CHAPTER 0 - Preliminaries

FREQUENCY DOMAIN ANALYSIS OF

DYNAMIC SYSTEMS

GRACE S. DEAECTO

DMC / School of Mechanical Engineering

UNICAMP, 13083 - 860, Campinas, SP, Brazil,
[email protected]

Campinas, Brazil, February 2014

1 / 26
CHAPTER 0 - Preliminaries

Contents

1 CHAPTER 0 - Preliminaries
Partial-fraction decomposition
State space realization
Bode diagram
Linear matrix inequalities

2 / 26
CHAPTER 0 - Preliminaries

Partial-fraction decomposition

Partial-fraction decomposition
Consider a linear continuous-time invariant system with
transfer function
Pm
ẑ(s) bi s i
H(s) = = Pi=0
n i
ŵ (s) i=0 ai s

with n ≥ m, ai ∈ R, ∀i = 0, · · · , n, an = 1 and
bi ∈ R, ∀i = 0, · · · , m. Whenever n = m, the function is said
to be proper and for n > m it is said to be strictly proper.
The partial-fraction decomposition consists of determining the
scalars αi such that
Pm M
bi s i X αi
Pi=0n i
= α 0 +
i=0 ia s (s − pi )ni
i=1
PM
where i=1 ni = n and pi are the poles with multiplicity ni .
3 / 26
CHAPTER 0 - Preliminaries

Partial-fraction decomposition

Partial-fraction decomposition

Decompose in partial-fraction the following rational functions

s +1
(s + 2)(s + 3)(s + 5)

s +1
(s + 2)2 (s + 3)

s +1
(s + 2)3 (s + 3)
s
(s + 1)(s 2 + 2s + 2)

4 / 26
CHAPTER 0 - Preliminaries

State space realization

State space realization

The transfer function H(s) can be written as the following
differential equation of n order
n m
X diz X diw
ai i (t) = bi i (t), ∀t ≥ 0
dt dt
i=0 i=0
with zero initial conditions. If the initial conditions are not
zero, then, applying the Laplace transform on both sides, we
obtain
ẑ(s) = H0 (s) + H(s)ŵ (s)
where H0 (s) depends only on the initial conditions. Let us
denote
n m
X diz X diw
D[z] = ai i (t) , N[z] = bi i (t)
dt dt
i=0 i=0
differential operators.
5 / 26
CHAPTER 0 - Preliminaries

State space realization

State space realization

Notice that any differential linear equation of order n can be

written as n differential equations of first order. This set of
equations is named state space realization of the original
differential one and has the matrix form

ẋ(t) = Ax(t) + Bw (t), x(0) = x0

z(t) = Cx(t) + Dw (t)

where A ∈ Rn×n , B ∈ Rn×1 , C ∈ R1×n and D ∈ R1×1 . Matrices

(A, B, C , D) and the initial condition x0 ∈ Rn must be determined
such that the function y (t) obtained from the state space
realization is identical to the one obtained from the differential
equation.

6 / 26
CHAPTER 0 - Preliminaries

State space realization

State space realization

For the equation D[z] = w with an = 1, defining the state

variables
 
x1 (t)
x(t) =  ...  , xi (t) = z (i−1) (t), i = 1, · · · , n
 

xn (t)

we obtain
 
  0
0 1 0 ··· 0  0
 0 0 1 ··· 0  .
 
A= .. .. .. .. ..  , B =  .. 
 
 . . . . .   
 0
−a0 −a1 −a2 · · · −an−1
1
 
C= 1 0 0 · · · 0 , D = [0]
7 / 26
CHAPTER 0 - Preliminaries

State space realization

State space realization

For the equation D[z] = N[w ] with N[·] a differential operator
of order m ≤ n − 1, we can rewrite
D[ξ] = w , z = N[ξ]
Defining the state variables
 
x1 (t)
x(t) =  ...  , xi (t) = ξ (i−1) (t), i = 1, · · · , n
 

xn (t)
matrices A and B are the same. Moreover,
m
X m
X
z(t) = E [ξ] = bi ξ (i) (t) = bi xj+1 (t) =
i=1 i=1
allows us to determine
 
C = b0 b1 b2 · · · 0 , D = [0]
8 / 26
CHAPTER 0 - Preliminaries

State space realization

State space realization

Obtain the state space realization of the following systems

s 2 + 5s + 3
H(s) =
s(s 2 + 5s + 6)

s 2 + 0.1s
H(s) =
s 2 + 0.1s + 10

Show that for an arbitrary nonsingular matrix T ∈ Rn×n the

state space realization (T −1 AT , T −1 B, CT , D) also represents
the transfer function H(s) with realization (A, B, C , D).

