Nonlinear Control - An Overview: Fernando Lobo Pereira, Flp@fe - Up.pt

1
NONLINEAR CONTROL - AN OVERVIEW
Fernando Lobo Pereira, [email protected]
Porto University, Portugal

Nonlinear Control, 2008, FEUP
References
1 M. Vidyasagar, Nonlinear Systems Analysis, Prentice-Hall, Englewood Cliffs, NJ,
1978.
2 A. Isidori, Nonlinear Control Systems, Springer-Verlag, 199X.
3 E. Sontag, Mathematical Control Theory: Deterministic Finite Dimensional
Systems, Springer-Verlag, 1998.
4 C. T. Chen, Introduction to Linear Systems Theory, Van Nostrand Reinhold
Company, New York, NY, 1970.
5 C. A. Desoer, Notes for a Second Course on Linear Systems, Holt Rhinehart,
Winston, 1970.
6 K. Ogata, State Space Analysis of Control Systems, Prentice-Hall, Englewood
Cliffs, NJ, 1965.
7 J. M. Carvalho, Dynamical Systems and Automatic Control, Prentice-Hall, New
York, NY, 1994.
Fernando Lobo Pereira, [email protected] 1

Introduction
Problem description - system components, objectives and control issues.
Example 1: telescope mirrors control
PICTURE
Most of real systems involve nonlinearities in one way or another.

However, many concepts for linear systems play an important role since some of the
techniques to deal with nonlinear systems are:
• Change of variable in phase space so that the resulting system is linear.
• Nonlinearities may be cancelled by picking up appropriate feedback control laws.
• Pervasive analogy ...

Introduction
Example 2: Pendulum
Modeling
PICTURE
Dynamics (Newton’s law for rotating objects):
mθ̈(t) + mgsin(θ(t)) = u(t)

State variables: θ, θ̇
Control/Input variable: u (external applied torque)
Let us assume m = 1, l = 1, g = 1.

Introduction
Control Design
½
(0, 0) stable equilibrium
Stationary positions: (1)
(π, 0) unstable equilibrium.
Consider the local control around the later.

Objective:
Apply u so that, for small (θ(0), θ̇(0)), (θ(t), θ̇(t)) → (π, 0) as t → ∞.
Step 1 - Analysis - Linearization

For θ close to π: sin(θ) = −(θ − π) + o(θ − π).
Let ϕ = θ − π, then:
ϕ̈(t) − ϕ(t) = u(t)

Introduction
Step 2 - Control Synthesis
Take ϕ > 0 and pick u(t) = −αϕ(t) with α > 0. Then
ϕ̈(t) + (α − 1)ϕ(t) = 0.

 α > 1 oscillatory behavior.
α = 1 only stable point: ϕ̇(0) = 0 √ (2)

α < 1 stable points: ϕ̇(0) = −ϕ(0) 1 − α
Conclusion: “proportional control” does not suffice.

Some anticipative control action is required.
This achieved by “proportional and derivative control”, i.e.,
u(t) = −αϕ(t) − β ϕ̇(t)
with α > 1 and β > 0.
Then
ϕ̈(t) + β ϕ̇(t) + (α − 1)ϕ(t) = 0.
√
−β± β 2 −4(α−1)
The roots of the characteristic polynomial are 2 .
Introduction
Linearization Principle
Linear designs apply to linearized nonlinear systems operating locally.
Extension: controller scheduling

Organize the phase space into regions and “patch” together local linear designs.

Introduction
Issues in control system design:
1 - Study the system of interest and decide on sensors and actuators

2 - Model the resulting system to be controlled
3 - Simplify the model so that it becomes tractable
4 - Analyze the system to determine its properties
5 - Define performance specifications
6 - Design the controller to meet the specifications
7 - Evaluate the design system via, say, simulation
8 - If not happy go to the beginning; otherwise
9 - Proceed to implementation
10 - Being the case, tune the controller on-line

Introduction
Strategy of the course:
• Easy introduction to the main concepts of nonlinear control by making as much use as
possible of linear systems theory and with as little as possible Mathematics.
• Example-oriented introduction to concepts.
• Matlab hands-on exercises

Introduction
Structure of the course:
• Selected topics of linear systems

• Issues and specific background topics for nonlinear systems
• Main issues Lyapunov Stability
• Examples of feedback linearization design
• Basic issues in optimal control
• Introduction to adaptive control

Selected topics of linear systems
State-space representation of Linear Dynamic Systems

½
ẋ(t) = A(t)x(t) + B(t)u(t) [t0, tf ] − a.e., x(t0) = x0
y(t) = C(t)x(t) + D(t)u(t) [t0, tf ] − a.e.
where
• x ∈ <n, y ∈ <q , and u ∈ <m
• A ∈ <n×n, B ∈ <n×m, C ∈ <q×n, and D ∈ <q×m
System’s state trajectory given by a closed form solution

Z t
x(t) = φ(t, t0)x0 + φ(t, τ )B(τ )u(τ )dτ
t0
where φ(t, s) is the State Transition Matrix from s to t.


Definition of State Transition Matrix Fundamental matrix X(·) of the pair (A(·), X0) is
the solution to Ẋ(t) = A(t)X(t), X(t0) = X0.
The State Transition Matrix φ(t, t0) is given by X(t) when X0 = I.
Thus, it satisfies the n × n matrix differential equation
∂
φ(t, t0) = A(t)φ(t, t0).
∂t
Main Properties
1. φ(t, t0) is uniquely defined
2. The solution to ẋ(t) = A(t)x(t), x(t0) = x0, is given by φ(t, t0)x0
3. For all t, t0, t1, φ(t, t0) = φ(t, t1)φ(t1, t0)
4. φ(t1, t0) is nonsingular and [φ(t, t0)]−1 = φ(t0, t)
5. φ(t, t0) = X(t, t1)[X(t0)]−1
Rt Rt R σ1
6. φ(t, t0) = I + t0 A(σ1)dσ1 + t0 A(σ1) t0 A(σ2)dσ2dσ1 +
Rt R σ1 R σ2
t0 A(σ1 ) t0 A(σ2 ) t0 A(σ3 )dσ3 dσ2 dσ1 + ...


