Chapter 1

Chapter 1
Introduction
Systems with static nonlinearities
Frequency-domain methods
Einführung, Systeme mit nichtlinearen Kennlinien,

Frequenzbereichsmethoden
Prof. Dr.-Ing. Steven Liu Nonlinear and Adaptive Control

SS 2023
Lehrstuhl für Regelungssysteme 1-1
Technische Universität Kaiserslautern Version 08.04.2023
General models of nonlinear systems
A nonlinear dynamic system can be generally described by the state-space differential equation
x (t )  f  t , x(t ), u (t )  (1.1)
with x as system state and u as system input
 x1   u1   f1 (t , x, u ) 
x  u   f (t , x, u ) 
x   2 ; u   2 ; f (t , x , u )   2 .
  
x  u   f (t , x, u ) 
 n  p  n 
In many cases we also need the output equation
y (t )  h  t , x (t ), u(t )  (1.2)
with y as system output and
 y1   h1 (t , x, u ) 
 y2   h2 (t , x, u ) 
y    ; h(t , x, u )   .
 
y   h (t , x, u ) 
 m  m 
SS 2023
Nonlinear system models in reduced form
If the state equation does not have explicit dependence of the input u, then we have an unforced
state equation:
x (t )  f  t , x (t )  (1.3)
It does not necessarily mean that there is no input to the system, since the input could be specified
as a given function of time contained in (1.3). It is also possible that the input u in (1.1) is generated
as a state feedback u(x,t) and therefore does not explicitly appear in (1.3). In this case (1.3)
describes the behavior of the closed-loop control.
A special case of (1.3) arises when the function f even does not depend explicitly on time t:
x  f ( x ) (1.4)
The system is then said to be autonomous or time-invariant. The solution of (1.4) in dependence of
the initial state is called eigentrajectory.
Unless explicitly otherwise specified we will keep our focus on autonomous systems only.

SS 2023
Equilibrium points
An extremely important concept in treating nonlinear control systems is that of an equilibrium point.
Definition 1.1 A point x = xR in the state space is said to be an equilibrium point of (1.3) if the
following property holds
x R  f  x R , t   0 t  0 (1.5)
It means that whenever the state of the system reaches the equilibrium point xR, it will remain at xR
for all future time, since no state changes can further occur.
In practice a nonlinear system will be often treated based on a linearized model, while the
linearization is usually done around some nominal operating point (mostly an equilibrium point)
Whenever possible and reasonable, it is also the preferable way since then the complete powerful
methods of system analysis and synthesis can be utilized. However, linearization alone is not enough
due to two basic limitations of linearization:
• the linearization is only a local approximation and not valid for "large signal behavior"
• the dynamics of a nonlinear system are much richer and more complex than those of a linear
system. There are essentially nonlinear phenomena that can take place only in the presence of
nonlinearity.

SS 2023
Phenomena of nonlinear dynamics (1)
Some essentially phenomena of nonlinear systems that cannot be described by linear system theory
are
• Finite escape time: The state of an unstable linear system goes to infinity as time approaches
infinity t  ∞; the state of a nonlinear system, however, can go to infinity in finite time.
• Multiple isolated equilibria: A linear system can have only one isolated equilibrium point; a
nonlinear system, however, can have one, multiple or infinitely many isolated equilibrium points.
• Input-depending stability: The stability of the equilibrium point is not always a system property,
but can be input-dependent.
• Irregular oscillations (Harmonics, subharmonics, almost-harmonics): A linear system responds to a
sinusoidal excitation always in form of a harmonic oscillation of same frequency. For nonlinear
systems oscillations of different frequencies can occur.
• Limit cycles: Marginally stable oscillations exist in linear systems only theoretically. They are
extremely parameter-sensitive and their amplitude depends on initial state only. A stable limit cycle
in nonlinear systems is of fixed amplitude and frequency, irrespective of the initial state and robust
against parameter variations or disturbances.

SS 2023
- Finite escape time example: x   x 2 with x (0)  0.5
dx dx 1
dt
  x2    x 2   dt  x  t  C
x (0)  0.5  C  2
x(t)
1 2 3
1 t
Solution: x (t )  0
t2 1
2
3
4
5
- Multiple isolated equilibria examples:
Linear systems:
1) d  x1   2 1  x1  2) d  x1   2 1  x1 
 
dt  x2 
 1 5  x2  dt  x2 
10 5  x2 
x  x   x 
  1    
0
  1   1 
 x2  R  0   x2  R  2 x1 
x2 unique
x2 connected
equilibrium equilibria
x1 x1
Nonlinear system:
d  x1   2 x2 
 
 x1 
  0,  π,  2π,
dt  x2   2sin x1   x2 
R
 0 
x2
multiple isolated equilibria
x1
- Input dependent stability example: x  ux with x (0)  1
x(t) x(t)
t t
u = const. > 0 u = const. < 0

- Irregular oscillations example:
Linear system: Nonlinear system:
A sin( t) A G (j ) sin  t   (j ) 

x  (2  3 x 2 )u
G(s)
No general closed-form solution!
Special solution for u  A sin( t) :

Linear frequency response:
Purely sinusoidal
A
2 6 (1 cos t )
2 e 
1
Same frequency, different x (t )   A2 6 (1 cos t )
amplitudes and phases 6 e 1
Periodic, mixed frequency oscillations
+ DC offset
A further example for “frequency response”:
x1   x1  x22 , y  x1
x 2   x2  u , u  A sin(t )
The solution for t   (stationary response) can be easily obtained:

A
x2  [sin(t )   cos(t )]
 1
2
A2
x1  
2 2 
2 2
( 2
 1)sin 2
( t )  2 ( 2
 2)sin(t ) cos(t )
(1  4 )(1   )
2
  2 (2 2  5) cos 2 (t ) 

A2 A2
  
2 2 
2 2
( 2
 2) sin(2 t )  (5 2
 1) cos(2t ) 
2(1   ) 2(1  4 )(1   )
2 2
i.e., for an excitation with frequency , the output will contain

- One DC component
- One 2-component
but no -component!
Phenomena of nonlinear dynamics (2)
• Chaos: A nonlinear system can have a more complicated steady-state behavior that is neither
equilibrium, nor periodic or almost-periodic oscillation. It is irregular and exhibits some “stochastic”
characters, despite the deterministic nature of the system, i.e., the trajectories never settle down
to fixed points or to period orbits. Such behavior is usually referred to as chaos.
• Bifurcations: In some nonlinear systems a (small) quantitative parameter variation can lead to
large qualitative changes of system behaviors.
• Multiple modes of behavior: While a linear system can only show some fixed characteristic mode
(periodic oscillation, exponential decay etc.) at once, it is not unusual for two or more modes of
behavior to be exhibited by the same nonlinear system with continuous or even discontinuous
change-overs between the modes.

