it is stable. Show further that this is the case
if and only if the following two conditions are met:
1. For all Spec[A], Re[] 0.

2. For all Spec[A] such that Re[] = 0, all Jordan blocks of have dimension 1.
Problem 6.9 (BIBO and BIBS stability) Consider the time varying linear system
x(t) = A(t)x(t) + B(t)u(t)

y(t) = C(t)x(t) + D(t)u(t).
with A(t), B(t), C(t), D(t) piecewise continuous. Assume that
1. The equilibrium solution is exponentially stable, i.e. there exists m, > 0 such that for all
(x0 , t0 ) Rn R+ and all t R+ , ks(t, t0 , x0 , U )k mkx0 ke(tt0 ) .
2. For all t0 R, kA()kt0 , , kB()kt0 , , kC()kt0 , , kD()kt0 , are bounded.
Show that:
1. For all (x0 , t0 ) Rn R and all u() : [t0 , ) Rm with ku()kt0 , bounded,
m
ks(, t0 , x0 , u)kt0 , mkx0 ket0 + kB()kt0 , ku()kt0 ,
hm i
k(, t0 , x0 , u)kt0 , mkC()kt0 , kx0 ket0 + kC()kt0 , kB()kt0 , + kD()kt0 , ku()kt0 , .

2. For all (x0 , t0 ) Rn R the functions
s(, t0 , x0 , ) : PC([t0 , ), Rm ) C([t0 , ), Rn )

{u() : [t0 , ) Rm } 7 {s(, t0 , x0 , u) : [t0 , ) Rn }
(, t0 , x0 , ) : PC([t0 , ), Rm ) PC([t0 , ), Rp )
{u() : [t0 , ) Rm } 7 {(, t0 , x0 , u) : [t0 , ) Rp }
are continuous.
Chapter 7
Inner product spaces
We return briefly to abstract vector spaces for a while to introduce the notion of inner products,
that will form the basis of our discussion on controllability.
7.1 Inner product

p
Consider a field F , either R or C. If F = C let |a| = a21 + a22 denote the magnitude and a = a1 ja2
denote the complex conjugate of a = a1 + ja2 F . If F = R let |a| denote the absolute value of
a F ; for completeness define a = a in this case.
Definition 7.1 Let (H, F ) be a linear space. A function h, i : H H F is called an inner product
if and only if for all x, y, z H, F ,
1. hx, y + zi = hx, yi + hx, zi.

2. hx, yi = hx, yi.
3. hx, xi is real and positive for all x 6= 0.
4. hx, yi = hy, xi (complex conjugate).
(H, F, h, i) is then called an inner product space.
Exercise 7.1 For all x, y, z H and all a F , hx + y, zi = hx, zi + hy, zi and hax, yi = ahx, yi.
Moreover, hx, 0i = h0, xi = 0 and hx, xi = 0 if and only if x = 0.
p
Fact 7.1 If (H, F, h, i) then the function k k : H F defined by kxk = hx, xi is a norm on
(H, F ).
Note that the function is well defined by property 3 of the inner product definition. The proof is
based on the following fact.
Theorem 7.1 (Schwarz inequality) With k k defined as in Fact 7.1, |hx, yi| kxk kyk.
Proof: If x = 0 the claim is obvious by Exercise 7.1. If x 6= 0, select F such that || = 1 and
hy,xi
hx, yi = |hx, yi|; i.e. if F = R take to be the sign of hx, yi and if F = C take = |hx,yi| . Then
99
for all R
kx + yk2 = hx + y, x + yi
= hx + y, xi + hx + y, yi
= hx + y, xi + hx + y, yi
= hx, x + yi + hy, x + yi
= hx, xi + hx, yi) + (hy, xi + hy, yi)
= 2 kxk2 + |hx, yi| + |hx, yi| + ||2 kyk2
= 2 kxk2 + 2|hx, yi| + kyk2 .
Since by definition hx + y, x + yi 0 we must have
2 kxk2 + 2|hx, yi| + kyk2 0.
This is a quadratic in that must be non-negative for all R. This will be the case if and only if
it is non-negative at its minimum point. Differentiating with respect to shows that the minimum
occurs when
|hx, yi|
=
kxk2
Substituting this back into the quadratic we see that
|hx, yi|2 2 |hx, yi|2

kxk 2 + kyk2 0 |hx, yi|2 kxk2 kyk2
kxk4 kxk2
Exercise 7.2 Prove Fact 7.1 using the Schwarz inequality.
Definition 7.2 Let (H, F, h, i) be an inner product space. The norm defined by kxk2 = hx, xi
for all x H is called the norm defined by the inner product. If the normed space (H, F, k k) is
complete (a Banach space) then (H, F, h, i) is called a Hilbert space.
To demonstrate the above definitions we consider two examples of inner products, one for finite
dimensional and one for infinite dimensional spaces.
Example (Finite dimensional inner product space) For F = R or F = C, consider the linear
space (F n , F ). Define the inner product h, i : F n F n F by
n
X
hx, yi = xi yi = xT y
i=1
for all x, y F n , where xT denotes complex conjugate (element-wise) transpose.
Exercise 7.3 Show that this satisfies the axioms of the inner product.
It is easy to see that the norm defined by this inner product

n
X
kxk2 = hx, xi = |xi |2
i=1
is the Euclidean norm. We have already seen that (F n , F, k k) is a complete normed space, hence
(F n , F, h, i) is a Hilbert space.
Example (Infinite dimensional inner product space) For F = R or F = C, t0 , t1 R with

t0 t1 consider the linear space L2 ([t0 , t1 ], F n ) of square integrable functions f () : [t0 , t1 ] F n ,
i.e. functions such that Z t1
kf (t)k22 dt < .
t0
On this space we can define the L2 inner product

Z t1
T
hf, gi = f (t) g(t)dt
t0
T
where as before f (t) denotes complex conjugate transposed of f (t) F n .
Exercise 7.4 Show that this satisfies the axioms of the inner product.
The norm defined by this inner product is the 2norm

Z t1 21
2
kf ()k2 = kf (t)k2 dt .
t0
As before, to make this norm well defined we need to identify functions that are zero almost every-
where (i.e. are zero over a dense set) with the zero function.
7.2 Orthogonal complement

Definition 7.3 Let (H, F, h, i) be an inner product space. x, y H are called orthogonal if and
only if hx, yi = 0.
Example (Orthogonal vectors) Consider the inner product space (Rn , R, h, i) with the inner
project hx, yi = xT y defined above. Given two non-zero vectors x, y Rn one can define the angle,
, between them by
hx, yi
= cos1
kxk kyk
x, y are orthogonal if and only if = (2k + 1)/2 for k Z.
Exercise 7.5 Show that is well defined by the Schwarz inequality). Show further that =
(2k + 1)/2 if and only if x and y are orthogonal.
This interpretation of the angle between two vectors motivates the following generalisation of a
well known theorem in geometry.
Theorem 7.2 (Pythagoras theorem) Let (H, F, h, i) be an inner product space. If x, y H are
orthogonal then kx + yk2 = kxk2 + kyk2 , where k k is the norm defined by the inner product.
Proof: Exercise.
Note that H does not need to be R3 (or even finite dimensional) for the Pythagoras theorem to hold.
Given a subspace M H of a linear space, consider now the set of all vectors y H which are
orthogonal to all vectors x M .
Definition 7.4 The orthogonal complement of a subspace, M , of an inner product space (H, F, h, i)
is the set
M = {y H | hx, yi = 0 x M }.
Exercise 7.6 Show that M is a subspace of H.
Consider now an inner product space (H, F, h, i) with the norm kyk2 = hy, yi defined by the
inner product. Recall that a sequence {yi } i=0 is said to converge to a point y H if an only if
limi ky yi k = 0. In this case y is called the limit point of the sequence {yi }
i=0 . Recall also that
a subset K H is called closed (Definition 3.3) if and only if it contains the limit points of all the
sequences {yi }
i=0 K.
Fact 7.2 Let M be a subspace of the inner product space (H, F, h, i). M is a closed subspace of
H and M M = {0}.
Proof: To show that M is a subspace consider y1 , y2 M , a1 , a2 F and show that a1 y1 +a2 y2

M . Indeed, for all x M ,
hx, a1 y1 + a2 y2 i = a1 hx, y1 i + a2 hx, y2 i = 0.
To show that M is closed, consider a sequence {yi }

i=0 M

and assume that it converges to
some y H. We need to show that y M . Note that, since yi M , by definition hx, yi i = 0 for

all x M . Consider an arbitrary x M and note that
hx, yi = hx, yi + y yi i
= hx, yi i + hx, y yi i
= hx, y yi i.
By the Schwarz inequality
0 |hx, yi| = |hx, y yi i| kxk ky yi k.
Since limi ky yi k = 0 we must have that |hx, yi| = 0. Hence y M .

Finally, consider y M M . Since y M , hx, yi = 0 for all x M . Since for y itself we have
that y M ,
hy, yi = kyk2 = 0.
By the axioms of the norm this is equivalent to y = 0.
Definition 7.5 Let M, N be subspaces of a linear space (H, F ). The sum of M and N is the set
M + N = {w | u M, v N such that w = u + v}.
If in addition M N = {0} then M + N is called the direct sum of M and N and is denoted by
M N.
Exercise 7.7 M + N is a subspace of H.
Fact 7.3 V = M N if and only if for all x V there exists unique u M and v N such that
x = u + v.
Proof: (exercise) (). Let V = M N . Then M N = {0} and for all x V there exist u M
and v N such that x = u + v. It remains to show that the u and v are unique. Assume, for the
sake of contradiction, that they are not. Then there exist u M , v N with (u , v ) 6= (u, v) such
that x = u + v . Then
u + v = u + v u u = v v .
Moreover, u u M and v v N (since M, N are subspaces). Hence M N {u u, 0} =
6 {0}
which is a contradiction.
(). Assume that for all x V there exist unique u M and v N such that x = u + v.
Then V = M + N by definition. We need to show that M N = {0}. Assume, for the sake of
contradiction that this is not the case, i.e. there exists y 6= 0 such that y M N . Consider and
arbitrary x V and the unique u M and v N such that x = u + v. Define u = u + y and
v = v y. Note that u M and v N since M and N are subspaces and y M N . Moreover,
u + v = u + y + v y = u + v = x but u 6= u and v 6= v since y 6= 0. This contradicts the
uniqueness of u and v.
Theorem 7.3 Let M be a closed subspace of a Hilbert space (H, F, h, i). Then:

1. H = M M (denoted by M M ).
2. For all x H there exists a unique y M such that x y M . This y is called the
orthogonal projection of x onto M .
3. For all x H the orthogonal projection y M is the unique element of M that achieves the
minimum
kx yk = inf{kx uk | u M }.
7.3 Adjoint of a linear map

Definition 7.6 Let (U, F, h, iU ) and (V, F, h, iV ) be Hilbert spaces and A : U V a continuous
linear map. The adjoint of A is the map A : V U defined by
hv, AuiV = hA v, uiU
for all u U , v V .
Theorem 7.4 Let (U, F, h, iU ), (V, F, h, iV ) and be (W, F, h, iW ) Hilbert spaces, A : U V ,

B : U V and C : W U be continuous linear maps and a F . The following hold:
1. A is well defined, linear and continuous.

2. (A + B) = A + B .
3. (aA) = aA .
4. (A C) = C A
5. If A is invertible then (A1 ) = (A )1 .
6. kA k = kAk where k k denotes the induced norms defined by the inner products.
7. (A ) = A.
Example (Finite dimensional adjoint) Let U = F n , V = F m , and A = [aij ] F mn and define

a linear map A : U V by A(x) = Ax for all x U . Then for all x U , y V
m
X m
X n
X n X
X m
hy, A(x)iF m = yT Ax = yi (Ax)i = yi aij xj = aij yi xj
i=1 i=1 j=1 j=1 i=1
n
X T T T
= (A y)j xj = Ay x = hA y, xiF n .
j=1
Therefore the adjoint of the linear map defined by the matrix A F mn is the linear map defined by
T
the matrix A = [aji ] F nm , the complex conjugate transpose (known as the Hermitian transpose)
of A. If in addition F = R, then A = AT = [aji ] F nm is simply the transpose of A.
As we have seen in Chapter 2, any linear map between finite dimensional vector spaces can be
represented by a matrix, by fixing bases for the domain and co-domain spaces. The above example
therefore demonstrates that when it comes to linear maps between finite dimensional spaces the
adjoint operation always involves taking the complex conjugate transpose of a matrix. Recall also
that linear maps between finite dimensional spaces are always continuous (Theorem 3.5). For infinite
dimensional spaces the situation is in general more complicated. In some cases, however, the adjoint
may be simple to represent.
Example (Infinite dimensional adjoint) Let U = (L2 ([t0 , t1 ], F m ), F, h, i2 denote the space of
square integrable functions u : [t0 , t1 ] F m and V = (F n , F, h, iF n . Consider G() L2 ([t0 , t1 ], F nm )
(think for example of G(t) = (t1 , t)B(t) for t [t0 , t1 ] for a linear system). Define a function
A : U V by Z t1
A(u()) = G(t)u(t)dt.
t0
It is easy to see that A is well defined and linear. Assume also that G(t) is well behaved enoug (e.g.
it is bounded) so that that A is also continuous.
For arbitrary x V , u() U
Z t1
T
hx, A(u())iF m = x G(t)u(t)dt
t0
Z t1
= xT G(t)u(t)dt
t0
Z t1 T
T
= G(t) x u(t)dt
t0
T
= hG x, ui2 .
Therefore the adjoint A : F n L2 ([t0 , t1 ], F m ) of the linear map A is given by

T
(A (x))(t) = G(t) x for all t [t0 , t1 ]
T
where as before G(t) F mn denotes the Hermitian transpose of G(t) F nm .
Definition 7.7 Let (H, F, h, i) be a Hilbert space and A : H H be linear and continuous. A is
called self-adjoint if and only if A = A, in other words for all x, y H
hx, A(y)i = hA(x), yi.

Example (Finite dimensional self-adjoint map) Let H = F n with the standard inner product
and A = [aij ] F nn . The linear map A : F n F n defined by A(x) = Ax is self adjoint if and
T
only if A = A, or in other words aij = aji for all i, j = 1, . . . n. Such matrices are called Hermitian.
If in addition F = R, A is self-adjoint if and only if aij = aji , i.e. AT = A. Such matrices are called
symmetric.
Example (Infinite dimensional self-adjoint map) Let H = L2 ([t0 , t1 ], R) and K(, ) : [t0 , t1 ]
[t0 , t1 ] R such that
Z t1 Z t1
|K(t, )|2 dtd <
t0 t0
(think for example of (t, )B( ).) Define A : L2 ([t0 , t1 ], R) L2 ([t0 , t1 ], R) by

Z t1
(A(u()))(t) = K(t, )u( )d for all t [t0 , t1 ].
t0
Exercise 7.8 Show that A is linear, continuous and self-adjoint.
Let (H, F ) be a linear space and A : H H a linear map. Recall that (Definition 2.14) an element
F is called an eigenvalue of A if and only if there exists v H with v 6= 0 such that A(v) = v;
in this case v is called an eigenvector corresponding to .
Fact 7.4 Let (H, F, h, i) be a Hilbert space and A : H H be linear, continuous and self-adjoint.
Then
1. All the eigenvalues of A are real.

