Stability of Nonlinear Systems

By
Lyapunov Stability

1
 Introduction to Stability

1. The concept of stability

2. Critical points
3. Linear stability analysis

2
 The Concept of Stability
• Imprecise definition
– Consider a nonlinear system with the origin as a steady-state point:
dy
 f (y )  f ( y )  0
dt

– Does the system return to the origin if perturbed away from the
dy If so, the system is stable. Otherwise, the system is unstable.
origin? f ( y ) y (0)  ε  lim y (t )  0
dt t 
y2

Precise definition
» Stability: produce a bound e on y(0) d
such that y(t) remains within a given
bound d e y1
y(0)
» Asymptotic stability: stable & y(t) y(t)  0
converges to the origin
» Commonly known as Lyapunov
stability
3
 Critical Points of a Linear System
• Two-dimensional system
dy1
dy  a11 y1  a12 y 2
 Ax  dt
dt dy 2
 a21 y1  a22 y 2
dt
• Divide equations
dy 2 dy 2 dt a21 y1  a22 y 2
 
dy1 dy1 dt a11 y1  a12 y 2

• Critical point
– Point where dy2/dy1 becomes undetermined
– Only the origin for a homogeneous linear system
– Five types of critical points depending on the geometric shape of
trajectories near the origin and eigenvalues of A matrix

4
 Types of Critical Points
• Proper node • Center
– Two identical real – Two imaginary
eigenvalues eigenvalues

• Improper node • Spiral point

– Two different real – Two complex
eigenvalues eigenvalues

• Saddle point • Degenerate

– Two real eigenvalues node
with different signs – No eigenvector
basis exists

5
 Linear Stability Analysis
• General solution form
for distinct eigenvalues
y (t )  c1x (1) e 1t  c2 x ( 2 ) e 2t    cn x ( n ) e nt
Imaginary
• Procedure
– Compute the eigenvalues Left-Half Plane Right-Half Plane

of A
– The system is
asymptotically stable if
and only if Re(l i) < 0 for i Stable Unstable
Real
= 1, 2, …, n Region 0 Region
– The origin is unstable if
Re(li) > 0 for any i
– Stability allows zero
eigenvalues

6
 Nonlinear Systems
• Steady-state points
– Nonlinear models can have multiple steady states
– Stability must be determined for each steady state
• Consider origin as a generic steady-state point
dy
 f (y )  f ( y )  0
dt
dy  dy
y'  y  y    f ( y   y )  g ( y )
dt dt
dy 
 g ( y )  g (0)  0
dt
• Nonlinear model linearization about origin
dy dy 
 f (y )   Ay 
dt dt

7
Lyapunov Stability

8
 Definition of Lyapunov Exponents
• Given a continuous dynamical system in an n-dimensional
phase space, we monitor the long-term evolution of an
infinitesimal n-sphere of initial conditions.
• The sphere will become an n-ellipsoid due to the locally
deforming nature of the flow.
• The i-th one-dimensional Lyapunov exponent is then defined
as following:

p1(t)
p1(0)
x(t)
x0 p (0)
2
t - time flow
p2(t)
9
 On more formal level

 The Multiplicative Ergodic Theorem of Oseledec states that

this limit exists for almost all points x0 and almost all
directions of infinitesimal displacement in the same basin of
attraction. 10
• Order: λ1> λ2 >…> λn
• The linear extent of the ellipsoid grows as 2λ1t
• The area defined by the first 2 principle axes grows as
2(λ1+λ2)t
• The volume defined by the first 3 principle axes grows
as 2(λ1+λ2+λ3)t and so on…
• The sum of the first j exponents is defined by the long-
term exponential growth rate of a j-volume element.

11
 Signs of the Lyapunov exponents
• Any continuous time-dependent DS without a
fixed point will have 1 zero exponents.
• The sum of the Lyapunov exponents must be
negative in dissipative DS   at least one
negative Lyapunov exponent.
• A positive Lyapunov exponent reflects a
“direction” of stretching and folding and
therefore determines chaos in the system.

12
The signs of the Lyapunov exponents provide a
qualitative picture of a system’s dynamics
• 1D maps: ! λ1=λ:
– λ=0 – a marginally stable orbit;
– λ<0 – a periodic orbit or a fixed point;
– λ>0 – chaos.
• 3D continuous dissipative DS: (λ1,λ2,λ3)
– (+,0,-) – a strange attractor;
– (0,0,-) – a two-torus;
– (0,-,-) – a limit cycle;
– (-,-,-) – a fixed point.