9 / 26
CHAPTER 0 - Preliminaries

Bode diagram

Bode diagram

The frequency response of an asymptotically stable system

with transfer function H(s) is simply given by H(jω).
It can be represented graphically by diagrams. The most used
is the Bode diagram of modulus and phase

A(ω)dB × log(ω), ∀ ω > 0

φ(ω) × log(ω), ∀ ω > 0

where A(ω)dB is the modulus of H(jω) in decibels and φ(ω) is
the phase in degrees or radians.
The modulos of H(jω) expressed in decibels is defined as

A(ω)dB = 20log(|H(jω)|)

10 / 26
CHAPTER 0 - Preliminaries

Bode diagram

Bode diagram
Bode diagram can be calculated numerically without any
difficulty. However, we can obtain important information
about the system by analysing an approximated Bode diagram
obtained from the asymptotes. Indeed, notice that writing
Qm
(γi s + 1)
H(s) = κ Qi=1
n
i=1 (τi s + 1)
for instance, we obtain
m
X n
X
A(ω) = κdB + |γi s + 1|dB − |τi s + 1|dB
i=1 i=1
m
X n
X
φ(ω) = ∠(γi s + 1) − ∠(τi s + 1)
i=1 i=1
for γi > 0, i = 1, · · · , m and τi > 0, i = 1, · · · , n.
11 / 26
CHAPTER 0 - Preliminaries

Bode diagram

Bode diagram
1
Considering that H(s) = τ s+1 , we obtain
q
|H(jω)|dB = −20log (τ ω)2 + 1

where for ω >> 1/τ we have

|H(jω)|dB ≈ − 20log(τ ω) = −20log(ω) − 20log(τ )

∠H(jω) = −90o
and for ω << 1/τ we have

|H(jω)|dB ≈ 0, ∠H(jω) = 0

At the frequency ωc = 1/τ occurs the intersection of both

asymptotes. This frequency is named cutoff frequency.
12 / 26
CHAPTER 0 - Preliminaries

Bode diagram

Bode diagram

For second order systems

ωn2
H(s) =
s 2 + 2ξωn s + ωn2

the same idea can be applied, that means, we can obtain the
asymptotes by making ω << ωn and ω >> ωn . However,
generally the approximation for frequencies near to the cuttof
one ωc = ωn may be not good because it depends sensibly on
the damping ratio ξ.

13 / 26
CHAPTER 0 - Preliminaries

Bode diagram

Phase and gain margins

Consider a closed-loop linear system with characteristic

equation 1 + C (s)G (s) = 0 where G (s) represents the system
to be controlled and C (s) the controller. Defining
H(s) = C (s)G (s) we can calculate, by using the Bode
diagram, the following margins:
Gain Margin (GM): In the the phase diagram, determine the
frequency ωf such that ∠H(ωf ) = −180o . By using this
frequency, the modulus diagram provides

GM = −|H(jωf )|dB

Phase Margin (PM): In the modulus diagram, determine the

frequency ωg such that |H(ωg )|dB = 0 dB. By using this
frequency, the phase diagram provides

PM = 180o + ∠H(jωg )

14 / 26
CHAPTER 0 - Preliminaries

Bode diagram

Bode diagram
Consider the asymptotic modulus Bode diagram of a
minimum phase system presenting one pair of complex poles
with ξ = 0.5 given by
18dB

20dB/dec −40dB/dec

20dB/dec

2 rad/s 4 rad/s 16 rad/s 20 rad/s

15 / 26
CHAPTER 0 - Preliminaries

Bode diagram

Bode diagram

Determine the approximated phase Bode diagram.

Determine the transfer function H(s).

Calculate the phase and gain margins (PM and GM).

Provide the output z(t) in steady state for an input

w (t) = 4 sin(8t).

16 / 26
CHAPTER 0 - Preliminaries

Linear matrix inequalities

Linear matrix inequalities

Linear matrix inequalities (LMIs) are essential in the analysis

and control design of dynamical systems and to several
optimization problems.
Linear Matrix Inequality
An LMI is expressed as
A(x) < 0
with
n
X
A(x) = A0 + A i xi
i=1

where Ai ∈ Rm×m , i = 0, · · · , n are symmetric matrices and

xi ∈ R is the i-th component of vector x.

17 / 26
CHAPTER 0 - Preliminaries

Linear matrix inequalities

Linear matrix inequalities

Notice that A(x) : Rn → Rm×m is a linear function of the
vector x ∈ R.
Convex set
The set of vectors x ∈ R satisfying the linear matrix inequality
A(x) < 0 is convex.