2. If i and j are eigenvalues with corresponding eigenvectors vi , vj H and i 6= j then vi is
orthogonal to vj .
Proof: Part 1: Let be an eigenvalue and v H the corresponding eigenvector. Then
hv, A(v)i = hv, vi = hv, vi = kvk2
Since A = A , however, we also have that
hv, A(v)i = hA(v), vi = hv, A(v)i = hv, vi = kvk2 .
Therefore = and is real.

Part 2. By definition

A(vi ) = i vi hvj , A(vi )i = i hvj , vi i
A(vj ) = j vj hA(vj ), vi i = j hvj , vi i (i j )hvj , vi i = 0.

A = A hvj , A(vi )i = hA(vj ), vi i
Therefore if i 6= j we must have hvj , vi i = 0 and vi and vj are orthogonal.

7.4 Finite rank lemma

Let U, V be linear spaces and consider a linear map A : U V . Recall that the range space,
Range(A), and null space, Null(A), of A defined by
Range(A) = {v V | u U, v = A(u)}
Null(A) = {u U | A(u) = 0}
are linear subspaces of V and U respectively.
Theorem 7.5 (Finite Rank Lemma) Let F = R or F = C, let (H, F, h, i) be a Hilbert space and
recall that (F m , F, h, iF m ) is a finite dimensional Hilbert space. Let A : H F m be a continuous
linear map and A : F m H be its adjoint. Then:
1. A A : F m F m and A A : H H are linear, continuous and self adjoint.

2. H = Range(A ) Null(A), i.e. Range(A )Null(A) = {0}, Range(A ) = (Null(A))

and H = Range(A) Null(A). Likewise, F m = Range(A) Null(A ).
3. The restriction of the linear map A to the range space of A ,
A|Range(A ) : Range(A ) F m
is a bijection between Range(A ) and Range(A).

4. Null(A A ) = Null(A ) and Range(A A ) = Range(A).
5. The restriction of the linear map A to the range space of A,
A |Range(A) : Range(A) H
is a bijection between Range(A) and Range(A ).

6. Null(A A) = Null(A) and Range(A A) = Range(A ).
Proof: Part 1: A is linear and continuous and hence A A and A A are linear, continuous and
self-adjoint by Theorem 7.4.
Part 2: Recall that Range(A) F m and therefore is finite dimensional. Moreover, Dim[Range(A )]
Dim[F m ] = m therefore Range(A ) is also finite dimensional. Therefore both Range(A) and
Range(A ) are closed (by Fact ??), hence by Theorem 7.3
H = Range(A ) Range(A ) and F m = Range(A) Range(A) .
But
x Range(A) hx, viF m = 0 v Range(A)

hx, A(y)iF m = 0 y H
hA (x), yiH = 0 y H
A (x) = 0 (e.g. take {yi } a basis for H)
x Null(A ).

Therefore, Null(A ) = Range(A) therefore F m = Range(A) Null(A ). The proof for H is
similar.
Part 3: Consider the restriction of the linear map A to the range space of A
A|Range(A ) : Range(A ) F m .
Clearly for all x Range(A ), A|Range(A ) (x) = A(x) Range(A) therefore
A|Range(A ) : Range(A ) Range(A).
We need to show that this map is injective and surjective.

To show that it is surjective, note that for all y Range(A) there exists x H such that y = A(x).

But since H = Range(A ) Null(A), x = x1 + x2 for some x1 Range(A ), x2 Null(A).
Then
y = A(x1 + x2 ) = A(x1 ) + A(x2 ) = A(x1 ) + 0 = A(x1 ).
Hence, for all y Range(A) there exists x1 Range(A ) such that y = A(x1 ) and the map is
surjective.
To show that the map is injective, recall that this is the case if and only if Null( A|Range(A ) ) =
{0}. Consider an arbitrary y Null( A|Range(A ) ). Then y Range(A ) and there exists
x F m such that y = A (x) and moreover A(y) = 0. Therefore y Null(A) Range(A ) = {0}

since H = Range(A ) Null(A) and the map is injective.
Part 4: To show that Null(A A ) = Null(A ) consider first an arbitrary x Null(A A ).
Then
A A (x) = 0 hx, A A (x)iF m = 0

hA (x), A (x)iH = 0
kA (x)k2H = 0
A (x) = 0
x Null(A ).
Hence Null(A A ) Null(A ). Conversely, consider an arbitrary x Null(A ). Then
A x = 0 A A (x) = 0 x Null(A A )
and hence Null(A ) Null(A A ). Overall, Null(A A ) = Null(A ).

Finally, to show that Range(A) = Range(A A ) note that
Range(A) = {y F m | x H, y = A(x)}
= {y F m | x Range(A ), y = A(x)} (by Part 3)
= {y F m | u F m , y = A(A (u))}
= {y F m | u F m , y = (A A )(u)}
= Range(A A ).
The proofs of Part 5 and Part 6 are analogous to those of Part 3 and Part 4 respectively.
Notice that the assumption that the co-domain of A is finite dimensional is only used to establish
that the ranges of A and A are closed sets, to be able to invoke Theorem 7.3. In fact, Theorem 7.5
holds more generally if we replace the spaces by their closures.
7.5 Application: Matrix pseudo-inverse

Problem 7.1 (An Example of a Hilbert Space) Let E = L2 ([t0 , t1 ], R) be the Hilbert space of
square integrable real-valued functions on [t0 , t1 ]. Let K : [t0 , t1 ] [t0 , t1 ] R be such that
Z t1 Z t1
2
|K(t, )| dt d < . ()
t0 t0
Define A : L2 ([t0 , t1 ], R) L2 ([t0 , t1 ], R) by

Z t1

A(u) (t) = K(t, )u( )d t [t0 , t1 ].
t0
Prove that A is linear and continuous. Also prove that A is self-adjoint if K is symmetric.
Chapter 8
Controllability and observability
Let us return now to time varying linear systems
x(t) = A(t)x(t) + B(t)u(t)

y(t) = C(t)x(t) + D(t)u(t).
We will investigate the following questions
Can the input u be used to steer the state of the system to an arbitrary value?
Can we infer the value of the state by observing the input and output?
We will again start by a general discussion of these questions for non-linear systems. We will then
specialize to the case of time varying linear systems and then further to the case of time invariant
linear systems.
8.1 Nonlinear systems

Consider the nonlinear system
x(t) = f (x(t), u(t), t) (8.1)

y(t) = r(x(t), u(t), t) (8.2)
where t R+ , x(t) Rn , u(t) Rm , y(t) Rp , f : Rn Rm R+ Rn and r : Rn Rm R+ Rp

and assume that r is continuous, that f is Lipschitz in x and continuous in u and t, and that
u : R+ Rm is piecewise continuous. Under these conditions there exist continuous solution maps
such that for (x0 , t0 ) Rn R+ , t1 t0 , u() : [t0 , t1 ] Rm and t [t0 , t1 ] return the solution
x(t) = s(t, t0 , x0 , u), y(t) = (t, t0 , x0 , u)
of system (8.1)(8.2). For simplicity, assume further that r is continuous in x and piecewise contin-
uous in t.
Definition 8.1 Consider t0 , t1 R with t0 t1 . The input trajectory u() PC([t0 , t1 ], Rm ) steers
(x0 , t0 ) Rn R to (x1 , t1 ) Rn R if and only if s(t1 , t0 , x0 , u) = x1 . The system (8.1)(8.2) is
controllable on [t0 , t1 ] if and only if for all x0 , x1 Rn there exists u() PC([t0 , t1 ], Rm ) that steers
(x0 , t0 ) to (x1 , t1 ).
109
v(t) u(t) y(t)

+ x(t) R x(t) r
f
FS
Figure 8.1: State feedback.
Notice that controllability has nothing to do with the output of the system, it is purely an input
to state relationship; we will therefore talk about controllability of the system (8.1) ignoring
Equation (8.2) for the time being. The following fact is an immediate consequence of the definition.
Fact 8.1 The system (8.1)(8.2) is controllable on [t0 , t1 ] if and only if for all x0 Rn the function
s(t1 , to , x0 , ) : PC([t0 , t1 ], Rm ) Rn is surjective.
It is easy to see that controllability is preserved under state and output feedback. Consider a state
feedback map FS : Rn R Rm and a new input variable v(t) Rm . Let u(t) = v(t) + FS (x(t), t)
and define the state feedback system
x(t) = f (x(t), v(t) + FS (x(t), t), t) = fS (x(t), v(t), t) (8.3)
To prevent technical difficulties assume that FS is continuous in x and piecewise continuous in t.

One can picture the action of the feedback map FS as shown in Figure 8.1.
Likewise, consider a output feedback map FO : Rp R Rm and assume that r does not explicitly
depend on u, i.e. y(t) = r(x(t), t). Consider a new input variable w(t) Rm , let u(t) = w(t) +
FO (y(t), t) and define the output feedback system
x(t) = f (x(t), w(t) + FO (y(t), t), t) = f (x(t), w(t) + FO (r(x(t), t), t), t) = fO (x(t), w(t), t) (8.4)
To prevent technical difficulties assume that FO is continuous in y and piecewise continuous in t.

One can picture the action of the feedback map FO as shown in Figure 8.2.
Theorem 8.1 Consider to , t1 R with t0 t0 . The following statements are equivalent:
1. The system of Equation (8.1) is controllable on [t0 , t1 ].

Proof: The proof requires only careful application of the definitions. For example, to show that if
the system of Equation (8.1) is controllable on [t0 , t1 ] then so is the system of Equation (8.3), consider
arbitrary x0 , x1 Rn and look for v() PC([t0 , t1 ], Rm ) that steers (x0 , t0 ) to (x1 , t1 ). Since (8.1) is
controllable on [t0 , t1 ] there exists u() PC([t0 , t1 ], Rm ) such that the unique solution, s(t, t0 , x0 , u),
of (8.1) satisfies s(t1 , t0 , x0 , u) = x1 . Define v(t) = u(t) FS (s(t, t0 , x0 , u), t); clearly v() : [t0 , t1 ]
Rm is piecewise continuous since u is piecewise continuous, FS is assumed to be continuous in its
replacemen
w(t) u(t)
+ x(t) R x(t) y(t)
f r
FO
Figure 8.2: Output feedback.
first argument and piecewise continuous in the second, and, by definition, s(, t0 , x0 , u) is continuous.
Moreover, under v() to system (8.3) evolves according to
x(t) = f (x(t), v(t) + FS (x(t), t), t) = f (x(t), u(t) FS (s(t, t0 , x0 , u), t) + FS (x(t), t), t)
starting at x(t0 ) = x0 . Clearly x(t) = s(t, t0 , x0 , u) satisfies both the initial condition and the
differential equation and therefore is also the unique solution for system (8.3) under the input v().
Hence, the proposed input v() steers (x0 , t0 ) to (x1 , t1 ).
The remaining parts are similar and are left as an exercise.
Dual to the definition of controllability is that of observability.
Definition 8.2 Consider t0 , t1 R with t0 t1 . The system (8.1)(8.2) is observable on [t0 , t1 ]

if and only if for all x0 Rn and all u() PC([t0 , t1 ], Rm ), given u() : [t0 , t1 ] Rm and the
corresponding output y() = (, t0 , x0 , u) : [t0 , t1 ] Rp the value of x0 can be uniquely determined.
A slight rephrasing of the definition leads to the following fact.
Fact 8.2 The system (8.1)(8.2) is observable on [t0 , t1 ] if and only if for all u() PC([t0 , t1 ], Rm )
the function
(, t0 , , u) : Rn PC([t0 , t1 ], Rp )
x0 7 (, t0 , x0 , u) : [t0 , t1 ] Rp
is injective.
It is easy to see that if we establish the value of x0 Rn we can in fact reconstruct the value of x(t)
for all t [t0 , t1 ]; this is because, by uniqueness, x(t) = s(t, t0 , x0 , u) is uniquely determined once we
know t, t0 , x0 and u : [t0 , t] Rm .
It is easy to see that observability is preserved under output feedback and input feed-forward.
Consider a input feed-forward map FF : Rm R Rp and a new output variable z(t) Rp . Let
z(t) = y(t) + FF (u(t), t) and define the input feed-forward system with
z(t) = r(x(t), u(t), t) + FF (u(t), t) = rF (x(t), u(t), t). (8.5)
To prevent technical difficulties assume that FF is continuous in u and piecewise continuous in t.

One can picture the action of the feed-forward map FF as shown in Figure 8.3.
1. The system (8.1)(8.2) is observable.

u(t) y(t) z(t)

x(t) R x(t) r +
f
FF
Figure 8.3: Input feed-forward.

Proof: Exercise.
Notice that observability is not preserved by state feedback. If system (8.1)(8.2) is observable,
system (8.3)(8.2) may or may not be observable and vice versa.
Exercise 8.1 Provide a simple example of a linear system whose observability properties are altered
by state feedback.
In nonlinear control, feedback linearization provides a dramatic demonstration of the fact that
state feedback can alter observability. In feedback linearization state feedback and a coordinate
transformation is used to alter the dynamics of some of the states, so that system behaves like a
linear system. The price to pay for this is that the remaining states (the so-called zero dynamics)
are disconnected from the output, i.e. are rendered unobservable; the reader is referred to [?, ?, ?]
for more information on this topic.
8.2 Linear time varying systems: Controllability

Let us now see how the additional structure afforded by linearity can be used to derive precise
conditions for controllability and observability. For most of the discussion we will need to assume
that the input trajectories of the system take values in an appropriate Hilbert space, to be able to
apply the Finite Rank Lemma and perform projections. Since the space of piecewise continuous
input trajectories used up to now (e.g. for the existence and uniqueness arguments) is not complete
we will assume that input trajectories take values in the somewhat larger Hilbert space of square
integrable functions. More specifically, given t0 , t1 R with t0 t1 we will consider input trajectories
u() : [t0 , t1 ] Rm and assume that u (L2 ([t0 , t1 ], Rm ), h, i2 ) with the inner product defined by
Z t1
hu, ui2 = u(t)T u(t)dt.
t0
2 m
for u, u L ([t0 , t1 ], R ). Note that by considering this extended space we can no longer assume
that the input trajectories will be piecewise continuous, and hence are no longer covered by the
existence uniqueness argument of Theorem 3.7. We will alleviate this difficulty, however, by showing
that the properties of interest here (e.g. steering the system from a given initial state to a desired
final state) can indeed be achieved by piecewise continuous trajectories.
Since controllability is an input to state property only the matrices A() and B() come into play.
To simplify the notation we give the following definition.
Definition 8.3 The pair (A(), B()) is controllable on [t0 , t1 ] if and only if for all x0 , x1 Rn there
exists u : [t0 , t1 ] Rm that steers (x0 , t0 ) to (x1 , t1 ), i.e.
Z t1
x1 = (t1 , t0 )x0 + (t1 , )B( )u( )d.
t0
1. (A(), B()) is controllable on [t0 , t1 ].