“Variations-by-parts” formula
Exercise: Check “variations-by-parts” formula by noting that
Z t
ẋ(t) = φ̇(t, t0)x0 + φ(t, t0)B(t)u(t) + φ̇(t, s)B(s)u(s)ds
t0
Heuristic derivation
Let u(σ) = u(s) when σ ∈ [s, s + ds] where t0 < s < t. Then
• x(s) = φ(s, t0)x0
• x(s + ds) ' x(s) + [A(s)x(s) + B(s)u(s)] ds
= [φ(s, t0) + A(s)φ(s, t0)ds] x0 + B(s)u(s)ds
' φ(s + ds, t0)x0 + B(s)u(s)ds
• x(t) = φ(t, s + ds)x(s + ds) ' φ(t, t0)x0 + φ(t, s + ds)B(s)u(s)ds
• Summing the input contributions for all s:
Z t
x(t) = φ(t, t0)x0 + φ(t, s)B(s)u(s)ds
t0


Computation of the State Transition Matrix
Z t X∞ Z t
1
φ(t, t0) = exp( A(s)ds) = ( A(s)ds)i
t0 i=0
i! t0
XN Z t
= αi(t, t0)( A(s)ds)i (3)
i=0 t0
for some N ≤ n − 1. The last equality holds due to the Cayley Hamilton Theorem:
p(A) = 0.
Here p is the characteristic polynomial of A.
Algorithm
Rt
• Compute eigenvalues and eigenvectors of t0 A(s)ds
• Compute coefficients αi(t, t0) by solving a system of linear functional equations
obtained by noting that A and its eigenvalues satisfy (3)
• Plug the αi(t, t0)’s in (3) in order to obtain φ(t, t0)


Observations:
• The above algorithm is more generic: it enables the computation of any sufficiently
regular function of the matrix A.
• When all the eigenvalues are distinct, N = n − 1 and one equation is written for each
eigenvalue. Otherwise, N is the degree of the Minimal polynomial.
• When an eigenvalue has multiplicity d > 1 and its greatest Jordan block of its
eigenspace has dimension m, then q equations associated with this eigenvalue have to
be added by taking the successive derivatives (from 0 to q − 1) w.r.t. the eigenvalue of
both sides.
• Geometric interpretation plays a key role.
Def. The minimal polynomial ψ is the one of least degree s. t. ψ(A) = 0.
Let A have σ distinct eigenvalues, λi whose multiplicity is di and let mi be the dimension
of the associated greatest invariant subspace. Then:
ψ(s) = Πσi=1(λ − λi)mi and p(s) = Πσi=1(λ − λi)di .


Jordan Representation
Let ℵk = ℵ[(A − λk I)mk ] and Bk a basis for ℵk . Then:
• dim(ℵk ) = dk
n
L L L
• < = ℵ1 ℵ2 ... ℵσ
Sσ
• A is represented by diag(Ai) in B = 1 Bi where Bi is a basis for ℵi
• det(A − λI) = Πσi=1det(Ai − λIi) = Πσk=1αk (λ − λk )dk
Function of a Matrix P P
m −1
f (A) = σk=1 l=0k f (l)(λk )pkl (A) where
(λ−λk )l φk (λ)
• pkl (λ) = l! , l = 0, ..., mk −1
ψ(λ)
• φk (λ) = nk (λ) (λ−λ m , and
k) k
1
• nk (λ) are the coefficients of a partial fraction expansion of ψ(λ) , i.e.,
1
Pσ nk (λ)
ψ(λ) = k=1 (λ−λ )mk k


Geometric Interpretation of Eigenvalues and Eigenvectors
Assume that the eigenvalues of A, λi, i = 1, ..., n are distinct and denote by vi the
associated eigenvectors.
Pn Pn
Let x(0) = i=1 αivi. Then x(t) = i=1 αieλitvi.
Exercise 1 Let A be a real valued square matrix. Show that, under the above
assumptions, the set of eigenvectors can be considered a basis for the state space.
Exercise 2 Let M be a matrix whose columns are the eigenvectors of A. Show that if ξ is
the representation of x in the basis formed by the eigenvalues of A, then
˙ = Λξ(t) + M −1Bu(t),
ξ(t) y(t) = CM ξ(t) + Du(t)
where Λ = M −1AM is a diagonal matrix.
Exercise 3 Let λi be a complex eigenvalue of A. Show that its conjugate is also an
eigenvalue of A and that if x(0) = Re(λi), then
x(t) = eRe(λi)t[cos(Im(λi)t)Re(vi) + sin(Im(λi)t)Im(vi)].

Trajectory graphical analysis

Representation in <2 suffices to cover all the situations.
a) Stable node λ1 < λ2 < 0
b) Unstable node λ1 > λ2 > 0
c) Saddle point λ1 > 0 > λ2
d) Unstable focus Re(λ1) > 0
e) Stable focus Re(λ1) < 0
f) Center Re(λ1) = 0

Realization theory: controllability

Definition - A system representation is completely controllable on a given time interval if
every initial state is controllable in that time interval, i.e., there is a control that transfers
the initial state to the origin at the final time.
Characterization R t1
Complete Controllability on [t0, t1] ⇐⇒ M (t0, t1):= t0 φ(t1, s)B(s)B T (s)φT (t1, s)ds > 0.
Exercise 1 Check that u(t) = −B T(t)φT(t1, t)M −1(t0, t1)φ(t1, t0)x0 drives x0 at t0 to 0 at
t1.
Observation: A contradiction argument shows that, for invariant systems, the
controllability condition is equivalent to rank of Q is n, where Q := [B|AB|...|An−1B].
Exercise 2 Show that the Range of Q, <(Q), is
a) the set of states that can be reached from the origin.
b) invariant under A.