13
 The sign of the Lyapunov Exponent
• <0 - the system attracts to a
fixed point or stable periodic
orbit. These systems are non
conservative (dissipative) and
exhibit asymptotic stability.
• =0 - the system is neutrally
stable. Such systems are
conservative and in a steady state
mode. They exhibit Lyapunov
stability.
• <0 - the system is chaotic and
unstable. Nearby points will
diverge irrespective of how close
they are.
14
 Computation of Lyapunov Exponents

• Obtaining the Lyapunov exponents from a system

with known differential equations is no real problem
and was dealt with by Wolf.
• In most real world situations we do not know the
differential equations and so we must calculate the
exponents from a time series of experimental data.
Extracting exponents from a time series is a complex
problem and requires care in its application and the
interpretation of its results.

15
 Calculation of Lyapunov spectra from
ODE (Wolf et al.)
• A “fiducial” trajectory (the center of the sphere) is
defined by the action of nonlinear equations of
motions on some initial condition.
• Trajectories of points on the surface of the sphere
are defined by the action of linearized equations on
points infinitesimally separated from the fiducial
trajectory.
• Thus the principle axis are defined by the evolution
via linearized equations of an initially orthonormal
vector frame {e1,e2,…,en} attached to the fiducial
trajectory
16
 Problems in implementing:
• Principal axis diverge in magnitude.
• In a chaotic system each vector tends to fall
along the local direction of most rapid growth.
(Due to the finite precision of computer calculations, the
collapse toward a common direction causes the tangent space
orientation of all axis vectors to become indistinguishable.)

Solution:
Gram-Schmidt reorthonormalization (GSR)
procedure!
17
• GSR never affects the direction of the first vector, so
v1 tends to seek out the direction in tangent space
which is most rapidly growing, |v1|~ 2λ1t;
• v2 has its component along v1 removed and then is
normalized, so v2 is not free to seek for direction,
however…
• {v’1,v’2} span the same 2D subspace as {v1,v2}, thus
this space continually seeks out the 2D subspace that
is most rapidly growing |S(v1,v2)|~2(λ1+λ2)t … |
S(v1,v2…,vk)|~2(λ1+λ2+…+λk)t k-volume
• So monitoring k-volume growth we can find first k
Lyapunov exponents.
18
 Lyapunov spectrum for experimental data
(Wolf et al.)

• Experimental data usually consist of discrete

measurements of a single observable.
• Need to reconstruct phase space with delay
coordinates and to obtain from such a time series an
attractor whose Lyapunov spectrum is identical to
that of the original one.

19
Procedure for 1

20
Procedure for 1+2

21
Lyapunov Simple Example
• A 2D map f: R2  R2.
f1 ( x)  x2
– (from Mathworld)
f 2 ( x)   x1  2 x2

V ( x)  ( x12  x22 ) / 2 • Define a Lyapunov function.

V * ( x )  V ( x )  f ( x )
• The derivative is negative so
V ( x)  x1 x2  x2 ( x1  2 x2 )
*
the origin is stable.
V * ( x)  2 x22

22
 Matlab Tools for Stability Analysis

• Matlab provides several functions for

linear stability analysis of nonlinear
systems
– fsolve – finds steady-state point for
nonlinear ODE system
– linmod – linearizes nonlinear ODE system
about given steady state to generate linear
ODE system
– Eigenvalue – computes eigenvalues of linear
ODE system
23
• Controllability
co  ctrb( A, B )
% uncontrolled states
unco  length( A)  rank (co)
• Observability
ob  obsv( A, C )
% unobserved states
unob  length( A)  rank (ob)

24
 Example
 x 1   7.1847  50.0415  x1  1
 x    50.0415      u
Model equations  2  0   x2  0
 x1 
y1  1.9558  0.04761  
 x2 
A  [7.1847  50.0415; 50.0415 0];
B  [ 1; 0];
Controllability
1.0 - 7.1847
co  ctrb( A, B)    ; rank (co)  2
 0 50.0415
A  [7.1847  50.0415; 50.0415 0];
Observability C  [1.9558 - 0.04761];
 1.9558 - 0.0476 
ob  obsv( A, C )   ; rank(ob)  2
-16.4343 - 97.8712