2. For all x0 Rn there exists u : [t0 , t1 ] Rm that steers (x0 , t0 ) to (0, t1 ).
3. For all x1 Rn there exists u : [t0 , t1 ] Rm that steers (0, t0 ) to (x1 , t1 ).
Proof: (Sketch). 1 2 and 1 3: Obvious.

2 1: For arbitrary x0 , x1 Rn let x0 = x0 (t0 , t1 )x1 and consider the input u : [t0 , t1 ] Rm
that steers (x0 , t0 ) to (0, t1 ). It is easy to see that the same input steers (x0 , t0 ) to (x1 , t1 ).
3 1: Exercise.
Exercise 8.2 The three statements of Theorem 8.3 are not equivalent for nonlinear systems. Where
does linearity come into play in the proof?
Theorem 8.3 states that for linear systems controllability is equivalent to controllability to zero and
to reachability from zero. In the subsequent discussion will use the last of the three equivalent
statements to simplify the analysis; Theorem 8.3 implies that this can be done without loss of
generality. In line with this, we introduce the following definition.
Definition 8.4 A state x1 Rn is reachable on [t0 , t1 ] by the pair (A(), B()) if and only if there
exists u() L2 ([t0 , t1 ], Rm ) that steers (0, t0 ) to (x1 , t1 ). The reachability map on [t0 , t1 ] of the pair
(A(), B()) is the function
Lr : L2 ([t0 , t1 ], Rm ) Rn
R t1
u 7 t0 (t1 , )B( )u( )d.
Fact 8.3 Lr is linear and continuous. The set of reachable states is the linear subspace Range(Lr ).
Proof: The fact that Lr is linear is obvious from the properties of the integral. The fact that
reachable states are equal to Range(Lr ) is immediate from the definition of reachable states. To
show that Lr is continuous let kk denote the 2-norm in Rn and note that for all u() L2 ([t0 , t1 ], Rm )
Z t1 2

2
kLr (u)k = (t1 , )B( )u( )d

t0
Z t1
2
k(t1 , )B( )k ku( )k2 d.
t0
Since (t1 , )B() is a piecewise continuous function on [t0 , t1 ], there exists some M > 0 such that
k(t1 , )B( )k M for all [t0 , t1 ]. Hence
Z t1
kLr (u)k2 M 2 ku( )k2 d = M 2 ku()k22
t0
i.e. the induced norm of Lr is finite and the function is continuous.

Definition 8.5 The controllability gramian of the pair (A(), B()) on [t0 , t1 ] is the matrix
Z t1
Wr (t0 , t1 ) = (t1 , )B( )B( )T (t1 , )T d Rnn .
t0
The controllability gramian will be our primary vehicle for studying the controllability of time
varying linear systems. We start by establishing some of its basic properties.
Fact 8.4 The controllability gramian, Wr (t0 , t1 ), has the following properties:
1. It is symmetric.
2. It is positive semi-definite.
3. For all t0 t0 , Wr (t0 , t1 ) Wr (t0 , t1 ) in the sense that xT [Wr (t0 , t1 ) Wr (t0 , t1 )]x 0 for
all x Rn .
Proof: Part 1 is straightforward, since

Z t1 T
T T T
Wr (t0 , t1 ) = (t1 , )B( )B( ) (t1 , ) d
t0
Z t1 T
= (t1 , )B( )B( )T (t1 , )T d
t0
Z t1
T T
= (t1 , )T B( )T B( )T (t1 , )T d
t0
= Wr (t0 , t1 ).
For Part 2 note that for all x Rn

Z t1 Z t1
xT Wr (t0 , t1 )x = xT (t1 , )B( )B( )T (t1 , )T xd = B( )T (t1 , )T x 2 d 0.
t0 t0
Finally, for Part 3, note that

Z t0
xT [Wr (t0 , t1 ) Wr (t0 , t1 )]x = xT (t1 , )B( )B( )T (t1 , )T xd 0
t0
(as shown in Part 2).
The following theorem allows us to answer controllability questions by simply checking the rank of
the controllability gramian.
1. (A(), B()) is controllable on [t0 , t1 ].

2. Range(Lr ) = Rn .
3. Range(Lr Lr ) = Rn .
4. Det[Wr (t0 , t1 )] 6= 0.
Proof: 1 2: By Theorem 8.3, (A(), B()) is controllable on [t0 , t1 ] if and only if all states are
reachable from 0 on [t0 , t1 ], i.e. if and only if Lr is surjective, i.e. if and only if Range(Lr ) = Rn .
2 3: By the finite rank lemma.
3 4: Notice that Lr Lr : Rn Rn is a linear map between two finite dimensional spaces.
Therefore it admits a matrix representation. We will show that Wr (t0 , t1 ) Rnn is the matrix
representation of this linear map. Then Lr Lr is surjective if and only if Wr (t0 , t1 ) is invertible i.e.
if and only if Det[Wr (t0 , t1 )] 6= 0.
Recall that Z t1
Lr : u 7 (t1 , )B( )u( )d
t0
and hLr x, ui = hx, Lr ui. We have already seen that (Section 7.3)
[Lr (x)]( ) = B( )T (t1 , )T x, for all [t0 , t1 ], x Rn
Therefore Z
t1
[Lr Lr ](x) = (t1 , )B( )B( )T (t1 , )T d x = Wr (t0 , t1 )x.
t0
Therefore the matrix representation of Lr Lr : Rn Rn is the matrix Wr (t0 , t1 ) Rnn .
Exercise 8.3 Show that (A(), B()) is controllable on [t0 , t1 ] if an only if Wr (t0 , t1 ) is positive
definite (i.e. xT Wr (t0 , t1 )x > 0 for all x Rn with x 6= 0).
8.3 Linear time varying systems: Minimum energy control

We have just seen that for all x0 , x1 Rn there exists u : [t0 , t1 ] Rm that steers (x0 , t0 ) to (x1 , t1 )
if and only if the matrix Wr (t0 , t1 ) Rnn is invertible. The question now becomes can we compute
such a u? And if so, can we find a piecewise continuous one, as required for existence and uniqueness
of solutions? Or even a continuous one? This would be nice in practice (most actuators do not like
applying discontinuous inputs) and in theory (the state trajectory will not only exist but would
generally also be differentiable).
First of all we note that if such an input exists it will not be unique. To start with let us restrict
our attention to the case where x0 = 0 (reachability); we have already seen this is equivalent to the
general controllability case (x0 6= 0) to which we will return shortly. Recall that by the Finite Rank
Lemma

L2 ([t0 , t1 ], Rm ) = Range(Lr ) Null(Lr ).
Recall also that L2 ([t0 , t1 ], Rm ) is infinite dimensional while Range(Lr ) is finite dimensional (of
dimension at most n), since it is a subspace of Rn which is finite dimensional. Therefore Null(Lr ) 6=
{0}, in fact it will be an infinite dimensional subspace of L2 ([t0 , t1 ], Rm ).
For a given x1 Rn consider u L2 ([t0 , t1 ], Rm ) that steers (0, t0 ) to (x1 , t1 ), i.e.
Lr (u) = x1 .
Then for every u L2 ([t0 , t1 ], Rm ) with u Null(Lr ) we have
Lr (u + u) = Lr (u) + Lr (u) = x1 + 0 = x1 .
Therefore for any input u that steers (0, t0 ) to (x1 , t1 ) any other input of the form u + u with
u Null(Lr ) will do the same. So in general there will be an infinite number of inputs that get
the job done. The question now becomes can we somehow find a canonical one? It turns out that
this can be done by a projection onto Range(Lr ).
L2 ([t0 , t1 ], Rp ) Rn
Range(Lr ) Lr Range(Lr )
u
x1 Lr Lr
u
Lr
Null(Lr ) = {0}
u
Null(Lr )
Figure 8.4: Linear space decomposition induced by reachability map.
Recall that by the finite rank lemma
Lr |Range(L ) : Range(Lr ) Range(Lr )

r
is a bijection. Therefore, for all x1 Range(Lr ) there exists a unique u Range(Lr ) such that
Lr (u) = x1 . We have seen that for any u Null(Lr ) if we let u = u + u then Lr (u) = x1 . However,
kuk22 = hu, ui2 = hu + u, u + ui2

= hu, ui2 + hu, ui2 + hu, ui2 + hu, ui2 .
But u Range(Lr ), u Null(Lr ) and by the finite rank lemma Range(Lr ) is orthogonal to
Null(Lr ). Therefore hu, ui2 = hu, ui2 = 0 and
kuk22 = kuk22 + kuk22 kuk22
since kuk22 0. Therefore, among all the u L2 ([t0 , t1 ], Rm ) that steer (0, t0 ) to (x1 , t1 ), the unique
u Range(Lr ) is the one with the minimum 2norm. The relation of the different spaces involved
is depicted in Figure 8.4.
We now return to the general controllability case and formally state our findings.
Theorem 8.5 Assume (A(), B()) are controllable in [t0 , t1 ]. Given x0 , x1 Rn define u : [t0 , t1 ]
Rm by u = Lr (Lr Lr )1 [x1 (t1 , t0 )x0 ], i.e.
u(t) = B(t)T (t1 , t)T Wr (t0 , t1 )1 [x1 (t1 , t0 )x0 ] t [t0 , t1 ].
Then
1. u steers (x0 , t0 ) to (x1 , t1 ).

2. u() : [t0 , t1 ] Rm is piecewise-continuous in general and continuous if B() : [t0 , t1 ] Rnm
is continuous.
T
3. kuk22 = [x1 (t1 , t0 )x0 ] Wr (t0 , t1 )1 [x1 (t1 , t0 )x0 ].
4. If u : [t0 , t1 ] Rm steers (x0 , t0 ) to (x1 , t1 ) then kuk2 kuk2 .
Proof:
Part 1:
Z t1
x(t1 ) = (t1 , t0 )x0 + (t1 , t)B(t)u(t)dt
t0
Z t1
= (t1 , t0 )x0 + (t1 , t)B(t)B(t)T (t1 , t)T Wr (t0 , t1 )1 [x1 (t1 , t0 )x0 ] dt
t0
Z t1
= (t1 , t0 )x0 + (t1 , t)B(t)B(t)T (t1 , t)T dt Wr (t0 , t1 )1 [x1 (t1 , t0 )x0 ]
t0
= (t1 , t0 )x0 + Wr (t0 , t1 )Wr (t0 , t1 )1 [x1 (t1 , t0 )x0 ]
= x1 .
Part 2 is obvious from the formula, since B(t) is piecewise continuous, (t1 , t) is continuous and
all other quantities are constant. In fact the only discontinuity points are those of B(t). Hence if
B(t) is continuous (e.g. for time invariant linear systems) then the minimum energy controls are
continuous.
Part 3:
Z t1
kuk22 = u(t)T u(t)dt
t0
Z t1
= [x1 (t1 , t0 )x0 ]T (Wr (t0 , t1 )1 )T (t1 , t)B(t)B(t)T (t1 , t)T dt
t0
Wr (t0 , t1 )1 [x1 (t1 , t0 )x0 ]
T
= [x1 (t1 , t0 )x0 ] (Wr (t0 , t1 )T )1 Wr (t0 , t1 )Wr (t0 , t1 )1 [x1 (t1 , t0 )x0 ]
= [x1 (t1 , t0 )x0 ]T Wr (t0 , t1 )1 [x1 (t1 , t0 )x0 ]
since Wr (t0 , t1 ) is self-adjoint by the finite rank lemma.

Part 4: Notice that by its definition u Range(Lr ). The claim follows by the discussion leading
up to the statement of the theorem.
The perceptive reader will have noticed that this calculation is effectively the same as the one carried
out in Section 7.5 to compute the pseudo-inverse of a matrix. The only difference is that here one
of the two spaces is infinite dimensional. Compare also Figure 8.4 with Figure ??.
The theorem provides several interesting insights into the structure of these so-called minimum
energy controls. For one, note that even though for mathematical reasons we were forced to allow
more general control inputs u L2 ([t0 , t1 ], Rm ) to apply the Finite Rank Lemma, the minimum
energy controls (the unique u Range(Lr ) that steers (x0 , t0 ) to (x1 , t1 )) happens to be piecewise
continuous. In other words, if there is a control in u L2 ([t0 , t1 ], Rm ) that will get the job done
there is also one in PC([t0 , t1 ], Rm ) (in fact, a better one!) that will do the same. This is good
to know, since we do not have to worry about existence of solutions which we only established for
piecewise continuous input functions in Chapter 3.
The structure of the minimum energy itself, kuk22 , is also revealing. kuk22 is a quadratic function of
x1 (t1 , t0 )x0 , a term which reflects the distance between where we want the system to be at time
t1 (namely x1 ) and where it would end up if we left it alone (namely (t1 , t0 )x0 ). In other words, the
further we want to push the system from its natural course the more energy we need to expend. The
term in the middle, Wr (t0 , t1 )1 , in a sense reflects how controllable the system is. By Theorem 8.4,
if the system is uncontrollable Wr (t0 , t1 ) will be singular and, loosely speaking, Wr (t0 , t1 )1 would
be infinite, suggesting that we may not be able to push the system where we want however hard
we try. If the system is weakly controllable Wr (t0 , t1 ) will be positive definite but small, hence
Wr (t0 , t1 )1 will be large. In other words, the less controllable a system is the more energy we
need to expend to get it where we want. The observation that Wr (t0 , t1 ) Wr (t0 , t1 ) for t0 t0 and
t1 t1 (Fact 8.4) also admits a natural interpretation: The more time we have to take the system
where we want the less energy we need to do so (i.e., the more relaxed we can be about it!)
The proof of the theorem reveals that an even more general statement is possible. Even if the system
is not controllable, if it happens that x1 (t1 , t0 )x0 Range(Lr ) then we can still find a minimum
energy input that steers (x0 , t0 ) to (x1 , t1 ). The construction will be somewhat different (in this case
Wr (t0 , t1 ) will not be invertible) but the general idea is the same: Consider the restricted map
Lr |Range(L ) : Range(Lr ) Range(Lr )
r
which is bijective even if Range(Lr ) 6= Rn and select the unique u Range(Lr for which Lr (u) =
x1 (t1 , t0 )x0 . Indeed, even if x1 (t1 , t0 )x0 6 Range(Lr ) we can still formulate a more
general problem where the objective is to minimize kx1 x(t1 )k using a minimum energy input.
The construction will be even more complicated in this case, but the idea is again similar to the
pseudo-inverse calculations: Project x1 (t1 , t0 )x0 onto Range(Lr ) and among the u that map
to the projection select the unique one that belongs to Range(Lr ).
Finally, we note that Theorem 8.5 provides a simple example of an optimal control problem. The u
of the theorem is the input that drives the system from (x0 , t0 ) to (x1 , t1 ) and minimizes
Z t1
2
kuk2 = u(t)T u(t)dt.
t0
This is the starting point for more general optimal control problems where costs of the form
Z t1

xT (t)Q(t)x(t) + u(t)T R(t)u(t) dt
t0
nn
for some Q(t) R and R(t) Rmm symmetric and positive definite are minimized, subject
to the constraints imposed by the dynamics of the system and the desired initial and final states.
An example are the so-called Linear Quadratic Regulator (LQR) problems. The interested reader
is referred to [?, ?, ?] for more information on this topic.
8.4 Linear time varying systems: Observability and duality