Realization theory: observability
Z t
y(t) = C(t)φ(t, t0)x0 + C(t) φ(t, s)B(s)u(s)ds + D(t)u(t)
t0
Definition - A state is unobservable if its input response is identically zero. A system
representation is completely observable on [t0, t1] if no state is unobservable on that time
interval.
Characterization R t1 T
Complete Observability on [t0, t1] ⇐⇒ N (t0, t1) := t0 φ (s, t0)C T (s)C(s)φ(s, t0)ds > 0.
R t1 T
Exercise 1 Check that x0 = N (t0, t1) t0 φ (s, t0)C T (s)y(s)ds.
−1
An invariant system is completely observable if and only if rank of R is n, where

R := col[C|CA|...|CAn−1].
Exercise 2 Let ℵ(R) denote the null space of R.
a) x0 is unobservable if and only if x0 ∈ ℵ(R).
b) ℵ(R) is invariant under A.
Minimal Realization
Definition - A representation is a minimal representation if it is completely observable
and completely controllable
Fact - Minimal representation ⇐⇒ rank(RQ) = n

Issues and specific background topics for nonlinear systems

Examples of nonlinear control systems
ẋ(t) = f (t, x(t), u(t)), x(0) = x0

Issues: Existence, Uniqueness
• ẋ(t) = −sign(x(t)), x(0) = 0
1
• ẋ(t) = 2x(t) , x(0) = 0
Definitions:
Autonomy - No t-dependence; even through u
Equilibrium Point - f (t, x0) = 0, ∀t > t0
Relevance: x(t) = x0 ∀t > t0
Isolated Equilibrium point - There is a neighborhood of x0 where no additional
equilibrium points can be found.
Exercise - List the Pendulum equilibrium points

Second order Systems
ẋ1(t) = f1(t, x1, x2)
ẋ2(t) = f2(t, x1, x2)
The direction of the vector field f : <2 → <2, also referred to as velocity vector field is
given by θf (x) = arctan( ff21(t,x 1 ,x2 )
(t,x1 ,x2 ) ).
Methods to “compute” state plane trajectories:
• Linearization - Consider the first order approximation to the dynamics and use results
of planar linear systems. The correspondence ok for all but center equilibrium points.
Why?
Example: ÿ(t) − µ[1 − y 2(t)]ẏ(t) + y(t) = 0
• Graphical Euler - forward Euler numerical integration
x2+ M x2 = x2 + s(x1, x2) M x1 where s(x1, x2) := ff21(t,x 1 ,x2 )
(t,x1 ,x2 )
• Isocline - Efficient when sketching trajectories from a set of initial points. For various
values of α, find (x1, x2) s.t. s(x1, x2) := α.
• Vector Field - As before but the resulting plot is the vector field itself.

Periodic Solutions
Consider the following pair of planar dynamic systems:
½ ½
ẋ1(t) = x2(t) ẋ1(t) = x2(t) + αx1(t)(β 2 − x21(t) − x22(t))
ẋ2(t) = −x1(t) ẋ2(t) = −x1(t) + αx2(t)(β 2 − x21(t) − x22(t))
Observation: Although both exhibit periodic solutions, they are of quite different nature.
Contrary to the linear system, the nonlinear one has only one isolated periodic solution
which is approached as t → ∞ regardless of the initial condition.
Bendixson’s Theorem - condition for the nonexistence of limit cycles
If D be a simply connected domain in <2 in which
∂f1 ∂f2
∇f (x) :=
(x1, x2) + (x1, x2)
∂x1 ∂x1
neither is identically null nor changes sign, then D does not contains closed trajectories.
Exercise: Test this result with the linear and nonlinear systems given above for different
domains.
Periodic Solutions
Poincare Bendixson Theorem - existence of periodic trajectories.
Let L be the set of limit points of a trajectory S such that L is contained in a closed,
bounded region M which does not contain equilibrium points.
Then: either L or S is a periodic trajectory.
By Limit set it is meant the set of all those points in the state space which are visited
infinitely often as time goes to ∞.
Exercise 1: Take M := {(x, y) : 1/2 ≤ x2 + y 2 ≤ 3/2} and the nonlinear system in the
previous slide.
Exercise 2: Analyze the system f (x) := col(−x1 + x2, −x1 − x2) in the closed unit disk.
Suggestion: consider a change to polar coordinates.


Index Theory
Let J denote a simple, closed, positively (counterclockwise) oriented Jordan curve in an
open, simply connected subset D of <n not passing through any equilibrium point of f .
1
R
Definitions - If (J) := 2π J dθf (x) - The index of a curve J w.r.t. f is the net change in
the direction of f as x traverses around J divided by 2π.
If p is the only equilibrium point lying in J, then the index of a p, If (p), is If (J).
Here, θf (x) := arctan(f2/f1)(x).
Facts:
• If J has no equilibrium points in its interior, then If (J) = 0.
• The index of a center, focus and node is 1, and the index of a saddle is −1.
PN
• If pi, i = 1, ..., N , are equilibrium points “in J”, then If (J) = i=1 If (pi).
• If J is a trajectory of f , then If (J) = 1.
• If f and g are such that |θf − θg | < π along a J that does not pass through any
equilibrium point of either f or g, then If (J) = Ig (J)

Index Theory (cont.)

Theorem
Every closed trajectory of f has at least one equilibrium point in its interior.
If f has only isolated equilibrium points, then it has only a finite number of such points
and the sum of their indices is 1.
Exercise
Consider the Volterra predator-prey equations:
½
ẋ1 = −x1 + x1x2
(4)
ẋ2 = x2 − x1x2
a) Which variable represents the number of prey and that of predators?
b) Determine and classify the system equilibrium points.
c) Sketch the vector field.


A simple method to compute closed trajectories
½
ẋ1 = f1(x1, x2)
Take and V : D → < s.t. D ⊂ <2 is open and
ẋ2 = f2(x1, x2)
∂V ∂V
V̇ (x1, x2) = f1(x1, x2) + f2(x1, x2) ≡ 0 on D.
∂x1 ∂x2
Let S̄ := {(x1(t), x2(t)) : trajectory for f, (x1(0), x2(0)) = (x01, x02), t > 0} and
S := {(x1, x2) : V (x1, x2) = V (x01, x02)}.
Certainly S̄ ⊂ S.
Under some assumptions: If S is a closed curve, then S̄ is a closed trajectory of f .
Furthermore: {(x1, x2) : V (x1, x2) = constant} may define a continuum of closed
trajectories of f
Exercise
Compute a family of closed trajectories of the prey-predator system by taking
V (x1, x2) = V1(x1) + V2(x2).

Lyapunov Stability
Concept of Stability
It pertains the property of how a dynamic system drives or not its state to a given initial
equilibrium point from which it was removed by some perturbation.
This is concept is independent of the control activity.
An important application is to support the design of a feedback control so that the closed
loop system has the desired stability properties at the equilibrium points of interest.
There are several ways of going back to the equilibrium point.

