Download as pdf or txt
Download as pdf or txt
You are on page 1of 39

Basic Concepts from Linear

Systems’ Theory and Control


A preparatory recapitulation (not only)
for the module Advanced Control
Winter Semester 2021/22

October 3, 2021

© PD Dr.-Ing. habil. Paul Kotyczka


Chair of Automatic Control
Technical University of Munich
Preface

The intention of this document to give you a brief recapitulation of the terms and concepts,
which are the prerequisites for the module Advanced Control. The document is for self-study.
It can be used, together with the corresponding teaching videos, to refresh basic concepts, if
necessary, before and while attending the course Advanced Control.
The document contains a Chapter Zero, which covers elementary properties of linear time-
invariant systems from linear algebra, differential equations and control. Most of the material
will be familiar to readers which have finished a Bachelor in Mechanical or Electrical Engineer-
ing, which contained a basic control module as well as a course on linear systems’ theory.
Only the topic of the last section, invariant zeros, might be not so familiar to some of you, as
this generalized definition of zeros comes up for systems with multiple in- and outputs (MIMO
systems). I encourage you to study this section to understand

• why invariant zeros are defined via the Rosenbrock system matrix,

• that pole-zero cancellations (compensation) are more delicate in the MIMO case, and
that

• transmission zeros form a subset of the invariant zeros.

In Advanced Control we will encounter invariant zeros mainly when a control approach involves
system inversion. In these cases, the locations of the invariant zeros are crucial to guarantee
stability of the closed-loop system.
Don’t hesitate to ask questions concerning the presented basic concepts

• via the forum on the learning platform moodle and

• during the weekly revision exercises.

I wish you a lot of pleasure with this recapitulation on linear systems!

Garching, in October 2021


Paul Kotyczka

3
4 © PD Dr.-Ing. habil. Paul Kotyczka
Chair of Automatic Control
Technical University of Munich
Contents

0 Linear Systems in Time and Frequency Domain 7


0.1 State Space Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
0.1.1 Modeling Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
0.1.2 Nonlinear State Space Representation . . . . . . . . . . . . . . . . . . 12
0.1.3 Linearization at an Equilibrium . . . . . . . . . . . . . . . . . . . . . . 13
0.1.4 State Space Representation of Linear Systems . . . . . . . . . . . . . . 15
0.2 Solution of the State Differential Equation . . . . . . . . . . . . . . . . . . . . 16
0.2.1 General Solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
0.2.2 Eigenvalues and Eigenvectors . . . . . . . . . . . . . . . . . . . . . . . 20
0.2.3 State Space Model in Diagonal Form . . . . . . . . . . . . . . . . . . . 22
0.3 System Representations in Frequency Domain . . . . . . . . . . . . . . . . . . 25
0.3.1 Laplace Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
0.3.2 Transfer Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
0.3.3 Poles and Zeros in SISO Systems . . . . . . . . . . . . . . . . . . . . . 28
0.3.4 Impulse and Step Response . . . . . . . . . . . . . . . . . . . . . . . . 29
0.3.5 Initial and Final Value Theorem . . . . . . . . . . . . . . . . . . . . . 31
0.4 Illustration of Controllability and Observability . . . . . . . . . . . . . . . . . . 31
0.5 Zeros of MIMO Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
0.5.1 Blocking Property and Rosenbrock Matrix . . . . . . . . . . . . . . . . 33
0.5.2 Invariant and Transmission Zeros . . . . . . . . . . . . . . . . . . . . . 34
0.5.3 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

5
Contents

6 © PD Dr.-Ing. habil. Paul Kotyczka


Chair of Automatic Control
Technical University of Munich
Chapter 0

Linear Systems in Time and


Frequency Domain

We start with the question how linear state space models are generated in Section 0.1 and
then discuss the solutions of the state differential equations in Section 0.2. In Section 0.3, we
establish the link to the frequency domain. Transfer functions and their discussion are typically
topic in introductory control courses. Based on the system description in diagonal/modal form,
we illustrate the concepts of controllability and observability in Section 0.4. This document
closes in Section 0.5 with the introduction and discussion of invariant zeros, which generalize
the notion of zeros from the SISO to the MIMO case1 .

0.1 State Space Models

The first step to obtain state space models is mathematical modeling, which means the tran-
scription of the physical behavior of a system – with an accuracy or modeling depth that
depends typically on the requirements on the model – to the language of differential equa-
tions. We deal with dynamical systems, which means that the system quantities evolve with
respect to time. If the spatial extent of the system properties does not play a role, and the
spatial distribution can be “lumped” into concentrated objects2 , we call such systems lumped
parameter dynamical systems, which are described by ordinary differential equations in time3 .
State variables are system quantities, whose time derivatives appear in the state differential
equations. A candidate for a state variable is a physical quantity that describes the content of
energy stored in a part of the system (a subsystem). The modeling examples in the following
subsection, and in the introductory exercise, illustrate the process of setting up the state
differential equations4 .
Typically, the modeling step results in a systems of nonlinear differential equations. Examples
1
SISO is short for “single input, single output”, while MIMO means “multiple input, multiple output.
2
Example: When we describe the translational motion of a rigid body, we can treat it as a point mass.
3
If the evolution of the system variables is governed also by spatial derivatives, then we have a spatially
distributed or distributed parameter system, which is described by partial differential equations.
4
For an energy-based, systematic way to model complex, multi-domain systems, please refer to the module
Modeling and Reduction of Complex Systems, which is taught in summer semester.

7
0.1. State Space Models

Figure 1: Magnetic levitation system and approximate inductance function over the distance
s.

for sources of nonlinearity are the kinematics in mechanical systems5 or nonlinear constitutive
equations or material laws6 . To obtain a linear state representation, which approximates
the nonlinear system behavior, the nonlinear functions in the state differential equations are
linearized. This happens in general at an equilibrium state of the system, i.e., at a state, where
all time derivatives are zero, and the system is at rest. The technique of linearization at an
equilibrium is summarized in the second subsection, which leads to the definition of linear state
space models at the end of this section.

0.1.1 Modeling Examples

Mathematical models can be derived in different ways. In electrical systems, typically the
network equations (Kirchhoff’s laws) are combined with the element equations, which char-
acterize the behavior of each component of the system. This approach usually directly leads
to first order differential equations. In mechanics, one way to obtain dynamical models is the
Euler-Lagrange formalism. It is based on Hamilton’s principle of stationary action and leads
to a set of second order differential equations. The following two nonlinear examples illustrate
the modeling process.

Magnetic Levitation System

We consider the magnetic levitation system depicted in Fig. 1 on the left. An iron ball shall
be lifted using an electromagnet and kept at a desired distance y ∗ = s∗ . The control input is
the terminal voltage in the electric circuit u(t) = vin (t). The inductance of the electromagnet
depends on the distance of the iron ball, which acts like an additional iron core, from the
electromagnet. L(s) has its maximum if the ball is attached to the magnet and its minimum
if the ball is completely removed. On the right in Fig. 1, the assumed inductance function
L(s) = ks , k > 0, is sketched, which represents a reasonably good approximation in the range
of distances s ∈ [smin , smax ].
5
Think of the trigonometric functions to describe the Cartesian coordinates of a robot tool center point
depending on the joint angles.
6
Think of the saturation of the current-flux characteristics of an electromagnet.

8 © PD Dr.-Ing. habil. Paul Kotyczka


Chair of Automatic Control
Technical University of Munich
CHAPTER 0. LINEAR SYSTEMS IN TIME AND FREQUENCY DOMAIN

The magnetic levitation system is an example for an electromechanical system, i.e., we have
to establish the electric equations, the mechanical equations and their coupling. We choose
as state variables x1 = s, x2 = ṡ and x3 = i, i.e., the distance and velocity of the ball (for the
mechanical part) and the current through the electromagnet (for the electrical part).
Mechanical equations. The first state differential equation

ẋ1 = x2 (1)

is trivial, as it only states that the time derivative of the distance x1 = s is defined as the
second state variable x2 = ṡ. The second differential equation comes from Newton’s second
law:
ms̈ = mg − Fmag , (2)
where Fmag is the attractive magnetic force, which always points into the direction of the
material with the lower magnetic permeability (here: air). An expression for the magnetic force
can be derived as follows, with the help of the two diagrams in Fig. 2. Consider a constant
current i0 through the coil, and assume that the ball is at a distance s0 . Assume furthermore
that under the action of the magnetic force Fmag , the ball is displaced by an increment ∆s < 0
towards the magnet. With the increased inductance L(s1 ) = L(s0 + ∆s) > L(s0 ) (don’t
forget that ∆s is negative), Figure 2 illustrates that the amount of energy of the magnetic
field increases:
∆Wmag = − Fmag |{z} ∆s > 0. (3)
| {z }
>0 <0

Letting ∆s → 0, and for arbitrary current i, we can write

∂ ∗
Fmag = − W (i, s), (4)
∂s mag

where Wmag
∗ (i, s) = 1 L(s)i2 is the magnetic (co-)energy at a given current.
2

Remark 0.1. While magnetic energy is the area left of the curve in the diagrams of Fig. 2,
and is parametrized by the magnetic flux linkage, Wmag (ψ, s), magnetic co-energy is the
∗ (i, s). For linear magnetic
area under the curve and expressed in terms of the current, Wmag
material, both areas are equal: Wmag (L(s)i, s) = Wmag
∗ (i, s) = 1 L(s)i2 .
2

With the ansatz L(s) = ks , we get

1 ∂ k 2
 
Fmag =− i
2 ∂s s
ki2
= 2 (5)
2s
kx2
= 23 ,
2x1

and (2) can be written as a first order differential equation in the chosen state variables:

kx23
ẋ2 = g − . (6)
2mx21

© PD Dr.-Ing. habil. Paul Kotyczka 9


Chair of Automatic Control
Technical University of Munich
0.1. State Space Models

Figure 2: Magnetic energy and co-energy when the distance of the ball is reduced from s0 to
s1 = s0 − ∆s at constant current i0 . The magnetic flux linkage at the same time increases
from ψ0 = L(s0 )i0 to ψ1 = L(s1 )i0 = ψ0 + ∆ψ.

