Module 4F2: Nonlinear

Systems and Control

Lectures 3 – 4: Invariant sets and Stability
Jan Maciejowski
[email protected]

Department of Engineering
University of Cambridge

Lent 2012 4F2: Nonlinear systems and Control, Lectures 3–4 – p. 1/32
Invariant sets
n
ẋ = f (x), x(0) = x0 , x ∈ R
Assume: f is Lipschitz continuous.
Unique trajectory: x(·).

Definition 1 (Invariant set) A set of states S ⊆ Rn of

ẋ = f (x) is an invariant set of ẋ = f (x) if ∀x0 ∈ S
and ∀t ≥ 0: x(t) ∈ S.

Equilibria: important special class of invariant sets:

Definition 2 (Equilibrium) A state x̂ is an
equilibrium of ẋ = f (x) if f (x̂) = 0.

Lent 2012 4F2: Nonlinear systems and Control, Lectures 3–4 – p. 2/32
Shift equilibrium to 0
w = x − x̂
In the w coordinates the dynamics of the system are
ẇ = ẋ = f (x) = f (w + x̂) = fˆ(w)

where fˆ(w) = f (w + x̂).

Since
fˆ(0) = f (0 + x̂) = f (x̂) = 0
the system in the new coordinates has an equilibrium
at ŵ = 0.
Useful trick!

Lent 2012 4F2: Nonlinear systems and Control, Lectures 3–4 – p. 3/32
Limit cycles
Another important class of invariant sets (n ≥ 2).
Limit cycle: A closed, non-trivial periodic trajectory.

Lent 2012 4F2: Nonlinear systems and Control, Lectures 3–4 – p. 4/32
Limit cycles
Another important class of invariant sets (n ≥ 2).
Limit cycle: A closed, non-trivial periodic trajectory.
Example: Van der Pol oscillator

θ̈ − ǫ(1 − θ2 )θ̇ + θ = 0

Lent 2012 4F2: Nonlinear systems and Control, Lectures 3–4 – p. 4/32
Limit cycles
Another important class of invariant sets (n ≥ 2).
Limit cycle: A closed, non-trivial periodic trajectory.
Example: Van der Pol oscillator

θ̈ − ǫ(1 − θ2 )θ̇ + θ = 0
Exercise: Write this in state-space form. Find any
equilibria.

Lent 2012 4F2: Nonlinear systems and Control, Lectures 3–4 – p. 4/32
Van der Pol (ǫ = 1)
5

x2
0

−1

−2

−3

−4
−3 −2 −1 0 1 2 3 4
x1

Lent 2012 4F2: Nonlinear systems and Control, Lectures 3–4 – p. 5/32
Exotic invariant sets
Invariant torus — like surface of a ring doughnut.
Chaotic attractor — eg Lorenz equations
ẋ1 = a(x2 − x1 )
ẋ2 = (1 + b)x1 − x2 − x1 x3
ẋ3 = x1 x2 − cx3

Lent 2012 4F2: Nonlinear systems and Control, Lectures 3–4 – p. 6/32
Lorenz attractor
25

20

15

10

x2 0

−5

−10

−15

−20

−25
−20 −15 −10 −5 0 5 10 15 20
x1

Don’t the crossings violate uniqueness of solution?

a = 10, b = 24, c = 2, x0 = [−5, −6, 20]T

Lent 2012 4F2: Nonlinear systems and Control, Lectures 3–4 – p. 7/32
Lorenz attractor

50

40

30
x3
20
30
10
20

10
0
−20 0
−15
−10
−5 −10
0
5 −20
10
15
20 −30 x2
x1

a = 10, b = 24, c = 2, x0 = [−5, −6, 20]T

Lent 2012 4F2: Nonlinear systems and Control, Lectures 3–4 – p. 7/32
Poincaré-Bendixson theorem
Theorem 1 Consider a two dimensional dynamical
system and let S ⊆ R2 be a compact (i.e. bounded
and closed) invariant set. If S contains no equilibrium
points then all trajectories starting in S are either
limit cycles themselves, or tend towards a limit cycle
as t → ∞.

