EECE 571M/491M, Spring 2008

Lecture 5

Stability lectures
Linear system stability

!
!

Stability of Continuous Systems

Nonlinear system stability

Dr. Meeko Oishi

Electrical and Computer Engineering

!
!
!

[email protected]

Khalil 4.3, Friedland A.10,

Tomlin LN 6
1

Definition of equilibrium
Multiple Lyapunov functions
Common Lyapunov function
Piecewise quadratic Lyapunov functions

EECE 571M / 491M Winter 2007

Stability of Linear Systems

Stability of Linear Systems

Linear System Asymptotic Stability Theorem:

!
The autonomous system

dx/dt = A x,
x(0) = x0

is asymptotically stable if and only if the eigenvalues of

A have strictly negative real part:
!i(A) < 0.
!

Lyapunovs indirect method

Lyapunovs direct method

Hybrid system stability

University of British Columbia, BC

http://courses.ece.ubc.ca/491m

Eigenvalues of A
Linear quadratic Lyapunov functions
Ellipses / LMIs

LTI systems can be represented in a variety of

coordinate systems, yet all representations share the
same stability properties.
Exercise: Consider the invertible transformation
z = T-1x.
!

What are the eigenvalues of the transformed system matrix?

(Use Cayley-Hamilton theorem.)

The trajectories of the system will follow

x(t) = exp(At) x0

which converges exponentially to 0 as x -> ! .

EECE 571M / 491M Winter 2007

EECE 571M / 491M Winter 2007

Linear system stability

!

Linear system stability

Example: Spring-mass-damper
system
!
!
!

Spring constant k
Damping coefficient b
Mass m

EECE 571M / 491M Winter 2007

Phase-plane analysis
!

Stable trajectories will tend towards the origin and can be classified
according to the types of eigenvalues:
!

!
!

!
!
!

Both negative real numbers (stable node)

Complex conjugate pair with negative real part (stable focus)

Unstable trajectories tend towards infinity and can be classified

according to the types of eigenvalues:
!

Eigenvalues occur where

And have negative real part for

EECE 571M / 491M Winter 2007

Lyapunov stability for LTI sys.

For 2-D linear systems, the phase plane plot is a plot of trajectories
in (x1, x2) space.

Example: Spring-mass-damper
system

For the dynamical system

consider the quadratic Lyapunov function

whose time-derivative

can be written as

Both positive real numbers (unstable node)

Complex conjugate pair with positive real part (unstable focus)
Positive and negative real numbers (saddle)

What happens when one eigenvalue has 0 real part?

What happens when both eigenvalues have 0 real part?
What happens when eigenvalues are repeated?

EECE 571M / 491M Winter 2007

EECE 571M / 491M Winter 2007

Lyapunov stability for LTI sys.

Positive definite matrices

Theorem: Lyapunov stability for linear systems

A matrix P is positive definite iff

A matrix P is positive semi-definite iff

A matrix P is negative definite iff

A matrix P is negative semi-definite iff

The system
is asymptotically stable about x=0 if and only if
for any positive definite Q, there exists a positive definite P
such that
This is known as the Lyapunov equation.

Further,
uniquely solves the Lyapunov equation.

EECE 571M / 491M Winter 2007

Positive definite matrices

!

EECE 571M / 491M Winter 2007

10

Positive definite matrices

Any real, symmetric positive definite matrix P=PT > 0 has real
eigenvalues !i > 0

Evaluate the following matrices for positive definiteness:

Note that P can be diagonalized by an orthonormal basis

where the columns of
are normalized, orthogonal
eigenvectors (even if there are repeated eigenvalues)

Values of V(x)=xTPx are bounded by

EECE 571M / 491M Winter 2007

11

EECE 571M / 491M Winter 2007

12

Positive definite matrices

!

Example: SMD system

Interpretation: The function

represents an ellipse in Rn with axes along the

eigenvectors of P

!
!

