INSE 6710: Fundamentals and Applications of

Cyber-Physical Systems

Lecture 3 - Mathematical Modeling of Physical Systems

Prof. Walter Lucia

Fall - 2022

Fall - 2022 1 / 58

Outline

1 Networked CPS - Plant Model (Background)

Continuous-Time Linear Systems
Classes of Dynamical Systems
Equilibrium Points and Stability

Fall - 2022 2 / 58

Lecture 2 - Recap

Fall - 2022 3 / 58
Attack Space: 3D modeling

Model

Communication Channels
Disruptive Disclosure
Attacker Policy

Networked Controller

3D ATTACK SPACE

Model Knowledge
Model Knowledge
- Plant Disclosure Resources
-Controller - Sensor measurements the attacker
-Detector can intercept
- Control Inputs the attacker can
intercept

Disclosure Resources

es
rc
ou
es
eR
Disruptive Resources

tiv
isr
up - Sensor measurements the attacker
can violate the integrity/availablity
D

- Control Inputs the attacker can

violate the integrity/availability

Fall - 2022 4 / 58

Networked CPS - Plant Model

(Background)

Fall - 2022 5 / 58

Networked CPS - Plant Model

Model→ Md = {P, C, E, D}

u(t) y(t)
A Plantstate x(t) S

Communication Channels

u(t) r(t)
control logic state
te estimator

anomaly detector

Networked Controller

• P =information set describing the plant’s dynamical model

Fall - 2022 6 / 58
Physical Plant in CPS

• In CPS, what we call “the physical system,” “the plant” or the

“system under control” is a system/entity that evolves over time,
possibly under external excitation

Input Output

Input Output
Input Output Input Output

Input

Output
Input Output Input

Output

Fall - 2022 7 / 58

Plant Dynamical Model

• The way the system evolves over time defines the dynamics of
the system
• A dynamical model of a system is a set of mathematical
equations that explain how the system evolves over time under the
effect of an external excitation u(t) or autonomously (if u(t) ≡ 0)

Fall - 2022 8 / 58

Why it is useful to have the plant’s model?

1 It helps us to understand how inputs influence outputs

2 It is useful to design the control strategy (i.e., u(t)) so that
y(t) ≈ r(t)
3 It helps us understanding if the system is behaving how we expect
• Do we have a physical fault in the system?
• Do we have a cyber-attack affecting the command inputs u(t) or the
measurements y(t) ?

Fall - 2022 9 / 58
How can we derive a dynamical model of a system?
1 Knowledge-based methods: A mathematical model is derived
by resorting to the physic laws governing the considered process,
e.g., Kirchhoff’s voltage and current laws, Newton’s second law
2 Identification based methods: The physical laws governing the
process dynamics are unknown (black box) or partially known
(gray box)
• We do experiments to find a model that explain how the input influences
the output.

Fall - 2022 10 / 58

Introduction to Linear Systems

Fall - 2022 11 / 58

What is it a linear system?

A linear dynamical system is a dynamical system whose mathematical
model y(t) = f (x(t), u(t)) satisfies the superposition principle.
Recap: Linear Function
• A function y = f (u) is linear if it enjoys two properties:
(
f (αu) = αf (u)
f (u1 + u2 ) = f (u1 ) + f (u2 )

• These two conditions can be combined, and they define the

so-called superposition principle

f (αu1 + βu2 ) = αf (u1 ) + βf (u2 )

Try yourself (at home):

• Is y(t)=3u(t) a linear function?
• Is y(t)=3u(t)+1 a linear function?
Fall - 2022 12 / 58
Dynamical Model of a Linear System

The model of any linear dynamical system can be represented as

a system of n 1st order Ordinary Differential Equations (ODEs).

• Example: If the system has 1 single-input u(t) ∈ IR, one

single-output y(t) ∈ IR and state vector x(t) ∈ IRn , then it can be
represented as follows:1

ẋ1 (t) = a11 x1 (t) + · · · + a1n xn (t) + b1 u(t)


..
 .