Electrical equation. For the differential equation of the electrical circuit we use Kirchhoff’s
voltage law, Ohm’s law and Faraday’s law. The sum of voltages in the electrical circuit amounts
for zero, the voltage across the resistor is proportional to the current, and the voltage across
the coil equals the rate of change of flux linkage:

−vin + vR + vL = 0
(7)
−vin + Ri + ψ̇ = 0.
∂L(s)
With ψ = L(s)i and ψ̇ = ∂s ṡi + L(s)i̇, we can establish the differential equation for the
current:
1 ∂L(s)
 
i̇ = −Ri − ṡi + vin , (8)
L(s) ∂s
or in terms of the chosen states and input, and after substitution of all expressions:
Rx1 x3 x2 x3 x1
ẋ3 = − + + u. (9)
k x1 k

Together with the definition of the distance as output variable, we obtain the nonlinear state
space representation of the magnetic levitation system:
 
x2
2
k x3 (10a)
 
ẋ = 
 g− 2m x21


x2 x3
−R
k x1 x3 +
1
x1 + k x 1 u
y = x1 . (10b)

Pendulum on a Cart

For our second example we consider the pendulum on a cart, which is depicted in Fig. 3. The
cart hass mass M , and a mathematical pendulum is mounted on its top. Mathematical means
here that the rod of length L is assumed massless, and the mass of the rod is concentrated in the
point mass m at its tip. The cart is driven by a horizontal force F , which is our control input.
We neglect friction and set up the equations of motion using the Euler-Lagrange formalism.

10 © PD Dr.-Ing. habil. Paul Kotyczka


Chair of Automatic Control
Technical University of Munich
CHAPTER 0. LINEAR SYSTEMS IN TIME AND FREQUENCY DOMAIN

Figure 3: Mathematical pendulum on a cart.

Degrees of freedom. The motion of the system can be fully described by the coordinates and
velocities for the two mechanical degrees of freedom. The translational motion is parametrized
by the cart position s and its velocity ṡ, while the rotational motion is described by the rod
angle φ and its angular velocity φ̇.
Energies and Lagrangian. To set up the second order Euler-Lagrange equations, we have to
express the Lagrangian
L(q, q̇) = T (q, q̇) − V (q), (11)
which is the difference between kinetic (co-)energy T and potential energy V in terms of the
h iT h iT
mechanical coordinates q = s φ and their velocities q̇ = ṡ φ̇ .

Remark 0.2. In classical (non-relativistic) mechanics, we do not need to distinguish between


kinetic energy and co-energy, and we set up the kinetic energy in terms of the velocities.

The translational and rotational kinetic energy are together


1 1
T (ṡ, φ̇) = (M + m)ṡ2 + mL2 φ̇2 − mL cos φṡφ̇, (12)
2 2
while the potential energy is
V (φ) = mLg cos φ. (13)

Euler-Lagrange equations. The Lagrangian is now substituted in the Euler Lagrange equations

d ∂L ∂L
 
− = fi , i = 1, . . . , N, (14)
dt ∂ q̇i ∂qi

for each of the N (here 2) degrees of freedom, where fi indicates the non-conservative (ex-
h iT
ternal) generalized forces. With f = u 0 , we obtain the set of second order differential
equations

(M + m)s̈ − mL cos φφ̈ + mL sin φφ̇2 = u (15a)


2
−mL cos φs̈ + mL φ̈ − mLg sin φ = 0. (15b)

© PD Dr.-Ing. habil. Paul Kotyczka 11


Chair of Automatic Control
Technical University of Munich
0.1. State Space Models

In matrix-vector form we can write

M (q)q̈ + C(q, q̇)q̇ + g(q) = Gu, (16)

where " #
M +m −mL cos φ
M (q) = (17)
−mL cos φ mL2
is the symmetric, positive definite mass or inertia matrix,
" # " #
mL sin φφ̇2 0
C(q, q̇)q̇ = , g(q) = (18)
0 −mLg sin φ
h iT
are the vectors of Coriolis and potential forces and G = 1 0 indicates that only the first
(the translational) degree of freedom is actuated.
First order state representation. Defining v = q̇ as the vector of generalized velocities, the
second order representation (16) can be cast to first order form

q̇ = v (19a)
−1 −1 −1
v̇ = −M (q)g(q) − M (q)C(q, v)v + M (q)Gu. (19b)

The nonlinear state representation is completed by defining an output y or output vector y.

Remark 0.3. To obtain a linear approximation of (19) (see the corresponding subsection below),
it is convenient for a mechanical system to first linearize the second order representation (16)
around some equilibrium configuration q ∗ , which goes along with zero velocity q̇ ∗ = 0. The
result is the representation of a linear mechanical system (the Coriolis term disappears) in the
form
M ∆q̈ + K∆q = G∆u (20)
with constant symmetric and positive (semi-)definite mass and stiffness matrices7 M and K.
To obtain a first order state differential equation, define ∆v = ∆q̇, and obtain
" # " #" # " #
∆q̇ 0 I ∆q 0
= + u. (21)
∆v̇ −M −1 K 0 ∆v M −1 G

0.1.2 Nonlinear State Space Representation

The first order differential equations for the two preceding examples, as well as the equations,
which express the controlled and/or measured quantities can be summarized on the state
(space) representation of the nonlinear system8 .

7
In the example of the linearized cart-pendulum system, K is only semi-definite, as the potential energy
does not depend on s.
8
We do not (yet) consider disturbances at this point. They are typically described by a disturbance input
vector z, which acts in the same way on the system equations as the control input vector. Therefore, the
disturbance input vector is treated, e.g., for the computation of the equilibrium and linearization as presented
below, in a complete analog manner.

12 © PD Dr.-Ing. habil. Paul Kotyczka


Chair of Automatic Control
Technical University of Munich
CHAPTER 0. LINEAR SYSTEMS IN TIME AND FREQUENCY DOMAIN

Definition 0.1 (Nonlinear state (space) representation). The (explicit) state representa-
tion of a (in general) nonlinear dynamical system consists of a vector-valued differential
equation, the state differential equation, as well as an algebraic output equation

ẋ(t) = f (x(t), u(t)), x(t0 ) = x0 (22)


y(t) = h(x(t), u(t)). (23)

Therein, x(t) ∈ Rn is the state vector, u(t) ∈ Rp is the input vector and the output
vector y(t) contains the controlled and/or measured variables.

The input vector u(t) contains the variables at our disposition, in general actuator signals,
while the output vector y(t) contains the variables we want to manipulate/control and/or
which are accessible by measurement. A usual assumption is that the set of control variables
coincides with the set of measured variables. If this should not be the case, it is indicated.

The mappings f : Rn ×Rp → Rn and h : Rn ×Rp → Rq are assumed to be sufficiently smooth,


i.e., at least twice continuously differentiable in x and u. The reason is the validity of the
linear approximation for ∆x, ∆u → 0, which is explained in the next subsection. x(t0 ) = x0
is the initial value for the state differential equations, therefore (22) defines an initial value
problem (IVP). In terms of the scalar values components of the state, output and input vector,
state differential and output equation can be written

ẋi (t) = fi (x1 (t), . . . , xn (t), u1 (t), . . . , up (t)), i = 1, . . . , n, (24)


yj (t) = hj (x1 (t), . . . , xn (t), u1 (t), . . . , up (t)), j = 1, . . . , q. (25)

Remark 0.4. Knowing that all system quantities evolve with time t, we frequently omit their
time argument, so in the rest of this subsection.