SS 2023
A chaotic
system
Source: https://doi.org/10.1155/2017/7871467
System with
bifurcation
Source: https://doi.org/10.1155/2017/7871467
Which concepts are no more valid?
• The principle of superposition does not hold any more

• Eigenvalue analysis is no more possible, since no eigenvalues can be defined for nonlinear systems
• System response cannot be determined using convolution integrals and no formal definition of
step or impulse response available
• No application of Laplace transform is possible, transfer function cannot be generally defined
• Frequency response is not generally defined
• Bibo-stability is not necessarily a system property only. Stability analysis based on Nyquist plot or
pole locations is not generally possible.
• Controllability and observability cannot be easily proven using simple matrix criteria
• Classical design methods for control (Bode diagram and root locus approaches) cannot be used for
nonlinear systems
Completely new methods are necessary!

SS 2023
Which concepts exist for nonlinear systems?
• There is no closed theory for nonlinear dynamic systems. General concepts are normally difficult to
concretize and often not easy to use
• For special classes of nonlinear systems there exist special analysis methods that lead to practically
useful algorithms
• “Classical concepts” are rather analysis tools than synthesis methods
• “Modern concepts” (since approximately 1990s) are often based on differential-algebraic or
differential-geometric theories and provide concrete design methods for nonlinear control. The
mathematical complexity behind these methods is, however, sometime pretty high
Methods need to be appropriately chosen depending on the required application;

Numerical tools are often useful.
There are still large needs of intensive research!

SS 2023
Which analysis methods for which behavior?
In part a) of this lecture notes following system behaviors will be treated. The analysis methods
provided belong to the “classical” methods of nonlinear system theory and can be, in principle, also
applied for control design.
• Periodic behavior of nonlinear systems, especially limit cycles. Many systems with static
nonlinearities exhibit such a behavior.
Analysis method: harmonic linearization (Describing function)
• Absolute stability of systems with static nonlinearities, especially those with sector functions
Analysis methods: Popov criterion, circle criterion
• State-space stability of equilibria in the sense of Lyapunov for systems with general dynamic
nonlinearities
Analysis methods: Lyapunov theory, invariance principle
We start our discussions with consideration of static nonlinearities.

SS 2023
Examples of static nonlinearities (1)
Exp. 1.1 Oscillator with a tunnel-diode iD

u
Tunnel-diodes show generally a nonlinear characteristic iD = h(uD). h
With x1 = uC = uD and x2 = i we obtain the state equation
1
x1  [h( x1 )  x2 ]
C uD
1
x 2  ( x1  Rx2  u ) .
L
To determine the equilibria we solve the equations
 h( x1 )  x2  0 i R L
 x1  Rx2  u  0
iC iD
together with the equation
1 u C
h( x1 )  (u  x1 ) . D
R
Obviously, there are up to three solutions possible.
In which equilibrium can the circuit work stably? 

SS 2023
Exp. 1.2 Negative-resistance oscillator iD

The anti-parallel connection of tunnel-diodes leads to a h
nonlinear characteristic iD = h(uD) and operates like a negative
resistance in the low-voltage region.
uD
From the Kirchhoff's laws we have
LCuD  uD  Lh(uD )u D  0 .
Applying the substitution   t / LC and the parameter
  L/C it yields
uD   h(uD )u D  uD  0 . iD
Will the nonlinear characteristic be approximated using iC iL
h(uD )  uD  13 u D3 , C L uD
then the system model is given by
uD   (1  uD2 )u D  uD  0 ,
which is in the general form known as Van-der-Pol equation.

SS 2023
By defining the state variables x1  uD and x2  u D the only equilibrium (0, 0) of the system can be
determined from the following equations:
x1  x2  0
x 2   x1   h( x1 ) x2  0
Of course, other definitions of state variables such as z1 = iL and z2 = uC are also possible. In this case
the system model is described by
1 1
z1  z2 z1  z2
L or  with respect to the time variable .
1 z2   [ z1  h( z2 )]
z2   [ z1  h( z2 )]
C
Both state models are equivalent and can be derived from each other by using the relationship
 z1     h( x1 )  (1/  ) x2 
 z2   x1 


SS 2023
Exp. 1.3 Mass-spring system with frictions and nonlinear spring characteristics
Different frictions can exist for the considered system:
static friction = adhesion
F
Coulomb friction = sliding friction m
y
viscous friction = air drag
FC FC FC FC
 s mg
 k mg kV
y y y y
  k mg
  s mg
Coulomb friction Coulomb friction + Static + Coulomb + linear Static + Coulomb + linear
linear viscous friction viscous friction viscous friction with
Stribeck effect
SS 2023
In the linear range (relatively small displacement) a spring obeys Hook’s law. For large
displacements or certain materials, however, the restoring force may depend nonlinearly on the
spring displacement, like in case of softening springs (e.g. packaging cushions with high elasticity
and small deflection) where the spring restoring force can be determined by the equation
FS ( y )  kS (1  a 2 y 2 ) y , ay  1
Some springs may also have “progressive” characteristics and are said to be hardening springs (e.g.
rubber in compression). The spring force can be described then by
FS ( y )  kS (1  a 2 y 2 ) y .
A combination of a (hardening) spring, linear viscous friction, and a periodic external force results,
without consideration of sliding and static frictions, in the equation
my  kV y  kS y  kS a 2 y 3  A cos t .
The generalized form of the above equation for arbitrary excitation is known as Duffing’s equation.