ẋ (t) = a x (t) + · · · + a x (t) + b u(t)
n n1 1 nn n n

y(t) = c1 x1 (t) + · · · + cn xn (t) + d1 u(t)

x1 (0) = x10 , · · · , xn (0) = xn0

1 dxi (t)
Note that ẋi (t) = dt
denotes the time derivative of xi (t).
Fall - 2022 13 / 58

Linear System: State-Space Representation

A linear dynamical system can always be represented in the following

compact matrix form, known as the state-space representation:
(
ẋ(t) = Ax(t) + Bu(t)
y(t) = Cx(t) + Du(t)

where:
• x(t) := [x1 , x2 , · · · , xn ]T ∈ IRn is state vector of the system
• u(t) := [u1 , u2 , · · · , um ]T ∈ IRm is input vector of the system
• y(t) := [y1 , y2 , · · · , yp ]T ∈ IRp is output vector of the system

A ∈ IRn×n , B ∈ IRn×m , C ∈ IRp×n , D ∈ IRp×m

Fall - 2022 14 / 58

Linear System: Block Scheme Representation2

(
ẋ(t) = Ax(t) + Bu(t)
y(t) = Cx(t) + Du(t)

2 R
the term 1/s denotes the integrator operator, i.e., x(t) = ẋ(t)dt
Fall - 2022 15 / 58
Example: mass-spring-damper system (1/4)
• Find the state-space model of the mass-spring-damper system

where
• M =mass
• K= spring constant
• β= damping coefficient

• We assume that the mass’s position z(t) is measured by a sensor:

y(t) = z(t)

Fall - 2022 16 / 58

Example: mass-spring-damper system (2/4)

• Use the Newton’s Second Law (knowledge-based method):3

X
F~i (t) = M z̈(t), y(t) = z(t)
i

⇒ u(t) − β ż(t) − Kz(t) = M z̈(t), y(t) = z(t)

|{z} | {z } | {z }
pulling force damper force spring force

u(t) = M z̈(t) + β ż(t) + Kz(t), y(t) = z(t)

• The above equation tells us that the dynamical model is described
by one ODE of order two.
3
the single and double dots on top of a variable/signal denote its first and second derivative, respectively.
Fall - 2022 17 / 58

Example: mass-spring-damper system (3/4)

u(t) = M z̈(t) + β ż(t) + Kz(t)

y(t) = z(t)

• Two obtain the state-space representation we need to re-write the

second order ODE into a system of first order ODEs
• We can define the states of this system as follows x1 (t) = z(t) and
x2 (t) = ż(t) and compute the first order derivatives4
 
ẋ1 (t) = ż(t)
 x˙1 (t) = x2 (t)

β K 1
ẋ2 (t) = z̈(t) ⇒ x˙2 (t) = − M x2 (t) − M x1 (t) + M u(t)

 

y(t) = x1 (t) y(t) = x1 (t)

4
other equivalent choices for x1 and x2 are possible
Fall - 2022 18 / 58
Example: mass-spring-damper system (4/4)

x˙1 (t) = x2 (t)

β K 1
x˙2 (t) = − M x2 (t) − M x1 (t) + M u(t)


y(t) = x1 (t)

• Now we can rewrite the above systems in a matrix form

        
ẋ1 (t) 0 1 x1 (t) 0   x1 (t)
= k β + 1 u(t), y(t) = 1 0
ẋ2 (t) −M −M x2 (t) M x2 (t)

• Defining x(t) = [x1 (t), x2 (t)]T , we obtain

    

 ẋ(t) = 0 1 0
 k β x(t) + 1 u(t)


 −M −M M
 | {z } | {z }
A B

  


 y(t) = 1 0 x(t) + |{z}
0 u(t)

 | {z }
D
C Fall - 2022 19 / 58

How can we obtain the state evolution of the system?

(
ẋ(t) = Ax(t) + Bu(t)
y(t) = Cx(t) + Du(t)

• We need to solve the above to know how x(t) and y(t) evolve
• The solution is given by the following Lagrange’s Formula
Proposition
Given the continuous-time linear system ẋ = Ax + Bu, with initial
condition x(0) = x0 ∈ IRn , there exist a unique solution (x(t), y(t)) for
any input u(t) :
forced response
natural response
zZ }| {
z }| { t
x(t) = eAt x0 + eA(t−τ ) Bu(τ )dτ
Z t 0 
y(t) = CeAt x(0) + C eA(t−τ ) Bu(τ )dτ + Du(t)
0