0.1.3 Linearization at an Equilibrium

We present the technique of linearization at an equilibrium, which provides a linear approxi-


mation of the state representation (22), (23) in the neighborhood of an equilibrium (x∗ , u∗ )

Equilibrium

An equilibrium is given by the state vector x∗ and the input vector u∗ for which the time
derivatives in the state differential equation (22) are zero: ẋ∗ = 0. We determine the equilibria
(x∗ , u∗ ) of a dynamical system by solving the set of nonlinear algebraic equations

0 = f (x∗ , u∗ ). (26)

Note that a nonlinear system has in general multiple equilibria, i.e., multiple solutions of (26).
Substitution of x∗ and u∗ in the output equation (23) gives the corresponding equilibrium
value of the output y ∗ .

© PD Dr.-Ing. habil. Paul Kotyczka 13


Chair of Automatic Control
Technical University of Munich
0.1. State Space Models

Linearization

To analyze the qualitative and quantitative behavior of a nonlinear dynamical system (22), (23)
in the neighborhood9 of an equilibrium (x∗ , u∗ ) and also to design controllers in a systematic,
model-based way10 , it is convenient to consider the deviations from the equilibrium/operating
point

∆xi := xi − x∗i i = 1, . . . , n
∆uj := uj − u∗j j = 1, . . . , p (27)
∆yk := yk − yk∗ k = 1, . . . , q.

and to derive linear approximations of the state differential equation (22) and the output
equation (23). This is done by Taylor series expansion of the nonlinear functions f (x, u) and
h(x, u).
As ẋ∗i = 0 holds in the equilibrium, we have ∆ẋi = ẋi , and the Taylor series expansion of the
i-th ODE in (24) is11

∆ẋi = fi (x∗1 , . . . , x∗n , u∗1 , . . . , u∗p ) +


| {z }
=0
n p
X ∂fi (·) X ∂fi (·) (28)
+ ∆xj + ∆uk
j=1
∂xj ∗ k=1
∂uk ∗

+ higher order terms (h.o.t.)

We obtain, written in matrix-vector form, the linear approximation of (22):

∆ẋ = A∆x + B∆u (29)

with  ∂f1
 ∂f ∂f1
 ∂f1 
∂x1
1
··· ∂xn ∂u1 ··· ∂up
∂f  . .. ..  ∂f  . .. .. 
. . . . (30)
∂u ∗  ∂f. . 
A= = .  , B= =  .
∂x ∗  ∂f.

∂fn ∂fn
∂x1
n
··· ∂xn ∂u1
n
··· ∂up
(x∗ ,u∗ ) (x∗ ,u∗ )

We obtain, in complete analogy, the linearization of the output equation (23):

∆y = C∆x + D∆u (31)

with  ∂h1
 ∂h ∂h1 ∂h1 
···

1
∂x1 ··· ∂xn ∂u1 ∂up
∂h  . .. ..  ∂h  . .. .. 
. . . . (32)
∂x ∗  ∂h. q .  ∂u ∗  ∂h. q . 
C= =  , D= =  .
∂hq ∂hq
∂x1 ··· ∂xn (x∗ ,u∗ ) ∂u1 ··· ∂up (x∗ ,u∗ )

Back to the beginning of this subsection: Why do the mappings f (x, u) and h(x, u) need
to be twice differentiable? The reason is that the boundedness of the second order partial
This means for states x(t) and inputs u(t), which are not too far away from x∗ and u∗ . What not too far
9

means concretely, depends on the particular case and the nonlinearity of the system.
10
This is the topic of Advanced Control!
11
The ∗ indicates that the partial derivatives have to be evaluated in the equilibrium (x∗ , u∗ ).

14 © PD Dr.-Ing. habil. Paul Kotyczka


Chair of Automatic Control
Technical University of Munich
CHAPTER 0. LINEAR SYSTEMS IN TIME AND FREQUENCY DOMAIN

derivatives guarantees that the second order residual terms, after the cut in the Taylor series,
tend to 0 for ∆x → 0 and ∆u → 0.
In many cases, the linearization of a nonlinear system at a given equilibrium point reflects the
system dynamics in a sufficient way and linear state feedback methods, as presented in the
course Advanced Control, are very powerful tools for systematic control design. In the case that
a system features dominant nonlinear effects in the considered operation domain, nonlinear
control methods may be necessary for stabilization and satisfactory closed-loop behavior. An
introduction into nonlinear state feedback control is given in the course Nonlinear Control.

0.1.4 State Space Representation of Linear Systems

We summarize the structure of the linear state representation, which is the basis for the
analysis of linear systems as presented in the following sections. We restrict the attention
here to undisturbed systems, i.e. effects like noise ore other disturbing exogeneous inputs are
excluded. Their systematic treatment is a topic in Advanced Control.

Definition 0.2 (Linear time-invariant state space representation). The state space rep-
resentation of linear time-invariant (LTI) dynamical systems is given in matrix-vector
notation by

ẋ(t) = Ax(t) + Bu(t), x(t0 ) = x0


(33)
y(t) = Cx(t) + Du(t).

It consists of a state differential equation and an output equation.


Linear refers to the linearity of the mappings, represented by the matrices A, B, C and
D. Time-invariant means that these matrices are constant, and do not contain the time
as parameter.
x(t0 ) = x0 is the initial state, which together with the state differential equation forms
an intial value problem (IVP). The solution of the IVP gives the evolution of the system
quantities.

The state x(t) ∈ Rn , the control input u(t) ∈ Rp , and the control output y(t) ∈ Rq are
vector-valued signals that evolve in time:

x1 (t) u1 (t) y1 (t)


     
 ..   ..   .. 
x(t) =  .  , u(t) =  .  , y(t) =  .  . (34)
xn (t) up (t) yq (t)

We call
a11 ··· a1n
 
 .. .. ..  ∈ Rn×n
A= . . .  (35)
an1 ··· ann
the state matrix,
b11 · · · b1p
 
 .. .. ..  ∈ Rn×p
B= . . .  (36)
bn1 · · · bnp

© PD Dr.-Ing. habil. Paul Kotyczka 15


Chair of Automatic Control
Technical University of Munich
0.2. Solution of the State Differential Equation

the input matrix,


c11 · · · c1n
 
 .. .. ..  ∈ Rq×n
C= . . .  (37)
cq1 · · · cqn
the output matrix and
d11 · · · d1p
 
 .. .. ..  ∈ Rq×p
D= . . .  (38)
dq1 · · · dqp
the feedthrough matrix.
In many physical systems, D = 0, which is the case we will consider in the rest of this
document and in Advanced Control. Note, however, that the methods for analysis and control
design, which are presented/developed for D = 0 can easily be reformulated for D ̸= 0.

Remark 0.5. Without loss of generality, we typically consider zero as the initial time, i.e.,
t0 = 0.

0.2 Solution of the State Differential Equation

The solution of the state differential equation is derived based on variation of constants, first
for the scalar-valued case, then for the vector-valued case, taking into account the computation
rules for matrices and vectors. In a second step, we show how a canonical state representation
– the modal or diagonal form – can be derived using eigenvalues and eigenvectors of the state
matrix. Finally, the solution of the diagonal state differential equation is discussed, where
we recognize eigenvectors as directions in state space, in which the solution evolves with a
dynamics that is given by the corresponding eigenvalue.

0.2.1 General Solution

We derive this solution formula using variation of constants, first for the scalar-valued case,
and then for the vector-valued case.

Scalar-valued case

Consider the inhomogeneous initial value problem

ẋ(t) = ax(t) + bu(t), x(t0 ) = x0 . (39)

We assume a solution of the form


x(t) = eat k(t) (40)
with a function k(t) that needs to be determined. We differentiate the ansatz and obtain

ẋ(t) = eat k̇(t) + aeat k(t). (41)

Substitution of the last two equations in (39) yields

eat k̇(t) +  at 
ae k(t) =  at 
ae k(t) + bu(t), (42)
 

16 © PD Dr.-Ing. habil. Paul Kotyczka


Chair of Automatic Control
Technical University of Munich
CHAPTER 0. LINEAR SYSTEMS IN TIME AND FREQUENCY DOMAIN

which is a differential equation for the unknown function k(t):

k̇(t) = e−at bu(t). (43)

Integration gives Z t
k(t) = k(t0 ) + e−aτ bu(τ ) dτ. (44)
t0

By evaluating (40) for t = t0 , we obtain the initial value k(t0 ) = e−at0 x0 . Replacing now (44)
in the ansatz of the solution gives
Z t
x(t) = e a(t−t0 )
x0 + ea(t−τ ) bu(τ ) dτ. (45)
t0

Example 0.1 (Scalar system). Given the linear system

ẋ(t) = −3x(t) + 2u(t), x(0) = 4. (46)


Compute the solution x(t) for the input signal u(t) = sin (t).
ax
e sin (bx)dx = a2e+b2 (a sin (bx) − b cos (bx)).
R ax
Hint:

Solution: Using equation (55) we get:


Z t
−3t
x(t) = e x(0) + e−3(t−τ ) 2 sin (τ )dτ =
0
Z t
−3t −3t
= 4e + 2e e3τ sin (τ )dτ
0
| {z }
(A)

The integral becomes


" #t
e3t e3t 1
(A) = 2 (3 sin τ − cos τ ) = (3 sin t − cos t) +
3 + 12 0
10 10

and finally we obtain the solution


21 −3t 3 1
x(t) = e + sin t − cos t
5 5 5

In the solution we can identify the terms that originate

• from the initial value x(0): 4e−3t

• from the input function u(t): 3


5 sin t − 15 cos t + 51 e−3t

Obviously,

• the solution tends to (the equilibrium) zero, when u(t) = 0 and x(0) ̸= 0,

• the solution remains bounded (at least for the given bounded input signal).