SS 2023
A model description as autonomous system without Stribeck effect is:

 k mg sgn( y ) y  0
my  kV y  kS y   ( y, y )  0 mit  ( y , y )   kS y y  0, y   s mg / kS
  s mg sgn( y ) y  0, y   s mg / kS
With the state variables x1  y and x2  y a first-order state-space model can be given
x1  x2
kS k 1
x 2   x1  V x2   ( x1 , x2 )
m m m
Which are equilibria of the system?

SS 2023
Exp. 1.4 Transmission with gap (backlash)

driving driven
In gear trains, there always exist small gaps between a gear gear
pair of mating gears. As a result of the gaps, when the
driving gear rotates a smaller angle than the gap a, the a
driven gear does not move at all, which corresponds to
the dead-zone OA; after contact has been established
between the two gears, the driven gear follows the
rotation of the driving gear in a linear fashion (AB).
When the driving gear rotates in the reverse direction
by a distance of 2a, the driven gear again does not y
move (BC). After the contact between the two gears is
re-established, the driven gear follows the rotation of C
the driving gear in the reverse direction (CD). B
Therefore, if the diving gear is in periodic motions, the O A
-a a u
driven gear will move in the fashion represented by the D
closed path EBCD. Note that the height of B, C, D, E E
depends on the amplitude of the input sinusoidal. 

SS 2023
y y y y
1 1 k
u  u  u  u
-1
relay relay with dead-zone saturation dead-zone
y y y
L
1
-S S u -a a u q/2 u
-L
hysteresis backlash quantization

SS 2023
Periodic behavior of nonlinear systems (1)
If, for the initial state x0 , there exist solutions of the state equation (1.3) with the property
x (t , x 0 )  x  t  T, x 0  , t  0 (1.6)
then we call them periodic with the period T. In the state-space they represent closed trajectories
and are called periodic orbit or closed orbit. An isolated closed trajectory is called limit oscillation or
limit cycle.
Definition 1.2 A limit oscillation (or limit cycle for second-order systems) is a closed trajectory in the
state-space, such that no other closed trajectory can be found arbitrarily close to it.
The undamped oscillation of a linear system is a periodic orbit, but not a limit oscillation, why?
We note a closed trajectory as {xG} and define as distance of a state x(t) from the trajectory
 ( x (t ),{x G }) : min x(t )  x G (1.7)

xG
Definition 1.3 A closed trajectory {xG} of the state equation (1.3) is said to be stable if there exists
a  > 0,   = () > 0, so that
 ( x 0 ,{x G })   ( )   ( x (t ),{x G })   t  0 (1.8)

SS 2023
Periodic behavior of nonlinear systems (2)
Definition 1.4 A closed trajectory {xG} of the state equation (1.3) is said to be asymptotically stable,
if {xG} is stable and, in addition, for all trajectories starting from a sufficiently small neighborhood of
{xG}
lim  ( x (t ),{x G })  0 (1.9)
t 
It is obviously that every asymptotically stable closed trajectory is also a limit oscillation.
The exact determination of limit oscillations is generally very difficult. However, they appear often in
control loops with a linear controller and linear plant, when sensors or actuators contain static
nonlinearities (e.g. saturation, hysteresis, dead-zone, etc.). In such situations the control system can
easily be formed to the so-called nonlinear standard control
loop (see figure), and the existence and stability of limit
r=0 u y
oscillations can be approximately investigated using method of G(s)
harmonic balance (also known as harmonic linearization or
describing function method).
yr
To simplify the description we derive here the method for SISO ()
systems only. Assuming that the control system can be divided

SS 2023
Van-der-Pol oscillator equation:
y   (1  y 2 ) y  y  0
  0 is a damping parameter ( = 0 : no damping, harmonic oscillation)
State-space model with the state variable vector ( x1  y , x2  x1  y ):
x1  x2
unique equilibrium:  
0
x 2   x1   (1  x12 ) x2  0
Separation into linear and nonlinear components:
Y ( s) 
y   y  y   u   y y  G ( s ) 
2
 2
U ( s) s   s  1
 yr   u   y 2 y   ()  y 2 y  yr
u linear low-pass block


0 G(s) y s2   s  1
yr s
(·)
nonlinear “feedback” block
(·)2
Assumption of an (approximate) permanent harmonic oscillation:

y  A sin(t )  y  A cos(t )
yr  y 2 y  A2 sin 2 (t )  A cos(t )
A3 A3
 1  cos(2t ) cos(t )  cos(t )  cos(3t )
2 4
Due to the low-pass property of G(s):
A3 A2 d A2
yr  cos(t )    A sin(t )  y
4 4 dt 4
Laplace transform (only applicable to a fixed A):
Yr ( s ) A2
 s  N (s)
Y ( s) 4
Describing Function of the nonlinear “feedback” block:

A2
N ( A,  )  (j )
4
Van-der-Pol oscillator as nonlinear standard control loop in the frequency
domain:
G(j)
u 
0 y  Asin(t)
(j ) 2   (j )  1
yr A2
(j )
4
N(A,)
Y (j )  G (j )U (j )  G (j ) N ( A,  )Y (j )

Harmonic balance  1  G (j ) N ( A,  ) Y (j )  0
 1  G (j ) N ( A,  )  0
Solution of the harmonic balance equation:
A2 
1 (j ) 0
4 (j )   (j )  1
2
 A2 2  2 A
2

(1   )  ( )  4 ( )   j (1   ) 4  
2 2 2
     0
(1   2 ) 2  ( ) 2
2  1    1
A2
 (1   )   0 
2
  0 
4
 do not make sense
A  0 
2
A
For  = 1: (1   2 ) 2  ( )2  ( ) 2  0
4
A2
 1 0  A2
4
Thus, the harmonic balance solution is: y  2sin( t ) (independent of  !)
Linear interpretation: for  = 1 and A = 2 the closed-loop becomes