We now turn to observability of time varying linear systems. As will soon become apparent, observ-
ability depends only on the properties of the matrices A() and C(); B() and D() play no role. We
therefore start by specializing Definition 8.2 to the linear time varying context.
Definition 8.6 The pair of matrices (C(), A()) is called observable on [t0 , t1 ] if and only if for
all x0 Rm and all u : [t0 , t1 ] Rm , one can uniquely determine x0 from the information
{(u(t), y(t)) | t [t0 , t1 ]}.
Note that once we know x0 and u : [t0 , t] Rm we can reconstruct x(t) for all t [t0 , t1 ] by
Z t
x(t) = (t, t0 )x0 + (t, )B( )u( )d.
t0
Moreover, since for all t [t0 , t1 ]

Z t
y(t) = C(t)(t, t0 )x0 + C(t) (t, )B( )u( )d + D(t)u(t),
t0
the last two terms in the summation can be reconstructed if we know u : [t0 , t1 ] Rm . There-
fore the difficult part in establishing observability is to determine x0 from the zero input response
C(t)(t, t0 )x0 with t [t0 , t1 ]. Without loss of generality we will therefore restrict our attention to
the case u(t) = 0 for all t [t0 , t1 ].
Definition 8.7 A state x0 Rn is called unobservable on [t0 , t1 ] if and only if C(t)(t, t0 )x0 = 0
for all t [t0 , t1 ].
Clearly the state x0 = 0 is unobservable. The question then becomes whether there are any x0 6= 0
that are also unobservable. This would be bad, since the zero input response from these states would
be indistinguishable from that of the 0 state.
Definition 8.8 The observability map of the pair (C(), A()) is the function
Lo : Rn L2 ([t0 , t1 ], Rp )
x0 7 C(t)(t, t0 )x0 t [t0 , t1 ].
The following fact is a direct consequence of the definition.
Fact 8.5 Lo is a linear, continuous function of x0 . Moreover, for all x0 Rn , Lo (x0 ) PC([t0 , t1 ], Rp .
The state x0 Rn is unobservable if and only if x0 Null(Lo ). The pair of matrices (C(), A())
is observable if and only if Null(Lo ) = {0}.
The proof is left as an exercise. Notice that to be able to apply the Finite Rank Lemma we are forced
to consider functions y() : [t0 , t1 ] Rp that belong to the space L2 ([t0 , t1 ], Rp ) in Definition 8.8.
Fact 8.5, however, clarifies that this is nothing to worry about: The only functions we will encounter
will live in Range(Lo ) and will be piecewise continuous, in fact the only discontinuity points will
be those of C(t).
Definition 8.9 The observability gramian of the pair (C(), A()) is the matrix
Z t1
Wo (t0 , t1 ) = (, t0 )T C( )T C( )(, t0 )d Rnn .
t0
The following theorem provides a complete characterization of observability of linear time varying
systems in terms of the properties of the matrix Wo (t0 , t1 ).
1. The pair of matrices (C(), A()) is observable on [t0 , t1 ].

2. Null(Lo ) = {0}.
3. Null(Lo Lo ) = {0}.
4. Det[Wo (t0 , t1 )] 6= 0.
Proof: The proof is effectively the same as that of Theorem 8.4 with Lo in place of Lr . Indeed,
Wo (t0 , t1 ) turns out to be the representation of the linear map Lo Lo : Rn Rn in the basis used
for the representation of the matrix A().
The similarity between the controllability and observability theorems is further highlighted by the
decomposition into subspaces induced by Lo , as shown in Figure 8.5. Comparing Figure 8.5 to
Figure 8.4, it becomes apparent that controllability and observability are in fact dual concepts, they
are just the two faces of the Finite Rank Lemma. To formalize this statement, note that given the
system
x(t) = A(t)x(t) + B(t)u(t) (8.6)

y(t) = C(t)x(t) + D(t)u(t) (8.7)
Rn L2 ([t0 , t1 ], Rm )
Range(Lo ) Lo Range(Lo )
y
Lo Lo x0
y
Lo
Null(Lo ) = {0}
y
Null(Lo )
Figure 8.5: Linear space decomposition induced by observability map.
with x(t) Rn , u(t) Rm , y(t) Rp , one can define the dual system

x(t) = A(t)T x(t) C(t)T u(t) (8.8)
T T
y(t) = B(t) x(t) + D(t) u(t) (8.9)
with x(t) Rn , u(t) Rp and y(t) Rm . The two systems are closely related in terms of their
controllability and observability properties.
Theorem 8.7 (Duality theorem) Let (t, t0 ), (t, t0 ) Rnn denote the state transition matri-
ces of system (8.6)(8.7) and system (8.8)(8.9) respectively.
1. (t, t0 ) = (t0 , t)T

2. System system (8.6)(8.7) is controllable on [t0 , t1 ] if and only if system (8.8)(8.9) is observ-
able on [t0 , t1 ].
3. System system (8.6)(8.7) is observable on [t0 , t1 ] if and only if system (8.8)(8.9) is control-
lable on [t0 , t1 ].

Proof: For Part 1, recall that by Theorem 4.2, (t0 , t0 ) = (t0 , t0 )T = I and t (t, t0 ) =
A(t)T (t, t0 ). Moreover

(t0 , t)(t, t0 ) = (t0 , t0 ) [(t0 , t)(t, t0 )] = [(t0 , t)] (t, t0 ) + (t0 , t) (t, t0 ) = 0
t t t

[(t0 , t)] (t, t0 ) = (t0 , t)A(t)(t, t0 )
t

(t0 , t)T = A(t)T (t0 , t)T A(t).
t
The conclusion follows by existence and uniqueness.
For Part 2, by Theorem 8.6 system (8.8)(8.9) is observable on [t0 , t1 ] if and only if the matrix
R t1 T
t0
(, t0 )T B( )T B( )T (, t0 )d is invertible. Substituting (t, t0 ) from Part 1 this matrix
is the same as
Z t1 Z t1
T T
(t0 , )B( )B( ) (t0 , ) d = (t0 , t1 )) (t1 , )B( )B( )T (t1 , )T d (t0 , t1 )T .
t0 t0
Rn L2 ([t0 , t1 ], Rp ) L2 ([t0 , t1 ], Rm ) Rn
Lo Lr
Lo Lo Range(L
o) Range(Lo ) Range(L
r) Range(Lr ) Lr Lr
Lo Lr
Null(L
o) Null(Lr )
Observability Controllability
Figure 8.6: duality.
Since (t0 , t1 ) is invertible by Theorem 4.2, the above is invertible if and only if
Z t1
(t1 , )B( )B( )T (t1 , )T d
t0
is invertible, i.e. if and only if system (8.6)(8.7) is controllable (by Theorem 8.4).
The proof of Part 3 is similar to that of Part 2 and is left as an exercise.
Duality is illustrated by the reflection symmetry of Figure 8.6.

Duality allows us to immediately extend results proved for controllability to related observability
results and vice versa. For example, the following dual statement of the minimum energy control
theorem allows us to estimate the value of the initial condition x0 that gave rise to a particular zero
input response trajectory y() : [t0 , t1 ] Rp .
Theorem 8.8 Assume that the pair of matrices (C(), A()) is observable and consider an arbitrary
y L2 ([t0 , t1 ], Rp ). Then
Z t1
x0 = (Lo Lo )1 Lo (y) = [Wo (t0 , t1 )]1 (, t0 )T C( )T y( )d
t0
is the unique minimizer of ky Lo (x)k2 over x Rn . Moreover,
min ky Lo (x)k22 = kyk22 xT0 Wo (t0 , t1 )x0

xRn
Proof: By the Finite Rank Lemma, y uniquely decomposes into a sum y = y+y with y Range(Lo )
and y Null(Lo ) (see Figure 8.5). For all x Rn , we clearly have Lo (x) Range(Lo ) and hence
y Lo (x) Range(Lo ) (recall that Range(Lo ) is a subspace of L2 ([t0 , t1 ], Rp )). Since Range(Lo )
and Null(Lo ) are orthogonal
ky Lo (x)k22 = ky Lo (x)k22 + kyk22 kyk22 . (8.10)
Therefore minxRn ky Lo (x)k2 is achieved for all x Rn such that Lo (x) = y.

Because the system is observable Null(Lo ) = {0} and the function Lo Lo : Rn Rn is invert-
ible (after all Wo (t0 , t1 ) is the representation of this map and Det[Wo (t0 , t1 )] 6= 0). Moreover,
Range(Lo ) = Rn ) and by the Finite Rank Lemma, the function
Lo (x) : Rn Range(Lo )
is bijective . Therefore, there exists a unique x0 Rn such that Lo (x0 ) = y. Note that
(Lo Lo )1 Lo (y) = (Lo Lo )1 Lo (y + y)

= (Lo Lo )1 Lo (y) (since y Null(Lo ))
= (Lo Lo )1 Lo Lo (x0 )
= x0 .
Writing the above in the coordinates used in the representation of A() and recalling the infinite
dimensional adjoint calculation in Section 7.3 shows that
Z t1
1
x0 = [Wo (t0 , t1 )] (, t0 )T C( )T y( )d
t0
is the unique minimizer of ky Lo (x)k2

To compute the minimum, note that y(t) = [Lo (x0 )](t) = C(t)(t, t0 )x0 . Hence,
Z t1 Z t1
kyk22 = T
y(t) y(t)dt = xT0 (t, t0 )T C(t)T C(t)(t, t0 )dtx0 = xT0 Wo (t0 , t1 )x0 .
t0 t0
The value of the minimum follows by substituting into (8.10).
Notice that the theorem does not require the observed output, y(), to lie in Range(Lo ), i.e. to
correspond exactly to the zero input response of some initial state x0 Rn . In fact, the observed
y() does even have to be piecewise continuous, as one would expect from a zero input response. The
formula provided in the theorem generates the x0 whose zero input res Ponce most closely matches
the observed y() in the 2-norm sense. This is very useful, since in practice the measured system
response will be corrupted by noise and other disturbances, hence is unlikely to correspond exactly
to some initial state.
As in the case of controllability, the above calculation can also be extended to the case where the
system is not observable. The minimizer will not be unique in this case and the computation will
be more involved since W0 (t0 , t1 ) will not be invertible, but the idea is similar: Find the projection
of the measured output y onto Range(Lo ) and the minimum norm x0 that corresponds to it.
8.5 Linear time invariant systems: Observability

Next, we restrict our attention further to the special case

y(t) = Cx(t) + Du(t).
Define the observability matrix by

C
CA

O= .. Rnpn .
.
CAn1
Theorem 8.9 For any [t0 , t1 ] the following hold:
1. Null(O) = Null(Lo ) = {x0 Rn | unobservable}.

2. The pair of matrices (C, A) is observable on [t0 , t1 ] if and only if Rank(O) = n.

3. Null(O) is an A invariant subspace, i.e. if x Null(O) then Ax Null(O).
4. The pair of matrices (C, A) is observable on [t0 , t1 ] if and only if for all Co

I A
Rank = n.
C
Proof: Consider any [t0 , t1 ] and without loss of generality assume that t0 = 0.
Part 1: x0 is unobservable if and only if x0 Null(Lo ), i.e. if and only if
C(t, 0)x0 = 0 t [0, t1 ] CeAt x0 = 0 t [0, t1 ]

A2 t2
C I + At + + . . . x0 = 0 t [0, t1 ]
2
CAk x0 = 0 k N (tk linearly independent by Fact 2.6)
CAk x0 = 0 k = 0, . . . , n 1 (by Cayley-Hamilton theorem)

C
CA

.. x0 = 0
.
n1
CA
x0 Null(O).
Part 2:
(C, A) observable Null(Lo ) = {0}

Null(O) = {0}
Dim(Null(O)) = Nullity(O) = 0
Rank(O) = n Nullity(O) = n ( by Theorem 2.2).
Part 3: Consider x0 Null(O), i.e. CAk x0 = 0 for all k = 0, . . . , n 1. We would like to show
that x = Ax0 Null(O). Indeed Cx = CAx0 = 0, CAx = CA2 x0 = 0, . . . , until
CAn1 x = CAn x0 = C(n I + n1 A + . . . + 1 An1 )x0 = 0
(where we make use of the Cayley Hamilton theorem).

Part 4: () By contraposition. Note that since A Rnn

I A
Rank n.
C
Assume that there exists C such that

I A
Rank < n.
C
Then the columns of the matrix are linearly dependent and there exists v Cn with v 6= 0 such that

I A (I A)v 0
v= = .
C Cv 0
Therefore, must be an eigenvalue of A with v the corresponding eigenvector and Cv = 0. Recall
that if v Cn is an eigenvector of A then
s(t, 0, v, U ) = eAt e = et v
(see, for example, the proof of Theorem 6.2). Therefore
CeAt e = et Cv = 0, t 0
and v 6= 0 is unobservable. Hence the pair of matrices (C, A) is unobservable1

Part 4: () Assume that C, A is unobservable. Then Null(O) 6= {0} and Dim(Null(O)) = r > 0
for some 0 < r n. Choose a basis for Null(O), say {v1 , . . . , vr }. Choose also a basis for Null(O) ,
say {w1 , . . . , wnr }. Since Null(O) is finite dimensional (hence closed) by Theorem 7.3,

Rn = Null(O) Null(O)
and {w1 , . . . , wnr , e1 , . . . , er } form a basis for Rn . The representation of x Rn with respect to
this basis decomposes into two parts, x = (x1 , x2 ), with x1 Rnr (observable) and x2 Rr
(unobservable). For the rest of this proof we assume that all vectors are represented with respect
to this basis.
Recall that, from Part 2, Null(O) is an Ainvariant subspace, i.e. for all x Null(O), Ax
Null(O). Note that
0
x Null(O) x = .
x2
Moreover,
A11 A12 0 A12 x2
Ax = =
A21 A22 x2 A22 x2
Therefore, since Ax Null(O) we must have that A12 x2 = 0 for all x2 Rr , which implies that
A12 = 0 R(nr)r (think of using the elements of the canonical basis of Rr as x2 one after the
other). Notice further that if x Null(O) then in particular x Null(C) and

0
Cx = C1 C2 = C2 x2 = 0 x2 Rr .
x2
Hence C2 = 0 Rpr .
In summary, in the new coordinates the system representation is
x1 (t) = A11 x1 (t) +B1 u(t)

x2 (t) = A21 x1 (t) +A22 x2 (t) +B2 u(t)
y(t) = C1 x1 (t) +Du(t).
Take C an eigenvalue of A22 Rrr (and note that will also be an eigenvalue of A Rnn ).
Let v Cr with v 6= 0 be the corresponding eigenvector. Then