© PD Dr.-Ing. habil. Paul Kotyczka 17
Chair of Automatic Control
Technical University of Munich
0.2. Solution of the State Differential Equation

Vector-valued case

For the solution of the vector-valued first order differential equation (49), we first recall the
Taylor series expansion of eat with a ∈ R around t = 0:
t t2 t3
eat = 1 + a + a2 + a3 + . . . . (47)
1! 2! 3!
In the same way, the matrix exponential function with A ∈ Rn×n is defined:

Definition 0.3 (Matrix exponential).

t t2 t3
eAt := I + A + A2 + A3 + . . . . (48)
1! 2! 3!

The following rules have to be considered for computations with the matrix exponential and
for differentiation the product of matrices (keep the order of the matrices):
Some matrix computation rules:
T
Transposition: (eAt )T = eA t

 −1
Inverse: eAt = e−At

Multiplication: eAt1 · eAt2 = eA(t1 +t2 )


d At
Chain rule: e = A · eAt = eAt · A
dt
d d d
Product rule for matrices: (X · Y ) = X · Y + X · Y
dt dt dt
Now we can set up the corresponding vector-valued ansatz for the solution of the vector-valued
initial value problem
ẋ(t) = Ax(t) + Bu(t), x(t0 ) = x0 , (49)
which is
x(t) = eAt k(t), k(t) ∈ Rn . (50)
The time derivative is
ẋ(t) = AeAt k(t) + eAt k̇(t). (51)
Substitution of these expressions for x(t) and ẋ(t) in (49) gives
eAt k̇(t) = Bu(t). (52)
The solution k(t) of this differential equation is
Z t
k(t) = k(t0 ) + e−Aτ Bu(τ ) dτ. (53)
t0

In complete analogy to the scalar case, we determine k(t0 ) = e−At0 x0 from the ansatz (50)
and finally obtain the solution formula
Z t
x(t) = e A(t−t0 )
x0 + eA(t−τ ) Bu(τ )dτ. (54)
t0

18 © PD Dr.-Ing. habil. Paul Kotyczka


Chair of Automatic Control
Technical University of Munich
CHAPTER 0. LINEAR SYSTEMS IN TIME AND FREQUENCY DOMAIN

Theorem 0.1 (General solution). The general solution of the vector-valued initial value
problem (49) is
Z t
x(t) = eA(t−t0 ) x0 + eA(t−τ ) Bu(τ )dτ. (55)
t0

Definition 0.4 (State transition matrix). The matrix

Φ(t, t0 ) = eA(t−t0 ) (56)

which appears in the general solution formula (55) is called state transition matrix, as it
describes the transition from the initial state x(t0 ) to the state at x(t) of the autonomous
system.

Remark 0.6. The fact that Φ̇(t, t0 ) = AΦ(t, t0 ) allows to compute the state transition matrix
by the solution of this matrix-valued differential equation with initial value Φ(t0 , t0 ) = I.

Example 0.2 (n = 2). Compute the solution x(t) of the initial value problem
" # " #
1 1 3
ẋ = x, x(0) = (57)
0 1 1

by determining the state transition matrix Φ(t, 0) = eAt .

Solution: The solution of a linear homogeneous ODE is

x(t) = eAt x(0).

In this simple example, we use the definition of the matrix exponential function (48). The
powers of the state matrix are
" # " # " #
1 1 1 2
2 1 3
3
A= , A = , A = , ....
0 1 0 1 0 1

Substituting these expressions in the definition of the matrix exponential gives


t2 t3 2t2 3t3
 
t t
1+ + + + ... 0 + + + + ... " #
At
1! 2! 3! 1! 2! 3! et tet
e = = .
 
t t2 t3 0 et
0 1+ 1! + 2! + 3! + ...

The solution of the differential equation (the initial value problem) is


" #" # " #
et tet 3 (3 + t)et
x(t) = = .
0 et 1 et

It is very obvious that the explicit calculation of a solution to the state differential equations
in the presented way becomes more difficult with the number of states!

© PD Dr.-Ing. habil. Paul Kotyczka 19


Chair of Automatic Control
Technical University of Munich
0.2. Solution of the State Differential Equation

The concept of eigenvalues and eigenvectors allows us to transform the state differential
equations into the so-called modal or diagonal form. The calculation of a solution is heavily
simplified and moreover, eigenvalues allow to assess the dynamic behavior of the system –
in particular the stability of its equilibrium – without the need to solve the state differential
equation.

0.2.2 Eigenvalues and Eigenvectors

Given a square matrix A ∈ Rn×n (or Cn×n ).

Definition 0.5. A number λ ∈ C is called an eigenvalue of A if there is at least one


non-zero column vector v ∈ Rn (or Cn ) such that the equation

Av = λv (58)

holds.

Computation of eigenvalues

Rearrange equation (58):


(A − λI)v = 0.
λ is an eigenvalue if it is a solution of this equation with v ̸= 0. Consider (A − λI) to be
invertible, then (A − λI)−1 0 = 0 is a solution → contradiction with the assumption v ̸= 0!
Consequently, for Eq. (58) to hold with v ̸= 0 the matrix (A − λI) must be singular (lose
rank) when λ is an eigenvalue. We know that
rank(A − λI) < n ⇔ det(A − λI) = 0.
which leads to the definition of the characteristic polynomial.

Definition 0.6 (Characteristic polynomial and characteristic equation). The characteristic


polynomial of the matrix A is the n-th order polynomial

p(λ) = det(λI − A). (59)

The eigenvalues of A are the n roots λi , i = 1, . . . , n of the characteristic equation


p(λ) = 0:
det(λi I − A) = 0 ⇔ λi is an eigenvalue of A. (60)

A has not necessarily different eigenvalues. If an eigenvalue λi occurs ki times, we say it has
an algebraic multiplicity µA (λi ) = ki .

Theorem of Cayley-Hamilton

For a matrix A ∈ Rn×n , the characteristic equation has the form


p(λ) = λn + an−1 λn−1 + . . . + a1 λ + a0 = 0 (61)
with a0 , . . . , an−1 the coefficients of the characteristic polynomial.

20 © PD Dr.-Ing. habil. Paul Kotyczka


Chair of Automatic Control
Technical University of Munich
CHAPTER 0. LINEAR SYSTEMS IN TIME AND FREQUENCY DOMAIN

Theorem 0.2. The matrix A satisfies its characteristic equation, i.e.

p(A) = An + an−1 An−1 + . . . + a1 A + a0 I = 0. (62)

By the Theorem of Cayley-Hamilton, An can be expressed in terms of I, A, . . . , An−1 . Also


every power greater than n of the matrix A can be written as a weighted sum of Ai , i =
0, . . . n − 1.

Example 0.3. The matrix An+1 can be written as follows:


n−1
X
An+1 = −( ai Ai )A
i=0
n−2 n−1
(63)
X X
i i
=− ai A + an−1 ai A .
i=0 i=0


Remark 0.7. For matrices A which can be diagonalized (see further below), the proof of the
Theorem of Cayley-Hamilton is straightforward using the transformation to diagonal form.

Homework: Prove the Theorem of Cayley-Hamilton for (i) a diagonal matrix and, after reading
of the next paragraph, (ii) for a diagonalizable matrix.

Computation of eigenvectors

Given an eigenvalue λi , we determine all the vectors v ji ̸= 0 for which

(A − λi I)v ji = 0 (64)

holds. The solutions v ji of this equation for a given eigenvalue span the eigenspace 12 cor-
responding to λi . The dimension of this eigenspace is called geometric multiplicity of the
corresponding eigenvalue γA (λi ).

Theorem 0.3. If a matrix A ∈ Rn×n (or Cn×n ) possesses n linearly independent eigen-
vectors v i then the transformation matrix

V = [v 1 . . . v n ]

12 γ (λ )
The eigenspace span {v 1i , . . . , v i A i } is the kernel or nullspace of the matrix (A − λi I). It is spanned by
all vectors v ji , which are mapped to the zero vector by right multiplication with the matrix (A − λi I).

© PD Dr.-Ing. habil. Paul Kotyczka 21


Chair of Automatic Control
Technical University of Munich
0.2. Solution of the State Differential Equation

brings the state matrix A into diagonal (or modal) form


 
λ1 0 0 ··· 0
0 λ
2 0 ··· 0
.. 
 