25
 Example
 x 1   7.1847  50.0415  x1  1
 x    50.0415      u
Model equations  2  0   x2  0 
 x1 
y1  1.9558  0.04761  
 x2 

A  [ 7.1847  50.0415; 50.0415 0];

Analysis
eig ( A)
- 3.5924  49.9124 i
- 3.5924  49.9124 i
asymptotically stable because all the eigenvalues of
the state matrix A have negative real parts

26
 Example
Continuous Biochemical Reactor
Exit Gas Flow

Fresh Media Feed

(substrates)

Agitator

Exit Liquid Flow

(cells & products)

27
 Cell Growth Modeling
• Specific growth rate
1 dX
 X  biomass concentration (g/L)
X dt
• Yield coefficients
– Biomass/substrate: YX/S = -DX/DS
– Product/substrate: YP/S = -DP/DS
– Product/biomass: YP/X = DP/DX
– Assumed to be constant
• Substrate limited growth

m S
 (S ) 
KS  S
– S = concentration of rate limiting substrate
– Ks = saturation constant
– mm = maximum specific growth rate (achieved when S >> Ks)
28
 Continuous Bioreactor Model
Assumptions
• Sterile feed
• Constant volume
• Perfect mixing
• Constant temperature &
pH
• Single rate limiting nutrient
• Constant yields
• Negligible cell death


Product formation rates
» Empirically related to specific growth rate
» Growth associated products: q = YP/Xm
» Nongrowth associated products: q = b
» Mixed growth associated products: q = YP/Xm+b
29
 Mass Balance Equations
• Cell mass
dX dX
VR   FX  VR X    DX  X
dt dt

– VR = reactor volume
– F = volumetric flow rate
– D = F/VR = dilution rate
• ProductdP dP
VR   FP  V R qX    DP  qX
dt dt

• Substrate
dS 1 dS 1
VR  FS 0  FS  V R X   D(S0  S )  X
dt YX / S dt YX / S

– S0 = feed concentration of rate limiting substrate 30

 Steady-State Solutions
• Simplified model equations
dX mS
  DX   ( S ) X  f1 ( X , S )  (S ) 
dt KS  S
dS 1
 D(S0  S )   (S ) X  f 2 ( X , S )
dt YX / S
• Steady-state equations
m S
 DX   ( S ) X  0  (S ) 
KS  S
1
D(S0  S )   (S ) X  0
YX / S
• Two steady-state points

KS D
Non - Trivial :  ( S )  D  S  X  YX / S ( S 0  S )
m  D
Washout : S  S0 X 0
31
 Model Linearization
• Biomass concentration equation
dX   f   f 
 f1 ( X , S )   1   X  X   1 S  S 
dt       X  X ,S   S  X ,S
zero
 m X m XS 
 
  S   D X     2 
S
K
 S  S  K S  S  
• Substrate concentration equation
dS   f   f 
 f2 ( X , S )   2   X  X    2  S  S 
dt      X  X , S  S  X , S
zero
1 mS  1   X  m X S  
 X   m
  DS
 Y X / S  K S  S  K S  S  
 2 
YX / S K S  S 
• Linear model structure:

dX 
 a11 X   a12 S 
dt
dS 
 a 21 X   a22 S 
dt
32
 Non-Trivial Steady State
• Parameter values
– KS = 1.2 g/L, mm = 0.48 h-1, YX/S = 0.4 g/g
– D = 0.15 h-1, S0 = 20 g/L
• Steady-state concentrations
KS D
S   0.545 g/L X  YX / S ( S 0  S )  7.78 g/L
m  D
• Linear model coefficients (units h-1)
m X  m XS
a11  0 a12    1.472
KS  S KS  S  2

1 m S 1   m X  m XS 
a21    0.375 a22     D  3.529
YX / S K S  S YX / S  K S  S  K S  S  
 2 

33
 Stability Analysis
• Matrix representation
 X  dx  0 1.472 
x     x  Ax
 S  dt   0.375  3.529 

• Eigenvalues (units h-1)

 1.472
A  I   1  0.164 1  3.365
 0.375  3.529  
• Conclusion
– The system is asymptotically stable

34