I A 0 0
= .
C v 0
Hence the columns of the matrix are linearly dependent and

I A
Rank < n.
C
The theorem provides two easy ways of checking the observability of a time invariant linear system,
by checking the rank of matrices. Notice that the conditions of the theorem are independent of
1 This argument is fine as long as is real. If is complex, strictly speaking, we have not constructed a real vector
x0 Rn which is unobservable. This can be done, however, by taking linear combinations of the complex eigenvectors,
as in the proof of Theorem 6.2. The construction is left as an exercise.
[t0 , t1 ]. Therefore, (C, A) is observable on some [t0 , t1 ] with t0 < t1 if and only if it is observable
on all [t0 , t1 ]. To put it another way, observations {(u(t), y(t) | t [0, ]} over an arbitrary small
interval are sufficient to reconstruct the initial condition x0 and hence all future states of the system
(at least in theory).
The most commonly used test for observability involves testing the rank of the matrix O. Though
easy to implement for small systems, this test may face numerical difficulties for larger systems, since
it requires the computation of powers of A up to An1 , which may lead to numerical conditioning
problems. Testing the rank of
I A
C
tends to be more robust numerically on the other hand. Notice that even though the theorem
suggests that this condition needs to be tested for all C (clearly an impossible task) the proof
indicates that the condition need only be checked for Spec[A], which is feasible.
The decomposition of the state developed in the proof of Part 4 also shows that for any (C, A),
the state of the system decomposes into an observable and and unobservable part. An immediate
corollary of the proof of the theorem is that the observable part is indeed observable.
Corollary 8.1 For any matrices C Rpn and A Rnn there exists a change of basis T Rnn
with Det[T ] 6= 0, such that in the new coordinates the representation of the matrices decomposes
into
A11 0
A = T AT 1 = and C = CT 1 = C1 0
A21 A22
and the pair of matrices (C1 , A11 ) is observable.
Proof: The decomposition was already established in the proof of Theorem 8.9. To show that
(C1 , A11 ) are observable assume, for the sake of contradiction, that they are not. Then there exists
x1 R(nr)r with x1 6= 0 such that

C1
C1 A11

.. x1 = 0.
.
nr1
C1 A11
The structure of the matrices implies that

C1 0
C1 A11 0
x1 x1
O = .. .. = 0.
0 . . 0
nr1
C1 A11 0
Hence (x1 , 0) Null(O). By the choice of basis, however, (x1 , 0) Span{w1 , . . . , wnr } =
Null(O) . Since (x1 , 0) 6= 0 this is a contradiction (by Fact 7.2).
Note also that because

I1 A11 0
Det = Det [I1 A11 ] Det [I2 A22 ]
A21 I2 A22
(where I1 and I2 are identity matrices of appropriate dimensions) the spectrum of A decomposes
into
Spec[A] = Spec[A11 ] Spec[A22 ]
where the former set contains eigenvalues whose eigenvectors are observable (known as observable
modes) while the latter contains eigenvalues whose eigenvectors are unobservable (known as unob-
servable modes). An immediate danger with unobservable systems is that the state may be diverging
to infinity without any indication of this at the output. Indeed, if the system is unobservable and one
of the eigenvalues of the matrix A22 above has positive real part the state x2 (t) may go to infinity
as t increases. Since, however, (0, x2 ) is in the nullspace of O, the output y(t) will be unaffected by
this (either directly or through the state x1 ) and may remain bounded. For unobservable systems
one would at least hope that the unobservable modes (the eigenvalues of the matrix A22 ) are stable.
In this case, if the output of the system is bounded one can be sure that the state is also bounded.
This requirement, which is weaker than observability, is known as detectability.
8.6 Linear time invariant systems: Controllability

As expected, the picture for controllability of time invariant linear systems is dual to that of observ-
ability. One can define the controllability matrix by

P = B AB . . . An1 B Rnnm
and use it to test whether the systems is controllable. The main facts in this direction are summarized
in the following theorem.
Theorem 8.10 For any [t0 , t1 ] the following hold:
1. Range(P ) = Range(Lr ) = {x Rn | reachable}.

2. The pair of matrices (A, B) is controllable on [t0 , t1 ] if and only if Rank(P ) = n.
3. Range(P ) is an A invariant subspace.
4. The pair of matrices (A, B) is controllable on [t0 , t1 ] if and only if for all C
Rank [I A B] = n.
The proof is effectively the same as that of Theorem 8.9 by duality. As before, the condition of
Part 4 need only be tested for Spec[A]. Moreover, for all (A, B) the state decomposes into a
controllable and an uncontrollable part. For an appropriate choice of basis the system representation
becomes:
x1 (t) = A11 x1 (t) +A12 x2 (t) +B1 u(t)
x2 (t) = +A22 x2 (t)
y(t) = C1 x1 (t) C2 x2 (t) +Du(t)
where the pair of matrices (A11 , B1 ) is controllable. Likewise, the spectrum of A decomposes into
Spec[A] = Spec[A11 ] Spec[A22 ]
where the former set contains eigenvalues whose eigenvectors are reachable (controllable modes) while
the latter contains eigenvalues whose eigenvectors are unreachable (uncontrollable modes).
Dually to observability, an immediate danger with uncontrollable systems is that the state may be
diverging to infinity without the input being able to prevent it. If the system is uncontrollable and
one of the eigenvalues of the matrix A22 above has positive real part the state x2 (t) may go to
infinity as t increases. Since the evolution of this state is not affected by the input (neither directly
nor through the state x1 ) there is nothing the input can do to prevent this. For uncontrollable
systems one would at least hope that the uncontrollable modes (the eigenvalues of the matrix A22 )
are stable. This requirement, which is weaker than controllability, is known as stabilizability.
8.7 Kalman decomposition

Applying the two theorems one after the other leads to the Kalman decomposition theorem.
Theorem 8.11 (Kalman decomposition) For an appropriate change of basis for Rn the state
vector is partitioned into

x1 controllable, observable
x2 controllable, unobservable
x=
x3 .
uncontrollable, observable
x4 uncontrollable, unobservable
In these coordinates the system matrices get partitioned into

A11 0 A13 0 B1
A21 A22 A23 A24
A= B = B2
0 0 A33 0 0
0 0 A43 A44 0

C = C1 0 C3 0 D
with

A11 0 B1 A11 A13
, controllable and C1 C3 , observable.
A21 A22 B2 0 A33
The theorem shows that the spectrum of A gets partitioned into
Spec[A] = Spec[A11 ] Spec[A22 ] Spec[A33 ] Spec[A44 ]
where the first set contains controllable and observable modes, the second set controllable and unob-
servable modes, the third set uncontrollable and observable modes and the fourth set uncontrollable
and unobservable modes.
Pictotially the interdependencies of the sub-systems revealed by the Kalman decomposition are
shown in Figure 8.7. Note that this figure is not a block diagram, the arrows represent dependencies
between subsystems rather than signals. The labels next to the arrows indicate which of the matrices
in Theorem 8.11 is responsible for the corresponding dependence.

Problem 8.1 (Controllability Gramian) Let
Z t1
Wr (t1 , t0 ) = (t1 , )B( )B( )T (t1 , )T d
t0
denote the controllability Gramian of the linear time varying system
x(t) = A(t)x(t) + B(t)u(t)
1. Show that Wr (, t0 ) is the unique solution of the matrix differential equation
X(t) = A(t)X(t) + X(t)AT (t) + B(t)B T (t) ()
with X(t0 ) = 0 Rnn .

u(t) B1 C1 y(t)
A11 +
A21 A13
B2 A23 C3
A22 A33
A33 A43
A44
Figure 8.7: Kalman decomposition.
2. Show that Wr (t1 , t0 ) 0 and that
Wr (t1 , t0 ) Wr (t1 , t0 ) 0 t0 t0 . ()
[Here P 0 means that P Rnn is positive semi-definite.]
Problem 8.2 (Controllability) Consider a harmonic oscillator with control input u, satisfying
the equation
x1 0 1 x1 0
= + u
x2 1 0 x2 1
Suppose we want to drive the system from the state [1, 0]T to the state [0, 0]T in 2 units of time.
1. Does there exist a control u which makes this transfer?

2. Now suppose u is to be a piecewise constant function of time of the form

u1 0 t < 2/3,

u(t) = u2 2/3 t < 4/3,

u3 4/3 t < 2
Do there exist constants u1 , u2 , u3 such that we can make the transfer from [1, 0]T at t = 0 to
[0, 0]T at time t = 2?
Problem 8.3 (Controllability) Let A be an n n matrix and B be an n m matrix, both with

real entries. Assume that the pair (A, B) is controllable. Prove or disprove the following statements
(a counterexample will suffices to disprove a statement):
1. The pair (A2 , B) is controllable.

2. Given that the system
x = Ax + Bu
has the initial condition x(0) = x0 6= 0, it is possible to find a piecewise continuous control
u : [0, +) Rm such that the system is brought to rest at t = 1 (i.e., x(t) = 0 for all t 1).
3. Suppose that the system is initially at rest, i.e. x(0)=0, and for x Rn we wish to find a
piecewise continuous control u : [0, +) Rm such that x(t) = x for all t 1. Such a control
can be found for all x Rn .
Problem 8.4 (Observability) For some matrices A Rnn , B Rn1 , C R1n and D R,
consider a single-input, single-output time-invariant system
x(t) = Ax(t) + Bu(t),

y(t) = Cx(t) + Du(t).
Assume that (A, C) is observable.

1. For a generic t 0, provide an expression for x(t) in terms of the derivatives y (i) (t); i =

0, 1, . . . , n 1 and u(i) (t); i = 0, 1, . . . , n 1 . [Hint: Consider the output equation y(t) =
Cx(t) + Du(t) and differentiate repeatedly.]
2. Discuss why for time-invariant systems the observation of y and u over an arbitrarily small
interval t [0, ], with > 0, suffices to reconstruct x(0).
Problem 8.5 (Duality) Let (t, ) denote the state-transition matrix of
x(t) = A(t)x(t) + B(t)u(t)

y(t) = C(t)x(t) + D(t)u(t)
and (t, ) that of the dual system
x = A(t) x
e(t) e(t) C(t) u
e(t)

ye(t) = B(t) xe(t) + D(t) ue(t)
Show that (t, ) = (, t) .

Chapter 9
State Feedback and Observer

Design
Knowing whether a system is controllable or observable is usually not an end in itself. One wants
to subsequently exploit this knowledge to design controllers that will make the system behave in
some desired way: Ensure stability, force the output to track some desired trajectory, force the state
to converge to an equilibrium at some desired rate, etc. We have already seen how for controllable
systems one can generate input trajectories to steer the system from its initial state to some desired
final state (Theorem 8.5). In practice, however, this may not be sufficient. Due to small deviations
in the initial state or in the matrices involved in the dynamics, these so-called open loop input
trajectories will invariably fail to steer the system as expected. Indeed, if the system is unstable
the deviation may get arbitrarily large. A much more robust way of steering the system is to
measure its state as we go along and if we find that it deviates from the desired trajectory adapt
the open loop inputs to correct the deviation. The simplest case of such a feedback control involves
stabilization, where the desired trajectory is simply an equilibrium solution and the objective is to
design a controller to make the system asymptotically stable. Similar (dual) comments can also be
made for the merits of closed loop state reconstruction by incorporating output measurements
on-line, as opposed to the batch processing approach of Theorem 8.8.
In this chapter we investigate how such controllers and state estimators can be designed. We first
consider the so-called state feedback controllers, where one assumes that the entire state of the
system can be measured and used when making decisions. We then extend the approach to output
feedback, where one assumes that only the outputs of the system are available for measurement.
In particular, we consider the case of observer based output feedback, where an observer is used to
reconstruct the value of the state from the output measurements; a state feedback controller then
uses the reconstructed state to steer the system.
The discussion is restricted to linear time invariant systems. In this case, one makes use of the
controllability and observability matrices derived in Chapter 8 to transform the system into special
forms for which controllers and observers can easily be designed. The transformations involve changes
of basis using invertible matrices derived from the controllability and observability matrices. For
clarity of exposition we first recall some facts about basis changes.
130
9.1 Revision: Change of basis

Consider a linear, time invariant system
x(t) = Ax(t) + Bu(t), (9.1)

y(t) = Cx(t) + Du(t), (9.2)
where as usual, t R+ , x(t) Rn , u(t) Rm , y(t) Rp , A Rnn , B Rnm , C Rpn and

D Rpm . Consider also a change of basis x(t) = T x(t) where x(t) Rn and T Rnn is an
invertible matrix. Recall that in the new coordinates the evolution of system is also described by
linear equations

x(t) = Ax(t) + Bu(t) (9.3)
y(t) = C x(t) + Du(t), (9.4)
with A Rnn , B Rnm , C Rpn and D Rpm . Throughout these notes several facts
regarding the relation between these two descriptions of the system have been established. We
summarize the most important ones in the following theorem.
Theorem 9.1 Consider the linear time invariant systems (9.1)(9.2) and (9.3)(9.4) related through
the change of coordinates x(t) = T x(t) for all t R+ , where T Rnn is an invertible matrix. The
following hold:
1. The matrices in (9.3)(9.4) are given by A = T AT 1, B = T B, C = CT 1 , D = D.

2. The matrices A and A have the same eigenvalues, i.e. Spec[A] = Spec[A].
3. The two systems have the same transfer function, i.e.
G(s) = C(sI A)1 B + D = C(sI A)1 B + D = G(s).
4. The two systems have the same impulse response matrix, i.e.
H(t) = CeAt B + D0 (t) = CeAt B + D0 (t).
5. (A, B) is controllable if and only if (A, B) is controllable.

6. (C, A) is observable if and only if (C, A) is observable.
The proof is left as an exercise.

The theorem shows that changes of coordinates do not affect the fundamental properties of the
system. This is to be expected, since a change of coordinates does not change the system itself,
just the representation of the system. So the only things affected by the change of coordinates are
those having to do with the representation of the matrices and not fundamental properties such as
stability, controllability, or observability.
Exercise 9.1 Are the state transition and impulse state transition matrices of the two systems the
same?
In this chapter we will repeatedly use the properties of coordinate transformations to bring the
system equations into special forms that make it easier to design controllers and observers. We will
then exploit the properties listed in Theorem 9.1 to transfer the design to the original coordinates.
The two forms we will especially be interested in are the so-called controllable canonical form and
observable canonical form. Both of these will first be presented for the easier case of single input,
single output systems and then generalized.
9.2 Linear state feedback for single input systems

Consider first a linear time invariant system with only one input
x(t) = Ax(t) + Bu(t), (9.5)
where u(t) R and B Rn is a vector. The output equations will play no role in this section and
will be omitted for simplicity, we will return to them in Section 9.3.
Recall that the system (9.5) is controllable if and only if the matrix

P = B AB . . . An1 B Rnn (9.6)
has full rank, or, equivalently (since the matrix is square), if and only if it is invertible, or, equiva-
lently, if and only if its columns are linearly independent. We will use the last fact to construct an
invertible matrix T Rnn to use as a change of coordinates. We will start by defining a matrix
S Rnn that will be invertible whenever the system is controllable and will eventually become
T 1 .
Let the characteristic polynomial of the matrix A Rnn be
Det[I A] = n + 1 n1 + . . . + n1 + n .
Consider a family of n vectors, si Rn for i = 1, . . . , n, defined last-to-first as follows:
sn = B
sn1 = Asn + 1 B = AB + 1 B
sn2 = Asn1 + 2 B = A2 B + 1 AB + 2 B
..
.
s1 = As2 + n1 B = An1 B + 1 An2 B + . . . + n1 B.
Lemma 9.1 As1 + n B = 0. Moreover, the matrix S = [s1 . . . sn ] Rnn is invertible if and only
if the system (9.5) is controllable.
Proof: By definition
As1 + n B = An B + 1 An1 B + . . . + n1 AB + n B

= An + 1 An1 + . . . + n1 A + n I B = 0
by the Cayley-Hamilton theorem (Theorem 5.7).