G( s) s 2
  s  1 4
T ( s)    2
1  G( s) N ( s) A 2
s 4 s  ( A2  4) s  4
1
4 s2   s  1
leading to the eigenvalues
 ( s )  4 s 2  ( A2  4) s  4  4( s 2  1)
 1,2  0  j  1
marginally stable harmonic oscillation!
-4
x2
 2
 1
-2
y = 2sin(t)  0
Van der Pol oscillator
for different values
of the parameter  0
x1
-2
-4
-4 -2 0 2 4
Harmonic Balance
into a linear time-invariant part represented by the transfer function G(s) and a nonlinear part (static
nonlinear characteristic) with the input-output relationship
yr (t )   ( y (t )) (1.10)
For the ease of analysis of possible limit oscillations we assume for the reference signal r = 0. There
is no difficulty to apply the method to control loops with for example constant reference signal by
introducing an appropriate structural transformation.
We start our discussion from a nonlinear standard control loop and suppose that a limit oscillation
does exist and can be approximately described by a harmonic oscillation. Thus, the following
structure in the frequency-domain can be given.
y  Asint
G(j)
yr  a1sint + b1cost N(A, )

SS 2023
Describing function
Here, the frequency-domain function N(A, ) is an approximated description of the nonlinearity ()
regarding the fundamental frequency and is therefore termed describing function of (). Contrary to
frequency response of a linear system the describing function depends on the amplitude of the input
sinusoid. It can be shown that for time-invariant static nonlinearity () the function N is even
independent of the frequency !
For deriving the describing function we write the output yr of the nonlinear element () in steady-
state as a Fourier series:
a0  a0 
yr (t )    an cos( nt )  bn sin( nt )    an2  bn2 sin(nt   n ) (1.11)
2 n 1 2 n 1
π π π
1 1 1
a0   yr (t )d(t ), an   yr (t )cos(nt )d(t ), bn   yr (t ) sin( nt )d(t ) (1.12)
π -π π -π π -π
For the output y of the linear subsystem it yields

a0
y (t )  G (0)   G ( jn )  an cos( nt  n )  bn sin(nt  n ), n  G ( jn )
2 n 1
We assume the existence of a sinusoidal limit oscillation which implies

SS 2023
Harmonic analysis (1)
a0G (0)  0, G ( jn ) an  0, G ( jn ) bn  0, (1.13)
To approximately meet the requirement (1.13), the system should satisfy the following conditions:
1) () is an odd function, i.e. point-symmetric, (or G(s) has a differential portion,) and
2) G(j) is a low-pass filter for n with n  2.
The describing function N(A) of the nonlinearity () can be calculated under these assumptions as
follows:
yr (t ) a12  b12 e j(t+1 ) 1 (1.14)

N (A) :   (b1  ja1 ) .
y (t ) Ae jt A
In general the describing function in (1.14) is complex-valued and depends on the input amplitude.
If, however, the nonlinear function () is single-valued (a hysteresis is for instance a multi-valued
function), the term a1 vanishes and the describing function is then real.
A sinusoidal limit cycle may exist if the following condition holds
1  G ( j ) N (A)  0 . (1.15)

SS 2023
Harmonic analysis (2)
Equation (1.15) is known as the harmonic balance equation and can be applied to approximately
predict and determine a limit cycle. The terms in (1.14) are calculated according to Fourier series
π π
1 1
a1   (t ) cos(t )d(t ), b1    (t )sin(t )d(t ) (1.16)
π -π π -π
Once the describing function N(A) and the transfer function G(s) are known, equation (1.15) can be
then evaluated for limit cycle prediction. There may be one, multiple or none solutions of (1.15).
Introducing the inverse describing function
1 (1.17)
N I (A) : 
N (A)
the equation of harmonic balance can be given in the form
G ( j )  N I (A) (1.18)
The evaluation of (1.18) can be done either analytically or, with help of a software tool,
graphically/numerically (today’s standard procedure).

SS 2023
Example 1.5 (1)
Exp. 1.5 Determine the describing function of ().

 k2
Solution: Due to the symmetry it is obviously
π k1
1
a1   (t ) cos(t )d(t )  0
π -π  y
For the b1 we need to distinguish between two cases:
|y|   : Here the function () is linear. Thus, we have
N (A)  k1 .
|y| >  : Here we divide the integration interval into two sections by
defining  = sin-1(/A):
π/2
4
N (A)  
πA 0
 (t )sin(t )d(t )
 π/2
4 4

πA 0
    k  +k (Asin - ) sin  d
2
k1A sin d 1 2
πA

SS 2023
Example 1.5 (2)
2k1  1  4 (k1  k2 )  π  2k2  π 1 1 

    sin 2     cos   cos     sin π- + sin 2  
π  2  πA  2 π 2 2 2 
2(k1  k2 )   
    cos    k2 .
π  A 
Finally the describing function can be given as
k1 A
N (A) 
2(k1  k2 )  1    
2 
 sin      
1      k2 A
π  A A  A  
 

SS 2023
Example 1.6 (1)
Exp. 1.6 Determine the describing function of () (relay with dead-zone and hysteresis).
Solution: It is reasonable to assume A  . 

k
We consider first the real part
S
2 - -q
2 2k
b1   k sin  d  cos 1  cos 2  q  y
π 1 π
where for the angles 1 and 2 -k
2
  
1
sin 1   cos1  1     A2   2
A A A
2
q  q  1
sin  2   cos 2   1      A 2  (q ) 2
A  A  A
2k  2
b1  A   2  A 2  (q ) 2 
πA  
SS 2023
Example 1.6 (2)
The imaginary part a1 can be obtained by evaluating the area S enclosed by the nonliearity (why?).
We have
S 2k (1  q )
a1   
πA πA
Therefore, the complete describing function is
2k  2
N (A)  2 
A   2
 A 2
 ( q ) 2
 j (1  q )  .
πA 
Note that A   and   1 must hold. 
As special cases of the describing function for a relay with dead-zone and hysteresis the describing
functions for the following nonlinearities can be easily derived by appropriately choosing the
parameters q and 
• a relay with dead-zone
• a relay
• a relay with hysteresis.