Fall - 2022 20 / 58

Math Recap: Eigenvalues and Eigenvectors

Fall - 2022 21 / 58
Eigenvalues and Eigenvectors (1/3)

Given a square matrix A ∈ IRn×n , the eigenvalues λi are the

roots of its characteristic polynomiala

det(λI − A) = λn + βn−1 λn−1 + · · · + β1 λ + a0 = 0

a
det = determinant, I = identity matrix

Example:
 
2 3
A= ∈ IR2×2 , det(λI − A) = (λ − 2)(λ − 4) = 0
0 4

The eigenvalues are:

λ1 = 2, λ2 = 4

Fall - 2022 22 / 58

Eigenvalues and Eigenvectors (2/3)

Given a square matrix A ∈ IRn×n , an eigenvector of A associ-

ated to the eigenvalue λi is any vector vi ∈ IRn such that

Avi = λi vi

Example:  
2 3
A= ∈ IR2×2
0 4
The eigenvectors associate to the eigenvalues λ1 = 2, λ1 = 4 are 5
   
1 0.8321
v1(λ1 =2) = , v2(λ2 =4) =
0 0.5547
Indeed (try yourself):
2 v1 , Av2 = |{z}
Av1 = |{z} 4 v2
λ1 λ2
5
You can use the Matlab function “eig(A)” to computer the eigenvalues and eigenvectors.
Fall - 2022 23 / 58

Eigenvalues and Eigenvectors (3/3)

The Algebraic Multiplicity (AM) of the eigenvalue λi is equal to
the number of times λ1 appears in the set of eigenvalues.

Example: If all the eigenvalues of A are λ̄ = {1, 2, 2, 3, 4, 5, 1}, then

AM(λ=1) = 2, AM(λ=2) = 2, AM(λ=3) = 1, AM(λ=4) = 1, etc.

The Geometric Multiplicity (GM) of the eigenvalue λi is the

number of linear independent eigenvectors vi associated to λi a .
a
A set of vectors is linearly independent when none of the vectors can be written as a linear combination of the
other vectors

Example: If for the eigenvalue λ1 = 1, two independent eigenvectors

can be found, e.g.:
   
1 2
v1(λ1 =1) = , v2(λ1 =1) =
3 4
then GM(λ=1) = 2 Fall - 2022 24 / 58
Diagonalizable Matrix

If for each eigenvalue λi of A, the algebraic multiplicity AM(λi ) is

equal to the geometric multiplicity GM(λi ) , then the matrix A is
said diagonalizable

• If A is diagonalizable, then the following decomposition of A exists:

A = T ∆T −1
 
λ1 0 . . . 0
 0 λ2 . . . 0 
 = T −1 AT, T = [v1 |v2 . . . |vn ]
 
∆= .. .. . . .
 . . . .. 
0 0 . . . λn

• λi are the eigenvalues of A

• vi are the eigenvectors of A
Fall - 2022 25 / 58

Why eigenvalues and eigenvectors are so important?

• Let us consider the following autonomous system (u ≡ 0 )

ẋ = Ax

• By resorting to the Lagrange’s formula, the state evolution of the

system is given only by the natural response (the forced response
is absent):
x(t) = eAt x0

From the Lagrange formula, without studying the matrix

exponential eAt , it is hard to understand what will be the
evolution of x(t)

Fall - 2022 26 / 58

Why eigenvalues and eigenvectors are so important?

• If A is diagonalizable, we can rewrite the response in function of
the system’s eigenvalues λi
 λt 
e 1 0 ... 0
 0 eλ2 t . . . 0 
x(t) = eAt x0 = T e∆t T −1 x0 = [v1 . . . vn ]  .
 
.. .. ..  α
| {z }  .. . . . 
α
0 0 . . . eλn t
 
α1 n
X
x(t) = [v1 eλ1 t . . . vn eλn t ]  ...  = αi eλi t vi
 

αn i=1
x(t)
At
• Instead of studying e we can now study e λ i t

• eλi t are called modes of the system

Example: how does x(t) evolve over time if x(t) = 2e−3t + 7e5t ?