Recall that S is invertible if and only if its columns are linearly independent. Notice that si is
the sum of Ani B plus a linear combination of Ani1 B, . . . , B. The Ani B terms are linearly
independent of each other if and only if the system is controllable, hence the claim follows. Or, more
formally, note that

n1 n2 . . . 1 1
n2 n3 . . . 1 0

.. .
S = B AB . . . A n1
B ... ..
.
..
. .
..
.

1 1 ... 0 0
1 0 ... 0 0
r(t) u(t) x(t) R x(t) y(t)

+ B + C +
Figure 9.1: Linear state feedback.
Therefore, recalling the definition of the matrix P in (9.6),

n1 n2 . . . 1 1
n2 n3 ... 1 0

Det[S] = Det[P ] Det ... ..
.
.. .
. ..
..
.,

1 1 ... 0 0
1 0 ... 0 0
which is not equal to zero if and only if the system is controllable; notice that the determinant of
the matrix on the right is email to 1.
The lemma shows that whenever system (9.5) is controllable we can use S to define a change of
coordinates. It turns out that in the new coordinates the system matrices take a particularly useful
form.
Theorem 9.2 The system (9.5) is controllable if and only if there exists a change of coordinates
x(t) = T x(t) with T Rnn invertible, such that the matrices A = T AT 1 and B = T B satisfy:

0 1 0 ... 0 0
0 0 1 ... 0 0

.. .
.. .
.. . .. .
.. , B = .
A = . .. (9.7)

0 0 0 ... 1 0
n n1 n2 . . . 1 1
Proof: (): Assume that the system is controllable. Define S Rnn as above and let T = S 1 .
Note that
0 0
.. ..

B = S . B = T B = S 1 B = . .
0 0
1 1
Consider next the matrix AS = A [s1 . . . sn ]. Reading the columns from the right we have

0 0
.. ..
. .

Asn = sn1 1 B = S 0 , Asn1 = sn2 2 B = S 1 , etc. until

1 0
1 2

0
..
.

As1 = n B = S 0 (by Lemma 9.1).

0
n
Hence,
0 1 0 ... 0
0 0 1 ... 0

.. .. .. .. ..
AS = S . . . . .

0 0 0 ... 1
n n1 n2 . . . 1
which implies that

0 1 0 ... 0
0 0 1 ... 0

.. .. .. .. ..
. . . . . = S 1 AS = T AT 1 = A.

0 0 0 ... 1
n n1 n2 . . . 1
(): Assume that there exists a change of coordinates bringing the systems in the form (9.7). Note
that in the new coordinates

0 0 ... 0 1
... 1
h i 0 0
. . .. ..
P = B . . . An1 B = .. .. . .

0 1 ...
1 1 . . .
h i
where stands for some number which depends on 1 , . . . , n . Hence Det P = 1 and the system
in the new coordinates is controllable. Theorem 9.1 then implies that the system in the original
coordinates is also controllable.
The system representation of equation (9.7) is known as the controllable canonical form. The term
canonical refers to the fact that any controllable system can be brought to this form through a
change of coordinates, as Theorem 9.2 shows. The controllable canonical form is useful for designing
state feedback controllers to modify the behavior of system (9.5). A state feedback controller is a
method for selecting the inputs of the system as a function of the state to force the system to exhibit
some desired behavior. Here we will restrict our attention to linear time invariant state feedback
controllers of the form
u(t) = Kx(t) + r(t), (9.8)
where K Rmn is known as the gain matrix, and r(t) Rm is an external input vector; in the
single input case considered in this section K = [k1 . . . kn ] R1n is just a row vector.
The system obtained by connecting a state feedback controller of the form (9.8) with a system of
the form (9.5) is known as the closed loop system (Figure 9.1). Its dynamics are described by the
linear state equations
x(t) = (A + BK)x(t) + Br(t). (9.9)
We would like to select the gain matrix K such that the closed loop systems exhibits some desired
behavior, for example is asymptotically stable and its state tends to zero sufficiently fast. To ensure
properties like these one typically needs to select K so that the eigenvalues of the closed loop system
matrix A + BK coincide with some desired set {1 , . . . , n } C. Clearly for this to be possible the
set must contain the complex conjugates of all its members, i.e. {1 , . . . , n } = {1 , . . . , n }; we
call such a set of complex numbers a complex conjugate set.
Theorem 9.3 System (9.5) is controllable if and only if for all complex conjugate sets {1 , . . . , n }
C there exists K Rmn such that Spec[A + BK] = {1 , . . . , n }.
Proof: (). Assume that (9.5) is controllable. Given a desired complex conjugate set {1 , . . . , n }
C we would like to select feedback gains K such that the closed loop matrix A + BK has this set as
eigenvalues. In other words, we would like
Det[I (A + BK)] = ( 1 )( 2 ) . . . ( n )
= n + d1 n1 + . . . + dn1 + dn (9.10)
where d1 , . . . , dn are real numbers uniquely determined by 1 , . . . , n .

By Theorem 9.2 there exists a change of coordinates x(t) = T x(t) to bring the system into control-
lable canonical form

0 1 0 ... 0 0
0 0 1 . . . 0 0

.. x(t) + ..

x(t) = ... ..
.
..
.
..
. . . u(t). (9.11)

0 0 0 ... 1 0
n n1 n2 . . . 1 1
In the new coordinates, the feedback function can be written as
u(t) = Kx(t) + r(t) = KT 1T x(t) + r(t) = K x(t) + r(t)

where K = kn . . . k1 R1n is the representation of the gain matrix in the new coordinates.
Substituting into (9.11) we obtain

x(t) = (A + B K)x(t) + Br(t)

0 1 0 ... 0 0
0 0 1 ... 0 0

.. .. .. .. .. ..
= . . . . . x(t) + . r(t).

0 0 0 ... 1 0
kn n kn1 n1 kn2 n2 . . . k1 1 1
Notice that the system is still in controllable canonical form. Therefore the terms in the last row
are the coefficients of the characteristic polynomial
Det[I (A + B K)] = n (k1 1 )n1 . . . (kn1 n1 ) (kn n ).
If we set
k1 = 1 d1 , . . . , kn = n dn
this becomes identical to the desired characteristic polynomial (9.10). Hence with this selection for
the matrix K we have made the eigenvalues of A+ B K equal to the desired set {1 , . . . , n }. Finally,
since eigenvalues are not affected by coordinate transformations (Theorem 9.1) setting K = KT
ensures that with the state feedback u(t) = KT x(t) + r(t) the eigenvalues of the matrix A + BK of
the closed loop system are also equal to the desired list.
(). By contraposition. Assume that the system is not controllable. Then, by Theorem 8.10, there
exists a change of coordinates x(t) = T x(t) such that

A11 A12 B1
x(t) = x(t) + u(t).
0 A22 0
In these coordinates a feedback function u(t) = Kx(t) + r(t) can be written as

u(t) = KT 1T x(t) + r(t) = K1 K2 x(t) + r(t).
The closed loop system then becomes

A11 + B1 K1 A12 + B1 K2 B1
x(t) = x(t) + r(t)
0 A22 0
and its characteristic polynomial

I1 (A11 + B1 K1 ) (A12 + B1 K2 )
Det = Det[I1 (A11 + B1 K1 )]Det[I2 A22 ],
0 I2 A22
where I1 and I2 denote identity matrices of appropriate dimensions. Hence
Spec[A + B K] = Spec[A11 + B1 K1 ] Spec(A22 ).
Notice that part of the spectrum (the eigenvalues of A22 ) is not affected by the feedback gains K.
Therefore given a list of desired eigenvalues {1 , . . . , n } C it is impossible, in general, to select K
such that Spec[A + B K] = {1 , . . . , n }; this is only possible if Spec[A22 ] {1 , . . . , n }. Setting
K = KT and noting that Spec[A + BK] = Spec[A + B K] (Theorem 9.1) completes the proof.
Theorem 9.3 provides a method for moving the eigenvalues of a controllable linear system to arbitrary,
complex conjugate values by selecting a feedback gain matrix K and applying linear state feedback
u(t) = Kx(t) + r(t). This method is known in the literature as pole placement, eigenvalue placement,
or eigenvalue assignment. Even though for the proof of the theorem we first have to bring the system
into controllable canonical form, in practice this is not necessary. Given matrices A Rnn and
B Rnm we can directly compute the characteristic polynomial Det[I (A + BK)], treating the
entries in the matrix K as variables. The coefficients of the characteristic polynomial will turn out
to be linear functions of the entries of the gain matrix K Rmn . Equating these coefficients with
those of the desired characteristic polynomial (9.10) leads to a system of linear equations where the
entries of K play the role of the unknowns. Theorem 9.3 then ensures that this system will have
a solution which, if used in the feedback function u(t) = Kx(t) + r(t), will lead to a closed loop
system with the desired eigenvalues. As we will see below, this procedure works more generally,
even if m > 1. The difference is that for single input systems (m = 1) the system of linear equations
has the same number of equations as unknowns (both n). Thus (assuming controllability) there
is a unique solution to this system of linear equations and hence a unique feedback gain matrix
K R1n that will get the job done. For multi-input systems (m > 1), on the other hand, there
are more unknowns (nm) than equations (n). Therefore there will be multiple choices of K that
will lead to the same eigenvalues for the close loop system; in fact a whole n(m 1) dimensional
subspace of them! Unless other considerations are present (minimizing some cost criterion, etc.) one
can eliminate the redundant degrees of freedom by setting some of the elements of K equal to zero.
This leads to a sparser feedback gain matrix, which both simplifies the calculations and makes the
feedback function easier to implement in practice.
Theorem 9.3 also shows that the procedure outlined above for designing feedback gain matrices will
in general fail if the system is not controllable. The same steps can again be executed, but in this case
r(t) u(t) x(t) x(t) y(t)

+ B C +
R
+
System (A, B, C, D)
Output feedback
Figure 9.2: Output feedback.
the system of linear equations in the entries of K will not have any solution in general. The reason
is the presence of the uncontrollable modes (the elements of Spec[A22 ] in the proof of Theorem 9.3)
which are not affected by our choice of K. If Spec[A22 ] contains some eigenvalues with positive real
part this is usually bad news: Whatever gain matrix K we select, it will not be possible to obtain a
stable closed loop system. If stability is what we are after, then drastic system re-design is needed.
We may, for example, try to add, or modify one of the system actuators to augment the matrix B
and make the offending modes controllable. If, on the other hand, Spec[A22 ] lies entirely on the left
half of the complex plane (i.e. the system is stabilizable) it is still possible to stabilize the system.
We just have to resign ourselves to the fact that the eigenvalues in Spec[A22 ] will stay put whatever
we do and select K to move the remaining eigenvalues (the controllable modes in Spec[A11 ]) to
desired locations. To do this, we first we have to determine the uncontrollable modes (e.g. using the
rank test in Theorem 8.10, condition 4), then construct a list of desired eigenvalues containing the
uncontrollable modes, compute the characteristic polynomial Det[I (A + BK)], and equate its
coefficients with those of the desired characteristic polynomial. The proof of Theorem 9.3 together
with the fact that the uncontrollable modes are included in the list of desired eigenvalues will ensure
that the resulting system of linear equations will have a solution for the entries of K. The solution
will not be unique, however, even for single input systems. Some of the degrees of freedom (those
corresponding to K2 in the proof of Theorem 9.3) will be redundant. Unless other considerations
are present, one can again select these redundant degrees of freedom to make the matrix K as sparse
as possible.
9.3 Linear state observers for single output systems

The method of pole placement described above is powerful and theoretically elegant. One can argue,
however, it is not entirely practical. In real systems it is often the case that not all states are available
for measurement. The set up of Figure 9.1 is therefore optimistic, because in reality one only had
access to information about the output of the system (measured through appropriate sensors) when
trying to make decisions about what inputs to apply. A more realistic feedback arrangement should
therefore look more like Figure 9.2. The problem now is how to design a feedback function that uses
the measurements of the outputs (past and present) to decide what input to apply to the system.
Since we already have a powerful method for designing state feedback controllers, one idea that
presents itself is to augment this by an algorithm that generates an estimate of the state using
past and present inputs and outputs. Then, instead of the true value of the state (which is not
available) we can use the estimate of the state and our state feedback controller to compute a value
for the input to apply. An algorithm that generates an estimate of the state using the past and
present inputs and outputs is known as a state estimator or a state observer. The resulting feedback
arrangement will look like Figure 9.3.
It turns out that for linear systems it suffices to consider linear state observers. Analogously to
the case of state feedback design we will start by describing the design of such observers for single
output systems (p = 1) and return to the general case of multi-output systems (p > 1) toward the
end of this chapter. We will consider the usual linear system equations
x(t) = Ax(t) + Bu(t) (9.12)
y(t) = Cx(t) + Du(t) (9.13)
where C R1n and D R1m are just row vectors.
A linear state observer generates an estimate of the state, x(t), at time t R+ using the values
of the input and output for all times up to t, {(u( ), y( )) | [0, t]}. The observer can itself be
thought of as a linear system, with n states (x(t) Rn ), m + p inputs ((u(t), y(t)) Rm+p ), and
n outputs (the state estimate itself x(t) Rn ). The equations of the observer try to mimic those
of the linear system (9.12)(9.13), with a correction term to account for any mismatch between the
measured output, y(t), at time t and the output predicted by the observer (denoted by y(t)) based
on its own state estimate, x(t). More precisely,

x(t) = Ax(t) + Bu(t) + L(y(t) y(t)) (9.14)
y(t) = C x(t) + Du(t). (9.15)
Notice that at time t the observer only makes use of its own estimate of the state, x(t), the input
u(t) and the output y(t) to update its estimate of the state; in particular the unknown value of the
state, x(t), is not used anywhere. The linear correction term, L(y(t) y(t)), introduces an indirect
estimate of the difference between the estimated and the true value of the state which, as we will
see below, can be used to make the two converge to each other. The matrix L Rnp is known
as the observer gain matrix. Note also that only the present values of y(t) and u(t) are used in
Equations (9.14)(9.15). The state estimate x(t) at time t, however, will depend on all of the past
values of the inputs and outputs, {(u( ), y( )) | [0, t]} (as well as the initial condition x0 that we
select). In a sense, x(t) encapsulates all the relevant information contained in the input and output
trajectories up to time t.
Under what conditions will x(t) be a good estimate of x(t)? To study this question we can look at the
estimation error, e(t) = x(t) x(t) Rn . If this error (or more precisely its norm) is small then the
state estimate generated by the observer is accurate. Clearly since initially we know nothing about
the true value of x(0) our initial estimate, x(0), and hence the initial error, e(0), will be arbitrary.
The question is whether the estimate will get better as time goes on and more information is collected
through u(t) and y(t). To answer this question we can look at how the error, e(t), evolves over time:

e(t) = x(t) x(t)
= Ax(t) + Bu(t) Ax(t) Bu(t) L(y(t) y(t))
= A(x(t) x(t)) L(Cx(t) + Du(t) Cx(t) Du(t))
= (A LC)(x(t) x(t))
= (A LC)e(t).
Notice that the evolution of the estimation error is also governed by a linear system without any
inputs. The estimation error will therefore converge to zero if and only if the eigenvalues of the
matrix A LC have negative real part. Otherwise, as time goes on our estimate of the state will
remain as bad as our initial guess, or get even worse, despite the fact that we have collected more
information through the output of the system.
The question now becomes can we select the observer gain matrix L so that the eigenvalues of ALC
have negative real parts? It is easy to see that this question is the dual of the question addressed
r(t) u(t) y(t)