Indeed, notice that for two generic points xa , xb ∈ Rn the

segment between them is x = αxa + (1 − α)xb for 0 ≤ α ≤ 1.
Assuming that A(xa ) < 0 and A(xb ) < 0, we have
A(x) = A(αxa + (1 − α)xb )
= αA(xa ) + (1 − α)A(xb )
< 0
where the second equality is due to the fact that A(x) is
linear.
18 / 26
CHAPTER 0 - Preliminaries

Linear matrix inequalities

Linear matrix inequalities

An important result used to linearise some nonlinear

constraints is the Schur Complement.
Schur Complement
A linear matrix inequality
 
S(x) V (x)
A(x) = <0
V (x)′ Q(x)

is equivalent to any of the two nonlinear inequalities

a) S(x) < 0 and Q(x) − V (x)′ S(x)−1 V (x) < 0
b) Q(x) < 0 and S(x) − V (x)Q(x)−1 V (x)′ < 0

19 / 26
CHAPTER 0 - Preliminaries

Linear matrix inequalities

Linear matrix inequalities

Indeed for part a), notice that S(x) < 0 also implies that
S(x)−1 . As a consequence, matrix
 
I 0
U(x) =
V (x)′ S(x)−1 I

is nonsingular and allows us to write A(x) = U(x)B(x)U(x)′ ,

where
 
S(x) 0
B(x) =
0 Q(x) − V (x)′ S(x)−1 V (x)

Hence matrix A(x) < 0 if and only if B(x) < 0. The proof of
part b) is similar.

20 / 26
CHAPTER 0 - Preliminaries

Linear matrix inequalities

Linear matrix inequalities

Example 1: Convert the linear inequalities 2x1 + 3x2 < 7,

−x1 + x2 < 5 and 2x1 − 4x2 < −4 in a matrix form.
Answer:  
−7 0 0
A0 = 0
 −5 0
0 0 5
 
2 0 0
A1 = 0 −1 0
0 0 2
 
3 0 0
A2 = 0 1 0
0 0 −4

21 / 26
CHAPTER 0 - Preliminaries

Linear matrix inequalities

Linear matrix inequalities

Example 2: Convert the nonlinear inequality
(x1 − 1)2 + 2(x2 − 2)2 < 52 , which is an ellipse with focus in
(1,2), in a linear matrix inequality.
Answer:
Performing the Schur Complement, we have that it is
equivalent to
2(x2 − 2)2 − 25 x1 − 1
 
<0
x1 − 1 −1
performing it again, we obtain
 
−25 x1 − 1 x2 − 2
 x1 − 1 −1 0 <0
x2 − 2 0 −1/2
where matrices A0 , A1 , A2 ∈ R3×3 can be directly determined
22 / 26
CHAPTER 0 - Preliminaries

Linear matrix inequalities

Linear matrix inequalities

The concepts we have just presented are important to solve

optimization problems described as

inf {c ′ x : A(x) < 0}

x

where c ∈ Rn .
In the specific context of control design, two very important
problems can be written as the optimization problem just
presented.
The first one is the H2 norm calculation kHk22 where
H(s) = C (sI − A)−1 B.

23 / 26
CHAPTER 0 - Preliminaries

Linear matrix inequalities

Linear matrix inequalities

The H2 norm of H(s) is equal to

kHk22 = Tr(B ′ PB)

where P > 0 is the solution of the Lyapunov equation

A′ P + PA + C ′ C = 0. It can calculated through the solution
of the optimization problem

kHk22 = inf {Tr(B ′ XB) : A′ X + XA + C ′ C < 0}

X >0

or, alternatively, through

kHk22 = inf {Tr(CYC ′ ) : AY + YA′ + BB ′ < 0}

Y >0

24 / 26
CHAPTER 0 - Preliminaries

Linear matrix inequalities

Linear matrix inequalities

The second problem of great importance is the H∞ norm.

It can be shown that the inequality kHk2∞ < µ holds if and
only if there exist P > 0 and µ > 0 satisfying the Riccati
inequality

A′ P + PA + C ′ C + (B ′ P + D ′ C )′ (µI − D ′ D)−1 (B ′ P + D ′ C ) < 0

which, applying the Schur Complement can be converted in

the following optimization problem
  ′  
2 A P + PA + C ′ C PB + C ′ D
kHk∞ = inf µ : <0
P>0,µ>0 B ′P + D ′C D ′ D − µI

which is a numerical procedure very efficient for this norm

calculation.
25 / 26
CHAPTER 0 - Preliminaries

Linear matrix inequalities

Linear matrix inequalities

Given a system with transfer function

s +2
H(s) =
s 3 + 2.4s 2 + 2.8s + 0.8

Obtain the system state space realization.

Using the LMILAB from Matlab, solve the optimization

problems already provided, in order to calculate H2 and H∞
norms.

Compare the results with the ones obtained by the commands

“normh2” and “normhinf” from Matlab.

26 / 26