Consider the 2D case:

For c=1, with

EECE 571M / 491M Winter 2007

13

Lyapunov stability
!
!
!

EECE 571M / 491M Winter 2007

14

is stable in the sense of Lyapunov if for all " > 0

there exists a # > 0 such that

V(x)

15

Definition:
The equilibrium point x* of
dx/dt = f(x), x(0) = x0

V(x)

(From Murray, Li, Sastry)

Assume m=1, b=2, k=1.

Choose Q = I.
By computing a P that is positive
definite, we prove asymptotic
stability.

Lyapunov stability

What if asymptotic stability is not possible?

If V always decreases, then the system is asymptotically stable
If V decreases or maintains a constant value as time increases,
then the system is stable in the sense of Lyapunov.

EECE 571M / 491M Winter 2007

Consider the standard autonomous

spring-mass-damper system.

Definition:
The equilibrium point x* is asymptotically stable if it
is stable and # can be chosen such that

EECE 571M / 491M Winter 2007

16

Lyapunov stability for LTI sys.

Lyapunov stability

Linear quadratic Lyapunov theorems

Converse linear quadratic Lyapunov theorems

!
!
!

If P > 0, Q > 0, then system is asymptotically stable

If P > 0, Q " 0, then system is stable in the sense of Lyapunov
If P > 0, Q " 0, and (Q, A) observable, then system is
asymptotically stable

If P " 0, Q " 0, then the system is not stable.

17

Example: SMD system

!

If A is asymptotically stable, then there exists P > 0, Q > 0 that

satisfy the Lyapunov equation
If A is stable and Q " 0, then P " 0
If A is stable, Q " 0, and (Q, A) is observable, then P > 0

If P > 0, Q " 0, the sublevel sets of { x | xTPx # a } are

invariant and are ellipsoids

EECE 571M / 491M Winter 2007

Case 1:
!
!

Case 1:
Proves stability in the sense of Lyapunov,
since

P > 0, Q > 0 --> therefore the system is globally

asymptotically stable.

Case 2:
Proves asymptotic stability since dV/dt < 0

P > 0, Q " 0 --> therefore the system is stable.

But is (Q, A) observable?

Case 2:
!

18

Example: SMD system

Consider the spring-mass-damper system

again.
Other Lyapunov functions are possible.
!

EECE 571M / 491M Winter 2007

Now evaluate each of these functions in

terms of the linear quadratic Lyapunov and
converse linear quadratic Lyapunov
theorems

EECE 571M / 491M Winter 2007

19

By the converse Lyapunov theorem, we know that

since eig(A) < 0 a quadratic Lyapunov function must
exist.

EECE 571M / 491M Winter 2007

20

Example 2
!

Consider the linear system with

!

Solving the Lyapunov equation

Integral solution to the Lyapunov equation

A = [-1 10; -100 1]

If dx/dt = Ax is asymptotically stable and Q = QT > 0, OR

if dx/dt = Ax is stable in the sense of Lyapunov & Q = QT " 0,

Does a quadratic Lyapunov function that

satisfies the Lyapunov equation exist?
is the unique solution to the Lyapunov equation

EECE 571M / 491M Winter 2007

21

Solving the Lyapunov equation

!

!
!
!

To show stability, we want to find a positive definite matrix P

such that

EECE 571M / 491M Winter 2007

Solving the Lyapunov equation

!

GUI to specify LMIs

The system of LMIs is encoded in lmisys

The optimization to find a feasible solution to the LMI is called through

>> lmiedit

More help can be found at

>> help lmidem

If the problem has a solution, the algorithm will find it.

If the problem does not have a solution, the algorithm will return a
certificate which indicates as such.

23

Or through the demo

>> help lmidem

The Matlab LMI Control Toolbox can solve this in O(n3).