+ System (A, B, C, D)
x(t)
K Observer
Figure 9.3: Linear feedback with state estimation. x(t) denotes the estimate of the state
in Section 9.2. If the system is observable, one can in fact mimic the pole placement method step
by step to develop a method for placing the eigenvalues of the estimation error dynamics in some
desired, complex conjugate values. First, one can write the system in observable canonical form
using a change of coordinates.
Theorem 9.4 The system (9.12)(9.13) is observable if and only if there exists a change of coor-
dinates x(t) = T x(t) with T Rnn invertible, such that the matrices A = T AT 1 and C = BT 1
satisfy:

0 0 ... 0 n
1 0 . . . 0 n1

A = . . .. . . .. ,
.. .. . . .
0 0 ... 1 1

C = 0 0 ... 0 1 .
The proof is the dual to that of Theorem 9.2: One uses the rows of the observability matrix

C
CA

O= .. Rnn
.
CAn1
which under the observability assumption are linearly independent to generate a change of coordi-
nates and bring the system to the desired form. One can then use the observable canonical form to
show the following.
Theorem 9.5 System (9.12)(9.13) is observable if and only if for all complex conjugate sets
{1 , . . . , n } C there exists L Rmn such that Spec[A LC] = {1 , . . . , n }.
The procedure for selecting the observer gain matrix is also similar. One uses the desired eigenvalues
{1 , . . . , n } to generate a desired characteristic polynomial. Then one computes the polynomial
Det[I (A LC)] and equates its coefficients to those of the desired polynomial, giving rise to a
system of n linear equations in np unknowns (the entries of the observer gain matrix L Rnp ). If
the system (9.12)(9.13) is observable, then the system of linear equations has a unique solution (if
p = 1) or multiple solutions (if p > 1), for which the eigenvalues of A LC are equal to the desired
ones. If the system (9.12)(9.13) is not observable, on the other hand, the system of linear equation
will in general have no solution. If the system is not even detectable (i.e. some of the unobservable
modes have non-negative real parts) then there is no hope of designing a stable observer that will
generate a correct estimate of the state, even asymptotically. If the system is detectable one can
still design a stable observer by computing the unobservable modes, including them in the list of
desired eigenvalues, and placing the remaining eigenvalues in some desired locations as above.
9.4 Output feedback and the separation principle

We have seen how, assuming that the value of the state is known, one can design a linear feedback
controller so that the eigenvalues of the closed loop system will be at some desired position. We
have also seen how, in cases where the value of the state is not fully measured, one can design an
observer that will asymptotically reconstruct the state. Presumably when one puts the two designs
(controller and observer) together as in Figure 9.3 the closed loop system will work as planned: In
the long run the state estimation error will converge to zero, the observer will provide the feedback
controller with the right value of the state, the controller will apply an appropriate input for this
value of the state, and everything will be fine. Or will it? Could something go catastrophically
wrong in the transients? Could the initial state estimation error cause the controller to make the
wrong decisions and destabilize the system, leading to greater estimation error, greater mistakes in
the control inputs, etc. so that eventually the whole thing breaks down?
To answer this question we now turn our attention to the dynamics of the closed loop system of
Figure 9.3. Collecting all the relationships from Sections 9.2 and 9.3 we have:

y(t) = Cx(t) + Du(t)
u(t) = K x(t) + r(t)

x(t) = Ax(t) + Bu(t) + L(y(t) y(t))
y(t) = C x(t) + Du(t).
Substituting one into the other leads to
x(t) = Ax(t) + BK x(t) + Br(t)

x(t) = LCx(t) + (A + BK LC)x(t) + Br(t)
y(t) = Cx(t) + DK x(t) + Dr(t).
The closed loop system is therefore itself a linear time invariant system, with 2n states (x(t), x(t))
R2n ), m inputs (r(t) Rm ) and p outputs (y(t) Rp ) and state equations

x(t) A BK x(t) B
= + r(t)
x(t) LC A + BK LC x(t) B

x(t)
y(t) = C DK + Dr(t).
x(t)
To determine whether this system is stable it is easier to start with a coordinate transformation,
making the state estimation error one of the states. Note that

x(t) x(t) I 0 x(t) x(t)
= = =T
e(t) x(t) x(t) I I x(t) x(t)
where I stands for the n n identity matrix.

I 0
Exercise 9.2 Show that T = is invertible and that T 1 = T .
I I
In the new coordinates

x(t) A + BK BK x(t) B
= + r(t)
e(t) 0 A LC e(t) 0

x(t)
y(t) = C + DK DK + Dr(t).
e(t)
Note that

I (A + BK) BK
Det = Det[I (A + BK)]Det[I (A LC)].
0 I (A LC)
hence
A + BK BK
Spec = Spec[A + BK] Spec[A LC].
0 A LC
In summary, the 2n eigenvalues of the closed loop system coincide with the n eigenvalues of the
system with perfect state feedback and the n eigenvalues of the state estimation error. This fact
greatly simplifies the design of output feedback controllers for linear systems. One can design the
observer and the state feedback gain matrix separately, put them together, and the resulting closed
loop system will work as planned. This separation of the controller from the observer design, known
as the principle of separation, is a fundamental property of linear systems. Unfortunately, even
though related principles have been established for other classes of systems, this nice property does
not always hold.
From the above discussion it appears that, assuming that the system is controllable and observable,
the eigenvalues of the closed loop system can be arbitrarily placed. Therefore in principle one can
make the closed loop system respond arbitrarily fast, by making the real part of all its eigenvalues
sufficiently negative. There are several good practical reasons, however, why this temptation should
be resisted. The most important is modeling inaccuracies. Any model of a system used for design
or analysis is an approximate mathematical abstraction of a physical process; in fact a linear model
is usually a rather crude approximation of reality, valid only for certain values of the state and
input.The real physical process is bound to exhibit nonlinearities, additional dynamics, noise in the
sensors and actuators, delays, and other phenomena not adequately captured by the model. Faster
response, small eigenvalues and high gains tend to make the closed loop system more sensitive to all
these unmodeled factors.
In the end, the choice of the eigenvalues of the closed loop system (and hence the gains of the
controller and observer) usually comes down to a trade-off between several such considerations.
For linear systems, methods that in some cases allow one to establish optimal trade-offs have been
developed, in areas such as optimal filtering, linear quadratic Gaussian control, robust control, etc.
Very often, however, insight, intuition, and trial-and-error play a central role.
9.5 The multi-input, multi-output case

Finally, we turn our attention to systems with multiple inputs (m > 1) and/or multiple outputs
(p > 1). The situation is very similar in this case. Careful consideration of the development in
the previous sections indeed reveals that the only place where the single input (respectively single
output) assumption is used is in the development of the controllable canonical form in Theorem 9.2
(respectively observable canonical form in Theorem 9.4). This in turn is used in the only if part
of Theorem 9.3, to establish that if the system is controllable then its eigenvalues can be moved
to arbitrary locations by state feedback) (respectively Theorem 9.5). The rest of the argument
(including the entire discussion in Section 9.4) remains unaffected, even if m > 1, or p > 1. In this
section we will discuss how the missing parts of Theorem 9.2 and Theorem 9.3 can be filled in for
multi-input systems; the discussion of observer design for multi-output systems is dual and will be
omitted.
nn
Consider again system
AR
(9.5) with and B Rnm , possibly with m > 1. Consider the
columns of B = b1 . . . bm and assume that they are linearly independent. This assumption
can be made without loss of generality. Indeed, if the columns of B are linearly dependent there
exists a change of coordinates in Rm u(t) = T u(t) with T Rmn invertible such that for B = B T 1
and an appropriate partition of u(t):

u1 (t)
Bu(t) = B u = B1 0
u2 (t)
and the columns of B1 are linearly independent; 0 denotes a matrix of appropriate dimensions all
of whose elements are equal to zero. Hence in the new coordinates the inputs u2 (t) are redundant
and we can work with just the reduced system
x(t) = Ax(t) + B1 u1 (t)
where the columns of B1 are linearly independent.

Assuming that the system is controllable, the matrix

P = B AB . . . An1 B Rnnm
has rank equal to n. As for the single input case, the idea is to use this matrix to generate a change of
coordinates, x(t) = T x(t) for some T Rnn invertible, so that the resulting matrices A = T AT 1
and B = T B are such that one can easily design feedback controllers for them. Since the matrix
has rank n, out of its nm columns,
{b1 , . . . , bm , Ab1 , . . . , Abm , . . . , An1 b1 , . . . , An1 bm }
we can select n linear independent ones. The way we do the selection is not particularly important.
Different choices will lead to somewhat different structures in the matrices A and B, hence the
resulting form we obtain in the end will be somewhat less canonical than the controllable canonical
form is for single input systems. The most common choice is to start selecting the columns from
the left, skipping any columns that are linearly dependent on the previously selected ones, until
n linearly independent columns have been selected. Since we have assumed that the columns of
B are linearly independent, b1 , . . ., bm will be among the selected columns. In addition, for each
j = 1, . . . , m there is a maximal integer kj = 1, . . . , n such that Akj 1 bj will be among the selected
columns. Or, equivalently, a minimal integer such that Akj bj can be written as a linear combination
of the previously selected columns.
Exercise 9.3 Show that if Akj 1 bj is among the selected columns for some kj then so is Akj i bj
for all i = 1, . . . , kj . Or, equivalently, that if Akj bj can be written as a linear combination of the
previously selected columns then so can Ai bj for all i kj .
In summary, the selected columns will come in chains of the form
b1 , . . . , Ak1 1 b1
b2 , . . . , Ak2 1 b2
..
.
bm , . . . , Akm 1 bm .
The integers k1 , . . . , km are known as the controllability indices of the system.
Exercise 9.4 Show that the sum of the controllability indices is k1 + . . . + km = n.

We will assume that the controllability indices come in increasing order k1 k2 . . . km ; if not,
reorder the columns of B and P below so that this is the case.
Arrange the selected columns in a new matrix

P = b1 . . . Ak1 1 b1 b2 . . . Ak2 1 b2 . . . bm . . . Akm 1 bm Rnn .
Note that P is invertible since, by construction, its columns are linearly independent. Compute the
inverse of P and consider its rows

p1
p2

P 1 = . Rnn , p1 , . . . , pn R1n .
..
pn
Among these rows select the ones corresponding to the controllability indices pk1 , pk1 +k2 , . . . ,
pk1 +...+km = pn and form the matrix

pk1
pk1 A

..
.

pk1 Ak1 1

..
T = . R(k1 +...+km )n = Rnn .

pn

p A
n
.
..
km 1
pn A
Lemma 9.2 pk1 Ak1 1 b1 = pk1 +k2 Ak2 1 b2 = . . . = pn Akm 1 bm = 1. Moreover, pk1 Ai b1 = 0 for
i = 0, . . . , k1 2 and pk1 Ai bj = 0 for j = 2, . . . , m and i = 0, . . . , kj 1. Likewise, pk1 +k2 Ai b2 = 0
for i = 0, . . . , k2 2 and pk1 +k2 Ai bj = 0 for j = 1, 3, . . . , m and i = 0, . . . , kj , etc. The matrix T is
invertible.
Proof: By definition

p1

P 1 P = ... b1 . . . Ak1 1 b1 . . . bm . . . Akm 1 bm
pn

p1 b1 . . . p1 Ak1 1 b1 . . . p1 Akm 1 bm
.. .. .. .. ..
. . . . .

= pk1 b1 . . . pk1 Ak1 1
b1 . . . pk1 A km 1
bm
= I.
. . . . .
.. .. .. .. ..
k1 1 km 1
pn b 1 . . . pn A b 1 . . . pn A bm
Therefore, the elements on the diagonal are all equal to 1; in particular,
pk1 Ak1 1 b1 = pk1 +k2 Ak2 1 b2 = . . . = pn Akm 1 bm = 1.
Likewise, the off-diagonal elements are all equal to 0; in particular
pk1 b1 = . . . = pk1 Ak1 2 b1 = pk1 b2 = . . . = pk1 Ak2 1 b2 = . . . = pk1 bm = . . . = pk1 Akm 1 bm = 0
..
.
pn b1 = . . . = pn Ak1 1 b1 = pn b2 = . . . = pn Ak2 1 b2 = . . . = pn bm = . . . = pn Akm 2 bm = 0.
To see that the matrix T is invertible, consider

pk1
..
.

pk1 Ak1 1

..
T P = . b1 . . . Ak1 1 b1 . . . bm . . . Akm 1 bm

pn

.
..
pn Akm 1

pk1 b1 ... pk1 Ak1 1 b1 ... pk1 Akm 1 bm
pk1 Ab1 ... pk1 Ak1 b1 ... pk1 Akm bm

.. .. .. .. ..
. . . . .
P11 ... P1m
pk1 Ak1 1 b1 ... pk1 A2(k1 2) b1 . . . pk1 Ak1 +km 2 bm
P21 ... P2m
.. .. .. .. ..
= . . . . . = .. .. .. .
. . .
p b1 ... pn Ak1 1 b1 ... pn Akm 1 bm
n Pm1 . . . Pmm
p Ab ... pn Ak1 b1 ... pn Akm bm
n 1
. .. .. .. ..
.
. . . . .
km 1
pn A b1 . . . pn Ak1 +km 2) b1 ... pn A2(km 1) bm
In the last row we have decomposed the n n matrix into blocks Pij Rki kj . By the first part of
the lemma, every block of the form Pii , for i = 1, . . . , m has 1 on its anti-diagonal (elements (1, ki ),
(2, ki 1), . . ., (ki , 1)) and 0 everywhere else. Every block of the form Pij for i 6= j, on the other hand
has all its elements equal to zero. Therefore the last matrix is invertible (indeed, its determinant is
equal to either +1 or to 1). Since Det[T P ] = Det[T ]Det[P ] and P is invertible by construction
(recall that its columns were selected to be linearly independent), we must have that Det[T ] 6= 0.
Hence the matrix T is invertible.
Consider now the coordinate transformation x(t) = T x(t). It is easier to think of the matrices
A = T AT 1 and B = T B as decomposed into blocks

A11 . . . A1m B1
.. and B = .. ,
A = ... ..
. . .
Am1 . . . Amm Bm
with Aij Rki kj and Bi Rki m , i, j = 1, . . . , m.