Fall - 2022 27 / 58
 Classes of Dynamical Systems
• Besides Linear Systems, other classes of dynamical systems exist

Fall - 2022 28 / 58

Other classes of dynamical systems (1/2)

• Let us consider the following linear continuous-time state-space

system
ẋ(t) = A(t)x(t) + B(t)u(t)
y(t) = C(t)x(t) + D(t)u(t)

• If A, B, C, D are all constant, the system is said linear-time invariant

(LTI)

LT I : P = {A, B, C, D}

• If any matrix A, B, C, D is not constant, the system is said linear-time

varying (LTV)

LT V : P(t) = {A(t), B(t), C(t), D(t)}

Fall - 2022 29 / 58

Other classes of dynamical systems (2/2)

• The system
ẋ(t) = f (x(t), u(t))
y(t) = g(x(t), u(t))
with f : IRn+m → IRn , g : IRn+m → IRp nonlinear functions, defines
a nonlinear time-invariant system (NLTI).

N LT I : P = {f (x, u), g(x, u)}

• The system
ẋ(t) = f (t, x(t), u(t))
y(t) = g(t, x(t), u(t))
defines a nonlinear time-varying system (NLTV).

N LT V : P(t) = {f (t, x, u), g(t, x, u)}

Fall - 2022 30 / 58
 Equilibrium Points and Stability

Fall - 2022 31 / 58

Equilibrium points for nonlinear systems

• What is an equilibrium point?

• Let us consider a NLTI system

ẋ(t) = f (x(t), u(t))

y(t) = g(x(t), u(t))

Definition
A state xe ∈ IRn and an input ue ∈ IRm are an equilibrium pair if for an
initial condition x(0) = xe and constant input u(t) ≡ ue , the state
remains constant, i.e.,
x(t) ≡ xe , ∀t ≥ 0

Fall - 2022 32 / 58

How to find equilibrium points?

ẋ(t) = f (x(t), u(t))
y(t) = g(x(t), u(t))

Definition
A state xe ∈ IRn and an input ue ∈ IRm are an equilibrium pair if

f (xe , ue ) = 0 =⇒ ẋ(t) = 0 ∀t

Definition
If xe ∈ IRn and an input ue ∈ IRm are an equilibrium pair then
• xe is called the equilibrium state
• ue is called the equilibrium input

• For nonlinear systems, we can have 0,1,..many...infinitely many

equilibrium pairs

Fall - 2022 33 / 58
Stability of Equilibrium Points

• How does the system behave when we start from an initial

condition non exactly on the equilibrium point?

What will happen if the system is on one equilibrium point and

an attack create a little perturbation?

Fall - 2022 34 / 58

Stability of Equilibrium Points

ẋ(t) = f (x(t), u(t))

y(t) = g(x(t), u(t))

• Different equilibria can have different stability

• Stability, in general, is not a system property, but it refers to a
specific equilibrium pair (local)
• The equilibrium pair (xe , ue ) can be
• Stable/Marginally Stable
• Asymptotically Stable
• Unstable

Fall - 2022 35 / 58

Stable and asymptotically equilibrium points stable6

Figure: Stable

Figure: Asymptotically stable

6
https://www.math24.net/stability-theory-basic-concepts/
Fall - 2022 36 / 58
 Equilibrium points and stability for LTI systems

Fall - 2022 37 / 58

Equilibrium points of linear systems

(
ẋ(t) = Ax(t) + Bu(t)
y(t) = Cx(t) + Du(t)
Equilibrium points:
• Pairs (xe , ue ) such that Axe + Bue = 0
• (xe , ue ) = (0, 0) is an evident equilibrium, valid for any linear
system

Fall - 2022 38 / 58

Stability of linear systems

(
ẋ(t) = Ax(t) + Bu(t)
y(t) = Cx(t) + Du(t)
Stability of Equilibrium Points:
• All the equilibria share the same type of stability (system’s
property)
• Stability/Instability depends only on the eigenvalues of A
Theorem
a Let λ1 , . . . , λm , m ≤ n be the eingenvalues of A ∈ IRn×n .
• Asymptotically stable if and only if (iff) Re[λi ] < 0, ∀i
• Stable or Marginally Stable if Re[λi ] ≤ 0, ∀i and the eigenvalues
with Re[λi ] = 0 have equal algebraic and geometric multiplicity
• Unstable if ∃ i such that Re[λi ] > 0 or ∃ Re[λi ] = 0 with algebraic
multiplicity different from the geometric multiplicity.
a
Re[λi ] = real part of the complex number λi