Exotic things cannot happen in R2

Lent 2012 4F2: Nonlinear systems and Control, Lectures 3–4 – p. 8/32
 2
Corollary in R
Corollary 1 Consider a two dimensional dynamical
system.
1. A compact region S ⊆ R2 that contains no
equilibria and on whose boundary the vector field
points towards the interior of S must contain a
stable limit cycle.
2. The region encircled by a limit cycle (in R2 )
contains either another limit cycle or an
equilibrium.

Lent 2012 4F2: Nonlinear systems and Control, Lectures 3–4 – p. 9/32
Stability
Definition 3 (Stable equilibrium) An equilibrium x̂
is stable if, ∀ǫ > 0, ∃δ > 0 such that
kx0 − x̂k < δ ⇒ kx(t) − x̂k < ǫ
Otherwise the equilibrium is said to be unstable.

Lent 2012 4F2: Nonlinear systems and Control, Lectures 3–4 – p. 10/32
Stability
Definition 4 (Stable equilibrium) An equilibrium x̂
is stable if, ∀ǫ > 0, ∃δ > 0 such that
kx0 − x̂k < δ ⇒ kx(t) − x̂k < ǫ
Otherwise the equilibrium is said to be unstable.
8

2
x2

−2

−4

−6

−8
−10 −8 −6 −4
x1 −2 0 2 4

Lent 2012 4F2: Nonlinear systems and Control, Lectures 3–4 – p. 10/32
Asymptotic stability
Definition 5 (Asymptotically stable equilibrium)
An equilibrium x̂ is locally asymptotically stable if
1. It is stable, and
2. There exists M > 0 such that
kx0 − x̂k < M ⇒ limt→∞ x(t) = x̂.

Lent 2012 4F2: Nonlinear systems and Control, Lectures 3–4 – p. 11/32
Asymptotic stability
Definition 6 (Asymptotically stable equilibrium)
An equilibrium x̂ is locally asymptotically stable if
1. It is stable, and
2. There exists M > 0 such that
kx0 − x̂k < M ⇒ limt→∞ x(t) = x̂.

The equilibrium is called globally asymptotically

stable if this holds for all M > 0.

Lent 2012 4F2: Nonlinear systems and Control, Lectures 3–4 – p. 11/32
Asymptotic stability
Definition 7 (Asymptotically stable equilibrium)
An equilibrium x̂ is locally asymptotically stable if
1. It is stable, and
2. There exists M > 0 such that
kx0 − x̂k < M ⇒ limt→∞ x(t) = x̂.

The equilibrium is called globally asymptotically

stable if this holds for all M > 0.
The domain of attraction of an asymptotically stable
equilibrium is the set of all x0 for which x(t) → x̂.

Lent 2012 4F2: Nonlinear systems and Control, Lectures 3–4 – p. 11/32
 Linearisation
Taylor series expansion:
f (x) = f (x̂) + A(x − x̂) + higher order terms in (x − x̂)
= A(x − x̂) + higher order terms in (x − x̂)
where  
∂f1 ∂f1
∂x1 (x̂) ... ∂xn (x̂)

A= .. .. .. 
. . . 
∂fn ∂fn
∂x1 (x̂) . . . ∂xn (x̂)
Letting δx = x − x̂ gives:
˙ = Aδx + higher order terms in δx ≈ Aδx
δx

Lent 2012 4F2: Nonlinear systems and Control, Lectures 3–4 – p. 12/32
Lyapunov’s indirect method
Theorem 2 (Lyapunov’s indirect method) The
equilibrium x̂ is
1. Locally asymptotically stable if the linearisation is
asymptotically stable;
2. Unstable if the linearisation is unstable.