Lyapunov Stability
Definitions of Stability Assume with no loss of generality that x0 = 0 is an equilibrium of
ẋ(t) = f (t, x(t)) at t = t0.
Types of stability
 

 Stable at t0 Uniformly Stable over [t0, ∞) 

 


 iff iff 

 
∀² ∃δ(t0, ²) such that ∀² ∃δ(²) such that
x0 is

 kx(t0)k ≤ δ(t0, ²) and kx(t1)k ≤ δ(²) and t1 ≥ t0  

 


 ⇓ ⇓ 

 
kx(t)k ≤ ² ∀t ≥ t0 kx(t)k ≤ ² ∀t ≥ t1
 

 Asymptotically Stable Uniformly Asymptotically Stable 


 


 at t0 iff over [t0, ∞) iff 


 

 it is stable at t0 and it is uniformly stable over [t0, ∞) and 
x0 is ∃γ(t0) > 0 such that and ∃γ > 0 such that

 


 kx(t0)k ≤ γ(t0) kx(t1)k ≤ γ and t1 ≥ t0 


 


 ⇓ ⇓ 

 
kx(t)k → 0 as t → ∞ kx(t)k → 0 as t → ∞

Lyapunov Stability
Definitions of Stability (cont.)
Exercise 1 - Check if (0, 0) is a stable equilibrium point of
½
ẋ1(t) = x2(t)
ẋ2(t) = −x1(t) + (1 − x21(t))x2(t)
Exercise 2 - Stability and Uniform Stability
a) Check that the solution to ẋ(t) = (6t sin(t) − 2t)x(t) is given by
ln(x(t)/x(t0)) = t20 − t2 + 6t0 cos(t) − 6t cos(t) − 6 sin(t0) + 6 sin(t).
b) Apply the definition of stability, i.e., show that you can pick δ(², t0) = c(t²0) for a
suitable c(t0). How would you choose c(t0)? (get a formula!)
c) Can such a constant c(t0) be chosen for any t0? Give a counterexample.
Observation:Stability and Uniform Stability coincide for Time invariant or periodic
systems. Why?
Exercise 3 - Check that (0, 0) is an asymptotically stable equilibrium point for the system:
½
ẋ1(t) = x1(t)(x21(t) + x22(t) − 1) − x2(t)
ẋ2(t) = x1(t) + x2(t)(x21(t) + x22(t) − 1)

Lyapunov Stability
Auxiliary Definitions
α is a Function of class K if α(·) is nondecreasing, α(0) = 0, and α(p) > 0, ∀p > 0.
V : <n × <+ → < is a Decrescent Function if ∃β(·) of class K s.t. V (t, x) ≤ β(kxk),
∀t ≥ 0, ∀x s.t. kxk ≤ r for some r.
V : <n × <+ → < is a Locally Positive Definite Function if V (t, 0) = 0, ∀t ≥ 0,
V (t, x) ≥ α(kxk), ∀t ≥ 0, ∀x s.t. kxk ≤ r for some r.
V is a Positive Definite Function if it is Locally Positive Definite Function with “r = ∞”
and α(r) ↑ ∞ as r ↑ ∞.
Observation: The time independent versions of the above definitions can be expressed
without the need of the K class functions, i.e.,
V̄ (0) = 0, V̄ (x) > 0 ∀x s.t. kxk ≤ r, ...
.
Fact A time dependent function is a l.p.d. (p.d.) iff it “dominates” a time independent
l.p.d.f. (p.d.f.).
Lyapunov Stability
Lyapunov’s Direct Method

This is a criterion to test the stability of equilibria whose Basic Idea is:
• Take 0 to be the identified equilibrium point

• Let V be an “energy function” which is 0 at 0 and positive everywhere else.
• The system is perturbed to a new nonzero initial point
• The system may be stable if the system’s dynamics are s.t. its energy level does not
increase with time.
• The system may is asymptotically stable if the system’s dynamics are s.t. its energy
level decreases monotonically with time.

Lyapunov Stability
Main Results The equilibrium point 0 at time t0 is (uniformly) stable (over [t0, ∞)) if ∃ a
C 1 (decrescent) l.p.d.f. V s.t.
V̇ (t, x) ≤ 0, ∀t ≥ 0, ∀x s.t. kxk ≤ r for some r.
Given ², let ²̄ := min{², r, s}, where s is such that V (t, x) ≥ α(kxk), ∀t > 0, ∀kxk ≤ s.
To check that δ > 0 s.t. β(t0, δ) := sup{V (t0, x) : kxk ≤ δ} < α(²̄) is as required in the
definition of stability, note that since V̇ (t, x) ≤ 0 whenever kxk < δ, we have
α(kx(t)k) ≤ V (t, x(t)) ≤ V (t0, x(t0)) ≤ α(²̄), and thus kx(t)k ≤ ².
1 2
R x1
Example Take V (x1, x2) = 2 x2 + 0 g(s)ds and apply Lyapunov theorem to the system:
½
ẋ1 = x2
ẋ2 = −f (x2) − g(x1)
where f (friction) and g (spring restoring force) are continuous and, ∀s ∈ [−s0, s0],
satisfy sf (s) ≥ 0 and sg(s) > 0 for s 6= 0.