Fall - 2022 39 / 58
Example - Mass-Spring-Damper
• Check the stability of the mass-spring-damper system if M = 25,
K = 24, and β = 8

   
0 1 0
ẋ(t) = k β x(t) + 1 u(t)
−M −M M
| {z } | {z }
A B
 
y(t) = 1 0 x(t)
| {z }
C

Fall - 2022 40 / 58

Example - Mass-Spring-Damper: Eigenvalues

• By using the MATLAB command “eig(A)” we obtain V and D :
• V, reading by column, contains the eigenvectors
• D, on the diagonal, contains the eigenvalues

Fall - 2022 41 / 58

Example - Mass-Spring-Damper

• By using MATLAB we have found the eigenvalues:

λ1 = −0.16 + 0.96i, λ2 = −0.16 − 0.96i

• Since all the eigenvalues have a negative real part (e.g., −0.16)
we can conclude that the system is asymptotically stable

Fall - 2022 42 / 58
Mass-Spring-Damper: Simulation

Starting from any initial state x(0) = [1, 0]T and for a constant input
u(t) = 1, ∀ t ≥ 0, the mass-spring-damper system will asymptotically
settle down (after a transitory period) to an equilibrium point

Fall - 2022 43 / 58

Stability of equilibrium points of NLTI systems

Fall - 2022 44 / 58

Stability of equilibrium points of nonlinear systems

ẋ(t) = f (x(t), u(t))

y(t) = g(x(t), u(t))

• Stability is hard to prove. For different equilibrium points, we can

have different stability results.
• Methods to prove stability of equilibrium points:
• Lyapunov indirect method
• Lyapunov direct method (we will not see this)
Fall - 2022 45 / 58
 Lyapunov Indirect Method

Fall - 2022 46 / 58

Lyapunov indirect method: Idea

ẋ(t) = f (x(t), u(t))

• We know how to check stability for Linear Systems (i.e., looking at

the eigenvalues of A)
• We can approximate the non-linear system as a linear system
around the equilibrium pair (xe , ue ) by using the Taylor expansion
truncated to the first order
• We study the stability of the linearized model and try to infer the
(local) stability of the original nonlinear system around the used
equilibrium point.

Fall - 2022 47 / 58

Linearization of a non-linear static and scalar function

• If x(t) is a scalar, then the linearization of a non-linear static

function y(t) = f (x(t)) around x = x0 , using a Taylor expansion
truncated to the first-order derivative, is:

df (x(t)) 
y(t) ≈ f (x(0)) + (x(t) − x0 )
dx(t) x(t)=x0

Fall - 2022 48 / 58
Linearization of nonlinear dynamical state-space
models around an equilibrium pair
• Given a nonlinear state-space model (set of n nonlinear
differential equation of order 1):

 f1 (x(t), u(t))

ẋ(t) = ..
 .

fn (x(t), u(t))
where x = [x1 , . . . , xn ]T , u = [u1 , . . . , um ]T
• The linearized model around an equilibrium pair (xe , ue ) is
˙
x̃(t) = Al x̃(t) + Bl ũ(t)
 ∂f1 ∂f1   ∂f1 ∂f1 
∂x1 ... ∂xn ∂u1 ... ∂um

Al =  .. ..  
Bl =  .. .. 
. ... .  . ... . 
∂fn ∂fn ∂fn ∂fn
∂x1 ... ∂xn (xe ,ue ) ∂u1 ... ∂um (xe ,ue )
where x̃ = x − xe and ũ = u − ue
Fall - 2022 49 / 58

Linearized Model vs Nonlinear Model

• Consider a NLTI system described by a set of n nonlinear

differential equations of order 1 :

ẋ(t) = f (x(t), u(t))

and its linearization around (xe , ue ),

˙
x̃(t) = Al x̃(t) + Bl ũ(t)

• Question:

Can we understand the stability of ẋ(t) = f (x(t), u(t)) around

˙
(xe , ue ) by studying the stability of x̃(t) = Al x̃(t) + Bl ũ(t)?