SS 2023
Limit oscillation and stability (1)
In order to graphically evaluate equation (1.18) we can plot both the frequency response function
G(j) (varying  > 0) and the negative inverse describing function NI (varying A  0) in the complex
plane, and determine the parameters of a possible limit oscillation (AG, G) as intersection of the two
curves. If such an intersection can be found, then a limit oscillation may exist with amplitude AG and
frequency G. However, if there is no solution of (1.18), limit oscillations may still exist. They are
then not harmonic and cannot be dealt using the method of harmonic balance. Fortunately, this
situation does not occur very often in practice.
Is the limit oscillation found this way also stable?
Theorem 1.1 Stability of limit oscillations (cycles) (AG2,G2)

The limit oscillation (AG, G) found by the method of
harmonic linearization is asymptotically stable if and

only if
G(j)
  (na  2nr )π A  AG A
 Soll  G  N I  . (1.19) NI(A)
 
 (na  2nr )π A  AG (AG1,G1) 
SS 2023
Limit oscillation and stability (2)
In (1.19) na is the number of poles of G(s) on the imaginary axis and nr the number of
poles of G(s) in the right half s-plane. For the correct application of this theorem it is
required that G(s) is strictly proper (numerator degree lower than denominator
degree). This condition, however, is usually satisfied in most practical systems with
low-pass characteristics.
If both conditions in (1.19) are satisfied interchanged, the limit oscillation is then unstable. Does the
equality or inequality sign in (1.19) apply both for A > AG as well as for A < AG, then the limit
oscillation is asymptotically semi-stable (i.e., one-sided stable).
Notes:
The application of method of harmonic balance can also be extended to MIMO systems and, under
some further conditions, to dynamic nonlinearities (nonlinearities with “memory”), especially in the
field of oscillation analysis.

SS 2023
Example 1.7 (1)
Exp. 1.7 Investigate the existence and stability of possible limit cycles
u 0.5 y
Solution: The frequency response of the linear part is
s(1  s)
0.5
G ( j ) 
j (1  j )
and the inverse describing function
1
1
N I (A)  
π
4
 
A2  1  j , A  1
of the nonlinear part are plotted in the complex plane.
Since both curves have an intersection point, a limit cycle
(AG,G) may exist. From the equation of harmonic balance
=
we have A=1
NI(A) π

0.5
j (1  j )

0.5  j0.5
 (1   2 )
 
π
4
 A2 1  j  A

4
G(j)

SS 2023
Example 1.7 (2)
Splitting the real and imaginary part we obtain

0.5 π 0.5 π
 A 2
1 , 
1  G 4 G (1  G2 ) 4
2 G
leading to the numerical solution (AG,G) = (1.13, 0.5). Thus, we predict that the control loop may
approximately result in a limit cycle
y (t )  1.13sin(0.5t )
This limit oscillation is asymptotically stable due to Theorem 1.1 as the continuous argument change
considered in this case is  and thus the Nyquist condition (1.19) is satisfied.

2
1
0
-1
-2 0 5 10 15 20 25 30 35 40 45 50 [s]

SS 2023
Summary: harmonic balance
• Useful for approximative investigations of limit oscillations, based on describing function and
harmonic analysis
• Can especially used for SISO systems with static nonlinearities, extendable to MIMO systems
and certain nonlinear dynamics
• Provides quantitative predictions with respect to amplitude, frequency, and stability of limit
oscillations
However:
• Quantitative prediction may be inaccurate (often)
• Qualitative prediction about existence or non-existence of limit oscillations may be incorrect
(relatively seldom)
• Suitable to only a limited extent for controller design (e.g. for suppression of limit oscillations)

SS 2023
Equilibrium stability in the sense of Lyapunov (1)
Definition 1.5 Stability of an equilibrium x2


The equilibrium point xR of the dynamic system
(1.1) is said to be stable (in the sense of Lyapunov) xR 
if, for each initial time t0   > 0   = (, t0) > 0 x0 x1

x(t)
such that
x (t0 )  x R   ( , t0 )  x(t )  x R    t  t0 (1.20)
Otherwise, the equilibrium is unstable.
Definition 1.6 Convergence of an equilibrium
The equilibrium point xR of the system (1.1) is said to be convergent if, for each initial
time t0  1(t0) > 0 such that
x (t0 )  x R  1 (t0 )  lim x (t )  x R (1.21)
t 
Definition 1.7 Asymptotical stability of an equilibrium

The equilibrium point xR of the dynamic system (1.1) is said to be asymptotically
stable (in the sense of Lyapunov) if it is both stable and convergent.

SS 2023
Equilibrium stability in the sense of Lyapunov (2)
Definition 1.8 Exponential stability of an equilibrium

The equilibrium point xR of the dynamic system (1.1) is said to be exponentially stable if
for each initial time t0 there exist two real positive constants (t0), (t0) > 0 such that
x (t0 )  x R   (t0 )  x (t )  x R   x(t0 )  x R e t  t  t0 (1.22)
Notes:
For autonomous systems the conditions (1.20) through (1.22) need to be checked for a special t0, for
example t0 = 0, only.
Every arbitrary equilibrium xR of a dynamic system can be transformed into the origin x = 0 by applying
an appropriate coordinate transformation. Thus, there is no loss of generality in assuming in the sequel
that the equilibrium point under study is the origin xR = 0, unless especially declared otherwise.
The stability properties of an equilibrium can hold in some limited neighborhood of the equilibrium
point (local stability) or for the whole definition set of the system equation (global stability). In case of
an asymptotically locally stable equilibrium point the set of all state-space points, from which the state
trajectory converges into the equilibrium point, is called region of attraction.

SS 2023
Standard control loop with static nonlinearity
We return to the control structure that we already have considered once before:
r=0 u y
G(s)
yr
()
The nonlinear function () will be assumed again to be static (i.e. without “memory”), although
generally possibly time-varying. Methods that are going to be considered in the following are also
applicable to MIMO systems. We describe, however, only the SISO version for simplicity.
We consider the case where the reference r = 0. It can be shown that the results are extendable to
control loops with a constant reference signal (fixed-value control).
A special role play the sector functions.