Lent 2012 4F2: Nonlinear systems and Control, Lectures 3–4 – p. 13/32
Lyapunov’s indirect method
Theorem 4 (Lyapunov’s indirect method) The
equilibrium x̂ is
1. Locally asymptotically stable if the linearisation is
asymptotically stable;
2. Unstable if the linearisation is unstable.
Theorem 5 (Stability of δx ˙ = Aδx)
1. Asymptotically stable if and only if all eigenvalues
of A have −ve real parts;
2. Stable if all eigenvalues of A have non-positive
real parts, and any eigenvalues on the imaginary
axis are distinct;
3. Unstable if ∃ an eigenvalue with +ve real part.
Lent 2012 4F2: Nonlinear systems and Control, Lectures 3–4 – p. 13/32
Behaviour near equilibrium

Stable Node Saddle point

Centre Stable Focus

Lent 2012 4F2: Nonlinear systems and Control, Lectures 3–4 – p. 14/32
Pendulum again
Linearisation about [θ, θ̇] = [0, 0]:
 
˙ 0 1
δx = g d δx
− ℓ −m
Eigenvalues: Solutions of
 
 λ −1  2 d g
det(λI − A) =  g d
=λ + λ+ =0
ℓ λ + m
 m ℓ
Both have negative real parts ⇒ asymptotically stable.

Lent 2012 4F2: Nonlinear systems and Control, Lectures 3–4 – p. 15/32
Pendulum again
Linearisation about [θ, θ̇] = [π, 0]:
 
˙ 0 1
δx = g d δx
+ ℓ −m
Eigenvalues: Solutions of
d 2 g
λ + λ− =0
m ℓ
One +ve real, one −ve real ⇒ unstable.

Lent 2012 4F2: Nonlinear systems and Control, Lectures 3–4 – p. 15/32
Sometimes inconclusive
Linearised system stable but not asymptotically stable
⇒ Inconclusive for nonlinear system.
Example:
3 3
ẋ = x
| {z } and ẋ = −x
| {z }
unstable stable
Equilibrium: x̂ = 0
˙ =0
Linearisation: δx stable — only

Lent 2012 4F2: Nonlinear systems and Control, Lectures 3–4 – p. 16/32
Lyapunov’s direct method
Suppose ẋ = f (x) and f (0) = 0.
Theorem 6 (Lyapunov Stability) Assume there
exists a differentiable function V : S → R defined on
some open region S ⊂ Rn containing 0, such that
1. V (0) = 0.
2. V (x) > 0 for all x ∈ S with x 6= 0.
3. V̇ (x) ≤ 0 for all x ∈ S.
Then x̂ = 0 is a stable equilibrium of ẋ = f (x).
Such a function V (.) is called a Lyapunov function.
It generalises idea of energy function.

Lent 2012 4F2: Nonlinear systems and Control, Lectures 3–4 – p. 17/32
Proof by picture
n n
X ∂V X ∂V
V̇ (x) = ẋi = fi (x) = ∇V · f (x)
i=1
∂xi i=1
∂xi

Lent 2012 4F2: Nonlinear systems and Control, Lectures 3–4 – p. 18/32
Proof by picture
n n
X ∂V X ∂V
V̇ (x) = ẋi = fi (x) = ∇V · f (x)
i=1
∂xi i=1
∂xi

ǫ {x ∈ Rn | V (x) ≤ c}

Lent 2012 4F2: Nonlinear systems and Control, Lectures 3–4 – p. 18/32
Asymptotic Stability
Theorem 7 (Lyapunov Asymptotic Stability)
Assume there exists a differentiable function
V : S → R defined on some open region S ⊂ Rn
containing the origin, such that
1. V (0) = 0.
2. V (x) > 0 for all x ∈ S with x 6= 0.
3. V̇ (x) < 0 for all x ∈ S with x 6= 0.
Then x̂ = 0 is a locally asymptotically stable
equilibrium of ẋ = f (x).

Lent 2012 4F2: Nonlinear systems and Control, Lectures 3–4 – p. 19/32
Easier condition to meet
Theorem 8 (LaSalle’s Theorem) Let S ⊆ Rn be a
compact invariant set. Assume there exists a
differentiable function V : S → R such that
V̇ (x) ≤ 0 ∀x ∈ S
Let M be the largest invariant set contained in
{x ∈ S | V̇ (x) = 0}. Then all trajectories starting in
S approach M as t → ∞.