Fall - 2022 50 / 58

Lyapunov indirect method: Theorem

Theorem
Consider a NLTI system ẋ = f (x), with f differentiable, and its
linearization
˙
x̃(t) = Al x̃(t)
obtained considering the equilibrium point xe = 0.
• If x̃˙ = Al x̃ is asymptotically stable, then x = 0 is a locally
asymptotically stable equilibrium point for the NLTI system
• If x̃˙ = Al x̃ is unstable, then x = 0 is an unstable equilibrium for
the NLTI system
• If x̃˙ = Al x̃ is marginally stable, nothing can be said about the
stability of x = 0 for the NLTI system

• Note: If we are in the third case, the indirect Lyapunov method

fails to assess stability. Other tools must be used (e.g., Lyapunov
direct method)
Fall - 2022 51 / 58
Example - Pendulum

The non-linear state-space representation of the autonomous

pendulum system is:7
 
x2 (t) h
ẋ(t) = g , H :=
− l sin x1 (t) − Hx2 (t) ml2
   
x1 (t) = y(t) x2 (t)
where x(t) = , f (x(t)) =
x2 (t) = ẏ(t) − gl sin x1 (t) − Hx2 (t)

7
g is the force per unit mass due to gravity
Fall - 2022 52 / 58

Pendulum - Equilibrium Points

• The equilibrium points of the autonomous pendulum system are

obtained by the following condition (f (xe ) = 0):
    
x2 0 x 2e = 0
= ⇒
− gl sin x1 − Hx2 0 x1e = ±kπ, k = 0, 1, . . .

Fall - 2022 53 / 58

Pendulum - Linearized Model

 
x2 (t)
• Nonlinear Model ẋ(t) = g
− l sin x1 (t) − Hx2 (t)
• Equilibrium points xe = [±kπ, 0]T
• Linearized Model x̃˙ = Al x̃(t), where
" #
∂f1 ∂f1
Al = ∂x1 ∂x2
∂f2 ∂f2
∂x1 ∂x2 xe
" ∂x2 ∂x2
#
∂x1 ∂x2
= ∂(− gl sin x1 −Hx2 ) ∂(− gl sin x1 −Hx2 )
∂x1 ∂x2 xe =[±kπ, 0]T
 
0 1
=
− gl cos x1 −H xe =[±kπ, 0]T
 
0 1
=
− gl cos(±kπ) −H
Fall - 2022 54 / 58
Pendulum - Stability of first kind of equilibrium points

First case x2e = 0 and

x1e = kπ, k = 0, 2, 4, . . . ⇒
Linearized Model:
 
˙x̃(t) = 0 1
x̃(t)
− gl −H

• The eigenvalues of the linearized system are

 r 
2 g 1 2
g
det(λI − Al ) = λ + Hλ + = 0 ⇒ λ1,2 = −H ± H − 4 )
l 2 l

• Re[λ1,2 ] < 0 ⇒ the equilibrium is asymptotically stable

Fall - 2022 55 / 58

Pendulum - Stability of second kind of equilibrium

points

Second case x2e = 0 and

x1e = kπ, k = 1, 3, 5, . . . ⇒
Linearized Model:
 
˙ 0 1
x̃(t) = g x̃(t)
l −H

• The eigenvalues of the linearized system are

 r 
2 g 1 2
g
det(λI − Al ) = λ + Hλ − = 0 ⇒ λ1,2 = −H ± H + 4 )
l 2 l

• Re[λ1 ] < 0 and Re[λ2 ] > 0 ⇒ the equilibrium is unstable

Fall - 2022 56 / 58

Thank you!

Fall - 2022 57 / 58
References I

Jean-Jacques E Slotine, Weiping Li, et. al

Applied nonlinear control
Prentice hall Englewood Cliffs, NJ 199(1), 1991.

Panos J Antsaklis, and Anthony N Michel

A linear systems primer
Birkhäuser Boston, 2007.

Emilio Frazzoli, and Munther Dahleh

Dynamic Systems and Control
Massachusetts Institute of Technology: MIT OpenCourseWare, spring 2011.

Alberto Bemporad
Lecture Slides: Identification, Analysis and Control of Dynamical Systems
http://dysco.imtlucca.it/courses/iacds/, 2018.

Fall - 2022 58 / 58