Lyapunov Stability

Main Results (cont.)
The equilibrium point 0 at time t0 is uniformly asymptotically stable over [t0, ∞) if ∃ a
C 1 (decrescent) l.p.d.f. V s.t. −V̇ is an l.p.d.f..
Global asymptotic stability requires the p.d.f.V to satisfy V̇ ≤ −γ(kxk), ∀t ≥ t0, x ∈ <n
for some K function γ.
Examples
Check the asymptotic stability of 
 ẋ1 = x1(x21 + x22 − 1) − x2
a) V (x1, x2) = x21 + x22 and
 2 2
½ ẋ 2 = x1 + x2 (x1 + x 2 − 1)
ẋ1 = −x1 − e−2tx2
b) V (t, x1, x2) = x21 + (1 + e−2t)x22 and
ẋ2 = x1 − x2

Lyapunov Stability
M ⊂ <n is an invariant set for ẋ = f (t, x) if x(t0) ∈ M for some t0 > 0 implies
x(t) ∈ M ∀t ≥ t0.
A set S ⊂ <n is the positive limit set for a trajectory x(·) if ∀x ∈ S, x = limtn x(tn) for
some sequence {tn} s.t. tn → ∞.
Facts
a) For periodic or autonomous systems, the positive set of any trajectory is an invariant
set.
b) The positive limit set of a bounded trajectory is closed and bounded.
c) Let x(·) be bounded and S be the system’s positive limit set
limt↑∞ supy∈S kx(t) − yk = 0.
Another Fact
Consider ẋ = f (x) and V : <n → < to be C 1 s.t.: Sc := {x ∈ <n : V (x) ≤ 0} is
bounded; V is bounded from below on Sc; and V̇ (x) ≤ 0 on Sc.
Then, ∀x0 ∈ Sc, limt↑∞ x(t; x0, 0) ∈ M where M is the largest invariant subset in
{x ∈ Sc : V̇ (x) = 0}.
Lyapunov Stability

Exercise - understanding the last fact
a) Show that x(t; t0, 0) ∈ Sc ∀t > 0.
b) Let L be the limit set of x(t; t0, 0). Show V (y) = limt↑∞ V (t; t0, 0) ∀y ∈ L.
c) Why is L an invariant set? Why does the above fact hold?
Notes
¥ x0 ∈ Ωc and V̇ (x) ≤ 0∀x ∈ Ωc =⇒ V (x(t; x0, 0)) nonincreasing
¥ V (x(t; x0, 0)) nonincreasing and bounded from below =⇒ ∃c0 s.t. V (x(t; x0, 0)) → c0
¥ L is the limit set of x(t; x0, 0) =⇒ V (y) = c0 ∀y ∈ L
¥ L is invariant =⇒ [ V̇ (y) = 0 ∀y ∈ L ⇒ L ⊂ S ]
¥ L ⊂ M =⇒ the conclusion

Lyapunov Stability
LaSalle’s local Theorem
Let ẋ = f (x) and V : <n → < be C 1 l.p.d s.t. V̇ (x) ≤ 0 ∀ kxk ≤ r. Assume
S := {x ∈ <n : V (x) ≤ m, V̇ (x) = 0}, m := supkxk≤r {V (x)} has only the trajectory
x ≡ 0. Then, 0 is asymptotically stable.
LaSalle’s global Theorem
Let ẋ = f (x) and V be autonomous or of period T. V be C 1, p.d. and s.t. V̇ (x) ≤ 0
∀x ∈ <n. Also S := {x ∈ <n : V̇ = 0 ∀t ≥ 0} contains no trivial trajectories. Then, 0 is
globally asymptotically stable.
Observation:
The advantage of LaSalle’s theorems is that Asymptotic Stability is concluded only by
requiring V̇ (x) ≤ 0 and not −V̇ (x) ≤ α(kxk).
The price to pay? The system has to be autonomous or periodic in time.
1
Ry
Exercise - Apply LaSalle’s theorem to ÿ + f (ẏ) + g(y) = 0 with V (y, ẏ) = 2 0 g(s)ds.
f, g are continuous, f (0) = g(0) = 0, sf (s) > 0, sg(s) > 0, ∀s 6= 0.
Lyapunov Stability
Conditions for equilibria instability
“0 is unstable at t0” if ∃ a C 1 V : < × <n → <,
(i) s.t.: V is decrescent, V̇ is l.p.d., V (t, 0) = 0 and ∀² > 0, ∃x ∈ ²B, V (t0, x) > 0.
(ii) s.t.: V is decrescent, V (t, 0) = 0, ∀² > 0, ∃x ∈ ²B, V (t0, x) > 0 and
V̇ (t, x) = λV (t, x) + W (t, x) with λ > 0 and W (t, x)∀t ≥ t0 kxk ≤ r.
(iii) closed Ω and open Ω̄ ⊂ Ω s.t. 0 ∈ intΩ, 0 ∈ ∂ Ω̄, and,
∀t ≥ t0, V is bounded above in Ω, uniformly in t, V (t, x) = 0, on ∂ Ω̄,
∀x ∈ Ω̄, V (t, x) > 0, and V̇ (t, x) ≥ γ(kxk) for some γ of class K.
Issue: How to construct a Lyapunov candidate function?

a) Physical considerations: Specify the system’s total energy function.
b) Trial method: Guess stability or instability and try functions of increasing order of
complexity until the criterion is met.

Lyapunov Stability
Linear Systems
For ẋ(t) = A(t)x(t), t > 0, 0 is an isolated equilibrium.
Let {λi : i = 1, ..., n} be the eigenvalues of A and Φ denote the State Transition Matrix.
Necessary and Sufficient Conditions for Stability

Stable Asymptotically Stable
(Also Stable ∀t1 ≥ t0) (Also Globally Asympt. Stable )
supt≥t0 kΦ(t, t0)k := m(t0) < ∞ limt→∞ kΦ(t, t0)k = 0
Uniformly Stable Uniformly Asympt. Stable over [0, ∞)

supt0≥0 m(t0) < ∞ supt0≥0 m(t0) < ∞ and
limt→∞ kΦ(t, t0)k = 0 unif. in t0
∃m, λ > 0 s.t., ∀t ≥ t0 ∀t0,

kΦ(t, t0)k ≤ me−λ(t−t0)

Lyapunov Stability
Linear Systems
Consider the autonomous system ẋ(t) = Ax(t).
Classical result
a) Glob. Asympt. Stability iff Re{λi} < 0 and
b) Stability iff Re{λi} = 0 only if λi is a minimal polynomial simple zero.
Lyapunov equation : A0P + P A + Q = 0 (LE)

1 0
V (x) = x P x =⇒ V̇ (x) = x0(A0P + P A)x = −x0Qx
2
(P, Q) s.t.: P > 0, Q > 0 ⇔ V and V̇ p.d. ⇔ 0 is G. A. S.
Q > 0, ∃λi(P ) ⇔ −V̇ p.d. and ∃² ⇔ 0 is unstable
Re(λi(P )) < 0 kxk < ² V (x) > 0