Lent 2012 4F2: Nonlinear systems and Control, Lectures 3–4 – p. 20/32
Easier condition to meet
Corollary 2 (LaSalle for Equilibria) Let S ⊆ ℜn be
a compact invariant set. Assume there exists a
differentiable function V : S → R such that
V̇ (x) ≤ 0 ∀x ∈ S
If the set {x ∈ S | V̇ (x) = 0} contains no trajectories
other than x(t) = 0, then 0 is locally asymptotically
stable. Moreover, all trajectories starting in S
converge to 0.

Lent 2012 4F2: Nonlinear systems and Control, Lectures 3–4 – p. 20/32
Notes about Lyapunov
Lyapunov’s Asymptotic Stability Theorem and
LaSalle’s Theorem for equilibria can be used to
estimate domains of attraction. System can never
leave a set {x ∈ Rn | V (x) ≤ c}, so any such set in S
must be in the domain of attraction.

Lent 2012 4F2: Nonlinear systems and Control, Lectures 3–4 – p. 21/32
Notes about Lyapunov
Lyapunov’s Asymptotic Stability Theorem and
LaSalle’s Theorem for equilibria can be used to
estimate domains of attraction. System can never
leave a set {x ∈ Rn | V (x) ≤ c}, so any such set in S
must be in the domain of attraction.
The conditions of the theorems are sufficient but not
necessary. If we can not find a Lyapunov function we
can not draw any immediate conclusion.

Lent 2012 4F2: Nonlinear systems and Control, Lectures 3–4 – p. 21/32
How to find suitable V ?
Use intuition about particular system.

Lent 2012 4F2: Nonlinear systems and Control, Lectures 3–4 – p. 22/32
How to find suitable V ?
Use intuition about particular system.
Try simplest quadratic function kxk2 .

Lent 2012 4F2: Nonlinear systems and Control, Lectures 3–4 – p. 22/32
How to find suitable V ?
Use intuition about particular system.
Try simplest quadratic function kxk2 .
Try xT P x where P is symmetric, positive-definite
matrix.

Lent 2012 4F2: Nonlinear systems and Control, Lectures 3–4 – p. 22/32
How to find suitable V ?
Use intuition about particular system.
Try simplest quadratic function kxk2 .
Try xT P x where P is symmetric, positive-definite
matrix.
Positive-definite matrix: xT P x > 0 for any x 6= 0.
P = P T is positive-definite if and only if every
eigenvalue > 0. Write P > 0.

Lent 2012 4F2: Nonlinear systems and Control, Lectures 3–4 – p. 22/32
How to find suitable V ?
Use intuition about particular system.
Try simplest quadratic function kxk2 .
Try xT P x where P is symmetric, positive-definite
matrix.
Positive-definite matrix: xT P x > 0 for any x 6= 0.
P = P T is positive-definite if and only if every
eigenvalue > 0. Write P > 0.
Positive-semidefinite matrix:
xT P x ≥ 0 for any x 6= 0.
P = P T is positive-semidefinite if and only if every
eigenvalue ≥ 0. Write P ≥ 0.

Lent 2012 4F2: Nonlinear systems and Control, Lectures 3–4 – p. 22/32
Quadratic V for linear systems
Consider
ẋ = Ax and V (x) = xT P x, (P = P T > 0)

Lent 2012 4F2: Nonlinear systems and Control, Lectures 3–4 – p. 23/32
Quadratic V for linear systems
Consider
ẋ = Ax and V (x) = xT P x, (P = P T > 0)
Then
V̇ (x) = ẋT P x + xT P ẋ
T T
= (Ax) P x + x P (Ax)
T T
= x (A P + P A)x

We want V̇ (x) < 0. So we want −(AT P + P A) > 0.

Lent 2012 4F2: Nonlinear systems and Control, Lectures 3–4 – p. 23/32
Quadratic V for linear systems
Consider
ẋ = Ax and V (x) = xT P x, (P = P T > 0)
Then
V̇ (x) = ẋT P x + xT P ẋ
T T
= (Ax) P x + x P (Ax)
T T
= x (A P + P A)x

We want V̇ (x) < 0. So we want −(AT P + P A) > 0.

Matrix Lyapunov Equation
AT P + P A = −Q for some Q > 0.