SS 2023
Sector function
Definition 1.9 Sector condition

A piecewise continuous function () is said to be a sector function (belongs to the
sector [k1, k2]) if there exist two real numbers k1 and k2 with k1 < k2 so that
 (t , y )
y  0  k1   k2 . (1.23)
y
k2y
(t, y)
k 1y
0 y

SS 2023
Absolute stability
Definition 1.10 Absolute stability (SISO case)

A nonlinear SISO system is given by
x  A x  bu (1.24)
y  c x  du (1.25)
u   (t , y ) (1.26)
where (A, b, c, d) is a controllable and observable LTI system and () is a sector
function in [k1, k2]. The system is absolutely stable if the origin x = 0 is globally
asymptotically stable for any nonlinearity  in the given sector.
The fact that the “dynamic part” of the system is linear and the r=0 u y
nonlinearity is “only” static (in “worst-case” time-varying), led G(s)
to intuitive conjectures that the absolute stability could be
analyzed based on a “linear equivalent structure”, shown in the
figure. k

SS 2023
The conjectures of Aiserman and Kalman
Aiserman's conjecture:
If the matrix [A  bck ] is stable for all fixed values of k  [k1 , k 2 ] (so-called Hurwitz sector), then the
nonlinear system (1.24) through (1.26) is globally asymptotically stable.
k2y
Kalman's conjecture: ’min

’max (t, y)
If the matrix [A  bck ] is stable for all values of k  [k1 , k 2 ] , where k 1y
0 y
d ( t , y )
k1   k 2 and
dy
then the nonlinear system (1.24) through (1.26) is globally asymptotically stable.
Unfortunately, both conjectures have been disproved to be false by several counter-examples.

However, the investigations on this conjecture led to very interesting results (Popov criterion, circle
criterion) that guarantee the absolute stability of nonlinear systems.
The failure of these two conjectures, proposed by two of the best minds of control theory, shows that
intuitive reasoning cannot be really relied on in nonlinear systems. An exact, analytical investigation is
indispensable.

SS 2023
Popov criterion
Theorem 1.2 Popov’s criterion

The system described by (1.24) through (1.26) is absolutely stable in the sector [0, k]
with k  , if the following conditions are satisfied:
a) The transfer function G(s) of the linear forward path has only left half-plane poles
except one possible pole in the origin,
b) The nonlinearity () is time-invariant and belongs to the sector [0, k],
c) There exists a real number , such that for an arbitrary small  > 0
1
Re  (1  j )G ( j )    0   0 (1.27)
k
Inequality (1.27) is known as Popov’s inequality.
Notes: 1) Popov’s criterion can be proven by constructing an appropriate Lyapunov function
and only applies to autonomous systems.
2) The application is restricted to a single memoryless nonlinearity in [0, k].
3) The stability of the nonlinear systems may be determined by examining alone the
frequency response function of a linear subsystem
4) It is “only" a sufficient condition

SS 2023
1
Popov inequality: Re (1  j )G (j )  0
k
for some real constant 
1
 Re G (j )   Im G (j )   0
k
1
 Re G (j )   Im G (j ) 
k
By defining the Popov locus GP (j )  Re G (j )  j Im G (j ):
1
 Re GP (j )   Im GP (j ) 
k
1
Popov line: Re GP    Im GP  
k
1
 x y 
k
Popov inequality in the complex plane:
1 1
  0  Re GP     x
k k
1 1 1 1
  0  Im GP    Re GP     y  x 
 k  k
Geometrical interpretation of the Popov inequality:
1 Popov line for  = 0

x y
k Im
1
 0  x
k
 GP(j) completely on the right -1/k 0 Re
side of the vertical line x = -1/k
GP (j )
Popov line for  > 0
Im
 1 1
  0  y    x  k 
  

  0  y  1  x  1  -1/k 0 Re
   k

GP (j )
 GP(j) always completely on the Popov line for  < 0

Im
right side of the Popov line !
-1/k 0 Re
GP (j )
Geometric interpretation of the Popov’s criterion
The Popov’s inequality can also be expressed as:

1
Re G ( j )     Im G ( j ) (1.28)
k
Let us now construct the so-called Popov locus GP(j) = Re{G(j)} + jIm{G(j)} together with the
Popov line
1 1
Im GP ( j )  Re G ( j  )  (1.29)
  P
k  Im
1
in the complex plane, then the following geometric tan  
interpretation of the Popov’s criterion can be made:


The system in the sector [0, k] is absolutely stable if 1
the complete Popov locus is below (or on the right  0 Re
k
side of) the Popov line. GP(j)

SS 2023
Example 1.8
Exp. 1.8 Determine the sector condition for the system described by (1.24) – (1.26) if the transfer
function of the linear subsystem is
1 Popov plot
G(s) 
( s  1)( s  2)(s  3) 0
Solution: The frequency response of the linear subsystem: -0.01
6(1   2 )  j(11   2 )
-0.02
G ( j )  -0.03
36(1   2 ) 2  (11   2 ) 2  2 -0.04
Thus the Popov plot is -0.05
6(1   2 )  j( 2  11) 2 -0.06
GP ( j )  -0.07
36(1   2 ) 2  (11   2 ) 2  2 -0.08
The largest sector for absolute stability is given -0.09
by the tangent, which let the complete frequency
-0.04 -0.02 0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16
response lie to the right and whose intersection
with the real axis has the minimum distance to
the origin. kmax  60,   0.55


SS 2023
Sector transformation
If the nonlinearity does not belong to the sector [0, k], but to [k1, k2] with k1 < k2, then it can be
transformed to the sector [0, k] by a sector transformation (also called loop transformation):
G0(s)
G(s) k1
*() G(s)
*()  [k1, k2] *()

k2y
*()
k1
0 y ()
k1y ()  [0,]
SS 2023
Example 1.9 [2]
Exp. 1.9 Determine the sector condition for the system described by (1.24) – (1.26) if the transfer
function of the linear subsystem is
1 Popov criterion
G(s) 
( s  1)( s  3) 2 0
Solution: The linear subsystem is obviously unstable, but kmax  14
-0.05
stabilizable using a P-type controller:
  5/3
9  k1  24 -0.1
We choose for example k1 = 10. For the -0.15

transformed subsystem it yields
-0.2
1
G0 ( s )  -0.25
s 3  5s 2  3s  1
from which the upper bound for the sector leading -0.3
to absolute stability can be determined: -0.35
k2  k1  14 -0.2 0 0.2 0.4 0.6 0.8 1
After an inverse transformation we obtain the sector condition: [10, 24]. 