Lyapunov Stability
Linear Systems
Given A, there are two approaches:
either pick P and study the resulting Q,
or pick Q and study the resulting P .
While the first requires an apriori guess on the stability (of 0), the second is a more
straightforward trial test.
However, there is a problem of nonuniqueness. Hence:
Theorem A
∀ Q, ∃1 P sol. to (LE) iff λi + λ∗j 6= 0 ∀ i, j.
Theorem B
Re(λi) < 0 ∀i iff ∃Q > 0 s.t. ∃1 P sol. to (LE), P > 0 iff ∀Q > 0, ∃1 P sol. to
(LE), P > 0.
Theorem C
Suppose λi + λ∗j 6= 0 and let (LE) be s.t. ∃1 solution P for each Q. If Q > 0, then P has
many negative eigenvalues as A has eigenvalues with positive real part.
Lyapunov Stability
Indirect Method
Key idea: derive (local) stability conclusions for nonlinear systems from results for linear
systems.
Take ẋ(t) = f (t, x(t)) where f is C 1 in x, f (t, 0) = 0 ∀t ≥ 0, and limkxk→0 kg(t,x)k
kxk = 0
where g(t, x) = f (t, x) − A(t)x with A(t) = ∂f ∂x (t, x)|x=0 .
Theorem Assume A(·) is bounded and limkxk→0 supt≥0 kg(t,x)kkxk = 0. Then, if 0 is uniformly
asymptotically stable (UAS) over [0, ∞) for the linearized system (LS), ż(t) = A(t)z(t),
so is for the nonlinear system ẋ(t) = f (t, x(t)).
Exercises (TakeRthe assumptions and definitions of the above theorem).
∞
1- Let P (t) := t Φ0(s, t)Φ(s, t)ds. Show that:
1.1 ∀t ≥ 0 P (t) > 0 and ∃b > a > 0 s.t. ax0x ≤ x0P (t)x ≤ bx0x.
1.2 Ṗ (t) + A0(t)P (t) + P (t)A(t) + I = 0.
2- Let V (t, x) := x0P (t)x. Show that V (t, x) is a decrescent p.d.f. with
V̇ (t, x) = −x0x + 2x0P (t)g(t, x).
3- Take r > 0 s.t. kxk ≤ r ⇒ kg(t, x)k ≤ kxk/(3b) ∀t ≥ 0. Show that
V̇ (t, x) ≤ −x0x/3.
Lyapunov Stability
Indirect Method
Theorem’
Take the data of the previous theorem and assume that A(t) = Ā ∀t ≥ 0. Then, if Ā has
at least an eigenvalue with positive real part, 0 is an unstable equilibrium point for the
nonlinear system.
Exercises Determine the stability of the origin for the following systems (including the
domain of attraction).
a) ÿ½ = (1 − y 2)ẏ − y with µ > 0
ẋ1 = x1 + x2 + x1x2
b)
ẋ2 = −x1 + x22

Lyapunov Stability
The Feedback Stabilization Problem
Take the control system ẋ(t) = f (x(t), u(t)) and specify a feedback control law
u(t) = g(x(t)) so that the reference equilibrium point of the closed loop system
ẋ(t) = f (x(t), g(x(t))) is asymptotically stable.
Assumptions
a) f is C 1 in <n × <m and f (0, 0) = 0.
∂f ∂f
b) rank[B|AB|...|An−1B] = n, with A = ∂x (x, u)|x=0,u=0 , B= ∂u (x, u)|x=0,u=0 .
Observation The assumptions imply that the linearized system around (0, 0),
ż(t) = Az(t) + Bv(t), is controllable.
Fact There is a matrix K s.t. all the eigenvalues of A − BK have negative real parts
and, thus 0 is G.A.S. for the closed loop system ż(t) = (A − BK)z(t).
Another Fact (how to compute such a K? Use LQ results) Given (A, B) as above,
K = Q−1B 0M where M is the solution to the Riccati equation,
−P − A0M − M A + M BQ−1B 0M = 0 for given P > 0 and Q > 0.
Lyapunov Stability
Feedback Stabilization
Theorem (Nonlinear stabilization)
Take ẋ(t) = f (x(t), u(t)) where f s.t. b) and let A and B defined as above.
Take K ∈ s.t all the eigenvalues of A − BK have negative real parts. Then,
u(t) = −Kx(t) =⇒ 0 is an asymptoptically stable equilibrium point of
ẋ(t) = f (x(t), −Kx(t)).
Approach
a) Linearize the nonlinear system
b) Compute K stabilizing the linear system
c) Feed the nonlinear system input with −Kx.
Exercise
Find a feedback stabilizing control for the system
½  2a + b if 2a + b ≤ 1
ẋ1 = 3x1 + x22 + g(x2, u)
where g(a, b) = 1 if 2a + b > 1 .
ẋ2 = sin(x1) − x2 + u 
−1 if 2a + b < −1

Input/Output Stability
The (dynamical) system regarded as an input-output transformation
Formalization requires:
Definition 1 Lp[0, ∞), (Lpe[0, ∞)),
R ∞ p = p1, ...∞ is the
R ∞set of all measurable
f (·)(fT (·)) : [0, ∞) → <) s.t. 0 |f (t)| dt < ∞ ( 0 |fT (t)|pdt < ∞ ∀T > 0).
Note: f might be vector valued.
Definition 2 A : Lnpe → Lm
pe is Lp -stable if
a) f ∈ Lnp implies that Af ∈ Lm p ; and
b) ∃k, c s.t. kAf kp ≤ kkf k + b ∀f ∈ Lp.
Note: Bounded input/Bounded output Stability - p = ∞
R t −α(t−τ )
Example 1 (Af )(t) = 0 e f (τ dτ
Find k and c.
Example 2 (Af )(t) = f 2(t)
Is it a Lp map? Can you find k and c fulfilling the definition of stability.