Λ := V −1 AV =  0 0 λ3 . . (65)
 
 . ..
 . ..

 . . . 0

0 0 ··· 0 λn

Proof: Rearrange (65):

V Λ = AV

i λ1
 
..
h h i
v1 · · · vn  .  = A v1 ··· vn
 

λn
h i h i
λ1 v 1 · · · λn v n = Av 1 · · · Av n

The last line contains the equations to determine the n eigenvalues and eigenvectors of A.
The requirement of linearly independent eigenvectors ensures the existence of the inverse V −1 .
Then, the last equation and (65) are equivalent, which completes the proof. □

Remark 0.8. Note that it is not necessary that all eigenvalues λi of A are different to diago-
nalize the matrix A. Rather their geometric and algebraic multiplicities must coincide.

0.2.3 State Space Model in Diagonal Form

Not only the autonomous state differential equation can be brought to diagonal or modal
form, but the complete linear state space model13 . This form is particularly useful for the
computation of the solution. Also the analysis of important system properties like controllability
and observability is very illustrative in diagonal form. See Section 0.4 and also Gilbert’s criterion
in Advanced Control.

Transformation to diagonal/modal form

Not only that state matrix, but the whole state space representation can be transformed to
modal form using the matrix of linearly independent eigenvectors.

Corollary 0.1. When the state matrix A has n linearly independent eigenvectors v i , the
linear coordinate change
h i
x=Vz ⇔ z = V −1 x, V = v1 · · · vn (66)

13
Note that diagonal form in this case refers to the diagonal state matrix.

22 © PD Dr.-Ing. habil. Paul Kotyczka


Chair of Automatic Control
Technical University of Munich
CHAPTER 0. LINEAR SYSTEMS IN TIME AND FREQUENCY DOMAIN

transforms the state space model

ẋ = Ax + Bu
(67)
y = Cx

into modal or diagonal form

ż = Λz + B̂u
(68)
y = Ĉz

with
Λ = diag{λ1 , . . . , λn }, B̂ = V −1 B, Ĉ = CV .

Homework: Prove the corollary.

Note that the new coordinates z = [z1 , . . . , zn ]T do not have the (physical) interpretation
of the original states x = [x1 , . . . , xn ]T any more. (This is a general property of coordinate
transformations, we will learn at least two more useful transformations!)

Remark 0.9. The transformed state differential equation consists of n ODEs which are now
decoupled with respect to the state variables:
T
żi = λi zi + b̂i u, i = 1, . . . , n.

Remark 0.10. If the geometric multiplicity γA (λi ) of λi is less than its algebraic multiplicity
µA (λi ), we can compute µA (λi ) − γA (λi ) generalized eigenvectors according to

(A − λi I)v ji = v j−1
i , j = µA (λi ) + 1, . . . , γA (λi ) (69)

and the matrix A can be transformed to Jordan canonical form.

Solution of the initial value problem

As the n scalar state differential equations in modal form are decoupled (with respect to zi ),
their solutions according to the general formula are
Z t
T
zi (t) = eλi t zi (0) + eλi (t−τ ) b̂i u(τ )dτ, i = 1, . . . , n.
0

Let us first consider the case u(t) ≡ 0. The solution of the vector-valued ODE

ż = Λz

is
 λ1 t
e

Λt ..
z(t) = e z(0) =  .  z(0).
 

e λn t

© PD Dr.-Ing. habil. Paul Kotyczka 23


Chair of Automatic Control
Technical University of Munich
0.2. Solution of the State Differential Equation

A transformation x = V z into the original coordinates yields


 λ1 t
i e

..
h
x(t) = v 1 · · · vn  .  z(0) =
 

eλn t (70)
n
X
= v i eλi t zi (0)
i=1

The zi (0) are the initial values in modal coordinates, i.e. the components of
 T
w1
−1  .. 
z(0) = V x(0) =  .  x(0).
wTn

In contrast to the right hand eigenvectors v i the row vectors wTi (rows of V −1 ) are called left
hand eigenvectors, as they solve the eigenvalue problem wTi A = wTi λi .
Homework: Show that the last statement is true.

In a compact form we can finally write:


Λt −1
| e {zV } x(0).
x(t) = V (71)
=eAt

Obviously the state transition matrix eAt can be easily calculated if the system can be trans-
formed in diagonal form. The transformation of the state transition matrix happens in complete
analogy to the transformation of the state matrix:

eAt = V eΛt V −1 ⇔ A = V ΛV −1 (72)

Example 0.4 (Solution of the state differential equation via diagonalization). Given the au-
tonomous system (i.e. no input)
" # " #
−1 1 −3
ẋ = x, x(0) = ,
0 −2 2

calculate the solution x(t) via the transformation to modal form.

Solution:

1. Determine the eigenvalues

λ + 1 −1 !
det(λI − A) = = (λ + 1)(λ + 2) = 0
0 λ+2

⇒ λ1 = −1 λ2 = −2.
Note that diagonal elements of a triangular matrix are exactly the eigenvalues, i.e. in
such a case, the eigenvalues can be determined by inspection.

24 © PD Dr.-Ing. habil. Paul Kotyczka


Chair of Automatic Control
Technical University of Munich
CHAPTER 0. LINEAR SYSTEMS IN TIME AND FREQUENCY DOMAIN

2. Compute the eigenvectors

(λi I − A)v i = 0, i = 1, 2.

λ1 = −1: " #" # " #" # " #


−1 + 1 −1 v11 0 −1 v11 0
= =
0 −1 + 2 v12 0 1 v12 0
" #
1
⇒ v11 = c1 = const., v12 = 0 ⇒ v 1 = c1
0
λ2 = −2: " #" # " #" # " #
−2 + 1 −1 v21 −1 −1 v21 0
= =
0 −2 + 2 v22 0 0 v22 0
" #
−1
⇒ −v21 − v22 = 0 ⇒ v 2 = c2
1

3. Transformation matrices (with c1 = c2 = 1)


" # " #−1 " #
h i 1 −1 −1 1 −1 1 1 1
V = v1 v2 = , V = = .
0 1 0 1 1 0 1

4. Solution

x(t) = V eΛt V −1 x(0) =


" #" #" #" # " #" #
1 −1 e−t 0 1 1 −3 e−t −e−2t −1
= = =
0 1 0 e−2t 0 1 2 0 e−2t 2
" #
−e−t − 2e−2t
= .
2e−2t

0.3 System Representations in Frequency Domain

We summarized state space modeling, linearization and the solution of the resulting linear time-
invariant state space models in time domain. Via an integral transformation, the underlying
differential equations can be transformed to algebraic equations in frequency domain, which, for
example, simplifies their solutions. To obtain solutions in time domain, a back-transformation
is necessary.
The integral transformation which is mainly used for control purposes is the Laplace transform
with the complex frequency variable s ∈ C. After a very short reminder on the Laplace
transform, we discuss how transfer functions, i.e., the frequency domain representations of
dynamical systems, which should be known from a basic control course, can be determined
from a given LTI state space model. We discuss poles and zeros of transfer functions in the
case of SISO systems, and we recall with the impulse and step response as well as the initial
and final value theorem some important connections between frequency and time domain.

© PD Dr.-Ing. habil. Paul Kotyczka 25


Chair of Automatic Control
Technical University of Munich
0.3. System Representations in Frequency Domain

0.3.1 Laplace Transform

The two-sided Laplace transform of a signal f (t) is defined as


Z ∞
f (t)e−st dt. (73)
−∞

This definition with the weighting factor e−st allows – in contrast to the Fourier transform –
to include also signals (of exponential type) which do not vanish for t → ∞.
Dealing with signals f (t) which are zero for t < 0, the lower integration limit can be set to
“0− ”, which refers to the instant “just before t = 0”. The one-sided Laplace transform (which
is the one we will use) Z ∞
F (s) := L{f (t)} = f (t)e−st dt (74)
0−
allows to include Dirac pulses at t = 0 in the integration. (For us this is important, as we
choose the Dirac pulse as an idealized input signal, based on which the impulse response of
an LTI system is defined.)
Thus, the Laplace transform of the Dirac pulse δ(t) (at t = 0) is well-defined:
Z ∞ Z ∞
−st
L{δ(t)} = δ(t)e dt = δ(t)e0 dt = 1. (75)
0− 0−
Other frequently used correspondences (as they can be found in tables) are
Z ∞ Z ∞
−st
L{σ(t)} = σ(t)e dt = e−st dt
0− 0−
∞ (76)
1 1 1

= − e−st = (e0 − lim e−st )) =
s 0− s t→∞ s
for σ(t) the Heaviside step function and
1
L{e−at } = (77)
s+a
for the exponential function.
An important property (see also the corresponding tables) is the algebraization of differential
equations in t by
d
L{ f (t)} = sF (s) − f (0− ) (78)
dt
where f (0− ) denotes the initial value of the function f . For more properties of the Laplace
transform we refer to text books and correspondence tables at this point.