Lent 2012 4F2: Nonlinear systems and Control, Lectures 3–4 – p. 23/32
Stability of Linear System
T
A P + P A = −Q
Theorem 9 (Linear Lyapunov Stability) For any
matrix A the following statements are equivalent
1. All eigenvalues of A have negative real parts.
2. We can find P = P T > 0 that solves Matrix
Lyapunov Equation for some Q = QT > 0.
3. We can find P = P T > 0 that solves Matrix
Lyapunov Equation for any Q = QT > 0.

Lent 2012 4F2: Nonlinear systems and Control, Lectures 3–4 – p. 24/32
Memoryless nonlinearities
h(0) = 0 and zh(z) > 0 for z 6= 0

Lent 2012 4F2: Nonlinear systems and Control, Lectures 3–4 – p. 25/32
Memoryless nonlinearities
h(0) = 0 and zh(z) > 0 for z 6= 0
Example:
1

0.8

0.6

0.4

0.2
h(z)
0

−0.2

−0.4

−0.6

−0.8

−1
−4 −3 −2 −1
z
0 1 2 3 4

Lent 2012 4F2: Nonlinear systems and Control, Lectures 3–4 – p. 25/32
Memoryless nonlinearities
h(0) = 0 and zh(z) > 0 for z 6= 0
Then Z xi
V (x) = h(z)dz > 0
0
and
∂V
(x) = h(xi )
∂xi
So try
n Z
X xi
V (x) = xT P x + hi (z)dz
i=1 0

Lent 2012 4F2: Nonlinear systems and Control, Lectures 3–4 – p. 25/32
Hopfield Neural Network

ui1 wi1

ui2 wi2
R zi xi
σ(·)
−

1
τ
ui(N +1) wi(N +1)

Model of one neuron

Lent 2012 4F2: Nonlinear systems and Control, Lectures 3–4 – p. 26/32
Sigmoid function
σ(·) : R → R : R 7→ (−1, 1)
eλz − e−λz
σ(z) = λz −λz
= tanh(λz)
e +e

Lent 2012 4F2: Nonlinear systems and Control, Lectures 3–4 – p. 27/32
Sigmoid function
σ(·) : R → R : R 7→ (−1, 1)
eλz − e−λz
σ(z) = λz −λz
= tanh(λz)
e +e
1

0.8

0.6

0.4
σ(zi )

0.2

−0.2

−0.4

−0.6

−0.8

−1
−4 −3 −2 −1
zi
0 1 2 3 4

Lent 2012 4F2: Nonlinear systems and Control, Lectures 3–4 – p. 27/32
Sigmoid function
σ(·) : R → R : R 7→ (−1, 1)
eλz − e−λz
σ(z) = λz −λz
= tanh(λz)
e +e
1

0.8

0.6

0.4

dσ
σ(zi )

0.2

0 >0
−0.2 dz
σ(·) invertible
−0.4

−0.6

−0.8

−1
−4 −3 −2 −1
zi
0 1 2 3 4

Lent 2012 4F2: Nonlinear systems and Control, Lectures 3–4 – p. 27/32
Dynamics of one neuron
Since xi = σ(zi ),
N +1
!
dσ dσ 1 X
ẋi = (zi )żi = (zi ) − zi + wik uik
dz dz τ
k=1
N +1
!
1 −1 X
= h(xi ) − σ (xi ) + wik uik
τ
k=1

where
dσ −1
h(xi ) = (σ (xi )) > 0
dz

Lent 2012 4F2: Nonlinear systems and Control, Lectures 3–4 – p. 28/32
Network of N neurons
Interconnection of neurons:
uik = xk , 1 ≤ k ≤ N

Lent 2012 4F2: Nonlinear systems and Control, Lectures 3–4 – p. 29/32
Network of N neurons
Interconnection of neurons:
uik = xk , 1 ≤ k ≤ N
Constant external inputs:
ui(N +1) = ui

Lent 2012 4F2: Nonlinear systems and Control, Lectures 3–4 – p. 29/32
Network of N neurons
Interconnection of neurons:
uik = xk , 1 ≤ k ≤ N
Constant external inputs:
ui(N +1) = ui
Symmetry of connection gains:
wik = wki , 1 ≤ i, k ≤ N