SS 2023
Circle criterion – “forbidden areas"
The nonlinear system is given by (1.24) – (1.26) and also a time-varying sector function ()  [k1, k2]
is now allowed. Depending on the sign of the sector boundaries the following “forbidden areas”
V(k1,k2) in the complex plane are defined for the frequency response G(j) of the linear subsystem:
V(k1,k2)
G(j) Im Im
V(k1,k2)
1 1 Re b) k1 = 0 1 1 Re d) k2 = 0
   
k1 k2 k2 G(j) k1 Im
Im
a) k1k2 > 0 c) k 1 < 0 < k2

1 1 Re
 Re G(j) 
V(k1,k2) k2 k1
G(j)
V(k1,k2)

SS 2023
The circle criterion
Theorem 1.3 circle criterion

If the system (1.24) – (1.26) satisfies the conditions
a) The linear subsystem G(s) has nr poles in the right half-plane and ni poles on the
imaginary axis,
b) The Nyquist plot G(j) does not enter the “forbidden area” V(k1,k2) and satisfies for
each arbitrary point from V(k1,k2) the Nyquist condition
Soll = (2nr+ni) (1.30)
Then the control loop is globally absolutely stable.
Notes 1) Obviously, the circle criterion is an generalization of the Nyquist criterion developed
for linear systems.
2) In the version formulated above the circle criterion is, in contrast to the Nyquist
criterion, a sufficient condition only!
3) In the case k1  0  k2 the stability of the open-loop linear subsystem must be
given. For k1k2 > 0 the stability of the closed-loop control system can also be
achieved by a nonlinear feedback.

SS 2023
Example 1.10 (1)
Exp. 1.10 We analyze now the stability of the nonlinear standard control structure with
4
G( s) 
( s  1)( 12 s  1)( 13 s  1)
Resolution: The linear subsystem is obviously stable. Nyquist Diagram
5
In order to examine a largest possible, absolutely
stable sector, we choose, in view of the Nyquist Im
plot G(j), the sector borders with different signs
(k1 < 0 < k2), and the right border of the “forbidden
area” coincides with the intersection point of the
0
Nyquist plot with the real axis (4,0). Thus we obtain -1.4 4
k2 = 1/1.4 and k1 = -1/4, so that the closed-loop G(j)

control is absolutely stable for every arbitrary
(and possibly time-varying) nonlinearity in the
sector [-0.25, 0.71].
-5
-5 0 Re 5

SS 2023
Example 1.10 (2)
Alternatively, we can also choose the sector boundaries to be k1 = 0, k2 > 0, which means that we only
allow nonlinearities in the first and third quadrants. The “forbidden area” will be left of -1/k2. By more
precisely inspecting the Nyquist plot we notice that the complete Nyquist plot stays to the right of a
vertical line Re{s}  -0.86. Thus we can conclude that the closed control loop is globally absolutely
stable for all nonlinearities within the sector [0, 1.163]. A practical application of this result could be
for example based on the following saturation characteristic:

y, y 1 1
 ( y )  sat( y ) 
sgn( y ), y 1 1 y
Together with the results developed for k1 < 0 < k2 we may expect that the absolute stability was
guaranteed within the whole sector [-0.25, 1.163]. This conclusion is, however, not correct. For a
nonlinearity, which is contained in the whole sector, but not completely in one of the both subsectors,
the closed-loop system is not necessarily stable!


SS 2023
Example 1.11
Exp. 1.11 We analyze now the stability of the nonlinear standard control structure with
4
G( s)  1
( s  1)( 12 s  1)( 13 s  1)
Solution: The linear subsystem is unstable and we 0.5 G(j)
must therefore choose k2 > k1 > 0. Applying
0
the circle criterion shows then that the
forbidden area must be completely enclosed -0.5
by the left loop of the Nyquist plot G(j).
We place now the center of the “forbidden -1-4 -3.5 -3 -2.5 -2 -1.5 -1 -0.5 0
disk” at (-3.5, 0). The maximum possible
radius is then ca. 0.18, such that the point
of intersection of the disk with the real axis is (-3.68, 0) and respectively (-3.32, 0). We
can thus conclude that the corresponding nonlinear standard control is absolutely
stable for all nonlinearities within the sector [0.272, 0.301]. 

SS 2023
Exercises (1)
1. For the following odd nonlinearity, verify the given expression of the describing function:
a)  ( y )  y 5 , N (A)  5A 4 / 8
b)  ( y )  y 3 y , N (A)  32A3 /15π
2. The figure shows a feedback connection of a linear time-invariant system represented by the
transfer function matrix G(s) and a nonlinear time-varying element defined by z = (t,y). The
variables r, u, y, and z are vectors of the same dimension, and (t,y) is a vector-valued function.
With r as input and y as output, find a state model.
u
r G ( s )  C( s I  A) 1 B y
z
 (t , y )