0.3.2 Transfer Functions

SISO Systems

First let us consider the case of a linear time invariant system in state representation with one
input and one output14 (SISO):
ẋ = Ax + bu
(79)
y = cT x.
14
Please note the convention that for a single output system, instead of writing C, we denote the output
matrix, which then becomes a row vector, cT ∈ R1×n .

26 © PD Dr.-Ing. habil. Paul Kotyczka


Chair of Automatic Control
Technical University of Munich
CHAPTER 0. LINEAR SYSTEMS IN TIME AND FREQUENCY DOMAIN

At this moment we are interested in the input-/output behavior, so we assume all initial values
x(0), to be zero. Using the Laplace transformation we get the following representation of the
above state representation in the frequency domain:
sX(s) = A X(s) + bU (s)
Y (s) = cT X(s)
By solving the first equation for X(s) and replacing the result in the second one, we obtain a
relation between the transformed input signal U (s) and the output Y (s).
(sI − A)X(s) = bU (s)
X(s) = (sI − A)−1 bU (s)
Y (s) = cT (sI − A)−1 bU (s)

The transfer function of the linear SISO system (79) is:

Y (s)
G(s) = = cT (sI − A)−1 b. (80)
U (s)
One way to perform the matrix inversion which is necessary to compute the transfer function,
is the following:
adj(sI − A)
(sI − A)−1 = (81)
det(sI − A)
Therein, adj denotes the adjugate of a matrix. For a general matrix M it can be computed
as follows, where [·]ij denotes the (i, j)-th element of the matrix:

[adj(M )]ij = (−1)i+j Dji .

Dji is the determinant of M after deleting the j-th row and i-th column.

If the state representation (79) is in modal form, then (sI −Λ) as well (sI −Λ)−1 are diagonal
matrix and the transfer function can be written as
 1 
 s−λ1 .. Xn
ĉi b̂i
G(s) = ĉT 

.  b̂ = =

1

i=1
s − λi
s−λn
Pn
i=1 ĉi b̂i j̸=i (s −
Q
λj ) N (s)
= Qn = .
i=1 (s − λi ) D(s)
Therein, D(s) is the denominator polynomial of order n and N (s) the numerator polynomial
of order m ≤ n − 1. At this stage we consider the orders of the polynomials before possible
cancellations (see further below). Depending on whether both polynomials are expanded or
factored out, the following two representations of G(s) are possible.
Polynomial and pole-zero representation of a transfer function:
bm sm + . . . + b1 s + b0
G(s) =
sn + an−1 sn−1 + . . . + a1 s + a0
Qm (82)
bm − ηj )
j=1 (s
= Qn .
i=1 (s − λi )

© PD Dr.-Ing. habil. Paul Kotyczka 27


Chair of Automatic Control
Technical University of Munich
0.3. System Representations in Frequency Domain

Remark 0.11. If we consider state space models with feedthrough,

ẋ = Ax + bu
(83)
y = cT x + du,

the transfer function is


G(s) = cT (sI − A)−1 b + d, (84)
and the orders of the numerator and the denominator polynomial, if G(s) is represented as a
rational function, are equal: m = n.

Remark 0.12. A given state space model has several transfer functions. Consider for example
(79) with additional disturbance input (see in Advanced Control),

ẋ = Ax + bu + ez
(85)
y = cT x,

the disturbance transfer function would be


Y (s)
Gz (s) = = cT (sI − A)−1 e. (86)
Z(s)
Besides the open-loop transfer functions of the uncontrolled system, the control engineer is
interested in closed-loop transfer functions. A control law of the form u = −r T x + f w for
example yields the closed-loop transfer function
Y (s)
Gw (s) = = cT (sI − A + br T )−1 bf. (87)
W (s)

MIMO Systems

If we consider the general case of a MIMO systems, then we can use the same arguments as
before and come up with the frequency domain relations between input vector U (s) ∈ Rp and
output vector Y (s) ∈ Rq :
Y (s) = G(s)U (s), (88)
where G(s) ∈ Rq×p is the transfer matrix

G(s) = C(sI − A)−1 B. (89)

The (i, j)-th element of G(s)

gij (s) = cTi (sI − A)−1 bj

denotes the scalar transfer function from the j-th scalar input uj to the i-th scalar output yi .

0.3.3 Poles and Zeros in SISO Systems

The roots of the denominator and the numerator polynomial in (82) are poles and zeros of
the transfer function G(s). We summarize some important facts on poles and zeros in SISO
systems.

28 © PD Dr.-Ing. habil. Paul Kotyczka


Chair of Automatic Control
Technical University of Munich
CHAPTER 0. LINEAR SYSTEMS IN TIME AND FREQUENCY DOMAIN

Figure 4: Step response of a non-minimum phase system with transfer function G(s) =
−s + 1
2
.
s + 2s + 1

• If there are no cancellations between the numerator polynomial N (s) and the denomi-
nator D(s), then all eigenvalues of A – which are the roots of det(sI − A) – appear as
poles of the transfer function.

• If zeros and eigenvalues coincide in the SISO case, the eigenvalues become uncontrollable
or unobservable (see later in this course). Poles and zeros compensate. A discussion
of zeros and the compensation with eigenvalues in the – more delicate – MIMO case
follows in Section 0.5.

• The poles of G(s) determine input-output stability of the transfer function, i.e., bound-
edness of the output up to a bounded input.

• The zeros do not affect stability as long as they do not compensate with unstable
eigenvalues. If this happens, internal instability occurs which is not visible in the transfer
behavior, yet it is present in the solution of the state space model.

• In the SISO case, all zeros which are not compensated (not cancelled) with poles are
transmission zeros: If the input signal u(t) contains a component eηi t , where ηi is a zero
of G(s), the stationary (or particular) part of the output signal will not contain eηi t .
In the MIMO case, this blocking property of transmission zeros is restricted to certain
directions in input and state space, see Section 0.5.

• However, zeros can influence the transient behavior, see Fig. 4 which shows a typical
undershoot in the step response of a system with a right half plane zero (Re ηj > 0).
Such systems are called non-minimum phase systems.

0.3.4 Impulse and Step Response

Given a transfer function G(s) (equal number of in- and outputs, p = q = 1). We are
interested in the response of the output y(t) to the typical reference input signals, see Fig. 5,

• ideal impulse (Dirac function)


u(t) = δ(t)

© PD Dr.-Ing. habil. Paul Kotyczka 29


Chair of Automatic Control
Technical University of Munich
0.3. System Representations in Frequency Domain

Figure 5: (Dirac) impulse and (Heaviside) step function

• and unit step (Heaviside function)

u(t) = σ(t).

Let g(t) = L{G(s)} be the inverse Laplace transform of G(s). Recall the correspondence
between multiplication and convolution
Z t
Y (s) = G(s)U (s) s c y(t) = (g ∗ u)(t) = g(t − τ )u(τ )dτ.
0

For the impulse input signal u(t) = δ(t) we obtain


Z t
y(t) = (g ∗ δ)(t) = g(t − τ )δ(τ )dτ = g(t),
0

which allows us to state the following.


The impulse response of a linear time-invariant system is

g(t) = L−1 {G(s)}. (90)


Now consider the Laplace transforms of the two standard inputs

c s 1, c s 1
δ(t) σ(t) .
s
Using the correspondence between multiplication and convolution, the output signal for a given
step input turns out to be

1 G(s) s c y(t) = (L−1 {H(s)} ∗ δ)(t) = h(t).


Y (s) = G(s) · = ·1
s | {z
s } | {z }
=h
=H(s)

With the correspondence Z t


c s F (s)
f (τ )dτ
0 s
finally we obtain the expression for h(t):
Z t
−1 −1 G(s)
h(t) = L {H(s)} = L { }= g(τ )dτ.
s 0

The step response of a linear time-invariant system is


Z t
G(s)
h(t) = L−1 { }= g(τ )dτ. (91)
s 0

30 © PD Dr.-Ing. habil. Paul Kotyczka


Chair of Automatic Control
Technical University of Munich
CHAPTER 0. LINEAR SYSTEMS IN TIME AND FREQUENCY DOMAIN

Using the correspondence


(sI − A)−1 s c eAt

it is easy to obtain the impulse response of an LTI system from its state representation:
Correspondence of transfer function and impulse response:

G(s) = cT (sI − A)−1 b s c g(t) = cT eAt b. (92)

0.3.5 Initial and Final Value Theorem

The Laplace transform of a signal can be used to determine the signal value for t → 0 and
t → ∞. The following theorem – which should be known from introductory control courses –
is stated without proof.