Lent 2012 4F2: Nonlinear systems and Control, Lectures 3–4 – p. 29/32
Network of N neurons
Interconnection of neurons:
uik = xk , 1 ≤ k ≤ N
Constant external inputs:
ui(N +1) = ui
Symmetry of connection gains:
wik = wki , 1 ≤ i, k ≤ N
Gain on external inputs:
wi(N +1) = 1

Lent 2012 4F2: Nonlinear systems and Control, Lectures 3–4 – p. 29/32
Network of N neurons
Interconnection of neurons:
uik = xk , 1 ≤ k ≤ N
Constant external inputs:
ui(N +1) = ui
Symmetry of connection gains:
wik = wki , 1 ≤ i, k ≤ N
Gain on external inputs:
wi(N +1) = 1
Will trajectories converge to (an) equilibrium?

Lent 2012 4F2: Nonlinear systems and Control, Lectures 3–4 – p. 29/32
 N
!
1 −1 X
ẋi = h(xi ) − σ (xi ) + wik xk + ui
τ
k=1

Lent 2012 4F2: Nonlinear systems and Control, Lectures 3–4 – p. 30/32
 N
!
1 −1 X
ẋi = h(xi ) − σ (xi ) + wik xk + ui
τ
k=1

Let
N Z xi N
1 T 1 X X
V (x) = − x W x + σ −1 (s)ds − xi ui
2 τ i=1 0 i=1

Note wik = wki ⇒ W = W T .

Lent 2012 4F2: Nonlinear systems and Control, Lectures 3–4 – p. 30/32
 N
!
1 −1 X
ẋi = h(xi ) − σ (xi ) + wik xk + ui
τ
k=1

Let
N Z xi N
1 T 1 X X
V (x) = − x W x + σ −1 (s)ds − xi ui
2 τ i=1 0 i=1

Note wik = wki ⇒ W = W T .

Then
∂V
ẋi = −h(xi ) (x)
∂xi

Lent 2012 4F2: Nonlinear systems and Control, Lectures 3–4 – p. 30/32
 N N  2
X ∂V X ∂V
V̇ (x) = (x)ẋi = − h(xi ) (x) ≤ 0
i=1
∂xi i=1
∂xi

since h(xi ) > 0.

Lent 2012 4F2: Nonlinear systems and Control, Lectures 3–4 – p. 31/32
 N N  2
X ∂V X ∂V
V̇ (x) = (x)ẋi = − h(xi ) (x) ≤ 0
i=1
∂xi i=1
∂xi

since h(xi ) > 0.

∂V
V̇ (x) = 0 ⇒ (x) = 0 ⇒ ẋi = 0
∂xi
So: V̇ (x̂) = 0 ⇒ x̂ is an equilibrium.

Lent 2012 4F2: Nonlinear systems and Control, Lectures 3–4 – p. 31/32
Now apply LaSalle
But recall:
Corollary 3 (LaSalle for Equilibria) Let S ⊆ ℜn be
a compact invariant set. Assume there exists a
differentiable function V : S → R such that
V̇ (x) ≤ 0 ∀x ∈ S
If the set {x ∈ S | V̇ (x) = 0} contains no trajectories
other than x(t) = 0, then 0 is locally asymptotically
stable. Moreover, all trajectories starting in S
converge to 0.

Lent 2012 4F2: Nonlinear systems and Control, Lectures 3–4 – p. 32/32
Now apply LaSalle
But recall:
Corollary 4 (LaSalle for Equilibria) Let S ⊆ ℜn be
a compact invariant set. Assume there exists a
differentiable function V : S → R such that
V̇ (x) ≤ 0 ∀x ∈ S
If the set {x ∈ S | V̇ (x) = 0} contains no trajectories
other than x(t) = 0, then 0 is locally asymptotically
stable. Moreover, all trajectories starting in S
converge to 0.
Hence deduce convergence to equilibrium.

Lent 2012 4F2: Nonlinear systems and Control, Lectures 3–4 – p. 32/32