Let us first turn our attention to the matrix B = T B:

pk1 pk1 b1 ... pk1 bm
.. .. .. ..
. . . .

pk1 Ak1 1 pk1 Ak1 1 b1 . . . pk1 Ak1 1 bm

.. .. .. ..
B = . b 1 . . . b m = . . . .

p p b1 ... pn b m
n n
.. .. .. ..
. . . .
pn Akm 1 pn Akm 1 b1 . . . pk1 Ak1 1 bm
Scanning the elements of B and comparing to Lemma 9.2 it becomes apparent that all of them are
equal to 0, except one element in each column, which is equal to 1, namely
bk1 1 = b(k1 +k2 )2 = . . . = bnm = 1.

For the matrix A, the fact that AT = T A implies that

pk1 pk1
.. ..
. .

A11 . . . A1m pk1 Ak1 1 pk1 Ak1 1

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

Am1 . . . Amm pn pn

.. ..
. .
km 1 km 1
pn A pn A

pk1 pn pk1 A
A ..
+ . . . + A
.. ..
11 . 1m . .

p A k1 1
p A km 1 k1
k1 n k1 A
p
.. ..
. = . .

p p pn A
k 1 n
.. .. ..
Am1 . + . . . + Amm . .
pk1 Ak1 1 pn Akm 1 pn Akm
Since the matrix T is invertible its rows (which can be thought of as the vectors pTk1 , . . ., (pk1 Ak1 1 )T ,
. . ., (pn Akm 1 )T ) are linearly independent. Reading the rows of the matrix one by one and equating
their coefficients on the left and on the right shows that:
The first row of A11 will have its second entry equal to 1 (the coefficient of pk1 A), and all other
entries equal to 0 (the coefficients of pk1 , pk1 A2 etc.).
The second row of A11 will have its third entry equal to 1 (the coefficient of pk1 A2 ) and all
other entries equal to 0, etc. until,
row k1 1 of A11 , which will have its last entry equal to 1 and all other entries equal to 0.
Likewise,
the first row of A12 will have all its entries equal to 0 (coefficients of pk1 +k2 , . . ., pk1 +k2 Ak2 1 ),
etc. until,
Row k1 1 of A1m , which will have all its entries equal to 0.
Row k1 is equal to pk1 Ak1 and (since the rows of T are linearly independent and span R1n ) can be
written as a linear combination
pk1 Ak1 = ak1 1 pk1 + ak1 2 pk1 A + . . . ak1 n pn Akm 1
for some ak1 1 , . . ., ak1 n R; we will not bother too much with the exact form of these coefficients.
Repeating the process for the remaining blocks shows that in the new coordinates the system matrices
A and B become respectively

0 1 ... 0 0 0 ... 0 0 ... 0
.. .. .. .. .. .. .. .. .. . . .

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

0 0 ... 1 0 0 ... 0 0 ... 0

ak1 1 ak1 2 . . . ak1 k1 ak1 (nkm +1) ak1 (nkm +2) . . . ak1 n
1 ... 0
.. .. . . .
. . .. , .. . . . .. .

0 0 ... 0 0 1 ... 0 0 ... 0

.. .. .. .. .. .. .. .. . . ..
. . . . . . . . . . . .
... .
0 0 ... 0 0 0 ... 1 0 ... 0
an1 an2 . . . ank1 an(nkm +1) an(nkm +2) . . . ann 0 ... 1
This special form is known as the Brunowski normal form and it the analogue of the controllable
canonical form for multi-input systems. We note, however, that, unlike the controllable canonical
form, the Brunowski normal form is not unique, since it depends on the choice of columns of the
matrix P , the way these columns are rearranged, etc.
How does the Brunowski normal form help us to design controllers for the system? Assume that we
are given a complex conjugate set of desired eigenvalues {1 , . . . , n }. As in the single input case
we would like to select a linear state feedback gain matrix K Rmn so that under the feedback
u(t) = Kx(t) + r(t)
the closed loop matrix A+ BK has the desired eigenvalues. To see that this is possible first bring the
system in Brunowski normal form through a coordinate transformation x(t) = T x(t) and determine
the controllability indices k1 , . . ., km . Write also the state feedback in the new coordinates
u(t) = KT 1 T x(t) + r(t) = K x(t) + r(t)
and consider the rows of K

K1 k11 . . . k1n
.. .
K = ... = ... ..
. .
Km km1 . . . kmn
Split the set of desired eigenvalues into m subsets {1 , . . . , k1 }, . . ., {nkm +1 , . . . , n }. For sim-
plicity assume that this can be done so that each of the subsets is itself complex conjugate. Form
the m characteristic polynomials
( 1 ) . . . ( k1 ) = k1 + d11 k1 1 + . . . + d1k1
...
( nkm +1 ) . . . ( n ) = km + dm1 km 1 + . . . + dmkm .
The idea is to use each row Ki of the feedback matrix to ensure that the subsystem Aii has the
corresponding characteristic polynomial. To do this select
k11 = (ak1 1 + d1k1 ), . . . , k1k1 = (ak1 k1 + d11 ), . . . , k1(k1 +1) = ak1 (k1 +1) , . . . , k1n = ak1 n
...
km1 = an1 , . . . , km(nkm ) = an(nkm ) , . . . , kmn = (ak1 n + dm1 ).
The elements of the off-diagonal blocks Aij for i 6= j are then eliminated and the resulting closed
loop system matrix becomes

0 1 ... 0 0 0 ... 0
.. .. . . .. .. .. .. .

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

0 0 . . . 1 0 0 ... 0

d1k1 d1(k1 1) . . . d11 0 0 ... 0

.. .. ..
A + B K = . . . .

0 0 ... 0 0 1 ... 0

.
.. .. . . . ...
. ..
.
..
.
..
.
..
.
...
0 0 ... 0 0 0 ... 1
0 0 ... 0 dmkm dm(km 1) . . . dm1
Notice that the matrix is block-diagonal and that each of the diagonal blocks is in controllable
canonical form. Therefore Det[I (A + B K)] decomposes into

1 1 ... 0 1 1 ... 0
.. .. . . .. .. .. .. ..
. . .
Det . . . . . Det . . .
0 0 . . . 1 0 0 . . . 1
d1k1 d1(k1 1) . . . d11 dmkm dm(km 1) . . . dm1
= (k1 + d11 k1 1 + . . . + d1k1 ) . . . (k1 + dm1 k1 1 + . . . + dmkm )
= ( 1 ) . . . ( k1 ) . . . ( nkm +1 ) . . . ( n )
which is the desired characteristic polynomial. Setting K = KT completes the design (recall that
coordinate transformations do not affect the eigenvalues by Theorem 9.1).
As one can imagine from the above discussion, bringing a system into Brunowski normal form can be
rather tedious. Fortunately this is not necessary if the end goal is to design feedback controllers for
the system. As for single input systems, to do this it suffices to form the characteristic polynomial
of the matrix (A + BK), treating the elements of the feedback gain matrix K Rmn as unknowns.
Equating the coefficients of this polynomial to those of the desired characteristic polynomial (
1 ) . . . ( n ) leads to a system of n linear equations with the nm elements of K as unknowns.
The Brunowski normal form construction guarantees that if the system is controllable this system
of equations will have a solution; for multi-input systems it will in fact have an infinite number of
solutions, since the number of equations is smaller than the number of unknowns. If the system
is not controllable but merely stabilizable then one has to make sure that the uncontrollable (but
stable) modes are included in the set of desired eigenvalues and repeat the process.
In a similar way one can also construct observers and implement output feedback controllers for the
system. The construction is just dual to that of the Brunowski normal form and will not be given
in detail.

Problem 9.1 (Pole placement) Consider the system

0 0 2 3
x(t) = 1 0 0 x(t) + 0 u(t) = Ax(t) + Bu(t),
0 2 1 0

y(t) = 0 0 2 x(t) = Cx(t). ()
1. Is the system observable? Is it controllable? Justify your answer in each case.

2. Design a state feedback u = Kx such that the closed loop system has three poles at s = 2.
3. Recall that a state-observer provides an estimate x(t) of the state x(t) of () by means of the
differential equation

x(t) = Ax(t) + Bu(t) + L y(t) y(t) ,
y(t) = C x(t).
Compute the observer gain L such that the dynamics of the error e := x x have three poles
at s = 3.
Problem 9.2 (Controllable canonical form) Use the procedure of Section 9.5 to re-derive the
controllable canonical form for single input systems. Explain how the coefficients in the last row of
matrix A arise.
Problem 9.3 (Controllability, observability and state feedback) Consider the linear state
feedback arrangement of Figure 9.1.
1. Show that the closed loop system is controllable if and only if the open loop system is control-
lable.
2. Assume now that the open loop system is observable. Is it true that the closed loop system
will always be observable? Provide a proof, or a counter-example.
Problem 9.4 (Controllability, observability and observer feedback) Consider the observer
feedback arrangement of Figure 9.3. Assume that the open loop system is observable and control-
lable. Will the closed loop system always be observable? Will it be controllable? In both cases
provide either a proof, or a counter-example.
Appendix A
Class composition and outline
A.1 Class composition

1. How many PhD students?
2. How many Masters students?
3. Other?
1. How many D-ITET?

2. How many D-MAVT?
3. Other?
A.2 Class outline

Topics:
1. Introduction: Notation and proof technique

2. Algebra: Rings, fields, linear spaces
3. Analysis: ODE, existence and uniqueness of solutions
4. Time domain solutions of linear ODE, stability
5. Controllability, observability, Kalman decomposition
6. Observers, state feedback, separation principle
7. Realization theory
Coverage:
Linear, continuous time, time varying systems

Time invariant systems treated as a special case
Discrete time treated sporadically
149
No nonlinear systems
No discrete event or hybrid systems
No infinite dimensional systems (no PDEs, occasional mention of delays)
Logistics:
Lecture: ETZ E7, Mondays 9:00-12:00, John Lygeros.

Examples classes: ETZ E7, alternate Thursdays, 17:00-19:00, Federico Ramponi (Head), De-
basish Chatterjee, and Stefan Almer.
6 examples papers, 25% of the final grade
90 minute written exam, 75% of the final grade
Material and style:
Hand written notes

Books:
Callier and Desoer
Michel and Antsaklis
Kailath
Excuse hand writing
Ask questions and answer questions!
Try to be rigorous
Formal statements, definitions, proofs.
Rather mathematical, not a lot of motivation!
Appendix B
Notation (C&D A1)
B.1 Shorthands
Def. = Definition
Thm. = Theorem
Ex./ = Example or exercise
iff = if and only if
wrt = with respect to
wlog = without loss of generality
ftsoc = for the sake of contradiction
= therefore
= contradiction
: = such that
= Q.E.D. (quod erat demonstrandum)
B.2 Sets
, 6, , 6, , ,
For A, B X, Ac stands for the complement of a set and \ for set difference, i.e. Ac = X \ A
and A \ B = A B c .
R real numbers, R+ non-negative real numbers, [a, b], [a, b), (a, b], (a, b) intervals
Q rational numbers
Z integers, N non negative integers (natural numbers)
C complex numbers, C+ complex number with non-negative real part
Sets usually defined through their properties as in: For a < b real
[a, b] = {x Rn | a x b}, or [a, b) = {x Rn | a x < b}, etc.
151
Cartesian products of sets: X, Y two set, X Y is the set of ordered pairs (x, y) such that
x X and y Y .
Example (Rn ) Rn = R R . . . R (n times).

x1
x2

x Rn x = (x1 , x2 , . . . , xn ) = .. , with x1 , . . . , xn R
.
xn
B.3 Logic
, , , , .
! = exists unique.
= and
= or
= not
Exercise B.1 Is the statement (! x R : x2 = 1) true or false? What about the statement
( x R : x2 = 1)?
Appendix C
Basic linear algebra
153
Bibliography
154

Lecture Notes On Linear System Theory

Uploaded by

Copyright:

Available Formats

Lecture Notes On Linear System Theory

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Lecture Notes On Linear System Theory

Uploaded by

Copyright:

Available Formats

What topics are covered in the lecture notes?

What topics are covered in the lecture notes?

What mathematical areas are the basis for linear systems theory?

What mathematical areas are the basis for linear systems theory?

Lecture Notes on Linear System Theory

John Lygeros and Federico Ramponi

Department of Information Engineering, University of Brescia

September 22, 2013

4 Time varying linear systems: Solutions 58

5 Time invariant linear systems: Solutions and transfer functions 69

7 Inner product spaces 99

8 Controllability and observability 109

9 State Feedback and Observer Design 130

B Notation (C&D A1) 151

C Basic linear algebra 153

1.1 Objectives of the course

x(t) = A(t)x(t) + B(t)u(t),

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

x(t) = Ax(t) + Bu(t) (1.1)

x(0) = x0 and x(t) = Ax(t) + Bu(t), for all t [0, T ]. (1.2)

1.2 Proof methods

(p implies q). p is called the hypothesis and q the consequence.

If there is smoke then there is a fire

If X is a Greek then X is a European

Theorem 1.2 If n is odd then n2 is odd.

Figure 1.1: Set theoretic interpretation of logical implication.

Theorem 1.3 For all m, k N, (m + k)! m!k!.

which completes the proof.

If X is not a European then X is not a Greek

In turn, this is equivalent to the set theoretic statement

People \ Europeans = Europeansc Greeksc = People \ Greeks.

A proof where we show that p q is true by showing q p is true is known as a proof by

Theorem 1.4 If n2 is odd then n is odd.

We show by contradiction that every Greek is not a Chinese, more formally

If X is a Greek then X is not a Chinese

On the other hand, by deduction,

(Greeks Chinese) (Europeans non-Europeans)

by the resulting numbers. Now

Therefore, n and m are both even and, according to Definition

Theorem 1.6 n2 is odd if and only if n is odd.

or, in other words, X A X B.

Theorem 1.7 Let N N be the largest natural number. Then N = 1.

Figure 1.2: Commutative diagram of function composition.

1.3 Functions and maps

is called the range of f (sometimes denoted by f (X)) and the set

is called the graph of f .

Definition 1.4 A function f : X Y is called:

1. Injective (or one-to-one) if and only if f (x1 ) = f (x2 ) implies that x1 = x2 .

Given two functions g : X Y and f : Y Z their composition is the function (f g) : X Z

Exercise 1.5 Show that the identity map is bijective.

Definition 1.6 Consider a function f : X Y .

1. The function gL : Y X is called a left inverse of f if and only if gL f = 1X .

f is called invertible if an inverse of f exists.

Fact 1.1 If f is invertible then its inverse is unique.

x1 = g1 (y) = 1X g1 (y) = (g2 f ) g1 (y) (g2 inverse of f )

This contradicts the assumption that x1 6= x2 .

Problems for chapter 1

1. f has a left inverse if and only if it is injective.