Theorem 0.4 (Initial and final value theorem). For a signal f (t) and its Laplace transform
F (s) the following holds (assuming the limits exist):

lim f (t) = lim (sF (s))


t→0 s→∞

lim f (t) = lim (sF (s))


t→∞ s→0

Remark 0.13. Note that F (s) in the initial/final value theorem denotes the Laplace transform
of a signal and not a transfer function.

Using the final value theorem, it becomes clear that G(0) is the steady state gain of a transfer
function G(s). Take a unit step as input signal:
c s 1
u(t) = σ(t) (amplitude 1) U (s) =
s
and express the corresponding output in the frequency domain:
G(s)
Y (s) = G(s)U (s) = .
s
Applying the final value theorem we get
G(s)
lim y(t) = lim (sY (s)) = lim (s ) = G(0).
t→∞ s→0 s→0 s

0.4 Illustration of Controllability and Observability

We consider the following particular system in diagonal form:


      
ẋ1 λ1 0 0 0 x1 b1
ẋ   0 λ 0 0  x  b 
 2  2   2  2
 =   +  u
ẋ3   0 0 λ3 0  x3   0  (93)
ẋ4 0 0 0 λ 4 x4 0
h i
y = c1 0 c3 0 x.

© PD Dr.-Ing. habil. Paul Kotyczka 31


Chair of Automatic Control
Technical University of Munich
0.4. Illustration of Controllability and Observability

x1 (0)
x1
b1 c1

λ1

x2 (0)
x2
b2

λ2

u y
x3 (0)
x3
c3

λ3

x4 (0)
x4

λ4

Figure 6: Illustration of the system (92). Only the eigenvalue λ1 is controllable and observable.

It is depicted in the block diagram in Fig. 6, and it illustrates the basic notions of controllability
and observability – before we will introduce them in Advanced Control. Also, this example
anticipates the controllability/observability criterion by Gilbert, which applies to diagonaliz-
able systems. The example helps for the understanding of the concepts of zeros and their
compensation with eigenvalues in the next section.
The system is composed of four first order systems
ẋi (t) = λi xi (t) + bi u(t), i = 1, . . . , 4, (94)
each of which has an initial value xi (0). The input u excites the first and the second subsystem
via the gains b1 ̸= 0 and b2 ̸= 0, while b3 = 0 and b4 = 0 disconnect the third and fourth
system from the input. On the side of the output, we recognize that only x1 and x3 contribute
to y (via c1 ̸= 0 and c3 ̸= 0), while the output remains unaffected by x2 and x4 (c2 = c4 = 0).
The first and the second subsystem (in diagonal form) can be influenced or controlled by the
input u1 (t) and u2 (t), respectively, which is not the case for the third and the fourth one. We
call the eigenvalues of the first two subsystems λ1 and λ2 controllable.

32 © PD Dr.-Ing. habil. Paul Kotyczka


Chair of Automatic Control
Technical University of Munich
CHAPTER 0. LINEAR SYSTEMS IN TIME AND FREQUENCY DOMAIN

The states x1 (t) and x3 (t) contribute to the output signal, their evolution can be observed in
y(t), which is not the case for x2 (t) and x4 (t). λ1 and λ3 are called observable eigenvalues.

0.5 Zeros of MIMO Systems

In this section we show how the notion of zeros can be generalized to the MIMO case. We
motivate the definition of invariant zeros via the Rosenbrock (system) matrix from the blocking
property of transmission zeros. It turns out that this definition is also meaningful, for zeros
which are not transmission zeros. The locations of invariant zeros are important for the
stability of inversion-based control designs, like decoupling control, which is a topic of the
lecture Advanced Control. We restrict our attention to cases where the number of inputs and
outputs of the LTI system (A, B, C) is identical: p = q > 1.

0.5.1 Blocking Property and Rosenbrock Matrix

In Section 0.3.3, we already mentioned the blocking property of transmission zeros, which we
now shall discuss in more detail for the MIMO case. We are interested in the question, if for
an LTI system (A, B, C) there exists a vector-valued input signal of the form

u(t) = u0 eηt , u0 ∈ Rp , (95)

whose stationary effect15 on the output y(t) is zero, which means that the input signal is
blocked. We assume at this point that η does not coincide with an eigenvalue of A: η ̸= λi (A),
i = 1, . . . , n.
We insert (95) in the solution formula (55), and obtain (note that e dt = X −1 eXt =
R Xt

eXt X −1 )
Z t
x(t) = eAt x0 + eA(t−τ ) Bu0 eητ dτ
0
Z t
= eAt x0 + eAt e(ηI−A)τ dτ Bu0
0 (96)
= eAt x0 + eAt e(ηI−A)t (ηI − A)−1 Bu0 − eAt (ηI − A)−1 Bu0
= eηt (ηI − A)−1 Bu0 +eAt (x0 − (ηI − A)−1 Bu0 ) .
| {z } | {z }
=:cp =:ch

The solution is composed of a particular or stationary part xp = eηt cp , in which the exitation
by the input (95) is visible, and a homogeneous or transient part xh = eAt ch , which contains
the initial value x0 . Substituting x(t) = eηt cp + eAt ch and its time derivative in the state
differential equation gives

AeAt ch + ηeηt cp = AeAt ch + Aeηt cp + Bu0 eηt , (97)

or
(A − ηI)cp eηt + Bu0 eηt = 0. (98)
15
We mean the effect when all other initial components of the solution have decayed to zero.

© PD Dr.-Ing. habil. Paul Kotyczka 33


Chair of Automatic Control
Technical University of Munich
0.5. Zeros of MIMO Systems

Blocking of the input signal means that the particular component of the output shall be
identically zero:
y p (t) = Cxp (t) = Ccp eηt = 0. (99)
Combining both equations (98) and (99) yields the condition
" #" #
A − ηI B cp ηt
e = 0, (100)
C 0 u0
h iT
which must be satisfied for all t. For a non-zero vector cTp uT0 to exist (in particular for
the blocking property, u0 shall be non-zero, because otherwise we consider zero input), the
(n + p) × (n + p) matrix on the left hand side must be rank deficient. This matrix plays
an important role in linear systems’ analysis and control design and is called the Rosenbrock
(system) matrix :

Definition 0.7 (Rosenbrock matrix). The matrix


" #
A − ηI B
P (η) = , η∈C (101)
C 0

is called the Rosenbrock matrix of the system (A, B, C).

0.5.2 Invariant and Transmission Zeros

We call those values of η, for which the Rosenbrock matrix becomes rank deficient, the invariant
zeros of the system.

Definition 0.8 (Invariant zeros). Let det(P (η)) = 0. Then η is called an invariant zero of
(A, B, C).

Why is such an η called invariant zero? In the context of state feedback control, we will see
that the location of invariant zeros remains unchanged under state feedback. This will be an
exercise in Advanced Control.
We now discuss three cases:

1. η ̸= λi : the invariant zero does not coincide with an eigenvalue,

2. η = λi : the invariant zero coincides with an uncontrollable eigenvalue,

3. η = λi : the invariant zero coincides with an controllable eigenvalue.

The result will be a characterization of invariant zeros that have an impact on the transmission
behavior and therefore can be defined as transmission zeros.
The first case will be discussed for general system matrices, for cases 2 and 3 (which are
completely analogue for the case of coincidence with an unobservable/observable eigenvalue),
it is convenient to consider the diagonal/modal form.

34 © PD Dr.-Ing. habil. Paul Kotyczka


Chair of Automatic Control
Technical University of Munich
CHAPTER 0. LINEAR SYSTEMS IN TIME AND FREQUENCY DOMAIN

Case 1: η ̸= λi (A)

In this case, det P (η) can be factorized as follows:

A − ηI B I 0 A − ηI B
=
C 0 −C(A − ηI)−1 I C 0
A − ηI B (102)
=
0 C(ηI − A)−1 B
= det(A − ηI) · det G(η).
Because η ̸= λi , we know that det(A − ηI) =
̸ 0, and all matrices are well-defined. Now the
two conditions
det P (η) = 0 ⇔ det G(η) = 0 (103)
are equivalent. det G(η) = 0 can be considered in this case a straightforward extension of the
characterization G(η) = 0 of a transmission zero in the SISO case.

Case 2: η = λi (A), λi uncontrollable

For simplicity, we consider the case of simple eigenvalues and that η is the only invariant zero.
Without loss of generality, we assume η = λ1 and λ1 uncontrollable. The modal form of the
state matrices, see Corollary 0.1, is then
" # " #
λ 0T bT h i
Λ= 1 , B= 1 , C = c1 C r , (104)
0 Λr Br

where λ1 , bT1 and c1 denote the first element, the first row and first column of the corresponding
matrices. Uncontrollability of λ1 according to Gilbert’s criterion (see Advanced Control, but
also compare with the illustration in the previous section) is expressed by
bT1 = 0T . (105)
The transfer matrix of the system in modal form can be written (verify this!)
n
ci bTi
(106)
X
G(s) = .
i=1
s − λi

Because of (105), which makes the first term of the sum vanish, the transfer matrix coincides
with the one of the reduced system (Λr , B r , C r )
n
ci bTi
(107)
X
Gr (s) = = G(s).
i=2
s − λi

Now consider the determinant of the Rosenbrock matrix of (Λ, B, C),

η − λ1 0T 0T
det P (η) = 0 ηI − Λr B r
c1 Cr 0
(108)
ηI − Λr B r
= (η − λ1 )
Cr 0
= (η − λ1 ) det P r (η) = 0.