EECE 571M / 491M Winter 2007

22

>> [tmin, pfeas]=feasp(lmisys)

>> p = dec2mat(lmisys, pfeas, p)

The variable in these matrix equations is the matrix P

This is known as a Linear Matrix Inequality
Efficient tools have been developed to quickly solve LMIs by
posing them as convex optimization problems.
!

If you know Q, you can also use P = lyap(A,Q) in Matlab

More on this later

EECE 571M / 491M Winter 2007

24

Summary: Linear Sys. Stability

!
!
!
!

Quadratic Lyapunov functions for linear systems

Positive definite matrix properties
Linear quadratic Lyapunov stability theorem for linear systems
Necessary and sufficient conditions for stability (special case for
linear systems)
!
!

Converse theorems
Asymptotic stability vs. stability in the sense of Lyapunov

Extensions to nonlinear sys.

!

Tools to solve the Lyapunov equation and LMIs

!

EECE 571M / 491M Winter 2007

25

Nonlinear System Stability

EECE 571M / 491M Winter 2007

26

Lyapunovs second method or direct method

Theorem:
Let x*=0 be the equilibrium of
dx/dt = f(x), x(0) = x0

and let
be a set containing x*. If V: D->R is a
continuously differentiable function such that

such that the linearized system is

Then
! x* is asymptotically stable if Re(!i)<0 for all
eigenvalues of A.
! x* is unstable if Re(! i)>0 for at least one eigenvalue of
A.

EECE 571M / 491M Winter 2007

As opposed to linear systems, nonlinear systems can have

multiple equilibria.
As opposed to linear systems, nonlinear system stability is
often only a local result (e.g., valid within some neighborhood
of the equilibrium point).
As opposed to linear systems, nonlinear systems rarely have
closed-form solutions (e.g., there is no x(t) = eAt x(0) ).
In addition to the behavior around equilibria that arose in linear
systems, nonlinear systems may exhibit orbits, limit cycles,
bifurcations, and other phenomena.

Nonlinear System Stability

Theorem: Lyapunovs indirect method

!
Let x*=0 be the equilibrium of the differentiable function
dx/dt = f(x), x(0) = x0
and let
be a set containing x*. Let

Nonlinear systems have significant differences that complicate

stability analysis.

then x* is stable. Further more, if x*=0 is stable and

then x* is asymptotically stable.
27

EECE 571M / 491M Winter 2007

28

Lyapunov stability

Summary: Nonlinear sys.

Lyapunov function is sufficient condition for stability

Evaluating eigenvalues is necessary and sufficient for stability

Allows trajectories which do not converge to the origin to be stable.

Stability in the sense of Lyapunov

!

Indirect method:
!

!
!

!
!

If the system is stable, then there exists a Lyapunov function.

If a Lyapunov function cannot be found, nothing is known about the
stability of the system.
For general nonlinear systems, these functions can be hard to find.
Recent computational tools in LMIs and polynomial functions can
provide numerical computations of Lyapunov functions.

Direct method:
!

EECE 571M / 491M Winter 2007

29

Next couple of weeks

!

Introduction to hybrid stability

!
!

Multiple Lyapunov functions

!

(Most general stability theory)

Global quadratic Lyapunov functions

!

Hybrid equilibrium
Hybrid stability

(specific to hybrid systems with linear dynamics and arbitrary

switching)

Piecewise quadratic Lyapunov functions

!

(hybrid systems with linear or affine dynamics and state-based

switching)

EECE 571M / 491M Winter 2007

31

If the linearization is asymptotically stable, then the nonlinear

system is locally asymptotically stable.
If the linearization is unstable, then the nonlinear system is
locally unstable.
In general, no conclusions are possible regarding the
nonlinear system if the eigenvalues have 0 real part. (Some
exceptions for 2D systems -- Hartman-Grobman theorem)
If you can find a Lyapunov function, then you know the system
is locally stable in the sense of Lyapunov.
Sufficient condition for stability

EECE 571M / 491M Winter 2007

30