Input/Output Stability
 ½

 e1 = u1 − y2 y1 = G1(u1 − y2)
 (2)
e2 = y1 + u2 % ½ y2 = G2(u2 + y1)
(1)

 y1 = G1e1 & e1 = u1 + G2e2

y2 = G2e2 e2 = u2 + G1e1
Theorem Rt
Take (1) with (G1x)(t) := 0 G(t, τ )n1(τ, x(τ ))dτ and (G2x)(t) := n2(t, x(t)).
Here, G(·, ·) and ni : <+ × <n → <n, i = 1, 2, are continuous with ni(·, 0) = 0 and the
Ki-Lipschitz continuity of ni(t, ·).
Then, for i = 1, 2, Gi : Lnpe → Lm pe and
∀u1, u2 ∈ Lnpe, ∃1(e1, e2, y1, y2) ∈ Lnpe s.t. (1) holds.
Definition 3 - (2) is Lp-stable if,
∀u1, u2 ∈ Lnpe, y1, y2 s.t. (2) holds are in Lnp, and
∃k, b s.t., for i = 1, 2, kyikp ≤ k(ku1kp + ku2kp) + b, whenever u1, u2, y1, y2 are s.t. (2)
holds.
Input/Output Stability and Lyapunov Stability

Theorem
½
ẋ(t) = Ax(t) − f (t, x(t))
Consider the system (1) ,
x(0) = x0
where Re{λi(A)} < 0 and f (·, ·) is continuous, and define the corresponding nonlinear
feedback system:
½ Rt
e(t) = u(t) − 0 eA(t−τ )y(τ )dτ
(2)
y(t) = f (t, e(t))
Then, if (2) is L2-stable, then the equilibrium 0 of (1) is GAS (Globally Asymptotically
Stable).
Observation: Input/Output techniques yield either GAS or nothing!
Difficult to estimate regions of attraction.

State Feedback Linearization

Addressed class of systems
m
X
ẋ(t) = f (x(t)) + gi(x(t))ui(t), x(0) = x0 ∈ <n
i=1
yi(t) = hi(x(t)) i = 1, . . . , p
General objectives:
Central issue: synthesis of state feedback control linearizing the dynamic system.
Nonlinear change of phase of coordinates so that only one dynamic component depends
nonlinearly on the state variable and affinely on the control.
The geometric insight provided by this framework allows to address various problems:
stabilization, tracking, regulation . . . .
Required background topics:
Differential calculus on vector fields,
Distributions,
Nonlinear change of coordinates,
Assorted dynamic systems results . . . .

Notation
m
X
ẋ = f (x) + gi(x)ui, x(0) = x0 ∈ <n
i=1
yi = hi(x) i = 1, . . . , p
Exercise: Linear systems as a special case.
f and gi : <n → <n i = 1, . . . , p are smooth maps on an open set U (in particular
U = <n) are referred to as Vector fields.
Geometrically, they define tangent directions to trajectories at x.
Covector field, the dual of a vector field, maps each x ∈ U with a point in the dual of the
image space of the vector field.
Covectors can be regarded as linear functionals on spaces of vector fields.
Example
dλ(x) = ∂λ n
∂x (x) is the differential/exact differential or gradient of λ : U → <, U ⊂ < .


Notation
Pn ∂λ
A) Derivative of λ along f - Lf λ(x):= dλ(x) · f (x) = i=1 ∂xi (x)fi (x).
∂(Lf λ(x))
Exercise: Check that Lg Lf λ(x) = ∂x g(x) and Lkf λ(x) = Lf (Lfk−1λ(x)).
B) Lie product or bracket of f and g - [f, g](x) := Lf g(x) − Lg f (x).
Note that adkf g(x) = [f, adk−1 0
f g](x), where adf g(x) = g(x).
Exercise: Show that [·, ·] is bilinear, skew commutative, and
[f, [g, h]] + [g, [h, f ]] + [h, [f, g]] = 0 (Jacobi identity).
0 ∂w0 0
C) Derivative of a covector w along f - Lf w(x) := + w ∂ff ( ∂x )
∂x .
Exercise: Check the following Properties
a) Lαf λ(x) = (Lf λ(x))α(x).
b) [αf, βg](x) = α(x)β(x)[f, g](x) + (Lf β(x))α(x)g(x) − (Lg α(x))β(x)f (x).
c) L[f,g]λ(x) = Lf Lg λ(x) − Lg Lf λ(x).
d) Lαf βw(x) = α(x)β(x)Lf w(x) + β(x)w(x) · f (x)dα(x) + (Lf β(x))α(x)β(x)
e) Lf dλ(x) = dLf λ(x).
f) Lf (w · g)(x) = (Lf w(x)) · g(x) + w(x) · [f, g](x).
State Space Change of Coordinates

Why?
highlights properties (observ., reachab.) or problem solution (decoupl., stabilaz.).
Example for linear systems:
Ā = T AT −1
, B̄ = T B and, C̄ = CT −1
where matrix T represents a linear change of phase coordinates.
A map z = Φ(x) is a nonlinear change of phase coordinates if it is a (global) (local)
diffeomorphism, i.e., invertible and, both Φ and Φ−1 are smooth.
Exercise:
a) det( ∂Φ
∂x ) 6= 0 at x0 implies that Φ is a local diffeomorphism.
b) Check that f¯ = ( ∂Φ −1 ∂Φ −1 −1
∂x f ) ◦ Φ (z), ḡ = ( ∂x g) ◦ Φ (z) and h̄ = h ◦ Φ (z).


Distributions
The map x → ∆(x) := span{fi(x) : i = 1, . . . , d}, where fi : U → <n is a (smooth)
vector field, is a (smooth) distribution.
Let F = col(f1, · · · , fd). Then, ∆(x) = Im{F (x)} and, dim(∆) = rank(F ).
x0 is a regular point of ∆ if ∃ open U , x0 ∈ U where ∆ is nonsingular, i.e., F is full rank.
Observations:
Distributions inherit pointwisely vector spaces operations and properties.
Pd
For any smooth vector field τ ∈ ∆, ∃ci(x) smooth s.t. τ (x) = i=1 ci(x)fi(x).
∆ is involutive if ∀τ, σ ∈ ∆, [τ, σ] ∈ ∆.
Criteria: rank(F ) = rank(col(F |[fi, fj ])), ∀i 6= j, ∀x(∈ U ).
Exercise: Show that the intersection of two involutive distributions is also involutive.
What about the sum?
Let ∆ be noninvolutive. The involutive closure of ∆ is the smallest involutive distribution
containing it.