© PD Dr.-Ing. habil. Paul Kotyczka 35


Chair of Automatic Control
Technical University of Munich
0.5. Zeros of MIMO Systems

Because η = λ1 is assumed a simple invariant zero, we can conclude that det P r (η) ̸= 0.
From (107) and (108) we recognize that

1. the eigenvalue λ1 does not contribute to the transfer matrix Gr (s) = G(s) and

2. η = λ1 is not an invariant zero of the reduced system (Λr , B r , C r ).

We conclude that both the invariant zero and the uncontrollable eigenvalue at the same location
have no effect on the transfer behavior. In this sense, η is not a transmission zero.

Remark 0.14. The same argumentation holds if λ1 is unobservable. The instead of (105), we
have c1 = 0.

We also recognize (at least for the considered case of simple eigenvalues) that unobservability
and uncontrollability always go along with the existence of an invariant zero at the same
location.

Case 3: η = λi (A), λi controllable

Neither can the transfer matrix be reduced, nor can a reduced system be constructed, in
which η is not invariant zero as in the previously considered case. Although the invariant zero
coincides with an eigenvalue, their effects remain in the transfer behavior. η is a transmission
zero.
Summarizing the three preceding paragraphs, we conclude:

Definition 0.9 (Transmission Zeros). Transmission zeros are invariant zeros that do not
coincide with uncontrollable or unobservable eigenvalues. They form a subset of the
above defined invariant zeros:

{Transmission Zeros} ⊆ {Invariant Zeros}

We state the following on compensation in MIMO systems.

Definition 0.10 (Compensation in MIMO systems.). Compensation of an eigenvalue and


an invariant zero occurs, if they are at the same location and the eigenvalue is uncontrol-
lable or unobservable.
In the case of compensation, the effects of the zero and the eigenvalue disappear from
the input-/output behavior of the system. The zero is no transmission zero.
Conversely, invariant zeros that coincide with controllable and observable eigenvalues, are
transmission zeros.

We made plausible that in the MIMO case the Rosenbrock matrix yields a decent definition of
(invariant) zeros and that the definition of MIMO transmission zeros is more delicate than in
the SISO case. Other than in the SISO case, a MIMO invariant zero is still a transmission zero
if it coincides with an eigenvalue, as long as this eigenvalue is observable and controllable.

36 © PD Dr.-Ing. habil. Paul Kotyczka


Chair of Automatic Control
Technical University of Munich
CHAPTER 0. LINEAR SYSTEMS IN TIME AND FREQUENCY DOMAIN

0.5.3 Examples

We close this section with three examples to illustrate the above general statements. The
examples of diagonal (or diagonalized) MIMO systems, in which the states are the modal
coordinates, illustrate what happens when invariant zeros coincide with eigenvalues, and when
compensation occurs or not. We do not discuss the particular case of non-diagonalizable
systems, or more general, the case of systems with multiple eigenvalues, as the additional
technical issues which have to be considered there do not contribute to the understanding of
the concepts of invariant zeros and compensation.

Example 0.5. Consider the diagonal MIMO system (A, B 1 , C 1 ) with


   
−1 0 0 1 −1 " #
1 −1 0
A =  0 −2 0  , B 1 = 0 1  , C1 = . (109)
   
0 1 −1
0 0 −3 0 2

The system has three eigenvalues λ1 = −1, λ2 = −2 and λ3 = −3 as well as one invariant
zero η = −1, which coincides with λ1 . The system is controllable and observable, as all rows
of B 1 and all columns of C 1 are non-zero16 .
The transfer matrix G1 (s) = C 1 (sI − A)−1 B 1 of the system is
" 1 2s+3
#
− (s+1)(s+2)
G1 (s) = s+1
s+1 . (110)
0 − (s+2)(s+3)

We recognize that, although η = λ1 , i.e., the invariant zero coincides with the first eigenvalue,
η = −1 appears as zero of the lower right transfer function G1,22 (s), and also λ1 = −1 appears
as pole of G1,11 (s) and G1,21 (s). Therefore, no compensation of invariant zero and eigenvalue
occurs and η = −1 is a transmission zero.
Moreover, −1.5, which is not an invariant zero of the MIMO systems, appears as a zero of the
upper right transfer function G1,12 (s).
We verify the blocking property of η = −1 by considering the solution (96), which we multiply
with the singular matrix (ηI − A). For an initial value x0 such that (ηI − A)x0 = B 1 u0 ,
the term with eAt vanishes, and what remains is

(ηI − A)x(t) = eηt B 1 u0 (111)

or, with numbers,    


0 0 0 1 −1
ηt 
0 1 0 x(t) = e 0 1  u0 . (112)
  
0 0 2 0 2
This equation can only be true17 for an input direction
" #
1
u0 = k1 , k1 ∈ R. (113)
1
16
Cf. Fig. 6. For the MIMO system (A, B 1 , C 1 ), each of the three scalar subsystems with eigenvalues λ1,2,3
is connected to at least one of the two inputs and one of the two outputs. Later in the course we will introduce
the corresponding conditions as Gilbert’s criterion for controllability/observability
17
Otherwise, the first component of the vector on the right hand side is zero, which contradicts the zero on
the left hand side.

© PD Dr.-Ing. habil. Paul Kotyczka 37


Chair of Automatic Control
Technical University of Munich
0.5. Zeros of MIMO Systems

Assuming k1 = 1, a solution which satisfies (111) has the form


 
k2
(114)
  ηt
x(t) =  1  e , k2 ∈ R,
1

where x1 (t) = k2 eηt follows from the form of the input18 u(t) = u0 eηt and linearity of the
system. Only for k2 = 1, i.e., a solution of the form
 
1
x(t) = cp eηt with cp = 1 , (115)
 
1

the condition C 1 x(t) = 0 is true. The effect of such a solution is not visible at the output.
u0 and cp therefore are directions in input and state space, for which an excitation with eηt is
blocked at the output. ◁

The example illustrates the following points.


• The zero of a scalar transfer function in a transfer matrix is not necessarily an
invariant zero of the MIMO system.

• A transmission zero does not necessarily appear as a zero of an element of the transfer
matrix19 .

• Blocking of a transmission zero in a MIMO system is a property, which is restricted


to certain directions in input and state space.

Example 0.6. Now consider the system (A, B 2 , C 2 ) with modified input and output matrices
 
0 0 " #
1 −1 0
B 2 = 1 0 , C2 = . (116)
 
0 0 1
0 1

This system has still a single invariant zero, which coincides with the first eigenvalue, η =
λ1 = −1. This eigenvalue – and therefore the system – is now, however, uncontrollable: The
zero row in B 1 avoids a modification of the first differential equation ẋ1 (t) = −x1 (t), which
is characterized by the eigenvalue λ1 = −1.
The computation of the transfer matrix gives
" #
1
− s+2 0
G2 (s) = 1 , (117)
0 s+3

and we note that the invariant zero η = −1 does not appear as a zero of any numerator
polynomial nor is the eigenvalue λ1 = −1 a pole of a scalar transfer function. The zero, as
well as the eigenvalue, have no effect on the input-output behavior. By compensation with an
uncontrollable eigenvalue, the invariant zero is not a transmission zero. ◁
18
Remember that the terms with eAt are excluded by restriction of the initial value x0 , see above.
19
In Example 0.5, the transmission zero appers in G1,22 (s). See the exercise for an example where this is not
true.

38 © PD Dr.-Ing. habil. Paul Kotyczka


Chair of Automatic Control
Technical University of Munich
CHAPTER 0. LINEAR SYSTEMS IN TIME AND FREQUENCY DOMAIN

Example 0.7. Setting


B 3 = C T2 , C 3 = B T2 , (118)
we construct an unobservable system. The zero first column of C 3 makes that the first
eigenmotion (mode) not visible at the output. The transfer matrix is

G3 (s) = B T2 (sI − A)−1 C T2


= (C 2 (sI − A)−1 B 2 )T (119)
= GT2 (s).

Compensation of the invariant zero with the unobservable eigenvalue, η = λ1 = −1 makes


their effects both disappear from the transfer matrix, see above. The invariant zero is not a
transmission zero. ◁

The last examples illustrated:


• Pole-zero cancellation in a transfer matrix, or in other words compensation, occurs
only if an invariant zero coincides with an uncontrollable or/and unobservable eigen-
value.

© PD Dr.-Ing. habil. Paul Kotyczka 39


Chair of Automatic Control
Technical University of Munich

You might also like