SS 2023
Exercises (2)
3. Using the describing function method, investigate the existence of periodic solutions and the
possible frequency and amplitude of oscillation in the “nonlinear standard control loop" for each of
the following cases. If there are limit oscillations, determine their stability behavior.
a) G ( s)  (1  s) / s( s  1),  ( y )  y 5  k=2
c)
b) G ( s )  1/( s  1)6 ,  ( y )  sgn( y ) 1
c) G ( s )  5( s  0.25) / s 2 ( s  2)2 and the nonlinearity in figure c) y
d) G ( s)  5( s  0.25) / s 2 ( s  2)2 and the nonlinearity in figure d)
e) G ( s)  5( s  0.25) / s 2 ( s  2)2 and the nonlinearity in figure e)
 
e)
d) 1 2
1 y 1 1.5 y

SS 2023
Exercises (3)
4. Using the Popov’s criterion, study absolute stability for each of the transfer functions given next. In
each case, find a sector [, ] for which the system is absolutely stable.
s 1
a) G(s)  b) G(s) 
s2  s  1 ( s  2)( s  3)
1 s2 1
c) G(s)  2 d) G ( s) 
s  s 1 ( s  1)( s 2  1)
1 s s 1
e) G(s)  f) G(s) 
( s  1) 2 ( s  1)( s  2) 2
5. Repeat Exercise 4 using the circle criterion.
6. Consider the system x1   x1  h( x1  x2 ), x 2  x1  x2  2h( x1  x2 )
where h is a smooth function satisfying

SS 2023
Exercises (4)
c, y  a2
yh( y )  0, y  , h( y )  0, y  a1
c, y   a2
h ( y )  c, for a1  y  a2 and  a2  y  a1
a) Show that the origin is the unique equilibrium point.

b) Show, using the circle criterion, that the origin is globally asymptotically stable.
7. The feedback system shown in the figure represents a control system where H(s) is the transfer
function of the plant and the inner loop models the actuator. Let
s6
H (s) 
( s  2)( s  3)
and suppose k > 0 and () belongs to a sector (0, ], where  could be arbitrarily large, but finite.

SS 2023
Exercises (5)
a) Show that the system can be represented as the feedback connection defined as the "nonlinear
standard control loop" with G(s) = [H(s)+k]/s.
b) Using the Popov criterion, find a lower bound kc such that the system is absolutely stable for all
k > kc .
r=0 e 1 u y
() s H(s)

SS 2023

Chapter 1

Uploaded by

Copyright:

Available Formats

Chapter 1

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Chapter 1

Uploaded by

Copyright:

Available Formats

Chapter 1

Einführung, Systeme mit nichtlinearen Kennlinien,

Prof. Dr.-Ing. Steven Liu Nonlinear and Adaptive Control

Prof. Dr.-Ing. Steven Liu Nonlinear and Adaptive Control

Prof. Dr.-Ing. Steven Liu Nonlinear and Adaptive Control

Prof. Dr.-Ing. Steven Liu Nonlinear and Adaptive Control

u = const. > 0 u = const. < 0

Linear system: Nonlinear system:

A sin( t) A G (j ) sin  t   (j ) 

Special solution for u  A sin( t) :

The solution for t   (stationary response) can be easily obtained:

  2 (2 2  5) cos 2 (t ) 

i.e., for an excitation with frequency , the output will contain

Prof. Dr.-Ing. Steven Liu Nonlinear and Adaptive Control

• The principle of superposition does not hold any more

Completely new methods are necessary!

Prof. Dr.-Ing. Steven Liu Nonlinear and Adaptive Control

Methods need to be appropriately chosen depending on the required application;

Prof. Dr.-Ing. Steven Liu Nonlinear and Adaptive Control

We start our discussions with consideration of static nonlinearities.

Prof. Dr.-Ing. Steven Liu Nonlinear and Adaptive Control

Exp. 1.1 Oscillator with a tunnel-diode iD

Prof. Dr.-Ing. Steven Liu Nonlinear and Adaptive Control

Exp. 1.2 Negative-resistance oscillator iD

Prof. Dr.-Ing. Steven Liu Nonlinear and Adaptive Control

Prof. Dr.-Ing. Steven Liu Nonlinear and Adaptive Control

Prof. Dr.-Ing. Steven Liu Nonlinear and Adaptive Control

A model description as autonomous system without Stribeck effect is:

Which are equilibria of the system?

Prof. Dr.-Ing. Steven Liu Nonlinear and Adaptive Control

Exp. 1.4 Transmission with gap (backlash)

Prof. Dr.-Ing. Steven Liu Nonlinear and Adaptive Control

relay relay with dead-zone saturation dead-zone

Prof. Dr.-Ing. Steven Liu Nonlinear and Adaptive Control

 ( x (t ),{x G }) : min x(t )  x G (1.7)

 ( x 0 ,{x G })   ( )   ( x (t ),{x G })   t  0 (1.8)

Prof. Dr.-Ing. Steven Liu Nonlinear and Adaptive Control

Prof. Dr.-Ing. Steven Liu Nonlinear and Adaptive Control

  0 is a damping parameter ( = 0 : no damping, harmonic oscillation)

State-space model with the state variable vector ( x1  y , x2  x1  y ):

u linear low-pass block

Assumption of an (approximate) permanent harmonic oscillation:

Laplace transform (only applicable to a fixed A):

Describing Function of the nonlinear “feedback” block:

Y (j )  G (j )U (j )  G (j ) N ( A,  )Y (j )

yr  a1sint + b1cost N(A, )

Prof. Dr.-Ing. Steven Liu Nonlinear and Adaptive Control

Prof. Dr.-Ing. Steven Liu Nonlinear and Adaptive Control

a0G (0)  0, G ( jn ) an  0, G ( jn ) bn  0, (1.13)

yr (t ) a12  b12 e j(t+1 ) 1 (1.14)

A sinusoidal limit cycle may exist if the following condition holds

Prof. Dr.-Ing. Steven Liu Nonlinear and Adaptive Control

Prof. Dr.-Ing. Steven Liu Nonlinear and Adaptive Control

Exp. 1.5 Determine the describing function of ().

Prof. Dr.-Ing. Steven Liu Nonlinear and Adaptive Control

2k1  1  4 (k1  k2 )  π  2k2  π 1 1 

Finally the describing function can be given as

Prof. Dr.-Ing. Steven Liu Nonlinear and Adaptive Control

()  [k1, k2] ()