Distributions
Codistribution, Ω, is the dual of distribution, ∆, in the sense that is spanned by the
associated covectors.
Example: The annihilator of ∆, ∆⊥(x) := {w∗ ∈ (<n)∗ : w∗ · v = 0 ∀v ∈ ∆}.
Properties (Let ∆ (Ω) is spanned by the rows of F (W )):
⊥
a) dim(∆)
T + dim(∆ ) = n ∀x.
b) [∆1 ∆2]⊥ = ∆⊥ ⊥
1 + ∆2 .
c) ∆⊥ = {w∗ : w∗F (x) = 0}.
d) W (x)v = 0 ∀v ∈ ∆ and Ω⊥(x) = Ker(W (x)).
Frobenius Theorem: A nonsingular distribution is completely integrable iff it is involutive.
∆, dim(∆) = d, is completely integrable at x ∈ U if ∃ λi : U → <, i = 1, . . . , n − d, s.t.
span{dλ1, . . . , dλn−d} = ∆⊥.
∂λ
In other words, the p.d.e. ∂x F (x) = 0 has n − d independent solutions.

Distributions
∂λi
Application of Frobenius theorem: Easy way to solve the p.d.e. ∂x F (x) = 0,
i = 1, . . . , n − d.
Procedure:
(1) Complete ∆ with additional n − d independent vector fields.
(2) Solve o.d.e.s ẋ = fi(x) with x(0) = x0, yielding xi(t) = Φft i (x0), t ∈ [0, zi], in
U² = {z ∈ <n : zi < ²}.
Qn
(3) Take Ψ : U² → < , Ψ(z) := i=1 Φfzii (x0) = Φfz11 ◦ · · · ◦ Φfznn (x0).
n
(4) The last n − d rows of Ψ−1 are a sol. to the p.d.e.

−1
To verify this, note that ( ∂Ψ∂x |x=Φ(z))( ∂Ψ
∂x ) = I and since the first d columns of
∂Ψ
∂x form a
−1
basis for ∆, the last n − d rows of ∂Ψ∂x |x=Φ(z) annihilate ∆.


Invariant Distributions - ∆ is invariant under the vector field f if [f, ∆] ⊂ ∆.
Extension of invariant sub-spaces for linear systems: Take fA(x) = Ax,
∆ = span{τ1(x), . . . , τd(x)} and τi(x) = vi. Then [fA, τi] = −Avi.
Advantages: For completely integrable distributions, it allows a simplified representation
of the dynamic system.
Proposition Let ∆ be a nonsingular, involutive smooth distribution of dimension d
invariant under f . Then,
(i) ∀x0 ∃ open U0, x0 ∈ U0 and a transformation z = Φ(x) on U0 for which f is
represented by f¯(z) = col(f¯1(z 1, z 2), f¯2(z 2)).
(ii) Furthermore, the codistribution ∆⊥ is also invariant under f . The converse is also
true.
Exercise Consider ∆ = span{v1, v2}, v1 = col(1, 0, 0, x2), v2 = col(0, 1, 0, x1), and
f = col(x2, x3, (x4 − x1x2)x3, sin(x3) + x22 + x1x3).
a) Check that ∆ is involutive and invariant under f .
b) Apply Frobenius theorem, i.e., find λ1, . . . λk s.t. span{dλ1, . . . , dλk ) = ∆⊥.
c) Form the change of coordinates mapping z = Φ(x) and compute f¯(z).

Nonlinear Control - An Overview: Fernando Lobo Pereira, Flp@fe - Up.pt

Uploaded by

Copyright:

Available Formats

Nonlinear Control - An Overview: Fernando Lobo Pereira, Flp@fe - Up.pt

Uploaded by

Document Information

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

Nonlinear Control - An Overview: Fernando Lobo Pereira, Flp@fe - Up.pt

Uploaded by

Copyright:

Available Formats

1

NONLINEAR CONTROL - AN OVERVIEW

Fernando Lobo Pereira, [email protected]

Porto University, Portugal

Fernando Lobo Pereira, [email protected] 1

Most of real systems involve nonlinearities in one way or another.

Fernando Lobo Pereira, [email protected] 2

Dynamics (Newton’s law for rotating objects):

mθ̈(t) + mgsin(θ(t)) = u(t)

Fernando Lobo Pereira, [email protected] 3

Consider the local control around the later.

Step 1 - Analysis - Linearization

ϕ̈(t) − ϕ(t) = u(t)

Fernando Lobo Pereira, [email protected] 4

Conclusion: “proportional control” does not suffice.

Linear designs apply to linearized nonlinear systems operating locally.

Extension: controller scheduling

Fernando Lobo Pereira, [email protected] 6

1 - Study the system of interest and decide on sensors and actuators

Fernando Lobo Pereira, [email protected] 7

Strategy of the course:

• Example-oriented introduction to concepts.

• Matlab hands-on exercises

Fernando Lobo Pereira, [email protected] 8

Structure of the course:

• Selected topics of linear systems

Fernando Lobo Pereira, [email protected] 9

Selected topics of linear systems

State-space representation of Linear Dynamic Systems

System’s state trajectory given by a closed form solution

Fernando Lobo Pereira, [email protected] 10

Selected topics of linear systems

Fernando Lobo Pereira, [email protected] 11

Selected topics of linear systems

Fernando Lobo Pereira, [email protected] 12

Selected topics of linear systems

Fernando Lobo Pereira, [email protected] 13

Selected topics of linear systems

ψ(s) = Πσi=1(λ − λi)mi and p(s) = Πσi=1(λ − λi)di .

Selected topics of linear systems

Fernando Lobo Pereira, [email protected] 15

Selected topics of linear systems

Fernando Lobo Pereira, [email protected] 16

Selected topics of linear systems

Trajectory graphical analysis

b) Unstable node λ1 > λ2 > 0

c) Saddle point λ1 > 0 > λ2

d) Unstable focus Re(λ1) > 0

e) Stable focus Re(λ1) < 0

Fernando Lobo Pereira, [email protected] 17

Selected topics of linear systems

Realization theory: controllability

Fernando Lobo Pereira, [email protected] 18

Selected topics of linear systems

An invariant system is completely observable if and only if rank of R is n, where

Selected topics of linear systems

Fernando Lobo Pereira, [email protected] 20

Issues and specific background topics for nonlinear systems

ẋ(t) = f (t, x(t), u(t)), x(0) = x0